1 /*
2 * Copyright (c) 2005, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shenandoah/brooksPointer.hpp"
28 #include "libadt/vectset.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/loopnode.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/node.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/phaseX.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/shenandoahSupport.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "runtime/sharedRuntime.hpp"
51
52
53 //
54 // Replace any references to "oldref" in inputs to "use" with "newref".
55 // Returns the number of replacements made.
56 //
57 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
58 int nreplacements = 0;
59 uint req = use->req();
60 for (uint j = 0; j < use->len(); j++) {
61 Node *uin = use->in(j);
62 if (uin == oldref) {
63 if (j < req)
64 use->set_req(j, newref);
65 else
66 use->set_prec(j, newref);
67 nreplacements++;
68 } else if (j >= req && uin == NULL) {
69 break;
70 }
71 }
72 return nreplacements;
73 }
74
75 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
76 // Copy debug information and adjust JVMState information
77 uint old_dbg_start = oldcall->tf()->domain()->cnt();
78 uint new_dbg_start = newcall->tf()->domain()->cnt();
79 int jvms_adj = new_dbg_start - old_dbg_start;
80 assert (new_dbg_start == newcall->req(), "argument count mismatch");
81
82 // SafePointScalarObject node could be referenced several times in debug info.
83 // Use Dict to record cloned nodes.
84 Dict* sosn_map = new Dict(cmpkey,hashkey);
85 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
86 Node* old_in = oldcall->in(i);
87 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
88 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
89 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
90 uint old_unique = C->unique();
91 Node* new_in = old_sosn->clone(sosn_map);
92 if (old_unique != C->unique()) { // New node?
93 new_in->set_req(0, C->root()); // reset control edge
94 new_in = transform_later(new_in); // Register new node.
95 }
96 old_in = new_in;
97 }
98 newcall->add_req(old_in);
99 }
100
101 // JVMS may be shared so clone it before we modify it
102 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
103 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
104 jvms->set_map(newcall);
105 jvms->set_locoff(jvms->locoff()+jvms_adj);
106 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
107 jvms->set_monoff(jvms->monoff()+jvms_adj);
108 jvms->set_scloff(jvms->scloff()+jvms_adj);
109 jvms->set_endoff(jvms->endoff()+jvms_adj);
110 }
111 }
112
113 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
114 Node* cmp;
115 if (mask != 0) {
116 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
117 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
118 } else {
119 cmp = word;
120 }
121 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
122 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
123 transform_later(iff);
124
125 // Fast path taken.
126 Node *fast_taken = transform_later(new IfFalseNode(iff));
127
128 // Fast path not-taken, i.e. slow path
129 Node *slow_taken = transform_later(new IfTrueNode(iff));
130
131 if (return_fast_path) {
132 region->init_req(edge, slow_taken); // Capture slow-control
133 return fast_taken;
134 } else {
135 region->init_req(edge, fast_taken); // Capture fast-control
136 return slow_taken;
137 }
138 }
139
140 //--------------------copy_predefined_input_for_runtime_call--------------------
141 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
142 // Set fixed predefined input arguments
143 call->init_req( TypeFunc::Control, ctrl );
144 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
145 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
146 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
147 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
148 }
149
150 //------------------------------make_slow_call---------------------------------
151 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
152 address slow_call, const char* leaf_name, Node* slow_path,
153 Node* parm0, Node* parm1, Node* parm2) {
154
155 // Slow-path call
156 CallNode *call = leaf_name
157 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
158 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
159
160 // Slow path call has no side-effects, uses few values
161 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
162 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
163 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
164 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
165 copy_call_debug_info(oldcall, call);
166 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
167 _igvn.replace_node(oldcall, call);
168 transform_later(call);
169
170 return call;
171 }
172
173 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
174 _fallthroughproj = NULL;
175 _fallthroughcatchproj = NULL;
176 _ioproj_fallthrough = NULL;
177 _ioproj_catchall = NULL;
178 _catchallcatchproj = NULL;
179 _memproj_fallthrough = NULL;
180 _memproj_catchall = NULL;
181 _resproj = NULL;
182 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
183 ProjNode *pn = call->fast_out(i)->as_Proj();
184 switch (pn->_con) {
185 case TypeFunc::Control:
186 {
187 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
188 _fallthroughproj = pn;
189 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
190 const Node *cn = pn->fast_out(j);
191 if (cn->is_Catch()) {
192 ProjNode *cpn = NULL;
193 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
194 cpn = cn->fast_out(k)->as_Proj();
195 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
196 if (cpn->_con == CatchProjNode::fall_through_index)
197 _fallthroughcatchproj = cpn;
198 else {
199 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
200 _catchallcatchproj = cpn;
201 }
202 }
203 }
204 break;
205 }
206 case TypeFunc::I_O:
207 if (pn->_is_io_use)
208 _ioproj_catchall = pn;
209 else
210 _ioproj_fallthrough = pn;
211 break;
212 case TypeFunc::Memory:
213 if (pn->_is_io_use)
214 _memproj_catchall = pn;
215 else
216 _memproj_fallthrough = pn;
217 break;
218 case TypeFunc::Parms:
219 _resproj = pn;
220 break;
221 default:
222 assert(false, "unexpected projection from allocation node.");
223 }
224 }
225
226 }
227
228 // Eliminate a card mark sequence. p2x is a ConvP2XNode
229 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
230 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
231 if (!UseG1GC) {
232 // vanilla/CMS post barrier
233 Node *shift = p2x->unique_out();
234 Node *addp = shift->unique_out();
235 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
236 Node *mem = addp->last_out(j);
237 if (UseCondCardMark && mem->is_Load()) {
238 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
239 // The load is checking if the card has been written so
240 // replace it with zero to fold the test.
241 _igvn.replace_node(mem, intcon(0));
242 continue;
243 }
244 assert(mem->is_Store(), "store required");
245 _igvn.replace_node(mem, mem->in(MemNode::Memory));
246 }
247 } else {
248 // G1 pre/post barriers
249 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
250 // It could be only one user, URShift node, in Object.clone() intrinsic
251 // but the new allocation is passed to arraycopy stub and it could not
252 // be scalar replaced. So we don't check the case.
253
254 // An other case of only one user (Xor) is when the value check for NULL
255 // in G1 post barrier is folded after CCP so the code which used URShift
256 // is removed.
257
258 // Take Region node before eliminating post barrier since it also
259 // eliminates CastP2X node when it has only one user.
260 Node* this_region = p2x->in(0);
261 assert(this_region != NULL, "");
262
263 // Remove G1 post barrier.
264
265 // Search for CastP2X->Xor->URShift->Cmp path which
266 // checks if the store done to a different from the value's region.
267 // And replace Cmp with #0 (false) to collapse G1 post barrier.
268 Node* xorx = p2x->find_out_with(Op_XorX);
269 if (xorx != NULL) {
270 Node* shift = xorx->unique_out();
271 Node* cmpx = shift->unique_out();
272 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
273 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
274 "missing region check in G1 post barrier");
275 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
276
277 // Remove G1 pre barrier.
278
279 // Search "if (marking != 0)" check and set it to "false".
280 // There is no G1 pre barrier if previous stored value is NULL
281 // (for example, after initialization).
282 if (this_region->is_Region() && this_region->req() == 3) {
283 int ind = 1;
284 if (!this_region->in(ind)->is_IfFalse()) {
285 ind = 2;
286 }
287 if (this_region->in(ind)->is_IfFalse()) {
288 Node* bol = this_region->in(ind)->in(0)->in(1);
289 assert(bol->is_Bool(), "");
290 cmpx = bol->in(1);
291 if (bol->as_Bool()->_test._test == BoolTest::ne &&
292 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
293 cmpx->in(1)->is_Load()) {
294 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
295 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
296 SATBMarkQueue::byte_offset_of_active());
297 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
298 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
299 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
300 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
301 }
302 }
303 }
304 }
305 } else {
306 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
307 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
308 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
309 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
310 Node* shift = p2x->find_out_with(Op_URShiftX);
311 assert(shift != NULL, "missing G1 post barrier");
312 Node* addp = shift->unique_out();
313 Node* load = addp->find_out_with(Op_LoadB);
314 assert(load != NULL, "missing G1 post barrier");
315 Node* cmpx = load->unique_out();
316 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
317 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
318 "missing card value check in G1 post barrier");
319 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
320 // There is no G1 pre barrier in this case
321 }
322 // Now CastP2X can be removed since it is used only on dead path
323 // which currently still alive until igvn optimize it.
324 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
325 _igvn.replace_node(p2x, top());
326 }
327 }
328
329 // Search for a memory operation for the specified memory slice.
330 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
331 Node *orig_mem = mem;
332 Node *alloc_mem = alloc->in(TypeFunc::Memory);
333 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
334 while (true) {
335 if (mem == alloc_mem || mem == start_mem ) {
336 return mem; // hit one of our sentinels
337 } else if (mem->is_MergeMem()) {
338 mem = mem->as_MergeMem()->memory_at(alias_idx);
339 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
340 Node *in = mem->in(0);
341 // we can safely skip over safepoints, calls, locks and membars because we
342 // already know that the object is safe to eliminate.
343 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
344 return in;
345 } else if (in->is_Call()) {
346 CallNode *call = in->as_Call();
347 if (call->may_modify(tinst, phase)) {
348 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
349 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
350 return in;
351 }
352 }
353 mem = in->in(TypeFunc::Memory);
354 } else if (in->is_MemBar()) {
355 ArrayCopyNode* ac = NULL;
356 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
357 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
358 return ac;
359 }
360 mem = in->in(TypeFunc::Memory);
361 } else {
362 assert(false, "unexpected projection");
363 }
364 } else if (mem->is_Store()) {
365 const TypePtr* atype = mem->as_Store()->adr_type();
366 int adr_idx = phase->C->get_alias_index(atype);
367 if (adr_idx == alias_idx) {
368 assert(atype->isa_oopptr(), "address type must be oopptr");
369 int adr_offset = atype->offset();
370 uint adr_iid = atype->is_oopptr()->instance_id();
371 // Array elements references have the same alias_idx
372 // but different offset and different instance_id.
373 if (adr_offset == offset && adr_iid == alloc->_idx)
374 return mem;
375 } else {
376 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
377 }
378 mem = mem->in(MemNode::Memory);
379 } else if (mem->is_ClearArray()) {
380 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
381 // Can not bypass initialization of the instance
382 // we are looking.
383 debug_only(intptr_t offset;)
384 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
385 InitializeNode* init = alloc->as_Allocate()->initialization();
386 // We are looking for stored value, return Initialize node
387 // or memory edge from Allocate node.
388 if (init != NULL)
389 return init;
390 else
391 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
392 }
393 // Otherwise skip it (the call updated 'mem' value).
