1 /*
2 * Copyright (c) 2005, 2015, 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 "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/macro.hpp"
38 #include "opto/memnode.hpp"
39 #include "opto/narrowptrnode.hpp"
40 #include "opto/node.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/phaseX.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46 #include "opto/type.hpp"
47 #include "runtime/sharedRuntime.hpp"
48
49
50 //
51 // Replace any references to "oldref" in inputs to "use" with "newref".
52 // Returns the number of replacements made.
53 //
54 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
55 int nreplacements = 0;
56 uint req = use->req();
57 for (uint j = 0; j < use->len(); j++) {
58 Node *uin = use->in(j);
59 if (uin == oldref) {
60 if (j < req)
61 use->set_req(j, newref);
62 else
63 use->set_prec(j, newref);
64 nreplacements++;
65 } else if (j >= req && uin == NULL) {
66 break;
67 }
68 }
69 return nreplacements;
70 }
71
72 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
73 // Copy debug information and adjust JVMState information
74 uint old_dbg_start = oldcall->tf()->domain()->cnt();
75 uint new_dbg_start = newcall->tf()->domain()->cnt();
76 int jvms_adj = new_dbg_start - old_dbg_start;
77 assert (new_dbg_start == newcall->req(), "argument count mismatch");
78
79 // SafePointScalarObject node could be referenced several times in debug info.
80 // Use Dict to record cloned nodes.
81 Dict* sosn_map = new Dict(cmpkey,hashkey);
82 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
83 Node* old_in = oldcall->in(i);
84 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
85 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
86 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
87 uint old_unique = C->unique();
88 Node* new_in = old_sosn->clone(sosn_map);
89 if (old_unique != C->unique()) { // New node?
90 new_in->set_req(0, C->root()); // reset control edge
91 new_in = transform_later(new_in); // Register new node.
92 }
93 old_in = new_in;
94 }
95 newcall->add_req(old_in);
96 }
97
98 // JVMS may be shared so clone it before we modify it
99 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
100 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
101 jvms->set_map(newcall);
102 jvms->set_locoff(jvms->locoff()+jvms_adj);
103 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
104 jvms->set_monoff(jvms->monoff()+jvms_adj);
105 jvms->set_scloff(jvms->scloff()+jvms_adj);
106 jvms->set_endoff(jvms->endoff()+jvms_adj);
107 }
108 }
109
110 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
111 Node* cmp;
112 if (mask != 0) {
113 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
114 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
115 } else {
116 cmp = word;
117 }
118 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
119 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
120 transform_later(iff);
121
122 // Fast path taken.
123 Node *fast_taken = transform_later(new IfFalseNode(iff));
124
125 // Fast path not-taken, i.e. slow path
126 Node *slow_taken = transform_later(new IfTrueNode(iff));
127
128 if (return_fast_path) {
129 region->init_req(edge, slow_taken); // Capture slow-control
130 return fast_taken;
131 } else {
132 region->init_req(edge, fast_taken); // Capture fast-control
133 return slow_taken;
134 }
135 }
136
137 //--------------------copy_predefined_input_for_runtime_call--------------------
138 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
139 // Set fixed predefined input arguments
140 call->init_req( TypeFunc::Control, ctrl );
141 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
142 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
143 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
144 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
145 }
146
147 //------------------------------make_slow_call---------------------------------
148 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
149 address slow_call, const char* leaf_name, Node* slow_path,
150 Node* parm0, Node* parm1, Node* parm2) {
151
152 // Slow-path call
153 CallNode *call = leaf_name
154 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
155 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
156
157 // Slow path call has no side-effects, uses few values
158 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
159 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
160 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
161 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
162 copy_call_debug_info(oldcall, call);
163 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
164 _igvn.replace_node(oldcall, call);
165 transform_later(call);
166
167 return call;
168 }
169
170 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
171 _fallthroughproj = NULL;
172 _fallthroughcatchproj = NULL;
173 _ioproj_fallthrough = NULL;
174 _ioproj_catchall = NULL;
175 _catchallcatchproj = NULL;
176 _memproj_fallthrough = NULL;
177 _memproj_catchall = NULL;
178 _resproj = NULL;
179 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
180 ProjNode *pn = call->fast_out(i)->as_Proj();
181 switch (pn->_con) {
182 case TypeFunc::Control:
183 {
184 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
185 _fallthroughproj = pn;
186 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
187 const Node *cn = pn->fast_out(j);
188 if (cn->is_Catch()) {
189 ProjNode *cpn = NULL;
190 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
191 cpn = cn->fast_out(k)->as_Proj();
192 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
193 if (cpn->_con == CatchProjNode::fall_through_index)
194 _fallthroughcatchproj = cpn;
195 else {
196 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
197 _catchallcatchproj = cpn;
198 }
199 }
200 }
201 break;
202 }
203 case TypeFunc::I_O:
204 if (pn->_is_io_use)
205 _ioproj_catchall = pn;
206 else
207 _ioproj_fallthrough = pn;
208 break;
209 case TypeFunc::Memory:
210 if (pn->_is_io_use)
211 _memproj_catchall = pn;
212 else
213 _memproj_fallthrough = pn;
214 break;
215 case TypeFunc::Parms:
216 _resproj = pn;
217 break;
218 default:
219 assert(false, "unexpected projection from allocation node.");
220 }
221 }
222
223 }
224
225 // Eliminate a card mark sequence. p2x is a ConvP2XNode
226 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
227 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
228 if (!UseG1GC) {
229 // vanilla/CMS post barrier
230 Node *shift = p2x->unique_out();
231 Node *addp = shift->unique_out();
232 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
233 Node *mem = addp->last_out(j);
234 if (UseCondCardMark && mem->is_Load()) {
235 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
236 // The load is checking if the card has been written so
237 // replace it with zero to fold the test.
238 _igvn.replace_node(mem, intcon(0));
239 continue;
240 }
241 assert(mem->is_Store(), "store required");
242 _igvn.replace_node(mem, mem->in(MemNode::Memory));
243 }
244 } else {
245 // G1 pre/post barriers
246 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
247 // It could be only one user, URShift node, in Object.clone() instrinsic
248 // but the new allocation is passed to arraycopy stub and it could not
249 // be scalar replaced. So we don't check the case.
250
251 // An other case of only one user (Xor) is when the value check for NULL
252 // in G1 post barrier is folded after CCP so the code which used URShift
253 // is removed.
254
255 // Take Region node before eliminating post barrier since it also
256 // eliminates CastP2X node when it has only one user.
257 Node* this_region = p2x->in(0);
258 assert(this_region != NULL, "");
259
260 // Remove G1 post barrier.
261
262 // Search for CastP2X->Xor->URShift->Cmp path which
263 // checks if the store done to a different from the value's region.
264 // And replace Cmp with #0 (false) to collapse G1 post barrier.
265 Node* xorx = p2x->find_out_with(Op_XorX);
266 assert(xorx != NULL, "missing G1 post barrier");
267 Node* shift = xorx->unique_out();
268 Node* cmpx = shift->unique_out();
269 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
270 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
271 "missing region check in G1 post barrier");
272 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
273
274 // Remove G1 pre barrier.
275
276 // Search "if (marking != 0)" check and set it to "false".
277 // There is no G1 pre barrier if previous stored value is NULL
278 // (for example, after initialization).
279 if (this_region->is_Region() && this_region->req() == 3) {
280 int ind = 1;
281 if (!this_region->in(ind)->is_IfFalse()) {
282 ind = 2;
283 }
284 if (this_region->in(ind)->is_IfFalse()) {
285 Node* bol = this_region->in(ind)->in(0)->in(1);
286 assert(bol->is_Bool(), "");
287 cmpx = bol->in(1);
288 if (bol->as_Bool()->_test._test == BoolTest::ne &&
289 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
290 cmpx->in(1)->is_Load()) {
291 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
292 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
293 PtrQueue::byte_offset_of_active());
294 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
295 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
296 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
297 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
298 }
299 }
300 }
301 }
302 // Now CastP2X can be removed since it is used only on dead path
303 // which currently still alive until igvn optimize it.
304 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
305 _igvn.replace_node(p2x, top());
306 }
307 }
308
309 // Search for a memory operation for the specified memory slice.
310 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
311 Node *orig_mem = mem;
312 Node *alloc_mem = alloc->in(TypeFunc::Memory);
313 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
314 while (true) {
315 if (mem == alloc_mem || mem == start_mem ) {
316 return mem; // hit one of our sentinels
317 } else if (mem->is_MergeMem()) {
318 mem = mem->as_MergeMem()->memory_at(alias_idx);
319 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
320 Node *in = mem->in(0);
321 // we can safely skip over safepoints, calls, locks and membars because we
322 // already know that the object is safe to eliminate.
323 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
324 return in;
325 } else if (in->is_Call()) {
326 CallNode *call = in->as_Call();
327 if (call->may_modify(tinst, phase)) {
328 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
329
330 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
331 return in;
332 }
333 }
334 mem = in->in(TypeFunc::Memory);
335 } else if (in->is_MemBar()) {
336 if (ArrayCopyNode::membar_for_arraycopy(tinst, in->as_MemBar(), phase)) {
337 assert(in->in(0)->is_Proj() && in->in(0)->in(0)->is_ArrayCopy(), "should be arraycopy");
338 ArrayCopyNode* ac = in->in(0)->in(0)->as_ArrayCopy();
339 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
340 return ac;
341 }
342 mem = in->in(TypeFunc::Memory);
343 } else {
344 assert(false, "unexpected projection");
345 }
346 } else if (mem->is_Store()) {
347 const TypePtr* atype = mem->as_Store()->adr_type();
348 int adr_idx = phase->C->get_alias_index(atype);
349 if (adr_idx == alias_idx) {
350 assert(atype->isa_oopptr(), "address type must be oopptr");
351 int adr_offset = atype->offset();
352 uint adr_iid = atype->is_oopptr()->instance_id();
353 // Array elements references have the same alias_idx
354 // but different offset and different instance_id.
355 if (adr_offset == offset && adr_iid == alloc->_idx)
356 return mem;
357 } else {
358 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
359 }
360 mem = mem->in(MemNode::Memory);
361 } else if (mem->is_ClearArray()) {
362 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
363 // Can not bypass initialization of the instance
364 // we are looking.
365 debug_only(intptr_t offset;)
366 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
367 InitializeNode* init = alloc->as_Allocate()->initialization();
368 // We are looking for stored value, return Initialize node
369 // or memory edge from Allocate node.
370 if (init != NULL)
371 return init;
372 else
373 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
374 }
375 // Otherwise skip it (the call updated 'mem' value).
