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));
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));
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->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));
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), to->instance_id());
1618 } else {
1619 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
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 }
1638
1639 ciKlass* klass = tk->klass();
1640 if (klass != NULL && klass->is_obj_array_klass()) {
1641 ciKlass* k = TypeAryPtr::OOPS->klass();
1642 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1643 k = TypeInstPtr::BOTTOM->klass();
1644 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1645 }
1646
1647 // Check for precise loads from the primary supertype array and force them
1648 // to the supertype cache alias index. Check for generic array loads from
1649 // the primary supertype array and also force them to the supertype cache
1650 // alias index. Since the same load can reach both, we need to merge
1651 // these 2 disparate memories into the same alias class. Since the
1652 // primary supertype array is read-only, there's no chance of confusion
1653 // where we bypass an array load and an array store.
1654 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1655 if (offset == Type::OffsetBot ||
1656 (offset >= primary_supers_offset &&
1657 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1658 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1659 offset = in_bytes(Klass::secondary_super_cache_offset());
1660 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1661 }
1662 }
1663
1664 // Flatten all Raw pointers together.
1665 if (tj->base() == Type::RawPtr)
1666 tj = TypeRawPtr::BOTTOM;
1667
1668 if (tj->base() == Type::AnyPtr)
1669 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1670
1671 // Flatten all to bottom for now
1672 switch( _AliasLevel ) {
1673 case 0:
1674 tj = TypePtr::BOTTOM;
1675 break;
1676 case 1: // Flatten to: oop, static, field or array
1677 switch (tj->base()) {
1678 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1679 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1680 case Type::AryPtr: // do not distinguish arrays at all
1681 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1682 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1683 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1684 default: ShouldNotReachHere();
1685 }
1686 break;
1687 case 2: // No collapsing at level 2; keep all splits
1688 case 3: // No collapsing at level 3; keep all splits
1689 break;
1690 default:
1691 Unimplemented();
1692 }
1693
1694 offset = tj->offset();
1695 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1696
1697 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1698 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1699 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1700 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1701 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1702 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1703 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1704 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1705 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1706 assert( tj->ptr() != TypePtr::TopPTR &&
1707 tj->ptr() != TypePtr::AnyNull &&
1708 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1709 // assert( tj->ptr() != TypePtr::Constant ||
1710 // tj->base() == Type::RawPtr ||
1711 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1712
1713 return tj;
1714 }
1715
1716 void Compile::AliasType::Init(int i, const TypePtr* at) {
1717 _index = i;
1718 _adr_type = at;
1719 _field = NULL;
1720 _element = NULL;
1721 _is_rewritable = true; // default
1722 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1723 if (atoop != NULL && atoop->is_known_instance()) {
1724 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1725 _general_index = Compile::current()->get_alias_index(gt);
1726 } else {
1727 _general_index = 0;
1728 }
1729 }
1730
1731 BasicType Compile::AliasType::basic_type() const {
1732 if (element() != NULL) {
1733 const Type* element = adr_type()->is_aryptr()->elem();
1734 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1735 } if (field() != NULL) {
1736 return field()->layout_type();
1737 } else {
1738 return T_ILLEGAL; // unknown
1739 }
1740 }
1741
1742 //---------------------------------print_on------------------------------------
1743 #ifndef PRODUCT
1744 void Compile::AliasType::print_on(outputStream* st) {
1745 if (index() < 10)
1746 st->print("@ <%d> ", index());
1747 else st->print("@ <%d>", index());
1748 st->print(is_rewritable() ? " " : " RO");
1749 int offset = adr_type()->offset();
1750 if (offset == Type::OffsetBot)
1751 st->print(" +any");
1752 else st->print(" +%-3d", offset);
1753 st->print(" in ");
1754 adr_type()->dump_on(st);
1755 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1756 if (field() != NULL && tjp) {
1757 if (tjp->klass() != field()->holder() ||
1758 tjp->offset() != field()->offset_in_bytes()) {
1759 st->print(" != ");
1760 field()->print();
1761 st->print(" ***");
1762 }
1763 }
1764 }
1765
1766 void print_alias_types() {
1767 Compile* C = Compile::current();
1768 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1769 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1770 C->alias_type(idx)->print_on(tty);
1771 tty->cr();
1772 }
1773 }
1774 #endif
1775
1776
1777 //----------------------------probe_alias_cache--------------------------------
1778 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1779 intptr_t key = (intptr_t) adr_type;
1780 key ^= key >> logAliasCacheSize;
1781 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1782 }
1783
1784
1785 //-----------------------------grow_alias_types--------------------------------
1786 void Compile::grow_alias_types() {
1787 const int old_ats = _max_alias_types; // how many before?
