1 /*
2 * Copyright (c) 2005, 2018, 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 "opto/addnode.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/loopnode.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/node.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/phaseX.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subnode.hpp"
48 #include "opto/type.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/macros.hpp"
51 #if INCLUDE_G1GC
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #endif // INCLUDE_G1GC
54 #if INCLUDE_SHENANDOAHGC
55 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
56 #endif
57
58
59 //
60 // Replace any references to "oldref" in inputs to "use" with "newref".
61 // Returns the number of replacements made.
62 //
63 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
64 int nreplacements = 0;
65 uint req = use->req();
66 for (uint j = 0; j < use->len(); j++) {
67 Node *uin = use->in(j);
68 if (uin == oldref) {
69 if (j < req)
70 use->set_req(j, newref);
71 else
72 use->set_prec(j, newref);
73 nreplacements++;
74 } else if (j >= req && uin == NULL) {
75 break;
76 }
77 }
78 return nreplacements;
79 }
80
81 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
82 // Copy debug information and adjust JVMState information
83 uint old_dbg_start = oldcall->tf()->domain()->cnt();
84 uint new_dbg_start = newcall->tf()->domain()->cnt();
85 int jvms_adj = new_dbg_start - old_dbg_start;
86 assert (new_dbg_start == newcall->req(), "argument count mismatch");
87
88 // SafePointScalarObject node could be referenced several times in debug info.
89 // Use Dict to record cloned nodes.
90 Dict* sosn_map = new Dict(cmpkey,hashkey);
91 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
92 Node* old_in = oldcall->in(i);
93 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
94 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
95 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
96 uint old_unique = C->unique();
97 Node* new_in = old_sosn->clone(sosn_map);
98 if (old_unique != C->unique()) { // New node?
99 new_in->set_req(0, C->root()); // reset control edge
100 new_in = transform_later(new_in); // Register new node.
101 }
102 old_in = new_in;
103 }
104 newcall->add_req(old_in);
105 }
106
107 // JVMS may be shared so clone it before we modify it
108 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
109 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
110 jvms->set_map(newcall);
111 jvms->set_locoff(jvms->locoff()+jvms_adj);
112 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
113 jvms->set_monoff(jvms->monoff()+jvms_adj);
114 jvms->set_scloff(jvms->scloff()+jvms_adj);
115 jvms->set_endoff(jvms->endoff()+jvms_adj);
116 }
117 }
118
119 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
120 Node* cmp;
121 if (mask != 0) {
122 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
123 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
124 } else {
125 cmp = word;
126 }
127 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
128 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
129 transform_later(iff);
130
131 // Fast path taken.
132 Node *fast_taken = transform_later(new IfFalseNode(iff));
133
134 // Fast path not-taken, i.e. slow path
135 Node *slow_taken = transform_later(new IfTrueNode(iff));
136
137 if (return_fast_path) {
138 region->init_req(edge, slow_taken); // Capture slow-control
139 return fast_taken;
140 } else {
141 region->init_req(edge, fast_taken); // Capture fast-control
142 return slow_taken;
143 }
144 }
145
146 //--------------------copy_predefined_input_for_runtime_call--------------------
147 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
148 // Set fixed predefined input arguments
149 call->init_req( TypeFunc::Control, ctrl );
150 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
151 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
152 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
153 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
154 }
155
156 //------------------------------make_slow_call---------------------------------
157 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
158 address slow_call, const char* leaf_name, Node* slow_path,
159 Node* parm0, Node* parm1, Node* parm2) {
160
161 // Slow-path call
162 CallNode *call = leaf_name
163 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
164 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
165
166 // Slow path call has no side-effects, uses few values
167 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
168 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
169 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
170 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
171 copy_call_debug_info(oldcall, call);
172 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
173 _igvn.replace_node(oldcall, call);
174 transform_later(call);
175
176 return call;
177 }
178
179 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
180 _fallthroughproj = NULL;
181 _fallthroughcatchproj = NULL;
182 _ioproj_fallthrough = NULL;
183 _ioproj_catchall = NULL;
184 _catchallcatchproj = NULL;
185 _memproj_fallthrough = NULL;
186 _memproj_catchall = NULL;
187 _resproj = NULL;
188 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
189 ProjNode *pn = call->fast_out(i)->as_Proj();
190 switch (pn->_con) {
191 case TypeFunc::Control:
192 {
193 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
194 _fallthroughproj = pn;
195 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
196 const Node *cn = pn->fast_out(j);
197 if (cn->is_Catch()) {
198 ProjNode *cpn = NULL;
199 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
200 cpn = cn->fast_out(k)->as_Proj();
201 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
202 if (cpn->_con == CatchProjNode::fall_through_index)
203 _fallthroughcatchproj = cpn;
204 else {
205 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
206 _catchallcatchproj = cpn;
207 }
208 }
209 }
210 break;
211 }
212 case TypeFunc::I_O:
213 if (pn->_is_io_use)
214 _ioproj_catchall = pn;
215 else
216 _ioproj_fallthrough = pn;
217 break;
218 case TypeFunc::Memory:
219 if (pn->_is_io_use)
220 _memproj_catchall = pn;
221 else
222 _memproj_fallthrough = pn;
223 break;
224 case TypeFunc::Parms:
225 _resproj = pn;
226 break;
227 default:
228 assert(false, "unexpected projection from allocation node.");
229 }
230 }
231
232 }
233
234 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
235 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
236 bs->eliminate_gc_barrier(this, p2x);
237 }
238
239 // Search for a memory operation for the specified memory slice.
240 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
241 Node *orig_mem = mem;
242 Node *alloc_mem = alloc->in(TypeFunc::Memory);
243 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
244 while (true) {
245 if (mem == alloc_mem || mem == start_mem ) {
246 return mem; // hit one of our sentinels
247 } else if (mem->is_MergeMem()) {
248 mem = mem->as_MergeMem()->memory_at(alias_idx);
249 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
250 Node *in = mem->in(0);
251 // we can safely skip over safepoints, calls, locks and membars because we
252 // already know that the object is safe to eliminate.
253 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
254 return in;
255 } else if (in->is_Call()) {
256 CallNode *call = in->as_Call();
257 if (call->may_modify(tinst, phase)) {
258 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
259 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
260 return in;
261 }
262 }
263 mem = in->in(TypeFunc::Memory);
264 } else if (in->is_MemBar()) {
265 ArrayCopyNode* ac = NULL;
266 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
267 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
268 return ac;
269 }
270 mem = in->in(TypeFunc::Memory);
271 } else {
272 #ifdef ASSERT
273 in->dump();
274 mem->dump();
275 assert(false, "unexpected projection");
276 #endif
277 }
278 } else if (mem->is_Store()) {
279 const TypePtr* atype = mem->as_Store()->adr_type();
280 int adr_idx = phase->C->get_alias_index(atype);
281 if (adr_idx == alias_idx) {
282 assert(atype->isa_oopptr(), "address type must be oopptr");
283 int adr_offset = atype->offset();
284 uint adr_iid = atype->is_oopptr()->instance_id();
285 // Array elements references have the same alias_idx
286 // but different offset and different instance_id.
287 if (adr_offset == offset && adr_iid == alloc->_idx) {
288 return mem;
289 }
290 } else {
291 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
292 }
293 mem = mem->in(MemNode::Memory);
294 } else if (mem->is_ClearArray()) {
295 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
296 // Can not bypass initialization of the instance
297 // we are looking.
298 debug_only(intptr_t offset;)
299 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
300 InitializeNode* init = alloc->as_Allocate()->initialization();
301 // We are looking for stored value, return Initialize node
302 // or memory edge from Allocate node.
303 if (init != NULL) {
304 return init;
305 } else {
306 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
307 }
308 }
309 // Otherwise skip it (the call updated 'mem' value).
310 } else if (mem->Opcode() == Op_SCMemProj) {
311 mem = mem->in(0);
312 Node* adr = NULL;
313 if (mem->is_LoadStore()) {
314 adr = mem->in(MemNode::Address);
315 } else {
316 assert(mem->Opcode() == Op_EncodeISOArray ||
317 mem->Opcode() == Op_StrCompressedCopy, "sanity");
318 adr = mem->in(3); // Destination array
319 }
320 const TypePtr* atype = adr->bottom_type()->is_ptr();
321 int adr_idx = phase->C->get_alias_index(atype);
322 if (adr_idx == alias_idx) {
323 DEBUG_ONLY(mem->dump();)
324 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
325 return NULL;
326 }
327 mem = mem->in(MemNode::Memory);
328 } else if (mem->Opcode() == Op_StrInflatedCopy) {
329 Node* adr = mem->in(3); // Destination array
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 StrInflatedCopy node accesses its field");
335 return NULL;
336 }
337 mem = mem->in(MemNode::Memory);
338 } else {
339 return mem;
340 }
341 assert(mem != orig_mem, "dead memory loop");
342 }
343 }
344
345 // Generate loads from source of the arraycopy for fields of
346 // destination needed at a deoptimization point
347 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
348 BasicType bt = ft;
349 const Type *type = ftype;
350 if (ft == T_NARROWOOP) {
351 bt = T_OBJECT;
352 type = ftype->make_oopptr();
353 }
354 Node* res = NULL;
355 if (ac->is_clonebasic()) {
356 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
357 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
358 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
359 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
360 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
361 } else {
362 if (ac->modifies(offset, offset, &_igvn, true)) {
363 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
364 uint shift = exact_log2(type2aelembytes(bt));
365 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
366 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
367 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
368 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
369
370 Node* adr = NULL;
371 const TypePtr* adr_type = NULL;
372 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
373 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
374 Node* base = ac->in(ArrayCopyNode::Src);
375 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
376 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
377 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
378 // Don't emit a new load from src if src == dst but try to get the value from memory instead
379 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
380 }
381 } else {
382 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
383 #ifdef _LP64
384 diff = _igvn.transform(new ConvI2LNode(diff));
385 #endif
386 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
387
388 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
389 Node* base = ac->in(ArrayCopyNode::Src);
390 adr = _igvn.transform(new AddPNode(base, base, off));
391 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
392 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
393 // Non constant offset in the array: we can't statically
394 // determine the value
395 return NULL;
396 }
397 }
398 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
399 }
400 }
401 if (res != NULL) {
402 res = _igvn.transform(res);
403 if (ftype->isa_narrowoop()) {
404 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
405 res = _igvn.transform(new EncodePNode(res, ftype));
406 }
407 return res;
408 }
409 return NULL;
410 }
411
412 //
413 // Given a Memory Phi, compute a value Phi containing the values from stores
414 // on the input paths.
415 // Note: this function is recursive, its depth is limited by the "level" argument
416 // Returns the computed Phi, or NULL if it cannot compute it.
417 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) {
418 assert(mem->is_Phi(), "sanity");
419 int alias_idx = C->get_alias_index(adr_t);
420 int offset = adr_t->offset();
421 int instance_id = adr_t->instance_id();
422
423 // Check if an appropriate value phi already exists.