376 } else if (mem->Opcode() == Op_SCMemProj) {
377 mem = mem->in(0);
378 Node* adr = NULL;
379 if (mem->is_LoadStore()) {
380 adr = mem->in(MemNode::Address);
381 } else {
382 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
383 adr = mem->in(3); // Destination array
384 }
385 const TypePtr* atype = adr->bottom_type()->is_ptr();
386 int adr_idx = phase->C->get_alias_index(atype);
387 if (adr_idx == alias_idx) {
388 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
389 return NULL;
390 }
391 mem = mem->in(MemNode::Memory);
392 } else {
393 return mem;
394 }
395 assert(mem != orig_mem, "dead memory loop");
396 }
397 }
398
399 // Generate loads from source of the arraycopy for fields of
400 // destination needed at a deoptimization point
401 Node* PhaseMacroExpand::handle_arraycopy(ArrayCopyNode* ac, intptr_t offset, Node* ctl, BasicType ft, const Type *ftype, AllocateNode *alloc) {
402 BasicType bt = ft;
403 const Type *type = ftype;
404 if (ft == T_NARROWOOP) {
405 bt = T_OBJECT;
406 type = ftype->make_oopptr();
407 }
408 Node* res = NULL;
409 if (ac->is_clonebasic()) {
410 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
411 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
412 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
413 Node* m = ac->in(TypeFunc::Memory);
414 while (m->is_MergeMem()) {
415 m = m->as_MergeMem()->memory_at(C->get_alias_index(adr_type));
416 if (m->is_Proj() && m->in(0)->is_MemBar()) {
417 m = m->in(0)->in(TypeFunc::Memory);
418 }
419 }
420 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
421 } else {
422 if (ac->modifies(offset, offset, &_igvn, true)) {
423 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
424 uint shift = exact_log2(type2aelembytes(bt));
425 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
426 #ifdef _LP64
427 diff = _igvn.transform(new ConvI2LNode(diff));
428 #endif
429 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
430
431 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
432 Node* base = ac->in(ArrayCopyNode::Src);
433 Node* adr = _igvn.transform(new AddPNode(base, base, off));
434 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
435 Node* m = ac->in(TypeFunc::Memory);
436 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
437 }
438 }
439 if (res != NULL) {
440 res = _igvn.transform(res);
441 if (ftype->isa_narrowoop()) {
442 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
443 res = _igvn.transform(new EncodePNode(res, ftype));
444 }
445 return res;
446 }
447 return NULL;
448 }
449
450 //
451 // Given a Memory Phi, compute a value Phi containing the values from stores
452 // on the input paths.
453 // Note: this function is recursive, its depth is limited by the "level" argument
454 // Returns the computed Phi, or NULL if it cannot compute it.
455 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) {
456 assert(mem->is_Phi(), "sanity");
457 int alias_idx = C->get_alias_index(adr_t);
458 int offset = adr_t->offset();
459 int instance_id = adr_t->instance_id();
460
461 // Check if an appropriate value phi already exists.
462 Node* region = mem->in(0);
463 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
464 Node* phi = region->fast_out(k);
465 if (phi->is_Phi() && phi != mem &&
466 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
467 return phi;
468 }
469 }
470 // Check if an appropriate new value phi already exists.
471 Node* new_phi = value_phis->find(mem->_idx);
472 if (new_phi != NULL)
473 return new_phi;
474
475 if (level <= 0) {
476 return NULL; // Give up: phi tree too deep
477 }
478 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
479 Node *alloc_mem = alloc->in(TypeFunc::Memory);
480
481 uint length = mem->req();
482 GrowableArray <Node *> values(length, length, NULL, false);
483
484 // create a new Phi for the value
485 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
486 transform_later(phi);
487 value_phis->push(phi, mem->_idx);
488
489 for (uint j = 1; j < length; j++) {
490 Node *in = mem->in(j);
491 if (in == NULL || in->is_top()) {
492 values.at_put(j, in);
493 } else {
494 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
495 if (val == start_mem || val == alloc_mem) {
496 // hit a sentinel, return appropriate 0 value
497 values.at_put(j, _igvn.zerocon(ft));
498 continue;
499 }
500 if (val->is_Initialize()) {
501 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
502 }
503 if (val == NULL) {
504 return NULL; // can't find a value on this path
505 }
506 if (val == mem) {
507 values.at_put(j, mem);
508 } else if (val->is_Store()) {
509 values.at_put(j, val->in(MemNode::ValueIn));
510 } else if(val->is_Proj() && val->in(0) == alloc) {
511 values.at_put(j, _igvn.zerocon(ft));
512 } else if (val->is_Phi()) {
513 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
514 if (val == NULL) {
515 return NULL;
516 }
517 values.at_put(j, val);
518 } else if (val->Opcode() == Op_SCMemProj) {
519 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
520 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
521 return NULL;
522 } else if (val->is_ArrayCopy()) {
523 Node* res = handle_arraycopy(val->as_ArrayCopy(), offset, val->in(0), ft, phi_type, alloc);
524 values.at_put(j, res);
525 } else {
526 #ifdef ASSERT
527 val->dump();
528 assert(false, "unknown node on this path");
529 #endif
530 return NULL; // unknown node on this path
531 }
532 }
533 }
534 // Set Phi's inputs
535 for (uint j = 1; j < length; j++) {
536 if (values.at(j) == mem) {
537 phi->init_req(j, phi);
538 } else {
539 phi->init_req(j, values.at(j));
540 }
541 }
542 return phi;
543 }
544
545 // Search the last value stored into the object's field.
546 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
547 assert(adr_t->is_known_instance_field(), "instance required");
548 int instance_id = adr_t->instance_id();
549 assert((uint)instance_id == alloc->_idx, "wrong allocation");
550
551 int alias_idx = C->get_alias_index(adr_t);
552 int offset = adr_t->offset();
553 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
554 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
555 Node *alloc_mem = alloc->in(TypeFunc::Memory);
556 Arena *a = Thread::current()->resource_area();
557 VectorSet visited(a);
558
559
560 bool done = sfpt_mem == alloc_mem;
561 Node *mem = sfpt_mem;
562 while (!done) {
563 if (visited.test_set(mem->_idx)) {
564 return NULL; // found a loop, give up
565 }
566 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
567 if (mem == start_mem || mem == alloc_mem) {
568 done = true; // hit a sentinel, return appropriate 0 value
569 } else if (mem->is_Initialize()) {
570 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
571 if (mem == NULL) {
572 done = true; // Something go wrong.
573 } else if (mem->is_Store()) {
574 const TypePtr* atype = mem->as_Store()->adr_type();
575 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
576 done = true;
577 }
578 } else if (mem->is_Store()) {
579 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
580 assert(atype != NULL, "address type must be oopptr");
581 assert(C->get_alias_index(atype) == alias_idx &&
582 atype->is_known_instance_field() && atype->offset() == offset &&
583 atype->instance_id() == instance_id, "store is correct memory slice");
584 done = true;
585 } else if (mem->is_Phi()) {
586 // try to find a phi's unique input
587 Node *unique_input = NULL;
588 Node *top = C->top();
589 for (uint i = 1; i < mem->req(); i++) {
590 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
591 if (n == NULL || n == top || n == mem) {
592 continue;
593 } else if (unique_input == NULL) {
594 unique_input = n;
595 } else if (unique_input != n) {
596 unique_input = top;
597 break;
598 }
599 }
600 if (unique_input != NULL && unique_input != top) {
601 mem = unique_input;
602 } else {
603 done = true;
604 }
605 } else if (mem->is_ArrayCopy()) {
606 done = true;
607 } else {
608 assert(false, "unexpected node");
609 }
610 }
611 if (mem != NULL) {
612 if (mem == start_mem || mem == alloc_mem) {
613 // hit a sentinel, return appropriate 0 value
614 return _igvn.zerocon(ft);
615 } else if (mem->is_Store()) {
616 return mem->in(MemNode::ValueIn);
617 } else if (mem->is_Phi()) {
618 // attempt to produce a Phi reflecting the values on the input paths of the Phi
619 Node_Stack value_phis(a, 8);
620 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
621 if (phi != NULL) {
622 return phi;
623 } else {
624 // Kill all new Phis
625 while(value_phis.is_nonempty()) {
626 Node* n = value_phis.node();
627 _igvn.replace_node(n, C->top());
628 value_phis.pop();
629 }
630 }
631 } else if (mem->is_ArrayCopy()) {
632 Node* ctl = mem->in(0);
633 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
634 // pin the loads in the uncommon trap path
635 ctl = sfpt_ctl;
636 }
637 return handle_arraycopy(mem->as_ArrayCopy(), offset, ctl, ft, ftype, alloc);
638 }
639 }
640 // Something go wrong.
641 return NULL;
642 }
643
644 // Check the possibility of scalar replacement.
645 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
646 // Scan the uses of the allocation to check for anything that would
647 // prevent us from eliminating it.
648 NOT_PRODUCT( const char* fail_eliminate = NULL; )
649 DEBUG_ONLY( Node* disq_node = NULL; )
650 bool can_eliminate = true;
651
652 Node* res = alloc->result_cast();
653 const TypeOopPtr* res_type = NULL;
654 if (res == NULL) {
655 // All users were eliminated.
656 } else if (!res->is_CheckCastPP()) {
657 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
658 can_eliminate = false;
659 } else {
660 res_type = _igvn.type(res)->isa_oopptr();
661 if (res_type == NULL) {
662 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
663 can_eliminate = false;
664 } else if (res_type->isa_aryptr()) {
665 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
666 if (length < 0) {
667 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
668 can_eliminate = false;
669 }
670 }
671 }
672
673 if (can_eliminate && res != NULL) {
674 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
675 j < jmax && can_eliminate; j++) {
676 Node* use = res->fast_out(j);
677
678 if (use->is_AddP()) {
679 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
680 int offset = addp_type->offset();
681
682 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
683 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
684 can_eliminate = false;
685 break;
686 }
687 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
688 k < kmax && can_eliminate; k++) {
689 Node* n = use->fast_out(k);
690 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
691 !(n->is_ArrayCopy() &&
692 n->as_ArrayCopy()->is_clonebasic() &&
693 n->in(ArrayCopyNode::Dest) == use)) {
694 DEBUG_ONLY(disq_node = n;)
695 if (n->is_Load() || n->is_LoadStore()) {
696 NOT_PRODUCT(fail_eliminate = "Field load";)
697 } else {
698 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
699 }
700 can_eliminate = false;
701 }
702 }
703 } else if (use->is_ArrayCopy() &&
704 (use->as_ArrayCopy()->is_arraycopy_validated() ||
705 use->as_ArrayCopy()->is_copyof_validated() ||
706 use->as_ArrayCopy()->is_copyofrange_validated()) &&
707 use->in(ArrayCopyNode::Dest) == res) {
708 // ok to eliminate
709 } else if (use->is_SafePoint()) {
710 SafePointNode* sfpt = use->as_SafePoint();
711 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
712 // Object is passed as argument.
