1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vectornode.hpp"
71 #include "runtime/arguments.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stubRoutines.hpp"
75 #include "runtime/timer.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/copy.hpp"
78 #include "utilities/macros.hpp"
79 #if INCLUDE_G1GC
80 #include "gc/g1/g1ThreadLocalData.hpp"
81 #endif // INCLUDE_G1GC
82 #if INCLUDE_ZGC
83 #include "gc/z/c2/zBarrierSetC2.hpp"
84 #endif
85
86
87 // -------------------- Compile::mach_constant_base_node -----------------------
88 // Constant table base node singleton.
89 MachConstantBaseNode* Compile::mach_constant_base_node() {
90 if (_mach_constant_base_node == NULL) {
91 _mach_constant_base_node = new MachConstantBaseNode();
92 _mach_constant_base_node->add_req(C->root());
93 }
94 return _mach_constant_base_node;
95 }
96
97
98 /// Support for intrinsics.
99
100 // Return the index at which m must be inserted (or already exists).
101 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
102 class IntrinsicDescPair {
103 private:
104 ciMethod* _m;
105 bool _is_virtual;
106 public:
107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
109 ciMethod* m= elt->method();
110 ciMethod* key_m = key->_m;
111 if (key_m < m) return -1;
112 else if (key_m > m) return 1;
113 else {
114 bool is_virtual = elt->is_virtual();
115 bool key_virtual = key->_is_virtual;
116 if (key_virtual < is_virtual) return -1;
117 else if (key_virtual > is_virtual) return 1;
118 else return 0;
119 }
120 }
121 };
122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
123 #ifdef ASSERT
124 for (int i = 1; i < _intrinsics->length(); i++) {
125 CallGenerator* cg1 = _intrinsics->at(i-1);
126 CallGenerator* cg2 = _intrinsics->at(i);
127 assert(cg1->method() != cg2->method()
128 ? cg1->method() < cg2->method()
129 : cg1->is_virtual() < cg2->is_virtual(),
130 "compiler intrinsics list must stay sorted");
131 }
132 #endif
133 IntrinsicDescPair pair(m, is_virtual);
134 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
135 }
136
137 void Compile::register_intrinsic(CallGenerator* cg) {
138 if (_intrinsics == NULL) {
139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
140 }
141 int len = _intrinsics->length();
142 bool found = false;
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
144 assert(!found, "registering twice");
145 _intrinsics->insert_before(index, cg);
146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
147 }
148
149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
150 assert(m->is_loaded(), "don't try this on unloaded methods");
151 if (_intrinsics != NULL) {
152 bool found = false;
153 int index = intrinsic_insertion_index(m, is_virtual, found);
154 if (found) {
155 return _intrinsics->at(index);
156 }
157 }
158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
159 if (m->intrinsic_id() != vmIntrinsics::_none &&
160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
162 if (cg != NULL) {
163 // Save it for next time:
164 register_intrinsic(cg);
165 return cg;
166 } else {
167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
168 }
169 }
170 return NULL;
171 }
172
173 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
174 // in library_call.cpp.
175
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
182
183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
185 int oflags = _intrinsic_hist_flags[id];
186 assert(flags != 0, "what happened?");
187 if (is_virtual) {
188 flags |= _intrinsic_virtual;
189 }
190 bool changed = (flags != oflags);
191 if ((flags & _intrinsic_worked) != 0) {
192 juint count = (_intrinsic_hist_count[id] += 1);
193 if (count == 1) {
194 changed = true; // first time
195 }
196 // increment the overall count also:
197 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
198 }
199 if (changed) {
200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
201 // Something changed about the intrinsic's virtuality.
202 if ((flags & _intrinsic_virtual) != 0) {
203 // This is the first use of this intrinsic as a virtual call.
204 if (oflags != 0) {
205 // We already saw it as a non-virtual, so note both cases.
206 flags |= _intrinsic_both;
207 }
208 } else if ((oflags & _intrinsic_both) == 0) {
209 // This is the first use of this intrinsic as a non-virtual
210 flags |= _intrinsic_both;
211 }
212 }
213 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
214 }
215 // update the overall flags also:
216 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
217 return changed;
218 }
219
220 static char* format_flags(int flags, char* buf) {
221 buf[0] = 0;
222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
227 if (buf[0] == 0) strcat(buf, ",");
228 assert(buf[0] == ',', "must be");
229 return &buf[1];
230 }
231
232 void Compile::print_intrinsic_statistics() {
233 char flagsbuf[100];
234 ttyLocker ttyl;
235 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
236 tty->print_cr("Compiler intrinsic usage:");
237 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
238 if (total == 0) total = 1; // avoid div0 in case of no successes
239 #define PRINT_STAT_LINE(name, c, f) \
240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
241 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
242 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
243 int flags = _intrinsic_hist_flags[id];
244 juint count = _intrinsic_hist_count[id];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
250 if (xtty != NULL) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != NULL) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseCCP::print_statistics();
258 PhaseRegAlloc::print_statistics();
259 Scheduling::print_statistics();
260 PhasePeephole::print_statistics();
261 PhaseIdealLoop::print_statistics();
262 if (xtty != NULL) xtty->tail("statistics");
263 }
264 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
265 // put this under its own <statistics> element.
266 print_intrinsic_statistics();
267 }
268 }
269 #endif //PRODUCT
270
271 // Support for bundling info
272 Bundle* Compile::node_bundling(const Node *n) {
273 assert(valid_bundle_info(n), "oob");
274 return &_node_bundling_base[n->_idx];
275 }
276
277 bool Compile::valid_bundle_info(const Node *n) {
278 return (_node_bundling_limit > n->_idx);
279 }
280
281
282 void Compile::gvn_replace_by(Node* n, Node* nn) {
283 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
284 Node* use = n->last_out(i);
285 bool is_in_table = initial_gvn()->hash_delete(use);
286 uint uses_found = 0;
287 for (uint j = 0; j < use->len(); j++) {
288 if (use->in(j) == n) {
289 if (j < use->req())
290 use->set_req(j, nn);
291 else
292 use->set_prec(j, nn);
293 uses_found++;
294 }
295 }
296 if (is_in_table) {
297 // reinsert into table
298 initial_gvn()->hash_find_insert(use);
299 }
300 record_for_igvn(use);
301 i -= uses_found; // we deleted 1 or more copies of this edge
302 }
303 }
304
305
306 static inline bool not_a_node(const Node* n) {
307 if (n == NULL) return true;
308 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
309 if (*(address*)n == badAddress) return true; // kill by Node::destruct
310 return false;
311 }
312
313 // Identify all nodes that are reachable from below, useful.
314 // Use breadth-first pass that records state in a Unique_Node_List,
315 // recursive traversal is slower.
316 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
317 int estimated_worklist_size = live_nodes();
318 useful.map( estimated_worklist_size, NULL ); // preallocate space
319
320 // Initialize worklist
321 if (root() != NULL) { useful.push(root()); }
322 // If 'top' is cached, declare it useful to preserve cached node
323 if( cached_top_node() ) { useful.push(cached_top_node()); }
324
325 // Push all useful nodes onto the list, breadthfirst
326 for( uint next = 0; next < useful.size(); ++next ) {
327 assert( next < unique(), "Unique useful nodes < total nodes");
328 Node *n = useful.at(next);
329 uint max = n->len();
330 for( uint i = 0; i < max; ++i ) {
331 Node *m = n->in(i);
332 if (not_a_node(m)) continue;
333 useful.push(m);
334 }
335 }
336 }
337
338 // Update dead_node_list with any missing dead nodes using useful
339 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
340 void Compile::update_dead_node_list(Unique_Node_List &useful) {
341 uint max_idx = unique();
342 VectorSet& useful_node_set = useful.member_set();
343
344 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
345 // If node with index node_idx is not in useful set,
346 // mark it as dead in dead node list.
347 if (! useful_node_set.test(node_idx) ) {
348 record_dead_node(node_idx);
349 }
350 }
351 }
352
353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
354 int shift = 0;
355 for (int i = 0; i < inlines->length(); i++) {
356 CallGenerator* cg = inlines->at(i);
357 CallNode* call = cg->call_node();
358 if (shift > 0) {
359 inlines->at_put(i-shift, cg);
360 }
361 if (!useful.member(call)) {
362 shift++;
363 }
364 }
365 inlines->trunc_to(inlines->length()-shift);
366 }
367
368 // Disconnect all useless nodes by disconnecting those at the boundary.
369 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
370 uint next = 0;
371 while (next < useful.size()) {
372 Node *n = useful.at(next++);
373 if (n->is_SafePoint()) {
374 // We're done with a parsing phase. Replaced nodes are not valid
375 // beyond that point.
376 n->as_SafePoint()->delete_replaced_nodes();
377 }
378 // Use raw traversal of out edges since this code removes out edges
379 int max = n->outcnt();
380 for (int j = 0; j < max; ++j) {
381 Node* child = n->raw_out(j);
382 if (! useful.member(child)) {
383 assert(!child->is_top() || child != top(),
384 "If top is cached in Compile object it is in useful list");
385 // Only need to remove this out-edge to the useless node
386 n->raw_del_out(j);
387 --j;
388 --max;
389 }
390 }
391 if (n->outcnt() == 1 && n->has_special_unique_user()) {
392 record_for_igvn(n->unique_out());
393 }
394 }
395 // Remove useless macro and predicate opaq nodes
396 for (int i = C->macro_count()-1; i >= 0; i--) {
397 Node* n = C->macro_node(i);
398 if (!useful.member(n)) {
399 remove_macro_node(n);
400 }
401 }
402 // Remove useless CastII nodes with range check dependency
403 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
404 Node* cast = range_check_cast_node(i);
405 if (!useful.member(cast)) {
406 remove_range_check_cast(cast);
407 }
408 }
409 // Remove useless expensive nodes
410 for (int i = C->expensive_count()-1; i >= 0; i--) {
411 Node* n = C->expensive_node(i);
412 if (!useful.member(n)) {
413 remove_expensive_node(n);
414 }
415 }
416 // Remove useless Opaque4 nodes
417 for (int i = opaque4_count() - 1; i >= 0; i--) {
418 Node* opaq = opaque4_node(i);
419 if (!useful.member(opaq)) {
420 remove_opaque4_node(opaq);
421 }
422 }
423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
424 bs->eliminate_useless_gc_barriers(useful, this);
425 // clean up the late inline lists
426 remove_useless_late_inlines(&_string_late_inlines, useful);
427 remove_useless_late_inlines(&_boxing_late_inlines, useful);
428 remove_useless_late_inlines(&_late_inlines, useful);
429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
430 }
431
432 //------------------------------frame_size_in_words-----------------------------
433 // frame_slots in units of words
434 int Compile::frame_size_in_words() const {
435 // shift is 0 in LP32 and 1 in LP64
436 const int shift = (LogBytesPerWord - LogBytesPerInt);
437 int words = _frame_slots >> shift;
438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
439 return words;
440 }
441
442 // To bang the stack of this compiled method we use the stack size
443 // that the interpreter would need in case of a deoptimization. This
444 // removes the need to bang the stack in the deoptimization blob which
445 // in turn simplifies stack overflow handling.
446 int Compile::bang_size_in_bytes() const {
447 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
448 }
449
450 // ============================================================================
451 //------------------------------CompileWrapper---------------------------------
452 class CompileWrapper : public StackObj {
453 Compile *const _compile;
454 public:
455 CompileWrapper(Compile* compile);
456
457 ~CompileWrapper();
458 };
459
460 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
461 // the Compile* pointer is stored in the current ciEnv:
462 ciEnv* env = compile->env();
463 assert(env == ciEnv::current(), "must already be a ciEnv active");
464 assert(env->compiler_data() == NULL, "compile already active?");
465 env->set_compiler_data(compile);
466 assert(compile == Compile::current(), "sanity");
467
468 compile->set_type_dict(NULL);
469 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
470 compile->clone_map().set_clone_idx(0);
471 compile->set_type_hwm(NULL);
472 compile->set_type_last_size(0);
473 compile->set_last_tf(NULL, NULL);
474 compile->set_indexSet_arena(NULL);
475 compile->set_indexSet_free_block_list(NULL);
476 compile->init_type_arena();
477 Type::Initialize(compile);
478 _compile->set_scratch_buffer_blob(NULL);
479 _compile->begin_method();
480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
481 }
482 CompileWrapper::~CompileWrapper() {
483 _compile->end_method();
484 if (_compile->scratch_buffer_blob() != NULL)
485 BufferBlob::free(_compile->scratch_buffer_blob());
486 _compile->env()->set_compiler_data(NULL);
487 }
488
489
490 //----------------------------print_compile_messages---------------------------
491 void Compile::print_compile_messages() {
492 #ifndef PRODUCT
493 // Check if recompiling
494 if (_subsume_loads == false && PrintOpto) {
495 // Recompiling without allowing machine instructions to subsume loads
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without subsuming loads **");
498 tty->print_cr("*********************************************************");
499 }
500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
501 // Recompiling without escape analysis
502 tty->print_cr("*********************************************************");
503 tty->print_cr("** Bailout: Recompile without escape analysis **");
504 tty->print_cr("*********************************************************");
505 }
506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
507 // Recompiling without boxing elimination
508 tty->print_cr("*********************************************************");
509 tty->print_cr("** Bailout: Recompile without boxing elimination **");
510 tty->print_cr("*********************************************************");
511 }
512 if (C->directive()->BreakAtCompileOption) {
513 // Open the debugger when compiling this method.
514 tty->print("### Breaking when compiling: ");
515 method()->print_short_name();
516 tty->cr();
517 BREAKPOINT;
518 }
519
520 if( PrintOpto ) {
521 if (is_osr_compilation()) {
522 tty->print("[OSR]%3d", _compile_id);
523 } else {
524 tty->print("%3d", _compile_id);
525 }
526 }
527 #endif
528 }
529
530
531 //-----------------------init_scratch_buffer_blob------------------------------
532 // Construct a temporary BufferBlob and cache it for this compile.
533 void Compile::init_scratch_buffer_blob(int const_size) {
534 // If there is already a scratch buffer blob allocated and the
535 // constant section is big enough, use it. Otherwise free the
536 // current and allocate a new one.
537 BufferBlob* blob = scratch_buffer_blob();
538 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
539 // Use the current blob.
540 } else {
541 if (blob != NULL) {
542 BufferBlob::free(blob);
543 }
544
545 ResourceMark rm;
546 _scratch_const_size = const_size;
547 int size = C2Compiler::initial_code_buffer_size(const_size);
548 blob = BufferBlob::create("Compile::scratch_buffer", size);
549 // Record the buffer blob for next time.
550 set_scratch_buffer_blob(blob);
551 // Have we run out of code space?
552 if (scratch_buffer_blob() == NULL) {
553 // Let CompilerBroker disable further compilations.
554 record_failure("Not enough space for scratch buffer in CodeCache");
555 return;
556 }
557 }
558
559 // Initialize the relocation buffers
560 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
561 set_scratch_locs_memory(locs_buf);
562 }
563
564
565 //-----------------------scratch_emit_size-------------------------------------
566 // Helper function that computes size by emitting code
567 uint Compile::scratch_emit_size(const Node* n) {
568 // Start scratch_emit_size section.
569 set_in_scratch_emit_size(true);
570
571 // Emit into a trash buffer and count bytes emitted.
572 // This is a pretty expensive way to compute a size,
573 // but it works well enough if seldom used.
574 // All common fixed-size instructions are given a size
575 // method by the AD file.
576 // Note that the scratch buffer blob and locs memory are
577 // allocated at the beginning of the compile task, and
578 // may be shared by several calls to scratch_emit_size.
579 // The allocation of the scratch buffer blob is particularly
580 // expensive, since it has to grab the code cache lock.
581 BufferBlob* blob = this->scratch_buffer_blob();
582 assert(blob != NULL, "Initialize BufferBlob at start");
583 assert(blob->size() > MAX_inst_size, "sanity");
584 relocInfo* locs_buf = scratch_locs_memory();
585 address blob_begin = blob->content_begin();
586 address blob_end = (address)locs_buf;
587 assert(blob->contains(blob_end), "sanity");
588 CodeBuffer buf(blob_begin, blob_end - blob_begin);
589 buf.initialize_consts_size(_scratch_const_size);
590 buf.initialize_stubs_size(MAX_stubs_size);
591 assert(locs_buf != NULL, "sanity");
592 int lsize = MAX_locs_size / 3;
593 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
594 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
595 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
596 // Mark as scratch buffer.
597 buf.consts()->set_scratch_emit();
598 buf.insts()->set_scratch_emit();
599 buf.stubs()->set_scratch_emit();
600
601 // Do the emission.
602
603 Label fakeL; // Fake label for branch instructions.
604 Label* saveL = NULL;
605 uint save_bnum = 0;
606 bool is_branch = n->is_MachBranch();
607 if (is_branch) {
608 MacroAssembler masm(&buf);
609 masm.bind(fakeL);
610 n->as_MachBranch()->save_label(&saveL, &save_bnum);
611 n->as_MachBranch()->label_set(&fakeL, 0);
612 }
613 n->emit(buf, this->regalloc());
614
615 // Emitting into the scratch buffer should not fail
616 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
617
618 if (is_branch) // Restore label.
619 n->as_MachBranch()->label_set(saveL, save_bnum);
620
621 // End scratch_emit_size section.
622 set_in_scratch_emit_size(false);
623
624 return buf.insts_size();
625 }
626
627
628 // ============================================================================
629 //------------------------------Compile standard-------------------------------
630 debug_only( int Compile::_debug_idx = 100000; )
631
632 // Compile a method. entry_bci is -1 for normal compilations and indicates
633 // the continuation bci for on stack replacement.
