1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "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 "runtime/sharedRuntime.hpp"
49
50
51 //
52 // Replace any references to "oldref" in inputs to "use" with "newref".
53 // Returns the number of replacements made.
54 //
55 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
56 int nreplacements = 0;
57 uint req = use->req();
58 for (uint j = 0; j < use->len(); j++) {
59 Node *uin = use->in(j);
60 if (uin == oldref) {
61 if (j < req)
62 use->set_req(j, newref);
63 else
64 use->set_prec(j, newref);
65 nreplacements++;
66 } else if (j >= req && uin == NULL) {
67 break;
68 }
69 }
70 return nreplacements;
71 }
72
73 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
74 // Copy debug information and adjust JVMState information
75 uint old_dbg_start = oldcall->tf()->domain()->cnt();
76 uint new_dbg_start = newcall->tf()->domain()->cnt();
77 int jvms_adj = new_dbg_start - old_dbg_start;
78 assert (new_dbg_start == newcall->req(), "argument count mismatch");
79
80 // SafePointScalarObject node could be referenced several times in debug info.
81 // Use Dict to record cloned nodes.
82 Dict* sosn_map = new Dict(cmpkey,hashkey);
83 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
84 Node* old_in = oldcall->in(i);
85 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
86 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
87 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
88 uint old_unique = C->unique();
89 Node* new_in = old_sosn->clone(sosn_map);
90 if (old_unique != C->unique()) { // New node?
91 new_in->set_req(0, C->root()); // reset control edge
92 new_in = transform_later(new_in); // Register new node.
93 }
94 old_in = new_in;
95 }
96 newcall->add_req(old_in);
97 }
98
99 // JVMS may be shared so clone it before we modify it
100 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
101 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
102 jvms->set_map(newcall);
103 jvms->set_locoff(jvms->locoff()+jvms_adj);
104 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
105 jvms->set_monoff(jvms->monoff()+jvms_adj);
106 jvms->set_scloff(jvms->scloff()+jvms_adj);
107 jvms->set_endoff(jvms->endoff()+jvms_adj);
108 }
109 }
110
111 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
112 Node* cmp;
113 if (mask != 0) {
114 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
115 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
116 } else {
117 cmp = word;
118 }
119 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
120 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
121 transform_later(iff);
122
123 // Fast path taken.
124 Node *fast_taken = transform_later(new IfFalseNode(iff));
125
126 // Fast path not-taken, i.e. slow path
127 Node *slow_taken = transform_later(new IfTrueNode(iff));
128
129 if (return_fast_path) {
130 region->init_req(edge, slow_taken); // Capture slow-control
131 return fast_taken;
132 } else {
133 region->init_req(edge, fast_taken); // Capture fast-control
134 return slow_taken;
135 }
136 }
137
138 //--------------------copy_predefined_input_for_runtime_call--------------------
139 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
140 // Set fixed predefined input arguments
141 call->init_req( TypeFunc::Control, ctrl );
142 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
143 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
144 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
145 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
146 }
147
148 //------------------------------make_slow_call---------------------------------
149 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
150 address slow_call, const char* leaf_name, Node* slow_path,
151 Node* parm0, Node* parm1, Node* parm2) {
152
153 // Slow-path call
154 CallNode *call = leaf_name
155 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
156 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
157
158 // Slow path call has no side-effects, uses few values
159 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
160 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
161 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
162 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
163 copy_call_debug_info(oldcall, call);
164 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
165 _igvn.replace_node(oldcall, call);
166 transform_later(call);
167
168 return call;
169 }
170
171 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
172 _fallthroughproj = NULL;
173 _fallthroughcatchproj = NULL;
174 _ioproj_fallthrough = NULL;
175 _ioproj_catchall = NULL;
176 _catchallcatchproj = NULL;
177 _memproj_fallthrough = NULL;
178 _memproj_catchall = NULL;
179 _resproj = NULL;
180 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
181 ProjNode *pn = call->fast_out(i)->as_Proj();
182 switch (pn->_con) {
183 case TypeFunc::Control:
184 {
185 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
186 _fallthroughproj = pn;
187 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
188 const Node *cn = pn->fast_out(j);
189 if (cn->is_Catch()) {
190 ProjNode *cpn = NULL;
191 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
192 cpn = cn->fast_out(k)->as_Proj();
193 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
194 if (cpn->_con == CatchProjNode::fall_through_index)
195 _fallthroughcatchproj = cpn;
196 else {
197 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
198 _catchallcatchproj = cpn;
199 }
200 }
201 }
202 break;
203 }
204 case TypeFunc::I_O:
205 if (pn->_is_io_use)
206 _ioproj_catchall = pn;
207 else
208 _ioproj_fallthrough = pn;
209 break;
210 case TypeFunc::Memory:
211 if (pn->_is_io_use)
212 _memproj_catchall = pn;
213 else
214 _memproj_fallthrough = pn;
215 break;
216 case TypeFunc::Parms:
217 _resproj = pn;
218 break;
219 default:
220 assert(false, "unexpected projection from allocation node.");
221 }
222 }
223
224 }
225
226 // Eliminate a card mark sequence. p2x is a ConvP2XNode
227 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
228 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
229 if (!UseG1GC) {
230 // vanilla/CMS post barrier
231 Node *shift = p2x->unique_out();
232 Node *addp = shift->unique_out();
233 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
234 Node *mem = addp->last_out(j);
235 if (UseCondCardMark && mem->is_Load()) {
236 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
237 // The load is checking if the card has been written so
238 // replace it with zero to fold the test.
239 _igvn.replace_node(mem, intcon(0));
240 continue;
241 }
242 assert(mem->is_Store(), "store required");
243 _igvn.replace_node(mem, mem->in(MemNode::Memory));
244 }
245 } else {
246 // G1 pre/post barriers
247 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
248 // It could be only one user, URShift node, in Object.clone() intrinsic
249 // but the new allocation is passed to arraycopy stub and it could not
250 // be scalar replaced. So we don't check the case.
251
252 // An other case of only one user (Xor) is when the value check for NULL
253 // in G1 post barrier is folded after CCP so the code which used URShift
254 // is removed.
255
256 // Take Region node before eliminating post barrier since it also
257 // eliminates CastP2X node when it has only one user.
258 Node* this_region = p2x->in(0);
259 assert(this_region != NULL, "");
260
261 // Remove G1 post barrier.
262
263 // Search for CastP2X->Xor->URShift->Cmp path which
264 // checks if the store done to a different from the value's region.
265 // And replace Cmp with #0 (false) to collapse G1 post barrier.
266 Node* xorx = p2x->find_out_with(Op_XorX);
267 if (xorx != NULL) {
268 Node* shift = xorx->unique_out();
269 Node* cmpx = shift->unique_out();
270 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
271 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
272 "missing region check in G1 post barrier");
273 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
274
275 // Remove G1 pre barrier.
276
277 // Search "if (marking != 0)" check and set it to "false".
278 // There is no G1 pre barrier if previous stored value is NULL
279 // (for example, after initialization).
280 if (this_region->is_Region() && this_region->req() == 3) {
281 int ind = 1;
282 if (!this_region->in(ind)->is_IfFalse()) {
283 ind = 2;
284 }
285 if (this_region->in(ind)->is_IfFalse() &&
286 this_region->in(ind)->in(0)->Opcode() == Op_If) {
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_or_null(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 int instance_id = adr_t->instance_id();
575 assert((uint)instance_id == alloc->_idx, "wrong allocation");
576
577 int alias_idx = C->get_alias_index(adr_t);
578 int offset = adr_t->offset();
579 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
580 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
581 Node *alloc_mem = alloc->in(TypeFunc::Memory);
582 Arena *a = Thread::current()->resource_area();
583 VectorSet visited(a);
584
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 go 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 go wrong.
669 return NULL;
670 }
671
672 // Check the possibility of scalar replacement.
673 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
674 // Scan the uses of the allocation to check for anything that would
675 // prevent us from eliminating it.
676 NOT_PRODUCT( const char* fail_eliminate = NULL; )
677 DEBUG_ONLY( Node* disq_node = NULL; )
678 bool can_eliminate = true;
679
680 Node* res = alloc->result_cast();
681 const TypeOopPtr* res_type = NULL;
682 if (res == NULL) {
683 // All users were eliminated.
684 } else if (!res->is_CheckCastPP()) {
685 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
686 can_eliminate = false;
687 } else {
688 res_type = _igvn.type(res)->isa_oopptr();
689 if (res_type == NULL) {
690 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
691 can_eliminate = false;
692 } else if (res_type->isa_aryptr()) {
693 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
694 if (length < 0) {
695 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
696 can_eliminate = false;
697 }
698 }
699 }
700
701 if (can_eliminate && res != NULL) {
702 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
703 j < jmax && can_eliminate; j++) {
704 Node* use = res->fast_out(j);
705
706 if (use->is_AddP()) {
707 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
708 int offset = addp_type->offset();
709
710 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
711 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
712 can_eliminate = false;
713 break;
714 }
715 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
716 k < kmax && can_eliminate; k++) {
717 Node* n = use->fast_out(k);
718 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
719 !(n->is_ArrayCopy() &&
720 n->as_ArrayCopy()->is_clonebasic() &&
721 n->in(ArrayCopyNode::Dest) == use)) {
722 DEBUG_ONLY(disq_node = n;)
723 if (n->is_Load() || n->is_LoadStore()) {
724 NOT_PRODUCT(fail_eliminate = "Field load";)
725 } else {
726 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
727 }
728 can_eliminate = false;
729 }
730 }
731 } else if (use->is_ArrayCopy() &&
732 (use->as_ArrayCopy()->is_arraycopy_validated() ||
733 use->as_ArrayCopy()->is_copyof_validated() ||
734 use->as_ArrayCopy()->is_copyofrange_validated()) &&
735 use->in(ArrayCopyNode::Dest) == res) {
736 // ok to eliminate
737 } else if (use->is_SafePoint()) {
738 SafePointNode* sfpt = use->as_SafePoint();
739 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
740 // Object is passed as argument.