713 DEBUG_ONLY(disq_node = use;)
714 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
715 can_eliminate = false;
716 }
717 Node* sfptMem = sfpt->memory();
718 if (sfptMem == NULL || sfptMem->is_top()) {
719 DEBUG_ONLY(disq_node = use;)
720 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
721 can_eliminate = false;
722 } else {
723 safepoints.append_if_missing(sfpt);
724 }
725 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
726 if (use->is_Phi()) {
727 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
728 NOT_PRODUCT(fail_eliminate = "Object is return value";)
729 } else {
730 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
731 }
732 DEBUG_ONLY(disq_node = use;)
733 } else {
734 if (use->Opcode() == Op_Return) {
735 NOT_PRODUCT(fail_eliminate = "Object is return value";)
736 }else {
737 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
738 }
739 DEBUG_ONLY(disq_node = use;)
740 }
741 can_eliminate = false;
742 }
743 }
744 }
745
746 #ifndef PRODUCT
747 if (PrintEliminateAllocations) {
748 if (can_eliminate) {
749 tty->print("Scalar ");
750 if (res == NULL)
751 alloc->dump();
752 else
753 res->dump();
754 } else if (alloc->_is_scalar_replaceable) {
755 tty->print("NotScalar (%s)", fail_eliminate);
756 if (res == NULL)
757 alloc->dump();
758 else
759 res->dump();
760 #ifdef ASSERT
761 if (disq_node != NULL) {
762 tty->print(" >>>> ");
763 disq_node->dump();
764 }
765 #endif /*ASSERT*/
766 }
767 }
768 #endif
769 return can_eliminate;
770 }
771
772 // Do scalar replacement.
773 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
774 GrowableArray <SafePointNode *> safepoints_done;
775
776 ciKlass* klass = NULL;
777 ciInstanceKlass* iklass = NULL;
778 int nfields = 0;
779 int array_base;
780 int element_size;
781 BasicType basic_elem_type;
782 ciType* elem_type;
783
784 Node* res = alloc->result_cast();
785 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
786 const TypeOopPtr* res_type = NULL;
787 if (res != NULL) { // Could be NULL when there are no users
788 res_type = _igvn.type(res)->isa_oopptr();
789 }
790
791 if (res != NULL) {
792 klass = res_type->klass();
793 if (res_type->isa_instptr()) {
794 // find the fields of the class which will be needed for safepoint debug information
795 assert(klass->is_instance_klass(), "must be an instance klass.");
796 iklass = klass->as_instance_klass();
797 nfields = iklass->nof_nonstatic_fields();
798 } else {
799 // find the array's elements which will be needed for safepoint debug information
800 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
801 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
802 elem_type = klass->as_array_klass()->element_type();
803 basic_elem_type = elem_type->basic_type();
804 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
805 element_size = type2aelembytes(basic_elem_type);
806 }
807 }
808 //
809 // Process the safepoint uses
810 //
811 while (safepoints.length() > 0) {
812 SafePointNode* sfpt = safepoints.pop();
813 Node* mem = sfpt->memory();
814 Node* ctl = sfpt->control();
815 assert(sfpt->jvms() != NULL, "missed JVMS");
816 // Fields of scalar objs are referenced only at the end
817 // of regular debuginfo at the last (youngest) JVMS.
818 // Record relative start index.
819 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
820 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
821 #ifdef ASSERT
822 alloc,
823 #endif
824 first_ind, nfields);
825 sobj->init_req(0, C->root());
826 transform_later(sobj);
827
828 // Scan object's fields adding an input to the safepoint for each field.
829 for (int j = 0; j < nfields; j++) {
830 intptr_t offset;
831 ciField* field = NULL;
832 if (iklass != NULL) {
833 field = iklass->nonstatic_field_at(j);
834 offset = field->offset();
835 elem_type = field->type();
836 basic_elem_type = field->layout_type();
837 } else {
838 offset = array_base + j * (intptr_t)element_size;
839 }
840
841 const Type *field_type;
842 // The next code is taken from Parse::do_get_xxx().
843 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
844 if (!elem_type->is_loaded()) {
845 field_type = TypeInstPtr::BOTTOM;
846 } else if (field != NULL && field->is_constant() && field->is_static()) {
847 // This can happen if the constant oop is non-perm.
848 ciObject* con = field->constant_value().as_object();
849 // Do not "join" in the previous type; it doesn't add value,
850 // and may yield a vacuous result if the field is of interface type.
851 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
852 assert(field_type != NULL, "field singleton type must be consistent");
853 } else {
854 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
855 }
856 if (UseCompressedOops) {
857 field_type = field_type->make_narrowoop();
858 basic_elem_type = T_NARROWOOP;
859 }
860 } else {
861 field_type = Type::get_const_basic_type(basic_elem_type);
862 }
863
864 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
865
866 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
867 if (field_val == NULL) {
868 // We weren't able to find a value for this field,
869 // give up on eliminating this allocation.
870
871 // Remove any extra entries we added to the safepoint.
872 uint last = sfpt->req() - 1;
873 for (int k = 0; k < j; k++) {
874 sfpt->del_req(last--);
875 }
876 _igvn._worklist.push(sfpt);
877 // rollback processed safepoints
878 while (safepoints_done.length() > 0) {
879 SafePointNode* sfpt_done = safepoints_done.pop();
880 // remove any extra entries we added to the safepoint
881 last = sfpt_done->req() - 1;
882 for (int k = 0; k < nfields; k++) {
883 sfpt_done->del_req(last--);
884 }
885 JVMState *jvms = sfpt_done->jvms();
886 jvms->set_endoff(sfpt_done->req());
887 // Now make a pass over the debug information replacing any references
888 // to SafePointScalarObjectNode with the allocated object.
889 int start = jvms->debug_start();
890 int end = jvms->debug_end();
891 for (int i = start; i < end; i++) {
892 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
893 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
894 if (scobj->first_index(jvms) == sfpt_done->req() &&
895 scobj->n_fields() == (uint)nfields) {
896 assert(scobj->alloc() == alloc, "sanity");
897 sfpt_done->set_req(i, res);
898 }
899 }
900 }
901 _igvn._worklist.push(sfpt_done);
902 }
903 #ifndef PRODUCT
904 if (PrintEliminateAllocations) {
905 if (field != NULL) {
906 tty->print("=== At SafePoint node %d can't find value of Field: ",
907 sfpt->_idx);
908 field->print();
909 int field_idx = C->get_alias_index(field_addr_type);
910 tty->print(" (alias_idx=%d)", field_idx);
911 } else { // Array's element
912 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
913 sfpt->_idx, j);
914 }
915 tty->print(", which prevents elimination of: ");
916 if (res == NULL)
917 alloc->dump();
918 else
919 res->dump();
920 }
921 #endif
922 return false;
923 }
924 if (UseCompressedOops && field_type->isa_narrowoop()) {
925 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
926 // to be able scalar replace the allocation.
927 if (field_val->is_EncodeP()) {
928 field_val = field_val->in(1);
929 } else {
930 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
931 }
932 }
933 sfpt->add_req(field_val);
934 }
935 JVMState *jvms = sfpt->jvms();
936 jvms->set_endoff(sfpt->req());
937 // Now make a pass over the debug information replacing any references
938 // to the allocated object with "sobj"
939 int start = jvms->debug_start();
940 int end = jvms->debug_end();
941 sfpt->replace_edges_in_range(res, sobj, start, end);
942 _igvn._worklist.push(sfpt);
943 safepoints_done.append_if_missing(sfpt); // keep it for rollback
944 }
945 return true;
946 }
947
948 // Process users of eliminated allocation.
949 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
950 Node* res = alloc->result_cast();
951 if (res != NULL) {
952 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
953 Node *use = res->last_out(j);
954 uint oc1 = res->outcnt();
955
956 if (use->is_AddP()) {
957 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
958 Node *n = use->last_out(k);
959 uint oc2 = use->outcnt();
960 if (n->is_Store()) {
961 #ifdef ASSERT
962 // Verify that there is no dependent MemBarVolatile nodes,
963 // they should be removed during IGVN, see MemBarNode::Ideal().
964 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
965 p < pmax; p++) {
966 Node* mb = n->fast_out(p);
967 assert(mb->is_Initialize() || !mb->is_MemBar() ||
968 mb->req() <= MemBarNode::Precedent ||
969 mb->in(MemBarNode::Precedent) != n,
970 "MemBarVolatile should be eliminated for non-escaping object");
971 }
972 #endif
973 _igvn.replace_node(n, n->in(MemNode::Memory));
974 } else if (n->is_ArrayCopy()) {
975 // Disconnect ArrayCopy node
976 ArrayCopyNode* ac = n->as_ArrayCopy();
977 assert(ac->is_clonebasic(), "unexpected array copy kind");
978 Node* ctl_proj = ac->proj_out(TypeFunc::Control);
979 Node* mem_proj = ac->proj_out(TypeFunc::Memory);
980 if (ctl_proj != NULL) {
981 _igvn.replace_node(ctl_proj, n->in(0));
982 }
983 if (mem_proj != NULL) {
984 _igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
985 }
986 } else {
987 eliminate_card_mark(n);
988 }
989 k -= (oc2 - use->outcnt());
990 }
991 } else if (use->is_ArrayCopy()) {
992 // Disconnect ArrayCopy node
993 ArrayCopyNode* ac = use->as_ArrayCopy();
994 assert(ac->is_arraycopy_validated() ||
995 ac->is_copyof_validated() ||
996 ac->is_copyofrange_validated(), "unsupported");
997 CallProjections callprojs;
998 ac->extract_projections(&callprojs, true);
999
1000 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1001 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1002 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1003
1004 // Set control to top. IGVN will remove the remaining projections
1005 ac->set_req(0, top());
1006 ac->replace_edge(res, top());
1007
1008 // Disconnect src right away: it can help find new
1009 // opportunities for allocation elimination
1010 Node* src = ac->in(ArrayCopyNode::Src);
1011 ac->replace_edge(src, top());
1012 if (src->outcnt() == 0) {
1013 _igvn.remove_dead_node(src);
1014 }
1015
1016 _igvn._worklist.push(ac);
1017 } else {
1018 eliminate_card_mark(use);
1019 }
1020 j -= (oc1 - res->outcnt());
1021 }
1022 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1023 _igvn.remove_dead_node(res);
1024 }
1025
1026 //
1027 // Process other users of allocation's projections
1028 //
1029 if (_resproj != NULL && _resproj->outcnt() != 0) {
1030 // First disconnect stores captured by Initialize node.
1031 // If Initialize node is eliminated first in the following code,
1032 // it will kill such stores and DUIterator_Last will assert.
1033 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1034 Node *use = _resproj->fast_out(j);
1035 if (use->is_AddP()) {
1036 // raw memory addresses used only by the initialization
1037 _igvn.replace_node(use, C->top());
1038 --j; --jmax;
1039 }
1040 }
1041 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1042 Node *use = _resproj->last_out(j);
1043 uint oc1 = _resproj->outcnt();
1044 if (use->is_Initialize()) {
1045 // Eliminate Initialize node.
1046 InitializeNode *init = use->as_Initialize();
1047 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1048 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
1049 if (ctrl_proj != NULL) {
1050 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1051 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1052 }
1053 Node *mem_proj = init->proj_out(TypeFunc::Memory);
1054 if (mem_proj != NULL) {
1055 Node *mem = init->in(TypeFunc::Memory);
1056 #ifdef ASSERT
1057 if (mem->is_MergeMem()) {
1058 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1059 } else {
1060 assert(mem == _memproj_fallthrough, "allocation memory projection");
1061 }
1062 #endif
1063 _igvn.replace_node(mem_proj, mem);
1064 }
1065 } else {
1066 assert(false, "only Initialize or AddP expected");
1067 }
1068 j -= (oc1 - _resproj->outcnt());
1069 }
1070 }
1071 if (_fallthroughcatchproj != NULL) {
1072 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1073 }
1074 if (_memproj_fallthrough != NULL) {
1075 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1076 }
1077 if (_memproj_catchall != NULL) {
1078 _igvn.replace_node(_memproj_catchall, C->top());
1079 }
1080 if (_ioproj_fallthrough != NULL) {
1081 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1082 }
1083 if (_ioproj_catchall != NULL) {
1084 _igvn.replace_node(_ioproj_catchall, C->top());
1085 }
1086 if (_catchallcatchproj != NULL) {
1087 _igvn.replace_node(_catchallcatchproj, C->top());
1088 }
1089 }
1090
1091 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1092 // Don't do scalar replacement if the frame can be popped by JVMTI:
1093 // if reallocation fails during deoptimization we'll pop all
1094 // interpreter frames for this compiled frame and that won't play
1095 // nice with JVMTI popframe.