394 } else if (mem->Opcode() == Op_SCMemProj) {
395 mem = mem->in(0);
396 Node* adr = NULL;
397 if (mem->is_LoadStore()) {
398 adr = mem->in(MemNode::Address);
399 } else {
400 assert(mem->Opcode() == Op_EncodeISOArray ||
401 mem->Opcode() == Op_StrCompressedCopy, "sanity");
402 adr = mem->in(3); // Destination array
403 }
404 const TypePtr* atype = adr->bottom_type()->is_ptr();
405 int adr_idx = phase->C->get_alias_index(atype);
406 if (adr_idx == alias_idx) {
407 DEBUG_ONLY(mem->dump();)
408 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
409 return NULL;
410 }
411 mem = mem->in(MemNode::Memory);
412 } else if (mem->Opcode() == Op_StrInflatedCopy) {
413 Node* adr = mem->in(3); // Destination array
414 const TypePtr* atype = adr->bottom_type()->is_ptr();
415 int adr_idx = phase->C->get_alias_index(atype);
416 if (adr_idx == alias_idx) {
417 DEBUG_ONLY(mem->dump();)
418 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
419 return NULL;
420 }
421 mem = mem->in(MemNode::Memory);
422 } else {
423 return mem;
424 }
425 assert(mem != orig_mem, "dead memory loop");
426 }
427 }
428
429 // Generate loads from source of the arraycopy for fields of
430 // destination needed at a deoptimization point
431 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, BasicType ft, const Type *ftype, AllocateNode *alloc) {
432 BasicType bt = ft;
433 const Type *type = ftype;
434 if (ft == T_NARROWOOP) {
435 bt = T_OBJECT;
436 type = ftype->make_oopptr();
437 }
438 Node* res = NULL;
439 if (ac->is_clonebasic()) {
440 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
441 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
442 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
443 Node* m = ac->in(TypeFunc::Memory);
444 while (m->is_MergeMem()) {
445 m = m->as_MergeMem()->memory_at(C->get_alias_index(adr_type));
446 if (m->is_Proj() && m->in(0)->is_MemBar()) {
447 m = m->in(0)->in(TypeFunc::Memory);
448 }
449 }
450 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
451 } else {
452 if (ac->modifies(offset, offset, &_igvn, true)) {
453 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
454 uint shift = exact_log2(type2aelembytes(bt));
455 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
456 #ifdef _LP64
457 diff = _igvn.transform(new ConvI2LNode(diff));
458 #endif
459 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
460
461 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
462 Node* base = ac->in(ArrayCopyNode::Src);
463 Node* adr = _igvn.transform(new AddPNode(base, base, off));
464 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
465 Node* m = ac->in(TypeFunc::Memory);
466 res = LoadNode::make(_igvn, ctl, m, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
467 }
468 }
469 if (res != NULL) {
470 res = _igvn.transform(res);
471 if (ftype->isa_narrowoop()) {
472 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
473 res = _igvn.transform(new EncodePNode(res, ftype));
474 }
475 return res;
476 }
477 return NULL;
478 }
479
480 //
481 // Given a Memory Phi, compute a value Phi containing the values from stores
482 // on the input paths.
483 // Note: this function is recursive, its depth is limited by the "level" argument
484 // Returns the computed Phi, or NULL if it cannot compute it.
485 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) {
486 assert(mem->is_Phi(), "sanity");
487 int alias_idx = C->get_alias_index(adr_t);
488 int offset = adr_t->offset();
489 int instance_id = adr_t->instance_id();
490
491 // Check if an appropriate value phi already exists.
492 Node* region = mem->in(0);
493 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
494 Node* phi = region->fast_out(k);
495 if (phi->is_Phi() && phi != mem &&
496 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
497 return phi;
498 }
499 }
500 // Check if an appropriate new value phi already exists.
501 Node* new_phi = value_phis->find(mem->_idx);
502 if (new_phi != NULL)
503 return new_phi;
504
505 if (level <= 0) {
506 return NULL; // Give up: phi tree too deep
507 }
508 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
509 Node *alloc_mem = alloc->in(TypeFunc::Memory);
510
511 uint length = mem->req();
512 GrowableArray <Node *> values(length, length, NULL, false);
513
514 // create a new Phi for the value
515 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
516 transform_later(phi);
517 value_phis->push(phi, mem->_idx);
518
519 for (uint j = 1; j < length; j++) {
520 Node *in = mem->in(j);
521 if (in == NULL || in->is_top()) {
522 values.at_put(j, in);
523 } else {
524 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
525 if (val == start_mem || val == alloc_mem) {
526 // hit a sentinel, return appropriate 0 value
527 values.at_put(j, _igvn.zerocon(ft));
528 continue;
529 }
530 if (val->is_Initialize()) {
531 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
532 }
533 if (val == NULL) {
534 return NULL; // can't find a value on this path
535 }
536 if (val == mem) {
537 values.at_put(j, mem);
538 } else if (val->is_Store()) {
539 values.at_put(j, ShenandoahBarrierNode::skip_through_barrier(val->in(MemNode::ValueIn)));
540 } else if(val->is_Proj() && val->in(0) == alloc) {
541 values.at_put(j, _igvn.zerocon(ft));
542 } else if (val->is_Phi()) {
543 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
544 if (val == NULL) {
545 return NULL;
546 }
547 values.at_put(j, val);
548 } else if (val->Opcode() == Op_SCMemProj) {
549 assert(val->in(0)->is_LoadStore() ||
550 val->in(0)->Opcode() == Op_EncodeISOArray ||
551 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
552 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
553 return NULL;
554 } else if (val->is_ArrayCopy()) {
555 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), ft, phi_type, alloc);
556 if (res == NULL) {
557 return NULL;
558 }
559 values.at_put(j, res);
560 } else {
561 #ifdef ASSERT
562 val->dump();
563 assert(false, "unknown node on this path");
564 #endif
565 return NULL; // unknown node on this path
566 }
567 }
568 }
569 // Set Phi's inputs
570 for (uint j = 1; j < length; j++) {
571 if (values.at(j) == mem) {
572 phi->init_req(j, phi);
573 } else {
574 phi->init_req(j, values.at(j));
575 }
576 }
577 return phi;
578 }
579
580 // Search the last value stored into the object's field.
581 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
582 assert(adr_t->is_known_instance_field(), "instance required");
583 int instance_id = adr_t->instance_id();
584 assert((uint)instance_id == alloc->_idx, "wrong allocation");
585
586 int alias_idx = C->get_alias_index(adr_t);
587 int offset = adr_t->offset();
588 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
589 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
590 Node *alloc_mem = alloc->in(TypeFunc::Memory);
591 Arena *a = Thread::current()->resource_area();
592 VectorSet visited(a);
593
594
595 bool done = sfpt_mem == alloc_mem;
596 Node *mem = sfpt_mem;
597 while (!done) {
598 if (visited.test_set(mem->_idx)) {
599 return NULL; // found a loop, give up
600 }
601 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
602 if (mem == start_mem || mem == alloc_mem) {
603 done = true; // hit a sentinel, return appropriate 0 value
604 } else if (mem->is_Initialize()) {
605 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
606 if (mem == NULL) {
607 done = true; // Something go wrong.
608 } else if (mem->is_Store()) {
609 const TypePtr* atype = mem->as_Store()->adr_type();
610 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
611 done = true;
612 }
613 } else if (mem->is_Store()) {
614 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
615 assert(atype != NULL, "address type must be oopptr");
616 assert(C->get_alias_index(atype) == alias_idx &&
617 atype->is_known_instance_field() && atype->offset() == offset &&
618 atype->instance_id() == instance_id, "store is correct memory slice");
619 done = true;
620 } else if (mem->is_Phi()) {
621 // try to find a phi's unique input
622 Node *unique_input = NULL;
623 Node *top = C->top();
624 for (uint i = 1; i < mem->req(); i++) {
625 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
626 if (n == NULL || n == top || n == mem) {
627 continue;
628 } else if (unique_input == NULL) {
629 unique_input = n;
630 } else if (unique_input != n) {
631 unique_input = top;
632 break;
633 }
634 }
635 if (unique_input != NULL && unique_input != top) {
636 mem = unique_input;
637 } else {
638 done = true;
639 }
640 } else if (mem->is_ArrayCopy()) {
641 done = true;
642 } else {
643 assert(false, "unexpected node");
644 }
645 }
646 if (mem != NULL) {
647 if (mem == start_mem || mem == alloc_mem) {
648 // hit a sentinel, return appropriate 0 value
649 return _igvn.zerocon(ft);
650 } else if (mem->is_Store()) {
651 return ShenandoahBarrierNode::skip_through_barrier(mem->in(MemNode::ValueIn));
652 } else if (mem->is_Phi()) {
653 // attempt to produce a Phi reflecting the values on the input paths of the Phi
654 Node_Stack value_phis(a, 8);
655 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
656 if (phi != NULL) {
657 return phi;
658 } else {
659 // Kill all new Phis
660 while(value_phis.is_nonempty()) {
661 Node* n = value_phis.node();
662 _igvn.replace_node(n, C->top());
663 value_phis.pop();
664 }
665 }
666 } else if (mem->is_ArrayCopy()) {
667 Node* ctl = mem->in(0);
668 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
669 // pin the loads in the uncommon trap path
670 ctl = sfpt_ctl;
671 }
672 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, ft, ftype, alloc);
673 }
674 }
675 // Something go wrong.
676 return NULL;
677 }
678
679 // Check the possibility of scalar replacement.
680 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
681 // Scan the uses of the allocation to check for anything that would
682 // prevent us from eliminating it.
683 NOT_PRODUCT( const char* fail_eliminate = NULL; )
684 DEBUG_ONLY( Node* disq_node = NULL; )
685 bool can_eliminate = true;
686
687 Node* res = alloc->result_cast();
688 const TypeOopPtr* res_type = NULL;
689 if (res == NULL) {
690 // All users were eliminated.
691 } else if (!res->is_CheckCastPP()) {
692 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
693 can_eliminate = false;
694 } else {
695 res_type = _igvn.type(res)->isa_oopptr();
696 if (res_type == NULL) {
697 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
698 can_eliminate = false;
699 } else if (res_type->isa_aryptr()) {
700 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
701 if (length < 0) {
702 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
703 can_eliminate = false;
704 }
705 }
706 }
707
708 if (can_eliminate && res != NULL) {
709 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
710 j < jmax && can_eliminate; j++) {
711 Node* use = res->fast_out(j);
712
713 if (use->is_AddP()) {
714 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
715 int offset = addp_type->offset();
716
717 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
718 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
719 can_eliminate = false;
720 break;
721 }
722 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
723 k < kmax && can_eliminate; k++) {
724 Node* n = use->fast_out(k);
725 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
726 !(n->is_ArrayCopy() &&
727 n->as_ArrayCopy()->is_clonebasic() &&
728 n->in(ArrayCopyNode::Dest) == use)) {
729 DEBUG_ONLY(disq_node = n;)
730 if (n->is_Load() || n->is_LoadStore()) {
731 NOT_PRODUCT(fail_eliminate = "Field load";)
732 } else {
733 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
734 }
735 can_eliminate = false;
736 }
737 }
738 } else if (use->is_ArrayCopy() &&
739 (use->as_ArrayCopy()->is_arraycopy_validated() ||
740 use->as_ArrayCopy()->is_copyof_validated() ||
741 use->as_ArrayCopy()->is_copyofrange_validated()) &&
742 use->in(ArrayCopyNode::Dest) == res) {
743 // ok to eliminate
744 } else if (use->is_SafePoint()) {
745 SafePointNode* sfpt = use->as_SafePoint();
746 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
747 // Object is passed as argument.
