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