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 "opto/valuetypenode.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50
51 //
52 // Replace any references to "oldref" in inputs to "use" with "newref".
53 // Returns the number of replacements made.
54 //
55 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
56 int nreplacements = 0;
57 uint req = use->req();
58 for (uint j = 0; j < use->len(); j++) {
59 Node *uin = use->in(j);
60 if (uin == oldref) {
61 if (j < req)
62 use->set_req(j, newref);
63 else
64 use->set_prec(j, newref);
65 nreplacements++;
66 } else if (j >= req && uin == NULL) {
67 break;
68 }
69 }
70 return nreplacements;
71 }
72
73 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
74 // Copy debug information and adjust JVMState information
75 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
76 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
77 int jvms_adj = new_dbg_start - old_dbg_start;
78 assert (new_dbg_start == newcall->req(), "argument count mismatch");
79
80 // SafePointScalarObject node could be referenced several times in debug info.
81 // Use Dict to record cloned nodes.
82 Dict* sosn_map = new Dict(cmpkey,hashkey);
83 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
84 Node* old_in = oldcall->in(i);
85 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
86 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
87 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
88 uint old_unique = C->unique();
89 Node* new_in = old_sosn->clone(sosn_map);
90 if (old_unique != C->unique()) { // New node?
91 new_in->set_req(0, C->root()); // reset control edge
92 new_in = transform_later(new_in); // Register new node.
93 }
94 old_in = new_in;
95 }
96 newcall->add_req(old_in);
97 }
98
99 // JVMS may be shared so clone it before we modify it
100 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
101 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
102 jvms->set_map(newcall);
103 jvms->set_locoff(jvms->locoff()+jvms_adj);
104 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
105 jvms->set_monoff(jvms->monoff()+jvms_adj);
106 jvms->set_scloff(jvms->scloff()+jvms_adj);
107 jvms->set_endoff(jvms->endoff()+jvms_adj);
108 }
109 }
110
111 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
112 Node* cmp;
113 if (mask != 0) {
114 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
115 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
116 } else {
117 cmp = word;
118 }
119 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
120 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
121 transform_later(iff);
122
123 // Fast path taken.
124 Node *fast_taken = transform_later(new IfFalseNode(iff));
125
126 // Fast path not-taken, i.e. slow path
127 Node *slow_taken = transform_later(new IfTrueNode(iff));
128
129 if (return_fast_path) {
130 region->init_req(edge, slow_taken); // Capture slow-control
131 return fast_taken;
132 } else {
133 region->init_req(edge, fast_taken); // Capture fast-control
134 return slow_taken;
135 }
136 }
137
138 //--------------------copy_predefined_input_for_runtime_call--------------------
139 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
140 // Set fixed predefined input arguments
141 call->init_req( TypeFunc::Control, ctrl );
142 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
143 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
144 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
145 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
146 }
147
148 //------------------------------make_slow_call---------------------------------
149 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
150 address slow_call, const char* leaf_name, Node* slow_path,
151 Node* parm0, Node* parm1, Node* parm2) {
152
153 // Slow-path call
154 CallNode *call = leaf_name
155 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
156 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
157
158 // Slow path call has no side-effects, uses few values
159 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
160 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
161 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
162 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
163 copy_call_debug_info(oldcall, call);
164 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
165 _igvn.replace_node(oldcall, call);
166 transform_later(call);
167
168 return call;
169 }
170
171 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
172 _fallthroughproj = NULL;
173 _fallthroughcatchproj = NULL;
174 _ioproj_fallthrough = NULL;
175 _ioproj_catchall = NULL;
176 _catchallcatchproj = NULL;
177 _memproj_fallthrough = NULL;
178 _memproj_catchall = NULL;
179 _resproj = NULL;
180 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
181 ProjNode *pn = call->fast_out(i)->as_Proj();
182 switch (pn->_con) {
183 case TypeFunc::Control:
184 {
185 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
186 _fallthroughproj = pn;
187 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
188 const Node *cn = pn->fast_out(j);
189 if (cn->is_Catch()) {
190 ProjNode *cpn = NULL;
191 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
192 cpn = cn->fast_out(k)->as_Proj();
193 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
194 if (cpn->_con == CatchProjNode::fall_through_index)
195 _fallthroughcatchproj = cpn;
196 else {
197 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
198 _catchallcatchproj = cpn;
199 }
200 }
201 }
202 break;
203 }
204 case TypeFunc::I_O:
205 if (pn->_is_io_use)
206 _ioproj_catchall = pn;
207 else
208 _ioproj_fallthrough = pn;
209 break;
210 case TypeFunc::Memory:
211 if (pn->_is_io_use)
212 _memproj_catchall = pn;
213 else
214 _memproj_fallthrough = pn;
215 break;
216 case TypeFunc::Parms:
217 _resproj = pn;
218 break;
219 default:
220 assert(false, "unexpected projection from allocation node.");
221 }
222 }
223
224 }
225
226 // Eliminate a card mark sequence. p2x is a ConvP2XNode
227 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
228 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
229 if (!UseG1GC) {
230 // vanilla/CMS post barrier
231 Node *shift = p2x->unique_out();
232 Node *addp = shift->unique_out();
233 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
234 Node *mem = addp->last_out(j);
235 if (UseCondCardMark && mem->is_Load()) {
236 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
237 // The load is checking if the card has been written so
238 // replace it with zero to fold the test.
239 _igvn.replace_node(mem, intcon(0));
240 continue;
241 }
242 assert(mem->is_Store(), "store required");
243 _igvn.replace_node(mem, mem->in(MemNode::Memory));
244 }
245 } else {
246 // G1 pre/post barriers
247 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
248 // It could be only one user, URShift node, in Object.clone() intrinsic
249 // but the new allocation is passed to arraycopy stub and it could not
250 // be scalar replaced. So we don't check the case.
251
252 // An other case of only one user (Xor) is when the value check for NULL
253 // in G1 post barrier is folded after CCP so the code which used URShift
254 // is removed.
255
256 // Take Region node before eliminating post barrier since it also
257 // eliminates CastP2X node when it has only one user.
258 Node* this_region = p2x->in(0);
259 assert(this_region != NULL, "");
260
261 // Remove G1 post barrier.
262
263 // Search for CastP2X->Xor->URShift->Cmp path which
264 // checks if the store done to a different from the value's region.
265 // And replace Cmp with #0 (false) to collapse G1 post barrier.
266 Node* xorx = p2x->find_out_with(Op_XorX);
267 assert(xorx != NULL, "missing G1 post barrier");
268 Node* shift = xorx->unique_out();
269 Node* cmpx = shift->unique_out();
270 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
271 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
272 "missing region check in G1 post barrier");
273 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
274
275 // Remove G1 pre barrier.
276
277 // Search "if (marking != 0)" check and set it to "false".
278 // There is no G1 pre barrier if previous stored value is NULL
279 // (for example, after initialization).
280 if (this_region->is_Region() && this_region->req() == 3) {
281 int ind = 1;
282 if (!this_region->in(ind)->is_IfFalse()) {
283 ind = 2;
284 }
285 if (this_region->in(ind)->is_IfFalse()) {
286 Node* bol = this_region->in(ind)->in(0)->in(1);
287 assert(bol->is_Bool(), "");
288 cmpx = bol->in(1);
289 if (bol->as_Bool()->_test._test == BoolTest::ne &&
290 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
291 cmpx->in(1)->is_Load()) {
292 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
293 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
294 SATBMarkQueue::byte_offset_of_active());
295 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
296 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
297 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
298 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
299 }
300 }
301 }
302 }
303 // Now CastP2X can be removed since it is used only on dead path
304 // which currently still alive until igvn optimize it.
305 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
306 _igvn.replace_node(p2x, top());
307 }
308 }
309
310 // Search for a memory operation for the specified memory slice.
311 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
312 Node *orig_mem = mem;
313 Node *alloc_mem = alloc->in(TypeFunc::Memory);
314 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
315 while (true) {
316 if (mem == alloc_mem || mem == start_mem ) {
317 return mem; // hit one of our sentinels
318 } else if (mem->is_MergeMem()) {
319 mem = mem->as_MergeMem()->memory_at(alias_idx);
320 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
321 Node *in = mem->in(0);
322 // we can safely skip over safepoints, calls, locks and membars because we
323 // already know that the object is safe to eliminate.
324 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
325 return in;
326 } else if (in->is_Call()) {
327 CallNode *call = in->as_Call();
328 if (call->may_modify(tinst, phase)) {
329 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
330
331 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
332 return in;
333 }
334 }
335 mem = in->in(TypeFunc::Memory);
336 } else if (in->is_MemBar()) {
337 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase)) {
338 assert(in->in(0)->is_Proj() && in->in(0)->in(0)->is_ArrayCopy(), "should be arraycopy");
339 ArrayCopyNode* ac = in->in(0)->in(0)->as_ArrayCopy();
340 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
341 return ac;
342 }
343 mem = in->in(TypeFunc::Memory);
344 } else {
345 assert(false, "unexpected projection");
346 }
347 } else if (mem->is_Store()) {
348 const TypePtr* atype = mem->as_Store()->adr_type();
349 int adr_idx = phase->C->get_alias_index(atype);
350 if (adr_idx == alias_idx) {
351 assert(atype->isa_oopptr(), "address type must be oopptr");
352 int adr_offset = atype->offset();
353 uint adr_iid = atype->is_oopptr()->instance_id();
354 // Array elements references have the same alias_idx
355 // but different offset and different instance_id.
356 if (adr_offset == offset && adr_iid == alloc->_idx)
357 return mem;
358 } else {
359 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
360 }
361 mem = mem->in(MemNode::Memory);
362 } else if (mem->is_ClearArray()) {
363 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
364 // Can not bypass initialization of the instance
365 // we are looking.
366 debug_only(intptr_t offset;)
367 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
368 InitializeNode* init = alloc->as_Allocate()->initialization();
369 // We are looking for stored value, return Initialize node
370 // or memory edge from Allocate node.
371 if (init != NULL)
372 return init;
373 else
374 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
375 }
376 // Otherwise skip it (the call updated 'mem' value).