634
635
636 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
637 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
638 : Phase(Compiler),
639 _compile_id(ci_env->compile_id()),
640 _save_argument_registers(false),
641 _subsume_loads(subsume_loads),
642 _do_escape_analysis(do_escape_analysis),
643 _eliminate_boxing(eliminate_boxing),
644 _method(target),
645 _entry_bci(osr_bci),
646 _stub_function(NULL),
647 _stub_name(NULL),
648 _stub_entry_point(NULL),
649 _max_node_limit(MaxNodeLimit),
650 _orig_pc_slot(0),
651 _orig_pc_slot_offset_in_bytes(0),
652 _inlining_progress(false),
653 _inlining_incrementally(false),
654 _has_reserved_stack_access(target->has_reserved_stack_access()),
655 #ifndef PRODUCT
656 _trace_opto_output(directive->TraceOptoOutputOption),
657 #endif
658 _has_method_handle_invokes(false),
659 _comp_arena(mtCompiler),
660 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
661 _env(ci_env),
662 _directive(directive),
663 _log(ci_env->log()),
664 _failure_reason(NULL),
665 _congraph(NULL),
666 #ifndef PRODUCT
667 _printer(IdealGraphPrinter::printer()),
668 #endif
669 _dead_node_list(comp_arena()),
670 _dead_node_count(0),
671 _node_arena(mtCompiler),
672 _old_arena(mtCompiler),
673 _mach_constant_base_node(NULL),
674 _Compile_types(mtCompiler),
675 _initial_gvn(NULL),
676 _for_igvn(NULL),
677 _warm_calls(NULL),
678 _late_inlines(comp_arena(), 2, 0, NULL),
679 _string_late_inlines(comp_arena(), 2, 0, NULL),
680 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
681 _late_inlines_pos(0),
682 _number_of_mh_late_inlines(0),
683 _print_inlining_stream(NULL),
684 _print_inlining_list(NULL),
685 _print_inlining_idx(0),
686 _print_inlining_output(NULL),
687 _replay_inline_data(NULL),
688 _java_calls(0),
689 _inner_loops(0),
690 _interpreter_frame_size(0),
691 _node_bundling_limit(0),
692 _node_bundling_base(NULL),
693 _code_buffer("Compile::Fill_buffer"),
694 _scratch_const_size(-1),
695 _in_scratch_emit_size(false)
696 #ifndef PRODUCT
697 , _in_dump_cnt(0)
698 #endif
699 {
700 C = this;
701 #ifndef PRODUCT
702 if (_printer != NULL) {
703 _printer->set_compile(this);
704 }
705 #endif
706 CompileWrapper cw(this);
707
708 if (CITimeVerbose) {
709 tty->print(" ");
710 target->holder()->name()->print();
711 tty->print(".");
712 target->print_short_name();
713 tty->print(" ");
714 }
715 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
716 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
717
718 #ifndef PRODUCT
719 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
720 if (!print_opto_assembly) {
721 bool print_assembly = directive->PrintAssemblyOption;
722 if (print_assembly && !Disassembler::can_decode()) {
723 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
724 print_opto_assembly = true;
725 }
726 }
727 set_print_assembly(print_opto_assembly);
728 set_parsed_irreducible_loop(false);
729
730 if (directive->ReplayInlineOption) {
731 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
732 }
733 #endif
734 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
735 set_print_intrinsics(directive->PrintIntrinsicsOption);
736 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
737
738 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
739 // Make sure the method being compiled gets its own MDO,
740 // so we can at least track the decompile_count().
741 // Need MDO to record RTM code generation state.
742 method()->ensure_method_data();
743 }
744
745 Init(::AliasLevel);
746
747
748 print_compile_messages();
749
750 _ilt = InlineTree::build_inline_tree_root();
751
752 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
753 assert(num_alias_types() >= AliasIdxRaw, "");
754
755 #define MINIMUM_NODE_HASH 1023
756 // Node list that Iterative GVN will start with
757 Unique_Node_List for_igvn(comp_arena());
758 set_for_igvn(&for_igvn);
759
760 // GVN that will be run immediately on new nodes
761 uint estimated_size = method()->code_size()*4+64;
762 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
763 PhaseGVN gvn(node_arena(), estimated_size);
764 set_initial_gvn(&gvn);
765
766 print_inlining_init();
767 { // Scope for timing the parser
768 TracePhase tp("parse", &timers[_t_parser]);
769
770 // Put top into the hash table ASAP.
771 initial_gvn()->transform_no_reclaim(top());
772
773 // Set up tf(), start(), and find a CallGenerator.
774 CallGenerator* cg = NULL;
775 if (is_osr_compilation()) {
776 const TypeTuple *domain = StartOSRNode::osr_domain();
777 const TypeTuple *range = TypeTuple::make_range(method()->signature());
778 init_tf(TypeFunc::make(domain, range));
779 StartNode* s = new StartOSRNode(root(), domain);
780 initial_gvn()->set_type_bottom(s);
781 init_start(s);
782 cg = CallGenerator::for_osr(method(), entry_bci());
783 } else {
784 // Normal case.
785 init_tf(TypeFunc::make(method()));
786 StartNode* s = new StartNode(root(), tf()->domain());
787 initial_gvn()->set_type_bottom(s);
788 init_start(s);
789 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
790 // With java.lang.ref.reference.get() we must go through the
791 // intrinsic - even when get() is the root
792 // method of the compile - so that, if necessary, the value in
793 // the referent field of the reference object gets recorded by
794 // the pre-barrier code.
795 cg = find_intrinsic(method(), false);
796 }
797 if (cg == NULL) {
798 float past_uses = method()->interpreter_invocation_count();
799 float expected_uses = past_uses;
800 cg = CallGenerator::for_inline(method(), expected_uses);
801 }
802 }
803 if (failing()) return;
804 if (cg == NULL) {
805 record_method_not_compilable("cannot parse method");
806 return;
807 }
808 JVMState* jvms = build_start_state(start(), tf());
809 if ((jvms = cg->generate(jvms)) == NULL) {
810 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
811 record_method_not_compilable("method parse failed");
812 }
813 return;
814 }
815 GraphKit kit(jvms);
816
817 if (!kit.stopped()) {
818 // Accept return values, and transfer control we know not where.
819 // This is done by a special, unique ReturnNode bound to root.
820 return_values(kit.jvms());
821 }
822
823 if (kit.has_exceptions()) {
824 // Any exceptions that escape from this call must be rethrown
825 // to whatever caller is dynamically above us on the stack.
826 // This is done by a special, unique RethrowNode bound to root.
827 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
828 }
829
830 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
831
832 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
833 inline_string_calls(true);
834 }
835
836 if (failing()) return;
837
838 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
839
840 // Remove clutter produced by parsing.
841 if (!failing()) {
842 ResourceMark rm;
843 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
844 }
845 }
846
847 // Note: Large methods are capped off in do_one_bytecode().
848 if (failing()) return;
849
850 // After parsing, node notes are no longer automagic.
851 // They must be propagated by register_new_node_with_optimizer(),
852 // clone(), or the like.
853 set_default_node_notes(NULL);
854
855 for (;;) {
856 int successes = Inline_Warm();
857 if (failing()) return;
858 if (successes == 0) break;
859 }
860
861 // Drain the list.
862 Finish_Warm();
863 #ifndef PRODUCT
864 if (_printer && _printer->should_print(1)) {
865 _printer->print_inlining();
866 }
867 #endif
868
869 if (failing()) return;
870 NOT_PRODUCT( verify_graph_edges(); )
871
872 // Now optimize
873 Optimize();
874 if (failing()) return;
875 NOT_PRODUCT( verify_graph_edges(); )
876
877 #ifndef PRODUCT
878 if (PrintIdeal) {
879 ttyLocker ttyl; // keep the following output all in one block
880 // This output goes directly to the tty, not the compiler log.
881 // To enable tools to match it up with the compilation activity,
882 // be sure to tag this tty output with the compile ID.
883 if (xtty != NULL) {
884 xtty->head("ideal compile_id='%d'%s", compile_id(),
885 is_osr_compilation() ? " compile_kind='osr'" :
886 "");
887 }
888 root()->dump(9999);
889 if (xtty != NULL) {
890 xtty->tail("ideal");
891 }
892 }
893 #endif
894
895 NOT_PRODUCT( verify_barriers(); )
896
897 // Dump compilation data to replay it.
898 if (directive->DumpReplayOption) {
899 env()->dump_replay_data(_compile_id);
900 }
901 if (directive->DumpInlineOption && (ilt() != NULL)) {
902 env()->dump_inline_data(_compile_id);
903 }
904
905 // Now that we know the size of all the monitors we can add a fixed slot
906 // for the original deopt pc.
907
908 _orig_pc_slot = fixed_slots();
909 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
910 set_fixed_slots(next_slot);
911
912 // Compute when to use implicit null checks. Used by matching trap based
913 // nodes and NullCheck optimization.
914 set_allowed_deopt_reasons();
915
916 // Now generate code
917 Code_Gen();
918 if (failing()) return;
919
920 // Check if we want to skip execution of all compiled code.
921 {
922 #ifndef PRODUCT
923 if (OptoNoExecute) {
924 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
925 return;
926 }
927 #endif
928 TracePhase tp("install_code", &timers[_t_registerMethod]);
929
930 if (is_osr_compilation()) {
931 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
932 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
933 } else {
934 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
935 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
936 }
937
938 env()->register_method(_method, _entry_bci,
939 &_code_offsets,
940 _orig_pc_slot_offset_in_bytes,
941 code_buffer(),
942 frame_size_in_words(), _oop_map_set,
943 &_handler_table, &_inc_table,
944 compiler,
945 has_unsafe_access(),
946 SharedRuntime::is_wide_vector(max_vector_size()),
947 rtm_state()
948 );
949
950 if (log() != NULL) // Print code cache state into compiler log
951 log()->code_cache_state();
952 }
953 }
954
955 //------------------------------Compile----------------------------------------
956 // Compile a runtime stub
957 Compile::Compile( ciEnv* ci_env,
958 TypeFunc_generator generator,
959 address stub_function,
960 const char *stub_name,
961 int is_fancy_jump,
962 bool pass_tls,
963 bool save_arg_registers,
964 bool return_pc,
965 DirectiveSet* directive)
966 : Phase(Compiler),
967 _compile_id(0),
968 _save_argument_registers(save_arg_registers),
969 _subsume_loads(true),
970 _do_escape_analysis(false),
971 _eliminate_boxing(false),
972 _method(NULL),
973 _entry_bci(InvocationEntryBci),
974 _stub_function(stub_function),
975 _stub_name(stub_name),
976 _stub_entry_point(NULL),
977 _max_node_limit(MaxNodeLimit),
978 _orig_pc_slot(0),
979 _orig_pc_slot_offset_in_bytes(0),
980 _inlining_progress(false),
981 _inlining_incrementally(false),
982 _has_reserved_stack_access(false),
983 #ifndef PRODUCT
984 _trace_opto_output(directive->TraceOptoOutputOption),
985 #endif
986 _has_method_handle_invokes(false),
987 _comp_arena(mtCompiler),
988 _env(ci_env),
989 _directive(directive),
990 _log(ci_env->log()),
991 _failure_reason(NULL),
992 _congraph(NULL),
993 #ifndef PRODUCT
994 _printer(NULL),
995 #endif
996 _dead_node_list(comp_arena()),
997 _dead_node_count(0),
998 _node_arena(mtCompiler),
999 _old_arena(mtCompiler),
1000 _mach_constant_base_node(NULL),
1001 _Compile_types(mtCompiler),
1002 _initial_gvn(NULL),
1003 _for_igvn(NULL),
1004 _warm_calls(NULL),
1005 _number_of_mh_late_inlines(0),
1006 _print_inlining_stream(NULL),
1007 _print_inlining_list(NULL),
1008 _print_inlining_idx(0),
1009 _print_inlining_output(NULL),
1010 _replay_inline_data(NULL),
1011 _java_calls(0),
1012 _inner_loops(0),
1013 _interpreter_frame_size(0),
1014 _node_bundling_limit(0),
1015 _node_bundling_base(NULL),
1016 _code_buffer("Compile::Fill_buffer"),
1017 #ifndef PRODUCT
1018 _in_dump_cnt(0),
1019 #endif
1020 _allowed_reasons(0) {
1021 C = this;
1022
1023 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1024 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1025
1026 #ifndef PRODUCT
1027 set_print_assembly(PrintFrameConverterAssembly);
1028 set_parsed_irreducible_loop(false);
1029 #endif
1030 set_has_irreducible_loop(false); // no loops
1031
1032 CompileWrapper cw(this);
1033 Init(/*AliasLevel=*/ 0);
1034 init_tf((*generator)());
1035
1036 {
1037 // The following is a dummy for the sake of GraphKit::gen_stub
1038 Unique_Node_List for_igvn(comp_arena());
1039 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1040 PhaseGVN gvn(Thread::current()->resource_area(),255);
1041 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1042 gvn.transform_no_reclaim(top());
1043
1044 GraphKit kit;
1045 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1046 }
1047
1048 NOT_PRODUCT( verify_graph_edges(); )
1049 Code_Gen();
1050 if (failing()) return;
1051
1052
1053 // Entry point will be accessed using compile->stub_entry_point();
1054 if (code_buffer() == NULL) {
1055 Matcher::soft_match_failure();
1056 } else {
1057 if (PrintAssembly && (WizardMode || Verbose))
1058 tty->print_cr("### Stub::%s", stub_name);
1059
1060 if (!failing()) {
1061 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1062
1063 // Make the NMethod
1064 // For now we mark the frame as never safe for profile stackwalking
1065 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1066 code_buffer(),
1067 CodeOffsets::frame_never_safe,
1068 // _code_offsets.value(CodeOffsets::Frame_Complete),
1069 frame_size_in_words(),
1070 _oop_map_set,
1071 save_arg_registers);
1072 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1073
1074 _stub_entry_point = rs->entry_point();
1075 }
1076 }
1077 }
1078
1079 //------------------------------Init-------------------------------------------
1080 // Prepare for a single compilation
1081 void Compile::Init(int aliaslevel) {
1082 _unique = 0;
1083 _regalloc = NULL;
1084
1085 _tf = NULL; // filled in later
1086 _top = NULL; // cached later
1087 _matcher = NULL; // filled in later
1088 _cfg = NULL; // filled in later
1089
1090 set_24_bit_selection_and_mode(Use24BitFP, false);
1091
1092 _node_note_array = NULL;
1093 _default_node_notes = NULL;
1094 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1095
1096 _immutable_memory = NULL; // filled in at first inquiry
1097
1098 // Globally visible Nodes
1099 // First set TOP to NULL to give safe behavior during creation of RootNode
1100 set_cached_top_node(NULL);
1101 set_root(new RootNode());
1102 // Now that you have a Root to point to, create the real TOP
1103 set_cached_top_node( new ConNode(Type::TOP) );
1104 set_recent_alloc(NULL, NULL);
1105
1106 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1107 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1108 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1109 env()->set_dependencies(new Dependencies(env()));
1110
1111 _fixed_slots = 0;
1112 set_has_split_ifs(false);
1113 set_has_loops(has_method() && method()->has_loops()); // first approximation
1114 set_has_stringbuilder(false);
1115 set_has_boxed_value(false);
1116 _trap_can_recompile = false; // no traps emitted yet
1117 _major_progress = true; // start out assuming good things will happen
1118 set_has_unsafe_access(false);
1119 set_max_vector_size(0);
1120 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1121 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1122 set_decompile_count(0);
1123
1124 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1125 set_num_loop_opts(LoopOptsCount);
1126 set_do_inlining(Inline);
1127 set_max_inline_size(MaxInlineSize);
1128 set_freq_inline_size(FreqInlineSize);
1129 set_do_scheduling(OptoScheduling);
1130 set_do_count_invocations(false);
1131 set_do_method_data_update(false);
1132
1133 set_do_vector_loop(false);
1134
1135 if (AllowVectorizeOnDemand) {
1136 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1137 set_do_vector_loop(true);
1138 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1139 } else if (has_method() && method()->name() != 0 &&
1140 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1141 set_do_vector_loop(true);
1142 }
1143 }
1144 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1145 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1146
1147 set_age_code(has_method() && method()->profile_aging());
1148 set_rtm_state(NoRTM); // No RTM lock eliding by default
1149 _max_node_limit = _directive->MaxNodeLimitOption;
1150
1151 #if INCLUDE_RTM_OPT
1152 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1153 int rtm_state = method()->method_data()->rtm_state();
1154 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1155 // Don't generate RTM lock eliding code.
1156 set_rtm_state(NoRTM);
1157 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1158 // Generate RTM lock eliding code without abort ratio calculation code.
1159 set_rtm_state(UseRTM);
1160 } else if (UseRTMDeopt) {
1161 // Generate RTM lock eliding code and include abort ratio calculation
1162 // code if UseRTMDeopt is on.
1163 set_rtm_state(ProfileRTM);
1164 }
1165 }
1166 #endif
1167 if (debug_info()->recording_non_safepoints()) {
1168 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1169 (comp_arena(), 8, 0, NULL));
1170 set_default_node_notes(Node_Notes::make(this));
1171 }
1172
1173 // // -- Initialize types before each compile --
1174 // // Update cached type information
1175 // if( _method && _method->constants() )
1176 // Type::update_loaded_types(_method, _method->constants());
1177
1178 // Init alias_type map.
1179 if (!_do_escape_analysis && aliaslevel == 3)
1180 aliaslevel = 2; // No unique types without escape analysis
1181 _AliasLevel = aliaslevel;
1182 const int grow_ats = 16;
1183 _max_alias_types = grow_ats;
1184 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1185 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1186 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1187 {
1188 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1189 }
1190 // Initialize the first few types.
1191 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1192 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1193 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1194 _num_alias_types = AliasIdxRaw+1;
1195 // Zero out the alias type cache.
1196 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1197 // A NULL adr_type hits in the cache right away. Preload the right answer.
1198 probe_alias_cache(NULL)->_index = AliasIdxTop;
1199
1200 _intrinsics = NULL;
1201 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1202 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1203 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1204 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1205 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1206 register_library_intrinsics();
1207 }
1208
1209 //---------------------------init_start----------------------------------------
1210 // Install the StartNode on this compile object.
1211 void Compile::init_start(StartNode* s) {
1212 if (failing())
1213 return; // already failing
1214 assert(s == start(), "");
1215 }
1216
1217 /**
1218 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1219 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1220 * the ideal graph.
1221 */
1222 StartNode* Compile::start() const {
1223 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1224 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1225 Node* start = root()->fast_out(i);
1226 if (start->is_Start()) {
1227 return start->as_Start();
1228 }
1229 }
1230 fatal("Did not find Start node!");
1231 return NULL;
1232 }
1233
1234 //-------------------------------immutable_memory-------------------------------------
1235 // Access immutable memory
1236 Node* Compile::immutable_memory() {
1237 if (_immutable_memory != NULL) {
1238 return _immutable_memory;
1239 }
1240 StartNode* s = start();
1241 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1242 Node *p = s->fast_out(i);
1243 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1244 _immutable_memory = p;
1245 return _immutable_memory;
1246 }
1247 }
1248 ShouldNotReachHere();
1249 return NULL;
1250 }
1251
1252 //----------------------set_cached_top_node------------------------------------
1253 // Install the cached top node, and make sure Node::is_top works correctly.
1254 void Compile::set_cached_top_node(Node* tn) {
1255 if (tn != NULL) verify_top(tn);
1256 Node* old_top = _top;
1257 _top = tn;
1258 // Calling Node::setup_is_top allows the nodes the chance to adjust
1259 // their _out arrays.
