1 /*
2 * Copyright (c) 2005, 2019, 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 "libadt/vectset.hpp"
29 #include "memory/universe.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/node.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/phaseX.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subnode.hpp"
50 #include "opto/subtypenode.hpp"
51 #include "opto/type.hpp"
52 #include "runtime/sharedRuntime.hpp"
53 #include "utilities/macros.hpp"
54 #include "utilities/powerOfTwo.hpp"
55 #if INCLUDE_G1GC
56 #include "gc/g1/g1ThreadLocalData.hpp"
57 #endif // INCLUDE_G1GC
58 #if INCLUDE_SHENANDOAHGC
59 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
60 #endif
61
62
63 //
64 // Replace any references to "oldref" in inputs to "use" with "newref".
65 // Returns the number of replacements made.
66 //
67 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
68 int nreplacements = 0;
69 uint req = use->req();
70 for (uint j = 0; j < use->len(); j++) {
71 Node *uin = use->in(j);
72 if (uin == oldref) {
73 if (j < req)
74 use->set_req(j, newref);
75 else
76 use->set_prec(j, newref);
77 nreplacements++;
78 } else if (j >= req && uin == NULL) {
79 break;
80 }
81 }
82 return nreplacements;
83 }
84
85 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
86 assert(old != NULL, "sanity");
87 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
88 Node* use = old->fast_out(i);
89 _igvn.rehash_node_delayed(use);
90 imax -= replace_input(use, old, target);
91 // back up iterator
92 --i;
93 }
94 assert(old->outcnt() == 0, "all uses must be deleted");
95 }
96
97 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
98 // Copy debug information and adjust JVMState information
99 uint old_dbg_start = oldcall->tf()->domain()->cnt();
100 uint new_dbg_start = newcall->tf()->domain()->cnt();
101 int jvms_adj = new_dbg_start - old_dbg_start;
102 assert (new_dbg_start == newcall->req(), "argument count mismatch");
103
104 // SafePointScalarObject node could be referenced several times in debug info.
105 // Use Dict to record cloned nodes.
106 Dict* sosn_map = new Dict(cmpkey,hashkey);
107 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
108 Node* old_in = oldcall->in(i);
109 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
110 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
111 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
112 uint old_unique = C->unique();
113 Node* new_in = old_sosn->clone(sosn_map);
114 if (old_unique != C->unique()) { // New node?
115 new_in->set_req(0, C->root()); // reset control edge
116 new_in = transform_later(new_in); // Register new node.
117 }
118 old_in = new_in;
119 }
120 newcall->add_req(old_in);
121 }
122
123 // JVMS may be shared so clone it before we modify it
124 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
125 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
126 jvms->set_map(newcall);
127 jvms->set_locoff(jvms->locoff()+jvms_adj);
128 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
129 jvms->set_monoff(jvms->monoff()+jvms_adj);
130 jvms->set_scloff(jvms->scloff()+jvms_adj);
131 jvms->set_endoff(jvms->endoff()+jvms_adj);
132 }
133 }
134
135 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
136 Node* cmp;
137 if (mask != 0) {
138 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
139 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
140 } else {
141 cmp = word;
142 }
143 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
144 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
145 transform_later(iff);
146
147 // Fast path taken.
148 Node *fast_taken = transform_later(new IfFalseNode(iff));
149
150 // Fast path not-taken, i.e. slow path
151 Node *slow_taken = transform_later(new IfTrueNode(iff));
152
153 if (return_fast_path) {
154 region->init_req(edge, slow_taken); // Capture slow-control
155 return fast_taken;
156 } else {
157 region->init_req(edge, fast_taken); // Capture fast-control
158 return slow_taken;
159 }
160 }
161
162 //--------------------copy_predefined_input_for_runtime_call--------------------
163 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
164 // Set fixed predefined input arguments
165 call->init_req( TypeFunc::Control, ctrl );
166 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
167 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
168 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
169 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
170 }
171
172 //------------------------------make_slow_call---------------------------------
173 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
174 address slow_call, const char* leaf_name, Node* slow_path,
175 Node* parm0, Node* parm1, Node* parm2) {
176
177 // Slow-path call
178 CallNode *call = leaf_name
179 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
180 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
181
182 // Slow path call has no side-effects, uses few values
183 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
184 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
185 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
186 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
187 copy_call_debug_info(oldcall, call);
188 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
189 _igvn.replace_node(oldcall, call);
190 transform_later(call);
191
192 return call;
193 }
194
195 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
196 _fallthroughproj = NULL;
197 _fallthroughcatchproj = NULL;
198 _ioproj_fallthrough = NULL;
199 _ioproj_catchall = NULL;
200 _catchallcatchproj = NULL;
201 _memproj_fallthrough = NULL;
202 _memproj_catchall = NULL;
203 _resproj = NULL;
204 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
205 ProjNode *pn = call->fast_out(i)->as_Proj();
206 switch (pn->_con) {
207 case TypeFunc::Control:
208 {
209 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
210 _fallthroughproj = pn;
211 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
212 const Node *cn = pn->fast_out(j);
213 if (cn->is_Catch()) {
214 ProjNode *cpn = NULL;
215 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
216 cpn = cn->fast_out(k)->as_Proj();
217 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
218 if (cpn->_con == CatchProjNode::fall_through_index)
219 _fallthroughcatchproj = cpn;
220 else {
221 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
222 _catchallcatchproj = cpn;
223 }
224 }
225 }
226 break;
227 }
228 case TypeFunc::I_O:
229 if (pn->_is_io_use)
230 _ioproj_catchall = pn;
231 else
232 _ioproj_fallthrough = pn;
233 break;
234 case TypeFunc::Memory:
235 if (pn->_is_io_use)
236 _memproj_catchall = pn;
237 else
238 _memproj_fallthrough = pn;
239 break;
240 case TypeFunc::Parms:
241 _resproj = pn;
242 break;
243 default:
244 assert(false, "unexpected projection from allocation node.");
245 }
246 }
247
248 }
249
250 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
251 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
252 bs->eliminate_gc_barrier(this, p2x);
253 }
254
255 // Search for a memory operation for the specified memory slice.
256 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
257 Node *orig_mem = mem;
258 Node *alloc_mem = alloc->in(TypeFunc::Memory);
259 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
260 while (true) {
261 if (mem == alloc_mem || mem == start_mem ) {
262 return mem; // hit one of our sentinels
263 } else if (mem->is_MergeMem()) {
264 mem = mem->as_MergeMem()->memory_at(alias_idx);
265 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
266 Node *in = mem->in(0);
267 // we can safely skip over safepoints, calls, locks and membars because we
268 // already know that the object is safe to eliminate.
269 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
270 return in;
271 } else if (in->is_Call()) {
272 CallNode *call = in->as_Call();
273 if (call->may_modify(tinst, phase)) {
274 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
275 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
276 return in;
277 }
278 }
279 mem = in->in(TypeFunc::Memory);
280 } else if (in->is_MemBar()) {
281 ArrayCopyNode* ac = NULL;
282 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
283 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
284 return ac;
285 }
286 mem = in->in(TypeFunc::Memory);
287 } else {
288 assert(false, "unexpected projection");
289 }
290 } else if (mem->is_Store()) {
291 const TypePtr* atype = mem->as_Store()->adr_type();
292 int adr_idx = phase->C->get_alias_index(atype);
293 if (adr_idx == alias_idx) {
294 assert(atype->isa_oopptr(), "address type must be oopptr");
295 int adr_offset = atype->offset();
296 uint adr_iid = atype->is_oopptr()->instance_id();
297 // Array elements references have the same alias_idx
298 // but different offset and different instance_id.
299 if (adr_offset == offset && adr_iid == alloc->_idx)
300 return mem;
301 } else {
302 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
303 }
304 mem = mem->in(MemNode::Memory);
305 } else if (mem->is_ClearArray()) {
306 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
307 // Can not bypass initialization of the instance
308 // we are looking.
309 debug_only(intptr_t offset;)
310 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
311 InitializeNode* init = alloc->as_Allocate()->initialization();
312 // We are looking for stored value, return Initialize node
313 // or memory edge from Allocate node.
314 if (init != NULL)
315 return init;
316 else
317 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
318 }
319 // Otherwise skip it (the call updated 'mem' value).
320 } else if (mem->Opcode() == Op_SCMemProj) {
321 mem = mem->in(0);
322 Node* adr = NULL;
323 if (mem->is_LoadStore()) {
324 adr = mem->in(MemNode::Address);
325 } else {
326 assert(mem->Opcode() == Op_EncodeISOArray ||
327 mem->Opcode() == Op_StrCompressedCopy, "sanity");
328 adr = mem->in(3); // Destination array
329 }
330 const TypePtr* atype = adr->bottom_type()->is_ptr();
331 int adr_idx = phase->C->get_alias_index(atype);
332 if (adr_idx == alias_idx) {
333 DEBUG_ONLY(mem->dump();)
334 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
335 return NULL;
336 }
337 mem = mem->in(MemNode::Memory);
338 } else if (mem->Opcode() == Op_StrInflatedCopy) {
339 Node* adr = mem->in(3); // Destination array
340 const TypePtr* atype = adr->bottom_type()->is_ptr();
341 int adr_idx = phase->C->get_alias_index(atype);
342 if (adr_idx == alias_idx) {
343 DEBUG_ONLY(mem->dump();)
344 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
345 return NULL;
346 }
347 mem = mem->in(MemNode::Memory);
348 } else {
349 return mem;
350 }
351 assert(mem != orig_mem, "dead memory loop");
352 }
353 }
354
355 // Generate loads from source of the arraycopy for fields of
356 // destination needed at a deoptimization point
357 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
358 BasicType bt = ft;
359 const Type *type = ftype;
360 if (ft == T_NARROWOOP) {
361 bt = T_OBJECT;
362 type = ftype->make_oopptr();
363 }
364 Node* res = NULL;
365 if (ac->is_clonebasic()) {
366 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
367 Node* base = ac->in(ArrayCopyNode::Src);
368 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
369 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
370 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
371 } else {
372 if (ac->modifies(offset, offset, &_igvn, true)) {
373 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
374 uint shift = exact_log2(type2aelembytes(bt));
375 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
376 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
377 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
378 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
379
380 Node* adr = NULL;
381 const TypePtr* adr_type = NULL;
382 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
383 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
384 Node* base = ac->in(ArrayCopyNode::Src);
385 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
386 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
387 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
388 // Don't emit a new load from src if src == dst but try to get the value from memory instead
389 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
390 }
391 } else {
392 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
393 #ifdef _LP64
394 diff = _igvn.transform(new ConvI2LNode(diff));
395 #endif
396 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
397
398 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
399 Node* base = ac->in(ArrayCopyNode::Src);
400 adr = _igvn.transform(new AddPNode(base, base, off));
401 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
402 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
403 // Non constant offset in the array: we can't statically
404 // determine the value
405 return NULL;
406 }
407 }
408 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
409 }
410 }
411 if (res != NULL) {
412 res = _igvn.transform(res);
413 if (ftype->isa_narrowoop()) {
414 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
415 res = _igvn.transform(new EncodePNode(res, ftype));
416 }
417 return res;
418 }
419 return NULL;
420 }
421
422 //
423 // Given a Memory Phi, compute a value Phi containing the values from stores
424 // on the input paths.
425 // Note: this function is recursive, its depth is limited by the "level" argument
426 // Returns the computed Phi, or NULL if it cannot compute it.
