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 "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/node.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subnode.hpp"
47 #include "opto/type.hpp"
48 #include "opto/valuetypenode.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51
52 //
53 // Replace any references to "oldref" in inputs to "use" with "newref".
54 // Returns the number of replacements made.
55 //
56 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
57 int nreplacements = 0;
58 uint req = use->req();
59 for (uint j = 0; j < use->len(); j++) {
60 Node *uin = use->in(j);
61 if (uin == oldref) {
62 if (j < req)
63 use->set_req(j, newref);
64 else
65 use->set_prec(j, newref);
66 nreplacements++;
67 } else if (j >= req && uin == NULL) {
68 break;
69 }
70 }
71 return nreplacements;
72 }
73
74 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
75 // Copy debug information and adjust JVMState information
76 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
77 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
78 int jvms_adj = new_dbg_start - old_dbg_start;
79 assert (new_dbg_start == newcall->req(), "argument count mismatch");
80
81 // SafePointScalarObject node could be referenced several times in debug info.
82 // Use Dict to record cloned nodes.
83 Dict* sosn_map = new Dict(cmpkey,hashkey);
84 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
85 Node* old_in = oldcall->in(i);
86 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
87 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
88 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
89 uint old_unique = C->unique();
90 Node* new_in = old_sosn->clone(sosn_map);
91 if (old_unique != C->unique()) { // New node?
92 new_in->set_req(0, C->root()); // reset control edge
93 new_in = transform_later(new_in); // Register new node.
94 }
95 old_in = new_in;
96 }
97 newcall->add_req(old_in);
98 }
99
100 // JVMS may be shared so clone it before we modify it
101 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
102 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
103 jvms->set_map(newcall);
104 jvms->set_locoff(jvms->locoff()+jvms_adj);
105 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
106 jvms->set_monoff(jvms->monoff()+jvms_adj);
107 jvms->set_scloff(jvms->scloff()+jvms_adj);
108 jvms->set_endoff(jvms->endoff()+jvms_adj);
109 }
110 }
111
112 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
113 Node* cmp;
114 if (mask != 0) {
115 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
116 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
117 } else {
118 cmp = word;
119 }
120 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
121 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
122 transform_later(iff);
123
124 // Fast path taken.
125 Node *fast_taken = transform_later(new IfFalseNode(iff));
126
127 // Fast path not-taken, i.e. slow path
128 Node *slow_taken = transform_later(new IfTrueNode(iff));
129
130 if (return_fast_path) {
131 region->init_req(edge, slow_taken); // Capture slow-control
132 return fast_taken;
133 } else {
134 region->init_req(edge, fast_taken); // Capture fast-control
135 return slow_taken;
136 }
137 }
138
139 //--------------------copy_predefined_input_for_runtime_call--------------------
140 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
141 // Set fixed predefined input arguments
142 call->init_req( TypeFunc::Control, ctrl );
143 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
144 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
145 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
146 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
147 }
148
149 //------------------------------make_slow_call---------------------------------
150 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
151 address slow_call, const char* leaf_name, Node* slow_path,
152 Node* parm0, Node* parm1, Node* parm2) {
153
154 // Slow-path call
155 CallNode *call = leaf_name
156 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
157 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
158
159 // Slow path call has no side-effects, uses few values
160 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
161 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
162 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
163 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
164 copy_call_debug_info(oldcall, call);
165 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
166 _igvn.replace_node(oldcall, call);
167 transform_later(call);
168
169 return call;
170 }
171
172 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
173 _fallthroughproj = NULL;
174 _fallthroughcatchproj = NULL;
175 _ioproj_fallthrough = NULL;
176 _ioproj_catchall = NULL;
177 _catchallcatchproj = NULL;
178 _memproj_fallthrough = NULL;
179 _memproj_catchall = NULL;
180 _resproj = NULL;
181 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
182 ProjNode *pn = call->fast_out(i)->as_Proj();
183 switch (pn->_con) {
184 case TypeFunc::Control:
185 {
186 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
187 _fallthroughproj = pn;
188 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
189 const Node *cn = pn->fast_out(j);
190 if (cn->is_Catch()) {
191 ProjNode *cpn = NULL;
192 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
193 cpn = cn->fast_out(k)->as_Proj();
194 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
195 if (cpn->_con == CatchProjNode::fall_through_index)
196 _fallthroughcatchproj = cpn;
197 else {
198 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
199 _catchallcatchproj = cpn;
200 }
201 }
202 }
203 break;
204 }
205 case TypeFunc::I_O:
206 if (pn->_is_io_use)
207 _ioproj_catchall = pn;
208 else
209 _ioproj_fallthrough = pn;
210 break;
211 case TypeFunc::Memory:
212 if (pn->_is_io_use)
213 _memproj_catchall = pn;
214 else
215 _memproj_fallthrough = pn;
216 break;
217 case TypeFunc::Parms:
218 _resproj = pn;
219 break;
220 default:
221 assert(false, "unexpected projection from allocation node.");
222 }
223 }
224
225 }
226
227 // Eliminate a card mark sequence. p2x is a ConvP2XNode
228 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
229 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
230 if (!UseG1GC) {
231 // vanilla/CMS post barrier
232 Node *shift = p2x->unique_out();
233 Node *addp = shift->unique_out();
234 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
235 Node *mem = addp->last_out(j);
236 if (UseCondCardMark && mem->is_Load()) {
237 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
238 // The load is checking if the card has been written so
239 // replace it with zero to fold the test.
240 _igvn.replace_node(mem, intcon(0));
241 continue;
242 }
243 assert(mem->is_Store(), "store required");
244 _igvn.replace_node(mem, mem->in(MemNode::Memory));
245 }
246 } else {
247 // G1 pre/post barriers
248 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
249 // It could be only one user, URShift node, in Object.clone() intrinsic
250 // but the new allocation is passed to arraycopy stub and it could not
251 // be scalar replaced. So we don't check the case.
252
253 // An other case of only one user (Xor) is when the value check for NULL
254 // in G1 post barrier is folded after CCP so the code which used URShift
255 // is removed.
256
257 // Take Region node before eliminating post barrier since it also
258 // eliminates CastP2X node when it has only one user.
259 Node* this_region = p2x->in(0);
260 assert(this_region != NULL, "");
261
262 // Remove G1 post barrier.
263
264 // Search for CastP2X->Xor->URShift->Cmp path which
265 // checks if the store done to a different from the value's region.
266 // And replace Cmp with #0 (false) to collapse G1 post barrier.
267 Node* xorx = p2x->find_out_with(Op_XorX);
268 if (xorx != NULL) {
269 Node* shift = xorx->unique_out();
270 Node* cmpx = shift->unique_out();
271 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
272 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
273 "missing region check in G1 post barrier");
274 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
275
276 // Remove G1 pre barrier.
277
278 // Search "if (marking != 0)" check and set it to "false".
279 // There is no G1 pre barrier if previous stored value is NULL
280 // (for example, after initialization).
281 if (this_region->is_Region() && this_region->req() == 3) {
282 int ind = 1;
283 if (!this_region->in(ind)->is_IfFalse()) {
284 ind = 2;
285 }
286 if (this_region->in(ind)->is_IfFalse()) {
287 Node* bol = this_region->in(ind)->in(0)->in(1);
288 assert(bol->is_Bool(), "");
289 cmpx = bol->in(1);
290 if (bol->as_Bool()->_test._test == BoolTest::ne &&
291 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
292 cmpx->in(1)->is_Load()) {
293 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
294 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
295 SATBMarkQueue::byte_offset_of_active());
296 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
297 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
298 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
299 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
300 }
301 }
302 }
303 }
304 } else {
305 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
306 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
307 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
308 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
309 Node* shift = p2x->find_out_with(Op_URShiftX);
310 assert(shift != NULL, "missing G1 post barrier");
311 Node* addp = shift->unique_out();
312 Node* load = addp->find_out_with(Op_LoadB);
313 assert(load != NULL, "missing G1 post barrier");
314 Node* cmpx = load->unique_out();
315 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
316 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
317 "missing card value check in G1 post barrier");
318 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
319 // There is no G1 pre barrier in this case
320 }
321 // Now CastP2X can be removed since it is used only on dead path
322 // which currently still alive until igvn optimize it.
323 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
324 _igvn.replace_node(p2x, top());
325 }
326 }
327
328 // Search for a memory operation for the specified memory slice.
329 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
330 Node *orig_mem = mem;
331 Node *alloc_mem = alloc->in(TypeFunc::Memory);
332 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
333 while (true) {
334 if (mem == alloc_mem || mem == start_mem ) {
335 return mem; // hit one of our sentinels
336 } else if (mem->is_MergeMem()) {
337 mem = mem->as_MergeMem()->memory_at(alias_idx);
338 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
339 Node *in = mem->in(0);
340 // we can safely skip over safepoints, calls, locks and membars because we
341 // already know that the object is safe to eliminate.
342 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
343 return in;
344 } else if (in->is_Call()) {
345 CallNode *call = in->as_Call();
346 if (call->may_modify(tinst, phase)) {
347 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
348 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
349 return in;
350 }
351 }
352 mem = in->in(TypeFunc::Memory);
353 } else if (in->is_MemBar()) {
354 ArrayCopyNode* ac = NULL;
355 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
356 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
357 return ac;
358 }
359 mem = in->in(TypeFunc::Memory);
360 } else {
361 assert(false, "unexpected projection");
362 }
363 } else if (mem->is_Store()) {
364 const TypePtr* atype = mem->as_Store()->adr_type();
365 int adr_idx = phase->C->get_alias_index(atype);
366 if (adr_idx == alias_idx) {
367 assert(atype->isa_oopptr(), "address type must be oopptr");
368 int adr_offset = atype->offset();
369 uint adr_iid = atype->is_oopptr()->instance_id();
370 // Array elements references have the same alias_idx
371 // but different offset and different instance_id.
372 if (adr_offset == offset && adr_iid == alloc->_idx)
373 return mem;
374 } else {
375 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
376 }
377 mem = mem->in(MemNode::Memory);
378 } else if (mem->is_ClearArray()) {
379 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
380 // Can not bypass initialization of the instance
381 // we are looking.
382 debug_only(intptr_t offset;)
383 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
384 InitializeNode* init = alloc->as_Allocate()->initialization();
385 // We are looking for stored value, return Initialize node
386 // or memory edge from Allocate node.
387 if (init != NULL)
388 return init;
389 else
390 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
391 }
392 // Otherwise skip it (the call updated 'mem' value).