1260 if (_top != NULL) _top->setup_is_top();
1261 if (old_top != NULL) old_top->setup_is_top();
1262 assert(_top == NULL || top()->is_top(), "");
1263 }
1264
1265 #ifdef ASSERT
1266 uint Compile::count_live_nodes_by_graph_walk() {
1267 Unique_Node_List useful(comp_arena());
1268 // Get useful node list by walking the graph.
1269 identify_useful_nodes(useful);
1270 return useful.size();
1271 }
1272
1273 void Compile::print_missing_nodes() {
1274
1275 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1276 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1277 return;
1278 }
1279
1280 // This is an expensive function. It is executed only when the user
1281 // specifies VerifyIdealNodeCount option or otherwise knows the
1282 // additional work that needs to be done to identify reachable nodes
1283 // by walking the flow graph and find the missing ones using
1284 // _dead_node_list.
1285
1286 Unique_Node_List useful(comp_arena());
1287 // Get useful node list by walking the graph.
1288 identify_useful_nodes(useful);
1289
1290 uint l_nodes = C->live_nodes();
1291 uint l_nodes_by_walk = useful.size();
1292
1293 if (l_nodes != l_nodes_by_walk) {
1294 if (_log != NULL) {
1295 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1296 _log->stamp();
1297 _log->end_head();
1298 }
1299 VectorSet& useful_member_set = useful.member_set();
1300 int last_idx = l_nodes_by_walk;
1301 for (int i = 0; i < last_idx; i++) {
1302 if (useful_member_set.test(i)) {
1303 if (_dead_node_list.test(i)) {
1304 if (_log != NULL) {
1305 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1306 }
1307 if (PrintIdealNodeCount) {
1308 // Print the log message to tty
1309 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1310 useful.at(i)->dump();
1311 }
1312 }
1313 }
1314 else if (! _dead_node_list.test(i)) {
1315 if (_log != NULL) {
1316 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1317 }
1318 if (PrintIdealNodeCount) {
1319 // Print the log message to tty
1320 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1321 }
1322 }
1323 }
1324 if (_log != NULL) {
1325 _log->tail("mismatched_nodes");
1326 }
1327 }
1328 }
1329 void Compile::record_modified_node(Node* n) {
1330 if (_modified_nodes != NULL && !_inlining_incrementally &&
1331 n->outcnt() != 0 && !n->is_Con()) {
1332 _modified_nodes->push(n);
1333 }
1334 }
1335
1336 void Compile::remove_modified_node(Node* n) {
1337 if (_modified_nodes != NULL) {
1338 _modified_nodes->remove(n);
1339 }
1340 }
1341 #endif
1342
1343 #ifndef PRODUCT
1344 void Compile::verify_top(Node* tn) const {
1345 if (tn != NULL) {
1346 assert(tn->is_Con(), "top node must be a constant");
1347 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1348 assert(tn->in(0) != NULL, "must have live top node");
1349 }
1350 }
1351 #endif
1352
1353
1354 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1355
1356 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1357 guarantee(arr != NULL, "");
1358 int num_blocks = arr->length();
1359 if (grow_by < num_blocks) grow_by = num_blocks;
1360 int num_notes = grow_by * _node_notes_block_size;
1361 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1362 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1363 while (num_notes > 0) {
1364 arr->append(notes);
1365 notes += _node_notes_block_size;
1366 num_notes -= _node_notes_block_size;
1367 }
1368 assert(num_notes == 0, "exact multiple, please");
1369 }
1370
1371 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1372 if (source == NULL || dest == NULL) return false;
1373
1374 if (dest->is_Con())
1375 return false; // Do not push debug info onto constants.
1376
1377 #ifdef ASSERT
1378 // Leave a bread crumb trail pointing to the original node:
1379 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1380 dest->set_debug_orig(source);
1381 }
1382 #endif
1383
1384 if (node_note_array() == NULL)
1385 return false; // Not collecting any notes now.
1386
1387 // This is a copy onto a pre-existing node, which may already have notes.
1388 // If both nodes have notes, do not overwrite any pre-existing notes.
1389 Node_Notes* source_notes = node_notes_at(source->_idx);
1390 if (source_notes == NULL || source_notes->is_clear()) return false;
1391 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1392 if (dest_notes == NULL || dest_notes->is_clear()) {
1393 return set_node_notes_at(dest->_idx, source_notes);
1394 }
1395
1396 Node_Notes merged_notes = (*source_notes);
1397 // The order of operations here ensures that dest notes will win...
1398 merged_notes.update_from(dest_notes);
1399 return set_node_notes_at(dest->_idx, &merged_notes);
1400 }
1401
1402
1403 //--------------------------allow_range_check_smearing-------------------------
1404 // Gating condition for coalescing similar range checks.
1405 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1406 // single covering check that is at least as strong as any of them.
1407 // If the optimization succeeds, the simplified (strengthened) range check
1408 // will always succeed. If it fails, we will deopt, and then give up
1409 // on the optimization.
1410 bool Compile::allow_range_check_smearing() const {
1411 // If this method has already thrown a range-check,
1412 // assume it was because we already tried range smearing
1413 // and it failed.
1414 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1415 return !already_trapped;
1416 }
1417
1418
1419 //------------------------------flatten_alias_type-----------------------------
1420 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1421 int offset = tj->offset();
1422 TypePtr::PTR ptr = tj->ptr();
1423
1424 // Known instance (scalarizable allocation) alias only with itself.
1425 bool is_known_inst = tj->isa_oopptr() != NULL &&
1426 tj->is_oopptr()->is_known_instance();
1427
1428 // Process weird unsafe references.
1429 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1430 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1431 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1432 tj = TypeOopPtr::BOTTOM;
1433 ptr = tj->ptr();
1434 offset = tj->offset();
1435 }
1436
1437 // Array pointers need some flattening
1438 const TypeAryPtr *ta = tj->isa_aryptr();
1439 if (ta && ta->is_stable()) {
1440 // Erase stability property for alias analysis.
1441 tj = ta = ta->cast_to_stable(false);
1442 }
1443 if( ta && is_known_inst ) {
1444 if ( offset != Type::OffsetBot &&
1445 offset > arrayOopDesc::length_offset_in_bytes() ) {
1446 offset = Type::OffsetBot; // Flatten constant access into array body only
1447 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1448 }
1449 } else if( ta && _AliasLevel >= 2 ) {
1450 // For arrays indexed by constant indices, we flatten the alias
1451 // space to include all of the array body. Only the header, klass
1452 // and array length can be accessed un-aliased.
1453 if( offset != Type::OffsetBot ) {
1454 if( ta->const_oop() ) { // MethodData* or Method*
1455 offset = Type::OffsetBot; // Flatten constant access into array body
1456 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1457 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1458 // range is OK as-is.
1459 tj = ta = TypeAryPtr::RANGE;
1460 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1461 tj = TypeInstPtr::KLASS; // all klass loads look alike
1462 ta = TypeAryPtr::RANGE; // generic ignored junk
1463 ptr = TypePtr::BotPTR;
1464 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1465 tj = TypeInstPtr::MARK;
1466 ta = TypeAryPtr::RANGE; // generic ignored junk
1467 ptr = TypePtr::BotPTR;
1468 } else { // Random constant offset into array body
1469 offset = Type::OffsetBot; // Flatten constant access into array body
1470 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1471 }
1472 }
1473 // Arrays of fixed size alias with arrays of unknown size.
1474 if (ta->size() != TypeInt::POS) {
1475 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1476 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1477 }
1478 // Arrays of known objects become arrays of unknown objects.
1479 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1480 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1481 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1482 }
1483 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1484 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1485 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1486 }
1487 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1488 // cannot be distinguished by bytecode alone.
1489 if (ta->elem() == TypeInt::BOOL) {
1490 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1491 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1492 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1493 }
1494 // During the 2nd round of IterGVN, NotNull castings are removed.
1495 // Make sure the Bottom and NotNull variants alias the same.
1496 // Also, make sure exact and non-exact variants alias the same.
1497 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1498 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1499 }
1500 }
1501
1502 // Oop pointers need some flattening
1503 const TypeInstPtr *to = tj->isa_instptr();
1504 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1505 ciInstanceKlass *k = to->klass()->as_instance_klass();
1506 if( ptr == TypePtr::Constant ) {
1507 if (to->klass() != ciEnv::current()->Class_klass() ||
1508 offset < k->size_helper() * wordSize) {
1509 // No constant oop pointers (such as Strings); they alias with
1510 // unknown strings.
1511 assert(!is_known_inst, "not scalarizable allocation");
1512 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1513 }
1514 } else if( is_known_inst ) {
1515 tj = to; // Keep NotNull and klass_is_exact for instance type
1516 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1517 // During the 2nd round of IterGVN, NotNull castings are removed.
1518 // Make sure the Bottom and NotNull variants alias the same.
1519 // Also, make sure exact and non-exact variants alias the same.
1520 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1521 }
1522 if (to->speculative() != NULL) {
1523 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1524 }
1525 // Canonicalize the holder of this field
1526 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1527 // First handle header references such as a LoadKlassNode, even if the
1528 // object's klass is unloaded at compile time (4965979).
1529 if (!is_known_inst) { // Do it only for non-instance types
1530 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1531 }
1532 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1533 // Static fields are in the space above the normal instance
1534 // fields in the java.lang.Class instance.
1535 if (to->klass() != ciEnv::current()->Class_klass()) {
1536 to = NULL;
1537 tj = TypeOopPtr::BOTTOM;
1538 offset = tj->offset();
1539 }
1540 } else {
1541 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1542 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1543 if( is_known_inst ) {
1544 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1545 } else {
1546 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1547 }
1548 }
1549 }
1550 }
1551
1552 // Klass pointers to object array klasses need some flattening
1553 const TypeKlassPtr *tk = tj->isa_klassptr();
1554 if( tk ) {
1555 // If we are referencing a field within a Klass, we need
1556 // to assume the worst case of an Object. Both exact and
1557 // inexact types must flatten to the same alias class so
1558 // use NotNull as the PTR.
1559 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1560
1561 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1562 TypeKlassPtr::OBJECT->klass(),
1563 offset);
1564 }
1565
1566 ciKlass* klass = tk->klass();
1567 if( klass->is_obj_array_klass() ) {
1568 ciKlass* k = TypeAryPtr::OOPS->klass();
1569 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1570 k = TypeInstPtr::BOTTOM->klass();
1571 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1572 }
1573
1574 // Check for precise loads from the primary supertype array and force them
1575 // to the supertype cache alias index. Check for generic array loads from
1576 // the primary supertype array and also force them to the supertype cache
1577 // alias index. Since the same load can reach both, we need to merge
1578 // these 2 disparate memories into the same alias class. Since the
1579 // primary supertype array is read-only, there's no chance of confusion
1580 // where we bypass an array load and an array store.
1581 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1582 if (offset == Type::OffsetBot ||
1583 (offset >= primary_supers_offset &&
1584 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1585 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1586 offset = in_bytes(Klass::secondary_super_cache_offset());
1587 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1588 }
1589 }
1590
1591 // Flatten all Raw pointers together.
1592 if (tj->base() == Type::RawPtr)
1593 tj = TypeRawPtr::BOTTOM;
1594
1595 if (tj->base() == Type::AnyPtr)
1596 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1597
1598 // Flatten all to bottom for now
1599 switch( _AliasLevel ) {
1600 case 0:
1601 tj = TypePtr::BOTTOM;
1602 break;
1603 case 1: // Flatten to: oop, static, field or array
1604 switch (tj->base()) {
1605 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1606 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1607 case Type::AryPtr: // do not distinguish arrays at all
1608 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1609 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1610 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1611 default: ShouldNotReachHere();
1612 }
1613 break;
1614 case 2: // No collapsing at level 2; keep all splits
1615 case 3: // No collapsing at level 3; keep all splits
1616 break;
1617 default:
1618 Unimplemented();
1619 }
1620
1621 offset = tj->offset();
1622 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1623
1624 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1625 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1626 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1627 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1628 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1629 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1630 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1631 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1632 assert( tj->ptr() != TypePtr::TopPTR &&
1633 tj->ptr() != TypePtr::AnyNull &&
1634 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1635 // assert( tj->ptr() != TypePtr::Constant ||
1636 // tj->base() == Type::RawPtr ||
1637 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1638
1639 return tj;
1640 }
1641
1642 void Compile::AliasType::Init(int i, const TypePtr* at) {
1643 _index = i;
1644 _adr_type = at;
1645 _field = NULL;
1646 _element = NULL;
1647 _is_rewritable = true; // default
1648 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1649 if (atoop != NULL && atoop->is_known_instance()) {
1650 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1651 _general_index = Compile::current()->get_alias_index(gt);
1652 } else {
1653 _general_index = 0;
1654 }
1655 }
1656
1657 BasicType Compile::AliasType::basic_type() const {
1658 if (element() != NULL) {
1659 const Type* element = adr_type()->is_aryptr()->elem();
1660 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1661 } if (field() != NULL) {
1662 return field()->layout_type();
1663 } else {
1664 return T_ILLEGAL; // unknown
1665 }
1666 }
1667
1668 //---------------------------------print_on------------------------------------
1669 #ifndef PRODUCT
1670 void Compile::AliasType::print_on(outputStream* st) {
1671 if (index() < 10)
1672 st->print("@ <%d> ", index());
1673 else st->print("@ <%d>", index());
1674 st->print(is_rewritable() ? " " : " RO");
1675 int offset = adr_type()->offset();
1676 if (offset == Type::OffsetBot)
1677 st->print(" +any");
1678 else st->print(" +%-3d", offset);
1679 st->print(" in ");
1680 adr_type()->dump_on(st);
1681 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1682 if (field() != NULL && tjp) {
1683 if (tjp->klass() != field()->holder() ||
1684 tjp->offset() != field()->offset_in_bytes()) {
1685 st->print(" != ");
1686 field()->print();
1687 st->print(" ***");
1688 }
1689 }
1690 }
1691
1692 void print_alias_types() {
1693 Compile* C = Compile::current();
1694 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1695 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1696 C->alias_type(idx)->print_on(tty);
1697 tty->cr();
1698 }
1699 }
1700 #endif
1701
1702
1703 //----------------------------probe_alias_cache--------------------------------
1704 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1705 intptr_t key = (intptr_t) adr_type;
1706 key ^= key >> logAliasCacheSize;
1707 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1708 }
1709
1710
1711 //-----------------------------grow_alias_types--------------------------------
1712 void Compile::grow_alias_types() {
1713 const int old_ats = _max_alias_types; // how many before?
1714 const int new_ats = old_ats; // how many more?
1715 const int grow_ats = old_ats+new_ats; // how many now?
1716 _max_alias_types = grow_ats;
1717 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1718 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1719 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1720 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1721 }
1722
1723
1724 //--------------------------------find_alias_type------------------------------
1725 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1726 if (_AliasLevel == 0)
1727 return alias_type(AliasIdxBot);
1728
1729 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1730 if (ace->_adr_type == adr_type) {
1731 return alias_type(ace->_index);
1732 }
1733
1734 // Handle special cases.
1735 if (adr_type == NULL) return alias_type(AliasIdxTop);
1736 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1737
1738 // Do it the slow way.
1739 const TypePtr* flat = flatten_alias_type(adr_type);
1740
1741 #ifdef ASSERT
1742 {
1743 ResourceMark rm;
1744 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1745 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1746 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1747 Type::str(adr_type));
1748 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1749 const TypeOopPtr* foop = flat->is_oopptr();
1750 // Scalarizable allocations have exact klass always.
1751 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1752 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1753 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1754 Type::str(foop), Type::str(xoop));
1755 }
1756 }
1757 #endif
1758
1759 int idx = AliasIdxTop;
1760 for (int i = 0; i < num_alias_types(); i++) {
1761 if (alias_type(i)->adr_type() == flat) {
1762 idx = i;
1763 break;
1764 }
1765 }
1766
1767 if (idx == AliasIdxTop) {
1768 if (no_create) return NULL;
1769 // Grow the array if necessary.
1770 if (_num_alias_types == _max_alias_types) grow_alias_types();
1771 // Add a new alias type.
1772 idx = _num_alias_types++;
1773 _alias_types[idx]->Init(idx, flat);
1774 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1775 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1776 if (flat->isa_instptr()) {
1777 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1778 && flat->is_instptr()->klass() == env()->Class_klass())
1779 alias_type(idx)->set_rewritable(false);
1780 }
1781 if (flat->isa_aryptr()) {
1782 #ifdef ASSERT
1783 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1784 // (T_BYTE has the weakest alignment and size restrictions...)
1785 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1786 #endif
1787 if (flat->offset() == TypePtr::OffsetBot) {
1788 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1789 }
1790 }
1791 if (flat->isa_klassptr()) {
1792 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1793 alias_type(idx)->set_rewritable(false);
1794 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1795 alias_type(idx)->set_rewritable(false);
1796 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1797 alias_type(idx)->set_rewritable(false);
1798 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1799 alias_type(idx)->set_rewritable(false);
1800 }
1801 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1802 // but the base pointer type is not distinctive enough to identify
1803 // references into JavaThread.)
1804
1805 // Check for final fields.
1806 const TypeInstPtr* tinst = flat->isa_instptr();
1807 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1808 ciField* field;
1809 if (tinst->const_oop() != NULL &&
1810 tinst->klass() == ciEnv::current()->Class_klass() &&
1811 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1812 // static field
1813 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1814 field = k->get_field_by_offset(tinst->offset(), true);
1815 } else {
1816 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1817 field = k->get_field_by_offset(tinst->offset(), false);
1818 }
1819 assert(field == NULL ||
1820 original_field == NULL ||
1821 (field->holder() == original_field->holder() &&
1822 field->offset() == original_field->offset() &&
1823 field->is_static() == original_field->is_static()), "wrong field?");
1824 // Set field() and is_rewritable() attributes.
1825 if (field != NULL) alias_type(idx)->set_field(field);
1826 }
1827 }
1828
1829 // Fill the cache for next time.
1830 ace->_adr_type = adr_type;
1831 ace->_index = idx;
1832 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1833
1834 // Might as well try to fill the cache for the flattened version, too.
1835 AliasCacheEntry* face = probe_alias_cache(flat);
1836 if (face->_adr_type == NULL) {
1837 face->_adr_type = flat;
1838 face->_index = idx;
1839 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1840 }
1841
1842 return alias_type(idx);
1843 }
1844
1845
1846 Compile::AliasType* Compile::alias_type(ciField* field) {
1847 const TypeOopPtr* t;
1848 if (field->is_static())
1849 t = TypeInstPtr::make(field->holder()->java_mirror());
1850 else
1851 t = TypeOopPtr::make_from_klass_raw(field->holder());
1852 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1853 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1854 return atp;
1855 }
1856
1857
1858 //------------------------------have_alias_type--------------------------------
1859 bool Compile::have_alias_type(const TypePtr* adr_type) {
1860 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1861 if (ace->_adr_type == adr_type) {
1862 return true;
1863 }
1864
1865 // Handle special cases.
