1 /*
2 * Copyright (c) 1997, 2019, 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/valuetypenode.hpp"
71 #include "opto/vectornode.hpp"
72 #include "runtime/arguments.hpp"
73 #include "runtime/sharedRuntime.hpp"
74 #include "runtime/signature.hpp"
75 #include "runtime/stubRoutines.hpp"
76 #include "runtime/timer.hpp"
77 #include "utilities/align.hpp"
78 #include "utilities/copy.hpp"
79 #include "utilities/macros.hpp"
80 #if INCLUDE_ZGC
81 #include "gc/z/c2/zBarrierSetC2.hpp"
82 #endif
83
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 class IntrinsicDescPair {
101 private:
102 ciMethod* _m;
103 bool _is_virtual;
104 public:
105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
107 ciMethod* m= elt->method();
108 ciMethod* key_m = key->_m;
109 if (key_m < m) return -1;
110 else if (key_m > m) return 1;
111 else {
112 bool is_virtual = elt->is_virtual();
113 bool key_virtual = key->_is_virtual;
114 if (key_virtual < is_virtual) return -1;
115 else if (key_virtual > is_virtual) return 1;
116 else return 0;
117 }
118 }
119 };
120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
121 #ifdef ASSERT
122 for (int i = 1; i < _intrinsics->length(); i++) {
123 CallGenerator* cg1 = _intrinsics->at(i-1);
124 CallGenerator* cg2 = _intrinsics->at(i);
125 assert(cg1->method() != cg2->method()
126 ? cg1->method() < cg2->method()
127 : cg1->is_virtual() < cg2->is_virtual(),
128 "compiler intrinsics list must stay sorted");
129 }
130 #endif
131 IntrinsicDescPair pair(m, is_virtual);
132 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
138 }
139 int len = _intrinsics->length();
140 bool found = false;
141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
142 assert(!found, "registering twice");
143 _intrinsics->insert_before(index, cg);
144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
145 }
146
147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
148 assert(m->is_loaded(), "don't try this on unloaded methods");
149 if (_intrinsics != NULL) {
150 bool found = false;
151 int index = intrinsic_insertion_index(m, is_virtual, found);
152 if (found) {
153 return _intrinsics->at(index);
154 }
155 }
156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
157 if (m->intrinsic_id() != vmIntrinsics::_none &&
158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
160 if (cg != NULL) {
161 // Save it for next time:
162 register_intrinsic(cg);
163 return cg;
164 } else {
165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
166 }
167 }
168 return NULL;
169 }
170
171 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
172 // in library_call.cpp.
173
174
175 #ifndef PRODUCT
176 // statistics gathering...
177
178 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
180
181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
182 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
183 int oflags = _intrinsic_hist_flags[id];
184 assert(flags != 0, "what happened?");
185 if (is_virtual) {
186 flags |= _intrinsic_virtual;
187 }
188 bool changed = (flags != oflags);
189 if ((flags & _intrinsic_worked) != 0) {
190 juint count = (_intrinsic_hist_count[id] += 1);
191 if (count == 1) {
192 changed = true; // first time
193 }
194 // increment the overall count also:
195 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
196 }
197 if (changed) {
198 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
199 // Something changed about the intrinsic's virtuality.
200 if ((flags & _intrinsic_virtual) != 0) {
201 // This is the first use of this intrinsic as a virtual call.
202 if (oflags != 0) {
203 // We already saw it as a non-virtual, so note both cases.
204 flags |= _intrinsic_both;
205 }
206 } else if ((oflags & _intrinsic_both) == 0) {
207 // This is the first use of this intrinsic as a non-virtual
208 flags |= _intrinsic_both;
209 }
210 }
211 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
212 }
213 // update the overall flags also:
214 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
215 return changed;
216 }
217
218 static char* format_flags(int flags, char* buf) {
219 buf[0] = 0;
220 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
221 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
222 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
223 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
224 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
225 if (buf[0] == 0) strcat(buf, ",");
226 assert(buf[0] == ',', "must be");
227 return &buf[1];
228 }
229
230 void Compile::print_intrinsic_statistics() {
231 char flagsbuf[100];
232 ttyLocker ttyl;
233 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
234 tty->print_cr("Compiler intrinsic usage:");
235 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
236 if (total == 0) total = 1; // avoid div0 in case of no successes
237 #define PRINT_STAT_LINE(name, c, f) \
238 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
239 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
240 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
241 int flags = _intrinsic_hist_flags[id];
242 juint count = _intrinsic_hist_count[id];
243 if ((flags | count) != 0) {
244 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
245 }
246 }
247 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
248 if (xtty != NULL) xtty->tail("statistics");
249 }
250
251 void Compile::print_statistics() {
252 { ttyLocker ttyl;
253 if (xtty != NULL) xtty->head("statistics type='opto'");
254 Parse::print_statistics();
255 PhaseCCP::print_statistics();
256 PhaseRegAlloc::print_statistics();
257 Scheduling::print_statistics();
258 PhasePeephole::print_statistics();
259 PhaseIdealLoop::print_statistics();
260 if (xtty != NULL) xtty->tail("statistics");
261 }
262 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
263 // put this under its own <statistics> element.
264 print_intrinsic_statistics();
265 }
266 }
267 #endif //PRODUCT
268
269 // Support for bundling info
270 Bundle* Compile::node_bundling(const Node *n) {
271 assert(valid_bundle_info(n), "oob");
272 return &_node_bundling_base[n->_idx];
273 }
274
275 bool Compile::valid_bundle_info(const Node *n) {
276 return (_node_bundling_limit > n->_idx);
277 }
278
279
280 void Compile::gvn_replace_by(Node* n, Node* nn) {
281 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
282 Node* use = n->last_out(i);
283 bool is_in_table = initial_gvn()->hash_delete(use);
284 uint uses_found = 0;
285 for (uint j = 0; j < use->len(); j++) {
286 if (use->in(j) == n) {
287 if (j < use->req())
288 use->set_req(j, nn);
289 else
290 use->set_prec(j, nn);
291 uses_found++;
292 }
293 }
294 if (is_in_table) {
295 // reinsert into table
296 initial_gvn()->hash_find_insert(use);
297 }
298 record_for_igvn(use);
299 i -= uses_found; // we deleted 1 or more copies of this edge
300 }
301 }
302
303
304 static inline bool not_a_node(const Node* n) {
305 if (n == NULL) return true;
306 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
307 if (*(address*)n == badAddress) return true; // kill by Node::destruct
308 return false;
309 }
310
311 // Identify all nodes that are reachable from below, useful.
312 // Use breadth-first pass that records state in a Unique_Node_List,
313 // recursive traversal is slower.
314 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
315 int estimated_worklist_size = live_nodes();
316 useful.map( estimated_worklist_size, NULL ); // preallocate space
317
318 // Initialize worklist
319 if (root() != NULL) { useful.push(root()); }
320 // If 'top' is cached, declare it useful to preserve cached node
321 if( cached_top_node() ) { useful.push(cached_top_node()); }
322
323 // Push all useful nodes onto the list, breadthfirst
324 for( uint next = 0; next < useful.size(); ++next ) {
325 assert( next < unique(), "Unique useful nodes < total nodes");
326 Node *n = useful.at(next);
327 uint max = n->len();
328 for( uint i = 0; i < max; ++i ) {
329 Node *m = n->in(i);
330 if (not_a_node(m)) continue;
331 useful.push(m);
332 }
333 }
334 }
335
336 // Update dead_node_list with any missing dead nodes using useful
337 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
338 void Compile::update_dead_node_list(Unique_Node_List &useful) {
339 uint max_idx = unique();
340 VectorSet& useful_node_set = useful.member_set();
341
342 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
343 // If node with index node_idx is not in useful set,
344 // mark it as dead in dead node list.
345 if (! useful_node_set.test(node_idx) ) {
346 record_dead_node(node_idx);
347 }
348 }
349 }
350
351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
352 int shift = 0;
353 for (int i = 0; i < inlines->length(); i++) {
354 CallGenerator* cg = inlines->at(i);
355 CallNode* call = cg->call_node();
356 if (shift > 0) {
357 inlines->at_put(i-shift, cg);
358 }
359 if (!useful.member(call)) {
360 shift++;
361 }
362 }
363 inlines->trunc_to(inlines->length()-shift);
364 }
365
366 // Disconnect all useless nodes by disconnecting those at the boundary.
367 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
368 uint next = 0;
369 while (next < useful.size()) {
370 Node *n = useful.at(next++);
371 if (n->is_SafePoint()) {
372 // We're done with a parsing phase. Replaced nodes are not valid
373 // beyond that point.
374 n->as_SafePoint()->delete_replaced_nodes();
375 }
376 // Use raw traversal of out edges since this code removes out edges
377 int max = n->outcnt();
378 for (int j = 0; j < max; ++j) {
379 Node* child = n->raw_out(j);
380 if (! useful.member(child)) {
381 assert(!child->is_top() || child != top(),
382 "If top is cached in Compile object it is in useful list");
383 // Only need to remove this out-edge to the useless node
384 n->raw_del_out(j);
385 --j;
386 --max;
387 }
388 }
389 if (n->outcnt() == 1 && n->has_special_unique_user()) {
390 record_for_igvn(n->unique_out());
391 }
392 }
393 // Remove useless macro and predicate opaq nodes
394 for (int i = C->macro_count()-1; i >= 0; i--) {
395 Node* n = C->macro_node(i);
396 if (!useful.member(n)) {
397 remove_macro_node(n);
398 }
399 }
400 // Remove useless CastII nodes with range check dependency
401 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
402 Node* cast = range_check_cast_node(i);
403 if (!useful.member(cast)) {
404 remove_range_check_cast(cast);
405 }
406 }
407 // Remove useless expensive nodes
408 for (int i = C->expensive_count()-1; i >= 0; i--) {
409 Node* n = C->expensive_node(i);
410 if (!useful.member(n)) {
411 remove_expensive_node(n);
412 }
413 }
414 // Remove useless Opaque4 nodes
415 for (int i = opaque4_count() - 1; i >= 0; i--) {
416 Node* opaq = opaque4_node(i);
417 if (!useful.member(opaq)) {
418 remove_opaque4_node(opaq);
419 }
420 }
421 // Remove useless value type nodes
422 if (_value_type_nodes != NULL) {
423 _value_type_nodes->remove_useless_nodes(useful.member_set());
424 }
425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
426 bs->eliminate_useless_gc_barriers(useful, this);
427 // clean up the late inline lists
428 remove_useless_late_inlines(&_string_late_inlines, useful);
429 remove_useless_late_inlines(&_boxing_late_inlines, useful);
430 remove_useless_late_inlines(&_late_inlines, useful);
431 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
432 }
433
434 //------------------------------frame_size_in_words-----------------------------
435 // frame_slots in units of words
436 int Compile::frame_size_in_words() const {
437 // shift is 0 in LP32 and 1 in LP64
438 const int shift = (LogBytesPerWord - LogBytesPerInt);
439 int words = _frame_slots >> shift;
440 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
441 return words;
442 }
443
444 // To bang the stack of this compiled method we use the stack size
445 // that the interpreter would need in case of a deoptimization. This
446 // removes the need to bang the stack in the deoptimization blob which
447 // in turn simplifies stack overflow handling.
448 int Compile::bang_size_in_bytes() const {
449 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
450 }
451
452 // ============================================================================
453 //------------------------------CompileWrapper---------------------------------
454 class CompileWrapper : public StackObj {
455 Compile *const _compile;
456 public:
457 CompileWrapper(Compile* compile);
458
459 ~CompileWrapper();
460 };
461
462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
463 // the Compile* pointer is stored in the current ciEnv:
464 ciEnv* env = compile->env();
465 assert(env == ciEnv::current(), "must already be a ciEnv active");
466 assert(env->compiler_data() == NULL, "compile already active?");
467 env->set_compiler_data(compile);
468 assert(compile == Compile::current(), "sanity");
469
470 compile->set_type_dict(NULL);
471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
472 compile->clone_map().set_clone_idx(0);
473 compile->set_type_last_size(0);
474 compile->set_last_tf(NULL, NULL);
475 compile->set_indexSet_arena(NULL);
476 compile->set_indexSet_free_block_list(NULL);
477 compile->init_type_arena();
478 Type::Initialize(compile);
479 _compile->set_scratch_buffer_blob(NULL);
480 _compile->begin_method();
481 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
482 }
483 CompileWrapper::~CompileWrapper() {
484 _compile->end_method();
485 if (_compile->scratch_buffer_blob() != NULL)
486 BufferBlob::free(_compile->scratch_buffer_blob());
487 _compile->env()->set_compiler_data(NULL);
488 }
489
490
491 //----------------------------print_compile_messages---------------------------
492 void Compile::print_compile_messages() {
493 #ifndef PRODUCT
494 // Check if recompiling
495 if (_subsume_loads == false && PrintOpto) {
496 // Recompiling without allowing machine instructions to subsume loads
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without subsuming loads **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
502 // Recompiling without escape analysis
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without escape analysis **");
505 tty->print_cr("*********************************************************");
506 }
507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
508 // Recompiling without boxing elimination
509 tty->print_cr("*********************************************************");
510 tty->print_cr("** Bailout: Recompile without boxing elimination **");
511 tty->print_cr("*********************************************************");
512 }
513 if (C->directive()->BreakAtCompileOption) {
514 // Open the debugger when compiling this method.
515 tty->print("### Breaking when compiling: ");
516 method()->print_short_name();
517 tty->cr();
518 BREAKPOINT;
519 }
520
521 if( PrintOpto ) {
522 if (is_osr_compilation()) {
523 tty->print("[OSR]%3d", _compile_id);
524 } else {
525 tty->print("%3d", _compile_id);
526 }
527 }
528 #endif
529 }
530
531
532 //-----------------------init_scratch_buffer_blob------------------------------
533 // Construct a temporary BufferBlob and cache it for this compile.
534 void Compile::init_scratch_buffer_blob(int const_size) {
535 // If there is already a scratch buffer blob allocated and the
536 // constant section is big enough, use it. Otherwise free the
537 // current and allocate a new one.
538 BufferBlob* blob = scratch_buffer_blob();
539 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
540 // Use the current blob.
541 } else {
542 if (blob != NULL) {
543 BufferBlob::free(blob);
544 }
545
546 ResourceMark rm;
547 _scratch_const_size = const_size;
548 int size = C2Compiler::initial_code_buffer_size(const_size);
549 #ifdef ASSERT
550 if (C->has_scalarized_args()) {
551 // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space
552 size += 5120;
553 }
554 #endif
555 blob = BufferBlob::create("Compile::scratch_buffer", size);
556 // Record the buffer blob for next time.
557 set_scratch_buffer_blob(blob);
558 // Have we run out of code space?
559 if (scratch_buffer_blob() == NULL) {
560 // Let CompilerBroker disable further compilations.
561 record_failure("Not enough space for scratch buffer in CodeCache");
562 return;
563 }
564 }
565
566 // Initialize the relocation buffers
567 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
568 set_scratch_locs_memory(locs_buf);
569 }
570
571
572 //-----------------------scratch_emit_size-------------------------------------
573 // Helper function that computes size by emitting code
574 uint Compile::scratch_emit_size(const Node* n) {
575 // Start scratch_emit_size section.
576 set_in_scratch_emit_size(true);
577
578 // Emit into a trash buffer and count bytes emitted.
579 // This is a pretty expensive way to compute a size,
580 // but it works well enough if seldom used.
581 // All common fixed-size instructions are given a size
582 // method by the AD file.
583 // Note that the scratch buffer blob and locs memory are
584 // allocated at the beginning of the compile task, and
585 // may be shared by several calls to scratch_emit_size.
586 // The allocation of the scratch buffer blob is particularly
587 // expensive, since it has to grab the code cache lock.
588 BufferBlob* blob = this->scratch_buffer_blob();
589 assert(blob != NULL, "Initialize BufferBlob at start");
590 assert(blob->size() > MAX_inst_size, "sanity");
591 relocInfo* locs_buf = scratch_locs_memory();
592 address blob_begin = blob->content_begin();
593 address blob_end = (address)locs_buf;
594 assert(blob->contains(blob_end), "sanity");
595 CodeBuffer buf(blob_begin, blob_end - blob_begin);
596 buf.initialize_consts_size(_scratch_const_size);
597 buf.initialize_stubs_size(MAX_stubs_size);
598 assert(locs_buf != NULL, "sanity");
599 int lsize = MAX_locs_size / 3;
600 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
601 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
602 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
603 // Mark as scratch buffer.
604 buf.consts()->set_scratch_emit();
605 buf.insts()->set_scratch_emit();
606 buf.stubs()->set_scratch_emit();
607
608 // Do the emission.
609
610 Label fakeL; // Fake label for branch instructions.
611 Label* saveL = NULL;
612 uint save_bnum = 0;
613 bool is_branch = n->is_MachBranch();
614 if (is_branch) {
615 MacroAssembler masm(&buf);
616 masm.bind(fakeL);
617 n->as_MachBranch()->save_label(&saveL, &save_bnum);
618 n->as_MachBranch()->label_set(&fakeL, 0);
619 } else if (n->is_MachProlog()) {
620 saveL = ((MachPrologNode*)n)->_verified_entry;
621 ((MachPrologNode*)n)->_verified_entry = &fakeL;
622 }
623 n->emit(buf, this->regalloc());
624
625 // Emitting into the scratch buffer should not fail
626 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
627
628 // Restore label.
629 if (is_branch) {
630 n->as_MachBranch()->label_set(saveL, save_bnum);
631 } else if (n->is_MachProlog()) {
632 ((MachPrologNode*)n)->_verified_entry = saveL;
633 }
634
635 // End scratch_emit_size section.
636 set_in_scratch_emit_size(false);
637
638 return buf.insts_size();
639 }
640
641
642 // ============================================================================
643 //------------------------------Compile standard-------------------------------
644 debug_only( int Compile::_debug_idx = 100000; )
645
646 // Compile a method. entry_bci is -1 for normal compilations and indicates
647 // the continuation bci for on stack replacement.
648
649
650 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
651 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
652 : Phase(Compiler),
653 _compile_id(ci_env->compile_id()),
654 _save_argument_registers(false),
655 _subsume_loads(subsume_loads),
656 _do_escape_analysis(do_escape_analysis),
657 _eliminate_boxing(eliminate_boxing),
658 _method(target),
659 _entry_bci(osr_bci),
660 _stub_function(NULL),
661 _stub_name(NULL),
662 _stub_entry_point(NULL),
663 _max_node_limit(MaxNodeLimit),
664 _orig_pc_slot(0),
665 _orig_pc_slot_offset_in_bytes(0),
666 _sp_inc_slot(0),
667 _sp_inc_slot_offset_in_bytes(0),
668 _inlining_progress(false),
669 _inlining_incrementally(false),
670 _do_cleanup(false),
671 _has_reserved_stack_access(target->has_reserved_stack_access()),
672 #ifndef PRODUCT
673 _trace_opto_output(directive->TraceOptoOutputOption),
674 #endif
675 _has_method_handle_invokes(false),
676 _clinit_barrier_on_entry(false),
677 _comp_arena(mtCompiler),
678 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
679 _env(ci_env),
680 _directive(directive),
681 _log(ci_env->log()),
682 _failure_reason(NULL),
683 _congraph(NULL),
684 #ifndef PRODUCT
685 _printer(IdealGraphPrinter::printer()),
686 #endif
687 _dead_node_list(comp_arena()),
688 _dead_node_count(0),
689 _node_arena(mtCompiler),
690 _old_arena(mtCompiler),
691 _mach_constant_base_node(NULL),
692 _Compile_types(mtCompiler),
693 _initial_gvn(NULL),
694 _for_igvn(NULL),
695 _warm_calls(NULL),
696 _late_inlines(comp_arena(), 2, 0, NULL),
697 _string_late_inlines(comp_arena(), 2, 0, NULL),
698 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
699 _late_inlines_pos(0),
700 _number_of_mh_late_inlines(0),
701 _print_inlining_stream(NULL),
702 _print_inlining_list(NULL),
703 _print_inlining_idx(0),
704 _print_inlining_output(NULL),
705 _replay_inline_data(NULL),
706 _java_calls(0),
707 _inner_loops(0),
708 _interpreter_frame_size(0),
709 _node_bundling_limit(0),
710 _node_bundling_base(NULL),
711 _code_buffer("Compile::Fill_buffer"),
712 _scratch_const_size(-1),
713 _in_scratch_emit_size(false)
714 #ifndef PRODUCT
715 , _in_dump_cnt(0)
716 #endif
717 {
718 C = this;
719 #ifndef PRODUCT
720 if (_printer != NULL) {
721 _printer->set_compile(this);
722 }
723 #endif
724 CompileWrapper cw(this);
725
726 if (CITimeVerbose) {
727 tty->print(" ");
728 target->holder()->name()->print();
729 tty->print(".");
730 target->print_short_name();
731 tty->print(" ");
732 }
733 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
734 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
735
736 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
737 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
738 // We can always print a disassembly, either abstract (hex dump) or
739 // with the help of a suitable hsdis library. Thus, we should not
740 // couple print_assembly and print_opto_assembly controls.
741 // But: always print opto and regular assembly on compile command 'print'.
742 bool print_assembly = directive->PrintAssemblyOption;
743 set_print_assembly(print_opto_assembly || print_assembly);
744 #else
745 set_print_assembly(false); // must initialize.