393 } else if (mem->Opcode() == Op_SCMemProj) {
394 mem = mem->in(0);
395 Node* adr = NULL;
396 if (mem->is_LoadStore()) {
397 adr = mem->in(MemNode::Address);
398 } else {
399 assert(mem->Opcode() == Op_EncodeISOArray ||
400 mem->Opcode() == Op_StrCompressedCopy, "sanity");
401 adr = mem->in(3); // Destination array
402 }
403 const TypePtr* atype = adr->bottom_type()->is_ptr();
404 int adr_idx = phase->C->get_alias_index(atype);
405 if (adr_idx == alias_idx) {
406 DEBUG_ONLY(mem->dump();)
407 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
408 return NULL;
409 }
410 mem = mem->in(MemNode::Memory);
411 } else if (mem->Opcode() == Op_StrInflatedCopy) {
412 Node* adr = mem->in(3); // Destination array
413 const TypePtr* atype = adr->bottom_type()->is_ptr();
414 int adr_idx = phase->C->get_alias_index(atype);
415 if (adr_idx == alias_idx) {
416 DEBUG_ONLY(mem->dump();)
417 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
418 return NULL;
419 }
420 mem = mem->in(MemNode::Memory);
421 } else {
422 return mem;
423 }
424 assert(mem != orig_mem, "dead memory loop");
425 }
426 }
427
428 // Generate loads from source of the arraycopy for fields of
429 // destination needed at a deoptimization point
430 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
431 BasicType bt = ft;
432 const Type *type = ftype;
433 if (ft == T_NARROWOOP) {
434 bt = T_OBJECT;
435 type = ftype->make_oopptr();
436 }
437 Node* res = NULL;
438 if (ac->is_clonebasic()) {
439 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
440 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
441 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
442 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
443 } else {
444 if (ac->modifies(offset, offset, &_igvn, true)) {
445 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
446 uint shift = exact_log2(type2aelembytes(bt));
447 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
448 #ifdef _LP64
449 diff = _igvn.transform(new ConvI2LNode(diff));
450 #endif
451 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
452
453 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
454 Node* base = ac->in(ArrayCopyNode::Src);
455 Node* adr = _igvn.transform(new AddPNode(base, base, off));
456 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
457 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
458 }
459 }
460 if (res != NULL) {
461 res = _igvn.transform(res);
462 if (ftype->isa_narrowoop()) {
463 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
464 res = _igvn.transform(new EncodePNode(res, ftype));
465 }
466 return res;
467 }
468 return NULL;
469 }
470
471 //
472 // Given a Memory Phi, compute a value Phi containing the values from stores
473 // on the input paths.
474 // Note: this function is recursive, its depth is limited by the "level" argument
475 // Returns the computed Phi, or NULL if it cannot compute it.
476 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) {
477 assert(mem->is_Phi(), "sanity");
478 int alias_idx = C->get_alias_index(adr_t);
479 int offset = adr_t->offset();
480 int instance_id = adr_t->instance_id();
481
482 // Check if an appropriate value phi already exists.
483 Node* region = mem->in(0);
484 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
485 Node* phi = region->fast_out(k);
486 if (phi->is_Phi() && phi != mem &&
487 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
488 return phi;
489 }
490 }
491 // Check if an appropriate new value phi already exists.
492 Node* new_phi = value_phis->find(mem->_idx);
493 if (new_phi != NULL)
494 return new_phi;
495
496 if (level <= 0) {
497 return NULL; // Give up: phi tree too deep
498 }
499 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
500 Node *alloc_mem = alloc->in(TypeFunc::Memory);
501
502 uint length = mem->req();
503 GrowableArray <Node *> values(length, length, NULL, false);
504
505 // create a new Phi for the value
506 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
507 transform_later(phi);
508 value_phis->push(phi, mem->_idx);
509
510 for (uint j = 1; j < length; j++) {
511 Node *in = mem->in(j);
512 if (in == NULL || in->is_top()) {
513 values.at_put(j, in);
514 } else {
515 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
516 if (val == start_mem || val == alloc_mem) {
517 // hit a sentinel, return appropriate 0 value
518 values.at_put(j, _igvn.zerocon(ft));
519 continue;
520 }
521 if (val->is_Initialize()) {
522 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
523 }
524 if (val == NULL) {
525 return NULL; // can't find a value on this path
526 }
527 if (val == mem) {
528 values.at_put(j, mem);
529 } else if (val->is_Store()) {
530 values.at_put(j, val->in(MemNode::ValueIn));
531 } else if(val->is_Proj() && val->in(0) == alloc) {
532 values.at_put(j, _igvn.zerocon(ft));
533 } else if (val->is_Phi()) {
534 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
535 if (val == NULL) {
536 return NULL;
537 }
538 values.at_put(j, val);
539 } else if (val->Opcode() == Op_SCMemProj) {
540 assert(val->in(0)->is_LoadStore() ||
541 val->in(0)->Opcode() == Op_EncodeISOArray ||
542 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
543 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
544 return NULL;
545 } else if (val->is_ArrayCopy()) {
546 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
547 if (res == NULL) {
548 return NULL;
549 }
550 values.at_put(j, res);
551 } else {
552 #ifdef ASSERT
553 val->dump();
554 assert(false, "unknown node on this path");
555 #endif
556 return NULL; // unknown node on this path
557 }
558 }
559 }
560 // Set Phi's inputs
561 for (uint j = 1; j < length; j++) {
562 if (values.at(j) == mem) {
563 phi->init_req(j, phi);
564 } else {
565 phi->init_req(j, values.at(j));
566 }
567 }
568 return phi;
569 }
570
571 // Search the last value stored into the object's field.
572 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
573 assert(adr_t->is_known_instance_field(), "instance required");
574 assert(ft != T_VALUETYPE, "should not be used for value type fields");
575 int instance_id = adr_t->instance_id();
576 assert((uint)instance_id == alloc->_idx, "wrong allocation");
577
578 int alias_idx = C->get_alias_index(adr_t);
579 int offset = adr_t->offset();
580 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
581 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
582 Node *alloc_mem = alloc->in(TypeFunc::Memory);
583 Arena *a = Thread::current()->resource_area();
584 VectorSet visited(a);
585
586 bool done = sfpt_mem == alloc_mem;
587 Node *mem = sfpt_mem;
588 while (!done) {
589 if (visited.test_set(mem->_idx)) {
590 return NULL; // found a loop, give up
591 }
592 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
593 if (mem == start_mem || mem == alloc_mem) {
594 done = true; // hit a sentinel, return appropriate 0 value
595 } else if (mem->is_Initialize()) {
596 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
597 if (mem == NULL) {
598 done = true; // Something went wrong.
599 } else if (mem->is_Store()) {
600 const TypePtr* atype = mem->as_Store()->adr_type();
601 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
602 done = true;
603 }
604 } else if (mem->is_Store()) {
605 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
606 assert(atype != NULL, "address type must be oopptr");
607 assert(C->get_alias_index(atype) == alias_idx &&
608 atype->is_known_instance_field() && atype->offset() == offset &&
609 atype->instance_id() == instance_id, "store is correct memory slice");
610 done = true;
611 } else if (mem->is_Phi()) {
612 // try to find a phi's unique input
613 Node *unique_input = NULL;
614 Node *top = C->top();
615 for (uint i = 1; i < mem->req(); i++) {
616 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
617 if (n == NULL || n == top || n == mem) {
618 continue;
619 } else if (unique_input == NULL) {
620 unique_input = n;
621 } else if (unique_input != n) {
622 unique_input = top;
623 break;
624 }
625 }
626 if (unique_input != NULL && unique_input != top) {
627 mem = unique_input;
628 } else {
629 done = true;
630 }
631 } else if (mem->is_ArrayCopy()) {
632 done = true;
633 } else {
634 assert(false, "unexpected node");
635 }
636 }
637 if (mem != NULL) {
638 if (mem == start_mem || mem == alloc_mem) {
639 // hit a sentinel, return appropriate 0 value
640 return _igvn.zerocon(ft);
641 } else if (mem->is_Store()) {
642 return mem->in(MemNode::ValueIn);
643 } else if (mem->is_Phi()) {
644 // attempt to produce a Phi reflecting the values on the input paths of the Phi
645 Node_Stack value_phis(a, 8);
646 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
647 if (phi != NULL) {
648 return phi;
649 } else {
650 // Kill all new Phis
651 while(value_phis.is_nonempty()) {
652 Node* n = value_phis.node();
653 _igvn.replace_node(n, C->top());
654 value_phis.pop();
655 }
656 }
657 } else if (mem->is_ArrayCopy()) {
658 Node* ctl = mem->in(0);
659 Node* m = mem->in(TypeFunc::Memory);
660 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
661 // pin the loads in the uncommon trap path
662 ctl = sfpt_ctl;
663 m = sfpt_mem;
664 }
665 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
666 }
667 }
668 // Something went wrong.
669 return NULL;
670 }
671
672 // Search the last value stored into the value type's fields.
673 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
674 // Subtract the offset of the first field to account for the missing oop header
675 offset -= vk->first_field_offset();
676 // Create a new ValueTypeNode and retrieve the field values from memory
677 ValueTypeNode* vt = ValueTypeNode::make(_igvn, vk)->as_ValueType();
678 for (int i = 0; i < vk->field_count(); ++i) {
679 ciType* field_type = vt->field_type(i);
680 int field_offset = offset + vt->field_offset(i);
681 // Each value type field has its own memory slice
682 adr_type = adr_type->with_field_offset(field_offset);
683 Node* value = NULL;
684 if (field_type->basic_type() == T_VALUETYPE) {
685 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
686 } else {
687 value = value_from_mem(mem, ctl, field_type->basic_type(), Type::get_const_type(field_type), adr_type, alloc);
688 }
689 vt->set_field_value(i, value);
690 }
691 return vt;
692 }
693
694 // Check the possibility of scalar replacement.
695 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
696 // Scan the uses of the allocation to check for anything that would
697 // prevent us from eliminating it.
698 NOT_PRODUCT( const char* fail_eliminate = NULL; )
699 DEBUG_ONLY( Node* disq_node = NULL; )
700 bool can_eliminate = true;
701
702 Node* res = alloc->result_cast();
703 const TypeOopPtr* res_type = NULL;
704 if (res == NULL) {
705 // All users were eliminated.
706 } else if (!res->is_CheckCastPP()) {
707 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
708 can_eliminate = false;
709 } else {
710 res_type = _igvn.type(res)->isa_oopptr();
711 if (res_type == NULL) {
712 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
713 can_eliminate = false;
714 } else if (res_type->isa_aryptr()) {
715 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
716 if (length < 0) {
717 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
718 can_eliminate = false;
719 }
720 }
721 }
722
723 if (can_eliminate && res != NULL) {
724 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
725 j < jmax && can_eliminate; j++) {
726 Node* use = res->fast_out(j);
727
728 if (use->is_AddP()) {
729 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
730 int offset = addp_type->offset();
731
732 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
733 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
734 can_eliminate = false;
735 break;
736 }
737 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
738 k < kmax && can_eliminate; k++) {
739 Node* n = use->fast_out(k);
740 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
741 !(n->is_ArrayCopy() &&
742 n->as_ArrayCopy()->is_clonebasic() &&
743 n->in(ArrayCopyNode::Dest) == use)) {
744 DEBUG_ONLY(disq_node = n;)
745 if (n->is_Load() || n->is_LoadStore()) {
746 NOT_PRODUCT(fail_eliminate = "Field load";)
747 } else {
748 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
749 }
750 can_eliminate = false;
751 }
752 }
753 } else if (use->is_ArrayCopy() &&
754 (use->as_ArrayCopy()->is_arraycopy_validated() ||
755 use->as_ArrayCopy()->is_copyof_validated() ||
756 use->as_ArrayCopy()->is_copyofrange_validated()) &&
757 use->in(ArrayCopyNode::Dest) == res) {
758 // ok to eliminate
759 } else if (use->is_SafePoint()) {
760 SafePointNode* sfpt = use->as_SafePoint();
761 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
762 // Object is passed as argument.
