1 /*
2 * Copyright (c) 2005, 2022, 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 "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/rootnode.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "opto/type.hpp"
53 #include "prims/jvmtiExport.hpp"
54 #include "runtime/continuation.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #if INCLUDE_G1GC
59 #include "gc/g1/g1ThreadLocalData.hpp"
60 #endif // INCLUDE_G1GC
61 #if INCLUDE_SHENANDOAHGC
62 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
63 #endif
64
65
66 //
67 // Replace any references to "oldref" in inputs to "use" with "newref".
68 // Returns the number of replacements made.
69 //
70 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
71 int nreplacements = 0;
72 uint req = use->req();
73 for (uint j = 0; j < use->len(); j++) {
74 Node *uin = use->in(j);
75 if (uin == oldref) {
76 if (j < req)
77 use->set_req(j, newref);
78 else
79 use->set_prec(j, newref);
80 nreplacements++;
81 } else if (j >= req && uin == NULL) {
82 break;
83 }
84 }
85 return nreplacements;
86 }
87
88 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
89 assert(old != NULL, "sanity");
90 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
91 Node* use = old->fast_out(i);
92 _igvn.rehash_node_delayed(use);
93 imax -= replace_input(use, old, target);
94 // back up iterator
95 --i;
96 }
97 assert(old->outcnt() == 0, "all uses must be deleted");
98 }
99
100 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
101 Node* cmp;
102 if (mask != 0) {
103 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
104 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
105 } else {
106 cmp = word;
107 }
108 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
109 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
110 transform_later(iff);
111
112 // Fast path taken.
113 Node *fast_taken = transform_later(new IfFalseNode(iff));
114
115 // Fast path not-taken, i.e. slow path
116 Node *slow_taken = transform_later(new IfTrueNode(iff));
117
118 if (return_fast_path) {
119 region->init_req(edge, slow_taken); // Capture slow-control
120 return fast_taken;
121 } else {
122 region->init_req(edge, fast_taken); // Capture fast-control
123 return slow_taken;
124 }
125 }
126
127 //--------------------copy_predefined_input_for_runtime_call--------------------
128 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
129 // Set fixed predefined input arguments
130 call->init_req( TypeFunc::Control, ctrl );
131 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
132 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
133 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
134 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
135 }
136
137 //------------------------------make_slow_call---------------------------------
138 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
139 address slow_call, const char* leaf_name, Node* slow_path,
140 Node* parm0, Node* parm1, Node* parm2) {
141
142 // Slow-path call
143 CallNode *call = leaf_name
144 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
145 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
146
147 // Slow path call has no side-effects, uses few values
148 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
149 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
150 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
151 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
152 call->copy_call_debug_info(&_igvn, oldcall);
153 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
154 _igvn.replace_node(oldcall, call);
155 transform_later(call);
156
157 return call;
158 }
159
160 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
161 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
162 bs->eliminate_gc_barrier(this, p2x);
163 #ifndef PRODUCT
164 if (PrintOptoStatistics) {
165 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
166 }
167 #endif
168 }
169
170 // Search for a memory operation for the specified memory slice.
171 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
172 Node *orig_mem = mem;
173 Node *alloc_mem = alloc->in(TypeFunc::Memory);
174 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
175 while (true) {
176 if (mem == alloc_mem || mem == start_mem ) {
177 return mem; // hit one of our sentinels
178 } else if (mem->is_MergeMem()) {
179 mem = mem->as_MergeMem()->memory_at(alias_idx);
180 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
181 Node *in = mem->in(0);
182 // we can safely skip over safepoints, calls, locks and membars because we
183 // already know that the object is safe to eliminate.
184 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
185 return in;
186 } else if (in->is_Call()) {
187 CallNode *call = in->as_Call();
188 if (call->may_modify(tinst, phase)) {
189 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
190 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
191 return in;
192 }
193 }
194 mem = in->in(TypeFunc::Memory);
195 } else if (in->is_MemBar()) {
196 ArrayCopyNode* ac = NULL;
197 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
198 if (ac != NULL) {
199 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
200 return ac;
201 }
202 }
203 mem = in->in(TypeFunc::Memory);
204 } else {
205 #ifdef ASSERT
206 in->dump();
207 mem->dump();
208 assert(false, "unexpected projection");
209 #endif
210 }
211 } else if (mem->is_Store()) {
212 const TypePtr* atype = mem->as_Store()->adr_type();
213 int adr_idx = phase->C->get_alias_index(atype);
214 if (adr_idx == alias_idx) {
215 assert(atype->isa_oopptr(), "address type must be oopptr");
216 int adr_offset = atype->offset();
217 uint adr_iid = atype->is_oopptr()->instance_id();
218 // Array elements references have the same alias_idx
219 // but different offset and different instance_id.
220 if (adr_offset == offset && adr_iid == alloc->_idx) {
221 return mem;
222 }
223 } else {
224 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
225 }
226 mem = mem->in(MemNode::Memory);
227 } else if (mem->is_ClearArray()) {
228 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
229 // Can not bypass initialization of the instance
230 // we are looking.
231 debug_only(intptr_t offset;)
232 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
233 InitializeNode* init = alloc->as_Allocate()->initialization();
234 // We are looking for stored value, return Initialize node
235 // or memory edge from Allocate node.
236 if (init != NULL) {
237 return init;
238 } else {
239 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
240 }
241 }
242 // Otherwise skip it (the call updated 'mem' value).
243 } else if (mem->Opcode() == Op_SCMemProj) {
244 mem = mem->in(0);
245 Node* adr = NULL;
246 if (mem->is_LoadStore()) {
247 adr = mem->in(MemNode::Address);
248 } else {
249 assert(mem->Opcode() == Op_EncodeISOArray ||
250 mem->Opcode() == Op_StrCompressedCopy, "sanity");
251 adr = mem->in(3); // Destination array
252 }
253 const TypePtr* atype = adr->bottom_type()->is_ptr();
254 int adr_idx = phase->C->get_alias_index(atype);
255 if (adr_idx == alias_idx) {
256 DEBUG_ONLY(mem->dump();)
257 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
258 return NULL;
259 }
260 mem = mem->in(MemNode::Memory);
261 } else if (mem->Opcode() == Op_StrInflatedCopy) {
262 Node* adr = mem->in(3); // Destination array
263 const TypePtr* atype = adr->bottom_type()->is_ptr();
264 int adr_idx = phase->C->get_alias_index(atype);
265 if (adr_idx == alias_idx) {
266 DEBUG_ONLY(mem->dump();)
267 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
268 return NULL;
269 }
270 mem = mem->in(MemNode::Memory);
271 } else {
272 return mem;
273 }
274 assert(mem != orig_mem, "dead memory loop");
275 }
276 }
277
278 // Generate loads from source of the arraycopy for fields of
279 // destination needed at a deoptimization point
280 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
281 BasicType bt = ft;
282 const Type *type = ftype;
283 if (ft == T_NARROWOOP) {
284 bt = T_OBJECT;
285 type = ftype->make_oopptr();
286 }
287 Node* res = NULL;
288 if (ac->is_clonebasic()) {
289 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
290 Node* base = ac->in(ArrayCopyNode::Src);
291 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
292 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
293 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
294 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
295 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
296 } else {
297 if (ac->modifies(offset, offset, &_igvn, true)) {
298 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
299 uint shift = exact_log2(type2aelembytes(bt));
300 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
301 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
302 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
303 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
304
305 Node* adr = NULL;
306 const TypePtr* adr_type = NULL;
307 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
308 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
309 Node* base = ac->in(ArrayCopyNode::Src);
310 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
311 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
312 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
313 // Don't emit a new load from src if src == dst but try to get the value from memory instead
314 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
315 }
316 } else {
317 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
318 #ifdef _LP64
319 diff = _igvn.transform(new ConvI2LNode(diff));
320 #endif
321 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
322
323 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
324 Node* base = ac->in(ArrayCopyNode::Src);
325 adr = _igvn.transform(new AddPNode(base, base, off));
326 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
327 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
328 // Non constant offset in the array: we can't statically
329 // determine the value
330 return NULL;
331 }
332 }
333 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
334 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
335 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
336 }
337 }
338 if (res != NULL) {
339 if (ftype->isa_narrowoop()) {
340 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
341 res = _igvn.transform(new EncodePNode(res, ftype));
342 }
343 return res;
344 }
345 return NULL;
346 }
347
348 //
349 // Given a Memory Phi, compute a value Phi containing the values from stores
350 // on the input paths.
351 // Note: this function is recursive, its depth is limited by the "level" argument
352 // Returns the computed Phi, or NULL if it cannot compute it.
353 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
354 assert(mem->is_Phi(), "sanity");
355 int alias_idx = C->get_alias_index(adr_t);
356 int offset = adr_t->offset();
357 int instance_id = adr_t->instance_id();
358
359 // Check if an appropriate value phi already exists.
360 Node* region = mem->in(0);
361 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
362 Node* phi = region->fast_out(k);
363 if (phi->is_Phi() && phi != mem &&
364 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
365 return phi;
366 }
367 }
368 // Check if an appropriate new value phi already exists.