746 #endif
747
748 #ifndef PRODUCT
749 set_parsed_irreducible_loop(false);
750
751 if (directive->ReplayInlineOption) {
752 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
753 }
754 #endif
755 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
756 set_print_intrinsics(directive->PrintIntrinsicsOption);
757 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
758
759 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
760 // Make sure the method being compiled gets its own MDO,
761 // so we can at least track the decompile_count().
762 // Need MDO to record RTM code generation state.
763 method()->ensure_method_data();
764 }
765
766 Init(::AliasLevel);
767
768
769 print_compile_messages();
770
771 _ilt = InlineTree::build_inline_tree_root();
772
773 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
774 assert(num_alias_types() >= AliasIdxRaw, "");
775
776 #define MINIMUM_NODE_HASH 1023
777 // Node list that Iterative GVN will start with
778 Unique_Node_List for_igvn(comp_arena());
779 set_for_igvn(&for_igvn);
780
781 // GVN that will be run immediately on new nodes
782 uint estimated_size = method()->code_size()*4+64;
783 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
784 PhaseGVN gvn(node_arena(), estimated_size);
785 set_initial_gvn(&gvn);
786
787 print_inlining_init();
788 { // Scope for timing the parser
789 TracePhase tp("parse", &timers[_t_parser]);
790
791 // Put top into the hash table ASAP.
792 initial_gvn()->transform_no_reclaim(top());
793
794 // Set up tf(), start(), and find a CallGenerator.
795 CallGenerator* cg = NULL;
796 if (is_osr_compilation()) {
797 const TypeTuple *domain = StartOSRNode::osr_domain();
798 const TypeTuple *range = TypeTuple::make_range(method()->signature());
799 init_tf(TypeFunc::make(domain, range));
800 StartNode* s = new StartOSRNode(root(), domain);
801 initial_gvn()->set_type_bottom(s);
802 init_start(s);
803 cg = CallGenerator::for_osr(method(), entry_bci());
804 } else {
805 // Normal case.
806 init_tf(TypeFunc::make(method()));
807 StartNode* s = new StartNode(root(), tf()->domain_cc());
808 initial_gvn()->set_type_bottom(s);
809 init_start(s);
810 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
811 // With java.lang.ref.reference.get() we must go through the
812 // intrinsic - even when get() is the root
813 // method of the compile - so that, if necessary, the value in
814 // the referent field of the reference object gets recorded by
815 // the pre-barrier code.
816 cg = find_intrinsic(method(), false);
817 }
818 if (cg == NULL) {
819 float past_uses = method()->interpreter_invocation_count();
820 float expected_uses = past_uses;
821 cg = CallGenerator::for_inline(method(), expected_uses);
822 }
823 }
824 if (failing()) return;
825 if (cg == NULL) {
826 record_method_not_compilable("cannot parse method");
827 return;
828 }
829 JVMState* jvms = build_start_state(start(), tf());
830 if ((jvms = cg->generate(jvms)) == NULL) {
831 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
832 record_method_not_compilable("method parse failed");
833 }
834 return;
835 }
836 GraphKit kit(jvms);
837
838 if (!kit.stopped()) {
839 // Accept return values, and transfer control we know not where.
840 // This is done by a special, unique ReturnNode bound to root.
841 return_values(kit.jvms());
842 }
843
844 if (kit.has_exceptions()) {
845 // Any exceptions that escape from this call must be rethrown
846 // to whatever caller is dynamically above us on the stack.
847 // This is done by a special, unique RethrowNode bound to root.
848 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
849 }
850
851 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
852
853 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
854 inline_string_calls(true);
855 }
856
857 if (failing()) return;
858
859 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
860
861 // Remove clutter produced by parsing.
862 if (!failing()) {
863 ResourceMark rm;
864 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
865 }
866 }
867
868 // Note: Large methods are capped off in do_one_bytecode().
869 if (failing()) return;
870
871 // After parsing, node notes are no longer automagic.
872 // They must be propagated by register_new_node_with_optimizer(),
873 // clone(), or the like.
874 set_default_node_notes(NULL);
875
876 for (;;) {
877 int successes = Inline_Warm();
878 if (failing()) return;
879 if (successes == 0) break;
880 }
881
882 // Drain the list.
883 Finish_Warm();
884 #ifndef PRODUCT
885 if (_printer && _printer->should_print(1)) {
886 _printer->print_inlining();
887 }
888 #endif
889
890 if (failing()) return;
891 NOT_PRODUCT( verify_graph_edges(); )
892
893 // Now optimize
894 Optimize();
895 if (failing()) return;
896 NOT_PRODUCT( verify_graph_edges(); )
897
898 #ifndef PRODUCT
899 if (PrintIdeal) {
900 ttyLocker ttyl; // keep the following output all in one block
901 // This output goes directly to the tty, not the compiler log.
902 // To enable tools to match it up with the compilation activity,
903 // be sure to tag this tty output with the compile ID.
904 if (xtty != NULL) {
905 xtty->head("ideal compile_id='%d'%s", compile_id(),
906 is_osr_compilation() ? " compile_kind='osr'" :
907 "");
908 }
909 root()->dump(9999);
910 if (xtty != NULL) {
911 xtty->tail("ideal");
912 }
913 }
914 #endif
915
916 #ifdef ASSERT
917 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
918 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
919 #endif
920
921 // Dump compilation data to replay it.
922 if (directive->DumpReplayOption) {
923 env()->dump_replay_data(_compile_id);
924 }
925 if (directive->DumpInlineOption && (ilt() != NULL)) {
926 env()->dump_inline_data(_compile_id);
927 }
928
929 // Now that we know the size of all the monitors we can add a fixed slot
930 // for the original deopt pc.
931
932 _orig_pc_slot = fixed_slots();
933 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
934
935 if (needs_stack_repair()) {
936 // One extra slot for the special stack increment value
937 _sp_inc_slot = next_slot;
938 next_slot += 2;
939 }
940
941 set_fixed_slots(next_slot);
942
943 // Compute when to use implicit null checks. Used by matching trap based
944 // nodes and NullCheck optimization.
945 set_allowed_deopt_reasons();
946
947 // Now generate code
948 Code_Gen();
949 if (failing()) return;
950
951 // Check if we want to skip execution of all compiled code.
952 {
953 #ifndef PRODUCT
954 if (OptoNoExecute) {
955 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
956 return;
957 }
958 #endif
959 TracePhase tp("install_code", &timers[_t_registerMethod]);
960
961 if (is_osr_compilation()) {
962 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
963 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
964 } else {
965 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
966 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) {
967 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size);
968 }
969 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry_RO) == -1) {
970 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, _first_block_size);
971 }
972 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
973 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
974 }
975 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
976 }
977
978 env()->register_method(_method, _entry_bci,
979 &_code_offsets,
980 _orig_pc_slot_offset_in_bytes,
981 code_buffer(),
982 frame_size_in_words(), _oop_map_set,
983 &_handler_table, &_inc_table,
984 compiler,
985 has_unsafe_access(),
986 SharedRuntime::is_wide_vector(max_vector_size()),
987 rtm_state()
988 );
989
990 if (log() != NULL) // Print code cache state into compiler log
991 log()->code_cache_state();
992 }
993 }
994
995 //------------------------------Compile----------------------------------------
996 // Compile a runtime stub
997 Compile::Compile( ciEnv* ci_env,
998 TypeFunc_generator generator,
999 address stub_function,
1000 const char *stub_name,
1001 int is_fancy_jump,
1002 bool pass_tls,
1003 bool save_arg_registers,
1004 bool return_pc,
1005 DirectiveSet* directive)
1006 : Phase(Compiler),
1007 _compile_id(0),
1008 _save_argument_registers(save_arg_registers),
1009 _subsume_loads(true),
1010 _do_escape_analysis(false),
1011 _eliminate_boxing(false),
1012 _method(NULL),
1013 _entry_bci(InvocationEntryBci),
1014 _stub_function(stub_function),
1015 _stub_name(stub_name),
1016 _stub_entry_point(NULL),
1017 _max_node_limit(MaxNodeLimit),
1018 _orig_pc_slot(0),
1019 _orig_pc_slot_offset_in_bytes(0),
1020 _sp_inc_slot(0),
1021 _sp_inc_slot_offset_in_bytes(0),
1022 _inlining_progress(false),
1023 _inlining_incrementally(false),
1024 _has_reserved_stack_access(false),
1025 #ifndef PRODUCT
1026 _trace_opto_output(directive->TraceOptoOutputOption),
1027 #endif
1028 _has_method_handle_invokes(false),
1029 _clinit_barrier_on_entry(false),
1030 _comp_arena(mtCompiler),
1031 _env(ci_env),
1032 _directive(directive),
1033 _log(ci_env->log()),
1034 _failure_reason(NULL),
1035 _congraph(NULL),
1036 #ifndef PRODUCT
1037 _printer(NULL),
1038 #endif
1039 _dead_node_list(comp_arena()),
1040 _dead_node_count(0),
1041 _node_arena(mtCompiler),
1042 _old_arena(mtCompiler),
1043 _mach_constant_base_node(NULL),
1044 _Compile_types(mtCompiler),
1045 _initial_gvn(NULL),
1046 _for_igvn(NULL),
1047 _warm_calls(NULL),
1048 _number_of_mh_late_inlines(0),
1049 _print_inlining_stream(NULL),
1050 _print_inlining_list(NULL),
1051 _print_inlining_idx(0),
1052 _print_inlining_output(NULL),
1053 _replay_inline_data(NULL),
1054 _java_calls(0),
1055 _inner_loops(0),
1056 _interpreter_frame_size(0),
1057 _node_bundling_limit(0),
1058 _node_bundling_base(NULL),
1059 _code_buffer("Compile::Fill_buffer"),
1060 #ifndef PRODUCT
1061 _in_dump_cnt(0),
1062 #endif
1063 _allowed_reasons(0) {
1064 C = this;
1065
1066 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1067 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1068
1069 #ifndef PRODUCT
1070 set_print_assembly(PrintFrameConverterAssembly);
1071 set_parsed_irreducible_loop(false);
1072 #else
1073 set_print_assembly(false); // Must initialize.
1074 #endif
1075 set_has_irreducible_loop(false); // no loops
1076
1077 CompileWrapper cw(this);
1078 Init(/*AliasLevel=*/ 0);
1079 init_tf((*generator)());
1080
1081 {
1082 // The following is a dummy for the sake of GraphKit::gen_stub
1083 Unique_Node_List for_igvn(comp_arena());
1084 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1085 PhaseGVN gvn(Thread::current()->resource_area(),255);
1086 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1087 gvn.transform_no_reclaim(top());
1088
1089 GraphKit kit;
1090 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1091 }
1092
1093 NOT_PRODUCT( verify_graph_edges(); )
1094 Code_Gen();
1095 if (failing()) return;
1096
1097
1098 // Entry point will be accessed using compile->stub_entry_point();
1099 if (code_buffer() == NULL) {
1100 Matcher::soft_match_failure();
1101 } else {
1102 if (PrintAssembly && (WizardMode || Verbose))
1103 tty->print_cr("### Stub::%s", stub_name);
1104
1105 if (!failing()) {
1106 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1107
1108 // Make the NMethod
1109 // For now we mark the frame as never safe for profile stackwalking
1110 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1111 code_buffer(),
1112 CodeOffsets::frame_never_safe,
1113 // _code_offsets.value(CodeOffsets::Frame_Complete),
1114 frame_size_in_words(),
1115 _oop_map_set,
1116 save_arg_registers);
1117 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1118
1119 _stub_entry_point = rs->entry_point();
1120 }
1121 }
1122 }
1123
1124 //------------------------------Init-------------------------------------------
1125 // Prepare for a single compilation
1126 void Compile::Init(int aliaslevel) {
1127 _unique = 0;
1128 _regalloc = NULL;
1129
1130 _tf = NULL; // filled in later
1131 _top = NULL; // cached later
1132 _matcher = NULL; // filled in later
1133 _cfg = NULL; // filled in later
1134
1135 set_24_bit_selection_and_mode(Use24BitFP, false);
1136
1137 _node_note_array = NULL;
1138 _default_node_notes = NULL;
1139 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1140
1141 _immutable_memory = NULL; // filled in at first inquiry
1142
1143 // Globally visible Nodes
1144 // First set TOP to NULL to give safe behavior during creation of RootNode
1145 set_cached_top_node(NULL);
1146 set_root(new RootNode());
1147 // Now that you have a Root to point to, create the real TOP
1148 set_cached_top_node( new ConNode(Type::TOP) );
1149 set_recent_alloc(NULL, NULL);
1150
1151 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1152 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1153 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1154 env()->set_dependencies(new Dependencies(env()));
1155
1156 _fixed_slots = 0;
1157 set_has_split_ifs(false);
1158 set_has_loops(has_method() && method()->has_loops()); // first approximation
1159 set_has_stringbuilder(false);
1160 set_has_boxed_value(false);
1161 _trap_can_recompile = false; // no traps emitted yet
1162 _major_progress = true; // start out assuming good things will happen
1163 set_has_unsafe_access(false);
1164 set_max_vector_size(0);
1165 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1166 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1167 set_decompile_count(0);
1168
1169 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1170 _loop_opts_cnt = LoopOptsCount;
1171 _has_flattened_accesses = false;
1172 _flattened_accesses_share_alias = true;
1173
1174 set_do_inlining(Inline);
1175 set_max_inline_size(MaxInlineSize);
1176 set_freq_inline_size(FreqInlineSize);
1177 set_do_scheduling(OptoScheduling);
1178 set_do_count_invocations(false);
1179 set_do_method_data_update(false);
1180
1181 set_do_vector_loop(false);
1182
1183 if (AllowVectorizeOnDemand) {
1184 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1185 set_do_vector_loop(true);
1186 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());})
1187 } else if (has_method() && method()->name() != 0 &&
1188 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1189 set_do_vector_loop(true);
1190 }
1191 }
1192 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1193 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());})
1194
1195 set_age_code(has_method() && method()->profile_aging());
1196 set_rtm_state(NoRTM); // No RTM lock eliding by default
1197 _max_node_limit = _directive->MaxNodeLimitOption;
1198
1199 #if INCLUDE_RTM_OPT
1200 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1201 int rtm_state = method()->method_data()->rtm_state();
1202 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1203 // Don't generate RTM lock eliding code.
1204 set_rtm_state(NoRTM);
1205 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1206 // Generate RTM lock eliding code without abort ratio calculation code.
1207 set_rtm_state(UseRTM);
1208 } else if (UseRTMDeopt) {
1209 // Generate RTM lock eliding code and include abort ratio calculation
1210 // code if UseRTMDeopt is on.
1211 set_rtm_state(ProfileRTM);
1212 }
1213 }
1214 #endif
1215 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1216 set_clinit_barrier_on_entry(true);
1217 }
1218 if (debug_info()->recording_non_safepoints()) {
1219 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1220 (comp_arena(), 8, 0, NULL));
1221 set_default_node_notes(Node_Notes::make(this));
1222 }
1223
1224 // // -- Initialize types before each compile --
1225 // // Update cached type information
1226 // if( _method && _method->constants() )
1227 // Type::update_loaded_types(_method, _method->constants());
1228
1229 // Init alias_type map.
1230 if (!_do_escape_analysis && aliaslevel == 3)
1231 aliaslevel = 2; // No unique types without escape analysis
1232 _AliasLevel = aliaslevel;
1233 const int grow_ats = 16;
1234 _max_alias_types = grow_ats;
1235 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1236 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1237 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1238 {
1239 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1240 }
1241 // Initialize the first few types.
1242 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1243 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1244 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1245 _num_alias_types = AliasIdxRaw+1;
1246 // Zero out the alias type cache.
1247 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1248 // A NULL adr_type hits in the cache right away. Preload the right answer.
1249 probe_alias_cache(NULL)->_index = AliasIdxTop;
1250
1251 _intrinsics = NULL;
1252 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1253 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1254 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1255 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1256 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1257 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
1258 register_library_intrinsics();
1259 }
1260
1261 //---------------------------init_start----------------------------------------
1262 // Install the StartNode on this compile object.
1263 void Compile::init_start(StartNode* s) {
1264 if (failing())
1265 return; // already failing
1266 assert(s == start(), "");
1267 }
1268
1269 /**
1270 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1271 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1272 * the ideal graph.
1273 */
1274 StartNode* Compile::start() const {
1275 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1276 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1277 Node* start = root()->fast_out(i);
1278 if (start->is_Start()) {
1279 return start->as_Start();
1280 }
1281 }
1282 fatal("Did not find Start node!");
1283 return NULL;
1284 }
1285
1286 //-------------------------------immutable_memory-------------------------------------
1287 // Access immutable memory
1288 Node* Compile::immutable_memory() {
1289 if (_immutable_memory != NULL) {
1290 return _immutable_memory;
1291 }
1292 StartNode* s = start();
1293 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1294 Node *p = s->fast_out(i);
1295 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1296 _immutable_memory = p;
1297 return _immutable_memory;
1298 }
1299 }
1300 ShouldNotReachHere();
1301 return NULL;
1302 }
1303
1304 //----------------------set_cached_top_node------------------------------------
1305 // Install the cached top node, and make sure Node::is_top works correctly.
1306 void Compile::set_cached_top_node(Node* tn) {
1307 if (tn != NULL) verify_top(tn);
1308 Node* old_top = _top;
1309 _top = tn;
1310 // Calling Node::setup_is_top allows the nodes the chance to adjust
1311 // their _out arrays.
1312 if (_top != NULL) _top->setup_is_top();
1313 if (old_top != NULL) old_top->setup_is_top();
1314 assert(_top == NULL || top()->is_top(), "");
1315 }
1316
1317 #ifdef ASSERT
1318 uint Compile::count_live_nodes_by_graph_walk() {
1319 Unique_Node_List useful(comp_arena());
1320 // Get useful node list by walking the graph.
1321 identify_useful_nodes(useful);
1322 return useful.size();
1323 }
1324
1325 void Compile::print_missing_nodes() {
1326
1327 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1328 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1329 return;
1330 }
1331
1332 // This is an expensive function. It is executed only when the user
1333 // specifies VerifyIdealNodeCount option or otherwise knows the
1334 // additional work that needs to be done to identify reachable nodes
1335 // by walking the flow graph and find the missing ones using
1336 // _dead_node_list.
1337
1338 Unique_Node_List useful(comp_arena());
1339 // Get useful node list by walking the graph.
1340 identify_useful_nodes(useful);
1341
1342 uint l_nodes = C->live_nodes();
1343 uint l_nodes_by_walk = useful.size();
1344
1345 if (l_nodes != l_nodes_by_walk) {
1346 if (_log != NULL) {
1347 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1348 _log->stamp();
1349 _log->end_head();
1350 }
1351 VectorSet& useful_member_set = useful.member_set();
1352 int last_idx = l_nodes_by_walk;
1353 for (int i = 0; i < last_idx; i++) {
1354 if (useful_member_set.test(i)) {
1355 if (_dead_node_list.test(i)) {
1356 if (_log != NULL) {
1357 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1358 }
1359 if (PrintIdealNodeCount) {
1360 // Print the log message to tty
1361 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1362 useful.at(i)->dump();
1363 }
1364 }
1365 }
1366 else if (! _dead_node_list.test(i)) {
1367 if (_log != NULL) {
1368 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1369 }
1370 if (PrintIdealNodeCount) {
1371 // Print the log message to tty
1372 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1373 }
1374 }
1375 }
1376 if (_log != NULL) {
1377 _log->tail("mismatched_nodes");
1378 }
1379 }
1380 }
1381 void Compile::record_modified_node(Node* n) {
1382 if (_modified_nodes != NULL && !_inlining_incrementally &&
1383 n->outcnt() != 0 && !n->is_Con()) {
1384 _modified_nodes->push(n);
1385 }
1386 }
1387
1388 void Compile::remove_modified_node(Node* n) {
1389 if (_modified_nodes != NULL) {
1390 _modified_nodes->remove(n);
1391 }
1392 }
1393 #endif
1394
1395 #ifndef PRODUCT
1396 void Compile::verify_top(Node* tn) const {
1397 if (tn != NULL) {
1398 assert(tn->is_Con(), "top node must be a constant");
1399 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1400 assert(tn->in(0) != NULL, "must have live top node");
1401 }
1402 }
1403 #endif
1404
1405
1406 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1407
1408 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1409 guarantee(arr != NULL, "");
1410 int num_blocks = arr->length();
1411 if (grow_by < num_blocks) grow_by = num_blocks;
1412 int num_notes = grow_by * _node_notes_block_size;
1413 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1414 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1415 while (num_notes > 0) {
1416 arr->append(notes);
1417 notes += _node_notes_block_size;
1418 num_notes -= _node_notes_block_size;
1419 }
1420 assert(num_notes == 0, "exact multiple, please");
1421 }
1422
1423 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1424 if (source == NULL || dest == NULL) return false;
1425
1426 if (dest->is_Con())
1427 return false; // Do not push debug info onto constants.
1428
1429 #ifdef ASSERT
1430 // Leave a bread crumb trail pointing to the original node:
1431 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1432 dest->set_debug_orig(source);
1433 }
1434 #endif
1435
1436 if (node_note_array() == NULL)
1437 return false; // Not collecting any notes now.
1438
1439 // This is a copy onto a pre-existing node, which may already have notes.
1440 // If both nodes have notes, do not overwrite any pre-existing notes.
