1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/compile.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/locknode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/macro.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/node.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/type.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #if INCLUDE_ALL_GCS
45 #include "gc_implementation/shenandoah/shenandoahForwarding.hpp"
46 #include "gc_implementation/shenandoah/c2/shenandoahBarrierSetC2.hpp"
47 #include "gc_implementation/shenandoah/c2/shenandoahSupport.hpp"
48 #endif
49
50
51 //
52 // Replace any references to "oldref" in inputs to "use" with "newref".
53 // Returns the number of replacements made.
54 //
55 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
56 int nreplacements = 0;
57 uint req = use->req();
58 for (uint j = 0; j < use->len(); j++) {
59 Node *uin = use->in(j);
60 if (uin == oldref) {
61 if (j < req)
62 use->set_req(j, newref);
63 else
64 use->set_prec(j, newref);
65 nreplacements++;
66 } else if (j >= req && uin == NULL) {
67 break;
68 }
69 }
70 return nreplacements;
71 }
72
73 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
74 // Copy debug information and adjust JVMState information
75 uint old_dbg_start = oldcall->tf()->domain()->cnt();
76 uint new_dbg_start = newcall->tf()->domain()->cnt();
77 int jvms_adj = new_dbg_start - old_dbg_start;
78 assert (new_dbg_start == newcall->req(), "argument count mismatch");
79
80 // SafePointScalarObject node could be referenced several times in debug info.
81 // Use Dict to record cloned nodes.
82 Dict* sosn_map = new Dict(cmpkey,hashkey);
83 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
84 Node* old_in = oldcall->in(i);
85 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
86 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
87 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
88 uint old_unique = C->unique();
89 Node* new_in = old_sosn->clone(sosn_map);
90 if (old_unique != C->unique()) { // New node?
91 new_in->set_req(0, C->root()); // reset control edge
92 new_in = transform_later(new_in); // Register new node.
93 }
94 old_in = new_in;
95 }
96 newcall->add_req(old_in);
97 }
98
99 newcall->set_jvms(oldcall->jvms());
100 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
101 jvms->set_map(newcall);
102 jvms->set_locoff(jvms->locoff()+jvms_adj);
103 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
104 jvms->set_monoff(jvms->monoff()+jvms_adj);
105 jvms->set_scloff(jvms->scloff()+jvms_adj);
106 jvms->set_endoff(jvms->endoff()+jvms_adj);
107 }
108 }
109
110 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
111 Node* cmp;
112 if (mask != 0) {
113 Node* and_node = transform_later(new (C) AndXNode(word, MakeConX(mask)));
114 cmp = transform_later(new (C) CmpXNode(and_node, MakeConX(bits)));
115 } else {
116 cmp = word;
117 }
118 Node* bol = transform_later(new (C) BoolNode(cmp, BoolTest::ne));
119 IfNode* iff = new (C) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
120 transform_later(iff);
121
122 // Fast path taken.
123 Node *fast_taken = transform_later( new (C) IfFalseNode(iff) );
124
125 // Fast path not-taken, i.e. slow path
126 Node *slow_taken = transform_later( new (C) IfTrueNode(iff) );
127
128 if (return_fast_path) {
129 region->init_req(edge, slow_taken); // Capture slow-control
130 return fast_taken;
131 } else {
132 region->init_req(edge, fast_taken); // Capture fast-control
133 return slow_taken;
134 }
135 }
136
137 //--------------------copy_predefined_input_for_runtime_call--------------------
138 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
139 // Set fixed predefined input arguments
140 call->init_req( TypeFunc::Control, ctrl );
141 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
142 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
143 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
144 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
145 }
146
147 //------------------------------make_slow_call---------------------------------
148 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) {
149
150 // Slow-path call
151 CallNode *call = leaf_name
152 ? (CallNode*)new (C) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
153 : (CallNode*)new (C) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
154
155 // Slow path call has no side-effects, uses few values
156 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
157 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
158 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
159 copy_call_debug_info(oldcall, call);
160 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
161 _igvn.replace_node(oldcall, call);
162 transform_later(call);
163
164 return call;
165 }
166
167 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
168 _fallthroughproj = NULL;
169 _fallthroughcatchproj = NULL;
170 _ioproj_fallthrough = NULL;
171 _ioproj_catchall = NULL;
172 _catchallcatchproj = NULL;
173 _memproj_fallthrough = NULL;
174 _memproj_catchall = NULL;
175 _resproj = NULL;
176 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
177 ProjNode *pn = call->fast_out(i)->as_Proj();
178 switch (pn->_con) {
179 case TypeFunc::Control:
180 {
181 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
182 _fallthroughproj = pn;
183 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
184 const Node *cn = pn->fast_out(j);
185 if (cn->is_Catch()) {
186 ProjNode *cpn = NULL;
187 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
188 cpn = cn->fast_out(k)->as_Proj();
189 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
190 if (cpn->_con == CatchProjNode::fall_through_index)
191 _fallthroughcatchproj = cpn;
192 else {
193 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
194 _catchallcatchproj = cpn;
195 }
196 }
197 }
198 break;
199 }
200 case TypeFunc::I_O:
201 if (pn->_is_io_use)
202 _ioproj_catchall = pn;
203 else
204 _ioproj_fallthrough = pn;
205 break;
206 case TypeFunc::Memory:
207 if (pn->_is_io_use)
208 _memproj_catchall = pn;
209 else
210 _memproj_fallthrough = pn;
211 break;
212 case TypeFunc::Parms:
213 _resproj = pn;
214 break;
215 default:
216 assert(false, "unexpected projection from allocation node.");
217 }
218 }
219
220 }
221
222 // Eliminate a card mark sequence. p2x is a ConvP2XNode
223 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
224 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
225 if (!UseG1GC) {
226 // vanilla/CMS post barrier
227 Node *shift = p2x->unique_out();
228 Node *addp = shift->unique_out();
229 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
230 Node *mem = addp->last_out(j);
231 if (UseCondCardMark && mem->is_Load()) {
232 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
233 // The load is checking if the card has been written so
234 // replace it with zero to fold the test.
235 _igvn.replace_node(mem, intcon(0));
236 continue;
237 }
238 assert(mem->is_Store(), "store required");
239 _igvn.replace_node(mem, mem->in(MemNode::Memory));
240 }
241 } else {
242 // G1 pre/post barriers
243 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
244 // It could be only one user, URShift node, in Object.clone() instrinsic
245 // but the new allocation is passed to arraycopy stub and it could not
246 // be scalar replaced. So we don't check the case.
247
248 // An other case of only one user (Xor) is when the value check for NULL
249 // in G1 post barrier is folded after CCP so the code which used URShift
250 // is removed.
251
252 // Take Region node before eliminating post barrier since it also
253 // eliminates CastP2X node when it has only one user.
254 Node* this_region = p2x->in(0);
255 assert(this_region != NULL, "");
256
257 // Remove G1 post barrier.
258
259 // Search for CastP2X->Xor->URShift->Cmp path which
260 // checks if the store done to a different from the value's region.
261 // And replace Cmp with #0 (false) to collapse G1 post barrier.
262 Node* xorx = NULL;
263 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
264 Node* u = p2x->fast_out(i);
265 if (u->Opcode() == Op_XorX) {
266 xorx = u;
267 break;
268 }
269 }
270 assert(xorx != NULL, "missing G1 post barrier");
271 Node* shift = xorx->unique_out();
272 Node* cmpx = shift->unique_out();
273 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
274 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
275 "missing region check in G1 post barrier");
276 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
277
278 // Remove G1 pre barrier.
279
280 // Search "if (marking != 0)" check and set it to "false".
281 // There is no G1 pre barrier if previous stored value is NULL
282 // (for example, after initialization).
283 if (this_region->is_Region() && this_region->req() == 3) {
284 int ind = 1;
285 if (!this_region->in(ind)->is_IfFalse()) {
286 ind = 2;
287 }
288 if (this_region->in(ind)->is_IfFalse()) {
289 Node* bol = this_region->in(ind)->in(0)->in(1);
290 assert(bol->is_Bool(), "");
291 cmpx = bol->in(1);
292 if (bol->as_Bool()->_test._test == BoolTest::ne &&
293 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
294 cmpx->in(1)->is_Load()) {
295 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
296 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
297 PtrQueue::byte_offset_of_active());
298 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
299 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
300 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
301 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
302 }
303 }
304 }
305 }
306 // Now CastP2X can be removed since it is used only on dead path
307 // which currently still alive until igvn optimize it.
308 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
309 _igvn.replace_node(p2x, top());
310 }
311 }
312
313 // Search for a memory operation for the specified memory slice.
314 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
315 Node *orig_mem = mem;
316 Node *alloc_mem = alloc->in(TypeFunc::Memory);
317 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
318 while (true) {
319 if (mem == alloc_mem || mem == start_mem ) {
320 return mem; // hit one of our sentinels
321 } else if (mem->is_MergeMem()) {
322 mem = mem->as_MergeMem()->memory_at(alias_idx);
323 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
324 Node *in = mem->in(0);
325 // we can safely skip over safepoints, calls, locks and membars because we
326 // already know that the object is safe to eliminate.
327 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
328 return in;
329 } else if (in->is_Call()) {
330 CallNode *call = in->as_Call();
331 if (!call->may_modify(tinst, phase)) {
332 mem = call->in(TypeFunc::Memory);
333 }
334 mem = in->in(TypeFunc::Memory);
335 } else if (in->is_MemBar()) {
336 mem = in->in(TypeFunc::Memory);
337 } else {
338 assert(false, "unexpected projection");
339 }
340 } else if (mem->is_Store()) {
341 const TypePtr* atype = mem->as_Store()->adr_type();
342 int adr_idx = Compile::current()->get_alias_index(atype);
343 if (adr_idx == alias_idx) {
344 assert(atype->isa_oopptr(), "address type must be oopptr");
345 int adr_offset = atype->offset();
346 uint adr_iid = atype->is_oopptr()->instance_id();
347 // Array elements references have the same alias_idx
348 // but different offset and different instance_id.
349 if (adr_offset == offset && adr_iid == alloc->_idx)
350 return mem;
351 } else {
352 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
353 }
354 mem = mem->in(MemNode::Memory);
355 } else if (mem->is_ClearArray()) {
356 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
357 // Can not bypass initialization of the instance
358 // we are looking.
359 debug_only(intptr_t offset;)
360 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
361 InitializeNode* init = alloc->as_Allocate()->initialization();
362 // We are looking for stored value, return Initialize node
363 // or memory edge from Allocate node.
364 if (init != NULL)
365 return init;
366 else
367 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
368 }
369 // Otherwise skip it (the call updated 'mem' value).
370 } else if (mem->Opcode() == Op_SCMemProj) {
371 mem = mem->in(0);
372 Node* adr = NULL;
373 if (mem->is_LoadStore()) {
374 adr = mem->in(MemNode::Address);
375 } else {
376 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
377 adr = mem->in(3); // Destination array
378 }
379 const TypePtr* atype = adr->bottom_type()->is_ptr();
380 int adr_idx = Compile::current()->get_alias_index(atype);
381 if (adr_idx == alias_idx) {
382 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
383 return NULL;
384 }
385 mem = mem->in(MemNode::Memory);
386 } else {
387 return mem;
388 }
389 assert(mem != orig_mem, "dead memory loop");
390 }
391 }
392
393 //
394 // Given a Memory Phi, compute a value Phi containing the values from stores
395 // on the input paths.
