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
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
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) {
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;
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(); )
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";)
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;
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());
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) {
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
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
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
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();
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 }
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) {
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
|
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 "ci/ciFlatArrayKlass.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/collectedHeap.inline.hpp"
29 #include "gc/shared/tlab_globals.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/universe.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/loopnode.hpp"
44 #include "opto/macro.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/phaseX.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/subtypenode.hpp"
54 #include "opto/type.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/continuation.hpp"
57 #include "runtime/sharedRuntime.hpp"
58 #include "runtime/stubRoutines.hpp"
59 #include "utilities/macros.hpp"
60 #include "utilities/powerOfTwo.hpp"
61 #if INCLUDE_G1GC
62 #include "gc/g1/g1ThreadLocalData.hpp"
63 #endif // INCLUDE_G1GC
64
65
66 //
67 // Replace any references to "oldref" in inputs to "use" with "newref".
68 // Returns the number of replacements made.
69 //
70 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
71 int nreplacements = 0;
72 uint req = use->req();
73 for (uint j = 0; j < use->len(); j++) {
74 Node *uin = use->in(j);
75 if (uin == oldref) {
76 if (j < req)
77 use->set_req(j, newref);
78 else
79 use->set_prec(j, newref);
80 nreplacements++;
81 } else if (j >= req && uin == NULL) {
82 break;
83 }
84 }
85 return nreplacements;
86 }
87
88 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
89 Node* cmp;
90 if (mask != 0) {
91 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
92 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
93 } else {
94 cmp = word;
95 }
96 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
97 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
98 transform_later(iff);
99
100 // Fast path taken.
101 Node *fast_taken = transform_later(new IfFalseNode(iff));
102
103 // Fast path not-taken, i.e. slow path
104 Node *slow_taken = transform_later(new IfTrueNode(iff));
105
106 if (return_fast_path) {
107 region->init_req(edge, slow_taken); // Capture slow-control
130 // Slow-path call
131 CallNode *call = leaf_name
132 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
133 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
134
135 // Slow path call has no side-effects, uses few values
136 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
137 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
138 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
139 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
140 call->copy_call_debug_info(&_igvn, oldcall);
141 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
142 _igvn.replace_node(oldcall, call);
143 transform_later(call);
144
145 return call;
146 }
147
148 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
149 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
150 bs->eliminate_gc_barrier(&_igvn, p2x);
151 #ifndef PRODUCT
152 if (PrintOptoStatistics) {
153 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
154 }
155 #endif
156 }
157
158 // Search for a memory operation for the specified memory slice.
159 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
160 Node *orig_mem = mem;
161 Node *alloc_mem = alloc->in(TypeFunc::Memory);
162 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
163 while (true) {
164 if (mem == alloc_mem || mem == start_mem ) {
165 return mem; // hit one of our sentinels
166 } else if (mem->is_MergeMem()) {
167 mem = mem->as_MergeMem()->memory_at(alias_idx);
168 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
169 Node *in = mem->in(0);
170 // we can safely skip over safepoints, calls, locks and membars because we
184 ArrayCopyNode* ac = NULL;
185 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
186 if (ac != NULL) {
187 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
188 return ac;
189 }
190 }
191 mem = in->in(TypeFunc::Memory);
192 } else {
193 #ifdef ASSERT
194 in->dump();
195 mem->dump();
196 assert(false, "unexpected projection");
197 #endif
198 }
199 } else if (mem->is_Store()) {
200 const TypePtr* atype = mem->as_Store()->adr_type();
201 int adr_idx = phase->C->get_alias_index(atype);
202 if (adr_idx == alias_idx) {
203 assert(atype->isa_oopptr(), "address type must be oopptr");
204 int adr_offset = atype->flattened_offset();
205 uint adr_iid = atype->is_oopptr()->instance_id();
206 // Array elements references have the same alias_idx
207 // but different offset and different instance_id.
208 if (adr_offset == offset && adr_iid == alloc->_idx) {
209 return mem;
210 }
211 } else {
212 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
213 }
214 mem = mem->in(MemNode::Memory);
215 } else if (mem->is_ClearArray()) {
216 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
217 // Can not bypass initialization of the instance
218 // we are looking.
219 debug_only(intptr_t offset;)
220 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
221 InitializeNode* init = alloc->as_Allocate()->initialization();
222 // We are looking for stored value, return Initialize node
223 // or memory edge from Allocate node.
224 if (init != NULL) {
229 }
230 // Otherwise skip it (the call updated 'mem' value).
231 } else if (mem->Opcode() == Op_SCMemProj) {
232 mem = mem->in(0);
233 Node* adr = NULL;
234 if (mem->is_LoadStore()) {
235 adr = mem->in(MemNode::Address);
236 } else {
237 assert(mem->Opcode() == Op_EncodeISOArray ||
238 mem->Opcode() == Op_StrCompressedCopy, "sanity");
239 adr = mem->in(3); // Destination array
240 }
241 const TypePtr* atype = adr->bottom_type()->is_ptr();
242 int adr_idx = phase->C->get_alias_index(atype);
243 if (adr_idx == alias_idx) {
244 DEBUG_ONLY(mem->dump();)
245 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
246 return NULL;
247 }
248 mem = mem->in(MemNode::Memory);
249 } else if (mem->Opcode() == Op_StrInflatedCopy) {
250 Node* adr = mem->in(3); // Destination array
251 const TypePtr* atype = adr->bottom_type()->is_ptr();
252 int adr_idx = phase->C->get_alias_index(atype);
253 if (adr_idx == alias_idx) {
254 DEBUG_ONLY(mem->dump();)
255 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
256 return NULL;
257 }
258 mem = mem->in(MemNode::Memory);
259 } else {
260 return mem;
261 }
262 assert(mem != orig_mem, "dead memory loop");
263 }
264 }
265
266 // Generate loads from source of the arraycopy for fields of
267 // destination needed at a deoptimization point
268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
269 BasicType bt = ft;
274 }
275 Node* res = NULL;
276 if (ac->is_clonebasic()) {
277 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
278 Node* base = ac->in(ArrayCopyNode::Src);
279 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
280 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
281 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
282 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
283 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
284 } else {
285 if (ac->modifies(offset, offset, &_igvn, true)) {
286 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
287 uint shift = exact_log2(type2aelembytes(bt));
288 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
289 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
290 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
291 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
292
293 Node* adr = NULL;
294 Node* base = ac->in(ArrayCopyNode::Src);
295 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
296 if (adr_type->is_flat()) {
297 shift = adr_type->flat_log_elem_size();
298 }
299 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
300 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
301 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
302 adr_type = _igvn.type(adr)->is_aryptr();
303 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Don't emit a new load from src if src == dst but try to get the value from memory instead
306 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
307 }
308 } else {
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Non constant offset in the array: we can't statically
311 // determine the value
312 return NULL;
313 }
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 adr = _igvn.transform(new AddPNode(base, base, off));
322 // In the case of a flattened inline type array, each field has its
323 // own slice so we need to extract the field being accessed from
324 // the address computation
325 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
326 adr = _igvn.transform(new CastPPNode(adr, adr_type));
327 }
328 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
329 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
330 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
331 }
332 }
333 if (res != NULL) {
334 if (ftype->isa_narrowoop()) {
335 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
336 assert(res->isa_DecodeN(), "should be narrow oop");
337 res = _igvn.transform(new EncodePNode(res, ftype));
338 }
339 return res;
340 }
341 return NULL;
342 }
343
344 //
345 // Given a Memory Phi, compute a value Phi containing the values from stores
346 // on the input paths.
347 // Note: this function is recursive, its depth is limited by the "level" argument
348 // Returns the computed Phi, or NULL if it cannot compute it.
349 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) {
350 assert(mem->is_Phi(), "sanity");
351 int alias_idx = C->get_alias_index(adr_t);
352 int offset = adr_t->flattened_offset();
353 int instance_id = adr_t->instance_id();
354
355 // Check if an appropriate value phi already exists.
356 Node* region = mem->in(0);
357 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
358 Node* phi = region->fast_out(k);
359 if (phi->is_Phi() && phi != mem &&
360 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
361 return phi;
362 }
363 }
364 // Check if an appropriate new value phi already exists.