1441 Node_Notes* source_notes = node_notes_at(source->_idx);
1442 if (source_notes == NULL || source_notes->is_clear()) return false;
1443 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1444 if (dest_notes == NULL || dest_notes->is_clear()) {
1445 return set_node_notes_at(dest->_idx, source_notes);
1446 }
1447
1448 Node_Notes merged_notes = (*source_notes);
1449 // The order of operations here ensures that dest notes will win...
1450 merged_notes.update_from(dest_notes);
1451 return set_node_notes_at(dest->_idx, &merged_notes);
1452 }
1453
1454
1455 //--------------------------allow_range_check_smearing-------------------------
1456 // Gating condition for coalescing similar range checks.
1457 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1458 // single covering check that is at least as strong as any of them.
1459 // If the optimization succeeds, the simplified (strengthened) range check
1460 // will always succeed. If it fails, we will deopt, and then give up
1461 // on the optimization.
1462 bool Compile::allow_range_check_smearing() const {
1463 // If this method has already thrown a range-check,
1464 // assume it was because we already tried range smearing
1465 // and it failed.
1466 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1467 return !already_trapped;
1468 }
1469
1470
1471 //------------------------------flatten_alias_type-----------------------------
1472 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1473 int offset = tj->offset();
1474 TypePtr::PTR ptr = tj->ptr();
1475
1476 // Known instance (scalarizable allocation) alias only with itself.
1477 bool is_known_inst = tj->isa_oopptr() != NULL &&
1478 tj->is_oopptr()->is_known_instance();
1479
1480 // Process weird unsafe references.
1481 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1482 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass();
1483 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops");
1484 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1485 tj = TypeOopPtr::BOTTOM;
1486 ptr = tj->ptr();
1487 offset = tj->offset();
1488 }
1489
1490 // Array pointers need some flattening
1491 const TypeAryPtr *ta = tj->isa_aryptr();
1492 if (ta && ta->is_stable()) {
1493 // Erase stability property for alias analysis.
1494 tj = ta = ta->cast_to_stable(false);
1495 }
1496 if (ta && ta->is_not_flat()) {
1497 // Erase not flat property for alias analysis.
1498 tj = ta = ta->cast_to_not_flat(false);
1499 }
1500 if (ta && ta->is_not_null_free()) {
1501 // Erase not null free property for alias analysis.
1502 tj = ta = ta->cast_to_not_null_free(false);
1503 }
1504
1505 if( ta && is_known_inst ) {
1506 if ( offset != Type::OffsetBot &&
1507 offset > arrayOopDesc::length_offset_in_bytes() ) {
1508 offset = Type::OffsetBot; // Flatten constant access into array body only
1509 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1510 }
1511 } else if( ta && _AliasLevel >= 2 ) {
1512 // For arrays indexed by constant indices, we flatten the alias
1513 // space to include all of the array body. Only the header, klass
1514 // and array length can be accessed un-aliased.
1515 // For flattened value type array, each field has its own slice so
1516 // we must include the field offset.
1517 if( offset != Type::OffsetBot ) {
1518 if( ta->const_oop() ) { // MethodData* or Method*
1519 offset = Type::OffsetBot; // Flatten constant access into array body
1520 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1521 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1522 // range is OK as-is.
1523 tj = ta = TypeAryPtr::RANGE;
1524 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1525 tj = TypeInstPtr::KLASS; // all klass loads look alike
1526 ta = TypeAryPtr::RANGE; // generic ignored junk
1527 ptr = TypePtr::BotPTR;
1528 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1529 tj = TypeInstPtr::MARK;
1530 ta = TypeAryPtr::RANGE; // generic ignored junk
1531 ptr = TypePtr::BotPTR;
1532 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1533 ta = tj->isa_aryptr();
1534 } else { // Random constant offset into array body
1535 offset = Type::OffsetBot; // Flatten constant access into array body
1536 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1537 }
1538 }
1539 // Arrays of fixed size alias with arrays of unknown size.
1540 if (ta->size() != TypeInt::POS) {
1541 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1542 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1543 }
1544 // Arrays of known objects become arrays of unknown objects.
1545 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1546 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1547 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1548 }
1549 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1550 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1551 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1552 }
1553 // Initially all flattened array accesses share a single slice
1554 if (ta->elem()->isa_valuetype() && ta->elem() != TypeValueType::BOTTOM && _flattened_accesses_share_alias) {
1555 const TypeAry *tary = TypeAry::make(TypeValueType::BOTTOM, ta->size());
1556 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1557 }
1558 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1559 // cannot be distinguished by bytecode alone.
1560 if (ta->elem() == TypeInt::BOOL) {
1561 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1562 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1563 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1564 }
1565 // During the 2nd round of IterGVN, NotNull castings are removed.
1566 // Make sure the Bottom and NotNull variants alias the same.
1567 // Also, make sure exact and non-exact variants alias the same.
1568 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1569 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1570 }
1571 }
1572
1573 // Oop pointers need some flattening
1574 const TypeInstPtr *to = tj->isa_instptr();
1575 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1576 ciInstanceKlass *k = to->klass()->as_instance_klass();
1577 if( ptr == TypePtr::Constant ) {
1578 if (to->klass() != ciEnv::current()->Class_klass() ||
1579 offset < k->size_helper() * wordSize) {
1580 // No constant oop pointers (such as Strings); they alias with
1581 // unknown strings.
1582 assert(!is_known_inst, "not scalarizable allocation");
1583 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset), to->klass()->flatten_array());
1584 }
1585 } else if( is_known_inst ) {
1586 tj = to; // Keep NotNull and klass_is_exact for instance type
1587 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1588 // During the 2nd round of IterGVN, NotNull castings are removed.
1589 // Make sure the Bottom and NotNull variants alias the same.
1590 // Also, make sure exact and non-exact variants alias the same.
1591 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset), to->klass()->flatten_array());
1592 }
1593 if (to->speculative() != NULL) {
1594 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->klass()->flatten_array(), to->instance_id());
1595 }
1596 // Canonicalize the holder of this field
1597 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1598 // First handle header references such as a LoadKlassNode, even if the
1599 // object's klass is unloaded at compile time (4965979).
1600 if (!is_known_inst) { // Do it only for non-instance types
1601 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset), false);
1602 }
1603 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1604 to = tj->is_instptr();
1605 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1606 // Static fields are in the space above the normal instance
1607 // fields in the java.lang.Class instance.
1608 if (to->klass() != ciEnv::current()->Class_klass()) {
1609 to = NULL;
1610 tj = TypeOopPtr::BOTTOM;
1611 offset = tj->offset();
1612 }
1613 } else {
1614 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1615 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1616 if( is_known_inst ) {
1617 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), canonical_holder->flatten_array(), to->instance_id());
1618 } else {
1619 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset), canonical_holder->flatten_array());
1620 }
1621 }
1622 }
1623 }
1624
1625 // Klass pointers to object array klasses need some flattening
1626 const TypeKlassPtr *tk = tj->isa_klassptr();
1627 if( tk ) {
1628 // If we are referencing a field within a Klass, we need
1629 // to assume the worst case of an Object. Both exact and
1630 // inexact types must flatten to the same alias class so
1631 // use NotNull as the PTR.
1632 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1633
1634 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1635 TypeKlassPtr::OBJECT->klass(),
1636 Type::Offset(offset),
1637 false);
1638 }
1639
1640 ciKlass* klass = tk->klass();
1641 if (klass != NULL && klass->is_obj_array_klass()) {
1642 ciKlass* k = TypeAryPtr::OOPS->klass();
1643 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1644 k = TypeInstPtr::BOTTOM->klass();
1645 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset), false);
1646 }
1647
1648 // Check for precise loads from the primary supertype array and force them
1649 // to the supertype cache alias index. Check for generic array loads from
1650 // the primary supertype array and also force them to the supertype cache
1651 // alias index. Since the same load can reach both, we need to merge
1652 // these 2 disparate memories into the same alias class. Since the
1653 // primary supertype array is read-only, there's no chance of confusion
1654 // where we bypass an array load and an array store.
1655 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1656 if (offset == Type::OffsetBot ||
1657 (offset >= primary_supers_offset &&
1658 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1659 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1660 offset = in_bytes(Klass::secondary_super_cache_offset());
1661 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset), tk->flatten_array());
1662 }
1663 }
1664
1665 // Flatten all Raw pointers together.
1666 if (tj->base() == Type::RawPtr)
1667 tj = TypeRawPtr::BOTTOM;
1668
1669 if (tj->base() == Type::AnyPtr)
1670 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1671
1672 // Flatten all to bottom for now
1673 switch( _AliasLevel ) {
1674 case 0:
1675 tj = TypePtr::BOTTOM;
1676 break;
1677 case 1: // Flatten to: oop, static, field or array
1678 switch (tj->base()) {
1679 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1680 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1681 case Type::AryPtr: // do not distinguish arrays at all
1682 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1683 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1684 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1685 default: ShouldNotReachHere();
1686 }
1687 break;
1688 case 2: // No collapsing at level 2; keep all splits
1689 case 3: // No collapsing at level 3; keep all splits
1690 break;
1691 default:
1692 Unimplemented();
1693 }
1694
1695 offset = tj->offset();
1696 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1697
1698 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1699 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1700 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1701 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1702 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1703 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1704 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1705 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1706 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1707 assert( tj->ptr() != TypePtr::TopPTR &&
1708 tj->ptr() != TypePtr::AnyNull &&
1709 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1710 // assert( tj->ptr() != TypePtr::Constant ||
1711 // tj->base() == Type::RawPtr ||
1712 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1713
1714 return tj;
1715 }
1716
1717 void Compile::AliasType::Init(int i, const TypePtr* at) {
1718 _index = i;
1719 _adr_type = at;
1720 _field = NULL;
1721 _element = NULL;
1722 _is_rewritable = true; // default
1723 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1724 if (atoop != NULL && atoop->is_known_instance()) {
1725 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1726 _general_index = Compile::current()->get_alias_index(gt);
1727 } else {
1728 _general_index = 0;
1729 }
1730 }
1731
1732 BasicType Compile::AliasType::basic_type() const {
1733 if (element() != NULL) {
1734 const Type* element = adr_type()->is_aryptr()->elem();
1735 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1736 } if (field() != NULL) {
1737 return field()->layout_type();
1738 } else {
1739 return T_ILLEGAL; // unknown
1740 }
1741 }
1742
1743 //---------------------------------print_on------------------------------------
1744 #ifndef PRODUCT
1745 void Compile::AliasType::print_on(outputStream* st) {
1746 if (index() < 10)
1747 st->print("@ <%d> ", index());
1748 else st->print("@ <%d>", index());
1749 st->print(is_rewritable() ? " " : " RO");
1750 int offset = adr_type()->offset();
1751 if (offset == Type::OffsetBot)
1752 st->print(" +any");
1753 else st->print(" +%-3d", offset);
1754 st->print(" in ");
1755 adr_type()->dump_on(st);
1756 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1757 if (field() != NULL && tjp) {
1758 if (tjp->klass() != field()->holder() ||
1759 tjp->offset() != field()->offset_in_bytes()) {
1760 st->print(" != ");
1761 field()->print();
1762 st->print(" ***");
1763 }
1764 }
1765 }
1766
1767 void print_alias_types() {
1768 Compile* C = Compile::current();
1769 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1770 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1771 C->alias_type(idx)->print_on(tty);
1772 tty->cr();
1773 }
1774 }
1775 #endif
1776
1777
1778 //----------------------------probe_alias_cache--------------------------------
1779 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1780 intptr_t key = (intptr_t) adr_type;
1781 key ^= key >> logAliasCacheSize;
1782 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1783 }
1784
1785
1786 //-----------------------------grow_alias_types--------------------------------
1787 void Compile::grow_alias_types() {
1788 const int old_ats = _max_alias_types; // how many before?
1789 const int new_ats = old_ats; // how many more?
1790 const int grow_ats = old_ats+new_ats; // how many now?
1791 _max_alias_types = grow_ats;
1792 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1793 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1794 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1795 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1796 }
1797
1798
1799 //--------------------------------find_alias_type------------------------------
1800 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1801 if (_AliasLevel == 0)
1802 return alias_type(AliasIdxBot);
1803
1804 AliasCacheEntry* ace = NULL;
1805 if (!uncached) {
1806 ace = probe_alias_cache(adr_type);
1807 if (ace->_adr_type == adr_type) {
1808 return alias_type(ace->_index);
1809 }
1810 }
1811
1812 // Handle special cases.
1813 if (adr_type == NULL) return alias_type(AliasIdxTop);
1814 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1815
1816 // Do it the slow way.
1817 const TypePtr* flat = flatten_alias_type(adr_type);
1818
1819 #ifdef ASSERT
1820 {
1821 ResourceMark rm;
1822 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1823 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1824 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1825 Type::str(adr_type));
1826 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1827 const TypeOopPtr* foop = flat->is_oopptr();
1828 // Scalarizable allocations have exact klass always.
1829 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1830 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1831 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1832 Type::str(foop), Type::str(xoop));
1833 }
1834 }
1835 #endif
1836
1837 int idx = AliasIdxTop;
1838 for (int i = 0; i < num_alias_types(); i++) {
1839 if (alias_type(i)->adr_type() == flat) {
1840 idx = i;
1841 break;
1842 }
1843 }
1844
1845 if (idx == AliasIdxTop) {
1846 if (no_create) return NULL;
1847 // Grow the array if necessary.
1848 if (_num_alias_types == _max_alias_types) grow_alias_types();
1849 // Add a new alias type.
1850 idx = _num_alias_types++;
1851 _alias_types[idx]->Init(idx, flat);
1852 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1853 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1854 if (flat->isa_instptr()) {
1855 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1856 && flat->is_instptr()->klass() == env()->Class_klass())
1857 alias_type(idx)->set_rewritable(false);
1858 }
1859 ciField* field = NULL;
1860 if (flat->isa_aryptr()) {
1861 #ifdef ASSERT
1862 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1863 // (T_BYTE has the weakest alignment and size restrictions...)
1864 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1865 #endif
1866 const Type* elemtype = flat->is_aryptr()->elem();
1867 if (flat->offset() == TypePtr::OffsetBot) {
1868 alias_type(idx)->set_element(elemtype);
1869 }
1870 int field_offset = flat->is_aryptr()->field_offset().get();
1871 if (elemtype->isa_valuetype() &&
1872 elemtype->value_klass() != NULL &&
1873 field_offset != Type::OffsetBot) {
1874 ciValueKlass* vk = elemtype->value_klass();
1875 field_offset += vk->first_field_offset();
1876 field = vk->get_field_by_offset(field_offset, false);
1877 }
1878 }
1879 if (flat->isa_klassptr()) {
1880 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1881 alias_type(idx)->set_rewritable(false);
1882 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1883 alias_type(idx)->set_rewritable(false);
1884 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1885 alias_type(idx)->set_rewritable(false);
1886 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1887 alias_type(idx)->set_rewritable(false);
1888 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1889 alias_type(idx)->set_rewritable(false);
1890 }
1891 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1892 // but the base pointer type is not distinctive enough to identify
1893 // references into JavaThread.)
1894
1895 // Check for final fields.
1896 const TypeInstPtr* tinst = flat->isa_instptr();
1897 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1898 if (tinst->const_oop() != NULL &&
1899 tinst->klass() == ciEnv::current()->Class_klass() &&
1900 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1901 // static field
1902 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1903 field = k->get_field_by_offset(tinst->offset(), true);
1904 } else if (tinst->klass()->is_valuetype()) {
1905 // Value type field
1906 ciValueKlass* vk = tinst->value_klass();
1907 field = vk->get_field_by_offset(tinst->offset(), false);
1908 } else {
1909 ciInstanceKlass* k = tinst->klass()->as_instance_klass();
1910 field = k->get_field_by_offset(tinst->offset(), false);
1911 }
1912 }
1913 assert(field == NULL ||
1914 original_field == NULL ||
1915 (field->holder() == original_field->holder() &&
1916 field->offset() == original_field->offset() &&
1917 field->is_static() == original_field->is_static()), "wrong field?");
1918 // Set field() and is_rewritable() attributes.
1919 if (field != NULL) {
1920 alias_type(idx)->set_field(field);
1921 if (flat->isa_aryptr()) {
1922 // Fields of flattened inline type arrays are rewritable although they are declared final
1923 assert(flat->is_aryptr()->elem()->isa_valuetype(), "must be a flattened value array");
1924 alias_type(idx)->set_rewritable(true);
1925 }
1926 }
1927 }
1928
1929 // Fill the cache for next time.
1930 if (!uncached) {
1931 ace->_adr_type = adr_type;
1932 ace->_index = idx;
1933 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1934
1935 // Might as well try to fill the cache for the flattened version, too.
1936 AliasCacheEntry* face = probe_alias_cache(flat);
1937 if (face->_adr_type == NULL) {
1938 face->_adr_type = flat;
1939 face->_index = idx;
1940 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1941 }
1942 }
1943
1944 return alias_type(idx);
1945 }
1946
1947
1948 Compile::AliasType* Compile::alias_type(ciField* field) {
1949 const TypeOopPtr* t;
1950 if (field->is_static())
1951 t = TypeInstPtr::make(field->holder()->java_mirror());
1952 else
1953 t = TypeOopPtr::make_from_klass_raw(field->holder());
1954 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1955 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1956 return atp;
1957 }
1958
1959
1960 //------------------------------have_alias_type--------------------------------
1961 bool Compile::have_alias_type(const TypePtr* adr_type) {
1962 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1963 if (ace->_adr_type == adr_type) {
1964 return true;
1965 }
1966
1967 // Handle special cases.
1968 if (adr_type == NULL) return true;
1969 if (adr_type == TypePtr::BOTTOM) return true;
1970
1971 return find_alias_type(adr_type, true, NULL) != NULL;
1972 }
1973
1974 //-----------------------------must_alias--------------------------------------
1975 // True if all values of the given address type are in the given alias category.
1976 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1977 if (alias_idx == AliasIdxBot) return true; // the universal category
1978 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1979 if (alias_idx == AliasIdxTop) return false; // the empty category
1980 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1981
1982 // the only remaining possible overlap is identity
1983 int adr_idx = get_alias_index(adr_type);
1984 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1985 assert(adr_idx == alias_idx ||
1986 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1987 && adr_type != TypeOopPtr::BOTTOM),
1988 "should not be testing for overlap with an unsafe pointer");
1989 return adr_idx == alias_idx;
1990 }
1991
1992 //------------------------------can_alias--------------------------------------
1993 // True if any values of the given address type are in the given alias category.
1994 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1995 if (alias_idx == AliasIdxTop) return false; // the empty category
1996 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1997 if (alias_idx == AliasIdxBot) return true; // the universal category
1998 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1999
2000 // the only remaining possible overlap is identity
2001 int adr_idx = get_alias_index(adr_type);
2002 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
2003 return adr_idx == alias_idx;
2004 }
2005
2006
2007
2008 //---------------------------pop_warm_call-------------------------------------
2009 WarmCallInfo* Compile::pop_warm_call() {
2010 WarmCallInfo* wci = _warm_calls;
2011 if (wci != NULL) _warm_calls = wci->remove_from(wci);
2012 return wci;
2013 }
2014
2015 //----------------------------Inline_Warm--------------------------------------
2016 int Compile::Inline_Warm() {
2017 // If there is room, try to inline some more warm call sites.
2018 // %%% Do a graph index compaction pass when we think we're out of space?
2019 if (!InlineWarmCalls) return 0;
2020
2021 int calls_made_hot = 0;
2022 int room_to_grow = NodeCountInliningCutoff - unique();
2023 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
2024 int amount_grown = 0;
2025 WarmCallInfo* call;
2026 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
2027 int est_size = (int)call->size();
2028 if (est_size > (room_to_grow - amount_grown)) {
2029 // This one won't fit anyway. Get rid of it.
2030 call->make_cold();
2031 continue;
2032 }
2033 call->make_hot();
2034 calls_made_hot++;
2035 amount_grown += est_size;
2036 amount_to_grow -= est_size;
2037 }
2038
2039 if (calls_made_hot > 0) set_major_progress();
2040 return calls_made_hot;
2041 }
2042
2043
2044 //----------------------------Finish_Warm--------------------------------------
2045 void Compile::Finish_Warm() {
2046 if (!InlineWarmCalls) return;
2047 if (failing()) return;
2048 if (warm_calls() == NULL) return;
2049
2050 // Clean up loose ends, if we are out of space for inlining.
2051 WarmCallInfo* call;
2052 while ((call = pop_warm_call()) != NULL) {
2053 call->make_cold();
2054 }
2055 }
2056
2057 //---------------------cleanup_loop_predicates-----------------------
2058 // Remove the opaque nodes that protect the predicates so that all unused
2059 // checks and uncommon_traps will be eliminated from the ideal graph
2060 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
2061 if (predicate_count()==0) return;
2062 for (int i = predicate_count(); i > 0; i--) {
2063 Node * n = predicate_opaque1_node(i-1);
2064 assert(n->Opcode() == Op_Opaque1, "must be");
2065 igvn.replace_node(n, n->in(1));
2066 }
2067 assert(predicate_count()==0, "should be clean!");
2068 }
2069
2070 void Compile::add_range_check_cast(Node* n) {
2071 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2072 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
2073 _range_check_casts->append(n);
2074 }
2075
2076 // Remove all range check dependent CastIINodes.