741 DEBUG_ONLY(disq_node = use;)
742 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
743 can_eliminate = false;
744 }
745 Node* sfptMem = sfpt->memory();
746 if (sfptMem == NULL || sfptMem->is_top()) {
747 DEBUG_ONLY(disq_node = use;)
748 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
749 can_eliminate = false;
750 } else {
751 safepoints.append_if_missing(sfpt);
752 }
753 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
754 if (use->is_Phi()) {
755 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
756 NOT_PRODUCT(fail_eliminate = "Object is return value";)
757 } else {
758 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
759 }
760 DEBUG_ONLY(disq_node = use;)
761 } else {
762 if (use->Opcode() == Op_Return) {
763 NOT_PRODUCT(fail_eliminate = "Object is return value";)
764 }else {
765 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
766 }
767 DEBUG_ONLY(disq_node = use;)
768 }
769 can_eliminate = false;
770 } else {
771 assert(use->Opcode() == Op_CastP2X, "should be");
772 if (use->has_out_with(Op_OrL)) {
773 // Fixme
774 // This is not a card mark!
775 can_eliminate = false;
776 }
777 }
778 }
779 }
780
781 #ifndef PRODUCT
782 if (PrintEliminateAllocations) {
783 if (can_eliminate) {
784 tty->print("Scalar ");
785 if (res == NULL)
786 alloc->dump();
787 else
788 res->dump();
789 } else if (alloc->_is_scalar_replaceable) {
790 tty->print("NotScalar (%s)", fail_eliminate);
791 if (res == NULL)
792 alloc->dump();
793 else
794 res->dump();
795 #ifdef ASSERT
796 if (disq_node != NULL) {
797 tty->print(" >>>> ");
798 disq_node->dump();
799 }
800 #endif /*ASSERT*/
801 }
802 }
803 #endif
804 return can_eliminate;
805 }
806
807 // Do scalar replacement.
808 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
809 GrowableArray <SafePointNode *> safepoints_done;
810
811 ciKlass* klass = NULL;
812 ciInstanceKlass* iklass = NULL;
813 int nfields = 0;
814 int array_base = 0;
815 int element_size = 0;
816 BasicType basic_elem_type = T_ILLEGAL;
817 ciType* elem_type = NULL;
818
819 Node* res = alloc->result_cast();
820 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
821 const TypeOopPtr* res_type = NULL;
822 if (res != NULL) { // Could be NULL when there are no users
823 res_type = _igvn.type(res)->isa_oopptr();
824 }
825
826 if (res != NULL) {
827 klass = res_type->klass();
828 if (res_type->isa_instptr()) {
829 // find the fields of the class which will be needed for safepoint debug information
830 assert(klass->is_instance_klass(), "must be an instance klass.");
831 iklass = klass->as_instance_klass();
832 nfields = iklass->nof_nonstatic_fields();
833 } else {
834 // find the array's elements which will be needed for safepoint debug information
835 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
836 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
837 elem_type = klass->as_array_klass()->element_type();
838 basic_elem_type = elem_type->basic_type();
839 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
840 element_size = type2aelembytes(basic_elem_type);
841 }
842 }
843 //
844 // Process the safepoint uses
845 //
846 while (safepoints.length() > 0) {
847 SafePointNode* sfpt = safepoints.pop();
848 Node* mem = sfpt->memory();
849 Node* ctl = sfpt->control();
850 assert(sfpt->jvms() != NULL, "missed JVMS");
851 // Fields of scalar objs are referenced only at the end
852 // of regular debuginfo at the last (youngest) JVMS.
853 // Record relative start index.
854 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
855 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
856 #ifdef ASSERT
857 alloc,
858 #endif
859 first_ind, nfields);
860 sobj->init_req(0, C->root());
861 transform_later(sobj);
862
863 // Scan object's fields adding an input to the safepoint for each field.
864 for (int j = 0; j < nfields; j++) {
865 intptr_t offset;
866 ciField* field = NULL;
867 if (iklass != NULL) {
868 field = iklass->nonstatic_field_at(j);
869 offset = field->offset();
870 elem_type = field->type();
871 basic_elem_type = field->layout_type();
872 } else {
873 offset = array_base + j * (intptr_t)element_size;
874 }
875
876 const Type *field_type;
877 // The next code is taken from Parse::do_get_xxx().
878 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
879 if (!elem_type->is_loaded()) {
880 field_type = TypeInstPtr::BOTTOM;
881 } else if (field != NULL && field->is_static_constant()) {
882 // This can happen if the constant oop is non-perm.
883 ciObject* con = field->constant_value().as_object();
884 // Do not "join" in the previous type; it doesn't add value,
885 // and may yield a vacuous result if the field is of interface type.
886 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
887 assert(field_type != NULL, "field singleton type must be consistent");
888 } else {
889 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
890 }
891 if (UseCompressedOops) {
892 field_type = field_type->make_narrowoop();
893 basic_elem_type = T_NARROWOOP;
894 }
895 } else {
896 field_type = Type::get_const_basic_type(basic_elem_type);
897 }
898
899 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
900
901 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
902 if (field_val == NULL) {
903 // We weren't able to find a value for this field,
904 // give up on eliminating this allocation.
905
906 // Remove any extra entries we added to the safepoint.
907 uint last = sfpt->req() - 1;
908 for (int k = 0; k < j; k++) {
909 sfpt->del_req(last--);
910 }
911 _igvn._worklist.push(sfpt);
912 // rollback processed safepoints
913 while (safepoints_done.length() > 0) {
914 SafePointNode* sfpt_done = safepoints_done.pop();
915 // remove any extra entries we added to the safepoint
916 last = sfpt_done->req() - 1;
917 for (int k = 0; k < nfields; k++) {
918 sfpt_done->del_req(last--);
919 }
920 JVMState *jvms = sfpt_done->jvms();
921 jvms->set_endoff(sfpt_done->req());
922 // Now make a pass over the debug information replacing any references
923 // to SafePointScalarObjectNode with the allocated object.
924 int start = jvms->debug_start();
925 int end = jvms->debug_end();
926 for (int i = start; i < end; i++) {
927 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
928 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
929 if (scobj->first_index(jvms) == sfpt_done->req() &&
930 scobj->n_fields() == (uint)nfields) {
931 assert(scobj->alloc() == alloc, "sanity");
932 sfpt_done->set_req(i, res);
933 }
934 }
935 }
936 _igvn._worklist.push(sfpt_done);
937 }
938 #ifndef PRODUCT
939 if (PrintEliminateAllocations) {
940 if (field != NULL) {
941 tty->print("=== At SafePoint node %d can't find value of Field: ",
942 sfpt->_idx);
943 field->print();
944 int field_idx = C->get_alias_index(field_addr_type);
945 tty->print(" (alias_idx=%d)", field_idx);
946 } else { // Array's element
947 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
948 sfpt->_idx, j);
949 }
950 tty->print(", which prevents elimination of: ");
951 if (res == NULL)
952 alloc->dump();
953 else
954 res->dump();
955 }
956 #endif
957 return false;
958 }
959 if (UseCompressedOops && field_type->isa_narrowoop()) {
960 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
961 // to be able scalar replace the allocation.
962 if (field_val->is_EncodeP()) {
963 field_val = field_val->in(1);
964 } else {
965 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
966 }
967 }
968 sfpt->add_req(field_val);
969 }
970 JVMState *jvms = sfpt->jvms();
971 jvms->set_endoff(sfpt->req());
972 // Now make a pass over the debug information replacing any references
973 // to the allocated object with "sobj"
974 int start = jvms->debug_start();
975 int end = jvms->debug_end();
976 sfpt->replace_edges_in_range(res, sobj, start, end);
977 _igvn._worklist.push(sfpt);
978 safepoints_done.append_if_missing(sfpt); // keep it for rollback
979 }
980 return true;
981 }
982
983 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
984 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
985 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
986 if (ctl_proj != NULL) {
987 igvn.replace_node(ctl_proj, n->in(0));
988 }
989 if (mem_proj != NULL) {
990 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
991 }
992 }
993
994 // Process users of eliminated allocation.
995 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
996 Node* res = alloc->result_cast();
997 if (res != NULL) {
998 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
999 Node *use = res->last_out(j);
1000 uint oc1 = res->outcnt();
1001
1002 if (use->is_AddP()) {
1003 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1004 Node *n = use->last_out(k);
1005 uint oc2 = use->outcnt();
1006 if (n->is_Store()) {
1007 #ifdef ASSERT
1008 // Verify that there is no dependent MemBarVolatile nodes,
1009 // they should be removed during IGVN, see MemBarNode::Ideal().
1010 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1011 p < pmax; p++) {
1012 Node* mb = n->fast_out(p);
1013 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1014 mb->req() <= MemBarNode::Precedent ||
1015 mb->in(MemBarNode::Precedent) != n,
1016 "MemBarVolatile should be eliminated for non-escaping object");
1017 }
1018 #endif
1019 _igvn.replace_node(n, n->in(MemNode::Memory));
1020 } else if (n->is_ArrayCopy()) {
1021 // Disconnect ArrayCopy node
1022 ArrayCopyNode* ac = n->as_ArrayCopy();
1023 assert(ac->is_clonebasic(), "unexpected array copy kind");
1024 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1025 disconnect_projections(ac, _igvn);
1026 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1027 Node* membar_before = alloc->in(0)->in(0);
1028 disconnect_projections(membar_before->as_MemBar(), _igvn);
1029 if (membar_after->is_MemBar()) {
1030 disconnect_projections(membar_after->as_MemBar(), _igvn);
1031 }
1032 } else {
1033 eliminate_card_mark(n);
1034 }
1035 k -= (oc2 - use->outcnt());
1036 }
1037 } else if (use->is_ArrayCopy()) {
1038 // Disconnect ArrayCopy node
1039 ArrayCopyNode* ac = use->as_ArrayCopy();
1040 assert(ac->is_arraycopy_validated() ||
1041 ac->is_copyof_validated() ||
1042 ac->is_copyofrange_validated(), "unsupported");
1043 CallProjections callprojs;
1044 ac->extract_projections(&callprojs, true);
1045
1046 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1047 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1048 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1049
1050 // Set control to top. IGVN will remove the remaining projections
1051 ac->set_req(0, top());
1052 ac->replace_edge(res, top());
1053
1054 // Disconnect src right away: it can help find new
1055 // opportunities for allocation elimination
1056 Node* src = ac->in(ArrayCopyNode::Src);
1057 ac->replace_edge(src, top());
1058 // src can be top at this point if src and dest of the
1059 // arraycopy were the same
1060 if (src->outcnt() == 0 && !src->is_top()) {
1061 _igvn.remove_dead_node(src);
1062 }
1063
1064 _igvn._worklist.push(ac);
1065 } else {
1066 eliminate_card_mark(use);
1067 }
1068 j -= (oc1 - res->outcnt());
1069 }
1070 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1071 _igvn.remove_dead_node(res);
1072 }
1073
1074 //
1075 // Process other users of allocation's projections
1076 //
1077 if (_resproj != NULL && _resproj->outcnt() != 0) {
1078 // First disconnect stores captured by Initialize node.
