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
62
63 //
64 // Replace any references to "oldref" in inputs to "use" with "newref".
65 // Returns the number of replacements made.
66 //
67 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
68 int nreplacements = 0;
69 uint req = use->req();
70 for (uint j = 0; j < use->len(); j++) {
71 Node *uin = use->in(j);
72 if (uin == oldref) {
73 if (j < req)
74 use->set_req(j, newref);
75 else
76 use->set_prec(j, newref);
77 nreplacements++;
78 } else if (j >= req && uin == NULL) {
79 break;
80 }
81 }
82 return nreplacements;
83 }
84
85 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
86 assert(old != NULL, "sanity");
87 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
88 Node* use = old->fast_out(i);
89 _igvn.rehash_node_delayed(use);
90 imax -= replace_input(use, old, target);
91 // back up iterator
92 --i;
93 }
94 assert(old->outcnt() == 0, "all uses must be deleted");
95 }
96
97 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
98 Node* cmp;
99 if (mask != 0) {
100 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
101 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
102 } else {
103 cmp = word;
104 }
105 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
106 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
107 transform_later(iff);
108
109 // Fast path taken.
110 Node *fast_taken = transform_later(new IfFalseNode(iff));
111
112 // Fast path not-taken, i.e. slow path
113 Node *slow_taken = transform_later(new IfTrueNode(iff));
114
115 if (return_fast_path) {
116 region->init_req(edge, slow_taken); // Capture slow-control
117 return fast_taken;
118 } else {
119 region->init_req(edge, fast_taken); // Capture fast-control
120 return slow_taken;
121 }
122 }
123
124 //--------------------copy_predefined_input_for_runtime_call--------------------
125 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
126 // Set fixed predefined input arguments
127 call->init_req( TypeFunc::Control, ctrl );
128 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
129 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
130 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
131 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
132 }
133
134 //------------------------------make_slow_call---------------------------------
135 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
136 address slow_call, const char* leaf_name, Node* slow_path,
137 Node* parm0, Node* parm1, Node* parm2) {
138
139 // Slow-path call
140 CallNode *call = leaf_name
141 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
142 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
143
144 // Slow path call has no side-effects, uses few values
145 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
146 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
147 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
148 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
149 call->copy_call_debug_info(&_igvn, oldcall);
150 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
151 _igvn.replace_node(oldcall, call);
152 transform_later(call);
153
154 return call;
155 }
156
157 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
158 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
159 bs->eliminate_gc_barrier(this, p2x);
160 #ifndef PRODUCT
161 if (PrintOptoStatistics) {
162 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
163 }
164 #endif
165 }
166
167 // Search for a memory operation for the specified memory slice.
168 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
169 Node *orig_mem = mem;
170 Node *alloc_mem = alloc->in(TypeFunc::Memory);
171 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
172 while (true) {
173 if (mem == alloc_mem || mem == start_mem ) {
174 return mem; // hit one of our sentinels
175 } else if (mem->is_MergeMem()) {
176 mem = mem->as_MergeMem()->memory_at(alias_idx);
177 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
178 Node *in = mem->in(0);
179 // we can safely skip over safepoints, calls, locks and membars because we
180 // already know that the object is safe to eliminate.
181 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
182 return in;
183 } else if (in->is_Call()) {
184 CallNode *call = in->as_Call();
185 if (call->may_modify(tinst, phase)) {
186 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
187 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
188 return in;
189 }
190 }
191 mem = in->in(TypeFunc::Memory);
192 } else if (in->is_MemBar()) {
193 ArrayCopyNode* ac = NULL;
194 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
195 if (ac != NULL) {
196 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
197 return ac;
198 }
199 }
200 mem = in->in(TypeFunc::Memory);
201 } else {
202 #ifdef ASSERT
203 in->dump();
204 mem->dump();
205 assert(false, "unexpected projection");
206 #endif
207 }
208 } else if (mem->is_Store()) {
209 const TypePtr* atype = mem->as_Store()->adr_type();
210 int adr_idx = phase->C->get_alias_index(atype);
211 if (adr_idx == alias_idx) {
212 assert(atype->isa_oopptr(), "address type must be oopptr");
213 int adr_offset = atype->offset();
214 uint adr_iid = atype->is_oopptr()->instance_id();
215 // Array elements references have the same alias_idx
216 // but different offset and different instance_id.
217 if (adr_offset == offset && adr_iid == alloc->_idx) {
218 return mem;
219 }
220 } else {
221 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
222 }
223 mem = mem->in(MemNode::Memory);
224 } else if (mem->is_ClearArray()) {
225 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
226 // Can not bypass initialization of the instance
227 // we are looking.
228 debug_only(intptr_t offset;)
229 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
230 InitializeNode* init = alloc->as_Allocate()->initialization();
231 // We are looking for stored value, return Initialize node
232 // or memory edge from Allocate node.
233 if (init != NULL) {
234 return init;
235 } else {
236 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
237 }
238 }
239 // Otherwise skip it (the call updated 'mem' value).
240 } else if (mem->Opcode() == Op_SCMemProj) {
241 mem = mem->in(0);
242 Node* adr = NULL;
243 if (mem->is_LoadStore()) {
244 adr = mem->in(MemNode::Address);
245 } else {
246 assert(mem->Opcode() == Op_EncodeISOArray ||
247 mem->Opcode() == Op_StrCompressedCopy, "sanity");
248 adr = mem->in(3); // Destination array
249 }
250 const TypePtr* atype = adr->bottom_type()->is_ptr();
251 int adr_idx = phase->C->get_alias_index(atype);
252 if (adr_idx == alias_idx) {
253 DEBUG_ONLY(mem->dump();)
254 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
255 return NULL;
256 }
257 mem = mem->in(MemNode::Memory);
258 } else if (mem->Opcode() == Op_StrInflatedCopy) {
259 Node* adr = mem->in(3); // Destination array
260 const TypePtr* atype = adr->bottom_type()->is_ptr();
261 int adr_idx = phase->C->get_alias_index(atype);
262 if (adr_idx == alias_idx) {
263 DEBUG_ONLY(mem->dump();)
264 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
265 return NULL;
266 }
267 mem = mem->in(MemNode::Memory);
268 } else {
269 return mem;
270 }
271 assert(mem != orig_mem, "dead memory loop");
272 }
273 }
274
275 // Generate loads from source of the arraycopy for fields of
276 // destination needed at a deoptimization point
277 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
278 BasicType bt = ft;
279 const Type *type = ftype;
280 if (ft == T_NARROWOOP) {
281 bt = T_OBJECT;
282 type = ftype->make_oopptr();
283 }
284 Node* res = NULL;
285 if (ac->is_clonebasic()) {
286 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
287 Node* base = ac->in(ArrayCopyNode::Src);
288 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
289 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
290 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
291 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
292 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
293 } else {
294 if (ac->modifies(offset, offset, &_igvn, true)) {
295 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
296 uint shift = exact_log2(type2aelembytes(bt));
297 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
298 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
299 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
300 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
301
302 Node* adr = NULL;
303 const TypePtr* adr_type = NULL;
304 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
305 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
306 Node* base = ac->in(ArrayCopyNode::Src);
307 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
308 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Don't emit a new load from src if src == dst but try to get the value from memory instead
311 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
312 }
313 } else {
314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
315 #ifdef _LP64
316 diff = _igvn.transform(new ConvI2LNode(diff));
317 #endif
318 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
319
320 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
321 Node* base = ac->in(ArrayCopyNode::Src);
322 adr = _igvn.transform(new AddPNode(base, base, off));
323 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
324 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
325 // Non constant offset in the array: we can't statically
326 // determine the value
327 return NULL;
328 }
329 }
330 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
331 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
332 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
333 }
334 }
335 if (res != NULL) {
336 if (ftype->isa_narrowoop()) {
337 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
338 res = _igvn.transform(new EncodePNode(res, ftype));
339 }
340 return res;
341 }
342 return NULL;
343 }
344
345 //
346 // Given a Memory Phi, compute a value Phi containing the values from stores
347 // on the input paths.
348 // Note: this function is recursive, its depth is limited by the "level" argument
349 // Returns the computed Phi, or NULL if it cannot compute it.
350 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) {
351 assert(mem->is_Phi(), "sanity");
352 int alias_idx = C->get_alias_index(adr_t);
353 int offset = adr_t->offset();
354 int instance_id = adr_t->instance_id();
355
356 // Check if an appropriate value phi already exists.
357 Node* region = mem->in(0);
358 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
359 Node* phi = region->fast_out(k);
360 if (phi->is_Phi() && phi != mem &&
361 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
362 return phi;
363 }
364 }
365 // Check if an appropriate new value phi already exists.
366 Node* new_phi = value_phis->find(mem->_idx);
367 if (new_phi != NULL)
368 return new_phi;
369
370 if (level <= 0) {
371 return NULL; // Give up: phi tree too deep
372 }
373 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
374 Node *alloc_mem = alloc->in(TypeFunc::Memory);
375
376 uint length = mem->req();
377 GrowableArray <Node *> values(length, length, NULL);
378
379 // create a new Phi for the value
380 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
381 transform_later(phi);
382 value_phis->push(phi, mem->_idx);
383
384 for (uint j = 1; j < length; j++) {
385 Node *in = mem->in(j);
386 if (in == NULL || in->is_top()) {
387 values.at_put(j, in);
388 } else {
389 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
390 if (val == start_mem || val == alloc_mem) {
391 // hit a sentinel, return appropriate 0 value
392 values.at_put(j, _igvn.zerocon(ft));
393 continue;
394 }
395 if (val->is_Initialize()) {
396 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
397 }
398 if (val == NULL) {
399 return NULL; // can't find a value on this path
400 }
401 if (val == mem) {
402 values.at_put(j, mem);
403 } else if (val->is_Store()) {
404 Node* n = val->in(MemNode::ValueIn);
405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
406 n = bs->step_over_gc_barrier(n);
407 if (is_subword_type(ft)) {
408 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
409 }
410 values.at_put(j, n);
411 } else if(val->is_Proj() && val->in(0) == alloc) {
412 values.at_put(j, _igvn.zerocon(ft));
413 } else if (val->is_Phi()) {
414 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
415 if (val == NULL) {
416 return NULL;
417 }
418 values.at_put(j, val);
419 } else if (val->Opcode() == Op_SCMemProj) {
420 assert(val->in(0)->is_LoadStore() ||
421 val->in(0)->Opcode() == Op_EncodeISOArray ||
422 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
423 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
424 return NULL;
425 } else if (val->is_ArrayCopy()) {
426 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
427 if (res == NULL) {
428 return NULL;
429 }
430 values.at_put(j, res);
431 } else {
432 DEBUG_ONLY( val->dump(); )
433 assert(false, "unknown node on this path");
434 return NULL; // unknown node on this path
435 }
436 }
437 }
438 // Set Phi's inputs
439 for (uint j = 1; j < length; j++) {
440 if (values.at(j) == mem) {
441 phi->init_req(j, phi);
442 } else {
443 phi->init_req(j, values.at(j));
444 }
445 }
446 return phi;
447 }
448
449 // Search the last value stored into the object's field.