2077 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
2078 for (int i = range_check_cast_count(); i > 0; i--) {
2079 Node* cast = range_check_cast_node(i-1);
2080 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2081 igvn.replace_node(cast, cast->in(1));
2082 }
2083 assert(range_check_cast_count() == 0, "should be empty");
2084 }
2085
2086 void Compile::add_opaque4_node(Node* n) {
2087 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2088 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2089 _opaque4_nodes->append(n);
2090 }
2091
2092 // Remove all Opaque4 nodes.
2093 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2094 for (int i = opaque4_count(); i > 0; i--) {
2095 Node* opaq = opaque4_node(i-1);
2096 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2097 igvn.replace_node(opaq, opaq->in(2));
2098 }
2099 assert(opaque4_count() == 0, "should be empty");
2100 }
2101
2102 void Compile::add_value_type(Node* n) {
2103 assert(n->is_ValueTypeBase(), "unexpected node");
2104 if (_value_type_nodes != NULL) {
2105 _value_type_nodes->push(n);
2106 }
2107 }
2108
2109 void Compile::remove_value_type(Node* n) {
2110 assert(n->is_ValueTypeBase(), "unexpected node");
2111 if (_value_type_nodes != NULL) {
2112 _value_type_nodes->remove(n);
2113 }
2114 }
2115
2116 // Does the return value keep otherwise useless value type allocations
2117 // alive?
2118 static bool return_val_keeps_allocations_alive(Node* ret_val) {
2119 ResourceMark rm;
2120 Unique_Node_List wq;
2121 wq.push(ret_val);
2122 bool some_allocations = false;
2123 for (uint i = 0; i < wq.size(); i++) {
2124 Node* n = wq.at(i);
2125 assert(!n->is_ValueTypeBase(), "chain of value type nodes");
2126 if (n->outcnt() > 1) {
2127 // Some other use for the allocation
2128 return false;
2129 } else if (n->is_Phi()) {
2130 for (uint j = 1; j < n->req(); j++) {
2131 wq.push(n->in(j));
2132 }
2133 } else if (n->is_CheckCastPP() &&
2134 n->in(1)->is_Proj() &&
2135 n->in(1)->in(0)->is_Allocate()) {
2136 some_allocations = true;
2137 }
2138 }
2139 return some_allocations;
2140 }
2141
2142 void Compile::process_value_types(PhaseIterGVN &igvn) {
2143 // Make value types scalar in safepoints
2144 while (_value_type_nodes->size() != 0) {
2145 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase();
2146 vt->make_scalar_in_safepoints(&igvn);
2147 if (vt->is_ValueTypePtr()) {
2148 igvn.replace_node(vt, vt->get_oop());
2149 } else if (vt->outcnt() == 0) {
2150 igvn.remove_dead_node(vt);
2151 }
2152 }
2153 _value_type_nodes = NULL;
2154 if (tf()->returns_value_type_as_fields()) {
2155 Node* ret = NULL;
2156 for (uint i = 1; i < root()->req(); i++){
2157 Node* in = root()->in(i);
2158 if (in->Opcode() == Op_Return) {
2159 assert(ret == NULL, "only one return");
2160 ret = in;
2161 }
2162 }
2163 if (ret != NULL) {
2164 Node* ret_val = ret->in(TypeFunc::Parms);
2165 if (igvn.type(ret_val)->isa_oopptr() &&
2166 return_val_keeps_allocations_alive(ret_val)) {
2167 igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn));
2168 assert(ret_val->outcnt() == 0, "should be dead now");
2169 igvn.remove_dead_node(ret_val);
2170 }
2171 }
2172 }
2173 igvn.optimize();
2174 }
2175
2176 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2177 if (!_has_flattened_accesses) {
2178 return;
2179 }
2180 // Initially, all flattened array accesses share the same slice to
2181 // keep dependencies with Object[] array accesses (that could be
2182 // to a flattened array) correct. We're done with parsing so we
2183 // now know all flattened array accesses in this compile
2184 // unit. Let's move flattened array accesses to their own slice,
2185 // one per element field. This should help memory access
2186 // optimizations.
2187 ResourceMark rm;
2188 Unique_Node_List wq;
2189 wq.push(root());
2190
2191 Node_List mergememnodes;
2192 Node_List memnodes;
2193
2194 // Alias index currently shared by all flattened memory accesses
2195 int index = get_alias_index(TypeAryPtr::VALUES);
2196
2197 // Find MergeMem nodes and flattened array accesses
2198 for (uint i = 0; i < wq.size(); i++) {
2199 Node* n = wq.at(i);
2200 if (n->is_Mem()) {
2201 const TypePtr* adr_type = NULL;
2202 if (n->Opcode() == Op_StoreCM) {
2203 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2204 } else {
2205 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2206 }
2207 if (adr_type == TypeAryPtr::VALUES) {
2208 memnodes.push(n);
2209 }
2210 } else if (n->is_MergeMem()) {
2211 MergeMemNode* mm = n->as_MergeMem();
2212 if (mm->memory_at(index) != mm->base_memory()) {
2213 mergememnodes.push(n);
2214 }
2215 }
2216 for (uint j = 0; j < n->req(); j++) {
2217 Node* m = n->in(j);
2218 if (m != NULL) {
2219 wq.push(m);
2220 }
2221 }
2222 }
2223
2224 if (memnodes.size() > 0) {
2225 _flattened_accesses_share_alias = false;
2226
2227 // We are going to change the slice for the flattened array
2228 // accesses so we need to clear the cache entries that refer to
2229 // them.
2230 for (uint i = 0; i < AliasCacheSize; i++) {
2231 AliasCacheEntry* ace = &_alias_cache[i];
2232 if (ace->_adr_type != NULL &&
2233 ace->_adr_type->isa_aryptr() &&
2234 ace->_adr_type->is_aryptr()->elem()->isa_valuetype()) {
2235 ace->_adr_type = NULL;
2236 ace->_index = 0;
2237 }
2238 }
2239
2240 // Find what aliases we are going to add
2241 int start_alias = num_alias_types()-1;
2242 int stop_alias = 0;
2243
2244 for (uint i = 0; i < memnodes.size(); i++) {
2245 Node* m = memnodes.at(i);
2246 const TypePtr* adr_type = NULL;
2247 if (m->Opcode() == Op_StoreCM) {
2248 adr_type = m->in(MemNode::OopStore)->adr_type();
2249 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2250 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2251 get_alias_index(adr_type));
2252 igvn.register_new_node_with_optimizer(clone);
2253 igvn.replace_node(m, clone);
2254 } else {
2255 adr_type = m->adr_type();
2256 #ifdef ASSERT
2257 m->as_Mem()->set_adr_type(adr_type);
2258 #endif
2259 }
2260 int idx = get_alias_index(adr_type);
2261 start_alias = MIN2(start_alias, idx);
2262 stop_alias = MAX2(stop_alias, idx);
2263 }
2264
2265 assert(stop_alias >= start_alias, "should have expanded aliases");
2266
2267 Node_Stack stack(0);
2268 #ifdef ASSERT
2269 VectorSet seen(Thread::current()->resource_area());
2270 #endif
2271 // Now let's fix the memory graph so each flattened array access
2272 // is moved to the right slice. Start from the MergeMem nodes.
2273 uint last = unique();
2274 for (uint i = 0; i < mergememnodes.size(); i++) {
2275 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2276 Node* n = current->memory_at(index);
2277 MergeMemNode* mm = NULL;
2278 do {
2279 // Follow memory edges through memory accesses, phis and
2280 // narrow membars and push nodes on the stack. Once we hit
2281 // bottom memory, we pop element off the stack one at a
2282 // time, in reverse order, and move them to the right slice
2283 // by changing their memory edges.
2284 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::VALUES) {
2285 assert(!seen.test_set(n->_idx), "");
2286 // Uses (a load for instance) will need to be moved to the
2287 // right slice as well and will get a new memory state
2288 // that we don't know yet. The use could also be the
2289 // backedge of a loop. We put a place holder node between
2290 // the memory node and its uses. We replace that place
2291 // holder with the correct memory state once we know it,
2292 // i.e. when nodes are popped off the stack. Using the
2293 // place holder make the logic work in the presence of
2294 // loops.
2295 if (n->outcnt() > 1) {
2296 Node* place_holder = NULL;
2297 assert(!n->has_out_with(Op_Node), "");
2298 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2299 Node* u = n->out(k);
2300 if (u != current && u->_idx < last) {
2301 bool success = false;
2302 for (uint l = 0; l < u->req(); l++) {
2303 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2304 continue;
2305 }
2306 Node* in = u->in(l);
2307 if (in == n) {
2308 if (place_holder == NULL) {
2309 place_holder = new Node(1);
2310 place_holder->init_req(0, n);
2311 }
2312 igvn.replace_input_of(u, l, place_holder);
2313 success = true;
2314 }
2315 }
2316 if (success) {
2317 --k;
2318 }
2319 }
2320 }
2321 }
2322 if (n->is_Phi()) {
2323 stack.push(n, 1);
2324 n = n->in(1);
2325 } else if (n->is_Mem()) {
2326 stack.push(n, n->req());
2327 n = n->in(MemNode::Memory);
2328 } else {
2329 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2330 stack.push(n, n->req());
2331 n = n->in(0)->in(TypeFunc::Memory);
2332 }
2333 } else {
2334 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2335 // Build a new MergeMem node to carry the new memory state
2336 // as we build it. IGVN should fold extraneous MergeMem
2337 // nodes.
2338 mm = MergeMemNode::make(n);
2339 igvn.register_new_node_with_optimizer(mm);
2340 while (stack.size() > 0) {
2341 Node* m = stack.node();
2342 uint idx = stack.index();
2343 if (m->is_Mem()) {
2344 // Move memory node to its new slice
2345 const TypePtr* adr_type = m->adr_type();
2346 int alias = get_alias_index(adr_type);
2347 Node* prev = mm->memory_at(alias);
2348 igvn.replace_input_of(m, MemNode::Memory, prev);
2349 mm->set_memory_at(alias, m);
2350 } else if (m->is_Phi()) {
2351 // We need as many new phis as there are new aliases
2352 igvn.replace_input_of(m, idx, mm);
2353 if (idx == m->req()-1) {
2354 Node* r = m->in(0);
2355 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2356 const Type* adr_type = get_adr_type(j);
2357 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2358 continue;
2359 }
2360 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2361 igvn.register_new_node_with_optimizer(phi);
2362 for (uint k = 1; k < m->req(); k++) {
2363 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2364 }
2365 mm->set_memory_at(j, phi);
2366 }
2367 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2368 igvn.register_new_node_with_optimizer(base_phi);
2369 for (uint k = 1; k < m->req(); k++) {
2370 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2371 }
2372 mm->set_base_memory(base_phi);
2373 }
2374 } else {
2375 // This is a MemBarCPUOrder node from
2376 // Parse::array_load()/Parse::array_store(), in the
2377 // branch that handles flattened arrays hidden under
2378 // an Object[] array. We also need one new membar per
2379 // new alias to keep the unknown access that the
2380 // membars protect properly ordered with accesses to
2381 // known flattened array.
2382 assert(m->is_Proj(), "projection expected");
2383 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2384 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2385 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2386 const Type* adr_type = get_adr_type(j);
2387 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2388 continue;
2389 }
2390 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2391 igvn.register_new_node_with_optimizer(mb);
2392 Node* mem = mm->memory_at(j);
2393 mb->init_req(TypeFunc::Control, ctrl);
2394 mb->init_req(TypeFunc::Memory, mem);
2395 ctrl = new ProjNode(mb, TypeFunc::Control);
2396 igvn.register_new_node_with_optimizer(ctrl);
2397 mem = new ProjNode(mb, TypeFunc::Memory);
2398 igvn.register_new_node_with_optimizer(mem);
2399 mm->set_memory_at(j, mem);
2400 }
2401 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2402 }
2403 if (idx < m->req()-1) {
2404 idx += 1;
2405 stack.set_index(idx);
2406 n = m->in(idx);
2407 break;
2408 }
2409 // Take care of place holder nodes
2410 if (m->has_out_with(Op_Node)) {
2411 Node* place_holder = m->find_out_with(Op_Node);
2412 if (place_holder != NULL) {
2413 Node* mm_clone = mm->clone();
2414 igvn.register_new_node_with_optimizer(mm_clone);
2415 Node* hook = new Node(1);
2416 hook->init_req(0, mm);
2417 igvn.replace_node(place_holder, mm_clone);
2418 hook->destruct();
2419 }
2420 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2421 }
2422 stack.pop();
2423 }
2424 }
2425 } while(stack.size() > 0);
2426 // Fix the memory state at the MergeMem we started from
2427 igvn.rehash_node_delayed(current);
2428 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2429 const Type* adr_type = get_adr_type(j);
2430 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2431 continue;
2432 }
2433 current->set_memory_at(j, mm);
2434 }
2435 current->set_memory_at(index, current->base_memory());
2436 }
2437 igvn.optimize();
2438 }
2439 print_method(PHASE_SPLIT_VALUES_ARRAY, 2);
2440 }
2441
2442
2443 // StringOpts and late inlining of string methods
2444 void Compile::inline_string_calls(bool parse_time) {
2445 {
2446 // remove useless nodes to make the usage analysis simpler
2447 ResourceMark rm;
2448 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2449 }
2450
2451 {
2452 ResourceMark rm;
2453 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2454 PhaseStringOpts pso(initial_gvn(), for_igvn());
2455 print_method(PHASE_AFTER_STRINGOPTS, 3);
2456 }
2457
2458 // now inline anything that we skipped the first time around
2459 if (!parse_time) {
2460 _late_inlines_pos = _late_inlines.length();
2461 }
2462
2463 while (_string_late_inlines.length() > 0) {
2464 CallGenerator* cg = _string_late_inlines.pop();
2465 cg->do_late_inline();
2466 if (failing()) return;
2467 }
2468 _string_late_inlines.trunc_to(0);
2469 }
2470
2471 // Late inlining of boxing methods
2472 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2473 if (_boxing_late_inlines.length() > 0) {
2474 assert(has_boxed_value(), "inconsistent");
2475
2476 PhaseGVN* gvn = initial_gvn();
2477 set_inlining_incrementally(true);
2478
2479 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2480 for_igvn()->clear();
2481 gvn->replace_with(&igvn);
2482
2483 _late_inlines_pos = _late_inlines.length();
2484
2485 while (_boxing_late_inlines.length() > 0) {
2486 CallGenerator* cg = _boxing_late_inlines.pop();
2487 cg->do_late_inline();
2488 if (failing()) return;
2489 }
2490 _boxing_late_inlines.trunc_to(0);
2491
2492 inline_incrementally_cleanup(igvn);
2493
2494 set_inlining_incrementally(false);
2495 }
2496 }
2497
2498 bool Compile::inline_incrementally_one() {
2499 assert(IncrementalInline, "incremental inlining should be on");
2500
2501 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2502 set_inlining_progress(false);
2503 set_do_cleanup(false);
2504 int i = 0;
2505 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2506 CallGenerator* cg = _late_inlines.at(i);
2507 _late_inlines_pos = i+1;
2508 cg->do_late_inline();
2509 if (failing()) return false;
2510 }
2511 int j = 0;
2512 for (; i < _late_inlines.length(); i++, j++) {
2513 _late_inlines.at_put(j, _late_inlines.at(i));
2514 }
2515 _late_inlines.trunc_to(j);
2516 assert(inlining_progress() || _late_inlines.length() == 0, "");
2517
2518 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2519
2520 set_inlining_progress(false);
2521 set_do_cleanup(false);
2522 return (_late_inlines.length() > 0) && !needs_cleanup;
2523 }
2524
2525 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2526 {
2527 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2528 ResourceMark rm;
2529 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2530 }
2531 {
2532 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2533 igvn = PhaseIterGVN(initial_gvn());
2534 igvn.optimize();
2535 }
2536 }
2537
2538 // Perform incremental inlining until bound on number of live nodes is reached
2539 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2540 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2541
2542 set_inlining_incrementally(true);
2543 uint low_live_nodes = 0;
2544
2545 while (_late_inlines.length() > 0) {
2546 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2547 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2548 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2549 // PhaseIdealLoop is expensive so we only try it once we are
2550 // out of live nodes and we only try it again if the previous
2551 // helped got the number of nodes down significantly
2552 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2553 if (failing()) return;
2554 low_live_nodes = live_nodes();
2555 _major_progress = true;
2556 }
2557
2558 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2559 break; // finish
2560 }
2561 }
2562
2563 for_igvn()->clear();
2564 initial_gvn()->replace_with(&igvn);
2565
2566 while (inline_incrementally_one()) {
2567 assert(!failing(), "inconsistent");
2568 }
2569
2570 if (failing()) return;
2571
2572 inline_incrementally_cleanup(igvn);
2573
2574 if (failing()) return;
2575 }
2576 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2577
2578 if (_string_late_inlines.length() > 0) {
2579 assert(has_stringbuilder(), "inconsistent");
2580 for_igvn()->clear();
2581 initial_gvn()->replace_with(&igvn);
2582
2583 inline_string_calls(false);
2584
2585 if (failing()) return;
2586
2587 inline_incrementally_cleanup(igvn);
2588 }
2589
2590 set_inlining_incrementally(false);
2591 }
2592
2593
2594 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2595 if(_loop_opts_cnt > 0) {
2596 debug_only( int cnt = 0; );
2597 while(major_progress() && (_loop_opts_cnt > 0)) {
2598 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2599 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2600 PhaseIdealLoop::optimize(igvn, mode);
2601 _loop_opts_cnt--;
2602 if (failing()) return false;
2603 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2604 }
2605 }
2606 return true;
2607 }
2608
2609 // Remove edges from "root" to each SafePoint at a backward branch.
2610 // They were inserted during parsing (see add_safepoint()) to make
2611 // infinite loops without calls or exceptions visible to root, i.e.,
2612 // useful.
2613 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2614 Node *r = root();
2615 if (r != NULL) {
2616 for (uint i = r->req(); i < r->len(); ++i) {
2617 Node *n = r->in(i);
2618 if (n != NULL && n->is_SafePoint()) {
2619 r->rm_prec(i);
2620 if (n->outcnt() == 0) {
2621 igvn.remove_dead_node(n);
2622 }
2623 --i;
2624 }
2625 }
2626 }
2627 }
2628
2629 //------------------------------Optimize---------------------------------------
2630 // Given a graph, optimize it.
2631 void Compile::Optimize() {
2632 TracePhase tp("optimizer", &timers[_t_optimizer]);
2633
2634 #ifndef PRODUCT
2635 if (_directive->BreakAtCompileOption) {
2636 BREAKPOINT;
2637 }
2638
2639 #endif
2640
2641 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2642 #ifdef ASSERT
2643 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2644 #endif
2645
2646 ResourceMark rm;
2647
2648 print_inlining_reinit();
2649
2650 NOT_PRODUCT( verify_graph_edges(); )
2651
2652 print_method(PHASE_AFTER_PARSING);
2653
2654 {
2655 // Iterative Global Value Numbering, including ideal transforms
2656 // Initialize IterGVN with types and values from parse-time GVN
2657 PhaseIterGVN igvn(initial_gvn());
2658 #ifdef ASSERT
2659 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2660 #endif
2661 {
2662 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2663 igvn.optimize();
2664 }
2665
2666 if (failing()) return;
2667
2668 print_method(PHASE_ITER_GVN1, 2);
2669
2670 inline_incrementally(igvn);
2671
2672 print_method(PHASE_INCREMENTAL_INLINE, 2);
2673
2674 if (failing()) return;
2675
2676 if (eliminate_boxing()) {
2677 // Inline valueOf() methods now.
2678 inline_boxing_calls(igvn);
2679
2680 if (AlwaysIncrementalInline) {
2681 inline_incrementally(igvn);
2682 }
2683
2684 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2685
2686 if (failing()) return;
2687 }
2688
2689 // Now that all inlining is over, cut edge from root to loop
2690 // safepoints
2691 remove_root_to_sfpts_edges(igvn);
2692
2693 // Remove the speculative part of types and clean up the graph from
2694 // the extra CastPP nodes whose only purpose is to carry them. Do
2695 // that early so that optimizations are not disrupted by the extra
2696 // CastPP nodes.
2697 remove_speculative_types(igvn);
2698
2699 // No more new expensive nodes will be added to the list from here
2700 // so keep only the actual candidates for optimizations.
2701 cleanup_expensive_nodes(igvn);
2702
2703 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2704 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2705 initial_gvn()->replace_with(&igvn);
2706 for_igvn()->clear();
2707 Unique_Node_List new_worklist(C->comp_arena());
2708 {
2709 ResourceMark rm;
2710 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2711 }
2712 set_for_igvn(&new_worklist);
2713 igvn = PhaseIterGVN(initial_gvn());
2714 igvn.optimize();
2715 }
2716
2717 if (_value_type_nodes->size() > 0) {
2718 // Do this once all inlining is over to avoid getting inconsistent debug info
2719 process_value_types(igvn);
2720 }
2721
2722 adjust_flattened_array_access_aliases(igvn);
2723
2724 // Perform escape analysis
2725 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2726 if (has_loops()) {
2727 // Cleanup graph (remove dead nodes).
2728 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2729 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2730 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2731 if (failing()) return;
2732 }
2733 ConnectionGraph::do_analysis(this, &igvn);
2734
2735 if (failing()) return;
2736
2737 // Optimize out fields loads from scalar replaceable allocations.