1079 // If Initialize node is eliminated first in the following code,
1080 // it will kill such stores and DUIterator_Last will assert.
1081 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1082 Node *use = _resproj->fast_out(j);
1083 if (use->is_AddP()) {
1084 // raw memory addresses used only by the initialization
1085 _igvn.replace_node(use, C->top());
1086 --j; --jmax;
1087 }
1088 }
1089 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1090 Node *use = _resproj->last_out(j);
1091 uint oc1 = _resproj->outcnt();
1092 if (use->is_Initialize()) {
1093 // Eliminate Initialize node.
1094 InitializeNode *init = use->as_Initialize();
1095 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1096 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1097 if (ctrl_proj != NULL) {
1098 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1099 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1100 }
1101 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1102 if (mem_proj != NULL) {
1103 Node *mem = init->in(TypeFunc::Memory);
1104 #ifdef ASSERT
1105 if (mem->is_MergeMem()) {
1106 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1107 } else {
1108 assert(mem == _memproj_fallthrough, "allocation memory projection");
1109 }
1110 #endif
1111 _igvn.replace_node(mem_proj, mem);
1112 }
1113 } else {
1114 assert(false, "only Initialize or AddP expected");
1115 }
1116 j -= (oc1 - _resproj->outcnt());
1117 }
1118 }
1119 if (_fallthroughcatchproj != NULL) {
1120 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1121 }
1122 if (_memproj_fallthrough != NULL) {
1123 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1124 }
1125 if (_memproj_catchall != NULL) {
1126 _igvn.replace_node(_memproj_catchall, C->top());
1127 }
1128 if (_ioproj_fallthrough != NULL) {
1129 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1130 }
1131 if (_ioproj_catchall != NULL) {
1132 _igvn.replace_node(_ioproj_catchall, C->top());
1133 }
1134 if (_catchallcatchproj != NULL) {
1135 _igvn.replace_node(_catchallcatchproj, C->top());
1136 }
1137 }
1138
1139 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1140 // Don't do scalar replacement if the frame can be popped by JVMTI:
1141 // if reallocation fails during deoptimization we'll pop all
1142 // interpreter frames for this compiled frame and that won't play
1143 // nice with JVMTI popframe.
1144 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1145 return false;
1146 }
1147 Node* klass = alloc->in(AllocateNode::KlassNode);
1148 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1149 Node* res = alloc->result_cast();
1150 // Eliminate boxing allocations which are not used
1151 // regardless scalar replacable status.
1152 bool boxing_alloc = C->eliminate_boxing() &&
1153 tklass->klass()->is_instance_klass() &&
1154 tklass->klass()->as_instance_klass()->is_box_klass();
1155 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1156 return false;
1157 }
1158
1159 extract_call_projections(alloc);
1160
1161 GrowableArray <SafePointNode *> safepoints;
1162 if (!can_eliminate_allocation(alloc, safepoints)) {
1163 return false;
1164 }
1165
1166 if (!alloc->_is_scalar_replaceable) {
1167 assert(res == NULL, "sanity");
1168 // We can only eliminate allocation if all debug info references
1169 // are already replaced with SafePointScalarObject because
1170 // we can't search for a fields value without instance_id.
1171 if (safepoints.length() > 0) {
1172 return false;
1173 }
1174 }
1175
1176 if (!scalar_replacement(alloc, safepoints)) {
1177 return false;
1178 }
1179
1180 CompileLog* log = C->log();
1181 if (log != NULL) {
1182 log->head("eliminate_allocation type='%d'",
1183 log->identify(tklass->klass()));
1184 JVMState* p = alloc->jvms();
1185 while (p != NULL) {
1186 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1187 p = p->caller();
1188 }
1189 log->tail("eliminate_allocation");
1190 }
1191
1192 process_users_of_allocation(alloc);
1193
1194 #ifndef PRODUCT
1195 if (PrintEliminateAllocations) {
1196 if (alloc->is_AllocateArray())
1197 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1198 else
1199 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1200 }
1201 #endif
1202
1203 return true;
1204 }
1205
1206 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1207 // EA should remove all uses of non-escaping boxing node.
1208 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1209 return false;
1210 }
1211
1212 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1213
1214 extract_call_projections(boxing);
1215
1216 const TypeTuple* r = boxing->tf()->range();
1217 assert(r->cnt() > TypeFunc::Parms, "sanity");
1218 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1219 assert(t != NULL, "sanity");
1220
1221 CompileLog* log = C->log();
1222 if (log != NULL) {
1223 log->head("eliminate_boxing type='%d'",
1224 log->identify(t->klass()));
1225 JVMState* p = boxing->jvms();
1226 while (p != NULL) {
1227 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1228 p = p->caller();
1229 }
1230 log->tail("eliminate_boxing");
1231 }
1232
1233 process_users_of_allocation(boxing);
1234
1235 #ifndef PRODUCT
1236 if (PrintEliminateAllocations) {
1237 tty->print("++++ Eliminated: %d ", boxing->_idx);
1238 boxing->method()->print_short_name(tty);
1239 tty->cr();
1240 }
1241 #endif
1242
1243 return true;
1244 }
1245
1246 //---------------------------set_eden_pointers-------------------------
1247 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1248 if (UseTLAB) { // Private allocation: load from TLS
1249 Node* thread = transform_later(new ThreadLocalNode());
1250 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1251 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1252 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1253 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1254 } else { // Shared allocation: load from globals
1255 CollectedHeap* ch = Universe::heap();
1256 address top_adr = (address)ch->top_addr();
1257 address end_adr = (address)ch->end_addr();
1258 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1259 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1260 }
1261 }
1262
1263
1264 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1265 Node* adr = basic_plus_adr(base, offset);
1266 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1267 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1268 transform_later(value);
1269 return value;
1270 }
1271
1272
1273 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1274 Node* adr = basic_plus_adr(base, offset);
1275 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1276 transform_later(mem);
1277 return mem;
1278 }
1279
1280 //=============================================================================
1281 //
1282 // A L L O C A T I O N
1283 //
1284 // Allocation attempts to be fast in the case of frequent small objects.
1285 // It breaks down like this:
1286 //
1287 // 1) Size in doublewords is computed. This is a constant for objects and
1288 // variable for most arrays. Doubleword units are used to avoid size
1289 // overflow of huge doubleword arrays. We need doublewords in the end for
1290 // rounding.
1291 //
1292 // 2) Size is checked for being 'too large'. Too-large allocations will go
1293 // the slow path into the VM. The slow path can throw any required
1294 // exceptions, and does all the special checks for very large arrays. The
1295 // size test can constant-fold away for objects. For objects with
1296 // finalizers it constant-folds the otherway: you always go slow with
1297 // finalizers.
1298 //
1299 // 3) If NOT using TLABs, this is the contended loop-back point.
1300 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1301 //
1302 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1303 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1304 // "size*8" we always enter the VM, where "largish" is a constant picked small
1305 // enough that there's always space between the eden max and 4Gig (old space is
1306 // there so it's quite large) and large enough that the cost of entering the VM
1307 // is dwarfed by the cost to initialize the space.
1308 //
1309 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1310 // down. If contended, repeat at step 3. If using TLABs normal-store
1311 // adjusted heap top back down; there is no contention.
1312 //
1313 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1314 // fields.
1315 //
1316 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1317 // oop flavor.
1318 //
1319 //=============================================================================
1320 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1321 // Allocations bigger than this always go the slow route.
1322 // This value must be small enough that allocation attempts that need to
1323 // trigger exceptions go the slow route. Also, it must be small enough so
1324 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1325 //=============================================================================j//
1326 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1327 // The allocator will coalesce int->oop copies away. See comment in
1328 // coalesce.cpp about how this works. It depends critically on the exact
1329 // code shape produced here, so if you are changing this code shape
1330 // make sure the GC info for the heap-top is correct in and around the
1331 // slow-path call.