365 Node* new_phi = value_phis->find(mem->_idx);
366 if (new_phi != NULL)
367 return new_phi;
368
369 if (level <= 0) {
370 return NULL; // Give up: phi tree too deep
371 }
372 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
373 Node *alloc_mem = alloc->in(TypeFunc::Memory);
374
375 uint length = mem->req();
376 GrowableArray <Node *> values(length, length, NULL);
377
378 // create a new Phi for the value
379 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
380 transform_later(phi);
381 value_phis->push(phi, mem->_idx);
382
383 for (uint j = 1; j < length; j++) {
384 Node *in = mem->in(j);
385 if (in == NULL || in->is_top()) {
386 values.at_put(j, in);
387 } else {
388 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
389 if (val == start_mem || val == alloc_mem) {
390 // hit a sentinel, return appropriate 0 value
391 Node* default_value = alloc->in(AllocateNode::DefaultValue);
392 if (default_value != NULL) {
393 values.at_put(j, default_value);
394 } else {
395 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
396 values.at_put(j, _igvn.zerocon(ft));
397 }
398 continue;
399 }
400 if (val->is_Initialize()) {
401 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
402 }
403 if (val == NULL) {
404 return NULL; // can't find a value on this path
405 }
406 if (val == mem) {
407 values.at_put(j, mem);
408 } else if (val->is_Store()) {
409 Node* n = val->in(MemNode::ValueIn);
410 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
411 n = bs->step_over_gc_barrier(n);
412 if (is_subword_type(ft)) {
413 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
414 }
415 values.at_put(j, n);
416 } else if(val->is_Proj() && val->in(0) == alloc) {
417 Node* default_value = alloc->in(AllocateNode::DefaultValue);
418 if (default_value != NULL) {
419 values.at_put(j, default_value);
420 } else {
421 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
422 values.at_put(j, _igvn.zerocon(ft));
423 }
424 } else if (val->is_Phi()) {
425 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
426 if (val == NULL) {
427 return NULL;
428 }
429 values.at_put(j, val);
430 } else if (val->Opcode() == Op_SCMemProj) {
431 assert(val->in(0)->is_LoadStore() ||
432 val->in(0)->Opcode() == Op_EncodeISOArray ||
433 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
434 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
435 return NULL;
436 } else if (val->is_ArrayCopy()) {
437 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
438 if (res == NULL) {
439 return NULL;
440 }
441 values.at_put(j, res);
442 } else {
443 DEBUG_ONLY( val->dump(); )
447 }
448 }
449 // Set Phi's inputs
450 for (uint j = 1; j < length; j++) {
451 if (values.at(j) == mem) {
452 phi->init_req(j, phi);
453 } else {
454 phi->init_req(j, values.at(j));
455 }
456 }
457 return phi;
458 }
459
460 // Search the last value stored into the object's field.
461 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
462 assert(adr_t->is_known_instance_field(), "instance required");
463 int instance_id = adr_t->instance_id();
464 assert((uint)instance_id == alloc->_idx, "wrong allocation");
465
466 int alias_idx = C->get_alias_index(adr_t);
467 int offset = adr_t->flattened_offset();
468 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
469 Node *alloc_mem = alloc->in(TypeFunc::Memory);
470 VectorSet visited;
471
472 bool done = sfpt_mem == alloc_mem;
473 Node *mem = sfpt_mem;
474 while (!done) {
475 if (visited.test_set(mem->_idx)) {
476 return NULL; // found a loop, give up
477 }
478 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
479 if (mem == start_mem || mem == alloc_mem) {
480 done = true; // hit a sentinel, return appropriate 0 value
481 } else if (mem->is_Initialize()) {
482 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
483 if (mem == NULL) {
484 done = true; // Something went wrong.
485 } else if (mem->is_Store()) {
486 const TypePtr* atype = mem->as_Store()->adr_type();
487 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
488 done = true;
489 }
490 } else if (mem->is_Store()) {
491 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
492 assert(atype != NULL, "address type must be oopptr");
493 assert(C->get_alias_index(atype) == alias_idx &&
494 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
495 atype->instance_id() == instance_id, "store is correct memory slice");
496 done = true;
497 } else if (mem->is_Phi()) {
498 // try to find a phi's unique input
499 Node *unique_input = NULL;
500 Node *top = C->top();
501 for (uint i = 1; i < mem->req(); i++) {
502 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
503 if (n == NULL || n == top || n == mem) {
504 continue;
505 } else if (unique_input == NULL) {
506 unique_input = n;
507 } else if (unique_input != n) {
508 unique_input = top;
509 break;
510 }
511 }
512 if (unique_input != NULL && unique_input != top) {
513 mem = unique_input;
514 } else {
515 done = true;
516 }
517 } else if (mem->is_ArrayCopy()) {
518 done = true;
519 } else {
520 DEBUG_ONLY( mem->dump(); )
521 assert(false, "unexpected node");
522 }
523 }
524 if (mem != NULL) {
525 if (mem == start_mem || mem == alloc_mem) {
526 // hit a sentinel, return appropriate 0 value
527 Node* default_value = alloc->in(AllocateNode::DefaultValue);
528 if (default_value != NULL) {
529 return default_value;
530 }
531 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
532 return _igvn.zerocon(ft);
533 } else if (mem->is_Store()) {
534 Node* n = mem->in(MemNode::ValueIn);
535 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
536 n = bs->step_over_gc_barrier(n);
537 return n;
538 } else if (mem->is_Phi()) {
539 // attempt to produce a Phi reflecting the values on the input paths of the Phi
540 Node_Stack value_phis(8);
541 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
542 if (phi != NULL) {
543 return phi;
544 } else {
545 // Kill all new Phis
546 while(value_phis.is_nonempty()) {
547 Node* n = value_phis.node();
548 _igvn.replace_node(n, C->top());
549 value_phis.pop();
550 }
551 }
552 } else if (mem->is_ArrayCopy()) {
553 Node* ctl = mem->in(0);
554 Node* m = mem->in(TypeFunc::Memory);
555 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
556 // pin the loads in the uncommon trap path
557 ctl = sfpt_ctl;
558 m = sfpt_mem;
559 }
560 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
561 }
562 }
563 // Something went wrong.
564 return NULL;
565 }
566
567 // Search the last value stored into the inline type's fields.
568 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
569 // Subtract the offset of the first field to account for the missing oop header
570 offset -= vk->first_field_offset();
571 // Create a new InlineTypeNode and retrieve the field values from memory
572 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
573 transform_later(vt);
574 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
575 ciType* field_type = vt->field_type(i);
576 int field_offset = offset + vt->field_offset(i);
577 Node* value = NULL;
578 if (vt->field_is_flattened(i)) {
579 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
580 } else {
581 const Type* ft = Type::get_const_type(field_type);
582 BasicType bt = type2field[field_type->basic_type()];
583 if (UseCompressedOops && !is_java_primitive(bt)) {
584 ft = ft->make_narrowoop();
585 bt = T_NARROWOOP;
586 }
587 // Each inline type field has its own memory slice
588 adr_type = adr_type->with_field_offset(field_offset);
589 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
590 if (value != NULL && ft->isa_narrowoop()) {
591 assert(UseCompressedOops, "unexpected narrow oop");
592 if (value->is_EncodeP()) {
593 value = value->in(1);
594 } else {
595 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
596 }
597 }
598 }
599 if (value != NULL) {
600 vt->set_field_value(i, value);
601 } else {
602 // We might have reached the TrackedInitializationLimit
603 return NULL;
604 }
605 }
606 return vt;
607 }
608
609 // Check the possibility of scalar replacement.
610 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
611 // Scan the uses of the allocation to check for anything that would
612 // prevent us from eliminating it.
613 NOT_PRODUCT( const char* fail_eliminate = NULL; )
614 DEBUG_ONLY( Node* disq_node = NULL; )
615 bool can_eliminate = true;
616
617 Unique_Node_List worklist;
618 Node* res = alloc->result_cast();
619 const TypeOopPtr* res_type = NULL;
620 if (res == NULL) {
621 // All users were eliminated.
622 } else if (!res->is_CheckCastPP()) {
623 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
624 can_eliminate = false;
625 } else {
626 worklist.push(res);
627 res_type = _igvn.type(res)->isa_oopptr();
628 if (res_type == NULL) {
629 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
630 can_eliminate = false;
631 } else if (res_type->isa_aryptr()) {
632 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
633 if (length < 0) {
634 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
635 can_eliminate = false;
636 }
637 }
638 }
639
640 while (can_eliminate && worklist.size() > 0) {
641 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
642 res = worklist.pop();
643 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
644 Node* use = res->fast_out(j);
645
646 if (use->is_AddP()) {
647 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
648 int offset = addp_type->offset();
649
650 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
651 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
652 can_eliminate = false;
653 break;
654 }
655 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
656 k < kmax && can_eliminate; k++) {
657 Node* n = use->fast_out(k);
658 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n)) {
659 DEBUG_ONLY(disq_node = n;)
660 if (n->is_Load() || n->is_LoadStore()) {
661 NOT_PRODUCT(fail_eliminate = "Field load";)
662 } else {
663 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
671 use->as_ArrayCopy()->is_copyof_validated() ||
672 use->as_ArrayCopy()->is_copyofrange_validated()) &&
673 use->in(ArrayCopyNode::Dest) == res) {
674 // ok to eliminate
675 } else if (use->is_SafePoint()) {
676 SafePointNode* sfpt = use->as_SafePoint();
677 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
678 // Object is passed as argument.