450 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
451 assert(adr_t->is_known_instance_field(), "instance required");
452 int instance_id = adr_t->instance_id();
453 assert((uint)instance_id == alloc->_idx, "wrong allocation");
454
455 int alias_idx = C->get_alias_index(adr_t);
456 int offset = adr_t->offset();
457 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
458 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
459 Node *alloc_mem = alloc->in(TypeFunc::Memory);
460 VectorSet visited;
461
462 bool done = sfpt_mem == alloc_mem;
463 Node *mem = sfpt_mem;
464 while (!done) {
465 if (visited.test_set(mem->_idx)) {
466 return NULL; // found a loop, give up
467 }
468 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
469 if (mem == start_mem || mem == alloc_mem) {
470 done = true; // hit a sentinel, return appropriate 0 value
471 } else if (mem->is_Initialize()) {
472 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
473 if (mem == NULL) {
474 done = true; // Something go wrong.
475 } else if (mem->is_Store()) {
476 const TypePtr* atype = mem->as_Store()->adr_type();
477 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
478 done = true;
479 }
480 } else if (mem->is_Store()) {
481 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
482 assert(atype != NULL, "address type must be oopptr");
483 assert(C->get_alias_index(atype) == alias_idx &&
484 atype->is_known_instance_field() && atype->offset() == offset &&
485 atype->instance_id() == instance_id, "store is correct memory slice");
486 done = true;
487 } else if (mem->is_Phi()) {
488 // try to find a phi's unique input
489 Node *unique_input = NULL;
490 Node *top = C->top();
491 for (uint i = 1; i < mem->req(); i++) {
492 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
493 if (n == NULL || n == top || n == mem) {
494 continue;
495 } else if (unique_input == NULL) {
496 unique_input = n;
497 } else if (unique_input != n) {
498 unique_input = top;
499 break;
500 }
501 }
502 if (unique_input != NULL && unique_input != top) {
503 mem = unique_input;
504 } else {
505 done = true;
506 }
507 } else if (mem->is_ArrayCopy()) {
508 done = true;
509 } else {
510 DEBUG_ONLY( mem->dump(); )
511 assert(false, "unexpected node");
512 }
513 }
514 if (mem != NULL) {
515 if (mem == start_mem || mem == alloc_mem) {
516 // hit a sentinel, return appropriate 0 value
517 return _igvn.zerocon(ft);
518 } else if (mem->is_Store()) {
519 Node* n = mem->in(MemNode::ValueIn);
520 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
521 n = bs->step_over_gc_barrier(n);
522 return n;
523 } else if (mem->is_Phi()) {
524 // attempt to produce a Phi reflecting the values on the input paths of the Phi
525 Node_Stack value_phis(8);
526 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
527 if (phi != NULL) {
528 return phi;
529 } else {
530 // Kill all new Phis
531 while(value_phis.is_nonempty()) {
532 Node* n = value_phis.node();
533 _igvn.replace_node(n, C->top());
534 value_phis.pop();
535 }
536 }
537 } else if (mem->is_ArrayCopy()) {
538 Node* ctl = mem->in(0);
539 Node* m = mem->in(TypeFunc::Memory);
540 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
541 // pin the loads in the uncommon trap path
542 ctl = sfpt_ctl;
543 m = sfpt_mem;
544 }
545 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
546 }
547 }
548 // Something go wrong.
549 return NULL;
550 }
551
552 // Check the possibility of scalar replacement.
553 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
554 // Scan the uses of the allocation to check for anything that would
555 // prevent us from eliminating it.
556 NOT_PRODUCT( const char* fail_eliminate = NULL; )
557 DEBUG_ONLY( Node* disq_node = NULL; )
558 bool can_eliminate = true;
559
560 Node* res = alloc->result_cast();
561 const TypeOopPtr* res_type = NULL;
562 if (res == NULL) {
563 // All users were eliminated.
564 } else if (!res->is_CheckCastPP()) {
565 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
566 can_eliminate = false;
567 } else {
568 res_type = _igvn.type(res)->isa_oopptr();
569 if (res_type == NULL) {
570 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
571 can_eliminate = false;
572 } else if (res_type->isa_aryptr()) {
573 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
574 if (length < 0) {
575 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
576 can_eliminate = false;
577 }
578 }
579 }
580
581 if (can_eliminate && res != NULL) {
582 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
583 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
584 j < jmax && can_eliminate; j++) {
585 Node* use = res->fast_out(j);
586
587 if (use->is_AddP()) {
588 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
589 int offset = addp_type->offset();
590
591 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
592 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
593 can_eliminate = false;
594 break;
595 }
596 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
597 k < kmax && can_eliminate; k++) {
598 Node* n = use->fast_out(k);
599 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n)) {
600 DEBUG_ONLY(disq_node = n;)
601 if (n->is_Load() || n->is_LoadStore()) {
602 NOT_PRODUCT(fail_eliminate = "Field load";)
603 } else {
604 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
605 }
606 can_eliminate = false;
607 }
608 }
609 } else if (use->is_ArrayCopy() &&
610 (use->as_ArrayCopy()->is_clonebasic() ||
611 use->as_ArrayCopy()->is_arraycopy_validated() ||
612 use->as_ArrayCopy()->is_copyof_validated() ||
613 use->as_ArrayCopy()->is_copyofrange_validated()) &&
614 use->in(ArrayCopyNode::Dest) == res) {
615 // ok to eliminate
616 } else if (use->is_SafePoint()) {
617 SafePointNode* sfpt = use->as_SafePoint();
618 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
619 // Object is passed as argument.
620 DEBUG_ONLY(disq_node = use;)
621 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
622 can_eliminate = false;
623 }
624 Node* sfptMem = sfpt->memory();
625 if (sfptMem == NULL || sfptMem->is_top()) {
626 DEBUG_ONLY(disq_node = use;)
627 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
628 can_eliminate = false;
629 } else {
630 safepoints.append_if_missing(sfpt);
631 }
632 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
633 if (use->is_Phi()) {
634 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
635 NOT_PRODUCT(fail_eliminate = "Object is return value";)
636 } else {
637 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
638 }
639 DEBUG_ONLY(disq_node = use;)
640 } else {
641 if (use->Opcode() == Op_Return) {
642 NOT_PRODUCT(fail_eliminate = "Object is return value";)
643 }else {
644 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
645 }
646 DEBUG_ONLY(disq_node = use;)
647 }
648 can_eliminate = false;
649 }
650 }
651 }
652
653 #ifndef PRODUCT
654 if (PrintEliminateAllocations) {
655 if (can_eliminate) {
656 tty->print("Scalar ");
657 if (res == NULL)
658 alloc->dump();
659 else
660 res->dump();
661 } else if (alloc->_is_scalar_replaceable) {
662 tty->print("NotScalar (%s)", fail_eliminate);
663 if (res == NULL)
664 alloc->dump();
665 else
666 res->dump();
667 #ifdef ASSERT
668 if (disq_node != NULL) {
669 tty->print(" >>>> ");
670 disq_node->dump();
671 }
672 #endif /*ASSERT*/
673 }
674 }
675 #endif
676 return can_eliminate;
677 }
678
679 // Do scalar replacement.
680 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
681 GrowableArray <SafePointNode *> safepoints_done;
682
683 ciInstanceKlass* iklass = NULL;
684 int nfields = 0;
685 int array_base = 0;
686 int element_size = 0;
687 BasicType basic_elem_type = T_ILLEGAL;
688 const Type* field_type = NULL;
689
690 Node* res = alloc->result_cast();
691 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
692 const TypeOopPtr* res_type = NULL;
693 if (res != NULL) { // Could be NULL when there are no users
694 res_type = _igvn.type(res)->isa_oopptr();
695 }
696
697 if (res != NULL) {
698 if (res_type->isa_instptr()) {
699 // find the fields of the class which will be needed for safepoint debug information
700 iklass = res_type->is_instptr()->instance_klass();
701 nfields = iklass->nof_nonstatic_fields();
702 } else {
703 // find the array's elements which will be needed for safepoint debug information
704 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
705 assert(nfields >= 0, "must be an array klass.");
706 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
707 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
708 element_size = type2aelembytes(basic_elem_type);
709 field_type = res_type->is_aryptr()->elem();
710 }
711 }
712 //
713 // Process the safepoint uses
714 //
715 while (safepoints.length() > 0) {
716 SafePointNode* sfpt = safepoints.pop();
717 Node* mem = sfpt->memory();
718 Node* ctl = sfpt->control();
719 assert(sfpt->jvms() != NULL, "missed JVMS");
720 // Fields of scalar objs are referenced only at the end
721 // of regular debuginfo at the last (youngest) JVMS.
722 // Record relative start index.
723 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
724 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
725 #ifdef ASSERT
726 alloc,
727 #endif
728 first_ind, nfields);
729 sobj->init_req(0, C->root());
730 transform_later(sobj);
731
732 // Scan object's fields adding an input to the safepoint for each field.
733 for (int j = 0; j < nfields; j++) {
734 intptr_t offset;
735 ciField* field = NULL;
736 if (iklass != NULL) {
737 field = iklass->nonstatic_field_at(j);
738 offset = field->offset();
739 ciType* elem_type = field->type();
740 basic_elem_type = field->layout_type();
741
742 // The next code is taken from Parse::do_get_xxx().
743 if (is_reference_type(basic_elem_type)) {
744 if (!elem_type->is_loaded()) {
745 field_type = TypeInstPtr::BOTTOM;
746 } else if (field != NULL && field->is_static_constant()) {
747 ciObject* con = field->constant_value().as_object();
748 // Do not "join" in the previous type; it doesn't add value,
749 // and may yield a vacuous result if the field is of interface type.