1332 //
1333
1334 void PhaseMacroExpand::expand_allocate_common(
1335 AllocateNode* alloc, // allocation node to be expanded
1336 Node* length, // array length for an array allocation
1337 const TypeFunc* slow_call_type, // Type of slow call
1338 address slow_call_address // Address of slow call
1339 )
1340 {
1341
1342 Node* ctrl = alloc->in(TypeFunc::Control);
1343 Node* mem = alloc->in(TypeFunc::Memory);
1344 Node* i_o = alloc->in(TypeFunc::I_O);
1345 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1346 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1347 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1348
1349 assert(ctrl != NULL, "must have control");
1350 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1351 // they will not be used if "always_slow" is set
1352 enum { slow_result_path = 1, fast_result_path = 2 };
1353 Node *result_region = NULL;
1354 Node *result_phi_rawmem = NULL;
1355 Node *result_phi_rawoop = NULL;
1356 Node *result_phi_i_o = NULL;
1357
1358 // The initial slow comparison is a size check, the comparison
1359 // we want to do is a BoolTest::gt
1360 bool always_slow = false;
1361 int tv = _igvn.find_int_con(initial_slow_test, -1);
1362 if (tv >= 0) {
1363 always_slow = (tv == 1);
1364 initial_slow_test = NULL;
1365 } else {
1366 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1367 }
1368
1369 if (C->env()->dtrace_alloc_probes() ||
1370 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1371 // Force slow-path allocation
1372 always_slow = true;
1373 initial_slow_test = NULL;
1374 }
1375
1376
1377 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1378 Node *slow_region = NULL;
1379 Node *toobig_false = ctrl;
1380
1381 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1382 // generate the initial test if necessary
1383 if (initial_slow_test != NULL ) {
1384 slow_region = new RegionNode(3);
1385
1386 // Now make the initial failure test. Usually a too-big test but
1387 // might be a TRUE for finalizers or a fancy class check for
1388 // newInstance0.
1389 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1390 transform_later(toobig_iff);
1391 // Plug the failing-too-big test into the slow-path region
1392 Node *toobig_true = new IfTrueNode( toobig_iff );
1393 transform_later(toobig_true);
1394 slow_region ->init_req( too_big_or_final_path, toobig_true );
1395 toobig_false = new IfFalseNode( toobig_iff );
1396 transform_later(toobig_false);
1397 } else { // No initial test, just fall into next case
1398 toobig_false = ctrl;
1399 debug_only(slow_region = NodeSentinel);
1400 }
1401
1402 Node *slow_mem = mem; // save the current memory state for slow path
1403 // generate the fast allocation code unless we know that the initial test will always go slow
1404 if (!always_slow) {
1405 // Fast path modifies only raw memory.
1406 if (mem->is_MergeMem()) {
1407 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1408 }
1409
1410 Node* eden_top_adr;
1411 Node* eden_end_adr;
1412
1413 set_eden_pointers(eden_top_adr, eden_end_adr);
1414
1415 // Load Eden::end. Loop invariant and hoisted.
1416 //
1417 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1418 // a TLAB to work around a bug where these values were being moved across
1419 // a safepoint. These are not oops, so they cannot be include in the oop
1420 // map, but they can be changed by a GC. The proper way to fix this would
1421 // be to set the raw memory state when generating a SafepointNode. However
1422 // this will require extensive changes to the loop optimization in order to
1423 // prevent a degradation of the optimization.
1424 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1425 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1426
1427 // allocate the Region and Phi nodes for the result
1428 result_region = new RegionNode(3);
1429 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1430 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1431 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1432
1433 // We need a Region for the loop-back contended case.
1434 enum { fall_in_path = 1, contended_loopback_path = 2 };
1435 Node *contended_region;
1436 Node *contended_phi_rawmem;
1437 if (UseTLAB) {
1438 contended_region = toobig_false;
1439 contended_phi_rawmem = mem;
1440 } else {
1441 contended_region = new RegionNode(3);
1442 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1443 // Now handle the passing-too-big test. We fall into the contended
1444 // loop-back merge point.
1445 contended_region ->init_req(fall_in_path, toobig_false);
1446 contended_phi_rawmem->init_req(fall_in_path, mem);
1447 transform_later(contended_region);
1448 transform_later(contended_phi_rawmem);
1449 }
1450
1451 // Load(-locked) the heap top.
1452 // See note above concerning the control input when using a TLAB
1453 Node *old_eden_top = UseTLAB
1454 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1455 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1456
1457 transform_later(old_eden_top);
1458 // Add to heap top to get a new heap top
1459 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1460 transform_later(new_eden_top);
1461 // Check for needing a GC; compare against heap end
1462 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1463 transform_later(needgc_cmp);
1464 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1465 transform_later(needgc_bol);
1466 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1467 transform_later(needgc_iff);
1468
1469 // Plug the failing-heap-space-need-gc test into the slow-path region
1470 Node *needgc_true = new IfTrueNode(needgc_iff);
1471 transform_later(needgc_true);
1472 if (initial_slow_test) {
1473 slow_region->init_req(need_gc_path, needgc_true);
1474 // This completes all paths into the slow merge point
1475 transform_later(slow_region);
1476 } else { // No initial slow path needed!
1477 // Just fall from the need-GC path straight into the VM call.
1478 slow_region = needgc_true;
1479 }
1480 // No need for a GC. Setup for the Store-Conditional
1481 Node *needgc_false = new IfFalseNode(needgc_iff);
1482 transform_later(needgc_false);
1483
1484 // Grab regular I/O before optional prefetch may change it.
1485 // Slow-path does no I/O so just set it to the original I/O.
1486 result_phi_i_o->init_req(slow_result_path, i_o);
1487
1488 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1489 old_eden_top, new_eden_top, length);
1490
1491 // Name successful fast-path variables
1492 Node* fast_oop = old_eden_top;
1493 Node* fast_oop_ctrl;
1494 Node* fast_oop_rawmem;
1495
1496 // Store (-conditional) the modified eden top back down.
1497 // StorePConditional produces flags for a test PLUS a modified raw
1498 // memory state.
1499 if (UseTLAB) {
1500 Node* store_eden_top =
1501 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1502 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1503 transform_later(store_eden_top);
1504 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1505 fast_oop_rawmem = store_eden_top;
1506 } else {
1507 Node* store_eden_top =
1508 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1509 new_eden_top, fast_oop/*old_eden_top*/);
1510 transform_later(store_eden_top);
1511 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1512 transform_later(contention_check);
1513 store_eden_top = new SCMemProjNode(store_eden_top);
1514 transform_later(store_eden_top);
1515
1516 // If not using TLABs, check to see if there was contention.
1517 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1518 transform_later(contention_iff);
1519 Node *contention_true = new IfTrueNode(contention_iff);
1520 transform_later(contention_true);
1521 // If contention, loopback and try again.
1522 contended_region->init_req(contended_loopback_path, contention_true);
1523 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1524
1525 // Fast-path succeeded with no contention!
1526 Node *contention_false = new IfFalseNode(contention_iff);
1527 transform_later(contention_false);
1528 fast_oop_ctrl = contention_false;
1529
1530 // Bump total allocated bytes for this thread
1531 Node* thread = new ThreadLocalNode();
1532 transform_later(thread);
1533 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1534 in_bytes(JavaThread::allocated_bytes_offset()));
1535 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1536 0, TypeLong::LONG, T_LONG);
1537 #ifdef _LP64
1538 Node* alloc_size = size_in_bytes;
1539 #else
1540 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1541 transform_later(alloc_size);
1542 #endif
1543 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1544 transform_later(new_alloc_bytes);
1545 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1546 0, new_alloc_bytes, T_LONG);
1547 }
1548
1549 InitializeNode* init = alloc->initialization();
1550 fast_oop_rawmem = initialize_object(alloc,
1551 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1552 klass_node, length, size_in_bytes);
1553
1554 // If initialization is performed by an array copy, any required
1555 // MemBarStoreStore was already added. If the object does not
1556 // escape no need for a MemBarStoreStore. If the object does not
1557 // escape in its initializer and memory barrier (MemBarStoreStore or
1558 // stronger) is already added at exit of initializer, also no need
1559 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1560 // so that stores that initialize this object can't be reordered
1561 // with a subsequent store that makes this object accessible by
1562 // other threads.
1563 // Other threads include java threads and JVM internal threads
1564 // (for example concurrent GC threads). Current concurrent GC
1565 // implementation: CMS and G1 will not scan newly created object,
1566 // so it's safe to skip storestore barrier when allocation does
1567 // not escape.
1568 if (!alloc->does_not_escape_thread() &&
1569 !alloc->is_allocation_MemBar_redundant() &&
1570 (init == NULL || !init->is_complete_with_arraycopy())) {
1571 if (init == NULL || init->req() < InitializeNode::RawStores) {
1572 // No InitializeNode or no stores captured by zeroing
1573 // elimination. Simply add the MemBarStoreStore after object
1574 // initialization.
1575 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1576 transform_later(mb);
1577
1578 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1579 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1580 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1581 transform_later(fast_oop_ctrl);
1582 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1583 transform_later(fast_oop_rawmem);
1584 } else {
1585 // Add the MemBarStoreStore after the InitializeNode so that
1586 // all stores performing the initialization that were moved
1587 // before the InitializeNode happen before the storestore
1588 // barrier.
1589
1590 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1591 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1592
1593 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1594 transform_later(mb);
1595
1596 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1597 transform_later(ctrl);
1598 Node* mem = new ProjNode(init,TypeFunc::Memory);
1599 transform_later(mem);
1600
1601 // The MemBarStoreStore depends on control and memory coming
1602 // from the InitializeNode
1603 mb->init_req(TypeFunc::Memory, mem);
1604 mb->init_req(TypeFunc::Control, ctrl);
1605
1606 ctrl = new ProjNode(mb,TypeFunc::Control);
1607 transform_later(ctrl);
1608 mem = new ProjNode(mb,TypeFunc::Memory);
1609 transform_later(mem);
1610
1611 // All nodes that depended on the InitializeNode for control
1612 // and memory must now depend on the MemBarNode that itself
1613 // depends on the InitializeNode
1614 if (init_ctrl != NULL) {
1615 _igvn.replace_node(init_ctrl, ctrl);
1616 }
1617 if (init_mem != NULL) {
1618 _igvn.replace_node(init_mem, mem);
1619 }
1620 }
1621 }
1622
1623 if (C->env()->dtrace_extended_probes()) {
1624 // Slow-path call
1625 int size = TypeFunc::Parms + 2;
1626 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1627 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1628 "dtrace_object_alloc",
1629 TypeRawPtr::BOTTOM);
1630
1631 // Get base of thread-local storage area
1632 Node* thread = new ThreadLocalNode();
1633 transform_later(thread);
1634
1635 call->init_req(TypeFunc::Parms+0, thread);
1636 call->init_req(TypeFunc::Parms+1, fast_oop);
1637 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1638 call->init_req(TypeFunc::I_O , top()); // does no i/o
1639 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1640 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1641 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1642 transform_later(call);
1643 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1644 transform_later(fast_oop_ctrl);
1645 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1646 transform_later(fast_oop_rawmem);
1647 }
1648
1649 // Plug in the successful fast-path into the result merge point
1650 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1651 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1652 result_phi_i_o ->init_req(fast_result_path, i_o);
1653 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1654 } else {
1655 slow_region = ctrl;
1656 result_phi_i_o = i_o; // Rename it to use in the following code.