679 DEBUG_ONLY(disq_node = use;)
680 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
681 can_eliminate = false;
682 }
683 Node* sfptMem = sfpt->memory();
684 if (sfptMem == NULL || sfptMem->is_top()) {
685 DEBUG_ONLY(disq_node = use;)
686 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
687 can_eliminate = false;
688 } else {
689 safepoints.append_if_missing(sfpt);
690 }
691 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
692 // Look at uses
693 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
694 Node* u = use->fast_out(k);
695 if (u->is_InlineType()) {
696 // Use in flat field can be eliminated
697 InlineTypeNode* vt = u->as_InlineType();
698 for (uint i = 0; i < vt->field_count(); ++i) {
699 if (vt->field_value(i) == use && !vt->field_is_flattened(i)) {
700 can_eliminate = false; // Use in non-flattened field
701 break;
702 }
703 }
704 } else {
705 // Add other uses to the worklist to process individually
706 worklist.push(u);
707 }
708 }
709 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
710 // Store to mark word of inline type larval buffer
711 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
712 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
713 if (use->is_Phi()) {
714 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
715 NOT_PRODUCT(fail_eliminate = "Object is return value";)
716 } else {
717 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
718 }
719 DEBUG_ONLY(disq_node = use;)
720 } else {
721 if (use->Opcode() == Op_Return) {
722 NOT_PRODUCT(fail_eliminate = "Object is return value";)
723 } else {
724 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
725 }
726 DEBUG_ONLY(disq_node = use;)
727 }
728 can_eliminate = false;
729 } else {
730 assert(use->Opcode() == Op_CastP2X, "should be");
731 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
732 }
733 }
734 }
735
736 #ifndef PRODUCT
737 if (PrintEliminateAllocations) {
738 if (can_eliminate) {
739 tty->print("Scalar ");
740 if (res == NULL)
741 alloc->dump();
742 else
743 res->dump();
744 } else {
745 tty->print("NotScalar (%s)", fail_eliminate);
746 if (res == NULL)
747 alloc->dump();
748 else
749 res->dump();
750 #ifdef ASSERT
751 if (disq_node != NULL) {
752 tty->print(" >>>> ");
753 disq_node->dump();
754 }
755 #endif /*ASSERT*/
756 }
757 }
758 #endif
759 return can_eliminate;
760 }
761
762 // Do scalar replacement.
763 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
764 GrowableArray <SafePointNode *> safepoints_done;
773 Node* res = alloc->result_cast();
774 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
775 const TypeOopPtr* res_type = NULL;
776 if (res != NULL) { // Could be NULL when there are no users
777 res_type = _igvn.type(res)->isa_oopptr();
778 }
779
780 if (res != NULL) {
781 if (res_type->isa_instptr()) {
782 // find the fields of the class which will be needed for safepoint debug information
783 iklass = res_type->is_instptr()->instance_klass();
784 nfields = iklass->nof_nonstatic_fields();
785 } else {
786 // find the array's elements which will be needed for safepoint debug information
787 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
788 assert(nfields >= 0, "must be an array klass.");
789 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
790 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
791 element_size = type2aelembytes(basic_elem_type);
792 field_type = res_type->is_aryptr()->elem();
793 if (res_type->is_flat()) {
794 // Flattened inline type array
795 element_size = res_type->is_aryptr()->flat_elem_size();
796 }
797 }
798 }
799 //
800 // Process the safepoint uses
801 //
802 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr(), "Inline type allocations should not have safepoint uses");
803 Unique_Node_List value_worklist;
804 while (safepoints.length() > 0) {
805 SafePointNode* sfpt = safepoints.pop();
806 Node* mem = sfpt->memory();
807 Node* ctl = sfpt->control();
808 assert(sfpt->jvms() != NULL, "missed JVMS");
809 // Fields of scalar objs are referenced only at the end
810 // of regular debuginfo at the last (youngest) JVMS.
811 // Record relative start index.
812 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
813 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
814 #ifdef ASSERT
815 alloc,
816 #endif
817 first_ind, nfields);
818 sobj->init_req(0, C->root());
819 transform_later(sobj);
820
821 // Scan object's fields adding an input to the safepoint for each field.
822 for (int j = 0; j < nfields; j++) {
823 intptr_t offset;
824 ciField* field = NULL;
825 if (iklass != NULL) {
826 field = iklass->nonstatic_field_at(j);
827 offset = field->offset();
828 ciType* elem_type = field->type();
829 basic_elem_type = field->layout_type();
830 assert(!field->is_flattened(), "flattened inline type fields should not have safepoint uses");
831
832 // The next code is taken from Parse::do_get_xxx().
833 if (is_reference_type(basic_elem_type)) {
834 if (!elem_type->is_loaded()) {
835 field_type = TypeInstPtr::BOTTOM;
836 } else if (field != NULL && field->is_static_constant()) {
837 ciObject* con = field->constant_value().as_object();
838 // Do not "join" in the previous type; it doesn't add value,
839 // and may yield a vacuous result if the field is of interface type.
840 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
841 assert(field_type != NULL, "field singleton type must be consistent");
842 } else {
843 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
844 }
845 if (UseCompressedOops) {
846 field_type = field_type->make_narrowoop();
847 basic_elem_type = T_NARROWOOP;
848 }
849 } else {
850 field_type = Type::get_const_basic_type(basic_elem_type);
851 }
852 } else {
853 offset = array_base + j * (intptr_t)element_size;
854 }
855
856 Node* field_val = NULL;
857 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
858 if (res_type->is_flat()) {
859 ciInlineKlass* vk = res_type->is_aryptr()->elem()->inline_klass();
860 assert(vk->flatten_array(), "must be flattened");
861 field_val = inline_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
862 } else {
863 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
864 }
865 if (field_val == NULL) {
866 // We weren't able to find a value for this field,
867 // give up on eliminating this allocation.
868
869 // Remove any extra entries we added to the safepoint.
870 uint last = sfpt->req() - 1;
871 for (int k = 0; k < j; k++) {
872 sfpt->del_req(last--);
873 }
874 _igvn._worklist.push(sfpt);
875 // rollback processed safepoints
876 while (safepoints_done.length() > 0) {
877 SafePointNode* sfpt_done = safepoints_done.pop();
878 // remove any extra entries we added to the safepoint
879 last = sfpt_done->req() - 1;
880 for (int k = 0; k < nfields; k++) {
881 sfpt_done->del_req(last--);
882 }
883 JVMState *jvms = sfpt_done->jvms();
884 jvms->set_endoff(sfpt_done->req());
907 int field_idx = C->get_alias_index(field_addr_type);
908 tty->print(" (alias_idx=%d)", field_idx);
909 } else { // Array's element
910 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
911 sfpt->_idx, j);
912 }
913 tty->print(", which prevents elimination of: ");
914 if (res == NULL)
915 alloc->dump();
916 else
917 res->dump();
918 }
919 #endif
920 return false;
921 }
922 if (UseCompressedOops && field_type->isa_narrowoop()) {
923 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
924 // to be able scalar replace the allocation.
925 if (field_val->is_EncodeP()) {
926 field_val = field_val->in(1);
927 } else if (!field_val->is_InlineType()) {
928 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
929 }
930 }
931 if (field_val->is_InlineType()) {
932 // Keep track of inline types to scalarize them later
933 value_worklist.push(field_val);
934 }
935 sfpt->add_req(field_val);
936 }
937 JVMState *jvms = sfpt->jvms();
938 jvms->set_endoff(sfpt->req());
939 // Now make a pass over the debug information replacing any references
940 // to the allocated object with "sobj"
941 int start = jvms->debug_start();
942 int end = jvms->debug_end();
943 sfpt->replace_edges_in_range(res, sobj, start, end, &_igvn);
944 _igvn._worklist.push(sfpt);
945 safepoints_done.append_if_missing(sfpt); // keep it for rollback
946 }
947 // Scalarize inline types that were added to the safepoint.
948 // Don't allow linking a constant oop (if available) for flat array elements
949 // because Deoptimization::reassign_flat_array_elements needs field values.
950 bool allow_oop = (res_type != NULL) && !res_type->is_flat();
951 for (uint i = 0; i < value_worklist.size(); ++i) {
952 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
953 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
954 }
955 return true;
956 }
957
958 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
959 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
960 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
961 if (ctl_proj != NULL) {
962 igvn.replace_node(ctl_proj, n->in(0));
963 }
964 if (mem_proj != NULL) {
965 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
966 }
967 }
968
969 // Process users of eliminated allocation.