748 DEBUG_ONLY(disq_node = use;)
749 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
750 can_eliminate = false;
751 }
752 Node* sfptMem = sfpt->memory();
753 if (sfptMem == NULL || sfptMem->is_top()) {
754 DEBUG_ONLY(disq_node = use;)
755 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
756 can_eliminate = false;
757 } else {
758 safepoints.append_if_missing(sfpt);
759 }
760 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
761 if (use->is_Phi()) {
762 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
763 NOT_PRODUCT(fail_eliminate = "Object is return value";)
764 } else {
765 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
766 }
767 DEBUG_ONLY(disq_node = use;)
768 } else {
769 if (use->Opcode() == Op_Return) {
770 NOT_PRODUCT(fail_eliminate = "Object is return value";)
771 }else {
772 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
773 }
774 DEBUG_ONLY(disq_node = use;)
775 }
776 can_eliminate = false;
777 }
778 }
779 }
780
781 #ifndef PRODUCT
782 if (PrintEliminateAllocations) {
783 if (can_eliminate) {
784 tty->print("Scalar ");
785 if (res == NULL)
786 alloc->dump();
787 else
788 res->dump();
789 } else if (alloc->_is_scalar_replaceable) {
790 tty->print("NotScalar (%s)", fail_eliminate);
791 if (res == NULL)
792 alloc->dump();
793 else
794 res->dump();
795 #ifdef ASSERT
796 if (disq_node != NULL) {
797 tty->print(" >>>> ");
798 disq_node->dump();
799 }
800 #endif /*ASSERT*/
801 }
802 }
803 #endif
804 return can_eliminate;
805 }
806
807 // Do scalar replacement.
808 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
809 GrowableArray <SafePointNode *> safepoints_done;
810
811 ciKlass* klass = NULL;
812 ciInstanceKlass* iklass = NULL;
813 int nfields = 0;
814 int array_base = 0;
815 int element_size = 0;
816 BasicType basic_elem_type = T_ILLEGAL;
817 ciType* elem_type = NULL;
818
819 Node* res = alloc->result_cast();
820 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
821 const TypeOopPtr* res_type = NULL;
822 if (res != NULL) { // Could be NULL when there are no users
823 res_type = _igvn.type(res)->isa_oopptr();
824 }
825
826 if (res != NULL) {
827 klass = res_type->klass();
828 if (res_type->isa_instptr()) {
829 // find the fields of the class which will be needed for safepoint debug information
830 assert(klass->is_instance_klass(), "must be an instance klass.");
831 iklass = klass->as_instance_klass();
832 nfields = iklass->nof_nonstatic_fields();
833 } else {
834 // find the array's elements which will be needed for safepoint debug information
835 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
836 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
837 elem_type = klass->as_array_klass()->element_type();
838 basic_elem_type = elem_type->basic_type();
839 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
840 element_size = type2aelembytes(basic_elem_type);
841 }
842 }
843 //
844 // Process the safepoint uses
845 //
846 while (safepoints.length() > 0) {
847 SafePointNode* sfpt = safepoints.pop();
848 Node* mem = sfpt->memory();
849 Node* ctl = sfpt->control();
850 assert(sfpt->jvms() != NULL, "missed JVMS");
851 // Fields of scalar objs are referenced only at the end
852 // of regular debuginfo at the last (youngest) JVMS.
853 // Record relative start index.
854 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
855 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
856 #ifdef ASSERT
857 alloc,
858 #endif
859 first_ind, nfields);
860 sobj->init_req(0, C->root());
861 transform_later(sobj);
862
863 // Scan object's fields adding an input to the safepoint for each field.
864 for (int j = 0; j < nfields; j++) {
865 intptr_t offset;
866 ciField* field = NULL;
867 if (iklass != NULL) {
868 field = iklass->nonstatic_field_at(j);
869 offset = field->offset();
870 elem_type = field->type();
871 basic_elem_type = field->layout_type();
872 } else {
873 offset = array_base + j * (intptr_t)element_size;
874 }
875
876 const Type *field_type;
877 // The next code is taken from Parse::do_get_xxx().
878 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
879 if (!elem_type->is_loaded()) {
880 field_type = TypeInstPtr::BOTTOM;
881 } else if (field != NULL && field->is_static_constant()) {
882 // This can happen if the constant oop is non-perm.
883 ciObject* con = field->constant_value().as_object();
884 // Do not "join" in the previous type; it doesn't add value,
885 // and may yield a vacuous result if the field is of interface type.
886 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
887 assert(field_type != NULL, "field singleton type must be consistent");
888 } else {
889 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
890 }
891 if (UseCompressedOops) {
892 field_type = field_type->make_narrowoop();
893 basic_elem_type = T_NARROWOOP;
894 }
895 } else {
896 field_type = Type::get_const_basic_type(basic_elem_type);
897 }
898
899 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
900
901 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
902 if (field_val == NULL) {
903 // We weren't able to find a value for this field,
904 // give up on eliminating this allocation.
905
906 // Remove any extra entries we added to the safepoint.
907 uint last = sfpt->req() - 1;
908 for (int k = 0; k < j; k++) {
909 sfpt->del_req(last--);
910 }
911 _igvn._worklist.push(sfpt);
912 // rollback processed safepoints
913 while (safepoints_done.length() > 0) {
914 SafePointNode* sfpt_done = safepoints_done.pop();
915 // remove any extra entries we added to the safepoint
916 last = sfpt_done->req() - 1;
917 for (int k = 0; k < nfields; k++) {
918 sfpt_done->del_req(last--);
919 }
920 JVMState *jvms = sfpt_done->jvms();
921 jvms->set_endoff(sfpt_done->req());
922 // Now make a pass over the debug information replacing any references
923 // to SafePointScalarObjectNode with the allocated object.
924 int start = jvms->debug_start();
925 int end = jvms->debug_end();
926 for (int i = start; i < end; i++) {
927 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
928 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
929 if (scobj->first_index(jvms) == sfpt_done->req() &&
930 scobj->n_fields() == (uint)nfields) {
931 assert(scobj->alloc() == alloc, "sanity");
932 sfpt_done->set_req(i, res);
933 }
934 }
935 }
936 _igvn._worklist.push(sfpt_done);
937 }
938 #ifndef PRODUCT
939 if (PrintEliminateAllocations) {
940 if (field != NULL) {
941 tty->print("=== At SafePoint node %d can't find value of Field: ",
942 sfpt->_idx);
943 field->print();
944 int field_idx = C->get_alias_index(field_addr_type);
945 tty->print(" (alias_idx=%d)", field_idx);
946 } else { // Array's element
947 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
948 sfpt->_idx, j);
949 }
950 tty->print(", which prevents elimination of: ");
951 if (res == NULL)
952 alloc->dump();
953 else
954 res->dump();
955 }
956 #endif
957 return false;
958 }
959 if (UseCompressedOops && field_type->isa_narrowoop()) {
960 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
961 // to be able scalar replace the allocation.
962 if (field_val->is_EncodeP()) {
963 field_val = field_val->in(1);
964 } else {
965 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
966 }
967 }
968 sfpt->add_req(field_val);
969 }
970 JVMState *jvms = sfpt->jvms();
971 jvms->set_endoff(sfpt->req());
972 // Now make a pass over the debug information replacing any references
973 // to the allocated object with "sobj"
974 int start = jvms->debug_start();
975 int end = jvms->debug_end();
976 sfpt->replace_edges_in_range(res, sobj, start, end);
977 _igvn._worklist.push(sfpt);
978 safepoints_done.append_if_missing(sfpt); // keep it for rollback
979 }
980 return true;
981 }
982
983 // Process users of eliminated allocation.
984 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
985 Node* res = alloc->result_cast();
986 if (res != NULL) {
987 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
988 Node *use = res->last_out(j);
989 uint oc1 = res->outcnt();
990
991 if (use->is_AddP()) {
992 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
993 Node *n = use->last_out(k);
994 uint oc2 = use->outcnt();
995 if (n->is_Store()) {
996 #ifdef ASSERT
997 // Verify that there is no dependent MemBarVolatile nodes,
998 // they should be removed during IGVN, see MemBarNode::Ideal().
999 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1000 p < pmax; p++) {
1001 Node* mb = n->fast_out(p);
1002 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1003 mb->req() <= MemBarNode::Precedent ||
1004 mb->in(MemBarNode::Precedent) != n,
1005 "MemBarVolatile should be eliminated for non-escaping object");
1006 }
1007 #endif
1008 _igvn.replace_node(n, n->in(MemNode::Memory));
1009 } else if (n->is_ArrayCopy()) {
1010 // Disconnect ArrayCopy node
1011 ArrayCopyNode* ac = n->as_ArrayCopy();
1012 assert(ac->is_clonebasic(), "unexpected array copy kind");
1013 Node* ctl_proj = ac->proj_out(TypeFunc::Control);
1014 Node* mem_proj = ac->proj_out(TypeFunc::Memory);
1015 if (ctl_proj != NULL) {
1016 _igvn.replace_node(ctl_proj, n->in(0));
1017 }
1018 if (mem_proj != NULL) {
1019 _igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1020 }
1021 } else {
1022 eliminate_card_mark(n);
1023 }
1024 k -= (oc2 - use->outcnt());
1025 }
1026 } else if (use->is_ArrayCopy()) {
1027 // Disconnect ArrayCopy node
1028 ArrayCopyNode* ac = use->as_ArrayCopy();
1029 assert(ac->is_arraycopy_validated() ||
1030 ac->is_copyof_validated() ||
1031 ac->is_copyofrange_validated(), "unsupported");
1032 CallProjections callprojs;
1033 ac->extract_projections(&callprojs, true);
1034
1035 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1036 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1037 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1038
1039 // Set control to top. IGVN will remove the remaining projections
1040 ac->set_req(0, top());
1041 ac->replace_edge(res, top());
1042
1043 // Disconnect src right away: it can help find new
1044 // opportunities for allocation elimination
1045 Node* src = ac->in(ArrayCopyNode::Src);
1046 ac->replace_edge(src, top());
1047 if (src->outcnt() == 0) {
1048 _igvn.remove_dead_node(src);
1049 }
1050
1051 _igvn._worklist.push(ac);
1052 } else {
1053 eliminate_card_mark(use);
1054 }
1055 j -= (oc1 - res->outcnt());
1056 }
1057 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1058 _igvn.remove_dead_node(res);
1059 }
1060
1061 //
1062 // Process other users of allocation's projections
1063 //
1064 if (_resproj != NULL && _resproj->outcnt() != 0) {
1065 // First disconnect stores captured by Initialize node.
1066 // If Initialize node is eliminated first in the following code,
1067 // it will kill such stores and DUIterator_Last will assert.
1068 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1069 Node *use = _resproj->fast_out(j);
1070 if (use->is_AddP()) {
1071 // raw memory addresses used only by the initialization
1072 _igvn.replace_node(use, C->top());
1073 --j; --jmax;
1074 }
1075 }
1076 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1077 Node *use = _resproj->last_out(j);
1078 uint oc1 = _resproj->outcnt();
1079 if (use->is_Initialize()) {
1080 // Eliminate Initialize node.