396 // Note: this function is recursive, its depth is limied by the "level" argument
397 // Returns the computed Phi, or NULL if it cannot compute it.
398 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) {
399 assert(mem->is_Phi(), "sanity");
400 int alias_idx = C->get_alias_index(adr_t);
401 int offset = adr_t->offset();
402 int instance_id = adr_t->instance_id();
403
404 // Check if an appropriate value phi already exists.
405 Node* region = mem->in(0);
406 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
407 Node* phi = region->fast_out(k);
408 if (phi->is_Phi() && phi != mem &&
409 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
410 return phi;
411 }
412 }
413 // Check if an appropriate new value phi already exists.
414 Node* new_phi = value_phis->find(mem->_idx);
415 if (new_phi != NULL)
416 return new_phi;
417
418 if (level <= 0) {
419 return NULL; // Give up: phi tree too deep
420 }
421 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
422 Node *alloc_mem = alloc->in(TypeFunc::Memory);
423
424 uint length = mem->req();
425 GrowableArray <Node *> values(length, length, NULL, false);
426
427 // create a new Phi for the value
428 PhiNode *phi = new (C) PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
429 transform_later(phi);
430 value_phis->push(phi, mem->_idx);
431
432 for (uint j = 1; j < length; j++) {
433 Node *in = mem->in(j);
434 if (in == NULL || in->is_top()) {
435 values.at_put(j, in);
436 } else {
437 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
438 if (val == start_mem || val == alloc_mem) {
439 // hit a sentinel, return appropriate 0 value
440 values.at_put(j, _igvn.zerocon(ft));
441 continue;
442 }
443 if (val->is_Initialize()) {
444 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
445 }
446 if (val == NULL) {
447 return NULL; // can't find a value on this path
448 }
449 if (val == mem) {
450 values.at_put(j, mem);
451 } else if (val->is_Store()) {
452 Node* n = val->in(MemNode::ValueIn);
453 #if INCLUDE_ALL_GCS
454 if (UseShenandoahGC) {
455 n = ShenandoahBarrierSetC2::bsc2()->step_over_gc_barrier(n);
456 }
457 #endif
458 values.at_put(j, n);
459 } else if(val->is_Proj() && val->in(0) == alloc) {
460 values.at_put(j, _igvn.zerocon(ft));
461 } else if (val->is_Phi()) {
462 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
463 if (val == NULL) {
464 return NULL;
465 }
466 values.at_put(j, val);
467 } else if (val->Opcode() == Op_SCMemProj) {
468 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
469 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
470 return NULL;
471 } else {
472 #ifdef ASSERT
473 val->dump();
474 assert(false, "unknown node on this path");
475 #endif
476 return NULL; // unknown node on this path
477 }
478 }
479 }
480 // Set Phi's inputs
481 for (uint j = 1; j < length; j++) {
482 if (values.at(j) == mem) {
483 phi->init_req(j, phi);
484 } else {
485 phi->init_req(j, values.at(j));
486 }
487 }
488 return phi;
489 }
490
491 // Search the last value stored into the object's field.
492 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
493 assert(adr_t->is_known_instance_field(), "instance required");
494 int instance_id = adr_t->instance_id();
495 assert((uint)instance_id == alloc->_idx, "wrong allocation");
496
497 int alias_idx = C->get_alias_index(adr_t);
498 int offset = adr_t->offset();
499 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
500 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
501 Node *alloc_mem = alloc->in(TypeFunc::Memory);
502 Arena *a = Thread::current()->resource_area();
503 VectorSet visited(a);
504
505
506 bool done = sfpt_mem == alloc_mem;
507 Node *mem = sfpt_mem;
508 while (!done) {
509 if (visited.test_set(mem->_idx)) {
510 return NULL; // found a loop, give up
511 }
512 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
513 if (mem == start_mem || mem == alloc_mem) {
514 done = true; // hit a sentinel, return appropriate 0 value
515 } else if (mem->is_Initialize()) {
516 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
517 if (mem == NULL) {
518 done = true; // Something go wrong.
519 } else if (mem->is_Store()) {
520 const TypePtr* atype = mem->as_Store()->adr_type();
521 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
522 done = true;
523 }
524 } else if (mem->is_Store()) {
525 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
526 assert(atype != NULL, "address type must be oopptr");
527 assert(C->get_alias_index(atype) == alias_idx &&
528 atype->is_known_instance_field() && atype->offset() == offset &&
529 atype->instance_id() == instance_id, "store is correct memory slice");
530 done = true;
531 } else if (mem->is_Phi()) {
532 // try to find a phi's unique input
533 Node *unique_input = NULL;
534 Node *top = C->top();
535 for (uint i = 1; i < mem->req(); i++) {
536 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
537 if (n == NULL || n == top || n == mem) {
538 continue;
539 } else if (unique_input == NULL) {
540 unique_input = n;
541 } else if (unique_input != n) {
542 unique_input = top;
543 break;
544 }
545 }
546 if (unique_input != NULL && unique_input != top) {
547 mem = unique_input;
548 } else {
549 done = true;
550 }
551 } else {
552 assert(false, "unexpected node");
553 }
554 }
555 if (mem != NULL) {
556 if (mem == start_mem || mem == alloc_mem) {
557 // hit a sentinel, return appropriate 0 value
558 return _igvn.zerocon(ft);
559 } else if (mem->is_Store()) {
560 Node* n = mem->in(MemNode::ValueIn);
561 #if INCLUDE_ALL_GCS
562 if (UseShenandoahGC) {
563 n = ShenandoahBarrierSetC2::bsc2()->step_over_gc_barrier(n);
564 }
565 #endif
566 return n;
567 } else if (mem->is_Phi()) {
568 // attempt to produce a Phi reflecting the values on the input paths of the Phi
569 Node_Stack value_phis(a, 8);
570 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
571 if (phi != NULL) {
572 return phi;
573 } else {
574 // Kill all new Phis
575 while(value_phis.is_nonempty()) {
576 Node* n = value_phis.node();
577 _igvn.replace_node(n, C->top());
578 value_phis.pop();
579 }
580 }
581 }
582 }
583 // Something go wrong.
584 return NULL;
585 }
586
587 // Check the possibility of scalar replacement.
588 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
589 // Scan the uses of the allocation to check for anything that would
590 // prevent us from eliminating it.
591 NOT_PRODUCT( const char* fail_eliminate = NULL; )
592 DEBUG_ONLY( Node* disq_node = NULL; )
593 bool can_eliminate = true;
594
595 Node* res = alloc->result_cast();
596 const TypeOopPtr* res_type = NULL;
597 if (res == NULL) {
598 // All users were eliminated.
599 } else if (!res->is_CheckCastPP()) {
600 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
601 can_eliminate = false;
602 } else {
603 res_type = _igvn.type(res)->isa_oopptr();
604 if (res_type == NULL) {
605 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
606 can_eliminate = false;
607 } else if (res_type->isa_aryptr()) {
608 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
609 if (length < 0) {
610 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
611 can_eliminate = false;
612 }
613 }
614 }
615
616 if (can_eliminate && res != NULL) {
617 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
618 j < jmax && can_eliminate; j++) {
619 Node* use = res->fast_out(j);
620
621 if (use->is_AddP()) {
622 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
623 int offset = addp_type->offset();
624
625 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
626 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
627 can_eliminate = false;
628 break;
629 }
630 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
631 k < kmax && can_eliminate; k++) {
632 Node* n = use->fast_out(k);
633 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
634 (!UseShenandoahGC || !n->is_g1_wb_pre_call())) {
635 DEBUG_ONLY(disq_node = n;)
636 if (n->is_Load() || n->is_LoadStore()) {
637 NOT_PRODUCT(fail_eliminate = "Field load";)
638 } else {
639 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
640 }
641 can_eliminate = false;
642 }
643 }
644 } else if (use->is_SafePoint()) {
645 SafePointNode* sfpt = use->as_SafePoint();
646 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
647 // Object is passed as argument.
648 DEBUG_ONLY(disq_node = use;)
649 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
650 can_eliminate = false;
651 }
652 Node* sfptMem = sfpt->memory();
653 if (sfptMem == NULL || sfptMem->is_top()) {
654 DEBUG_ONLY(disq_node = use;)
655 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
656 can_eliminate = false;
657 } else {
658 safepoints.append_if_missing(sfpt);
659 }
660 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
661 if (use->is_Phi()) {
662 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
663 NOT_PRODUCT(fail_eliminate = "Object is return value";)
664 } else {
665 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
666 }
667 DEBUG_ONLY(disq_node = use;)
668 } else {
669 if (use->Opcode() == Op_Return) {
670 NOT_PRODUCT(fail_eliminate = "Object is return value";)
671 }else {
672 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
673 }
674 DEBUG_ONLY(disq_node = use;)
675 }
676 can_eliminate = false;
677 }
678 }
679 }
680
681 #ifndef PRODUCT
682 if (PrintEliminateAllocations) {
683 if (can_eliminate) {
684 tty->print("Scalar ");
685 if (res == NULL)
686 alloc->dump();
687 else
688 res->dump();
689 } else if (alloc->_is_scalar_replaceable) {
690 tty->print("NotScalar (%s)", fail_eliminate);
691 if (res == NULL)
692 alloc->dump();
693 else
694 res->dump();
695 #ifdef ASSERT
696 if (disq_node != NULL) {
697 tty->print(" >>>> ");
698 disq_node->dump();
699 }
700 #endif /*ASSERT*/
701 }
702 }
703 #endif
704 return can_eliminate;
705 }
706
707 // Do scalar replacement.
708 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
709 GrowableArray <SafePointNode *> safepoints_done;
710
711 ciKlass* klass = NULL;
712 ciInstanceKlass* iklass = NULL;
713 int nfields = 0;
714 int array_base = 0;
715 int element_size = 0;
716 BasicType basic_elem_type = T_ILLEGAL;
717 ciType* elem_type = NULL;
718
719 Node* res = alloc->result_cast();
720 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
721 const TypeOopPtr* res_type = NULL;
722 if (res != NULL) { // Could be NULL when there are no users
723 res_type = _igvn.type(res)->isa_oopptr();
724 }
725
726 if (res != NULL) {
727 klass = res_type->klass();
728 if (res_type->isa_instptr()) {
729 // find the fields of the class which will be needed for safepoint debug information
730 assert(klass->is_instance_klass(), "must be an instance klass.");
731 iklass = klass->as_instance_klass();
732 nfields = iklass->nof_nonstatic_fields();
733 } else {
734 // find the array's elements which will be needed for safepoint debug information
735 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
736 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
737 elem_type = klass->as_array_klass()->element_type();
738 basic_elem_type = elem_type->basic_type();
739 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
740 element_size = type2aelembytes(basic_elem_type);
741 }
742 }
743 //
744 // Process the safepoint uses
745 //
746 while (safepoints.length() > 0) {
747 SafePointNode* sfpt = safepoints.pop();
748 Node* mem = sfpt->memory();
749 assert(sfpt->jvms() != NULL, "missed JVMS");
750 // Fields of scalar objs are referenced only at the end
751 // of regular debuginfo at the last (youngest) JVMS.