750 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
751 assert(field_type != NULL, "field singleton type must be consistent");
752 } else {
753 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
754 }
755 if (UseCompressedOops) {
756 field_type = field_type->make_narrowoop();
757 basic_elem_type = T_NARROWOOP;
758 }
759 } else {
760 field_type = Type::get_const_basic_type(basic_elem_type);
761 }
762 } else {
763 offset = array_base + j * (intptr_t)element_size;
764 }
765
766 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
767
768 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
769 if (field_val == NULL) {
770 // We weren't able to find a value for this field,
771 // give up on eliminating this allocation.
772
773 // Remove any extra entries we added to the safepoint.
774 uint last = sfpt->req() - 1;
775 for (int k = 0; k < j; k++) {
776 sfpt->del_req(last--);
777 }
778 _igvn._worklist.push(sfpt);
779 // rollback processed safepoints
780 while (safepoints_done.length() > 0) {
781 SafePointNode* sfpt_done = safepoints_done.pop();
782 // remove any extra entries we added to the safepoint
783 last = sfpt_done->req() - 1;
784 for (int k = 0; k < nfields; k++) {
785 sfpt_done->del_req(last--);
786 }
787 JVMState *jvms = sfpt_done->jvms();
788 jvms->set_endoff(sfpt_done->req());
789 // Now make a pass over the debug information replacing any references
790 // to SafePointScalarObjectNode with the allocated object.
791 int start = jvms->debug_start();
792 int end = jvms->debug_end();
793 for (int i = start; i < end; i++) {
794 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
795 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
796 if (scobj->first_index(jvms) == sfpt_done->req() &&
797 scobj->n_fields() == (uint)nfields) {
798 assert(scobj->alloc() == alloc, "sanity");
799 sfpt_done->set_req(i, res);
800 }
801 }
802 }
803 _igvn._worklist.push(sfpt_done);
804 }
805 #ifndef PRODUCT
806 if (PrintEliminateAllocations) {
807 if (field != NULL) {
808 tty->print("=== At SafePoint node %d can't find value of Field: ",
809 sfpt->_idx);
810 field->print();
811 int field_idx = C->get_alias_index(field_addr_type);
812 tty->print(" (alias_idx=%d)", field_idx);
813 } else { // Array's element
814 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
815 sfpt->_idx, j);
816 }
817 tty->print(", which prevents elimination of: ");
818 if (res == NULL)
819 alloc->dump();
820 else
821 res->dump();
822 }
823 #endif
824 return false;
825 }
826 if (UseCompressedOops && field_type->isa_narrowoop()) {
827 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
828 // to be able scalar replace the allocation.
829 if (field_val->is_EncodeP()) {
830 field_val = field_val->in(1);
831 } else {
832 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
833 }
834 }
835 sfpt->add_req(field_val);
836 }
837 JVMState *jvms = sfpt->jvms();
838 jvms->set_endoff(sfpt->req());
839 // Now make a pass over the debug information replacing any references
840 // to the allocated object with "sobj"
841 int start = jvms->debug_start();
842 int end = jvms->debug_end();
843 sfpt->replace_edges_in_range(res, sobj, start, end, &_igvn);
844 _igvn._worklist.push(sfpt);
845 safepoints_done.append_if_missing(sfpt); // keep it for rollback
846 }
847 return true;
848 }
849
850 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
851 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
852 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
853 if (ctl_proj != NULL) {
854 igvn.replace_node(ctl_proj, n->in(0));
855 }
856 if (mem_proj != NULL) {
857 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
858 }
859 }
860
861 // Process users of eliminated allocation.
862 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
863 Node* res = alloc->result_cast();
864 if (res != NULL) {
865 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
866 Node *use = res->last_out(j);
867 uint oc1 = res->outcnt();
868
869 if (use->is_AddP()) {
870 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
871 Node *n = use->last_out(k);
872 uint oc2 = use->outcnt();
873 if (n->is_Store()) {
874 #ifdef ASSERT
875 // Verify that there is no dependent MemBarVolatile nodes,
876 // they should be removed during IGVN, see MemBarNode::Ideal().
877 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
878 p < pmax; p++) {
879 Node* mb = n->fast_out(p);
880 assert(mb->is_Initialize() || !mb->is_MemBar() ||
881 mb->req() <= MemBarNode::Precedent ||
882 mb->in(MemBarNode::Precedent) != n,
883 "MemBarVolatile should be eliminated for non-escaping object");
884 }
885 #endif
886 _igvn.replace_node(n, n->in(MemNode::Memory));
887 } else {
888 eliminate_gc_barrier(n);
889 }
890 k -= (oc2 - use->outcnt());
891 }
892 _igvn.remove_dead_node(use);
893 } else if (use->is_ArrayCopy()) {
894 // Disconnect ArrayCopy node
895 ArrayCopyNode* ac = use->as_ArrayCopy();
896 if (ac->is_clonebasic()) {
897 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
898 disconnect_projections(ac, _igvn);
899 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
900 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
901 disconnect_projections(membar_before->as_MemBar(), _igvn);
902 if (membar_after->is_MemBar()) {
903 disconnect_projections(membar_after->as_MemBar(), _igvn);
904 }
905 } else {
906 assert(ac->is_arraycopy_validated() ||
907 ac->is_copyof_validated() ||
908 ac->is_copyofrange_validated(), "unsupported");
909 CallProjections callprojs;
910 ac->extract_projections(&callprojs, true);
911
912 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
913 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
914 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
915
916 // Set control to top. IGVN will remove the remaining projections
917 ac->set_req(0, top());
918 ac->replace_edge(res, top(), &_igvn);
919
920 // Disconnect src right away: it can help find new
921 // opportunities for allocation elimination
922 Node* src = ac->in(ArrayCopyNode::Src);
923 ac->replace_edge(src, top(), &_igvn);
924 // src can be top at this point if src and dest of the
925 // arraycopy were the same
926 if (src->outcnt() == 0 && !src->is_top()) {
927 _igvn.remove_dead_node(src);
928 }
929 }
930 _igvn._worklist.push(ac);
931 } else {
932 eliminate_gc_barrier(use);
933 }
934 j -= (oc1 - res->outcnt());
935 }
936 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
937 _igvn.remove_dead_node(res);
938 }
939
940 //
941 // Process other users of allocation's projections
942 //
943 if (_callprojs.resproj != NULL && _callprojs.resproj->outcnt() != 0) {
944 // First disconnect stores captured by Initialize node.
945 // If Initialize node is eliminated first in the following code,
946 // it will kill such stores and DUIterator_Last will assert.
947 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
948 Node* use = _callprojs.resproj->fast_out(j);
949 if (use->is_AddP()) {
950 // raw memory addresses used only by the initialization
951 _igvn.replace_node(use, C->top());
952 --j; --jmax;
953 }
954 }
955 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
956 Node* use = _callprojs.resproj->last_out(j);
957 uint oc1 = _callprojs.resproj->outcnt();
958 if (use->is_Initialize()) {
959 // Eliminate Initialize node.
960 InitializeNode *init = use->as_Initialize();
961 assert(init->outcnt() <= 2, "only a control and memory projection expected");
962 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
963 if (ctrl_proj != NULL) {
964 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
965 #ifdef ASSERT
966 // If the InitializeNode has no memory out, it will die, and tmp will become NULL
967 Node* tmp = init->in(TypeFunc::Control);
968 assert(tmp == NULL || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
969 #endif
970 }
971 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
972 if (mem_proj != NULL) {
973 Node *mem = init->in(TypeFunc::Memory);
974 #ifdef ASSERT
975 if (mem->is_MergeMem()) {
976 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
977 } else {
978 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
979 }
980 #endif
981 _igvn.replace_node(mem_proj, mem);
982 }
983 } else {
984 assert(false, "only Initialize or AddP expected");
985 }
986 j -= (oc1 - _callprojs.resproj->outcnt());
987 }
988 }
989 if (_callprojs.fallthrough_catchproj != NULL) {
990 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
991 }
992 if (_callprojs.fallthrough_memproj != NULL) {
993 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
994 }
995 if (_callprojs.catchall_memproj != NULL) {
996 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
997 }
998 if (_callprojs.fallthrough_ioproj != NULL) {
999 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1000 }
1001 if (_callprojs.catchall_ioproj != NULL) {
1002 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1003 }
1004 if (_callprojs.catchall_catchproj != NULL) {
1005 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1006 }
1007 }
1008
1009 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1010 // If reallocation fails during deoptimization we'll pop all
1011 // interpreter frames for this compiled frame and that won't play
1012 // nice with JVMTI popframe.
1013 // We avoid this issue by eager reallocation when the popframe request
1014 // is received.
1015 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1016 return false;
1017 }
1018 Node* klass = alloc->in(AllocateNode::KlassNode);
1019 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1020 Node* res = alloc->result_cast();
1021 // Eliminate boxing allocations which are not used
1022 // regardless scalar replaceable status.
1023 bool boxing_alloc = C->eliminate_boxing() &&
1024 tklass->isa_instklassptr() &&
1025 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1026 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1027 return false;
1028 }
1029
1030 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1031
1032 GrowableArray <SafePointNode *> safepoints;
1033 if (!can_eliminate_allocation(alloc, safepoints)) {
1034 return false;
1035 }
1036
1037 if (!alloc->_is_scalar_replaceable) {
1038 assert(res == NULL, "sanity");
1039 // We can only eliminate allocation if all debug info references
1040 // are already replaced with SafePointScalarObject because
1041 // we can't search for a fields value without instance_id.
1042 if (safepoints.length() > 0) {
1043 return false;
1044 }
1045 }
1046
1047 if (!scalar_replacement(alloc, safepoints)) {
1048 return false;
1049 }
1050
1051 CompileLog* log = C->log();
1052 if (log != NULL) {
1053 log->head("eliminate_allocation type='%d'",
1054 log->identify(tklass->exact_klass()));
1055 JVMState* p = alloc->jvms();
1056 while (p != NULL) {
1057 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1058 p = p->caller();
1059 }
1060 log->tail("eliminate_allocation");
1061 }
1062
1063 process_users_of_allocation(alloc);
1064
1065 #ifndef PRODUCT
1066 if (PrintEliminateAllocations) {
1067 if (alloc->is_AllocateArray())
1068 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1069 else
1070 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1071 }
1072 #endif
1073
1074 return true;
1075 }
1076
1077 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1078 // EA should remove all uses of non-escaping boxing node.