1657 }
1658
1659 // Generate slow-path call
1660 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1661 OptoRuntime::stub_name(slow_call_address),
1662 alloc->jvms()->bci(),
1663 TypePtr::BOTTOM);
1664 call->init_req( TypeFunc::Control, slow_region );
1665 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1666 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1667 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1668 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1669
1670 call->init_req(TypeFunc::Parms+0, klass_node);
1671 if (length != NULL) {
1672 call->init_req(TypeFunc::Parms+1, length);
1673 }
1674
1675 // Copy debug information and adjust JVMState information, then replace
1676 // allocate node with the call
1677 copy_call_debug_info((CallNode *) alloc, call);
1678 if (!always_slow) {
1679 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1680 } else {
1681 // Hook i_o projection to avoid its elimination during allocation
1682 // replacement (when only a slow call is generated).
1683 call->set_req(TypeFunc::I_O, result_phi_i_o);
1684 }
1685 _igvn.replace_node(alloc, call);
1686 transform_later(call);
1687
1688 // Identify the output projections from the allocate node and
1689 // adjust any references to them.
1690 // The control and io projections look like:
1691 //
1692 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1693 // Allocate Catch
1694 // ^---Proj(io) <-------+ ^---CatchProj(io)
1695 //
1696 // We are interested in the CatchProj nodes.
1697 //
1698 extract_call_projections(call);
1699
1700 // An allocate node has separate memory projections for the uses on
1701 // the control and i_o paths. Replace the control memory projection with
1702 // result_phi_rawmem (unless we are only generating a slow call when
1703 // both memory projections are combined)
1704 if (!always_slow && _memproj_fallthrough != NULL) {
1705 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1706 Node *use = _memproj_fallthrough->fast_out(i);
1707 _igvn.rehash_node_delayed(use);
1708 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1709 // back up iterator
1710 --i;
1711 }
1712 }
1713 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1714 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1715 if (_memproj_catchall != NULL ) {
1716 if (_memproj_fallthrough == NULL) {
1717 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1718 transform_later(_memproj_fallthrough);
1719 }
1720 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1721 Node *use = _memproj_catchall->fast_out(i);
1722 _igvn.rehash_node_delayed(use);
1723 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1724 // back up iterator
1725 --i;
1726 }
1727 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1728 _igvn.remove_dead_node(_memproj_catchall);
1729 }
1730
1731 // An allocate node has separate i_o projections for the uses on the control
1732 // and i_o paths. Always replace the control i_o projection with result i_o
1733 // otherwise incoming i_o become dead when only a slow call is generated
1734 // (it is different from memory projections where both projections are
1735 // combined in such case).
1736 if (_ioproj_fallthrough != NULL) {
1737 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1738 Node *use = _ioproj_fallthrough->fast_out(i);
1739 _igvn.rehash_node_delayed(use);
1740 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1741 // back up iterator
1742 --i;
1743 }
1744 }
1745 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1746 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1747 if (_ioproj_catchall != NULL ) {
1748 if (_ioproj_fallthrough == NULL) {
1749 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1750 transform_later(_ioproj_fallthrough);
1751 }
1752 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1753 Node *use = _ioproj_catchall->fast_out(i);
1754 _igvn.rehash_node_delayed(use);
1755 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1756 // back up iterator
1757 --i;
1758 }
1759 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1760 _igvn.remove_dead_node(_ioproj_catchall);
1761 }
1762
1763 // if we generated only a slow call, we are done
1764 if (always_slow) {
1765 // Now we can unhook i_o.
1766 if (result_phi_i_o->outcnt() > 1) {
1767 call->set_req(TypeFunc::I_O, top());
1768 } else {
1769 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1770 // Case of new array with negative size known during compilation.
1771 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1772 // following code since call to runtime will throw exception.
1773 // As result there will be no users of i_o after the call.
1774 // Leave i_o attached to this call to avoid problems in preceding graph.
1775 }
1776 return;
1777 }
1778
1779
1780 if (_fallthroughcatchproj != NULL) {
1781 ctrl = _fallthroughcatchproj->clone();
1782 transform_later(ctrl);
1783 _igvn.replace_node(_fallthroughcatchproj, result_region);
1784 } else {
1785 ctrl = top();
1786 }
1787 Node *slow_result;
1788 if (_resproj == NULL) {
1789 // no uses of the allocation result
1790 slow_result = top();
1791 } else {
1792 slow_result = _resproj->clone();
1793 transform_later(slow_result);
1794 _igvn.replace_node(_resproj, result_phi_rawoop);
1795 }
1796
1797 // Plug slow-path into result merge point
1798 result_region ->init_req( slow_result_path, ctrl );
1799 result_phi_rawoop->init_req( slow_result_path, slow_result);
1800 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1801 transform_later(result_region);
1802 transform_later(result_phi_rawoop);
1803 transform_later(result_phi_rawmem);
1804 transform_later(result_phi_i_o);
1805 // This completes all paths into the result merge point
1806 }
1807
1808
1809 // Helper for PhaseMacroExpand::expand_allocate_common.
1810 // Initializes the newly-allocated storage.
1811 Node*
1812 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1813 Node* control, Node* rawmem, Node* object,
1814 Node* klass_node, Node* length,
1815 Node* size_in_bytes) {
1816 InitializeNode* init = alloc->initialization();
1817 // Store the klass & mark bits
1818 Node* mark_node = NULL;
1819 // For now only enable fast locking for non-array types
1820 if (UseBiasedLocking && (length == NULL)) {
1821 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1822 } else {
1823 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1824 }
1825 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1826
1827 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1828 int header_size = alloc->minimum_header_size(); // conservatively small
1829
1830 // Array length
1831 if (length != NULL) { // Arrays need length field
1832 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1833 // conservatively small header size:
1834 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1835 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1836 if (k->is_array_klass()) // we know the exact header size in most cases:
1837 header_size = Klass::layout_helper_header_size(k->layout_helper());
1838 }
1839
1840 // Clear the object body, if necessary.
1841 if (init == NULL) {
1842 // The init has somehow disappeared; be cautious and clear everything.
1843 //
1844 // This can happen if a node is allocated but an uncommon trap occurs
1845 // immediately. In this case, the Initialize gets associated with the
1846 // trap, and may be placed in a different (outer) loop, if the Allocate
1847 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1848 // there can be two Allocates to one Initialize. The answer in all these
1849 // edge cases is safety first. It is always safe to clear immediately
1850 // within an Allocate, and then (maybe or maybe not) clear some more later.
1851 if (!(UseTLAB && ZeroTLAB)) {
1852 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1853 header_size, size_in_bytes,
1854 &_igvn);
1855 }
1856 } else {
1857 if (!init->is_complete()) {
1858 // Try to win by zeroing only what the init does not store.
1859 // We can also try to do some peephole optimizations,
1860 // such as combining some adjacent subword stores.
1861 rawmem = init->complete_stores(control, rawmem, object,
1862 header_size, size_in_bytes, &_igvn);
1863 }
1864 // We have no more use for this link, since the AllocateNode goes away:
1865 init->set_req(InitializeNode::RawAddress, top());
1866 // (If we keep the link, it just confuses the register allocator,
1867 // who thinks he sees a real use of the address by the membar.)
1868 }
1869
1870 return rawmem;
1871 }
1872
1873 // Generate prefetch instructions for next allocations.
1874 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1875 Node*& contended_phi_rawmem,
1876 Node* old_eden_top, Node* new_eden_top,
1877 Node* length) {
1878 enum { fall_in_path = 1, pf_path = 2 };
1879 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1880 // Generate prefetch allocation with watermark check.
1881 // As an allocation hits the watermark, we will prefetch starting
1882 // at a "distance" away from watermark.
1883
1884 Node *pf_region = new RegionNode(3);
1885 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1886 TypeRawPtr::BOTTOM );
1887 // I/O is used for Prefetch
1888 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1889
1890 Node *thread = new ThreadLocalNode();
1891 transform_later(thread);
1892
1893 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1894 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1895 transform_later(eden_pf_adr);
1896
1897 Node *old_pf_wm = new LoadPNode(needgc_false,
1898 contended_phi_rawmem, eden_pf_adr,
1899 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1900 MemNode::unordered);
1901 transform_later(old_pf_wm);
1902
1903 // check against new_eden_top
1904 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1905 transform_later(need_pf_cmp);
1906 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1907 transform_later(need_pf_bol);
1908 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1909 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1910 transform_later(need_pf_iff);
1911
1912 // true node, add prefetchdistance
1913 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1914 transform_later(need_pf_true);
1915
1916 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1917 transform_later(need_pf_false);
1918
1919 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1920 _igvn.MakeConX(AllocatePrefetchDistance) );
1921 transform_later(new_pf_wmt );
1922 new_pf_wmt->set_req(0, need_pf_true);
1923
1924 Node *store_new_wmt = new StorePNode(need_pf_true,
1925 contended_phi_rawmem, eden_pf_adr,
1926 TypeRawPtr::BOTTOM, new_pf_wmt,
1927 MemNode::unordered);
1928 transform_later(store_new_wmt);
1929
1930 // adding prefetches
1931 pf_phi_abio->init_req( fall_in_path, i_o );
1932
1933 Node *prefetch_adr;
1934 Node *prefetch;
1935 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1936 uint step_size = AllocatePrefetchStepSize;
1937 uint distance = 0;
1938
1939 for ( uint i = 0; i < lines; i++ ) {
1940 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1941 _igvn.MakeConX(distance) );
1942 transform_later(prefetch_adr);
1943 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1944 transform_later(prefetch);
1945 distance += step_size;
1946 i_o = prefetch;
1947 }
1948 pf_phi_abio->set_req( pf_path, i_o );
1949
1950 pf_region->init_req( fall_in_path, need_pf_false );
1951 pf_region->init_req( pf_path, need_pf_true );
1952
1953 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1954 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1955
1956 transform_later(pf_region);
1957 transform_later(pf_phi_rawmem);
1958 transform_later(pf_phi_abio);
1959
1960 needgc_false = pf_region;
1961 contended_phi_rawmem = pf_phi_rawmem;
1962 i_o = pf_phi_abio;
1963 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1964 // Insert a prefetch instruction for each allocation.