369 Node* new_phi = value_phis->find(mem->_idx);
370 if (new_phi != NULL)
371 return new_phi;
372
373 if (level <= 0) {
374 return NULL; // Give up: phi tree too deep
375 }
376 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
377 Node *alloc_mem = alloc->in(TypeFunc::Memory);
378
379 uint length = mem->req();
380 GrowableArray <Node *> values(length, length, NULL);
381
382 // create a new Phi for the value
383 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
384 transform_later(phi);
385 value_phis->push(phi, mem->_idx);
386
387 for (uint j = 1; j < length; j++) {
388 Node *in = mem->in(j);
389 if (in == NULL || in->is_top()) {
390 values.at_put(j, in);
391 } else {
392 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
393 if (val == start_mem || val == alloc_mem) {
394 // hit a sentinel, return appropriate 0 value
395 values.at_put(j, _igvn.zerocon(ft));
396 continue;
397 }
398 if (val->is_Initialize()) {
399 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
400 }
401 if (val == NULL) {
402 return NULL; // can't find a value on this path
403 }
404 if (val == mem) {
405 values.at_put(j, mem);
406 } else if (val->is_Store()) {
407 Node* n = val->in(MemNode::ValueIn);
408 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
409 n = bs->step_over_gc_barrier(n);
410 if (is_subword_type(ft)) {
411 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
412 }
413 values.at_put(j, n);
414 } else if(val->is_Proj() && val->in(0) == alloc) {
415 values.at_put(j, _igvn.zerocon(ft));
416 } else if (val->is_Phi()) {
417 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
418 if (val == NULL) {
419 return NULL;
420 }
421 values.at_put(j, val);
422 } else if (val->Opcode() == Op_SCMemProj) {
423 assert(val->in(0)->is_LoadStore() ||
424 val->in(0)->Opcode() == Op_EncodeISOArray ||
425 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
426 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
427 return NULL;
428 } else if (val->is_ArrayCopy()) {
429 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
430 if (res == NULL) {
431 return NULL;
432 }
433 values.at_put(j, res);
434 } else {
435 DEBUG_ONLY( val->dump(); )
436 assert(false, "unknown node on this path");
437 return NULL; // unknown node on this path
438 }
439 }
440 }
441 // Set Phi's inputs
442 for (uint j = 1; j < length; j++) {
443 if (values.at(j) == mem) {
444 phi->init_req(j, phi);
445 } else {
446 phi->init_req(j, values.at(j));
447 }
448 }
449 return phi;
450 }
451
452 // Search the last value stored into the object's field.
453 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
454 assert(adr_t->is_known_instance_field(), "instance required");
455 int instance_id = adr_t->instance_id();
456 assert((uint)instance_id == alloc->_idx, "wrong allocation");
457
458 int alias_idx = C->get_alias_index(adr_t);
459 int offset = adr_t->offset();
460 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
461 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
462 Node *alloc_mem = alloc->in(TypeFunc::Memory);
463 VectorSet visited;
464
465 bool done = sfpt_mem == alloc_mem;
466 Node *mem = sfpt_mem;
467 while (!done) {
468 if (visited.test_set(mem->_idx)) {
469 return NULL; // found a loop, give up
470 }
471 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
472 if (mem == start_mem || mem == alloc_mem) {
473 done = true; // hit a sentinel, return appropriate 0 value
474 } else if (mem->is_Initialize()) {
475 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
476 if (mem == NULL) {
477 done = true; // Something go wrong.
478 } else if (mem->is_Store()) {
479 const TypePtr* atype = mem->as_Store()->adr_type();
480 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
481 done = true;
482 }
483 } else if (mem->is_Store()) {
484 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
485 assert(atype != NULL, "address type must be oopptr");
486 assert(C->get_alias_index(atype) == alias_idx &&
487 atype->is_known_instance_field() && atype->offset() == offset &&
488 atype->instance_id() == instance_id, "store is correct memory slice");
489 done = true;
490 } else if (mem->is_Phi()) {
491 // try to find a phi's unique input
492 Node *unique_input = NULL;
493 Node *top = C->top();
494 for (uint i = 1; i < mem->req(); i++) {
495 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
496 if (n == NULL || n == top || n == mem) {
497 continue;
498 } else if (unique_input == NULL) {
499 unique_input = n;
500 } else if (unique_input != n) {
501 unique_input = top;
502 break;
503 }
504 }
505 if (unique_input != NULL && unique_input != top) {
506 mem = unique_input;
507 } else {
508 done = true;
509 }
510 } else if (mem->is_ArrayCopy()) {
511 done = true;
512 } else {
513 DEBUG_ONLY( mem->dump(); )
514 assert(false, "unexpected node");
515 }
516 }
517 if (mem != NULL) {
518 if (mem == start_mem || mem == alloc_mem) {
519 // hit a sentinel, return appropriate 0 value
520 return _igvn.zerocon(ft);
521 } else if (mem->is_Store()) {
522 Node* n = mem->in(MemNode::ValueIn);
523 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
524 n = bs->step_over_gc_barrier(n);
525 return n;
526 } else if (mem->is_Phi()) {
527 // attempt to produce a Phi reflecting the values on the input paths of the Phi
528 Node_Stack value_phis(8);
529 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
530 if (phi != NULL) {
531 return phi;
532 } else {
533 // Kill all new Phis
534 while(value_phis.is_nonempty()) {
535 Node* n = value_phis.node();
536 _igvn.replace_node(n, C->top());
537 value_phis.pop();
538 }
539 }
540 } else if (mem->is_ArrayCopy()) {
541 Node* ctl = mem->in(0);
542 Node* m = mem->in(TypeFunc::Memory);
543 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
544 // pin the loads in the uncommon trap path
545 ctl = sfpt_ctl;
546 m = sfpt_mem;
547 }
548 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
549 }
550 }
551 // Something go wrong.
552 return NULL;
553 }
554
555 // Check the possibility of scalar replacement.
556 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
557 // Scan the uses of the allocation to check for anything that would
558 // prevent us from eliminating it.
559 NOT_PRODUCT( const char* fail_eliminate = NULL; )
560 DEBUG_ONLY( Node* disq_node = NULL; )
561 bool can_eliminate = true;
562
563 Node* res = alloc->result_cast();
564 const TypeOopPtr* res_type = NULL;
565 if (res == NULL) {
566 // All users were eliminated.
567 } else if (!res->is_CheckCastPP()) {
568 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
569 can_eliminate = false;
570 } else {
571 res_type = _igvn.type(res)->isa_oopptr();
572 if (res_type == NULL) {
573 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
574 can_eliminate = false;
575 } else if (res_type->isa_aryptr()) {
576 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
577 if (length < 0) {
578 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
579 can_eliminate = false;
580 }
581 }
582 }
583
584 if (can_eliminate && res != NULL) {
585 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
586 j < jmax && can_eliminate; j++) {
587 Node* use = res->fast_out(j);
588
589 if (use->is_AddP()) {
590 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
591 int offset = addp_type->offset();
592
593 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
594 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
595 can_eliminate = false;
596 break;
597 }
598 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
599 k < kmax && can_eliminate; k++) {
600 Node* n = use->fast_out(k);
601 if (!n->is_Store() && n->Opcode() != Op_CastP2X
602 SHENANDOAHGC_ONLY(&& (!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n))) ) {
603 DEBUG_ONLY(disq_node = n;)
604 if (n->is_Load() || n->is_LoadStore()) {
605 NOT_PRODUCT(fail_eliminate = "Field load";)
606 } else {
607 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
608 }
609 can_eliminate = false;
610 }
611 }
612 } else if (use->is_ArrayCopy() &&
613 (use->as_ArrayCopy()->is_clonebasic() ||
614 use->as_ArrayCopy()->is_arraycopy_validated() ||
615 use->as_ArrayCopy()->is_copyof_validated() ||
616 use->as_ArrayCopy()->is_copyofrange_validated()) &&
617 use->in(ArrayCopyNode::Dest) == res) {
618 // ok to eliminate
619 } else if (use->is_SafePoint()) {
620 SafePointNode* sfpt = use->as_SafePoint();
621 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
622 // Object is passed as argument.
623 DEBUG_ONLY(disq_node = use;)
624 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
625 can_eliminate = false;
626 }
627 Node* sfptMem = sfpt->memory();
628 if (sfptMem == NULL || sfptMem->is_top()) {
629 DEBUG_ONLY(disq_node = use;)
630 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
631 can_eliminate = false;
632 } else {
633 safepoints.append_if_missing(sfpt);
634 }
635 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
636 if (use->is_Phi()) {
637 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
638 NOT_PRODUCT(fail_eliminate = "Object is return value";)
639 } else {
640 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
641 }
642 DEBUG_ONLY(disq_node = use;)
643 } else {
644 if (use->Opcode() == Op_Return) {
645 NOT_PRODUCT(fail_eliminate = "Object is return value";)
646 }else {
647 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
648 }
649 DEBUG_ONLY(disq_node = use;)
650 }
651 can_eliminate = false;
652 }
653 }
654 }
655
656 #ifndef PRODUCT
657 if (PrintEliminateAllocations) {
658 if (can_eliminate) {
659 tty->print("Scalar ");
660 if (res == NULL)
661 alloc->dump();
662 else
663 res->dump();
664 } else if (alloc->_is_scalar_replaceable) {
665 tty->print("NotScalar (%s)", fail_eliminate);
666 if (res == NULL)
667 alloc->dump();
668 else
669 res->dump();
670 #ifdef ASSERT
671 if (disq_node != NULL) {
672 tty->print(" >>>> ");
673 disq_node->dump();
674 }
675 #endif /*ASSERT*/
676 }
677 }
678 #endif
679 return can_eliminate;
680 }
681
682 // Do scalar replacement.
683 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
684 GrowableArray <SafePointNode *> safepoints_done;
685
686 ciInstanceKlass* iklass = NULL;
687 int nfields = 0;
688 int array_base = 0;
689 int element_size = 0;
690 BasicType basic_elem_type = T_ILLEGAL;
691 const Type* field_type = NULL;
692
693 Node* res = alloc->result_cast();
694 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
695 const TypeOopPtr* res_type = NULL;
696 if (res != NULL) { // Could be NULL when there are no users
697 res_type = _igvn.type(res)->isa_oopptr();
698 }
699
700 if (res != NULL) {
701 if (res_type->isa_instptr()) {
702 // find the fields of the class which will be needed for safepoint debug information
703 iklass = res_type->is_instptr()->instance_klass();
704 nfields = iklass->nof_nonstatic_fields();
705 } else {
706 // find the array's elements which will be needed for safepoint debug information
707 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
708 assert(nfields >= 0, "must be an array klass.");
709 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
710 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
711 element_size = type2aelembytes(basic_elem_type);
712 field_type = res_type->is_aryptr()->elem();
713 }
714 }
715 //
716 // Process the safepoint uses
717 //
718 while (safepoints.length() > 0) {
719 SafePointNode* sfpt = safepoints.pop();
720 Node* mem = sfpt->memory();
721 Node* ctl = sfpt->control();
722 assert(sfpt->jvms() != NULL, "missed JVMS");
723 // Fields of scalar objs are referenced only at the end
724 // of regular debuginfo at the last (youngest) JVMS.