1079 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1080 return false;
1081 }
1082
1083 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1084
1085 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1086
1087 const TypeTuple* r = boxing->tf()->range();
1088 assert(r->cnt() > TypeFunc::Parms, "sanity");
1089 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1090 assert(t != NULL, "sanity");
1091
1092 CompileLog* log = C->log();
1093 if (log != NULL) {
1094 log->head("eliminate_boxing type='%d'",
1095 log->identify(t->instance_klass()));
1096 JVMState* p = boxing->jvms();
1097 while (p != NULL) {
1098 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1099 p = p->caller();
1100 }
1101 log->tail("eliminate_boxing");
1102 }
1103
1104 process_users_of_allocation(boxing);
1105
1106 #ifndef PRODUCT
1107 if (PrintEliminateAllocations) {
1108 tty->print("++++ Eliminated: %d ", boxing->_idx);
1109 boxing->method()->print_short_name(tty);
1110 tty->cr();
1111 }
1112 #endif
1113
1114 return true;
1115 }
1116
1117
1118 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1119 Node* adr = basic_plus_adr(base, offset);
1120 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1121 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1122 transform_later(value);
1123 return value;
1124 }
1125
1126
1127 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1128 Node* adr = basic_plus_adr(base, offset);
1129 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1130 transform_later(mem);
1131 return mem;
1132 }
1133
1134 //=============================================================================
1135 //
1136 // A L L O C A T I O N
1137 //
1138 // Allocation attempts to be fast in the case of frequent small objects.
1139 // It breaks down like this:
1140 //
1141 // 1) Size in doublewords is computed. This is a constant for objects and
1142 // variable for most arrays. Doubleword units are used to avoid size
1143 // overflow of huge doubleword arrays. We need doublewords in the end for
1144 // rounding.
1145 //
1146 // 2) Size is checked for being 'too large'. Too-large allocations will go
1147 // the slow path into the VM. The slow path can throw any required
1148 // exceptions, and does all the special checks for very large arrays. The
1149 // size test can constant-fold away for objects. For objects with
1150 // finalizers it constant-folds the otherway: you always go slow with
1151 // finalizers.
1152 //
1153 // 3) If NOT using TLABs, this is the contended loop-back point.
1154 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1155 //
1156 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1157 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1158 // "size*8" we always enter the VM, where "largish" is a constant picked small
1159 // enough that there's always space between the eden max and 4Gig (old space is
1160 // there so it's quite large) and large enough that the cost of entering the VM
1161 // is dwarfed by the cost to initialize the space.
1162 //
1163 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1164 // down. If contended, repeat at step 3. If using TLABs normal-store
1165 // adjusted heap top back down; there is no contention.
1166 //
1167 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1168 // fields.
1169 //
1170 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1171 // oop flavor.
1172 //
1173 //=============================================================================
1174 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1175 // Allocations bigger than this always go the slow route.
1176 // This value must be small enough that allocation attempts that need to
1177 // trigger exceptions go the slow route. Also, it must be small enough so
1178 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1179 //=============================================================================j//
1180 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1181 // The allocator will coalesce int->oop copies away. See comment in
1182 // coalesce.cpp about how this works. It depends critically on the exact
1183 // code shape produced here, so if you are changing this code shape
1184 // make sure the GC info for the heap-top is correct in and around the
1185 // slow-path call.
1186 //
1187
1188 void PhaseMacroExpand::expand_allocate_common(
1189 AllocateNode* alloc, // allocation node to be expanded
1190 Node* length, // array length for an array allocation
1191 const TypeFunc* slow_call_type, // Type of slow call
1192 address slow_call_address, // Address of slow call
1193 Node* valid_length_test // whether length is valid or not
1194 )
1195 {
1196 Node* ctrl = alloc->in(TypeFunc::Control);
1197 Node* mem = alloc->in(TypeFunc::Memory);
1198 Node* i_o = alloc->in(TypeFunc::I_O);
1199 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1200 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1201 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1202 assert(ctrl != NULL, "must have control");
1203
1204 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1205 // they will not be used if "always_slow" is set
1206 enum { slow_result_path = 1, fast_result_path = 2 };
1207 Node *result_region = NULL;
1208 Node *result_phi_rawmem = NULL;
1209 Node *result_phi_rawoop = NULL;
1210 Node *result_phi_i_o = NULL;
1211
1212 // The initial slow comparison is a size check, the comparison
1213 // we want to do is a BoolTest::gt
1214 bool expand_fast_path = true;
1215 int tv = _igvn.find_int_con(initial_slow_test, -1);
1216 if (tv >= 0) {
1217 // InitialTest has constant result
1218 // 0 - can fit in TLAB
1219 // 1 - always too big or negative
1220 assert(tv <= 1, "0 or 1 if a constant");
1221 expand_fast_path = (tv == 0);
1222 initial_slow_test = NULL;
1223 } else {
1224 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1225 }
1226
1227 if (!UseTLAB) {
1228 // Force slow-path allocation
1229 expand_fast_path = false;
1230 initial_slow_test = NULL;
1231 }
1232
1233 bool allocation_has_use = (alloc->result_cast() != NULL);
1234 if (!allocation_has_use) {
1235 InitializeNode* init = alloc->initialization();
1236 if (init != NULL) {
1237 init->remove(&_igvn);
1238 }
1239 if (expand_fast_path && (initial_slow_test == NULL)) {
1240 // Remove allocation node and return.
1241 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1242 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1243 #ifndef PRODUCT
1244 if (PrintEliminateAllocations) {
1245 tty->print("NotUsed ");
1246 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1247 if (res != NULL) {
1248 res->dump();
1249 } else {
1250 alloc->dump();
1251 }
1252 }
1253 #endif
1254 yank_alloc_node(alloc);
1255 return;
1256 }
1257 }
1258
1259 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1260 Node *slow_region = NULL;
1261 Node *toobig_false = ctrl;
1262
1263 // generate the initial test if necessary
1264 if (initial_slow_test != NULL ) {
1265 assert (expand_fast_path, "Only need test if there is a fast path");
1266 slow_region = new RegionNode(3);
1267
1268 // Now make the initial failure test. Usually a too-big test but
1269 // might be a TRUE for finalizers or a fancy class check for
1270 // newInstance0.
1271 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1272 transform_later(toobig_iff);
1273 // Plug the failing-too-big test into the slow-path region
1274 Node *toobig_true = new IfTrueNode( toobig_iff );
1275 transform_later(toobig_true);
1276 slow_region ->init_req( too_big_or_final_path, toobig_true );
1277 toobig_false = new IfFalseNode( toobig_iff );
1278 transform_later(toobig_false);
1279 } else {
1280 // No initial test, just fall into next case
1281 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1282 toobig_false = ctrl;
1283 debug_only(slow_region = NodeSentinel);
1284 }
1285
1286 // If we are here there are several possibilities
1287 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1288 // no_initial_check means a constant allocation.
1289 // - If check always evaluates to false -> expand_fast_path is false (see above)
1290 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1291 // if !allocation_has_use the fast path is empty
1292 // if !allocation_has_use && no_initial_check
1293 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1294 // removed by yank_alloc_node above.
1295
1296 Node *slow_mem = mem; // save the current memory state for slow path
1297 // generate the fast allocation code unless we know that the initial test will always go slow
1298 if (expand_fast_path) {
1299 // Fast path modifies only raw memory.
1300 if (mem->is_MergeMem()) {
1301 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1302 }
1303
1304 // allocate the Region and Phi nodes for the result
1305 result_region = new RegionNode(3);
1306 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1307 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1308
1309 // Grab regular I/O before optional prefetch may change it.
1310 // Slow-path does no I/O so just set it to the original I/O.
1311 result_phi_i_o->init_req(slow_result_path, i_o);
1312
1313 // Name successful fast-path variables
1314 Node* fast_oop_ctrl;
1315 Node* fast_oop_rawmem;
1316 if (allocation_has_use) {
1317 Node* needgc_ctrl = NULL;
1318 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1319
1320 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1321 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1322 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1323 fast_oop_ctrl, fast_oop_rawmem,
1324 prefetch_lines);
1325
1326 if (initial_slow_test != NULL) {
1327 // This completes all paths into the slow merge point
1328 slow_region->init_req(need_gc_path, needgc_ctrl);
1329 transform_later(slow_region);
1330 } else {
1331 // No initial slow path needed!
1332 // Just fall from the need-GC path straight into the VM call.
1333 slow_region = needgc_ctrl;
1334 }
1335
1336 InitializeNode* init = alloc->initialization();
1337 fast_oop_rawmem = initialize_object(alloc,
1338 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1339 klass_node, length, size_in_bytes);
1340 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1341 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1342
1343 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1344 } else {
1345 assert (initial_slow_test != NULL, "sanity");
1346 fast_oop_ctrl = toobig_false;
1347 fast_oop_rawmem = mem;
1348 transform_later(slow_region);
1349 }
1350
1351 // Plug in the successful fast-path into the result merge point
1352 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1353 result_phi_i_o ->init_req(fast_result_path, i_o);
1354 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1355 } else {
1356 slow_region = ctrl;
1357 result_phi_i_o = i_o; // Rename it to use in the following code.
1358 }
1359
1360 // Generate slow-path call
1361 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1362 OptoRuntime::stub_name(slow_call_address),
1363 TypePtr::BOTTOM);
1364 call->init_req(TypeFunc::Control, slow_region);
1365 call->init_req(TypeFunc::I_O, top()); // does no i/o
1366 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1367 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1368 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1369
1370 call->init_req(TypeFunc::Parms+0, klass_node);
1371 if (length != NULL) {
1372 call->init_req(TypeFunc::Parms+1, length);
1373 }
1374
1375 // Copy debug information and adjust JVMState information, then replace
1376 // allocate node with the call
1377 call->copy_call_debug_info(&_igvn, alloc);
1378 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1379 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1380 // path dies).
1381 if (valid_length_test != NULL) {
1382 call->add_req(valid_length_test);
1383 }
1384 if (expand_fast_path) {
1385 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1386 } else {
1387 // Hook i_o projection to avoid its elimination during allocation
1388 // replacement (when only a slow call is generated).
1389 call->set_req(TypeFunc::I_O, result_phi_i_o);
1390 }
1391 _igvn.replace_node(alloc, call);
1392 transform_later(call);
1393
1394 // Identify the output projections from the allocate node and
1395 // adjust any references to them.