1965 // This code is used to generate 1 prefetch instruction per cache line.
1966
1967 // Generate several prefetch instructions.
1968 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1969 uint step_size = AllocatePrefetchStepSize;
1970 uint distance = AllocatePrefetchDistance;
1971
1972 // Next cache address.
1973 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1974 _igvn.MakeConX(step_size + distance));
1975 transform_later(cache_adr);
1976 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1977 transform_later(cache_adr);
1978 // Address is aligned to execute prefetch to the beginning of cache line size
1979 // (it is important when BIS instruction is used on SPARC as prefetch).
1980 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1981 cache_adr = new AndXNode(cache_adr, mask);
1982 transform_later(cache_adr);
1983 cache_adr = new CastX2PNode(cache_adr);
1984 transform_later(cache_adr);
1985
1986 // Prefetch
1987 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1988 prefetch->set_req(0, needgc_false);
1989 transform_later(prefetch);
1990 contended_phi_rawmem = prefetch;
1991 Node *prefetch_adr;
1992 distance = step_size;
1993 for ( uint i = 1; i < lines; i++ ) {
1994 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1995 _igvn.MakeConX(distance) );
1996 transform_later(prefetch_adr);
1997 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1998 transform_later(prefetch);
1999 distance += step_size;
2000 contended_phi_rawmem = prefetch;
2001 }
2002 } else if( AllocatePrefetchStyle > 0 ) {
2003 // Insert a prefetch for each allocation only on the fast-path
2004 Node *prefetch_adr;
2005 Node *prefetch;
2006 // Generate several prefetch instructions.
2007 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2008 uint step_size = AllocatePrefetchStepSize;
2009 uint distance = AllocatePrefetchDistance;
2010 for ( uint i = 0; i < lines; i++ ) {
2011 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2012 _igvn.MakeConX(distance) );
2013 transform_later(prefetch_adr);
2014 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2015 // Do not let it float too high, since if eden_top == eden_end,
2016 // both might be null.
2017 if( i == 0 ) { // Set control for first prefetch, next follows it
2018 prefetch->init_req(0, needgc_false);
2019 }
2020 transform_later(prefetch);
2021 distance += step_size;
2022 i_o = prefetch;
2023 }
2024 }
2025 return i_o;
2026 }
2027
2028
2029 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2030 expand_allocate_common(alloc, NULL,
2031 OptoRuntime::new_instance_Type(),
2032 OptoRuntime::new_instance_Java());
2033 }
2034
2035 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2036 Node* length = alloc->in(AllocateNode::ALength);
2037 InitializeNode* init = alloc->initialization();
2038 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2039 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2040 address slow_call_address; // Address of slow call
2041 if (init != NULL && init->is_complete_with_arraycopy() &&
2042 k->is_type_array_klass()) {
2043 // Don't zero type array during slow allocation in VM since
2044 // it will be initialized later by arraycopy in compiled code.
2045 slow_call_address = OptoRuntime::new_array_nozero_Java();
2046 } else {
2047 slow_call_address = OptoRuntime::new_array_Java();
2048 }
2049 expand_allocate_common(alloc, length,
2050 OptoRuntime::new_array_Type(),
2051 slow_call_address);
2052 }
2053
2054 //-------------------mark_eliminated_box----------------------------------
2055 //
2056 // During EA obj may point to several objects but after few ideal graph
2057 // transformations (CCP) it may point to only one non escaping object
2058 // (but still using phi), corresponding locks and unlocks will be marked
2059 // for elimination. Later obj could be replaced with a new node (new phi)
2060 // and which does not have escape information. And later after some graph
2061 // reshape other locks and unlocks (which were not marked for elimination
2062 // before) are connected to this new obj (phi) but they still will not be
2063 // marked for elimination since new obj has no escape information.
2064 // Mark all associated (same box and obj) lock and unlock nodes for
2065 // elimination if some of them marked already.
2066 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2067 if (oldbox->as_BoxLock()->is_eliminated())
2068 return; // This BoxLock node was processed already.
2069
2070 // New implementation (EliminateNestedLocks) has separate BoxLock
2071 // node for each locked region so mark all associated locks/unlocks as
2072 // eliminated even if different objects are referenced in one locked region
2073 // (for example, OSR compilation of nested loop inside locked scope).
2074 if (EliminateNestedLocks ||
2075 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2076 // Box is used only in one lock region. Mark this box as eliminated.
2077 _igvn.hash_delete(oldbox);
2078 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2079 _igvn.hash_insert(oldbox);
2080
2081 for (uint i = 0; i < oldbox->outcnt(); i++) {
2082 Node* u = oldbox->raw_out(i);
2083 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2084 AbstractLockNode* alock = u->as_AbstractLock();
2085 // Check lock's box since box could be referenced by Lock's debug info.
2086 if (alock->box_node() == oldbox) {
2087 // Mark eliminated all related locks and unlocks.
2088 #ifdef ASSERT
2089 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2090 #endif
2091 alock->set_non_esc_obj();
2092 }
2093 }
2094 }
2095 return;
2096 }
2097
2098 // Create new "eliminated" BoxLock node and use it in monitor debug info
2099 // instead of oldbox for the same object.
2100 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2101
2102 // Note: BoxLock node is marked eliminated only here and it is used
2103 // to indicate that all associated lock and unlock nodes are marked
2104 // for elimination.
2105 newbox->set_eliminated();
2106 transform_later(newbox);
2107
2108 // Replace old box node with new box for all users of the same object.
2109 for (uint i = 0; i < oldbox->outcnt();) {
2110 bool next_edge = true;
2111
2112 Node* u = oldbox->raw_out(i);
2113 if (u->is_AbstractLock()) {
2114 AbstractLockNode* alock = u->as_AbstractLock();
2115 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2116 // Replace Box and mark eliminated all related locks and unlocks.
2117 #ifdef ASSERT
2118 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2119 #endif
2120 alock->set_non_esc_obj();
2121 _igvn.rehash_node_delayed(alock);
2122 alock->set_box_node(newbox);
2123 next_edge = false;
2124 }
2125 }
2126 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2127 FastLockNode* flock = u->as_FastLock();
2128 assert(flock->box_node() == oldbox, "sanity");
2129 _igvn.rehash_node_delayed(flock);
2130 flock->set_box_node(newbox);
2131 next_edge = false;
2132 }
2133
2134 // Replace old box in monitor debug info.
2135 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2136 SafePointNode* sfn = u->as_SafePoint();
2137 JVMState* youngest_jvms = sfn->jvms();
2138 int max_depth = youngest_jvms->depth();
2139 for (int depth = 1; depth <= max_depth; depth++) {
2140 JVMState* jvms = youngest_jvms->of_depth(depth);
2141 int num_mon = jvms->nof_monitors();
2142 // Loop over monitors
2143 for (int idx = 0; idx < num_mon; idx++) {
2144 Node* obj_node = sfn->monitor_obj(jvms, idx);
2145 Node* box_node = sfn->monitor_box(jvms, idx);
2146 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2147 int j = jvms->monitor_box_offset(idx);
2148 _igvn.replace_input_of(u, j, newbox);
2149 next_edge = false;
2150 }
2151 }
2152 }
2153 }
2154 if (next_edge) i++;
2155 }
2156 }
2157
2158 //-----------------------mark_eliminated_locking_nodes-----------------------
2159 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2160 if (EliminateNestedLocks) {
2161 if (alock->is_nested()) {
2162 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2163 return;
2164 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2165 // Only Lock node has JVMState needed here.
2166 // Not that preceding claim is documented anywhere else.
2167 if (alock->jvms() != NULL) {
2168 if (alock->as_Lock()->is_nested_lock_region()) {
2169 // Mark eliminated related nested locks and unlocks.
2170 Node* obj = alock->obj_node();
2171 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2172 assert(!box_node->is_eliminated(), "should not be marked yet");
2173 // Note: BoxLock node is marked eliminated only here
2174 // and it is used to indicate that all associated lock
2175 // and unlock nodes are marked for elimination.
2176 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2177 for (uint i = 0; i < box_node->outcnt(); i++) {
2178 Node* u = box_node->raw_out(i);
2179 if (u->is_AbstractLock()) {
2180 alock = u->as_AbstractLock();
2181 if (alock->box_node() == box_node) {
2182 // Verify that this Box is referenced only by related locks.
2183 assert(alock->obj_node()->eqv_uncast(obj), "");
2184 // Mark all related locks and unlocks.
2185 #ifdef ASSERT
2186 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2187 #endif
2188 alock->set_nested();
2189 }
2190 }
2191 }
2192 } else {
2193 #ifdef ASSERT
2194 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2195 if (C->log() != NULL)
2196 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2197 #endif
2198 }
2199 }
2200 return;
2201 }
2202 // Process locks for non escaping object
2203 assert(alock->is_non_esc_obj(), "");
2204 } // EliminateNestedLocks
2205
2206 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2207 // Look for all locks of this object and mark them and
2208 // corresponding BoxLock nodes as eliminated.