725 // Record relative start index.
726 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
727 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
728 #ifdef ASSERT
729 alloc,
730 #endif
731 first_ind, nfields);
732 sobj->init_req(0, C->root());
733 transform_later(sobj);
734
735 // Scan object's fields adding an input to the safepoint for each field.
736 for (int j = 0; j < nfields; j++) {
737 intptr_t offset;
738 ciField* field = NULL;
739 if (iklass != NULL) {
740 field = iklass->nonstatic_field_at(j);
741 offset = field->offset();
742 ciType* elem_type = field->type();
743 basic_elem_type = field->layout_type();
744
745 // The next code is taken from Parse::do_get_xxx().
746 if (is_reference_type(basic_elem_type)) {
747 if (!elem_type->is_loaded()) {
748 field_type = TypeInstPtr::BOTTOM;
749 } else if (field != NULL && field->is_static_constant()) {
750 ciObject* con = field->constant_value().as_object();
751 // Do not "join" in the previous type; it doesn't add value,
752 // and may yield a vacuous result if the field is of interface type.
753 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
754 assert(field_type != NULL, "field singleton type must be consistent");
755 } else {
756 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
757 }
758 if (UseCompressedOops) {
759 field_type = field_type->make_narrowoop();
760 basic_elem_type = T_NARROWOOP;
761 }
762 } else {
763 field_type = Type::get_const_basic_type(basic_elem_type);
764 }
765 } else {
766 offset = array_base + j * (intptr_t)element_size;
767 }
768
769 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
770
771 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
772 if (field_val == NULL) {
773 // We weren't able to find a value for this field,
774 // give up on eliminating this allocation.
775
776 // Remove any extra entries we added to the safepoint.
777 uint last = sfpt->req() - 1;
778 for (int k = 0; k < j; k++) {
779 sfpt->del_req(last--);
780 }
781 _igvn._worklist.push(sfpt);
782 // rollback processed safepoints
783 while (safepoints_done.length() > 0) {
784 SafePointNode* sfpt_done = safepoints_done.pop();
785 // remove any extra entries we added to the safepoint
786 last = sfpt_done->req() - 1;
787 for (int k = 0; k < nfields; k++) {
788 sfpt_done->del_req(last--);
789 }
790 JVMState *jvms = sfpt_done->jvms();
791 jvms->set_endoff(sfpt_done->req());
792 // Now make a pass over the debug information replacing any references
793 // to SafePointScalarObjectNode with the allocated object.
794 int start = jvms->debug_start();
795 int end = jvms->debug_end();
796 for (int i = start; i < end; i++) {
797 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
798 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
799 if (scobj->first_index(jvms) == sfpt_done->req() &&
800 scobj->n_fields() == (uint)nfields) {
801 assert(scobj->alloc() == alloc, "sanity");
802 sfpt_done->set_req(i, res);
803 }
804 }
805 }
806 _igvn._worklist.push(sfpt_done);
807 }
808 #ifndef PRODUCT
809 if (PrintEliminateAllocations) {
810 if (field != NULL) {
811 tty->print("=== At SafePoint node %d can't find value of Field: ",
812 sfpt->_idx);
813 field->print();
814 int field_idx = C->get_alias_index(field_addr_type);
815 tty->print(" (alias_idx=%d)", field_idx);
816 } else { // Array's element
817 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
818 sfpt->_idx, j);
819 }
820 tty->print(", which prevents elimination of: ");
821 if (res == NULL)
822 alloc->dump();
823 else
824 res->dump();
825 }
826 #endif
827 return false;
828 }
829 if (UseCompressedOops && field_type->isa_narrowoop()) {
830 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
831 // to be able scalar replace the allocation.
832 if (field_val->is_EncodeP()) {
833 field_val = field_val->in(1);
834 } else {
835 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
836 }
837 }
838 sfpt->add_req(field_val);
839 }
840 JVMState *jvms = sfpt->jvms();
841 jvms->set_endoff(sfpt->req());
842 // Now make a pass over the debug information replacing any references
843 // to the allocated object with "sobj"
844 int start = jvms->debug_start();
845 int end = jvms->debug_end();
846 sfpt->replace_edges_in_range(res, sobj, start, end, &_igvn);
847 _igvn._worklist.push(sfpt);
848 safepoints_done.append_if_missing(sfpt); // keep it for rollback
849 }
850 return true;
851 }
852
853 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
854 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
855 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
856 if (ctl_proj != NULL) {
857 igvn.replace_node(ctl_proj, n->in(0));
858 }
859 if (mem_proj != NULL) {
860 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
861 }
862 }
863
864 // Process users of eliminated allocation.
865 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
866 Node* res = alloc->result_cast();
867 if (res != NULL) {
868 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
869 Node *use = res->last_out(j);
870 uint oc1 = res->outcnt();
871
872 if (use->is_AddP()) {
873 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
874 Node *n = use->last_out(k);
875 uint oc2 = use->outcnt();
876 if (n->is_Store()) {
877 #ifdef ASSERT
878 // Verify that there is no dependent MemBarVolatile nodes,
879 // they should be removed during IGVN, see MemBarNode::Ideal().
880 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
881 p < pmax; p++) {
882 Node* mb = n->fast_out(p);
883 assert(mb->is_Initialize() || !mb->is_MemBar() ||
884 mb->req() <= MemBarNode::Precedent ||
885 mb->in(MemBarNode::Precedent) != n,
886 "MemBarVolatile should be eliminated for non-escaping object");
887 }
888 #endif
889 _igvn.replace_node(n, n->in(MemNode::Memory));
890 } else {
891 eliminate_gc_barrier(n);
892 }
893 k -= (oc2 - use->outcnt());
894 }
895 _igvn.remove_dead_node(use);
896 } else if (use->is_ArrayCopy()) {
897 // Disconnect ArrayCopy node
898 ArrayCopyNode* ac = use->as_ArrayCopy();
899 if (ac->is_clonebasic()) {
900 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
901 disconnect_projections(ac, _igvn);
902 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
903 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
904 disconnect_projections(membar_before->as_MemBar(), _igvn);
905 if (membar_after->is_MemBar()) {
906 disconnect_projections(membar_after->as_MemBar(), _igvn);
907 }
908 } else {
909 assert(ac->is_arraycopy_validated() ||
910 ac->is_copyof_validated() ||
911 ac->is_copyofrange_validated(), "unsupported");
912 CallProjections callprojs;
913 ac->extract_projections(&callprojs, true);
914
915 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
916 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
917 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
918
919 // Set control to top. IGVN will remove the remaining projections
920 ac->set_req(0, top());
921 ac->replace_edge(res, top(), &_igvn);
922
923 // Disconnect src right away: it can help find new
924 // opportunities for allocation elimination
925 Node* src = ac->in(ArrayCopyNode::Src);
926 ac->replace_edge(src, top(), &_igvn);
927 // src can be top at this point if src and dest of the
928 // arraycopy were the same
929 if (src->outcnt() == 0 && !src->is_top()) {
930 _igvn.remove_dead_node(src);
931 }
932 }
933 _igvn._worklist.push(ac);
934 } else {
935 eliminate_gc_barrier(use);
936 }
937 j -= (oc1 - res->outcnt());
938 }
939 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
940 _igvn.remove_dead_node(res);
941 }
942
943 //
944 // Process other users of allocation's projections
945 //
946 if (_callprojs.resproj != NULL && _callprojs.resproj->outcnt() != 0) {
947 // First disconnect stores captured by Initialize node.
948 // If Initialize node is eliminated first in the following code,
949 // it will kill such stores and DUIterator_Last will assert.
950 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
951 Node* use = _callprojs.resproj->fast_out(j);
952 if (use->is_AddP()) {
953 // raw memory addresses used only by the initialization
954 _igvn.replace_node(use, C->top());
955 --j; --jmax;
956 }
957 }
958 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
959 Node* use = _callprojs.resproj->last_out(j);
960 uint oc1 = _callprojs.resproj->outcnt();
961 if (use->is_Initialize()) {
962 // Eliminate Initialize node.
963 InitializeNode *init = use->as_Initialize();
964 assert(init->outcnt() <= 2, "only a control and memory projection expected");
965 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
966 if (ctrl_proj != NULL) {
967 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
968 #ifdef ASSERT
969 // If the InitializeNode has no memory out, it will die, and tmp will become NULL
970 Node* tmp = init->in(TypeFunc::Control);
971 assert(tmp == NULL || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
972 #endif
973 }
974 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
975 if (mem_proj != NULL) {
976 Node *mem = init->in(TypeFunc::Memory);
977 #ifdef ASSERT
978 if (mem->is_MergeMem()) {
979 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
980 } else {
981 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
982 }
983 #endif
984 _igvn.replace_node(mem_proj, mem);
985 }
986 } else {
987 assert(false, "only Initialize or AddP expected");
988 }
989 j -= (oc1 - _callprojs.resproj->outcnt());
990 }
991 }
992 if (_callprojs.fallthrough_catchproj != NULL) {
993 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
994 }
995 if (_callprojs.fallthrough_memproj != NULL) {
996 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
997 }
998 if (_callprojs.catchall_memproj != NULL) {
999 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1000 }
1001 if (_callprojs.fallthrough_ioproj != NULL) {
1002 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1003 }
1004 if (_callprojs.catchall_ioproj != NULL) {
1005 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1006 }
1007 if (_callprojs.catchall_catchproj != NULL) {
1008 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1009 }
1010 }
1011
1012 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1013 // If reallocation fails during deoptimization we'll pop all
1014 // interpreter frames for this compiled frame and that won't play
1015 // nice with JVMTI popframe.
1016 // We avoid this issue by eager reallocation when the popframe request
1017 // is received.
1018 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1019 return false;
1020 }
1021 Node* klass = alloc->in(AllocateNode::KlassNode);
1022 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1023 Node* res = alloc->result_cast();
1024 // Eliminate boxing allocations which are not used
1025 // regardless scalar replaceable status.