427 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) {
428 assert(mem->is_Phi(), "sanity");
429 int alias_idx = C->get_alias_index(adr_t);
430 int offset = adr_t->offset();
431 int instance_id = adr_t->instance_id();
432
433 // Check if an appropriate value phi already exists.
434 Node* region = mem->in(0);
435 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
436 Node* phi = region->fast_out(k);
437 if (phi->is_Phi() && phi != mem &&
438 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
439 return phi;
440 }
441 }
442 // Check if an appropriate new value phi already exists.
443 Node* new_phi = value_phis->find(mem->_idx);
444 if (new_phi != NULL)
445 return new_phi;
446
447 if (level <= 0) {
448 return NULL; // Give up: phi tree too deep
449 }
450 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
451 Node *alloc_mem = alloc->in(TypeFunc::Memory);
452
453 uint length = mem->req();
454 GrowableArray <Node *> values(length, length, NULL);
455
456 // create a new Phi for the value
457 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
458 transform_later(phi);
459 value_phis->push(phi, mem->_idx);
460
461 for (uint j = 1; j < length; j++) {
462 Node *in = mem->in(j);
463 if (in == NULL || in->is_top()) {
464 values.at_put(j, in);
465 } else {
466 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
467 if (val == start_mem || val == alloc_mem) {
468 // hit a sentinel, return appropriate 0 value
469 values.at_put(j, _igvn.zerocon(ft));
470 continue;
471 }
472 if (val->is_Initialize()) {
473 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
474 }
475 if (val == NULL) {
476 return NULL; // can't find a value on this path
477 }
478 if (val == mem) {
479 values.at_put(j, mem);
480 } else if (val->is_Store()) {
481 Node* n = val->in(MemNode::ValueIn);
482 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
483 n = bs->step_over_gc_barrier(n);
484 values.at_put(j, n);
485 } else if(val->is_Proj() && val->in(0) == alloc) {
486 values.at_put(j, _igvn.zerocon(ft));
487 } else if (val->is_Phi()) {
488 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
489 if (val == NULL) {
490 return NULL;
491 }
492 values.at_put(j, val);
493 } else if (val->Opcode() == Op_SCMemProj) {
494 assert(val->in(0)->is_LoadStore() ||
495 val->in(0)->Opcode() == Op_EncodeISOArray ||
496 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
497 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
498 return NULL;
499 } else if (val->is_ArrayCopy()) {
500 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
501 if (res == NULL) {
502 return NULL;
503 }
504 values.at_put(j, res);
505 } else {
506 #ifdef ASSERT
507 val->dump();
508 assert(false, "unknown node on this path");
509 #endif
510 return NULL; // unknown node on this path
511 }
512 }
513 }
514 // Set Phi's inputs
515 for (uint j = 1; j < length; j++) {
516 if (values.at(j) == mem) {
517 phi->init_req(j, phi);
518 } else {
519 phi->init_req(j, values.at(j));
520 }
521 }
522 return phi;
523 }
524
525 // Search the last value stored into the object's field.
526 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
527 assert(adr_t->is_known_instance_field(), "instance required");
528 int instance_id = adr_t->instance_id();
529 assert((uint)instance_id == alloc->_idx, "wrong allocation");
530
531 int alias_idx = C->get_alias_index(adr_t);
532 int offset = adr_t->offset();
533 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
534 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
535 Node *alloc_mem = alloc->in(TypeFunc::Memory);
536 Arena *a = Thread::current()->resource_area();
537 VectorSet visited(a);
538
539 bool done = sfpt_mem == alloc_mem;
540 Node *mem = sfpt_mem;
541 while (!done) {
542 if (visited.test_set(mem->_idx)) {
543 return NULL; // found a loop, give up
544 }
545 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
546 if (mem == start_mem || mem == alloc_mem) {
547 done = true; // hit a sentinel, return appropriate 0 value
548 } else if (mem->is_Initialize()) {
549 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
550 if (mem == NULL) {
551 done = true; // Something go wrong.
552 } else if (mem->is_Store()) {
553 const TypePtr* atype = mem->as_Store()->adr_type();
554 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
555 done = true;
556 }
557 } else if (mem->is_Store()) {
558 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
559 assert(atype != NULL, "address type must be oopptr");
560 assert(C->get_alias_index(atype) == alias_idx &&
561 atype->is_known_instance_field() && atype->offset() == offset &&
562 atype->instance_id() == instance_id, "store is correct memory slice");
563 done = true;
564 } else if (mem->is_Phi()) {
565 // try to find a phi's unique input
566 Node *unique_input = NULL;
567 Node *top = C->top();
568 for (uint i = 1; i < mem->req(); i++) {
569 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
570 if (n == NULL || n == top || n == mem) {
571 continue;
572 } else if (unique_input == NULL) {
573 unique_input = n;
574 } else if (unique_input != n) {
575 unique_input = top;
576 break;
577 }
578 }
579 if (unique_input != NULL && unique_input != top) {
580 mem = unique_input;
581 } else {
582 done = true;
583 }
584 } else if (mem->is_ArrayCopy()) {
585 done = true;
586 } else {
587 assert(false, "unexpected node");
588 }
589 }
590 if (mem != NULL) {
591 if (mem == start_mem || mem == alloc_mem) {
592 // hit a sentinel, return appropriate 0 value
593 return _igvn.zerocon(ft);
594 } else if (mem->is_Store()) {
595 Node* n = mem->in(MemNode::ValueIn);
596 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
597 n = bs->step_over_gc_barrier(n);
598 return n;
599 } else if (mem->is_Phi()) {
600 // attempt to produce a Phi reflecting the values on the input paths of the Phi
601 Node_Stack value_phis(a, 8);
602 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
603 if (phi != NULL) {
604 return phi;
605 } else {
606 // Kill all new Phis
607 while(value_phis.is_nonempty()) {
608 Node* n = value_phis.node();
609 _igvn.replace_node(n, C->top());
610 value_phis.pop();
611 }
612 }
613 } else if (mem->is_ArrayCopy()) {
614 Node* ctl = mem->in(0);
615 Node* m = mem->in(TypeFunc::Memory);
616 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
617 // pin the loads in the uncommon trap path
618 ctl = sfpt_ctl;
619 m = sfpt_mem;
620 }
621 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
622 }
623 }
624 // Something go wrong.
625 return NULL;
626 }
627
628 // Check the possibility of scalar replacement.
629 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
630 // Scan the uses of the allocation to check for anything that would
631 // prevent us from eliminating it.
632 NOT_PRODUCT( const char* fail_eliminate = NULL; )
633 DEBUG_ONLY( Node* disq_node = NULL; )
634 bool can_eliminate = true;
635
636 Node* res = alloc->result_cast();
637 const TypeOopPtr* res_type = NULL;
638 if (res == NULL) {
639 // All users were eliminated.
640 } else if (!res->is_CheckCastPP()) {
641 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
642 can_eliminate = false;
643 } else {
644 res_type = _igvn.type(res)->isa_oopptr();
645 if (res_type == NULL) {
646 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
647 can_eliminate = false;
648 } else if (res_type->isa_aryptr()) {
649 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
650 if (length < 0) {
651 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
652 can_eliminate = false;
653 }
654 }
655 }
656
657 if (can_eliminate && res != NULL) {
658 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
659 j < jmax && can_eliminate; j++) {
660 Node* use = res->fast_out(j);
661
662 if (use->is_AddP()) {
663 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
664 int offset = addp_type->offset();
665
666 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
667 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
668 can_eliminate = false;
669 break;
670 }
671 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
672 k < kmax && can_eliminate; k++) {
673 Node* n = use->fast_out(k);
674 if (!n->is_Store() && n->Opcode() != Op_CastP2X
675 SHENANDOAHGC_ONLY(&& (!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n))) ) {
676 DEBUG_ONLY(disq_node = n;)
677 if (n->is_Load() || n->is_LoadStore()) {
678 NOT_PRODUCT(fail_eliminate = "Field load";)
679 } else {
680 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
681 }
682 can_eliminate = false;
683 }
684 }
685 } else if (use->is_ArrayCopy() &&
686 (use->as_ArrayCopy()->is_clonebasic() ||
687 use->as_ArrayCopy()->is_arraycopy_validated() ||
688 use->as_ArrayCopy()->is_copyof_validated() ||
689 use->as_ArrayCopy()->is_copyofrange_validated()) &&
690 use->in(ArrayCopyNode::Dest) == res) {
691 // ok to eliminate
692 } else if (use->is_SafePoint()) {
693 SafePointNode* sfpt = use->as_SafePoint();
694 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
695 // Object is passed as argument.
696 DEBUG_ONLY(disq_node = use;)
697 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
698 can_eliminate = false;
699 }
700 Node* sfptMem = sfpt->memory();
701 if (sfptMem == NULL || sfptMem->is_top()) {
702 DEBUG_ONLY(disq_node = use;)
703 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
704 can_eliminate = false;
705 } else {
706 safepoints.append_if_missing(sfpt);
707 }
708 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
709 if (use->is_Phi()) {
710 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
711 NOT_PRODUCT(fail_eliminate = "Object is return value";)
712 } else {
713 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
714 }
715 DEBUG_ONLY(disq_node = use;)
716 } else {
717 if (use->Opcode() == Op_Return) {
718 NOT_PRODUCT(fail_eliminate = "Object is return value";)
719 }else {
720 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
721 }
722 DEBUG_ONLY(disq_node = use;)
723 }
724 can_eliminate = false;
725 }
726 }
727 }
728
729 #ifndef PRODUCT
730 if (PrintEliminateAllocations) {
731 if (can_eliminate) {
732 tty->print("Scalar ");
733 if (res == NULL)
734 alloc->dump();
735 else
736 res->dump();
737 } else if (alloc->_is_scalar_replaceable) {
738 tty->print("NotScalar (%s)", fail_eliminate);
739 if (res == NULL)
740 alloc->dump();
741 else
742 res->dump();
743 #ifdef ASSERT
744 if (disq_node != NULL) {
745 tty->print(" >>>> ");
746 disq_node->dump();
747 }
748 #endif /*ASSERT*/
749 }
750 }
751 #endif
752 return can_eliminate;
753 }
754
755 // Do scalar replacement.
756 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
757 GrowableArray <SafePointNode *> safepoints_done;
758
759 ciKlass* klass = NULL;
760 ciInstanceKlass* iklass = NULL;
761 int nfields = 0;
762 int array_base = 0;
763 int element_size = 0;
764 BasicType basic_elem_type = T_ILLEGAL;
765 ciType* elem_type = NULL;
766
767 Node* res = alloc->result_cast();
768 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
769 const TypeOopPtr* res_type = NULL;
770 if (res != NULL) { // Could be NULL when there are no users
771 res_type = _igvn.type(res)->isa_oopptr();
772 }
773
774 if (res != NULL) {
775 klass = res_type->klass();
776 if (res_type->isa_instptr()) {
777 // find the fields of the class which will be needed for safepoint debug information
778 assert(klass->is_instance_klass(), "must be an instance klass.");
779 iklass = klass->as_instance_klass();
780 nfields = iklass->nof_nonstatic_fields();
781 } else {
782 // find the array's elements which will be needed for safepoint debug information
783 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
784 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
785 elem_type = klass->as_array_klass()->element_type();
786 basic_elem_type = elem_type->basic_type();
787 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
788 element_size = type2aelembytes(basic_elem_type);
789 }
790 }
791 //
792 // Process the safepoint uses
793 //
794 while (safepoints.length() > 0) {
795 SafePointNode* sfpt = safepoints.pop();
796 Node* mem = sfpt->memory();
797 Node* ctl = sfpt->control();
798 assert(sfpt->jvms() != NULL, "missed JVMS");
799 // Fields of scalar objs are referenced only at the end
800 // of regular debuginfo at the last (youngest) JVMS.