1788 const int new_ats = old_ats; // how many more?
1789 const int grow_ats = old_ats+new_ats; // how many now?
1790 _max_alias_types = grow_ats;
1791 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1792 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1793 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1794 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1795 }
1796
1797
1798 //--------------------------------find_alias_type------------------------------
1799 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1800 if (_AliasLevel == 0)
1801 return alias_type(AliasIdxBot);
1802
1803 AliasCacheEntry* ace = NULL;
1804 if (!uncached) {
1805 ace = probe_alias_cache(adr_type);
1806 if (ace->_adr_type == adr_type) {
1807 return alias_type(ace->_index);
1808 }
1809 }
1810
1811 // Handle special cases.
1812 if (adr_type == NULL) return alias_type(AliasIdxTop);
1813 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1814
1815 // Do it the slow way.
1816 const TypePtr* flat = flatten_alias_type(adr_type);
1817
1818 #ifdef ASSERT
1819 {
1820 ResourceMark rm;
1821 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1822 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1823 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1824 Type::str(adr_type));
1825 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1826 const TypeOopPtr* foop = flat->is_oopptr();
1827 // Scalarizable allocations have exact klass always.
1828 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1829 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1830 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1831 Type::str(foop), Type::str(xoop));
1832 }
1833 }
1834 #endif
1835
1836 int idx = AliasIdxTop;
1837 for (int i = 0; i < num_alias_types(); i++) {
1838 if (alias_type(i)->adr_type() == flat) {
1839 idx = i;
1840 break;
1841 }
1842 }
1843
1844 if (idx == AliasIdxTop) {
1845 if (no_create) return NULL;
1846 // Grow the array if necessary.
1847 if (_num_alias_types == _max_alias_types) grow_alias_types();
1848 // Add a new alias type.
1849 idx = _num_alias_types++;
1850 _alias_types[idx]->Init(idx, flat);
1851 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1852 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1853 if (flat->isa_instptr()) {
1854 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1855 && flat->is_instptr()->klass() == env()->Class_klass())
1856 alias_type(idx)->set_rewritable(false);
1857 }
1858 ciField* field = NULL;
1859 if (flat->isa_aryptr()) {
1860 #ifdef ASSERT
1861 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1862 // (T_BYTE has the weakest alignment and size restrictions...)
1863 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1864 #endif
1865 const Type* elemtype = flat->is_aryptr()->elem();
1866 if (flat->offset() == TypePtr::OffsetBot) {
1867 alias_type(idx)->set_element(elemtype);
1868 }
1869 int field_offset = flat->is_aryptr()->field_offset().get();
1870 if (elemtype->isa_valuetype() &&
1871 elemtype->value_klass() != NULL &&
1872 field_offset != Type::OffsetBot) {
1873 ciValueKlass* vk = elemtype->value_klass();
1874 field_offset += vk->first_field_offset();
1875 field = vk->get_field_by_offset(field_offset, false);
1876 }
1877 }
1878 if (flat->isa_klassptr()) {
1879 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1880 alias_type(idx)->set_rewritable(false);
1881 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1882 alias_type(idx)->set_rewritable(false);
1883 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1884 alias_type(idx)->set_rewritable(false);
1885 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1886 alias_type(idx)->set_rewritable(false);
1887 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1888 alias_type(idx)->set_rewritable(false);
1889 }
1890 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1891 // but the base pointer type is not distinctive enough to identify
1892 // references into JavaThread.)
1893
1894 // Check for final fields.
1895 const TypeInstPtr* tinst = flat->isa_instptr();
1896 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1897 if (tinst->const_oop() != NULL &&
1898 tinst->klass() == ciEnv::current()->Class_klass() &&
1899 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1900 // static field
1901 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1902 field = k->get_field_by_offset(tinst->offset(), true);
1903 } else if (tinst->klass()->is_valuetype()) {
1904 // Value type field
1905 ciValueKlass* vk = tinst->value_klass();
1906 field = vk->get_field_by_offset(tinst->offset(), false);
1907 } else {
1908 ciInstanceKlass* k = tinst->klass()->as_instance_klass();
1909 field = k->get_field_by_offset(tinst->offset(), false);
1910 }
1911 }
1912 assert(field == NULL ||
1913 original_field == NULL ||
1914 (field->holder() == original_field->holder() &&
1915 field->offset() == original_field->offset() &&
1916 field->is_static() == original_field->is_static()), "wrong field?");
1917 // Set field() and is_rewritable() attributes.