1026 bool boxing_alloc = C->eliminate_boxing() &&
1027 tklass->isa_instklassptr() &&
1028 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1029 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1030 return false;
1031 }
1032
1033 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1034
1035 GrowableArray <SafePointNode *> safepoints;
1036 if (!can_eliminate_allocation(alloc, safepoints)) {
1037 return false;
1038 }
1039
1040 if (!alloc->_is_scalar_replaceable) {
1041 assert(res == NULL, "sanity");
1042 // We can only eliminate allocation if all debug info references
1043 // are already replaced with SafePointScalarObject because
1044 // we can't search for a fields value without instance_id.
1045 if (safepoints.length() > 0) {
1046 return false;
1047 }
1048 }
1049
1050 if (!scalar_replacement(alloc, safepoints)) {
1051 return false;
1052 }
1053
1054 CompileLog* log = C->log();
1055 if (log != NULL) {
1056 log->head("eliminate_allocation type='%d'",
1057 log->identify(tklass->exact_klass()));
1058 JVMState* p = alloc->jvms();
1059 while (p != NULL) {
1060 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1061 p = p->caller();
1062 }
1063 log->tail("eliminate_allocation");
1064 }
1065
1066 process_users_of_allocation(alloc);
1067
1068 #ifndef PRODUCT
1069 if (PrintEliminateAllocations) {
1070 if (alloc->is_AllocateArray())
1071 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1072 else
1073 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1074 }
1075 #endif
1076
1077 return true;
1078 }
1079
1080 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1081 // EA should remove all uses of non-escaping boxing node.
1082 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1083 return false;
1084 }
1085
1086 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1087
1088 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1089
1090 const TypeTuple* r = boxing->tf()->range();
1091 assert(r->cnt() > TypeFunc::Parms, "sanity");
1092 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1093 assert(t != NULL, "sanity");
1094
1095 CompileLog* log = C->log();
1096 if (log != NULL) {
1097 log->head("eliminate_boxing type='%d'",
1098 log->identify(t->instance_klass()));
1099 JVMState* p = boxing->jvms();
1100 while (p != NULL) {
1101 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1102 p = p->caller();
1103 }
1104 log->tail("eliminate_boxing");
1105 }
1106
1107 process_users_of_allocation(boxing);
1108
1109 #ifndef PRODUCT
1110 if (PrintEliminateAllocations) {
1111 tty->print("++++ Eliminated: %d ", boxing->_idx);
1112 boxing->method()->print_short_name(tty);
1113 tty->cr();
1114 }
1115 #endif
1116
1117 return true;
1118 }
1119
1120
1121 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1122 Node* adr = basic_plus_adr(base, offset);
1123 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1124 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1125 transform_later(value);
1126 return value;
1127 }
1128
1129
1130 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1131 Node* adr = basic_plus_adr(base, offset);
1132 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1133 transform_later(mem);
1134 return mem;
1135 }
1136
1137 //=============================================================================
1138 //
1139 // A L L O C A T I O N
1140 //
1141 // Allocation attempts to be fast in the case of frequent small objects.
1142 // It breaks down like this:
1143 //
1144 // 1) Size in doublewords is computed. This is a constant for objects and
1145 // variable for most arrays. Doubleword units are used to avoid size
1146 // overflow of huge doubleword arrays. We need doublewords in the end for
1147 // rounding.
1148 //
1149 // 2) Size is checked for being 'too large'. Too-large allocations will go
1150 // the slow path into the VM. The slow path can throw any required
1151 // exceptions, and does all the special checks for very large arrays. The
1152 // size test can constant-fold away for objects. For objects with
1153 // finalizers it constant-folds the otherway: you always go slow with
1154 // finalizers.
1155 //
1156 // 3) If NOT using TLABs, this is the contended loop-back point.
1157 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1158 //
1159 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1160 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1161 // "size*8" we always enter the VM, where "largish" is a constant picked small
1162 // enough that there's always space between the eden max and 4Gig (old space is
1163 // there so it's quite large) and large enough that the cost of entering the VM
1164 // is dwarfed by the cost to initialize the space.
1165 //
1166 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1167 // down. If contended, repeat at step 3. If using TLABs normal-store
1168 // adjusted heap top back down; there is no contention.
1169 //
1170 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1171 // fields.
1172 //
1173 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1174 // oop flavor.
1175 //
1176 //=============================================================================
1177 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1178 // Allocations bigger than this always go the slow route.
1179 // This value must be small enough that allocation attempts that need to
1180 // trigger exceptions go the slow route. Also, it must be small enough so
1181 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1182 //=============================================================================j//
1183 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1184 // The allocator will coalesce int->oop copies away. See comment in
1185 // coalesce.cpp about how this works. It depends critically on the exact
1186 // code shape produced here, so if you are changing this code shape
1187 // make sure the GC info for the heap-top is correct in and around the
1188 // slow-path call.
1189 //
1190
1191 void PhaseMacroExpand::expand_allocate_common(
1192 AllocateNode* alloc, // allocation node to be expanded
1193 Node* length, // array length for an array allocation
1194 const TypeFunc* slow_call_type, // Type of slow call
1195 address slow_call_address // Address of slow call
1196 )
1197 {
1198 Node* ctrl = alloc->in(TypeFunc::Control);
1199 Node* mem = alloc->in(TypeFunc::Memory);
1200 Node* i_o = alloc->in(TypeFunc::I_O);
1201 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1202 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1203 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1204 assert(ctrl != NULL, "must have control");
1205
1206 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1207 // they will not be used if "always_slow" is set
1208 enum { slow_result_path = 1, fast_result_path = 2 };
1209 Node *result_region = NULL;
1210 Node *result_phi_rawmem = NULL;
1211 Node *result_phi_rawoop = NULL;
1212 Node *result_phi_i_o = NULL;
1213
1214 // The initial slow comparison is a size check, the comparison
1215 // we want to do is a BoolTest::gt
1216 bool expand_fast_path = true;
1217 int tv = _igvn.find_int_con(initial_slow_test, -1);
1218 if (tv >= 0) {
1219 // InitialTest has constant result
1220 // 0 - can fit in TLAB
1221 // 1 - always too big or negative
1222 assert(tv <= 1, "0 or 1 if a constant");
1223 expand_fast_path = (tv == 0);
1224 initial_slow_test = NULL;
1225 } else {
1226 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1227 }
1228
1229 if (!UseTLAB) {
1230 // Force slow-path allocation
1231 expand_fast_path = false;
1232 initial_slow_test = NULL;
1233 }
1234
1235 bool allocation_has_use = (alloc->result_cast() != NULL);
1236 if (!allocation_has_use) {
1237 InitializeNode* init = alloc->initialization();
1238 if (init != NULL) {
1239 init->remove(&_igvn);
1240 }
1241 if (expand_fast_path && (initial_slow_test == NULL)) {
1242 // Remove allocation node and return.
1243 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1244 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1245 #ifndef PRODUCT
1246 if (PrintEliminateAllocations) {
1247 tty->print("NotUsed ");
1248 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1249 if (res != NULL) {
1250 res->dump();
1251 } else {
1252 alloc->dump();
1253 }
1254 }
1255 #endif
1256 yank_alloc_node(alloc);
1257 return;
1258 }
1259 }
1260
1261 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1262 Node *slow_region = NULL;
1263 Node *toobig_false = ctrl;
1264
1265 // generate the initial test if necessary
1266 if (initial_slow_test != NULL ) {
1267 assert (expand_fast_path, "Only need test if there is a fast path");
1268 slow_region = new RegionNode(3);
1269
1270 // Now make the initial failure test. Usually a too-big test but
1271 // might be a TRUE for finalizers or a fancy class check for
1272 // newInstance0.
1273 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1274 transform_later(toobig_iff);
1275 // Plug the failing-too-big test into the slow-path region
1276 Node *toobig_true = new IfTrueNode( toobig_iff );
1277 transform_later(toobig_true);
1278 slow_region ->init_req( too_big_or_final_path, toobig_true );
1279 toobig_false = new IfFalseNode( toobig_iff );
1280 transform_later(toobig_false);
1281 } else {
1282 // No initial test, just fall into next case
1283 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1284 toobig_false = ctrl;
1285 debug_only(slow_region = NodeSentinel);
1286 }
1287
1288 // If we are here there are several possibilities
1289 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1290 // no_initial_check means a constant allocation.
1291 // - If check always evaluates to false -> expand_fast_path is false (see above)
1292 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1293 // if !allocation_has_use the fast path is empty
1294 // if !allocation_has_use && no_initial_check
1295 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1296 // removed by yank_alloc_node above.
1297
1298 Node *slow_mem = mem; // save the current memory state for slow path
1299 // generate the fast allocation code unless we know that the initial test will always go slow
1300 if (expand_fast_path) {
1301 // Fast path modifies only raw memory.
1302 if (mem->is_MergeMem()) {
1303 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1304 }
1305
1306 // allocate the Region and Phi nodes for the result
1307 result_region = new RegionNode(3);
1308 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1309 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1310
1311 // Grab regular I/O before optional prefetch may change it.
1312 // Slow-path does no I/O so just set it to the original I/O.
1313 result_phi_i_o->init_req(slow_result_path, i_o);
1314
1315 // Name successful fast-path variables
1316 Node* fast_oop_ctrl;
1317 Node* fast_oop_rawmem;
1318 if (allocation_has_use) {
1319 Node* needgc_ctrl = NULL;
1320 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1321
1322 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1323 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1324 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1325 fast_oop_ctrl, fast_oop_rawmem,
1326 prefetch_lines);
1327
1328 if (initial_slow_test != NULL) {
1329 // This completes all paths into the slow merge point
1330 slow_region->init_req(need_gc_path, needgc_ctrl);
1331 transform_later(slow_region);
1332 } else {
1333 // No initial slow path needed!
1334 // Just fall from the need-GC path straight into the VM call.
1335 slow_region = needgc_ctrl;
1336 }
1337
1338 InitializeNode* init = alloc->initialization();
1339 fast_oop_rawmem = initialize_object(alloc,
1340 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1341 klass_node, length, size_in_bytes);
1342 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1343 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1344
1345 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1346 } else {
1347 assert (initial_slow_test != NULL, "sanity");
1348 fast_oop_ctrl = toobig_false;
1349 fast_oop_rawmem = mem;
1350 transform_later(slow_region);
1351 }
1352
1353 // Plug in the successful fast-path into the result merge point
1354 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1355 result_phi_i_o ->init_req(fast_result_path, i_o);
1356 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1357 } else {
1358 slow_region = ctrl;
1359 result_phi_i_o = i_o; // Rename it to use in the following code.
