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