763 DEBUG_ONLY(disq_node = use;)
764 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
765 can_eliminate = false;
766 }
767 Node* sfptMem = sfpt->memory();
768 if (sfptMem == NULL || sfptMem->is_top()) {
769 DEBUG_ONLY(disq_node = use;)
770 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
771 can_eliminate = false;
772 } else {
773 safepoints.append_if_missing(sfpt);
774 }
775 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
776 if (use->is_Phi()) {
777 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
778 NOT_PRODUCT(fail_eliminate = "Object is return value";)
779 } else {
780 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
781 }
782 DEBUG_ONLY(disq_node = use;)
783 } else {
784 if (use->Opcode() == Op_Return) {
785 NOT_PRODUCT(fail_eliminate = "Object is return value";)
786 } else {
787 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
788 }
789 DEBUG_ONLY(disq_node = use;)
790 }
791 can_eliminate = false;
792 }
793 }
794 }
795
796 #ifndef PRODUCT
797 if (PrintEliminateAllocations) {
798 if (can_eliminate) {
799 tty->print("Scalar ");
800 if (res == NULL)
801 alloc->dump();
802 else
803 res->dump();
804 } else if (alloc->_is_scalar_replaceable) {
805 tty->print("NotScalar (%s)", fail_eliminate);
806 if (res == NULL)
807 alloc->dump();
808 else
809 res->dump();
810 #ifdef ASSERT
811 if (disq_node != NULL) {
812 tty->print(" >>>> ");
813 disq_node->dump();
814 }
815 #endif /*ASSERT*/
816 }
817 }
818 #endif
819 return can_eliminate;
820 }
821
822 // Do scalar replacement.
823 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
824 GrowableArray <SafePointNode *> safepoints_done;
825
826 ciKlass* klass = NULL;
827 ciInstanceKlass* iklass = NULL;
828 int nfields = 0;
829 int array_base = 0;
830 int element_size = 0;
831 BasicType basic_elem_type = T_ILLEGAL;
832 ciType* elem_type = NULL;
833
834 Node* res = alloc->result_cast();
835 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
836 const TypeOopPtr* res_type = NULL;
837 if (res != NULL) { // Could be NULL when there are no users
838 res_type = _igvn.type(res)->isa_oopptr();
839 }
840
841 if (res != NULL) {
842 klass = res_type->klass();
843 // Value types are only allocated on demand
844 if (res_type->isa_instptr() || res_type->isa_valuetypeptr()) {
845 // find the fields of the class which will be needed for safepoint debug information
846 assert(klass->is_instance_klass(), "must be an instance klass.");
847 iklass = klass->as_instance_klass();
848 nfields = iklass->nof_nonstatic_fields();
849 } else {
850 // find the array's elements which will be needed for safepoint debug information
851 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
852 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
853 elem_type = klass->as_array_klass()->element_type();
854 basic_elem_type = elem_type->basic_type();
855 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
856 element_size = type2aelembytes(basic_elem_type);
857 }
858 }
859 //
860 // Process the safepoint uses
861 //
862 while (safepoints.length() > 0) {
863 SafePointNode* sfpt = safepoints.pop();
864 Node* mem = sfpt->memory();
865 Node* ctl = sfpt->control();
866 assert(sfpt->jvms() != NULL, "missed JVMS");
867 // Fields of scalar objs are referenced only at the end
868 // of regular debuginfo at the last (youngest) JVMS.
869 // Record relative start index.
870 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
871 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
872 #ifdef ASSERT
873 alloc,
874 #endif
875 first_ind, nfields);
876 sobj->init_req(0, C->root());
877 transform_later(sobj);
878
879 // Scan object's fields adding an input to the safepoint for each field.
880 for (int j = 0; j < nfields; j++) {
881 intptr_t offset;
882 ciField* field = NULL;
883 if (iklass != NULL) {
884 field = iklass->nonstatic_field_at(j);
885 offset = field->offset();
886 elem_type = field->type();
887 basic_elem_type = field->layout_type();
888 // Value type fields should not have safepoint uses
889 assert(basic_elem_type != T_VALUETYPE, "value type fields are flattened");
890 } else {
891 offset = array_base + j * (intptr_t)element_size;
892 }
893
894 const Type *field_type;
895 // The next code is taken from Parse::do_get_xxx().
896 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
897 if (!elem_type->is_loaded()) {
898 field_type = TypeInstPtr::BOTTOM;
899 } else if (field != NULL && field->is_static_constant()) {
900 // This can happen if the constant oop is non-perm.
901 ciObject* con = field->constant_value().as_object();
902 // Do not "join" in the previous type; it doesn't add value,
903 // and may yield a vacuous result if the field is of interface type.
904 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
905 assert(field_type != NULL, "field singleton type must be consistent");
906 } else {
907 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
908 }
909 if (UseCompressedOops) {
910 field_type = field_type->make_narrowoop();
911 basic_elem_type = T_NARROWOOP;
912 }
913 } else {
914 field_type = Type::get_const_basic_type(basic_elem_type);
915 }
916
917 Node* field_val = NULL;
918 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
919 if (klass->is_value_array_klass()) {
920 ciValueKlass* vk = elem_type->as_value_klass();
921 assert(vk->flatten_array(), "must be flattened");
922 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
923 } else {
924 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
925 }
926 if (field_val == NULL) {
927 // We weren't able to find a value for this field,
928 // give up on eliminating this allocation.
929
930 // Remove any extra entries we added to the safepoint.
931 uint last = sfpt->req() - 1;
932 for (int k = 0; k < j; k++) {
933 sfpt->del_req(last--);
934 }
935 _igvn._worklist.push(sfpt);
936 // rollback processed safepoints
937 while (safepoints_done.length() > 0) {
938 SafePointNode* sfpt_done = safepoints_done.pop();
939 // remove any extra entries we added to the safepoint
940 last = sfpt_done->req() - 1;
941 for (int k = 0; k < nfields; k++) {
942 sfpt_done->del_req(last--);
943 }
944 JVMState *jvms = sfpt_done->jvms();
945 jvms->set_endoff(sfpt_done->req());
946 // Now make a pass over the debug information replacing any references
947 // to SafePointScalarObjectNode with the allocated object.
948 int start = jvms->debug_start();
949 int end = jvms->debug_end();
950 for (int i = start; i < end; i++) {
951 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
952 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
953 if (scobj->first_index(jvms) == sfpt_done->req() &&
954 scobj->n_fields() == (uint)nfields) {
955 assert(scobj->alloc() == alloc, "sanity");
956 sfpt_done->set_req(i, res);
957 }
958 }
959 }
960 _igvn._worklist.push(sfpt_done);
961 }
962 #ifndef PRODUCT
963 if (PrintEliminateAllocations) {
964 if (field != NULL) {
965 tty->print("=== At SafePoint node %d can't find value of Field: ",
966 sfpt->_idx);
967 field->print();
968 int field_idx = C->get_alias_index(field_addr_type);
969 tty->print(" (alias_idx=%d)", field_idx);
970 } else { // Array's element
971 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
972 sfpt->_idx, j);
973 }
974 tty->print(", which prevents elimination of: ");
975 if (res == NULL)
976 alloc->dump();
977 else
978 res->dump();
979 }
980 #endif
981 return false;
982 }
983 if (UseCompressedOops && field_type->isa_narrowoop()) {
984 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
985 // to be able scalar replace the allocation.
986 if (field_val->is_EncodeP()) {
987 field_val = field_val->in(1);
988 } else {
989 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
990 }
991 }
992 sfpt->add_req(field_val);
993 }
994 JVMState *jvms = sfpt->jvms();
995 jvms->set_endoff(sfpt->req());
996 // Now make a pass over the debug information replacing any references
997 // to the allocated object with "sobj"
998 int start = jvms->debug_start();
999 int end = jvms->debug_end();
1000 sfpt->replace_edges_in_range(res, sobj, start, end);
1001 _igvn._worklist.push(sfpt);
1002 safepoints_done.append_if_missing(sfpt); // keep it for rollback
1003 }
1004 return true;
1005 }
1006
1007 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1008 Node* ctl_proj = n->proj_out(TypeFunc::Control);
1009 Node* mem_proj = n->proj_out(TypeFunc::Memory);
1010 if (ctl_proj != NULL) {
1011 igvn.replace_node(ctl_proj, n->in(0));
1012 }
1013 if (mem_proj != NULL) {
1014 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1015 }
1016 }
1017
1018 // Process users of eliminated allocation.
1019 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
1020 Node* res = alloc->result_cast();
1021 if (res != NULL) {
1022 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1023 Node *use = res->last_out(j);
1024 uint oc1 = res->outcnt();
1025
1026 if (use->is_AddP()) {
1027 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1028 Node *n = use->last_out(k);
1029 uint oc2 = use->outcnt();
1030 if (n->is_Store()) {
1031 #ifdef ASSERT
1032 // Verify that there is no dependent MemBarVolatile nodes,
1033 // they should be removed during IGVN, see MemBarNode::Ideal().
1034 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1035 p < pmax; p++) {
1036 Node* mb = n->fast_out(p);
1037 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1038 mb->req() <= MemBarNode::Precedent ||
1039 mb->in(MemBarNode::Precedent) != n,
1040 "MemBarVolatile should be eliminated for non-escaping object");
1041 }
1042 #endif
1043 _igvn.replace_node(n, n->in(MemNode::Memory));
1044 } else if (n->is_ArrayCopy()) {
1045 // Disconnect ArrayCopy node
1046 ArrayCopyNode* ac = n->as_ArrayCopy();
1047 assert(ac->is_clonebasic(), "unexpected array copy kind");
1048 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1049 disconnect_projections(ac, _igvn);
1050 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1051 Node* membar_before = alloc->in(0)->in(0);
1052 disconnect_projections(membar_before->as_MemBar(), _igvn);
1053 if (membar_after->is_MemBar()) {
1054 disconnect_projections(membar_after->as_MemBar(), _igvn);
1055 }
1056 } else {
1057 eliminate_card_mark(n);
1058 }
1059 k -= (oc2 - use->outcnt());
1060 }
1061 } else if (use->is_ArrayCopy()) {
1062 // Disconnect ArrayCopy node
1063 ArrayCopyNode* ac = use->as_ArrayCopy();
1064 assert(ac->is_arraycopy_validated() ||
1065 ac->is_copyof_validated() ||
1066 ac->is_copyofrange_validated(), "unsupported");
1067 CallProjections callprojs;
1068 ac->extract_projections(&callprojs, true);
1069
1070 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1071 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1072 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1073
1074 // Set control to top. IGVN will remove the remaining projections
1075 ac->set_req(0, top());
1076 ac->replace_edge(res, top());
1077
1078 // Disconnect src right away: it can help find new
1079 // opportunities for allocation elimination
1080 Node* src = ac->in(ArrayCopyNode::Src);
1081 ac->replace_edge(src, top());
1082 if (src->outcnt() == 0) {
1083 _igvn.remove_dead_node(src);
1084 }
1085
1086 _igvn._worklist.push(ac);
1087 } else {
1088 eliminate_card_mark(use);
1089 }
1090 j -= (oc1 - res->outcnt());
1091 }
1092 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1093 _igvn.remove_dead_node(res);
1094 }
1095
1096 //
1097 // Process other users of allocation's projections
1098 //
1099 if (_resproj != NULL && _resproj->outcnt() != 0) {
1100 // First disconnect stores captured by Initialize node.
1101 // If Initialize node is eliminated first in the following code,
1102 // it will kill such stores and DUIterator_Last will assert.
1103 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1104 Node *use = _resproj->fast_out(j);
1105 if (use->is_AddP()) {
1106 // raw memory addresses used only by the initialization
1107 _igvn.replace_node(use, C->top());
1108 --j; --jmax;
1109 }
1110 }
1111 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1112 Node *use = _resproj->last_out(j);
1113 uint oc1 = _resproj->outcnt();
1114 if (use->is_Initialize()) {
1115 // Eliminate Initialize node.