1360 }
1361
1362 // Generate slow-path call
1363 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1364 OptoRuntime::stub_name(slow_call_address),
1365 TypePtr::BOTTOM);
1366 call->init_req(TypeFunc::Control, slow_region);
1367 call->init_req(TypeFunc::I_O, top()); // does no i/o
1368 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1369 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1370 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1371
1372 call->init_req(TypeFunc::Parms+0, klass_node);
1373 if (length != NULL) {
1374 call->init_req(TypeFunc::Parms+1, length);
1375 }
1376
1377 // Copy debug information and adjust JVMState information, then replace
1378 // allocate node with the call
1379 call->copy_call_debug_info(&_igvn, alloc);
1380 if (expand_fast_path) {
1381 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1382 } else {
1383 // Hook i_o projection to avoid its elimination during allocation
1384 // replacement (when only a slow call is generated).
1385 call->set_req(TypeFunc::I_O, result_phi_i_o);
1386 }
1387 _igvn.replace_node(alloc, call);
1388 transform_later(call);
1389
1390 // Identify the output projections from the allocate node and
1391 // adjust any references to them.
1392 // The control and io projections look like:
1393 //
1394 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1395 // Allocate Catch
1396 // ^---Proj(io) <-------+ ^---CatchProj(io)
1397 //
1398 // We are interested in the CatchProj nodes.
1399 //
1400 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1401
1402 // An allocate node has separate memory projections for the uses on
1403 // the control and i_o paths. Replace the control memory projection with
1404 // result_phi_rawmem (unless we are only generating a slow call when
1405 // both memory projections are combined)
1406 if (expand_fast_path && _callprojs.fallthrough_memproj != NULL) {
1407 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1408 }
1409 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1410 // catchall_memproj so we end up with a call that has only 1 memory projection.
1411 if (_callprojs.catchall_memproj != NULL ) {
1412 if (_callprojs.fallthrough_memproj == NULL) {
1413 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1414 transform_later(_callprojs.fallthrough_memproj);
1415 }
1416 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1417 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1418 }
1419
1420 // An allocate node has separate i_o projections for the uses on the control
1421 // and i_o paths. Always replace the control i_o projection with result i_o
1422 // otherwise incoming i_o become dead when only a slow call is generated
1423 // (it is different from memory projections where both projections are
1424 // combined in such case).
1425 if (_callprojs.fallthrough_ioproj != NULL) {
1426 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1427 }
1428 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1429 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1430 if (_callprojs.catchall_ioproj != NULL ) {
1431 if (_callprojs.fallthrough_ioproj == NULL) {
1432 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1433 transform_later(_callprojs.fallthrough_ioproj);
1434 }
1435 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1436 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1437 }
1438
1439 // if we generated only a slow call, we are done
1440 if (!expand_fast_path) {
1441 // Now we can unhook i_o.
1442 if (result_phi_i_o->outcnt() > 1) {
1443 call->set_req(TypeFunc::I_O, top());
1444 } else {
1445 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1446 // Case of new array with negative size known during compilation.
1447 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1448 // following code since call to runtime will throw exception.
1449 // As result there will be no users of i_o after the call.
1450 // Leave i_o attached to this call to avoid problems in preceding graph.
1451 }
1452 return;
1453 }
1454
1455 if (_callprojs.fallthrough_catchproj != NULL) {
1456 ctrl = _callprojs.fallthrough_catchproj->clone();
1457 transform_later(ctrl);
1458 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1459 } else {
1460 ctrl = top();
1461 }
1462 Node *slow_result;
1463 if (_callprojs.resproj == NULL) {
1464 // no uses of the allocation result
1465 slow_result = top();
1466 } else {
1467 slow_result = _callprojs.resproj->clone();
1468 transform_later(slow_result);
1469 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1470 }
1471
1472 // Plug slow-path into result merge point
1473 result_region->init_req( slow_result_path, ctrl);
1474 transform_later(result_region);
1475 if (allocation_has_use) {
1476 result_phi_rawoop->init_req(slow_result_path, slow_result);
1477 transform_later(result_phi_rawoop);
1478 }
1479 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1480 transform_later(result_phi_rawmem);
1481 transform_later(result_phi_i_o);
1482 // This completes all paths into the result merge point
1483 }
1484
1485 // Remove alloc node that has no uses.
1486 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1487 Node* ctrl = alloc->in(TypeFunc::Control);
1488 Node* mem = alloc->in(TypeFunc::Memory);
1489 Node* i_o = alloc->in(TypeFunc::I_O);
1490
1491 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1492 if (_callprojs.resproj != NULL) {
1493 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1494 Node* use = _callprojs.resproj->fast_out(i);
1495 use->isa_MemBar()->remove(&_igvn);
1496 --imax;
1497 --i; // back up iterator
1498 }
1499 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1500 _igvn.remove_dead_node(_callprojs.resproj);
1501 }
1502 if (_callprojs.fallthrough_catchproj != NULL) {
1503 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1504 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1505 }
1506 if (_callprojs.catchall_catchproj != NULL) {
1507 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1508 _callprojs.catchall_catchproj->set_req(0, top());
1509 }
1510 if (_callprojs.fallthrough_proj != NULL) {
1511 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1512 _igvn.remove_dead_node(catchnode);
1513 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1514 }
1515 if (_callprojs.fallthrough_memproj != NULL) {
1516 migrate_outs(_callprojs.fallthrough_memproj, mem);
1517 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1518 }
1519 if (_callprojs.fallthrough_ioproj != NULL) {
1520 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1521 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1522 }
1523 if (_callprojs.catchall_memproj != NULL) {
1524 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1525 _callprojs.catchall_memproj->set_req(0, top());
1526 }
1527 if (_callprojs.catchall_ioproj != NULL) {
1528 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1529 _callprojs.catchall_ioproj->set_req(0, top());
1530 }
1531 #ifndef PRODUCT
1532 if (PrintEliminateAllocations) {
1533 if (alloc->is_AllocateArray()) {
1534 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1535 } else {
1536 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1537 }
1538 }
1539 #endif
1540 _igvn.remove_dead_node(alloc);
1541 }
1542
1543 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1544 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1545 // If initialization is performed by an array copy, any required
1546 // MemBarStoreStore was already added. If the object does not
1547 // escape no need for a MemBarStoreStore. If the object does not
1548 // escape in its initializer and memory barrier (MemBarStoreStore or
1549 // stronger) is already added at exit of initializer, also no need
1550 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1551 // so that stores that initialize this object can't be reordered
1552 // with a subsequent store that makes this object accessible by
1553 // other threads.
1554 // Other threads include java threads and JVM internal threads
1555 // (for example concurrent GC threads). Current concurrent GC
1556 // implementation: G1 will not scan newly created object,
1557 // so it's safe to skip storestore barrier when allocation does
1558 // not escape.
1559 if (!alloc->does_not_escape_thread() &&
1560 !alloc->is_allocation_MemBar_redundant() &&
1561 (init == NULL || !init->is_complete_with_arraycopy())) {
1562 if (init == NULL || init->req() < InitializeNode::RawStores) {
1563 // No InitializeNode or no stores captured by zeroing
1564 // elimination. Simply add the MemBarStoreStore after object
1565 // initialization.
1566 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1567 transform_later(mb);
1568
1569 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1570 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1571 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1572 transform_later(fast_oop_ctrl);
1573 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1574 transform_later(fast_oop_rawmem);
1575 } else {
1576 // Add the MemBarStoreStore after the InitializeNode so that
1577 // all stores performing the initialization that were moved
1578 // before the InitializeNode happen before the storestore
1579 // barrier.
1580
1581 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1582 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1583
1584 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1585 transform_later(mb);
1586
1587 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1588 transform_later(ctrl);
1589 Node* mem = new ProjNode(init, TypeFunc::Memory);
1590 transform_later(mem);
1591
1592 // The MemBarStoreStore depends on control and memory coming
1593 // from the InitializeNode
1594 mb->init_req(TypeFunc::Memory, mem);
1595 mb->init_req(TypeFunc::Control, ctrl);
1596
1597 ctrl = new ProjNode(mb, TypeFunc::Control);
1598 transform_later(ctrl);
1599 mem = new ProjNode(mb, TypeFunc::Memory);
1600 transform_later(mem);
1601
1602 // All nodes that depended on the InitializeNode for control
1603 // and memory must now depend on the MemBarNode that itself
1604 // depends on the InitializeNode
1605 if (init_ctrl != NULL) {
1606 _igvn.replace_node(init_ctrl, ctrl);
1607 }
1608 if (init_mem != NULL) {
1609 _igvn.replace_node(init_mem, mem);
1610 }
1611 }
1612 }
1613 }
1614
1615 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1616 Node*& ctrl, Node*& rawmem) {
1617 if (C->env()->dtrace_alloc_probes()) {
1618 // Slow-path call
1619 int size = TypeFunc::Parms + 2;
1620 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1621 CAST_FROM_FN_PTR(address,
1622 static_cast<int (*)(Thread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1623 "dtrace_object_alloc",
1624 TypeRawPtr::BOTTOM);
1625
1626 // Get base of thread-local storage area
1627 Node* thread = new ThreadLocalNode();
1628 transform_later(thread);
1629
1630 call->init_req(TypeFunc::Parms + 0, thread);
1631 call->init_req(TypeFunc::Parms + 1, oop);
1632 call->init_req(TypeFunc::Control, ctrl);
1633 call->init_req(TypeFunc::I_O , top()); // does no i/o
1634 call->init_req(TypeFunc::Memory , rawmem);
1635 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1636 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1637 transform_later(call);
1638 ctrl = new ProjNode(call, TypeFunc::Control);
1639 transform_later(ctrl);
1640 rawmem = new ProjNode(call, TypeFunc::Memory);
1641 transform_later(rawmem);
1642 }
1643 }
1644
1645 // Helper for PhaseMacroExpand::expand_allocate_common.