801 // Record relative start index.
802 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
803 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
804 #ifdef ASSERT
805 alloc,
806 #endif
807 first_ind, nfields);
808 sobj->init_req(0, C->root());
809 transform_later(sobj);
810
811 // Scan object's fields adding an input to the safepoint for each field.
812 for (int j = 0; j < nfields; j++) {
813 intptr_t offset;
814 ciField* field = NULL;
815 if (iklass != NULL) {
816 field = iklass->nonstatic_field_at(j);
817 offset = field->offset();
818 elem_type = field->type();
819 basic_elem_type = field->layout_type();
820 } else {
821 offset = array_base + j * (intptr_t)element_size;
822 }
823
824 const Type *field_type;
825 // The next code is taken from Parse::do_get_xxx().
826 if (is_reference_type(basic_elem_type)) {
827 if (!elem_type->is_loaded()) {
828 field_type = TypeInstPtr::BOTTOM;
829 } else if (field != NULL && field->is_static_constant()) {
830 // This can happen if the constant oop is non-perm.
831 ciObject* con = field->constant_value().as_object();
832 // Do not "join" in the previous type; it doesn't add value,
833 // and may yield a vacuous result if the field is of interface type.
834 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
835 assert(field_type != NULL, "field singleton type must be consistent");
836 } else {
837 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
838 }
839 if (UseCompressedOops) {
840 field_type = field_type->make_narrowoop();
841 basic_elem_type = T_NARROWOOP;
842 }
843 } else {
844 field_type = Type::get_const_basic_type(basic_elem_type);
845 }
846
847 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
848
849 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
850 if (field_val == NULL) {
851 // We weren't able to find a value for this field,
852 // give up on eliminating this allocation.
853
854 // Remove any extra entries we added to the safepoint.
855 uint last = sfpt->req() - 1;
856 for (int k = 0; k < j; k++) {
857 sfpt->del_req(last--);
858 }
859 _igvn._worklist.push(sfpt);
860 // rollback processed safepoints
861 while (safepoints_done.length() > 0) {
862 SafePointNode* sfpt_done = safepoints_done.pop();
863 // remove any extra entries we added to the safepoint
864 last = sfpt_done->req() - 1;
865 for (int k = 0; k < nfields; k++) {
866 sfpt_done->del_req(last--);
867 }
868 JVMState *jvms = sfpt_done->jvms();
869 jvms->set_endoff(sfpt_done->req());
870 // Now make a pass over the debug information replacing any references
871 // to SafePointScalarObjectNode with the allocated object.
872 int start = jvms->debug_start();
873 int end = jvms->debug_end();
874 for (int i = start; i < end; i++) {
875 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
876 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
877 if (scobj->first_index(jvms) == sfpt_done->req() &&
878 scobj->n_fields() == (uint)nfields) {
879 assert(scobj->alloc() == alloc, "sanity");
880 sfpt_done->set_req(i, res);
881 }
882 }
883 }
884 _igvn._worklist.push(sfpt_done);
885 }
886 #ifndef PRODUCT
887 if (PrintEliminateAllocations) {
888 if (field != NULL) {
889 tty->print("=== At SafePoint node %d can't find value of Field: ",
890 sfpt->_idx);
891 field->print();
892 int field_idx = C->get_alias_index(field_addr_type);
893 tty->print(" (alias_idx=%d)", field_idx);
894 } else { // Array's element
895 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
896 sfpt->_idx, j);
897 }
898 tty->print(", which prevents elimination of: ");
899 if (res == NULL)
900 alloc->dump();
901 else
902 res->dump();
903 }
904 #endif
905 return false;
906 }
907 if (UseCompressedOops && field_type->isa_narrowoop()) {
908 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
909 // to be able scalar replace the allocation.
910 if (field_val->is_EncodeP()) {
911 field_val = field_val->in(1);
912 } else {
913 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
914 }
915 }
916 sfpt->add_req(field_val);
917 }
918 JVMState *jvms = sfpt->jvms();
919 jvms->set_endoff(sfpt->req());
920 // Now make a pass over the debug information replacing any references
921 // to the allocated object with "sobj"
922 int start = jvms->debug_start();
923 int end = jvms->debug_end();
924 sfpt->replace_edges_in_range(res, sobj, start, end);
925 _igvn._worklist.push(sfpt);
926 safepoints_done.append_if_missing(sfpt); // keep it for rollback
927 }
928 return true;
929 }
930
931 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
932 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
933 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
934 if (ctl_proj != NULL) {
935 igvn.replace_node(ctl_proj, n->in(0));
936 }
937 if (mem_proj != NULL) {
938 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
939 }
940 }
941
942 // Process users of eliminated allocation.
943 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
944 Node* res = alloc->result_cast();
945 if (res != NULL) {
946 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
947 Node *use = res->last_out(j);
948 uint oc1 = res->outcnt();
949
950 if (use->is_AddP()) {
951 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
952 Node *n = use->last_out(k);
953 uint oc2 = use->outcnt();
954 if (n->is_Store()) {
955 #ifdef ASSERT
956 // Verify that there is no dependent MemBarVolatile nodes,
957 // they should be removed during IGVN, see MemBarNode::Ideal().
958 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
959 p < pmax; p++) {
960 Node* mb = n->fast_out(p);
961 assert(mb->is_Initialize() || !mb->is_MemBar() ||
962 mb->req() <= MemBarNode::Precedent ||
963 mb->in(MemBarNode::Precedent) != n,
964 "MemBarVolatile should be eliminated for non-escaping object");
965 }
966 #endif
967 _igvn.replace_node(n, n->in(MemNode::Memory));
968 } else {
969 eliminate_gc_barrier(n);
970 }
971 k -= (oc2 - use->outcnt());
972 }
973 _igvn.remove_dead_node(use);
974 } else if (use->is_ArrayCopy()) {
975 // Disconnect ArrayCopy node
976 ArrayCopyNode* ac = use->as_ArrayCopy();
977 if (ac->is_clonebasic()) {
978 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
979 disconnect_projections(ac, _igvn);
980 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
981 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
982 disconnect_projections(membar_before->as_MemBar(), _igvn);
983 if (membar_after->is_MemBar()) {
984 disconnect_projections(membar_after->as_MemBar(), _igvn);
985 }
986 } else {
987 assert(ac->is_arraycopy_validated() ||
988 ac->is_copyof_validated() ||
989 ac->is_copyofrange_validated(), "unsupported");
990 CallProjections callprojs;
991 ac->extract_projections(&callprojs, true);
992
993 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
994 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
995 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
996
997 // Set control to top. IGVN will remove the remaining projections
998 ac->set_req(0, top());
999 ac->replace_edge(res, top());
1000
1001 // Disconnect src right away: it can help find new
1002 // opportunities for allocation elimination
1003 Node* src = ac->in(ArrayCopyNode::Src);
1004 ac->replace_edge(src, top());
1005 // src can be top at this point if src and dest of the
1006 // arraycopy were the same
1007 if (src->outcnt() == 0 && !src->is_top()) {
1008 _igvn.remove_dead_node(src);
1009 }
1010 }
1011 _igvn._worklist.push(ac);
1012 } else {
1013 eliminate_gc_barrier(use);
1014 }
1015 j -= (oc1 - res->outcnt());
1016 }
1017 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1018 _igvn.remove_dead_node(res);
1019 }
1020
1021 //
1022 // Process other users of allocation's projections
1023 //
1024 if (_resproj != NULL && _resproj->outcnt() != 0) {
1025 // First disconnect stores captured by Initialize node.
1026 // If Initialize node is eliminated first in the following code,
1027 // it will kill such stores and DUIterator_Last will assert.
1028 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1029 Node *use = _resproj->fast_out(j);
1030 if (use->is_AddP()) {
1031 // raw memory addresses used only by the initialization
1032 _igvn.replace_node(use, C->top());
1033 --j; --jmax;
1034 }
1035 }
1036 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1037 Node *use = _resproj->last_out(j);
1038 uint oc1 = _resproj->outcnt();
1039 if (use->is_Initialize()) {
1040 // Eliminate Initialize node.
1041 InitializeNode *init = use->as_Initialize();
1042 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1043 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1044 if (ctrl_proj != NULL) {
1045 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1046 #ifdef ASSERT
1047 Node* tmp = init->in(TypeFunc::Control);
1048 assert(tmp == _fallthroughcatchproj, "allocation control projection");
1049 #endif
1050 }
1051 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1052 if (mem_proj != NULL) {
1053 Node *mem = init->in(TypeFunc::Memory);
1054 #ifdef ASSERT
1055 if (mem->is_MergeMem()) {
1056 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1057 } else {
1058 assert(mem == _memproj_fallthrough, "allocation memory projection");
1059 }
1060 #endif
1061 _igvn.replace_node(mem_proj, mem);
1062 }
1063 } else {
1064 assert(false, "only Initialize or AddP expected");
1065 }
1066 j -= (oc1 - _resproj->outcnt());
1067 }
1068 }
1069 if (_fallthroughcatchproj != NULL) {
1070 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1071 }
1072 if (_memproj_fallthrough != NULL) {
1073 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1074 }
1075 if (_memproj_catchall != NULL) {
1076 _igvn.replace_node(_memproj_catchall, C->top());
1077 }
1078 if (_ioproj_fallthrough != NULL) {
1079 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1080 }
1081 if (_ioproj_catchall != NULL) {
1082 _igvn.replace_node(_ioproj_catchall, C->top());
1083 }
1084 if (_catchallcatchproj != NULL) {
1085 _igvn.replace_node(_catchallcatchproj, C->top());
1086 }
1087 }
1088
1089 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1090 // Don't do scalar replacement if the frame can be popped by JVMTI:
1091 // if reallocation fails during deoptimization we'll pop all
1092 // interpreter frames for this compiled frame and that won't play
1093 // nice with JVMTI popframe.
1094 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1095 return false;
1096 }
1097 Node* klass = alloc->in(AllocateNode::KlassNode);
1098 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1099 Node* res = alloc->result_cast();
1100 // Eliminate boxing allocations which are not used
1101 // regardless scalar replacable status.
1102 bool boxing_alloc = C->eliminate_boxing() &&
1103 tklass->klass()->is_instance_klass() &&
1104 tklass->klass()->as_instance_klass()->is_box_klass();
1105 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1106 return false;
1107 }
1108
1109 extract_call_projections(alloc);
1110
1111 GrowableArray <SafePointNode *> safepoints;
1112 if (!can_eliminate_allocation(alloc, safepoints)) {
1113 return false;
1114 }
1115
1116 if (!alloc->_is_scalar_replaceable) {
1117 assert(res == NULL, "sanity");
1118 // We can only eliminate allocation if all debug info references
1119 // are already replaced with SafePointScalarObject because
1120 // we can't search for a fields value without instance_id.
1121 if (safepoints.length() > 0) {
1122 return false;
1123 }
1124 }
1125
1126 if (!scalar_replacement(alloc, safepoints)) {
1127 return false;
1128 }
1129
1130 CompileLog* log = C->log();
1131 if (log != NULL) {
1132 log->head("eliminate_allocation type='%d'",
1133 log->identify(tklass->klass()));
1134 JVMState* p = alloc->jvms();
1135 while (p != NULL) {
1136 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1137 p = p->caller();
1138 }
1139 log->tail("eliminate_allocation");
1140 }
1141
1142 process_users_of_allocation(alloc);
1143
1144 #ifndef PRODUCT
1145 if (PrintEliminateAllocations) {
1146 if (alloc->is_AllocateArray())
1147 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1148 else
1149 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1150 }
1151 #endif
1152
1153 return true;
1154 }
1155
1156 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1157 // EA should remove all uses of non-escaping boxing node.