2209 Node* obj = alock->obj_node();
2210 for (uint j = 0; j < obj->outcnt(); j++) {
2211 Node* o = obj->raw_out(j);
2212 if (o->is_AbstractLock() &&
2213 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2214 alock = o->as_AbstractLock();
2215 Node* box = alock->box_node();
2216 // Replace old box node with new eliminated box for all users
2217 // of the same object and mark related locks as eliminated.
2218 mark_eliminated_box(box, obj);
2219 }
2220 }
2221 }
2222 }
2223
2224 // we have determined that this lock/unlock can be eliminated, we simply
2225 // eliminate the node without expanding it.
2226 //
2227 // Note: The membar's associated with the lock/unlock are currently not
2228 // eliminated. This should be investigated as a future enhancement.
2229 //
2230 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2231
2232 if (!alock->is_eliminated()) {
2233 return false;
2234 }
2235 #ifdef ASSERT
2236 if (!alock->is_coarsened()) {
2237 // Check that new "eliminated" BoxLock node is created.
2238 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2239 assert(oldbox->is_eliminated(), "should be done already");
2240 }
2241 #endif
2242
2243 alock->log_lock_optimization(C, "eliminate_lock");
2244
2245 #ifndef PRODUCT
2246 if (PrintEliminateLocks) {
2247 if (alock->is_Lock()) {
2248 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2249 } else {
2250 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2251 }
2252 }
2253 #endif
2254
2255 Node* mem = alock->in(TypeFunc::Memory);
2256 Node* ctrl = alock->in(TypeFunc::Control);
2257
2258 extract_call_projections(alock);
2259 // There are 2 projections from the lock. The lock node will
2260 // be deleted when its last use is subsumed below.
2261 assert(alock->outcnt() == 2 &&
2262 _fallthroughproj != NULL &&
2263 _memproj_fallthrough != NULL,
2264 "Unexpected projections from Lock/Unlock");
2265
2266 Node* fallthroughproj = _fallthroughproj;
2267 Node* memproj_fallthrough = _memproj_fallthrough;
2268
2269 // The memory projection from a lock/unlock is RawMem
2270 // The input to a Lock is merged memory, so extract its RawMem input
2271 // (unless the MergeMem has been optimized away.)
2272 if (alock->is_Lock()) {
2273 // Seach for MemBarAcquireLock node and delete it also.
2274 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2275 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2276 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2277 Node* memproj = membar->proj_out(TypeFunc::Memory);
2278 _igvn.replace_node(ctrlproj, fallthroughproj);
2279 _igvn.replace_node(memproj, memproj_fallthrough);
2280
2281 // Delete FastLock node also if this Lock node is unique user
2282 // (a loop peeling may clone a Lock node).
2283 Node* flock = alock->as_Lock()->fastlock_node();
2284 if (flock->outcnt() == 1) {
2285 assert(flock->unique_out() == alock, "sanity");
2286 _igvn.replace_node(flock, top());
2287 }
2288 }
2289
2290 // Seach for MemBarReleaseLock node and delete it also.
2291 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2292 ctrl->in(0)->is_MemBar()) {
2293 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2294 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2295 mem->is_Proj() && membar == mem->in(0), "");
2296 _igvn.replace_node(fallthroughproj, ctrl);
2297 _igvn.replace_node(memproj_fallthrough, mem);
2298 fallthroughproj = ctrl;
2299 memproj_fallthrough = mem;
2300 ctrl = membar->in(TypeFunc::Control);
2301 mem = membar->in(TypeFunc::Memory);
2302 }
2303
2304 _igvn.replace_node(fallthroughproj, ctrl);
2305 _igvn.replace_node(memproj_fallthrough, mem);
2306 return true;
2307 }
2308
2309
2310 //------------------------------expand_lock_node----------------------
2311 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2312
2313 Node* ctrl = lock->in(TypeFunc::Control);
2314 Node* mem = lock->in(TypeFunc::Memory);
2315 Node* obj = lock->obj_node();
2316 Node* box = lock->box_node();
2317 Node* flock = lock->fastlock_node();
2318
2319 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2320
2321 // Make the merge point
2322 Node *region;
2323 Node *mem_phi;
2324 Node *slow_path;
2325
2326 if (UseOptoBiasInlining) {
2327 /*
2328 * See the full description in MacroAssembler::biased_locking_enter().
2329 *
2330 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2331 * // The object is biased.
2332 * proto_node = klass->prototype_header;
2333 * o_node = thread | proto_node;
2334 * x_node = o_node ^ mark_word;
2335 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2336 * // Done.
2337 * } else {
2338 * if( (x_node & biased_lock_mask) != 0 ) {
2339 * // The klass's prototype header is no longer biased.
2340 * cas(&mark_word, mark_word, proto_node)
2341 * goto cas_lock;
2342 * } else {
2343 * // The klass's prototype header is still biased.
2344 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2345 * old = mark_word;
2346 * new = o_node;
2347 * } else {
2348 * // Different thread or anonymous biased.
2349 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2350 * new = thread | old;
2351 * }
2352 * // Try to rebias.
2353 * if( cas(&mark_word, old, new) == 0 ) {
2354 * // Done.
2355 * } else {
2356 * goto slow_path; // Failed.
2357 * }
2358 * }
2359 * }
2360 * } else {
2361 * // The object is not biased.
2362 * cas_lock:
2363 * if( FastLock(obj) == 0 ) {
2364 * // Done.
2365 * } else {
2366 * slow_path:
2367 * OptoRuntime::complete_monitor_locking_Java(obj);
2368 * }
2369 * }
2370 */
2371
2372 region = new RegionNode(5);
2373 // create a Phi for the memory state
2374 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2375
2376 Node* fast_lock_region = new RegionNode(3);
2377 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2378
2379 // First, check mark word for the biased lock pattern.
2380 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2381
2382 // Get fast path - mark word has the biased lock pattern.
2383 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2384 markOopDesc::biased_lock_mask_in_place,
2385 markOopDesc::biased_lock_pattern, true);
2386 // fast_lock_region->in(1) is set to slow path.
2387 fast_lock_mem_phi->init_req(1, mem);
2388
2389 // Now check that the lock is biased to the current thread and has
2390 // the same epoch and bias as Klass::_prototype_header.
2391
2392 // Special-case a fresh allocation to avoid building nodes:
2393 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2394 if (klass_node == NULL) {
2395 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2396 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2397 #ifdef _LP64
2398 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2399 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2400 klass_node->in(1)->init_req(0, ctrl);
2401 } else
2402 #endif
2403 klass_node->init_req(0, ctrl);
2404 }
2405 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2406
2407 Node* thread = transform_later(new ThreadLocalNode());
2408 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2409 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2410 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2411
2412 // Get slow path - mark word does NOT match the value.
2413 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2414 (~markOopDesc::age_mask_in_place), 0);
2415 // region->in(3) is set to fast path - the object is biased to the current thread.
2416 mem_phi->init_req(3, mem);
2417
2418
2419 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2420
2421
2422 // First, check biased pattern.
2423 // Get fast path - _prototype_header has the same biased lock pattern.
2424 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2425 markOopDesc::biased_lock_mask_in_place, 0, true);
2426
2427 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2428 // fast_lock_region->in(2) - the prototype header is no longer biased
2429 // and we have to revoke the bias on this object.
2430 // We are going to try to reset the mark of this object to the prototype
2431 // value and fall through to the CAS-based locking scheme.
2432 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2433 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2434 proto_node, mark_node);
2435 transform_later(cas);
2436 Node* proj = transform_later(new SCMemProjNode(cas));
2437 fast_lock_mem_phi->init_req(2, proj);
2438
2439
2440 // Second, check epoch bits.
2441 Node* rebiased_region = new RegionNode(3);
2442 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2443 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2444
2445 // Get slow path - mark word does NOT match epoch bits.
2446 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2447 markOopDesc::epoch_mask_in_place, 0);
2448 // The epoch of the current bias is not valid, attempt to rebias the object
2449 // toward the current thread.
2450 rebiased_region->init_req(2, epoch_ctrl);
2451 old_phi->init_req(2, mark_node);
2452 new_phi->init_req(2, o_node);
2453
2454 // rebiased_region->in(1) is set to fast path.
2455 // The epoch of the current bias is still valid but we know
2456 // nothing about the owner; it might be set or it might be clear.
2457 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2458 markOopDesc::age_mask_in_place |
2459 markOopDesc::epoch_mask_in_place);
2460 Node* old = transform_later(new AndXNode(mark_node, cmask));
2461 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2462 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2463 old_phi->init_req(1, old);
2464 new_phi->init_req(1, new_mark);
2465
2466 transform_later(rebiased_region);
2467 transform_later(old_phi);
2468 transform_later(new_phi);
2469
2470 // Try to acquire the bias of the object using an atomic operation.
2471 // If this fails we will go in to the runtime to revoke the object's bias.
2472 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2473 transform_later(cas);
2474 proj = transform_later(new SCMemProjNode(cas));
2475
2476 // Get slow path - Failed to CAS.
2477 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2478 mem_phi->init_req(4, proj);
2479 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2480
2481 // Failed to CAS.
2482 slow_path = new RegionNode(3);
2483 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2484
2485 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2486 slow_mem->init_req(1, proj);
2487
2488 // Call CAS-based locking scheme (FastLock node).
2489
2490 transform_later(fast_lock_region);
2491 transform_later(fast_lock_mem_phi);
2492
2493 // Get slow path - FastLock failed to lock the object.
2494 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2495 mem_phi->init_req(2, fast_lock_mem_phi);
2496 // region->in(2) is set to fast path - the object is locked to the current thread.
2497
2498 slow_path->init_req(2, ctrl); // Capture slow-control
2499 slow_mem->init_req(2, fast_lock_mem_phi);
2500
2501 transform_later(slow_path);
2502 transform_later(slow_mem);
2503 // Reset lock's memory edge.