1116 InitializeNode *init = use->as_Initialize();
1117 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1118 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
1119 if (ctrl_proj != NULL) {
1120 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1121 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1122 }
1123 Node *mem_proj = init->proj_out(TypeFunc::Memory);
1124 if (mem_proj != NULL) {
1125 Node *mem = init->in(TypeFunc::Memory);
1126 #ifdef ASSERT
1127 if (mem->is_MergeMem()) {
1128 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1129 } else {
1130 assert(mem == _memproj_fallthrough, "allocation memory projection");
1131 }
1132 #endif
1133 _igvn.replace_node(mem_proj, mem);
1134 }
1135 } else {
1136 assert(false, "only Initialize or AddP expected");
1137 }
1138 j -= (oc1 - _resproj->outcnt());
1139 }
1140 }
1141 if (_fallthroughcatchproj != NULL) {
1142 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1143 }
1144 if (_memproj_fallthrough != NULL) {
1145 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1146 }
1147 if (_memproj_catchall != NULL) {
1148 _igvn.replace_node(_memproj_catchall, C->top());
1149 }
1150 if (_ioproj_fallthrough != NULL) {
1151 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1152 }
1153 if (_ioproj_catchall != NULL) {
1154 _igvn.replace_node(_ioproj_catchall, C->top());
1155 }
1156 if (_catchallcatchproj != NULL) {
1157 _igvn.replace_node(_catchallcatchproj, C->top());
1158 }
1159 }
1160
1161 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1162 // Don't do scalar replacement if the frame can be popped by JVMTI:
1163 // if reallocation fails during deoptimization we'll pop all
1164 // interpreter frames for this compiled frame and that won't play
1165 // nice with JVMTI popframe.
1166 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1167 return false;
1168 }
1169 Node* klass = alloc->in(AllocateNode::KlassNode);
1170 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1171 Node* res = alloc->result_cast();
1172 // Eliminate boxing allocations which are not used
1173 // regardless scalar replacable status.
1174 bool boxing_alloc = C->eliminate_boxing() &&
1175 tklass->klass()->is_instance_klass() &&
1176 tklass->klass()->as_instance_klass()->is_box_klass();
1177 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1178 return false;
1179 }
1180
1181 extract_call_projections(alloc);
1182
1183 GrowableArray <SafePointNode *> safepoints;
1184 if (!can_eliminate_allocation(alloc, safepoints)) {
1185 return false;
1186 }
1187
1188 if (!alloc->_is_scalar_replaceable) {
1189 assert(res == NULL, "sanity");
1190 // We can only eliminate allocation if all debug info references
1191 // are already replaced with SafePointScalarObject because
1192 // we can't search for a fields value without instance_id.
1193 if (safepoints.length() > 0) {
1194 return false;
1195 }
1196 }
1197
1198 if (!scalar_replacement(alloc, safepoints)) {
1199 return false;
1200 }
1201
1202 CompileLog* log = C->log();
1203 if (log != NULL) {
1204 log->head("eliminate_allocation type='%d'",
1205 log->identify(tklass->klass()));
1206 JVMState* p = alloc->jvms();
1207 while (p != NULL) {
1208 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1209 p = p->caller();
1210 }
1211 log->tail("eliminate_allocation");
1212 }
1213
1214 process_users_of_allocation(alloc);
1215
1216 #ifndef PRODUCT
1217 if (PrintEliminateAllocations) {
1218 if (alloc->is_AllocateArray())
1219 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1220 else
1221 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1222 }
1223 #endif
1224
1225 return true;
1226 }
1227
1228 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1229 // EA should remove all uses of non-escaping boxing node.
1230 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1231 return false;
1232 }
1233
1234 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1235
1236 extract_call_projections(boxing);
1237
1238 const TypeTuple* r = boxing->tf()->range_sig();
1239 assert(r->cnt() > TypeFunc::Parms, "sanity");
1240 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1241 assert(t != NULL, "sanity");
1242
1243 CompileLog* log = C->log();
1244 if (log != NULL) {
1245 log->head("eliminate_boxing type='%d'",
1246 log->identify(t->klass()));
1247 JVMState* p = boxing->jvms();
1248 while (p != NULL) {
1249 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1250 p = p->caller();
1251 }
1252 log->tail("eliminate_boxing");
1253 }
1254
1255 process_users_of_allocation(boxing);
1256
1257 #ifndef PRODUCT
1258 if (PrintEliminateAllocations) {
1259 tty->print("++++ Eliminated: %d ", boxing->_idx);
1260 boxing->method()->print_short_name(tty);
1261 tty->cr();
1262 }
1263 #endif
1264
1265 return true;
1266 }
1267
1268 //---------------------------set_eden_pointers-------------------------
1269 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1270 if (UseTLAB) { // Private allocation: load from TLS
1271 Node* thread = transform_later(new ThreadLocalNode());
1272 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1273 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1274 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1275 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1276 } else { // Shared allocation: load from globals
1277 CollectedHeap* ch = Universe::heap();
1278 address top_adr = (address)ch->top_addr();
1279 address end_adr = (address)ch->end_addr();
1280 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1281 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1282 }
1283 }
1284
1285
1286 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1287 Node* adr = basic_plus_adr(base, offset);
1288 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1289 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1290 transform_later(value);
1291 return value;
1292 }
1293
1294
1295 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1296 Node* adr = basic_plus_adr(base, offset);
1297 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1298 transform_later(mem);
1299 return mem;
1300 }
1301
1302 //=============================================================================
1303 //
1304 // A L L O C A T I O N
1305 //
1306 // Allocation attempts to be fast in the case of frequent small objects.
1307 // It breaks down like this:
1308 //
1309 // 1) Size in doublewords is computed. This is a constant for objects and
1310 // variable for most arrays. Doubleword units are used to avoid size
1311 // overflow of huge doubleword arrays. We need doublewords in the end for
1312 // rounding.
1313 //
1314 // 2) Size is checked for being 'too large'. Too-large allocations will go
1315 // the slow path into the VM. The slow path can throw any required
1316 // exceptions, and does all the special checks for very large arrays. The
1317 // size test can constant-fold away for objects. For objects with
1318 // finalizers it constant-folds the otherway: you always go slow with
1319 // finalizers.
1320 //
1321 // 3) If NOT using TLABs, this is the contended loop-back point.
1322 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1323 //
1324 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1325 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1326 // "size*8" we always enter the VM, where "largish" is a constant picked small
1327 // enough that there's always space between the eden max and 4Gig (old space is
1328 // there so it's quite large) and large enough that the cost of entering the VM
1329 // is dwarfed by the cost to initialize the space.
1330 //
1331 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1332 // down. If contended, repeat at step 3. If using TLABs normal-store
1333 // adjusted heap top back down; there is no contention.
1334 //
1335 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1336 // fields.
1337 //
1338 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1339 // oop flavor.
1340 //
1341 //=============================================================================
1342 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1343 // Allocations bigger than this always go the slow route.
1344 // This value must be small enough that allocation attempts that need to
1345 // trigger exceptions go the slow route. Also, it must be small enough so
1346 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1347 //=============================================================================j//
1348 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1349 // The allocator will coalesce int->oop copies away. See comment in
1350 // coalesce.cpp about how this works. It depends critically on the exact
1351 // code shape produced here, so if you are changing this code shape
1352 // make sure the GC info for the heap-top is correct in and around the
1353 // slow-path call.
1354 //
1355
1356 void PhaseMacroExpand::expand_allocate_common(
1357 AllocateNode* alloc, // allocation node to be expanded
1358 Node* length, // array length for an array allocation
1359 const TypeFunc* slow_call_type, // Type of slow call
1360 address slow_call_address // Address of slow call
1361 )
1362 {
1363
1364 Node* ctrl = alloc->in(TypeFunc::Control);
1365 Node* mem = alloc->in(TypeFunc::Memory);
1366 Node* i_o = alloc->in(TypeFunc::I_O);
1367 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1368 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1369 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1370
1371 assert(ctrl != NULL, "must have control");
1372 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1373 // they will not be used if "always_slow" is set
1374 enum { slow_result_path = 1, fast_result_path = 2 };
1375 Node *result_region = NULL;
1376 Node *result_phi_rawmem = NULL;
1377 Node *result_phi_rawoop = NULL;
1378 Node *result_phi_i_o = NULL;
1379
1380 // The initial slow comparison is a size check, the comparison
1381 // we want to do is a BoolTest::gt
1382 bool always_slow = false;
1383 int tv = _igvn.find_int_con(initial_slow_test, -1);
1384 if (tv >= 0) {
1385 always_slow = (tv == 1);
1386 initial_slow_test = NULL;
1387 } else {
1388 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1389 }
1390
1391 if (C->env()->dtrace_alloc_probes() ||
1392 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1393 // Force slow-path allocation
1394 always_slow = true;
1395 initial_slow_test = NULL;
1396 }
1397
1398
1399 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1400 Node *slow_region = NULL;
1401 Node *toobig_false = ctrl;
1402
1403 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1404 // generate the initial test if necessary
1405 if (initial_slow_test != NULL ) {
1406 slow_region = new RegionNode(3);
1407
1408 // Now make the initial failure test. Usually a too-big test but
1409 // might be a TRUE for finalizers or a fancy class check for
1410 // newInstance0.
1411 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1412 transform_later(toobig_iff);
1413 // Plug the failing-too-big test into the slow-path region
1414 Node *toobig_true = new IfTrueNode( toobig_iff );
1415 transform_later(toobig_true);
1416 slow_region ->init_req( too_big_or_final_path, toobig_true );
1417 toobig_false = new IfFalseNode( toobig_iff );
1418 transform_later(toobig_false);
1419 } else { // No initial test, just fall into next case
1420 toobig_false = ctrl;
1421 debug_only(slow_region = NodeSentinel);
1422 }
1423
1424 Node *slow_mem = mem; // save the current memory state for slow path
1425 // generate the fast allocation code unless we know that the initial test will always go slow
1426 if (!always_slow) {
1427 // Fast path modifies only raw memory.
1428 if (mem->is_MergeMem()) {
1429 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1430 }
1431
1432 Node* eden_top_adr;
1433 Node* eden_end_adr;
1434
1435 set_eden_pointers(eden_top_adr, eden_end_adr);
1436
1437 // Load Eden::end. Loop invariant and hoisted.
1438 //
1439 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1440 // a TLAB to work around a bug where these values were being moved across
1441 // a safepoint. These are not oops, so they cannot be include in the oop
1442 // map, but they can be changed by a GC. The proper way to fix this would
1443 // be to set the raw memory state when generating a SafepointNode. However
1444 // this will require extensive changes to the loop optimization in order to
1445 // prevent a degradation of the optimization.
1446 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1447 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1448
1449 // allocate the Region and Phi nodes for the result
1450 result_region = new RegionNode(3);
1451 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1452 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1453 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1454
1455 // We need a Region for the loop-back contended case.
1456 enum { fall_in_path = 1, contended_loopback_path = 2 };
1457 Node *contended_region;
1458 Node *contended_phi_rawmem;
1459 if (UseTLAB) {
1460 contended_region = toobig_false;
1461 contended_phi_rawmem = mem;
1462 } else {
1463 contended_region = new RegionNode(3);
1464 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1465 // Now handle the passing-too-big test. We fall into the contended
1466 // loop-back merge point.
1467 contended_region ->init_req(fall_in_path, toobig_false);
1468 contended_phi_rawmem->init_req(fall_in_path, mem);
1469 transform_later(contended_region);
1470 transform_later(contended_phi_rawmem);
1471 }
1472
1473 // Load(-locked) the heap top.