424 Node* region = mem->in(0);
425 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
426 Node* phi = region->fast_out(k);
427 if (phi->is_Phi() && phi != mem &&
428 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
429 return phi;
430 }
431 }
432 // Check if an appropriate new value phi already exists.
433 Node* new_phi = value_phis->find(mem->_idx);
434 if (new_phi != NULL)
435 return new_phi;
436
437 if (level <= 0) {
438 return NULL; // Give up: phi tree too deep
439 }
440 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
441 Node *alloc_mem = alloc->in(TypeFunc::Memory);
442
443 uint length = mem->req();
444 GrowableArray <Node *> values(length, length, NULL, false);
445
446 // create a new Phi for the value
447 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
448 transform_later(phi);
449 value_phis->push(phi, mem->_idx);
450
451 for (uint j = 1; j < length; j++) {
452 Node *in = mem->in(j);
453 if (in == NULL || in->is_top()) {
454 values.at_put(j, in);
455 } else {
456 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
457 if (val == start_mem || val == alloc_mem) {
458 // hit a sentinel, return appropriate 0 value
459 values.at_put(j, _igvn.zerocon(ft));
460 continue;
461 }
462 if (val->is_Initialize()) {
463 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
464 }
465 if (val == NULL) {
466 return NULL; // can't find a value on this path
467 }
468 if (val == mem) {
469 values.at_put(j, mem);
470 } else if (val->is_Store()) {
471 Node* n = val->in(MemNode::ValueIn);
472 #if INCLUDE_SHENANDOAHGC
473 if (UseShenandoahGC) {
474 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
475 n = bs->step_over_gc_barrier(n);
476 }
477 #endif
478 if (is_subword_type(ft)) {
479 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
480 }
481 values.at_put(j, n);
482 } else if(val->is_Proj() && val->in(0) == alloc) {
483 values.at_put(j, _igvn.zerocon(ft));
484 } else if (val->is_Phi()) {
485 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
486 if (val == NULL) {
487 return NULL;
488 }
489 values.at_put(j, val);
490 } else if (val->Opcode() == Op_SCMemProj) {
491 assert(val->in(0)->is_LoadStore() ||
492 val->in(0)->Opcode() == Op_EncodeISOArray ||
493 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
494 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
495 return NULL;
496 } else if (val->is_ArrayCopy()) {
497 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
498 if (res == NULL) {
499 return NULL;
500 }
501 values.at_put(j, res);
502 } else {
503 DEBUG_ONLY( val->dump(); )
504 assert(false, "unknown node on this path");
505 return NULL; // unknown node on this path
506 }
507 }
508 }
509 // Set Phi's inputs
510 for (uint j = 1; j < length; j++) {
511 if (values.at(j) == mem) {
512 phi->init_req(j, phi);
513 } else {
514 phi->init_req(j, values.at(j));
515 }
516 }
517 return phi;
518 }
519
520 // Search the last value stored into the object's field.
521 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
522 assert(adr_t->is_known_instance_field(), "instance required");
523 int instance_id = adr_t->instance_id();
524 assert((uint)instance_id == alloc->_idx, "wrong allocation");
525
526 int alias_idx = C->get_alias_index(adr_t);
527 int offset = adr_t->offset();
528 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
529 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
530 Node *alloc_mem = alloc->in(TypeFunc::Memory);
531 Arena *a = Thread::current()->resource_area();
532 VectorSet visited(a);
533
534 bool done = sfpt_mem == alloc_mem;
535 Node *mem = sfpt_mem;
536 while (!done) {
537 if (visited.test_set(mem->_idx)) {
538 return NULL; // found a loop, give up
539 }
540 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
541 if (mem == start_mem || mem == alloc_mem) {
542 done = true; // hit a sentinel, return appropriate 0 value
543 } else if (mem->is_Initialize()) {
544 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
545 if (mem == NULL) {
546 done = true; // Something go wrong.
547 } else if (mem->is_Store()) {
548 const TypePtr* atype = mem->as_Store()->adr_type();
549 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
550 done = true;
551 }
552 } else if (mem->is_Store()) {
553 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
554 assert(atype != NULL, "address type must be oopptr");
555 assert(C->get_alias_index(atype) == alias_idx &&
556 atype->is_known_instance_field() && atype->offset() == offset &&
557 atype->instance_id() == instance_id, "store is correct memory slice");
558 done = true;
559 } else if (mem->is_Phi()) {
560 // try to find a phi's unique input
561 Node *unique_input = NULL;
562 Node *top = C->top();
563 for (uint i = 1; i < mem->req(); i++) {
564 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
565 if (n == NULL || n == top || n == mem) {
566 continue;
567 } else if (unique_input == NULL) {
568 unique_input = n;
569 } else if (unique_input != n) {
570 unique_input = top;
571 break;
572 }
573 }
574 if (unique_input != NULL && unique_input != top) {
575 mem = unique_input;
576 } else {
577 done = true;
578 }
579 } else if (mem->is_ArrayCopy()) {
580 done = true;
581 } else {
582 DEBUG_ONLY( mem->dump(); )
583 assert(false, "unexpected node");
584 }
585 }
586 if (mem != NULL) {
587 if (mem == start_mem || mem == alloc_mem) {
588 // hit a sentinel, return appropriate 0 value
589 return _igvn.zerocon(ft);
590 } else if (mem->is_Store()) {
591 #if INCLUDE_SHENANDOAHGC
592 Node* n = mem->in(MemNode::ValueIn);
593 if (UseShenandoahGC) {
594 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
595 n = bs->step_over_gc_barrier(n);
596 }
597 return n;
598 #else
599 return mem->in(MemNode::ValueIn);
600 #endif
601 } else if (mem->is_Phi()) {
602 // attempt to produce a Phi reflecting the values on the input paths of the Phi
603 Node_Stack value_phis(a, 8);
604 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
605 if (phi != NULL) {
606 return phi;
607 } else {
608 // Kill all new Phis
609 while(value_phis.is_nonempty()) {
610 Node* n = value_phis.node();
611 _igvn.replace_node(n, C->top());
612 value_phis.pop();
613 }
614 }
615 } else if (mem->is_ArrayCopy()) {
616 Node* ctl = mem->in(0);
617 Node* m = mem->in(TypeFunc::Memory);
618 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
619 // pin the loads in the uncommon trap path
620 ctl = sfpt_ctl;
621 m = sfpt_mem;
622 }
623 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
624 }
625 }
626 // Something go wrong.
627 return NULL;
628 }
629
630 // Check the possibility of scalar replacement.
631 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
632 // Scan the uses of the allocation to check for anything that would
633 // prevent us from eliminating it.
634 NOT_PRODUCT( const char* fail_eliminate = NULL; )
635 DEBUG_ONLY( Node* disq_node = NULL; )
636 bool can_eliminate = true;
637
638 Node* res = alloc->result_cast();
639 const TypeOopPtr* res_type = NULL;
640 if (res == NULL) {
641 // All users were eliminated.
642 } else if (!res->is_CheckCastPP()) {
643 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
644 can_eliminate = false;
645 } else {
646 res_type = _igvn.type(res)->isa_oopptr();
647 if (res_type == NULL) {
648 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
649 can_eliminate = false;
650 } else if (res_type->isa_aryptr()) {
651 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
652 if (length < 0) {
653 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
654 can_eliminate = false;
655 }
656 }
657 }
658
659 if (can_eliminate && res != NULL) {
660 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
661 j < jmax && can_eliminate; j++) {
662 Node* use = res->fast_out(j);
663
664 if (use->is_AddP()) {
665 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
666 int offset = addp_type->offset();
667
668 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
669 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
670 can_eliminate = false;
671 break;
672 }
673 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
674 k < kmax && can_eliminate; k++) {
675 Node* n = use->fast_out(k);
676 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
677 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
678 !(n->is_ArrayCopy() &&
679 n->as_ArrayCopy()->is_clonebasic() &&
680 n->in(ArrayCopyNode::Dest) == use)) {
681 DEBUG_ONLY(disq_node = n;)
682 if (n->is_Load() || n->is_LoadStore()) {
683 NOT_PRODUCT(fail_eliminate = "Field load";)
684 } else {
685 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
686 }
687 can_eliminate = false;
688 }
689 }
690 } else if (use->is_ArrayCopy() &&
691 (use->as_ArrayCopy()->is_arraycopy_validated() ||
692 use->as_ArrayCopy()->is_copyof_validated() ||
693 use->as_ArrayCopy()->is_copyofrange_validated()) &&
694 use->in(ArrayCopyNode::Dest) == res) {
695 // ok to eliminate
696 } else if (use->is_SafePoint()) {
697 SafePointNode* sfpt = use->as_SafePoint();
698 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
699 // Object is passed as argument.
700 DEBUG_ONLY(disq_node = use;)
701 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
702 can_eliminate = false;
703 }
704 Node* sfptMem = sfpt->memory();
705 if (sfptMem == NULL || sfptMem->is_top()) {
706 DEBUG_ONLY(disq_node = use;)
707 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
708 can_eliminate = false;
709 } else {
710 safepoints.append_if_missing(sfpt);
711 }
712 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
713 if (use->is_Phi()) {
714 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
715 NOT_PRODUCT(fail_eliminate = "Object is return value";)
716 } else {
717 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
718 }
719 DEBUG_ONLY(disq_node = use;)
720 } else {
721 if (use->Opcode() == Op_Return) {
722 NOT_PRODUCT(fail_eliminate = "Object is return value";)
723 }else {
724 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
725 }
726 DEBUG_ONLY(disq_node = use;)
727 }
728 can_eliminate = false;
729 }
730 }
731 }
732
733 #ifndef PRODUCT
734 if (PrintEliminateAllocations) {
735 if (can_eliminate) {
736 tty->print("Scalar ");
737 if (res == NULL)
738 alloc->dump();
739 else
740 res->dump();
741 } else if (alloc->_is_scalar_replaceable) {
742 tty->print("NotScalar (%s)", fail_eliminate);
743 if (res == NULL)
744 alloc->dump();
745 else
746 res->dump();
747 #ifdef ASSERT
748 if (disq_node != NULL) {
749 tty->print(" >>>> ");
750 disq_node->dump();
751 }
752 #endif /*ASSERT*/
753 }
754 }
755 #endif
756 return can_eliminate;
757 }
758
759 // Do scalar replacement.
760 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
761 GrowableArray <SafePointNode *> safepoints_done;
762
763 ciKlass* klass = NULL;
764 ciInstanceKlass* iklass = NULL;
765 int nfields = 0;
766 int array_base = 0;
767 int element_size = 0;
768 BasicType basic_elem_type = T_ILLEGAL;
769 ciType* elem_type = NULL;
770
771 Node* res = alloc->result_cast();
772 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
773 const TypeOopPtr* res_type = NULL;
774 if (res != NULL) { // Could be NULL when there are no users
775 res_type = _igvn.type(res)->isa_oopptr();
776 }
777
778 if (res != NULL) {
779 klass = res_type->klass();
780 if (res_type->isa_instptr()) {
781 // find the fields of the class which will be needed for safepoint debug information
782 assert(klass->is_instance_klass(), "must be an instance klass.");
783 iklass = klass->as_instance_klass();
784 nfields = iklass->nof_nonstatic_fields();
785 } else {
786 // find the array's elements which will be needed for safepoint debug information
787 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
788 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
789 elem_type = klass->as_array_klass()->element_type();
790 basic_elem_type = elem_type->basic_type();
791 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
792 element_size = type2aelembytes(basic_elem_type);
793 }
794 }
795 //
796 // Process the safepoint uses
797 //
798 while (safepoints.length() > 0) {
799 SafePointNode* sfpt = safepoints.pop();
800 Node* mem = sfpt->memory();
801 Node* ctl = sfpt->control();
802 assert(sfpt->jvms() != NULL, "missed JVMS");
803 // Fields of scalar objs are referenced only at the end
804 // of regular debuginfo at the last (youngest) JVMS.