1096 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1097 return false;
1098 }
1099 Node* klass = alloc->in(AllocateNode::KlassNode);
1100 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1101 Node* res = alloc->result_cast();
1102 // Eliminate boxing allocations which are not used
1103 // regardless scalar replacable status.
1104 bool boxing_alloc = C->eliminate_boxing() &&
1105 tklass->klass()->is_instance_klass() &&
1106 tklass->klass()->as_instance_klass()->is_box_klass();
1107 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1108 return false;
1109 }
1110
1111 extract_call_projections(alloc);
1112
1113 GrowableArray <SafePointNode *> safepoints;
1114 if (!can_eliminate_allocation(alloc, safepoints)) {
1115 return false;
1116 }
1117
1118 if (!alloc->_is_scalar_replaceable) {
1119 assert(res == NULL, "sanity");
1120 // We can only eliminate allocation if all debug info references
1121 // are already replaced with SafePointScalarObject because
1122 // we can't search for a fields value without instance_id.
1123 if (safepoints.length() > 0) {
1124 return false;
1125 }
1126 }
1127
1128 if (!scalar_replacement(alloc, safepoints)) {
1129 return false;
1130 }
1131
1132 CompileLog* log = C->log();
1133 if (log != NULL) {
1134 log->head("eliminate_allocation type='%d'",
1135 log->identify(tklass->klass()));
1136 JVMState* p = alloc->jvms();
1137 while (p != NULL) {
1138 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1139 p = p->caller();
1140 }
1141 log->tail("eliminate_allocation");
1142 }
1143
1144 process_users_of_allocation(alloc);
1145
1146 #ifndef PRODUCT
1147 if (PrintEliminateAllocations) {
1148 if (alloc->is_AllocateArray())
1149 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1150 else
1151 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1152 }
1153 #endif
1154
1155 return true;
1156 }
1157
1158 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1159 // EA should remove all uses of non-escaping boxing node.
1160 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1161 return false;
1162 }
1163
1164 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1165
1166 extract_call_projections(boxing);
1167
1168 const TypeTuple* r = boxing->tf()->range();
1169 assert(r->cnt() > TypeFunc::Parms, "sanity");
1170 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1171 assert(t != NULL, "sanity");
1172
1173 CompileLog* log = C->log();
1174 if (log != NULL) {
1175 log->head("eliminate_boxing type='%d'",
1176 log->identify(t->klass()));
1177 JVMState* p = boxing->jvms();
1178 while (p != NULL) {
1179 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1180 p = p->caller();
1181 }
1182 log->tail("eliminate_boxing");
1183 }
1184
1185 process_users_of_allocation(boxing);
1186
1187 #ifndef PRODUCT
1188 if (PrintEliminateAllocations) {
1189 tty->print("++++ Eliminated: %d ", boxing->_idx);
1190 boxing->method()->print_short_name(tty);
1191 tty->cr();
1192 }
1193 #endif
1194
1195 return true;
1196 }
1197
1198 //---------------------------set_eden_pointers-------------------------
1199 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1200 if (UseTLAB) { // Private allocation: load from TLS
1201 Node* thread = transform_later(new ThreadLocalNode());
1202 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1203 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1204 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1205 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1206 } else { // Shared allocation: load from globals
1207 CollectedHeap* ch = Universe::heap();
1208 address top_adr = (address)ch->top_addr();
1209 address end_adr = (address)ch->end_addr();
1210 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1211 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1212 }
1213 }
1214
1215
1216 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1217 Node* adr = basic_plus_adr(base, offset);
1218 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1219 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1220 transform_later(value);
1221 return value;
1222 }
1223
1224
1225 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1226 Node* adr = basic_plus_adr(base, offset);
1227 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1228 transform_later(mem);
1229 return mem;
1230 }
1231
1232 //=============================================================================
1233 //
1234 // A L L O C A T I O N
1235 //
1236 // Allocation attempts to be fast in the case of frequent small objects.
1237 // It breaks down like this:
1238 //
1239 // 1) Size in doublewords is computed. This is a constant for objects and
1240 // variable for most arrays. Doubleword units are used to avoid size
1241 // overflow of huge doubleword arrays. We need doublewords in the end for
1242 // rounding.
1243 //
1244 // 2) Size is checked for being 'too large'. Too-large allocations will go
1245 // the slow path into the VM. The slow path can throw any required
1246 // exceptions, and does all the special checks for very large arrays. The
1247 // size test can constant-fold away for objects. For objects with
1248 // finalizers it constant-folds the otherway: you always go slow with
1249 // finalizers.
1250 //
1251 // 3) If NOT using TLABs, this is the contended loop-back point.
1252 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1253 //
1254 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1255 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1256 // "size*8" we always enter the VM, where "largish" is a constant picked small
1257 // enough that there's always space between the eden max and 4Gig (old space is
1258 // there so it's quite large) and large enough that the cost of entering the VM
1259 // is dwarfed by the cost to initialize the space.
1260 //
1261 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1262 // down. If contended, repeat at step 3. If using TLABs normal-store
1263 // adjusted heap top back down; there is no contention.
1264 //
1265 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1266 // fields.
1267 //
1268 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1269 // oop flavor.
1270 //
1271 //=============================================================================
1272 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1273 // Allocations bigger than this always go the slow route.
1274 // This value must be small enough that allocation attempts that need to
1275 // trigger exceptions go the slow route. Also, it must be small enough so
1276 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1277 //=============================================================================j//
1278 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1279 // The allocator will coalesce int->oop copies away. See comment in
1280 // coalesce.cpp about how this works. It depends critically on the exact
1281 // code shape produced here, so if you are changing this code shape
1282 // make sure the GC info for the heap-top is correct in and around the
1283 // slow-path call.
1284 //
1285
1286 void PhaseMacroExpand::expand_allocate_common(
1287 AllocateNode* alloc, // allocation node to be expanded
1288 Node* length, // array length for an array allocation
1289 const TypeFunc* slow_call_type, // Type of slow call
1290 address slow_call_address // Address of slow call
1291 )
1292 {
1293
1294 Node* ctrl = alloc->in(TypeFunc::Control);
1295 Node* mem = alloc->in(TypeFunc::Memory);
1296 Node* i_o = alloc->in(TypeFunc::I_O);
1297 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1298 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1299 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1300
1301 assert(ctrl != NULL, "must have control");
1302 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1303 // they will not be used if "always_slow" is set
1304 enum { slow_result_path = 1, fast_result_path = 2 };
1305 Node *result_region;
1306 Node *result_phi_rawmem;
1307 Node *result_phi_rawoop;
1308 Node *result_phi_i_o;
1309
1310 // The initial slow comparison is a size check, the comparison
1311 // we want to do is a BoolTest::gt
1312 bool always_slow = false;
1313 int tv = _igvn.find_int_con(initial_slow_test, -1);
1314 if (tv >= 0) {
1315 always_slow = (tv == 1);
1316 initial_slow_test = NULL;
1317 } else {
1318 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1319 }
1320
1321 if (C->env()->dtrace_alloc_probes() ||
1322 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1323 // Force slow-path allocation
1324 always_slow = true;
1325 initial_slow_test = NULL;
1326 }
1327
1328
1329 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1330 Node *slow_region = NULL;
1331 Node *toobig_false = ctrl;
1332
1333 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1334 // generate the initial test if necessary
1335 if (initial_slow_test != NULL ) {
1336 slow_region = new RegionNode(3);
1337
1338 // Now make the initial failure test. Usually a too-big test but
1339 // might be a TRUE for finalizers or a fancy class check for
1340 // newInstance0.
1341 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1342 transform_later(toobig_iff);
1343 // Plug the failing-too-big test into the slow-path region
1344 Node *toobig_true = new IfTrueNode( toobig_iff );
1345 transform_later(toobig_true);
1346 slow_region ->init_req( too_big_or_final_path, toobig_true );
1347 toobig_false = new IfFalseNode( toobig_iff );
1348 transform_later(toobig_false);
1349 } else { // No initial test, just fall into next case
1350 toobig_false = ctrl;
1351 debug_only(slow_region = NodeSentinel);
1352 }
1353
1354 Node *slow_mem = mem; // save the current memory state for slow path
1355 // generate the fast allocation code unless we know that the initial test will always go slow
1356 if (!always_slow) {
1357 // Fast path modifies only raw memory.
1358 if (mem->is_MergeMem()) {
1359 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1360 }
1361
1362 Node* eden_top_adr;
1363 Node* eden_end_adr;
1364
1365 set_eden_pointers(eden_top_adr, eden_end_adr);
1366
1367 // Load Eden::end. Loop invariant and hoisted.
1368 //
1369 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1370 // a TLAB to work around a bug where these values were being moved across
1371 // a safepoint. These are not oops, so they cannot be include in the oop
1372 // map, but they can be changed by a GC. The proper way to fix this would
1373 // be to set the raw memory state when generating a SafepointNode. However
1374 // this will require extensive changes to the loop optimization in order to
1375 // prevent a degradation of the optimization.
1376 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1377 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1378
1379 // allocate the Region and Phi nodes for the result
1380 result_region = new RegionNode(3);
1381 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1382 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1383 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1384
1385 // We need a Region for the loop-back contended case.
1386 enum { fall_in_path = 1, contended_loopback_path = 2 };
1387 Node *contended_region;
1388 Node *contended_phi_rawmem;
1389 if (UseTLAB) {
1390 contended_region = toobig_false;
1391 contended_phi_rawmem = mem;
1392 } else {
1393 contended_region = new RegionNode(3);
1394 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1395 // Now handle the passing-too-big test. We fall into the contended
1396 // loop-back merge point.
1397 contended_region ->init_req(fall_in_path, toobig_false);
1398 contended_phi_rawmem->init_req(fall_in_path, mem);
1399 transform_later(contended_region);
1400 transform_later(contended_phi_rawmem);
1401 }
1402
1403 // Load(-locked) the heap top.
1404 // See note above concerning the control input when using a TLAB
1405 Node *old_eden_top = UseTLAB
1406 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1407 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1408
1409 transform_later(old_eden_top);
1410 // Add to heap top to get a new heap top
1411 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1412 transform_later(new_eden_top);
1413 // Check for needing a GC; compare against heap end
1414 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1415 transform_later(needgc_cmp);
1416 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1417 transform_later(needgc_bol);
1418 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1419 transform_later(needgc_iff);
1420
1421 // Plug the failing-heap-space-need-gc test into the slow-path region
1422 Node *needgc_true = new IfTrueNode(needgc_iff);
1423 transform_later(needgc_true);
1424 if (initial_slow_test) {
1425 slow_region->init_req(need_gc_path, needgc_true);
1426 // This completes all paths into the slow merge point
1427 transform_later(slow_region);
1428 } else { // No initial slow path needed!