752 // Record relative start index.
753 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
754 SafePointScalarObjectNode* sobj = new (C) SafePointScalarObjectNode(res_type,
755 #ifdef ASSERT
756 alloc,
757 #endif
758 first_ind, nfields);
759 sobj->init_req(0, C->root());
760 transform_later(sobj);
761
762 // Scan object's fields adding an input to the safepoint for each field.
763 for (int j = 0; j < nfields; j++) {
764 intptr_t offset;
765 ciField* field = NULL;
766 if (iklass != NULL) {
767 field = iklass->nonstatic_field_at(j);
768 offset = field->offset();
769 elem_type = field->type();
770 basic_elem_type = field->layout_type();
771 } else {
772 offset = array_base + j * (intptr_t)element_size;
773 }
774
775 const Type *field_type;
776 // The next code is taken from Parse::do_get_xxx().
777 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
778 if (!elem_type->is_loaded()) {
779 field_type = TypeInstPtr::BOTTOM;
780 } else if (field != NULL && field->is_constant() && field->is_static()) {
781 // This can happen if the constant oop is non-perm.
782 ciObject* con = field->constant_value().as_object();
783 // Do not "join" in the previous type; it doesn't add value,
784 // and may yield a vacuous result if the field is of interface type.
785 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
786 assert(field_type != NULL, "field singleton type must be consistent");
787 } else {
788 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
789 }
790 if (UseCompressedOops) {
791 field_type = field_type->make_narrowoop();
792 basic_elem_type = T_NARROWOOP;
793 }
794 } else {
795 field_type = Type::get_const_basic_type(basic_elem_type);
796 }
797
798 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
799
800 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
801 if (field_val == NULL) {
802 // We weren't able to find a value for this field,
803 // give up on eliminating this allocation.
804
805 // Remove any extra entries we added to the safepoint.
806 uint last = sfpt->req() - 1;
807 for (int k = 0; k < j; k++) {
808 sfpt->del_req(last--);
809 }
810 // rollback processed safepoints
811 while (safepoints_done.length() > 0) {
812 SafePointNode* sfpt_done = safepoints_done.pop();
813 // remove any extra entries we added to the safepoint
814 last = sfpt_done->req() - 1;
815 for (int k = 0; k < nfields; k++) {
816 sfpt_done->del_req(last--);
817 }
818 JVMState *jvms = sfpt_done->jvms();
819 jvms->set_endoff(sfpt_done->req());
820 // Now make a pass over the debug information replacing any references
821 // to SafePointScalarObjectNode with the allocated object.
822 int start = jvms->debug_start();
823 int end = jvms->debug_end();
824 for (int i = start; i < end; i++) {
825 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
826 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
827 if (scobj->first_index(jvms) == sfpt_done->req() &&
828 scobj->n_fields() == (uint)nfields) {
829 assert(scobj->alloc() == alloc, "sanity");
830 sfpt_done->set_req(i, res);
831 }
832 }
833 }
834 }
835 #ifndef PRODUCT
836 if (PrintEliminateAllocations) {
837 if (field != NULL) {
838 tty->print("=== At SafePoint node %d can't find value of Field: ",
839 sfpt->_idx);
840 field->print();
841 int field_idx = C->get_alias_index(field_addr_type);
842 tty->print(" (alias_idx=%d)", field_idx);
843 } else { // Array's element
844 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
845 sfpt->_idx, j);
846 }
847 tty->print(", which prevents elimination of: ");
848 if (res == NULL)
849 alloc->dump();
850 else
851 res->dump();
852 }
853 #endif
854 return false;
855 }
856 if (UseCompressedOops && field_type->isa_narrowoop()) {
857 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
858 // to be able scalar replace the allocation.
859 if (field_val->is_EncodeP()) {
860 field_val = field_val->in(1);
861 } else {
862 field_val = transform_later(new (C) DecodeNNode(field_val, field_val->get_ptr_type()));
863 }
864 }
865 sfpt->add_req(field_val);
866 }
867 JVMState *jvms = sfpt->jvms();
868 jvms->set_endoff(sfpt->req());
869 // Now make a pass over the debug information replacing any references
870 // to the allocated object with "sobj"
871 int start = jvms->debug_start();
872 int end = jvms->debug_end();
873 sfpt->replace_edges_in_range(res, sobj, start, end);
874 safepoints_done.append_if_missing(sfpt); // keep it for rollback
875 }
876 return true;
877 }
878
879 // Process users of eliminated allocation.
880 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
881 Node* res = alloc->result_cast();
882 if (res != NULL) {
883 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
884 Node *use = res->last_out(j);
885 uint oc1 = res->outcnt();
886
887 if (use->is_AddP()) {
888 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
889 Node *n = use->last_out(k);
890 uint oc2 = use->outcnt();
891 if (n->is_Store()) {
892 #ifdef ASSERT
893 // Verify that there is no dependent MemBarVolatile nodes,
894 // they should be removed during IGVN, see MemBarNode::Ideal().
895 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
896 p < pmax; p++) {
897 Node* mb = n->fast_out(p);
898 assert(mb->is_Initialize() || !mb->is_MemBar() ||
899 mb->req() <= MemBarNode::Precedent ||
900 mb->in(MemBarNode::Precedent) != n,
901 "MemBarVolatile should be eliminated for non-escaping object");
902 }
903 #endif
904 _igvn.replace_node(n, n->in(MemNode::Memory));
905 } else if (UseShenandoahGC && n->is_g1_wb_pre_call()) {
906 C->shenandoah_eliminate_g1_wb_pre(n, &_igvn);
907 } else {
908 eliminate_card_mark(n);
909 }
910 k -= (oc2 - use->outcnt());
911 }
912 _igvn.remove_dead_node(use);
913 } else {
914 eliminate_card_mark(use);
915 }
916 j -= (oc1 - res->outcnt());
917 }
918 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
919 _igvn.remove_dead_node(res);
920 }
921
922 //
923 // Process other users of allocation's projections
924 //
925 if (_resproj != NULL && _resproj->outcnt() != 0) {
926 // First disconnect stores captured by Initialize node.
927 // If Initialize node is eliminated first in the following code,
928 // it will kill such stores and DUIterator_Last will assert.
929 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
930 Node *use = _resproj->fast_out(j);
931 if (use->is_AddP()) {
932 // raw memory addresses used only by the initialization
933 _igvn.replace_node(use, C->top());
934 --j; --jmax;
935 }
936 }
937 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
938 Node *use = _resproj->last_out(j);
939 uint oc1 = _resproj->outcnt();
940 if (use->is_Initialize()) {
941 // Eliminate Initialize node.
942 InitializeNode *init = use->as_Initialize();
943 assert(init->outcnt() <= 2, "only a control and memory projection expected");
944 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
945 if (ctrl_proj != NULL) {
946 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
947 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
948 }
949 Node *mem_proj = init->proj_out(TypeFunc::Memory);
950 if (mem_proj != NULL) {
951 Node *mem = init->in(TypeFunc::Memory);
952 #ifdef ASSERT
953 if (mem->is_MergeMem()) {
954 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
955 } else {
956 assert(mem == _memproj_fallthrough, "allocation memory projection");
957 }
958 #endif
959 _igvn.replace_node(mem_proj, mem);
960 }
961 } else {
962 assert(false, "only Initialize or AddP expected");
963 }
964 j -= (oc1 - _resproj->outcnt());
965 }
966 }
967 if (_fallthroughcatchproj != NULL) {
968 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
969 }
970 if (_memproj_fallthrough != NULL) {
971 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
972 }
973 if (_memproj_catchall != NULL) {
974 _igvn.replace_node(_memproj_catchall, C->top());
975 }
976 if (_ioproj_fallthrough != NULL) {
977 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
978 }
979 if (_ioproj_catchall != NULL) {
980 _igvn.replace_node(_ioproj_catchall, C->top());
981 }
982 if (_catchallcatchproj != NULL) {
983 _igvn.replace_node(_catchallcatchproj, C->top());
984 }
985 }
986
987 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
988 // Don't do scalar replacement if the frame can be popped by JVMTI:
989 // if reallocation fails during deoptimization we'll pop all
990 // interpreter frames for this compiled frame and that won't play
991 // nice with JVMTI popframe.
992 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
993 return false;
994 }
995 Node* klass = alloc->in(AllocateNode::KlassNode);
996 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
997 Node* res = alloc->result_cast();
998 // Eliminate boxing allocations which are not used
999 // regardless scalar replacable status.
1000 bool boxing_alloc = C->eliminate_boxing() &&
1001 tklass->klass()->is_instance_klass() &&
1002 tklass->klass()->as_instance_klass()->is_box_klass();
1003 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1004 return false;
1005 }
1006
1007 extract_call_projections(alloc);
1008
1009 GrowableArray <SafePointNode *> safepoints;
1010 if (!can_eliminate_allocation(alloc, safepoints)) {
1011 return false;
1012 }
1013
1014 if (!alloc->_is_scalar_replaceable) {
1015 assert(res == NULL, "sanity");
1016 // We can only eliminate allocation if all debug info references
1017 // are already replaced with SafePointScalarObject because
1018 // we can't search for a fields value without instance_id.
1019 if (safepoints.length() > 0) {
1020 return false;
1021 }
1022 }
1023
1024 if (!scalar_replacement(alloc, safepoints)) {
1025 return false;
1026 }
1027
1028 CompileLog* log = C->log();
1029 if (log != NULL) {
1030 log->head("eliminate_allocation type='%d'",
1031 log->identify(tklass->klass()));
1032 JVMState* p = alloc->jvms();
1033 while (p != NULL) {
1034 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1035 p = p->caller();
1036 }
1037 log->tail("eliminate_allocation");
1038 }
1039
1040 process_users_of_allocation(alloc);
1041
1042 #ifndef PRODUCT
1043 if (PrintEliminateAllocations) {
1044 if (alloc->is_AllocateArray())
1045 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1046 else
1047 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1048 }
1049 #endif
1050
1051 return true;
1052 }
1053
1054 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1055 // EA should remove all uses of non-escaping boxing node.