805 // Record relative start index.
806 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
807 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
808 #ifdef ASSERT
809 alloc,
810 #endif
811 first_ind, nfields);
812 sobj->init_req(0, C->root());
813 transform_later(sobj);
814
815 // Scan object's fields adding an input to the safepoint for each field.
816 for (int j = 0; j < nfields; j++) {
817 intptr_t offset;
818 ciField* field = NULL;
819 if (iklass != NULL) {
820 field = iklass->nonstatic_field_at(j);
821 offset = field->offset();
822 elem_type = field->type();
823 basic_elem_type = field->layout_type();
824 } else {
825 offset = array_base + j * (intptr_t)element_size;
826 }
827
828 const Type *field_type;
829 // The next code is taken from Parse::do_get_xxx().
830 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
831 if (!elem_type->is_loaded()) {
832 field_type = TypeInstPtr::BOTTOM;
833 } else if (field != NULL && field->is_static_constant()) {
834 // This can happen if the constant oop is non-perm.
835 ciObject* con = field->constant_value().as_object();
836 // Do not "join" in the previous type; it doesn't add value,
837 // and may yield a vacuous result if the field is of interface type.
838 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
839 assert(field_type != NULL, "field singleton type must be consistent");
840 } else {
841 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
842 }
843 if (UseCompressedOops) {
844 field_type = field_type->make_narrowoop();
845 basic_elem_type = T_NARROWOOP;
846 }
847 } else {
848 field_type = Type::get_const_basic_type(basic_elem_type);
849 }
850
851 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
852
853 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
854 if (field_val == NULL) {
855 // We weren't able to find a value for this field,
856 // give up on eliminating this allocation.
857
858 // Remove any extra entries we added to the safepoint.
859 uint last = sfpt->req() - 1;
860 for (int k = 0; k < j; k++) {
861 sfpt->del_req(last--);
862 }
863 _igvn._worklist.push(sfpt);
864 // rollback processed safepoints
865 while (safepoints_done.length() > 0) {
866 SafePointNode* sfpt_done = safepoints_done.pop();
867 // remove any extra entries we added to the safepoint
868 last = sfpt_done->req() - 1;
869 for (int k = 0; k < nfields; k++) {
870 sfpt_done->del_req(last--);
871 }
872 JVMState *jvms = sfpt_done->jvms();
873 jvms->set_endoff(sfpt_done->req());
874 // Now make a pass over the debug information replacing any references
875 // to SafePointScalarObjectNode with the allocated object.
876 int start = jvms->debug_start();
877 int end = jvms->debug_end();
878 for (int i = start; i < end; i++) {
879 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
880 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
881 if (scobj->first_index(jvms) == sfpt_done->req() &&
882 scobj->n_fields() == (uint)nfields) {
883 assert(scobj->alloc() == alloc, "sanity");
884 sfpt_done->set_req(i, res);
885 }
886 }
887 }
888 _igvn._worklist.push(sfpt_done);
889 }
890 #ifndef PRODUCT
891 if (PrintEliminateAllocations) {
892 if (field != NULL) {
893 tty->print("=== At SafePoint node %d can't find value of Field: ",
894 sfpt->_idx);
895 field->print();
896 int field_idx = C->get_alias_index(field_addr_type);
897 tty->print(" (alias_idx=%d)", field_idx);
898 } else { // Array's element
899 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
900 sfpt->_idx, j);
901 }
902 tty->print(", which prevents elimination of: ");
903 if (res == NULL)
904 alloc->dump();
905 else
906 res->dump();
907 }
908 #endif
909 return false;
910 }
911 if (UseCompressedOops && field_type->isa_narrowoop()) {
912 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
913 // to be able scalar replace the allocation.
914 if (field_val->is_EncodeP()) {
915 field_val = field_val->in(1);
916 } else {
917 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
918 }
919 }
920 sfpt->add_req(field_val);
921 }
922 JVMState *jvms = sfpt->jvms();
923 jvms->set_endoff(sfpt->req());
924 // Now make a pass over the debug information replacing any references
925 // to the allocated object with "sobj"
926 int start = jvms->debug_start();
927 int end = jvms->debug_end();
928 sfpt->replace_edges_in_range(res, sobj, start, end);
929 _igvn._worklist.push(sfpt);
930 safepoints_done.append_if_missing(sfpt); // keep it for rollback
931 }
932 return true;
933 }
934
935 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
936 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
937 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
938 if (ctl_proj != NULL) {
939 igvn.replace_node(ctl_proj, n->in(0));
940 }
941 if (mem_proj != NULL) {
942 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
943 }
944 }
945
946 // Process users of eliminated allocation.
947 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
948 Node* res = alloc->result_cast();
949 if (res != NULL) {
950 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
951 Node *use = res->last_out(j);
952 uint oc1 = res->outcnt();
953
954 if (use->is_AddP()) {
955 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
956 Node *n = use->last_out(k);
957 uint oc2 = use->outcnt();
958 if (n->is_Store()) {
959 #ifdef ASSERT
960 // Verify that there is no dependent MemBarVolatile nodes,
961 // they should be removed during IGVN, see MemBarNode::Ideal().
962 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
963 p < pmax; p++) {
964 Node* mb = n->fast_out(p);
965 assert(mb->is_Initialize() || !mb->is_MemBar() ||
966 mb->req() <= MemBarNode::Precedent ||
967 mb->in(MemBarNode::Precedent) != n,
968 "MemBarVolatile should be eliminated for non-escaping object");
969 }
970 #endif
971 _igvn.replace_node(n, n->in(MemNode::Memory));
972 } else if (n->is_ArrayCopy()) {
973 // Disconnect ArrayCopy node
974 ArrayCopyNode* ac = n->as_ArrayCopy();
975 assert(ac->is_clonebasic(), "unexpected array copy kind");
976 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
977 disconnect_projections(ac, _igvn);
978 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
979 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
980 disconnect_projections(membar_before->as_MemBar(), _igvn);
981 if (membar_after->is_MemBar()) {
982 disconnect_projections(membar_after->as_MemBar(), _igvn);
983 }
984 } else {
985 eliminate_gc_barrier(n);
986 }
987 k -= (oc2 - use->outcnt());
988 }
989 _igvn.remove_dead_node(use);
990 } else if (use->is_ArrayCopy()) {
991 // Disconnect ArrayCopy node
992 ArrayCopyNode* ac = use->as_ArrayCopy();
993 assert(ac->is_arraycopy_validated() ||
994 ac->is_copyof_validated() ||
995 ac->is_copyofrange_validated(), "unsupported");
996 CallProjections callprojs;
997 ac->extract_projections(&callprojs, true);
998
999 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1000 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1001 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1002
1003 // Set control to top. IGVN will remove the remaining projections
1004 ac->set_req(0, top());
1005 ac->replace_edge(res, top());
1006
1007 // Disconnect src right away: it can help find new
1008 // opportunities for allocation elimination
1009 Node* src = ac->in(ArrayCopyNode::Src);
1010 ac->replace_edge(src, top());
1011 // src can be top at this point if src and dest of the
1012 // arraycopy were the same
1013 if (src->outcnt() == 0 && !src->is_top()) {
1014 _igvn.remove_dead_node(src);
1015 }
1016
1017 _igvn._worklist.push(ac);
1018 } else {
1019 eliminate_gc_barrier(use);
1020 }
1021 j -= (oc1 - res->outcnt());
1022 }
1023 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1024 _igvn.remove_dead_node(res);
1025 }
1026
1027 //
1028 // Process other users of allocation's projections
1029 //
1030 if (_resproj != NULL && _resproj->outcnt() != 0) {
1031 // First disconnect stores captured by Initialize node.
1032 // If Initialize node is eliminated first in the following code,
1033 // it will kill such stores and DUIterator_Last will assert.
1034 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1035 Node *use = _resproj->fast_out(j);
1036 if (use->is_AddP()) {
1037 // raw memory addresses used only by the initialization
1038 _igvn.replace_node(use, C->top());
1039 --j; --jmax;
1040 }
1041 }
1042 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1043 Node *use = _resproj->last_out(j);
1044 uint oc1 = _resproj->outcnt();
1045 if (use->is_Initialize()) {
1046 // Eliminate Initialize node.
1047 InitializeNode *init = use->as_Initialize();
1048 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1049 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1050 if (ctrl_proj != NULL) {
1051 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1052 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1053 }
1054 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1055 if (mem_proj != NULL) {
1056 Node *mem = init->in(TypeFunc::Memory);
1057 #ifdef ASSERT
1058 if (mem->is_MergeMem()) {
1059 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1060 } else {
1061 assert(mem == _memproj_fallthrough, "allocation memory projection");
1062 }
1063 #endif
1064 _igvn.replace_node(mem_proj, mem);
1065 }
1066 } else {
1067 assert(false, "only Initialize or AddP expected");
1068 }
1069 j -= (oc1 - _resproj->outcnt());
1070 }
1071 }
1072 if (_fallthroughcatchproj != NULL) {
1073 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1074 }
1075 if (_memproj_fallthrough != NULL) {
1076 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1077 }
1078 if (_memproj_catchall != NULL) {
1079 _igvn.replace_node(_memproj_catchall, C->top());
1080 }
1081 if (_ioproj_fallthrough != NULL) {
1082 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1083 }
1084 if (_ioproj_catchall != NULL) {
1085 _igvn.replace_node(_ioproj_catchall, C->top());
1086 }
1087 if (_catchallcatchproj != NULL) {
1088 _igvn.replace_node(_catchallcatchproj, C->top());
1089 }
1090 }
1091
1092 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1093 // Don't do scalar replacement if the frame can be popped by JVMTI:
1094 // if reallocation fails during deoptimization we'll pop all
1095 // interpreter frames for this compiled frame and that won't play
1096 // nice with JVMTI popframe.
1097 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1098 return false;
1099 }
1100 Node* klass = alloc->in(AllocateNode::KlassNode);
1101 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1102 Node* res = alloc->result_cast();
1103 // Eliminate boxing allocations which are not used
1104 // regardless scalar replacable status.