1158 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1159 return false;
1160 }
1161
1162 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1163
1164 extract_call_projections(boxing);
1165
1166 const TypeTuple* r = boxing->tf()->range();
1167 assert(r->cnt() > TypeFunc::Parms, "sanity");
1168 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1169 assert(t != NULL, "sanity");
1170
1171 CompileLog* log = C->log();
1172 if (log != NULL) {
1173 log->head("eliminate_boxing type='%d'",
1174 log->identify(t->klass()));
1175 JVMState* p = boxing->jvms();
1176 while (p != NULL) {
1177 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1178 p = p->caller();
1179 }
1180 log->tail("eliminate_boxing");
1181 }
1182
1183 process_users_of_allocation(boxing);
1184
1185 #ifndef PRODUCT
1186 if (PrintEliminateAllocations) {
1187 tty->print("++++ Eliminated: %d ", boxing->_idx);
1188 boxing->method()->print_short_name(tty);
1189 tty->cr();
1190 }
1191 #endif
1192
1193 return true;
1194 }
1195
1196 //---------------------------set_eden_pointers-------------------------
1197 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1198 if (UseTLAB) { // Private allocation: load from TLS
1199 Node* thread = transform_later(new ThreadLocalNode());
1200 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1201 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1202 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1203 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1204 } else { // Shared allocation: load from globals
1205 CollectedHeap* ch = Universe::heap();
1206 address top_adr = (address)ch->top_addr();
1207 address end_adr = (address)ch->end_addr();
1208 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1209 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1210 }
1211 }
1212
1213
1214 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1215 Node* adr = basic_plus_adr(base, offset);
1216 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1217 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1218 transform_later(value);
1219 return value;
1220 }
1221
1222
1223 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1224 Node* adr = basic_plus_adr(base, offset);
1225 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1226 transform_later(mem);
1227 return mem;
1228 }
1229
1230 //=============================================================================
1231 //
1232 // A L L O C A T I O N
1233 //
1234 // Allocation attempts to be fast in the case of frequent small objects.
1235 // It breaks down like this:
1236 //
1237 // 1) Size in doublewords is computed. This is a constant for objects and
1238 // variable for most arrays. Doubleword units are used to avoid size
1239 // overflow of huge doubleword arrays. We need doublewords in the end for
1240 // rounding.
1241 //
1242 // 2) Size is checked for being 'too large'. Too-large allocations will go
1243 // the slow path into the VM. The slow path can throw any required
1244 // exceptions, and does all the special checks for very large arrays. The
1245 // size test can constant-fold away for objects. For objects with
1246 // finalizers it constant-folds the otherway: you always go slow with
1247 // finalizers.
1248 //
1249 // 3) If NOT using TLABs, this is the contended loop-back point.
1250 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1251 //
1252 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1253 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1254 // "size*8" we always enter the VM, where "largish" is a constant picked small
1255 // enough that there's always space between the eden max and 4Gig (old space is
1256 // there so it's quite large) and large enough that the cost of entering the VM
1257 // is dwarfed by the cost to initialize the space.
1258 //
1259 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1260 // down. If contended, repeat at step 3. If using TLABs normal-store
1261 // adjusted heap top back down; there is no contention.
1262 //
1263 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1264 // fields.
1265 //
1266 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1267 // oop flavor.
1268 //
1269 //=============================================================================
1270 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1271 // Allocations bigger than this always go the slow route.
1272 // This value must be small enough that allocation attempts that need to
1273 // trigger exceptions go the slow route. Also, it must be small enough so
1274 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1275 //=============================================================================j//
1276 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1277 // The allocator will coalesce int->oop copies away. See comment in
1278 // coalesce.cpp about how this works. It depends critically on the exact
1279 // code shape produced here, so if you are changing this code shape
1280 // make sure the GC info for the heap-top is correct in and around the
1281 // slow-path call.
1282 //
1283
1284 void PhaseMacroExpand::expand_allocate_common(
1285 AllocateNode* alloc, // allocation node to be expanded
1286 Node* length, // array length for an array allocation
1287 const TypeFunc* slow_call_type, // Type of slow call
1288 address slow_call_address // Address of slow call
1289 )
1290 {
1291 Node* ctrl = alloc->in(TypeFunc::Control);
1292 Node* mem = alloc->in(TypeFunc::Memory);
1293 Node* i_o = alloc->in(TypeFunc::I_O);
1294 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1295 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1296 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1297 assert(ctrl != NULL, "must have control");
1298
1299 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1300 // they will not be used if "always_slow" is set
1301 enum { slow_result_path = 1, fast_result_path = 2 };
1302 Node *result_region = NULL;
1303 Node *result_phi_rawmem = NULL;
1304 Node *result_phi_rawoop = NULL;
1305 Node *result_phi_i_o = NULL;
1306
1307 // The initial slow comparison is a size check, the comparison
1308 // we want to do is a BoolTest::gt
1309 bool expand_fast_path = true;
1310 int tv = _igvn.find_int_con(initial_slow_test, -1);
1311 if (tv >= 0) {
1312 // InitialTest has constant result
1313 // 0 - can fit in TLAB
1314 // 1 - always too big or negative
1315 assert(tv <= 1, "0 or 1 if a constant");
1316 expand_fast_path = (tv == 0);
1317 initial_slow_test = NULL;
1318 } else {
1319 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1320 }
1321
1322 if (C->env()->dtrace_alloc_probes() ||
1323 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1324 // Force slow-path allocation
1325 expand_fast_path = false;
1326 initial_slow_test = NULL;
1327 }
1328
1329 bool allocation_has_use = (alloc->result_cast() != NULL);
1330 if (!allocation_has_use) {
1331 InitializeNode* init = alloc->initialization();
1332 if (init != NULL) {
1333 init->remove(&_igvn);
1334 }
1335 if (expand_fast_path && (initial_slow_test == NULL)) {
1336 // Remove allocation node and return.
1337 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1338 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1339 #ifndef PRODUCT
1340 if (PrintEliminateAllocations) {
1341 tty->print("NotUsed ");
1342 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1343 if (res != NULL) {
1344 res->dump();
1345 } else {
1346 alloc->dump();
1347 }
1348 }
1349 #endif
1350 yank_alloc_node(alloc);
1351 return;
1352 }
1353 }
1354
1355 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1356 Node *slow_region = NULL;
1357 Node *toobig_false = ctrl;
1358
1359 // generate the initial test if necessary
1360 if (initial_slow_test != NULL ) {
1361 assert (expand_fast_path, "Only need test if there is a fast path");
1362 slow_region = new RegionNode(3);
1363
1364 // Now make the initial failure test. Usually a too-big test but
1365 // might be a TRUE for finalizers or a fancy class check for
1366 // newInstance0.
1367 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1368 transform_later(toobig_iff);
1369 // Plug the failing-too-big test into the slow-path region
1370 Node *toobig_true = new IfTrueNode( toobig_iff );
1371 transform_later(toobig_true);
1372 slow_region ->init_req( too_big_or_final_path, toobig_true );
1373 toobig_false = new IfFalseNode( toobig_iff );
1374 transform_later(toobig_false);
1375 } else {
1376 // No initial test, just fall into next case
1377 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1378 toobig_false = ctrl;
1379 debug_only(slow_region = NodeSentinel);
1380 }
1381
1382 // If we are here there are several possibilities
1383 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1384 // no_initial_check means a constant allocation.
1385 // - If check always evaluates to false -> expand_fast_path is false (see above)
1386 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1387 // if !allocation_has_use the fast path is empty
1388 // if !allocation_has_use && no_initial_check
1389 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1390 // removed by yank_alloc_node above.
1391
1392 Node *slow_mem = mem; // save the current memory state for slow path
1393 // generate the fast allocation code unless we know that the initial test will always go slow
1394 if (expand_fast_path) {
1395 // Fast path modifies only raw memory.
1396 if (mem->is_MergeMem()) {
1397 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1398 }
1399
1400 // allocate the Region and Phi nodes for the result
1401 result_region = new RegionNode(3);
1402 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1403 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1404
1405 // Grab regular I/O before optional prefetch may change it.
1406 // Slow-path does no I/O so just set it to the original I/O.
1407 result_phi_i_o->init_req(slow_result_path, i_o);
1408
1409 // Name successful fast-path variables
1410 Node* fast_oop_ctrl;
1411 Node* fast_oop_rawmem;
1412 if (allocation_has_use) {
1413 Node* needgc_ctrl = NULL;
1414 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1415
1416 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1418 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1419 fast_oop_ctrl, fast_oop_rawmem,
1420 prefetch_lines);
1421
1422 if (initial_slow_test != NULL) {
1423 // This completes all paths into the slow merge point
1424 slow_region->init_req(need_gc_path, needgc_ctrl);
1425 transform_later(slow_region);
1426 } else {
1427 // No initial slow path needed!
1428 // Just fall from the need-GC path straight into the VM call.
1429 slow_region = needgc_ctrl;
1430 }
1431
1432 InitializeNode* init = alloc->initialization();
1433 fast_oop_rawmem = initialize_object(alloc,
1434 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1435 klass_node, length, size_in_bytes);
1436 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1437 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1438
1439 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1440 } else {
1441 assert (initial_slow_test != NULL, "sanity");
1442 fast_oop_ctrl = toobig_false;
1443 fast_oop_rawmem = mem;
1444 transform_later(slow_region);
1445 }
1446
1447 // Plug in the successful fast-path into the result merge point
1448 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1449 result_phi_i_o ->init_req(fast_result_path, i_o);
1450 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1451 } else {
1452 slow_region = ctrl;
1453 result_phi_i_o = i_o; // Rename it to use in the following code.
1454 }
1455
1456 // Generate slow-path call
1457 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1458 OptoRuntime::stub_name(slow_call_address),
1459 alloc->jvms()->bci(),
1460 TypePtr::BOTTOM);
1461 call->init_req(TypeFunc::Control, slow_region);
1462 call->init_req(TypeFunc::I_O, top()); // does no i/o
1463 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1464 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1465 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1466
1467 call->init_req(TypeFunc::Parms+0, klass_node);
1468 if (length != NULL) {
1469 call->init_req(TypeFunc::Parms+1, length);
1470 }
1471
1472 // Copy debug information and adjust JVMState information, then replace
1473 // allocate node with the call
1474 copy_call_debug_info((CallNode *) alloc, call);
1475 if (expand_fast_path) {
1476 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1477 } else {
1478 // Hook i_o projection to avoid its elimination during allocation
1479 // replacement (when only a slow call is generated).
1480 call->set_req(TypeFunc::I_O, result_phi_i_o);
1481 }
1482 _igvn.replace_node(alloc, call);
1483 transform_later(call);
1484
1485 // Identify the output projections from the allocate node and
1486 // adjust any references to them.
1487 // The control and io projections look like:
1488 //
1489 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1490 // Allocate Catch
1491 // ^---Proj(io) <-------+ ^---CatchProj(io)
1492 //
1493 // We are interested in the CatchProj nodes.