1429 // Just fall from the need-GC path straight into the VM call.
1430 slow_region = needgc_true;
1431 }
1432 // No need for a GC. Setup for the Store-Conditional
1433 Node *needgc_false = new IfFalseNode(needgc_iff);
1434 transform_later(needgc_false);
1435
1436 // Grab regular I/O before optional prefetch may change it.
1437 // Slow-path does no I/O so just set it to the original I/O.
1438 result_phi_i_o->init_req(slow_result_path, i_o);
1439
1440 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1441 old_eden_top, new_eden_top, length);
1442
1443 // Name successful fast-path variables
1444 Node* fast_oop = old_eden_top;
1445 Node* fast_oop_ctrl;
1446 Node* fast_oop_rawmem;
1447
1448 // Store (-conditional) the modified eden top back down.
1449 // StorePConditional produces flags for a test PLUS a modified raw
1450 // memory state.
1451 if (UseTLAB) {
1452 Node* store_eden_top =
1453 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1454 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1455 transform_later(store_eden_top);
1456 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1457 fast_oop_rawmem = store_eden_top;
1458 } else {
1459 Node* store_eden_top =
1460 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1461 new_eden_top, fast_oop/*old_eden_top*/);
1462 transform_later(store_eden_top);
1463 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1464 transform_later(contention_check);
1465 store_eden_top = new SCMemProjNode(store_eden_top);
1466 transform_later(store_eden_top);
1467
1468 // If not using TLABs, check to see if there was contention.
1469 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1470 transform_later(contention_iff);
1471 Node *contention_true = new IfTrueNode(contention_iff);
1472 transform_later(contention_true);
1473 // If contention, loopback and try again.
1474 contended_region->init_req(contended_loopback_path, contention_true);
1475 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1476
1477 // Fast-path succeeded with no contention!
1478 Node *contention_false = new IfFalseNode(contention_iff);
1479 transform_later(contention_false);
1480 fast_oop_ctrl = contention_false;
1481
1482 // Bump total allocated bytes for this thread
1483 Node* thread = new ThreadLocalNode();
1484 transform_later(thread);
1485 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1486 in_bytes(JavaThread::allocated_bytes_offset()));
1487 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1488 0, TypeLong::LONG, T_LONG);
1489 #ifdef _LP64
1490 Node* alloc_size = size_in_bytes;
1491 #else
1492 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1493 transform_later(alloc_size);
1494 #endif
1495 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1496 transform_later(new_alloc_bytes);
1497 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1498 0, new_alloc_bytes, T_LONG);
1499 }
1500
1501 InitializeNode* init = alloc->initialization();
1502 fast_oop_rawmem = initialize_object(alloc,
1503 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1504 klass_node, length, size_in_bytes);
1505
1506 // If initialization is performed by an array copy, any required
1507 // MemBarStoreStore was already added. If the object does not
1508 // escape no need for a MemBarStoreStore. Otherwise we need a
1509 // MemBarStoreStore so that stores that initialize this object
1510 // can't be reordered with a subsequent store that makes this
1511 // object accessible by other threads.
1512 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1513 if (init == NULL || init->req() < InitializeNode::RawStores) {
1514 // No InitializeNode or no stores captured by zeroing
1515 // elimination. Simply add the MemBarStoreStore after object
1516 // initialization.
1517 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1518 transform_later(mb);
1519
1520 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1521 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1522 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1523 transform_later(fast_oop_ctrl);
1524 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1525 transform_later(fast_oop_rawmem);
1526 } else {
1527 // Add the MemBarStoreStore after the InitializeNode so that
1528 // all stores performing the initialization that were moved
1529 // before the InitializeNode happen before the storestore
1530 // barrier.
1531
1532 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1533 Node* init_mem = init->proj_out(TypeFunc::Memory);
1534
1535 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1536 transform_later(mb);
1537
1538 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1539 transform_later(ctrl);
1540 Node* mem = new ProjNode(init,TypeFunc::Memory);
1541 transform_later(mem);
1542
1543 // The MemBarStoreStore depends on control and memory coming
1544 // from the InitializeNode
1545 mb->init_req(TypeFunc::Memory, mem);
1546 mb->init_req(TypeFunc::Control, ctrl);
1547
1548 ctrl = new ProjNode(mb,TypeFunc::Control);
1549 transform_later(ctrl);
1550 mem = new ProjNode(mb,TypeFunc::Memory);
1551 transform_later(mem);
1552
1553 // All nodes that depended on the InitializeNode for control
1554 // and memory must now depend on the MemBarNode that itself
1555 // depends on the InitializeNode
1556 _igvn.replace_node(init_ctrl, ctrl);
1557 _igvn.replace_node(init_mem, mem);
1558 }
1559 }
1560
1561 if (C->env()->dtrace_extended_probes()) {
1562 // Slow-path call
1563 int size = TypeFunc::Parms + 2;
1564 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1565 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1566 "dtrace_object_alloc",
1567 TypeRawPtr::BOTTOM);
1568
1569 // Get base of thread-local storage area
1570 Node* thread = new ThreadLocalNode();
1571 transform_later(thread);
1572
1573 call->init_req(TypeFunc::Parms+0, thread);
1574 call->init_req(TypeFunc::Parms+1, fast_oop);
1575 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1576 call->init_req(TypeFunc::I_O , top()); // does no i/o
1577 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1578 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1579 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1580 transform_later(call);
1581 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1582 transform_later(fast_oop_ctrl);
1583 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1584 transform_later(fast_oop_rawmem);
1585 }
1586
1587 // Plug in the successful fast-path into the result merge point
1588 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1589 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1590 result_phi_i_o ->init_req(fast_result_path, i_o);
1591 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1592 } else {
1593 slow_region = ctrl;
1594 result_phi_i_o = i_o; // Rename it to use in the following code.
1595 }
1596
1597 // Generate slow-path call
1598 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1599 OptoRuntime::stub_name(slow_call_address),
1600 alloc->jvms()->bci(),
1601 TypePtr::BOTTOM);
1602 call->init_req( TypeFunc::Control, slow_region );
1603 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1604 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1605 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1606 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1607
1608 call->init_req(TypeFunc::Parms+0, klass_node);
1609 if (length != NULL) {
1610 call->init_req(TypeFunc::Parms+1, length);
1611 }
1612
1613 // Copy debug information and adjust JVMState information, then replace
1614 // allocate node with the call
1615 copy_call_debug_info((CallNode *) alloc, call);
1616 if (!always_slow) {
1617 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1618 } else {
1619 // Hook i_o projection to avoid its elimination during allocation
1620 // replacement (when only a slow call is generated).
1621 call->set_req(TypeFunc::I_O, result_phi_i_o);
1622 }
1623 _igvn.replace_node(alloc, call);
1624 transform_later(call);
1625
1626 // Identify the output projections from the allocate node and
1627 // adjust any references to them.
1628 // The control and io projections look like:
1629 //
1630 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1631 // Allocate Catch
1632 // ^---Proj(io) <-------+ ^---CatchProj(io)
1633 //
1634 // We are interested in the CatchProj nodes.
1635 //
1636 extract_call_projections(call);
1637
1638 // An allocate node has separate memory projections for the uses on
1639 // the control and i_o paths. Replace the control memory projection with
1640 // result_phi_rawmem (unless we are only generating a slow call when
1641 // both memory projections are combined)
1642 if (!always_slow && _memproj_fallthrough != NULL) {
1643 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1644 Node *use = _memproj_fallthrough->fast_out(i);
1645 _igvn.rehash_node_delayed(use);
1646 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1647 // back up iterator
1648 --i;
1649 }
1650 }
1651 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1652 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1653 if (_memproj_catchall != NULL ) {
1654 if (_memproj_fallthrough == NULL) {
1655 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1656 transform_later(_memproj_fallthrough);
1657 }
1658 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1659 Node *use = _memproj_catchall->fast_out(i);
1660 _igvn.rehash_node_delayed(use);
1661 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1662 // back up iterator
1663 --i;
1664 }
1665 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1666 _igvn.remove_dead_node(_memproj_catchall);
1667 }
1668
1669 // An allocate node has separate i_o projections for the uses on the control
1670 // and i_o paths. Always replace the control i_o projection with result i_o
1671 // otherwise incoming i_o become dead when only a slow call is generated
1672 // (it is different from memory projections where both projections are
1673 // combined in such case).
1674 if (_ioproj_fallthrough != NULL) {
1675 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1676 Node *use = _ioproj_fallthrough->fast_out(i);
1677 _igvn.rehash_node_delayed(use);
1678 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1679 // back up iterator
1680 --i;
1681 }
1682 }
1683 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1684 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1685 if (_ioproj_catchall != NULL ) {
1686 if (_ioproj_fallthrough == NULL) {
1687 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1688 transform_later(_ioproj_fallthrough);
1689 }
1690 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1691 Node *use = _ioproj_catchall->fast_out(i);
1692 _igvn.rehash_node_delayed(use);
1693 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1694 // back up iterator
1695 --i;
1696 }
1697 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1698 _igvn.remove_dead_node(_ioproj_catchall);
1699 }
1700
1701 // if we generated only a slow call, we are done
1702 if (always_slow) {
1703 // Now we can unhook i_o.
1704 if (result_phi_i_o->outcnt() > 1) {
1705 call->set_req(TypeFunc::I_O, top());
1706 } else {
1707 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1708 // Case of new array with negative size known during compilation.
1709 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1710 // following code since call to runtime will throw exception.
1711 // As result there will be no users of i_o after the call.
1712 // Leave i_o attached to this call to avoid problems in preceding graph.
1713 }
1714 return;
1715 }
1716
1717
1718 if (_fallthroughcatchproj != NULL) {
1719 ctrl = _fallthroughcatchproj->clone();
1720 transform_later(ctrl);
1721 _igvn.replace_node(_fallthroughcatchproj, result_region);
1722 } else {
1723 ctrl = top();
1724 }
1725 Node *slow_result;
1726 if (_resproj == NULL) {
1727 // no uses of the allocation result
1728 slow_result = top();
1729 } else {
1730 slow_result = _resproj->clone();
1731 transform_later(slow_result);
1732 _igvn.replace_node(_resproj, result_phi_rawoop);
1733 }
1734
1735 // Plug slow-path into result merge point
1736 result_region ->init_req( slow_result_path, ctrl );
1737 result_phi_rawoop->init_req( slow_result_path, slow_result);
1738 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1739 transform_later(result_region);
1740 transform_later(result_phi_rawoop);
1741 transform_later(result_phi_rawmem);
1742 transform_later(result_phi_i_o);
1743 // This completes all paths into the result merge point
1744 }
1745
1746
1747 // Helper for PhaseMacroExpand::expand_allocate_common.
1748 // Initializes the newly-allocated storage.