1105 bool boxing_alloc = C->eliminate_boxing() &&
1106 tklass->klass()->is_instance_klass() &&
1107 tklass->klass()->as_instance_klass()->is_box_klass();
1108 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1109 return false;
1110 }
1111
1112 extract_call_projections(alloc);
1113
1114 GrowableArray <SafePointNode *> safepoints;
1115 if (!can_eliminate_allocation(alloc, safepoints)) {
1116 return false;
1117 }
1118
1119 if (!alloc->_is_scalar_replaceable) {
1120 assert(res == NULL, "sanity");
1121 // We can only eliminate allocation if all debug info references
1122 // are already replaced with SafePointScalarObject because
1123 // we can't search for a fields value without instance_id.
1124 if (safepoints.length() > 0) {
1125 return false;
1126 }
1127 }
1128
1129 if (!scalar_replacement(alloc, safepoints)) {
1130 return false;
1131 }
1132
1133 CompileLog* log = C->log();
1134 if (log != NULL) {
1135 log->head("eliminate_allocation type='%d'",
1136 log->identify(tklass->klass()));
1137 JVMState* p = alloc->jvms();
1138 while (p != NULL) {
1139 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1140 p = p->caller();
1141 }
1142 log->tail("eliminate_allocation");
1143 }
1144
1145 process_users_of_allocation(alloc);
1146
1147 #ifndef PRODUCT
1148 if (PrintEliminateAllocations) {
1149 if (alloc->is_AllocateArray())
1150 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1151 else
1152 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1153 }
1154 #endif
1155
1156 return true;
1157 }
1158
1159 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1160 // EA should remove all uses of non-escaping boxing node.
1161 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1162 return false;
1163 }
1164
1165 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1166
1167 extract_call_projections(boxing);
1168
1169 const TypeTuple* r = boxing->tf()->range();
1170 assert(r->cnt() > TypeFunc::Parms, "sanity");
1171 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1172 assert(t != NULL, "sanity");
1173
1174 CompileLog* log = C->log();
1175 if (log != NULL) {
1176 log->head("eliminate_boxing type='%d'",
1177 log->identify(t->klass()));
1178 JVMState* p = boxing->jvms();
1179 while (p != NULL) {
1180 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1181 p = p->caller();
1182 }
1183 log->tail("eliminate_boxing");
1184 }
1185
1186 process_users_of_allocation(boxing);
1187
1188 #ifndef PRODUCT
1189 if (PrintEliminateAllocations) {
1190 tty->print("++++ Eliminated: %d ", boxing->_idx);
1191 boxing->method()->print_short_name(tty);
1192 tty->cr();
1193 }
1194 #endif
1195
1196 return true;
1197 }
1198
1199 //---------------------------set_eden_pointers-------------------------
1200 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1201 if (UseTLAB) { // Private allocation: load from TLS
1202 Node* thread = transform_later(new ThreadLocalNode());
1203 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1204 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1205 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1206 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1207 } else { // Shared allocation: load from globals
1208 CollectedHeap* ch = Universe::heap();
1209 address top_adr = (address)ch->top_addr();
1210 address end_adr = (address)ch->end_addr();
1211 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1212 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1213 }
1214 }
1215
1216
1217 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1218 Node* adr = basic_plus_adr(base, offset);
1219 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1220 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1221 transform_later(value);
1222 return value;
1223 }
1224
1225
1226 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1227 Node* adr = basic_plus_adr(base, offset);
1228 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1229 transform_later(mem);
1230 return mem;
1231 }
1232
1233 //=============================================================================
1234 //
1235 // A L L O C A T I O N
1236 //
1237 // Allocation attempts to be fast in the case of frequent small objects.
1238 // It breaks down like this:
1239 //
1240 // 1) Size in doublewords is computed. This is a constant for objects and
1241 // variable for most arrays. Doubleword units are used to avoid size
1242 // overflow of huge doubleword arrays. We need doublewords in the end for
1243 // rounding.
1244 //
1245 // 2) Size is checked for being 'too large'. Too-large allocations will go
1246 // the slow path into the VM. The slow path can throw any required
1247 // exceptions, and does all the special checks for very large arrays. The
1248 // size test can constant-fold away for objects. For objects with
1249 // finalizers it constant-folds the otherway: you always go slow with
1250 // finalizers.
1251 //
1252 // 3) If NOT using TLABs, this is the contended loop-back point.
1253 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1254 //
1255 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1256 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1257 // "size*8" we always enter the VM, where "largish" is a constant picked small
1258 // enough that there's always space between the eden max and 4Gig (old space is
1259 // there so it's quite large) and large enough that the cost of entering the VM
1260 // is dwarfed by the cost to initialize the space.
1261 //
1262 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1263 // down. If contended, repeat at step 3. If using TLABs normal-store
1264 // adjusted heap top back down; there is no contention.
1265 //
1266 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1267 // fields.
1268 //
1269 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1270 // oop flavor.
1271 //
1272 //=============================================================================
1273 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1274 // Allocations bigger than this always go the slow route.
1275 // This value must be small enough that allocation attempts that need to
1276 // trigger exceptions go the slow route. Also, it must be small enough so
1277 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1278 //=============================================================================j//
1279 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1280 // The allocator will coalesce int->oop copies away. See comment in
1281 // coalesce.cpp about how this works. It depends critically on the exact
1282 // code shape produced here, so if you are changing this code shape
1283 // make sure the GC info for the heap-top is correct in and around the
1284 // slow-path call.
1285 //
1286
1287 void PhaseMacroExpand::expand_allocate_common(
1288 AllocateNode* alloc, // allocation node to be expanded
1289 Node* length, // array length for an array allocation
1290 const TypeFunc* slow_call_type, // Type of slow call
1291 address slow_call_address // Address of slow call
1292 )
1293 {
1294
1295 Node* ctrl = alloc->in(TypeFunc::Control);
1296 Node* mem = alloc->in(TypeFunc::Memory);
1297 Node* i_o = alloc->in(TypeFunc::I_O);
1298 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1299 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1300 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1301
1302 assert(ctrl != NULL, "must have control");
1303 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1304 // they will not be used if "always_slow" is set
1305 enum { slow_result_path = 1, fast_result_path = 2 };
1306 Node *result_region = NULL;
1307 Node *result_phi_rawmem = NULL;
1308 Node *result_phi_rawoop = NULL;
1309 Node *result_phi_i_o = NULL;
1310
1311 // The initial slow comparison is a size check, the comparison
1312 // we want to do is a BoolTest::gt
1313 bool always_slow = false;
1314 int tv = _igvn.find_int_con(initial_slow_test, -1);
1315 if (tv >= 0) {
1316 always_slow = (tv == 1);
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 always_slow = true;
1326 initial_slow_test = NULL;
1327 }
1328
1329
1330 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1331 Node *slow_region = NULL;
1332 Node *toobig_false = ctrl;
1333
1334 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1335 // generate the initial test if necessary
1336 if (initial_slow_test != NULL ) {
1337 slow_region = new RegionNode(3);
1338
1339 // Now make the initial failure test. Usually a too-big test but
1340 // might be a TRUE for finalizers or a fancy class check for
1341 // newInstance0.
1342 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1343 transform_later(toobig_iff);
1344 // Plug the failing-too-big test into the slow-path region
1345 Node *toobig_true = new IfTrueNode( toobig_iff );
1346 transform_later(toobig_true);
1347 slow_region ->init_req( too_big_or_final_path, toobig_true );
1348 toobig_false = new IfFalseNode( toobig_iff );
1349 transform_later(toobig_false);
1350 } else { // No initial test, just fall into next case
1351 toobig_false = ctrl;
1352 debug_only(slow_region = NodeSentinel);
1353 }
1354
1355 Node *slow_mem = mem; // save the current memory state for slow path
1356 // generate the fast allocation code unless we know that the initial test will always go slow
1357 if (!always_slow) {
1358 // Fast path modifies only raw memory.
1359 if (mem->is_MergeMem()) {
1360 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1361 }
1362
1363 Node* eden_top_adr;
1364 Node* eden_end_adr;
1365
1366 set_eden_pointers(eden_top_adr, eden_end_adr);
1367
1368 // Load Eden::end. Loop invariant and hoisted.
1369 //
1370 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1371 // a TLAB to work around a bug where these values were being moved across
1372 // a safepoint. These are not oops, so they cannot be include in the oop
1373 // map, but they can be changed by a GC. The proper way to fix this would
1374 // be to set the raw memory state when generating a SafepointNode. However
1375 // this will require extensive changes to the loop optimization in order to
1376 // prevent a degradation of the optimization.
1377 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1378 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1379
1380 // allocate the Region and Phi nodes for the result
1381 result_region = new RegionNode(3);
1382 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1383 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1384 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1385
1386 // We need a Region for the loop-back contended case.
1387 enum { fall_in_path = 1, contended_loopback_path = 2 };
1388 Node *contended_region;
1389 Node *contended_phi_rawmem;
1390 if (UseTLAB) {
1391 contended_region = toobig_false;
1392 contended_phi_rawmem = mem;
1393 } else {
1394 contended_region = new RegionNode(3);
1395 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1396 // Now handle the passing-too-big test. We fall into the contended
1397 // loop-back merge point.
1398 contended_region ->init_req(fall_in_path, toobig_false);
1399 contended_phi_rawmem->init_req(fall_in_path, mem);
1400 transform_later(contended_region);
1401 transform_later(contended_phi_rawmem);
1402 }
1403
1404 // Load(-locked) the heap top.
1405 // See note above concerning the control input when using a TLAB
1406 Node *old_eden_top = UseTLAB
1407 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1408 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1409
1410 transform_later(old_eden_top);
1411 // Add to heap top to get a new heap top
1412 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1413 transform_later(new_eden_top);
1414 // Check for needing a GC; compare against heap end
1415 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1416 transform_later(needgc_cmp);
1417 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1418 transform_later(needgc_bol);
1419 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1420 transform_later(needgc_iff);
1421
1422 // Plug the failing-heap-space-need-gc test into the slow-path region
1423 Node *needgc_true = new IfTrueNode(needgc_iff);
1424 transform_later(needgc_true);
1425 if (initial_slow_test) {
1426 slow_region->init_req(need_gc_path, needgc_true);
1427 // This completes all paths into the slow merge point
1428 transform_later(slow_region);
1429 } else { // No initial slow path needed!
1430 // Just fall from the need-GC path straight into the VM call.
1431 slow_region = needgc_true;
1432 }
1433 // No need for a GC. Setup for the Store-Conditional
1434 Node *needgc_false = new IfFalseNode(needgc_iff);
1435 transform_later(needgc_false);
1436
1437 // Grab regular I/O before optional prefetch may change it.
1438 // Slow-path does no I/O so just set it to the original I/O.
1439 result_phi_i_o->init_req(slow_result_path, i_o);
1440
1441 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1442 old_eden_top, new_eden_top, length);
1443
1444 // Name successful fast-path variables
1445 Node* fast_oop = old_eden_top;
1446 Node* fast_oop_ctrl;
1447 Node* fast_oop_rawmem;
1448
1449 // Store (-conditional) the modified eden top back down.
1450 // StorePConditional produces flags for a test PLUS a modified raw
1451 // memory state.