1474 // See note above concerning the control input when using a TLAB
1475 Node *old_eden_top = UseTLAB
1476 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1477 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1478
1479 transform_later(old_eden_top);
1480 // Add to heap top to get a new heap top
1481 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1482 transform_later(new_eden_top);
1483 // Check for needing a GC; compare against heap end
1484 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1485 transform_later(needgc_cmp);
1486 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1487 transform_later(needgc_bol);
1488 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1489 transform_later(needgc_iff);
1490
1491 // Plug the failing-heap-space-need-gc test into the slow-path region
1492 Node *needgc_true = new IfTrueNode(needgc_iff);
1493 transform_later(needgc_true);
1494 if (initial_slow_test) {
1495 slow_region->init_req(need_gc_path, needgc_true);
1496 // This completes all paths into the slow merge point
1497 transform_later(slow_region);
1498 } else { // No initial slow path needed!
1499 // Just fall from the need-GC path straight into the VM call.
1500 slow_region = needgc_true;
1501 }
1502 // No need for a GC. Setup for the Store-Conditional
1503 Node *needgc_false = new IfFalseNode(needgc_iff);
1504 transform_later(needgc_false);
1505
1506 // Grab regular I/O before optional prefetch may change it.
1507 // Slow-path does no I/O so just set it to the original I/O.
1508 result_phi_i_o->init_req(slow_result_path, i_o);
1509
1510 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1511 old_eden_top, new_eden_top, length);
1512
1513 // Name successful fast-path variables
1514 Node* fast_oop = old_eden_top;
1515 Node* fast_oop_ctrl;
1516 Node* fast_oop_rawmem;
1517
1518 // Store (-conditional) the modified eden top back down.
1519 // StorePConditional produces flags for a test PLUS a modified raw
1520 // memory state.
1521 if (UseTLAB) {
1522 Node* store_eden_top =
1523 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1524 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1525 transform_later(store_eden_top);
1526 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1527 fast_oop_rawmem = store_eden_top;
1528 } else {
1529 Node* store_eden_top =
1530 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1531 new_eden_top, fast_oop/*old_eden_top*/);
1532 transform_later(store_eden_top);
1533 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1534 transform_later(contention_check);
1535 store_eden_top = new SCMemProjNode(store_eden_top);
1536 transform_later(store_eden_top);
1537
1538 // If not using TLABs, check to see if there was contention.
1539 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1540 transform_later(contention_iff);
1541 Node *contention_true = new IfTrueNode(contention_iff);
1542 transform_later(contention_true);
1543 // If contention, loopback and try again.
1544 contended_region->init_req(contended_loopback_path, contention_true);
1545 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1546
1547 // Fast-path succeeded with no contention!
1548 Node *contention_false = new IfFalseNode(contention_iff);
1549 transform_later(contention_false);
1550 fast_oop_ctrl = contention_false;
1551
1552 // Bump total allocated bytes for this thread
1553 Node* thread = new ThreadLocalNode();
1554 transform_later(thread);
1555 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1556 in_bytes(JavaThread::allocated_bytes_offset()));
1557 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1558 0, TypeLong::LONG, T_LONG);
1559 #ifdef _LP64
1560 Node* alloc_size = size_in_bytes;
1561 #else
1562 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1563 transform_later(alloc_size);
1564 #endif
1565 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1566 transform_later(new_alloc_bytes);
1567 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1568 0, new_alloc_bytes, T_LONG);
1569 }
1570
1571 InitializeNode* init = alloc->initialization();
1572 fast_oop_rawmem = initialize_object(alloc,
1573 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1574 klass_node, length, size_in_bytes);
1575
1576 // If initialization is performed by an array copy, any required
1577 // MemBarStoreStore was already added. If the object does not
1578 // escape no need for a MemBarStoreStore. If the object does not
1579 // escape in its initializer and memory barrier (MemBarStoreStore or
1580 // stronger) is already added at exit of initializer, also no need
1581 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1582 // so that stores that initialize this object can't be reordered
1583 // with a subsequent store that makes this object accessible by
1584 // other threads.
1585 // Other threads include java threads and JVM internal threads
1586 // (for example concurrent GC threads). Current concurrent GC
1587 // implementation: CMS and G1 will not scan newly created object,
1588 // so it's safe to skip storestore barrier when allocation does
1589 // not escape.
1590 if (!alloc->does_not_escape_thread() &&
1591 !alloc->is_allocation_MemBar_redundant() &&
1592 (init == NULL || !init->is_complete_with_arraycopy())) {
1593 if (init == NULL || init->req() < InitializeNode::RawStores) {
1594 // No InitializeNode or no stores captured by zeroing
1595 // elimination. Simply add the MemBarStoreStore after object
1596 // initialization.
1597 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1598 transform_later(mb);
1599
1600 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1601 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1602 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1603 transform_later(fast_oop_ctrl);
1604 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1605 transform_later(fast_oop_rawmem);
1606 } else {
1607 // Add the MemBarStoreStore after the InitializeNode so that
1608 // all stores performing the initialization that were moved
1609 // before the InitializeNode happen before the storestore
1610 // barrier.
1611
1612 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1613 Node* init_mem = init->proj_out(TypeFunc::Memory);
1614
1615 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1616 transform_later(mb);
1617
1618 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1619 transform_later(ctrl);
1620 Node* mem = new ProjNode(init,TypeFunc::Memory);
1621 transform_later(mem);
1622
1623 // The MemBarStoreStore depends on control and memory coming
1624 // from the InitializeNode
1625 mb->init_req(TypeFunc::Memory, mem);
1626 mb->init_req(TypeFunc::Control, ctrl);
1627
1628 ctrl = new ProjNode(mb,TypeFunc::Control);
1629 transform_later(ctrl);
1630 mem = new ProjNode(mb,TypeFunc::Memory);
1631 transform_later(mem);
1632
1633 // All nodes that depended on the InitializeNode for control
1634 // and memory must now depend on the MemBarNode that itself
1635 // depends on the InitializeNode
1636 if (init_ctrl != NULL) {
1637 _igvn.replace_node(init_ctrl, ctrl);
1638 }
1639 if (init_mem != NULL) {
1640 _igvn.replace_node(init_mem, mem);
1641 }
1642 }
1643 }
1644
1645 if (C->env()->dtrace_extended_probes()) {
1646 // Slow-path call
1647 int size = TypeFunc::Parms + 2;
1648 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1649 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1650 "dtrace_object_alloc",
1651 TypeRawPtr::BOTTOM);
1652
1653 // Get base of thread-local storage area
1654 Node* thread = new ThreadLocalNode();
1655 transform_later(thread);
1656
1657 call->init_req(TypeFunc::Parms+0, thread);
1658 call->init_req(TypeFunc::Parms+1, fast_oop);
1659 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1660 call->init_req(TypeFunc::I_O , top()); // does no i/o
1661 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1662 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1663 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1664 transform_later(call);
1665 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1666 transform_later(fast_oop_ctrl);
1667 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1668 transform_later(fast_oop_rawmem);
1669 }
1670
1671 // Plug in the successful fast-path into the result merge point
1672 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1673 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1674 result_phi_i_o ->init_req(fast_result_path, i_o);
1675 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1676 } else {
1677 slow_region = ctrl;
1678 result_phi_i_o = i_o; // Rename it to use in the following code.
1679 }
1680
1681 // Generate slow-path call
1682 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1683 OptoRuntime::stub_name(slow_call_address),
1684 alloc->jvms()->bci(),
1685 TypePtr::BOTTOM);
1686 call->init_req( TypeFunc::Control, slow_region );
1687 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1688 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1689 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1690 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1691
1692 call->init_req(TypeFunc::Parms+0, klass_node);
1693 if (length != NULL) {
1694 call->init_req(TypeFunc::Parms+1, length);
1695 }
1696
1697 // Copy debug information and adjust JVMState information, then replace
1698 // allocate node with the call
1699 copy_call_debug_info((CallNode *) alloc, call);
1700 if (!always_slow) {
1701 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1702 } else {
1703 // Hook i_o projection to avoid its elimination during allocation
1704 // replacement (when only a slow call is generated).
1705 call->set_req(TypeFunc::I_O, result_phi_i_o);
1706 }
1707 _igvn.replace_node(alloc, call);
1708 transform_later(call);
1709
1710 // Identify the output projections from the allocate node and
1711 // adjust any references to them.
1712 // The control and io projections look like:
1713 //
1714 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1715 // Allocate Catch
1716 // ^---Proj(io) <-------+ ^---CatchProj(io)
1717 //
1718 // We are interested in the CatchProj nodes.
1719 //
1720 extract_call_projections(call);
1721
1722 // An allocate node has separate memory projections for the uses on
1723 // the control and i_o paths. Replace the control memory projection with
1724 // result_phi_rawmem (unless we are only generating a slow call when
1725 // both memory projections are combined)
1726 if (!always_slow && _memproj_fallthrough != NULL) {
1727 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1728 Node *use = _memproj_fallthrough->fast_out(i);
1729 _igvn.rehash_node_delayed(use);
1730 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1731 // back up iterator
1732 --i;
1733 }
1734 }
1735 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1736 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1737 if (_memproj_catchall != NULL ) {
1738 if (_memproj_fallthrough == NULL) {
1739 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1740 transform_later(_memproj_fallthrough);
1741 }
1742 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1743 Node *use = _memproj_catchall->fast_out(i);
1744 _igvn.rehash_node_delayed(use);
1745 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1746 // back up iterator
1747 --i;
1748 }
1749 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1750 _igvn.remove_dead_node(_memproj_catchall);
1751 }
1752
1753 // An allocate node has separate i_o projections for the uses on the control
1754 // and i_o paths. Always replace the control i_o projection with result i_o
1755 // otherwise incoming i_o become dead when only a slow call is generated
1756 // (it is different from memory projections where both projections are
1757 // combined in such case).
1758 if (_ioproj_fallthrough != NULL) {
1759 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1760 Node *use = _ioproj_fallthrough->fast_out(i);
1761 _igvn.rehash_node_delayed(use);
1762 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1763 // back up iterator
1764 --i;
1765 }
1766 }
1767 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1768 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1769 if (_ioproj_catchall != NULL ) {
1770 if (_ioproj_fallthrough == NULL) {
1771 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1772 transform_later(_ioproj_fallthrough);
1773 }
1774 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1775 Node *use = _ioproj_catchall->fast_out(i);
1776 _igvn.rehash_node_delayed(use);
1777 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1778 // back up iterator
1779 --i;
1780 }
1781 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1782 _igvn.remove_dead_node(_ioproj_catchall);
1783 }
1784
1785 // if we generated only a slow call, we are done
1786 if (always_slow) {
1787 // Now we can unhook i_o.
1788 if (result_phi_i_o->outcnt() > 1) {
1789 call->set_req(TypeFunc::I_O, top());
1790 } else {
1791 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1792 // Case of new array with negative size known during compilation.
1793 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1794 // following code since call to runtime will throw exception.
1795 // As result there will be no users of i_o after the call.
1796 // Leave i_o attached to this call to avoid problems in preceding graph.