1081 InitializeNode *init = use->as_Initialize();
1082 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1083 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
1084 if (ctrl_proj != NULL) {
1085 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1086 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1087 }
1088 Node *mem_proj = init->proj_out(TypeFunc::Memory);
1089 if (mem_proj != NULL) {
1090 Node *mem = init->in(TypeFunc::Memory);
1091 #ifdef ASSERT
1092 if (mem->is_MergeMem()) {
1093 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1094 } else {
1095 assert(mem == _memproj_fallthrough, "allocation memory projection");
1096 }
1097 #endif
1098 _igvn.replace_node(mem_proj, mem);
1099 }
1100 } else {
1101 assert(false, "only Initialize or AddP expected");
1102 }
1103 j -= (oc1 - _resproj->outcnt());
1104 }
1105 }
1106 if (_fallthroughcatchproj != NULL) {
1107 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1108 }
1109 if (_memproj_fallthrough != NULL) {
1110 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1111 }
1112 if (_memproj_catchall != NULL) {
1113 _igvn.replace_node(_memproj_catchall, C->top());
1114 }
1115 if (_ioproj_fallthrough != NULL) {
1116 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1117 }
1118 if (_ioproj_catchall != NULL) {
1119 _igvn.replace_node(_ioproj_catchall, C->top());
1120 }
1121 if (_catchallcatchproj != NULL) {
1122 _igvn.replace_node(_catchallcatchproj, C->top());
1123 }
1124 }
1125
1126 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1127 // Don't do scalar replacement if the frame can be popped by JVMTI:
1128 // if reallocation fails during deoptimization we'll pop all
1129 // interpreter frames for this compiled frame and that won't play
1130 // nice with JVMTI popframe.
1131 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1132 return false;
1133 }
1134 Node* klass = alloc->in(AllocateNode::KlassNode);
1135 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1136 Node* res = alloc->result_cast();
1137 // Eliminate boxing allocations which are not used
1138 // regardless scalar replacable status.
1139 bool boxing_alloc = C->eliminate_boxing() &&
1140 tklass->klass()->is_instance_klass() &&
1141 tklass->klass()->as_instance_klass()->is_box_klass();
1142 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1143 return false;
1144 }
1145
1146 extract_call_projections(alloc);
1147
1148 GrowableArray <SafePointNode *> safepoints;
1149 if (!can_eliminate_allocation(alloc, safepoints)) {
1150 return false;
1151 }
1152
1153 if (!alloc->_is_scalar_replaceable) {
1154 assert(res == NULL, "sanity");
1155 // We can only eliminate allocation if all debug info references
1156 // are already replaced with SafePointScalarObject because
1157 // we can't search for a fields value without instance_id.
1158 if (safepoints.length() > 0) {
1159 return false;
1160 }
1161 }
1162
1163 if (!scalar_replacement(alloc, safepoints)) {
1164 return false;
1165 }
1166
1167 CompileLog* log = C->log();
1168 if (log != NULL) {
1169 log->head("eliminate_allocation type='%d'",
1170 log->identify(tklass->klass()));
1171 JVMState* p = alloc->jvms();
1172 while (p != NULL) {
1173 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1174 p = p->caller();
1175 }
1176 log->tail("eliminate_allocation");
1177 }
1178
1179 process_users_of_allocation(alloc);
1180
1181 #ifndef PRODUCT
1182 if (PrintEliminateAllocations) {
1183 if (alloc->is_AllocateArray())
1184 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1185 else
1186 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1187 }
1188 #endif
1189
1190 return true;
1191 }
1192
1193 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1194 // EA should remove all uses of non-escaping boxing node.
1195 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1196 return false;
1197 }
1198
1199 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1200
1201 extract_call_projections(boxing);
1202
1203 const TypeTuple* r = boxing->tf()->range();
1204 assert(r->cnt() > TypeFunc::Parms, "sanity");
1205 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1206 assert(t != NULL, "sanity");
1207
1208 CompileLog* log = C->log();
1209 if (log != NULL) {
1210 log->head("eliminate_boxing type='%d'",
1211 log->identify(t->klass()));
1212 JVMState* p = boxing->jvms();
1213 while (p != NULL) {
1214 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1215 p = p->caller();
1216 }
1217 log->tail("eliminate_boxing");
1218 }
1219
1220 process_users_of_allocation(boxing);
1221
1222 #ifndef PRODUCT
1223 if (PrintEliminateAllocations) {
1224 tty->print("++++ Eliminated: %d ", boxing->_idx);
1225 boxing->method()->print_short_name(tty);
1226 tty->cr();
1227 }
1228 #endif
1229
1230 return true;
1231 }
1232
1233 //---------------------------set_eden_pointers-------------------------
1234 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1235 if (UseTLAB) { // Private allocation: load from TLS
1236 Node* thread = transform_later(new ThreadLocalNode());
1237 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1238 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1239 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1240 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1241 } else { // Shared allocation: load from globals
1242 CollectedHeap* ch = Universe::heap();
1243 address top_adr = (address)ch->top_addr();
1244 address end_adr = (address)ch->end_addr();
1245 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1246 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1247 }
1248 }
1249
1250
1251 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1252 Node* adr = basic_plus_adr(base, offset);
1253 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1254 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1255 transform_later(value);
1256 return value;
1257 }
1258
1259
1260 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1261 Node* adr = basic_plus_adr(base, offset);
1262 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1263 transform_later(mem);
1264 return mem;
1265 }
1266
1267 //=============================================================================
1268 //
1269 // A L L O C A T I O N
1270 //
1271 // Allocation attempts to be fast in the case of frequent small objects.
1272 // It breaks down like this:
1273 //
1274 // 1) Size in doublewords is computed. This is a constant for objects and
1275 // variable for most arrays. Doubleword units are used to avoid size
1276 // overflow of huge doubleword arrays. We need doublewords in the end for
1277 // rounding.
1278 //
1279 // 2) Size is checked for being 'too large'. Too-large allocations will go
1280 // the slow path into the VM. The slow path can throw any required
1281 // exceptions, and does all the special checks for very large arrays. The
1282 // size test can constant-fold away for objects. For objects with
1283 // finalizers it constant-folds the otherway: you always go slow with
1284 // finalizers.
1285 //
1286 // 3) If NOT using TLABs, this is the contended loop-back point.
1287 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1288 //
1289 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1290 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1291 // "size*8" we always enter the VM, where "largish" is a constant picked small
1292 // enough that there's always space between the eden max and 4Gig (old space is
1293 // there so it's quite large) and large enough that the cost of entering the VM
1294 // is dwarfed by the cost to initialize the space.
1295 //
1296 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1297 // down. If contended, repeat at step 3. If using TLABs normal-store
1298 // adjusted heap top back down; there is no contention.
1299 //
1300 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1301 // fields.
1302 //
1303 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1304 // oop flavor.
1305 //
1306 //=============================================================================
1307 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1308 // Allocations bigger than this always go the slow route.
1309 // This value must be small enough that allocation attempts that need to
1310 // trigger exceptions go the slow route. Also, it must be small enough so
1311 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1312 //=============================================================================j//
1313 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1314 // The allocator will coalesce int->oop copies away. See comment in
1315 // coalesce.cpp about how this works. It depends critically on the exact
1316 // code shape produced here, so if you are changing this code shape
1317 // make sure the GC info for the heap-top is correct in and around the
1318 // slow-path call.
1319 //
1320
1321 void PhaseMacroExpand::expand_allocate_common(
1322 AllocateNode* alloc, // allocation node to be expanded
1323 Node* length, // array length for an array allocation
1324 const TypeFunc* slow_call_type, // Type of slow call
1325 address slow_call_address // Address of slow call
1326 )
1327 {
1328
1329 Node* ctrl = alloc->in(TypeFunc::Control);
1330 Node* mem = alloc->in(TypeFunc::Memory);
1331 Node* i_o = alloc->in(TypeFunc::I_O);
1332 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1333 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1334 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1335
1336 assert(ctrl != NULL, "must have control");
1337 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1338 // they will not be used if "always_slow" is set
1339 enum { slow_result_path = 1, fast_result_path = 2 };
1340 Node *result_region = NULL;
1341 Node *result_phi_rawmem = NULL;
1342 Node *result_phi_rawoop = NULL;
1343 Node *result_phi_i_o = NULL;
1344
1345 // The initial slow comparison is a size check, the comparison
1346 // we want to do is a BoolTest::gt
1347 bool always_slow = false;
1348 int tv = _igvn.find_int_con(initial_slow_test, -1);
1349 if (tv >= 0) {
1350 always_slow = (tv == 1);
1351 initial_slow_test = NULL;
1352 } else {
1353 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1354 }
1355
1356 if (C->env()->dtrace_alloc_probes() ||
1357 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1358 // Force slow-path allocation
1359 always_slow = true;
1360 initial_slow_test = NULL;
1361 }
1362
1363
1364 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1365 Node *slow_region = NULL;
1366 Node *toobig_false = ctrl;
1367
1368 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1369 // generate the initial test if necessary
1370 if (initial_slow_test != NULL ) {
1371 slow_region = new RegionNode(3);
1372
1373 // Now make the initial failure test. Usually a too-big test but
1374 // might be a TRUE for finalizers or a fancy class check for
1375 // newInstance0.
1376 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1377 transform_later(toobig_iff);
1378 // Plug the failing-too-big test into the slow-path region
1379 Node *toobig_true = new IfTrueNode( toobig_iff );
1380 transform_later(toobig_true);
1381 slow_region ->init_req( too_big_or_final_path, toobig_true );
1382 toobig_false = new IfFalseNode( toobig_iff );
1383 transform_later(toobig_false);
1384 } else { // No initial test, just fall into next case
1385 toobig_false = ctrl;
1386 debug_only(slow_region = NodeSentinel);
1387 }
1388
1389 Node *slow_mem = mem; // save the current memory state for slow path
1390 // generate the fast allocation code unless we know that the initial test will always go slow
1391 if (!always_slow) {
1392 // Fast path modifies only raw memory.
1393 if (mem->is_MergeMem()) {
1394 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1395 }
1396
1397 Node* eden_top_adr;
1398 Node* eden_end_adr;
1399
1400 set_eden_pointers(eden_top_adr, eden_end_adr);
1401
1402 // Load Eden::end. Loop invariant and hoisted.
1403 //
1404 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1405 // a TLAB to work around a bug where these values were being moved across
1406 // a safepoint. These are not oops, so they cannot be include in the oop
1407 // map, but they can be changed by a GC. The proper way to fix this would
1408 // be to set the raw memory state when generating a SafepointNode. However
1409 // this will require extensive changes to the loop optimization in order to
1410 // prevent a degradation of the optimization.
1411 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1412 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1413
1414 // allocate the Region and Phi nodes for the result
1415 result_region = new RegionNode(3);
1416 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1417 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1418 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1419
1420 // We need a Region for the loop-back contended case.
1421 enum { fall_in_path = 1, contended_loopback_path = 2 };
1422 Node *contended_region;
1423 Node *contended_phi_rawmem;
1424 if (UseTLAB) {
1425 contended_region = toobig_false;
1426 contended_phi_rawmem = mem;
1427 } else {
1428 contended_region = new RegionNode(3);
1429 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1430 // Now handle the passing-too-big test. We fall into the contended
1431 // loop-back merge point.
1432 contended_region ->init_req(fall_in_path, toobig_false);
1433 contended_phi_rawmem->init_req(fall_in_path, mem);
1434 transform_later(contended_region);
1435 transform_later(contended_phi_rawmem);
1436 }
1437
1438 // Load(-locked) the heap top.
1439 // See note above concerning the control input when using a TLAB
1440 Node *old_eden_top = UseTLAB
1441 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1442 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1443
1444 transform_later(old_eden_top);
1445 // Add to heap top to get a new heap top
1446
1447 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1448 transform_later(new_eden_top);
1449 // Check for needing a GC; compare against heap end
1450 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1451 transform_later(needgc_cmp);
1452 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1453 transform_later(needgc_bol);
1454 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1455 transform_later(needgc_iff);
1456
1457 // Plug the failing-heap-space-need-gc test into the slow-path region
1458 Node *needgc_true = new IfTrueNode(needgc_iff);
1459 transform_later(needgc_true);
1460 if (initial_slow_test) {
1461 slow_region->init_req(need_gc_path, needgc_true);
1462 // This completes all paths into the slow merge point
1463 transform_later(slow_region);
1464 } else { // No initial slow path needed!