1452 if (UseTLAB) {
1453 Node* store_eden_top =
1454 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1455 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1456 transform_later(store_eden_top);
1457 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1458 fast_oop_rawmem = store_eden_top;
1459 } else {
1460 Node* store_eden_top =
1461 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1462 new_eden_top, fast_oop/*old_eden_top*/);
1463 transform_later(store_eden_top);
1464 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1465 transform_later(contention_check);
1466 store_eden_top = new SCMemProjNode(store_eden_top);
1467 transform_later(store_eden_top);
1468
1469 // If not using TLABs, check to see if there was contention.
1470 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1471 transform_later(contention_iff);
1472 Node *contention_true = new IfTrueNode(contention_iff);
1473 transform_later(contention_true);
1474 // If contention, loopback and try again.
1475 contended_region->init_req(contended_loopback_path, contention_true);
1476 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1477
1478 // Fast-path succeeded with no contention!
1479 Node *contention_false = new IfFalseNode(contention_iff);
1480 transform_later(contention_false);
1481 fast_oop_ctrl = contention_false;
1482
1483 // Bump total allocated bytes for this thread
1484 Node* thread = new ThreadLocalNode();
1485 transform_later(thread);
1486 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1487 in_bytes(JavaThread::allocated_bytes_offset()));
1488 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1489 0, TypeLong::LONG, T_LONG);
1490 #ifdef _LP64
1491 Node* alloc_size = size_in_bytes;
1492 #else
1493 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1494 transform_later(alloc_size);
1495 #endif
1496 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1497 transform_later(new_alloc_bytes);
1498 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1499 0, new_alloc_bytes, T_LONG);
1500 }
1501
1502 InitializeNode* init = alloc->initialization();
1503 fast_oop_rawmem = initialize_object(alloc,
1504 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1505 klass_node, length, size_in_bytes);
1506
1507 // If initialization is performed by an array copy, any required
1508 // MemBarStoreStore was already added. If the object does not
1509 // escape no need for a MemBarStoreStore. If the object does not
1510 // escape in its initializer and memory barrier (MemBarStoreStore or
1511 // stronger) is already added at exit of initializer, also no need
1512 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1513 // so that stores that initialize this object can't be reordered
1514 // with a subsequent store that makes this object accessible by
1515 // other threads.
1516 // Other threads include java threads and JVM internal threads
1517 // (for example concurrent GC threads). Current concurrent GC
1518 // implementation: CMS and G1 will not scan newly created object,
1519 // so it's safe to skip storestore barrier when allocation does
1520 // not escape.
1521 if (!alloc->does_not_escape_thread() &&
1522 !alloc->is_allocation_MemBar_redundant() &&
1523 (init == NULL || !init->is_complete_with_arraycopy())) {
1524 if (init == NULL || init->req() < InitializeNode::RawStores) {
1525 // No InitializeNode or no stores captured by zeroing
1526 // elimination. Simply add the MemBarStoreStore after object
1527 // initialization.
1528 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1529 transform_later(mb);
1530
1531 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1532 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1533 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1534 transform_later(fast_oop_ctrl);
1535 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1536 transform_later(fast_oop_rawmem);
1537 } else {
1538 // Add the MemBarStoreStore after the InitializeNode so that
1539 // all stores performing the initialization that were moved
1540 // before the InitializeNode happen before the storestore
1541 // barrier.
1542
1543 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1544 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1545
1546 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1547 transform_later(mb);
1548
1549 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1550 transform_later(ctrl);
1551 Node* mem = new ProjNode(init,TypeFunc::Memory);
1552 transform_later(mem);
1553
1554 // The MemBarStoreStore depends on control and memory coming
1555 // from the InitializeNode
1556 mb->init_req(TypeFunc::Memory, mem);
1557 mb->init_req(TypeFunc::Control, ctrl);
1558
1559 ctrl = new ProjNode(mb,TypeFunc::Control);
1560 transform_later(ctrl);
1561 mem = new ProjNode(mb,TypeFunc::Memory);
1562 transform_later(mem);
1563
1564 // All nodes that depended on the InitializeNode for control
1565 // and memory must now depend on the MemBarNode that itself
1566 // depends on the InitializeNode
1567 if (init_ctrl != NULL) {
1568 _igvn.replace_node(init_ctrl, ctrl);
1569 }
1570 if (init_mem != NULL) {
1571 _igvn.replace_node(init_mem, mem);
1572 }
1573 }
1574 }
1575
1576 if (C->env()->dtrace_extended_probes()) {
1577 // Slow-path call
1578 int size = TypeFunc::Parms + 2;
1579 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1580 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1581 "dtrace_object_alloc",
1582 TypeRawPtr::BOTTOM);
1583
1584 // Get base of thread-local storage area
1585 Node* thread = new ThreadLocalNode();
1586 transform_later(thread);
1587
1588 call->init_req(TypeFunc::Parms+0, thread);
1589 call->init_req(TypeFunc::Parms+1, fast_oop);
1590 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1591 call->init_req(TypeFunc::I_O , top()); // does no i/o
1592 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1593 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1594 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1595 transform_later(call);
1596 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1597 transform_later(fast_oop_ctrl);
1598 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1599 transform_later(fast_oop_rawmem);
1600 }
1601
1602 // Plug in the successful fast-path into the result merge point
1603 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1604 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1605 result_phi_i_o ->init_req(fast_result_path, i_o);
1606 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1607 } else {
1608 slow_region = ctrl;
1609 result_phi_i_o = i_o; // Rename it to use in the following code.
1610 }
1611
1612 // Generate slow-path call
1613 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1614 OptoRuntime::stub_name(slow_call_address),
1615 alloc->jvms()->bci(),
1616 TypePtr::BOTTOM);
1617 call->init_req( TypeFunc::Control, slow_region );
1618 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1619 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1620 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1621 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1622
1623 call->init_req(TypeFunc::Parms+0, klass_node);
1624 if (length != NULL) {
1625 call->init_req(TypeFunc::Parms+1, length);
1626 }
1627
1628 // Copy debug information and adjust JVMState information, then replace
1629 // allocate node with the call
1630 copy_call_debug_info((CallNode *) alloc, call);
1631 if (!always_slow) {
1632 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1633 } else {
1634 // Hook i_o projection to avoid its elimination during allocation
1635 // replacement (when only a slow call is generated).
1636 call->set_req(TypeFunc::I_O, result_phi_i_o);
1637 }
1638 _igvn.replace_node(alloc, call);
1639 transform_later(call);
1640
1641 // Identify the output projections from the allocate node and
1642 // adjust any references to them.
1643 // The control and io projections look like:
1644 //
1645 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1646 // Allocate Catch
1647 // ^---Proj(io) <-------+ ^---CatchProj(io)
1648 //
1649 // We are interested in the CatchProj nodes.
1650 //
1651 extract_call_projections(call);
1652
1653 // An allocate node has separate memory projections for the uses on
1654 // the control and i_o paths. Replace the control memory projection with
1655 // result_phi_rawmem (unless we are only generating a slow call when
1656 // both memory projections are combined)
1657 if (!always_slow && _memproj_fallthrough != NULL) {
1658 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1659 Node *use = _memproj_fallthrough->fast_out(i);
1660 _igvn.rehash_node_delayed(use);
1661 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1662 // back up iterator
1663 --i;
1664 }
1665 }
1666 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1667 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1668 if (_memproj_catchall != NULL ) {
1669 if (_memproj_fallthrough == NULL) {
1670 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1671 transform_later(_memproj_fallthrough);
1672 }
1673 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1674 Node *use = _memproj_catchall->fast_out(i);
1675 _igvn.rehash_node_delayed(use);
1676 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1677 // back up iterator
1678 --i;
1679 }
1680 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1681 _igvn.remove_dead_node(_memproj_catchall);
1682 }
1683
1684 // An allocate node has separate i_o projections for the uses on the control
1685 // and i_o paths. Always replace the control i_o projection with result i_o
1686 // otherwise incoming i_o become dead when only a slow call is generated
1687 // (it is different from memory projections where both projections are
1688 // combined in such case).
1689 if (_ioproj_fallthrough != NULL) {
1690 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1691 Node *use = _ioproj_fallthrough->fast_out(i);
1692 _igvn.rehash_node_delayed(use);
1693 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1694 // back up iterator
1695 --i;
1696 }
1697 }
1698 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1699 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1700 if (_ioproj_catchall != NULL ) {
1701 if (_ioproj_fallthrough == NULL) {
1702 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1703 transform_later(_ioproj_fallthrough);
1704 }
1705 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1706 Node *use = _ioproj_catchall->fast_out(i);
1707 _igvn.rehash_node_delayed(use);
1708 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1709 // back up iterator
1710 --i;
1711 }
1712 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1713 _igvn.remove_dead_node(_ioproj_catchall);
1714 }
1715
1716 // if we generated only a slow call, we are done
1717 if (always_slow) {
1718 // Now we can unhook i_o.
1719 if (result_phi_i_o->outcnt() > 1) {
1720 call->set_req(TypeFunc::I_O, top());
1721 } else {
1722 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1723 // Case of new array with negative size known during compilation.
1724 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1725 // following code since call to runtime will throw exception.
1726 // As result there will be no users of i_o after the call.
1727 // Leave i_o attached to this call to avoid problems in preceding graph.
1728 }
1729 return;
1730 }
1731
1732
1733 if (_fallthroughcatchproj != NULL) {
1734 ctrl = _fallthroughcatchproj->clone();
1735 transform_later(ctrl);
1736 _igvn.replace_node(_fallthroughcatchproj, result_region);
1737 } else {
1738 ctrl = top();
1739 }
1740 Node *slow_result;
1741 if (_resproj == NULL) {
1742 // no uses of the allocation result
1743 slow_result = top();
1744 } else {
1745 slow_result = _resproj->clone();
1746 transform_later(slow_result);
1747 _igvn.replace_node(_resproj, result_phi_rawoop);
1748 }
1749
1750 // Plug slow-path into result merge point
1751 result_region ->init_req( slow_result_path, ctrl );
1752 result_phi_rawoop->init_req( slow_result_path, slow_result);
1753 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1754 transform_later(result_region);
1755 transform_later(result_phi_rawoop);
1756 transform_later(result_phi_rawmem);
1757 transform_later(result_phi_i_o);
1758 // This completes all paths into the result merge point
1759 }
1760
1761
1762 // Helper for PhaseMacroExpand::expand_allocate_common.
1763 // Initializes the newly-allocated storage.
1764 Node*
1765 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1766 Node* control, Node* rawmem, Node* object,
1767 Node* klass_node, Node* length,
1768 Node* size_in_bytes) {
1769 InitializeNode* init = alloc->initialization();
1770 // Store the klass & mark bits
1771 Node* mark_node = NULL;
1772 // For now only enable fast locking for non-array types
1773 if (UseBiasedLocking && (length == NULL)) {
1774 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1775 } else {
1776 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1777 }
1778 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1779
1780 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1781 int header_size = alloc->minimum_header_size(); // conservatively small
1782
1783 // Array length
1784 if (length != NULL) { // Arrays need length field
1785 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1786 // conservatively small header size:
1787 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1788 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1789 if (k->is_array_klass()) // we know the exact header size in most cases:
1790 header_size = Klass::layout_helper_header_size(k->layout_helper());
1791 }
1792
1793 // Clear the object body, if necessary.