1396 // The control and io projections look like:
1397 //
1398 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1399 // Allocate Catch
1400 // ^---Proj(io) <-------+ ^---CatchProj(io)
1401 //
1402 // We are interested in the CatchProj nodes.
1403 //
1404 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1405
1406 // An allocate node has separate memory projections for the uses on
1407 // the control and i_o paths. Replace the control memory projection with
1408 // result_phi_rawmem (unless we are only generating a slow call when
1409 // both memory projections are combined)
1410 if (expand_fast_path && _callprojs.fallthrough_memproj != NULL) {
1411 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1412 }
1413 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1414 // catchall_memproj so we end up with a call that has only 1 memory projection.
1415 if (_callprojs.catchall_memproj != NULL ) {
1416 if (_callprojs.fallthrough_memproj == NULL) {
1417 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1418 transform_later(_callprojs.fallthrough_memproj);
1419 }
1420 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1421 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1422 }
1423
1424 // An allocate node has separate i_o projections for the uses on the control
1425 // and i_o paths. Always replace the control i_o projection with result i_o
1426 // otherwise incoming i_o become dead when only a slow call is generated
1427 // (it is different from memory projections where both projections are
1428 // combined in such case).
1429 if (_callprojs.fallthrough_ioproj != NULL) {
1430 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1431 }
1432 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1433 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1434 if (_callprojs.catchall_ioproj != NULL ) {
1435 if (_callprojs.fallthrough_ioproj == NULL) {
1436 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1437 transform_later(_callprojs.fallthrough_ioproj);
1438 }
1439 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1440 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1441 }
1442
1443 // if we generated only a slow call, we are done
1444 if (!expand_fast_path) {
1445 // Now we can unhook i_o.
1446 if (result_phi_i_o->outcnt() > 1) {
1447 call->set_req(TypeFunc::I_O, top());
1448 } else {
1449 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1450 // Case of new array with negative size known during compilation.
1451 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1452 // following code since call to runtime will throw exception.
1453 // As result there will be no users of i_o after the call.
1454 // Leave i_o attached to this call to avoid problems in preceding graph.
1455 }
1456 return;
1457 }
1458
1459 if (_callprojs.fallthrough_catchproj != NULL) {
1460 ctrl = _callprojs.fallthrough_catchproj->clone();
1461 transform_later(ctrl);
1462 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1463 } else {
1464 ctrl = top();
1465 }
1466 Node *slow_result;
1467 if (_callprojs.resproj == NULL) {
1468 // no uses of the allocation result
1469 slow_result = top();
1470 } else {
1471 slow_result = _callprojs.resproj->clone();
1472 transform_later(slow_result);
1473 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1474 }
1475
1476 // Plug slow-path into result merge point
1477 result_region->init_req( slow_result_path, ctrl);
1478 transform_later(result_region);
1479 if (allocation_has_use) {
1480 result_phi_rawoop->init_req(slow_result_path, slow_result);
1481 transform_later(result_phi_rawoop);
1482 }
1483 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1484 transform_later(result_phi_rawmem);
1485 transform_later(result_phi_i_o);
1486 // This completes all paths into the result merge point
1487 }
1488
1489 // Remove alloc node that has no uses.
1490 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1491 Node* ctrl = alloc->in(TypeFunc::Control);
1492 Node* mem = alloc->in(TypeFunc::Memory);
1493 Node* i_o = alloc->in(TypeFunc::I_O);
1494
1495 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1496 if (_callprojs.resproj != NULL) {
1497 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1498 Node* use = _callprojs.resproj->fast_out(i);
1499 use->isa_MemBar()->remove(&_igvn);
1500 --imax;
1501 --i; // back up iterator
1502 }
1503 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1504 _igvn.remove_dead_node(_callprojs.resproj);
1505 }
1506 if (_callprojs.fallthrough_catchproj != NULL) {
1507 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1508 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1509 }
1510 if (_callprojs.catchall_catchproj != NULL) {
1511 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1512 _callprojs.catchall_catchproj->set_req(0, top());
1513 }
1514 if (_callprojs.fallthrough_proj != NULL) {
1515 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1516 _igvn.remove_dead_node(catchnode);
1517 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1518 }
1519 if (_callprojs.fallthrough_memproj != NULL) {
1520 migrate_outs(_callprojs.fallthrough_memproj, mem);
1521 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1522 }
1523 if (_callprojs.fallthrough_ioproj != NULL) {
1524 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1525 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1526 }
1527 if (_callprojs.catchall_memproj != NULL) {
1528 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1529 _callprojs.catchall_memproj->set_req(0, top());
1530 }
1531 if (_callprojs.catchall_ioproj != NULL) {
1532 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1533 _callprojs.catchall_ioproj->set_req(0, top());
1534 }
1535 #ifndef PRODUCT
1536 if (PrintEliminateAllocations) {
1537 if (alloc->is_AllocateArray()) {
1538 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1539 } else {
1540 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1541 }
1542 }
1543 #endif
1544 _igvn.remove_dead_node(alloc);
1545 }
1546
1547 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1548 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1549 // If initialization is performed by an array copy, any required
1550 // MemBarStoreStore was already added. If the object does not
1551 // escape no need for a MemBarStoreStore. If the object does not
1552 // escape in its initializer and memory barrier (MemBarStoreStore or
1553 // stronger) is already added at exit of initializer, also no need
1554 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1555 // so that stores that initialize this object can't be reordered
1556 // with a subsequent store that makes this object accessible by
1557 // other threads.
1558 // Other threads include java threads and JVM internal threads
1559 // (for example concurrent GC threads). Current concurrent GC
1560 // implementation: G1 will not scan newly created object,
1561 // so it's safe to skip storestore barrier when allocation does
1562 // not escape.
1563 if (!alloc->does_not_escape_thread() &&
1564 !alloc->is_allocation_MemBar_redundant() &&
1565 (init == NULL || !init->is_complete_with_arraycopy())) {
1566 if (init == NULL || init->req() < InitializeNode::RawStores) {
1567 // No InitializeNode or no stores captured by zeroing
1568 // elimination. Simply add the MemBarStoreStore after object
1569 // initialization.
1570 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1571 transform_later(mb);
1572
1573 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1574 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1575 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1576 transform_later(fast_oop_ctrl);
1577 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1578 transform_later(fast_oop_rawmem);
1579 } else {
1580 // Add the MemBarStoreStore after the InitializeNode so that
1581 // all stores performing the initialization that were moved
1582 // before the InitializeNode happen before the storestore
1583 // barrier.
1584
1585 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1586 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1587
1588 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1589 transform_later(mb);
1590
1591 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1592 transform_later(ctrl);
1593 Node* mem = new ProjNode(init, TypeFunc::Memory);
1594 transform_later(mem);
1595
1596 // The MemBarStoreStore depends on control and memory coming
1597 // from the InitializeNode
1598 mb->init_req(TypeFunc::Memory, mem);
1599 mb->init_req(TypeFunc::Control, ctrl);
1600
1601 ctrl = new ProjNode(mb, TypeFunc::Control);
1602 transform_later(ctrl);
1603 mem = new ProjNode(mb, TypeFunc::Memory);
1604 transform_later(mem);
1605
1606 // All nodes that depended on the InitializeNode for control
1607 // and memory must now depend on the MemBarNode that itself
1608 // depends on the InitializeNode
1609 if (init_ctrl != NULL) {
1610 _igvn.replace_node(init_ctrl, ctrl);
1611 }
1612 if (init_mem != NULL) {
1613 _igvn.replace_node(init_mem, mem);
1614 }
1615 }
1616 }
1617 }
1618
1619 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1620 Node*& ctrl, Node*& rawmem) {
1621 if (C->env()->dtrace_alloc_probes()) {
1622 // Slow-path call
1623 int size = TypeFunc::Parms + 2;
1624 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1625 CAST_FROM_FN_PTR(address,
1626 static_cast<int (*)(JavaThread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1627 "dtrace_object_alloc",
1628 TypeRawPtr::BOTTOM);
1629
1630 // Get base of thread-local storage area
1631 Node* thread = new ThreadLocalNode();
1632 transform_later(thread);
1633
1634 call->init_req(TypeFunc::Parms + 0, thread);
1635 call->init_req(TypeFunc::Parms + 1, oop);
1636 call->init_req(TypeFunc::Control, ctrl);
1637 call->init_req(TypeFunc::I_O , top()); // does no i/o
1638 call->init_req(TypeFunc::Memory , rawmem);
1639 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1640 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1641 transform_later(call);
1642 ctrl = new ProjNode(call, TypeFunc::Control);
1643 transform_later(ctrl);
1644 rawmem = new ProjNode(call, TypeFunc::Memory);
1645 transform_later(rawmem);
1646 }
1647 }
1648
1649 // Helper for PhaseMacroExpand::expand_allocate_common.
1650 // Initializes the newly-allocated storage.
1651 Node*
1652 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1653 Node* control, Node* rawmem, Node* object,
1654 Node* klass_node, Node* length,
1655 Node* size_in_bytes) {
1656 InitializeNode* init = alloc->initialization();
1657 // Store the klass & mark bits
1658 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1659 if (!mark_node->is_Con()) {
1660 transform_later(mark_node);
1661 }
1662 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1663
1664 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1665 int header_size = alloc->minimum_header_size(); // conservatively small
1666
1667 // Array length
1668 if (length != NULL) { // Arrays need length field
1669 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1670 // conservatively small header size:
1671 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1672 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1673 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1674 if (is_reference_type(elem, true)) {
1675 elem = T_OBJECT;
1676 }
1677 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1678 }
1679 }
1680
1681 // Clear the object body, if necessary.
1682 if (init == NULL) {
1683 // The init has somehow disappeared; be cautious and clear everything.
1684 //
1685 // This can happen if a node is allocated but an uncommon trap occurs
1686 // immediately. In this case, the Initialize gets associated with the
1687 // trap, and may be placed in a different (outer) loop, if the Allocate
1688 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1689 // there can be two Allocates to one Initialize. The answer in all these
1690 // edge cases is safety first. It is always safe to clear immediately
1691 // within an Allocate, and then (maybe or maybe not) clear some more later.
1692 if (!(UseTLAB && ZeroTLAB)) {
1693 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1694 header_size, size_in_bytes,
1695 &_igvn);
1696 }
1697 } else {
1698 if (!init->is_complete()) {
1699 // Try to win by zeroing only what the init does not store.
1700 // We can also try to do some peephole optimizations,
1701 // such as combining some adjacent subword stores.