1797 }
1798 return;
1799 }
1800
1801
1802 if (_fallthroughcatchproj != NULL) {
1803 ctrl = _fallthroughcatchproj->clone();
1804 transform_later(ctrl);
1805 _igvn.replace_node(_fallthroughcatchproj, result_region);
1806 } else {
1807 ctrl = top();
1808 }
1809 Node *slow_result;
1810 if (_resproj == NULL) {
1811 // no uses of the allocation result
1812 slow_result = top();
1813 } else {
1814 slow_result = _resproj->clone();
1815 transform_later(slow_result);
1816 _igvn.replace_node(_resproj, result_phi_rawoop);
1817 }
1818
1819 // Plug slow-path into result merge point
1820 result_region ->init_req( slow_result_path, ctrl );
1821 result_phi_rawoop->init_req( slow_result_path, slow_result);
1822 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1823 transform_later(result_region);
1824 transform_later(result_phi_rawoop);
1825 transform_later(result_phi_rawmem);
1826 transform_later(result_phi_i_o);
1827 // This completes all paths into the result merge point
1828 }
1829
1830
1831 // Helper for PhaseMacroExpand::expand_allocate_common.
1832 // Initializes the newly-allocated storage.
1833 Node*
1834 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1835 Node* control, Node* rawmem, Node* object,
1836 Node* klass_node, Node* length,
1837 Node* size_in_bytes) {
1838 InitializeNode* init = alloc->initialization();
1839 // Store the klass & mark bits
1840 Node* mark_node = NULL;
1841 // For now only enable fast locking for non-array types
1842 if (UseBiasedLocking && (length == NULL)) {
1843 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1844 } else {
1845 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1846 }
1847 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1848
1849 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1850 int header_size = alloc->minimum_header_size(); // conservatively small
1851
1852 // Array length
1853 if (length != NULL) { // Arrays need length field
1854 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1855 // conservatively small header size:
1856 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1857 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1858 if (k->is_array_klass()) // we know the exact header size in most cases:
1859 header_size = Klass::layout_helper_header_size(k->layout_helper());
1860 }
1861
1862 // Clear the object body, if necessary.
1863 if (init == NULL) {
1864 // The init has somehow disappeared; be cautious and clear everything.
1865 //
1866 // This can happen if a node is allocated but an uncommon trap occurs
1867 // immediately. In this case, the Initialize gets associated with the
1868 // trap, and may be placed in a different (outer) loop, if the Allocate
1869 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1870 // there can be two Allocates to one Initialize. The answer in all these
1871 // edge cases is safety first. It is always safe to clear immediately
1872 // within an Allocate, and then (maybe or maybe not) clear some more later.
1873 if (!(UseTLAB && ZeroTLAB)) {
1874 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1875 header_size, size_in_bytes,
1876 &_igvn);
1877 }
1878 } else {
1879 if (!init->is_complete()) {
1880 // Try to win by zeroing only what the init does not store.
1881 // We can also try to do some peephole optimizations,
1882 // such as combining some adjacent subword stores.
1883 rawmem = init->complete_stores(control, rawmem, object,
1884 header_size, size_in_bytes, &_igvn);
1885 }
1886 // We have no more use for this link, since the AllocateNode goes away:
1887 init->set_req(InitializeNode::RawAddress, top());
1888 // (If we keep the link, it just confuses the register allocator,
1889 // who thinks he sees a real use of the address by the membar.)
1890 }
1891
1892 return rawmem;
1893 }
1894
1895 // Generate prefetch instructions for next allocations.
1896 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1897 Node*& contended_phi_rawmem,
1898 Node* old_eden_top, Node* new_eden_top,
1899 Node* length) {
1900 enum { fall_in_path = 1, pf_path = 2 };
1901 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1902 // Generate prefetch allocation with watermark check.
1903 // As an allocation hits the watermark, we will prefetch starting
1904 // at a "distance" away from watermark.
1905
1906 Node *pf_region = new RegionNode(3);
1907 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1908 TypeRawPtr::BOTTOM );
1909 // I/O is used for Prefetch
1910 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1911
1912 Node *thread = new ThreadLocalNode();
1913 transform_later(thread);
1914
1915 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1916 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1917 transform_later(eden_pf_adr);
1918
1919 Node *old_pf_wm = new LoadPNode(needgc_false,
1920 contended_phi_rawmem, eden_pf_adr,
1921 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1922 MemNode::unordered);
1923 transform_later(old_pf_wm);
1924
1925 // check against new_eden_top
1926 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1927 transform_later(need_pf_cmp);
1928 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1929 transform_later(need_pf_bol);
1930 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1931 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1932 transform_later(need_pf_iff);
1933
1934 // true node, add prefetchdistance
1935 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1936 transform_later(need_pf_true);
1937
1938 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1939 transform_later(need_pf_false);
1940
1941 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1942 _igvn.MakeConX(AllocatePrefetchDistance) );
1943 transform_later(new_pf_wmt );
1944 new_pf_wmt->set_req(0, need_pf_true);
1945
1946 Node *store_new_wmt = new StorePNode(need_pf_true,
1947 contended_phi_rawmem, eden_pf_adr,
1948 TypeRawPtr::BOTTOM, new_pf_wmt,
1949 MemNode::unordered);
1950 transform_later(store_new_wmt);
1951
1952 // adding prefetches
1953 pf_phi_abio->init_req( fall_in_path, i_o );
1954
1955 Node *prefetch_adr;
1956 Node *prefetch;
1957 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1958 uint step_size = AllocatePrefetchStepSize;
1959 uint distance = 0;
1960
1961 for ( uint i = 0; i < lines; i++ ) {
1962 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1963 _igvn.MakeConX(distance) );
1964 transform_later(prefetch_adr);
1965 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1966 transform_later(prefetch);
1967 distance += step_size;
1968 i_o = prefetch;
1969 }
1970 pf_phi_abio->set_req( pf_path, i_o );
1971
1972 pf_region->init_req( fall_in_path, need_pf_false );
1973 pf_region->init_req( pf_path, need_pf_true );
1974
1975 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1976 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1977
1978 transform_later(pf_region);
1979 transform_later(pf_phi_rawmem);
1980 transform_later(pf_phi_abio);
1981
1982 needgc_false = pf_region;
1983 contended_phi_rawmem = pf_phi_rawmem;
1984 i_o = pf_phi_abio;
1985 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1986 // Insert a prefetch instruction for each allocation.
1987 // This code is used to generate 1 prefetch instruction per cache line.
1988
1989 // Generate several prefetch instructions.
1990 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1991 uint step_size = AllocatePrefetchStepSize;
1992 uint distance = AllocatePrefetchDistance;
1993
1994 // Next cache address.
1995 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1996 _igvn.MakeConX(step_size + distance));
1997 transform_later(cache_adr);
1998 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1999 transform_later(cache_adr);
2000 // Address is aligned to execute prefetch to the beginning of cache line size
2001 // (it is important when BIS instruction is used on SPARC as prefetch).
2002 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2003 cache_adr = new AndXNode(cache_adr, mask);
2004 transform_later(cache_adr);
2005 cache_adr = new CastX2PNode(cache_adr);
2006 transform_later(cache_adr);
2007
2008 // Prefetch
2009 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2010 prefetch->set_req(0, needgc_false);
2011 transform_later(prefetch);
2012 contended_phi_rawmem = prefetch;
2013 Node *prefetch_adr;
2014 distance = step_size;
2015 for ( uint i = 1; i < lines; i++ ) {
2016 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2017 _igvn.MakeConX(distance) );
2018 transform_later(prefetch_adr);
2019 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2020 transform_later(prefetch);
2021 distance += step_size;
2022 contended_phi_rawmem = prefetch;
2023 }
2024 } else if( AllocatePrefetchStyle > 0 ) {
2025 // Insert a prefetch for each allocation only on the fast-path
2026 Node *prefetch_adr;
2027 Node *prefetch;
2028 // Generate several prefetch instructions.
2029 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2030 uint step_size = AllocatePrefetchStepSize;
2031 uint distance = AllocatePrefetchDistance;
2032 for ( uint i = 0; i < lines; i++ ) {
2033 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2034 _igvn.MakeConX(distance) );
2035 transform_later(prefetch_adr);
2036 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2037 // Do not let it float too high, since if eden_top == eden_end,
2038 // both might be null.
2039 if( i == 0 ) { // Set control for first prefetch, next follows it
2040 prefetch->init_req(0, needgc_false);
2041 }
2042 transform_later(prefetch);
2043 distance += step_size;
2044 i_o = prefetch;
2045 }
2046 }
2047 return i_o;
2048 }
2049
2050
2051 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2052 expand_allocate_common(alloc, NULL,
2053 OptoRuntime::new_instance_Type(),
2054 OptoRuntime::new_instance_Java());
2055 }
2056
2057 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2058 Node* length = alloc->in(AllocateNode::ALength);
2059 InitializeNode* init = alloc->initialization();
2060 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2061 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2062 address slow_call_address; // Address of slow call
2063 if (init != NULL && init->is_complete_with_arraycopy() &&
2064 k->is_type_array_klass()) {
2065 // Don't zero type array during slow allocation in VM since
2066 // it will be initialized later by arraycopy in compiled code.
2067 slow_call_address = OptoRuntime::new_array_nozero_Java();
2068 } else {
2069 slow_call_address = OptoRuntime::new_array_Java();
2070 }
2071 expand_allocate_common(alloc, length,
2072 OptoRuntime::new_array_Type(),
2073 slow_call_address);
2074 }
2075
2076 //-------------------mark_eliminated_box----------------------------------
2077 //
2078 // During EA obj may point to several objects but after few ideal graph
2079 // transformations (CCP) it may point to only one non escaping object
2080 // (but still using phi), corresponding locks and unlocks will be marked
2081 // for elimination. Later obj could be replaced with a new node (new phi)
2082 // and which does not have escape information. And later after some graph
2083 // reshape other locks and unlocks (which were not marked for elimination
2084 // before) are connected to this new obj (phi) but they still will not be
2085 // marked for elimination since new obj has no escape information.
2086 // Mark all associated (same box and obj) lock and unlock nodes for
2087 // elimination if some of them marked already.
2088 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2089 if (oldbox->as_BoxLock()->is_eliminated())
2090 return; // This BoxLock node was processed already.
2091
2092 // New implementation (EliminateNestedLocks) has separate BoxLock
2093 // node for each locked region so mark all associated locks/unlocks as
2094 // eliminated even if different objects are referenced in one locked region
2095 // (for example, OSR compilation of nested loop inside locked scope).
2096 if (EliminateNestedLocks ||
2097 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2098 // Box is used only in one lock region. Mark this box as eliminated.
2099 _igvn.hash_delete(oldbox);
2100 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2101 _igvn.hash_insert(oldbox);
2102
2103 for (uint i = 0; i < oldbox->outcnt(); i++) {
2104 Node* u = oldbox->raw_out(i);
2105 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2106 AbstractLockNode* alock = u->as_AbstractLock();
2107 // Check lock's box since box could be referenced by Lock's debug info.
2108 if (alock->box_node() == oldbox) {
2109 // Mark eliminated all related locks and unlocks.
2110 #ifdef ASSERT
2111 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2112 #endif
2113 alock->set_non_esc_obj();
2114 }
2115 }
2116 }
2117 return;
2118 }
2119
2120 // Create new "eliminated" BoxLock node and use it in monitor debug info
2121 // instead of oldbox for the same object.
2122 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2123
2124 // Note: BoxLock node is marked eliminated only here and it is used
2125 // to indicate that all associated lock and unlock nodes are marked
2126 // for elimination.
2127 newbox->set_eliminated();
2128 transform_later(newbox);
2129
2130 // Replace old box node with new box for all users of the same object.