1494 //
1495 extract_call_projections(call);
1496
1497 // An allocate node has separate memory projections for the uses on
1498 // the control and i_o paths. Replace the control memory projection with
1499 // result_phi_rawmem (unless we are only generating a slow call when
1500 // both memory projections are combined)
1501 if (expand_fast_path && _memproj_fallthrough != NULL) {
1502 migrate_outs(_memproj_fallthrough, result_phi_rawmem);
1503 }
1504 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1505 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1506 if (_memproj_catchall != NULL ) {
1507 if (_memproj_fallthrough == NULL) {
1508 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1509 transform_later(_memproj_fallthrough);
1510 }
1511 migrate_outs(_memproj_catchall, _memproj_fallthrough);
1512 _igvn.remove_dead_node(_memproj_catchall);
1513 }
1514
1515 // An allocate node has separate i_o projections for the uses on the control
1516 // and i_o paths. Always replace the control i_o projection with result i_o
1517 // otherwise incoming i_o become dead when only a slow call is generated
1518 // (it is different from memory projections where both projections are
1519 // combined in such case).
1520 if (_ioproj_fallthrough != NULL) {
1521 migrate_outs(_ioproj_fallthrough, result_phi_i_o);
1522 }
1523 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1524 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1525 if (_ioproj_catchall != NULL ) {
1526 if (_ioproj_fallthrough == NULL) {
1527 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1528 transform_later(_ioproj_fallthrough);
1529 }
1530 migrate_outs(_ioproj_catchall, _ioproj_fallthrough);
1531 _igvn.remove_dead_node(_ioproj_catchall);
1532 }
1533
1534 // if we generated only a slow call, we are done
1535 if (!expand_fast_path) {
1536 // Now we can unhook i_o.
1537 if (result_phi_i_o->outcnt() > 1) {
1538 call->set_req(TypeFunc::I_O, top());
1539 } else {
1540 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1541 // Case of new array with negative size known during compilation.
1542 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1543 // following code since call to runtime will throw exception.
1544 // As result there will be no users of i_o after the call.
1545 // Leave i_o attached to this call to avoid problems in preceding graph.
1546 }
1547 return;
1548 }
1549
1550 if (_fallthroughcatchproj != NULL) {
1551 ctrl = _fallthroughcatchproj->clone();
1552 transform_later(ctrl);
1553 _igvn.replace_node(_fallthroughcatchproj, result_region);
1554 } else {
1555 ctrl = top();
1556 }
1557 Node *slow_result;
1558 if (_resproj == NULL) {
1559 // no uses of the allocation result
1560 slow_result = top();
1561 } else {
1562 slow_result = _resproj->clone();
1563 transform_later(slow_result);
1564 _igvn.replace_node(_resproj, result_phi_rawoop);
1565 }
1566
1567 // Plug slow-path into result merge point
1568 result_region->init_req( slow_result_path, ctrl);
1569 transform_later(result_region);
1570 if (allocation_has_use) {
1571 result_phi_rawoop->init_req(slow_result_path, slow_result);
1572 transform_later(result_phi_rawoop);
1573 }
1574 result_phi_rawmem->init_req(slow_result_path, _memproj_fallthrough);
1575 transform_later(result_phi_rawmem);
1576 transform_later(result_phi_i_o);
1577 // This completes all paths into the result merge point
1578 }
1579
1580 // Remove alloc node that has no uses.
1581 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1582 Node* ctrl = alloc->in(TypeFunc::Control);
1583 Node* mem = alloc->in(TypeFunc::Memory);
1584 Node* i_o = alloc->in(TypeFunc::I_O);
1585
1586 extract_call_projections(alloc);
1587 if (_resproj != NULL) {
1588 for (DUIterator_Fast imax, i = _resproj->fast_outs(imax); i < imax; i++) {
1589 Node* use = _resproj->fast_out(i);
1590 use->isa_MemBar()->remove(&_igvn);
1591 --imax;
1592 --i; // back up iterator
1593 }
1594 assert(_resproj->outcnt() == 0, "all uses must be deleted");
1595 _igvn.remove_dead_node(_resproj);
1596 }
1597 if (_fallthroughcatchproj != NULL) {
1598 migrate_outs(_fallthroughcatchproj, ctrl);
1599 _igvn.remove_dead_node(_fallthroughcatchproj);
1600 }
1601 if (_catchallcatchproj != NULL) {
1602 _igvn.rehash_node_delayed(_catchallcatchproj);
1603 _catchallcatchproj->set_req(0, top());
1604 }
1605 if (_fallthroughproj != NULL) {
1606 Node* catchnode = _fallthroughproj->unique_ctrl_out();
1607 _igvn.remove_dead_node(catchnode);
1608 _igvn.remove_dead_node(_fallthroughproj);
1609 }
1610 if (_memproj_fallthrough != NULL) {
1611 migrate_outs(_memproj_fallthrough, mem);
1612 _igvn.remove_dead_node(_memproj_fallthrough);
1613 }
1614 if (_ioproj_fallthrough != NULL) {
1615 migrate_outs(_ioproj_fallthrough, i_o);
1616 _igvn.remove_dead_node(_ioproj_fallthrough);
1617 }
1618 if (_memproj_catchall != NULL) {
1619 _igvn.rehash_node_delayed(_memproj_catchall);
1620 _memproj_catchall->set_req(0, top());
1621 }
1622 if (_ioproj_catchall != NULL) {
1623 _igvn.rehash_node_delayed(_ioproj_catchall);
1624 _ioproj_catchall->set_req(0, top());
1625 }
1626 #ifndef PRODUCT
1627 if (PrintEliminateAllocations) {
1628 if (alloc->is_AllocateArray()) {
1629 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1630 } else {
1631 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1632 }
1633 }
1634 #endif
1635 _igvn.remove_dead_node(alloc);
1636 }
1637
1638 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1639 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1640 // If initialization is performed by an array copy, any required
1641 // MemBarStoreStore was already added. If the object does not
1642 // escape no need for a MemBarStoreStore. If the object does not
1643 // escape in its initializer and memory barrier (MemBarStoreStore or
1644 // stronger) is already added at exit of initializer, also no need
1645 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1646 // so that stores that initialize this object can't be reordered
1647 // with a subsequent store that makes this object accessible by
1648 // other threads.
1649 // Other threads include java threads and JVM internal threads
1650 // (for example concurrent GC threads). Current concurrent GC
1651 // implementation: G1 will not scan newly created object,
1652 // so it's safe to skip storestore barrier when allocation does
1653 // not escape.
1654 if (!alloc->does_not_escape_thread() &&
1655 !alloc->is_allocation_MemBar_redundant() &&
1656 (init == NULL || !init->is_complete_with_arraycopy())) {
1657 if (init == NULL || init->req() < InitializeNode::RawStores) {
1658 // No InitializeNode or no stores captured by zeroing
1659 // elimination. Simply add the MemBarStoreStore after object
1660 // initialization.
1661 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1662 transform_later(mb);
1663
1664 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1665 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1666 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1667 transform_later(fast_oop_ctrl);
1668 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1669 transform_later(fast_oop_rawmem);
1670 } else {
1671 // Add the MemBarStoreStore after the InitializeNode so that
1672 // all stores performing the initialization that were moved
1673 // before the InitializeNode happen before the storestore
1674 // barrier.
1675
1676 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1677 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1678
1679 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1680 transform_later(mb);
1681
1682 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1683 transform_later(ctrl);
1684 Node* mem = new ProjNode(init, TypeFunc::Memory);
1685 transform_later(mem);
1686
1687 // The MemBarStoreStore depends on control and memory coming
1688 // from the InitializeNode
1689 mb->init_req(TypeFunc::Memory, mem);
1690 mb->init_req(TypeFunc::Control, ctrl);
1691
1692 ctrl = new ProjNode(mb, TypeFunc::Control);
1693 transform_later(ctrl);
1694 mem = new ProjNode(mb, TypeFunc::Memory);
1695 transform_later(mem);
1696
1697 // All nodes that depended on the InitializeNode for control
1698 // and memory must now depend on the MemBarNode that itself
1699 // depends on the InitializeNode
1700 if (init_ctrl != NULL) {
1701 _igvn.replace_node(init_ctrl, ctrl);
1702 }
1703 if (init_mem != NULL) {
1704 _igvn.replace_node(init_mem, mem);
1705 }
1706 }
1707 }
1708 }
1709
1710 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1711 Node*& ctrl, Node*& rawmem) {
1712 if (C->env()->dtrace_extended_probes()) {
1713 // Slow-path call
1714 int size = TypeFunc::Parms + 2;
1715 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1716 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1717 "dtrace_object_alloc",
1718 TypeRawPtr::BOTTOM);
1719
1720 // Get base of thread-local storage area
1721 Node* thread = new ThreadLocalNode();
1722 transform_later(thread);
1723
1724 call->init_req(TypeFunc::Parms + 0, thread);
1725 call->init_req(TypeFunc::Parms + 1, oop);
1726 call->init_req(TypeFunc::Control, ctrl);
1727 call->init_req(TypeFunc::I_O , top()); // does no i/o
1728 call->init_req(TypeFunc::Memory , ctrl);
1729 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1730 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1731 transform_later(call);
1732 ctrl = new ProjNode(call, TypeFunc::Control);
1733 transform_later(ctrl);
1734 rawmem = new ProjNode(call, TypeFunc::Memory);
1735 transform_later(rawmem);
1736 }
1737 }
1738
1739 // Helper for PhaseMacroExpand::expand_allocate_common.
1740 // Initializes the newly-allocated storage.
1741 Node*
1742 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1743 Node* control, Node* rawmem, Node* object,
1744 Node* klass_node, Node* length,
1745 Node* size_in_bytes) {
1746 InitializeNode* init = alloc->initialization();
1747 // Store the klass & mark bits
1748 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1749 if (!mark_node->is_Con()) {
1750 transform_later(mark_node);
1751 }
1752 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1753
1754 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1755 int header_size = alloc->minimum_header_size(); // conservatively small
1756
1757 // Array length
1758 if (length != NULL) { // Arrays need length field
1759 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1760 // conservatively small header size:
1761 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1762 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1763 if (k->is_array_klass()) // we know the exact header size in most cases:
1764 header_size = Klass::layout_helper_header_size(k->layout_helper());
1765 }
1766
1767 // Clear the object body, if necessary.
1768 if (init == NULL) {
1769 // The init has somehow disappeared; be cautious and clear everything.
1770 //
1771 // This can happen if a node is allocated but an uncommon trap occurs
1772 // immediately. In this case, the Initialize gets associated with the
1773 // trap, and may be placed in a different (outer) loop, if the Allocate
1774 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1775 // there can be two Allocates to one Initialize. The answer in all these
1776 // edge cases is safety first. It is always safe to clear immediately
1777 // within an Allocate, and then (maybe or maybe not) clear some more later.
1778 if (!(UseTLAB && ZeroTLAB)) {
1779 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1780 header_size, size_in_bytes,
1781 &_igvn);
1782 }
1783 } else {
1784 if (!init->is_complete()) {
1785 // Try to win by zeroing only what the init does not store.
1786 // We can also try to do some peephole optimizations,
1787 // such as combining some adjacent subword stores.
1788 rawmem = init->complete_stores(control, rawmem, object,
1789 header_size, size_in_bytes, &_igvn);
1790 }
1791 // We have no more use for this link, since the AllocateNode goes away:
1792 init->set_req(InitializeNode::RawAddress, top());
1793 // (If we keep the link, it just confuses the register allocator,
1794 // who thinks he sees a real use of the address by the membar.)
1795 }
1796
1797 return rawmem;
1798 }
1799
1800 // Generate prefetch instructions for next allocations.