1866 if (adr_type == NULL) return true;
1867 if (adr_type == TypePtr::BOTTOM) return true;
1868
1869 return find_alias_type(adr_type, true, NULL) != NULL;
1870 }
1871
1872 //-----------------------------must_alias--------------------------------------
1873 // True if all values of the given address type are in the given alias category.
1874 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1875 if (alias_idx == AliasIdxBot) return true; // the universal category
1876 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1877 if (alias_idx == AliasIdxTop) return false; // the empty category
1878 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1879
1880 // the only remaining possible overlap is identity
1881 int adr_idx = get_alias_index(adr_type);
1882 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1883 assert(adr_idx == alias_idx ||
1884 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1885 && adr_type != TypeOopPtr::BOTTOM),
1886 "should not be testing for overlap with an unsafe pointer");
1887 return adr_idx == alias_idx;
1888 }
1889
1890 //------------------------------can_alias--------------------------------------
1891 // True if any values of the given address type are in the given alias category.
1892 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1893 if (alias_idx == AliasIdxTop) return false; // the empty category
1894 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1895 if (alias_idx == AliasIdxBot) return true; // the universal category
1896 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1897
1898 // the only remaining possible overlap is identity
1899 int adr_idx = get_alias_index(adr_type);
1900 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1901 return adr_idx == alias_idx;
1902 }
1903
1904
1905
1906 //---------------------------pop_warm_call-------------------------------------
1907 WarmCallInfo* Compile::pop_warm_call() {
1908 WarmCallInfo* wci = _warm_calls;
1909 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1910 return wci;
1911 }
1912
1913 //----------------------------Inline_Warm--------------------------------------
1914 int Compile::Inline_Warm() {
1915 // If there is room, try to inline some more warm call sites.
1916 // %%% Do a graph index compaction pass when we think we're out of space?
1917 if (!InlineWarmCalls) return 0;
1918
1919 int calls_made_hot = 0;
1920 int room_to_grow = NodeCountInliningCutoff - unique();
1921 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1922 int amount_grown = 0;
1923 WarmCallInfo* call;
1924 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1925 int est_size = (int)call->size();
1926 if (est_size > (room_to_grow - amount_grown)) {
1927 // This one won't fit anyway. Get rid of it.
1928 call->make_cold();
1929 continue;
1930 }
1931 call->make_hot();
1932 calls_made_hot++;
1933 amount_grown += est_size;
1934 amount_to_grow -= est_size;
1935 }
1936
1937 if (calls_made_hot > 0) set_major_progress();
1938 return calls_made_hot;
1939 }
1940
1941
1942 //----------------------------Finish_Warm--------------------------------------
1943 void Compile::Finish_Warm() {
1944 if (!InlineWarmCalls) return;
1945 if (failing()) return;
1946 if (warm_calls() == NULL) return;
1947
1948 // Clean up loose ends, if we are out of space for inlining.
1949 WarmCallInfo* call;
1950 while ((call = pop_warm_call()) != NULL) {
1951 call->make_cold();
1952 }
1953 }
1954
1955 //---------------------cleanup_loop_predicates-----------------------
1956 // Remove the opaque nodes that protect the predicates so that all unused
1957 // checks and uncommon_traps will be eliminated from the ideal graph
1958 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1959 if (predicate_count()==0) return;
1960 for (int i = predicate_count(); i > 0; i--) {
1961 Node * n = predicate_opaque1_node(i-1);
1962 assert(n->Opcode() == Op_Opaque1, "must be");
1963 igvn.replace_node(n, n->in(1));
1964 }
1965 assert(predicate_count()==0, "should be clean!");
1966 }
1967
1968 void Compile::add_range_check_cast(Node* n) {
1969 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1970 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1971 _range_check_casts->append(n);
1972 }
1973
1974 // Remove all range check dependent CastIINodes.
1975 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1976 for (int i = range_check_cast_count(); i > 0; i--) {
1977 Node* cast = range_check_cast_node(i-1);
1978 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1979 igvn.replace_node(cast, cast->in(1));
1980 }
1981 assert(range_check_cast_count() == 0, "should be empty");
1982 }
1983
1984 void Compile::add_opaque4_node(Node* n) {
1985 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1986 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1987 _opaque4_nodes->append(n);
1988 }
1989
1990 // Remove all Opaque4 nodes.
1991 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1992 for (int i = opaque4_count(); i > 0; i--) {
1993 Node* opaq = opaque4_node(i-1);
1994 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
1995 igvn.replace_node(opaq, opaq->in(2));
1996 }
1997 assert(opaque4_count() == 0, "should be empty");
1998 }
1999
2000 // StringOpts and late inlining of string methods
2001 void Compile::inline_string_calls(bool parse_time) {
2002 {
2003 // remove useless nodes to make the usage analysis simpler
2004 ResourceMark rm;
2005 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2006 }
2007
2008 {
2009 ResourceMark rm;
2010 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2011 PhaseStringOpts pso(initial_gvn(), for_igvn());
2012 print_method(PHASE_AFTER_STRINGOPTS, 3);
2013 }
2014
2015 // now inline anything that we skipped the first time around
2016 if (!parse_time) {
2017 _late_inlines_pos = _late_inlines.length();
2018 }
2019
2020 while (_string_late_inlines.length() > 0) {
2021 CallGenerator* cg = _string_late_inlines.pop();
2022 cg->do_late_inline();
2023 if (failing()) return;
2024 }
2025 _string_late_inlines.trunc_to(0);
2026 }
2027
2028 // Late inlining of boxing methods
2029 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2030 if (_boxing_late_inlines.length() > 0) {
2031 assert(has_boxed_value(), "inconsistent");
2032
2033 PhaseGVN* gvn = initial_gvn();
2034 set_inlining_incrementally(true);
2035
2036 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2037 for_igvn()->clear();
2038 gvn->replace_with(&igvn);
2039
2040 _late_inlines_pos = _late_inlines.length();
2041
2042 while (_boxing_late_inlines.length() > 0) {
2043 CallGenerator* cg = _boxing_late_inlines.pop();
2044 cg->do_late_inline();
2045 if (failing()) return;
2046 }
2047 _boxing_late_inlines.trunc_to(0);
2048
2049 {
2050 ResourceMark rm;
2051 PhaseRemoveUseless pru(gvn, for_igvn());
2052 }
2053
2054 igvn = PhaseIterGVN(gvn);
2055 igvn.optimize();
2056
2057 set_inlining_progress(false);
2058 set_inlining_incrementally(false);
2059 }
2060 }
2061
2062 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2063 assert(IncrementalInline, "incremental inlining should be on");
2064 PhaseGVN* gvn = initial_gvn();
2065
2066 set_inlining_progress(false);
2067 for_igvn()->clear();
2068 gvn->replace_with(&igvn);
2069
2070 {
2071 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2072 int i = 0;
2073 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2074 CallGenerator* cg = _late_inlines.at(i);
2075 _late_inlines_pos = i+1;
2076 cg->do_late_inline();
2077 if (failing()) return;
2078 }
2079 int j = 0;
2080 for (; i < _late_inlines.length(); i++, j++) {
2081 _late_inlines.at_put(j, _late_inlines.at(i));
2082 }
2083 _late_inlines.trunc_to(j);
2084 }
2085
2086 {
2087 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2088 ResourceMark rm;
2089 PhaseRemoveUseless pru(gvn, for_igvn());
2090 }
2091
2092 {
2093 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2094 igvn = PhaseIterGVN(gvn);
2095 }
2096 }
2097
2098 // Perform incremental inlining until bound on number of live nodes is reached
2099 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2100 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2101
2102 PhaseGVN* gvn = initial_gvn();
2103
2104 set_inlining_incrementally(true);
2105 set_inlining_progress(true);
2106 uint low_live_nodes = 0;
2107
2108 while(inlining_progress() && _late_inlines.length() > 0) {
2109
2110 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2111 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2112 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2113 // PhaseIdealLoop is expensive so we only try it once we are
2114 // out of live nodes and we only try it again if the previous
2115 // helped got the number of nodes down significantly
2116 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2117 if (failing()) return;
2118 low_live_nodes = live_nodes();
2119 _major_progress = true;
2120 }
2121
2122 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2123 break;
2124 }
2125 }
2126
2127 inline_incrementally_one(igvn);
2128
2129 if (failing()) return;
2130
2131 {
2132 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2133 igvn.optimize();
2134 }
2135
2136 if (failing()) return;
2137 }
2138
2139 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2140
2141 if (_string_late_inlines.length() > 0) {
2142 assert(has_stringbuilder(), "inconsistent");
2143 for_igvn()->clear();
2144 initial_gvn()->replace_with(&igvn);
2145
2146 inline_string_calls(false);
2147
2148 if (failing()) return;
2149
2150 {
2151 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2152 ResourceMark rm;
2153 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2154 }
2155
2156 {
2157 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2158 igvn = PhaseIterGVN(gvn);
2159 igvn.optimize();
2160 }
2161 }
2162
2163 set_inlining_incrementally(false);
2164 }
2165
2166
2167 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2168 if(loop_opts_cnt > 0) {
2169 debug_only( int cnt = 0; );
2170 while(major_progress() && (loop_opts_cnt > 0)) {
2171 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2172 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2173 PhaseIdealLoop ideal_loop(igvn, mode);
2174 loop_opts_cnt--;
2175 if (failing()) return false;
2176 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2177 }
2178 }
2179 return true;
2180 }
2181
2182 //------------------------------Optimize---------------------------------------
2183 // Given a graph, optimize it.
2184 void Compile::Optimize() {
2185 TracePhase tp("optimizer", &timers[_t_optimizer]);
2186
2187 #ifndef PRODUCT
2188 if (_directive->BreakAtCompileOption) {
2189 BREAKPOINT;
2190 }
2191
2192 #endif
2193
2194 #ifdef ASSERT
2195 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2196 bs->verify_gc_barriers(true);
2197 #endif
2198
2199 ResourceMark rm;
2200 int loop_opts_cnt;
2201
2202 print_inlining_reinit();
2203
2204 NOT_PRODUCT( verify_graph_edges(); )
2205
2206 print_method(PHASE_AFTER_PARSING);
2207
2208 {
2209 // Iterative Global Value Numbering, including ideal transforms
2210 // Initialize IterGVN with types and values from parse-time GVN
2211 PhaseIterGVN igvn(initial_gvn());
2212 #ifdef ASSERT
2213 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2214 #endif
2215 {
2216 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2217 igvn.optimize();
2218 }
2219
2220 if (failing()) return;
2221
2222 print_method(PHASE_ITER_GVN1, 2);
2223
2224 inline_incrementally(igvn);
2225
2226 print_method(PHASE_INCREMENTAL_INLINE, 2);
2227
2228 if (failing()) return;
2229
2230 if (eliminate_boxing()) {
2231 // Inline valueOf() methods now.
2232 inline_boxing_calls(igvn);
2233
2234 if (AlwaysIncrementalInline) {
2235 inline_incrementally(igvn);
2236 }
2237
2238 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2239
2240 if (failing()) return;
2241 }
2242
2243 // Remove the speculative part of types and clean up the graph from
2244 // the extra CastPP nodes whose only purpose is to carry them. Do
2245 // that early so that optimizations are not disrupted by the extra
2246 // CastPP nodes.
2247 remove_speculative_types(igvn);
2248
2249 // No more new expensive nodes will be added to the list from here
2250 // so keep only the actual candidates for optimizations.
2251 cleanup_expensive_nodes(igvn);
2252
2253 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2254 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2255 initial_gvn()->replace_with(&igvn);
2256 for_igvn()->clear();
2257 Unique_Node_List new_worklist(C->comp_arena());
2258 {
2259 ResourceMark rm;
2260 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2261 }
2262 set_for_igvn(&new_worklist);
2263 igvn = PhaseIterGVN(initial_gvn());
2264 igvn.optimize();
2265 }
2266
2267 // Perform escape analysis
2268 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2269 if (has_loops()) {
2270 // Cleanup graph (remove dead nodes).
2271 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2272 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2273 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2274 if (failing()) return;
2275 }
2276 ConnectionGraph::do_analysis(this, &igvn);
2277
2278 if (failing()) return;
2279
2280 // Optimize out fields loads from scalar replaceable allocations.
2281 igvn.optimize();
2282 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2283
2284 if (failing()) return;
2285
2286 if (congraph() != NULL && macro_count() > 0) {
2287 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2288 PhaseMacroExpand mexp(igvn);
2289 mexp.eliminate_macro_nodes();
2290 igvn.set_delay_transform(false);
2291
2292 igvn.optimize();
2293 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2294
2295 if (failing()) return;
2296 }
2297 }
2298
2299 // Loop transforms on the ideal graph. Range Check Elimination,
2300 // peeling, unrolling, etc.
2301
2302 // Set loop opts counter
2303 loop_opts_cnt = num_loop_opts();
2304 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2305 {
2306 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2307 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2308 loop_opts_cnt--;
2309 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2310 if (failing()) return;
2311 }
2312 // Loop opts pass if partial peeling occurred in previous pass
2313 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2314 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2315 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2316 loop_opts_cnt--;
2317 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2318 if (failing()) return;
2319 }
2320 // Loop opts pass for loop-unrolling before CCP
2321 if(major_progress() && (loop_opts_cnt > 0)) {
2322 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2323 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2324 loop_opts_cnt--;
2325 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2326 }
2327 if (!failing()) {
2328 // Verify that last round of loop opts produced a valid graph
2329 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2330 PhaseIdealLoop::verify(igvn);
2331 }
2332 }
2333 if (failing()) return;
2334
2335 // Conditional Constant Propagation;
2336 PhaseCCP ccp( &igvn );
2337 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2338 {
2339 TracePhase tp("ccp", &timers[_t_ccp]);
2340 ccp.do_transform();
2341 }
2342 print_method(PHASE_CPP1, 2);
2343
2344 assert( true, "Break here to ccp.dump_old2new_map()");
2345
2346 // Iterative Global Value Numbering, including ideal transforms
2347 {
2348 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2349 igvn = ccp;
2350 igvn.optimize();
2351 }
2352
2353 print_method(PHASE_ITER_GVN2, 2);
2354
2355 if (failing()) return;
2356
2357 // Loop transforms on the ideal graph. Range Check Elimination,
2358 // peeling, unrolling, etc.
2359 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2360 return;
2361 }
2362
2363 #if INCLUDE_ZGC
2364 if (UseZGC) {
2365 ZBarrierSetC2::find_dominating_barriers(igvn);
2366 }
2367 #endif
2368
2369 if (failing()) return;
2370
2371 // Ensure that major progress is now clear
2372 C->clear_major_progress();
2373
2374 {
2375 // Verify that all previous optimizations produced a valid graph
2376 // at least to this point, even if no loop optimizations were done.
2377 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2378 PhaseIdealLoop::verify(igvn);
2379 }
2380
2381 if (range_check_cast_count() > 0) {
2382 // No more loop optimizations. Remove all range check dependent CastIINodes.
2383 C->remove_range_check_casts(igvn);
2384 igvn.optimize();
2385 }
2386
2387 #ifdef ASSERT
2388 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2389 bs->verify_gc_barriers(false);
2390 #endif
2391
2392 {
2393 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2394 PhaseMacroExpand mex(igvn);
2395 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2396 if (mex.expand_macro_nodes()) {
2397 assert(failing(), "must bail out w/ explicit message");
2398 return;
2399 }
2400 }
2401
2402 if (opaque4_count() > 0) {
2403 C->remove_opaque4_nodes(igvn);
2404 igvn.optimize();
2405 }
2406
2407 DEBUG_ONLY( _modified_nodes = NULL; )
2408 } // (End scope of igvn; run destructor if necessary for asserts.)
2409
2410 process_print_inlining();
2411 // A method with only infinite loops has no edges entering loops from root
2412 {
2413 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2414 if (final_graph_reshaping()) {
2415 assert(failing(), "must bail out w/ explicit message");
2416 return;
2417 }
2418 }
2419
2420 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2421 }
2422
2423
2424 //------------------------------Code_Gen---------------------------------------
2425 // Given a graph, generate code for it
2426 void Compile::Code_Gen() {
2427 if (failing()) {
2428 return;
2429 }
2430
2431 // Perform instruction selection. You might think we could reclaim Matcher
2432 // memory PDQ, but actually the Matcher is used in generating spill code.
2433 // Internals of the Matcher (including some VectorSets) must remain live
2434 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2435 // set a bit in reclaimed memory.
2436
2437 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2438 // nodes. Mapping is only valid at the root of each matched subtree.
2439 NOT_PRODUCT( verify_graph_edges(); )
2440
2441 Matcher matcher;
2442 _matcher = &matcher;
2443 {
2444 TracePhase tp("matcher", &timers[_t_matcher]);
2445 matcher.match();
2446 }
2447 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2448 // nodes. Mapping is only valid at the root of each matched subtree.
2449 NOT_PRODUCT( verify_graph_edges(); )
2450
2451 // If you have too many nodes, or if matching has failed, bail out
2452 check_node_count(0, "out of nodes matching instructions");
2453 if (failing()) {
2454 return;
2455 }
2456
2457 print_method(PHASE_MATCHING, 2);
2458
2459 // Build a proper-looking CFG
2460 PhaseCFG cfg(node_arena(), root(), matcher);
2461 _cfg = &cfg;
2462 {
2463 TracePhase tp("scheduler", &timers[_t_scheduler]);
2464 bool success = cfg.do_global_code_motion();
2465 if (!success) {
2466 return;
2467 }
2468
2469 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2470 NOT_PRODUCT( verify_graph_edges(); )
2471 debug_only( cfg.verify(); )
2472 }
2473
2474 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2475 _regalloc = ®alloc;
2476 {
2477 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2478 // Perform register allocation. After Chaitin, use-def chains are
2479 // no longer accurate (at spill code) and so must be ignored.
2480 // Node->LRG->reg mappings are still accurate.
2481 _regalloc->Register_Allocate();
2482
2483 // Bail out if the allocator builds too many nodes
2484 if (failing()) {
2485 return;
2486 }
2487 }
2488
2489 // Prior to register allocation we kept empty basic blocks in case the
2490 // the allocator needed a place to spill. After register allocation we
2491 // are not adding any new instructions. If any basic block is empty, we
2492 // can now safely remove it.
2493 {
2494 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2495 cfg.remove_empty_blocks();
2496 if (do_freq_based_layout()) {
2497 PhaseBlockLayout layout(cfg);
2498 } else {
2499 cfg.set_loop_alignment();
2500 }
2501 cfg.fixup_flow();
2502 }
2503
2504 // Apply peephole optimizations
2505 if( OptoPeephole ) {
2506 TracePhase tp("peephole", &timers[_t_peephole]);
2507 PhasePeephole peep( _regalloc, cfg);
2508 peep.do_transform();
2509 }
2510
2511 // Do late expand if CPU requires this.