1646 // Initializes the newly-allocated storage.
1647 Node*
1648 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1649 Node* control, Node* rawmem, Node* object,
1650 Node* klass_node, Node* length,
1651 Node* size_in_bytes) {
1652 InitializeNode* init = alloc->initialization();
1653 // Store the klass & mark bits
1654 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1655 if (!mark_node->is_Con()) {
1656 transform_later(mark_node);
1657 }
1658 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1659
1660 #ifndef _LP64
1661 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1662 #endif
1663 int header_size = alloc->minimum_header_size(); // conservatively small
1664
1665 // Array length
1666 if (length != NULL) { // Arrays need length field
1667 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1668 // conservatively small header size:
1669 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1670 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1671 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1672 if (is_reference_type(elem, true)) {
1673 elem = T_OBJECT;
1674 }
1675 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1676 }
1677 }
1678
1679 // Clear the object body, if necessary.
1680 if (init == NULL) {
1681 // The init has somehow disappeared; be cautious and clear everything.
1682 //
1683 // This can happen if a node is allocated but an uncommon trap occurs
1684 // immediately. In this case, the Initialize gets associated with the
1685 // trap, and may be placed in a different (outer) loop, if the Allocate
1686 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1687 // there can be two Allocates to one Initialize. The answer in all these
1688 // edge cases is safety first. It is always safe to clear immediately
1689 // within an Allocate, and then (maybe or maybe not) clear some more later.
1690 if (!(UseTLAB && ZeroTLAB)) {
1691 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1692 header_size, size_in_bytes,
1693 &_igvn);
1694 }
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 intx lines) {
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 RegionNode(3);
1724 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1725 TypeRawPtr::BOTTOM );
1726 // I/O is used for Prefetch
1727 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1728
1729 Node *thread = new ThreadLocalNode();
1730 transform_later(thread);
1731
1732 Node *eden_pf_adr = new 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 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 CmpPNode( new_eden_top, old_pf_wm );
1744 transform_later(need_pf_cmp);
1745 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1746 transform_later(need_pf_bol);
1747 IfNode *need_pf_iff = new 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 IfTrueNode( need_pf_iff );
1753 transform_later(need_pf_true);
1754
1755 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1756 transform_later(need_pf_false);
1757
1758 Node *new_pf_wmt = new 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 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 step_size = AllocatePrefetchStepSize;
1775 uint distance = 0;
1776
1777 for ( intx i = 0; i < lines; i++ ) {
1778 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1779 _igvn.MakeConX(distance) );
1780 transform_later(prefetch_adr);
1781 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1782 transform_later(prefetch);
1783 distance += step_size;
1784 i_o = prefetch;
1785 }
1786 pf_phi_abio->set_req( pf_path, i_o );
1787
1788 pf_region->init_req( fall_in_path, need_pf_false );
1789 pf_region->init_req( pf_path, need_pf_true );
1790
1791 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1792 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1793
1794 transform_later(pf_region);
1795 transform_later(pf_phi_rawmem);
1796 transform_later(pf_phi_abio);
1797
1798 needgc_false = pf_region;
1799 contended_phi_rawmem = pf_phi_rawmem;
1800 i_o = pf_phi_abio;
1801 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1802 // Insert a prefetch instruction for each allocation.
1803 // This code is used to generate 1 prefetch instruction per cache line.
1804
1805 // Generate several prefetch instructions.
1806 uint step_size = AllocatePrefetchStepSize;
1807 uint distance = AllocatePrefetchDistance;
1808
1809 // Next cache address.
1810 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1811 _igvn.MakeConX(step_size + distance));
1812 transform_later(cache_adr);
1813 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1814 transform_later(cache_adr);
1815 // Address is aligned to execute prefetch to the beginning of cache line size
1816 // (it is important when BIS instruction is used on SPARC as prefetch).
1817 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1818 cache_adr = new AndXNode(cache_adr, mask);
1819 transform_later(cache_adr);
1820 cache_adr = new CastX2PNode(cache_adr);
1821 transform_later(cache_adr);
1822
1823 // Prefetch
1824 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1825 prefetch->set_req(0, needgc_false);
1826 transform_later(prefetch);
1827 contended_phi_rawmem = prefetch;
1828 Node *prefetch_adr;
1829 distance = step_size;
1830 for ( intx i = 1; i < lines; i++ ) {
1831 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1832 _igvn.MakeConX(distance) );
1833 transform_later(prefetch_adr);
1834 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1835 transform_later(prefetch);
1836 distance += step_size;
1837 contended_phi_rawmem = prefetch;
1838 }
1839 } else if( AllocatePrefetchStyle > 0 ) {
1840 // Insert a prefetch for each allocation only on the fast-path
1841 Node *prefetch_adr;
1842 Node *prefetch;
1843 // Generate several prefetch instructions.
1844 uint step_size = AllocatePrefetchStepSize;
1845 uint distance = AllocatePrefetchDistance;
1846 for ( intx i = 0; i < lines; i++ ) {
1847 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1848 _igvn.MakeConX(distance) );
1849 transform_later(prefetch_adr);
1850 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1851 // Do not let it float too high, since if eden_top == eden_end,
1852 // both might be null.
1853 if( i == 0 ) { // Set control for first prefetch, next follows it
1854 prefetch->init_req(0, needgc_false);
1855 }
1856 transform_later(prefetch);
1857 distance += step_size;
1858 i_o = prefetch;
1859 }
1860 }
1861 return i_o;
1862 }
1863
1864
1865 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1866 expand_allocate_common(alloc, NULL,
1867 OptoRuntime::new_instance_Type(),
1868 OptoRuntime::new_instance_Java());
1869 }
1870
1871 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1872 Node* length = alloc->in(AllocateNode::ALength);
1873 InitializeNode* init = alloc->initialization();
1874 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1875 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1876 address slow_call_address; // Address of slow call
1877 if (init != NULL && init->is_complete_with_arraycopy() &&
1878 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == NULL) {
1879 // Don't zero type array during slow allocation in VM since
1880 // it will be initialized later by arraycopy in compiled code.
1881 slow_call_address = OptoRuntime::new_array_nozero_Java();
1882 } else {
1883 slow_call_address = OptoRuntime::new_array_Java();
1884 }
1885 expand_allocate_common(alloc, length,
1886 OptoRuntime::new_array_Type(),
1887 slow_call_address);
1888 }
1889
1890 //-------------------mark_eliminated_box----------------------------------
1891 //
1892 // During EA obj may point to several objects but after few ideal graph
1893 // transformations (CCP) it may point to only one non escaping object
1894 // (but still using phi), corresponding locks and unlocks will be marked
1895 // for elimination. Later obj could be replaced with a new node (new phi)
1896 // and which does not have escape information. And later after some graph
1897 // reshape other locks and unlocks (which were not marked for elimination
1898 // before) are connected to this new obj (phi) but they still will not be
1899 // marked for elimination since new obj has no escape information.
1900 // Mark all associated (same box and obj) lock and unlock nodes for
1901 // elimination if some of them marked already.
1902 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1903 if (oldbox->as_BoxLock()->is_eliminated()) {
1904 return; // This BoxLock node was processed already.
1905 }
1906 // New implementation (EliminateNestedLocks) has separate BoxLock
1907 // node for each locked region so mark all associated locks/unlocks as
1908 // eliminated even if different objects are referenced in one locked region
1909 // (for example, OSR compilation of nested loop inside locked scope).
1910 if (EliminateNestedLocks ||
1911 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj, NULL)) {
1912 // Box is used only in one lock region. Mark this box as eliminated.
1913 _igvn.hash_delete(oldbox);
1914 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1915 _igvn.hash_insert(oldbox);
1916
1917 for (uint i = 0; i < oldbox->outcnt(); i++) {
1918 Node* u = oldbox->raw_out(i);
1919 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1920 AbstractLockNode* alock = u->as_AbstractLock();
1921 // Check lock's box since box could be referenced by Lock's debug info.
1922 if (alock->box_node() == oldbox) {
1923 // Mark eliminated all related locks and unlocks.
1924 #ifdef ASSERT
1925 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1926 #endif
1927 alock->set_non_esc_obj();
1928 }
1929 }
1930 }
1931 return;
1932 }
1933
1934 // Create new "eliminated" BoxLock node and use it in monitor debug info
1935 // instead of oldbox for the same object.
1936 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1937
1938 // Note: BoxLock node is marked eliminated only here and it is used
1939 // to indicate that all associated lock and unlock nodes are marked
1940 // for elimination.
1941 newbox->set_eliminated();
1942 transform_later(newbox);
1943
1944 // Replace old box node with new box for all users of the same object.
1945 for (uint i = 0; i < oldbox->outcnt();) {
1946 bool next_edge = true;
1947
1948 Node* u = oldbox->raw_out(i);
1949 if (u->is_AbstractLock()) {
1950 AbstractLockNode* alock = u->as_AbstractLock();
1951 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1952 // Replace Box and mark eliminated all related locks and unlocks.
1953 #ifdef ASSERT
1954 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1955 #endif
1956 alock->set_non_esc_obj();
1957 _igvn.rehash_node_delayed(alock);
1958 alock->set_box_node(newbox);
1959 next_edge = false;
1960 }
1961 }
1962 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1963 FastLockNode* flock = u->as_FastLock();
1964 assert(flock->box_node() == oldbox, "sanity");
1965 _igvn.rehash_node_delayed(flock);
1966 flock->set_box_node(newbox);
1967 next_edge = false;
1968 }
1969
1970 // Replace old box in monitor debug info.