970 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
971 Unique_Node_List worklist;
972 Node* res = alloc->result_cast();
973 if (res != NULL) {
974 worklist.push(res);
975 }
976 while (worklist.size() > 0) {
977 res = worklist.pop();
978 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
979 Node *use = res->last_out(j);
980 uint oc1 = res->outcnt();
981
982 if (use->is_AddP()) {
983 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
984 Node *n = use->last_out(k);
985 uint oc2 = use->outcnt();
986 if (n->is_Store()) {
987 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
988 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
989 if (mb != NULL && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
990 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
991 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
992 mb->remove(&_igvn);
993 }
994 }
995 _igvn.replace_node(n, n->in(MemNode::Memory));
996 } else {
997 eliminate_gc_barrier(n);
998 }
999 k -= (oc2 - use->outcnt());
1000 }
1001 _igvn.remove_dead_node(use);
1002 } else if (use->is_ArrayCopy()) {
1003 // Disconnect ArrayCopy node
1004 ArrayCopyNode* ac = use->as_ArrayCopy();
1005 if (ac->is_clonebasic()) {
1006 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1007 disconnect_projections(ac, _igvn);
1008 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1009 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1010 disconnect_projections(membar_before->as_MemBar(), _igvn);
1011 if (membar_after->is_MemBar()) {
1012 disconnect_projections(membar_after->as_MemBar(), _igvn);
1013 }
1014 } else {
1015 assert(ac->is_arraycopy_validated() ||
1016 ac->is_copyof_validated() ||
1017 ac->is_copyofrange_validated(), "unsupported");
1018 CallProjections* callprojs = ac->extract_projections(true);
1019
1020 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1021 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1022 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1023
1024 // Set control to top. IGVN will remove the remaining projections
1025 ac->set_req(0, top());
1026 ac->replace_edge(res, top(), &_igvn);
1027
1028 // Disconnect src right away: it can help find new
1029 // opportunities for allocation elimination
1030 Node* src = ac->in(ArrayCopyNode::Src);
1031 ac->replace_edge(src, top(), &_igvn);
1032 // src can be top at this point if src and dest of the
1033 // arraycopy were the same
1034 if (src->outcnt() == 0 && !src->is_top()) {
1035 _igvn.remove_dead_node(src);
1036 }
1037 }
1038 _igvn._worklist.push(ac);
1039 } else if (use->is_InlineType()) {
1040 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1041 // Cut off oop input and remove known instance id from type
1042 _igvn.rehash_node_delayed(use);
1043 use->as_InlineType()->set_oop(_igvn.zerocon(T_PRIMITIVE_OBJECT));
1044 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1045 _igvn.set_type(use, toop);
1046 use->as_InlineType()->set_type(toop);
1047 // Process users
1048 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1049 Node* u = use->fast_out(k);
1050 if (!u->is_InlineType()) {
1051 worklist.push(u);
1052 }
1053 }
1054 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1055 // Store to mark word of inline type larval buffer
1056 assert(inline_alloc, "Unexpected store to mark word");
1057 _igvn.replace_node(use, use->in(MemNode::Memory));
1058 } else {
1059 eliminate_gc_barrier(use);
1060 }
1061 j -= (oc1 - res->outcnt());
1062 }
1063 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1064 _igvn.remove_dead_node(res);
1065 }
1066
1067 //
1068 // Process other users of allocation's projections
1069 //
1070 if (_callprojs->resproj[0] != NULL && _callprojs->resproj[0]->outcnt() != 0) {
1071 // First disconnect stores captured by Initialize node.
1072 // If Initialize node is eliminated first in the following code,
1073 // it will kill such stores and DUIterator_Last will assert.
1074 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1075 Node* use = _callprojs->resproj[0]->fast_out(j);
1076 if (use->is_AddP()) {
1077 // raw memory addresses used only by the initialization
1078 _igvn.replace_node(use, C->top());
1079 --j; --jmax;
1080 }
1081 }
1082 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1083 Node* use = _callprojs->resproj[0]->last_out(j);
1084 uint oc1 = _callprojs->resproj[0]->outcnt();
1085 if (use->is_Initialize()) {
1086 // Eliminate Initialize node.
1087 InitializeNode *init = use->as_Initialize();
1088 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1089 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1090 if (ctrl_proj != NULL) {
1091 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1092 #ifdef ASSERT
1093 // If the InitializeNode has no memory out, it will die, and tmp will become NULL
1094 Node* tmp = init->in(TypeFunc::Control);
1095 assert(tmp == NULL || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1096 #endif
1097 }
1098 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1099 if (mem_proj != NULL) {
1100 Node *mem = init->in(TypeFunc::Memory);
1101 #ifdef ASSERT
1102 if (mem->is_MergeMem()) {
1103 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1104 } else {
1105 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1106 }
1107 #endif
1108 _igvn.replace_node(mem_proj, mem);
1109 }
1110 } else if (use->Opcode() == Op_MemBarStoreStore) {
1111 // Inline type buffer allocations are followed by a membar
1112 assert(inline_alloc, "Unexpected MemBarStoreStore");
1113 use->as_MemBar()->remove(&_igvn);
1114 } else {
1115 assert(false, "only Initialize or AddP expected");
1116 }
1117 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1118 }
1119 }
1120 if (_callprojs->fallthrough_catchproj != NULL) {
1121 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1122 }
1123 if (_callprojs->fallthrough_memproj != NULL) {
1124 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1125 }
1126 if (_callprojs->catchall_memproj != NULL) {
1127 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1128 }
1129 if (_callprojs->fallthrough_ioproj != NULL) {
1130 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1131 }
1132 if (_callprojs->catchall_ioproj != NULL) {
1133 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1134 }
1135 if (_callprojs->catchall_catchproj != NULL) {
1136 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1137 }
1138 }
1139
1140 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1141 // If reallocation fails during deoptimization we'll pop all
1142 // interpreter frames for this compiled frame and that won't play
1143 // nice with JVMTI popframe.
1144 // We avoid this issue by eager reallocation when the popframe request
1145 // is received.
1146 if (!EliminateAllocations) {
1147 return false;
1148 }
1149 Node* klass = alloc->in(AllocateNode::KlassNode);
1150 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1151
1152 // Attempt to eliminate inline type buffer allocations
1153 // regardless of usage and escape/replaceable status.
1154 bool inline_alloc = tklass->isa_instklassptr() &&
1155 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1156 if (!alloc->_is_non_escaping && !inline_alloc) {
1157 return false;
1158 }
1159 // Eliminate boxing allocations which are not used
1160 // regardless scalar replaceable status.
1161 Node* res = alloc->result_cast();
1162 bool boxing_alloc = (res == NULL) && C->eliminate_boxing() &&
1163 tklass->isa_instklassptr() &&
1164 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1165 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1166 return false;
1167 }
1168
1169 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1170
1171 GrowableArray <SafePointNode *> safepoints;
1172 if (!can_eliminate_allocation(alloc, safepoints)) {
1173 return false;
1174 }
1175
1176 if (!alloc->_is_scalar_replaceable) {
1177 assert(res == NULL || inline_alloc, "sanity");
1178 // We can only eliminate allocation if all debug info references
1179 // are already replaced with SafePointScalarObject because
1180 // we can't search for a fields value without instance_id.
1181 if (safepoints.length() > 0) {
1182 assert(!inline_alloc, "Inline type allocations should not have safepoint uses");
1183 return false;
1184 }
1185 }
1186
1187 if (!scalar_replacement(alloc, safepoints)) {
1188 return false;
1189 }
1190
1191 CompileLog* log = C->log();
1192 if (log != NULL) {
1193 log->head("eliminate_allocation type='%d'",
1194 log->identify(tklass->exact_klass()));
1195 JVMState* p = alloc->jvms();
1196 while (p != NULL) {
1197 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1198 p = p->caller();
1199 }
1200 log->tail("eliminate_allocation");
1201 }
1202
1203 process_users_of_allocation(alloc, inline_alloc);
1204
1205 #ifndef PRODUCT
1206 if (PrintEliminateAllocations) {
1207 if (alloc->is_AllocateArray())
1208 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1209 else
1210 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1211 }
1212 #endif
1213
1214 return true;
1215 }
1216
1217 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1218 // EA should remove all uses of non-escaping boxing node.
1219 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1220 return false;
1221 }
1222
1223 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1224
1225 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1226
1227 const TypeTuple* r = boxing->tf()->range_sig();
1228 assert(r->cnt() > TypeFunc::Parms, "sanity");
1229 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1230 assert(t != NULL, "sanity");
1231
1232 CompileLog* log = C->log();
1233 if (log != NULL) {
1234 log->head("eliminate_boxing type='%d'",
1235 log->identify(t->instance_klass()));
1236 JVMState* p = boxing->jvms();
1237 while (p != NULL) {
1238 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1239 p = p->caller();
1240 }
1241 log->tail("eliminate_boxing");
1242 }
1243
1244 process_users_of_allocation(boxing);
1245
1246 #ifndef PRODUCT
1247 if (PrintEliminateAllocations) {
1391 }
1392 }
1393 #endif
1394 yank_alloc_node(alloc);
1395 return;
1396 }
1397 }
1398
1399 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1400 Node *slow_region = NULL;
1401 Node *toobig_false = ctrl;
1402
1403 // generate the initial test if necessary
1404 if (initial_slow_test != NULL ) {
1405 assert (expand_fast_path, "Only need test if there is a fast path");
1406 slow_region = new RegionNode(3);
1407
1408 // Now make the initial failure test. Usually a too-big test but
1409 // might be a TRUE for finalizers or a fancy class check for
1410 // newInstance0.
1411 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1412 transform_later(toobig_iff);
1413 // Plug the failing-too-big test into the slow-path region
1414 Node* toobig_true = new IfTrueNode(toobig_iff);
1415 transform_later(toobig_true);
1416 slow_region ->init_req( too_big_or_final_path, toobig_true );
1417 toobig_false = new IfFalseNode(toobig_iff);
1418 transform_later(toobig_false);
1419 } else {
1420 // No initial test, just fall into next case
1421 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1422 toobig_false = ctrl;
1423 debug_only(slow_region = NodeSentinel);
1424 }
1425
1426 // If we are here there are several possibilities
1427 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1428 // no_initial_check means a constant allocation.
1429 // - If check always evaluates to false -> expand_fast_path is false (see above)
1430 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1431 // if !allocation_has_use the fast path is empty
1432 // if !allocation_has_use && no_initial_check
1433 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1434 // removed by yank_alloc_node above.
1435
1436 Node *slow_mem = mem; // save the current memory state for slow path
1437 // generate the fast allocation code unless we know that the initial test will always go slow
1438 if (expand_fast_path) {
1439 // Fast path modifies only raw memory.