2512 if (Matcher::require_postalloc_expand) {
2513 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2514 cfg.postalloc_expand(_regalloc);
2515 }
2516
2517 // Convert Nodes to instruction bits in a buffer
2518 {
2519 TraceTime tp("output", &timers[_t_output], CITime);
2520 Output();
2521 }
2522
2523 print_method(PHASE_FINAL_CODE);
2524
2525 // He's dead, Jim.
2526 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2527 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2528 }
2529
2530
2531 //------------------------------dump_asm---------------------------------------
2532 // Dump formatted assembly
2533 #ifndef PRODUCT
2534 void Compile::dump_asm(int *pcs, uint pc_limit) {
2535 bool cut_short = false;
2536 tty->print_cr("#");
2537 tty->print("# "); _tf->dump(); tty->cr();
2538 tty->print_cr("#");
2539
2540 // For all blocks
2541 int pc = 0x0; // Program counter
2542 char starts_bundle = ' ';
2543 _regalloc->dump_frame();
2544
2545 Node *n = NULL;
2546 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2547 if (VMThread::should_terminate()) {
2548 cut_short = true;
2549 break;
2550 }
2551 Block* block = _cfg->get_block(i);
2552 if (block->is_connector() && !Verbose) {
2553 continue;
2554 }
2555 n = block->head();
2556 if (pcs && n->_idx < pc_limit) {
2557 tty->print("%3.3x ", pcs[n->_idx]);
2558 } else {
2559 tty->print(" ");
2560 }
2561 block->dump_head(_cfg);
2562 if (block->is_connector()) {
2563 tty->print_cr(" # Empty connector block");
2564 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2565 tty->print_cr(" # Block is sole successor of call");
2566 }
2567
2568 // For all instructions
2569 Node *delay = NULL;
2570 for (uint j = 0; j < block->number_of_nodes(); j++) {
2571 if (VMThread::should_terminate()) {
2572 cut_short = true;
2573 break;
2574 }
2575 n = block->get_node(j);
2576 if (valid_bundle_info(n)) {
2577 Bundle* bundle = node_bundling(n);
2578 if (bundle->used_in_unconditional_delay()) {
2579 delay = n;
2580 continue;
2581 }
2582 if (bundle->starts_bundle()) {
2583 starts_bundle = '+';
2584 }
2585 }
2586
2587 if (WizardMode) {
2588 n->dump();
2589 }
2590
2591 if( !n->is_Region() && // Dont print in the Assembly
2592 !n->is_Phi() && // a few noisely useless nodes
2593 !n->is_Proj() &&
2594 !n->is_MachTemp() &&
2595 !n->is_SafePointScalarObject() &&
2596 !n->is_Catch() && // Would be nice to print exception table targets
2597 !n->is_MergeMem() && // Not very interesting
2598 !n->is_top() && // Debug info table constants
2599 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2600 ) {
2601 if (pcs && n->_idx < pc_limit)
2602 tty->print("%3.3x", pcs[n->_idx]);
2603 else
2604 tty->print(" ");
2605 tty->print(" %c ", starts_bundle);
2606 starts_bundle = ' ';
2607 tty->print("\t");
2608 n->format(_regalloc, tty);
2609 tty->cr();
2610 }
2611
2612 // If we have an instruction with a delay slot, and have seen a delay,
2613 // then back up and print it
2614 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2615 assert(delay != NULL, "no unconditional delay instruction");
2616 if (WizardMode) delay->dump();
2617
2618 if (node_bundling(delay)->starts_bundle())
2619 starts_bundle = '+';
2620 if (pcs && n->_idx < pc_limit)
2621 tty->print("%3.3x", pcs[n->_idx]);
2622 else
2623 tty->print(" ");
2624 tty->print(" %c ", starts_bundle);
2625 starts_bundle = ' ';
2626 tty->print("\t");
2627 delay->format(_regalloc, tty);
2628 tty->cr();
2629 delay = NULL;
2630 }
2631
2632 // Dump the exception table as well
2633 if( n->is_Catch() && (Verbose || WizardMode) ) {
2634 // Print the exception table for this offset
2635 _handler_table.print_subtable_for(pc);
2636 }
2637 }
2638
2639 if (pcs && n->_idx < pc_limit)
2640 tty->print_cr("%3.3x", pcs[n->_idx]);
2641 else
2642 tty->cr();
2643
2644 assert(cut_short || delay == NULL, "no unconditional delay branch");
2645
2646 } // End of per-block dump
2647 tty->cr();
2648
2649 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2650 }
2651 #endif
2652
2653 //------------------------------Final_Reshape_Counts---------------------------
2654 // This class defines counters to help identify when a method
2655 // may/must be executed using hardware with only 24-bit precision.
2656 struct Final_Reshape_Counts : public StackObj {
2657 int _call_count; // count non-inlined 'common' calls
2658 int _float_count; // count float ops requiring 24-bit precision
2659 int _double_count; // count double ops requiring more precision
2660 int _java_call_count; // count non-inlined 'java' calls
2661 int _inner_loop_count; // count loops which need alignment
2662 VectorSet _visited; // Visitation flags
2663 Node_List _tests; // Set of IfNodes & PCTableNodes
2664
2665 Final_Reshape_Counts() :
2666 _call_count(0), _float_count(0), _double_count(0),
2667 _java_call_count(0), _inner_loop_count(0),
2668 _visited( Thread::current()->resource_area() ) { }
2669
2670 void inc_call_count () { _call_count ++; }
2671 void inc_float_count () { _float_count ++; }
2672 void inc_double_count() { _double_count++; }
2673 void inc_java_call_count() { _java_call_count++; }
2674 void inc_inner_loop_count() { _inner_loop_count++; }
2675
2676 int get_call_count () const { return _call_count ; }
2677 int get_float_count () const { return _float_count ; }
2678 int get_double_count() const { return _double_count; }
2679 int get_java_call_count() const { return _java_call_count; }
2680 int get_inner_loop_count() const { return _inner_loop_count; }
2681 };
2682
2683 #ifdef ASSERT
2684 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2685 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2686 // Make sure the offset goes inside the instance layout.
2687 return k->contains_field_offset(tp->offset());
2688 // Note that OffsetBot and OffsetTop are very negative.
2689 }
2690 #endif
2691
2692 // Eliminate trivially redundant StoreCMs and accumulate their
2693 // precedence edges.
2694 void Compile::eliminate_redundant_card_marks(Node* n) {
2695 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2696 if (n->in(MemNode::Address)->outcnt() > 1) {
2697 // There are multiple users of the same address so it might be
2698 // possible to eliminate some of the StoreCMs
2699 Node* mem = n->in(MemNode::Memory);
2700 Node* adr = n->in(MemNode::Address);
2701 Node* val = n->in(MemNode::ValueIn);
2702 Node* prev = n;
2703 bool done = false;
2704 // Walk the chain of StoreCMs eliminating ones that match. As
2705 // long as it's a chain of single users then the optimization is
2706 // safe. Eliminating partially redundant StoreCMs would require
2707 // cloning copies down the other paths.
2708 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2709 if (adr == mem->in(MemNode::Address) &&
2710 val == mem->in(MemNode::ValueIn)) {
2711 // redundant StoreCM
2712 if (mem->req() > MemNode::OopStore) {
2713 // Hasn't been processed by this code yet.
2714 n->add_prec(mem->in(MemNode::OopStore));
2715 } else {
2716 // Already converted to precedence edge
2717 for (uint i = mem->req(); i < mem->len(); i++) {
2718 // Accumulate any precedence edges
2719 if (mem->in(i) != NULL) {
2720 n->add_prec(mem->in(i));
2721 }
2722 }
2723 // Everything above this point has been processed.
2724 done = true;
2725 }
2726 // Eliminate the previous StoreCM
2727 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2728 assert(mem->outcnt() == 0, "should be dead");
2729 mem->disconnect_inputs(NULL, this);
2730 } else {
2731 prev = mem;
2732 }
2733 mem = prev->in(MemNode::Memory);
2734 }
2735 }
2736 }
2737
2738 //------------------------------final_graph_reshaping_impl----------------------
2739 // Implement items 1-5 from final_graph_reshaping below.
2740 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2741
2742 if ( n->outcnt() == 0 ) return; // dead node
2743 uint nop = n->Opcode();
2744
2745 // Check for 2-input instruction with "last use" on right input.
2746 // Swap to left input. Implements item (2).
2747 if( n->req() == 3 && // two-input instruction
2748 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2749 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2750 n->in(2)->outcnt() == 1 &&// right use IS a last use
2751 !n->in(2)->is_Con() ) { // right use is not a constant
2752 // Check for commutative opcode
2753 switch( nop ) {
2754 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2755 case Op_MaxI: case Op_MinI:
2756 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2757 case Op_AndL: case Op_XorL: case Op_OrL:
2758 case Op_AndI: case Op_XorI: case Op_OrI: {
2759 // Move "last use" input to left by swapping inputs
2760 n->swap_edges(1, 2);
2761 break;
2762 }
2763 default:
2764 break;
2765 }
2766 }
2767
2768 #ifdef ASSERT
2769 if( n->is_Mem() ) {
2770 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2771 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2772 // oop will be recorded in oop map if load crosses safepoint
2773 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2774 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2775 "raw memory operations should have control edge");
2776 }
2777 if (n->is_MemBar()) {
2778 MemBarNode* mb = n->as_MemBar();
2779 if (mb->trailing_store() || mb->trailing_load_store()) {
2780 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2781 Node* mem = mb->in(MemBarNode::Precedent);
2782 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2783 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2784 } else if (mb->leading()) {
2785 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2786 }
2787 }
2788 #endif
2789 // Count FPU ops and common calls, implements item (3)
2790 switch( nop ) {
2791 // Count all float operations that may use FPU
2792 case Op_AddF:
2793 case Op_SubF:
2794 case Op_MulF:
2795 case Op_DivF:
2796 case Op_NegF:
2797 case Op_ModF:
2798 case Op_ConvI2F:
2799 case Op_ConF:
2800 case Op_CmpF:
2801 case Op_CmpF3:
2802 // case Op_ConvL2F: // longs are split into 32-bit halves
2803 frc.inc_float_count();
2804 break;
2805
2806 case Op_ConvF2D:
2807 case Op_ConvD2F:
2808 frc.inc_float_count();
2809 frc.inc_double_count();
2810 break;
2811
2812 // Count all double operations that may use FPU
2813 case Op_AddD:
2814 case Op_SubD:
2815 case Op_MulD:
2816 case Op_DivD:
2817 case Op_NegD:
2818 case Op_ModD:
2819 case Op_ConvI2D:
2820 case Op_ConvD2I:
2821 // case Op_ConvL2D: // handled by leaf call
2822 // case Op_ConvD2L: // handled by leaf call
2823 case Op_ConD:
2824 case Op_CmpD:
2825 case Op_CmpD3:
2826 frc.inc_double_count();
2827 break;
2828 case Op_Opaque1: // Remove Opaque Nodes before matching
2829 case Op_Opaque2: // Remove Opaque Nodes before matching
2830 case Op_Opaque3:
2831 n->subsume_by(n->in(1), this);
2832 break;
2833 case Op_CallStaticJava:
2834 case Op_CallJava:
2835 case Op_CallDynamicJava:
2836 frc.inc_java_call_count(); // Count java call site;
2837 case Op_CallRuntime:
2838 case Op_CallLeaf:
2839 case Op_CallLeafNoFP: {
2840 assert (n->is_Call(), "");
2841 CallNode *call = n->as_Call();
2842 // Count call sites where the FP mode bit would have to be flipped.
2843 // Do not count uncommon runtime calls:
2844 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2845 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2846 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2847 frc.inc_call_count(); // Count the call site
2848 } else { // See if uncommon argument is shared
2849 Node *n = call->in(TypeFunc::Parms);
2850 int nop = n->Opcode();
2851 // Clone shared simple arguments to uncommon calls, item (1).
2852 if (n->outcnt() > 1 &&
2853 !n->is_Proj() &&
2854 nop != Op_CreateEx &&
2855 nop != Op_CheckCastPP &&
2856 nop != Op_DecodeN &&
2857 nop != Op_DecodeNKlass &&
2858 !n->is_Mem() &&
2859 !n->is_Phi()) {
2860 Node *x = n->clone();
2861 call->set_req(TypeFunc::Parms, x);
2862 }
2863 }
2864 break;
2865 }
2866
2867 case Op_StoreD:
2868 case Op_LoadD:
2869 case Op_LoadD_unaligned:
2870 frc.inc_double_count();
2871 goto handle_mem;
2872 case Op_StoreF:
2873 case Op_LoadF:
2874 frc.inc_float_count();
2875 goto handle_mem;
2876
2877 case Op_StoreCM:
2878 {
2879 // Convert OopStore dependence into precedence edge
2880 Node* prec = n->in(MemNode::OopStore);
2881 n->del_req(MemNode::OopStore);
2882 n->add_prec(prec);
2883 eliminate_redundant_card_marks(n);
2884 }
2885
2886 // fall through
2887
2888 case Op_StoreB:
2889 case Op_StoreC:
2890 case Op_StorePConditional:
2891 case Op_StoreI:
2892 case Op_StoreL:
2893 case Op_StoreIConditional:
2894 case Op_StoreLConditional:
2895 case Op_CompareAndSwapB:
2896 case Op_CompareAndSwapS:
2897 case Op_CompareAndSwapI:
2898 case Op_CompareAndSwapL:
2899 case Op_CompareAndSwapP:
2900 case Op_CompareAndSwapN:
2901 case Op_WeakCompareAndSwapB:
2902 case Op_WeakCompareAndSwapS:
2903 case Op_WeakCompareAndSwapI:
2904 case Op_WeakCompareAndSwapL:
2905 case Op_WeakCompareAndSwapP:
2906 case Op_WeakCompareAndSwapN:
2907 case Op_CompareAndExchangeB:
2908 case Op_CompareAndExchangeS:
2909 case Op_CompareAndExchangeI:
2910 case Op_CompareAndExchangeL:
2911 case Op_CompareAndExchangeP:
2912 case Op_CompareAndExchangeN:
2913 case Op_GetAndAddS:
2914 case Op_GetAndAddB:
2915 case Op_GetAndAddI:
2916 case Op_GetAndAddL:
2917 case Op_GetAndSetS:
2918 case Op_GetAndSetB:
2919 case Op_GetAndSetI:
2920 case Op_GetAndSetL:
2921 case Op_GetAndSetP:
2922 case Op_GetAndSetN:
2923 case Op_StoreP:
2924 case Op_StoreN:
2925 case Op_StoreNKlass:
2926 case Op_LoadB:
2927 case Op_LoadUB:
2928 case Op_LoadUS:
2929 case Op_LoadI:
2930 case Op_LoadKlass:
2931 case Op_LoadNKlass:
2932 case Op_LoadL:
2933 case Op_LoadL_unaligned:
2934 case Op_LoadPLocked:
2935 case Op_LoadP:
2936 #if INCLUDE_ZGC
2937 case Op_LoadBarrierSlowReg:
2938 case Op_LoadBarrierWeakSlowReg:
2939 #endif
2940 case Op_LoadN:
2941 case Op_LoadRange:
2942 case Op_LoadS: {
2943 handle_mem:
2944 #ifdef ASSERT
2945 if( VerifyOptoOopOffsets ) {
2946 assert( n->is_Mem(), "" );
2947 MemNode *mem = (MemNode*)n;
2948 // Check to see if address types have grounded out somehow.
2949 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2950 assert( !tp || oop_offset_is_sane(tp), "" );
2951 }
2952 #endif
2953 break;
2954 }
2955
2956 case Op_AddP: { // Assert sane base pointers
2957 Node *addp = n->in(AddPNode::Address);
2958 assert( !addp->is_AddP() ||
2959 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2960 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2961 "Base pointers must match (addp %u)", addp->_idx );
2962 #ifdef _LP64
2963 if ((UseCompressedOops || UseCompressedClassPointers) &&
2964 addp->Opcode() == Op_ConP &&
2965 addp == n->in(AddPNode::Base) &&
2966 n->in(AddPNode::Offset)->is_Con()) {
2967 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2968 // on the platform and on the compressed oops mode.
2969 // Use addressing with narrow klass to load with offset on x86.
2970 // Some platforms can use the constant pool to load ConP.
2971 // Do this transformation here since IGVN will convert ConN back to ConP.
2972 const Type* t = addp->bottom_type();
2973 bool is_oop = t->isa_oopptr() != NULL;
2974 bool is_klass = t->isa_klassptr() != NULL;
2975
2976 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2977 (is_klass && Matcher::const_klass_prefer_decode())) {
2978 Node* nn = NULL;
2979
2980 int op = is_oop ? Op_ConN : Op_ConNKlass;
2981
2982 // Look for existing ConN node of the same exact type.
2983 Node* r = root();
2984 uint cnt = r->outcnt();
2985 for (uint i = 0; i < cnt; i++) {
2986 Node* m = r->raw_out(i);
2987 if (m!= NULL && m->Opcode() == op &&
2988 m->bottom_type()->make_ptr() == t) {
2989 nn = m;
2990 break;
2991 }
2992 }
2993 if (nn != NULL) {
2994 // Decode a narrow oop to match address
2995 // [R12 + narrow_oop_reg<<3 + offset]
2996 if (is_oop) {
2997 nn = new DecodeNNode(nn, t);
2998 } else {
2999 nn = new DecodeNKlassNode(nn, t);
3000 }
3001 // Check for succeeding AddP which uses the same Base.
3002 // Otherwise we will run into the assertion above when visiting that guy.
3003 for (uint i = 0; i < n->outcnt(); ++i) {
3004 Node *out_i = n->raw_out(i);
3005 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3006 out_i->set_req(AddPNode::Base, nn);
3007 #ifdef ASSERT
3008 for (uint j = 0; j < out_i->outcnt(); ++j) {
3009 Node *out_j = out_i->raw_out(j);
3010 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3011 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3012 }
3013 #endif
3014 }
3015 }
3016 n->set_req(AddPNode::Base, nn);
3017 n->set_req(AddPNode::Address, nn);
3018 if (addp->outcnt() == 0) {
3019 addp->disconnect_inputs(NULL, this);
3020 }
3021 }
3022 }
3023 }
3024 #endif
3025 // platform dependent reshaping of the address expression
3026 reshape_address(n->as_AddP());
3027 break;
3028 }
3029
3030 case Op_CastPP: {
3031 // Remove CastPP nodes to gain more freedom during scheduling but
3032 // keep the dependency they encode as control or precedence edges
3033 // (if control is set already) on memory operations. Some CastPP
3034 // nodes don't have a control (don't carry a dependency): skip
3035 // those.