2131 for (uint i = 0; i < oldbox->outcnt();) {
2132 bool next_edge = true;
2133
2134 Node* u = oldbox->raw_out(i);
2135 if (u->is_AbstractLock()) {
2136 AbstractLockNode* alock = u->as_AbstractLock();
2137 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2138 // Replace Box and mark eliminated all related locks and unlocks.
2139 #ifdef ASSERT
2140 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2141 #endif
2142 alock->set_non_esc_obj();
2143 _igvn.rehash_node_delayed(alock);
2144 alock->set_box_node(newbox);
2145 next_edge = false;
2146 }
2147 }
2148 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2149 FastLockNode* flock = u->as_FastLock();
2150 assert(flock->box_node() == oldbox, "sanity");
2151 _igvn.rehash_node_delayed(flock);
2152 flock->set_box_node(newbox);
2153 next_edge = false;
2154 }
2155
2156 // Replace old box in monitor debug info.
2157 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2158 SafePointNode* sfn = u->as_SafePoint();
2159 JVMState* youngest_jvms = sfn->jvms();
2160 int max_depth = youngest_jvms->depth();
2161 for (int depth = 1; depth <= max_depth; depth++) {
2162 JVMState* jvms = youngest_jvms->of_depth(depth);
2163 int num_mon = jvms->nof_monitors();
2164 // Loop over monitors
2165 for (int idx = 0; idx < num_mon; idx++) {
2166 Node* obj_node = sfn->monitor_obj(jvms, idx);
2167 Node* box_node = sfn->monitor_box(jvms, idx);
2168 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2169 int j = jvms->monitor_box_offset(idx);
2170 _igvn.replace_input_of(u, j, newbox);
2171 next_edge = false;
2172 }
2173 }
2174 }
2175 }
2176 if (next_edge) i++;
2177 }
2178 }
2179
2180 //-----------------------mark_eliminated_locking_nodes-----------------------
2181 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2182 if (EliminateNestedLocks) {
2183 if (alock->is_nested()) {
2184 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2185 return;
2186 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2187 // Only Lock node has JVMState needed here.
2188 // Not that preceding claim is documented anywhere else.
2189 if (alock->jvms() != NULL) {
2190 if (alock->as_Lock()->is_nested_lock_region()) {
2191 // Mark eliminated related nested locks and unlocks.
2192 Node* obj = alock->obj_node();
2193 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2194 assert(!box_node->is_eliminated(), "should not be marked yet");
2195 // Note: BoxLock node is marked eliminated only here
2196 // and it is used to indicate that all associated lock
2197 // and unlock nodes are marked for elimination.
2198 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2199 for (uint i = 0; i < box_node->outcnt(); i++) {
2200 Node* u = box_node->raw_out(i);
2201 if (u->is_AbstractLock()) {
2202 alock = u->as_AbstractLock();
2203 if (alock->box_node() == box_node) {
2204 // Verify that this Box is referenced only by related locks.
2205 assert(alock->obj_node()->eqv_uncast(obj), "");
2206 // Mark all related locks and unlocks.
2207 #ifdef ASSERT
2208 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2209 #endif
2210 alock->set_nested();
2211 }
2212 }
2213 }
2214 } else {
2215 #ifdef ASSERT
2216 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2217 if (C->log() != NULL)
2218 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2219 #endif
2220 }
2221 }
2222 return;
2223 }
2224 // Process locks for non escaping object
2225 assert(alock->is_non_esc_obj(), "");
2226 } // EliminateNestedLocks
2227
2228 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2229 // Look for all locks of this object and mark them and
2230 // corresponding BoxLock nodes as eliminated.
2231 Node* obj = alock->obj_node();
2232 for (uint j = 0; j < obj->outcnt(); j++) {
2233 Node* o = obj->raw_out(j);
2234 if (o->is_AbstractLock() &&
2235 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2236 alock = o->as_AbstractLock();
2237 Node* box = alock->box_node();
2238 // Replace old box node with new eliminated box for all users
2239 // of the same object and mark related locks as eliminated.
2240 mark_eliminated_box(box, obj);
2241 }
2242 }
2243 }
2244 }
2245
2246 // we have determined that this lock/unlock can be eliminated, we simply
2247 // eliminate the node without expanding it.
2248 //
2249 // Note: The membar's associated with the lock/unlock are currently not
2250 // eliminated. This should be investigated as a future enhancement.
2251 //
2252 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2253
2254 if (!alock->is_eliminated()) {
2255 return false;
2256 }
2257 #ifdef ASSERT
2258 if (!alock->is_coarsened()) {
2259 // Check that new "eliminated" BoxLock node is created.
2260 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2261 assert(oldbox->is_eliminated(), "should be done already");
2262 }
2263 #endif
2264
2265 alock->log_lock_optimization(C, "eliminate_lock");
2266
2267 #ifndef PRODUCT
2268 if (PrintEliminateLocks) {
2269 if (alock->is_Lock()) {
2270 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2271 } else {
2272 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2273 }
2274 }
2275 #endif
2276
2277 Node* mem = alock->in(TypeFunc::Memory);
2278 Node* ctrl = alock->in(TypeFunc::Control);
2279
2280 extract_call_projections(alock);
2281 // There are 2 projections from the lock. The lock node will
2282 // be deleted when its last use is subsumed below.
2283 assert(alock->outcnt() == 2 &&
2284 _fallthroughproj != NULL &&
2285 _memproj_fallthrough != NULL,
2286 "Unexpected projections from Lock/Unlock");
2287
2288 Node* fallthroughproj = _fallthroughproj;
2289 Node* memproj_fallthrough = _memproj_fallthrough;
2290
2291 // The memory projection from a lock/unlock is RawMem
2292 // The input to a Lock is merged memory, so extract its RawMem input
2293 // (unless the MergeMem has been optimized away.)
2294 if (alock->is_Lock()) {
2295 // Seach for MemBarAcquireLock node and delete it also.
2296 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2297 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2298 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2299 Node* memproj = membar->proj_out(TypeFunc::Memory);
2300 _igvn.replace_node(ctrlproj, fallthroughproj);
2301 _igvn.replace_node(memproj, memproj_fallthrough);
2302
2303 // Delete FastLock node also if this Lock node is unique user
2304 // (a loop peeling may clone a Lock node).
2305 Node* flock = alock->as_Lock()->fastlock_node();
2306 if (flock->outcnt() == 1) {
2307 assert(flock->unique_out() == alock, "sanity");
2308 _igvn.replace_node(flock, top());
2309 }
2310 }
2311
2312 // Seach for MemBarReleaseLock node and delete it also.
2313 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2314 ctrl->in(0)->is_MemBar()) {
2315 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2316 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2317 mem->is_Proj() && membar == mem->in(0), "");
2318 _igvn.replace_node(fallthroughproj, ctrl);
2319 _igvn.replace_node(memproj_fallthrough, mem);
2320 fallthroughproj = ctrl;
2321 memproj_fallthrough = mem;
2322 ctrl = membar->in(TypeFunc::Control);
2323 mem = membar->in(TypeFunc::Memory);
2324 }
2325
2326 _igvn.replace_node(fallthroughproj, ctrl);
2327 _igvn.replace_node(memproj_fallthrough, mem);
2328 return true;
2329 }
2330
2331
2332 //------------------------------expand_lock_node----------------------
2333 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2334
2335 Node* ctrl = lock->in(TypeFunc::Control);
2336 Node* mem = lock->in(TypeFunc::Memory);
2337 Node* obj = lock->obj_node();
2338 Node* box = lock->box_node();
2339 Node* flock = lock->fastlock_node();
2340
2341 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2342
2343 // Make the merge point
2344 Node *region;
2345 Node *mem_phi;
2346 Node *slow_path;
2347
2348 if (UseOptoBiasInlining) {
2349 /*
2350 * See the full description in MacroAssembler::biased_locking_enter().
2351 *
2352 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2353 * // The object is biased.
2354 * proto_node = klass->prototype_header;
2355 * o_node = thread | proto_node;
2356 * x_node = o_node ^ mark_word;
2357 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2358 * // Done.
2359 * } else {
2360 * if( (x_node & biased_lock_mask) != 0 ) {
2361 * // The klass's prototype header is no longer biased.
2362 * cas(&mark_word, mark_word, proto_node)
2363 * goto cas_lock;
2364 * } else {
2365 * // The klass's prototype header is still biased.
2366 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2367 * old = mark_word;
2368 * new = o_node;
2369 * } else {
2370 * // Different thread or anonymous biased.
2371 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2372 * new = thread | old;
2373 * }
2374 * // Try to rebias.
2375 * if( cas(&mark_word, old, new) == 0 ) {
2376 * // Done.
2377 * } else {
2378 * goto slow_path; // Failed.
2379 * }
2380 * }
2381 * }
2382 * } else {
2383 * // The object is not biased.
2384 * cas_lock:
2385 * if( FastLock(obj) == 0 ) {
2386 * // Done.
2387 * } else {
2388 * slow_path:
2389 * OptoRuntime::complete_monitor_locking_Java(obj);
2390 * }
2391 * }
2392 */
2393
2394 region = new RegionNode(5);
2395 // create a Phi for the memory state
2396 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2397
2398 Node* fast_lock_region = new RegionNode(3);
2399 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2400
2401 // First, check mark word for the biased lock pattern.
2402 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2403
2404 // Get fast path - mark word has the biased lock pattern.
2405 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2406 markOopDesc::biased_lock_mask_in_place,
2407 markOopDesc::biased_lock_pattern, true);
2408 // fast_lock_region->in(1) is set to slow path.
2409 fast_lock_mem_phi->init_req(1, mem);
2410
2411 // Now check that the lock is biased to the current thread and has
2412 // the same epoch and bias as Klass::_prototype_header.
2413
2414 // Special-case a fresh allocation to avoid building nodes:
2415 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2416 if (klass_node == NULL) {
2417 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2418 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2419 #ifdef _LP64
2420 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2421 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2422 klass_node->in(1)->init_req(0, ctrl);
2423 } else
2424 #endif
2425 klass_node->init_req(0, ctrl);
2426 }
2427 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2428
2429 Node* thread = transform_later(new ThreadLocalNode());
2430 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2431 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2432 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2433
2434 // Get slow path - mark word does NOT match the value.
2435 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2436 (~markOopDesc::age_mask_in_place), 0);
2437 // region->in(3) is set to fast path - the object is biased to the current thread.
2438 mem_phi->init_req(3, mem);
2439
2440
2441 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2442
2443
2444 // First, check biased pattern.
2445 // Get fast path - _prototype_header has the same biased lock pattern.
2446 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2447 markOopDesc::biased_lock_mask_in_place, 0, true);
2448
2449 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2450 // fast_lock_region->in(2) - the prototype header is no longer biased
2451 // and we have to revoke the bias on this object.
2452 // We are going to try to reset the mark of this object to the prototype
2453 // value and fall through to the CAS-based locking scheme.
2454 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2455 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2456 proto_node, mark_node);
2457 transform_later(cas);
2458 Node* proj = transform_later(new SCMemProjNode(cas));
2459 fast_lock_mem_phi->init_req(2, proj);
2460
2461
2462 // Second, check epoch bits.
2463 Node* rebiased_region = new RegionNode(3);
2464 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2465 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2466
2467 // Get slow path - mark word does NOT match epoch bits.
2468 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2469 markOopDesc::epoch_mask_in_place, 0);
2470 // The epoch of the current bias is not valid, attempt to rebias the object
2471 // toward the current thread.
2472 rebiased_region->init_req(2, epoch_ctrl);
2473 old_phi->init_req(2, mark_node);
2474 new_phi->init_req(2, o_node);
2475
2476 // rebiased_region->in(1) is set to fast path.
