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