2738 igvn.optimize();
2739 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2740
2741 if (failing()) return;
2742
2743 if (congraph() != NULL && macro_count() > 0) {
2744 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2745 PhaseMacroExpand mexp(igvn);
2746 mexp.eliminate_macro_nodes();
2747 igvn.set_delay_transform(false);
2748
2749 igvn.optimize();
2750 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2751
2752 if (failing()) return;
2753 }
2754 }
2755
2756 // Loop transforms on the ideal graph. Range Check Elimination,
2757 // peeling, unrolling, etc.
2758
2759 // Set loop opts counter
2760 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2761 {
2762 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2763 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2764 _loop_opts_cnt--;
2765 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2766 if (failing()) return;
2767 }
2768 // Loop opts pass if partial peeling occurred in previous pass
2769 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2770 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2771 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2772 _loop_opts_cnt--;
2773 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2774 if (failing()) return;
2775 }
2776 // Loop opts pass for loop-unrolling before CCP
2777 if(major_progress() && (_loop_opts_cnt > 0)) {
2778 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2779 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2780 _loop_opts_cnt--;
2781 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2782 }
2783 if (!failing()) {
2784 // Verify that last round of loop opts produced a valid graph
2785 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2786 PhaseIdealLoop::verify(igvn);
2787 }
2788 }
2789 if (failing()) return;
2790
2791 // Conditional Constant Propagation;
2792 PhaseCCP ccp( &igvn );
2793 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2794 {
2795 TracePhase tp("ccp", &timers[_t_ccp]);
2796 ccp.do_transform();
2797 }
2798 print_method(PHASE_CPP1, 2);
2799
2800 assert( true, "Break here to ccp.dump_old2new_map()");
2801
2802 // Iterative Global Value Numbering, including ideal transforms
2803 {
2804 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2805 igvn = ccp;
2806 igvn.optimize();
2807 }
2808 print_method(PHASE_ITER_GVN2, 2);
2809
2810 if (failing()) return;
2811
2812 // Loop transforms on the ideal graph. Range Check Elimination,
2813 // peeling, unrolling, etc.
2814 if (!optimize_loops(igvn, LoopOptsDefault)) {
2815 return;
2816 }
2817
2818 if (failing()) return;
2819
2820 // Ensure that major progress is now clear
2821 C->clear_major_progress();
2822
2823 {
2824 // Verify that all previous optimizations produced a valid graph
2825 // at least to this point, even if no loop optimizations were done.
2826 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2827 PhaseIdealLoop::verify(igvn);
2828 }
2829
2830 if (range_check_cast_count() > 0) {
2831 // No more loop optimizations. Remove all range check dependent CastIINodes.
2832 C->remove_range_check_casts(igvn);
2833 igvn.optimize();
2834 }
2835
2836 #ifdef ASSERT
2837 bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion);
2838 #endif
2839
2840 bs->barrier_insertion_phase(C, igvn);
2841 if (failing()) return;
2842
2843 #ifdef ASSERT
2844 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2845 #endif
2846
2847 {
2848 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2849 PhaseMacroExpand mex(igvn);
2850 if (mex.expand_macro_nodes()) {
2851 assert(failing(), "must bail out w/ explicit message");
2852 return;
2853 }
2854 print_method(PHASE_MACRO_EXPANSION, 2);
2855 }
2856
2857 {
2858 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2859 if (bs->expand_barriers(this, igvn)) {
2860 assert(failing(), "must bail out w/ explicit message");
2861 return;
2862 }
2863 print_method(PHASE_BARRIER_EXPANSION, 2);
2864 }
2865
2866 if (opaque4_count() > 0) {
2867 C->remove_opaque4_nodes(igvn);
2868 igvn.optimize();
2869 }
2870
2871 DEBUG_ONLY( _modified_nodes = NULL; )
2872 } // (End scope of igvn; run destructor if necessary for asserts.)
2873
2874 process_print_inlining();
2875 // A method with only infinite loops has no edges entering loops from root
2876 {
2877 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2878 if (final_graph_reshaping()) {
2879 assert(failing(), "must bail out w/ explicit message");
2880 return;
2881 }
2882 }
2883
2884 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2885 }
2886
2887 //------------------------------Code_Gen---------------------------------------
2888 // Given a graph, generate code for it
2889 void Compile::Code_Gen() {
2890 if (failing()) {
2891 return;
2892 }
2893
2894 // Perform instruction selection. You might think we could reclaim Matcher
2895 // memory PDQ, but actually the Matcher is used in generating spill code.
2896 // Internals of the Matcher (including some VectorSets) must remain live
2897 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2898 // set a bit in reclaimed memory.
2899
2900 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2901 // nodes. Mapping is only valid at the root of each matched subtree.
2902 NOT_PRODUCT( verify_graph_edges(); )
2903
2904 Matcher matcher;
2905 _matcher = &matcher;
2906 {
2907 TracePhase tp("matcher", &timers[_t_matcher]);
2908 matcher.match();
2909 }
2910 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2911 // nodes. Mapping is only valid at the root of each matched subtree.
2912 NOT_PRODUCT( verify_graph_edges(); )
2913
2914 // If you have too many nodes, or if matching has failed, bail out
2915 check_node_count(0, "out of nodes matching instructions");
2916 if (failing()) {
2917 return;
2918 }
2919
2920 print_method(PHASE_MATCHING, 2);
2921
2922 // Build a proper-looking CFG
2923 PhaseCFG cfg(node_arena(), root(), matcher);
2924 _cfg = &cfg;
2925 {
2926 TracePhase tp("scheduler", &timers[_t_scheduler]);
2927 bool success = cfg.do_global_code_motion();
2928 if (!success) {
2929 return;
2930 }
2931
2932 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2933 NOT_PRODUCT( verify_graph_edges(); )
2934 debug_only( cfg.verify(); )
2935 }
2936
2937 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2938 _regalloc = ®alloc;
2939 {
2940 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2941 // Perform register allocation. After Chaitin, use-def chains are
2942 // no longer accurate (at spill code) and so must be ignored.
2943 // Node->LRG->reg mappings are still accurate.
2944 _regalloc->Register_Allocate();
2945
2946 // Bail out if the allocator builds too many nodes
2947 if (failing()) {
2948 return;
2949 }
2950 }
2951
2952 // Prior to register allocation we kept empty basic blocks in case the
2953 // the allocator needed a place to spill. After register allocation we
2954 // are not adding any new instructions. If any basic block is empty, we
2955 // can now safely remove it.
2956 {
2957 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2958 cfg.remove_empty_blocks();
2959 if (do_freq_based_layout()) {
2960 PhaseBlockLayout layout(cfg);
2961 } else {
2962 cfg.set_loop_alignment();
2963 }
2964 cfg.fixup_flow();
2965 }
2966
2967 // Apply peephole optimizations
2968 if( OptoPeephole ) {
2969 TracePhase tp("peephole", &timers[_t_peephole]);
2970 PhasePeephole peep( _regalloc, cfg);
2971 peep.do_transform();
2972 }
2973
2974 // Do late expand if CPU requires this.
2975 if (Matcher::require_postalloc_expand) {
2976 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2977 cfg.postalloc_expand(_regalloc);
2978 }
2979
2980 // Convert Nodes to instruction bits in a buffer
2981 {
2982 TraceTime tp("output", &timers[_t_output], CITime);
2983 Output();
2984 }
2985
2986 print_method(PHASE_FINAL_CODE);
2987
2988 // He's dead, Jim.
2989 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2990 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2991 }
2992
2993
2994 //------------------------------dump_asm---------------------------------------
2995 // Dump formatted assembly
2996 #if defined(SUPPORT_OPTO_ASSEMBLY)
2997 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2998
2999 int pc_digits = 3; // #chars required for pc
3000 int sb_chars = 3; // #chars for "start bundle" indicator
3001 int tab_size = 8;
3002 if (pcs != NULL) {
3003 int max_pc = 0;
3004 for (uint i = 0; i < pc_limit; i++) {
3005 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3006 }
3007 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3008 }
3009 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3010
3011 bool cut_short = false;
3012 st->print_cr("#");
3013 st->print("# "); _tf->dump_on(st); st->cr();
3014 st->print_cr("#");
3015
3016 // For all blocks
3017 int pc = 0x0; // Program counter
3018 char starts_bundle = ' ';
3019 _regalloc->dump_frame();
3020
3021 Node *n = NULL;
3022 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
3023 if (VMThread::should_terminate()) {
3024 cut_short = true;
3025 break;
3026 }
3027 Block* block = _cfg->get_block(i);
3028 if (block->is_connector() && !Verbose) {
3029 continue;
3030 }
3031 n = block->head();
3032 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3033 pc = pcs[n->_idx];
3034 st->print("%*.*x", pc_digits, pc_digits, pc);
3035 }
3036 st->fill_to(prefix_len);
3037 block->dump_head(_cfg, st);
3038 if (block->is_connector()) {
3039 st->fill_to(prefix_len);
3040 st->print_cr("# Empty connector block");
3041 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3042 st->fill_to(prefix_len);
3043 st->print_cr("# Block is sole successor of call");
3044 }
3045
3046 // For all instructions
3047 Node *delay = NULL;
3048 for (uint j = 0; j < block->number_of_nodes(); j++) {
3049 if (VMThread::should_terminate()) {
3050 cut_short = true;
3051 break;
3052 }
3053 n = block->get_node(j);
3054 if (valid_bundle_info(n)) {
3055 Bundle* bundle = node_bundling(n);
3056 if (bundle->used_in_unconditional_delay()) {
3057 delay = n;
3058 continue;
3059 }
3060 if (bundle->starts_bundle()) {
3061 starts_bundle = '+';
3062 }
3063 }
3064
3065 if (WizardMode) {
3066 n->dump();
3067 }
3068
3069 if( !n->is_Region() && // Dont print in the Assembly
3070 !n->is_Phi() && // a few noisely useless nodes
3071 !n->is_Proj() &&
3072 !n->is_MachTemp() &&
3073 !n->is_SafePointScalarObject() &&
3074 !n->is_Catch() && // Would be nice to print exception table targets
3075 !n->is_MergeMem() && // Not very interesting
3076 !n->is_top() && // Debug info table constants
3077 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3078 ) {
3079 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3080 pc = pcs[n->_idx];
3081 st->print("%*.*x", pc_digits, pc_digits, pc);
3082 } else {
3083 st->fill_to(pc_digits);
3084 }
3085 st->print(" %c ", starts_bundle);
3086 starts_bundle = ' ';
3087 st->fill_to(prefix_len);
3088 n->format(_regalloc, st);
3089 st->cr();
3090 }
3091
3092 // If we have an instruction with a delay slot, and have seen a delay,
3093 // then back up and print it
3094 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3095 // Coverity finding - Explicit null dereferenced.
3096 guarantee(delay != NULL, "no unconditional delay instruction");
3097 if (WizardMode) delay->dump();
3098
3099 if (node_bundling(delay)->starts_bundle())
3100 starts_bundle = '+';
3101 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3102 pc = pcs[n->_idx];
3103 st->print("%*.*x", pc_digits, pc_digits, pc);
3104 } else {
3105 st->fill_to(pc_digits);
3106 }
3107 st->print(" %c ", starts_bundle);
3108 starts_bundle = ' ';
3109 st->fill_to(prefix_len);
3110 delay->format(_regalloc, st);
3111 st->cr();
3112 delay = NULL;
3113 }
3114
3115 // Dump the exception table as well
3116 if( n->is_Catch() && (Verbose || WizardMode) ) {
3117 // Print the exception table for this offset
3118 _handler_table.print_subtable_for(pc);
3119 }
3120 st->bol(); // Make sure we start on a new line
3121 }
3122 st->cr(); // one empty line between blocks
3123 assert(cut_short || delay == NULL, "no unconditional delay branch");
3124 } // End of per-block dump
3125
3126 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3127 }
3128 #endif
3129
3130 //------------------------------Final_Reshape_Counts---------------------------
3131 // This class defines counters to help identify when a method
3132 // may/must be executed using hardware with only 24-bit precision.
3133 struct Final_Reshape_Counts : public StackObj {
3134 int _call_count; // count non-inlined 'common' calls
3135 int _float_count; // count float ops requiring 24-bit precision
3136 int _double_count; // count double ops requiring more precision
3137 int _java_call_count; // count non-inlined 'java' calls
3138 int _inner_loop_count; // count loops which need alignment
3139 VectorSet _visited; // Visitation flags
3140 Node_List _tests; // Set of IfNodes & PCTableNodes
3141
3142 Final_Reshape_Counts() :
3143 _call_count(0), _float_count(0), _double_count(0),
3144 _java_call_count(0), _inner_loop_count(0),
3145 _visited( Thread::current()->resource_area() ) { }
3146
3147 void inc_call_count () { _call_count ++; }
3148 void inc_float_count () { _float_count ++; }
3149 void inc_double_count() { _double_count++; }
3150 void inc_java_call_count() { _java_call_count++; }
3151 void inc_inner_loop_count() { _inner_loop_count++; }
3152
3153 int get_call_count () const { return _call_count ; }
3154 int get_float_count () const { return _float_count ; }
3155 int get_double_count() const { return _double_count; }
3156 int get_java_call_count() const { return _java_call_count; }
3157 int get_inner_loop_count() const { return _inner_loop_count; }
3158 };
3159
3160 #ifdef ASSERT
3161 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
3162 ciInstanceKlass *k = tp->klass()->as_instance_klass();
3163 // Make sure the offset goes inside the instance layout.
3164 return k->contains_field_offset(tp->offset());
3165 // Note that OffsetBot and OffsetTop are very negative.
3166 }
3167 #endif
3168
3169 // Eliminate trivially redundant StoreCMs and accumulate their
3170 // precedence edges.
3171 void Compile::eliminate_redundant_card_marks(Node* n) {
3172 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3173 if (n->in(MemNode::Address)->outcnt() > 1) {
3174 // There are multiple users of the same address so it might be
3175 // possible to eliminate some of the StoreCMs
3176 Node* mem = n->in(MemNode::Memory);
3177 Node* adr = n->in(MemNode::Address);
3178 Node* val = n->in(MemNode::ValueIn);
3179 Node* prev = n;
3180 bool done = false;
3181 // Walk the chain of StoreCMs eliminating ones that match. As
3182 // long as it's a chain of single users then the optimization is
3183 // safe. Eliminating partially redundant StoreCMs would require
3184 // cloning copies down the other paths.
3185 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3186 if (adr == mem->in(MemNode::Address) &&
3187 val == mem->in(MemNode::ValueIn)) {
3188 // redundant StoreCM
3189 if (mem->req() > MemNode::OopStore) {
3190 // Hasn't been processed by this code yet.
3191 n->add_prec(mem->in(MemNode::OopStore));
3192 } else {
3193 // Already converted to precedence edge
3194 for (uint i = mem->req(); i < mem->len(); i++) {
3195 // Accumulate any precedence edges
3196 if (mem->in(i) != NULL) {
3197 n->add_prec(mem->in(i));
3198 }
3199 }
3200 // Everything above this point has been processed.
3201 done = true;
3202 }
3203 // Eliminate the previous StoreCM
3204 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3205 assert(mem->outcnt() == 0, "should be dead");
3206 mem->disconnect_inputs(NULL, this);
3207 } else {
3208 prev = mem;
3209 }
3210 mem = prev->in(MemNode::Memory);
3211 }
3212 }
3213 }
3214
3215
3216 //------------------------------final_graph_reshaping_impl----------------------
3217 // Implement items 1-5 from final_graph_reshaping below.
3218 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3219
3220 if ( n->outcnt() == 0 ) return; // dead node
3221 uint nop = n->Opcode();
3222
3223 // Check for 2-input instruction with "last use" on right input.
3224 // Swap to left input. Implements item (2).
3225 if( n->req() == 3 && // two-input instruction
3226 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3227 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3228 n->in(2)->outcnt() == 1 &&// right use IS a last use
3229 !n->in(2)->is_Con() ) { // right use is not a constant
3230 // Check for commutative opcode
3231 switch( nop ) {
3232 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3233 case Op_MaxI: case Op_MinI:
3234 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3235 case Op_AndL: case Op_XorL: case Op_OrL:
3236 case Op_AndI: case Op_XorI: case Op_OrI: {
3237 // Move "last use" input to left by swapping inputs
3238 n->swap_edges(1, 2);
3239 break;
3240 }
3241 default:
3242 break;
3243 }
3244 }
3245
3246 #ifdef ASSERT
3247 if( n->is_Mem() ) {
3248 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3249 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3250 // oop will be recorded in oop map if load crosses safepoint
3251 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3252 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3253 "raw memory operations should have control edge");
3254 }
3255 if (n->is_MemBar()) {
3256 MemBarNode* mb = n->as_MemBar();
3257 if (mb->trailing_store() || mb->trailing_load_store()) {
3258 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3259 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3260 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3261 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3262 } else if (mb->leading()) {
3263 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3264 }
3265 }
3266 #endif
3267 // Count FPU ops and common calls, implements item (3)
3268 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3269 if (!gc_handled) {
3270 final_graph_reshaping_main_switch(n, frc, nop);
3271 }
3272
3273 // Collect CFG split points
3274 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3275 frc._tests.push(n);
3276 }
3277 }
3278
3279 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3280 switch( nop ) {
3281 // Count all float operations that may use FPU
3282 case Op_AddF:
3283 case Op_SubF:
3284 case Op_MulF:
3285 case Op_DivF:
3286 case Op_NegF:
3287 case Op_ModF:
3288 case Op_ConvI2F:
3289 case Op_ConF:
3290 case Op_CmpF:
3291 case Op_CmpF3:
3292 // case Op_ConvL2F: // longs are split into 32-bit halves
3293 frc.inc_float_count();
3294 break;
3295
3296 case Op_ConvF2D:
3297 case Op_ConvD2F:
3298 frc.inc_float_count();
3299 frc.inc_double_count();
3300 break;
3301
3302 // Count all double operations that may use FPU
3303 case Op_AddD:
3304 case Op_SubD:
3305 case Op_MulD:
3306 case Op_DivD:
3307 case Op_NegD:
3308 case Op_ModD:
3309 case Op_ConvI2D:
3310 case Op_ConvD2I:
3311 // case Op_ConvL2D: // handled by leaf call
3312 // case Op_ConvD2L: // handled by leaf call
3313 case Op_ConD:
3314 case Op_CmpD:
3315 case Op_CmpD3:
3316 frc.inc_double_count();
3317 break;
3318 case Op_Opaque1: // Remove Opaque Nodes before matching
3319 case Op_Opaque2: // Remove Opaque Nodes before matching
3320 case Op_Opaque3:
3321 n->subsume_by(n->in(1), this);
3322 break;
3323 case Op_CallStaticJava:
3324 case Op_CallJava:
3325 case Op_CallDynamicJava:
3326 frc.inc_java_call_count(); // Count java call site;
3327 case Op_CallRuntime:
3328 case Op_CallLeaf:
3329 case Op_CallLeafNoFP: {
3330 assert (n->is_Call(), "");
3331 CallNode *call = n->as_Call();
3332 // Count call sites where the FP mode bit would have to be flipped.
3333 // Do not count uncommon runtime calls:
3334 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3335 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3336 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3337 frc.inc_call_count(); // Count the call site
3338 } else { // See if uncommon argument is shared
3339 Node *n = call->in(TypeFunc::Parms);
3340 int nop = n->Opcode();
3341 // Clone shared simple arguments to uncommon calls, item (1).
3342 if (n->outcnt() > 1 &&
3343 !n->is_Proj() &&
3344 nop != Op_CreateEx &&
3345 nop != Op_CheckCastPP &&
3346 nop != Op_DecodeN &&
3347 nop != Op_DecodeNKlass &&
3348 !n->is_Mem() &&
3349 !n->is_Phi()) {
3350 Node *x = n->clone();
3351 call->set_req(TypeFunc::Parms, x);
3352 }
3353 }
3354 break;
3355 }
3356
3357 case Op_StoreD:
3358 case Op_LoadD:
3359 case Op_LoadD_unaligned:
3360 frc.inc_double_count();
3361 goto handle_mem;
3362 case Op_StoreF:
3363 case Op_LoadF:
3364 frc.inc_float_count();
3365 goto handle_mem;
3366
3367 case Op_StoreCM:
3368 {
3369 // Convert OopStore dependence into precedence edge
3370 Node* prec = n->in(MemNode::OopStore);
3371 n->del_req(MemNode::OopStore);
3372 n->add_prec(prec);
3373 eliminate_redundant_card_marks(n);
3374 }
3375
3376 // fall through
3377
3378 case Op_StoreB:
3379 case Op_StoreC:
3380 case Op_StorePConditional:
3381 case Op_StoreI:
3382 case Op_StoreL:
3383 case Op_StoreIConditional:
3384 case Op_StoreLConditional:
3385 case Op_CompareAndSwapB:
3386 case Op_CompareAndSwapS:
3387 case Op_CompareAndSwapI:
3388 case Op_CompareAndSwapL:
3389 case Op_CompareAndSwapP:
3390 case Op_CompareAndSwapN:
3391 case Op_WeakCompareAndSwapB:
3392 case Op_WeakCompareAndSwapS:
3393 case Op_WeakCompareAndSwapI:
3394 case Op_WeakCompareAndSwapL:
3395 case Op_WeakCompareAndSwapP:
3396 case Op_WeakCompareAndSwapN:
3397 case Op_CompareAndExchangeB:
3398 case Op_CompareAndExchangeS:
3399 case Op_CompareAndExchangeI:
3400 case Op_CompareAndExchangeL:
3401 case Op_CompareAndExchangeP:
3402 case Op_CompareAndExchangeN:
3403 case Op_GetAndAddS:
3404 case Op_GetAndAddB:
3405 case Op_GetAndAddI:
3406 case Op_GetAndAddL:
3407 case Op_GetAndSetS:
3408 case Op_GetAndSetB:
3409 case Op_GetAndSetI:
3410 case Op_GetAndSetL:
3411 case Op_GetAndSetP:
3412 case Op_GetAndSetN:
3413 case Op_StoreP:
3414 case Op_StoreN:
3415 case Op_StoreNKlass:
3416 case Op_LoadB:
3417 case Op_LoadUB:
3418 case Op_LoadUS:
3419 case Op_LoadI:
3420 case Op_LoadKlass:
3421 case Op_LoadNKlass:
3422 case Op_LoadL:
3423 case Op_LoadL_unaligned:
3424 case Op_LoadPLocked:
3425 case Op_LoadP:
3426 case Op_LoadN:
3427 case Op_LoadRange:
3428 case Op_LoadS: {
3429 handle_mem:
3430 #ifdef ASSERT
3431 if( VerifyOptoOopOffsets ) {
3432 MemNode* mem = n->as_Mem();
3433 // Check to see if address types have grounded out somehow.