377 } else if (mem->Opcode() == Op_SCMemProj) {
378 mem = mem->in(0);
379 Node* adr = NULL;
380 if (mem->is_LoadStore()) {
381 adr = mem->in(MemNode::Address);
382 } else {
383 assert(mem->Opcode() == Op_EncodeISOArray ||
384 mem->Opcode() == Op_StrCompressedCopy, "sanity");
385 adr = mem->in(3); // Destination array
386 }
387 const TypePtr* atype = adr->bottom_type()->is_ptr();
388 int adr_idx = phase->C->get_alias_index(atype);
389 if (adr_idx == alias_idx) {
390 DEBUG_ONLY(mem->dump();)
391 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
392 return NULL;
393 }
394 mem = mem->in(MemNode::Memory);
395 } else if (mem->Opcode() == Op_StrInflatedCopy) {
396 Node* adr = mem->in(3); // Destination array
397 const TypePtr* atype = adr->bottom_type()->is_ptr();
398 int adr_idx = phase->C->get_alias_index(atype);
399 if (adr_idx == alias_idx) {
400 DEBUG_ONLY(mem->dump();)
401 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
402 return NULL;
403 }
404 mem = mem->in(MemNode::Memory);
405 } else {
406 return mem;
407 }
408 assert(mem != orig_mem, "dead memory loop");
409 }
410 }
411
412 // Generate loads from source of the arraycopy for fields of
413 // destination needed at a deoptimization point
414 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, BasicType ft, const Type *ftype, AllocateNode *alloc) {
415 BasicType bt = ft;
416 const Type *type = ftype;
417 if (ft == T_NARROWOOP) {
418 bt = T_OBJECT;
419 type = ftype->make_oopptr();
420 }
421 Node* res = NULL;
422 if (ac->is_clonebasic()) {
423 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
424 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
425 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
426 Node* m = ac->in(TypeFunc::Memory);
427 while (m->is_MergeMem()) {
428 m = m->as_MergeMem()->memory_at(C->get_alias_index(adr_type));
429 if (m->is_Proj() && m->in(0)->is_MemBar()) {
430 m = m->in(0)->in(TypeFunc::Memory);
431 }
432 }
433 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
434 } else {
435 if (ac->modifies(offset, offset, &_igvn, true)) {
436 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
437 uint shift = exact_log2(type2aelembytes(bt));
438 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
439 #ifdef _LP64
440 diff = _igvn.transform(new ConvI2LNode(diff));
441 #endif
442 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
443
444 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
445 Node* base = ac->in(ArrayCopyNode::Src);
446 Node* adr = _igvn.transform(new AddPNode(base, base, off));
447 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
448 Node* m = ac->in(TypeFunc::Memory);
449 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
450 }
451 }
452 if (res != NULL) {
453 res = _igvn.transform(res);
454 if (ftype->isa_narrowoop()) {
455 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
456 res = _igvn.transform(new EncodePNode(res, ftype));
457 }
458 return res;
459 }
460 return NULL;
461 }
462
463 //
464 // Given a Memory Phi, compute a value Phi containing the values from stores
465 // on the input paths.
466 // Note: this function is recursive, its depth is limited by the "level" argument
467 // Returns the computed Phi, or NULL if it cannot compute it.
468 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) {
469 assert(mem->is_Phi(), "sanity");
470 int alias_idx = C->get_alias_index(adr_t);
471 int offset = adr_t->offset();
472 int instance_id = adr_t->instance_id();
473
474 // Check if an appropriate value phi already exists.
475 Node* region = mem->in(0);
476 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
477 Node* phi = region->fast_out(k);
478 if (phi->is_Phi() && phi != mem &&
479 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
480 return phi;
481 }
482 }
483 // Check if an appropriate new value phi already exists.
484 Node* new_phi = value_phis->find(mem->_idx);
485 if (new_phi != NULL)
486 return new_phi;
487
488 if (level <= 0) {
489 return NULL; // Give up: phi tree too deep
490 }
491 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
492 Node *alloc_mem = alloc->in(TypeFunc::Memory);
493
494 uint length = mem->req();
495 GrowableArray <Node *> values(length, length, NULL, false);
496
497 // create a new Phi for the value
498 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
499 transform_later(phi);
500 value_phis->push(phi, mem->_idx);
501
502 for (uint j = 1; j < length; j++) {
503 Node *in = mem->in(j);
504 if (in == NULL || in->is_top()) {
505 values.at_put(j, in);
506 } else {
507 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
508 if (val == start_mem || val == alloc_mem) {
509 // hit a sentinel, return appropriate 0 value
510 values.at_put(j, _igvn.zerocon(ft));
511 continue;
512 }
513 if (val->is_Initialize()) {
514 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
515 }
516 if (val == NULL) {
517 return NULL; // can't find a value on this path
518 }
519 if (val == mem) {
520 values.at_put(j, mem);
521 } else if (val->is_Store()) {
522 values.at_put(j, val->in(MemNode::ValueIn));
523 } else if(val->is_Proj() && val->in(0) == alloc) {
524 values.at_put(j, _igvn.zerocon(ft));
525 } else if (val->is_Phi()) {
526 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
527 if (val == NULL) {
528 return NULL;
529 }
530 values.at_put(j, val);
531 } else if (val->Opcode() == Op_SCMemProj) {
532 assert(val->in(0)->is_LoadStore() ||
533 val->in(0)->Opcode() == Op_EncodeISOArray ||
534 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
535 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
536 return NULL;
537 } else if (val->is_ArrayCopy()) {
538 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), ft, phi_type, alloc);
539 if (res == NULL) {
540 return NULL;
541 }
542 values.at_put(j, res);
543 } else {
544 #ifdef ASSERT
545 val->dump();
546 assert(false, "unknown node on this path");
547 #endif
548 return NULL; // unknown node on this path
549 }
550 }
551 }
552 // Set Phi's inputs
553 for (uint j = 1; j < length; j++) {
554 if (values.at(j) == mem) {
555 phi->init_req(j, phi);
556 } else {
557 phi->init_req(j, values.at(j));
558 }
559 }
560 return phi;
561 }
562
563 // Search the last value stored into the object's field.
564 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
565 assert(adr_t->is_known_instance_field(), "instance required");
566 assert(ft != T_VALUETYPE, "should not be used for value type fields");
567 int instance_id = adr_t->instance_id();
568 assert((uint)instance_id == alloc->_idx, "wrong allocation");
569
570 int alias_idx = C->get_alias_index(adr_t);
571 int offset = adr_t->offset();
572 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
573 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
574 Node *alloc_mem = alloc->in(TypeFunc::Memory);
575 Arena *a = Thread::current()->resource_area();
576 VectorSet visited(a);
577
578 bool done = sfpt_mem == alloc_mem;
579 Node *mem = sfpt_mem;
580 while (!done) {
581 if (visited.test_set(mem->_idx)) {
582 return NULL; // found a loop, give up
583 }
584 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
585 if (mem == start_mem || mem == alloc_mem) {
586 done = true; // hit a sentinel, return appropriate 0 value
587 } else if (mem->is_Initialize()) {
588 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
589 if (mem == NULL) {
590 done = true; // Something went wrong.
591 } else if (mem->is_Store()) {
592 const TypePtr* atype = mem->as_Store()->adr_type();
593 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
594 done = true;
595 }
596 } else if (mem->is_Store()) {
597 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
598 assert(atype != NULL, "address type must be oopptr");
599 assert(C->get_alias_index(atype) == alias_idx &&
600 atype->is_known_instance_field() && atype->offset() == offset &&
601 atype->instance_id() == instance_id, "store is correct memory slice");
602 done = true;
603 } else if (mem->is_Phi()) {
604 // try to find a phi's unique input
605 Node *unique_input = NULL;
606 Node *top = C->top();
607 for (uint i = 1; i < mem->req(); i++) {
608 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
609 if (n == NULL || n == top || n == mem) {
610 continue;
611 } else if (unique_input == NULL) {
612 unique_input = n;
613 } else if (unique_input != n) {
614 unique_input = top;
615 break;
616 }
617 }
618 if (unique_input != NULL && unique_input != top) {
619 mem = unique_input;
620 } else {
621 done = true;
622 }
623 } else if (mem->is_ArrayCopy()) {
624 done = true;
625 } else {
626 assert(false, "unexpected node");
627 }
628 }
629 if (mem != NULL) {
630 if (mem == start_mem || mem == alloc_mem) {
631 // hit a sentinel, return appropriate 0 value
632 return _igvn.zerocon(ft);
633 } else if (mem->is_Store()) {
634 return mem->in(MemNode::ValueIn);
635 } else if (mem->is_Phi()) {
636 // attempt to produce a Phi reflecting the values on the input paths of the Phi
637 Node_Stack value_phis(a, 8);
638 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
639 if (phi != NULL) {
640 return phi;
641 } else {
642 // Kill all new Phis
643 while(value_phis.is_nonempty()) {
644 Node* n = value_phis.node();
645 _igvn.replace_node(n, C->top());
646 value_phis.pop();
647 }
648 }
649 } else if (mem->is_ArrayCopy()) {
650 Node* ctl = mem->in(0);
651 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
652 // pin the loads in the uncommon trap path
653 ctl = sfpt_ctl;
654 }
655 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, ft, ftype, alloc);
656 }
657 }
658 // Something went wrong.
659 return NULL;
660 }
661
662 // Search the last value stored into the value type's fields.
663 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
664 // Subtract the offset of the first field to account for the missing oop header
665 offset -= vk->first_field_offset();
666 // Create a new ValueTypeNode and retrieve the field values from memory
667 ValueTypeNode* vt = ValueTypeNode::make(_igvn, vk)->as_ValueType();
668 for (int i = 0; i < vk->field_count(); ++i) {
669 ciType* field_type = vt->field_type(i);
670 int field_offset = offset + vt->field_offset(i);
671 // Each value type field has its own memory slice
672 adr_type = adr_type->with_field_offset(field_offset);
673 Node* value = NULL;
674 if (field_type->basic_type() == T_VALUETYPE) {
675 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
676 } else {
677 value = value_from_mem(mem, ctl, field_type->basic_type(), Type::get_const_type(field_type), adr_type, alloc);
678 }
679 vt->set_field_value(i, value);
680 }
681 return vt;
682 }
683
684 // Check the possibility of scalar replacement.
685 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
686 // Scan the uses of the allocation to check for anything that would
687 // prevent us from eliminating it.
688 NOT_PRODUCT( const char* fail_eliminate = NULL; )
689 DEBUG_ONLY( Node* disq_node = NULL; )
690 bool can_eliminate = true;
691
692 Node* res = alloc->result_cast();
693 const TypeOopPtr* res_type = NULL;
694 if (res == NULL) {
695 // All users were eliminated.
696 } else if (!res->is_CheckCastPP()) {
697 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
698 can_eliminate = false;
699 } else {
700 res_type = _igvn.type(res)->isa_oopptr();
701 if (res_type == NULL) {
702 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
703 can_eliminate = false;
704 } else if (res_type->isa_aryptr()) {
705 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
706 if (length < 0) {
707 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
708 can_eliminate = false;
709 }
710 }
711 }
712
713 if (can_eliminate && res != NULL) {
714 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
715 j < jmax && can_eliminate; j++) {
716 Node* use = res->fast_out(j);
717
718 if (use->is_AddP()) {
719 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
720 int offset = addp_type->offset();
721
722 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
723 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
724 can_eliminate = false;
725 break;
726 }
727 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
728 k < kmax && can_eliminate; k++) {
729 Node* n = use->fast_out(k);
730 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
731 !(n->is_ArrayCopy() &&
732 n->as_ArrayCopy()->is_clonebasic() &&
733 n->in(ArrayCopyNode::Dest) == use)) {
734 DEBUG_ONLY(disq_node = n;)
735 if (n->is_Load() || n->is_LoadStore()) {
736 NOT_PRODUCT(fail_eliminate = "Field load";)
737 } else {
738 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
739 }
740 can_eliminate = false;
741 }
742 }
743 } else if (use->is_ArrayCopy() &&
744 (use->as_ArrayCopy()->is_arraycopy_validated() ||
745 use->as_ArrayCopy()->is_copyof_validated() ||
746 use->as_ArrayCopy()->is_copyofrange_validated()) &&
747 use->in(ArrayCopyNode::Dest) == res) {
748 // ok to eliminate
749 } else if (use->is_SafePoint()) {
750 SafePointNode* sfpt = use->as_SafePoint();
751 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
752 // Object is passed as argument.