1440 if (mem->is_MergeMem()) {
1441 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1442 }
1443
1444 // allocate the Region and Phi nodes for the result
1445 result_region = new RegionNode(3);
1446 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1447 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1448
1449 // Grab regular I/O before optional prefetch may change it.
1450 // Slow-path does no I/O so just set it to the original I/O.
1451 result_phi_i_o->init_req(slow_result_path, i_o);
1452
1453 // Name successful fast-path variables
1454 Node* fast_oop_ctrl;
1455 Node* fast_oop_rawmem;
1456
1457 if (allocation_has_use) {
1458 Node* needgc_ctrl = NULL;
1459 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1460
1461 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1462 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1463 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1464 fast_oop_ctrl, fast_oop_rawmem,
1465 prefetch_lines);
1466
1467 if (initial_slow_test != NULL) {
1468 // This completes all paths into the slow merge point
1469 slow_region->init_req(need_gc_path, needgc_ctrl);
1470 transform_later(slow_region);
1471 } else {
1472 // No initial slow path needed!
1473 // Just fall from the need-GC path straight into the VM call.
1474 slow_region = needgc_ctrl;
1475 }
1476
1494 result_phi_i_o ->init_req(fast_result_path, i_o);
1495 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1496 } else {
1497 slow_region = ctrl;
1498 result_phi_i_o = i_o; // Rename it to use in the following code.
1499 }
1500
1501 // Generate slow-path call
1502 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1503 OptoRuntime::stub_name(slow_call_address),
1504 TypePtr::BOTTOM);
1505 call->init_req(TypeFunc::Control, slow_region);
1506 call->init_req(TypeFunc::I_O, top()); // does no i/o
1507 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1508 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1509 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1510
1511 call->init_req(TypeFunc::Parms+0, klass_node);
1512 if (length != NULL) {
1513 call->init_req(TypeFunc::Parms+1, length);
1514 } else {
1515 // Let the runtime know if this is a larval allocation
1516 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1517 }
1518
1519 // Copy debug information and adjust JVMState information, then replace
1520 // allocate node with the call
1521 call->copy_call_debug_info(&_igvn, alloc);
1522 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1523 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1524 // path dies).
1525 if (valid_length_test != NULL) {
1526 call->add_req(valid_length_test);
1527 }
1528 if (expand_fast_path) {
1529 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1530 } else {
1531 // Hook i_o projection to avoid its elimination during allocation
1532 // replacement (when only a slow call is generated).
1533 call->set_req(TypeFunc::I_O, result_phi_i_o);
1534 }
1535 _igvn.replace_node(alloc, call);
1536 transform_later(call);
1537
1538 // Identify the output projections from the allocate node and
1539 // adjust any references to them.
1540 // The control and io projections look like:
1541 //
1542 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1543 // Allocate Catch
1544 // ^---Proj(io) <-------+ ^---CatchProj(io)
1545 //
1546 // We are interested in the CatchProj nodes.
1547 //
1548 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1549
1550 // An allocate node has separate memory projections for the uses on
1551 // the control and i_o paths. Replace the control memory projection with
1552 // result_phi_rawmem (unless we are only generating a slow call when
1553 // both memory projections are combined)
1554 if (expand_fast_path && _callprojs->fallthrough_memproj != NULL) {
1555 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1556 }
1557 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1558 // catchall_memproj so we end up with a call that has only 1 memory projection.
1559 if (_callprojs->catchall_memproj != NULL) {
1560 if (_callprojs->fallthrough_memproj == NULL) {
1561 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1562 transform_later(_callprojs->fallthrough_memproj);
1563 }
1564 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1565 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1566 }
1567
1568 // An allocate node has separate i_o projections for the uses on the control
1569 // and i_o paths. Always replace the control i_o projection with result i_o
1570 // otherwise incoming i_o become dead when only a slow call is generated
1571 // (it is different from memory projections where both projections are
1572 // combined in such case).
1573 if (_callprojs->fallthrough_ioproj != NULL) {
1574 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1575 }
1576 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1577 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1578 if (_callprojs->catchall_ioproj != NULL) {
1579 if (_callprojs->fallthrough_ioproj == NULL) {
1580 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1581 transform_later(_callprojs->fallthrough_ioproj);
1582 }
1583 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1584 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1585 }
1586
1587 // if we generated only a slow call, we are done
1588 if (!expand_fast_path) {
1589 // Now we can unhook i_o.
1590 if (result_phi_i_o->outcnt() > 1) {
1591 call->set_req(TypeFunc::I_O, top());
1592 } else {
1593 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1594 // Case of new array with negative size known during compilation.
1595 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1596 // following code since call to runtime will throw exception.
1597 // As result there will be no users of i_o after the call.
1598 // Leave i_o attached to this call to avoid problems in preceding graph.
1599 }
1600 return;
1601 }
1602
1603 if (_callprojs->fallthrough_catchproj != NULL) {
1604 ctrl = _callprojs->fallthrough_catchproj->clone();
1605 transform_later(ctrl);
1606 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1607 } else {
1608 ctrl = top();
1609 }
1610 Node *slow_result;
1611 if (_callprojs->resproj[0] == NULL) {
1612 // no uses of the allocation result
1613 slow_result = top();
1614 } else {
1615 slow_result = _callprojs->resproj[0]->clone();
1616 transform_later(slow_result);
1617 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1618 }
1619
1620 // Plug slow-path into result merge point
1621 result_region->init_req( slow_result_path, ctrl);
1622 transform_later(result_region);
1623 if (allocation_has_use) {
1624 result_phi_rawoop->init_req(slow_result_path, slow_result);
1625 transform_later(result_phi_rawoop);
1626 }
1627 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1628 transform_later(result_phi_rawmem);
1629 transform_later(result_phi_i_o);
1630 // This completes all paths into the result merge point
1631 }
1632
1633 // Remove alloc node that has no uses.
1634 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1635 Node* ctrl = alloc->in(TypeFunc::Control);
1636 Node* mem = alloc->in(TypeFunc::Memory);
1637 Node* i_o = alloc->in(TypeFunc::I_O);
1638
1639 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1640 if (_callprojs->resproj[0] != NULL) {
1641 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1642 Node* use = _callprojs->resproj[0]->fast_out(i);
1643 use->isa_MemBar()->remove(&_igvn);
1644 --imax;
1645 --i; // back up iterator
1646 }
1647 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1648 _igvn.remove_dead_node(_callprojs->resproj[0]);
1649 }
1650 if (_callprojs->fallthrough_catchproj != NULL) {
1651 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1652 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1653 }
1654 if (_callprojs->catchall_catchproj != NULL) {
1655 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1656 _callprojs->catchall_catchproj->set_req(0, top());
1657 }
1658 if (_callprojs->fallthrough_proj != NULL) {
1659 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1660 _igvn.remove_dead_node(catchnode);
1661 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1662 }
1663 if (_callprojs->fallthrough_memproj != NULL) {
1664 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1665 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1666 }
1667 if (_callprojs->fallthrough_ioproj != NULL) {
1668 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1669 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1670 }
1671 if (_callprojs->catchall_memproj != NULL) {
1672 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1673 _callprojs->catchall_memproj->set_req(0, top());
1674 }
1675 if (_callprojs->catchall_ioproj != NULL) {
1676 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1677 _callprojs->catchall_ioproj->set_req(0, top());
1678 }
1679 #ifndef PRODUCT
1680 if (PrintEliminateAllocations) {
1681 if (alloc->is_AllocateArray()) {
1682 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1683 } else {
1684 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1685 }
1686 }
1687 #endif
1688 _igvn.remove_dead_node(alloc);
1689 }
1690
1691 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1692 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1693 // If initialization is performed by an array copy, any required
1694 // MemBarStoreStore was already added. If the object does not
1695 // escape no need for a MemBarStoreStore. If the object does not
1696 // escape in its initializer and memory barrier (MemBarStoreStore or
1697 // stronger) is already added at exit of initializer, also no need
1775 Node* thread = new ThreadLocalNode();
1776 transform_later(thread);
1777
1778 call->init_req(TypeFunc::Parms + 0, thread);
1779 call->init_req(TypeFunc::Parms + 1, oop);
1780 call->init_req(TypeFunc::Control, ctrl);
1781 call->init_req(TypeFunc::I_O , top()); // does no i/o
1782 call->init_req(TypeFunc::Memory , rawmem);
1783 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1784 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1785 transform_later(call);
1786 ctrl = new ProjNode(call, TypeFunc::Control);
1787 transform_later(ctrl);
1788 rawmem = new ProjNode(call, TypeFunc::Memory);
1789 transform_later(rawmem);
1790 }
1791 }
1792
1793 // Helper for PhaseMacroExpand::expand_allocate_common.
1794 // Initializes the newly-allocated storage.
1795 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1796 Node* control, Node* rawmem, Node* object,
1797 Node* klass_node, Node* length,
1798 Node* size_in_bytes) {
1799 InitializeNode* init = alloc->initialization();
1800 // Store the klass & mark bits
1801 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1802 if (!mark_node->is_Con()) {
1803 transform_later(mark_node);
1804 }
1805 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1806
1807 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1808 int header_size = alloc->minimum_header_size(); // conservatively small
1809
1810 // Array length
1811 if (length != NULL) { // Arrays need length field
1812 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1813 // conservatively small header size:
1814 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1815 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1816 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1817 if (is_reference_type(elem, true)) {
1818 elem = T_OBJECT;
1819 }
1820 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1821 }
1822 }
1823
1824 // Clear the object body, if necessary.