1056 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1057 return false;
1058 }
1059
1060 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1061
1062 extract_call_projections(boxing);
1063
1064 const TypeTuple* r = boxing->tf()->range();
1065 assert(r->cnt() > TypeFunc::Parms, "sanity");
1066 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1067 assert(t != NULL, "sanity");
1068
1069 CompileLog* log = C->log();
1070 if (log != NULL) {
1071 log->head("eliminate_boxing type='%d'",
1072 log->identify(t->klass()));
1073 JVMState* p = boxing->jvms();
1074 while (p != NULL) {
1075 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1076 p = p->caller();
1077 }
1078 log->tail("eliminate_boxing");
1079 }
1080
1081 process_users_of_allocation(boxing);
1082
1083 #ifndef PRODUCT
1084 if (PrintEliminateAllocations) {
1085 tty->print("++++ Eliminated: %d ", boxing->_idx);
1086 boxing->method()->print_short_name(tty);
1087 tty->cr();
1088 }
1089 #endif
1090
1091 return true;
1092 }
1093
1094 //---------------------------set_eden_pointers-------------------------
1095 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1096 if (UseTLAB) { // Private allocation: load from TLS
1097 Node* thread = transform_later(new (C) ThreadLocalNode());
1098 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1099 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1100 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1101 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1102 } else { // Shared allocation: load from globals
1103 CollectedHeap* ch = Universe::heap();
1104 address top_adr = (address)ch->top_addr();
1105 address end_adr = (address)ch->end_addr();
1106 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1107 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1108 }
1109 }
1110
1111
1112 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1113 Node* adr = basic_plus_adr(base, offset);
1114 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1115 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1116 transform_later(value);
1117 return value;
1118 }
1119
1120
1121 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1122 Node* adr = basic_plus_adr(base, offset);
1123 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1124 transform_later(mem);
1125 return mem;
1126 }
1127
1128 //=============================================================================
1129 //
1130 // A L L O C A T I O N
1131 //
1132 // Allocation attempts to be fast in the case of frequent small objects.
1133 // It breaks down like this:
1134 //
1135 // 1) Size in doublewords is computed. This is a constant for objects and
1136 // variable for most arrays. Doubleword units are used to avoid size
1137 // overflow of huge doubleword arrays. We need doublewords in the end for
1138 // rounding.
1139 //
1140 // 2) Size is checked for being 'too large'. Too-large allocations will go
1141 // the slow path into the VM. The slow path can throw any required
1142 // exceptions, and does all the special checks for very large arrays. The
1143 // size test can constant-fold away for objects. For objects with
1144 // finalizers it constant-folds the otherway: you always go slow with
1145 // finalizers.
1146 //
1147 // 3) If NOT using TLABs, this is the contended loop-back point.
1148 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1149 //
1150 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1151 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1152 // "size*8" we always enter the VM, where "largish" is a constant picked small
1153 // enough that there's always space between the eden max and 4Gig (old space is
1154 // there so it's quite large) and large enough that the cost of entering the VM
1155 // is dwarfed by the cost to initialize the space.
1156 //
1157 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1158 // down. If contended, repeat at step 3. If using TLABs normal-store
1159 // adjusted heap top back down; there is no contention.
1160 //
1161 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1162 // fields.
1163 //
1164 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1165 // oop flavor.
1166 //
1167 //=============================================================================
1168 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1169 // Allocations bigger than this always go the slow route.
1170 // This value must be small enough that allocation attempts that need to
1171 // trigger exceptions go the slow route. Also, it must be small enough so
1172 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1173 //=============================================================================j//
1174 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1175 // The allocator will coalesce int->oop copies away. See comment in
1176 // coalesce.cpp about how this works. It depends critically on the exact
1177 // code shape produced here, so if you are changing this code shape
1178 // make sure the GC info for the heap-top is correct in and around the
1179 // slow-path call.
1180 //
1181
1182 void PhaseMacroExpand::expand_allocate_common(
1183 AllocateNode* alloc, // allocation node to be expanded
1184 Node* length, // array length for an array allocation
1185 const TypeFunc* slow_call_type, // Type of slow call
1186 address slow_call_address // Address of slow call
1187 )
1188 {
1189
1190 Node* ctrl = alloc->in(TypeFunc::Control);
1191 Node* mem = alloc->in(TypeFunc::Memory);
1192 Node* i_o = alloc->in(TypeFunc::I_O);
1193 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1194 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1195 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1196
1197 assert(ctrl != NULL, "must have control");
1198 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1199 // they will not be used if "always_slow" is set
1200 enum { slow_result_path = 1, fast_result_path = 2 };
1201 Node *result_region = NULL;
1202 Node *result_phi_rawmem = NULL;
1203 Node *result_phi_rawoop = NULL;
1204 Node *result_phi_i_o = NULL;
1205
1206 // The initial slow comparison is a size check, the comparison
1207 // we want to do is a BoolTest::gt
1208 bool always_slow = false;
1209 int tv = _igvn.find_int_con(initial_slow_test, -1);
1210 if (tv >= 0) {
1211 always_slow = (tv == 1);
1212 initial_slow_test = NULL;
1213 } else {
1214 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1215 }
1216
1217 if (C->env()->dtrace_alloc_probes() ||
1218 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1219 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1220 // Force slow-path allocation
1221 always_slow = true;
1222 initial_slow_test = NULL;
1223 }
1224
1225
1226 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1227 Node *slow_region = NULL;
1228 Node *toobig_false = ctrl;
1229
1230 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1231 // generate the initial test if necessary
1232 if (initial_slow_test != NULL ) {
1233 slow_region = new (C) RegionNode(3);
1234
1235 // Now make the initial failure test. Usually a too-big test but
1236 // might be a TRUE for finalizers or a fancy class check for
1237 // newInstance0.
1238 IfNode *toobig_iff = new (C) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1239 transform_later(toobig_iff);
1240 // Plug the failing-too-big test into the slow-path region
1241 Node *toobig_true = new (C) IfTrueNode( toobig_iff );
1242 transform_later(toobig_true);
1243 slow_region ->init_req( too_big_or_final_path, toobig_true );
1244 toobig_false = new (C) IfFalseNode( toobig_iff );
1245 transform_later(toobig_false);
1246 } else { // No initial test, just fall into next case
1247 toobig_false = ctrl;
1248 debug_only(slow_region = NodeSentinel);
1249 }
1250
1251 Node *slow_mem = mem; // save the current memory state for slow path
1252 // generate the fast allocation code unless we know that the initial test will always go slow
1253 if (!always_slow) {
1254 // Fast path modifies only raw memory.
1255 if (mem->is_MergeMem()) {
1256 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1257 }
1258
1259 Node* eden_top_adr;
1260 Node* eden_end_adr;
1261
1262 set_eden_pointers(eden_top_adr, eden_end_adr);
1263
1264 // Load Eden::end. Loop invariant and hoisted.
1265 //
1266 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1267 // a TLAB to work around a bug where these values were being moved across
1268 // a safepoint. These are not oops, so they cannot be include in the oop
1269 // map, but they can be changed by a GC. The proper way to fix this would
1270 // be to set the raw memory state when generating a SafepointNode. However
1271 // this will require extensive changes to the loop optimization in order to
1272 // prevent a degradation of the optimization.
1273 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1274 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1275
1276 // allocate the Region and Phi nodes for the result
1277 result_region = new (C) RegionNode(3);
1278 result_phi_rawmem = new (C) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1279 result_phi_rawoop = new (C) PhiNode(result_region, TypeRawPtr::BOTTOM);
1280 result_phi_i_o = new (C) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1281
1282 // We need a Region for the loop-back contended case.
1283 enum { fall_in_path = 1, contended_loopback_path = 2 };
1284 Node *contended_region;
1285 Node *contended_phi_rawmem;
1286 if (UseTLAB) {
1287 contended_region = toobig_false;
1288 contended_phi_rawmem = mem;
1289 } else {
1290 contended_region = new (C) RegionNode(3);
1291 contended_phi_rawmem = new (C) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1292 // Now handle the passing-too-big test. We fall into the contended
1293 // loop-back merge point.
1294 contended_region ->init_req(fall_in_path, toobig_false);
1295 contended_phi_rawmem->init_req(fall_in_path, mem);
1296 transform_later(contended_region);
1297 transform_later(contended_phi_rawmem);
1298 }
1299
1300 // Load(-locked) the heap top.
1301 // See note above concerning the control input when using a TLAB
1302 Node *old_eden_top = UseTLAB
1303 ? new (C) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1304 : new (C) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1305
1306 transform_later(old_eden_top);
1307 // Add to heap top to get a new heap top
1308 Node *new_eden_top = new (C) AddPNode(top(), old_eden_top, size_in_bytes);
1309 transform_later(new_eden_top);
1310 // Check for needing a GC; compare against heap end
1311 Node *needgc_cmp = new (C) CmpPNode(new_eden_top, eden_end);
1312 transform_later(needgc_cmp);
1313 Node *needgc_bol = new (C) BoolNode(needgc_cmp, BoolTest::ge);
1314 transform_later(needgc_bol);
1315 IfNode *needgc_iff = new (C) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1316 transform_later(needgc_iff);
1317
1318 // Plug the failing-heap-space-need-gc test into the slow-path region
1319 Node *needgc_true = new (C) IfTrueNode(needgc_iff);
1320 transform_later(needgc_true);
1321 if (initial_slow_test) {
1322 slow_region->init_req(need_gc_path, needgc_true);
1323 // This completes all paths into the slow merge point
1324 transform_later(slow_region);
1325 } else { // No initial slow path needed!
1326 // Just fall from the need-GC path straight into the VM call.
1327 slow_region = needgc_true;
1328 }
1329 // No need for a GC. Setup for the Store-Conditional
1330 Node *needgc_false = new (C) IfFalseNode(needgc_iff);
1331 transform_later(needgc_false);
1332
1333 // Grab regular I/O before optional prefetch may change it.
1334 // Slow-path does no I/O so just set it to the original I/O.
1335 result_phi_i_o->init_req(slow_result_path, i_o);
1336
1337 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1338 old_eden_top, new_eden_top, length);
1339
1340 // Name successful fast-path variables
1341 Node* fast_oop = old_eden_top;
1342 Node* fast_oop_ctrl;
1343 Node* fast_oop_rawmem;
1344
1345 // Store (-conditional) the modified eden top back down.
1346 // StorePConditional produces flags for a test PLUS a modified raw
1347 // memory state.
1348 if (UseTLAB) {
1349 Node* store_eden_top =
1350 new (C) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1351 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1352 transform_later(store_eden_top);
1353 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1354 fast_oop_rawmem = store_eden_top;
1355 } else {
1356 Node* store_eden_top =
1357 new (C) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1358 new_eden_top, fast_oop/*old_eden_top*/);
1359 transform_later(store_eden_top);
1360 Node *contention_check = new (C) BoolNode(store_eden_top, BoolTest::ne);
1361 transform_later(contention_check);
1362 store_eden_top = new (C) SCMemProjNode(store_eden_top);
1363 transform_later(store_eden_top);
1364
1365 // If not using TLABs, check to see if there was contention.
1366 IfNode *contention_iff = new (C) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1367 transform_later(contention_iff);
1368 Node *contention_true = new (C) IfTrueNode(contention_iff);
1369 transform_later(contention_true);
1370 // If contention, loopback and try again.
1371 contended_region->init_req(contended_loopback_path, contention_true);
1372 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1373
1374 // Fast-path succeeded with no contention!