3434 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3435 assert( !tp || oop_offset_is_sane(tp), "" );
3436 }
3437 #endif
3438 if (EnableValhalla && (nop == Op_LoadKlass || nop == Op_LoadNKlass)) {
3439 const TypeKlassPtr* tk = n->bottom_type()->make_ptr()->is_klassptr();
3440 assert(!tk->klass_is_exact(), "should have been folded");
3441 if (tk->klass()->can_be_value_array_klass() && n->as_Mem()->adr_type()->offset() == oopDesc::klass_offset_in_bytes()) {
3442 // Array load klass needs to filter out property bits (but not
3443 // GetNullFreePropertyNode or GetFlattenedPropertyNode which
3444 // needs to extract the storage property bits)
3445 uint last = unique();
3446 Node* pointer = NULL;
3447 if (nop == Op_LoadKlass) {
3448 Node* cast = new CastP2XNode(NULL, n);
3449 Node* masked = new LShiftXNode(cast, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3450 masked = new RShiftXNode(masked, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3451 pointer = new CastX2PNode(masked);
3452 pointer = new CheckCastPPNode(NULL, pointer, n->bottom_type());
3453 } else {
3454 Node* cast = new CastN2INode(n);
3455 Node* masked = new AndINode(cast, new ConINode(TypeInt::make(oopDesc::compressed_klass_mask())));
3456 pointer = new CastI2NNode(masked, n->bottom_type());
3457 }
3458 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3459 Node* u = n->fast_out(i);
3460 if (u->_idx < last && u->Opcode() != Op_GetNullFreeProperty && u->Opcode() != Op_GetFlattenedProperty) {
3461 // If user is a comparison with a klass that can't be a value type
3462 // array klass, we don't need to clear the storage property bits.
3463 Node* cmp = (u->is_DecodeNKlass() && u->outcnt() == 1) ? u->unique_out() : u;
3464 if (cmp->is_Cmp()) {
3465 const TypeKlassPtr* kp1 = cmp->in(1)->bottom_type()->make_ptr()->isa_klassptr();
3466 const TypeKlassPtr* kp2 = cmp->in(2)->bottom_type()->make_ptr()->isa_klassptr();
3467 if ((kp1 != NULL && !kp1->klass()->can_be_value_array_klass()) ||
3468 (kp2 != NULL && !kp2->klass()->can_be_value_array_klass())) {
3469 continue;
3470 }
3471 }
3472 int nb = u->replace_edge(n, pointer);
3473 --i, imax -= nb;
3474 }
3475 }
3476 }
3477 }
3478 break;
3479 }
3480
3481 case Op_AddP: { // Assert sane base pointers
3482 Node *addp = n->in(AddPNode::Address);
3483 assert( !addp->is_AddP() ||
3484 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3485 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3486 "Base pointers must match (addp %u)", addp->_idx );
3487 #ifdef _LP64
3488 if ((UseCompressedOops || UseCompressedClassPointers) &&
3489 addp->Opcode() == Op_ConP &&
3490 addp == n->in(AddPNode::Base) &&
3491 n->in(AddPNode::Offset)->is_Con()) {
3492 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3493 // on the platform and on the compressed oops mode.
3494 // Use addressing with narrow klass to load with offset on x86.
3495 // Some platforms can use the constant pool to load ConP.
3496 // Do this transformation here since IGVN will convert ConN back to ConP.
3497 const Type* t = addp->bottom_type();
3498 bool is_oop = t->isa_oopptr() != NULL;
3499 bool is_klass = t->isa_klassptr() != NULL;
3500
3501 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3502 (is_klass && Matcher::const_klass_prefer_decode())) {
3503 Node* nn = NULL;
3504
3505 int op = is_oop ? Op_ConN : Op_ConNKlass;
3506
3507 // Look for existing ConN node of the same exact type.
3508 Node* r = root();
3509 uint cnt = r->outcnt();
3510 for (uint i = 0; i < cnt; i++) {
3511 Node* m = r->raw_out(i);
3512 if (m!= NULL && m->Opcode() == op &&
3513 m->bottom_type()->make_ptr() == t) {
3514 nn = m;
3515 break;
3516 }
3517 }
3518 if (nn != NULL) {
3519 // Decode a narrow oop to match address
3520 // [R12 + narrow_oop_reg<<3 + offset]
3521 if (is_oop) {
3522 nn = new DecodeNNode(nn, t);
3523 } else {
3524 nn = new DecodeNKlassNode(nn, t);
3525 }
3526 // Check for succeeding AddP which uses the same Base.
3527 // Otherwise we will run into the assertion above when visiting that guy.
3528 for (uint i = 0; i < n->outcnt(); ++i) {
3529 Node *out_i = n->raw_out(i);
3530 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3531 out_i->set_req(AddPNode::Base, nn);
3532 #ifdef ASSERT
3533 for (uint j = 0; j < out_i->outcnt(); ++j) {
3534 Node *out_j = out_i->raw_out(j);
3535 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3536 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3537 }
3538 #endif
3539 }
3540 }
3541 n->set_req(AddPNode::Base, nn);
3542 n->set_req(AddPNode::Address, nn);
3543 if (addp->outcnt() == 0) {
3544 addp->disconnect_inputs(NULL, this);
3545 }
3546 }
3547 }
3548 }
3549 #endif
3550 // platform dependent reshaping of the address expression
3551 reshape_address(n->as_AddP());
3552 break;
3553 }
3554
3555 case Op_CastPP: {
3556 // Remove CastPP nodes to gain more freedom during scheduling but
3557 // keep the dependency they encode as control or precedence edges
3558 // (if control is set already) on memory operations. Some CastPP
3559 // nodes don't have a control (don't carry a dependency): skip
3560 // those.
3561 if (n->in(0) != NULL) {
3562 ResourceMark rm;
3563 Unique_Node_List wq;
3564 wq.push(n);
3565 for (uint next = 0; next < wq.size(); ++next) {
3566 Node *m = wq.at(next);
3567 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3568 Node* use = m->fast_out(i);
3569 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3570 use->ensure_control_or_add_prec(n->in(0));
3571 } else {
3572 switch(use->Opcode()) {
3573 case Op_AddP:
3574 case Op_DecodeN:
3575 case Op_DecodeNKlass:
3576 case Op_CheckCastPP:
3577 case Op_CastPP:
3578 wq.push(use);
3579 break;
3580 }
3581 }
3582 }
3583 }
3584 }
3585 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3586 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3587 Node* in1 = n->in(1);
3588 const Type* t = n->bottom_type();
3589 Node* new_in1 = in1->clone();
3590 new_in1->as_DecodeN()->set_type(t);
3591
3592 if (!Matcher::narrow_oop_use_complex_address()) {
3593 //
3594 // x86, ARM and friends can handle 2 adds in addressing mode
3595 // and Matcher can fold a DecodeN node into address by using
3596 // a narrow oop directly and do implicit NULL check in address:
3597 //
3598 // [R12 + narrow_oop_reg<<3 + offset]
3599 // NullCheck narrow_oop_reg
3600 //
3601 // On other platforms (Sparc) we have to keep new DecodeN node and
3602 // use it to do implicit NULL check in address:
3603 //
3604 // decode_not_null narrow_oop_reg, base_reg
3605 // [base_reg + offset]
3606 // NullCheck base_reg
3607 //
3608 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3609 // to keep the information to which NULL check the new DecodeN node
3610 // corresponds to use it as value in implicit_null_check().
3611 //
3612 new_in1->set_req(0, n->in(0));
3613 }
3614
3615 n->subsume_by(new_in1, this);
3616 if (in1->outcnt() == 0) {
3617 in1->disconnect_inputs(NULL, this);
3618 }
3619 } else {
3620 n->subsume_by(n->in(1), this);
3621 if (n->outcnt() == 0) {
3622 n->disconnect_inputs(NULL, this);
3623 }
3624 }
3625 break;
3626 }
3627 #ifdef _LP64
3628 case Op_CmpP:
3629 // Do this transformation here to preserve CmpPNode::sub() and
3630 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3631 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3632 Node* in1 = n->in(1);
3633 Node* in2 = n->in(2);
3634 if (!in1->is_DecodeNarrowPtr()) {
3635 in2 = in1;
3636 in1 = n->in(2);
3637 }
3638 assert(in1->is_DecodeNarrowPtr(), "sanity");
3639
3640 Node* new_in2 = NULL;
3641 if (in2->is_DecodeNarrowPtr()) {
3642 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3643 new_in2 = in2->in(1);
3644 } else if (in2->Opcode() == Op_ConP) {
3645 const Type* t = in2->bottom_type();
3646 if (t == TypePtr::NULL_PTR) {
3647 assert(in1->is_DecodeN(), "compare klass to null?");
3648 // Don't convert CmpP null check into CmpN if compressed
3649 // oops implicit null check is not generated.
3650 // This will allow to generate normal oop implicit null check.
3651 if (Matcher::gen_narrow_oop_implicit_null_checks())
3652 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3653 //
3654 // This transformation together with CastPP transformation above
3655 // will generated code for implicit NULL checks for compressed oops.
3656 //
3657 // The original code after Optimize()
3658 //
3659 // LoadN memory, narrow_oop_reg
3660 // decode narrow_oop_reg, base_reg
3661 // CmpP base_reg, NULL
3662 // CastPP base_reg // NotNull
3663 // Load [base_reg + offset], val_reg
3664 //
3665 // after these transformations will be
3666 //
3667 // LoadN memory, narrow_oop_reg
3668 // CmpN narrow_oop_reg, NULL
3669 // decode_not_null narrow_oop_reg, base_reg
3670 // Load [base_reg + offset], val_reg
3671 //
3672 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3673 // since narrow oops can be used in debug info now (see the code in
3674 // final_graph_reshaping_walk()).
3675 //
3676 // At the end the code will be matched to
3677 // on x86:
3678 //
3679 // Load_narrow_oop memory, narrow_oop_reg
3680 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3681 // NullCheck narrow_oop_reg
3682 //
3683 // and on sparc:
3684 //
3685 // Load_narrow_oop memory, narrow_oop_reg
3686 // decode_not_null narrow_oop_reg, base_reg
3687 // Load [base_reg + offset], val_reg
3688 // NullCheck base_reg
3689 //
3690 } else if (t->isa_oopptr()) {
3691 new_in2 = ConNode::make(t->make_narrowoop());
3692 } else if (t->isa_klassptr()) {
3693 new_in2 = ConNode::make(t->make_narrowklass());
3694 }
3695 }
3696 if (new_in2 != NULL) {
3697 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3698 n->subsume_by(cmpN, this);
3699 if (in1->outcnt() == 0) {
3700 in1->disconnect_inputs(NULL, this);
3701 }
3702 if (in2->outcnt() == 0) {
3703 in2->disconnect_inputs(NULL, this);
3704 }
3705 }
3706 }
3707 break;
3708
3709 case Op_DecodeN:
3710 case Op_DecodeNKlass:
3711 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3712 // DecodeN could be pinned when it can't be fold into
3713 // an address expression, see the code for Op_CastPP above.
3714 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3715 break;
3716
3717 case Op_EncodeP:
3718 case Op_EncodePKlass: {
3719 Node* in1 = n->in(1);
3720 if (in1->is_DecodeNarrowPtr()) {
3721 n->subsume_by(in1->in(1), this);
3722 } else if (in1->Opcode() == Op_ConP) {
3723 const Type* t = in1->bottom_type();
3724 if (t == TypePtr::NULL_PTR) {
3725 assert(t->isa_oopptr(), "null klass?");
3726 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3727 } else if (t->isa_oopptr()) {
3728 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3729 } else if (t->isa_klassptr()) {
3730 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3731 }
3732 }
3733 if (in1->outcnt() == 0) {
3734 in1->disconnect_inputs(NULL, this);
3735 }
3736 break;
3737 }
3738
3739 case Op_Proj: {
3740 if (OptimizeStringConcat) {
3741 ProjNode* p = n->as_Proj();
3742 if (p->_is_io_use) {
3743 // Separate projections were used for the exception path which
3744 // are normally removed by a late inline. If it wasn't inlined
3745 // then they will hang around and should just be replaced with
3746 // the original one.
3747 Node* proj = NULL;
3748 // Replace with just one
3749 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3750 Node *use = i.get();
3751 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3752 proj = use;
3753 break;
3754 }
3755 }
3756 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3757 if (proj != NULL) {
3758 p->subsume_by(proj, this);
3759 }
3760 }
3761 }
3762 break;
3763 }
3764
3765 case Op_Phi:
3766 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3767 // The EncodeP optimization may create Phi with the same edges
3768 // for all paths. It is not handled well by Register Allocator.
3769 Node* unique_in = n->in(1);
3770 assert(unique_in != NULL, "");
3771 uint cnt = n->req();
3772 for (uint i = 2; i < cnt; i++) {
3773 Node* m = n->in(i);
3774 assert(m != NULL, "");
3775 if (unique_in != m)
3776 unique_in = NULL;
3777 }
3778 if (unique_in != NULL) {
3779 n->subsume_by(unique_in, this);
3780 }
3781 }
3782 break;
3783
3784 #endif
3785
3786 #ifdef ASSERT
3787 case Op_CastII:
3788 // Verify that all range check dependent CastII nodes were removed.
3789 if (n->isa_CastII()->has_range_check()) {
3790 n->dump(3);
3791 assert(false, "Range check dependent CastII node was not removed");
3792 }
3793 break;
3794 #endif
3795
3796 case Op_ModI:
3797 if (UseDivMod) {
3798 // Check if a%b and a/b both exist
3799 Node* d = n->find_similar(Op_DivI);
3800 if (d) {
3801 // Replace them with a fused divmod if supported
3802 if (Matcher::has_match_rule(Op_DivModI)) {
3803 DivModINode* divmod = DivModINode::make(n);
3804 d->subsume_by(divmod->div_proj(), this);
3805 n->subsume_by(divmod->mod_proj(), this);
3806 } else {
3807 // replace a%b with a-((a/b)*b)
3808 Node* mult = new MulINode(d, d->in(2));
3809 Node* sub = new SubINode(d->in(1), mult);
3810 n->subsume_by(sub, this);
3811 }
3812 }
3813 }
3814 break;
3815
3816 case Op_ModL:
3817 if (UseDivMod) {
3818 // Check if a%b and a/b both exist
3819 Node* d = n->find_similar(Op_DivL);
3820 if (d) {
3821 // Replace them with a fused divmod if supported
3822 if (Matcher::has_match_rule(Op_DivModL)) {
3823 DivModLNode* divmod = DivModLNode::make(n);
3824 d->subsume_by(divmod->div_proj(), this);
3825 n->subsume_by(divmod->mod_proj(), this);
3826 } else {
3827 // replace a%b with a-((a/b)*b)
3828 Node* mult = new MulLNode(d, d->in(2));
3829 Node* sub = new SubLNode(d->in(1), mult);
3830 n->subsume_by(sub, this);
3831 }
3832 }
3833 }
3834 break;
3835
3836 case Op_LoadVector:
3837 case Op_StoreVector:
3838 break;
3839
3840 case Op_AddReductionVI:
3841 case Op_AddReductionVL:
3842 case Op_AddReductionVF:
3843 case Op_AddReductionVD:
3844 case Op_MulReductionVI:
3845 case Op_MulReductionVL:
3846 case Op_MulReductionVF:
3847 case Op_MulReductionVD:
3848 case Op_MinReductionV:
3849 case Op_MaxReductionV:
3850 break;
3851
3852 case Op_PackB:
3853 case Op_PackS:
3854 case Op_PackI:
3855 case Op_PackF:
3856 case Op_PackL:
3857 case Op_PackD:
3858 if (n->req()-1 > 2) {
3859 // Replace many operand PackNodes with a binary tree for matching
3860 PackNode* p = (PackNode*) n;
3861 Node* btp = p->binary_tree_pack(1, n->req());
3862 n->subsume_by(btp, this);
3863 }
3864 break;
3865 case Op_Loop:
3866 case Op_CountedLoop:
3867 case Op_OuterStripMinedLoop:
3868 if (n->as_Loop()->is_inner_loop()) {
3869 frc.inc_inner_loop_count();
3870 }
3871 n->as_Loop()->verify_strip_mined(0);
3872 break;
3873 case Op_LShiftI:
3874 case Op_RShiftI:
3875 case Op_URShiftI:
3876 case Op_LShiftL:
3877 case Op_RShiftL:
3878 case Op_URShiftL:
3879 if (Matcher::need_masked_shift_count) {
3880 // The cpu's shift instructions don't restrict the count to the
3881 // lower 5/6 bits. We need to do the masking ourselves.
3882 Node* in2 = n->in(2);
3883 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3884 const TypeInt* t = in2->find_int_type();
3885 if (t != NULL && t->is_con()) {
3886 juint shift = t->get_con();
3887 if (shift > mask) { // Unsigned cmp
3888 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3889 }
3890 } else {
3891 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3892 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3893 n->set_req(2, shift);
3894 }
3895 }
3896 if (in2->outcnt() == 0) { // Remove dead node
3897 in2->disconnect_inputs(NULL, this);
3898 }
3899 }
3900 break;
3901 case Op_MemBarStoreStore:
3902 case Op_MemBarRelease:
3903 // Break the link with AllocateNode: it is no longer useful and
3904 // confuses register allocation.
3905 if (n->req() > MemBarNode::Precedent) {
3906 n->set_req(MemBarNode::Precedent, top());
3907 }
3908 break;
3909 case Op_MemBarAcquire: {
3910 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3911 // At parse time, the trailing MemBarAcquire for a volatile load
3912 // is created with an edge to the load. After optimizations,
3913 // that input may be a chain of Phis. If those phis have no
3914 // other use, then the MemBarAcquire keeps them alive and
3915 // register allocation can be confused.
3916 ResourceMark rm;
3917 Unique_Node_List wq;
3918 wq.push(n->in(MemBarNode::Precedent));
3919 n->set_req(MemBarNode::Precedent, top());
3920 while (wq.size() > 0) {
3921 Node* m = wq.pop();
3922 if (m->outcnt() == 0) {
3923 for (uint j = 0; j < m->req(); j++) {
3924 Node* in = m->in(j);
3925 if (in != NULL) {
3926 wq.push(in);
3927 }
3928 }
3929 m->disconnect_inputs(NULL, this);
3930 }
3931 }
3932 }
3933 break;
3934 }
3935 case Op_RangeCheck: {
3936 RangeCheckNode* rc = n->as_RangeCheck();
3937 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3938 n->subsume_by(iff, this);
3939 frc._tests.push(iff);
3940 break;
3941 }
3942 case Op_ConvI2L: {
3943 if (!Matcher::convi2l_type_required) {
3944 // Code generation on some platforms doesn't need accurate
3945 // ConvI2L types. Widening the type can help remove redundant
3946 // address computations.