1918 if (field != NULL) {
1919 alias_type(idx)->set_field(field);
1920 if (flat->isa_aryptr()) {
1921 // Fields of flattened inline type arrays are rewritable although they are declared final
1922 assert(flat->is_aryptr()->elem()->isa_valuetype(), "must be a flattened value array");
1923 alias_type(idx)->set_rewritable(true);
1924 }
1925 }
1926 }
1927
1928 // Fill the cache for next time.
1929 if (!uncached) {
1930 ace->_adr_type = adr_type;
1931 ace->_index = idx;
1932 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1933
1934 // Might as well try to fill the cache for the flattened version, too.
1935 AliasCacheEntry* face = probe_alias_cache(flat);
1936 if (face->_adr_type == NULL) {
1937 face->_adr_type = flat;
1938 face->_index = idx;
1939 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1940 }
1941 }
1942
1943 return alias_type(idx);
1944 }
1945
1946
1947 Compile::AliasType* Compile::alias_type(ciField* field) {
1948 const TypeOopPtr* t;
1949 if (field->is_static())
1950 t = TypeInstPtr::make(field->holder()->java_mirror());
1951 else
1952 t = TypeOopPtr::make_from_klass_raw(field->holder());
1953 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1954 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1955 return atp;
1956 }
1957
1958
1959 //------------------------------have_alias_type--------------------------------
1960 bool Compile::have_alias_type(const TypePtr* adr_type) {
1961 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1962 if (ace->_adr_type == adr_type) {
1963 return true;
1964 }
1965
1966 // Handle special cases.
1967 if (adr_type == NULL) return true;
1968 if (adr_type == TypePtr::BOTTOM) return true;
1969
1970 return find_alias_type(adr_type, true, NULL) != NULL;
1971 }
1972
1973 //-----------------------------must_alias--------------------------------------
1974 // True if all values of the given address type are in the given alias category.
1975 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1976 if (alias_idx == AliasIdxBot) return true; // the universal category
1977 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1978 if (alias_idx == AliasIdxTop) return false; // the empty category
1979 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1980
1981 // the only remaining possible overlap is identity
1982 int adr_idx = get_alias_index(adr_type);
1983 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1984 assert(adr_idx == alias_idx ||
1985 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1986 && adr_type != TypeOopPtr::BOTTOM),
1987 "should not be testing for overlap with an unsafe pointer");
1988 return adr_idx == alias_idx;
1989 }
1990
1991 //------------------------------can_alias--------------------------------------
1992 // True if any values of the given address type are in the given alias category.
1993 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1994 if (alias_idx == AliasIdxTop) return false; // the empty category
1995 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1996 if (alias_idx == AliasIdxBot) return true; // the universal category
1997 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1998
1999 // the only remaining possible overlap is identity
2000 int adr_idx = get_alias_index(adr_type);
2001 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
2002 return adr_idx == alias_idx;
2003 }
2004
2005
2006
2007 //---------------------------pop_warm_call-------------------------------------
2008 WarmCallInfo* Compile::pop_warm_call() {
2009 WarmCallInfo* wci = _warm_calls;
2010 if (wci != NULL) _warm_calls = wci->remove_from(wci);
2011 return wci;
2012 }
2013
2014 //----------------------------Inline_Warm--------------------------------------
2015 int Compile::Inline_Warm() {
2016 // If there is room, try to inline some more warm call sites.
2017 // %%% Do a graph index compaction pass when we think we're out of space?
2018 if (!InlineWarmCalls) return 0;
2019
2020 int calls_made_hot = 0;
2021 int room_to_grow = NodeCountInliningCutoff - unique();
2022 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
2023 int amount_grown = 0;
2024 WarmCallInfo* call;
2025 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
2026 int est_size = (int)call->size();
2027 if (est_size > (room_to_grow - amount_grown)) {
2028 // This one won't fit anyway. Get rid of it.
2029 call->make_cold();
2030 continue;
2031 }
2032 call->make_hot();
2033 calls_made_hot++;
2034 amount_grown += est_size;
2035 amount_to_grow -= est_size;
2036 }
2037
2038 if (calls_made_hot > 0) set_major_progress();
2039 return calls_made_hot;
2040 }
2041
2042
2043 //----------------------------Finish_Warm--------------------------------------
2044 void Compile::Finish_Warm() {
2045 if (!InlineWarmCalls) return;
2046 if (failing()) return;
2047 if (warm_calls() == NULL) return;
2048
2049 // Clean up loose ends, if we are out of space for inlining.