1702 rawmem = init->complete_stores(control, rawmem, object,
1703 header_size, size_in_bytes, &_igvn);
1704 }
1705 // We have no more use for this link, since the AllocateNode goes away:
1706 init->set_req(InitializeNode::RawAddress, top());
1707 // (If we keep the link, it just confuses the register allocator,
1708 // who thinks he sees a real use of the address by the membar.)
1709 }
1710
1711 return rawmem;
1712 }
1713
1714 // Generate prefetch instructions for next allocations.
1715 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1716 Node*& contended_phi_rawmem,
1717 Node* old_eden_top, Node* new_eden_top,
1718 intx lines) {
1719 enum { fall_in_path = 1, pf_path = 2 };
1720 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1721 // Generate prefetch allocation with watermark check.
1722 // As an allocation hits the watermark, we will prefetch starting
1723 // at a "distance" away from watermark.
1724
1725 Node *pf_region = new RegionNode(3);
1726 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1727 TypeRawPtr::BOTTOM );
1728 // I/O is used for Prefetch
1729 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1730
1731 Node *thread = new ThreadLocalNode();
1732 transform_later(thread);
1733
1734 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1735 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1736 transform_later(eden_pf_adr);
1737
1738 Node *old_pf_wm = new LoadPNode(needgc_false,
1739 contended_phi_rawmem, eden_pf_adr,
1740 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1741 MemNode::unordered);
1742 transform_later(old_pf_wm);
1743
1744 // check against new_eden_top
1745 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1746 transform_later(need_pf_cmp);
1747 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1748 transform_later(need_pf_bol);
1749 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1750 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1751 transform_later(need_pf_iff);
1752
1753 // true node, add prefetchdistance
1754 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1755 transform_later(need_pf_true);
1756
1757 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1758 transform_later(need_pf_false);
1759
1760 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1761 _igvn.MakeConX(AllocatePrefetchDistance) );
1762 transform_later(new_pf_wmt );
1763 new_pf_wmt->set_req(0, need_pf_true);
1764
1765 Node *store_new_wmt = new StorePNode(need_pf_true,
1766 contended_phi_rawmem, eden_pf_adr,
1767 TypeRawPtr::BOTTOM, new_pf_wmt,
1768 MemNode::unordered);
1769 transform_later(store_new_wmt);
1770
1771 // adding prefetches
1772 pf_phi_abio->init_req( fall_in_path, i_o );
1773
1774 Node *prefetch_adr;
1775 Node *prefetch;
1776 uint step_size = AllocatePrefetchStepSize;
1777 uint distance = 0;
1778
1779 for ( intx i = 0; i < lines; i++ ) {
1780 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1781 _igvn.MakeConX(distance) );
1782 transform_later(prefetch_adr);
1783 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1784 transform_later(prefetch);
1785 distance += step_size;
1786 i_o = prefetch;
1787 }
1788 pf_phi_abio->set_req( pf_path, i_o );
1789
1790 pf_region->init_req( fall_in_path, need_pf_false );
1791 pf_region->init_req( pf_path, need_pf_true );
1792
1793 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1794 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1795
1796 transform_later(pf_region);
1797 transform_later(pf_phi_rawmem);
1798 transform_later(pf_phi_abio);
1799
1800 needgc_false = pf_region;
1801 contended_phi_rawmem = pf_phi_rawmem;
1802 i_o = pf_phi_abio;
1803 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1804 // Insert a prefetch instruction for each allocation.
1805 // This code is used to generate 1 prefetch instruction per cache line.
1806
1807 // Generate several prefetch instructions.
1808 uint step_size = AllocatePrefetchStepSize;
1809 uint distance = AllocatePrefetchDistance;
1810
1811 // Next cache address.
1812 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1813 _igvn.MakeConX(step_size + distance));
1814 transform_later(cache_adr);
1815 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1816 transform_later(cache_adr);
1817 // Address is aligned to execute prefetch to the beginning of cache line size
1818 // (it is important when BIS instruction is used on SPARC as prefetch).
1819 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1820 cache_adr = new AndXNode(cache_adr, mask);
1821 transform_later(cache_adr);
1822 cache_adr = new CastX2PNode(cache_adr);
1823 transform_later(cache_adr);
1824
1825 // Prefetch
1826 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1827 prefetch->set_req(0, needgc_false);
1828 transform_later(prefetch);
1829 contended_phi_rawmem = prefetch;
1830 Node *prefetch_adr;
1831 distance = step_size;
1832 for ( intx i = 1; i < lines; i++ ) {
1833 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1834 _igvn.MakeConX(distance) );
1835 transform_later(prefetch_adr);
1836 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1837 transform_later(prefetch);
1838 distance += step_size;
1839 contended_phi_rawmem = prefetch;
1840 }
1841 } else if( AllocatePrefetchStyle > 0 ) {
1842 // Insert a prefetch for each allocation only on the fast-path
1843 Node *prefetch_adr;
1844 Node *prefetch;
1845 // Generate several prefetch instructions.
1846 uint step_size = AllocatePrefetchStepSize;
1847 uint distance = AllocatePrefetchDistance;
1848 for ( intx i = 0; i < lines; i++ ) {
1849 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1850 _igvn.MakeConX(distance) );
1851 transform_later(prefetch_adr);
1852 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1853 // Do not let it float too high, since if eden_top == eden_end,
1854 // both might be null.
1855 if( i == 0 ) { // Set control for first prefetch, next follows it
1856 prefetch->init_req(0, needgc_false);
1857 }
1858 transform_later(prefetch);
1859 distance += step_size;
1860 i_o = prefetch;
1861 }
1862 }
1863 return i_o;
1864 }
1865
1866
1867 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1868 expand_allocate_common(alloc, NULL,
1869 OptoRuntime::new_instance_Type(),
1870 OptoRuntime::new_instance_Java(), NULL);
1871 }
1872
1873 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1874 Node* length = alloc->in(AllocateNode::ALength);
1875 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1876 InitializeNode* init = alloc->initialization();
1877 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1878 const TypeAryKlassPtr* ary_klass_t = _igvn.type(klass_node)->isa_aryklassptr();
1879 address slow_call_address; // Address of slow call
1880 if (init != NULL && init->is_complete_with_arraycopy() &&
1881 ary_klass_t && ary_klass_t->elem()->isa_klassptr() == NULL) {
1882 // Don't zero type array during slow allocation in VM since
1883 // it will be initialized later by arraycopy in compiled code.
1884 slow_call_address = OptoRuntime::new_array_nozero_Java();
1885 } else {
1886 slow_call_address = OptoRuntime::new_array_Java();
1887 }
1888 expand_allocate_common(alloc, length,
1889 OptoRuntime::new_array_Type(),
1890 slow_call_address, valid_length_test);
1891 }
1892
1893 //-------------------mark_eliminated_box----------------------------------
1894 //
1895 // During EA obj may point to several objects but after few ideal graph
1896 // transformations (CCP) it may point to only one non escaping object
1897 // (but still using phi), corresponding locks and unlocks will be marked
1898 // for elimination. Later obj could be replaced with a new node (new phi)
1899 // and which does not have escape information. And later after some graph
1900 // reshape other locks and unlocks (which were not marked for elimination
1901 // before) are connected to this new obj (phi) but they still will not be
1902 // marked for elimination since new obj has no escape information.
1903 // Mark all associated (same box and obj) lock and unlock nodes for
1904 // elimination if some of them marked already.
1905 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1906 if (oldbox->as_BoxLock()->is_eliminated()) {
1907 return; // This BoxLock node was processed already.
1908 }
1909 // New implementation (EliminateNestedLocks) has separate BoxLock
1910 // node for each locked region so mark all associated locks/unlocks as
1911 // eliminated even if different objects are referenced in one locked region
1912 // (for example, OSR compilation of nested loop inside locked scope).
1913 if (EliminateNestedLocks ||
1914 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj, NULL)) {
1915 // Box is used only in one lock region. Mark this box as eliminated.
1916 _igvn.hash_delete(oldbox);
1917 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1918 _igvn.hash_insert(oldbox);
1919
1920 for (uint i = 0; i < oldbox->outcnt(); i++) {
1921 Node* u = oldbox->raw_out(i);
1922 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1923 AbstractLockNode* alock = u->as_AbstractLock();
1924 // Check lock's box since box could be referenced by Lock's debug info.
1925 if (alock->box_node() == oldbox) {
1926 // Mark eliminated all related locks and unlocks.
1927 #ifdef ASSERT
1928 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1929 #endif
1930 alock->set_non_esc_obj();
1931 }
1932 }
1933 }
1934 return;
1935 }
1936
1937 // Create new "eliminated" BoxLock node and use it in monitor debug info
1938 // instead of oldbox for the same object.
1939 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1940
1941 // Note: BoxLock node is marked eliminated only here and it is used
1942 // to indicate that all associated lock and unlock nodes are marked
1943 // for elimination.
1944 newbox->set_eliminated();
1945 transform_later(newbox);
1946
1947 // Replace old box node with new box for all users of the same object.
1948 for (uint i = 0; i < oldbox->outcnt();) {
1949 bool next_edge = true;
1950
1951 Node* u = oldbox->raw_out(i);
1952 if (u->is_AbstractLock()) {
1953 AbstractLockNode* alock = u->as_AbstractLock();
1954 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1955 // Replace Box and mark eliminated all related locks and unlocks.
1956 #ifdef ASSERT
1957 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1958 #endif
1959 alock->set_non_esc_obj();
1960 _igvn.rehash_node_delayed(alock);
1961 alock->set_box_node(newbox);
1962 next_edge = false;
1963 }
1964 }
1965 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1966 FastLockNode* flock = u->as_FastLock();
1967 assert(flock->box_node() == oldbox, "sanity");
1968 _igvn.rehash_node_delayed(flock);
1969 flock->set_box_node(newbox);
1970 next_edge = false;
1971 }
1972
1973 // Replace old box in monitor debug info.
