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