1825 if (init == NULL) {
1826 // The init has somehow disappeared; be cautious and clear everything.
1827 //
1828 // This can happen if a node is allocated but an uncommon trap occurs
1829 // immediately. In this case, the Initialize gets associated with the
1830 // trap, and may be placed in a different (outer) loop, if the Allocate
1831 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1832 // there can be two Allocates to one Initialize. The answer in all these
1833 // edge cases is safety first. It is always safe to clear immediately
1834 // within an Allocate, and then (maybe or maybe not) clear some more later.
1835 if (!(UseTLAB && ZeroTLAB)) {
1836 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1837 alloc->in(AllocateNode::DefaultValue),
1838 alloc->in(AllocateNode::RawDefaultValue),
1839 header_size, size_in_bytes,
1840 &_igvn);
1841 }
1842 } else {
1843 if (!init->is_complete()) {
1844 // Try to win by zeroing only what the init does not store.
1845 // We can also try to do some peephole optimizations,
1846 // such as combining some adjacent subword stores.
1847 rawmem = init->complete_stores(control, rawmem, object,
1848 header_size, size_in_bytes, &_igvn);
1849 }
1850 // We have no more use for this link, since the AllocateNode goes away:
1851 init->set_req(InitializeNode::RawAddress, top());
1852 // (If we keep the link, it just confuses the register allocator,
1853 // who thinks he sees a real use of the address by the membar.)
1854 }
1855
1856 return rawmem;
1857 }
1858
2188 } // EliminateNestedLocks
2189
2190 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2191 // Look for all locks of this object and mark them and
2192 // corresponding BoxLock nodes as eliminated.
2193 Node* obj = alock->obj_node();
2194 for (uint j = 0; j < obj->outcnt(); j++) {
2195 Node* o = obj->raw_out(j);
2196 if (o->is_AbstractLock() &&
2197 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2198 alock = o->as_AbstractLock();
2199 Node* box = alock->box_node();
2200 // Replace old box node with new eliminated box for all users
2201 // of the same object and mark related locks as eliminated.
2202 mark_eliminated_box(box, obj);
2203 }
2204 }
2205 }
2206 }
2207
2208 void PhaseMacroExpand::inline_type_guard(Node** ctrl, LockNode* lock) {
2209 Node* obj = lock->obj_node();
2210 const TypePtr* obj_type = _igvn.type(obj)->make_ptr();
2211 if (!obj_type->can_be_inline_type()) {
2212 return;
2213 }
2214 Node* mark = make_load(*ctrl, lock->memory(), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2215 Node* value_mask = _igvn.MakeConX(markWord::inline_type_pattern);
2216 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2217 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2218 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2219 Node* unc_ctrl = generate_slow_guard(ctrl, bol, NULL);
2220
2221 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2222 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2223 const TypePtr* no_memory_effects = NULL;
2224 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2225 no_memory_effects);
2226 unc->init_req(TypeFunc::Control, unc_ctrl);
2227 unc->init_req(TypeFunc::I_O, lock->i_o());
2228 unc->init_req(TypeFunc::Memory, lock->memory());
2229 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2230 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2231 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2232 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2233 unc->copy_call_debug_info(&_igvn, lock);
2234
2235 assert(unc->peek_monitor_box() == lock->box_node(), "wrong monitor");
2236 assert((obj_type->is_inlinetypeptr() && unc->peek_monitor_obj()->is_SafePointScalarObject()) ||
2237 (obj->is_InlineType() && obj->in(1) == unc->peek_monitor_obj()) ||
2238 (obj == unc->peek_monitor_obj()), "wrong monitor");
2239
2240 // pop monitor and push obj back on stack: we trap before the monitorenter
2241 unc->pop_monitor();
2242 unc->grow_stack(unc->jvms(), 1);
2243 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2244 _igvn.register_new_node_with_optimizer(unc);
2245
2246 unc_ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2247 Node* halt = _igvn.transform(new HaltNode(unc_ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on inline type"));
2248 _igvn.add_input_to(C->root(), halt);
2249 }
2250
2251 // we have determined that this lock/unlock can be eliminated, we simply
2252 // eliminate the node without expanding it.
2253 //
2254 // Note: The membar's associated with the lock/unlock are currently not
2255 // eliminated. This should be investigated as a future enhancement.
2256 //
2257 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2258
2259 if (!alock->is_eliminated()) {
2260 return false;
2261 }
2262 #ifdef ASSERT
2263 if (!alock->is_coarsened()) {
2264 // Check that new "eliminated" BoxLock node is created.
2265 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2266 assert(oldbox->is_eliminated(), "should be done already");
2267 }
2268 #endif
2269
2270 alock->log_lock_optimization(C, "eliminate_lock");
2271
2272 #ifndef PRODUCT
2273 if (PrintEliminateLocks) {
2274 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2275 }
2276 #endif
2277
2278 Node* mem = alock->in(TypeFunc::Memory);
2279 Node* ctrl = alock->in(TypeFunc::Control);
2280 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2281
2282 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2283 // There are 2 projections from the lock. The lock node will
2284 // be deleted when its last use is subsumed below.
2285 assert(alock->outcnt() == 2 &&
2286 _callprojs->fallthrough_proj != NULL &&
2287 _callprojs->fallthrough_memproj != NULL,
2288 "Unexpected projections from Lock/Unlock");
2289
2290 Node* fallthroughproj = _callprojs->fallthrough_proj;
2291 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2292
2293 // The memory projection from a lock/unlock is RawMem
2294 // The input to a Lock is merged memory, so extract its RawMem input
2295 // (unless the MergeMem has been optimized away.)
2296 if (alock->is_Lock()) {
2297 // Deoptimize and re-execute if object is an inline type
2298 inline_type_guard(&ctrl, alock->as_Lock());
2299
2300 // Search for MemBarAcquireLock node and delete it also.
2301 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2302 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2303 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2304 Node* memproj = membar->proj_out(TypeFunc::Memory);
2305 _igvn.replace_node(ctrlproj, fallthroughproj);
2306 _igvn.replace_node(memproj, memproj_fallthrough);
2307
2308 // Delete FastLock node also if this Lock node is unique user
2309 // (a loop peeling may clone a Lock node).
2310 Node* flock = alock->as_Lock()->fastlock_node();
2311 if (flock->outcnt() == 1) {
2312 assert(flock->unique_out() == alock, "sanity");
2313 _igvn.replace_node(flock, top());
2314 }
2315 }
2316
2317 // Search for MemBarReleaseLock node and delete it also.
2318 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2319 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2340 Node* mem = lock->in(TypeFunc::Memory);
2341 Node* obj = lock->obj_node();
2342 Node* box = lock->box_node();
2343 Node* flock = lock->fastlock_node();
2344
2345 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2346
2347 // Make the merge point
2348 Node *region;
2349 Node *mem_phi;
2350 Node *slow_path;
2351
2352 region = new RegionNode(3);
2353 // create a Phi for the memory state
2354 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2355
2356 // Optimize test; set region slot 2
2357 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2358 mem_phi->init_req(2, mem);
2359
2360 // Deoptimize and re-execute if object is an inline type
2361 inline_type_guard(&slow_path, lock);
2362
2363 // Make slow path call
2364 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2365 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2366 obj, box, NULL);
2367
2368 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2369
2370 // Slow path can only throw asynchronous exceptions, which are always
2371 // de-opted. So the compiler thinks the slow-call can never throw an
2372 // exception. If it DOES throw an exception we would need the debug
2373 // info removed first (since if it throws there is no monitor).