1794 if (init == NULL) {
1795 // The init has somehow disappeared; be cautious and clear everything.
1796 //
1797 // This can happen if a node is allocated but an uncommon trap occurs
1798 // immediately. In this case, the Initialize gets associated with the
1799 // trap, and may be placed in a different (outer) loop, if the Allocate
1800 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1801 // there can be two Allocates to one Initialize. The answer in all these
1802 // edge cases is safety first. It is always safe to clear immediately
1803 // within an Allocate, and then (maybe or maybe not) clear some more later.
1804 if (!(UseTLAB && ZeroTLAB)) {
1805 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1806 header_size, size_in_bytes,
1807 &_igvn);
1808 }
1809 } else {
1810 if (!init->is_complete()) {
1811 // Try to win by zeroing only what the init does not store.
1812 // We can also try to do some peephole optimizations,
1813 // such as combining some adjacent subword stores.
1814 rawmem = init->complete_stores(control, rawmem, object,
1815 header_size, size_in_bytes, &_igvn);
1816 }
1817 // We have no more use for this link, since the AllocateNode goes away:
1818 init->set_req(InitializeNode::RawAddress, top());
1819 // (If we keep the link, it just confuses the register allocator,
1820 // who thinks he sees a real use of the address by the membar.)
1821 }
1822
1823 return rawmem;
1824 }
1825
1826 // Generate prefetch instructions for next allocations.
1827 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1828 Node*& contended_phi_rawmem,
1829 Node* old_eden_top, Node* new_eden_top,
1830 Node* length) {
1831 enum { fall_in_path = 1, pf_path = 2 };
1832 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1833 // Generate prefetch allocation with watermark check.
1834 // As an allocation hits the watermark, we will prefetch starting
1835 // at a "distance" away from watermark.
1836
1837 Node *pf_region = new RegionNode(3);
1838 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1839 TypeRawPtr::BOTTOM );
1840 // I/O is used for Prefetch
1841 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1842
1843 Node *thread = new ThreadLocalNode();
1844 transform_later(thread);
1845
1846 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1847 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1848 transform_later(eden_pf_adr);
1849
1850 Node *old_pf_wm = new LoadPNode(needgc_false,
1851 contended_phi_rawmem, eden_pf_adr,
1852 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1853 MemNode::unordered);
1854 transform_later(old_pf_wm);
1855
1856 // check against new_eden_top
1857 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1858 transform_later(need_pf_cmp);
1859 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1860 transform_later(need_pf_bol);
1861 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1862 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1863 transform_later(need_pf_iff);
1864
1865 // true node, add prefetchdistance
1866 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1867 transform_later(need_pf_true);
1868
1869 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1870 transform_later(need_pf_false);
1871
1872 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1873 _igvn.MakeConX(AllocatePrefetchDistance) );
1874 transform_later(new_pf_wmt );
1875 new_pf_wmt->set_req(0, need_pf_true);
1876
1877 Node *store_new_wmt = new StorePNode(need_pf_true,
1878 contended_phi_rawmem, eden_pf_adr,
1879 TypeRawPtr::BOTTOM, new_pf_wmt,
1880 MemNode::unordered);
1881 transform_later(store_new_wmt);
1882
1883 // adding prefetches
1884 pf_phi_abio->init_req( fall_in_path, i_o );
1885
1886 Node *prefetch_adr;
1887 Node *prefetch;
1888 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1889 uint step_size = AllocatePrefetchStepSize;
1890 uint distance = 0;
1891
1892 for ( uint i = 0; i < lines; i++ ) {
1893 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1894 _igvn.MakeConX(distance) );
1895 transform_later(prefetch_adr);
1896 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1897 transform_later(prefetch);
1898 distance += step_size;
1899 i_o = prefetch;
1900 }
1901 pf_phi_abio->set_req( pf_path, i_o );
1902
1903 pf_region->init_req( fall_in_path, need_pf_false );
1904 pf_region->init_req( pf_path, need_pf_true );
1905
1906 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1907 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1908
1909 transform_later(pf_region);
1910 transform_later(pf_phi_rawmem);
1911 transform_later(pf_phi_abio);
1912
1913 needgc_false = pf_region;
1914 contended_phi_rawmem = pf_phi_rawmem;
1915 i_o = pf_phi_abio;
1916 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1917 // Insert a prefetch instruction for each allocation.
1918 // This code is used to generate 1 prefetch instruction per cache line.
1919
1920 // Generate several prefetch instructions.
1921 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1922 uint step_size = AllocatePrefetchStepSize;
1923 uint distance = AllocatePrefetchDistance;
1924
1925 // Next cache address.
1926 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1927 _igvn.MakeConX(step_size + distance));
1928 transform_later(cache_adr);
1929 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1930 transform_later(cache_adr);
1931 // Address is aligned to execute prefetch to the beginning of cache line size
1932 // (it is important when BIS instruction is used on SPARC as prefetch).
1933 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1934 cache_adr = new AndXNode(cache_adr, mask);
1935 transform_later(cache_adr);
1936 cache_adr = new CastX2PNode(cache_adr);
1937 transform_later(cache_adr);
1938
1939 // Prefetch
1940 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1941 prefetch->set_req(0, needgc_false);
1942 transform_later(prefetch);
1943 contended_phi_rawmem = prefetch;
1944 Node *prefetch_adr;
1945 distance = step_size;
1946 for ( uint i = 1; i < lines; i++ ) {
1947 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1948 _igvn.MakeConX(distance) );
1949 transform_later(prefetch_adr);
1950 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1951 transform_later(prefetch);
1952 distance += step_size;
1953 contended_phi_rawmem = prefetch;
1954 }
1955 } else if( AllocatePrefetchStyle > 0 ) {
1956 // Insert a prefetch for each allocation only on the fast-path
1957 Node *prefetch_adr;
1958 Node *prefetch;
1959 // Generate several prefetch instructions.
1960 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1961 uint step_size = AllocatePrefetchStepSize;
1962 uint distance = AllocatePrefetchDistance;
1963 for ( uint i = 0; i < lines; i++ ) {
1964 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1965 _igvn.MakeConX(distance) );
1966 transform_later(prefetch_adr);
1967 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1968 // Do not let it float too high, since if eden_top == eden_end,
1969 // both might be null.
1970 if( i == 0 ) { // Set control for first prefetch, next follows it
1971 prefetch->init_req(0, needgc_false);
1972 }
1973 transform_later(prefetch);
1974 distance += step_size;
1975 i_o = prefetch;
1976 }
1977 }
1978 return i_o;
1979 }
1980
1981
1982 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1983 expand_allocate_common(alloc, NULL,
1984 OptoRuntime::new_instance_Type(),
1985 OptoRuntime::new_instance_Java());
1986 }
1987
1988 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1989 Node* length = alloc->in(AllocateNode::ALength);
1990 InitializeNode* init = alloc->initialization();
1991 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1992 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1993 address slow_call_address; // Address of slow call
1994 if (init != NULL && init->is_complete_with_arraycopy() &&
1995 k->is_type_array_klass()) {
1996 // Don't zero type array during slow allocation in VM since
1997 // it will be initialized later by arraycopy in compiled code.
1998 slow_call_address = OptoRuntime::new_array_nozero_Java();
1999 } else {
2000 slow_call_address = OptoRuntime::new_array_Java();
2001 }
2002 expand_allocate_common(alloc, length,
2003 OptoRuntime::new_array_Type(),
2004 slow_call_address);
2005 }
2006
2007 //-------------------mark_eliminated_box----------------------------------
2008 //
2009 // During EA obj may point to several objects but after few ideal graph
2010 // transformations (CCP) it may point to only one non escaping object
2011 // (but still using phi), corresponding locks and unlocks will be marked
2012 // for elimination. Later obj could be replaced with a new node (new phi)
2013 // and which does not have escape information. And later after some graph
2014 // reshape other locks and unlocks (which were not marked for elimination
2015 // before) are connected to this new obj (phi) but they still will not be
2016 // marked for elimination since new obj has no escape information.
2017 // Mark all associated (same box and obj) lock and unlock nodes for
2018 // elimination if some of them marked already.
2019 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2020 if (oldbox->as_BoxLock()->is_eliminated())
2021 return; // This BoxLock node was processed already.
2022
2023 // New implementation (EliminateNestedLocks) has separate BoxLock
2024 // node for each locked region so mark all associated locks/unlocks as
2025 // eliminated even if different objects are referenced in one locked region
2026 // (for example, OSR compilation of nested loop inside locked scope).
2027 if (EliminateNestedLocks ||
2028 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2029 // Box is used only in one lock region. Mark this box as eliminated.
2030 _igvn.hash_delete(oldbox);
2031 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2032 _igvn.hash_insert(oldbox);
2033
2034 for (uint i = 0; i < oldbox->outcnt(); i++) {
2035 Node* u = oldbox->raw_out(i);
2036 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2037 AbstractLockNode* alock = u->as_AbstractLock();
2038 // Check lock's box since box could be referenced by Lock's debug info.
2039 if (alock->box_node() == oldbox) {
2040 // Mark eliminated all related locks and unlocks.
2041 #ifdef ASSERT
2042 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2043 #endif
2044 alock->set_non_esc_obj();
2045 }
2046 }
2047 }
2048 return;
2049 }
2050
2051 // Create new "eliminated" BoxLock node and use it in monitor debug info
2052 // instead of oldbox for the same object.
2053 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2054
2055 // Note: BoxLock node is marked eliminated only here and it is used
2056 // to indicate that all associated lock and unlock nodes are marked
2057 // for elimination.
2058 newbox->set_eliminated();
2059 transform_later(newbox);
2060
2061 // Replace old box node with new box for all users of the same object.
2062 for (uint i = 0; i < oldbox->outcnt();) {
2063 bool next_edge = true;
2064
2065 Node* u = oldbox->raw_out(i);
2066 if (u->is_AbstractLock()) {
2067 AbstractLockNode* alock = u->as_AbstractLock();
2068 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2069 // Replace Box and mark eliminated all related locks and unlocks.
2070 #ifdef ASSERT
2071 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2072 #endif
2073 alock->set_non_esc_obj();
2074 _igvn.rehash_node_delayed(alock);
2075 alock->set_box_node(newbox);
2076 next_edge = false;
2077 }
2078 }
2079 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2080 FastLockNode* flock = u->as_FastLock();
2081 assert(flock->box_node() == oldbox, "sanity");
2082 _igvn.rehash_node_delayed(flock);
2083 flock->set_box_node(newbox);
2084 next_edge = false;
2085 }
2086
2087 // Replace old box in monitor debug info.