1375 Node *contention_false = new (C) IfFalseNode(contention_iff);
1376 transform_later(contention_false);
1377 fast_oop_ctrl = contention_false;
1378
1379 // Bump total allocated bytes for this thread
1380 Node* thread = new (C) ThreadLocalNode();
1381 transform_later(thread);
1382 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1383 in_bytes(JavaThread::allocated_bytes_offset()));
1384 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1385 0, TypeLong::LONG, T_LONG);
1386 #ifdef _LP64
1387 Node* alloc_size = size_in_bytes;
1388 #else
1389 Node* alloc_size = new (C) ConvI2LNode(size_in_bytes);
1390 transform_later(alloc_size);
1391 #endif
1392 Node* new_alloc_bytes = new (C) AddLNode(alloc_bytes, alloc_size);
1393 transform_later(new_alloc_bytes);
1394 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1395 0, new_alloc_bytes, T_LONG);
1396 }
1397
1398 InitializeNode* init = alloc->initialization();
1399 fast_oop_rawmem = initialize_object(alloc,
1400 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1401 klass_node, length, size_in_bytes);
1402
1403 // If initialization is performed by an array copy, any required
1404 // MemBarStoreStore was already added. If the object does not
1405 // escape no need for a MemBarStoreStore. Otherwise we need a
1406 // MemBarStoreStore so that stores that initialize this object
1407 // can't be reordered with a subsequent store that makes this
1408 // object accessible by other threads.
1409 #ifndef AARCH64
1410 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1411 #else
1412 if (!alloc->does_not_escape_thread() &&
1413 (init == NULL || !init->is_complete_with_arraycopy())) {
1414 #endif
1415 if (init == NULL || init->req() < InitializeNode::RawStores) {
1416 // No InitializeNode or no stores captured by zeroing
1417 // elimination. Simply add the MemBarStoreStore after object
1418 // initialization.
1419 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1420 transform_later(mb);
1421
1422 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1423 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1424 fast_oop_ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1425 transform_later(fast_oop_ctrl);
1426 fast_oop_rawmem = new (C) ProjNode(mb,TypeFunc::Memory);
1427 transform_later(fast_oop_rawmem);
1428 } else {
1429 // Add the MemBarStoreStore after the InitializeNode so that
1430 // all stores performing the initialization that were moved
1431 // before the InitializeNode happen before the storestore
1432 // barrier.
1433
1434 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1435 Node* init_mem = init->proj_out(TypeFunc::Memory);
1436
1437 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1438 transform_later(mb);
1439
1440 Node* ctrl = new (C) ProjNode(init,TypeFunc::Control);
1441 transform_later(ctrl);
1442 Node* mem = new (C) ProjNode(init,TypeFunc::Memory);
1443 transform_later(mem);
1444
1445 // The MemBarStoreStore depends on control and memory coming
1446 // from the InitializeNode
1447 mb->init_req(TypeFunc::Memory, mem);
1448 mb->init_req(TypeFunc::Control, ctrl);
1449
1450 ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1451 transform_later(ctrl);
1452 mem = new (C) ProjNode(mb,TypeFunc::Memory);
1453 transform_later(mem);
1454
1455 // All nodes that depended on the InitializeNode for control
1456 // and memory must now depend on the MemBarNode that itself
1457 // depends on the InitializeNode
1458 _igvn.replace_node(init_ctrl, ctrl);
1459 _igvn.replace_node(init_mem, mem);
1460 }
1461 }
1462
1463 if (C->env()->dtrace_extended_probes()) {
1464 // Slow-path call
1465 int size = TypeFunc::Parms + 2;
1466 CallLeafNode *call = new (C) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1467 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1468 "dtrace_object_alloc",
1469 TypeRawPtr::BOTTOM);
1470
1471 // Get base of thread-local storage area
1472 Node* thread = new (C) ThreadLocalNode();
1473 transform_later(thread);
1474
1475 call->init_req(TypeFunc::Parms+0, thread);
1476 call->init_req(TypeFunc::Parms+1, fast_oop);
1477 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1478 call->init_req(TypeFunc::I_O , top()); // does no i/o
1479 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1480 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1481 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1482 transform_later(call);
1483 fast_oop_ctrl = new (C) ProjNode(call,TypeFunc::Control);
1484 transform_later(fast_oop_ctrl);
1485 fast_oop_rawmem = new (C) ProjNode(call,TypeFunc::Memory);
1486 transform_later(fast_oop_rawmem);
1487 }
1488
1489 // Plug in the successful fast-path into the result merge point
1490 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1491 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1492 result_phi_i_o ->init_req(fast_result_path, i_o);
1493 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1494 } else {
1495 slow_region = ctrl;
1496 result_phi_i_o = i_o; // Rename it to use in the following code.
1497 }
1498
1499 // Generate slow-path call
1500 CallNode *call = new (C) CallStaticJavaNode(slow_call_type, slow_call_address,
1501 OptoRuntime::stub_name(slow_call_address),
1502 alloc->jvms()->bci(),
1503 TypePtr::BOTTOM);
1504 call->init_req( TypeFunc::Control, slow_region );
1505 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1506 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1507 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1508 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1509
1510 call->init_req(TypeFunc::Parms+0, klass_node);
1511 if (length != NULL) {
1512 call->init_req(TypeFunc::Parms+1, length);
1513 }
1514
1515 // Copy debug information and adjust JVMState information, then replace
1516 // allocate node with the call
1517 copy_call_debug_info((CallNode *) alloc, call);
1518 if (!always_slow) {
1519 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1520 } else {
1521 // Hook i_o projection to avoid its elimination during allocation
1522 // replacement (when only a slow call is generated).
1523 call->set_req(TypeFunc::I_O, result_phi_i_o);
1524 }
1525 _igvn.replace_node(alloc, call);
1526 transform_later(call);
1527
1528 // Identify the output projections from the allocate node and
1529 // adjust any references to them.
1530 // The control and io projections look like:
1531 //
1532 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1533 // Allocate Catch
1534 // ^---Proj(io) <-------+ ^---CatchProj(io)
1535 //
1536 // We are interested in the CatchProj nodes.
1537 //
1538 extract_call_projections(call);
1539
1540 // An allocate node has separate memory projections for the uses on
1541 // the control and i_o paths. Replace the control memory projection with
1542 // result_phi_rawmem (unless we are only generating a slow call when
1543 // both memory projections are combined)
1544 if (!always_slow && _memproj_fallthrough != NULL) {
1545 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1546 Node *use = _memproj_fallthrough->fast_out(i);
1547 _igvn.rehash_node_delayed(use);
1548 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1549 // back up iterator
1550 --i;
1551 }
1552 }
1553 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1554 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1555 if (_memproj_catchall != NULL ) {
1556 if (_memproj_fallthrough == NULL) {
1557 _memproj_fallthrough = new (C) ProjNode(call, TypeFunc::Memory);
1558 transform_later(_memproj_fallthrough);
1559 }
1560 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1561 Node *use = _memproj_catchall->fast_out(i);
1562 _igvn.rehash_node_delayed(use);
1563 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1564 // back up iterator
1565 --i;
1566 }
1567 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1568 _igvn.remove_dead_node(_memproj_catchall);
1569 }
1570
1571 // An allocate node has separate i_o projections for the uses on the control
1572 // and i_o paths. Always replace the control i_o projection with result i_o
1573 // otherwise incoming i_o become dead when only a slow call is generated
1574 // (it is different from memory projections where both projections are
1575 // combined in such case).
1576 if (_ioproj_fallthrough != NULL) {
1577 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1578 Node *use = _ioproj_fallthrough->fast_out(i);
1579 _igvn.rehash_node_delayed(use);
1580 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1581 // back up iterator
1582 --i;
1583 }
1584 }
1585 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1586 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1587 if (_ioproj_catchall != NULL ) {
1588 if (_ioproj_fallthrough == NULL) {
1589 _ioproj_fallthrough = new (C) ProjNode(call, TypeFunc::I_O);
1590 transform_later(_ioproj_fallthrough);
1591 }
1592 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1593 Node *use = _ioproj_catchall->fast_out(i);
1594 _igvn.rehash_node_delayed(use);
1595 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1596 // back up iterator
1597 --i;
1598 }
1599 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1600 _igvn.remove_dead_node(_ioproj_catchall);
1601 }
1602
1603 // if we generated only a slow call, we are done
1604 if (always_slow) {
1605 // Now we can unhook i_o.
1606 if (result_phi_i_o->outcnt() > 1) {
1607 call->set_req(TypeFunc::I_O, top());
1608 } else {
1609 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1610 // Case of new array with negative size known during compilation.
1611 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1612 // following code since call to runtime will throw exception.
1613 // As result there will be no users of i_o after the call.
1614 // Leave i_o attached to this call to avoid problems in preceding graph.
1615 }
1616 return;
1617 }
1618
1619
1620 if (_fallthroughcatchproj != NULL) {
1621 ctrl = _fallthroughcatchproj->clone();
1622 transform_later(ctrl);
1623 _igvn.replace_node(_fallthroughcatchproj, result_region);
1624 } else {
1625 ctrl = top();
1626 }
1627 Node *slow_result;
1628 if (_resproj == NULL) {
1629 // no uses of the allocation result
1630 slow_result = top();
1631 } else {
1632 slow_result = _resproj->clone();
1633 transform_later(slow_result);
1634 _igvn.replace_node(_resproj, result_phi_rawoop);
1635 }
1636
1637 // Plug slow-path into result merge point
1638 result_region ->init_req( slow_result_path, ctrl );
1639 result_phi_rawoop->init_req( slow_result_path, slow_result);
1640 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1641 transform_later(result_region);
1642 transform_later(result_phi_rawoop);
1643 transform_later(result_phi_rawmem);
1644 transform_later(result_phi_i_o);
1645 // This completes all paths into the result merge point
1646 }
1647
1648
1649 // Helper for PhaseMacroExpand::expand_allocate_common.
1650 // Initializes the newly-allocated storage.
1651 Node*
1652 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1653 Node* control, Node* rawmem, Node* object,
1654 Node* klass_node, Node* length,
1655 Node* size_in_bytes) {
1656 InitializeNode* init = alloc->initialization();
1657 // Store the klass & mark bits
1658 Node* mark_node = NULL;
1659 // For now only enable fast locking for non-array types
1660 if (UseBiasedLocking && (length == NULL)) {
1661 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1662 } else {
1663 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1664 }
1665 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1666
1667 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1668 int header_size = alloc->minimum_header_size(); // conservatively small
1669
1670 // Array length
1671 if (length != NULL) { // Arrays need length field
1672 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1673 // conservatively small header size:
1674 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1675 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1676 if (k->is_array_klass()) // we know the exact header size in most cases:
1677 header_size = Klass::layout_helper_header_size(k->layout_helper());
1678 }
1679
1680 // Clear the object body, if necessary.
1681 if (init == NULL) {
1682 // The init has somehow disappeared; be cautious and clear everything.
1683 //
1684 // This can happen if a node is allocated but an uncommon trap occurs
1685 // immediately. In this case, the Initialize gets associated with the
1686 // trap, and may be placed in a different (outer) loop, if the Allocate
1687 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1688 // there can be two Allocates to one Initialize. The answer in all these
1689 // edge cases is safety first. It is always safe to clear immediately
1690 // within an Allocate, and then (maybe or maybe not) clear some more later.