1749 Node*
1750 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1751 Node* control, Node* rawmem, Node* object,
1752 Node* klass_node, Node* length,
1753 Node* size_in_bytes) {
1754 InitializeNode* init = alloc->initialization();
1755 // Store the klass & mark bits
1756 Node* mark_node = NULL;
1757 // For now only enable fast locking for non-array types
1758 if (UseBiasedLocking && (length == NULL)) {
1759 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1760 } else {
1761 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1762 }
1763 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1764
1765 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1766 int header_size = alloc->minimum_header_size(); // conservatively small
1767
1768 // Array length
1769 if (length != NULL) { // Arrays need length field
1770 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1771 // conservatively small header size:
1772 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1773 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1774 if (k->is_array_klass()) // we know the exact header size in most cases:
1775 header_size = Klass::layout_helper_header_size(k->layout_helper());
1776 }
1777
1778 // Clear the object body, if necessary.
1779 if (init == NULL) {
1780 // The init has somehow disappeared; be cautious and clear everything.
1781 //
1782 // This can happen if a node is allocated but an uncommon trap occurs
1783 // immediately. In this case, the Initialize gets associated with the
1784 // trap, and may be placed in a different (outer) loop, if the Allocate
1785 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1786 // there can be two Allocates to one Initialize. The answer in all these
1787 // edge cases is safety first. It is always safe to clear immediately
1788 // within an Allocate, and then (maybe or maybe not) clear some more later.
1789 if (!ZeroTLAB)
1790 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1791 header_size, size_in_bytes,
1792 &_igvn);
1793 } else {
1794 if (!init->is_complete()) {
1795 // Try to win by zeroing only what the init does not store.
1796 // We can also try to do some peephole optimizations,
1797 // such as combining some adjacent subword stores.
1798 rawmem = init->complete_stores(control, rawmem, object,
1799 header_size, size_in_bytes, &_igvn);
1800 }
1801 // We have no more use for this link, since the AllocateNode goes away:
1802 init->set_req(InitializeNode::RawAddress, top());
1803 // (If we keep the link, it just confuses the register allocator,
1804 // who thinks he sees a real use of the address by the membar.)
1805 }
1806
1807 return rawmem;
1808 }
1809
1810 // Generate prefetch instructions for next allocations.
1811 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1812 Node*& contended_phi_rawmem,
1813 Node* old_eden_top, Node* new_eden_top,
1814 Node* length) {
1815 enum { fall_in_path = 1, pf_path = 2 };
1816 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1817 // Generate prefetch allocation with watermark check.
1818 // As an allocation hits the watermark, we will prefetch starting
1819 // at a "distance" away from watermark.
1820
1821 Node *pf_region = new RegionNode(3);
1822 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1823 TypeRawPtr::BOTTOM );
1824 // I/O is used for Prefetch
1825 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1826
1827 Node *thread = new ThreadLocalNode();
1828 transform_later(thread);
1829
1830 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1831 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1832 transform_later(eden_pf_adr);
1833
1834 Node *old_pf_wm = new LoadPNode(needgc_false,
1835 contended_phi_rawmem, eden_pf_adr,
1836 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1837 MemNode::unordered);
1838 transform_later(old_pf_wm);
1839
1840 // check against new_eden_top
1841 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1842 transform_later(need_pf_cmp);
1843 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1844 transform_later(need_pf_bol);
1845 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1846 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1847 transform_later(need_pf_iff);
1848
1849 // true node, add prefetchdistance
1850 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1851 transform_later(need_pf_true);
1852
1853 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1854 transform_later(need_pf_false);
1855
1856 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1857 _igvn.MakeConX(AllocatePrefetchDistance) );
1858 transform_later(new_pf_wmt );
1859 new_pf_wmt->set_req(0, need_pf_true);
1860
1861 Node *store_new_wmt = new StorePNode(need_pf_true,
1862 contended_phi_rawmem, eden_pf_adr,
1863 TypeRawPtr::BOTTOM, new_pf_wmt,
1864 MemNode::unordered);
1865 transform_later(store_new_wmt);
1866
1867 // adding prefetches
1868 pf_phi_abio->init_req( fall_in_path, i_o );
1869
1870 Node *prefetch_adr;
1871 Node *prefetch;
1872 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1873 uint step_size = AllocatePrefetchStepSize;
1874 uint distance = 0;
1875
1876 for ( uint i = 0; i < lines; i++ ) {
1877 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1878 _igvn.MakeConX(distance) );
1879 transform_later(prefetch_adr);
1880 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1881 transform_later(prefetch);
1882 distance += step_size;
1883 i_o = prefetch;
1884 }
1885 pf_phi_abio->set_req( pf_path, i_o );
1886
1887 pf_region->init_req( fall_in_path, need_pf_false );
1888 pf_region->init_req( pf_path, need_pf_true );
1889
1890 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1891 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1892
1893 transform_later(pf_region);
1894 transform_later(pf_phi_rawmem);
1895 transform_later(pf_phi_abio);
1896
1897 needgc_false = pf_region;
1898 contended_phi_rawmem = pf_phi_rawmem;
1899 i_o = pf_phi_abio;
1900 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1901 // Insert a prefetch for each allocation.
1902 // This code is used for Sparc with BIS.
1903 Node *pf_region = new RegionNode(3);
1904 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1905 TypeRawPtr::BOTTOM );
1906 transform_later(pf_region);
1907
1908 // Generate several prefetch instructions.
1909 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1910 uint step_size = AllocatePrefetchStepSize;
1911 uint distance = AllocatePrefetchDistance;
1912
1913 // Next cache address.
1914 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1915 _igvn.MakeConX(distance));
1916 transform_later(cache_adr);
1917 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1918 transform_later(cache_adr);
1919 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1920 cache_adr = new AndXNode(cache_adr, mask);
1921 transform_later(cache_adr);
1922 cache_adr = new CastX2PNode(cache_adr);
1923 transform_later(cache_adr);
1924
1925 // Prefetch
1926 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1927 prefetch->set_req(0, needgc_false);
1928 transform_later(prefetch);
1929 contended_phi_rawmem = prefetch;
1930 Node *prefetch_adr;
1931 distance = step_size;
1932 for ( uint i = 1; i < lines; i++ ) {
1933 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1934 _igvn.MakeConX(distance) );
1935 transform_later(prefetch_adr);
1936 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1937 transform_later(prefetch);
1938 distance += step_size;
1939 contended_phi_rawmem = prefetch;
1940 }
1941 } else if( AllocatePrefetchStyle > 0 ) {
1942 // Insert a prefetch for each allocation only on the fast-path
1943 Node *prefetch_adr;
1944 Node *prefetch;
1945 // Generate several prefetch instructions.
1946 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1947 uint step_size = AllocatePrefetchStepSize;
1948 uint distance = AllocatePrefetchDistance;
1949 for ( uint i = 0; i < lines; i++ ) {
1950 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1951 _igvn.MakeConX(distance) );
1952 transform_later(prefetch_adr);
1953 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1954 // Do not let it float too high, since if eden_top == eden_end,
1955 // both might be null.
1956 if( i == 0 ) { // Set control for first prefetch, next follows it
1957 prefetch->init_req(0, needgc_false);
1958 }
1959 transform_later(prefetch);
1960 distance += step_size;
1961 i_o = prefetch;
1962 }
1963 }
1964 return i_o;
1965 }
1966
1967
1968 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1969 expand_allocate_common(alloc, NULL,
1970 OptoRuntime::new_instance_Type(),
1971 OptoRuntime::new_instance_Java());
1972 }
1973
1974 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1975 Node* length = alloc->in(AllocateNode::ALength);
1976 InitializeNode* init = alloc->initialization();
1977 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1978 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1979 address slow_call_address; // Address of slow call
1980 if (init != NULL && init->is_complete_with_arraycopy() &&
1981 k->is_type_array_klass()) {
1982 // Don't zero type array during slow allocation in VM since
1983 // it will be initialized later by arraycopy in compiled code.
1984 slow_call_address = OptoRuntime::new_array_nozero_Java();
1985 } else {
1986 slow_call_address = OptoRuntime::new_array_Java();
1987 }
1988 expand_allocate_common(alloc, length,
1989 OptoRuntime::new_array_Type(),
1990 slow_call_address);
1991 }
1992
1993 //-------------------mark_eliminated_box----------------------------------
1994 //
1995 // During EA obj may point to several objects but after few ideal graph
1996 // transformations (CCP) it may point to only one non escaping object
1997 // (but still using phi), corresponding locks and unlocks will be marked
1998 // for elimination. Later obj could be replaced with a new node (new phi)
1999 // and which does not have escape information. And later after some graph
2000 // reshape other locks and unlocks (which were not marked for elimination
2001 // before) are connected to this new obj (phi) but they still will not be
2002 // marked for elimination since new obj has no escape information.
2003 // Mark all associated (same box and obj) lock and unlock nodes for
2004 // elimination if some of them marked already.
2005 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2006 if (oldbox->as_BoxLock()->is_eliminated())
2007 return; // This BoxLock node was processed already.
2008
2009 // New implementation (EliminateNestedLocks) has separate BoxLock
2010 // node for each locked region so mark all associated locks/unlocks as
2011 // eliminated even if different objects are referenced in one locked region
2012 // (for example, OSR compilation of nested loop inside locked scope).
2013 if (EliminateNestedLocks ||
2014 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2015 // Box is used only in one lock region. Mark this box as eliminated.
2016 _igvn.hash_delete(oldbox);
2017 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2018 _igvn.hash_insert(oldbox);
2019
2020 for (uint i = 0; i < oldbox->outcnt(); i++) {
2021 Node* u = oldbox->raw_out(i);
2022 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2023 AbstractLockNode* alock = u->as_AbstractLock();
2024 // Check lock's box since box could be referenced by Lock's debug info.
2025 if (alock->box_node() == oldbox) {
2026 // Mark eliminated all related locks and unlocks.
2027 #ifdef ASSERT
2028 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2029 #endif
2030 alock->set_non_esc_obj();
2031 }
2032 }
2033 }
2034 return;
2035 }
2036
2037 // Create new "eliminated" BoxLock node and use it in monitor debug info
2038 // instead of oldbox for the same object.
2039 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2040
2041 // Note: BoxLock node is marked eliminated only here and it is used
2042 // to indicate that all associated lock and unlock nodes are marked
2043 // for elimination.
2044 newbox->set_eliminated();
2045 transform_later(newbox);
2046
2047 // Replace old box node with new box for all users of the same object.
2048 for (uint i = 0; i < oldbox->outcnt();) {
2049 bool next_edge = true;
2050
2051 Node* u = oldbox->raw_out(i);
2052 if (u->is_AbstractLock()) {
2053 AbstractLockNode* alock = u->as_AbstractLock();
2054 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2055 // Replace Box and mark eliminated all related locks and unlocks.
2056 #ifdef ASSERT
2057 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2058 #endif
2059 alock->set_non_esc_obj();
2060 _igvn.rehash_node_delayed(alock);
2061 alock->set_box_node(newbox);
2062 next_edge = false;
2063 }
2064 }
2065 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2066 FastLockNode* flock = u->as_FastLock();
2067 assert(flock->box_node() == oldbox, "sanity");
2068 _igvn.rehash_node_delayed(flock);
2069 flock->set_box_node(newbox);
2070 next_edge = false;
2071 }
2072
2073 // Replace old box in monitor debug info.