2088 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2089 SafePointNode* sfn = u->as_SafePoint();
2090 JVMState* youngest_jvms = sfn->jvms();
2091 int max_depth = youngest_jvms->depth();
2092 for (int depth = 1; depth <= max_depth; depth++) {
2093 JVMState* jvms = youngest_jvms->of_depth(depth);
2094 int num_mon = jvms->nof_monitors();
2095 // Loop over monitors
2096 for (int idx = 0; idx < num_mon; idx++) {
2097 Node* obj_node = sfn->monitor_obj(jvms, idx);
2098 Node* box_node = sfn->monitor_box(jvms, idx);
2099 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2100 int j = jvms->monitor_box_offset(idx);
2101 _igvn.replace_input_of(u, j, newbox);
2102 next_edge = false;
2103 }
2104 }
2105 }
2106 }
2107 if (next_edge) i++;
2108 }
2109 }
2110
2111 //-----------------------mark_eliminated_locking_nodes-----------------------
2112 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2113 if (EliminateNestedLocks) {
2114 if (alock->is_nested()) {
2115 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2116 return;
2117 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2118 // Only Lock node has JVMState needed here.
2119 // Not that preceding claim is documented anywhere else.
2120 if (alock->jvms() != NULL) {
2121 if (alock->as_Lock()->is_nested_lock_region()) {
2122 // Mark eliminated related nested locks and unlocks.
2123 Node* obj = alock->obj_node();
2124 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2125 assert(!box_node->is_eliminated(), "should not be marked yet");
2126 // Note: BoxLock node is marked eliminated only here
2127 // and it is used to indicate that all associated lock
2128 // and unlock nodes are marked for elimination.
2129 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2130 for (uint i = 0; i < box_node->outcnt(); i++) {
2131 Node* u = box_node->raw_out(i);
2132 if (u->is_AbstractLock()) {
2133 alock = u->as_AbstractLock();
2134 if (alock->box_node() == box_node) {
2135 // Verify that this Box is referenced only by related locks.
2136 assert(alock->obj_node()->eqv_uncast(obj), "");
2137 // Mark all related locks and unlocks.
2138 #ifdef ASSERT
2139 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2140 #endif
2141 alock->set_nested();
2142 }
2143 }
2144 }
2145 } else {
2146 #ifdef ASSERT
2147 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2148 if (C->log() != NULL)
2149 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2150 #endif
2151 }
2152 }
2153 return;
2154 }
2155 // Process locks for non escaping object
2156 assert(alock->is_non_esc_obj(), "");
2157 } // EliminateNestedLocks
2158
2159 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2160 // Look for all locks of this object and mark them and
2161 // corresponding BoxLock nodes as eliminated.
2162 Node* obj = alock->obj_node();
2163 for (uint j = 0; j < obj->outcnt(); j++) {
2164 Node* o = obj->raw_out(j);
2165 if (o->is_AbstractLock() &&
2166 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2167 alock = o->as_AbstractLock();
2168 Node* box = alock->box_node();
2169 // Replace old box node with new eliminated box for all users
2170 // of the same object and mark related locks as eliminated.
2171 mark_eliminated_box(box, obj);
2172 }
2173 }
2174 }
2175 }
2176
2177 // we have determined that this lock/unlock can be eliminated, we simply
2178 // eliminate the node without expanding it.
2179 //
2180 // Note: The membar's associated with the lock/unlock are currently not
2181 // eliminated. This should be investigated as a future enhancement.
2182 //
2183 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2184
2185 if (!alock->is_eliminated()) {
2186 return false;
2187 }
2188 #ifdef ASSERT
2189 if (!alock->is_coarsened()) {
2190 // Check that new "eliminated" BoxLock node is created.
2191 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2192 assert(oldbox->is_eliminated(), "should be done already");
2193 }
2194 #endif
2195
2196 alock->log_lock_optimization(C, "eliminate_lock");
2197
2198 #ifndef PRODUCT
2199 if (PrintEliminateLocks) {
2200 if (alock->is_Lock()) {
2201 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2202 } else {
2203 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2204 }
2205 }
2206 #endif
2207
2208 Node* mem = alock->in(TypeFunc::Memory);
2209 Node* ctrl = alock->in(TypeFunc::Control);
2210 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2211
2212 extract_call_projections(alock);
2213 // There are 2 projections from the lock. The lock node will
2214 // be deleted when its last use is subsumed below.
2215 assert(alock->outcnt() == 2 &&
2216 _fallthroughproj != NULL &&
2217 _memproj_fallthrough != NULL,
2218 "Unexpected projections from Lock/Unlock");
2219
2220 Node* fallthroughproj = _fallthroughproj;
2221 Node* memproj_fallthrough = _memproj_fallthrough;
2222
2223 // The memory projection from a lock/unlock is RawMem
2224 // The input to a Lock is merged memory, so extract its RawMem input
2225 // (unless the MergeMem has been optimized away.)
2226 if (alock->is_Lock()) {
2227 // Seach for MemBarAcquireLock node and delete it also.
2228 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2229 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2230 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2231 Node* memproj = membar->proj_out(TypeFunc::Memory);
2232 _igvn.replace_node(ctrlproj, fallthroughproj);
2233 _igvn.replace_node(memproj, memproj_fallthrough);
2234
2235 // Delete FastLock node also if this Lock node is unique user
2236 // (a loop peeling may clone a Lock node).
2237 Node* flock = alock->as_Lock()->fastlock_node();
2238 if (flock->outcnt() == 1) {
2239 assert(flock->unique_out() == alock, "sanity");
2240 _igvn.replace_node(flock, top());
2241 }
2242 }
2243
2244 // Seach for MemBarReleaseLock node and delete it also.
2245 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2246 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2247 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2248 mem->is_Proj() && membar == mem->in(0), "");
2249 _igvn.replace_node(fallthroughproj, ctrl);
2250 _igvn.replace_node(memproj_fallthrough, mem);
2251 fallthroughproj = ctrl;
2252 memproj_fallthrough = mem;
2253 ctrl = membar->in(TypeFunc::Control);
2254 mem = membar->in(TypeFunc::Memory);
2255 }
2256
2257 _igvn.replace_node(fallthroughproj, ctrl);
2258 _igvn.replace_node(memproj_fallthrough, mem);
2259 return true;
2260 }
2261
2262
2263 //------------------------------expand_lock_node----------------------
2264 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2265
2266 Node* ctrl = lock->in(TypeFunc::Control);
2267 Node* mem = lock->in(TypeFunc::Memory);
2268 Node* obj = lock->obj_node();
2269 Node* box = lock->box_node();
2270 Node* flock = lock->fastlock_node();
2271
2272 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2273
2274 // Make the merge point
2275 Node *region;
2276 Node *mem_phi;
2277 Node *slow_path;
2278
2279 if (UseOptoBiasInlining) {
2280 /*
2281 * See the full description in MacroAssembler::biased_locking_enter().
2282 *
2283 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2284 * // The object is biased.
2285 * proto_node = klass->prototype_header;
2286 * o_node = thread | proto_node;
2287 * x_node = o_node ^ mark_word;
2288 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2289 * // Done.
2290 * } else {
2291 * if( (x_node & biased_lock_mask) != 0 ) {
2292 * // The klass's prototype header is no longer biased.
2293 * cas(&mark_word, mark_word, proto_node)
2294 * goto cas_lock;
2295 * } else {
2296 * // The klass's prototype header is still biased.
2297 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2298 * old = mark_word;
2299 * new = o_node;
2300 * } else {
2301 * // Different thread or anonymous biased.
2302 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2303 * new = thread | old;
2304 * }
2305 * // Try to rebias.
2306 * if( cas(&mark_word, old, new) == 0 ) {
2307 * // Done.
2308 * } else {
2309 * goto slow_path; // Failed.
2310 * }
2311 * }
2312 * }
2313 * } else {
2314 * // The object is not biased.
2315 * cas_lock:
2316 * if( FastLock(obj) == 0 ) {
2317 * // Done.
2318 * } else {
2319 * slow_path:
2320 * OptoRuntime::complete_monitor_locking_Java(obj);
2321 * }
2322 * }
2323 */
2324
2325 region = new RegionNode(5);
2326 // create a Phi for the memory state
2327 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2328
2329 Node* fast_lock_region = new RegionNode(3);
2330 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2331
2332 // First, check mark word for the biased lock pattern.
2333 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2334
2335 // Get fast path - mark word has the biased lock pattern.
2336 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2337 markOopDesc::biased_lock_mask_in_place,
2338 markOopDesc::biased_lock_pattern, true);
2339 // fast_lock_region->in(1) is set to slow path.
2340 fast_lock_mem_phi->init_req(1, mem);
2341
2342 // Now check that the lock is biased to the current thread and has
2343 // the same epoch and bias as Klass::_prototype_header.
2344
2345 // Special-case a fresh allocation to avoid building nodes:
2346 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2347 if (klass_node == NULL) {
2348 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2349 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2350 #ifdef _LP64
2351 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2352 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2353 klass_node->in(1)->init_req(0, ctrl);
2354 } else
2355 #endif
2356 klass_node->init_req(0, ctrl);
2357 }
2358 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2359
2360 Node* thread = transform_later(new ThreadLocalNode());
2361 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2362 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2363 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2364
2365 // Get slow path - mark word does NOT match the value.
2366 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2367 (~markOopDesc::age_mask_in_place), 0);
2368 // region->in(3) is set to fast path - the object is biased to the current thread.
2369 mem_phi->init_req(3, mem);
2370
2371
2372 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2373
2374
2375 // First, check biased pattern.
2376 // Get fast path - _prototype_header has the same biased lock pattern.
2377 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2378 markOopDesc::biased_lock_mask_in_place, 0, true);
2379
2380 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2381 // fast_lock_region->in(2) - the prototype header is no longer biased
2382 // and we have to revoke the bias on this object.
2383 // We are going to try to reset the mark of this object to the prototype
2384 // value and fall through to the CAS-based locking scheme.
2385 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2386 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2387 proto_node, mark_node);
2388 transform_later(cas);
2389 Node* proj = transform_later(new SCMemProjNode(cas));
2390 fast_lock_mem_phi->init_req(2, proj);
2391
2392
2393 // Second, check epoch bits.
2394 Node* rebiased_region = new RegionNode(3);
2395 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2396 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2397
2398 // Get slow path - mark word does NOT match epoch bits.
2399 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2400 markOopDesc::epoch_mask_in_place, 0);
2401 // The epoch of the current bias is not valid, attempt to rebias the object
2402 // toward the current thread.
2403 rebiased_region->init_req(2, epoch_ctrl);
2404 old_phi->init_req(2, mark_node);
2405 new_phi->init_req(2, o_node);
2406
2407 // rebiased_region->in(1) is set to fast path.
2408 // The epoch of the current bias is still valid but we know
2409 // nothing about the owner; it might be set or it might be clear.
2410 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2411 markOopDesc::age_mask_in_place |
2412 markOopDesc::epoch_mask_in_place);
2413 Node* old = transform_later(new AndXNode(mark_node, cmask));
2414 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2415 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2416 old_phi->init_req(1, old);
2417 new_phi->init_req(1, new_mark);
2418
2419 transform_later(rebiased_region);
2420 transform_later(old_phi);
2421 transform_later(new_phi);
2422
2423 // Try to acquire the bias of the object using an atomic operation.
2424 // If this fails we will go in to the runtime to revoke the object's bias.