1691 if (!ZeroTLAB)
1692 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1693 header_size, size_in_bytes,
1694 &_igvn);
1695 } else {
1696 if (!init->is_complete()) {
1697 // Try to win by zeroing only what the init does not store.
1698 // We can also try to do some peephole optimizations,
1699 // such as combining some adjacent subword stores.
1700 rawmem = init->complete_stores(control, rawmem, object,
1701 header_size, size_in_bytes, &_igvn);
1702 }
1703 // We have no more use for this link, since the AllocateNode goes away:
1704 init->set_req(InitializeNode::RawAddress, top());
1705 // (If we keep the link, it just confuses the register allocator,
1706 // who thinks he sees a real use of the address by the membar.)
1707 }
1708
1709 return rawmem;
1710 }
1711
1712 // Generate prefetch instructions for next allocations.
1713 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1714 Node*& contended_phi_rawmem,
1715 Node* old_eden_top, Node* new_eden_top,
1716 Node* length) {
1717 enum { fall_in_path = 1, pf_path = 2 };
1718 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1719 // Generate prefetch allocation with watermark check.
1720 // As an allocation hits the watermark, we will prefetch starting
1721 // at a "distance" away from watermark.
1722
1723 Node *pf_region = new (C) RegionNode(3);
1724 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1725 TypeRawPtr::BOTTOM );
1726 // I/O is used for Prefetch
1727 Node *pf_phi_abio = new (C) PhiNode( pf_region, Type::ABIO );
1728
1729 Node *thread = new (C) ThreadLocalNode();
1730 transform_later(thread);
1731
1732 Node *eden_pf_adr = new (C) AddPNode( top()/*not oop*/, thread,
1733 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1734 transform_later(eden_pf_adr);
1735
1736 Node *old_pf_wm = new (C) LoadPNode(needgc_false,
1737 contended_phi_rawmem, eden_pf_adr,
1738 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1739 MemNode::unordered);
1740 transform_later(old_pf_wm);
1741
1742 // check against new_eden_top
1743 Node *need_pf_cmp = new (C) CmpPNode( new_eden_top, old_pf_wm );
1744 transform_later(need_pf_cmp);
1745 Node *need_pf_bol = new (C) BoolNode( need_pf_cmp, BoolTest::ge );
1746 transform_later(need_pf_bol);
1747 IfNode *need_pf_iff = new (C) IfNode( needgc_false, need_pf_bol,
1748 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1749 transform_later(need_pf_iff);
1750
1751 // true node, add prefetchdistance
1752 Node *need_pf_true = new (C) IfTrueNode( need_pf_iff );
1753 transform_later(need_pf_true);
1754
1755 Node *need_pf_false = new (C) IfFalseNode( need_pf_iff );
1756 transform_later(need_pf_false);
1757
1758 Node *new_pf_wmt = new (C) AddPNode( top(), old_pf_wm,
1759 _igvn.MakeConX(AllocatePrefetchDistance) );
1760 transform_later(new_pf_wmt );
1761 new_pf_wmt->set_req(0, need_pf_true);
1762
1763 Node *store_new_wmt = new (C) StorePNode(need_pf_true,
1764 contended_phi_rawmem, eden_pf_adr,
1765 TypeRawPtr::BOTTOM, new_pf_wmt,
1766 MemNode::unordered);
1767 transform_later(store_new_wmt);
1768
1769 // adding prefetches
1770 pf_phi_abio->init_req( fall_in_path, i_o );
1771
1772 Node *prefetch_adr;
1773 Node *prefetch;
1774 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1775 uint step_size = AllocatePrefetchStepSize;
1776 uint distance = 0;
1777
1778 for ( uint i = 0; i < lines; i++ ) {
1779 prefetch_adr = new (C) AddPNode( old_pf_wm, new_pf_wmt,
1780 _igvn.MakeConX(distance) );
1781 transform_later(prefetch_adr);
1782 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1783 transform_later(prefetch);
1784 distance += step_size;
1785 i_o = prefetch;
1786 }
1787 pf_phi_abio->set_req( pf_path, i_o );
1788
1789 pf_region->init_req( fall_in_path, need_pf_false );
1790 pf_region->init_req( pf_path, need_pf_true );
1791
1792 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1793 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1794
1795 transform_later(pf_region);
1796 transform_later(pf_phi_rawmem);
1797 transform_later(pf_phi_abio);
1798
1799 needgc_false = pf_region;
1800 contended_phi_rawmem = pf_phi_rawmem;
1801 i_o = pf_phi_abio;
1802 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1803 // Insert a prefetch for each allocation.
1804 // This code is used to generate 1 prefetch instruction per cache line.
1805 Node *pf_region = new (C) RegionNode(3);
1806 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1807 TypeRawPtr::BOTTOM );
1808
1809 // Generate several prefetch instructions.
1810 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1811 uint step_size = AllocatePrefetchStepSize;
1812 uint distance = AllocatePrefetchDistance;
1813
1814 // Next cache address.
1815 Node *cache_adr = new (C) AddPNode(old_eden_top, old_eden_top,
1816 _igvn.MakeConX(distance));
1817 transform_later(cache_adr);
1818 cache_adr = new (C) CastP2XNode(needgc_false, cache_adr);
1819 transform_later(cache_adr);
1820 // Address is aligned to execute prefetch to the beginning of cache line size
1821 // (it is important when BIS instruction is used on SPARC as prefetch).
1822 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1823 cache_adr = new (C) AndXNode(cache_adr, mask);
1824 transform_later(cache_adr);
1825 cache_adr = new (C) CastX2PNode(cache_adr);
1826 transform_later(cache_adr);
1827
1828 // Prefetch
1829 Node *prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1830 prefetch->set_req(0, needgc_false);
1831 transform_later(prefetch);
1832 contended_phi_rawmem = prefetch;
1833 Node *prefetch_adr;
1834 distance = step_size;
1835 for ( uint i = 1; i < lines; i++ ) {
1836 prefetch_adr = new (C) AddPNode( cache_adr, cache_adr,
1837 _igvn.MakeConX(distance) );
1838 transform_later(prefetch_adr);
1839 prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1840 transform_later(prefetch);
1841 distance += step_size;
1842 contended_phi_rawmem = prefetch;
1843 }
1844 } else if( AllocatePrefetchStyle > 0 ) {
1845 // Insert a prefetch for each allocation only on the fast-path
1846 Node *prefetch_adr;
1847 Node *prefetch;
1848 // Generate several prefetch instructions.
1849 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1850 uint step_size = AllocatePrefetchStepSize;
1851 uint distance = AllocatePrefetchDistance;
1852 for ( uint i = 0; i < lines; i++ ) {
1853 prefetch_adr = new (C) AddPNode( old_eden_top, new_eden_top,
1854 _igvn.MakeConX(distance) );
1855 transform_later(prefetch_adr);
1856 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1857 // Do not let it float too high, since if eden_top == eden_end,
1858 // both might be null.
1859 if( i == 0 ) { // Set control for first prefetch, next follows it
1860 prefetch->init_req(0, needgc_false);
1861 }
1862 transform_later(prefetch);
1863 distance += step_size;
1864 i_o = prefetch;
1865 }
1866 }
1867 return i_o;
1868 }
1869
1870
1871 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1872 expand_allocate_common(alloc, NULL,
1873 OptoRuntime::new_instance_Type(),
1874 OptoRuntime::new_instance_Java());
1875 }
1876
1877 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1878 Node* length = alloc->in(AllocateNode::ALength);
1879 InitializeNode* init = alloc->initialization();
1880 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1881 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1882 address slow_call_address; // Address of slow call
1883 if (init != NULL && init->is_complete_with_arraycopy() &&
1884 k->is_type_array_klass()) {
1885 // Don't zero type array during slow allocation in VM since
1886 // it will be initialized later by arraycopy in compiled code.
1887 slow_call_address = OptoRuntime::new_array_nozero_Java();
1888 } else {
1889 slow_call_address = OptoRuntime::new_array_Java();
1890 }
1891 expand_allocate_common(alloc, length,
1892 OptoRuntime::new_array_Type(),
1893 slow_call_address);
1894 }
1895
1896 //-------------------mark_eliminated_box----------------------------------
1897 //
1898 // During EA obj may point to several objects but after few ideal graph
1899 // transformations (CCP) it may point to only one non escaping object
1900 // (but still using phi), corresponding locks and unlocks will be marked
1901 // for elimination. Later obj could be replaced with a new node (new phi)
1902 // and which does not have escape information. And later after some graph
1903 // reshape other locks and unlocks (which were not marked for elimination
1904 // before) are connected to this new obj (phi) but they still will not be
1905 // marked for elimination since new obj has no escape information.
1906 // Mark all associated (same box and obj) lock and unlock nodes for
1907 // elimination if some of them marked already.
1908 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1909 if (oldbox->as_BoxLock()->is_eliminated())
1910 return; // This BoxLock node was processed already.
1911
1912 // New implementation (EliminateNestedLocks) has separate BoxLock
1913 // node for each locked region so mark all associated locks/unlocks as
1914 // eliminated even if different objects are referenced in one locked region
1915 // (for example, OSR compilation of nested loop inside locked scope).
1916 if (EliminateNestedLocks ||
1917 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1918 // Box is used only in one lock region. Mark this box as eliminated.
1919 _igvn.hash_delete(oldbox);
1920 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1921 _igvn.hash_insert(oldbox);
1922
1923 for (uint i = 0; i < oldbox->outcnt(); i++) {
1924 Node* u = oldbox->raw_out(i);
1925 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1926 AbstractLockNode* alock = u->as_AbstractLock();
1927 // Check lock's box since box could be referenced by Lock's debug info.
1928 if (alock->box_node() == oldbox) {
1929 // Mark eliminated all related locks and unlocks.
1930 #ifdef ASSERT
1931 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1932 #endif
1933 alock->set_non_esc_obj();
1934 }
1935 }
1936 }
1937 return;
1938 }
1939
1940 // Create new "eliminated" BoxLock node and use it in monitor debug info
1941 // instead of oldbox for the same object.
1942 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1943
1944 // Note: BoxLock node is marked eliminated only here and it is used
1945 // to indicate that all associated lock and unlock nodes are marked
1946 // for elimination.
1947 newbox->set_eliminated();
1948 transform_later(newbox);
1949
1950 // Replace old box node with new box for all users of the same object.
1951 for (uint i = 0; i < oldbox->outcnt();) {
1952 bool next_edge = true;
1953
1954 Node* u = oldbox->raw_out(i);
1955 if (u->is_AbstractLock()) {
1956 AbstractLockNode* alock = u->as_AbstractLock();
1957 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1958 // Replace Box and mark eliminated all related locks and unlocks.
1959 #ifdef ASSERT
1960 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1961 #endif
1962 alock->set_non_esc_obj();
1963 _igvn.rehash_node_delayed(alock);
1964 alock->set_box_node(newbox);
1965 next_edge = false;
1966 }
1967 }
1968 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1969 FastLockNode* flock = u->as_FastLock();
1970 assert(flock->box_node() == oldbox, "sanity");
1971 _igvn.rehash_node_delayed(flock);
1972 flock->set_box_node(newbox);
1973 next_edge = false;
1974 }
1975
1976 // Replace old box in monitor debug info.