1801 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1802 Node*& contended_phi_rawmem,
1803 Node* old_eden_top, Node* new_eden_top,
1804 intx lines) {
1805 enum { fall_in_path = 1, pf_path = 2 };
1806 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1807 // Generate prefetch allocation with watermark check.
1808 // As an allocation hits the watermark, we will prefetch starting
1809 // at a "distance" away from watermark.
1810
1811 Node *pf_region = new RegionNode(3);
1812 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1813 TypeRawPtr::BOTTOM );
1814 // I/O is used for Prefetch
1815 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1816
1817 Node *thread = new ThreadLocalNode();
1818 transform_later(thread);
1819
1820 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1821 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1822 transform_later(eden_pf_adr);
1823
1824 Node *old_pf_wm = new LoadPNode(needgc_false,
1825 contended_phi_rawmem, eden_pf_adr,
1826 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1827 MemNode::unordered);
1828 transform_later(old_pf_wm);
1829
1830 // check against new_eden_top
1831 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1832 transform_later(need_pf_cmp);
1833 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1834 transform_later(need_pf_bol);
1835 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1836 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1837 transform_later(need_pf_iff);
1838
1839 // true node, add prefetchdistance
1840 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1841 transform_later(need_pf_true);
1842
1843 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1844 transform_later(need_pf_false);
1845
1846 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1847 _igvn.MakeConX(AllocatePrefetchDistance) );
1848 transform_later(new_pf_wmt );
1849 new_pf_wmt->set_req(0, need_pf_true);
1850
1851 Node *store_new_wmt = new StorePNode(need_pf_true,
1852 contended_phi_rawmem, eden_pf_adr,
1853 TypeRawPtr::BOTTOM, new_pf_wmt,
1854 MemNode::unordered);
1855 transform_later(store_new_wmt);
1856
1857 // adding prefetches
1858 pf_phi_abio->init_req( fall_in_path, i_o );
1859
1860 Node *prefetch_adr;
1861 Node *prefetch;
1862 uint step_size = AllocatePrefetchStepSize;
1863 uint distance = 0;
1864
1865 for ( intx i = 0; i < lines; i++ ) {
1866 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1867 _igvn.MakeConX(distance) );
1868 transform_later(prefetch_adr);
1869 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1870 transform_later(prefetch);
1871 distance += step_size;
1872 i_o = prefetch;
1873 }
1874 pf_phi_abio->set_req( pf_path, i_o );
1875
1876 pf_region->init_req( fall_in_path, need_pf_false );
1877 pf_region->init_req( pf_path, need_pf_true );
1878
1879 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1880 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1881
1882 transform_later(pf_region);
1883 transform_later(pf_phi_rawmem);
1884 transform_later(pf_phi_abio);
1885
1886 needgc_false = pf_region;
1887 contended_phi_rawmem = pf_phi_rawmem;
1888 i_o = pf_phi_abio;
1889 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1890 // Insert a prefetch instruction for each allocation.
1891 // This code is used to generate 1 prefetch instruction per cache line.
1892
1893 // Generate several prefetch instructions.
1894 uint step_size = AllocatePrefetchStepSize;
1895 uint distance = AllocatePrefetchDistance;
1896
1897 // Next cache address.
1898 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1899 _igvn.MakeConX(step_size + distance));
1900 transform_later(cache_adr);
1901 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1902 transform_later(cache_adr);
1903 // Address is aligned to execute prefetch to the beginning of cache line size
1904 // (it is important when BIS instruction is used on SPARC as prefetch).
1905 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1906 cache_adr = new AndXNode(cache_adr, mask);
1907 transform_later(cache_adr);
1908 cache_adr = new CastX2PNode(cache_adr);
1909 transform_later(cache_adr);
1910
1911 // Prefetch
1912 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1913 prefetch->set_req(0, needgc_false);
1914 transform_later(prefetch);
1915 contended_phi_rawmem = prefetch;
1916 Node *prefetch_adr;
1917 distance = step_size;
1918 for ( intx i = 1; i < lines; i++ ) {
1919 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1920 _igvn.MakeConX(distance) );
1921 transform_later(prefetch_adr);
1922 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1923 transform_later(prefetch);
1924 distance += step_size;
1925 contended_phi_rawmem = prefetch;
1926 }
1927 } else if( AllocatePrefetchStyle > 0 ) {
1928 // Insert a prefetch for each allocation only on the fast-path
1929 Node *prefetch_adr;
1930 Node *prefetch;
1931 // Generate several prefetch instructions.
1932 uint step_size = AllocatePrefetchStepSize;
1933 uint distance = AllocatePrefetchDistance;
1934 for ( intx i = 0; i < lines; i++ ) {
1935 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1936 _igvn.MakeConX(distance) );
1937 transform_later(prefetch_adr);
1938 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1939 // Do not let it float too high, since if eden_top == eden_end,
1940 // both might be null.
1941 if( i == 0 ) { // Set control for first prefetch, next follows it
1942 prefetch->init_req(0, needgc_false);
1943 }
1944 transform_later(prefetch);
1945 distance += step_size;
1946 i_o = prefetch;
1947 }
1948 }
1949 return i_o;
1950 }
1951
1952
1953 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1954 expand_allocate_common(alloc, NULL,
1955 OptoRuntime::new_instance_Type(),
1956 OptoRuntime::new_instance_Java());
1957 }
1958
1959 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1960 Node* length = alloc->in(AllocateNode::ALength);
1961 InitializeNode* init = alloc->initialization();
1962 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1963 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1964 address slow_call_address; // Address of slow call
1965 if (init != NULL && init->is_complete_with_arraycopy() &&
1966 k->is_type_array_klass()) {
1967 // Don't zero type array during slow allocation in VM since
1968 // it will be initialized later by arraycopy in compiled code.
1969 slow_call_address = OptoRuntime::new_array_nozero_Java();
1970 } else {
1971 slow_call_address = OptoRuntime::new_array_Java();
1972 }
1973 expand_allocate_common(alloc, length,
1974 OptoRuntime::new_array_Type(),
1975 slow_call_address);
1976 }
1977
1978 //-------------------mark_eliminated_box----------------------------------
1979 //
1980 // During EA obj may point to several objects but after few ideal graph
1981 // transformations (CCP) it may point to only one non escaping object
1982 // (but still using phi), corresponding locks and unlocks will be marked
1983 // for elimination. Later obj could be replaced with a new node (new phi)
1984 // and which does not have escape information. And later after some graph
1985 // reshape other locks and unlocks (which were not marked for elimination
1986 // before) are connected to this new obj (phi) but they still will not be
1987 // marked for elimination since new obj has no escape information.
1988 // Mark all associated (same box and obj) lock and unlock nodes for
1989 // elimination if some of them marked already.
1990 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1991 if (oldbox->as_BoxLock()->is_eliminated())
1992 return; // This BoxLock node was processed already.
1993
1994 // New implementation (EliminateNestedLocks) has separate BoxLock
1995 // node for each locked region so mark all associated locks/unlocks as
1996 // eliminated even if different objects are referenced in one locked region
1997 // (for example, OSR compilation of nested loop inside locked scope).
1998 if (EliminateNestedLocks ||
1999 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2000 // Box is used only in one lock region. Mark this box as eliminated.
2001 _igvn.hash_delete(oldbox);
2002 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2003 _igvn.hash_insert(oldbox);
2004
2005 for (uint i = 0; i < oldbox->outcnt(); i++) {
2006 Node* u = oldbox->raw_out(i);
2007 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2008 AbstractLockNode* alock = u->as_AbstractLock();
2009 // Check lock's box since box could be referenced by Lock's debug info.
2010 if (alock->box_node() == oldbox) {
2011 // Mark eliminated all related locks and unlocks.
2012 #ifdef ASSERT
2013 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2014 #endif
2015 alock->set_non_esc_obj();
2016 }
2017 }
2018 }
2019 return;
2020 }
2021
2022 // Create new "eliminated" BoxLock node and use it in monitor debug info
2023 // instead of oldbox for the same object.
2024 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2025
2026 // Note: BoxLock node is marked eliminated only here and it is used
2027 // to indicate that all associated lock and unlock nodes are marked
2028 // for elimination.
2029 newbox->set_eliminated();
2030 transform_later(newbox);
2031
2032 // Replace old box node with new box for all users of the same object.
2033 for (uint i = 0; i < oldbox->outcnt();) {
2034 bool next_edge = true;
2035
2036 Node* u = oldbox->raw_out(i);
2037 if (u->is_AbstractLock()) {
2038 AbstractLockNode* alock = u->as_AbstractLock();
2039 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2040 // Replace Box and mark eliminated all related locks and unlocks.
2041 #ifdef ASSERT
2042 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2043 #endif
2044 alock->set_non_esc_obj();
2045 _igvn.rehash_node_delayed(alock);
2046 alock->set_box_node(newbox);
2047 next_edge = false;
2048 }
2049 }
2050 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2051 FastLockNode* flock = u->as_FastLock();
2052 assert(flock->box_node() == oldbox, "sanity");
2053 _igvn.rehash_node_delayed(flock);
2054 flock->set_box_node(newbox);
2055 next_edge = false;
2056 }
2057
2058 // Replace old box in monitor debug info.
2059 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2060 SafePointNode* sfn = u->as_SafePoint();
2061 JVMState* youngest_jvms = sfn->jvms();
2062 int max_depth = youngest_jvms->depth();
2063 for (int depth = 1; depth <= max_depth; depth++) {
2064 JVMState* jvms = youngest_jvms->of_depth(depth);
2065 int num_mon = jvms->nof_monitors();
2066 // Loop over monitors
2067 for (int idx = 0; idx < num_mon; idx++) {
2068 Node* obj_node = sfn->monitor_obj(jvms, idx);
2069 Node* box_node = sfn->monitor_box(jvms, idx);
2070 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2071 int j = jvms->monitor_box_offset(idx);
2072 _igvn.replace_input_of(u, j, newbox);
2073 next_edge = false;
2074 }
2075 }
2076 }
2077 }
2078 if (next_edge) i++;
2079 }
2080 }
2081
2082 //-----------------------mark_eliminated_locking_nodes-----------------------
2083 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2084 if (EliminateNestedLocks) {
2085 if (alock->is_nested()) {
2086 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2087 return;
2088 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2089 // Only Lock node has JVMState needed here.
2090 // Not that preceding claim is documented anywhere else.
2091 if (alock->jvms() != NULL) {
2092 if (alock->as_Lock()->is_nested_lock_region()) {
2093 // Mark eliminated related nested locks and unlocks.
2094 Node* obj = alock->obj_node();
2095 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2096 assert(!box_node->is_eliminated(), "should not be marked yet");
2097 // Note: BoxLock node is marked eliminated only here
2098 // and it is used to indicate that all associated lock
2099 // and unlock nodes are marked for elimination.
2100 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2101 for (uint i = 0; i < box_node->outcnt(); i++) {
2102 Node* u = box_node->raw_out(i);
2103 if (u->is_AbstractLock()) {
2104 alock = u->as_AbstractLock();
2105 if (alock->box_node() == box_node) {
2106 // Verify that this Box is referenced only by related locks.
2107 assert(alock->obj_node()->eqv_uncast(obj), "");
2108 // Mark all related locks and unlocks.
2109 #ifdef ASSERT
2110 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2111 #endif
2112 alock->set_nested();
2113 }
2114 }
2115 }
2116 } else {
2117 #ifdef ASSERT
2118 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2119 if (C->log() != NULL)
2120 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2121 #endif
2122 }
2123 }
2124 return;
2125 }
2126 // Process locks for non escaping object
2127 assert(alock->is_non_esc_obj(), "");
2128 } // EliminateNestedLocks
2129
2130 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2131 // Look for all locks of this object and mark them and
2132 // corresponding BoxLock nodes as eliminated.