1974 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1975 SafePointNode* sfn = u->as_SafePoint();
1976 JVMState* youngest_jvms = sfn->jvms();
1977 int max_depth = youngest_jvms->depth();
1978 for (int depth = 1; depth <= max_depth; depth++) {
1979 JVMState* jvms = youngest_jvms->of_depth(depth);
1980 int num_mon = jvms->nof_monitors();
1981 // Loop over monitors
1982 for (int idx = 0; idx < num_mon; idx++) {
1983 Node* obj_node = sfn->monitor_obj(jvms, idx);
1984 Node* box_node = sfn->monitor_box(jvms, idx);
1985 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1986 int j = jvms->monitor_box_offset(idx);
1987 _igvn.replace_input_of(u, j, newbox);
1988 next_edge = false;
1989 }
1990 }
1991 }
1992 }
1993 if (next_edge) i++;
1994 }
1995 }
1996
1997 //-----------------------mark_eliminated_locking_nodes-----------------------
1998 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1999 if (EliminateNestedLocks) {
2000 if (alock->is_nested()) {
2001 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2002 return;
2003 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2004 // Only Lock node has JVMState needed here.
2005 // Not that preceding claim is documented anywhere else.
2006 if (alock->jvms() != NULL) {
2007 if (alock->as_Lock()->is_nested_lock_region()) {
2008 // Mark eliminated related nested locks and unlocks.
2009 Node* obj = alock->obj_node();
2010 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2011 assert(!box_node->is_eliminated(), "should not be marked yet");
2012 // Note: BoxLock node is marked eliminated only here
2013 // and it is used to indicate that all associated lock
2014 // and unlock nodes are marked for elimination.
2015 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2016 for (uint i = 0; i < box_node->outcnt(); i++) {
2017 Node* u = box_node->raw_out(i);
2018 if (u->is_AbstractLock()) {
2019 alock = u->as_AbstractLock();
2020 if (alock->box_node() == box_node) {
2021 // Verify that this Box is referenced only by related locks.
2022 assert(alock->obj_node()->eqv_uncast(obj), "");
2023 // Mark all related locks and unlocks.
2024 #ifdef ASSERT
2025 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2026 #endif
2027 alock->set_nested();
2028 }
2029 }
2030 }
2031 } else {
2032 #ifdef ASSERT
2033 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2034 if (C->log() != NULL)
2035 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2036 #endif
2037 }
2038 }
2039 return;
2040 }
2041 // Process locks for non escaping object
2042 assert(alock->is_non_esc_obj(), "");
2043 } // EliminateNestedLocks
2044
2045 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2046 // Look for all locks of this object and mark them and
2047 // corresponding BoxLock nodes as eliminated.
2048 Node* obj = alock->obj_node();
2049 for (uint j = 0; j < obj->outcnt(); j++) {
2050 Node* o = obj->raw_out(j);
2051 if (o->is_AbstractLock() &&
2052 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2053 alock = o->as_AbstractLock();
2054 Node* box = alock->box_node();
2055 // Replace old box node with new eliminated box for all users
2056 // of the same object and mark related locks as eliminated.
2057 mark_eliminated_box(box, obj);
2058 }
2059 }
2060 }
2061 }
2062
2063 // we have determined that this lock/unlock can be eliminated, we simply
2064 // eliminate the node without expanding it.
2065 //
2066 // Note: The membar's associated with the lock/unlock are currently not
2067 // eliminated. This should be investigated as a future enhancement.
2068 //
2069 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2070
2071 if (!alock->is_eliminated()) {
2072 return false;
2073 }
2074 #ifdef ASSERT
2075 if (!alock->is_coarsened()) {
2076 // Check that new "eliminated" BoxLock node is created.
2077 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2078 assert(oldbox->is_eliminated(), "should be done already");
2079 }
2080 #endif
2081
2082 alock->log_lock_optimization(C, "eliminate_lock");
2083
2084 #ifndef PRODUCT
2085 if (PrintEliminateLocks) {
2086 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2087 }
2088 #endif
2089
2090 Node* mem = alock->in(TypeFunc::Memory);
2091 Node* ctrl = alock->in(TypeFunc::Control);
2092 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2093
2094 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2095 // There are 2 projections from the lock. The lock node will
2096 // be deleted when its last use is subsumed below.
2097 assert(alock->outcnt() == 2 &&
2098 _callprojs.fallthrough_proj != NULL &&
2099 _callprojs.fallthrough_memproj != NULL,
2100 "Unexpected projections from Lock/Unlock");
2101
2102 Node* fallthroughproj = _callprojs.fallthrough_proj;
2103 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2104
2105 // The memory projection from a lock/unlock is RawMem
2106 // The input to a Lock is merged memory, so extract its RawMem input
2107 // (unless the MergeMem has been optimized away.)
2108 if (alock->is_Lock()) {
2109 // Search for MemBarAcquireLock node and delete it also.
2110 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2111 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2112 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2113 Node* memproj = membar->proj_out(TypeFunc::Memory);
2114 _igvn.replace_node(ctrlproj, fallthroughproj);
2115 _igvn.replace_node(memproj, memproj_fallthrough);
2116
2117 // Delete FastLock node also if this Lock node is unique user
2118 // (a loop peeling may clone a Lock node).
2119 Node* flock = alock->as_Lock()->fastlock_node();
2120 if (flock->outcnt() == 1) {
2121 assert(flock->unique_out() == alock, "sanity");
2122 _igvn.replace_node(flock, top());
2123 }
2124 }
2125
2126 // Search for MemBarReleaseLock node and delete it also.
2127 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2128 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2129 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2130 mem->is_Proj() && membar == mem->in(0), "");
2131 _igvn.replace_node(fallthroughproj, ctrl);
2132 _igvn.replace_node(memproj_fallthrough, mem);
2133 fallthroughproj = ctrl;
2134 memproj_fallthrough = mem;
2135 ctrl = membar->in(TypeFunc::Control);
2136 mem = membar->in(TypeFunc::Memory);
2137 }
2138
2139 _igvn.replace_node(fallthroughproj, ctrl);
2140 _igvn.replace_node(memproj_fallthrough, mem);
2141 return true;
2142 }
2143
2144
2145 //------------------------------expand_lock_node----------------------
2146 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2147
2148 Node* ctrl = lock->in(TypeFunc::Control);
2149 Node* mem = lock->in(TypeFunc::Memory);
2150 Node* obj = lock->obj_node();
2151 Node* box = lock->box_node();
2152 Node* flock = lock->fastlock_node();
2153
2154 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2155
2156 // Make the merge point
2157 Node *region;
2158 Node *mem_phi;
2159 Node *slow_path;
2160
2161 region = new RegionNode(3);
2162 // create a Phi for the memory state
2163 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2164
2165 // Optimize test; set region slot 2
2166 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2167 mem_phi->init_req(2, mem);
2168
2169 // Make slow path call
2170 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2171 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2172 obj, box, NULL);
2173
2174 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2175
2176 // Slow path can only throw asynchronous exceptions, which are always
2177 // de-opted. So the compiler thinks the slow-call can never throw an
2178 // exception. If it DOES throw an exception we would need the debug
2179 // info removed first (since if it throws there is no monitor).
2180 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2181 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2182
2183 // Capture slow path
2184 // disconnect fall-through projection from call and create a new one
2185 // hook up users of fall-through projection to region
2186 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2187 transform_later(slow_ctrl);
2188 _igvn.hash_delete(_callprojs.fallthrough_proj);
2189 _callprojs.fallthrough_proj->disconnect_inputs(C);
2190 region->init_req(1, slow_ctrl);
2191 // region inputs are now complete
2192 transform_later(region);
2193 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2194
2195 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2196
2197 mem_phi->init_req(1, memproj);
2198
2199 transform_later(mem_phi);
2200
2201 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2202 }
2203
2204 //------------------------------expand_unlock_node----------------------
2205 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2206
2207 Node* ctrl = unlock->in(TypeFunc::Control);
2208 Node* mem = unlock->in(TypeFunc::Memory);
2209 Node* obj = unlock->obj_node();
2210 Node* box = unlock->box_node();
2211
2212 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2213
2214 // No need for a null check on unlock
2215
2216 // Make the merge point
2217 Node *region;
2218 Node *mem_phi;
2219
2220 region = new RegionNode(3);
2221 // create a Phi for the memory state
2222 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2223
2224 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2225 funlock = transform_later( funlock )->as_FastUnlock();
2226 // Optimize test; set region slot 2
2227 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2228 Node *thread = transform_later(new ThreadLocalNode());
2229
2230 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2231 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2232 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2233
2234 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2235 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2236 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2237
2238 // No exceptions for unlocking
2239 // Capture slow path
2240 // disconnect fall-through projection from call and create a new one
2241 // hook up users of fall-through projection to region
2242 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2243 transform_later(slow_ctrl);
2244 _igvn.hash_delete(_callprojs.fallthrough_proj);
2245 _callprojs.fallthrough_proj->disconnect_inputs(C);
2246 region->init_req(1, slow_ctrl);
2247 // region inputs are now complete
2248 transform_later(region);
2249 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2250
2251 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2252 mem_phi->init_req(1, memproj );
2253 mem_phi->init_req(2, mem);
2254 transform_later(mem_phi);
2255
2256 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2257 }
2258
2259 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2260 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2261 Node* bol = check->unique_out();
2262 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2263 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2264 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2265
2266 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2267 Node* iff = bol->last_out(i);
2268 assert(iff->is_If(), "where's the if?");
2269
2270 if (iff->in(0)->is_top()) {
2271 _igvn.replace_input_of(iff, 1, C->top());
2272 continue;
2273 }
2274
2275 Node* iftrue = iff->as_If()->proj_out(1);
2276 Node* iffalse = iff->as_If()->proj_out(0);
2277 Node* ctrl = iff->in(0);
2278
2279 Node* subklass = NULL;
2280 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2281 subklass = obj_or_subklass;
2282 } else {
2283 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2284 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2285 }
2286
2287 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2288
2289 _igvn.replace_input_of(iff, 0, C->top());
2290 _igvn.replace_node(iftrue, not_subtype_ctrl);
2291 _igvn.replace_node(iffalse, ctrl);
2292 }
2293 _igvn.replace_node(check, C->top());
2294 }
2295
2296 //---------------------------eliminate_macro_nodes----------------------
2297 // Eliminate scalar replaced allocations and associated locks.
2298 void PhaseMacroExpand::eliminate_macro_nodes() {
2299 if (C->macro_count() == 0)
2300 return;
2301 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2302
2303 // Before elimination may re-mark (change to Nested or NonEscObj)
2304 // all associated (same box and obj) lock and unlock nodes.
2305 int cnt = C->macro_count();
2306 for (int i=0; i < cnt; i++) {
2307 Node *n = C->macro_node(i);
2308 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2309 mark_eliminated_locking_nodes(n->as_AbstractLock());
2310 }
2311 }
2312 // Re-marking may break consistency of Coarsened locks.