3036 if (n->in(0) != NULL) {
3037 ResourceMark rm;
3038 Unique_Node_List wq;
3039 wq.push(n);
3040 for (uint next = 0; next < wq.size(); ++next) {
3041 Node *m = wq.at(next);
3042 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3043 Node* use = m->fast_out(i);
3044 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3045 use->ensure_control_or_add_prec(n->in(0));
3046 } else {
3047 switch(use->Opcode()) {
3048 case Op_AddP:
3049 case Op_DecodeN:
3050 case Op_DecodeNKlass:
3051 case Op_CheckCastPP:
3052 case Op_CastPP:
3053 wq.push(use);
3054 break;
3055 }
3056 }
3057 }
3058 }
3059 }
3060 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3061 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3062 Node* in1 = n->in(1);
3063 const Type* t = n->bottom_type();
3064 Node* new_in1 = in1->clone();
3065 new_in1->as_DecodeN()->set_type(t);
3066
3067 if (!Matcher::narrow_oop_use_complex_address()) {
3068 //
3069 // x86, ARM and friends can handle 2 adds in addressing mode
3070 // and Matcher can fold a DecodeN node into address by using
3071 // a narrow oop directly and do implicit NULL check in address:
3072 //
3073 // [R12 + narrow_oop_reg<<3 + offset]
3074 // NullCheck narrow_oop_reg
3075 //
3076 // On other platforms (Sparc) we have to keep new DecodeN node and
3077 // use it to do implicit NULL check in address:
3078 //
3079 // decode_not_null narrow_oop_reg, base_reg
3080 // [base_reg + offset]
3081 // NullCheck base_reg
3082 //
3083 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3084 // to keep the information to which NULL check the new DecodeN node
3085 // corresponds to use it as value in implicit_null_check().
3086 //
3087 new_in1->set_req(0, n->in(0));
3088 }
3089
3090 n->subsume_by(new_in1, this);
3091 if (in1->outcnt() == 0) {
3092 in1->disconnect_inputs(NULL, this);
3093 }
3094 } else {
3095 n->subsume_by(n->in(1), this);
3096 if (n->outcnt() == 0) {
3097 n->disconnect_inputs(NULL, this);
3098 }
3099 }
3100 break;
3101 }
3102 #ifdef _LP64
3103 case Op_CmpP:
3104 // Do this transformation here to preserve CmpPNode::sub() and
3105 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3106 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3107 Node* in1 = n->in(1);
3108 Node* in2 = n->in(2);
3109 if (!in1->is_DecodeNarrowPtr()) {
3110 in2 = in1;
3111 in1 = n->in(2);
3112 }
3113 assert(in1->is_DecodeNarrowPtr(), "sanity");
3114
3115 Node* new_in2 = NULL;
3116 if (in2->is_DecodeNarrowPtr()) {
3117 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3118 new_in2 = in2->in(1);
3119 } else if (in2->Opcode() == Op_ConP) {
3120 const Type* t = in2->bottom_type();
3121 if (t == TypePtr::NULL_PTR) {
3122 assert(in1->is_DecodeN(), "compare klass to null?");
3123 // Don't convert CmpP null check into CmpN if compressed
3124 // oops implicit null check is not generated.
3125 // This will allow to generate normal oop implicit null check.
3126 if (Matcher::gen_narrow_oop_implicit_null_checks())
3127 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3128 //
3129 // This transformation together with CastPP transformation above
3130 // will generated code for implicit NULL checks for compressed oops.
3131 //
3132 // The original code after Optimize()
3133 //
3134 // LoadN memory, narrow_oop_reg
3135 // decode narrow_oop_reg, base_reg
3136 // CmpP base_reg, NULL
3137 // CastPP base_reg // NotNull
3138 // Load [base_reg + offset], val_reg
3139 //
3140 // after these transformations will be
3141 //
3142 // LoadN memory, narrow_oop_reg
3143 // CmpN narrow_oop_reg, NULL
3144 // decode_not_null narrow_oop_reg, base_reg
3145 // Load [base_reg + offset], val_reg
3146 //
3147 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3148 // since narrow oops can be used in debug info now (see the code in
3149 // final_graph_reshaping_walk()).
3150 //
3151 // At the end the code will be matched to
3152 // on x86:
3153 //
3154 // Load_narrow_oop memory, narrow_oop_reg
3155 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3156 // NullCheck narrow_oop_reg
3157 //
3158 // and on sparc:
3159 //
3160 // Load_narrow_oop memory, narrow_oop_reg
3161 // decode_not_null narrow_oop_reg, base_reg
3162 // Load [base_reg + offset], val_reg
3163 // NullCheck base_reg
3164 //
3165 } else if (t->isa_oopptr()) {
3166 new_in2 = ConNode::make(t->make_narrowoop());
3167 } else if (t->isa_klassptr()) {
3168 new_in2 = ConNode::make(t->make_narrowklass());
3169 }
3170 }
3171 if (new_in2 != NULL) {
3172 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3173 n->subsume_by(cmpN, this);
3174 if (in1->outcnt() == 0) {
3175 in1->disconnect_inputs(NULL, this);
3176 }
3177 if (in2->outcnt() == 0) {
3178 in2->disconnect_inputs(NULL, this);
3179 }
3180 }
3181 }
3182 break;
3183
3184 case Op_DecodeN:
3185 case Op_DecodeNKlass:
3186 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3187 // DecodeN could be pinned when it can't be fold into
3188 // an address expression, see the code for Op_CastPP above.
3189 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3190 break;
3191
3192 case Op_EncodeP:
3193 case Op_EncodePKlass: {
3194 Node* in1 = n->in(1);
3195 if (in1->is_DecodeNarrowPtr()) {
3196 n->subsume_by(in1->in(1), this);
3197 } else if (in1->Opcode() == Op_ConP) {
3198 const Type* t = in1->bottom_type();
3199 if (t == TypePtr::NULL_PTR) {
3200 assert(t->isa_oopptr(), "null klass?");
3201 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3202 } else if (t->isa_oopptr()) {
3203 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3204 } else if (t->isa_klassptr()) {
3205 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3206 }
3207 }
3208 if (in1->outcnt() == 0) {
3209 in1->disconnect_inputs(NULL, this);
3210 }
3211 break;
3212 }
3213
3214 case Op_Proj: {
3215 if (OptimizeStringConcat) {
3216 ProjNode* p = n->as_Proj();
3217 if (p->_is_io_use) {
3218 // Separate projections were used for the exception path which
3219 // are normally removed by a late inline. If it wasn't inlined
3220 // then they will hang around and should just be replaced with
3221 // the original one.
3222 Node* proj = NULL;
3223 // Replace with just one
3224 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3225 Node *use = i.get();
3226 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3227 proj = use;
3228 break;
3229 }
3230 }
3231 assert(proj != NULL, "must be found");
3232 p->subsume_by(proj, this);
3233 }
3234 }
3235 break;
3236 }
3237
3238 case Op_Phi:
3239 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3240 // The EncodeP optimization may create Phi with the same edges
3241 // for all paths. It is not handled well by Register Allocator.
3242 Node* unique_in = n->in(1);
3243 assert(unique_in != NULL, "");
3244 uint cnt = n->req();
3245 for (uint i = 2; i < cnt; i++) {
3246 Node* m = n->in(i);
3247 assert(m != NULL, "");
3248 if (unique_in != m)
3249 unique_in = NULL;
3250 }
3251 if (unique_in != NULL) {
3252 n->subsume_by(unique_in, this);
3253 }
3254 }
3255 break;
3256
3257 #endif
3258
3259 #ifdef ASSERT
3260 case Op_CastII:
3261 // Verify that all range check dependent CastII nodes were removed.
3262 if (n->isa_CastII()->has_range_check()) {
3263 n->dump(3);
3264 assert(false, "Range check dependent CastII node was not removed");
3265 }
3266 break;
3267 #endif
3268
3269 case Op_ModI:
3270 if (UseDivMod) {
3271 // Check if a%b and a/b both exist
3272 Node* d = n->find_similar(Op_DivI);
3273 if (d) {
3274 // Replace them with a fused divmod if supported
3275 if (Matcher::has_match_rule(Op_DivModI)) {
3276 DivModINode* divmod = DivModINode::make(n);
3277 d->subsume_by(divmod->div_proj(), this);
3278 n->subsume_by(divmod->mod_proj(), this);
3279 } else {
3280 // replace a%b with a-((a/b)*b)
3281 Node* mult = new MulINode(d, d->in(2));
3282 Node* sub = new SubINode(d->in(1), mult);
3283 n->subsume_by(sub, this);
3284 }
3285 }
3286 }
3287 break;
3288
3289 case Op_ModL:
3290 if (UseDivMod) {
3291 // Check if a%b and a/b both exist
3292 Node* d = n->find_similar(Op_DivL);
3293 if (d) {
3294 // Replace them with a fused divmod if supported
3295 if (Matcher::has_match_rule(Op_DivModL)) {
3296 DivModLNode* divmod = DivModLNode::make(n);
3297 d->subsume_by(divmod->div_proj(), this);
3298 n->subsume_by(divmod->mod_proj(), this);
3299 } else {
3300 // replace a%b with a-((a/b)*b)
3301 Node* mult = new MulLNode(d, d->in(2));
3302 Node* sub = new SubLNode(d->in(1), mult);
3303 n->subsume_by(sub, this);
3304 }
3305 }
3306 }
3307 break;
3308
3309 case Op_LoadVector:
3310 case Op_StoreVector:
3311 break;
3312
3313 case Op_AddReductionVI:
3314 case Op_AddReductionVL:
3315 case Op_AddReductionVF:
3316 case Op_AddReductionVD:
3317 case Op_MulReductionVI:
3318 case Op_MulReductionVL:
3319 case Op_MulReductionVF:
3320 case Op_MulReductionVD:
3321 break;
3322
3323 case Op_PackB:
3324 case Op_PackS:
3325 case Op_PackI:
3326 case Op_PackF:
3327 case Op_PackL:
3328 case Op_PackD:
3329 if (n->req()-1 > 2) {
3330 // Replace many operand PackNodes with a binary tree for matching
3331 PackNode* p = (PackNode*) n;
3332 Node* btp = p->binary_tree_pack(1, n->req());
3333 n->subsume_by(btp, this);
3334 }
3335 break;
3336 case Op_Loop:
3337 case Op_CountedLoop:
3338 case Op_OuterStripMinedLoop:
3339 if (n->as_Loop()->is_inner_loop()) {
3340 frc.inc_inner_loop_count();
3341 }
3342 n->as_Loop()->verify_strip_mined(0);
3343 break;
3344 case Op_LShiftI:
3345 case Op_RShiftI:
3346 case Op_URShiftI:
3347 case Op_LShiftL:
3348 case Op_RShiftL:
3349 case Op_URShiftL:
3350 if (Matcher::need_masked_shift_count) {
3351 // The cpu's shift instructions don't restrict the count to the
3352 // lower 5/6 bits. We need to do the masking ourselves.
3353 Node* in2 = n->in(2);
3354 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3355 const TypeInt* t = in2->find_int_type();
3356 if (t != NULL && t->is_con()) {
3357 juint shift = t->get_con();
3358 if (shift > mask) { // Unsigned cmp
3359 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3360 }
3361 } else {
3362 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3363 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3364 n->set_req(2, shift);
3365 }
3366 }
3367 if (in2->outcnt() == 0) { // Remove dead node
3368 in2->disconnect_inputs(NULL, this);
3369 }
3370 }
3371 break;
3372 case Op_MemBarStoreStore:
3373 case Op_MemBarRelease:
3374 // Break the link with AllocateNode: it is no longer useful and
3375 // confuses register allocation.
3376 if (n->req() > MemBarNode::Precedent) {
3377 n->set_req(MemBarNode::Precedent, top());
3378 }
3379 break;
3380 case Op_MemBarAcquire: {
3381 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3382 // At parse time, the trailing MemBarAcquire for a volatile load
3383 // is created with an edge to the load. After optimizations,
3384 // that input may be a chain of Phis. If those phis have no
3385 // other use, then the MemBarAcquire keeps them alive and
3386 // register allocation can be confused.
3387 ResourceMark rm;
3388 Unique_Node_List wq;
3389 wq.push(n->in(MemBarNode::Precedent));
3390 n->set_req(MemBarNode::Precedent, top());
3391 while (wq.size() > 0) {
3392 Node* m = wq.pop();
3393 if (m->outcnt() == 0) {
3394 for (uint j = 0; j < m->req(); j++) {
3395 Node* in = m->in(j);
3396 if (in != NULL) {
3397 wq.push(in);
3398 }
3399 }
3400 m->disconnect_inputs(NULL, this);
3401 }
3402 }
3403 }
3404 break;
3405 }
3406 case Op_RangeCheck: {
3407 RangeCheckNode* rc = n->as_RangeCheck();
3408 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3409 n->subsume_by(iff, this);
3410 frc._tests.push(iff);
3411 break;
3412 }
3413 case Op_ConvI2L: {
3414 if (!Matcher::convi2l_type_required) {
3415 // Code generation on some platforms doesn't need accurate
3416 // ConvI2L types. Widening the type can help remove redundant
3417 // address computations.
3418 n->as_Type()->set_type(TypeLong::INT);
3419 ResourceMark rm;
3420 Node_List wq;
3421 wq.push(n);
3422 for (uint next = 0; next < wq.size(); next++) {
3423 Node *m = wq.at(next);
3424
3425 for(;;) {
3426 // Loop over all nodes with identical inputs edges as m
3427 Node* k = m->find_similar(m->Opcode());
3428 if (k == NULL) {
3429 break;
3430 }
3431 // Push their uses so we get a chance to remove node made
3432 // redundant
3433 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3434 Node* u = k->fast_out(i);
3435 assert(!wq.contains(u), "shouldn't process one node several times");
3436 if (u->Opcode() == Op_LShiftL ||
3437 u->Opcode() == Op_AddL ||
3438 u->Opcode() == Op_SubL ||
3439 u->Opcode() == Op_AddP) {
3440 wq.push(u);
3441 }
3442 }
3443 // Replace all nodes with identical edges as m with m
3444 k->subsume_by(m, this);
3445 }
3446 }
3447 }
3448 break;
3449 }
3450 case Op_CmpUL: {
3451 if (!Matcher::has_match_rule(Op_CmpUL)) {
3452 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3453 // to max_julong if < 0 to make the signed comparison fail.
3454 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3455 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3456 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3457 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3458 Node* andl = new AndLNode(orl, remove_sign_mask);
3459 Node* cmp = new CmpLNode(andl, n->in(2));
3460 n->subsume_by(cmp, this);
3461 }
3462 break;
3463 }
3464 default:
3465 assert( !n->is_Call(), "" );
3466 assert( !n->is_Mem(), "" );
3467 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3468 break;
3469 }
3470
3471 // Collect CFG split points
3472 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3473 frc._tests.push(n);
3474 }
3475 }
3476
3477 //------------------------------final_graph_reshaping_walk---------------------
3478 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3479 // requires that the walk visits a node's inputs before visiting the node.
3480 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3481 ResourceArea *area = Thread::current()->resource_area();
3482 Unique_Node_List sfpt(area);
3483
3484 frc._visited.set(root->_idx); // first, mark node as visited
3485 uint cnt = root->req();
3486 Node *n = root;
3487 uint i = 0;
3488 while (true) {
3489 if (i < cnt) {
3490 // Place all non-visited non-null inputs onto stack
3491 Node* m = n->in(i);
3492 ++i;
3493 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3494 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3495 // compute worst case interpreter size in case of a deoptimization
3496 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3497
3498 sfpt.push(m);
3499 }
3500 cnt = m->req();
3501 nstack.push(n, i); // put on stack parent and next input's index
3502 n = m;
3503 i = 0;
3504 }
3505 } else {
3506 // Now do post-visit work
3507 final_graph_reshaping_impl( n, frc );
3508 if (nstack.is_empty())
3509 break; // finished
3510 n = nstack.node(); // Get node from stack
3511 cnt = n->req();
3512 i = nstack.index();
3513 nstack.pop(); // Shift to the next node on stack
3514 }
3515 }
3516
3517 // Skip next transformation if compressed oops are not used.
3518 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3519 (!UseCompressedOops && !UseCompressedClassPointers))
3520 return;
3521
3522 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3523 // It could be done for an uncommon traps or any safepoints/calls
3524 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3525 while (sfpt.size() > 0) {
3526 n = sfpt.pop();
3527 JVMState *jvms = n->as_SafePoint()->jvms();
3528 assert(jvms != NULL, "sanity");
3529 int start = jvms->debug_start();
3530 int end = n->req();
3531 bool is_uncommon = (n->is_CallStaticJava() &&
3532 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3533 for (int j = start; j < end; j++) {
3534 Node* in = n->in(j);
3535 if (in->is_DecodeNarrowPtr()) {
3536 bool safe_to_skip = true;
3537 if (!is_uncommon ) {
3538 // Is it safe to skip?
3539 for (uint i = 0; i < in->outcnt(); i++) {
3540 Node* u = in->raw_out(i);
3541 if (!u->is_SafePoint() ||
3542 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3543 safe_to_skip = false;
3544 }
3545 }
3546 }
3547 if (safe_to_skip) {
3548 n->set_req(j, in->in(1));
3549 }
3550 if (in->outcnt() == 0) {
3551 in->disconnect_inputs(NULL, this);
3552 }
3553 }
3554 }
3555 }
3556 }
3557
3558 //------------------------------final_graph_reshaping--------------------------
3559 // Final Graph Reshaping.
3560 //
3561 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3562 // and not commoned up and forced early. Must come after regular
3563 // optimizations to avoid GVN undoing the cloning. Clone constant
3564 // inputs to Loop Phis; these will be split by the allocator anyways.
3565 // Remove Opaque nodes.
3566 // (2) Move last-uses by commutative operations to the left input to encourage
3567 // Intel update-in-place two-address operations and better register usage
3568 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3569 // calls canonicalizing them back.
3570 // (3) Count the number of double-precision FP ops, single-precision FP ops
3571 // and call sites. On Intel, we can get correct rounding either by
3572 // forcing singles to memory (requires extra stores and loads after each
3573 // FP bytecode) or we can set a rounding mode bit (requires setting and
3574 // clearing the mode bit around call sites). The mode bit is only used
3575 // if the relative frequency of single FP ops to calls is low enough.
3576 // This is a key transform for SPEC mpeg_audio.
3577 // (4) Detect infinite loops; blobs of code reachable from above but not
3578 // below. Several of the Code_Gen algorithms fail on such code shapes,
3579 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3580 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3581 // Detection is by looking for IfNodes where only 1 projection is
3582 // reachable from below or CatchNodes missing some targets.