2133 Node* obj = alock->obj_node();
2134 for (uint j = 0; j < obj->outcnt(); j++) {
2135 Node* o = obj->raw_out(j);
2136 if (o->is_AbstractLock() &&
2137 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2138 alock = o->as_AbstractLock();
2139 Node* box = alock->box_node();
2140 // Replace old box node with new eliminated box for all users
2141 // of the same object and mark related locks as eliminated.
2142 mark_eliminated_box(box, obj);
2143 }
2144 }
2145 }
2146 }
2147
2148 // we have determined that this lock/unlock can be eliminated, we simply
2149 // eliminate the node without expanding it.
2150 //
2151 // Note: The membar's associated with the lock/unlock are currently not
2152 // eliminated. This should be investigated as a future enhancement.
2153 //
2154 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2155
2156 if (!alock->is_eliminated()) {
2157 return false;
2158 }
2159 #ifdef ASSERT
2160 if (!alock->is_coarsened()) {
2161 // Check that new "eliminated" BoxLock node is created.
2162 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2163 assert(oldbox->is_eliminated(), "should be done already");
2164 }
2165 #endif
2166
2167 alock->log_lock_optimization(C, "eliminate_lock");
2168
2169 #ifndef PRODUCT
2170 if (PrintEliminateLocks) {
2171 if (alock->is_Lock()) {
2172 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2173 } else {
2174 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2175 }
2176 }
2177 #endif
2178
2179 Node* mem = alock->in(TypeFunc::Memory);
2180 Node* ctrl = alock->in(TypeFunc::Control);
2181 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2182
2183 extract_call_projections(alock);
2184 // There are 2 projections from the lock. The lock node will
2185 // be deleted when its last use is subsumed below.
2186 assert(alock->outcnt() == 2 &&
2187 _fallthroughproj != NULL &&
2188 _memproj_fallthrough != NULL,
2189 "Unexpected projections from Lock/Unlock");
2190
2191 Node* fallthroughproj = _fallthroughproj;
2192 Node* memproj_fallthrough = _memproj_fallthrough;
2193
2194 // The memory projection from a lock/unlock is RawMem
2195 // The input to a Lock is merged memory, so extract its RawMem input
2196 // (unless the MergeMem has been optimized away.)
2197 if (alock->is_Lock()) {
2198 // Seach for MemBarAcquireLock node and delete it also.
2199 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2200 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2201 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2202 Node* memproj = membar->proj_out(TypeFunc::Memory);
2203 _igvn.replace_node(ctrlproj, fallthroughproj);
2204 _igvn.replace_node(memproj, memproj_fallthrough);
2205
2206 // Delete FastLock node also if this Lock node is unique user
2207 // (a loop peeling may clone a Lock node).
2208 Node* flock = alock->as_Lock()->fastlock_node();
2209 if (flock->outcnt() == 1) {
2210 assert(flock->unique_out() == alock, "sanity");
2211 _igvn.replace_node(flock, top());
2212 }
2213 }
2214
2215 // Seach for MemBarReleaseLock node and delete it also.
2216 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2217 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2218 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2219 mem->is_Proj() && membar == mem->in(0), "");
2220 _igvn.replace_node(fallthroughproj, ctrl);
2221 _igvn.replace_node(memproj_fallthrough, mem);
2222 fallthroughproj = ctrl;
2223 memproj_fallthrough = mem;
2224 ctrl = membar->in(TypeFunc::Control);
2225 mem = membar->in(TypeFunc::Memory);
2226 }
2227
2228 _igvn.replace_node(fallthroughproj, ctrl);
2229 _igvn.replace_node(memproj_fallthrough, mem);
2230 return true;
2231 }
2232
2233
2234 //------------------------------expand_lock_node----------------------
2235 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2236
2237 Node* ctrl = lock->in(TypeFunc::Control);
2238 Node* mem = lock->in(TypeFunc::Memory);
2239 Node* obj = lock->obj_node();
2240 Node* box = lock->box_node();
2241 Node* flock = lock->fastlock_node();
2242
2243 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2244
2245 // Make the merge point
2246 Node *region;
2247 Node *mem_phi;
2248 Node *slow_path;
2249
2250 if (UseOptoBiasInlining) {
2251 /*
2252 * See the full description in MacroAssembler::biased_locking_enter().
2253 *
2254 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2255 * // The object is biased.
2256 * proto_node = klass->prototype_header;
2257 * o_node = thread | proto_node;
2258 * x_node = o_node ^ mark_word;
2259 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2260 * // Done.
2261 * } else {
2262 * if( (x_node & biased_lock_mask) != 0 ) {
2263 * // The klass's prototype header is no longer biased.
2264 * cas(&mark_word, mark_word, proto_node)
2265 * goto cas_lock;
2266 * } else {
2267 * // The klass's prototype header is still biased.
2268 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2269 * old = mark_word;
2270 * new = o_node;
2271 * } else {
2272 * // Different thread or anonymous biased.
2273 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2274 * new = thread | old;
2275 * }
2276 * // Try to rebias.
2277 * if( cas(&mark_word, old, new) == 0 ) {
2278 * // Done.
2279 * } else {
2280 * goto slow_path; // Failed.
2281 * }
2282 * }
2283 * }
2284 * } else {
2285 * // The object is not biased.
2286 * cas_lock:
2287 * if( FastLock(obj) == 0 ) {
2288 * // Done.
2289 * } else {
2290 * slow_path:
2291 * OptoRuntime::complete_monitor_locking_Java(obj);
2292 * }
2293 * }
2294 */
2295
2296 region = new RegionNode(5);
2297 // create a Phi for the memory state
2298 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2299
2300 Node* fast_lock_region = new RegionNode(3);
2301 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2302
2303 // First, check mark word for the biased lock pattern.
2304 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2305
2306 // Get fast path - mark word has the biased lock pattern.
2307 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2308 markWord::biased_lock_mask_in_place,
2309 markWord::biased_lock_pattern, true);
2310 // fast_lock_region->in(1) is set to slow path.
2311 fast_lock_mem_phi->init_req(1, mem);
2312
2313 // Now check that the lock is biased to the current thread and has
2314 // the same epoch and bias as Klass::_prototype_header.
2315
2316 // Special-case a fresh allocation to avoid building nodes:
2317 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2318 if (klass_node == NULL) {
2319 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2320 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2321 #ifdef _LP64
2322 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2323 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2324 klass_node->in(1)->init_req(0, ctrl);
2325 } else
2326 #endif
2327 klass_node->init_req(0, ctrl);
2328 }
2329 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2330
2331 Node* thread = transform_later(new ThreadLocalNode());
2332 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2333 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2334 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2335
2336 // Get slow path - mark word does NOT match the value.
2337 STATIC_ASSERT(markWord::age_mask_in_place <= INT_MAX);
2338 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2339 (~(int)markWord::age_mask_in_place), 0);
2340 // region->in(3) is set to fast path - the object is biased to the current thread.
2341 mem_phi->init_req(3, mem);
2342
2343
2344 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2345
2346
2347 // First, check biased pattern.
2348 // Get fast path - _prototype_header has the same biased lock pattern.
2349 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2350 markWord::biased_lock_mask_in_place, 0, true);
2351
2352 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2353 // fast_lock_region->in(2) - the prototype header is no longer biased
2354 // and we have to revoke the bias on this object.
2355 // We are going to try to reset the mark of this object to the prototype
2356 // value and fall through to the CAS-based locking scheme.
2357 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2358 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2359 proto_node, mark_node);
2360 transform_later(cas);
2361 Node* proj = transform_later(new SCMemProjNode(cas));
2362 fast_lock_mem_phi->init_req(2, proj);
2363
2364
2365 // Second, check epoch bits.
2366 Node* rebiased_region = new RegionNode(3);
2367 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2368 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2369
2370 // Get slow path - mark word does NOT match epoch bits.
2371 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2372 markWord::epoch_mask_in_place, 0);
2373 // The epoch of the current bias is not valid, attempt to rebias the object
2374 // toward the current thread.
2375 rebiased_region->init_req(2, epoch_ctrl);
2376 old_phi->init_req(2, mark_node);
2377 new_phi->init_req(2, o_node);
2378
2379 // rebiased_region->in(1) is set to fast path.
2380 // The epoch of the current bias is still valid but we know
2381 // nothing about the owner; it might be set or it might be clear.
2382 Node* cmask = MakeConX(markWord::biased_lock_mask_in_place |
2383 markWord::age_mask_in_place |
2384 markWord::epoch_mask_in_place);
2385 Node* old = transform_later(new AndXNode(mark_node, cmask));
2386 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2387 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2388 old_phi->init_req(1, old);
2389 new_phi->init_req(1, new_mark);
2390
2391 transform_later(rebiased_region);
2392 transform_later(old_phi);
2393 transform_later(new_phi);
2394
2395 // Try to acquire the bias of the object using an atomic operation.
2396 // If this fails we will go in to the runtime to revoke the object's bias.
2397 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2398 transform_later(cas);
2399 proj = transform_later(new SCMemProjNode(cas));
2400
2401 // Get slow path - Failed to CAS.
2402 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2403 mem_phi->init_req(4, proj);
2404 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2405
2406 // Failed to CAS.
2407 slow_path = new RegionNode(3);
2408 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2409
2410 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2411 slow_mem->init_req(1, proj);
2412
2413 // Call CAS-based locking scheme (FastLock node).
2414
2415 transform_later(fast_lock_region);
2416 transform_later(fast_lock_mem_phi);
2417
2418 // Get slow path - FastLock failed to lock the object.
2419 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2420 mem_phi->init_req(2, fast_lock_mem_phi);
2421 // region->in(2) is set to fast path - the object is locked to the current thread.
2422
2423 slow_path->init_req(2, ctrl); // Capture slow-control
2424 slow_mem->init_req(2, fast_lock_mem_phi);
2425
2426 transform_later(slow_path);
2427 transform_later(slow_mem);
2428 // Reset lock's memory edge.
2429 lock->set_req(TypeFunc::Memory, slow_mem);
2430
2431 } else {
2432 region = new RegionNode(3);
2433 // create a Phi for the memory state
2434 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2435
2436 // Optimize test; set region slot 2
2437 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2438 mem_phi->init_req(2, mem);
2439 }
2440
2441 // Make slow path call
2442 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2443 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2444 obj, box, NULL);
2445
2446 extract_call_projections(call);
2447
2448 // Slow path can only throw asynchronous exceptions, which are always
2449 // de-opted. So the compiler thinks the slow-call can never throw an
2450 // exception. If it DOES throw an exception we would need the debug
2451 // info removed first (since if it throws there is no monitor).
2452 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2453 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2454
2455 // Capture slow path
2456 // disconnect fall-through projection from call and create a new one
2457 // hook up users of fall-through projection to region
2458 Node *slow_ctrl = _fallthroughproj->clone();
2459 transform_later(slow_ctrl);
2460 _igvn.hash_delete(_fallthroughproj);
2461 _fallthroughproj->disconnect_inputs(NULL, C);
2462 region->init_req(1, slow_ctrl);
2463 // region inputs are now complete
2464 transform_later(region);
2465 _igvn.replace_node(_fallthroughproj, region);
2466
2467 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2468 mem_phi->init_req(1, memproj );
2469 transform_later(mem_phi);
2470 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2471 }
2472
2473 //------------------------------expand_unlock_node----------------------
2474 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2475
2476 Node* ctrl = unlock->in(TypeFunc::Control);
2477 Node* mem = unlock->in(TypeFunc::Memory);
2478 Node* obj = unlock->obj_node();
2479 Node* box = unlock->box_node();
2480
2481 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2482
2483 // No need for a null check on unlock
2484
2485 // Make the merge point
2486 Node *region;
2487 Node *mem_phi;
2488
2489 if (UseOptoBiasInlining) {
2490 // Check for biased locking unlock case, which is a no-op.