2425 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2426 transform_later(cas);
2427 proj = transform_later(new SCMemProjNode(cas));
2428
2429 // Get slow path - Failed to CAS.
2430 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2431 mem_phi->init_req(4, proj);
2432 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2433
2434 // Failed to CAS.
2435 slow_path = new RegionNode(3);
2436 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2437
2438 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2439 slow_mem->init_req(1, proj);
2440
2441 // Call CAS-based locking scheme (FastLock node).
2442
2443 transform_later(fast_lock_region);
2444 transform_later(fast_lock_mem_phi);
2445
2446 // Get slow path - FastLock failed to lock the object.
2447 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2448 mem_phi->init_req(2, fast_lock_mem_phi);
2449 // region->in(2) is set to fast path - the object is locked to the current thread.
2450
2451 slow_path->init_req(2, ctrl); // Capture slow-control
2452 slow_mem->init_req(2, fast_lock_mem_phi);
2453
2454 transform_later(slow_path);
2455 transform_later(slow_mem);
2456 // Reset lock's memory edge.
2457 lock->set_req(TypeFunc::Memory, slow_mem);
2458
2459 } else {
2460 region = new RegionNode(3);
2461 // create a Phi for the memory state
2462 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2463
2464 // Optimize test; set region slot 2
2465 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2466 mem_phi->init_req(2, mem);
2467 }
2468
2469 // Make slow path call
2470 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2471 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2472 obj, box, NULL);
2473
2474 extract_call_projections(call);
2475
2476 // Slow path can only throw asynchronous exceptions, which are always
2477 // de-opted. So the compiler thinks the slow-call can never throw an
2478 // exception. If it DOES throw an exception we would need the debug
2479 // info removed first (since if it throws there is no monitor).
2480 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2481 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2482
2483 // Capture slow path
2484 // disconnect fall-through projection from call and create a new one
2485 // hook up users of fall-through projection to region
2486 Node *slow_ctrl = _fallthroughproj->clone();
2487 transform_later(slow_ctrl);
2488 _igvn.hash_delete(_fallthroughproj);
2489 _fallthroughproj->disconnect_inputs(NULL, C);
2490 region->init_req(1, slow_ctrl);
2491 // region inputs are now complete
2492 transform_later(region);
2493 _igvn.replace_node(_fallthroughproj, region);
2494
2495 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2496 mem_phi->init_req(1, memproj );
2497 transform_later(mem_phi);
2498 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2499 }
2500
2501 //------------------------------expand_unlock_node----------------------
2502 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2503
2504 Node* ctrl = unlock->in(TypeFunc::Control);
2505 Node* mem = unlock->in(TypeFunc::Memory);
2506 Node* obj = unlock->obj_node();
2507 Node* box = unlock->box_node();
2508
2509 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2510
2511 // No need for a null check on unlock
2512
2513 // Make the merge point
2514 Node *region;
2515 Node *mem_phi;
2516
2517 if (UseOptoBiasInlining) {
2518 // Check for biased locking unlock case, which is a no-op.
2519 // See the full description in MacroAssembler::biased_locking_exit().
2520 region = new RegionNode(4);
2521 // create a Phi for the memory state
2522 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2523 mem_phi->init_req(3, mem);
2524
2525 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2526 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2527 markOopDesc::biased_lock_mask_in_place,
2528 markOopDesc::biased_lock_pattern);
2529 } else {
2530 region = new RegionNode(3);
2531 // create a Phi for the memory state
2532 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2533 }
2534
2535 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2536 funlock = transform_later( funlock )->as_FastUnlock();
2537 // Optimize test; set region slot 2
2538 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2539 Node *thread = transform_later(new ThreadLocalNode());
2540
2541 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2542 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2543 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2544
2545 extract_call_projections(call);
2546
2547 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2548 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2549
2550 // No exceptions for unlocking
2551 // Capture slow path
2552 // disconnect fall-through projection from call and create a new one
2553 // hook up users of fall-through projection to region
2554 Node *slow_ctrl = _fallthroughproj->clone();
2555 transform_later(slow_ctrl);
2556 _igvn.hash_delete(_fallthroughproj);
2557 _fallthroughproj->disconnect_inputs(NULL, C);
2558 region->init_req(1, slow_ctrl);
2559 // region inputs are now complete
2560 transform_later(region);
2561 _igvn.replace_node(_fallthroughproj, region);
2562
2563 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2564 mem_phi->init_req(1, memproj );
2565 mem_phi->init_req(2, mem);
2566 transform_later(mem_phi);
2567 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2568 }
2569
2570 //---------------------------eliminate_macro_nodes----------------------
2571 // Eliminate scalar replaced allocations and associated locks.
2572 void PhaseMacroExpand::eliminate_macro_nodes() {
2573 if (C->macro_count() == 0)
2574 return;
2575
2576 // First, attempt to eliminate locks
2577 int cnt = C->macro_count();
2578 for (int i=0; i < cnt; i++) {
2579 Node *n = C->macro_node(i);
2580 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2581 // Before elimination mark all associated (same box and obj)
2582 // lock and unlock nodes.
2583 mark_eliminated_locking_nodes(n->as_AbstractLock());
2584 }
2585 }
2586 bool progress = true;
2587 while (progress) {
2588 progress = false;
2589 for (int i = C->macro_count(); i > 0; i--) {
2590 Node * n = C->macro_node(i-1);
2591 bool success = false;
2592 debug_only(int old_macro_count = C->macro_count(););
2593 if (n->is_AbstractLock()) {
2594 success = eliminate_locking_node(n->as_AbstractLock());
2595 }
2596 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2597 progress = progress || success;
2598 }
2599 }
2600 // Next, attempt to eliminate allocations
2601 _has_locks = false;
2602 progress = true;
2603 while (progress) {
2604 progress = false;
2605 for (int i = C->macro_count(); i > 0; i--) {
2606 Node * n = C->macro_node(i-1);
2607 bool success = false;
2608 debug_only(int old_macro_count = C->macro_count(););
2609 switch (n->class_id()) {
2610 case Node::Class_Allocate:
2611 case Node::Class_AllocateArray:
2612 success = eliminate_allocate_node(n->as_Allocate());
2613 break;
2614 case Node::Class_CallStaticJava:
2615 success = eliminate_boxing_node(n->as_CallStaticJava());
2616 break;
2617 case Node::Class_Lock:
2618 case Node::Class_Unlock:
2619 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2620 _has_locks = true;
2621 break;
2622 case Node::Class_ArrayCopy:
2623 break;
2624 case Node::Class_OuterStripMinedLoop:
2625 break;
2626 case Node::Class_Opaque1:
2627 break;
2628 default:
2629 assert(n->Opcode() == Op_LoopLimit ||
2630 n->Opcode() == Op_Opaque2 ||
2631 n->Opcode() == Op_Opaque3 ||
2632 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2633 "unknown node type in macro list");
2634 }
2635 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2636 progress = progress || success;
2637 }
2638 }
2639 }
2640
2641 //------------------------------expand_macro_nodes----------------------
2642 // Returns true if a failure occurred.
2643 bool PhaseMacroExpand::expand_macro_nodes() {
2644 // Last attempt to eliminate macro nodes.
2645 eliminate_macro_nodes();
2646
2647 // Make sure expansion will not cause node limit to be exceeded.
2648 // Worst case is a macro node gets expanded into about 200 nodes.
2649 // Allow 50% more for optimization.
2650 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2651 return true;
2652
2653 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2654 bool progress = true;
2655 while (progress) {
2656 progress = false;
2657 for (int i = C->macro_count(); i > 0; i--) {
2658 Node * n = C->macro_node(i-1);
2659 bool success = false;
2660 debug_only(int old_macro_count = C->macro_count(););
2661 if (n->Opcode() == Op_LoopLimit) {
2662 // Remove it from macro list and put on IGVN worklist to optimize.
2663 C->remove_macro_node(n);
2664 _igvn._worklist.push(n);
2665 success = true;
2666 } else if (n->Opcode() == Op_CallStaticJava) {
2667 // Remove it from macro list and put on IGVN worklist to optimize.
2668 C->remove_macro_node(n);
2669 _igvn._worklist.push(n);
2670 success = true;
2671 } else if (n->is_Opaque1() || n->Opcode() == Op_Opaque2) {
2672 _igvn.replace_node(n, n->in(1));
2673 success = true;
2674 #if INCLUDE_RTM_OPT
2675 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2676 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2677 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2678 Node* cmp = n->unique_out();
2679 #ifdef ASSERT
2680 // Validate graph.
2681 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2682 BoolNode* bol = cmp->unique_out()->as_Bool();
2683 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2684 (bol->_test._test == BoolTest::ne), "");
2685 IfNode* ifn = bol->unique_out()->as_If();
2686 assert((ifn->outcnt() == 2) &&
2687 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2688 #endif
2689 Node* repl = n->in(1);
2690 if (!_has_locks) {
2691 // Remove RTM state check if there are no locks in the code.
2692 // Replace input to compare the same value.
2693 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2694 }
2695 _igvn.replace_node(n, repl);
2696 success = true;
2697 #endif
2698 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2699 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2700 C->remove_macro_node(n);
2701 success = true;
2702 }
2703 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2704 progress = progress || success;
2705 }
2706 }
2707
2708 // expand arraycopy "macro" nodes first
2709 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2710 // before the allocate nodes are expanded.
2711 int macro_idx = C->macro_count() - 1;
2712 while (macro_idx >= 0) {
2713 Node * n = C->macro_node(macro_idx);
2714 assert(n->is_macro(), "only macro nodes expected here");
2715 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2716 // node is unreachable, so don't try to expand it
2717 C->remove_macro_node(n);
2718 } else if (n->is_ArrayCopy()){
2719 int macro_count = C->macro_count();
2720 expand_arraycopy_node(n->as_ArrayCopy());
2721 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2722 }
2723 if (C->failing()) return true;
2724 macro_idx --;
2725 }
2726
2727 // expand "macro" nodes
2728 // nodes are removed from the macro list as they are processed
2729 while (C->macro_count() > 0) {
2730 int macro_count = C->macro_count();
2731 Node * n = C->macro_node(macro_count-1);
2732 assert(n->is_macro(), "only macro nodes expected here");
2733 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2734 // node is unreachable, so don't try to expand it
2735 C->remove_macro_node(n);
2736 continue;
2737 }
2738 switch (n->class_id()) {
2739 case Node::Class_Allocate:
2740 expand_allocate(n->as_Allocate());
2741 break;
2742 case Node::Class_AllocateArray:
2743 expand_allocate_array(n->as_AllocateArray());
2744 break;
2745 case Node::Class_Lock:
2746 expand_lock_node(n->as_Lock());
2747 break;
2748 case Node::Class_Unlock:
2749 expand_unlock_node(n->as_Unlock());
2750 break;
2751 default:
2752 assert(false, "unknown node type in macro list");
2753 }
2754 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2755 if (C->failing()) return true;
2756 }
2757
2758 _igvn.set_delay_transform(false);
2759 _igvn.optimize();
2760 if (C->failing()) return true;
2761 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2762 return bs->expand_macro_nodes(this);
2763 }