2074 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2075 SafePointNode* sfn = u->as_SafePoint();
2076 JVMState* youngest_jvms = sfn->jvms();
2077 int max_depth = youngest_jvms->depth();
2078 for (int depth = 1; depth <= max_depth; depth++) {
2079 JVMState* jvms = youngest_jvms->of_depth(depth);
2080 int num_mon = jvms->nof_monitors();
2081 // Loop over monitors
2082 for (int idx = 0; idx < num_mon; idx++) {
2083 Node* obj_node = sfn->monitor_obj(jvms, idx);
2084 Node* box_node = sfn->monitor_box(jvms, idx);
2085 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2086 int j = jvms->monitor_box_offset(idx);
2087 _igvn.replace_input_of(u, j, newbox);
2088 next_edge = false;
2089 }
2090 }
2091 }
2092 }
2093 if (next_edge) i++;
2094 }
2095 }
2096
2097 //-----------------------mark_eliminated_locking_nodes-----------------------
2098 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2099 if (EliminateNestedLocks) {
2100 if (alock->is_nested()) {
2101 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2102 return;
2103 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2104 // Only Lock node has JVMState needed here.
2105 // Not that preceding claim is documented anywhere else.
2106 if (alock->jvms() != NULL) {
2107 if (alock->as_Lock()->is_nested_lock_region()) {
2108 // Mark eliminated related nested locks and unlocks.
2109 Node* obj = alock->obj_node();
2110 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2111 assert(!box_node->is_eliminated(), "should not be marked yet");
2112 // Note: BoxLock node is marked eliminated only here
2113 // and it is used to indicate that all associated lock
2114 // and unlock nodes are marked for elimination.
2115 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2116 for (uint i = 0; i < box_node->outcnt(); i++) {
2117 Node* u = box_node->raw_out(i);
2118 if (u->is_AbstractLock()) {
2119 alock = u->as_AbstractLock();
2120 if (alock->box_node() == box_node) {
2121 // Verify that this Box is referenced only by related locks.
2122 assert(alock->obj_node()->eqv_uncast(obj), "");
2123 // Mark all related locks and unlocks.
2124 #ifdef ASSERT
2125 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2126 #endif
2127 alock->set_nested();
2128 }
2129 }
2130 }
2131 } else {
2132 #ifdef ASSERT
2133 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2134 if (C->log() != NULL)
2135 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2136 #endif
2137 }
2138 }
2139 return;
2140 }
2141 // Process locks for non escaping object
2142 assert(alock->is_non_esc_obj(), "");
2143 } // EliminateNestedLocks
2144
2145 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2146 // Look for all locks of this object and mark them and
2147 // corresponding BoxLock nodes as eliminated.
2148 Node* obj = alock->obj_node();
2149 for (uint j = 0; j < obj->outcnt(); j++) {
2150 Node* o = obj->raw_out(j);
2151 if (o->is_AbstractLock() &&
2152 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2153 alock = o->as_AbstractLock();
2154 Node* box = alock->box_node();
2155 // Replace old box node with new eliminated box for all users
2156 // of the same object and mark related locks as eliminated.
2157 mark_eliminated_box(box, obj);
2158 }
2159 }
2160 }
2161 }
2162
2163 // we have determined that this lock/unlock can be eliminated, we simply
2164 // eliminate the node without expanding it.
2165 //
2166 // Note: The membar's associated with the lock/unlock are currently not
2167 // eliminated. This should be investigated as a future enhancement.
2168 //
2169 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2170
2171 if (!alock->is_eliminated()) {
2172 return false;
2173 }
2174 #ifdef ASSERT
2175 if (!alock->is_coarsened()) {
2176 // Check that new "eliminated" BoxLock node is created.
2177 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2178 assert(oldbox->is_eliminated(), "should be done already");
2179 }
2180 #endif
2181
2182 alock->log_lock_optimization(C, "eliminate_lock");
2183
2184 #ifndef PRODUCT
2185 if (PrintEliminateLocks) {
2186 if (alock->is_Lock()) {
2187 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2188 } else {
2189 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2190 }
2191 }
2192 #endif
2193
2194 Node* mem = alock->in(TypeFunc::Memory);
2195 Node* ctrl = alock->in(TypeFunc::Control);
2196
2197 extract_call_projections(alock);
2198 // There are 2 projections from the lock. The lock node will
2199 // be deleted when its last use is subsumed below.
2200 assert(alock->outcnt() == 2 &&
2201 _fallthroughproj != NULL &&
2202 _memproj_fallthrough != NULL,
2203 "Unexpected projections from Lock/Unlock");
2204
2205 Node* fallthroughproj = _fallthroughproj;
2206 Node* memproj_fallthrough = _memproj_fallthrough;
2207
2208 // The memory projection from a lock/unlock is RawMem
2209 // The input to a Lock is merged memory, so extract its RawMem input
2210 // (unless the MergeMem has been optimized away.)
2211 if (alock->is_Lock()) {
2212 // Seach for MemBarAcquireLock node and delete it also.
2213 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2214 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2215 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2216 Node* memproj = membar->proj_out(TypeFunc::Memory);
2217 _igvn.replace_node(ctrlproj, fallthroughproj);
2218 _igvn.replace_node(memproj, memproj_fallthrough);
2219
2220 // Delete FastLock node also if this Lock node is unique user
2221 // (a loop peeling may clone a Lock node).
2222 Node* flock = alock->as_Lock()->fastlock_node();
2223 if (flock->outcnt() == 1) {
2224 assert(flock->unique_out() == alock, "sanity");
2225 _igvn.replace_node(flock, top());
2226 }
2227 }
2228
2229 // Seach for MemBarReleaseLock node and delete it also.
2230 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2231 ctrl->in(0)->is_MemBar()) {
2232 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2233 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2234 mem->is_Proj() && membar == mem->in(0), "");
2235 _igvn.replace_node(fallthroughproj, ctrl);
2236 _igvn.replace_node(memproj_fallthrough, mem);
2237 fallthroughproj = ctrl;
2238 memproj_fallthrough = mem;
2239 ctrl = membar->in(TypeFunc::Control);
2240 mem = membar->in(TypeFunc::Memory);
2241 }
2242
2243 _igvn.replace_node(fallthroughproj, ctrl);
2244 _igvn.replace_node(memproj_fallthrough, mem);
2245 return true;
2246 }
2247
2248
2249 //------------------------------expand_lock_node----------------------
2250 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2251
2252 Node* ctrl = lock->in(TypeFunc::Control);
2253 Node* mem = lock->in(TypeFunc::Memory);
2254 Node* obj = lock->obj_node();
2255 Node* box = lock->box_node();
2256 Node* flock = lock->fastlock_node();
2257
2258 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2259
2260 // Make the merge point
2261 Node *region;
2262 Node *mem_phi;
2263 Node *slow_path;
2264
2265 if (UseOptoBiasInlining) {
2266 /*
2267 * See the full description in MacroAssembler::biased_locking_enter().
2268 *
2269 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2270 * // The object is biased.
2271 * proto_node = klass->prototype_header;
2272 * o_node = thread | proto_node;
2273 * x_node = o_node ^ mark_word;
2274 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2275 * // Done.
2276 * } else {
2277 * if( (x_node & biased_lock_mask) != 0 ) {
2278 * // The klass's prototype header is no longer biased.
2279 * cas(&mark_word, mark_word, proto_node)
2280 * goto cas_lock;
2281 * } else {
2282 * // The klass's prototype header is still biased.
2283 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2284 * old = mark_word;
2285 * new = o_node;
2286 * } else {
2287 * // Different thread or anonymous biased.
2288 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2289 * new = thread | old;
2290 * }
2291 * // Try to rebias.
2292 * if( cas(&mark_word, old, new) == 0 ) {
2293 * // Done.
2294 * } else {
2295 * goto slow_path; // Failed.
2296 * }
2297 * }
2298 * }
2299 * } else {
2300 * // The object is not biased.
2301 * cas_lock:
2302 * if( FastLock(obj) == 0 ) {
2303 * // Done.
2304 * } else {
2305 * slow_path:
2306 * OptoRuntime::complete_monitor_locking_Java(obj);
2307 * }
2308 * }
2309 */
2310
2311 region = new RegionNode(5);
2312 // create a Phi for the memory state
2313 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2314
2315 Node* fast_lock_region = new RegionNode(3);
2316 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2317
2318 // First, check mark word for the biased lock pattern.
2319 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2320
2321 // Get fast path - mark word has the biased lock pattern.
2322 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2323 markOopDesc::biased_lock_mask_in_place,
2324 markOopDesc::biased_lock_pattern, true);
2325 // fast_lock_region->in(1) is set to slow path.
2326 fast_lock_mem_phi->init_req(1, mem);
2327
2328 // Now check that the lock is biased to the current thread and has
2329 // the same epoch and bias as Klass::_prototype_header.
2330
2331 // Special-case a fresh allocation to avoid building nodes:
2332 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2333 if (klass_node == NULL) {
2334 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2335 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2336 #ifdef _LP64
2337 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2338 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2339 klass_node->in(1)->init_req(0, ctrl);
2340 } else
2341 #endif
2342 klass_node->init_req(0, ctrl);
2343 }
2344 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2345
2346 Node* thread = transform_later(new ThreadLocalNode());
2347 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2348 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2349 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2350
2351 // Get slow path - mark word does NOT match the value.
2352 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2353 (~markOopDesc::age_mask_in_place), 0);
2354 // region->in(3) is set to fast path - the object is biased to the current thread.
2355 mem_phi->init_req(3, mem);
2356
2357
2358 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2359
2360
2361 // First, check biased pattern.
2362 // Get fast path - _prototype_header has the same biased lock pattern.
2363 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2364 markOopDesc::biased_lock_mask_in_place, 0, true);
2365
2366 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2367 // fast_lock_region->in(2) - the prototype header is no longer biased
2368 // and we have to revoke the bias on this object.
2369 // We are going to try to reset the mark of this object to the prototype
2370 // value and fall through to the CAS-based locking scheme.
2371 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2372 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2373 proto_node, mark_node);
2374 transform_later(cas);
2375 Node* proj = transform_later(new SCMemProjNode(cas));
2376 fast_lock_mem_phi->init_req(2, proj);
2377
2378
2379 // Second, check epoch bits.
2380 Node* rebiased_region = new RegionNode(3);
2381 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2382 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2383
2384 // Get slow path - mark word does NOT match epoch bits.
2385 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2386 markOopDesc::epoch_mask_in_place, 0);
2387 // The epoch of the current bias is not valid, attempt to rebias the object
2388 // toward the current thread.
2389 rebiased_region->init_req(2, epoch_ctrl);
2390 old_phi->init_req(2, mark_node);
2391 new_phi->init_req(2, o_node);
2392
2393 // rebiased_region->in(1) is set to fast path.
2394 // The epoch of the current bias is still valid but we know
2395 // nothing about the owner; it might be set or it might be clear.