1971 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1972 SafePointNode* sfn = u->as_SafePoint();
1973 JVMState* youngest_jvms = sfn->jvms();
1974 int max_depth = youngest_jvms->depth();
1975 for (int depth = 1; depth <= max_depth; depth++) {
1976 JVMState* jvms = youngest_jvms->of_depth(depth);
1977 int num_mon = jvms->nof_monitors();
1978 // Loop over monitors
1979 for (int idx = 0; idx < num_mon; idx++) {
1980 Node* obj_node = sfn->monitor_obj(jvms, idx);
1981 Node* box_node = sfn->monitor_box(jvms, idx);
1982 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1983 int j = jvms->monitor_box_offset(idx);
1984 _igvn.replace_input_of(u, j, newbox);
1985 next_edge = false;
1986 }
1987 }
1988 }
1989 }
1990 if (next_edge) i++;
1991 }
1992 }
1993
1994 //-----------------------mark_eliminated_locking_nodes-----------------------
1995 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1996 if (EliminateNestedLocks) {
1997 if (alock->is_nested()) {
1998 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1999 return;
2000 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2001 // Only Lock node has JVMState needed here.
2002 // Not that preceding claim is documented anywhere else.
2003 if (alock->jvms() != NULL) {
2004 if (alock->as_Lock()->is_nested_lock_region()) {
2005 // Mark eliminated related nested locks and unlocks.
2006 Node* obj = alock->obj_node();
2007 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2008 assert(!box_node->is_eliminated(), "should not be marked yet");
2009 // Note: BoxLock node is marked eliminated only here
2010 // and it is used to indicate that all associated lock
2011 // and unlock nodes are marked for elimination.
2012 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2013 for (uint i = 0; i < box_node->outcnt(); i++) {
2014 Node* u = box_node->raw_out(i);
2015 if (u->is_AbstractLock()) {
2016 alock = u->as_AbstractLock();
2017 if (alock->box_node() == box_node) {
2018 // Verify that this Box is referenced only by related locks.
2019 assert(alock->obj_node()->eqv_uncast(obj), "");
2020 // Mark all related locks and unlocks.
2021 #ifdef ASSERT
2022 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2023 #endif
2024 alock->set_nested();
2025 }
2026 }
2027 }
2028 } else {
2029 #ifdef ASSERT
2030 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2031 if (C->log() != NULL)
2032 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2033 #endif
2034 }
2035 }
2036 return;
2037 }
2038 // Process locks for non escaping object
2039 assert(alock->is_non_esc_obj(), "");
2040 } // EliminateNestedLocks
2041
2042 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2043 // Look for all locks of this object and mark them and
2044 // corresponding BoxLock nodes as eliminated.
2045 Node* obj = alock->obj_node();
2046 for (uint j = 0; j < obj->outcnt(); j++) {
2047 Node* o = obj->raw_out(j);
2048 if (o->is_AbstractLock() &&
2049 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2050 alock = o->as_AbstractLock();
2051 Node* box = alock->box_node();
2052 // Replace old box node with new eliminated box for all users
2053 // of the same object and mark related locks as eliminated.
2054 mark_eliminated_box(box, obj);
2055 }
2056 }
2057 }
2058 }
2059
2060 // we have determined that this lock/unlock can be eliminated, we simply
2061 // eliminate the node without expanding it.
2062 //
2063 // Note: The membar's associated with the lock/unlock are currently not
2064 // eliminated. This should be investigated as a future enhancement.
2065 //
2066 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2067
2068 if (!alock->is_eliminated()) {
2069 return false;
2070 }
2071 #ifdef ASSERT
2072 if (!alock->is_coarsened()) {
2073 // Check that new "eliminated" BoxLock node is created.
2074 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2075 assert(oldbox->is_eliminated(), "should be done already");
2076 }
2077 #endif
2078
2079 alock->log_lock_optimization(C, "eliminate_lock");
2080
2081 #ifndef PRODUCT
2082 if (PrintEliminateLocks) {
2083 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2084 }
2085 #endif
2086
2087 Node* mem = alock->in(TypeFunc::Memory);
2088 Node* ctrl = alock->in(TypeFunc::Control);
2089 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2090
2091 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2092 // There are 2 projections from the lock. The lock node will
2093 // be deleted when its last use is subsumed below.
2094 assert(alock->outcnt() == 2 &&
2095 _callprojs.fallthrough_proj != NULL &&
2096 _callprojs.fallthrough_memproj != NULL,
2097 "Unexpected projections from Lock/Unlock");
2098
2099 Node* fallthroughproj = _callprojs.fallthrough_proj;
2100 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2101
2102 // The memory projection from a lock/unlock is RawMem
2103 // The input to a Lock is merged memory, so extract its RawMem input
2104 // (unless the MergeMem has been optimized away.)
2105 if (alock->is_Lock()) {
2106 // Search for MemBarAcquireLock node and delete it also.
2107 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2108 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2109 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2110 Node* memproj = membar->proj_out(TypeFunc::Memory);
2111 _igvn.replace_node(ctrlproj, fallthroughproj);
2112 _igvn.replace_node(memproj, memproj_fallthrough);
2113
2114 // Delete FastLock node also if this Lock node is unique user
2115 // (a loop peeling may clone a Lock node).
2116 Node* flock = alock->as_Lock()->fastlock_node();
2117 if (flock->outcnt() == 1) {
2118 assert(flock->unique_out() == alock, "sanity");
2119 _igvn.replace_node(flock, top());
2120 }
2121 }
2122
2123 // Search for MemBarReleaseLock node and delete it also.
2124 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2125 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2126 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2127 mem->is_Proj() && membar == mem->in(0), "");
2128 _igvn.replace_node(fallthroughproj, ctrl);
2129 _igvn.replace_node(memproj_fallthrough, mem);
2130 fallthroughproj = ctrl;
2131 memproj_fallthrough = mem;
2132 ctrl = membar->in(TypeFunc::Control);
2133 mem = membar->in(TypeFunc::Memory);
2134 }
2135
2136 _igvn.replace_node(fallthroughproj, ctrl);
2137 _igvn.replace_node(memproj_fallthrough, mem);
2138 return true;
2139 }
2140
2141
2142 //------------------------------expand_lock_node----------------------
2143 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2144
2145 Node* ctrl = lock->in(TypeFunc::Control);
2146 Node* mem = lock->in(TypeFunc::Memory);
2147 Node* obj = lock->obj_node();
2148 Node* box = lock->box_node();
2149 Node* flock = lock->fastlock_node();
2150
2151 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2152
2153 // Make the merge point
2154 Node *region;
2155 Node *mem_phi;
2156 Node *slow_path;
2157
2158 region = new RegionNode(3);
2159 // create a Phi for the memory state
2160 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2161
2162 // Optimize test; set region slot 2
2163 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2164 mem_phi->init_req(2, mem);
2165
2166 // Make slow path call
2167 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2168 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2169 obj, NULL, NULL);
2170
2171 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2172
2173 // Slow path can only throw asynchronous exceptions, which are always
2174 // de-opted. So the compiler thinks the slow-call can never throw an
2175 // exception. If it DOES throw an exception we would need the debug
2176 // info removed first (since if it throws there is no monitor).
2177 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2178 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2179
2180 // Capture slow path
2181 // disconnect fall-through projection from call and create a new one
2182 // hook up users of fall-through projection to region
2183 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2184 transform_later(slow_ctrl);
2185 _igvn.hash_delete(_callprojs.fallthrough_proj);
2186 _callprojs.fallthrough_proj->disconnect_inputs(C);
2187 region->init_req(1, slow_ctrl);
2188 // region inputs are now complete
2189 transform_later(region);
2190 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2191
2192 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2193
2194 mem_phi->init_req(1, memproj);
2195
2196 transform_later(mem_phi);
2197
2198 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2199 }
2200
2201 //------------------------------expand_unlock_node----------------------
2202 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2203
2204 Node* ctrl = unlock->in(TypeFunc::Control);
2205 Node* mem = unlock->in(TypeFunc::Memory);
2206 Node* obj = unlock->obj_node();
2207 Node* box = unlock->box_node();
2208
2209 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2210
2211 // No need for a null check on unlock
2212
2213 // Make the merge point
2214 Node *region;
2215 Node *mem_phi;
2216
2217 region = new RegionNode(3);
2218 // create a Phi for the memory state
2219 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2220
2221 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2222 funlock = transform_later( funlock )->as_FastUnlock();
2223 // Optimize test; set region slot 2
2224 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2225 Node *thread = transform_later(new ThreadLocalNode());
2226
2227 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2228 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2229 "complete_monitor_unlocking_C", slow_path, obj, thread, NULL);
2230
2231 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2232 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2233 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2234
2235 // No exceptions for unlocking
2236 // Capture slow path
2237 // disconnect fall-through projection from call and create a new one
2238 // hook up users of fall-through projection to region
2239 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2240 transform_later(slow_ctrl);
2241 _igvn.hash_delete(_callprojs.fallthrough_proj);
2242 _callprojs.fallthrough_proj->disconnect_inputs(C);
2243 region->init_req(1, slow_ctrl);
2244 // region inputs are now complete
2245 transform_later(region);
2246 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2247
2248 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2249 mem_phi->init_req(1, memproj );
2250 mem_phi->init_req(2, mem);
2251 transform_later(mem_phi);
2252
2253 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2254 }
2255
2256 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2257 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2258 Node* bol = check->unique_out();
2259 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2260 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2261 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2262
2263 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2264 Node* iff = bol->last_out(i);
2265 assert(iff->is_If(), "where's the if?");
2266
2267 if (iff->in(0)->is_top()) {
2268 _igvn.replace_input_of(iff, 1, C->top());
2269 continue;
2270 }
2271
2272 Node* iftrue = iff->as_If()->proj_out(1);
2273 Node* iffalse = iff->as_If()->proj_out(0);
2274 Node* ctrl = iff->in(0);
2275
2276 Node* subklass = NULL;
2277 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2278 subklass = obj_or_subklass;
2279 } else {
2280 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2281 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2282 }
2283
2284 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2285
2286 _igvn.replace_input_of(iff, 0, C->top());
2287 _igvn.replace_node(iftrue, not_subtype_ctrl);
2288 _igvn.replace_node(iffalse, ctrl);
2289 }
2290 _igvn.replace_node(check, C->top());
2291 }
2292
2293 //---------------------------eliminate_macro_nodes----------------------
2294 // Eliminate scalar replaced allocations and associated locks.
2295 void PhaseMacroExpand::eliminate_macro_nodes() {
2296 if (C->macro_count() == 0)
2297 return;
2298 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2299
2300 // Before elimination may re-mark (change to Nested or NonEscObj)
2301 // all associated (same box and obj) lock and unlock nodes.