2050 WarmCallInfo* call;
2051 while ((call = pop_warm_call()) != NULL) {
2052 call->make_cold();
2053 }
2054 }
2055
2056 //---------------------cleanup_loop_predicates-----------------------
2057 // Remove the opaque nodes that protect the predicates so that all unused
2058 // checks and uncommon_traps will be eliminated from the ideal graph
2059 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
2060 if (predicate_count()==0) return;
2061 for (int i = predicate_count(); i > 0; i--) {
2062 Node * n = predicate_opaque1_node(i-1);
2063 assert(n->Opcode() == Op_Opaque1, "must be");
2064 igvn.replace_node(n, n->in(1));
2065 }
2066 assert(predicate_count()==0, "should be clean!");
2067 }
2068
2069 void Compile::add_range_check_cast(Node* n) {
2070 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2071 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
2072 _range_check_casts->append(n);
2073 }
2074
2075 // Remove all range check dependent CastIINodes.
2076 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
2077 for (int i = range_check_cast_count(); i > 0; i--) {
2078 Node* cast = range_check_cast_node(i-1);
2079 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2080 igvn.replace_node(cast, cast->in(1));
2081 }
2082 assert(range_check_cast_count() == 0, "should be empty");
2083 }
2084
2085 void Compile::add_opaque4_node(Node* n) {
2086 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2087 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2088 _opaque4_nodes->append(n);
2089 }
2090
2091 // Remove all Opaque4 nodes.
2092 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2093 for (int i = opaque4_count(); i > 0; i--) {
2094 Node* opaq = opaque4_node(i-1);
2095 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2096 igvn.replace_node(opaq, opaq->in(2));
2097 }
2098 assert(opaque4_count() == 0, "should be empty");
2099 }
2100
2101 void Compile::add_value_type(Node* n) {
2102 assert(n->is_ValueTypeBase(), "unexpected node");
2103 if (_value_type_nodes != NULL) {
2104 _value_type_nodes->push(n);
2105 }
2106 }
2107
2108 void Compile::remove_value_type(Node* n) {
2109 assert(n->is_ValueTypeBase(), "unexpected node");
2110 if (_value_type_nodes != NULL) {
2111 _value_type_nodes->remove(n);
2112 }
2113 }
2114
2115 // Does the return value keep otherwise useless value type allocations
2116 // alive?
2117 static bool return_val_keeps_allocations_alive(Node* ret_val) {
2118 ResourceMark rm;
2119 Unique_Node_List wq;
2120 wq.push(ret_val);
2121 bool some_allocations = false;
2122 for (uint i = 0; i < wq.size(); i++) {
2123 Node* n = wq.at(i);
2124 assert(!n->is_ValueTypeBase(), "chain of value type nodes");
2125 if (n->outcnt() > 1) {
2126 // Some other use for the allocation
2127 return false;
2128 } else if (n->is_Phi()) {
2129 for (uint j = 1; j < n->req(); j++) {
2130 wq.push(n->in(j));
2131 }
2132 } else if (n->is_CheckCastPP() &&
2133 n->in(1)->is_Proj() &&
2134 n->in(1)->in(0)->is_Allocate()) {
2135 some_allocations = true;
2136 }
2137 }
2138 return some_allocations;
2139 }
2140
2141 void Compile::process_value_types(PhaseIterGVN &igvn) {
2142 // Make value types scalar in safepoints
2143 while (_value_type_nodes->size() != 0) {
2144 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase();
2145 vt->make_scalar_in_safepoints(&igvn);
2146 if (vt->is_ValueTypePtr()) {
2147 igvn.replace_node(vt, vt->get_oop());
2148 } else if (vt->outcnt() == 0) {
2149 igvn.remove_dead_node(vt);
2150 }
2151 }
2152 _value_type_nodes = NULL;
2153 if (tf()->returns_value_type_as_fields()) {
2154 Node* ret = NULL;
2155 for (uint i = 1; i < root()->req(); i++){
2156 Node* in = root()->in(i);
2157 if (in->Opcode() == Op_Return) {
2158 assert(ret == NULL, "only one return");
2159 ret = in;
2160 }
2161 }
2162 if (ret != NULL) {
2163 Node* ret_val = ret->in(TypeFunc::Parms);
2164 if (igvn.type(ret_val)->isa_oopptr() &&
2165 return_val_keeps_allocations_alive(ret_val)) {
2166 igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn));
2167 assert(ret_val->outcnt() == 0, "should be dead now");
2168 igvn.remove_dead_node(ret_val);
2169 }
2170 }
2171 }
2172 igvn.optimize();
2173 }
2174
2175 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2176 if (!_has_flattened_accesses) {
2177 return;
2178 }
2179 // Initially, all flattened array accesses share the same slice to
2180 // keep dependencies with Object[] array accesses (that could be
2181 // to a flattened array) correct. We're done with parsing so we
2182 // now know all flattened array accesses in this compile
2183 // unit. Let's move flattened array accesses to their own slice,
2184 // one per element field. This should help memory access
2185 // optimizations.