2491 // See the full description in MacroAssembler::biased_locking_exit().
2492 region = new RegionNode(4);
2493 // create a Phi for the memory state
2494 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2495 mem_phi->init_req(3, mem);
2496
2497 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2498 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2499 markWord::biased_lock_mask_in_place,
2500 markWord::biased_lock_pattern);
2501 } else {
2502 region = new RegionNode(3);
2503 // create a Phi for the memory state
2504 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2505 }
2506
2507 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2508 funlock = transform_later( funlock )->as_FastUnlock();
2509 // Optimize test; set region slot 2
2510 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2511 Node *thread = transform_later(new ThreadLocalNode());
2512
2513 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2514 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2515 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2516
2517 extract_call_projections(call);
2518
2519 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2520 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2521
2522 // No exceptions for unlocking
2523 // Capture slow path
2524 // disconnect fall-through projection from call and create a new one
2525 // hook up users of fall-through projection to region
2526 Node *slow_ctrl = _fallthroughproj->clone();
2527 transform_later(slow_ctrl);
2528 _igvn.hash_delete(_fallthroughproj);
2529 _fallthroughproj->disconnect_inputs(NULL, C);
2530 region->init_req(1, slow_ctrl);
2531 // region inputs are now complete
2532 transform_later(region);
2533 _igvn.replace_node(_fallthroughproj, region);
2534
2535 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2536 mem_phi->init_req(1, memproj );
2537 mem_phi->init_req(2, mem);
2538 transform_later(mem_phi);
2539 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2540 }
2541
2542 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2543 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2544 Node* bol = check->unique_out();
2545 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2546 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2547 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2548
2549 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2550 Node* iff = bol->last_out(i);
2551 assert(iff->is_If(), "where's the if?");
2552
2553 if (iff->in(0)->is_top()) {
2554 _igvn.replace_input_of(iff, 1, C->top());
2555 continue;
2556 }
2557
2558 Node* iftrue = iff->as_If()->proj_out(1);
2559 Node* iffalse = iff->as_If()->proj_out(0);
2560 Node* ctrl = iff->in(0);
2561
2562 Node* subklass = NULL;
2563 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2564 subklass = obj_or_subklass;
2565 } else {
2566 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2567 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2568 }
2569
2570 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2571
2572 _igvn.replace_input_of(iff, 0, C->top());
2573 _igvn.replace_node(iftrue, not_subtype_ctrl);
2574 _igvn.replace_node(iffalse, ctrl);
2575 }
2576 _igvn.replace_node(check, C->top());
2577 }
2578
2579 //---------------------------eliminate_macro_nodes----------------------
2580 // Eliminate scalar replaced allocations and associated locks.
2581 void PhaseMacroExpand::eliminate_macro_nodes() {
2582 if (C->macro_count() == 0)
2583 return;
2584
2585 // First, attempt to eliminate locks
2586 int cnt = C->macro_count();
2587 for (int i=0; i < cnt; i++) {
2588 Node *n = C->macro_node(i);
2589 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2590 // Before elimination mark all associated (same box and obj)
2591 // lock and unlock nodes.
2592 mark_eliminated_locking_nodes(n->as_AbstractLock());
2593 }
2594 }
2595 bool progress = true;
2596 while (progress) {
2597 progress = false;
2598 for (int i = C->macro_count(); i > 0; i--) {
2599 Node * n = C->macro_node(i-1);
2600 bool success = false;
2601 debug_only(int old_macro_count = C->macro_count(););
2602 if (n->is_AbstractLock()) {
2603 success = eliminate_locking_node(n->as_AbstractLock());
2604 }
2605 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2606 progress = progress || success;
2607 }
2608 }
2609 // Next, attempt to eliminate allocations
2610 _has_locks = false;
2611 progress = true;
2612 while (progress) {
2613 progress = false;
2614 for (int i = C->macro_count(); i > 0; i--) {
2615 Node * n = C->macro_node(i-1);
2616 bool success = false;
2617 debug_only(int old_macro_count = C->macro_count(););
2618 switch (n->class_id()) {
2619 case Node::Class_Allocate:
2620 case Node::Class_AllocateArray:
2621 success = eliminate_allocate_node(n->as_Allocate());
2622 break;
2623 case Node::Class_CallStaticJava:
2624 success = eliminate_boxing_node(n->as_CallStaticJava());
2625 break;
2626 case Node::Class_Lock:
2627 case Node::Class_Unlock:
2628 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2629 _has_locks = true;
2630 break;
2631 case Node::Class_ArrayCopy:
2632 break;
2633 case Node::Class_OuterStripMinedLoop:
2634 break;
2635 case Node::Class_SubTypeCheck:
2636 break;
2637 default:
2638 assert(n->Opcode() == Op_LoopLimit ||
2639 n->Opcode() == Op_Opaque1 ||
2640 n->Opcode() == Op_Opaque2 ||
2641 n->Opcode() == Op_Opaque3 ||
2642 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2643 "unknown node type in macro list");
2644 }
2645 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2646 progress = progress || success;
2647 }
2648 }
2649 }
2650
2651 //------------------------------expand_macro_nodes----------------------
2652 // Returns true if a failure occurred.
2653 bool PhaseMacroExpand::expand_macro_nodes() {
2654 // Last attempt to eliminate macro nodes.
2655 eliminate_macro_nodes();
2656
2657 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2658 bool progress = true;
2659 while (progress) {
2660 progress = false;
2661 for (int i = C->macro_count(); i > 0; i--) {
2662 Node* n = C->macro_node(i-1);
2663 bool success = false;
2664 debug_only(int old_macro_count = C->macro_count(););
2665 if (n->Opcode() == Op_LoopLimit) {
2666 // Remove it from macro list and put on IGVN worklist to optimize.
2667 C->remove_macro_node(n);
2668 _igvn._worklist.push(n);
2669 success = true;
2670 } else if (n->Opcode() == Op_CallStaticJava) {
2671 // Remove it from macro list and put on IGVN worklist to optimize.
2672 C->remove_macro_node(n);
2673 _igvn._worklist.push(n);
2674 success = true;
2675 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2676 _igvn.replace_node(n, n->in(1));
2677 success = true;
2678 #if INCLUDE_RTM_OPT
2679 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2680 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2681 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2682 Node* cmp = n->unique_out();
2683 #ifdef ASSERT
2684 // Validate graph.
2685 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2686 BoolNode* bol = cmp->unique_out()->as_Bool();
2687 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2688 (bol->_test._test == BoolTest::ne), "");
2689 IfNode* ifn = bol->unique_out()->as_If();
2690 assert((ifn->outcnt() == 2) &&
2691 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2692 #endif
2693 Node* repl = n->in(1);
2694 if (!_has_locks) {
2695 // Remove RTM state check if there are no locks in the code.
2696 // Replace input to compare the same value.
2697 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2698 }
2699 _igvn.replace_node(n, repl);
2700 success = true;
2701 #endif
2702 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2703 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2704 C->remove_macro_node(n);
2705 success = true;
2706 }
2707 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2708 progress = progress || success;
2709 }
2710 }
2711
2712 // Clean up the graph so we're less likely to hit the maximum node
2713 // limit
2714 _igvn.set_delay_transform(false);
2715 _igvn.optimize();
2716 if (C->failing()) return true;
2717 _igvn.set_delay_transform(true);
2718
2719
2720 // Because we run IGVN after each expansion, some macro nodes may go
2721 // dead and be removed from the list as we iterate over it. Move
2722 // Allocate nodes (processed in a second pass) at the beginning of
2723 // the list and then iterate from the last element of the list until
2724 // an Allocate node is seen. This is robust to random deletion in
2725 // the list due to nodes going dead.
2726 C->sort_macro_nodes();
2727
2728 // expand arraycopy "macro" nodes first
2729 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2730 // before the allocate nodes are expanded.
2731 while (C->macro_count() > 0) {
2732 int macro_count = C->macro_count();
2733 Node * n = C->macro_node(macro_count-1);
2734 assert(n->is_macro(), "only macro nodes expected here");
2735 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2736 // node is unreachable, so don't try to expand it
2737 C->remove_macro_node(n);
2738 continue;
2739 }
2740 if (n->is_Allocate()) {
2741 break;
2742 }
2743 // Make sure expansion will not cause node limit to be exceeded.
2744 // Worst case is a macro node gets expanded into about 200 nodes.
2745 // Allow 50% more for optimization.
2746 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2747 return true;
2748 }
2749
2750 debug_only(int old_macro_count = C->macro_count(););
2751 switch (n->class_id()) {
2752 case Node::Class_Lock:
2753 expand_lock_node(n->as_Lock());
2754 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2755 break;
2756 case Node::Class_Unlock:
2757 expand_unlock_node(n->as_Unlock());
2758 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2759 break;
2760 case Node::Class_ArrayCopy:
2761 expand_arraycopy_node(n->as_ArrayCopy());
2762 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2763 break;
2764 case Node::Class_SubTypeCheck:
2765 expand_subtypecheck_node(n->as_SubTypeCheck());
2766 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2767 break;
2768 default:
2769 assert(false, "unknown node type in macro list");
2770 }
2771 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2772 if (C->failing()) return true;
2773
2774 // Clean up the graph so we're less likely to hit the maximum node
2775 // limit
2776 _igvn.set_delay_transform(false);
2777 _igvn.optimize();
2778 if (C->failing()) return true;
2779 _igvn.set_delay_transform(true);
2780 }
2781
2782 // All nodes except Allocate nodes are expanded now. There could be
2783 // new optimization opportunities (such as folding newly created
2784 // load from a just allocated object). Run IGVN.
2785
2786 // expand "macro" nodes
2787 // nodes are removed from the macro list as they are processed
2788 while (C->macro_count() > 0) {
2789 int macro_count = C->macro_count();
2790 Node * n = C->macro_node(macro_count-1);
2791 assert(n->is_macro(), "only macro nodes expected here");
2792 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2793 // node is unreachable, so don't try to expand it
2794 C->remove_macro_node(n);
2795 continue;
2796 }
2797 // Make sure expansion will not cause node limit to be exceeded.
2798 // Worst case is a macro node gets expanded into about 200 nodes.
2799 // Allow 50% more for optimization.
2800 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2801 return true;
2802 }
2803 switch (n->class_id()) {
2804 case Node::Class_Allocate:
2805 expand_allocate(n->as_Allocate());
2806 break;
2807 case Node::Class_AllocateArray:
2808 expand_allocate_array(n->as_AllocateArray());
2809 break;
2810 default:
2811 assert(false, "unknown node type in macro list");
2812 }
2813 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2814 if (C->failing()) return true;
2815
2816 // Clean up the graph so we're less likely to hit the maximum node
2817 // limit
2818 _igvn.set_delay_transform(false);
2819 _igvn.optimize();
2820 if (C->failing()) return true;
2821 _igvn.set_delay_transform(true);
2822 }
2823
2824 _igvn.set_delay_transform(false);
2825 return false;
2826 }