753 DEBUG_ONLY(disq_node = use;)
754 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
755 can_eliminate = false;
756 }
757 Node* sfptMem = sfpt->memory();
758 if (sfptMem == NULL || sfptMem->is_top()) {
759 DEBUG_ONLY(disq_node = use;)
760 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
761 can_eliminate = false;
762 } else {
763 safepoints.append_if_missing(sfpt);
764 }
765 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
766 if (use->is_Phi()) {
767 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
768 NOT_PRODUCT(fail_eliminate = "Object is return value";)
769 } else {
770 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
771 }
772 DEBUG_ONLY(disq_node = use;)
773 } else {
774 if (use->Opcode() == Op_Return) {
775 NOT_PRODUCT(fail_eliminate = "Object is return value";)
776 } else {
777 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
778 }
779 DEBUG_ONLY(disq_node = use;)
780 }
781 can_eliminate = false;
782 }
783 }
784 }
785
786 #ifndef PRODUCT
787 if (PrintEliminateAllocations) {
788 if (can_eliminate) {
789 tty->print("Scalar ");
790 if (res == NULL)
791 alloc->dump();
792 else
793 res->dump();
794 } else if (alloc->_is_scalar_replaceable) {
795 tty->print("NotScalar (%s)", fail_eliminate);
796 if (res == NULL)
797 alloc->dump();
798 else
799 res->dump();
800 #ifdef ASSERT
801 if (disq_node != NULL) {
802 tty->print(" >>>> ");
803 disq_node->dump();
804 }
805 #endif /*ASSERT*/
806 }
807 }
808 #endif
809 return can_eliminate;
810 }
811
812 // Do scalar replacement.
813 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
814 GrowableArray <SafePointNode *> safepoints_done;
815
816 ciKlass* klass = NULL;
817 ciInstanceKlass* iklass = NULL;
818 int nfields = 0;
819 int array_base = 0;
820 int element_size = 0;
821 BasicType basic_elem_type = T_ILLEGAL;
822 ciType* elem_type = NULL;
823
824 Node* res = alloc->result_cast();
825 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
826 const TypeOopPtr* res_type = NULL;
827 if (res != NULL) { // Could be NULL when there are no users
828 res_type = _igvn.type(res)->isa_oopptr();
829 }
830
831 if (res != NULL) {
832 klass = res_type->klass();
833 // Value types are only allocated on demand
834 if (res_type->isa_instptr() || res_type->isa_valuetypeptr()) {
835 // find the fields of the class which will be needed for safepoint debug information
836 assert(klass->is_instance_klass(), "must be an instance klass.");
837 iklass = klass->as_instance_klass();
838 nfields = iklass->nof_nonstatic_fields();
839 } else {
840 // find the array's elements which will be needed for safepoint debug information
841 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
842 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
843 elem_type = klass->as_array_klass()->element_type();
844 basic_elem_type = elem_type->basic_type();
845 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
846 element_size = type2aelembytes(basic_elem_type);
847 }
848 }
849 //
850 // Process the safepoint uses
851 //
852 while (safepoints.length() > 0) {
853 SafePointNode* sfpt = safepoints.pop();
854 Node* mem = sfpt->memory();
855 Node* ctl = sfpt->control();
856 assert(sfpt->jvms() != NULL, "missed JVMS");
857 // Fields of scalar objs are referenced only at the end
858 // of regular debuginfo at the last (youngest) JVMS.
859 // Record relative start index.
860 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
861 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
862 #ifdef ASSERT
863 alloc,
864 #endif
865 first_ind, nfields);
866 sobj->init_req(0, C->root());
867 transform_later(sobj);
868
869 // Scan object's fields adding an input to the safepoint for each field.
870 for (int j = 0; j < nfields; j++) {
871 intptr_t offset;
872 ciField* field = NULL;
873 if (iklass != NULL) {
874 field = iklass->nonstatic_field_at(j);
875 offset = field->offset();
876 elem_type = field->type();
877 basic_elem_type = field->layout_type();
878 // Value type fields should not have safepoint uses
879 assert(basic_elem_type != T_VALUETYPE, "value type fields are flattened");
880 } else {
881 offset = array_base + j * (intptr_t)element_size;
882 }
883
884 const Type *field_type;
885 // The next code is taken from Parse::do_get_xxx().
886 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
887 if (!elem_type->is_loaded()) {
888 field_type = TypeInstPtr::BOTTOM;
889 } else if (field != NULL && field->is_constant() && field->is_static()) {
890 // This can happen if the constant oop is non-perm.
891 ciObject* con = field->constant_value().as_object();
892 // Do not "join" in the previous type; it doesn't add value,
893 // and may yield a vacuous result if the field is of interface type.
894 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
895 assert(field_type != NULL, "field singleton type must be consistent");
896 } else {
897 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
898 }
899 if (UseCompressedOops) {
900 field_type = field_type->make_narrowoop();
901 basic_elem_type = T_NARROWOOP;
902 }
903 } else {
904 field_type = Type::get_const_basic_type(basic_elem_type);
905 }
906
907 Node* field_val = NULL;
908 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
909 if (klass->is_value_array_klass()) {
910 ciValueKlass* vk = elem_type->as_value_klass();
911 assert(vk->flatten_array(), "must be flattened");
912 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
913 } else {
914 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
915 }
916 if (field_val == NULL) {
917 // We weren't able to find a value for this field,
918 // give up on eliminating this allocation.
919
920 // Remove any extra entries we added to the safepoint.
921 uint last = sfpt->req() - 1;
922 for (int k = 0; k < j; k++) {
923 sfpt->del_req(last--);
924 }
925 _igvn._worklist.push(sfpt);
926 // rollback processed safepoints
927 while (safepoints_done.length() > 0) {
928 SafePointNode* sfpt_done = safepoints_done.pop();
929 // remove any extra entries we added to the safepoint
930 last = sfpt_done->req() - 1;
931 for (int k = 0; k < nfields; k++) {
932 sfpt_done->del_req(last--);
933 }
934 JVMState *jvms = sfpt_done->jvms();
935 jvms->set_endoff(sfpt_done->req());
936 // Now make a pass over the debug information replacing any references
937 // to SafePointScalarObjectNode with the allocated object.
938 int start = jvms->debug_start();
939 int end = jvms->debug_end();
940 for (int i = start; i < end; i++) {
941 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
942 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
943 if (scobj->first_index(jvms) == sfpt_done->req() &&
944 scobj->n_fields() == (uint)nfields) {
945 assert(scobj->alloc() == alloc, "sanity");
946 sfpt_done->set_req(i, res);
947 }
948 }
949 }
950 _igvn._worklist.push(sfpt_done);
951 }
952 #ifndef PRODUCT
953 if (PrintEliminateAllocations) {
954 if (field != NULL) {
955 tty->print("=== At SafePoint node %d can't find value of Field: ",
956 sfpt->_idx);
957 field->print();
958 int field_idx = C->get_alias_index(field_addr_type);
959 tty->print(" (alias_idx=%d)", field_idx);
960 } else { // Array's element
961 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
962 sfpt->_idx, j);
963 }
964 tty->print(", which prevents elimination of: ");
965 if (res == NULL)
966 alloc->dump();
967 else
968 res->dump();
969 }
970 #endif
971 return false;
972 }
973 if (UseCompressedOops && field_type->isa_narrowoop()) {
974 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
975 // to be able scalar replace the allocation.
976 if (field_val->is_EncodeP()) {
977 field_val = field_val->in(1);
978 } else {
979 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
980 }
981 }
982 sfpt->add_req(field_val);
983 }
984 JVMState *jvms = sfpt->jvms();
985 jvms->set_endoff(sfpt->req());
986 // Now make a pass over the debug information replacing any references
987 // to the allocated object with "sobj"
988 int start = jvms->debug_start();
989 int end = jvms->debug_end();
990 sfpt->replace_edges_in_range(res, sobj, start, end);
991 _igvn._worklist.push(sfpt);
992 safepoints_done.append_if_missing(sfpt); // keep it for rollback
993 }
994 return true;
995 }
996
997 // Process users of eliminated allocation.
998 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
999 Node* res = alloc->result_cast();
1000 if (res != NULL) {
1001 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1002 Node *use = res->last_out(j);
1003 uint oc1 = res->outcnt();
1004
1005 if (use->is_AddP()) {
1006 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1007 Node *n = use->last_out(k);
1008 uint oc2 = use->outcnt();
1009 if (n->is_Store()) {
1010 #ifdef ASSERT
1011 // Verify that there is no dependent MemBarVolatile nodes,
1012 // they should be removed during IGVN, see MemBarNode::Ideal().
1013 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1014 p < pmax; p++) {
1015 Node* mb = n->fast_out(p);
1016 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1017 mb->req() <= MemBarNode::Precedent ||
1018 mb->in(MemBarNode::Precedent) != n,
1019 "MemBarVolatile should be eliminated for non-escaping object");
1020 }
1021 #endif
1022 _igvn.replace_node(n, n->in(MemNode::Memory));
1023 } else if (n->is_ArrayCopy()) {
1024 // Disconnect ArrayCopy node
1025 ArrayCopyNode* ac = n->as_ArrayCopy();
1026 assert(ac->is_clonebasic(), "unexpected array copy kind");
1027 Node* ctl_proj = ac->proj_out(TypeFunc::Control);
1028 Node* mem_proj = ac->proj_out(TypeFunc::Memory);
1029 if (ctl_proj != NULL) {
1030 _igvn.replace_node(ctl_proj, n->in(0));
1031 }
1032 if (mem_proj != NULL) {
1033 _igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1034 }
1035 } else {
1036 eliminate_card_mark(n);
1037 }
1038 k -= (oc2 - use->outcnt());
1039 }
1040 } else if (use->is_ArrayCopy()) {
1041 // Disconnect ArrayCopy node
1042 ArrayCopyNode* ac = use->as_ArrayCopy();
1043 assert(ac->is_arraycopy_validated() ||
1044 ac->is_copyof_validated() ||
1045 ac->is_copyofrange_validated(), "unsupported");
1046 CallProjections callprojs;
1047 ac->extract_projections(&callprojs, true);
1048
1049 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1050 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1051 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1052
1053 // Set control to top. IGVN will remove the remaining projections
1054 ac->set_req(0, top());
1055 ac->replace_edge(res, top());
1056
1057 // Disconnect src right away: it can help find new
1058 // opportunities for allocation elimination
1059 Node* src = ac->in(ArrayCopyNode::Src);
1060 ac->replace_edge(src, top());
1061 if (src->outcnt() == 0) {
1062 _igvn.remove_dead_node(src);
1063 }
1064
1065 _igvn._worklist.push(ac);
1066 } else {
1067 eliminate_card_mark(use);
1068 }
1069 j -= (oc1 - res->outcnt());
1070 }
1071 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1072 _igvn.remove_dead_node(res);
1073 }
1074
1075 //
1076 // Process other users of allocation's projections
1077 //
1078 if (_resproj != NULL && _resproj->outcnt() != 0) {
1079 // First disconnect stores captured by Initialize node.
1080 // If Initialize node is eliminated first in the following code,
1081 // it will kill such stores and DUIterator_Last will assert.
1082 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1083 Node *use = _resproj->fast_out(j);
1084 if (use->is_AddP()) {
1085 // raw memory addresses used only by the initialization
1086 _igvn.replace_node(use, C->top());
1087 --j; --jmax;
1088 }
1089 }
1090 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1091 Node *use = _resproj->last_out(j);
1092 uint oc1 = _resproj->outcnt();
1093 if (use->is_Initialize()) {
1094 // Eliminate Initialize node.