3947 n->as_Type()->set_type(TypeLong::INT);
3948 ResourceMark rm;
3949 Node_List wq;
3950 wq.push(n);
3951 for (uint next = 0; next < wq.size(); next++) {
3952 Node *m = wq.at(next);
3953
3954 for(;;) {
3955 // Loop over all nodes with identical inputs edges as m
3956 Node* k = m->find_similar(m->Opcode());
3957 if (k == NULL) {
3958 break;
3959 }
3960 // Push their uses so we get a chance to remove node made
3961 // redundant
3962 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3963 Node* u = k->fast_out(i);
3964 assert(!wq.contains(u), "shouldn't process one node several times");
3965 if (u->Opcode() == Op_LShiftL ||
3966 u->Opcode() == Op_AddL ||
3967 u->Opcode() == Op_SubL ||
3968 u->Opcode() == Op_AddP) {
3969 wq.push(u);
3970 }
3971 }
3972 // Replace all nodes with identical edges as m with m
3973 k->subsume_by(m, this);
3974 }
3975 }
3976 }
3977 break;
3978 }
3979 case Op_CmpUL: {
3980 if (!Matcher::has_match_rule(Op_CmpUL)) {
3981 // No support for unsigned long comparisons
3982 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3983 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3984 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3985 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3986 Node* andl = new AndLNode(orl, remove_sign_mask);
3987 Node* cmp = new CmpLNode(andl, n->in(2));
3988 n->subsume_by(cmp, this);
3989 }
3990 break;
3991 }
3992 #ifdef ASSERT
3993 case Op_ValueTypePtr:
3994 case Op_ValueType: {
3995 n->dump(-1);
3996 assert(false, "value type node was not removed");
3997 break;
3998 }
3999 #endif
4000 case Op_GetNullFreeProperty:
4001 case Op_GetFlattenedProperty: {
4002 // Extract the null free bits
4003 uint last = unique();
4004 Node* null_free = NULL;
4005 int bit = nop == Op_GetNullFreeProperty ? ArrayStorageProperties::null_free_bit : ArrayStorageProperties::flattened_bit;
4006 if (n->in(1)->Opcode() == Op_LoadKlass) {
4007 Node* cast = new CastP2XNode(NULL, n->in(1));
4008 null_free = new AndLNode(cast, new ConLNode(TypeLong::make(((jlong)1)<<(oopDesc::wide_storage_props_shift + bit))));
4009 } else {
4010 assert(n->in(1)->Opcode() == Op_LoadNKlass, "not a compressed klass?");
4011 Node* cast = new CastN2INode(n->in(1));
4012 null_free = new AndINode(cast, new ConINode(TypeInt::make(1<<(oopDesc::narrow_storage_props_shift + bit))));
4013 }
4014 n->subsume_by(null_free, this);
4015 break;
4016 }
4017 default:
4018 assert(!n->is_Call(), "");
4019 assert(!n->is_Mem(), "");
4020 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4021 break;
4022 }
4023 }
4024
4025 //------------------------------final_graph_reshaping_walk---------------------
4026 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4027 // requires that the walk visits a node's inputs before visiting the node.
4028 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4029 ResourceArea *area = Thread::current()->resource_area();
4030 Unique_Node_List sfpt(area);
4031
4032 frc._visited.set(root->_idx); // first, mark node as visited
4033 uint cnt = root->req();
4034 Node *n = root;
4035 uint i = 0;
4036 while (true) {
4037 if (i < cnt) {
4038 // Place all non-visited non-null inputs onto stack
4039 Node* m = n->in(i);
4040 ++i;
4041 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4042 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4043 // compute worst case interpreter size in case of a deoptimization
4044 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4045
4046 sfpt.push(m);
4047 }
4048 cnt = m->req();
4049 nstack.push(n, i); // put on stack parent and next input's index
4050 n = m;
4051 i = 0;
4052 }
4053 } else {
4054 // Now do post-visit work
4055 final_graph_reshaping_impl( n, frc );
4056 if (nstack.is_empty())
4057 break; // finished
4058 n = nstack.node(); // Get node from stack
4059 cnt = n->req();
4060 i = nstack.index();
4061 nstack.pop(); // Shift to the next node on stack
4062 }
4063 }
4064
4065 // Skip next transformation if compressed oops are not used.
4066 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4067 (!UseCompressedOops && !UseCompressedClassPointers))
4068 return;
4069
4070 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4071 // It could be done for an uncommon traps or any safepoints/calls
4072 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4073 while (sfpt.size() > 0) {
4074 n = sfpt.pop();
4075 JVMState *jvms = n->as_SafePoint()->jvms();
4076 assert(jvms != NULL, "sanity");
4077 int start = jvms->debug_start();
4078 int end = n->req();
4079 bool is_uncommon = (n->is_CallStaticJava() &&
4080 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4081 for (int j = start; j < end; j++) {
4082 Node* in = n->in(j);
4083 if (in->is_DecodeNarrowPtr()) {
4084 bool safe_to_skip = true;
4085 if (!is_uncommon ) {
4086 // Is it safe to skip?
4087 for (uint i = 0; i < in->outcnt(); i++) {
4088 Node* u = in->raw_out(i);
4089 if (!u->is_SafePoint() ||
4090 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4091 safe_to_skip = false;
4092 }
4093 }
4094 }
4095 if (safe_to_skip) {
4096 n->set_req(j, in->in(1));
4097 }
4098 if (in->outcnt() == 0) {
4099 in->disconnect_inputs(NULL, this);
4100 }
4101 }
4102 }
4103 }
4104 }
4105
4106 //------------------------------final_graph_reshaping--------------------------
4107 // Final Graph Reshaping.
4108 //
4109 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4110 // and not commoned up and forced early. Must come after regular
4111 // optimizations to avoid GVN undoing the cloning. Clone constant
4112 // inputs to Loop Phis; these will be split by the allocator anyways.
4113 // Remove Opaque nodes.
4114 // (2) Move last-uses by commutative operations to the left input to encourage
4115 // Intel update-in-place two-address operations and better register usage
4116 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4117 // calls canonicalizing them back.
4118 // (3) Count the number of double-precision FP ops, single-precision FP ops
4119 // and call sites. On Intel, we can get correct rounding either by
4120 // forcing singles to memory (requires extra stores and loads after each
4121 // FP bytecode) or we can set a rounding mode bit (requires setting and
4122 // clearing the mode bit around call sites). The mode bit is only used
4123 // if the relative frequency of single FP ops to calls is low enough.
4124 // This is a key transform for SPEC mpeg_audio.
4125 // (4) Detect infinite loops; blobs of code reachable from above but not
4126 // below. Several of the Code_Gen algorithms fail on such code shapes,
4127 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4128 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4129 // Detection is by looking for IfNodes where only 1 projection is
4130 // reachable from below or CatchNodes missing some targets.
4131 // (5) Assert for insane oop offsets in debug mode.
4132
4133 bool Compile::final_graph_reshaping() {
4134 // an infinite loop may have been eliminated by the optimizer,
4135 // in which case the graph will be empty.
4136 if (root()->req() == 1) {
4137 record_method_not_compilable("trivial infinite loop");
4138 return true;
4139 }
4140
4141 // Expensive nodes have their control input set to prevent the GVN
4142 // from freely commoning them. There's no GVN beyond this point so
4143 // no need to keep the control input. We want the expensive nodes to
4144 // be freely moved to the least frequent code path by gcm.
4145 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4146 for (int i = 0; i < expensive_count(); i++) {
4147 _expensive_nodes->at(i)->set_req(0, NULL);
4148 }
4149
4150 Final_Reshape_Counts frc;
4151
4152 // Visit everybody reachable!
4153 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4154 Node_Stack nstack(live_nodes() >> 1);
4155 final_graph_reshaping_walk(nstack, root(), frc);
4156
4157 // Check for unreachable (from below) code (i.e., infinite loops).
4158 for( uint i = 0; i < frc._tests.size(); i++ ) {
4159 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4160 // Get number of CFG targets.
4161 // Note that PCTables include exception targets after calls.
4162 uint required_outcnt = n->required_outcnt();
4163 if (n->outcnt() != required_outcnt) {
4164 // Check for a few special cases. Rethrow Nodes never take the
4165 // 'fall-thru' path, so expected kids is 1 less.
4166 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4167 if (n->in(0)->in(0)->is_Call()) {
4168 CallNode *call = n->in(0)->in(0)->as_Call();
4169 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4170 required_outcnt--; // Rethrow always has 1 less kid
4171 } else if (call->req() > TypeFunc::Parms &&
4172 call->is_CallDynamicJava()) {
4173 // Check for null receiver. In such case, the optimizer has
4174 // detected that the virtual call will always result in a null
4175 // pointer exception. The fall-through projection of this CatchNode
4176 // will not be populated.
4177 Node *arg0 = call->in(TypeFunc::Parms);
4178 if (arg0->is_Type() &&
4179 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4180 required_outcnt--;
4181 }
4182 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4183 call->req() > TypeFunc::Parms+1 &&
4184 call->is_CallStaticJava()) {
4185 // Check for negative array length. In such case, the optimizer has
4186 // detected that the allocation attempt will always result in an
4187 // exception. There is no fall-through projection of this CatchNode .
4188 Node *arg1 = call->in(TypeFunc::Parms+1);
4189 if (arg1->is_Type() &&
4190 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4191 required_outcnt--;
4192 }
4193 }
4194 }
4195 }
4196 // Recheck with a better notion of 'required_outcnt'
4197 if (n->outcnt() != required_outcnt) {
4198 record_method_not_compilable("malformed control flow");
4199 return true; // Not all targets reachable!
4200 }
4201 }
4202 // Check that I actually visited all kids. Unreached kids
4203 // must be infinite loops.
4204 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4205 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4206 record_method_not_compilable("infinite loop");
4207 return true; // Found unvisited kid; must be unreach
4208 }
4209
4210 // Here so verification code in final_graph_reshaping_walk()
4211 // always see an OuterStripMinedLoopEnd
4212 if (n->is_OuterStripMinedLoopEnd()) {
4213 IfNode* init_iff = n->as_If();
4214 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4215 n->subsume_by(iff, this);
4216 }
4217 }
4218
4219 // If original bytecodes contained a mixture of floats and doubles
4220 // check if the optimizer has made it homogenous, item (3).
4221 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
4222 frc.get_float_count() > 32 &&
4223 frc.get_double_count() == 0 &&
4224 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4225 set_24_bit_selection_and_mode( false, true );
4226 }
4227
4228 set_java_calls(frc.get_java_call_count());
4229 set_inner_loops(frc.get_inner_loop_count());
4230
4231 // No infinite loops, no reason to bail out.
4232 return false;
4233 }
4234
4235 //-----------------------------too_many_traps----------------------------------
4236 // Report if there are too many traps at the current method and bci.
4237 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4238 bool Compile::too_many_traps(ciMethod* method,
4239 int bci,
4240 Deoptimization::DeoptReason reason) {
4241 ciMethodData* md = method->method_data();
4242 if (md->is_empty()) {
4243 // Assume the trap has not occurred, or that it occurred only
4244 // because of a transient condition during start-up in the interpreter.
4245 return false;
4246 }
4247 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4248 if (md->has_trap_at(bci, m, reason) != 0) {
4249 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4250 // Also, if there are multiple reasons, or if there is no per-BCI record,
4251 // assume the worst.
4252 if (log())
4253 log()->elem("observe trap='%s' count='%d'",
4254 Deoptimization::trap_reason_name(reason),
4255 md->trap_count(reason));
4256 return true;
4257 } else {
4258 // Ignore method/bci and see if there have been too many globally.
4259 return too_many_traps(reason, md);
4260 }
4261 }
4262
4263 // Less-accurate variant which does not require a method and bci.
4264 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4265 ciMethodData* logmd) {
4266 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4267 // Too many traps globally.
4268 // Note that we use cumulative trap_count, not just md->trap_count.
4269 if (log()) {
4270 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4271 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4272 Deoptimization::trap_reason_name(reason),
4273 mcount, trap_count(reason));
4274 }
4275 return true;
4276 } else {
4277 // The coast is clear.
4278 return false;
4279 }
4280 }
4281
4282 //--------------------------too_many_recompiles--------------------------------
4283 // Report if there are too many recompiles at the current method and bci.
4284 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4285 // Is not eager to return true, since this will cause the compiler to use
4286 // Action_none for a trap point, to avoid too many recompilations.
4287 bool Compile::too_many_recompiles(ciMethod* method,
4288 int bci,
4289 Deoptimization::DeoptReason reason) {
4290 ciMethodData* md = method->method_data();
4291 if (md->is_empty()) {
4292 // Assume the trap has not occurred, or that it occurred only
4293 // because of a transient condition during start-up in the interpreter.
4294 return false;
4295 }
4296 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4297 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4298 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4299 Deoptimization::DeoptReason per_bc_reason
4300 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4301 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4302 if ((per_bc_reason == Deoptimization::Reason_none
4303 || md->has_trap_at(bci, m, reason) != 0)
4304 // The trap frequency measure we care about is the recompile count:
4305 && md->trap_recompiled_at(bci, m)
4306 && md->overflow_recompile_count() >= bc_cutoff) {
4307 // Do not emit a trap here if it has already caused recompilations.
4308 // Also, if there are multiple reasons, or if there is no per-BCI record,
4309 // assume the worst.
4310 if (log())
4311 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4312 Deoptimization::trap_reason_name(reason),
4313 md->trap_count(reason),
4314 md->overflow_recompile_count());
4315 return true;
4316 } else if (trap_count(reason) != 0
4317 && decompile_count() >= m_cutoff) {
4318 // Too many recompiles globally, and we have seen this sort of trap.
4319 // Use cumulative decompile_count, not just md->decompile_count.
4320 if (log())
4321 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4322 Deoptimization::trap_reason_name(reason),
4323 md->trap_count(reason), trap_count(reason),
4324 md->decompile_count(), decompile_count());
4325 return true;
4326 } else {
4327 // The coast is clear.
4328 return false;
4329 }
4330 }
4331
4332 // Compute when not to trap. Used by matching trap based nodes and
4333 // NullCheck optimization.
4334 void Compile::set_allowed_deopt_reasons() {
4335 _allowed_reasons = 0;
4336 if (is_method_compilation()) {
4337 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4338 assert(rs < BitsPerInt, "recode bit map");
4339 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4340 _allowed_reasons |= nth_bit(rs);
4341 }
4342 }
4343 }
4344 }
4345
4346 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4347 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4348 }
4349
4350 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4351 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4352 }
4353
4354 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4355 if (holder->is_initialized()) {
4356 return false;
4357 }
4358 if (holder->is_being_initialized()) {
4359 if (accessing_method->holder() == holder) {
4360 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4361 // <init>, or a static method. In all those cases, there was an initialization
4362 // barrier on the holder klass passed.
4363 if (accessing_method->is_class_initializer() ||
4364 accessing_method->is_object_constructor() ||
4365 accessing_method->is_static()) {
4366 return false;
4367 }
4368 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4369 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4370 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4371 // child class can become fully initialized while its parent class is still being initialized.
4372 if (accessing_method->is_class_initializer()) {
4373 return false;
4374 }
4375 }
4376 ciMethod* root = method(); // the root method of compilation
4377 if (root != accessing_method) {
4378 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4379 }
4380 }
4381 return true;
4382 }
4383
4384 #ifndef PRODUCT
4385 //------------------------------verify_graph_edges---------------------------
4386 // Walk the Graph and verify that there is a one-to-one correspondence
4387 // between Use-Def edges and Def-Use edges in the graph.
4388 void Compile::verify_graph_edges(bool no_dead_code) {
4389 if (VerifyGraphEdges) {
4390 ResourceArea *area = Thread::current()->resource_area();
4391 Unique_Node_List visited(area);
4392 // Call recursive graph walk to check edges
4393 _root->verify_edges(visited);
4394 if (no_dead_code) {
4395 // Now make sure that no visited node is used by an unvisited node.
4396 bool dead_nodes = false;
4397 Unique_Node_List checked(area);
4398 while (visited.size() > 0) {
4399 Node* n = visited.pop();
4400 checked.push(n);
4401 for (uint i = 0; i < n->outcnt(); i++) {
4402 Node* use = n->raw_out(i);
4403 if (checked.member(use)) continue; // already checked
4404 if (visited.member(use)) continue; // already in the graph
4405 if (use->is_Con()) continue; // a dead ConNode is OK
4406 // At this point, we have found a dead node which is DU-reachable.
4407 if (!dead_nodes) {
4408 tty->print_cr("*** Dead nodes reachable via DU edges:");
4409 dead_nodes = true;
4410 }
4411 use->dump(2);
4412 tty->print_cr("---");
4413 checked.push(use); // No repeats; pretend it is now checked.
4414 }
4415 }
4416 assert(!dead_nodes, "using nodes must be reachable from root");
4417 }
4418 }
4419 }
4420 #endif
4421
4422 // The Compile object keeps track of failure reasons separately from the ciEnv.
4423 // This is required because there is not quite a 1-1 relation between the
4424 // ciEnv and its compilation task and the Compile object. Note that one
4425 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4426 // to backtrack and retry without subsuming loads. Other than this backtracking
4427 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4428 // by the logic in C2Compiler.
4429 void Compile::record_failure(const char* reason) {
4430 if (log() != NULL) {
4431 log()->elem("failure reason='%s' phase='compile'", reason);
4432 }
4433 if (_failure_reason == NULL) {
4434 // Record the first failure reason.
4435 _failure_reason = reason;
4436 }
4437
4438 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4439 C->print_method(PHASE_FAILURE);
4440 }
4441 _root = NULL; // flush the graph, too
4442 }
4443
4444 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4445 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4446 _phase_name(name), _dolog(CITimeVerbose)
4447 {
4448 if (_dolog) {
4449 C = Compile::current();
4450 _log = C->log();
4451 } else {
4452 C = NULL;
4453 _log = NULL;
4454 }
4455 if (_log != NULL) {
4456 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4457 _log->stamp();
4458 _log->end_head();
4459 }
4460 }
4461
4462 Compile::TracePhase::~TracePhase() {
4463
4464 C = Compile::current();
4465 if (_dolog) {
4466 _log = C->log();
4467 } else {
4468 _log = NULL;
4469 }
4470
4471 #ifdef ASSERT
4472 if (PrintIdealNodeCount) {
4473 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4474 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4475 }
4476
4477 if (VerifyIdealNodeCount) {
4478 Compile::current()->print_missing_nodes();
4479 }
4480 #endif
4481
4482 if (_log != NULL) {
4483 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4484 }
4485 }
4486
4487 //=============================================================================
4488 // Two Constant's are equal when the type and the value are equal.
4489 bool Compile::Constant::operator==(const Constant& other) {
4490 if (type() != other.type() ) return false;
4491 if (can_be_reused() != other.can_be_reused()) return false;
4492 // For floating point values we compare the bit pattern.
4493 switch (type()) {
4494 case T_INT:
4495 case T_FLOAT: return (_v._value.i == other._v._value.i);
4496 case T_LONG:
4497 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4498 case T_OBJECT:
4499 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4500 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4501 case T_METADATA: return (_v._metadata == other._v._metadata);
4502 default: ShouldNotReachHere(); return false;
4503 }
4504 }
4505
4506 static int type_to_size_in_bytes(BasicType t) {
4507 switch (t) {
4508 case T_INT: return sizeof(jint );
4509 case T_LONG: return sizeof(jlong );
4510 case T_FLOAT: return sizeof(jfloat );
4511 case T_DOUBLE: return sizeof(jdouble);
4512 case T_METADATA: return sizeof(Metadata*);
4513 // We use T_VOID as marker for jump-table entries (labels) which
4514 // need an internal word relocation.
4515 case T_VOID:
4516 case T_ADDRESS:
4517 case T_OBJECT: return sizeof(jobject);
4518 default:
4519 ShouldNotReachHere();
4520 return -1;
4521 }
4522 }
4523
4524 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4525 // sort descending
4526 if (a->freq() > b->freq()) return -1;
4527 if (a->freq() < b->freq()) return 1;
4528 return 0;
4529 }
4530
4531 void Compile::ConstantTable::calculate_offsets_and_size() {
4532 // First, sort the array by frequencies.
4533 _constants.sort(qsort_comparator);
4534
4535 #ifdef ASSERT
4536 // Make sure all jump-table entries were sorted to the end of the
4537 // array (they have a negative frequency).
4538 bool found_void = false;
4539 for (int i = 0; i < _constants.length(); i++) {
4540 Constant con = _constants.at(i);
4541 if (con.type() == T_VOID)
4542 found_void = true; // jump-tables
4543 else
4544 assert(!found_void, "wrong sorting");
4545 }
4546 #endif
4547
4548 int offset = 0;
4549 for (int i = 0; i < _constants.length(); i++) {
4550 Constant* con = _constants.adr_at(i);
4551
4552 // Align offset for type.
4553 int typesize = type_to_size_in_bytes(con->type());
4554 offset = align_up(offset, typesize);
4555 con->set_offset(offset); // set constant's offset
4556
4557 if (con->type() == T_VOID) {
4558 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4559 offset = offset + typesize * n->outcnt(); // expand jump-table
4560 } else {
4561 offset = offset + typesize;
4562 }
4563 }
4564
4565 // Align size up to the next section start (which is insts; see
4566 // CodeBuffer::align_at_start).
4567 assert(_size == -1, "already set?");
4568 _size = align_up(offset, (int)CodeEntryAlignment);
4569 }
4570
4571 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4572 MacroAssembler _masm(&cb);
4573 for (int i = 0; i < _constants.length(); i++) {
4574 Constant con = _constants.at(i);
4575 address constant_addr = NULL;
4576 switch (con.type()) {
4577 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4578 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4579 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4580 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4581 case T_OBJECT: {
4582 jobject obj = con.get_jobject();
4583 int oop_index = _masm.oop_recorder()->find_index(obj);
4584 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4585 break;
4586 }
4587 case T_ADDRESS: {
4588 address addr = (address) con.get_jobject();
4589 constant_addr = _masm.address_constant(addr);
4590 break;
4591 }
4592 // We use T_VOID as marker for jump-table entries (labels) which
4593 // need an internal word relocation.
4594 case T_VOID: {
4595 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4596 // Fill the jump-table with a dummy word. The real value is
4597 // filled in later in fill_jump_table.