2374 assert(_callprojs->fallthrough_ioproj == NULL && _callprojs->catchall_ioproj == NULL &&
2375 _callprojs->catchall_memproj == NULL && _callprojs->catchall_catchproj == NULL, "Unexpected projection from Lock");
2376
2377 // Capture slow path
2378 // disconnect fall-through projection from call and create a new one
2379 // hook up users of fall-through projection to region
2380 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2381 transform_later(slow_ctrl);
2382 _igvn.hash_delete(_callprojs->fallthrough_proj);
2383 _callprojs->fallthrough_proj->disconnect_inputs(C);
2384 region->init_req(1, slow_ctrl);
2385 // region inputs are now complete
2386 transform_later(region);
2387 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2388
2389 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2390
2391 mem_phi->init_req(1, memproj);
2392
2393 transform_later(mem_phi);
2394
2395 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2396 }
2397
2398 //------------------------------expand_unlock_node----------------------
2399 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2400
2401 Node* ctrl = unlock->in(TypeFunc::Control);
2402 Node* mem = unlock->in(TypeFunc::Memory);
2403 Node* obj = unlock->obj_node();
2404 Node* box = unlock->box_node();
2405
2406 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2407
2408 // No need for a null check on unlock
2409
2410 // Make the merge point
2411 Node *region;
2412 Node *mem_phi;
2413
2414 region = new RegionNode(3);
2415 // create a Phi for the memory state
2416 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2417
2418 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2419 funlock = transform_later( funlock )->as_FastUnlock();
2420 // Optimize test; set region slot 2
2421 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2422 Node *thread = transform_later(new ThreadLocalNode());
2423
2424 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2425 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2426 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2427
2428 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2429 assert(_callprojs->fallthrough_ioproj == NULL && _callprojs->catchall_ioproj == NULL &&
2430 _callprojs->catchall_memproj == NULL && _callprojs->catchall_catchproj == NULL, "Unexpected projection from Lock");
2431
2432 // No exceptions for unlocking
2433 // Capture slow path
2434 // disconnect fall-through projection from call and create a new one
2435 // hook up users of fall-through projection to region
2436 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2437 transform_later(slow_ctrl);
2438 _igvn.hash_delete(_callprojs->fallthrough_proj);
2439 _callprojs->fallthrough_proj->disconnect_inputs(C);
2440 region->init_req(1, slow_ctrl);
2441 // region inputs are now complete
2442 transform_later(region);
2443 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2444
2445 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2446 mem_phi->init_req(1, memproj );
2447 mem_phi->init_req(2, mem);
2448 transform_later(mem_phi);
2449
2450 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2451 }
2452
2453 // An inline type might be returned from the call but we don't know its
2454 // type. Either we get a buffered inline type (and nothing needs to be done)
2455 // or one of the inlines being returned is the klass of the inline type
2456 // and we need to allocate an inline type instance of that type and
2457 // initialize it with other values being returned. In that case, we
2458 // first try a fast path allocation and initialize the value with the
2459 // inline klass's pack handler or we fall back to a runtime call.
2460 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2461 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2462 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2463 if (ret == NULL) {
2464 return;
2465 }
2466 const TypeFunc* tf = call->_tf;
2467 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2468 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2469 call->_tf = new_tf;
2470 // Make sure the change of type is applied before projections are processed by igvn
2471 _igvn.set_type(call, call->Value(&_igvn));
2472 _igvn.set_type(ret, ret->Value(&_igvn));
2473
2474 // Before any new projection is added:
2475 CallProjections* projs = call->extract_projections(true, true);
2476
2477 // Create temporary hook nodes that will be replaced below.
2478 // Add an input to prevent hook nodes from being dead.
2479 Node* ctl = new Node(call);
2480 Node* mem = new Node(ctl);
2481 Node* io = new Node(ctl);
2482 Node* ex_ctl = new Node(ctl);
2483 Node* ex_mem = new Node(ctl);
2484 Node* ex_io = new Node(ctl);
2485 Node* res = new Node(ctl);
2486
2487 // Allocate a new buffered inline type only if a new one is not returned
2488 Node* cast = transform_later(new CastP2XNode(ctl, res));
2489 Node* mask = MakeConX(0x1);
2490 Node* masked = transform_later(new AndXNode(cast, mask));
2491 Node* cmp = transform_later(new CmpXNode(masked, mask));
2492 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2493 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2494 transform_later(allocation_iff);
2495 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2496 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2497 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2498
2499 // Try to allocate a new buffered inline instance either from TLAB or eden space
2500 Node* needgc_ctrl = NULL; // needgc means slowcase, i.e. allocation failed
2501 CallLeafNoFPNode* handler_call;
2502 const bool alloc_in_place = UseTLAB;
2503 if (alloc_in_place) {
2504 Node* fast_oop_ctrl = NULL;
2505 Node* fast_oop_rawmem = NULL;
2506 Node* mask2 = MakeConX(-2);
2507 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2508 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2509 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2510 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2511 Node* size_in_bytes = ConvI2X(layout_val);
2512 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2513 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2514 fast_oop_ctrl, fast_oop_rawmem,
2515 AllocateInstancePrefetchLines);
2516 // Allocation succeed, initialize buffered inline instance header firstly,
2517 // and then initialize its fields with an inline class specific handler
2518 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2519 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2520 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2521 if (UseCompressedClassPointers) {
2522 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2523 }
2524 Node* fixed_block = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2525 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2526 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2527 NULL,
2528 "pack handler",
2529 TypeRawPtr::BOTTOM);
2530 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2531 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2532 handler_call->init_req(TypeFunc::I_O, top());
2533 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2534 handler_call->init_req(TypeFunc::ReturnAdr, top());
2535 handler_call->init_req(TypeFunc::Parms, pack_handler);
2536 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2537 } else {
2538 needgc_ctrl = allocation_ctl;
2539 }
2540
2541 // Allocation failed, fall back to a runtime call
2542 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2543 StubRoutines::store_inline_type_fields_to_buf(),
2544 "store_inline_type_fields",
2545 TypePtr::BOTTOM);
2546 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2547 slow_call->init_req(TypeFunc::Memory, mem);
2548 slow_call->init_req(TypeFunc::I_O, io);
2549 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2550 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2551 slow_call->init_req(TypeFunc::Parms, res);
2552
2553 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2554 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2555 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2556 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2557 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2558 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2559 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2560
2561 Node* ex_r = new RegionNode(3);
2562 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2563 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2564 ex_r->init_req(1, slow_excp);
2565 ex_mem_phi->init_req(1, slow_mem);
2566 ex_io_phi->init_req(1, slow_io);
2567 ex_r->init_req(2, ex_ctl);
2568 ex_mem_phi->init_req(2, ex_mem);
2569 ex_io_phi->init_req(2, ex_io);
2570 transform_later(ex_r);
2571 transform_later(ex_mem_phi);
2572 transform_later(ex_io_phi);
2573
2574 // We don't know how many values are returned. This assumes the
2575 // worst case, that all available registers are used.
2576 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2577 if (domain->field_at(i) == Type::HALF) {
2578 slow_call->init_req(i, top());
2579 if (alloc_in_place) {
2580 handler_call->init_req(i+1, top());
2581 }
2582 continue;
2583 }
2584 Node* proj = transform_later(new ProjNode(call, i));
2585 slow_call->init_req(i, proj);
2586 if (alloc_in_place) {
2587 handler_call->init_req(i+1, proj);
2588 }
2589 }
2590 // We can safepoint at that new call
2591 slow_call->copy_call_debug_info(&_igvn, call);
2592 transform_later(slow_call);
2593 if (alloc_in_place) {
2594 transform_later(handler_call);
2595 }
2596
2597 Node* fast_ctl = NULL;
2598 Node* fast_res = NULL;
2599 MergeMemNode* fast_mem = NULL;
2600 if (alloc_in_place) {
2601 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2602 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2603 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2604 fast_mem = MergeMemNode::make(mem);
2605 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2606 transform_later(fast_mem);
2607 }
2608
2609 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2610 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2611 Node* io_phi = new PhiNode(r, Type::ABIO);
2612 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2613 r->init_req(1, no_allocation_ctl);
2614 mem_phi->init_req(1, mem);
2615 io_phi->init_req(1, io);
2616 res_phi->init_req(1, no_allocation_res);
2617 r->init_req(2, slow_norm);
2618 mem_phi->init_req(2, slow_mem);
2619 io_phi->init_req(2, slow_io);
2620 res_phi->init_req(2, slow_res);
2621 if (alloc_in_place) {
2622 r->init_req(3, fast_ctl);
2623 mem_phi->init_req(3, fast_mem);
2624 io_phi->init_req(3, io);
2625 res_phi->init_req(3, fast_res);
2626 }
2627 transform_later(r);
2628 transform_later(mem_phi);
2629 transform_later(io_phi);
2630 transform_later(res_phi);
2631
2632 // Do not let stores that initialize this buffer be reordered with a subsequent
2633 // store that would make this buffer accessible by other threads.
2634 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2635 transform_later(mb);
2636 mb->init_req(TypeFunc::Memory, mem_phi);
2637 mb->init_req(TypeFunc::Control, r);
2638 r = new ProjNode(mb, TypeFunc::Control);
2639 transform_later(r);
2640 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2641 transform_later(mem_phi);
2642
2643 assert(projs->nb_resproj == 1, "unexpected number of results");
2644 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2645 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2646 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2647 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2648 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2649 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2650 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2651 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2652 Node* cn = projs->fallthrough_catchproj->in(0);
2653 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2654
2655 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2656 _igvn.replace_node(mem, projs->fallthrough_memproj);
2657 _igvn.replace_node(io, projs->fallthrough_ioproj);
2658 _igvn.replace_node(res, projs->resproj[0]);
2659 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2660 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2661 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2662 }
2663
2664 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2665 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2666 Node* bol = check->unique_out();
2667 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2668 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2669 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2670
2671 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2672 Node* iff = bol->last_out(i);
2673 assert(iff->is_If(), "where's the if?");
2674
2675 if (iff->in(0)->is_top()) {
2676 _igvn.replace_input_of(iff, 1, C->top());
2677 continue;
2678 }
2679
2680 Node* iftrue = iff->as_If()->proj_out(1);
2681 Node* iffalse = iff->as_If()->proj_out(0);
2682 Node* ctrl = iff->in(0);
2683
2684 Node* subklass = NULL;
2685 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2686 subklass = obj_or_subklass;
2687 } else {
2688 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2689 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2690 }
2691
2692 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2693
2694 _igvn.replace_input_of(iff, 0, C->top());
2695 _igvn.replace_node(iftrue, not_subtype_ctrl);
2696 _igvn.replace_node(iffalse, ctrl);
2697 }
2698 _igvn.replace_node(check, C->top());
2699 }
2700
2701 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2702 //
2703 // long mark = array1.mark | array2.mark | ...;
2704 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2705 // if (locked_bit == 0) {
2706 // // One array is locked, load prototype header from the klass
2707 // mark = array1.klass.proto | array2.klass.proto | ...