1465 // Just fall from the need-GC path straight into the VM call.
1466 slow_region = needgc_true;
1467 }
1468 // No need for a GC. Setup for the Store-Conditional
1469 Node *needgc_false = new IfFalseNode(needgc_iff);
1470 transform_later(needgc_false);
1471
1472 // Grab regular I/O before optional prefetch may change it.
1473 // Slow-path does no I/O so just set it to the original I/O.
1474 result_phi_i_o->init_req(slow_result_path, i_o);
1475
1476 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1477 old_eden_top, new_eden_top, length);
1478
1479 // Name successful fast-path variables
1480 Node* fast_oop = old_eden_top;
1481 Node* fast_oop_ctrl;
1482 Node* fast_oop_rawmem;
1483
1484 // Store (-conditional) the modified eden top back down.
1485 // StorePConditional produces flags for a test PLUS a modified raw
1486 // memory state.
1487 if (UseTLAB) {
1488 Node* store_eden_top =
1489 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1490 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1491 transform_later(store_eden_top);
1492 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1493 fast_oop_rawmem = store_eden_top;
1494 } else {
1495 Node* store_eden_top =
1496 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1497 new_eden_top, fast_oop/*old_eden_top*/);
1498 transform_later(store_eden_top);
1499 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1500 transform_later(contention_check);
1501 store_eden_top = new SCMemProjNode(store_eden_top);
1502 transform_later(store_eden_top);
1503
1504 // If not using TLABs, check to see if there was contention.
1505 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1506 transform_later(contention_iff);
1507 Node *contention_true = new IfTrueNode(contention_iff);
1508 transform_later(contention_true);
1509 // If contention, loopback and try again.
1510 contended_region->init_req(contended_loopback_path, contention_true);
1511 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1512
1513 // Fast-path succeeded with no contention!
1514 Node *contention_false = new IfFalseNode(contention_iff);
1515 transform_later(contention_false);
1516 fast_oop_ctrl = contention_false;
1517
1518 // Bump total allocated bytes for this thread
1519 Node* thread = new ThreadLocalNode();
1520 transform_later(thread);
1521 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1522 in_bytes(JavaThread::allocated_bytes_offset()));
1523 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1524 0, TypeLong::LONG, T_LONG);
1525 #ifdef _LP64
1526 Node* alloc_size = size_in_bytes;
1527 #else
1528 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1529 transform_later(alloc_size);
1530 #endif
1531 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1532 transform_later(new_alloc_bytes);
1533 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1534 0, new_alloc_bytes, T_LONG);
1535 }
1536
1537 InitializeNode* init = alloc->initialization();
1538 fast_oop_rawmem = initialize_object(alloc,
1539 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1540 klass_node, length, size_in_bytes);
1541
1542 // If initialization is performed by an array copy, any required
1543 // MemBarStoreStore was already added. If the object does not
1544 // escape no need for a MemBarStoreStore. If the object does not
1545 // escape in its initializer and memory barrier (MemBarStoreStore or
1546 // stronger) is already added at exit of initializer, also no need
1547 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1548 // so that stores that initialize this object can't be reordered
1549 // with a subsequent store that makes this object accessible by
1550 // other threads.
1551 // Other threads include java threads and JVM internal threads
1552 // (for example concurrent GC threads). Current concurrent GC
1553 // implementation: CMS and G1 will not scan newly created object,
1554 // so it's safe to skip storestore barrier when allocation does
1555 // not escape.
1556 if (!alloc->does_not_escape_thread() &&
1557 !alloc->is_allocation_MemBar_redundant() &&
1558 (init == NULL || !init->is_complete_with_arraycopy())) {
1559 if (init == NULL || init->req() < InitializeNode::RawStores) {
1560 // No InitializeNode or no stores captured by zeroing
1561 // elimination. Simply add the MemBarStoreStore after object
1562 // initialization.
1563 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1564 transform_later(mb);
1565
1566 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1567 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1568 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1569 transform_later(fast_oop_ctrl);
1570 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1571 transform_later(fast_oop_rawmem);
1572 } else {
1573 // Add the MemBarStoreStore after the InitializeNode so that
1574 // all stores performing the initialization that were moved
1575 // before the InitializeNode happen before the storestore
1576 // barrier.
1577
1578 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1579 Node* init_mem = init->proj_out(TypeFunc::Memory);
1580
1581 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1582 transform_later(mb);
1583
1584 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1585 transform_later(ctrl);
1586 Node* mem = new ProjNode(init,TypeFunc::Memory);
1587 transform_later(mem);
1588
1589 // The MemBarStoreStore depends on control and memory coming
1590 // from the InitializeNode
1591 mb->init_req(TypeFunc::Memory, mem);
1592 mb->init_req(TypeFunc::Control, ctrl);
1593
1594 ctrl = new ProjNode(mb,TypeFunc::Control);
1595 transform_later(ctrl);
1596 mem = new ProjNode(mb,TypeFunc::Memory);
1597 transform_later(mem);
1598
1599 // All nodes that depended on the InitializeNode for control
1600 // and memory must now depend on the MemBarNode that itself
1601 // depends on the InitializeNode
1602 if (init_ctrl != NULL) {
1603 _igvn.replace_node(init_ctrl, ctrl);
1604 }
1605 if (init_mem != NULL) {
1606 _igvn.replace_node(init_mem, mem);
1607 }
1608 }
1609 }
1610
1611 if (C->env()->dtrace_extended_probes()) {
1612 // Slow-path call
1613 int size = TypeFunc::Parms + 2;
1614 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1615 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1616 "dtrace_object_alloc",
1617 TypeRawPtr::BOTTOM);
1618
1619 // Get base of thread-local storage area
1620 Node* thread = new ThreadLocalNode();
1621 transform_later(thread);
1622
1623 call->init_req(TypeFunc::Parms+0, thread);
1624 call->init_req(TypeFunc::Parms+1, fast_oop);
1625 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1626 call->init_req(TypeFunc::I_O , top()); // does no i/o
1627 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1628 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1629 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1630 transform_later(call);
1631 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1632 transform_later(fast_oop_ctrl);
1633 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1634 transform_later(fast_oop_rawmem);
1635 }
1636
1637 // Plug in the successful fast-path into the result merge point
1638 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1639 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1640 result_phi_i_o ->init_req(fast_result_path, i_o);
1641 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1642 } else {
1643 slow_region = ctrl;
1644 result_phi_i_o = i_o; // Rename it to use in the following code.
1645 }
1646
1647 // Generate slow-path call
1648 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1649 OptoRuntime::stub_name(slow_call_address),
1650 alloc->jvms()->bci(),
1651 TypePtr::BOTTOM);
1652 call->init_req( TypeFunc::Control, slow_region );
1653 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1654 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1655 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1656 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1657
1658 call->init_req(TypeFunc::Parms+0, klass_node);
1659 if (length != NULL) {
1660 call->init_req(TypeFunc::Parms+1, length);
1661 }
1662
1663 // Copy debug information and adjust JVMState information, then replace
1664 // allocate node with the call
1665 copy_call_debug_info((CallNode *) alloc, call);
1666 if (!always_slow) {
1667 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1668 } else {
1669 // Hook i_o projection to avoid its elimination during allocation
1670 // replacement (when only a slow call is generated).
1671 call->set_req(TypeFunc::I_O, result_phi_i_o);
1672 }
1673 _igvn.replace_node(alloc, call);
1674 transform_later(call);
1675
1676 // Identify the output projections from the allocate node and
1677 // adjust any references to them.
1678 // The control and io projections look like:
1679 //
1680 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1681 // Allocate Catch
1682 // ^---Proj(io) <-------+ ^---CatchProj(io)
1683 //
1684 // We are interested in the CatchProj nodes.
1685 //
1686 extract_call_projections(call);
1687
1688 // An allocate node has separate memory projections for the uses on
1689 // the control and i_o paths. Replace the control memory projection with
1690 // result_phi_rawmem (unless we are only generating a slow call when
1691 // both memory projections are combined)
1692 if (!always_slow && _memproj_fallthrough != NULL) {
1693 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1694 Node *use = _memproj_fallthrough->fast_out(i);
1695 _igvn.rehash_node_delayed(use);
1696 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1697 // back up iterator
1698 --i;
1699 }
1700 }
1701 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1702 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1703 if (_memproj_catchall != NULL ) {
1704 if (_memproj_fallthrough == NULL) {
1705 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1706 transform_later(_memproj_fallthrough);
1707 }
1708 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1709 Node *use = _memproj_catchall->fast_out(i);
1710 _igvn.rehash_node_delayed(use);
1711 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1712 // back up iterator
1713 --i;
1714 }
1715 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1716 _igvn.remove_dead_node(_memproj_catchall);
1717 }
1718
1719 // An allocate node has separate i_o projections for the uses on the control
1720 // and i_o paths. Always replace the control i_o projection with result i_o
1721 // otherwise incoming i_o become dead when only a slow call is generated
1722 // (it is different from memory projections where both projections are
1723 // combined in such case).
1724 if (_ioproj_fallthrough != NULL) {
1725 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1726 Node *use = _ioproj_fallthrough->fast_out(i);
1727 _igvn.rehash_node_delayed(use);
1728 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1729 // back up iterator
1730 --i;
1731 }
1732 }
1733 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1734 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1735 if (_ioproj_catchall != NULL ) {
1736 if (_ioproj_fallthrough == NULL) {
1737 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1738 transform_later(_ioproj_fallthrough);
1739 }
1740 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1741 Node *use = _ioproj_catchall->fast_out(i);
1742 _igvn.rehash_node_delayed(use);
1743 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1744 // back up iterator
1745 --i;
1746 }
1747 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1748 _igvn.remove_dead_node(_ioproj_catchall);
1749 }
1750
1751 // if we generated only a slow call, we are done
1752 if (always_slow) {
1753 // Now we can unhook i_o.
1754 if (result_phi_i_o->outcnt() > 1) {
1755 call->set_req(TypeFunc::I_O, top());
1756 } else {
1757 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1758 // Case of new array with negative size known during compilation.
1759 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1760 // following code since call to runtime will throw exception.
1761 // As result there will be no users of i_o after the call.
1762 // Leave i_o attached to this call to avoid problems in preceding graph.
1763 }
1764 return;
1765 }
1766
1767
1768 if (_fallthroughcatchproj != NULL) {
1769 ctrl = _fallthroughcatchproj->clone();
1770 transform_later(ctrl);
1771 _igvn.replace_node(_fallthroughcatchproj, result_region);
1772 } else {
1773 ctrl = top();
1774 }
1775 Node *slow_result;
1776 if (_resproj == NULL) {
1777 // no uses of the allocation result
1778 slow_result = top();
1779 } else {
1780 slow_result = _resproj->clone();
1781 transform_later(slow_result);
1782 _igvn.replace_node(_resproj, result_phi_rawoop);
1783 }
1784
1785 // Plug slow-path into result merge point
1786 result_region ->init_req( slow_result_path, ctrl );
1787 result_phi_rawoop->init_req( slow_result_path, slow_result);
1788 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1789 transform_later(result_region);
1790 transform_later(result_phi_rawoop);
1791 transform_later(result_phi_rawmem);
1792 transform_later(result_phi_i_o);
1793 // This completes all paths into the result merge point
1794 }
1795
1796
1797 // Helper for PhaseMacroExpand::expand_allocate_common.