2504 lock->set_req(TypeFunc::Memory, slow_mem);
2505
2506 } else {
2507 region = new RegionNode(3);
2508 // create a Phi for the memory state
2509 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2510
2511 // Optimize test; set region slot 2
2512 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2513 mem_phi->init_req(2, mem);
2514 }
2515
2516 // Make slow path call
2517 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2518 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2519 obj, box, NULL);
2520
2521 extract_call_projections(call);
2522
2523 // Slow path can only throw asynchronous exceptions, which are always
2524 // de-opted. So the compiler thinks the slow-call can never throw an
2525 // exception. If it DOES throw an exception we would need the debug
2526 // info removed first (since if it throws there is no monitor).
2527 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2528 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2529
2530 // Capture slow path
2531 // disconnect fall-through projection from call and create a new one
2532 // hook up users of fall-through projection to region
2533 Node *slow_ctrl = _fallthroughproj->clone();
2534 transform_later(slow_ctrl);
2535 _igvn.hash_delete(_fallthroughproj);
2536 _fallthroughproj->disconnect_inputs(NULL, C);
2537 region->init_req(1, slow_ctrl);
2538 // region inputs are now complete
2539 transform_later(region);
2540 _igvn.replace_node(_fallthroughproj, region);
2541
2542 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2543 mem_phi->init_req(1, memproj );
2544 transform_later(mem_phi);
2545 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2546 }
2547
2548 //------------------------------expand_unlock_node----------------------
2549 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2550
2551 Node* ctrl = unlock->in(TypeFunc::Control);
2552 Node* mem = unlock->in(TypeFunc::Memory);
2553 Node* obj = unlock->obj_node();
2554 Node* box = unlock->box_node();
2555
2556 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2557
2558 // No need for a null check on unlock
2559
2560 // Make the merge point
2561 Node *region;
2562 Node *mem_phi;
2563
2564 if (UseOptoBiasInlining) {
2565 // Check for biased locking unlock case, which is a no-op.
2566 // See the full description in MacroAssembler::biased_locking_exit().
2567 region = new RegionNode(4);
2568 // create a Phi for the memory state
2569 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2570 mem_phi->init_req(3, mem);
2571
2572 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2573 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2574 markOopDesc::biased_lock_mask_in_place,
2575 markOopDesc::biased_lock_pattern);
2576 } else {
2577 region = new RegionNode(3);
2578 // create a Phi for the memory state
2579 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2580 }
2581
2582 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2583 funlock = transform_later( funlock )->as_FastUnlock();
2584 // Optimize test; set region slot 2
2585 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2586 Node *thread = transform_later(new ThreadLocalNode());
2587
2588 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2589 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2590 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2591
2592 extract_call_projections(call);
2593
2594 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2595 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2596
2597 // No exceptions for unlocking
2598 // Capture slow path
2599 // disconnect fall-through projection from call and create a new one
2600 // hook up users of fall-through projection to region
2601 Node *slow_ctrl = _fallthroughproj->clone();
2602 transform_later(slow_ctrl);
2603 _igvn.hash_delete(_fallthroughproj);
2604 _fallthroughproj->disconnect_inputs(NULL, C);
2605 region->init_req(1, slow_ctrl);
2606 // region inputs are now complete
2607 transform_later(region);
2608 _igvn.replace_node(_fallthroughproj, region);
2609
2610 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2611 mem_phi->init_req(1, memproj );
2612 mem_phi->init_req(2, mem);
2613 transform_later(mem_phi);
2614 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2615 }
2616
2617 //---------------------------eliminate_macro_nodes----------------------
2618 // Eliminate scalar replaced allocations and associated locks.
2619 void PhaseMacroExpand::eliminate_macro_nodes() {
2620 if (C->macro_count() == 0)
2621 return;
2622
2623 // First, attempt to eliminate locks
2624 int cnt = C->macro_count();
2625 for (int i=0; i < cnt; i++) {
2626 Node *n = C->macro_node(i);
2627 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2628 // Before elimination mark all associated (same box and obj)
2629 // lock and unlock nodes.
2630 mark_eliminated_locking_nodes(n->as_AbstractLock());
2631 }
2632 }
2633 bool progress = true;
2634 while (progress) {
2635 progress = false;
2636 for (int i = C->macro_count(); i > 0; i--) {
2637 Node * n = C->macro_node(i-1);
2638 bool success = false;
2639 debug_only(int old_macro_count = C->macro_count(););
2640 if (n->is_AbstractLock()) {
2641 success = eliminate_locking_node(n->as_AbstractLock());
2642 }
2643 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2644 progress = progress || success;
2645 }
2646 }
2647 // Next, attempt to eliminate allocations
2648 _has_locks = false;
2649 progress = true;
2650 while (progress) {
2651 progress = false;
2652 for (int i = C->macro_count(); i > 0; i--) {
2653 Node * n = C->macro_node(i-1);
2654 bool success = false;
2655 debug_only(int old_macro_count = C->macro_count(););
2656 switch (n->class_id()) {
2657 case Node::Class_Allocate:
2658 case Node::Class_AllocateArray:
2659 success = eliminate_allocate_node(n->as_Allocate());
2660 break;
2661 case Node::Class_CallStaticJava:
2662 success = eliminate_boxing_node(n->as_CallStaticJava());
2663 break;
2664 case Node::Class_Lock:
2665 case Node::Class_Unlock:
2666 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2667 _has_locks = true;
2668 break;
2669 case Node::Class_ArrayCopy:
2670 break;
2671 case Node::Class_OuterStripMinedLoop:
2672 break;
2673 default:
2674 assert(n->Opcode() == Op_LoopLimit ||
2675 n->Opcode() == Op_Opaque1 ||
2676 n->Opcode() == Op_Opaque2 ||
2677 n->Opcode() == Op_Opaque3 ||
2678 n->Opcode() == Op_Opaque4, "unknown node type in macro list");
2679 }
2680 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2681 progress = progress || success;
2682 }
2683 }
2684 }
2685
2686 //------------------------------expand_macro_nodes----------------------
2687 // Returns true if a failure occurred.
2688 bool PhaseMacroExpand::expand_macro_nodes() {
2689 // Last attempt to eliminate macro nodes.
2690 eliminate_macro_nodes();
2691
2692 // Make sure expansion will not cause node limit to be exceeded.
2693 // Worst case is a macro node gets expanded into about 200 nodes.
2694 // Allow 50% more for optimization.
2695 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2696 return true;
2697
2698 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2699 bool progress = true;
2700 while (progress) {
2701 progress = false;
2702 for (int i = C->macro_count(); i > 0; i--) {
2703 Node * n = C->macro_node(i-1);
2704 bool success = false;
2705 debug_only(int old_macro_count = C->macro_count(););
2706 if (n->Opcode() == Op_LoopLimit) {
2707 // Remove it from macro list and put on IGVN worklist to optimize.
2708 C->remove_macro_node(n);
2709 _igvn._worklist.push(n);
2710 success = true;
2711 } else if (n->Opcode() == Op_CallStaticJava) {
2712 // Remove it from macro list and put on IGVN worklist to optimize.
2713 C->remove_macro_node(n);
2714 _igvn._worklist.push(n);
2715 success = true;
2716 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2717 _igvn.replace_node(n, n->in(1));
2718 success = true;
2719 #if INCLUDE_RTM_OPT
2720 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2721 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2722 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2723 Node* cmp = n->unique_out();
2724 #ifdef ASSERT
2725 // Validate graph.
2726 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2727 BoolNode* bol = cmp->unique_out()->as_Bool();
2728 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2729 (bol->_test._test == BoolTest::ne), "");
2730 IfNode* ifn = bol->unique_out()->as_If();
2731 assert((ifn->outcnt() == 2) &&
2732 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2733 #endif
2734 Node* repl = n->in(1);
2735 if (!_has_locks) {
2736 // Remove RTM state check if there are no locks in the code.
2737 // Replace input to compare the same value.
2738 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2739 }
2740 _igvn.replace_node(n, repl);
2741 success = true;
2742 #endif
2743 } else if (n->Opcode() == Op_Opaque4) {
2744 _igvn.replace_node(n, n->in(2));
2745 success = true;
2746 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2747 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2748 C->remove_macro_node(n);
2749 success = true;
2750 }
2751 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2752 progress = progress || success;
2753 }
2754 }
2755
2756 // expand arraycopy "macro" nodes first
2757 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2758 // before the allocate nodes are expanded.
2759 int macro_idx = C->macro_count() - 1;
2760 while (macro_idx >= 0) {
2761 Node * n = C->macro_node(macro_idx);
2762 assert(n->is_macro(), "only macro nodes expected here");
2763 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2764 // node is unreachable, so don't try to expand it
2765 C->remove_macro_node(n);
2766 } else if (n->is_ArrayCopy()){
2767 int macro_count = C->macro_count();
2768 expand_arraycopy_node(n->as_ArrayCopy());
2769 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2770 }
2771 if (C->failing()) return true;
2772 macro_idx --;
2773 }
2774
2775 // expand "macro" nodes
2776 // nodes are removed from the macro list as they are processed
2777 while (C->macro_count() > 0) {
2778 int macro_count = C->macro_count();
2779 Node * n = C->macro_node(macro_count-1);
2780 assert(n->is_macro(), "only macro nodes expected here");
2781 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2782 // node is unreachable, so don't try to expand it
2783 C->remove_macro_node(n);
2784 continue;
2785 }
2786 switch (n->class_id()) {
2787 case Node::Class_Allocate:
2788 expand_allocate(n->as_Allocate());
2789 break;
2790 case Node::Class_AllocateArray:
2791 expand_allocate_array(n->as_AllocateArray());
2792 break;
2793 case Node::Class_Lock:
2794 expand_lock_node(n->as_Lock());
2795 break;
2796 case Node::Class_Unlock:
2797 expand_unlock_node(n->as_Unlock());
2798 break;
2799 default:
2800 assert(false, "unknown node type in macro list");
2801 }
2802 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2803 if (C->failing()) return true;
2804 }
2805
2806 _igvn.set_delay_transform(false);
2807 _igvn.optimize();
2808 if (C->failing()) return true;
2809 return false;
2810 }