1095 InitializeNode *init = use->as_Initialize();
1096 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1097 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
1098 if (ctrl_proj != NULL) {
1099 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1100 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1101 }
1102 Node *mem_proj = init->proj_out(TypeFunc::Memory);
1103 if (mem_proj != NULL) {
1104 Node *mem = init->in(TypeFunc::Memory);
1105 #ifdef ASSERT
1106 if (mem->is_MergeMem()) {
1107 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1108 } else {
1109 assert(mem == _memproj_fallthrough, "allocation memory projection");
1110 }
1111 #endif
1112 _igvn.replace_node(mem_proj, mem);
1113 }
1114 } else {
1115 assert(false, "only Initialize or AddP expected");
1116 }
1117 j -= (oc1 - _resproj->outcnt());
1118 }
1119 }
1120 if (_fallthroughcatchproj != NULL) {
1121 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1122 }
1123 if (_memproj_fallthrough != NULL) {
1124 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1125 }
1126 if (_memproj_catchall != NULL) {
1127 _igvn.replace_node(_memproj_catchall, C->top());
1128 }
1129 if (_ioproj_fallthrough != NULL) {
1130 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1131 }
1132 if (_ioproj_catchall != NULL) {
1133 _igvn.replace_node(_ioproj_catchall, C->top());
1134 }
1135 if (_catchallcatchproj != NULL) {
1136 _igvn.replace_node(_catchallcatchproj, C->top());
1137 }
1138 }
1139
1140 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1141 // Don't do scalar replacement if the frame can be popped by JVMTI:
1142 // if reallocation fails during deoptimization we'll pop all
1143 // interpreter frames for this compiled frame and that won't play
1144 // nice with JVMTI popframe.
1145 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1146 return false;
1147 }
1148 Node* klass = alloc->in(AllocateNode::KlassNode);
1149 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1150 Node* res = alloc->result_cast();
1151 // Eliminate boxing allocations which are not used
1152 // regardless scalar replacable status.
1153 bool boxing_alloc = C->eliminate_boxing() &&
1154 tklass->klass()->is_instance_klass() &&
1155 tklass->klass()->as_instance_klass()->is_box_klass();
1156 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1157 return false;
1158 }
1159
1160 extract_call_projections(alloc);
1161
1162 GrowableArray <SafePointNode *> safepoints;
1163 if (!can_eliminate_allocation(alloc, safepoints)) {
1164 return false;
1165 }
1166
1167 if (!alloc->_is_scalar_replaceable) {
1168 assert(res == NULL, "sanity");
1169 // We can only eliminate allocation if all debug info references
1170 // are already replaced with SafePointScalarObject because
1171 // we can't search for a fields value without instance_id.
1172 if (safepoints.length() > 0) {
1173 return false;
1174 }
1175 }
1176
1177 if (!scalar_replacement(alloc, safepoints)) {
1178 return false;
1179 }
1180
1181 CompileLog* log = C->log();
1182 if (log != NULL) {
1183 log->head("eliminate_allocation type='%d'",
1184 log->identify(tklass->klass()));
1185 JVMState* p = alloc->jvms();
1186 while (p != NULL) {
1187 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1188 p = p->caller();
1189 }
1190 log->tail("eliminate_allocation");
1191 }
1192
1193 process_users_of_allocation(alloc);
1194
1195 #ifndef PRODUCT
1196 if (PrintEliminateAllocations) {
1197 if (alloc->is_AllocateArray())
1198 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1199 else
1200 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1201 }
1202 #endif
1203
1204 return true;
1205 }
1206
1207 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1208 // EA should remove all uses of non-escaping boxing node.
1209 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1210 return false;
1211 }
1212
1213 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1214
1215 extract_call_projections(boxing);
1216
1217 const TypeTuple* r = boxing->tf()->range();
1218 assert(r->cnt() > TypeFunc::Parms, "sanity");
1219 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1220 assert(t != NULL, "sanity");
1221
1222 CompileLog* log = C->log();
1223 if (log != NULL) {
1224 log->head("eliminate_boxing type='%d'",
1225 log->identify(t->klass()));
1226 JVMState* p = boxing->jvms();
1227 while (p != NULL) {
1228 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1229 p = p->caller();
1230 }
1231 log->tail("eliminate_boxing");
1232 }
1233
1234 process_users_of_allocation(boxing);
1235
1236 #ifndef PRODUCT
1237 if (PrintEliminateAllocations) {
1238 tty->print("++++ Eliminated: %d ", boxing->_idx);
1239 boxing->method()->print_short_name(tty);
1240 tty->cr();
1241 }
1242 #endif
1243
1244 return true;
1245 }
1246
1247 //---------------------------set_eden_pointers-------------------------
1248 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1249 if (UseTLAB) { // Private allocation: load from TLS
1250 Node* thread = transform_later(new ThreadLocalNode());
1251 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1252 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1253 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1254 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1255 } else { // Shared allocation: load from globals
1256 CollectedHeap* ch = Universe::heap();
1257 address top_adr = (address)ch->top_addr();
1258 address end_adr = (address)ch->end_addr();
1259 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1260 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1261 }
1262 }
1263
1264
1265 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1266 Node* adr = basic_plus_adr(base, offset);
1267 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1268 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1269 transform_later(value);
1270 return value;
1271 }
1272
1273
1274 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1275 Node* adr = basic_plus_adr(base, offset);
1276 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1277 transform_later(mem);
1278 return mem;
1279 }
1280
1281 //=============================================================================
1282 //
1283 // A L L O C A T I O N
1284 //
1285 // Allocation attempts to be fast in the case of frequent small objects.
1286 // It breaks down like this:
1287 //
1288 // 1) Size in doublewords is computed. This is a constant for objects and
1289 // variable for most arrays. Doubleword units are used to avoid size
1290 // overflow of huge doubleword arrays. We need doublewords in the end for
1291 // rounding.
1292 //
1293 // 2) Size is checked for being 'too large'. Too-large allocations will go
1294 // the slow path into the VM. The slow path can throw any required
1295 // exceptions, and does all the special checks for very large arrays. The
1296 // size test can constant-fold away for objects. For objects with
1297 // finalizers it constant-folds the otherway: you always go slow with
1298 // finalizers.
1299 //
1300 // 3) If NOT using TLABs, this is the contended loop-back point.
1301 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1302 //
1303 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1304 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1305 // "size*8" we always enter the VM, where "largish" is a constant picked small
1306 // enough that there's always space between the eden max and 4Gig (old space is
1307 // there so it's quite large) and large enough that the cost of entering the VM
1308 // is dwarfed by the cost to initialize the space.
1309 //
1310 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1311 // down. If contended, repeat at step 3. If using TLABs normal-store
1312 // adjusted heap top back down; there is no contention.
1313 //
1314 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1315 // fields.
1316 //
1317 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1318 // oop flavor.
1319 //
1320 //=============================================================================
1321 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1322 // Allocations bigger than this always go the slow route.
1323 // This value must be small enough that allocation attempts that need to
1324 // trigger exceptions go the slow route. Also, it must be small enough so
1325 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1326 //=============================================================================j//
1327 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1328 // The allocator will coalesce int->oop copies away. See comment in
1329 // coalesce.cpp about how this works. It depends critically on the exact
1330 // code shape produced here, so if you are changing this code shape
1331 // make sure the GC info for the heap-top is correct in and around the
1332 // slow-path call.
1333 //
1334
1335 void PhaseMacroExpand::expand_allocate_common(
1336 AllocateNode* alloc, // allocation node to be expanded
1337 Node* length, // array length for an array allocation
1338 const TypeFunc* slow_call_type, // Type of slow call
1339 address slow_call_address // Address of slow call
1340 )
1341 {
1342
1343 Node* ctrl = alloc->in(TypeFunc::Control);
1344 Node* mem = alloc->in(TypeFunc::Memory);
1345 Node* i_o = alloc->in(TypeFunc::I_O);
1346 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1347 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1348 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1349
1350 assert(ctrl != NULL, "must have control");
1351 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1352 // they will not be used if "always_slow" is set
1353 enum { slow_result_path = 1, fast_result_path = 2 };
1354 Node *result_region = NULL;
1355 Node *result_phi_rawmem = NULL;
1356 Node *result_phi_rawoop = NULL;
1357 Node *result_phi_i_o = NULL;
1358
1359 // The initial slow comparison is a size check, the comparison
1360 // we want to do is a BoolTest::gt
1361 bool always_slow = false;
1362 int tv = _igvn.find_int_con(initial_slow_test, -1);
1363 if (tv >= 0) {
1364 always_slow = (tv == 1);
1365 initial_slow_test = NULL;
1366 } else {
1367 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1368 }
1369
1370 if (C->env()->dtrace_alloc_probes() ||
1371 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1372 // Force slow-path allocation
1373 always_slow = true;
1374 initial_slow_test = NULL;
1375 }
1376
1377
1378 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1379 Node *slow_region = NULL;
1380 Node *toobig_false = ctrl;
1381
1382 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1383 // generate the initial test if necessary
1384 if (initial_slow_test != NULL ) {
1385 slow_region = new RegionNode(3);
1386
1387 // Now make the initial failure test. Usually a too-big test but
1388 // might be a TRUE for finalizers or a fancy class check for
1389 // newInstance0.
1390 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1391 transform_later(toobig_iff);
1392 // Plug the failing-too-big test into the slow-path region
1393 Node *toobig_true = new IfTrueNode( toobig_iff );
1394 transform_later(toobig_true);
1395 slow_region ->init_req( too_big_or_final_path, toobig_true );
1396 toobig_false = new IfFalseNode( toobig_iff );
1397 transform_later(toobig_false);
1398 } else { // No initial test, just fall into next case
1399 toobig_false = ctrl;
1400 debug_only(slow_region = NodeSentinel);
1401 }
1402
1403 Node *slow_mem = mem; // save the current memory state for slow path
1404 // generate the fast allocation code unless we know that the initial test will always go slow
1405 if (!always_slow) {
1406 // Fast path modifies only raw memory.
1407 if (mem->is_MergeMem()) {
1408 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1409 }
1410
1411 Node* eden_top_adr;
1412 Node* eden_end_adr;
1413
1414 set_eden_pointers(eden_top_adr, eden_end_adr);
1415
1416 // Load Eden::end. Loop invariant and hoisted.
1417 //
1418 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1419 // a TLAB to work around a bug where these values were being moved across
1420 // a safepoint. These are not oops, so they cannot be include in the oop
1421 // map, but they can be changed by a GC. The proper way to fix this would
1422 // be to set the raw memory state when generating a SafepointNode. However
1423 // this will require extensive changes to the loop optimization in order to
1424 // prevent a degradation of the optimization.
1425 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1426 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1427
1428 // allocate the Region and Phi nodes for the result
1429 result_region = new RegionNode(3);
1430 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1431 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1432 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1433
1434 // We need a Region for the loop-back contended case.
1435 enum { fall_in_path = 1, contended_loopback_path = 2 };
1436 Node *contended_region;
1437 Node *contended_phi_rawmem;
1438 if (UseTLAB) {
1439 contended_region = toobig_false;
1440 contended_phi_rawmem = mem;
1441 } else {
1442 contended_region = new RegionNode(3);
1443 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1444 // Now handle the passing-too-big test. We fall into the contended
1445 // loop-back merge point.
1446 contended_region ->init_req(fall_in_path, toobig_false);
1447 contended_phi_rawmem->init_req(fall_in_path, mem);
1448 transform_later(contended_region);
1449 transform_later(contended_phi_rawmem);
1450 }
1451
1452 // Load(-locked) the heap top.