1798 // Initializes the newly-allocated storage.
1799 Node*
1800 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1801 Node* control, Node* rawmem, Node* object,
1802 Node* klass_node, Node* length,
1803 Node* size_in_bytes) {
1804 InitializeNode* init = alloc->initialization();
1805 // Store the klass & mark bits
1806 Node* mark_node = NULL;
1807 // For now only enable fast locking for non-array types
1808 if (UseBiasedLocking && (length == NULL)) {
1809 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1810 } else {
1811 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1812 }
1813 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1814
1815 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1816 int header_size = alloc->minimum_header_size(); // conservatively small
1817
1818 // Array length
1819 if (length != NULL) { // Arrays need length field
1820 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1821 // conservatively small header size:
1822 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1823 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1824 if (k->is_array_klass()) // we know the exact header size in most cases:
1825 header_size = Klass::layout_helper_header_size(k->layout_helper());
1826 }
1827
1828 if (UseShenandoahGC) {
1829 // Initialize Shenandoah brooks pointer to point to the object itself.
1830 rawmem = make_store(control, rawmem, object, BrooksPointer::byte_offset(), object, T_OBJECT);
1831 }
1832
1833 // Clear the object body, if necessary.
1834 if (init == NULL) {
1835 // The init has somehow disappeared; be cautious and clear everything.
1836 //
1837 // This can happen if a node is allocated but an uncommon trap occurs
1838 // immediately. In this case, the Initialize gets associated with the
1839 // trap, and may be placed in a different (outer) loop, if the Allocate
1840 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1841 // there can be two Allocates to one Initialize. The answer in all these
1842 // edge cases is safety first. It is always safe to clear immediately
1843 // within an Allocate, and then (maybe or maybe not) clear some more later.
1844 if (!(UseTLAB && ZeroTLAB)) {
1845 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1846 header_size, size_in_bytes,
1847 &_igvn);
1848 }
1849 } else {
1850 if (!init->is_complete()) {
1851 // Try to win by zeroing only what the init does not store.
1852 // We can also try to do some peephole optimizations,
1853 // such as combining some adjacent subword stores.
1854 rawmem = init->complete_stores(control, rawmem, object,
1855 header_size, size_in_bytes, &_igvn);
1856 }
1857 // We have no more use for this link, since the AllocateNode goes away:
1858 init->set_req(InitializeNode::RawAddress, top());
1859 // (If we keep the link, it just confuses the register allocator,
1860 // who thinks he sees a real use of the address by the membar.)
1861 }
1862
1863 return rawmem;
1864 }
1865
1866 // Generate prefetch instructions for next allocations.
1867 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1868 Node*& contended_phi_rawmem,
1869 Node* old_eden_top, Node* new_eden_top,
1870 Node* length) {
1871 enum { fall_in_path = 1, pf_path = 2 };
1872 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1873 // Generate prefetch allocation with watermark check.
1874 // As an allocation hits the watermark, we will prefetch starting
1875 // at a "distance" away from watermark.
1876
1877 Node *pf_region = new RegionNode(3);
1878 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1879 TypeRawPtr::BOTTOM );
1880 // I/O is used for Prefetch
1881 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1882
1883 Node *thread = new ThreadLocalNode();
1884 transform_later(thread);
1885
1886 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1887 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1888 transform_later(eden_pf_adr);
1889
1890 Node *old_pf_wm = new LoadPNode(needgc_false,
1891 contended_phi_rawmem, eden_pf_adr,
1892 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1893 MemNode::unordered);
1894 transform_later(old_pf_wm);
1895
1896 // check against new_eden_top
1897 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1898 transform_later(need_pf_cmp);
1899 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1900 transform_later(need_pf_bol);
1901 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1902 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1903 transform_later(need_pf_iff);
1904
1905 // true node, add prefetchdistance
1906 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1907 transform_later(need_pf_true);
1908
1909 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1910 transform_later(need_pf_false);
1911
1912 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1913 _igvn.MakeConX(AllocatePrefetchDistance) );
1914 transform_later(new_pf_wmt );
1915 new_pf_wmt->set_req(0, need_pf_true);
1916
1917 Node *store_new_wmt = new StorePNode(need_pf_true,
1918 contended_phi_rawmem, eden_pf_adr,
1919 TypeRawPtr::BOTTOM, new_pf_wmt,
1920 MemNode::unordered);
1921 transform_later(store_new_wmt);
1922
1923 // adding prefetches
1924 pf_phi_abio->init_req( fall_in_path, i_o );
1925
1926 Node *prefetch_adr;
1927 Node *prefetch;
1928 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1929 uint step_size = AllocatePrefetchStepSize;
1930 uint distance = 0;
1931
1932 for ( uint i = 0; i < lines; i++ ) {
1933 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1934 _igvn.MakeConX(distance) );
1935 transform_later(prefetch_adr);
1936 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1937 transform_later(prefetch);
1938 distance += step_size;
1939 i_o = prefetch;
1940 }
1941 pf_phi_abio->set_req( pf_path, i_o );
1942
1943 pf_region->init_req( fall_in_path, need_pf_false );
1944 pf_region->init_req( pf_path, need_pf_true );
1945
1946 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1947 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1948
1949 transform_later(pf_region);
1950 transform_later(pf_phi_rawmem);
1951 transform_later(pf_phi_abio);
1952
1953 needgc_false = pf_region;
1954 contended_phi_rawmem = pf_phi_rawmem;
1955 i_o = pf_phi_abio;
1956 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1957 // Insert a prefetch instruction for each allocation.
1958 // This code is used for SPARC with BIS.
1959
1960 // Generate several prefetch instructions.
1961 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1962 uint step_size = AllocatePrefetchStepSize;
1963 uint distance = AllocatePrefetchDistance;
1964
1965 // Next cache address.
1966 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1967 _igvn.MakeConX(step_size + distance));
1968 transform_later(cache_adr);
1969 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1970 transform_later(cache_adr);
1971 // For BIS instructions to be emitted, the address must be aligned at cache line size.
1972 // (The VM sets AllocatePrefetchStepSize to the cache line size, unless a value is
1973 // specified at the command line.) If the address is not aligned at cache line size
1974 // boundary, a standard store instruction is triggered (instead of the BIS). For the
1975 // latter, 8-byte alignment is necessary.
1976 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1977 cache_adr = new AndXNode(cache_adr, mask);
1978 transform_later(cache_adr);
1979 cache_adr = new CastX2PNode(cache_adr);
1980 transform_later(cache_adr);
1981
1982 // Prefetch
1983 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1984 prefetch->set_req(0, needgc_false);
1985 transform_later(prefetch);
1986 contended_phi_rawmem = prefetch;
1987 Node *prefetch_adr;
1988 distance = step_size;
1989 for ( uint i = 1; i < lines; i++ ) {
1990 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1991 _igvn.MakeConX(distance) );
1992 transform_later(prefetch_adr);
1993 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1994 transform_later(prefetch);
1995 distance += step_size;
1996 contended_phi_rawmem = prefetch;
1997 }
1998 } else if( AllocatePrefetchStyle > 0 ) {
1999 // Insert a prefetch for each allocation only on the fast-path
2000 Node *prefetch_adr;
2001 Node *prefetch;
2002 // Generate several prefetch instructions.
2003 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2004 uint step_size = AllocatePrefetchStepSize;
2005 uint distance = AllocatePrefetchDistance;
2006 for ( uint i = 0; i < lines; i++ ) {
2007 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2008 _igvn.MakeConX(distance) );
2009 transform_later(prefetch_adr);
2010 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2011 // Do not let it float too high, since if eden_top == eden_end,
2012 // both might be null.
2013 if( i == 0 ) { // Set control for first prefetch, next follows it
2014 prefetch->init_req(0, needgc_false);
2015 }
2016 transform_later(prefetch);
2017 distance += step_size;
2018 i_o = prefetch;
2019 }
2020 }
2021 return i_o;
2022 }
2023
2024
2025 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2026 expand_allocate_common(alloc, NULL,
2027 OptoRuntime::new_instance_Type(),
2028 OptoRuntime::new_instance_Java());
2029 }
2030
2031 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2032 Node* length = alloc->in(AllocateNode::ALength);
2033 InitializeNode* init = alloc->initialization();
2034 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2035 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2036 address slow_call_address; // Address of slow call
2037 if (init != NULL && init->is_complete_with_arraycopy() &&
2038 k->is_type_array_klass()) {
2039 // Don't zero type array during slow allocation in VM since
2040 // it will be initialized later by arraycopy in compiled code.
2041 slow_call_address = OptoRuntime::new_array_nozero_Java();
2042 } else {
2043 slow_call_address = OptoRuntime::new_array_Java();
2044 }
2045 expand_allocate_common(alloc, length,
2046 OptoRuntime::new_array_Type(),
2047 slow_call_address);
2048 }
2049
2050 //-------------------mark_eliminated_box----------------------------------
2051 //
2052 // During EA obj may point to several objects but after few ideal graph
2053 // transformations (CCP) it may point to only one non escaping object
2054 // (but still using phi), corresponding locks and unlocks will be marked
2055 // for elimination. Later obj could be replaced with a new node (new phi)
2056 // and which does not have escape information. And later after some graph
2057 // reshape other locks and unlocks (which were not marked for elimination
2058 // before) are connected to this new obj (phi) but they still will not be
2059 // marked for elimination since new obj has no escape information.
2060 // Mark all associated (same box and obj) lock and unlock nodes for
2061 // elimination if some of them marked already.
2062 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2063 if (oldbox->as_BoxLock()->is_eliminated())
2064 return; // This BoxLock node was processed already.
2065
2066 // New implementation (EliminateNestedLocks) has separate BoxLock
2067 // node for each locked region so mark all associated locks/unlocks as
2068 // eliminated even if different objects are referenced in one locked region
2069 // (for example, OSR compilation of nested loop inside locked scope).
2070 if (EliminateNestedLocks ||
2071 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2072 // Box is used only in one lock region. Mark this box as eliminated.
2073 _igvn.hash_delete(oldbox);
2074 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2075 _igvn.hash_insert(oldbox);
2076
2077 for (uint i = 0; i < oldbox->outcnt(); i++) {
2078 Node* u = oldbox->raw_out(i);
2079 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2080 AbstractLockNode* alock = u->as_AbstractLock();
2081 // Check lock's box since box could be referenced by Lock's debug info.
2082 if (alock->box_node() == oldbox) {
2083 // Mark eliminated all related locks and unlocks.
2084 #ifdef ASSERT
2085 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2086 #endif
2087 alock->set_non_esc_obj();
2088 }
2089 }
2090 }
2091 return;
2092 }
2093
2094 // Create new "eliminated" BoxLock node and use it in monitor debug info
2095 // instead of oldbox for the same object.
2096 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2097
2098 // Note: BoxLock node is marked eliminated only here and it is used
2099 // to indicate that all associated lock and unlock nodes are marked
2100 // for elimination.
2101 newbox->set_eliminated();
2102 transform_later(newbox);
2103
2104 // Replace old box node with new box for all users of the same object.
2105 for (uint i = 0; i < oldbox->outcnt();) {
2106 bool next_edge = true;
2107
2108 Node* u = oldbox->raw_out(i);
2109 if (u->is_AbstractLock()) {
2110 AbstractLockNode* alock = u->as_AbstractLock();
2111 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2112 // Replace Box and mark eliminated all related locks and unlocks.