1977 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1978 SafePointNode* sfn = u->as_SafePoint();
1979 JVMState* youngest_jvms = sfn->jvms();
1980 int max_depth = youngest_jvms->depth();
1981 for (int depth = 1; depth <= max_depth; depth++) {
1982 JVMState* jvms = youngest_jvms->of_depth(depth);
1983 int num_mon = jvms->nof_monitors();
1984 // Loop over monitors
1985 for (int idx = 0; idx < num_mon; idx++) {
1986 Node* obj_node = sfn->monitor_obj(jvms, idx);
1987 Node* box_node = sfn->monitor_box(jvms, idx);
1988 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1989 int j = jvms->monitor_box_offset(idx);
1990 _igvn.replace_input_of(u, j, newbox);
1991 next_edge = false;
1992 }
1993 }
1994 }
1995 }
1996 if (next_edge) i++;
1997 }
1998 }
1999
2000 //-----------------------mark_eliminated_locking_nodes-----------------------
2001 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2002 if (EliminateNestedLocks) {
2003 if (alock->is_nested()) {
2004 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2005 return;
2006 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2007 // Only Lock node has JVMState needed here.
2008 // Not that preceding claim is documented anywhere else.
2009 if (alock->jvms() != NULL) {
2010 if (alock->as_Lock()->is_nested_lock_region()) {
2011 // Mark eliminated related nested locks and unlocks.
2012 Node* obj = alock->obj_node();
2013 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2014 assert(!box_node->is_eliminated(), "should not be marked yet");
2015 // Note: BoxLock node is marked eliminated only here
2016 // and it is used to indicate that all associated lock
2017 // and unlock nodes are marked for elimination.
2018 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2019 for (uint i = 0; i < box_node->outcnt(); i++) {
2020 Node* u = box_node->raw_out(i);
2021 if (u->is_AbstractLock()) {
2022 alock = u->as_AbstractLock();
2023 if (alock->box_node() == box_node) {
2024 // Verify that this Box is referenced only by related locks.
2025 assert(alock->obj_node()->eqv_uncast(obj), "");
2026 // Mark all related locks and unlocks.
2027 #ifdef ASSERT
2028 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2029 #endif
2030 alock->set_nested();
2031 }
2032 }
2033 }
2034 } else {
2035 #ifdef ASSERT
2036 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2037 if (C->log() != NULL)
2038 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2039 #endif
2040 }
2041 }
2042 return;
2043 }
2044 // Process locks for non escaping object
2045 assert(alock->is_non_esc_obj(), "");
2046 } // EliminateNestedLocks
2047
2048 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2049 // Look for all locks of this object and mark them and
2050 // corresponding BoxLock nodes as eliminated.
2051 Node* obj = alock->obj_node();
2052 for (uint j = 0; j < obj->outcnt(); j++) {
2053 Node* o = obj->raw_out(j);
2054 if (o->is_AbstractLock() &&
2055 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2056 alock = o->as_AbstractLock();
2057 Node* box = alock->box_node();
2058 // Replace old box node with new eliminated box for all users
2059 // of the same object and mark related locks as eliminated.
2060 mark_eliminated_box(box, obj);
2061 }
2062 }
2063 }
2064 }
2065
2066 // we have determined that this lock/unlock can be eliminated, we simply
2067 // eliminate the node without expanding it.
2068 //
2069 // Note: The membar's associated with the lock/unlock are currently not
2070 // eliminated. This should be investigated as a future enhancement.
2071 //
2072 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2073
2074 if (!alock->is_eliminated()) {
2075 return false;
2076 }
2077 #ifdef ASSERT
2078 if (!alock->is_coarsened()) {
2079 // Check that new "eliminated" BoxLock node is created.
2080 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2081 assert(oldbox->is_eliminated(), "should be done already");
2082 }
2083 #endif
2084
2085 alock->log_lock_optimization(C, "eliminate_lock");
2086
2087 #ifndef PRODUCT
2088 if (PrintEliminateLocks) {
2089 if (alock->is_Lock()) {
2090 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2091 } else {
2092 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2093 }
2094 }
2095 #endif
2096
2097 Node* mem = alock->in(TypeFunc::Memory);
2098 Node* ctrl = alock->in(TypeFunc::Control);
2099
2100 extract_call_projections(alock);
2101 // There are 2 projections from the lock. The lock node will
2102 // be deleted when its last use is subsumed below.
2103 assert(alock->outcnt() == 2 &&
2104 _fallthroughproj != NULL &&
2105 _memproj_fallthrough != NULL,
2106 "Unexpected projections from Lock/Unlock");
2107
2108 Node* fallthroughproj = _fallthroughproj;
2109 Node* memproj_fallthrough = _memproj_fallthrough;
2110
2111 // The memory projection from a lock/unlock is RawMem
2112 // The input to a Lock is merged memory, so extract its RawMem input
2113 // (unless the MergeMem has been optimized away.)
2114 if (alock->is_Lock()) {
2115 // Seach for MemBarAcquireLock node and delete it also.
2116 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2117 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2118 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2119 Node* memproj = membar->proj_out(TypeFunc::Memory);
2120 _igvn.replace_node(ctrlproj, fallthroughproj);
2121 _igvn.replace_node(memproj, memproj_fallthrough);
2122
2123 // Delete FastLock node also if this Lock node is unique user
2124 // (a loop peeling may clone a Lock node).
2125 Node* flock = alock->as_Lock()->fastlock_node();
2126 if (flock->outcnt() == 1) {
2127 assert(flock->unique_out() == alock, "sanity");
2128 _igvn.replace_node(flock, top());
2129 }
2130 }
2131
2132 // Seach for MemBarReleaseLock node and delete it also.
2133 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2134 ctrl->in(0)->is_MemBar()) {
2135 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2136 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2137 mem->is_Proj() && membar == mem->in(0), "");
2138 _igvn.replace_node(fallthroughproj, ctrl);
2139 _igvn.replace_node(memproj_fallthrough, mem);
2140 fallthroughproj = ctrl;
2141 memproj_fallthrough = mem;
2142 ctrl = membar->in(TypeFunc::Control);
2143 mem = membar->in(TypeFunc::Memory);
2144 }
2145
2146 _igvn.replace_node(fallthroughproj, ctrl);
2147 _igvn.replace_node(memproj_fallthrough, mem);
2148 return true;
2149 }
2150
2151
2152 //------------------------------expand_lock_node----------------------
2153 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2154
2155 Node* ctrl = lock->in(TypeFunc::Control);
2156 Node* mem = lock->in(TypeFunc::Memory);
2157 Node* obj = lock->obj_node();
2158 Node* box = lock->box_node();
2159 Node* flock = lock->fastlock_node();
2160
2161 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2162
2163 // Make the merge point
2164 Node *region;
2165 Node *mem_phi;
2166 Node *slow_path;
2167
2168 if (UseOptoBiasInlining) {
2169 /*
2170 * See the full description in MacroAssembler::biased_locking_enter().
2171 *
2172 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2173 * // The object is biased.
2174 * proto_node = klass->prototype_header;
2175 * o_node = thread | proto_node;
2176 * x_node = o_node ^ mark_word;
2177 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2178 * // Done.
2179 * } else {
2180 * if( (x_node & biased_lock_mask) != 0 ) {
2181 * // The klass's prototype header is no longer biased.
2182 * cas(&mark_word, mark_word, proto_node)
2183 * goto cas_lock;
2184 * } else {
2185 * // The klass's prototype header is still biased.
2186 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2187 * old = mark_word;
2188 * new = o_node;
2189 * } else {
2190 * // Different thread or anonymous biased.
2191 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2192 * new = thread | old;
2193 * }
2194 * // Try to rebias.
2195 * if( cas(&mark_word, old, new) == 0 ) {
2196 * // Done.
2197 * } else {
2198 * goto slow_path; // Failed.
2199 * }
2200 * }
2201 * }
2202 * } else {
2203 * // The object is not biased.
2204 * cas_lock:
2205 * if( FastLock(obj) == 0 ) {
2206 * // Done.
2207 * } else {
2208 * slow_path:
2209 * OptoRuntime::complete_monitor_locking_Java(obj);
2210 * }
2211 * }
2212 */
2213
2214 region = new (C) RegionNode(5);
2215 // create a Phi for the memory state
2216 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2217
2218 Node* fast_lock_region = new (C) RegionNode(3);
2219 Node* fast_lock_mem_phi = new (C) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2220
2221 // First, check mark word for the biased lock pattern.
2222 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2223
2224 // Get fast path - mark word has the biased lock pattern.
2225 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2226 markOopDesc::biased_lock_mask_in_place,
2227 markOopDesc::biased_lock_pattern, true);
2228 // fast_lock_region->in(1) is set to slow path.
2229 fast_lock_mem_phi->init_req(1, mem);
2230
2231 // Now check that the lock is biased to the current thread and has
2232 // the same epoch and bias as Klass::_prototype_header.
2233
2234 // Special-case a fresh allocation to avoid building nodes:
2235 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2236 if (klass_node == NULL) {
2237 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2238 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2239 #ifdef _LP64
2240 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2241 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2242 klass_node->in(1)->init_req(0, ctrl);
2243 } else
2244 #endif
2245 klass_node->init_req(0, ctrl);
2246 }
2247 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2248
2249 Node* thread = transform_later(new (C) ThreadLocalNode());
2250 Node* cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2251 Node* o_node = transform_later(new (C) OrXNode(cast_thread, proto_node));
2252 Node* x_node = transform_later(new (C) XorXNode(o_node, mark_node));
2253
2254 // Get slow path - mark word does NOT match the value.
2255 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2256 (~markOopDesc::age_mask_in_place), 0);
2257 // region->in(3) is set to fast path - the object is biased to the current thread.
2258 mem_phi->init_req(3, mem);
2259
2260
2261 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2262
2263
2264 // First, check biased pattern.
2265 // Get fast path - _prototype_header has the same biased lock pattern.
2266 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2267 markOopDesc::biased_lock_mask_in_place, 0, true);
2268
2269 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2270 // fast_lock_region->in(2) - the prototype header is no longer biased
2271 // and we have to revoke the bias on this object.
2272 // We are going to try to reset the mark of this object to the prototype
2273 // value and fall through to the CAS-based locking scheme.
2274 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2275 Node* cas = new (C) StoreXConditionalNode(not_biased_ctrl, mem, adr,
2276 proto_node, mark_node);
2277 transform_later(cas);
2278 Node* proj = transform_later( new (C) SCMemProjNode(cas));
2279 fast_lock_mem_phi->init_req(2, proj);
2280
2281
2282 // Second, check epoch bits.
2283 Node* rebiased_region = new (C) RegionNode(3);
2284 Node* old_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2285 Node* new_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2286
2287 // Get slow path - mark word does NOT match epoch bits.
2288 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2289 markOopDesc::epoch_mask_in_place, 0);
2290 // The epoch of the current bias is not valid, attempt to rebias the object
2291 // toward the current thread.