1453 // See note above concerning the control input when using a TLAB
1454 Node *old_eden_top = UseTLAB
1455 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1456 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1457
1458 transform_later(old_eden_top);
1459 // Add to heap top to get a new heap top
1460 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1461 transform_later(new_eden_top);
1462 // Check for needing a GC; compare against heap end
1463 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1464 transform_later(needgc_cmp);
1465 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1466 transform_later(needgc_bol);
1467 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1468 transform_later(needgc_iff);
1469
1470 // Plug the failing-heap-space-need-gc test into the slow-path region
1471 Node *needgc_true = new IfTrueNode(needgc_iff);
1472 transform_later(needgc_true);
1473 if (initial_slow_test) {
1474 slow_region->init_req(need_gc_path, needgc_true);
1475 // This completes all paths into the slow merge point
1476 transform_later(slow_region);
1477 } else { // No initial slow path needed!
1478 // Just fall from the need-GC path straight into the VM call.
1479 slow_region = needgc_true;
1480 }
1481 // No need for a GC. Setup for the Store-Conditional
1482 Node *needgc_false = new IfFalseNode(needgc_iff);
1483 transform_later(needgc_false);
1484
1485 // Grab regular I/O before optional prefetch may change it.
1486 // Slow-path does no I/O so just set it to the original I/O.
1487 result_phi_i_o->init_req(slow_result_path, i_o);
1488
1489 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1490 old_eden_top, new_eden_top, length);
1491
1492 // Name successful fast-path variables
1493 Node* fast_oop = old_eden_top;
1494 Node* fast_oop_ctrl;
1495 Node* fast_oop_rawmem;
1496
1497 // Store (-conditional) the modified eden top back down.
1498 // StorePConditional produces flags for a test PLUS a modified raw
1499 // memory state.
1500 if (UseTLAB) {
1501 Node* store_eden_top =
1502 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1503 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1504 transform_later(store_eden_top);
1505 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1506 fast_oop_rawmem = store_eden_top;
1507 } else {
1508 Node* store_eden_top =
1509 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1510 new_eden_top, fast_oop/*old_eden_top*/);
1511 transform_later(store_eden_top);
1512 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1513 transform_later(contention_check);
1514 store_eden_top = new SCMemProjNode(store_eden_top);
1515 transform_later(store_eden_top);
1516
1517 // If not using TLABs, check to see if there was contention.
1518 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1519 transform_later(contention_iff);
1520 Node *contention_true = new IfTrueNode(contention_iff);
1521 transform_later(contention_true);
1522 // If contention, loopback and try again.
1523 contended_region->init_req(contended_loopback_path, contention_true);
1524 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1525
1526 // Fast-path succeeded with no contention!
1527 Node *contention_false = new IfFalseNode(contention_iff);
1528 transform_later(contention_false);
1529 fast_oop_ctrl = contention_false;
1530
1531 // Bump total allocated bytes for this thread
1532 Node* thread = new ThreadLocalNode();
1533 transform_later(thread);
1534 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1535 in_bytes(JavaThread::allocated_bytes_offset()));
1536 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1537 0, TypeLong::LONG, T_LONG);
1538 #ifdef _LP64
1539 Node* alloc_size = size_in_bytes;
1540 #else
1541 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1542 transform_later(alloc_size);
1543 #endif
1544 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1545 transform_later(new_alloc_bytes);
1546 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1547 0, new_alloc_bytes, T_LONG);
1548 }
1549
1550 InitializeNode* init = alloc->initialization();
1551 fast_oop_rawmem = initialize_object(alloc,
1552 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1553 klass_node, length, size_in_bytes);
1554
1555 // If initialization is performed by an array copy, any required
1556 // MemBarStoreStore was already added. If the object does not
1557 // escape no need for a MemBarStoreStore. If the object does not
1558 // escape in its initializer and memory barrier (MemBarStoreStore or
1559 // stronger) is already added at exit of initializer, also no need
1560 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1561 // so that stores that initialize this object can't be reordered
1562 // with a subsequent store that makes this object accessible by
1563 // other threads.
1564 // Other threads include java threads and JVM internal threads
1565 // (for example concurrent GC threads). Current concurrent GC
1566 // implementation: CMS and G1 will not scan newly created object,
1567 // so it's safe to skip storestore barrier when allocation does
1568 // not escape.
1569 if (!alloc->does_not_escape_thread() &&
1570 !alloc->is_allocation_MemBar_redundant() &&
1571 (init == NULL || !init->is_complete_with_arraycopy())) {
1572 if (init == NULL || init->req() < InitializeNode::RawStores) {
1573 // No InitializeNode or no stores captured by zeroing
1574 // elimination. Simply add the MemBarStoreStore after object
1575 // initialization.
1576 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1577 transform_later(mb);
1578
1579 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1580 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1581 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1582 transform_later(fast_oop_ctrl);
1583 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1584 transform_later(fast_oop_rawmem);
1585 } else {
1586 // Add the MemBarStoreStore after the InitializeNode so that
1587 // all stores performing the initialization that were moved
1588 // before the InitializeNode happen before the storestore
1589 // barrier.
1590
1591 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1592 Node* init_mem = init->proj_out(TypeFunc::Memory);
1593
1594 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1595 transform_later(mb);
1596
1597 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1598 transform_later(ctrl);
1599 Node* mem = new ProjNode(init,TypeFunc::Memory);
1600 transform_later(mem);
1601
1602 // The MemBarStoreStore depends on control and memory coming
1603 // from the InitializeNode
1604 mb->init_req(TypeFunc::Memory, mem);
1605 mb->init_req(TypeFunc::Control, ctrl);
1606
1607 ctrl = new ProjNode(mb,TypeFunc::Control);
1608 transform_later(ctrl);
1609 mem = new ProjNode(mb,TypeFunc::Memory);
1610 transform_later(mem);
1611
1612 // All nodes that depended on the InitializeNode for control
1613 // and memory must now depend on the MemBarNode that itself
1614 // depends on the InitializeNode
1615 _igvn.replace_node(init_ctrl, ctrl);
1616 _igvn.replace_node(init_mem, mem);
1617 }
1618 }
1619
1620 if (C->env()->dtrace_extended_probes()) {
1621 // Slow-path call
1622 int size = TypeFunc::Parms + 2;
1623 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1624 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1625 "dtrace_object_alloc",
1626 TypeRawPtr::BOTTOM);
1627
1628 // Get base of thread-local storage area
1629 Node* thread = new ThreadLocalNode();
1630 transform_later(thread);
1631
1632 call->init_req(TypeFunc::Parms+0, thread);
1633 call->init_req(TypeFunc::Parms+1, fast_oop);
1634 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1635 call->init_req(TypeFunc::I_O , top()); // does no i/o
1636 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1637 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1638 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1639 transform_later(call);
1640 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1641 transform_later(fast_oop_ctrl);
1642 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1643 transform_later(fast_oop_rawmem);
1644 }
1645
1646 // Plug in the successful fast-path into the result merge point
1647 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1648 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1649 result_phi_i_o ->init_req(fast_result_path, i_o);
1650 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1651 } else {
1652 slow_region = ctrl;
1653 result_phi_i_o = i_o; // Rename it to use in the following code.
1654 }
1655
1656 // Generate slow-path call
1657 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1658 OptoRuntime::stub_name(slow_call_address),
1659 alloc->jvms()->bci(),
1660 TypePtr::BOTTOM);
1661 call->init_req( TypeFunc::Control, slow_region );
1662 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1663 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1664 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1665 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1666
1667 call->init_req(TypeFunc::Parms+0, klass_node);
1668 if (length != NULL) {
1669 call->init_req(TypeFunc::Parms+1, length);
1670 }
1671
1672 // Copy debug information and adjust JVMState information, then replace
1673 // allocate node with the call
1674 copy_call_debug_info((CallNode *) alloc, call);
1675 if (!always_slow) {
1676 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1677 } else {
1678 // Hook i_o projection to avoid its elimination during allocation
1679 // replacement (when only a slow call is generated).
1680 call->set_req(TypeFunc::I_O, result_phi_i_o);
1681 }
1682 _igvn.replace_node(alloc, call);
1683 transform_later(call);
1684
1685 // Identify the output projections from the allocate node and
1686 // adjust any references to them.
1687 // The control and io projections look like:
1688 //
1689 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1690 // Allocate Catch
1691 // ^---Proj(io) <-------+ ^---CatchProj(io)
1692 //
1693 // We are interested in the CatchProj nodes.
1694 //
1695 extract_call_projections(call);
1696
1697 // An allocate node has separate memory projections for the uses on
1698 // the control and i_o paths. Replace the control memory projection with
1699 // result_phi_rawmem (unless we are only generating a slow call when
1700 // both memory projections are combined)
1701 if (!always_slow && _memproj_fallthrough != NULL) {
1702 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1703 Node *use = _memproj_fallthrough->fast_out(i);
1704 _igvn.rehash_node_delayed(use);
1705 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1706 // back up iterator
1707 --i;
1708 }
1709 }
1710 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1711 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1712 if (_memproj_catchall != NULL ) {
1713 if (_memproj_fallthrough == NULL) {
1714 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1715 transform_later(_memproj_fallthrough);
1716 }
1717 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1718 Node *use = _memproj_catchall->fast_out(i);
1719 _igvn.rehash_node_delayed(use);
1720 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1721 // back up iterator
1722 --i;
1723 }
1724 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1725 _igvn.remove_dead_node(_memproj_catchall);
1726 }
1727
1728 // An allocate node has separate i_o projections for the uses on the control
1729 // and i_o paths. Always replace the control i_o projection with result i_o
1730 // otherwise incoming i_o become dead when only a slow call is generated
1731 // (it is different from memory projections where both projections are
1732 // combined in such case).
1733 if (_ioproj_fallthrough != NULL) {
1734 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1735 Node *use = _ioproj_fallthrough->fast_out(i);
1736 _igvn.rehash_node_delayed(use);
1737 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1738 // back up iterator
1739 --i;
1740 }
1741 }
1742 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1743 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1744 if (_ioproj_catchall != NULL ) {
1745 if (_ioproj_fallthrough == NULL) {
1746 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1747 transform_later(_ioproj_fallthrough);
1748 }
1749 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1750 Node *use = _ioproj_catchall->fast_out(i);
1751 _igvn.rehash_node_delayed(use);
1752 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1753 // back up iterator
1754 --i;
1755 }
1756 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1757 _igvn.remove_dead_node(_ioproj_catchall);
1758 }
1759
1760 // if we generated only a slow call, we are done
1761 if (always_slow) {
1762 // Now we can unhook i_o.
1763 if (result_phi_i_o->outcnt() > 1) {
1764 call->set_req(TypeFunc::I_O, top());
1765 } else {
1766 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1767 // Case of new array with negative size known during compilation.
1768 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1769 // following code since call to runtime will throw exception.
1770 // As result there will be no users of i_o after the call.
1771 // Leave i_o attached to this call to avoid problems in preceding graph.