2113 #ifdef ASSERT
2114 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2115 #endif
2116 alock->set_non_esc_obj();
2117 _igvn.rehash_node_delayed(alock);
2118 alock->set_box_node(newbox);
2119 next_edge = false;
2120 }
2121 }
2122 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2123 FastLockNode* flock = u->as_FastLock();
2124 assert(flock->box_node() == oldbox, "sanity");
2125 _igvn.rehash_node_delayed(flock);
2126 flock->set_box_node(newbox);
2127 next_edge = false;
2128 }
2129
2130 // Replace old box in monitor debug info.
2131 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2132 SafePointNode* sfn = u->as_SafePoint();
2133 JVMState* youngest_jvms = sfn->jvms();
2134 int max_depth = youngest_jvms->depth();
2135 for (int depth = 1; depth <= max_depth; depth++) {
2136 JVMState* jvms = youngest_jvms->of_depth(depth);
2137 int num_mon = jvms->nof_monitors();
2138 // Loop over monitors
2139 for (int idx = 0; idx < num_mon; idx++) {
2140 Node* obj_node = sfn->monitor_obj(jvms, idx);
2141 Node* box_node = sfn->monitor_box(jvms, idx);
2142 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2143 int j = jvms->monitor_box_offset(idx);
2144 _igvn.replace_input_of(u, j, newbox);
2145 next_edge = false;
2146 }
2147 }
2148 }
2149 }
2150 if (next_edge) i++;
2151 }
2152 }
2153
2154 //-----------------------mark_eliminated_locking_nodes-----------------------
2155 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2156 if (EliminateNestedLocks) {
2157 if (alock->is_nested()) {
2158 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2159 return;
2160 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2161 // Only Lock node has JVMState needed here.
2162 // Not that preceding claim is documented anywhere else.
2163 if (alock->jvms() != NULL) {
2164 if (alock->as_Lock()->is_nested_lock_region()) {
2165 // Mark eliminated related nested locks and unlocks.
2166 Node* obj = alock->obj_node();
2167 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2168 assert(!box_node->is_eliminated(), "should not be marked yet");
2169 // Note: BoxLock node is marked eliminated only here
2170 // and it is used to indicate that all associated lock
2171 // and unlock nodes are marked for elimination.
2172 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2173 for (uint i = 0; i < box_node->outcnt(); i++) {
2174 Node* u = box_node->raw_out(i);
2175 if (u->is_AbstractLock()) {
2176 alock = u->as_AbstractLock();
2177 if (alock->box_node() == box_node) {
2178 // Verify that this Box is referenced only by related locks.
2179 assert(alock->obj_node()->eqv_uncast(obj), "");
2180 // Mark all related locks and unlocks.
2181 #ifdef ASSERT
2182 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2183 #endif
2184 alock->set_nested();
2185 }
2186 }
2187 }
2188 } else {
2189 #ifdef ASSERT
2190 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2191 if (C->log() != NULL)
2192 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2193 #endif
2194 }
2195 }
2196 return;
2197 }
2198 // Process locks for non escaping object
2199 assert(alock->is_non_esc_obj(), "");
2200 } // EliminateNestedLocks
2201
2202 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2203 // Look for all locks of this object and mark them and
2204 // corresponding BoxLock nodes as eliminated.
2205 Node* obj = alock->obj_node();
2206 for (uint j = 0; j < obj->outcnt(); j++) {
2207 Node* o = obj->raw_out(j);
2208 if (o->is_AbstractLock() &&
2209 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2210 alock = o->as_AbstractLock();
2211 Node* box = alock->box_node();
2212 // Replace old box node with new eliminated box for all users
2213 // of the same object and mark related locks as eliminated.
2214 mark_eliminated_box(box, obj);
2215 }
2216 }
2217 }
2218 }
2219
2220 // we have determined that this lock/unlock can be eliminated, we simply
2221 // eliminate the node without expanding it.
2222 //
2223 // Note: The membar's associated with the lock/unlock are currently not
2224 // eliminated. This should be investigated as a future enhancement.
2225 //
2226 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2227
2228 if (!alock->is_eliminated()) {
2229 return false;
2230 }
2231 #ifdef ASSERT
2232 if (!alock->is_coarsened()) {
2233 // Check that new "eliminated" BoxLock node is created.
2234 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2235 assert(oldbox->is_eliminated(), "should be done already");
2236 }
2237 #endif
2238
2239 alock->log_lock_optimization(C, "eliminate_lock");
2240
2241 #ifndef PRODUCT
2242 if (PrintEliminateLocks) {
2243 if (alock->is_Lock()) {
2244 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2245 } else {
2246 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2247 }
2248 }
2249 #endif
2250
2251 Node* mem = alock->in(TypeFunc::Memory);
2252 Node* ctrl = alock->in(TypeFunc::Control);
2253
2254 extract_call_projections(alock);
2255 // There are 2 projections from the lock. The lock node will
2256 // be deleted when its last use is subsumed below.
2257 assert(alock->outcnt() == 2 &&
2258 _fallthroughproj != NULL &&
2259 _memproj_fallthrough != NULL,
2260 "Unexpected projections from Lock/Unlock");
2261
2262 Node* fallthroughproj = _fallthroughproj;
2263 Node* memproj_fallthrough = _memproj_fallthrough;
2264
2265 // The memory projection from a lock/unlock is RawMem
2266 // The input to a Lock is merged memory, so extract its RawMem input
2267 // (unless the MergeMem has been optimized away.)
2268 if (alock->is_Lock()) {
2269 // Seach for MemBarAcquireLock node and delete it also.
2270 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2271 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2272 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2273 Node* memproj = membar->proj_out(TypeFunc::Memory);
2274 _igvn.replace_node(ctrlproj, fallthroughproj);
2275 _igvn.replace_node(memproj, memproj_fallthrough);
2276
2277 // Delete FastLock node also if this Lock node is unique user
2278 // (a loop peeling may clone a Lock node).
2279 Node* flock = alock->as_Lock()->fastlock_node();
2280 if (flock->outcnt() == 1) {
2281 assert(flock->unique_out() == alock, "sanity");
2282 _igvn.replace_node(flock, top());
2283 }
2284 }
2285
2286 // Seach for MemBarReleaseLock node and delete it also.
2287 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2288 ctrl->in(0)->is_MemBar()) {
2289 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2290 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2291 mem->is_Proj() && membar == mem->in(0), "");
2292 _igvn.replace_node(fallthroughproj, ctrl);
2293 _igvn.replace_node(memproj_fallthrough, mem);
2294 fallthroughproj = ctrl;
2295 memproj_fallthrough = mem;
2296 ctrl = membar->in(TypeFunc::Control);
2297 mem = membar->in(TypeFunc::Memory);
2298 }
2299
2300 _igvn.replace_node(fallthroughproj, ctrl);
2301 _igvn.replace_node(memproj_fallthrough, mem);
2302 return true;
2303 }
2304
2305
2306 //------------------------------expand_lock_node----------------------
2307 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2308
2309 Node* ctrl = lock->in(TypeFunc::Control);
2310 Node* mem = lock->in(TypeFunc::Memory);
2311 Node* obj = lock->obj_node();
2312 Node* box = lock->box_node();
2313 Node* flock = lock->fastlock_node();
2314
2315 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2316
2317 // Make the merge point
2318 Node *region;
2319 Node *mem_phi;
2320 Node *slow_path;
2321
2322 if (UseOptoBiasInlining) {
2323 /*
2324 * See the full description in MacroAssembler::biased_locking_enter().
2325 *
2326 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2327 * // The object is biased.
2328 * proto_node = klass->prototype_header;
2329 * o_node = thread | proto_node;
2330 * x_node = o_node ^ mark_word;
2331 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2332 * // Done.
2333 * } else {
2334 * if( (x_node & biased_lock_mask) != 0 ) {
2335 * // The klass's prototype header is no longer biased.
2336 * cas(&mark_word, mark_word, proto_node)
2337 * goto cas_lock;
2338 * } else {
2339 * // The klass's prototype header is still biased.
2340 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2341 * old = mark_word;
2342 * new = o_node;
2343 * } else {
2344 * // Different thread or anonymous biased.
2345 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2346 * new = thread | old;
2347 * }
2348 * // Try to rebias.
2349 * if( cas(&mark_word, old, new) == 0 ) {
2350 * // Done.
2351 * } else {
2352 * goto slow_path; // Failed.
2353 * }
2354 * }
2355 * }
2356 * } else {
2357 * // The object is not biased.
2358 * cas_lock:
2359 * if( FastLock(obj) == 0 ) {
2360 * // Done.
2361 * } else {
2362 * slow_path:
2363 * OptoRuntime::complete_monitor_locking_Java(obj);
2364 * }
2365 * }
2366 */
2367
2368 region = new RegionNode(5);
2369 // create a Phi for the memory state
2370 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2371
2372 Node* fast_lock_region = new RegionNode(3);
2373 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2374
2375 // First, check mark word for the biased lock pattern.
2376 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2377
2378 // Get fast path - mark word has the biased lock pattern.
2379 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2380 markOopDesc::biased_lock_mask_in_place,
2381 markOopDesc::biased_lock_pattern, true);
2382 // fast_lock_region->in(1) is set to slow path.
2383 fast_lock_mem_phi->init_req(1, mem);
2384
2385 // Now check that the lock is biased to the current thread and has
2386 // the same epoch and bias as Klass::_prototype_header.
2387
2388 // Special-case a fresh allocation to avoid building nodes:
2389 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2390 if (klass_node == NULL) {
2391 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2392 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2393 #ifdef _LP64
2394 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2395 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2396 klass_node->in(1)->init_req(0, ctrl);
2397 } else
2398 #endif
2399 klass_node->init_req(0, ctrl);
2400 }
2401 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2402
2403 Node* thread = transform_later(new ThreadLocalNode());
2404 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2405 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2406 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2407
2408 // Get slow path - mark word does NOT match the value.
2409 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2410 (~markOopDesc::age_mask_in_place), 0);
2411 // region->in(3) is set to fast path - the object is biased to the current thread.
2412 mem_phi->init_req(3, mem);
2413
2414
2415 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2416
2417
2418 // First, check biased pattern.
2419 // Get fast path - _prototype_header has the same biased lock pattern.
2420 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2421 markOopDesc::biased_lock_mask_in_place, 0, true);
2422
2423 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2424 // fast_lock_region->in(2) - the prototype header is no longer biased
2425 // and we have to revoke the bias on this object.
2426 // We are going to try to reset the mark of this object to the prototype
2427 // value and fall through to the CAS-based locking scheme.
2428 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2429 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2430 proto_node, mark_node);
2431 transform_later(cas);
2432 Node* proj = transform_later(new SCMemProjNode(cas));
2433 fast_lock_mem_phi->init_req(2, proj);
2434
2435
2436 // Second, check epoch bits.
2437 Node* rebiased_region = new RegionNode(3);
2438 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2439 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2440
2441 // Get slow path - mark word does NOT match epoch bits.
2442 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2443 markOopDesc::epoch_mask_in_place, 0);
2444 // The epoch of the current bias is not valid, attempt to rebias the object
2445 // toward the current thread.
2446 rebiased_region->init_req(2, epoch_ctrl);
2447 old_phi->init_req(2, mark_node);
2448 new_phi->init_req(2, o_node);
2449
2450 // rebiased_region->in(1) is set to fast path.