2313 if (!C->coarsened_locks_consistent()) {
2314 return; // recompile without Coarsened locks if broken
2315 }
2316
2317 // First, attempt to eliminate locks
2318 bool progress = true;
2319 while (progress) {
2320 progress = false;
2321 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2322 Node* n = C->macro_node(i - 1);
2323 bool success = false;
2324 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2325 if (n->is_AbstractLock()) {
2326 success = eliminate_locking_node(n->as_AbstractLock());
2327 #ifndef PRODUCT
2328 if (success && PrintOptoStatistics) {
2329 Atomic::inc(&PhaseMacroExpand::_monitor_objects_removed_counter);
2330 }
2331 #endif
2332 }
2333 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2334 progress = progress || success;
2335 }
2336 }
2337 // Next, attempt to eliminate allocations
2338 _has_locks = false;
2339 progress = true;
2340 while (progress) {
2341 progress = false;
2342 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2343 Node* n = C->macro_node(i - 1);
2344 bool success = false;
2345 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2346 switch (n->class_id()) {
2347 case Node::Class_Allocate:
2348 case Node::Class_AllocateArray:
2349 success = eliminate_allocate_node(n->as_Allocate());
2350 #ifndef PRODUCT
2351 if (success && PrintOptoStatistics) {
2352 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2353 }
2354 #endif
2355 break;
2356 case Node::Class_CallStaticJava:
2357 success = eliminate_boxing_node(n->as_CallStaticJava());
2358 break;
2359 case Node::Class_Lock:
2360 case Node::Class_Unlock:
2361 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2362 _has_locks = true;
2363 break;
2364 case Node::Class_ArrayCopy:
2365 break;
2366 case Node::Class_OuterStripMinedLoop:
2367 break;
2368 case Node::Class_SubTypeCheck:
2369 break;
2370 case Node::Class_Opaque1:
2371 break;
2372 default:
2373 assert(n->Opcode() == Op_LoopLimit ||
2374 n->Opcode() == Op_Opaque3 ||
2375 n->Opcode() == Op_Opaque4 ||
2376 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2377 "unknown node type in macro list");
2378 }
2379 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2380 progress = progress || success;
2381 }
2382 }
2383 #ifndef PRODUCT
2384 if (PrintOptoStatistics) {
2385 int membar_after = count_MemBar(C);
2386 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2387 }
2388 #endif
2389 }
2390
2391 //------------------------------expand_macro_nodes----------------------
2392 // Returns true if a failure occurred.
2393 bool PhaseMacroExpand::expand_macro_nodes() {
2394 // Last attempt to eliminate macro nodes.
2395 eliminate_macro_nodes();
2396 if (C->failing()) return true;
2397
2398 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2399 bool progress = true;
2400 while (progress) {
2401 progress = false;
2402 for (int i = C->macro_count(); i > 0; i--) {
2403 Node* n = C->macro_node(i-1);
2404 bool success = false;
2405 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2406 if (n->Opcode() == Op_LoopLimit) {
2407 // Remove it from macro list and put on IGVN worklist to optimize.
2408 C->remove_macro_node(n);
2409 _igvn._worklist.push(n);
2410 success = true;
2411 } else if (n->Opcode() == Op_CallStaticJava) {
2412 // Remove it from macro list and put on IGVN worklist to optimize.
2413 C->remove_macro_node(n);
2414 _igvn._worklist.push(n);
2415 success = true;
2416 } else if (n->is_Opaque1()) {
2417 _igvn.replace_node(n, n->in(1));
2418 success = true;
2419 #if INCLUDE_RTM_OPT
2420 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2421 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2422 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2423 Node* cmp = n->unique_out();
2424 #ifdef ASSERT
2425 // Validate graph.
2426 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2427 BoolNode* bol = cmp->unique_out()->as_Bool();
2428 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2429 (bol->_test._test == BoolTest::ne), "");
2430 IfNode* ifn = bol->unique_out()->as_If();
2431 assert((ifn->outcnt() == 2) &&
2432 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2433 #endif
2434 Node* repl = n->in(1);
2435 if (!_has_locks) {
2436 // Remove RTM state check if there are no locks in the code.
2437 // Replace input to compare the same value.
2438 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2439 }
2440 _igvn.replace_node(n, repl);
2441 success = true;
2442 #endif
2443 } else if (n->Opcode() == Op_Opaque4) {
2444 // With Opaque4 nodes, the expectation is that the test of input 1
2445 // is always equal to the constant value of input 2. So we can
2446 // remove the Opaque4 and replace it by input 2. In debug builds,
2447 // leave the non constant test in instead to sanity check that it
2448 // never fails (if it does, that subgraph was constructed so, at
2449 // runtime, a Halt node is executed).
2450 #ifdef ASSERT
2451 _igvn.replace_node(n, n->in(1));
2452 #else
2453 _igvn.replace_node(n, n->in(2));
2454 #endif
2455 success = true;
2456 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2457 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2458 C->remove_macro_node(n);
2459 success = true;
2460 }
2461 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2462 progress = progress || success;
2463 }
2464 }
2465
2466 // Clean up the graph so we're less likely to hit the maximum node
2467 // limit
2468 _igvn.set_delay_transform(false);
2469 _igvn.optimize();
2470 if (C->failing()) return true;
2471 _igvn.set_delay_transform(true);
2472
2473
2474 // Because we run IGVN after each expansion, some macro nodes may go
2475 // dead and be removed from the list as we iterate over it. Move
2476 // Allocate nodes (processed in a second pass) at the beginning of
2477 // the list and then iterate from the last element of the list until
2478 // an Allocate node is seen. This is robust to random deletion in
2479 // the list due to nodes going dead.
2480 C->sort_macro_nodes();
2481
2482 // expand arraycopy "macro" nodes first
2483 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2484 // before the allocate nodes are expanded.
2485 while (C->macro_count() > 0) {
2486 int macro_count = C->macro_count();
2487 Node * n = C->macro_node(macro_count-1);
2488 assert(n->is_macro(), "only macro nodes expected here");
2489 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2490 // node is unreachable, so don't try to expand it
2491 C->remove_macro_node(n);
2492 continue;
2493 }
2494 if (n->is_Allocate()) {
2495 break;
2496 }
2497 // Make sure expansion will not cause node limit to be exceeded.
2498 // Worst case is a macro node gets expanded into about 200 nodes.
2499 // Allow 50% more for optimization.
2500 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2501 return true;
2502 }
2503
2504 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2505 switch (n->class_id()) {
2506 case Node::Class_Lock:
2507 expand_lock_node(n->as_Lock());
2508 break;
2509 case Node::Class_Unlock:
2510 expand_unlock_node(n->as_Unlock());
2511 break;
2512 case Node::Class_ArrayCopy:
2513 expand_arraycopy_node(n->as_ArrayCopy());
2514 break;
2515 case Node::Class_SubTypeCheck:
2516 expand_subtypecheck_node(n->as_SubTypeCheck());
2517 break;
2518 default:
2519 assert(false, "unknown node type in macro list");
2520 }
2521 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2522 if (C->failing()) return true;
2523
2524 // Clean up the graph so we're less likely to hit the maximum node
2525 // limit
2526 _igvn.set_delay_transform(false);
2527 _igvn.optimize();
2528 if (C->failing()) return true;
2529 _igvn.set_delay_transform(true);
2530 }
2531
2532 // All nodes except Allocate nodes are expanded now. There could be
2533 // new optimization opportunities (such as folding newly created
2534 // load from a just allocated object). Run IGVN.
2535
2536 // expand "macro" nodes
2537 // nodes are removed from the macro list as they are processed
2538 while (C->macro_count() > 0) {
2539 int macro_count = C->macro_count();
2540 Node * n = C->macro_node(macro_count-1);
2541 assert(n->is_macro(), "only macro nodes expected here");
2542 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2543 // node is unreachable, so don't try to expand it
2544 C->remove_macro_node(n);
2545 continue;
2546 }
2547 // Make sure expansion will not cause node limit to be exceeded.
2548 // Worst case is a macro node gets expanded into about 200 nodes.
2549 // Allow 50% more for optimization.
2550 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2551 return true;
2552 }
2553 switch (n->class_id()) {
2554 case Node::Class_Allocate:
2555 expand_allocate(n->as_Allocate());
2556 break;
2557 case Node::Class_AllocateArray:
2558 expand_allocate_array(n->as_AllocateArray());
2559 break;
2560 default:
2561 assert(false, "unknown node type in macro list");
2562 }
2563 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2564 if (C->failing()) return true;
2565
2566 // Clean up the graph so we're less likely to hit the maximum node
2567 // limit
2568 _igvn.set_delay_transform(false);
2569 _igvn.optimize();
2570 if (C->failing()) return true;
2571 _igvn.set_delay_transform(true);
2572 }
2573
2574 _igvn.set_delay_transform(false);
2575 return false;
2576 }
2577
2578 #ifndef PRODUCT
2579 int PhaseMacroExpand::_objs_scalar_replaced_counter = 0;
2580 int PhaseMacroExpand::_monitor_objects_removed_counter = 0;
2581 int PhaseMacroExpand::_GC_barriers_removed_counter = 0;
2582 int PhaseMacroExpand::_memory_barriers_removed_counter = 0;
2583
2584 void PhaseMacroExpand::print_statistics() {
2585 tty->print("Objects scalar replaced = %d, ", Atomic::load(&_objs_scalar_replaced_counter));
2586 tty->print("Monitor objects removed = %d, ", Atomic::load(&_monitor_objects_removed_counter));
2587 tty->print("GC barriers removed = %d, ", Atomic::load(&_GC_barriers_removed_counter));
2588 tty->print_cr("Memory barriers removed = %d", Atomic::load(&_memory_barriers_removed_counter));
2589 }
2590
2591 int PhaseMacroExpand::count_MemBar(Compile *C) {
2592 if (!PrintOptoStatistics) {
2593 return 0;
2594 }
2595 Unique_Node_List ideal_nodes;
2596 int total = 0;
2597 ideal_nodes.map(C->live_nodes(), NULL);
2598 ideal_nodes.push(C->root());
2599 for (uint next = 0; next < ideal_nodes.size(); ++next) {
2600 Node* n = ideal_nodes.at(next);
2601 if (n->is_MemBar()) {
2602 total++;
2603 }
2604 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2605 Node* m = n->fast_out(i);
2606 ideal_nodes.push(m);
2607 }
2608 }
2609 return total;
2610 }
2611 #endif