1772 }
1773 return;
1774 }
1775
1776
1777 if (_fallthroughcatchproj != NULL) {
1778 ctrl = _fallthroughcatchproj->clone();
1779 transform_later(ctrl);
1780 _igvn.replace_node(_fallthroughcatchproj, result_region);
1781 } else {
1782 ctrl = top();
1783 }
1784 Node *slow_result;
1785 if (_resproj == NULL) {
1786 // no uses of the allocation result
1787 slow_result = top();
1788 } else {
1789 slow_result = _resproj->clone();
1790 transform_later(slow_result);
1791 _igvn.replace_node(_resproj, result_phi_rawoop);
1792 }
1793
1794 // Plug slow-path into result merge point
1795 result_region ->init_req( slow_result_path, ctrl );
1796 result_phi_rawoop->init_req( slow_result_path, slow_result);
1797 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1798 transform_later(result_region);
1799 transform_later(result_phi_rawoop);
1800 transform_later(result_phi_rawmem);
1801 transform_later(result_phi_i_o);
1802 // This completes all paths into the result merge point
1803 }
1804
1805
1806 // Helper for PhaseMacroExpand::expand_allocate_common.
1807 // Initializes the newly-allocated storage.
1808 Node*
1809 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1810 Node* control, Node* rawmem, Node* object,
1811 Node* klass_node, Node* length,
1812 Node* size_in_bytes) {
1813 InitializeNode* init = alloc->initialization();
1814 // Store the klass & mark bits
1815 Node* mark_node = NULL;
1816 // For now only enable fast locking for non-array types
1817 if (UseBiasedLocking && (length == NULL)) {
1818 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1819 } else {
1820 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1821 }
1822 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1823
1824 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1825 int header_size = alloc->minimum_header_size(); // conservatively small
1826
1827 // Array length
1828 if (length != NULL) { // Arrays need length field
1829 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1830 // conservatively small header size:
1831 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1832 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1833 if (k->is_array_klass()) // we know the exact header size in most cases:
1834 header_size = Klass::layout_helper_header_size(k->layout_helper());
1835 }
1836
1837 // Clear the object body, if necessary.
1838 if (init == NULL) {
1839 // The init has somehow disappeared; be cautious and clear everything.
1840 //
1841 // This can happen if a node is allocated but an uncommon trap occurs
1842 // immediately. In this case, the Initialize gets associated with the
1843 // trap, and may be placed in a different (outer) loop, if the Allocate
1844 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1845 // there can be two Allocates to one Initialize. The answer in all these
1846 // edge cases is safety first. It is always safe to clear immediately
1847 // within an Allocate, and then (maybe or maybe not) clear some more later.
1848 if (!(UseTLAB && ZeroTLAB)) {
1849 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1850 header_size, size_in_bytes,
1851 &_igvn);
1852 }
1853 } else {
1854 if (!init->is_complete()) {
1855 // Try to win by zeroing only what the init does not store.
1856 // We can also try to do some peephole optimizations,
1857 // such as combining some adjacent subword stores.
1858 rawmem = init->complete_stores(control, rawmem, object,
1859 header_size, size_in_bytes, &_igvn);
1860 }
1861 // We have no more use for this link, since the AllocateNode goes away:
1862 init->set_req(InitializeNode::RawAddress, top());
1863 // (If we keep the link, it just confuses the register allocator,
1864 // who thinks he sees a real use of the address by the membar.)
1865 }
1866
1867 return rawmem;
1868 }
1869
1870 // Generate prefetch instructions for next allocations.
1871 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1872 Node*& contended_phi_rawmem,
1873 Node* old_eden_top, Node* new_eden_top,
1874 Node* length) {
1875 enum { fall_in_path = 1, pf_path = 2 };
1876 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1877 // Generate prefetch allocation with watermark check.
1878 // As an allocation hits the watermark, we will prefetch starting
1879 // at a "distance" away from watermark.
1880
1881 Node *pf_region = new RegionNode(3);
1882 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1883 TypeRawPtr::BOTTOM );
1884 // I/O is used for Prefetch
1885 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1886
1887 Node *thread = new ThreadLocalNode();
1888 transform_later(thread);
1889
1890 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1891 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1892 transform_later(eden_pf_adr);
1893
1894 Node *old_pf_wm = new LoadPNode(needgc_false,
1895 contended_phi_rawmem, eden_pf_adr,
1896 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1897 MemNode::unordered);
1898 transform_later(old_pf_wm);
1899
1900 // check against new_eden_top
1901 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1902 transform_later(need_pf_cmp);
1903 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1904 transform_later(need_pf_bol);
1905 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1906 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1907 transform_later(need_pf_iff);
1908
1909 // true node, add prefetchdistance
1910 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1911 transform_later(need_pf_true);
1912
1913 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1914 transform_later(need_pf_false);
1915
1916 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1917 _igvn.MakeConX(AllocatePrefetchDistance) );
1918 transform_later(new_pf_wmt );
1919 new_pf_wmt->set_req(0, need_pf_true);
1920
1921 Node *store_new_wmt = new StorePNode(need_pf_true,
1922 contended_phi_rawmem, eden_pf_adr,
1923 TypeRawPtr::BOTTOM, new_pf_wmt,
1924 MemNode::unordered);
1925 transform_later(store_new_wmt);
1926
1927 // adding prefetches
1928 pf_phi_abio->init_req( fall_in_path, i_o );
1929
1930 Node *prefetch_adr;
1931 Node *prefetch;
1932 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1933 uint step_size = AllocatePrefetchStepSize;
1934 uint distance = 0;
1935
1936 for ( uint i = 0; i < lines; i++ ) {
1937 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1938 _igvn.MakeConX(distance) );
1939 transform_later(prefetch_adr);
1940 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1941 transform_later(prefetch);
1942 distance += step_size;
1943 i_o = prefetch;
1944 }
1945 pf_phi_abio->set_req( pf_path, i_o );
1946
1947 pf_region->init_req( fall_in_path, need_pf_false );
1948 pf_region->init_req( pf_path, need_pf_true );
1949
1950 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1951 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1952
1953 transform_later(pf_region);
1954 transform_later(pf_phi_rawmem);
1955 transform_later(pf_phi_abio);
1956
1957 needgc_false = pf_region;
1958 contended_phi_rawmem = pf_phi_rawmem;
1959 i_o = pf_phi_abio;
1960 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1961 // Insert a prefetch for each allocation.
1962 // This code is used for Sparc with BIS.
1963 Node *pf_region = new RegionNode(3);
1964 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1965 TypeRawPtr::BOTTOM );
1966 transform_later(pf_region);
1967
1968 // Generate several prefetch instructions.
1969 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1970 uint step_size = AllocatePrefetchStepSize;
1971 uint distance = AllocatePrefetchDistance;
1972
1973 // Next cache address.
1974 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1975 _igvn.MakeConX(distance));
1976 transform_later(cache_adr);
1977 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1978 transform_later(cache_adr);
1979 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1980 cache_adr = new AndXNode(cache_adr, mask);
1981 transform_later(cache_adr);
1982 cache_adr = new CastX2PNode(cache_adr);
1983 transform_later(cache_adr);
1984
1985 // Prefetch
1986 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1987 prefetch->set_req(0, needgc_false);
1988 transform_later(prefetch);
1989 contended_phi_rawmem = prefetch;
1990 Node *prefetch_adr;
1991 distance = step_size;
1992 for ( uint i = 1; i < lines; i++ ) {
1993 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1994 _igvn.MakeConX(distance) );
1995 transform_later(prefetch_adr);
1996 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1997 transform_later(prefetch);
1998 distance += step_size;
1999 contended_phi_rawmem = prefetch;
2000 }
2001 } else if( AllocatePrefetchStyle > 0 ) {
2002 // Insert a prefetch for each allocation only on the fast-path
2003 Node *prefetch_adr;
2004 Node *prefetch;
2005 // Generate several prefetch instructions.
2006 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2007 uint step_size = AllocatePrefetchStepSize;
2008 uint distance = AllocatePrefetchDistance;
2009 for ( uint i = 0; i < lines; i++ ) {
2010 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2011 _igvn.MakeConX(distance) );
2012 transform_later(prefetch_adr);
2013 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2014 // Do not let it float too high, since if eden_top == eden_end,
2015 // both might be null.
2016 if( i == 0 ) { // Set control for first prefetch, next follows it
2017 prefetch->init_req(0, needgc_false);
2018 }
2019 transform_later(prefetch);
2020 distance += step_size;
2021 i_o = prefetch;
2022 }
2023 }
2024 return i_o;
2025 }
2026
2027
2028 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2029 expand_allocate_common(alloc, NULL,
2030 OptoRuntime::new_instance_Type(),
2031 OptoRuntime::new_instance_Java());
2032 }
2033
2034 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2035 Node* length = alloc->in(AllocateNode::ALength);
2036 InitializeNode* init = alloc->initialization();
2037 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2038 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2039 address slow_call_address; // Address of slow call
2040 if (init != NULL && init->is_complete_with_arraycopy() &&
2041 k->is_type_array_klass()) {
2042 // Don't zero type array during slow allocation in VM since
2043 // it will be initialized later by arraycopy in compiled code.
2044 slow_call_address = OptoRuntime::new_array_nozero_Java();
2045 } else {
2046 slow_call_address = OptoRuntime::new_array_Java();
2047 }
2048 expand_allocate_common(alloc, length,
2049 OptoRuntime::new_array_Type(),
2050 slow_call_address);
2051 }
2052
2053 //-------------------mark_eliminated_box----------------------------------
2054 //
2055 // During EA obj may point to several objects but after few ideal graph
2056 // transformations (CCP) it may point to only one non escaping object
2057 // (but still using phi), corresponding locks and unlocks will be marked
2058 // for elimination. Later obj could be replaced with a new node (new phi)
2059 // and which does not have escape information. And later after some graph
2060 // reshape other locks and unlocks (which were not marked for elimination
2061 // before) are connected to this new obj (phi) but they still will not be
2062 // marked for elimination since new obj has no escape information.
2063 // Mark all associated (same box and obj) lock and unlock nodes for
2064 // elimination if some of them marked already.
2065 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2066 if (oldbox->as_BoxLock()->is_eliminated())
2067 return; // This BoxLock node was processed already.
2068
2069 // New implementation (EliminateNestedLocks) has separate BoxLock
2070 // node for each locked region so mark all associated locks/unlocks as
2071 // eliminated even if different objects are referenced in one locked region
2072 // (for example, OSR compilation of nested loop inside locked scope).
2073 if (EliminateNestedLocks ||
2074 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2075 // Box is used only in one lock region. Mark this box as eliminated.
2076 _igvn.hash_delete(oldbox);
2077 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2078 _igvn.hash_insert(oldbox);
2079
2080 for (uint i = 0; i < oldbox->outcnt(); i++) {
2081 Node* u = oldbox->raw_out(i);
2082 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2083 AbstractLockNode* alock = u->as_AbstractLock();
2084 // Check lock's box since box could be referenced by Lock's debug info.
2085 if (alock->box_node() == oldbox) {
2086 // Mark eliminated all related locks and unlocks.
2087 #ifdef ASSERT
2088 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2089 #endif
2090 alock->set_non_esc_obj();
2091 }
2092 }
2093 }
2094 return;
2095 }
2096
2097 // Create new "eliminated" BoxLock node and use it in monitor debug info
2098 // instead of oldbox for the same object.
2099 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2100
2101 // Note: BoxLock node is marked eliminated only here and it is used
2102 // to indicate that all associated lock and unlock nodes are marked
2103 // for elimination.
2104 newbox->set_eliminated();
2105 transform_later(newbox);
2106
2107 // Replace old box node with new box for all users of the same object.
2108 for (uint i = 0; i < oldbox->outcnt();) {
2109 bool next_edge = true;
2110
2111 Node* u = oldbox->raw_out(i);
2112 if (u->is_AbstractLock()) {
2113 AbstractLockNode* alock = u->as_AbstractLock();
2114 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2115 // Replace Box and mark eliminated all related locks and unlocks.