2396 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2397 markOopDesc::age_mask_in_place |
2398 markOopDesc::epoch_mask_in_place);
2399 Node* old = transform_later(new AndXNode(mark_node, cmask));
2400 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2401 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2402 old_phi->init_req(1, old);
2403 new_phi->init_req(1, new_mark);
2404
2405 transform_later(rebiased_region);
2406 transform_later(old_phi);
2407 transform_later(new_phi);
2408
2409 // Try to acquire the bias of the object using an atomic operation.
2410 // If this fails we will go in to the runtime to revoke the object's bias.
2411 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2412 transform_later(cas);
2413 proj = transform_later(new SCMemProjNode(cas));
2414
2415 // Get slow path - Failed to CAS.
2416 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2417 mem_phi->init_req(4, proj);
2418 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2419
2420 // Failed to CAS.
2421 slow_path = new RegionNode(3);
2422 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2423
2424 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2425 slow_mem->init_req(1, proj);
2426
2427 // Call CAS-based locking scheme (FastLock node).
2428
2429 transform_later(fast_lock_region);
2430 transform_later(fast_lock_mem_phi);
2431
2432 // Get slow path - FastLock failed to lock the object.
2433 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2434 mem_phi->init_req(2, fast_lock_mem_phi);
2435 // region->in(2) is set to fast path - the object is locked to the current thread.
2436
2437 slow_path->init_req(2, ctrl); // Capture slow-control
2438 slow_mem->init_req(2, fast_lock_mem_phi);
2439
2440 transform_later(slow_path);
2441 transform_later(slow_mem);
2442 // Reset lock's memory edge.
2443 lock->set_req(TypeFunc::Memory, slow_mem);
2444
2445 } else {
2446 region = new RegionNode(3);
2447 // create a Phi for the memory state
2448 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2449
2450 // Optimize test; set region slot 2
2451 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2452 mem_phi->init_req(2, mem);
2453 }
2454
2455 // Make slow path call
2456 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2457 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2458 obj, box, NULL);
2459
2460 extract_call_projections(call);
2461
2462 // Slow path can only throw asynchronous exceptions, which are always
2463 // de-opted. So the compiler thinks the slow-call can never throw an
2464 // exception. If it DOES throw an exception we would need the debug
2465 // info removed first (since if it throws there is no monitor).
2466 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2467 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2468
2469 // Capture slow path
2470 // disconnect fall-through projection from call and create a new one
2471 // hook up users of fall-through projection to region
2472 Node *slow_ctrl = _fallthroughproj->clone();
2473 transform_later(slow_ctrl);
2474 _igvn.hash_delete(_fallthroughproj);
2475 _fallthroughproj->disconnect_inputs(NULL, C);
2476 region->init_req(1, slow_ctrl);
2477 // region inputs are now complete
2478 transform_later(region);
2479 _igvn.replace_node(_fallthroughproj, region);
2480
2481 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2482 mem_phi->init_req(1, memproj );
2483 transform_later(mem_phi);
2484 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2485 }
2486
2487 //------------------------------expand_unlock_node----------------------
2488 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2489
2490 Node* ctrl = unlock->in(TypeFunc::Control);
2491 Node* mem = unlock->in(TypeFunc::Memory);
2492 Node* obj = unlock->obj_node();
2493 Node* box = unlock->box_node();
2494
2495 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2496
2497 // No need for a null check on unlock
2498
2499 // Make the merge point
2500 Node *region;
2501 Node *mem_phi;
2502
2503 if (UseOptoBiasInlining) {
2504 // Check for biased locking unlock case, which is a no-op.
2505 // See the full description in MacroAssembler::biased_locking_exit().
2506 region = new RegionNode(4);
2507 // create a Phi for the memory state
2508 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2509 mem_phi->init_req(3, mem);
2510
2511 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2512 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2513 markOopDesc::biased_lock_mask_in_place,
2514 markOopDesc::biased_lock_pattern);
2515 } else {
2516 region = new RegionNode(3);
2517 // create a Phi for the memory state
2518 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2519 }
2520
2521 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2522 funlock = transform_later( funlock )->as_FastUnlock();
2523 // Optimize test; set region slot 2
2524 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2525 Node *thread = transform_later(new ThreadLocalNode());
2526
2527 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2528 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2529 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2530
2531 extract_call_projections(call);
2532
2533 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2534 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2535
2536 // No exceptions for unlocking
2537 // Capture slow path
2538 // disconnect fall-through projection from call and create a new one
2539 // hook up users of fall-through projection to region
2540 Node *slow_ctrl = _fallthroughproj->clone();
2541 transform_later(slow_ctrl);
2542 _igvn.hash_delete(_fallthroughproj);
2543 _fallthroughproj->disconnect_inputs(NULL, C);
2544 region->init_req(1, slow_ctrl);
2545 // region inputs are now complete
2546 transform_later(region);
2547 _igvn.replace_node(_fallthroughproj, region);
2548
2549 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2550 mem_phi->init_req(1, memproj );
2551 mem_phi->init_req(2, mem);
2552 transform_later(mem_phi);
2553 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2554 }
2555
2556 //---------------------------eliminate_macro_nodes----------------------
2557 // Eliminate scalar replaced allocations and associated locks.
2558 void PhaseMacroExpand::eliminate_macro_nodes() {
2559 if (C->macro_count() == 0)
2560 return;
2561
2562 // First, attempt to eliminate locks
2563 int cnt = C->macro_count();
2564 for (int i=0; i < cnt; i++) {
2565 Node *n = C->macro_node(i);
2566 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2567 // Before elimination mark all associated (same box and obj)
2568 // lock and unlock nodes.
2569 mark_eliminated_locking_nodes(n->as_AbstractLock());
2570 }
2571 }
2572 bool progress = true;
2573 while (progress) {
2574 progress = false;
2575 for (int i = C->macro_count(); i > 0; i--) {
2576 Node * n = C->macro_node(i-1);
2577 bool success = false;
2578 debug_only(int old_macro_count = C->macro_count(););
2579 if (n->is_AbstractLock()) {
2580 success = eliminate_locking_node(n->as_AbstractLock());
2581 }
2582 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2583 progress = progress || success;
2584 }
2585 }
2586 // Next, attempt to eliminate allocations
2587 _has_locks = false;
2588 progress = true;
2589 while (progress) {
2590 progress = false;
2591 for (int i = C->macro_count(); i > 0; i--) {
2592 Node * n = C->macro_node(i-1);
2593 bool success = false;
2594 debug_only(int old_macro_count = C->macro_count(););
2595 switch (n->class_id()) {
2596 case Node::Class_Allocate:
2597 case Node::Class_AllocateArray:
2598 success = eliminate_allocate_node(n->as_Allocate());
2599 break;
2600 case Node::Class_CallStaticJava:
2601 success = eliminate_boxing_node(n->as_CallStaticJava());
2602 break;
2603 case Node::Class_Lock:
2604 case Node::Class_Unlock:
2605 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2606 _has_locks = true;
2607 break;
2608 case Node::Class_ArrayCopy:
2609 break;
2610 default:
2611 assert(n->Opcode() == Op_LoopLimit ||
2612 n->Opcode() == Op_Opaque1 ||
2613 n->Opcode() == Op_Opaque2 ||
2614 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2615 }
2616 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2617 progress = progress || success;
2618 }
2619 }
2620 }
2621
2622 //------------------------------expand_macro_nodes----------------------
2623 // Returns true if a failure occurred.
2624 bool PhaseMacroExpand::expand_macro_nodes() {
2625 // Last attempt to eliminate macro nodes.
2626 eliminate_macro_nodes();
2627
2628 // Make sure expansion will not cause node limit to be exceeded.
2629 // Worst case is a macro node gets expanded into about 50 nodes.
2630 // Allow 50% more for optimization.
2631 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2632 return true;
2633
2634 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2635 bool progress = true;
2636 while (progress) {
2637 progress = false;
2638 for (int i = C->macro_count(); i > 0; i--) {
2639 Node * n = C->macro_node(i-1);
2640 bool success = false;
2641 debug_only(int old_macro_count = C->macro_count(););
2642 if (n->Opcode() == Op_LoopLimit) {
2643 // Remove it from macro list and put on IGVN worklist to optimize.
2644 C->remove_macro_node(n);
2645 _igvn._worklist.push(n);
2646 success = true;
2647 } else if (n->Opcode() == Op_CallStaticJava) {
2648 // Remove it from macro list and put on IGVN worklist to optimize.
2649 C->remove_macro_node(n);
2650 _igvn._worklist.push(n);
2651 success = true;
2652 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2653 _igvn.replace_node(n, n->in(1));
2654 success = true;
2655 #if INCLUDE_RTM_OPT
2656 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2657 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2658 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2659 Node* cmp = n->unique_out();
2660 #ifdef ASSERT
2661 // Validate graph.
2662 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2663 BoolNode* bol = cmp->unique_out()->as_Bool();
2664 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2665 (bol->_test._test == BoolTest::ne), "");
2666 IfNode* ifn = bol->unique_out()->as_If();
2667 assert((ifn->outcnt() == 2) &&
2668 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2669 #endif
2670 Node* repl = n->in(1);
2671 if (!_has_locks) {
2672 // Remove RTM state check if there are no locks in the code.
2673 // Replace input to compare the same value.
2674 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2675 }
2676 _igvn.replace_node(n, repl);
2677 success = true;
2678 #endif
2679 }
2680 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2681 progress = progress || success;
2682 }
2683 }
2684
2685 // expand arraycopy "macro" nodes first
2686 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2687 // before the allocate nodes are expanded.
2688 int macro_idx = C->macro_count() - 1;
2689 while (macro_idx >= 0) {
2690 Node * n = C->macro_node(macro_idx);
2691 assert(n->is_macro(), "only macro nodes expected here");
2692 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2693 // node is unreachable, so don't try to expand it
2694 C->remove_macro_node(n);
2695 } else if (n->is_ArrayCopy()){
2696 int macro_count = C->macro_count();
2697 expand_arraycopy_node(n->as_ArrayCopy());
2698 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2699 }
2700 if (C->failing()) return true;
2701 macro_idx --;
2702 }
2703
2704 // expand "macro" nodes
2705 // nodes are removed from the macro list as they are processed
2706 while (C->macro_count() > 0) {
2707 int macro_count = C->macro_count();
2708 Node * n = C->macro_node(macro_count-1);
2709 assert(n->is_macro(), "only macro nodes expected here");
2710 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2711 // node is unreachable, so don't try to expand it
2712 C->remove_macro_node(n);
2713 continue;
2714 }
2715 switch (n->class_id()) {
2716 case Node::Class_Allocate:
2717 expand_allocate(n->as_Allocate());
2718 break;
2719 case Node::Class_AllocateArray:
2720 expand_allocate_array(n->as_AllocateArray());
2721 break;
2722 case Node::Class_Lock:
2723 expand_lock_node(n->as_Lock());
2724 break;
2725 case Node::Class_Unlock:
2726 expand_unlock_node(n->as_Unlock());
2727 break;
2728 default:
2729 assert(false, "unknown node type in macro list");
2730 }
2731 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2732 if (C->failing()) return true;
2733 }
2734
2735 _igvn.set_delay_transform(false);
2736 _igvn.optimize();
2737 if (C->failing()) return true;
2738 return false;
2739 }