2302 int cnt = C->macro_count();
2303 for (int i=0; i < cnt; i++) {
2304 Node *n = C->macro_node(i);
2305 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2306 mark_eliminated_locking_nodes(n->as_AbstractLock());
2307 }
2308 }
2309 // Re-marking may break consistency of Coarsened locks.
2310 if (!C->coarsened_locks_consistent()) {
2311 return; // recompile without Coarsened locks if broken
2312 }
2313
2314 // First, attempt to eliminate locks
2315 bool progress = true;
2316 while (progress) {
2317 progress = false;
2318 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2319 Node* n = C->macro_node(i - 1);
2320 bool success = false;
2321 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2322 if (n->is_AbstractLock()) {
2323 success = eliminate_locking_node(n->as_AbstractLock());
2324 #ifndef PRODUCT
2325 if (success && PrintOptoStatistics) {
2326 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2327 }
2328 #endif
2329 }
2330 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2331 progress = progress || success;
2332 }
2333 }
2334 // Next, attempt to eliminate allocations
2335 _has_locks = false;
2336 progress = true;
2337 while (progress) {
2338 progress = false;
2339 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2340 Node* n = C->macro_node(i - 1);
2341 bool success = false;
2342 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2343 switch (n->class_id()) {
2344 case Node::Class_Allocate:
2345 case Node::Class_AllocateArray:
2346 success = eliminate_allocate_node(n->as_Allocate());
2347 #ifndef PRODUCT
2348 if (success && PrintOptoStatistics) {
2349 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2350 }
2351 #endif
2352 break;
2353 case Node::Class_CallStaticJava:
2354 success = eliminate_boxing_node(n->as_CallStaticJava());
2355 break;
2356 case Node::Class_Lock:
2357 case Node::Class_Unlock:
2358 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2359 _has_locks = true;
2360 break;
2361 case Node::Class_ArrayCopy:
2362 break;
2363 case Node::Class_OuterStripMinedLoop:
2364 break;
2365 case Node::Class_SubTypeCheck:
2366 break;
2367 case Node::Class_Opaque1:
2368 break;
2369 default:
2370 assert(n->Opcode() == Op_LoopLimit ||
2371 n->Opcode() == Op_Opaque2 ||
2372 n->Opcode() == Op_Opaque3 ||
2373 n->Opcode() == Op_Opaque4 ||
2374 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2375 "unknown node type in macro list");
2376 }
2377 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2378 progress = progress || success;
2379 }
2380 }
2381 #ifndef PRODUCT
2382 if (PrintOptoStatistics) {
2383 int membar_after = count_MemBar(C);
2384 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2385 }
2386 #endif
2387 }
2388
2389 //------------------------------expand_macro_nodes----------------------
2390 // Returns true if a failure occurred.
2391 bool PhaseMacroExpand::expand_macro_nodes() {
2392 // Last attempt to eliminate macro nodes.
2393 eliminate_macro_nodes();
2394 if (C->failing()) return true;
2395
2396 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2397 bool progress = true;
2398 while (progress) {
2399 progress = false;
2400 for (int i = C->macro_count(); i > 0; i--) {
2401 Node* n = C->macro_node(i-1);
2402 bool success = false;
2403 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2404 if (n->Opcode() == Op_LoopLimit) {
2405 // Remove it from macro list and put on IGVN worklist to optimize.
2406 C->remove_macro_node(n);
2407 _igvn._worklist.push(n);
2408 success = true;
2409 } else if (n->Opcode() == Op_CallStaticJava) {
2410 // Remove it from macro list and put on IGVN worklist to optimize.
2411 C->remove_macro_node(n);
2412 _igvn._worklist.push(n);
2413 success = true;
2414 } else if (n->is_Opaque1() || n->Opcode() == Op_Opaque2) {
2415 _igvn.replace_node(n, n->in(1));
2416 success = true;
2417 #if INCLUDE_RTM_OPT
2418 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2419 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2420 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2421 Node* cmp = n->unique_out();
2422 #ifdef ASSERT
2423 // Validate graph.
2424 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2425 BoolNode* bol = cmp->unique_out()->as_Bool();
2426 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2427 (bol->_test._test == BoolTest::ne), "");
2428 IfNode* ifn = bol->unique_out()->as_If();
2429 assert((ifn->outcnt() == 2) &&
2430 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2431 #endif
2432 Node* repl = n->in(1);
2433 if (!_has_locks) {
2434 // Remove RTM state check if there are no locks in the code.
2435 // Replace input to compare the same value.
2436 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2437 }
2438 _igvn.replace_node(n, repl);
2439 success = true;
2440 #endif
2441 } else if (n->Opcode() == Op_Opaque4) {
2442 // With Opaque4 nodes, the expectation is that the test of input 1
2443 // is always equal to the constant value of input 2. So we can
2444 // remove the Opaque4 and replace it by input 2. In debug builds,
2445 // leave the non constant test in instead to sanity check that it
2446 // never fails (if it does, that subgraph was constructed so, at
2447 // runtime, a Halt node is executed).
2448 #ifdef ASSERT
2449 _igvn.replace_node(n, n->in(1));
2450 #else
2451 _igvn.replace_node(n, n->in(2));
2452 #endif
2453 success = true;
2454 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2455 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2456 C->remove_macro_node(n);
2457 success = true;
2458 }
2459 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2460 progress = progress || success;
2461 }
2462 }
2463
2464 // Clean up the graph so we're less likely to hit the maximum node
2465 // limit
2466 _igvn.set_delay_transform(false);
2467 _igvn.optimize();
2468 if (C->failing()) return true;
2469 _igvn.set_delay_transform(true);
2470
2471
2472 // Because we run IGVN after each expansion, some macro nodes may go
2473 // dead and be removed from the list as we iterate over it. Move
2474 // Allocate nodes (processed in a second pass) at the beginning of
2475 // the list and then iterate from the last element of the list until
2476 // an Allocate node is seen. This is robust to random deletion in
2477 // the list due to nodes going dead.
2478 C->sort_macro_nodes();
2479
2480 // expand arraycopy "macro" nodes first
2481 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2482 // before the allocate nodes are expanded.
2483 while (C->macro_count() > 0) {
2484 int macro_count = C->macro_count();
2485 Node * n = C->macro_node(macro_count-1);
2486 assert(n->is_macro(), "only macro nodes expected here");
2487 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2488 // node is unreachable, so don't try to expand it
2489 C->remove_macro_node(n);
2490 continue;
2491 }
2492 if (n->is_Allocate()) {
2493 break;
2494 }
2495 // Make sure expansion will not cause node limit to be exceeded.
2496 // Worst case is a macro node gets expanded into about 200 nodes.
2497 // Allow 50% more for optimization.
2498 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2499 return true;
2500 }
2501
2502 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2503 switch (n->class_id()) {
2504 case Node::Class_Lock:
2505 expand_lock_node(n->as_Lock());
2506 break;
2507 case Node::Class_Unlock:
2508 expand_unlock_node(n->as_Unlock());
2509 break;
2510 case Node::Class_ArrayCopy:
2511 expand_arraycopy_node(n->as_ArrayCopy());
2512 break;
2513 case Node::Class_SubTypeCheck:
2514 expand_subtypecheck_node(n->as_SubTypeCheck());
2515 break;
2516 default:
2517 assert(false, "unknown node type in macro list");
2518 }
2519 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2520 if (C->failing()) return true;
2521
2522 // Clean up the graph so we're less likely to hit the maximum node
2523 // limit
2524 _igvn.set_delay_transform(false);
2525 _igvn.optimize();
2526 if (C->failing()) return true;
2527 _igvn.set_delay_transform(true);
2528 }
2529
2530 // All nodes except Allocate nodes are expanded now. There could be
2531 // new optimization opportunities (such as folding newly created
2532 // load from a just allocated object). Run IGVN.
2533
2534 // expand "macro" nodes
2535 // nodes are removed from the macro list as they are processed
2536 while (C->macro_count() > 0) {
2537 int macro_count = C->macro_count();
2538 Node * n = C->macro_node(macro_count-1);
2539 assert(n->is_macro(), "only macro nodes expected here");
2540 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2541 // node is unreachable, so don't try to expand it
2542 C->remove_macro_node(n);
2543 continue;
2544 }
2545 // Make sure expansion will not cause node limit to be exceeded.
2546 // Worst case is a macro node gets expanded into about 200 nodes.
2547 // Allow 50% more for optimization.
2548 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2549 return true;
2550 }
2551 switch (n->class_id()) {
2552 case Node::Class_Allocate:
2553 expand_allocate(n->as_Allocate());
2554 break;
2555 case Node::Class_AllocateArray:
2556 expand_allocate_array(n->as_AllocateArray());
2557 break;
2558 default:
2559 assert(false, "unknown node type in macro list");
2560 }
2561 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2562 if (C->failing()) return true;
2563
2564 // Clean up the graph so we're less likely to hit the maximum node
2565 // limit
2566 _igvn.set_delay_transform(false);
2567 _igvn.optimize();
2568 if (C->failing()) return true;
2569 _igvn.set_delay_transform(true);
2570 }
2571
2572 _igvn.set_delay_transform(false);
2573 return false;
2574 }
2575
2576 #ifndef PRODUCT
2577 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
2578 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
2579 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
2580 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
2581
2582 void PhaseMacroExpand::print_statistics() {
2583 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
2584 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
2585 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
2586 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
2587 }
2588
2589 int PhaseMacroExpand::count_MemBar(Compile *C) {
2590 if (!PrintOptoStatistics) {
2591 return 0;
2592 }
2593 Unique_Node_List ideal_nodes;
2594 int total = 0;
2595 ideal_nodes.map(C->live_nodes(), NULL);
2596 ideal_nodes.push(C->root());
2597 for (uint next = 0; next < ideal_nodes.size(); ++next) {
2598 Node* n = ideal_nodes.at(next);
2599 if (n->is_MemBar()) {
2600 total++;
2601 }
2602 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2603 Node* m = n->fast_out(i);
2604 ideal_nodes.push(m);
2605 }
2606 }
2607 return total;
2608 }
2609 #endif