2477 // The epoch of the current bias is still valid but we know
2478 // nothing about the owner; it might be set or it might be clear.
2479 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2480 markOopDesc::age_mask_in_place |
2481 markOopDesc::epoch_mask_in_place);
2482 Node* old = transform_later(new AndXNode(mark_node, cmask));
2483 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2484 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2485 old_phi->init_req(1, old);
2486 new_phi->init_req(1, new_mark);
2487
2488 transform_later(rebiased_region);
2489 transform_later(old_phi);
2490 transform_later(new_phi);
2491
2492 // Try to acquire the bias of the object using an atomic operation.
2493 // If this fails we will go in to the runtime to revoke the object's bias.
2494 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2495 transform_later(cas);
2496 proj = transform_later(new SCMemProjNode(cas));
2497
2498 // Get slow path - Failed to CAS.
2499 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2500 mem_phi->init_req(4, proj);
2501 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2502
2503 // Failed to CAS.
2504 slow_path = new RegionNode(3);
2505 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2506
2507 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2508 slow_mem->init_req(1, proj);
2509
2510 // Call CAS-based locking scheme (FastLock node).
2511
2512 transform_later(fast_lock_region);
2513 transform_later(fast_lock_mem_phi);
2514
2515 // Get slow path - FastLock failed to lock the object.
2516 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2517 mem_phi->init_req(2, fast_lock_mem_phi);
2518 // region->in(2) is set to fast path - the object is locked to the current thread.
2519
2520 slow_path->init_req(2, ctrl); // Capture slow-control
2521 slow_mem->init_req(2, fast_lock_mem_phi);
2522
2523 transform_later(slow_path);
2524 transform_later(slow_mem);
2525 // Reset lock's memory edge.
2526 lock->set_req(TypeFunc::Memory, slow_mem);
2527
2528 } else {
2529 region = new RegionNode(3);
2530 // create a Phi for the memory state
2531 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2532
2533 // Optimize test; set region slot 2
2534 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2535 mem_phi->init_req(2, mem);
2536 }
2537
2538 // Make slow path call
2539 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2540 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2541 obj, box, NULL);
2542
2543 extract_call_projections(call);
2544
2545 // Slow path can only throw asynchronous exceptions, which are always
2546 // de-opted. So the compiler thinks the slow-call can never throw an
2547 // exception. If it DOES throw an exception we would need the debug
2548 // info removed first (since if it throws there is no monitor).
2549 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2550 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2551
2552 // Capture slow path
2553 // disconnect fall-through projection from call and create a new one
2554 // hook up users of fall-through projection to region
2555 Node *slow_ctrl = _fallthroughproj->clone();
2556 transform_later(slow_ctrl);
2557 _igvn.hash_delete(_fallthroughproj);
2558 _fallthroughproj->disconnect_inputs(NULL, C);
2559 region->init_req(1, slow_ctrl);
2560 // region inputs are now complete
2561 transform_later(region);
2562 _igvn.replace_node(_fallthroughproj, region);
2563
2564 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2565 mem_phi->init_req(1, memproj );
2566 transform_later(mem_phi);
2567 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2568 }
2569
2570 //------------------------------expand_unlock_node----------------------
2571 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2572
2573 Node* ctrl = unlock->in(TypeFunc::Control);
2574 Node* mem = unlock->in(TypeFunc::Memory);
2575 Node* obj = unlock->obj_node();
2576 Node* box = unlock->box_node();
2577
2578 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2579
2580 // No need for a null check on unlock
2581
2582 // Make the merge point
2583 Node *region;
2584 Node *mem_phi;
2585
2586 if (UseOptoBiasInlining) {
2587 // Check for biased locking unlock case, which is a no-op.
2588 // See the full description in MacroAssembler::biased_locking_exit().
2589 region = new RegionNode(4);
2590 // create a Phi for the memory state
2591 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2592 mem_phi->init_req(3, mem);
2593
2594 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2595 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2596 markOopDesc::biased_lock_mask_in_place,
2597 markOopDesc::biased_lock_pattern);
2598 } else {
2599 region = new RegionNode(3);
2600 // create a Phi for the memory state
2601 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2602 }
2603
2604 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2605 funlock = transform_later( funlock )->as_FastUnlock();
2606 // Optimize test; set region slot 2
2607 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2608 Node *thread = transform_later(new ThreadLocalNode());
2609
2610 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2611 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2612 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2613
2614 extract_call_projections(call);
2615
2616 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2617 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2618
2619 // No exceptions for unlocking
2620 // Capture slow path
2621 // disconnect fall-through projection from call and create a new one
2622 // hook up users of fall-through projection to region
2623 Node *slow_ctrl = _fallthroughproj->clone();
2624 transform_later(slow_ctrl);
2625 _igvn.hash_delete(_fallthroughproj);
2626 _fallthroughproj->disconnect_inputs(NULL, C);
2627 region->init_req(1, slow_ctrl);
2628 // region inputs are now complete
2629 transform_later(region);
2630 _igvn.replace_node(_fallthroughproj, region);
2631
2632 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2633 mem_phi->init_req(1, memproj );
2634 mem_phi->init_req(2, mem);
2635 transform_later(mem_phi);
2636 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2637 }
2638
2639 //---------------------------eliminate_macro_nodes----------------------
2640 // Eliminate scalar replaced allocations and associated locks.
2641 void PhaseMacroExpand::eliminate_macro_nodes() {
2642 if (C->macro_count() == 0)
2643 return;
2644
2645 // First, attempt to eliminate locks
2646 int cnt = C->macro_count();
2647 for (int i=0; i < cnt; i++) {
2648 Node *n = C->macro_node(i);
2649 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2650 // Before elimination mark all associated (same box and obj)
2651 // lock and unlock nodes.
2652 mark_eliminated_locking_nodes(n->as_AbstractLock());
2653 }
2654 }
2655 bool progress = true;
2656 while (progress) {
2657 progress = false;
2658 for (int i = C->macro_count(); i > 0; i--) {
2659 Node * n = C->macro_node(i-1);
2660 bool success = false;
2661 debug_only(int old_macro_count = C->macro_count(););
2662 if (n->is_AbstractLock()) {
2663 success = eliminate_locking_node(n->as_AbstractLock());
2664 }
2665 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2666 progress = progress || success;
2667 }
2668 }
2669 // Next, attempt to eliminate allocations
2670 _has_locks = false;
2671 progress = true;
2672 while (progress) {
2673 progress = false;
2674 for (int i = C->macro_count(); i > 0; i--) {
2675 Node * n = C->macro_node(i-1);
2676 bool success = false;
2677 debug_only(int old_macro_count = C->macro_count(););
2678 switch (n->class_id()) {
2679 case Node::Class_Allocate:
2680 case Node::Class_AllocateArray:
2681 success = eliminate_allocate_node(n->as_Allocate());
2682 break;
2683 case Node::Class_CallStaticJava:
2684 success = eliminate_boxing_node(n->as_CallStaticJava());
2685 break;
2686 case Node::Class_Lock:
2687 case Node::Class_Unlock:
2688 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2689 _has_locks = true;
2690 break;
2691 case Node::Class_ArrayCopy:
2692 break;
2693 default:
2694 assert(n->Opcode() == Op_LoopLimit ||
2695 n->Opcode() == Op_Opaque1 ||
2696 n->Opcode() == Op_Opaque2 ||
2697 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2698 }
2699 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2700 progress = progress || success;
2701 }
2702 }
2703 }
2704
2705 //------------------------------expand_macro_nodes----------------------
2706 // Returns true if a failure occurred.
2707 bool PhaseMacroExpand::expand_macro_nodes() {
2708 // Last attempt to eliminate macro nodes.
2709 eliminate_macro_nodes();
2710
2711 // Make sure expansion will not cause node limit to be exceeded.
2712 // Worst case is a macro node gets expanded into about 200 nodes.
2713 // Allow 50% more for optimization.
2714 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2715 return true;
2716
2717 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2718 bool progress = true;
2719 while (progress) {
2720 progress = false;
2721 for (int i = C->macro_count(); i > 0; i--) {
2722 Node * n = C->macro_node(i-1);
2723 bool success = false;
2724 debug_only(int old_macro_count = C->macro_count(););
2725 if (n->Opcode() == Op_LoopLimit) {
2726 // Remove it from macro list and put on IGVN worklist to optimize.
2727 C->remove_macro_node(n);
2728 _igvn._worklist.push(n);
2729 success = true;
2730 } else if (n->Opcode() == Op_CallStaticJava) {
2731 // Remove it from macro list and put on IGVN worklist to optimize.
2732 C->remove_macro_node(n);
2733 _igvn._worklist.push(n);
2734 success = true;
2735 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2736 _igvn.replace_node(n, n->in(1));
2737 success = true;
2738 #if INCLUDE_RTM_OPT
2739 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2740 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2741 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2742 Node* cmp = n->unique_out();
2743 #ifdef ASSERT
2744 // Validate graph.
2745 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2746 BoolNode* bol = cmp->unique_out()->as_Bool();
2747 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2748 (bol->_test._test == BoolTest::ne), "");
2749 IfNode* ifn = bol->unique_out()->as_If();
2750 assert((ifn->outcnt() == 2) &&
2751 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2752 #endif
2753 Node* repl = n->in(1);
2754 if (!_has_locks) {
2755 // Remove RTM state check if there are no locks in the code.
2756 // Replace input to compare the same value.
2757 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2758 }
2759 _igvn.replace_node(n, repl);
2760 success = true;
2761 #endif
2762 }
2763 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2764 progress = progress || success;
2765 }
2766 }
2767
2768 // expand arraycopy "macro" nodes first
2769 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2770 // before the allocate nodes are expanded.
2771 int macro_idx = C->macro_count() - 1;
2772 while (macro_idx >= 0) {
2773 Node * n = C->macro_node(macro_idx);
2774 assert(n->is_macro(), "only macro nodes expected here");
2775 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2776 // node is unreachable, so don't try to expand it
2777 C->remove_macro_node(n);
2778 } else if (n->is_ArrayCopy()){
2779 int macro_count = C->macro_count();
2780 expand_arraycopy_node(n->as_ArrayCopy());
2781 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2782 }
2783 if (C->failing()) return true;
2784 macro_idx --;
2785 }
2786
2787 // expand "macro" nodes
2788 // nodes are removed from the macro list as they are processed
2789 while (C->macro_count() > 0) {
2790 int macro_count = C->macro_count();
2791 Node * n = C->macro_node(macro_count-1);
2792 assert(n->is_macro(), "only macro nodes expected here");
2793 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2794 // node is unreachable, so don't try to expand it
2795 C->remove_macro_node(n);
2796 continue;
2797 }
2798 switch (n->class_id()) {
2799 case Node::Class_Allocate:
2800 expand_allocate(n->as_Allocate());
2801 break;
2802 case Node::Class_AllocateArray:
2803 expand_allocate_array(n->as_AllocateArray());
2804 break;
2805 case Node::Class_Lock:
2806 expand_lock_node(n->as_Lock());
2807 break;
2808 case Node::Class_Unlock:
2809 expand_unlock_node(n->as_Unlock());
2810 break;
2811 default:
2812 assert(false, "unknown node type in macro list");
2813 }
2814 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2815 if (C->failing()) return true;
2816 }
2817
2818 _igvn.set_delay_transform(false);
2819 _igvn.optimize();
2820 if (C->failing()) return true;
2821 return false;
2822 }