2708 // }
2709 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2710 // ...
2711 // }
2712 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2713 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != NULL;
2714 if (UseArrayMarkWordCheck && array_inputs) {
2715 Node* mark = MakeConX(0);
2716 Node* locked_bit = MakeConX(markWord::unlocked_value);
2717 Node* mem = check->in(FlatArrayCheckNode::Memory);
2718 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2719 Node* ary = check->in(i);
2720 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2721 assert(t != NULL, "Mixing array and klass inputs");
2722 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2723 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2724 Node* mark_load = _igvn.transform(LoadNode::make(_igvn, NULL, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2725 mark = _igvn.transform(new OrXNode(mark, mark_load));
2726 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2727 }
2728 assert(!mark->is_Con(), "Should have been optimized out");
2729 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2730 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2731
2732 // BoolNode might be shared, replace each if user
2733 Node* old_bol = check->unique_out();
2734 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2735 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2736 IfNode* old_iff = old_bol->last_out(i)->as_If();
2737 Node* ctrl = old_iff->in(0);
2738 RegionNode* region = new RegionNode(3);
2739 Node* mark_phi = new PhiNode(region, TypeX_X);
2740
2741 // Check if array is unlocked
2742 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2743
2744 // Unlocked: Use bits from mark word
2745 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2746 mark_phi->init_req(1, mark);
2747
2748 // Locked: Load prototype header from klass
2749 ctrl = _igvn.transform(new IfFalseNode(iff));
2750 Node* proto = MakeConX(0);
2751 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2752 Node* ary = check->in(i);
2753 // Make loads control dependent to make sure they are only executed if array is locked
2754 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2755 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2756 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2757 Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2758 proto = _igvn.transform(new OrXNode(proto, proto_load));
2759 }
2760 region->init_req(2, ctrl);
2761 mark_phi->init_req(2, proto);
2762
2763 // Check if flat array bits are set
2764 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2765 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2766 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2767 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2768
2769 ctrl = _igvn.transform(region);
2770 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2771 _igvn.replace_node(old_iff, iff);
2772 }
2773 _igvn.replace_node(check, C->top());
2774 } else {
2775 // Fall back to layout helper check
2776 Node* lhs = intcon(0);
2777 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2778 Node* array_or_klass = check->in(i);
2779 Node* klass = NULL;
2780 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2781 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2782 if (t->isa_oopptr() != NULL) {
2783 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2784 klass = transform_later(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2785 } else {
2786 assert(t->isa_aryklassptr(), "Unexpected input type");
2787 klass = array_or_klass;
2788 }
2789 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2790 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, NULL, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
2791 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2792 }
2793 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2794 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2795 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2796 Node* old_bol = check->unique_out();
2797 _igvn.replace_node(old_bol, bol);
2798 _igvn.replace_node(check, C->top());
2799 }
2800 }
2801
2802 //---------------------------eliminate_macro_nodes----------------------
2803 // Eliminate scalar replaced allocations and associated locks.
2804 void PhaseMacroExpand::eliminate_macro_nodes() {
2805 if (C->macro_count() == 0)
2806 return;
2807 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2808
2809 // Before elimination may re-mark (change to Nested or NonEscObj)
2810 // all associated (same box and obj) lock and unlock nodes.
2811 int cnt = C->macro_count();
2812 for (int i=0; i < cnt; i++) {
2813 Node *n = C->macro_node(i);
2814 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2815 mark_eliminated_locking_nodes(n->as_AbstractLock());
2816 }
2817 }
2818 // Re-marking may break consistency of Coarsened locks.
2819 if (!C->coarsened_locks_consistent()) {
2820 return; // recompile without Coarsened locks if broken
2821 }
2842 }
2843 // Next, attempt to eliminate allocations
2844 _has_locks = false;
2845 progress = true;
2846 while (progress) {
2847 progress = false;
2848 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2849 Node* n = C->macro_node(i - 1);
2850 bool success = false;
2851 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2852 switch (n->class_id()) {
2853 case Node::Class_Allocate:
2854 case Node::Class_AllocateArray:
2855 success = eliminate_allocate_node(n->as_Allocate());
2856 #ifndef PRODUCT
2857 if (success && PrintOptoStatistics) {
2858 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2859 }
2860 #endif
2861 break;
2862 case Node::Class_CallStaticJava: {
2863 CallStaticJavaNode* call = n->as_CallStaticJava();
2864 if (!call->method()->is_method_handle_intrinsic()) {
2865 success = eliminate_boxing_node(n->as_CallStaticJava());
2866 }
2867 break;
2868 }
2869 case Node::Class_Lock:
2870 case Node::Class_Unlock:
2871 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2872 _has_locks = true;
2873 break;
2874 case Node::Class_ArrayCopy:
2875 break;
2876 case Node::Class_OuterStripMinedLoop:
2877 break;
2878 case Node::Class_SubTypeCheck:
2879 break;
2880 case Node::Class_Opaque1:
2881 break;
2882 case Node::Class_FlatArrayCheck:
2883 break;
2884 default:
2885 assert(n->Opcode() == Op_LoopLimit ||
2886 n->Opcode() == Op_Opaque3 ||
2887 n->Opcode() == Op_Opaque4 ||
2888 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2889 "unknown node type in macro list");
2890 }
2891 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2892 progress = progress || success;
2893 }
2894 }
2895 #ifndef PRODUCT
2896 if (PrintOptoStatistics) {
2897 int membar_after = count_MemBar(C);
2898 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2899 }
2900 #endif
2901 }
2902
2903 //------------------------------expand_macro_nodes----------------------
2904 // Returns true if a failure occurred.
2905 bool PhaseMacroExpand::expand_macro_nodes() {
2906 // Last attempt to eliminate macro nodes.
2907 eliminate_macro_nodes();
2908 if (C->failing()) return true;
2909
2910 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2911 bool progress = true;
2912 while (progress) {
2913 progress = false;
2914 for (int i = C->macro_count(); i > 0; i--) {
2915 Node* n = C->macro_node(i-1);
2916 bool success = false;
2917 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2918 if (n->Opcode() == Op_LoopLimit) {
2919 // Remove it from macro list and put on IGVN worklist to optimize.
2920 C->remove_macro_node(n);
2921 _igvn._worklist.push(n);
2922 success = true;
2923 } else if (n->Opcode() == Op_CallStaticJava) {
2924 CallStaticJavaNode* call = n->as_CallStaticJava();
2925 if (!call->method()->is_method_handle_intrinsic()) {
2926 // Remove it from macro list and put on IGVN worklist to optimize.
2927 C->remove_macro_node(n);
2928 _igvn._worklist.push(n);
2929 success = true;
2930 }
2931 } else if (n->is_Opaque1()) {
2932 _igvn.replace_node(n, n->in(1));
2933 success = true;
2934 #if INCLUDE_RTM_OPT
2935 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2936 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2937 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2938 Node* cmp = n->unique_out();
2939 #ifdef ASSERT
2940 // Validate graph.
2941 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2942 BoolNode* bol = cmp->unique_out()->as_Bool();
2943 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2944 (bol->_test._test == BoolTest::ne), "");
2945 IfNode* ifn = bol->unique_out()->as_If();
2946 assert((ifn->outcnt() == 2) &&
2947 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2948 #endif
2949 Node* repl = n->in(1);
2950 if (!_has_locks) {
3013 // Worst case is a macro node gets expanded into about 200 nodes.
3014 // Allow 50% more for optimization.
3015 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3016 return true;
3017 }
3018
3019 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3020 switch (n->class_id()) {
3021 case Node::Class_Lock:
3022 expand_lock_node(n->as_Lock());
3023 break;
3024 case Node::Class_Unlock:
3025 expand_unlock_node(n->as_Unlock());
3026 break;
3027 case Node::Class_ArrayCopy:
3028 expand_arraycopy_node(n->as_ArrayCopy());
3029 break;
3030 case Node::Class_SubTypeCheck:
3031 expand_subtypecheck_node(n->as_SubTypeCheck());
3032 break;
3033 case Node::Class_CallStaticJava:
3034 expand_mh_intrinsic_return(n->as_CallStaticJava());
3035 C->remove_macro_node(n);
3036 break;
3037 case Node::Class_FlatArrayCheck:
3038 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3039 break;
3040 default:
3041 assert(false, "unknown node type in macro list");
3042 }
3043 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3044 if (C->failing()) return true;
3045
3046 // Clean up the graph so we're less likely to hit the maximum node
3047 // limit
3048 _igvn.set_delay_transform(false);
3049 _igvn.optimize();
3050 if (C->failing()) return true;
3051 _igvn.set_delay_transform(true);
3052 }
3053
3054 // All nodes except Allocate nodes are expanded now. There could be
3055 // new optimization opportunities (such as folding newly created
3056 // load from a just allocated object). Run IGVN.
3057
3058 // expand "macro" nodes
3059 // nodes are removed from the macro list as they are processed
|