4598 address dummy = (address) n;
4599 constant_addr = _masm.address_constant(dummy);
4600 // Expand jump-table
4601 for (uint i = 1; i < n->outcnt(); i++) {
4602 address temp_addr = _masm.address_constant(dummy + i);
4603 assert(temp_addr, "consts section too small");
4604 }
4605 break;
4606 }
4607 case T_METADATA: {
4608 Metadata* obj = con.get_metadata();
4609 int metadata_index = _masm.oop_recorder()->find_index(obj);
4610 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4611 break;
4612 }
4613 default: ShouldNotReachHere();
4614 }
4615 assert(constant_addr, "consts section too small");
4616 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4617 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4618 }
4619 }
4620
4621 int Compile::ConstantTable::find_offset(Constant& con) const {
4622 int idx = _constants.find(con);
4623 guarantee(idx != -1, "constant must be in constant table");
4624 int offset = _constants.at(idx).offset();
4625 guarantee(offset != -1, "constant table not emitted yet?");
4626 return offset;
4627 }
4628
4629 void Compile::ConstantTable::add(Constant& con) {
4630 if (con.can_be_reused()) {
4631 int idx = _constants.find(con);
4632 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4633 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4634 return;
4635 }
4636 }
4637 (void) _constants.append(con);
4638 }
4639
4640 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4641 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4642 Constant con(type, value, b->_freq);
4643 add(con);
4644 return con;
4645 }
4646
4647 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4648 Constant con(metadata);
4649 add(con);
4650 return con;
4651 }
4652
4653 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4654 jvalue value;
4655 BasicType type = oper->type()->basic_type();
4656 switch (type) {
4657 case T_LONG: value.j = oper->constantL(); break;
4658 case T_FLOAT: value.f = oper->constantF(); break;
4659 case T_DOUBLE: value.d = oper->constantD(); break;
4660 case T_OBJECT:
4661 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4662 case T_METADATA: return add((Metadata*)oper->constant()); break;
4663 default: guarantee(false, "unhandled type: %s", type2name(type));
4664 }
4665 return add(n, type, value);
4666 }
4667
4668 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4669 jvalue value;
4670 // We can use the node pointer here to identify the right jump-table
4671 // as this method is called from Compile::Fill_buffer right before
4672 // the MachNodes are emitted and the jump-table is filled (means the
4673 // MachNode pointers do not change anymore).
4674 value.l = (jobject) n;
4675 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4676 add(con);
4677 return con;
4678 }
4679
4680 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4681 // If called from Compile::scratch_emit_size do nothing.
4682 if (Compile::current()->in_scratch_emit_size()) return;
4683
4684 assert(labels.is_nonempty(), "must be");
4685 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4686
4687 // Since MachConstantNode::constant_offset() also contains
4688 // table_base_offset() we need to subtract the table_base_offset()
4689 // to get the plain offset into the constant table.
4690 int offset = n->constant_offset() - table_base_offset();
4691
4692 MacroAssembler _masm(&cb);
4693 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4694
4695 for (uint i = 0; i < n->outcnt(); i++) {
4696 address* constant_addr = &jump_table_base[i];
4697 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));
4698 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4699 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4700 }
4701 }
4702
4703 //----------------------------static_subtype_check-----------------------------
4704 // Shortcut important common cases when superklass is exact:
4705 // (0) superklass is java.lang.Object (can occur in reflective code)
4706 // (1) subklass is already limited to a subtype of superklass => always ok
4707 // (2) subklass does not overlap with superklass => always fail
4708 // (3) superklass has NO subtypes and we can check with a simple compare.
4709 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4710 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4711 return SSC_full_test; // Let caller generate the general case.
4712 }
4713
4714 if (superk == env()->Object_klass()) {
4715 return SSC_always_true; // (0) this test cannot fail
4716 }
4717
4718 ciType* superelem = superk;
4719 if (superelem->is_array_klass()) {
4720 ciArrayKlass* ak = superelem->as_array_klass();
4721 superelem = superelem->as_array_klass()->base_element_type();
4722 }
4723
4724 if (!subk->is_interface()) { // cannot trust static interface types yet
4725 if (subk->is_subtype_of(superk)) {
4726 return SSC_always_true; // (1) false path dead; no dynamic test needed
4727 }
4728 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4729 !superk->is_subtype_of(subk)) {
4730 return SSC_always_false;
4731 }
4732 }
4733
4734 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4735 // [V is a subtype of [V? but the klass for [V is not equal to the klass for [V?. Perform a full test.
4736 if (superk->is_obj_array_klass() && !superk->as_array_klass()->storage_properties().is_null_free() && superk->as_array_klass()->element_klass()->is_valuetype()) {
4737 return SSC_full_test;
4738 }
4739 // If casting to an instance klass, it must have no subtypes
4740 if (superk->is_interface()) {
4741 // Cannot trust interfaces yet.
4742 // %%% S.B. superk->nof_implementors() == 1
4743 } else if (superelem->is_instance_klass()) {
4744 ciInstanceKlass* ik = superelem->as_instance_klass();
4745 if (!ik->has_subklass() && !ik->is_interface()) {
4746 if (!ik->is_final()) {
4747 // Add a dependency if there is a chance of a later subclass.
4748 dependencies()->assert_leaf_type(ik);
4749 }
4750 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4751 }
4752 } else {
4753 // A primitive array type has no subtypes.
4754 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4755 }
4756
4757 return SSC_full_test;
4758 }
4759
4760 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4761 #ifdef _LP64
4762 // The scaled index operand to AddP must be a clean 64-bit value.
4763 // Java allows a 32-bit int to be incremented to a negative
4764 // value, which appears in a 64-bit register as a large
4765 // positive number. Using that large positive number as an
4766 // operand in pointer arithmetic has bad consequences.
4767 // On the other hand, 32-bit overflow is rare, and the possibility
4768 // can often be excluded, if we annotate the ConvI2L node with
4769 // a type assertion that its value is known to be a small positive
4770 // number. (The prior range check has ensured this.)
4771 // This assertion is used by ConvI2LNode::Ideal.
4772 int index_max = max_jint - 1; // array size is max_jint, index is one less
4773 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4774 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4775 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4776 #endif
4777 return idx;
4778 }
4779
4780 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4781 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4782 if (ctrl != NULL) {
4783 // Express control dependency by a CastII node with a narrow type.
4784 value = new CastIINode(value, itype, false, true /* range check dependency */);
4785 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4786 // node from floating above the range check during loop optimizations. Otherwise, the
4787 // ConvI2L node may be eliminated independently of the range check, causing the data path
4788 // to become TOP while the control path is still there (although it's unreachable).
4789 value->set_req(0, ctrl);
4790 // Save CastII node to remove it after loop optimizations.
4791 phase->C->add_range_check_cast(value);
4792 value = phase->transform(value);
4793 }
4794 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4795 return phase->transform(new ConvI2LNode(value, ltype));
4796 }
4797
4798 // The message about the current inlining is accumulated in
4799 // _print_inlining_stream and transfered into the _print_inlining_list
4800 // once we know whether inlining succeeds or not. For regular
4801 // inlining, messages are appended to the buffer pointed by
4802 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4803 // a new buffer is added after _print_inlining_idx in the list. This
4804 // way we can update the inlining message for late inlining call site
4805 // when the inlining is attempted again.
4806 void Compile::print_inlining_init() {
4807 if (print_inlining() || print_intrinsics()) {
4808 _print_inlining_stream = new stringStream();
4809 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4810 }
4811 }
4812
4813 void Compile::print_inlining_reinit() {
4814 if (print_inlining() || print_intrinsics()) {
4815 // Re allocate buffer when we change ResourceMark
4816 _print_inlining_stream = new stringStream();
4817 }
4818 }
4819
4820 void Compile::print_inlining_reset() {
4821 _print_inlining_stream->reset();
4822 }
4823
4824 void Compile::print_inlining_commit() {
4825 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4826 // Transfer the message from _print_inlining_stream to the current
4827 // _print_inlining_list buffer and clear _print_inlining_stream.
4828 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4829 print_inlining_reset();
4830 }
4831
4832 void Compile::print_inlining_push() {
4833 // Add new buffer to the _print_inlining_list at current position
4834 _print_inlining_idx++;
4835 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4836 }
4837
4838 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4839 return _print_inlining_list->at(_print_inlining_idx);
4840 }
4841
4842 void Compile::print_inlining_update(CallGenerator* cg) {
4843 if (print_inlining() || print_intrinsics()) {
4844 if (!cg->is_late_inline()) {
4845 if (print_inlining_current().cg() != NULL) {
4846 print_inlining_push();
4847 }
4848 print_inlining_commit();
4849 } else {
4850 if (print_inlining_current().cg() != cg &&
4851 (print_inlining_current().cg() != NULL ||
4852 print_inlining_current().ss()->size() != 0)) {
4853 print_inlining_push();
4854 }
4855 print_inlining_commit();
4856 print_inlining_current().set_cg(cg);
4857 }
4858 }
4859 }
4860
4861 void Compile::print_inlining_move_to(CallGenerator* cg) {
4862 // We resume inlining at a late inlining call site. Locate the
4863 // corresponding inlining buffer so that we can update it.
4864 if (print_inlining()) {
4865 for (int i = 0; i < _print_inlining_list->length(); i++) {
4866 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4867 _print_inlining_idx = i;
4868 return;
4869 }
4870 }
4871 ShouldNotReachHere();
4872 }
4873 }
4874
4875 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4876 if (print_inlining()) {
4877 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4878 assert(print_inlining_current().cg() == cg, "wrong entry");
4879 // replace message with new message
4880 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4881 print_inlining_commit();
4882 print_inlining_current().set_cg(cg);
4883 }
4884 }
4885
4886 void Compile::print_inlining_assert_ready() {
4887 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4888 }
4889
4890 void Compile::process_print_inlining() {
4891 bool do_print_inlining = print_inlining() || print_intrinsics();
4892 if (do_print_inlining || log() != NULL) {
4893 // Print inlining message for candidates that we couldn't inline
4894 // for lack of space
4895 for (int i = 0; i < _late_inlines.length(); i++) {
4896 CallGenerator* cg = _late_inlines.at(i);
4897 if (!cg->is_mh_late_inline()) {
4898 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4899 if (do_print_inlining) {
4900 cg->print_inlining_late(msg);
4901 }
4902 log_late_inline_failure(cg, msg);
4903 }
4904 }
4905 }
4906 if (do_print_inlining) {
4907 ResourceMark rm;
4908 stringStream ss;
4909 for (int i = 0; i < _print_inlining_list->length(); i++) {
4910 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4911 }
4912 size_t end = ss.size();
4913 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4914 strncpy(_print_inlining_output, ss.base(), end+1);
4915 _print_inlining_output[end] = 0;
4916 }
4917 }
4918
4919 void Compile::dump_print_inlining() {
4920 if (_print_inlining_output != NULL) {
4921 tty->print_raw(_print_inlining_output);
4922 }
4923 }
4924
4925 void Compile::log_late_inline(CallGenerator* cg) {
4926 if (log() != NULL) {
4927 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4928 cg->unique_id());
4929 JVMState* p = cg->call_node()->jvms();
4930 while (p != NULL) {
4931 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4932 p = p->caller();
4933 }
4934 log()->tail("late_inline");
4935 }
4936 }
4937
4938 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4939 log_late_inline(cg);
4940 if (log() != NULL) {
4941 log()->inline_fail(msg);
4942 }
4943 }
4944
4945 void Compile::log_inline_id(CallGenerator* cg) {
4946 if (log() != NULL) {
4947 // The LogCompilation tool needs a unique way to identify late
4948 // inline call sites. This id must be unique for this call site in
4949 // this compilation. Try to have it unique across compilations as
4950 // well because it can be convenient when grepping through the log
4951 // file.
4952 // Distinguish OSR compilations from others in case CICountOSR is
4953 // on.
4954 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4955 cg->set_unique_id(id);
4956 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4957 }
4958 }
4959
4960 void Compile::log_inline_failure(const char* msg) {
4961 if (C->log() != NULL) {
4962 C->log()->inline_fail(msg);
4963 }
4964 }
4965
4966
4967 // Dump inlining replay data to the stream.
4968 // Don't change thread state and acquire any locks.
4969 void Compile::dump_inline_data(outputStream* out) {
4970 InlineTree* inl_tree = ilt();
4971 if (inl_tree != NULL) {
4972 out->print(" inline %d", inl_tree->count());
4973 inl_tree->dump_replay_data(out);
4974 }
4975 }
4976
4977 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4978 if (n1->Opcode() < n2->Opcode()) return -1;
4979 else if (n1->Opcode() > n2->Opcode()) return 1;
4980
4981 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4982 for (uint i = 1; i < n1->req(); i++) {
4983 if (n1->in(i) < n2->in(i)) return -1;
4984 else if (n1->in(i) > n2->in(i)) return 1;
4985 }
4986
4987 return 0;
4988 }
4989
4990 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4991 Node* n1 = *n1p;
4992 Node* n2 = *n2p;
4993
4994 return cmp_expensive_nodes(n1, n2);
4995 }
4996
4997 void Compile::sort_expensive_nodes() {
4998 if (!expensive_nodes_sorted()) {
4999 _expensive_nodes->sort(cmp_expensive_nodes);
5000 }
5001 }
5002
5003 bool Compile::expensive_nodes_sorted() const {
5004 for (int i = 1; i < _expensive_nodes->length(); i++) {
5005 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
5006 return false;
5007 }
5008 }
5009 return true;
5010 }
5011
5012 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5013 if (_expensive_nodes->length() == 0) {
5014 return false;
5015 }
5016
5017 assert(OptimizeExpensiveOps, "optimization off?");
5018
5019 // Take this opportunity to remove dead nodes from the list
5020 int j = 0;
5021 for (int i = 0; i < _expensive_nodes->length(); i++) {
5022 Node* n = _expensive_nodes->at(i);
5023 if (!n->is_unreachable(igvn)) {
5024 assert(n->is_expensive(), "should be expensive");
5025 _expensive_nodes->at_put(j, n);
5026 j++;
5027 }
5028 }
5029 _expensive_nodes->trunc_to(j);
5030
5031 // Then sort the list so that similar nodes are next to each other
5032 // and check for at least two nodes of identical kind with same data
5033 // inputs.
5034 sort_expensive_nodes();
5035
5036 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
5037 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
5038 return true;
5039 }
5040 }
5041
5042 return false;
5043 }
5044
5045 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5046 if (_expensive_nodes->length() == 0) {
5047 return;
5048 }
5049
5050 assert(OptimizeExpensiveOps, "optimization off?");
5051
5052 // Sort to bring similar nodes next to each other and clear the
5053 // control input of nodes for which there's only a single copy.
5054 sort_expensive_nodes();
5055
5056 int j = 0;
5057 int identical = 0;
5058 int i = 0;
5059 bool modified = false;
5060 for (; i < _expensive_nodes->length()-1; i++) {
5061 assert(j <= i, "can't write beyond current index");
5062 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
5063 identical++;
5064 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5065 continue;
5066 }
5067 if (identical > 0) {
5068 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5069 identical = 0;
5070 } else {
5071 Node* n = _expensive_nodes->at(i);
5072 igvn.replace_input_of(n, 0, NULL);
5073 igvn.hash_insert(n);
5074 modified = true;
5075 }
5076 }
5077 if (identical > 0) {
5078 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5079 } else if (_expensive_nodes->length() >= 1) {
5080 Node* n = _expensive_nodes->at(i);
5081 igvn.replace_input_of(n, 0, NULL);
5082 igvn.hash_insert(n);
5083 modified = true;
5084 }
5085 _expensive_nodes->trunc_to(j);
5086 if (modified) {
5087 igvn.optimize();
5088 }
5089 }
5090
5091 void Compile::add_expensive_node(Node * n) {
5092 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
5093 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5094 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5095 if (OptimizeExpensiveOps) {
5096 _expensive_nodes->append(n);
5097 } else {
5098 // Clear control input and let IGVN optimize expensive nodes if
5099 // OptimizeExpensiveOps is off.
5100 n->set_req(0, NULL);
5101 }
5102 }
5103
5104 /**
5105 * Remove the speculative part of types and clean up the graph
5106 */
5107 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5108 if (UseTypeSpeculation) {
5109 Unique_Node_List worklist;
5110 worklist.push(root());
5111 int modified = 0;
5112 // Go over all type nodes that carry a speculative type, drop the
5113 // speculative part of the type and enqueue the node for an igvn
5114 // which may optimize it out.
5115 for (uint next = 0; next < worklist.size(); ++next) {
5116 Node *n = worklist.at(next);
5117 if (n->is_Type()) {
5118 TypeNode* tn = n->as_Type();
5119 const Type* t = tn->type();
5120 const Type* t_no_spec = t->remove_speculative();
5121 if (t_no_spec != t) {
5122 bool in_hash = igvn.hash_delete(n);
5123 assert(in_hash, "node should be in igvn hash table");
5124 tn->set_type(t_no_spec);
5125 igvn.hash_insert(n);
5126 igvn._worklist.push(n); // give it a chance to go away
5127 modified++;
5128 }
5129 }
5130 uint max = n->len();
5131 for( uint i = 0; i < max; ++i ) {
5132 Node *m = n->in(i);
5133 if (not_a_node(m)) continue;
5134 worklist.push(m);
5135 }
5136 }
5137 // Drop the speculative part of all types in the igvn's type table
5138 igvn.remove_speculative_types();
5139 if (modified > 0) {
5140 igvn.optimize();
5141 }
5142 #ifdef ASSERT
5143 // Verify that after the IGVN is over no speculative type has resurfaced
5144 worklist.clear();
5145 worklist.push(root());
5146 for (uint next = 0; next < worklist.size(); ++next) {
5147 Node *n = worklist.at(next);
5148 const Type* t = igvn.type_or_null(n);
5149 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5150 if (n->is_Type()) {
5151 t = n->as_Type()->type();
5152 assert(t == t->remove_speculative(), "no more speculative types");
5153 }
5154 uint max = n->len();
5155 for( uint i = 0; i < max; ++i ) {
5156 Node *m = n->in(i);
5157 if (not_a_node(m)) continue;
5158 worklist.push(m);
5159 }
5160 }
5161 igvn.check_no_speculative_types();
5162 #endif
5163 }
5164 }
5165
5166 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5167 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5168 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5169 if (!EnableValhalla || ta == NULL || tb == NULL ||
5170 ta->is_zero_type() || tb->is_zero_type() ||
5171 !ta->can_be_value_type() || !tb->can_be_value_type()) {
5172 // Use old acmp if one operand is null or not a value type
5173 return new CmpPNode(a, b);
5174 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) {
5175 // We know that one operand is a value type. Therefore,
5176 // new acmp will only return true if both operands are NULL.
5177 // Check if both operands are null by or'ing the oops.
5178 a = phase->transform(new CastP2XNode(NULL, a));
5179 b = phase->transform(new CastP2XNode(NULL, b));
5180 a = phase->transform(new OrXNode(a, b));
5181 return new CmpXNode(a, phase->MakeConX(0));
5182 }
5183 // Use new acmp
5184 return NULL;
5185 }
5186
5187 // Auxiliary method to support randomized stressing/fuzzing.
5188 //
5189 // This method can be called the arbitrary number of times, with current count
5190 // as the argument. The logic allows selecting a single candidate from the
5191 // running list of candidates as follows:
5192 // int count = 0;
5193 // Cand* selected = null;
5194 // while(cand = cand->next()) {
5195 // if (randomized_select(++count)) {
5196 // selected = cand;
5197 // }
5198 // }
5199 //
5200 // Including count equalizes the chances any candidate is "selected".
5201 // This is useful when we don't have the complete list of candidates to choose
5202 // from uniformly. In this case, we need to adjust the randomicity of the
5203 // selection, or else we will end up biasing the selection towards the latter
5204 // candidates.
5205 //
5206 // Quick back-envelope calculation shows that for the list of n candidates
5207 // the equal probability for the candidate to persist as "best" can be
5208 // achieved by replacing it with "next" k-th candidate with the probability
5209 // of 1/k. It can be easily shown that by the end of the run, the
5210 // probability for any candidate is converged to 1/n, thus giving the
5211 // uniform distribution among all the candidates.
5212 //
5213 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5214 #define RANDOMIZED_DOMAIN_POW 29
5215 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5216 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5217 bool Compile::randomized_select(int count) {
5218 assert(count > 0, "only positive");
5219 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5220 }
5221
5222 CloneMap& Compile::clone_map() { return _clone_map; }
5223 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5224
5225 void NodeCloneInfo::dump() const {
5226 tty->print(" {%d:%d} ", idx(), gen());
5227 }
5228
5229 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5230 uint64_t val = value(old->_idx);
5231 NodeCloneInfo cio(val);
5232 assert(val != 0, "old node should be in the map");
5233 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5234 insert(nnn->_idx, cin.get());
5235 #ifndef PRODUCT
5236 if (is_debug()) {
5237 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5238 }
5239 #endif
5240 }
5241
5242 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5243 NodeCloneInfo cio(value(old->_idx));
5244 if (cio.get() == 0) {
5245 cio.set(old->_idx, 0);
5246 insert(old->_idx, cio.get());
5247 #ifndef PRODUCT
5248 if (is_debug()) {
5249 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5250 }
5251 #endif
5252 }
5253 clone(old, nnn, gen);
5254 }
5255
5256 int CloneMap::max_gen() const {
5257 int g = 0;
5258 DictI di(_dict);
5259 for(; di.test(); ++di) {
5260 int t = gen(di._key);
5261 if (g < t) {
5262 g = t;
5263 #ifndef PRODUCT
5264 if (is_debug()) {
5265 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5266 }
5267 #endif
5268 }
5269 }
5270 return g;
5271 }
5272
5273 void CloneMap::dump(node_idx_t key) const {
5274 uint64_t val = value(key);
5275 if (val != 0) {
5276 NodeCloneInfo ni(val);
5277 ni.dump();
5278 }
5279 }