3583 // (5) Assert for insane oop offsets in debug mode.
3584
3585 bool Compile::final_graph_reshaping() {
3586 // an infinite loop may have been eliminated by the optimizer,
3587 // in which case the graph will be empty.
3588 if (root()->req() == 1) {
3589 record_method_not_compilable("trivial infinite loop");
3590 return true;
3591 }
3592
3593 // Expensive nodes have their control input set to prevent the GVN
3594 // from freely commoning them. There's no GVN beyond this point so
3595 // no need to keep the control input. We want the expensive nodes to
3596 // be freely moved to the least frequent code path by gcm.
3597 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3598 for (int i = 0; i < expensive_count(); i++) {
3599 _expensive_nodes->at(i)->set_req(0, NULL);
3600 }
3601
3602 Final_Reshape_Counts frc;
3603
3604 // Visit everybody reachable!
3605 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3606 Node_Stack nstack(live_nodes() >> 1);
3607 final_graph_reshaping_walk(nstack, root(), frc);
3608
3609 // Check for unreachable (from below) code (i.e., infinite loops).
3610 for( uint i = 0; i < frc._tests.size(); i++ ) {
3611 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3612 // Get number of CFG targets.
3613 // Note that PCTables include exception targets after calls.
3614 uint required_outcnt = n->required_outcnt();
3615 if (n->outcnt() != required_outcnt) {
3616 // Check for a few special cases. Rethrow Nodes never take the
3617 // 'fall-thru' path, so expected kids is 1 less.
3618 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3619 if (n->in(0)->in(0)->is_Call()) {
3620 CallNode *call = n->in(0)->in(0)->as_Call();
3621 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3622 required_outcnt--; // Rethrow always has 1 less kid
3623 } else if (call->req() > TypeFunc::Parms &&
3624 call->is_CallDynamicJava()) {
3625 // Check for null receiver. In such case, the optimizer has
3626 // detected that the virtual call will always result in a null
3627 // pointer exception. The fall-through projection of this CatchNode
3628 // will not be populated.
3629 Node *arg0 = call->in(TypeFunc::Parms);
3630 if (arg0->is_Type() &&
3631 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3632 required_outcnt--;
3633 }
3634 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3635 call->req() > TypeFunc::Parms+1 &&
3636 call->is_CallStaticJava()) {
3637 // Check for negative array length. In such case, the optimizer has
3638 // detected that the allocation attempt will always result in an
3639 // exception. There is no fall-through projection of this CatchNode .
3640 Node *arg1 = call->in(TypeFunc::Parms+1);
3641 if (arg1->is_Type() &&
3642 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3643 required_outcnt--;
3644 }
3645 }
3646 }
3647 }
3648 // Recheck with a better notion of 'required_outcnt'
3649 if (n->outcnt() != required_outcnt) {
3650 record_method_not_compilable("malformed control flow");
3651 return true; // Not all targets reachable!
3652 }
3653 }
3654 // Check that I actually visited all kids. Unreached kids
3655 // must be infinite loops.
3656 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3657 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3658 record_method_not_compilable("infinite loop");
3659 return true; // Found unvisited kid; must be unreach
3660 }
3661
3662 // Here so verification code in final_graph_reshaping_walk()
3663 // always see an OuterStripMinedLoopEnd
3664 if (n->is_OuterStripMinedLoopEnd()) {
3665 IfNode* init_iff = n->as_If();
3666 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3667 n->subsume_by(iff, this);
3668 }
3669 }
3670
3671 // If original bytecodes contained a mixture of floats and doubles
3672 // check if the optimizer has made it homogenous, item (3).
3673 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3674 frc.get_float_count() > 32 &&
3675 frc.get_double_count() == 0 &&
3676 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3677 set_24_bit_selection_and_mode( false, true );
3678 }
3679
3680 set_java_calls(frc.get_java_call_count());
3681 set_inner_loops(frc.get_inner_loop_count());
3682
3683 // No infinite loops, no reason to bail out.
3684 return false;
3685 }
3686
3687 //-----------------------------too_many_traps----------------------------------
3688 // Report if there are too many traps at the current method and bci.
3689 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3690 bool Compile::too_many_traps(ciMethod* method,
3691 int bci,
3692 Deoptimization::DeoptReason reason) {
3693 ciMethodData* md = method->method_data();
3694 if (md->is_empty()) {
3695 // Assume the trap has not occurred, or that it occurred only
3696 // because of a transient condition during start-up in the interpreter.
3697 return false;
3698 }
3699 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3700 if (md->has_trap_at(bci, m, reason) != 0) {
3701 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3702 // Also, if there are multiple reasons, or if there is no per-BCI record,
3703 // assume the worst.
3704 if (log())
3705 log()->elem("observe trap='%s' count='%d'",
3706 Deoptimization::trap_reason_name(reason),
3707 md->trap_count(reason));
3708 return true;
3709 } else {
3710 // Ignore method/bci and see if there have been too many globally.
3711 return too_many_traps(reason, md);
3712 }
3713 }
3714
3715 // Less-accurate variant which does not require a method and bci.
3716 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3717 ciMethodData* logmd) {
3718 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3719 // Too many traps globally.
3720 // Note that we use cumulative trap_count, not just md->trap_count.
3721 if (log()) {
3722 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3723 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3724 Deoptimization::trap_reason_name(reason),
3725 mcount, trap_count(reason));
3726 }
3727 return true;
3728 } else {
3729 // The coast is clear.
3730 return false;
3731 }
3732 }
3733
3734 //--------------------------too_many_recompiles--------------------------------
3735 // Report if there are too many recompiles at the current method and bci.
3736 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3737 // Is not eager to return true, since this will cause the compiler to use
3738 // Action_none for a trap point, to avoid too many recompilations.
3739 bool Compile::too_many_recompiles(ciMethod* method,
3740 int bci,
3741 Deoptimization::DeoptReason reason) {
3742 ciMethodData* md = method->method_data();
3743 if (md->is_empty()) {
3744 // Assume the trap has not occurred, or that it occurred only
3745 // because of a transient condition during start-up in the interpreter.
3746 return false;
3747 }
3748 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3749 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3750 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3751 Deoptimization::DeoptReason per_bc_reason
3752 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3753 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3754 if ((per_bc_reason == Deoptimization::Reason_none
3755 || md->has_trap_at(bci, m, reason) != 0)
3756 // The trap frequency measure we care about is the recompile count:
3757 && md->trap_recompiled_at(bci, m)
3758 && md->overflow_recompile_count() >= bc_cutoff) {
3759 // Do not emit a trap here if it has already caused recompilations.
3760 // Also, if there are multiple reasons, or if there is no per-BCI record,
3761 // assume the worst.
3762 if (log())
3763 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3764 Deoptimization::trap_reason_name(reason),
3765 md->trap_count(reason),
3766 md->overflow_recompile_count());
3767 return true;
3768 } else if (trap_count(reason) != 0
3769 && decompile_count() >= m_cutoff) {
3770 // Too many recompiles globally, and we have seen this sort of trap.
3771 // Use cumulative decompile_count, not just md->decompile_count.
3772 if (log())
3773 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3774 Deoptimization::trap_reason_name(reason),
3775 md->trap_count(reason), trap_count(reason),
3776 md->decompile_count(), decompile_count());
3777 return true;
3778 } else {
3779 // The coast is clear.
3780 return false;
3781 }
3782 }
3783
3784 // Compute when not to trap. Used by matching trap based nodes and
3785 // NullCheck optimization.
3786 void Compile::set_allowed_deopt_reasons() {
3787 _allowed_reasons = 0;
3788 if (is_method_compilation()) {
3789 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3790 assert(rs < BitsPerInt, "recode bit map");
3791 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3792 _allowed_reasons |= nth_bit(rs);
3793 }
3794 }
3795 }
3796 }
3797
3798 #ifndef PRODUCT
3799 //------------------------------verify_graph_edges---------------------------
3800 // Walk the Graph and verify that there is a one-to-one correspondence
3801 // between Use-Def edges and Def-Use edges in the graph.
3802 void Compile::verify_graph_edges(bool no_dead_code) {
3803 if (VerifyGraphEdges) {
3804 ResourceArea *area = Thread::current()->resource_area();
3805 Unique_Node_List visited(area);
3806 // Call recursive graph walk to check edges
3807 _root->verify_edges(visited);
3808 if (no_dead_code) {
3809 // Now make sure that no visited node is used by an unvisited node.
3810 bool dead_nodes = false;
3811 Unique_Node_List checked(area);
3812 while (visited.size() > 0) {
3813 Node* n = visited.pop();
3814 checked.push(n);
3815 for (uint i = 0; i < n->outcnt(); i++) {
3816 Node* use = n->raw_out(i);
3817 if (checked.member(use)) continue; // already checked
3818 if (visited.member(use)) continue; // already in the graph
3819 if (use->is_Con()) continue; // a dead ConNode is OK
3820 // At this point, we have found a dead node which is DU-reachable.
3821 if (!dead_nodes) {
3822 tty->print_cr("*** Dead nodes reachable via DU edges:");
3823 dead_nodes = true;
3824 }
3825 use->dump(2);
3826 tty->print_cr("---");
3827 checked.push(use); // No repeats; pretend it is now checked.
3828 }
3829 }
3830 assert(!dead_nodes, "using nodes must be reachable from root");
3831 }
3832 }
3833 }
3834
3835 // Verify GC barriers consistency
3836 // Currently supported:
3837 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3838 void Compile::verify_barriers() {
3839 #if INCLUDE_G1GC
3840 if (UseG1GC) {
3841 // Verify G1 pre-barriers
3842 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
3843
3844 ResourceArea *area = Thread::current()->resource_area();
3845 Unique_Node_List visited(area);
3846 Node_List worklist(area);
3847 // We're going to walk control flow backwards starting from the Root
3848 worklist.push(_root);
3849 while (worklist.size() > 0) {
3850 Node* x = worklist.pop();
3851 if (x == NULL || x == top()) continue;
3852 if (visited.member(x)) {
3853 continue;
3854 } else {
3855 visited.push(x);
3856 }
3857
3858 if (x->is_Region()) {
3859 for (uint i = 1; i < x->req(); i++) {
3860 worklist.push(x->in(i));
3861 }
3862 } else {
3863 worklist.push(x->in(0));
3864 // We are looking for the pattern:
3865 // /->ThreadLocal
3866 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3867 // \->ConI(0)
3868 // We want to verify that the If and the LoadB have the same control
3869 // See GraphKit::g1_write_barrier_pre()
3870 if (x->is_If()) {
3871 IfNode *iff = x->as_If();
3872 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3873 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3874 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3875 && cmp->in(1)->is_Load()) {
3876 LoadNode* load = cmp->in(1)->as_Load();
3877 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3878 && load->in(2)->in(3)->is_Con()
3879 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3880
3881 Node* if_ctrl = iff->in(0);
3882 Node* load_ctrl = load->in(0);
3883
3884 if (if_ctrl != load_ctrl) {
3885 // Skip possible CProj->NeverBranch in infinite loops
3886 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3887 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3888 if_ctrl = if_ctrl->in(0)->in(0);
3889 }
3890 }
3891 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3892 }
3893 }
3894 }
3895 }
3896 }
3897 }
3898 }
3899 #endif
3900 }
3901
3902 #endif
3903
3904 // The Compile object keeps track of failure reasons separately from the ciEnv.
3905 // This is required because there is not quite a 1-1 relation between the
3906 // ciEnv and its compilation task and the Compile object. Note that one
3907 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3908 // to backtrack and retry without subsuming loads. Other than this backtracking
3909 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3910 // by the logic in C2Compiler.
3911 void Compile::record_failure(const char* reason) {
3912 if (log() != NULL) {
3913 log()->elem("failure reason='%s' phase='compile'", reason);
3914 }
3915 if (_failure_reason == NULL) {
3916 // Record the first failure reason.
3917 _failure_reason = reason;
3918 }
3919
3920 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3921 C->print_method(PHASE_FAILURE);
3922 }
3923 _root = NULL; // flush the graph, too
3924 }
3925
3926 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3927 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3928 _phase_name(name), _dolog(CITimeVerbose)
3929 {
3930 if (_dolog) {
3931 C = Compile::current();
3932 _log = C->log();
3933 } else {
3934 C = NULL;
3935 _log = NULL;
3936 }
3937 if (_log != NULL) {
3938 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3939 _log->stamp();
3940 _log->end_head();
3941 }
3942 }
3943
3944 Compile::TracePhase::~TracePhase() {
3945
3946 C = Compile::current();
3947 if (_dolog) {
3948 _log = C->log();
3949 } else {
3950 _log = NULL;
3951 }
3952
3953 #ifdef ASSERT
3954 if (PrintIdealNodeCount) {
3955 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3956 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3957 }
3958
3959 if (VerifyIdealNodeCount) {
3960 Compile::current()->print_missing_nodes();
3961 }
3962 #endif
3963
3964 if (_log != NULL) {
3965 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3966 }
3967 }
3968
3969 //=============================================================================
3970 // Two Constant's are equal when the type and the value are equal.
3971 bool Compile::Constant::operator==(const Constant& other) {
3972 if (type() != other.type() ) return false;
3973 if (can_be_reused() != other.can_be_reused()) return false;
3974 // For floating point values we compare the bit pattern.
3975 switch (type()) {
3976 case T_INT:
3977 case T_FLOAT: return (_v._value.i == other._v._value.i);
3978 case T_LONG:
3979 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3980 case T_OBJECT:
3981 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3982 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3983 case T_METADATA: return (_v._metadata == other._v._metadata);
3984 default: ShouldNotReachHere(); return false;
3985 }
3986 }
3987
3988 static int type_to_size_in_bytes(BasicType t) {
3989 switch (t) {
3990 case T_INT: return sizeof(jint );
3991 case T_LONG: return sizeof(jlong );
3992 case T_FLOAT: return sizeof(jfloat );
3993 case T_DOUBLE: return sizeof(jdouble);
3994 case T_METADATA: return sizeof(Metadata*);
3995 // We use T_VOID as marker for jump-table entries (labels) which
3996 // need an internal word relocation.
3997 case T_VOID:
3998 case T_ADDRESS:
3999 case T_OBJECT: return sizeof(jobject);
4000 default:
4001 ShouldNotReachHere();
4002 return -1;
4003 }
4004 }
4005
4006 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4007 // sort descending
4008 if (a->freq() > b->freq()) return -1;
4009 if (a->freq() < b->freq()) return 1;
4010 return 0;
4011 }
4012
4013 void Compile::ConstantTable::calculate_offsets_and_size() {
4014 // First, sort the array by frequencies.
4015 _constants.sort(qsort_comparator);
4016
4017 #ifdef ASSERT
4018 // Make sure all jump-table entries were sorted to the end of the
4019 // array (they have a negative frequency).
4020 bool found_void = false;
4021 for (int i = 0; i < _constants.length(); i++) {
4022 Constant con = _constants.at(i);
4023 if (con.type() == T_VOID)
4024 found_void = true; // jump-tables
4025 else
4026 assert(!found_void, "wrong sorting");
4027 }
4028 #endif
4029
4030 int offset = 0;
4031 for (int i = 0; i < _constants.length(); i++) {
4032 Constant* con = _constants.adr_at(i);
4033
4034 // Align offset for type.
4035 int typesize = type_to_size_in_bytes(con->type());
4036 offset = align_up(offset, typesize);
4037 con->set_offset(offset); // set constant's offset
4038
4039 if (con->type() == T_VOID) {
4040 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4041 offset = offset + typesize * n->outcnt(); // expand jump-table
4042 } else {
4043 offset = offset + typesize;
4044 }
4045 }
4046
4047 // Align size up to the next section start (which is insts; see
4048 // CodeBuffer::align_at_start).
4049 assert(_size == -1, "already set?");
4050 _size = align_up(offset, (int)CodeEntryAlignment);
4051 }
4052
4053 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4054 MacroAssembler _masm(&cb);
4055 for (int i = 0; i < _constants.length(); i++) {
4056 Constant con = _constants.at(i);
4057 address constant_addr = NULL;
4058 switch (con.type()) {
4059 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4060 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4061 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4062 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4063 case T_OBJECT: {
4064 jobject obj = con.get_jobject();
4065 int oop_index = _masm.oop_recorder()->find_index(obj);
4066 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4067 break;
4068 }
4069 case T_ADDRESS: {
4070 address addr = (address) con.get_jobject();
4071 constant_addr = _masm.address_constant(addr);
4072 break;
4073 }
4074 // We use T_VOID as marker for jump-table entries (labels) which
4075 // need an internal word relocation.
4076 case T_VOID: {
4077 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4078 // Fill the jump-table with a dummy word. The real value is
4079 // filled in later in fill_jump_table.
4080 address dummy = (address) n;
4081 constant_addr = _masm.address_constant(dummy);
4082 // Expand jump-table
4083 for (uint i = 1; i < n->outcnt(); i++) {
4084 address temp_addr = _masm.address_constant(dummy + i);
4085 assert(temp_addr, "consts section too small");
4086 }
4087 break;
4088 }
4089 case T_METADATA: {
4090 Metadata* obj = con.get_metadata();
4091 int metadata_index = _masm.oop_recorder()->find_index(obj);
4092 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4093 break;
4094 }
4095 default: ShouldNotReachHere();
4096 }
4097 assert(constant_addr, "consts section too small");
4098 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4099 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4100 }
4101 }
4102
4103 int Compile::ConstantTable::find_offset(Constant& con) const {
4104 int idx = _constants.find(con);
4105 guarantee(idx != -1, "constant must be in constant table");
4106 int offset = _constants.at(idx).offset();
4107 guarantee(offset != -1, "constant table not emitted yet?");
4108 return offset;
4109 }
4110
4111 void Compile::ConstantTable::add(Constant& con) {
4112 if (con.can_be_reused()) {
4113 int idx = _constants.find(con);
4114 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4115 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4116 return;
4117 }
4118 }
4119 (void) _constants.append(con);
4120 }
4121
4122 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4123 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4124 Constant con(type, value, b->_freq);
4125 add(con);
4126 return con;
4127 }
4128
4129 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4130 Constant con(metadata);
4131 add(con);
4132 return con;
4133 }
4134
4135 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4136 jvalue value;
4137 BasicType type = oper->type()->basic_type();
4138 switch (type) {
4139 case T_LONG: value.j = oper->constantL(); break;
4140 case T_FLOAT: value.f = oper->constantF(); break;
4141 case T_DOUBLE: value.d = oper->constantD(); break;
4142 case T_OBJECT:
4143 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4144 case T_METADATA: return add((Metadata*)oper->constant()); break;
4145 default: guarantee(false, "unhandled type: %s", type2name(type));
4146 }
4147 return add(n, type, value);
4148 }
4149
4150 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4151 jvalue value;
4152 // We can use the node pointer here to identify the right jump-table
4153 // as this method is called from Compile::Fill_buffer right before
4154 // the MachNodes are emitted and the jump-table is filled (means the
4155 // MachNode pointers do not change anymore).