2451 // The epoch of the current bias is still valid but we know
2452 // nothing about the owner; it might be set or it might be clear.
2453 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2454 markOopDesc::age_mask_in_place |
2455 markOopDesc::epoch_mask_in_place);
2456 Node* old = transform_later(new AndXNode(mark_node, cmask));
2457 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2458 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2459 old_phi->init_req(1, old);
2460 new_phi->init_req(1, new_mark);
2461
2462 transform_later(rebiased_region);
2463 transform_later(old_phi);
2464 transform_later(new_phi);
2465
2466 // Try to acquire the bias of the object using an atomic operation.
2467 // If this fails we will go in to the runtime to revoke the object's bias.
2468 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2469 transform_later(cas);
2470 proj = transform_later(new SCMemProjNode(cas));
2471
2472 // Get slow path - Failed to CAS.
2473 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2474 mem_phi->init_req(4, proj);
2475 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2476
2477 // Failed to CAS.
2478 slow_path = new RegionNode(3);
2479 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2480
2481 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2482 slow_mem->init_req(1, proj);
2483
2484 // Call CAS-based locking scheme (FastLock node).
2485
2486 transform_later(fast_lock_region);
2487 transform_later(fast_lock_mem_phi);
2488
2489 // Get slow path - FastLock failed to lock the object.
2490 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2491 mem_phi->init_req(2, fast_lock_mem_phi);
2492 // region->in(2) is set to fast path - the object is locked to the current thread.
2493
2494 slow_path->init_req(2, ctrl); // Capture slow-control
2495 slow_mem->init_req(2, fast_lock_mem_phi);
2496
2497 transform_later(slow_path);
2498 transform_later(slow_mem);
2499 // Reset lock's memory edge.
2500 lock->set_req(TypeFunc::Memory, slow_mem);
2501
2502 } else {
2503 region = new RegionNode(3);
2504 // create a Phi for the memory state
2505 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2506
2507 // Optimize test; set region slot 2
2508 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2509 mem_phi->init_req(2, mem);
2510 }
2511
2512 // Make slow path call
2513 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2514 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2515 obj, box, NULL);
2516
2517 extract_call_projections(call);
2518
2519 // Slow path can only throw asynchronous exceptions, which are always
2520 // de-opted. So the compiler thinks the slow-call can never throw an
2521 // exception. If it DOES throw an exception we would need the debug
2522 // info removed first (since if it throws there is no monitor).
2523 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2524 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2525
2526 // Capture slow path
2527 // disconnect fall-through projection from call and create a new one
2528 // hook up users of fall-through projection to region
2529 Node *slow_ctrl = _fallthroughproj->clone();
2530 transform_later(slow_ctrl);
2531 _igvn.hash_delete(_fallthroughproj);
2532 _fallthroughproj->disconnect_inputs(NULL, C);
2533 region->init_req(1, slow_ctrl);
2534 // region inputs are now complete
2535 transform_later(region);
2536 _igvn.replace_node(_fallthroughproj, region);
2537
2538 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2539 mem_phi->init_req(1, memproj );
2540 transform_later(mem_phi);
2541 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2542 }
2543
2544 //------------------------------expand_unlock_node----------------------
2545 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2546
2547 Node* ctrl = unlock->in(TypeFunc::Control);
2548 Node* mem = unlock->in(TypeFunc::Memory);
2549 Node* obj = unlock->obj_node();
2550 Node* box = unlock->box_node();
2551
2552 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2553
2554 // No need for a null check on unlock
2555
2556 // Make the merge point
2557 Node *region;
2558 Node *mem_phi;
2559
2560 if (UseOptoBiasInlining) {
2561 // Check for biased locking unlock case, which is a no-op.
2562 // See the full description in MacroAssembler::biased_locking_exit().
2563 region = new RegionNode(4);
2564 // create a Phi for the memory state
2565 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2566 mem_phi->init_req(3, mem);
2567
2568 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2569 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2570 markOopDesc::biased_lock_mask_in_place,
2571 markOopDesc::biased_lock_pattern);
2572 } else {
2573 region = new RegionNode(3);
2574 // create a Phi for the memory state
2575 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2576 }
2577
2578 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2579 funlock = transform_later( funlock )->as_FastUnlock();
2580 // Optimize test; set region slot 2
2581 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2582 Node *thread = transform_later(new ThreadLocalNode());
2583
2584 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2585 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2586 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2587
2588 extract_call_projections(call);
2589
2590 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2591 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2592
2593 // No exceptions for unlocking
2594 // Capture slow path
2595 // disconnect fall-through projection from call and create a new one
2596 // hook up users of fall-through projection to region
2597 Node *slow_ctrl = _fallthroughproj->clone();
2598 transform_later(slow_ctrl);
2599 _igvn.hash_delete(_fallthroughproj);
2600 _fallthroughproj->disconnect_inputs(NULL, C);
2601 region->init_req(1, slow_ctrl);
2602 // region inputs are now complete
2603 transform_later(region);
2604 _igvn.replace_node(_fallthroughproj, region);
2605
2606 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2607 mem_phi->init_req(1, memproj );
2608 mem_phi->init_req(2, mem);
2609 transform_later(mem_phi);
2610 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2611 }
2612
2613 //---------------------------eliminate_macro_nodes----------------------
2614 // Eliminate scalar replaced allocations and associated locks.
2615 void PhaseMacroExpand::eliminate_macro_nodes() {
2616 if (C->macro_count() == 0)
2617 return;
2618
2619 // First, attempt to eliminate locks
2620 int cnt = C->macro_count();
2621 for (int i=0; i < cnt; i++) {
2622 Node *n = C->macro_node(i);
2623 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2624 // Before elimination mark all associated (same box and obj)
2625 // lock and unlock nodes.
2626 mark_eliminated_locking_nodes(n->as_AbstractLock());
2627 }
2628 }
2629 bool progress = true;
2630 while (progress) {
2631 progress = false;
2632 for (int i = C->macro_count(); i > 0; i--) {
2633 Node * n = C->macro_node(i-1);
2634 bool success = false;
2635 debug_only(int old_macro_count = C->macro_count(););
2636 if (n->is_AbstractLock()) {
2637 success = eliminate_locking_node(n->as_AbstractLock());
2638 }
2639 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2640 progress = progress || success;
2641 }
2642 }
2643 // Next, attempt to eliminate allocations
2644 _has_locks = false;
2645 progress = true;
2646 while (progress) {
2647 progress = false;
2648 for (int i = C->macro_count(); i > 0; i--) {
2649 Node * n = C->macro_node(i-1);
2650 bool success = false;
2651 debug_only(int old_macro_count = C->macro_count(););
2652 switch (n->class_id()) {
2653 case Node::Class_Allocate:
2654 case Node::Class_AllocateArray:
2655 success = eliminate_allocate_node(n->as_Allocate());
2656 break;
2657 case Node::Class_CallStaticJava:
2658 success = eliminate_boxing_node(n->as_CallStaticJava());
2659 break;
2660 case Node::Class_Lock:
2661 case Node::Class_Unlock:
2662 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2663 _has_locks = true;
2664 break;
2665 case Node::Class_ArrayCopy:
2666 break;
2667 default:
2668 assert(n->Opcode() == Op_LoopLimit ||
2669 n->Opcode() == Op_Opaque1 ||
2670 n->Opcode() == Op_Opaque2 ||
2671 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2672 }
2673 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2674 progress = progress || success;
2675 }
2676 }
2677 }
2678
2679 //------------------------------expand_macro_nodes----------------------
2680 // Returns true if a failure occurred.
2681 bool PhaseMacroExpand::expand_macro_nodes() {
2682 // Last attempt to eliminate macro nodes.
2683 eliminate_macro_nodes();
2684
2685 // Make sure expansion will not cause node limit to be exceeded.
2686 // Worst case is a macro node gets expanded into about 200 nodes.
2687 // Allow 50% more for optimization.
2688 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2689 return true;
2690
2691 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2692 bool progress = true;
2693 while (progress) {
2694 progress = false;
2695 for (int i = C->macro_count(); i > 0; i--) {
2696 Node * n = C->macro_node(i-1);
2697 bool success = false;
2698 debug_only(int old_macro_count = C->macro_count(););
2699 if (n->Opcode() == Op_LoopLimit) {
2700 // Remove it from macro list and put on IGVN worklist to optimize.
2701 C->remove_macro_node(n);
2702 _igvn._worklist.push(n);
2703 success = true;
2704 } else if (n->Opcode() == Op_CallStaticJava) {
2705 // Remove it from macro list and put on IGVN worklist to optimize.
2706 C->remove_macro_node(n);
2707 _igvn._worklist.push(n);
2708 success = true;
2709 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2710 _igvn.replace_node(n, n->in(1));
2711 success = true;
2712 #if INCLUDE_RTM_OPT
2713 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2714 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2715 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2716 Node* cmp = n->unique_out();
2717 #ifdef ASSERT
2718 // Validate graph.
2719 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2720 BoolNode* bol = cmp->unique_out()->as_Bool();
2721 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2722 (bol->_test._test == BoolTest::ne), "");
2723 IfNode* ifn = bol->unique_out()->as_If();
2724 assert((ifn->outcnt() == 2) &&
2725 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2726 #endif
2727 Node* repl = n->in(1);
2728 if (!_has_locks) {
2729 // Remove RTM state check if there are no locks in the code.
2730 // Replace input to compare the same value.
2731 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2732 }
2733 _igvn.replace_node(n, repl);
2734 success = true;
2735 #endif
2736 }
2737 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2738 progress = progress || success;
2739 }
2740 }
2741
2742 // expand arraycopy "macro" nodes first
2743 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2744 // before the allocate nodes are expanded.
2745 int macro_idx = C->macro_count() - 1;
2746 while (macro_idx >= 0) {
2747 Node * n = C->macro_node(macro_idx);
2748 assert(n->is_macro(), "only macro nodes expected here");
2749 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2750 // node is unreachable, so don't try to expand it
2751 C->remove_macro_node(n);
2752 } else if (n->is_ArrayCopy()){
2753 int macro_count = C->macro_count();
2754 expand_arraycopy_node(n->as_ArrayCopy());
2755 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2756 }
2757 if (C->failing()) return true;
2758 macro_idx --;
2759 }
2760
2761 // expand "macro" nodes
2762 // nodes are removed from the macro list as they are processed
2763 while (C->macro_count() > 0) {
2764 int macro_count = C->macro_count();
2765 Node * n = C->macro_node(macro_count-1);
2766 assert(n->is_macro(), "only macro nodes expected here");
2767 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2768 // node is unreachable, so don't try to expand it
2769 C->remove_macro_node(n);
2770 continue;
2771 }
2772 switch (n->class_id()) {
2773 case Node::Class_Allocate:
2774 expand_allocate(n->as_Allocate());
2775 break;
2776 case Node::Class_AllocateArray:
2777 expand_allocate_array(n->as_AllocateArray());
2778 break;
2779 case Node::Class_Lock:
2780 expand_lock_node(n->as_Lock());
2781 break;
2782 case Node::Class_Unlock:
2783 expand_unlock_node(n->as_Unlock());
2784 break;
2785 default:
2786 assert(false, "unknown node type in macro list");
2787 }
2788 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2789 if (C->failing()) return true;
2790 }
2791
2792 _igvn.set_delay_transform(false);
2793 _igvn.optimize();
2794 if (C->failing()) return true;
2795 return false;
2796 }