2116 #ifdef ASSERT
2117 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2118 #endif
2119 alock->set_non_esc_obj();
2120 _igvn.rehash_node_delayed(alock);
2121 alock->set_box_node(newbox);
2122 next_edge = false;
2123 }
2124 }
2125 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2126 FastLockNode* flock = u->as_FastLock();
2127 assert(flock->box_node() == oldbox, "sanity");
2128 _igvn.rehash_node_delayed(flock);
2129 flock->set_box_node(newbox);
2130 next_edge = false;
2131 }
2132
2133 // Replace old box in monitor debug info.
2134 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2135 SafePointNode* sfn = u->as_SafePoint();
2136 JVMState* youngest_jvms = sfn->jvms();
2137 int max_depth = youngest_jvms->depth();
2138 for (int depth = 1; depth <= max_depth; depth++) {
2139 JVMState* jvms = youngest_jvms->of_depth(depth);
2140 int num_mon = jvms->nof_monitors();
2141 // Loop over monitors
2142 for (int idx = 0; idx < num_mon; idx++) {
2143 Node* obj_node = sfn->monitor_obj(jvms, idx);
2144 Node* box_node = sfn->monitor_box(jvms, idx);
2145 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2146 int j = jvms->monitor_box_offset(idx);
2147 _igvn.replace_input_of(u, j, newbox);
2148 next_edge = false;
2149 }
2150 }
2151 }
2152 }
2153 if (next_edge) i++;
2154 }
2155 }
2156
2157 //-----------------------mark_eliminated_locking_nodes-----------------------
2158 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2159 if (EliminateNestedLocks) {
2160 if (alock->is_nested()) {
2161 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2162 return;
2163 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2164 // Only Lock node has JVMState needed here.
2165 // Not that preceding claim is documented anywhere else.
2166 if (alock->jvms() != NULL) {
2167 if (alock->as_Lock()->is_nested_lock_region()) {
2168 // Mark eliminated related nested locks and unlocks.
2169 Node* obj = alock->obj_node();
2170 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2171 assert(!box_node->is_eliminated(), "should not be marked yet");
2172 // Note: BoxLock node is marked eliminated only here
2173 // and it is used to indicate that all associated lock
2174 // and unlock nodes are marked for elimination.
2175 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2176 for (uint i = 0; i < box_node->outcnt(); i++) {
2177 Node* u = box_node->raw_out(i);
2178 if (u->is_AbstractLock()) {
2179 alock = u->as_AbstractLock();
2180 if (alock->box_node() == box_node) {
2181 // Verify that this Box is referenced only by related locks.
2182 assert(alock->obj_node()->eqv_uncast(obj), "");
2183 // Mark all related locks and unlocks.
2184 #ifdef ASSERT
2185 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2186 #endif
2187 alock->set_nested();
2188 }
2189 }
2190 }
2191 } else {
2192 #ifdef ASSERT
2193 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2194 if (C->log() != NULL)
2195 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2196 #endif
2197 }
2198 }
2199 return;
2200 }
2201 // Process locks for non escaping object
2202 assert(alock->is_non_esc_obj(), "");
2203 } // EliminateNestedLocks
2204
2205 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2206 // Look for all locks of this object and mark them and
2207 // corresponding BoxLock nodes as eliminated.
2208 Node* obj = alock->obj_node();
2209 for (uint j = 0; j < obj->outcnt(); j++) {
2210 Node* o = obj->raw_out(j);
2211 if (o->is_AbstractLock() &&
2212 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2213 alock = o->as_AbstractLock();
2214 Node* box = alock->box_node();
2215 // Replace old box node with new eliminated box for all users
2216 // of the same object and mark related locks as eliminated.
2217 mark_eliminated_box(box, obj);
2218 }
2219 }
2220 }
2221 }
2222
2223 // we have determined that this lock/unlock can be eliminated, we simply
2224 // eliminate the node without expanding it.
2225 //
2226 // Note: The membar's associated with the lock/unlock are currently not
2227 // eliminated. This should be investigated as a future enhancement.
2228 //
2229 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2230
2231 if (!alock->is_eliminated()) {
2232 return false;
2233 }
2234 #ifdef ASSERT
2235 if (!alock->is_coarsened()) {
2236 // Check that new "eliminated" BoxLock node is created.
2237 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2238 assert(oldbox->is_eliminated(), "should be done already");
2239 }
2240 #endif
2241
2242 alock->log_lock_optimization(C, "eliminate_lock");
2243
2244 #ifndef PRODUCT
2245 if (PrintEliminateLocks) {
2246 if (alock->is_Lock()) {
2247 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2248 } else {
2249 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2250 }
2251 }
2252 #endif
2253
2254 Node* mem = alock->in(TypeFunc::Memory);
2255 Node* ctrl = alock->in(TypeFunc::Control);
2256
2257 extract_call_projections(alock);
2258 // There are 2 projections from the lock. The lock node will
2259 // be deleted when its last use is subsumed below.
2260 assert(alock->outcnt() == 2 &&
2261 _fallthroughproj != NULL &&
2262 _memproj_fallthrough != NULL,
2263 "Unexpected projections from Lock/Unlock");
2264
2265 Node* fallthroughproj = _fallthroughproj;
2266 Node* memproj_fallthrough = _memproj_fallthrough;
2267
2268 // The memory projection from a lock/unlock is RawMem
2269 // The input to a Lock is merged memory, so extract its RawMem input
2270 // (unless the MergeMem has been optimized away.)
2271 if (alock->is_Lock()) {
2272 // Seach for MemBarAcquireLock node and delete it also.
2273 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2274 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2275 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2276 Node* memproj = membar->proj_out(TypeFunc::Memory);
2277 _igvn.replace_node(ctrlproj, fallthroughproj);
2278 _igvn.replace_node(memproj, memproj_fallthrough);
2279
2280 // Delete FastLock node also if this Lock node is unique user
2281 // (a loop peeling may clone a Lock node).
2282 Node* flock = alock->as_Lock()->fastlock_node();
2283 if (flock->outcnt() == 1) {
2284 assert(flock->unique_out() == alock, "sanity");
2285 _igvn.replace_node(flock, top());
2286 }
2287 }
2288
2289 // Seach for MemBarReleaseLock node and delete it also.
2290 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2291 ctrl->in(0)->is_MemBar()) {
2292 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2293 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2294 mem->is_Proj() && membar == mem->in(0), "");
2295 _igvn.replace_node(fallthroughproj, ctrl);
2296 _igvn.replace_node(memproj_fallthrough, mem);
2297 fallthroughproj = ctrl;
2298 memproj_fallthrough = mem;
2299 ctrl = membar->in(TypeFunc::Control);
2300 mem = membar->in(TypeFunc::Memory);
2301 }
2302
2303 _igvn.replace_node(fallthroughproj, ctrl);
2304 _igvn.replace_node(memproj_fallthrough, mem);
2305 return true;
2306 }
2307
2308
2309 //------------------------------expand_lock_node----------------------
2310 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2311
2312 Node* ctrl = lock->in(TypeFunc::Control);
2313 Node* mem = lock->in(TypeFunc::Memory);
2314 Node* obj = lock->obj_node();
2315 Node* box = lock->box_node();
2316 Node* flock = lock->fastlock_node();
2317
2318 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2319
2320 // Make the merge point
2321 Node *region;
2322 Node *mem_phi;
2323 Node *slow_path;
2324
2325 if (UseOptoBiasInlining) {
2326 /*
2327 * See the full description in MacroAssembler::biased_locking_enter().
2328 *
2329 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2330 * // The object is biased.
2331 * proto_node = klass->prototype_header;
2332 * o_node = thread | proto_node;
2333 * x_node = o_node ^ mark_word;
2334 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2335 * // Done.
2336 * } else {
2337 * if( (x_node & biased_lock_mask) != 0 ) {
2338 * // The klass's prototype header is no longer biased.
2339 * cas(&mark_word, mark_word, proto_node)
2340 * goto cas_lock;
2341 * } else {
2342 * // The klass's prototype header is still biased.
2343 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2344 * old = mark_word;
2345 * new = o_node;
2346 * } else {
2347 * // Different thread or anonymous biased.
2348 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2349 * new = thread | old;
2350 * }
2351 * // Try to rebias.
2352 * if( cas(&mark_word, old, new) == 0 ) {
2353 * // Done.
2354 * } else {
2355 * goto slow_path; // Failed.
2356 * }
2357 * }
2358 * }
2359 * } else {
2360 * // The object is not biased.
2361 * cas_lock:
2362 * if( FastLock(obj) == 0 ) {
2363 * // Done.
2364 * } else {
2365 * slow_path:
2366 * OptoRuntime::complete_monitor_locking_Java(obj);
2367 * }
2368 * }
2369 */
2370
2371 region = new RegionNode(5);
2372 // create a Phi for the memory state
2373 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2374
2375 Node* fast_lock_region = new RegionNode(3);
2376 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2377
2378 // First, check mark word for the biased lock pattern.
2379 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2380
2381 // Get fast path - mark word has the biased lock pattern.
2382 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2383 markOopDesc::biased_lock_mask_in_place,
2384 markOopDesc::biased_lock_pattern, true);
2385 // fast_lock_region->in(1) is set to slow path.
2386 fast_lock_mem_phi->init_req(1, mem);
2387
2388 // Now check that the lock is biased to the current thread and has
2389 // the same epoch and bias as Klass::_prototype_header.
2390
2391 // Special-case a fresh allocation to avoid building nodes:
2392 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2393 if (klass_node == NULL) {
2394 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2395 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2396 #ifdef _LP64
2397 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2398 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2399 klass_node->in(1)->init_req(0, ctrl);
2400 } else
2401 #endif
2402 klass_node->init_req(0, ctrl);
2403 }
2404 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2405
2406 Node* thread = transform_later(new ThreadLocalNode());
2407 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2408 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2409 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2410
2411 // Get slow path - mark word does NOT match the value.
2412 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2413 (~markOopDesc::age_mask_in_place), 0);
2414 // region->in(3) is set to fast path - the object is biased to the current thread.
2415 mem_phi->init_req(3, mem);
2416
2417
2418 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2419
2420
2421 // First, check biased pattern.
2422 // Get fast path - _prototype_header has the same biased lock pattern.
2423 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2424 markOopDesc::biased_lock_mask_in_place, 0, true);
2425
2426 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2427 // fast_lock_region->in(2) - the prototype header is no longer biased
2428 // and we have to revoke the bias on this object.
2429 // We are going to try to reset the mark of this object to the prototype
2430 // value and fall through to the CAS-based locking scheme.
2431 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2432 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2433 proto_node, mark_node);
2434 transform_later(cas);
2435 Node* proj = transform_later(new SCMemProjNode(cas));
2436 fast_lock_mem_phi->init_req(2, proj);
2437
2438
2439 // Second, check epoch bits.
2440 Node* rebiased_region = new RegionNode(3);
2441 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2442 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2443
2444 // Get slow path - mark word does NOT match epoch bits.
2445 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2446 markOopDesc::epoch_mask_in_place, 0);
2447 // The epoch of the current bias is not valid, attempt to rebias the object
2448 // toward the current thread.
2449 rebiased_region->init_req(2, epoch_ctrl);
2450 old_phi->init_req(2, mark_node);
2451 new_phi->init_req(2, o_node);
2452
2453 // rebiased_region->in(1) is set to fast path.