4156 value.l = (jobject) n;
4157 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4158 add(con);
4159 return con;
4160 }
4161
4162 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4163 // If called from Compile::scratch_emit_size do nothing.
4164 if (Compile::current()->in_scratch_emit_size()) return;
4165
4166 assert(labels.is_nonempty(), "must be");
4167 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4168
4169 // Since MachConstantNode::constant_offset() also contains
4170 // table_base_offset() we need to subtract the table_base_offset()
4171 // to get the plain offset into the constant table.
4172 int offset = n->constant_offset() - table_base_offset();
4173
4174 MacroAssembler _masm(&cb);
4175 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4176
4177 for (uint i = 0; i < n->outcnt(); i++) {
4178 address* constant_addr = &jump_table_base[i];
4179 assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i));
4180 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4181 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4182 }
4183 }
4184
4185 //----------------------------static_subtype_check-----------------------------
4186 // Shortcut important common cases when superklass is exact:
4187 // (0) superklass is java.lang.Object (can occur in reflective code)
4188 // (1) subklass is already limited to a subtype of superklass => always ok
4189 // (2) subklass does not overlap with superklass => always fail
4190 // (3) superklass has NO subtypes and we can check with a simple compare.
4191 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4192 if (StressReflectiveCode) {
4193 return SSC_full_test; // Let caller generate the general case.
4194 }
4195
4196 if (superk == env()->Object_klass()) {
4197 return SSC_always_true; // (0) this test cannot fail
4198 }
4199
4200 ciType* superelem = superk;
4201 if (superelem->is_array_klass())
4202 superelem = superelem->as_array_klass()->base_element_type();
4203
4204 if (!subk->is_interface()) { // cannot trust static interface types yet
4205 if (subk->is_subtype_of(superk)) {
4206 return SSC_always_true; // (1) false path dead; no dynamic test needed
4207 }
4208 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4209 !superk->is_subtype_of(subk)) {
4210 return SSC_always_false;
4211 }
4212 }
4213
4214 // If casting to an instance klass, it must have no subtypes
4215 if (superk->is_interface()) {
4216 // Cannot trust interfaces yet.
4217 // %%% S.B. superk->nof_implementors() == 1
4218 } else if (superelem->is_instance_klass()) {
4219 ciInstanceKlass* ik = superelem->as_instance_klass();
4220 if (!ik->has_subklass() && !ik->is_interface()) {
4221 if (!ik->is_final()) {
4222 // Add a dependency if there is a chance of a later subclass.
4223 dependencies()->assert_leaf_type(ik);
4224 }
4225 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4226 }
4227 } else {
4228 // A primitive array type has no subtypes.
4229 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4230 }
4231
4232 return SSC_full_test;
4233 }
4234
4235 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4236 #ifdef _LP64
4237 // The scaled index operand to AddP must be a clean 64-bit value.
4238 // Java allows a 32-bit int to be incremented to a negative
4239 // value, which appears in a 64-bit register as a large
4240 // positive number. Using that large positive number as an
4241 // operand in pointer arithmetic has bad consequences.
4242 // On the other hand, 32-bit overflow is rare, and the possibility
4243 // can often be excluded, if we annotate the ConvI2L node with
4244 // a type assertion that its value is known to be a small positive
4245 // number. (The prior range check has ensured this.)
4246 // This assertion is used by ConvI2LNode::Ideal.
4247 int index_max = max_jint - 1; // array size is max_jint, index is one less
4248 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4249 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4250 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4251 #endif
4252 return idx;
4253 }
4254
4255 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4256 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4257 if (ctrl != NULL) {
4258 // Express control dependency by a CastII node with a narrow type.
4259 value = new CastIINode(value, itype, false, true /* range check dependency */);
4260 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4261 // node from floating above the range check during loop optimizations. Otherwise, the
4262 // ConvI2L node may be eliminated independently of the range check, causing the data path
4263 // to become TOP while the control path is still there (although it's unreachable).
4264 value->set_req(0, ctrl);
4265 // Save CastII node to remove it after loop optimizations.
4266 phase->C->add_range_check_cast(value);
4267 value = phase->transform(value);
4268 }
4269 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4270 return phase->transform(new ConvI2LNode(value, ltype));
4271 }
4272
4273 // The message about the current inlining is accumulated in
4274 // _print_inlining_stream and transfered into the _print_inlining_list
4275 // once we know whether inlining succeeds or not. For regular
4276 // inlining, messages are appended to the buffer pointed by
4277 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4278 // a new buffer is added after _print_inlining_idx in the list. This
4279 // way we can update the inlining message for late inlining call site
4280 // when the inlining is attempted again.
4281 void Compile::print_inlining_init() {
4282 if (print_inlining() || print_intrinsics()) {
4283 _print_inlining_stream = new stringStream();
4284 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4285 }
4286 }
4287
4288 void Compile::print_inlining_reinit() {
4289 if (print_inlining() || print_intrinsics()) {
4290 // Re allocate buffer when we change ResourceMark
4291 _print_inlining_stream = new stringStream();
4292 }
4293 }
4294
4295 void Compile::print_inlining_reset() {
4296 _print_inlining_stream->reset();
4297 }
4298
4299 void Compile::print_inlining_commit() {
4300 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4301 // Transfer the message from _print_inlining_stream to the current
4302 // _print_inlining_list buffer and clear _print_inlining_stream.
4303 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4304 print_inlining_reset();
4305 }
4306
4307 void Compile::print_inlining_push() {
4308 // Add new buffer to the _print_inlining_list at current position
4309 _print_inlining_idx++;
4310 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4311 }
4312
4313 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4314 return _print_inlining_list->at(_print_inlining_idx);
4315 }
4316
4317 void Compile::print_inlining_update(CallGenerator* cg) {
4318 if (print_inlining() || print_intrinsics()) {
4319 if (!cg->is_late_inline()) {
4320 if (print_inlining_current().cg() != NULL) {
4321 print_inlining_push();
4322 }
4323 print_inlining_commit();
4324 } else {
4325 if (print_inlining_current().cg() != cg &&
4326 (print_inlining_current().cg() != NULL ||
4327 print_inlining_current().ss()->size() != 0)) {
4328 print_inlining_push();
4329 }
4330 print_inlining_commit();
4331 print_inlining_current().set_cg(cg);
4332 }
4333 }
4334 }
4335
4336 void Compile::print_inlining_move_to(CallGenerator* cg) {
4337 // We resume inlining at a late inlining call site. Locate the
4338 // corresponding inlining buffer so that we can update it.
4339 if (print_inlining()) {
4340 for (int i = 0; i < _print_inlining_list->length(); i++) {
4341 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4342 _print_inlining_idx = i;
4343 return;
4344 }
4345 }
4346 ShouldNotReachHere();
4347 }
4348 }
4349
4350 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4351 if (print_inlining()) {
4352 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4353 assert(print_inlining_current().cg() == cg, "wrong entry");
4354 // replace message with new message
4355 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4356 print_inlining_commit();
4357 print_inlining_current().set_cg(cg);
4358 }
4359 }
4360
4361 void Compile::print_inlining_assert_ready() {
4362 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4363 }
4364
4365 void Compile::process_print_inlining() {
4366 bool do_print_inlining = print_inlining() || print_intrinsics();
4367 if (do_print_inlining || log() != NULL) {
4368 // Print inlining message for candidates that we couldn't inline
4369 // for lack of space
4370 for (int i = 0; i < _late_inlines.length(); i++) {
4371 CallGenerator* cg = _late_inlines.at(i);
4372 if (!cg->is_mh_late_inline()) {
4373 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4374 if (do_print_inlining) {
4375 cg->print_inlining_late(msg);
4376 }
4377 log_late_inline_failure(cg, msg);
4378 }
4379 }
4380 }
4381 if (do_print_inlining) {
4382 ResourceMark rm;
4383 stringStream ss;
4384 for (int i = 0; i < _print_inlining_list->length(); i++) {
4385 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4386 }
4387 size_t end = ss.size();
4388 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4389 strncpy(_print_inlining_output, ss.base(), end+1);
4390 _print_inlining_output[end] = 0;
4391 }
4392 }
4393
4394 void Compile::dump_print_inlining() {
4395 if (_print_inlining_output != NULL) {
4396 tty->print_raw(_print_inlining_output);
4397 }
4398 }
4399
4400 void Compile::log_late_inline(CallGenerator* cg) {
4401 if (log() != NULL) {
4402 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4403 cg->unique_id());
4404 JVMState* p = cg->call_node()->jvms();
4405 while (p != NULL) {
4406 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4407 p = p->caller();
4408 }
4409 log()->tail("late_inline");
4410 }
4411 }
4412
4413 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4414 log_late_inline(cg);
4415 if (log() != NULL) {
4416 log()->inline_fail(msg);
4417 }
4418 }
4419
4420 void Compile::log_inline_id(CallGenerator* cg) {
4421 if (log() != NULL) {
4422 // The LogCompilation tool needs a unique way to identify late
4423 // inline call sites. This id must be unique for this call site in
4424 // this compilation. Try to have it unique across compilations as
4425 // well because it can be convenient when grepping through the log
4426 // file.
4427 // Distinguish OSR compilations from others in case CICountOSR is
4428 // on.
4429 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4430 cg->set_unique_id(id);
4431 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4432 }
4433 }
4434
4435 void Compile::log_inline_failure(const char* msg) {
4436 if (C->log() != NULL) {
4437 C->log()->inline_fail(msg);
4438 }
4439 }
4440
4441
4442 // Dump inlining replay data to the stream.
4443 // Don't change thread state and acquire any locks.
4444 void Compile::dump_inline_data(outputStream* out) {
4445 InlineTree* inl_tree = ilt();
4446 if (inl_tree != NULL) {
4447 out->print(" inline %d", inl_tree->count());
4448 inl_tree->dump_replay_data(out);
4449 }
4450 }
4451
4452 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4453 if (n1->Opcode() < n2->Opcode()) return -1;
4454 else if (n1->Opcode() > n2->Opcode()) return 1;
4455
4456 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4457 for (uint i = 1; i < n1->req(); i++) {
4458 if (n1->in(i) < n2->in(i)) return -1;
4459 else if (n1->in(i) > n2->in(i)) return 1;
4460 }
4461
4462 return 0;
4463 }
4464
4465 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4466 Node* n1 = *n1p;
4467 Node* n2 = *n2p;
4468
4469 return cmp_expensive_nodes(n1, n2);
4470 }
4471
4472 void Compile::sort_expensive_nodes() {
4473 if (!expensive_nodes_sorted()) {
4474 _expensive_nodes->sort(cmp_expensive_nodes);
4475 }
4476 }
4477
4478 bool Compile::expensive_nodes_sorted() const {
4479 for (int i = 1; i < _expensive_nodes->length(); i++) {
4480 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4481 return false;
4482 }
4483 }
4484 return true;
4485 }
4486
4487 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4488 if (_expensive_nodes->length() == 0) {
4489 return false;
4490 }
4491
4492 assert(OptimizeExpensiveOps, "optimization off?");
4493
4494 // Take this opportunity to remove dead nodes from the list
4495 int j = 0;
4496 for (int i = 0; i < _expensive_nodes->length(); i++) {
4497 Node* n = _expensive_nodes->at(i);
4498 if (!n->is_unreachable(igvn)) {
4499 assert(n->is_expensive(), "should be expensive");
4500 _expensive_nodes->at_put(j, n);
4501 j++;
4502 }
4503 }
4504 _expensive_nodes->trunc_to(j);
4505
4506 // Then sort the list so that similar nodes are next to each other
4507 // and check for at least two nodes of identical kind with same data
4508 // inputs.
4509 sort_expensive_nodes();
4510
4511 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4512 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4513 return true;
4514 }
4515 }
4516
4517 return false;
4518 }
4519
4520 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4521 if (_expensive_nodes->length() == 0) {
4522 return;
4523 }
4524
4525 assert(OptimizeExpensiveOps, "optimization off?");
4526
4527 // Sort to bring similar nodes next to each other and clear the
4528 // control input of nodes for which there's only a single copy.
4529 sort_expensive_nodes();
4530
4531 int j = 0;
4532 int identical = 0;
4533 int i = 0;
4534 bool modified = false;
4535 for (; i < _expensive_nodes->length()-1; i++) {
4536 assert(j <= i, "can't write beyond current index");
4537 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4538 identical++;
4539 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4540 continue;
4541 }
4542 if (identical > 0) {
4543 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4544 identical = 0;
4545 } else {
4546 Node* n = _expensive_nodes->at(i);
4547 igvn.replace_input_of(n, 0, NULL);
4548 igvn.hash_insert(n);
4549 modified = true;
4550 }
4551 }
4552 if (identical > 0) {
4553 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4554 } else if (_expensive_nodes->length() >= 1) {
4555 Node* n = _expensive_nodes->at(i);
4556 igvn.replace_input_of(n, 0, NULL);
4557 igvn.hash_insert(n);
4558 modified = true;
4559 }
4560 _expensive_nodes->trunc_to(j);
4561 if (modified) {
4562 igvn.optimize();
4563 }
4564 }
4565
4566 void Compile::add_expensive_node(Node * n) {
4567 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4568 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4569 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4570 if (OptimizeExpensiveOps) {
4571 _expensive_nodes->append(n);
4572 } else {
4573 // Clear control input and let IGVN optimize expensive nodes if
4574 // OptimizeExpensiveOps is off.
4575 n->set_req(0, NULL);
4576 }
4577 }
4578
4579 /**
4580 * Remove the speculative part of types and clean up the graph
4581 */
4582 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4583 if (UseTypeSpeculation) {
4584 Unique_Node_List worklist;
4585 worklist.push(root());
4586 int modified = 0;
4587 // Go over all type nodes that carry a speculative type, drop the
4588 // speculative part of the type and enqueue the node for an igvn
4589 // which may optimize it out.
4590 for (uint next = 0; next < worklist.size(); ++next) {
4591 Node *n = worklist.at(next);
4592 if (n->is_Type()) {
4593 TypeNode* tn = n->as_Type();
4594 const Type* t = tn->type();
4595 const Type* t_no_spec = t->remove_speculative();
4596 if (t_no_spec != t) {
4597 bool in_hash = igvn.hash_delete(n);
4598 assert(in_hash, "node should be in igvn hash table");
4599 tn->set_type(t_no_spec);
4600 igvn.hash_insert(n);
4601 igvn._worklist.push(n); // give it a chance to go away
4602 modified++;
4603 }
4604 }
4605 uint max = n->len();
4606 for( uint i = 0; i < max; ++i ) {
4607 Node *m = n->in(i);
4608 if (not_a_node(m)) continue;
4609 worklist.push(m);
4610 }
4611 }
4612 // Drop the speculative part of all types in the igvn's type table
4613 igvn.remove_speculative_types();
4614 if (modified > 0) {
4615 igvn.optimize();
4616 }
4617 #ifdef ASSERT
4618 // Verify that after the IGVN is over no speculative type has resurfaced
4619 worklist.clear();
4620 worklist.push(root());
4621 for (uint next = 0; next < worklist.size(); ++next) {
4622 Node *n = worklist.at(next);
4623 const Type* t = igvn.type_or_null(n);
4624 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4625 if (n->is_Type()) {
4626 t = n->as_Type()->type();
4627 assert(t == t->remove_speculative(), "no more speculative types");
4628 }
4629 uint max = n->len();
4630 for( uint i = 0; i < max; ++i ) {
4631 Node *m = n->in(i);
4632 if (not_a_node(m)) continue;
4633 worklist.push(m);
4634 }
4635 }
4636 igvn.check_no_speculative_types();
4637 #endif
4638 }
4639 }
4640
4641 // Auxiliary method to support randomized stressing/fuzzing.
4642 //
4643 // This method can be called the arbitrary number of times, with current count
4644 // as the argument. The logic allows selecting a single candidate from the
4645 // running list of candidates as follows:
4646 // int count = 0;
4647 // Cand* selected = null;
4648 // while(cand = cand->next()) {
4649 // if (randomized_select(++count)) {
4650 // selected = cand;
4651 // }
4652 // }
4653 //
4654 // Including count equalizes the chances any candidate is "selected".
4655 // This is useful when we don't have the complete list of candidates to choose
4656 // from uniformly. In this case, we need to adjust the randomicity of the
4657 // selection, or else we will end up biasing the selection towards the latter
4658 // candidates.
4659 //
4660 // Quick back-envelope calculation shows that for the list of n candidates
4661 // the equal probability for the candidate to persist as "best" can be
4662 // achieved by replacing it with "next" k-th candidate with the probability
4663 // of 1/k. It can be easily shown that by the end of the run, the
4664 // probability for any candidate is converged to 1/n, thus giving the
4665 // uniform distribution among all the candidates.
4666 //
4667 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4668 #define RANDOMIZED_DOMAIN_POW 29
4669 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4670 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4671 bool Compile::randomized_select(int count) {
4672 assert(count > 0, "only positive");
4673 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4674 }
4675
4676 CloneMap& Compile::clone_map() { return _clone_map; }
4677 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4678
4679 void NodeCloneInfo::dump() const {
4680 tty->print(" {%d:%d} ", idx(), gen());
4681 }
4682
4683 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4684 uint64_t val = value(old->_idx);
4685 NodeCloneInfo cio(val);
4686 assert(val != 0, "old node should be in the map");
4687 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4688 insert(nnn->_idx, cin.get());
4689 #ifndef PRODUCT
4690 if (is_debug()) {
4691 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4692 }
4693 #endif
4694 }
4695
4696 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4697 NodeCloneInfo cio(value(old->_idx));
4698 if (cio.get() == 0) {
4699 cio.set(old->_idx, 0);
4700 insert(old->_idx, cio.get());
4701 #ifndef PRODUCT
4702 if (is_debug()) {
4703 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4704 }
4705 #endif
4706 }
4707 clone(old, nnn, gen);
4708 }
4709
4710 int CloneMap::max_gen() const {
4711 int g = 0;
4712 DictI di(_dict);
4713 for(; di.test(); ++di) {
4714 int t = gen(di._key);
4715 if (g < t) {
4716 g = t;
4717 #ifndef PRODUCT
4718 if (is_debug()) {
4719 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4720 }
4721 #endif
4722 }
4723 }
4724 return g;
4725 }
4726
4727 void CloneMap::dump(node_idx_t key) const {
4728 uint64_t val = value(key);
4729 if (val != 0) {
4730 NodeCloneInfo ni(val);
4731 ni.dump();
4732 }
4733 }