2292 rebiased_region->init_req(2, epoch_ctrl);
2293 old_phi->init_req(2, mark_node);
2294 new_phi->init_req(2, o_node);
2295
2296 // rebiased_region->in(1) is set to fast path.
2297 // The epoch of the current bias is still valid but we know
2298 // nothing about the owner; it might be set or it might be clear.
2299 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2300 markOopDesc::age_mask_in_place |
2301 markOopDesc::epoch_mask_in_place);
2302 Node* old = transform_later(new (C) AndXNode(mark_node, cmask));
2303 cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2304 Node* new_mark = transform_later(new (C) OrXNode(cast_thread, old));
2305 old_phi->init_req(1, old);
2306 new_phi->init_req(1, new_mark);
2307
2308 transform_later(rebiased_region);
2309 transform_later(old_phi);
2310 transform_later(new_phi);
2311
2312 // Try to acquire the bias of the object using an atomic operation.
2313 // If this fails we will go in to the runtime to revoke the object's bias.
2314 cas = new (C) StoreXConditionalNode(rebiased_region, mem, adr,
2315 new_phi, old_phi);
2316 transform_later(cas);
2317 proj = transform_later( new (C) SCMemProjNode(cas));
2318
2319 // Get slow path - Failed to CAS.
2320 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2321 mem_phi->init_req(4, proj);
2322 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2323
2324 // Failed to CAS.
2325 slow_path = new (C) RegionNode(3);
2326 Node *slow_mem = new (C) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2327
2328 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2329 slow_mem->init_req(1, proj);
2330
2331 // Call CAS-based locking scheme (FastLock node).
2332
2333 transform_later(fast_lock_region);
2334 transform_later(fast_lock_mem_phi);
2335
2336 // Get slow path - FastLock failed to lock the object.
2337 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2338 mem_phi->init_req(2, fast_lock_mem_phi);
2339 // region->in(2) is set to fast path - the object is locked to the current thread.
2340
2341 slow_path->init_req(2, ctrl); // Capture slow-control
2342 slow_mem->init_req(2, fast_lock_mem_phi);
2343
2344 transform_later(slow_path);
2345 transform_later(slow_mem);
2346 // Reset lock's memory edge.
2347 lock->set_req(TypeFunc::Memory, slow_mem);
2348
2349 } else {
2350 region = new (C) RegionNode(3);
2351 // create a Phi for the memory state
2352 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2353
2354 // Optimize test; set region slot 2
2355 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2356 mem_phi->init_req(2, mem);
2357 }
2358
2359 // Make slow path call
2360 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
2361
2362 extract_call_projections(call);
2363
2364 // Slow path can only throw asynchronous exceptions, which are always
2365 // de-opted. So the compiler thinks the slow-call can never throw an
2366 // exception. If it DOES throw an exception we would need the debug
2367 // info removed first (since if it throws there is no monitor).
2368 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2369 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2370
2371 // Capture slow path
2372 // disconnect fall-through projection from call and create a new one
2373 // hook up users of fall-through projection to region
2374 Node *slow_ctrl = _fallthroughproj->clone();
2375 transform_later(slow_ctrl);
2376 _igvn.hash_delete(_fallthroughproj);
2377 _fallthroughproj->disconnect_inputs(NULL, C);
2378 region->init_req(1, slow_ctrl);
2379 // region inputs are now complete
2380 transform_later(region);
2381 _igvn.replace_node(_fallthroughproj, region);
2382
2383 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2384 mem_phi->init_req(1, memproj );
2385 transform_later(mem_phi);
2386 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2387 }
2388
2389 //------------------------------expand_unlock_node----------------------
2390 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2391
2392 Node* ctrl = unlock->in(TypeFunc::Control);
2393 Node* mem = unlock->in(TypeFunc::Memory);
2394 Node* obj = unlock->obj_node();
2395 Node* box = unlock->box_node();
2396
2397 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2398
2399 // No need for a null check on unlock
2400
2401 // Make the merge point
2402 Node *region;
2403 Node *mem_phi;
2404
2405 if (UseOptoBiasInlining) {
2406 // Check for biased locking unlock case, which is a no-op.
2407 // See the full description in MacroAssembler::biased_locking_exit().
2408 region = new (C) RegionNode(4);
2409 // create a Phi for the memory state
2410 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2411 mem_phi->init_req(3, mem);
2412
2413 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2414 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2415 markOopDesc::biased_lock_mask_in_place,
2416 markOopDesc::biased_lock_pattern);
2417 } else {
2418 region = new (C) RegionNode(3);
2419 // create a Phi for the memory state
2420 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2421 }
2422
2423 FastUnlockNode *funlock = new (C) FastUnlockNode( ctrl, obj, box );
2424 funlock = transform_later( funlock )->as_FastUnlock();
2425 // Optimize test; set region slot 2
2426 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2427
2428 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 );
2429
2430 extract_call_projections(call);
2431
2432 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2433 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2434
2435 // No exceptions for unlocking
2436 // Capture slow path
2437 // disconnect fall-through projection from call and create a new one
2438 // hook up users of fall-through projection to region
2439 Node *slow_ctrl = _fallthroughproj->clone();
2440 transform_later(slow_ctrl);
2441 _igvn.hash_delete(_fallthroughproj);
2442 _fallthroughproj->disconnect_inputs(NULL, C);
2443 region->init_req(1, slow_ctrl);
2444 // region inputs are now complete
2445 transform_later(region);
2446 _igvn.replace_node(_fallthroughproj, region);
2447
2448 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2449 mem_phi->init_req(1, memproj );
2450 mem_phi->init_req(2, mem);
2451 transform_later(mem_phi);
2452 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2453 }
2454
2455 //---------------------------eliminate_macro_nodes----------------------
2456 // Eliminate scalar replaced allocations and associated locks.
2457 void PhaseMacroExpand::eliminate_macro_nodes() {
2458 if (C->macro_count() == 0)
2459 return;
2460
2461 // First, attempt to eliminate locks
2462 int cnt = C->macro_count();
2463 for (int i=0; i < cnt; i++) {
2464 Node *n = C->macro_node(i);
2465 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2466 // Before elimination mark all associated (same box and obj)
2467 // lock and unlock nodes.
2468 mark_eliminated_locking_nodes(n->as_AbstractLock());
2469 }
2470 }
2471 bool progress = true;
2472 while (progress) {
2473 progress = false;
2474 for (int i = C->macro_count(); i > 0; i--) {
2475 Node * n = C->macro_node(i-1);
2476 bool success = false;
2477 debug_only(int old_macro_count = C->macro_count(););
2478 if (n->is_AbstractLock()) {
2479 success = eliminate_locking_node(n->as_AbstractLock());
2480 }
2481 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2482 progress = progress || success;
2483 }
2484 }
2485 // Next, attempt to eliminate allocations
2486 _has_locks = false;
2487 progress = true;
2488 while (progress) {
2489 progress = false;
2490 for (int i = C->macro_count(); i > 0; i--) {
2491 Node * n = C->macro_node(i-1);
2492 bool success = false;
2493 debug_only(int old_macro_count = C->macro_count(););
2494 switch (n->class_id()) {
2495 case Node::Class_Allocate:
2496 case Node::Class_AllocateArray:
2497 success = eliminate_allocate_node(n->as_Allocate());
2498 break;
2499 case Node::Class_CallStaticJava:
2500 success = eliminate_boxing_node(n->as_CallStaticJava());
2501 break;
2502 case Node::Class_Lock:
2503 case Node::Class_Unlock:
2504 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2505 _has_locks = true;
2506 break;
2507 default:
2508 assert(n->Opcode() == Op_LoopLimit ||
2509 n->Opcode() == Op_Opaque1 ||
2510 n->Opcode() == Op_Opaque2 ||
2511 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2512 }
2513 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2514 progress = progress || success;
2515 }
2516 }
2517 }
2518
2519 //------------------------------expand_macro_nodes----------------------
2520 // Returns true if a failure occurred.
2521 bool PhaseMacroExpand::expand_macro_nodes() {
2522 // Last attempt to eliminate macro nodes.
2523 eliminate_macro_nodes();
2524
2525 // Make sure expansion will not cause node limit to be exceeded.
2526 // Worst case is a macro node gets expanded into about 50 nodes.
2527 // Allow 50% more for optimization.
2528 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2529 return true;
2530
2531 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2532 bool progress = true;
2533 while (progress) {
2534 progress = false;
2535 for (int i = C->macro_count(); i > 0; i--) {
2536 Node * n = C->macro_node(i-1);
2537 bool success = false;
2538 debug_only(int old_macro_count = C->macro_count(););
2539 if (n->Opcode() == Op_LoopLimit) {
2540 // Remove it from macro list and put on IGVN worklist to optimize.
2541 C->remove_macro_node(n);
2542 _igvn._worklist.push(n);
2543 success = true;
2544 } else if (n->Opcode() == Op_CallStaticJava) {
2545 // Remove it from macro list and put on IGVN worklist to optimize.
2546 C->remove_macro_node(n);
2547 _igvn._worklist.push(n);
2548 success = true;
2549 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2550 _igvn.replace_node(n, n->in(1));
2551 success = true;
2552 #if INCLUDE_RTM_OPT
2553 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2554 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2555 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2556 Node* cmp = n->unique_out();
2557 #ifdef ASSERT
2558 // Validate graph.
2559 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2560 BoolNode* bol = cmp->unique_out()->as_Bool();
2561 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2562 (bol->_test._test == BoolTest::ne), "");
2563 IfNode* ifn = bol->unique_out()->as_If();
2564 assert((ifn->outcnt() == 2) &&
2565 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change), "");
2566 #endif
2567 Node* repl = n->in(1);
2568 if (!_has_locks) {
2569 // Remove RTM state check if there are no locks in the code.
2570 // Replace input to compare the same value.
2571 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2572 }
2573 _igvn.replace_node(n, repl);
2574 success = true;
2575 #endif
2576 }
2577 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2578 progress = progress || success;
2579 }
2580 }
2581
2582 // expand "macro" nodes
2583 // nodes are removed from the macro list as they are processed
2584 while (C->macro_count() > 0) {
2585 int macro_count = C->macro_count();
2586 Node * n = C->macro_node(macro_count-1);
2587 assert(n->is_macro(), "only macro nodes expected here");
2588 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2589 // node is unreachable, so don't try to expand it
2590 C->remove_macro_node(n);
2591 continue;
2592 }
2593 switch (n->class_id()) {
2594 case Node::Class_Allocate:
2595 expand_allocate(n->as_Allocate());
2596 break;
2597 case Node::Class_AllocateArray:
2598 expand_allocate_array(n->as_AllocateArray());
2599 break;
2600 case Node::Class_Lock:
2601 expand_lock_node(n->as_Lock());
2602 break;
2603 case Node::Class_Unlock:
2604 expand_unlock_node(n->as_Unlock());
2605 break;
2606 default:
2607 assert(false, "unknown node type in macro list");
2608 }
2609 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2610 if (C->failing()) return true;
2611 }
2612
2613 _igvn.set_delay_transform(false);
2614 _igvn.optimize();
2615 if (C->failing()) return true;
2616 return false;
2617 }