2454 // The epoch of the current bias is still valid but we know
2455 // nothing about the owner; it might be set or it might be clear.
2456 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2457 markOopDesc::age_mask_in_place |
2458 markOopDesc::epoch_mask_in_place);
2459 Node* old = transform_later(new AndXNode(mark_node, cmask));
2460 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2461 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2462 old_phi->init_req(1, old);
2463 new_phi->init_req(1, new_mark);
2464
2465 transform_later(rebiased_region);
2466 transform_later(old_phi);
2467 transform_later(new_phi);
2468
2469 // Try to acquire the bias of the object using an atomic operation.
2470 // If this fails we will go in to the runtime to revoke the object's bias.
2471 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2472 transform_later(cas);
2473 proj = transform_later(new SCMemProjNode(cas));
2474
2475 // Get slow path - Failed to CAS.
2476 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2477 mem_phi->init_req(4, proj);
2478 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2479
2480 // Failed to CAS.
2481 slow_path = new RegionNode(3);
2482 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2483
2484 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2485 slow_mem->init_req(1, proj);
2486
2487 // Call CAS-based locking scheme (FastLock node).
2488
2489 transform_later(fast_lock_region);
2490 transform_later(fast_lock_mem_phi);
2491
2492 // Get slow path - FastLock failed to lock the object.
2493 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2494 mem_phi->init_req(2, fast_lock_mem_phi);
2495 // region->in(2) is set to fast path - the object is locked to the current thread.
2496
2497 slow_path->init_req(2, ctrl); // Capture slow-control
2498 slow_mem->init_req(2, fast_lock_mem_phi);
2499
2500 transform_later(slow_path);
2501 transform_later(slow_mem);
2502 // Reset lock's memory edge.
2503 lock->set_req(TypeFunc::Memory, slow_mem);
2504
2505 } else {
2506 region = new RegionNode(3);
2507 // create a Phi for the memory state
2508 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2509
2510 // Optimize test; set region slot 2
2511 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2512 mem_phi->init_req(2, mem);
2513 }
2514
2515 // Make slow path call
2516 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2517 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2518 obj, box, NULL);
2519
2520 extract_call_projections(call);
2521
2522 // Slow path can only throw asynchronous exceptions, which are always
2523 // de-opted. So the compiler thinks the slow-call can never throw an
2524 // exception. If it DOES throw an exception we would need the debug
2525 // info removed first (since if it throws there is no monitor).
2526 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2527 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2528
2529 // Capture slow path
2530 // disconnect fall-through projection from call and create a new one
2531 // hook up users of fall-through projection to region
2532 Node *slow_ctrl = _fallthroughproj->clone();
2533 transform_later(slow_ctrl);
2534 _igvn.hash_delete(_fallthroughproj);
2535 _fallthroughproj->disconnect_inputs(NULL, C);
2536 region->init_req(1, slow_ctrl);
2537 // region inputs are now complete
2538 transform_later(region);
2539 _igvn.replace_node(_fallthroughproj, region);
2540
2541 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2542 mem_phi->init_req(1, memproj );
2543 transform_later(mem_phi);
2544 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2545 }
2546
2547 //------------------------------expand_unlock_node----------------------
2548 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2549
2550 Node* ctrl = unlock->in(TypeFunc::Control);
2551 Node* mem = unlock->in(TypeFunc::Memory);
2552 Node* obj = unlock->obj_node();
2553 Node* box = unlock->box_node();
2554
2555 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2556
2557 // No need for a null check on unlock
2558
2559 // Make the merge point
2560 Node *region;
2561 Node *mem_phi;
2562
2563 if (UseOptoBiasInlining) {
2564 // Check for biased locking unlock case, which is a no-op.
2565 // See the full description in MacroAssembler::biased_locking_exit().
2566 region = new RegionNode(4);
2567 // create a Phi for the memory state
2568 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2569 mem_phi->init_req(3, mem);
2570
2571 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2572 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2573 markOopDesc::biased_lock_mask_in_place,
2574 markOopDesc::biased_lock_pattern);
2575 } else {
2576 region = new RegionNode(3);
2577 // create a Phi for the memory state
2578 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2579 }
2580
2581 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2582 funlock = transform_later( funlock )->as_FastUnlock();
2583 // Optimize test; set region slot 2
2584 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2585 Node *thread = transform_later(new ThreadLocalNode());
2586
2587 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2588 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2589 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2590
2591 extract_call_projections(call);
2592
2593 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2594 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2595
2596 // No exceptions for unlocking
2597 // Capture slow path
2598 // disconnect fall-through projection from call and create a new one
2599 // hook up users of fall-through projection to region
2600 Node *slow_ctrl = _fallthroughproj->clone();
2601 transform_later(slow_ctrl);
2602 _igvn.hash_delete(_fallthroughproj);
2603 _fallthroughproj->disconnect_inputs(NULL, C);
2604 region->init_req(1, slow_ctrl);
2605 // region inputs are now complete
2606 transform_later(region);
2607 _igvn.replace_node(_fallthroughproj, region);
2608
2609 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2610 mem_phi->init_req(1, memproj );
2611 mem_phi->init_req(2, mem);
2612 transform_later(mem_phi);
2613 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2614 }
2615
2616 //---------------------------eliminate_macro_nodes----------------------
2617 // Eliminate scalar replaced allocations and associated locks.
2618 void PhaseMacroExpand::eliminate_macro_nodes() {
2619 if (C->macro_count() == 0)
2620 return;
2621
2622 // First, attempt to eliminate locks
2623 int cnt = C->macro_count();
2624 for (int i=0; i < cnt; i++) {
2625 Node *n = C->macro_node(i);
2626 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2627 // Before elimination mark all associated (same box and obj)
2628 // lock and unlock nodes.
2629 mark_eliminated_locking_nodes(n->as_AbstractLock());
2630 }
2631 }
2632 bool progress = true;
2633 while (progress) {
2634 progress = false;
2635 for (int i = C->macro_count(); i > 0; i--) {
2636 Node * n = C->macro_node(i-1);
2637 bool success = false;
2638 debug_only(int old_macro_count = C->macro_count(););
2639 if (n->is_AbstractLock()) {
2640 success = eliminate_locking_node(n->as_AbstractLock());
2641 }
2642 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2643 progress = progress || success;
2644 }
2645 }
2646 // Next, attempt to eliminate allocations
2647 _has_locks = false;
2648 progress = true;
2649 while (progress) {
2650 progress = false;
2651 for (int i = C->macro_count(); i > 0; i--) {
2652 Node * n = C->macro_node(i-1);
2653 bool success = false;
2654 debug_only(int old_macro_count = C->macro_count(););
2655 switch (n->class_id()) {
2656 case Node::Class_Allocate:
2657 case Node::Class_AllocateArray:
2658 success = eliminate_allocate_node(n->as_Allocate());
2659 break;
2660 case Node::Class_CallStaticJava:
2661 success = eliminate_boxing_node(n->as_CallStaticJava());
2662 break;
2663 case Node::Class_Lock:
2664 case Node::Class_Unlock:
2665 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2666 _has_locks = true;
2667 break;
2668 case Node::Class_ArrayCopy:
2669 break;
2670 default:
2671 assert(n->Opcode() == Op_LoopLimit ||
2672 n->Opcode() == Op_Opaque1 ||
2673 n->Opcode() == Op_Opaque2 ||
2674 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2675 }
2676 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2677 progress = progress || success;
2678 }
2679 }
2680 }
2681
2682 //------------------------------expand_macro_nodes----------------------
2683 // Returns true if a failure occurred.
2684 bool PhaseMacroExpand::expand_macro_nodes() {
2685 // Last attempt to eliminate macro nodes.
2686 eliminate_macro_nodes();
2687
2688 // Make sure expansion will not cause node limit to be exceeded.
2689 // Worst case is a macro node gets expanded into about 200 nodes.
2690 // Allow 50% more for optimization.
2691 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2692 return true;
2693
2694 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2695 bool progress = true;
2696 while (progress) {
2697 progress = false;
2698 for (int i = C->macro_count(); i > 0; i--) {
2699 Node * n = C->macro_node(i-1);
2700 bool success = false;
2701 debug_only(int old_macro_count = C->macro_count(););
2702 if (n->Opcode() == Op_LoopLimit) {
2703 // Remove it from macro list and put on IGVN worklist to optimize.
2704 C->remove_macro_node(n);
2705 _igvn._worklist.push(n);
2706 success = true;
2707 } else if (n->Opcode() == Op_CallStaticJava) {
2708 // Remove it from macro list and put on IGVN worklist to optimize.
2709 C->remove_macro_node(n);
2710 _igvn._worklist.push(n);
2711 success = true;
2712 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2713 _igvn.replace_node(n, n->in(1));
2714 success = true;
2715 #if INCLUDE_RTM_OPT
2716 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2717 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2718 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2719 Node* cmp = n->unique_out();
2720 #ifdef ASSERT
2721 // Validate graph.
2722 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2723 BoolNode* bol = cmp->unique_out()->as_Bool();
2724 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2725 (bol->_test._test == BoolTest::ne), "");
2726 IfNode* ifn = bol->unique_out()->as_If();
2727 assert((ifn->outcnt() == 2) &&
2728 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2729 #endif
2730 Node* repl = n->in(1);
2731 if (!_has_locks) {
2732 // Remove RTM state check if there are no locks in the code.
2733 // Replace input to compare the same value.
2734 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2735 }
2736 _igvn.replace_node(n, repl);
2737 success = true;
2738 #endif
2739 }
2740 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2741 progress = progress || success;
2742 }
2743 }
2744
2745 // expand arraycopy "macro" nodes first
2746 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2747 // before the allocate nodes are expanded.
2748 int macro_idx = C->macro_count() - 1;
2749 while (macro_idx >= 0) {
2750 Node * n = C->macro_node(macro_idx);
2751 assert(n->is_macro(), "only macro nodes expected here");
2752 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2753 // node is unreachable, so don't try to expand it
2754 C->remove_macro_node(n);
2755 } else if (n->is_ArrayCopy()){
2756 int macro_count = C->macro_count();
2757 expand_arraycopy_node(n->as_ArrayCopy());
2758 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2759 }
2760 if (C->failing()) return true;
2761 macro_idx --;
2762 }
2763
2764 // expand "macro" nodes
2765 // nodes are removed from the macro list as they are processed
2766 while (C->macro_count() > 0) {
2767 int macro_count = C->macro_count();
2768 Node * n = C->macro_node(macro_count-1);
2769 assert(n->is_macro(), "only macro nodes expected here");
2770 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2771 // node is unreachable, so don't try to expand it
2772 C->remove_macro_node(n);
2773 continue;
2774 }
2775 switch (n->class_id()) {
2776 case Node::Class_Allocate:
2777 expand_allocate(n->as_Allocate());
2778 break;
2779 case Node::Class_AllocateArray:
2780 expand_allocate_array(n->as_AllocateArray());
2781 break;
2782 case Node::Class_Lock:
2783 expand_lock_node(n->as_Lock());
2784 break;
2785 case Node::Class_Unlock:
2786 expand_unlock_node(n->as_Unlock());
2787 break;
2788 default:
2789 assert(false, "unknown node type in macro list");
2790 }
2791 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2792 if (C->failing()) return true;
2793 }
2794
2795 _igvn.set_delay_transform(false);
2796 _igvn.optimize();
2797 if (C->failing()) return true;
2798 return false;
2799 }