2186 ResourceMark rm;
2187 Unique_Node_List wq;
2188 wq.push(root());
2189
2190 Node_List mergememnodes;
2191 Node_List memnodes;
2192
2193 // Alias index currently shared by all flattened memory accesses
2194 int index = get_alias_index(TypeAryPtr::VALUES);
2195
2196 // Find MergeMem nodes and flattened array accesses
2197 for (uint i = 0; i < wq.size(); i++) {
2198 Node* n = wq.at(i);
2199 if (n->is_Mem()) {
2200 const TypePtr* adr_type = NULL;
2201 if (n->Opcode() == Op_StoreCM) {
2202 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2203 } else {
2204 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2205 }
2206 if (adr_type == TypeAryPtr::VALUES) {
2207 memnodes.push(n);
2208 }
2209 } else if (n->is_MergeMem()) {
2210 MergeMemNode* mm = n->as_MergeMem();
2211 if (mm->memory_at(index) != mm->base_memory()) {
2212 mergememnodes.push(n);
2213 }
2214 }
2215 for (uint j = 0; j < n->req(); j++) {
2216 Node* m = n->in(j);
2217 if (m != NULL) {
2218 wq.push(m);
2219 }
2220 }
2221 }
2222
2223 if (memnodes.size() > 0) {
2224 _flattened_accesses_share_alias = false;
2225
2226 // We are going to change the slice for the flattened array
2227 // accesses so we need to clear the cache entries that refer to
2228 // them.
2229 for (uint i = 0; i < AliasCacheSize; i++) {
2230 AliasCacheEntry* ace = &_alias_cache[i];
2231 if (ace->_adr_type != NULL &&
2232 ace->_adr_type->isa_aryptr() &&
2233 ace->_adr_type->is_aryptr()->elem()->isa_valuetype()) {
2234 ace->_adr_type = NULL;
2235 ace->_index = 0;
2236 }
2237 }
2238
2239 // Find what aliases we are going to add
2240 int start_alias = num_alias_types()-1;
2241 int stop_alias = 0;
2242
2243 for (uint i = 0; i < memnodes.size(); i++) {
2244 Node* m = memnodes.at(i);
2245 const TypePtr* adr_type = NULL;
2246 if (m->Opcode() == Op_StoreCM) {
2247 adr_type = m->in(MemNode::OopStore)->adr_type();
2248 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2249 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2250 get_alias_index(adr_type));
2251 igvn.register_new_node_with_optimizer(clone);
2252 igvn.replace_node(m, clone);
2253 } else {
2254 adr_type = m->adr_type();
2255 #ifdef ASSERT
2256 m->as_Mem()->set_adr_type(adr_type);
2257 #endif
2258 }
2259 int idx = get_alias_index(adr_type);
2260 start_alias = MIN2(start_alias, idx);
2261 stop_alias = MAX2(stop_alias, idx);
2262 }
2263
2264 assert(stop_alias >= start_alias, "should have expanded aliases");
2265
2266 Node_Stack stack(0);
2267 #ifdef ASSERT
2268 VectorSet seen(Thread::current()->resource_area());
2269 #endif
2270 // Now let's fix the memory graph so each flattened array access
2271 // is moved to the right slice. Start from the MergeMem nodes.
2272 uint last = unique();
2273 for (uint i = 0; i < mergememnodes.size(); i++) {
2274 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2275 Node* n = current->memory_at(index);
2276 MergeMemNode* mm = NULL;
2277 do {
2278 // Follow memory edges through memory accesses, phis and
2279 // narrow membars and push nodes on the stack. Once we hit
2280 // bottom memory, we pop element off the stack one at a
2281 // time, in reverse order, and move them to the right slice
2282 // by changing their memory edges.
2283 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::VALUES) {
2284 assert(!seen.test_set(n->_idx), "");
2285 // Uses (a load for instance) will need to be moved to the
2286 // right slice as well and will get a new memory state
2287 // that we don't know yet. The use could also be the
2288 // backedge of a loop. We put a place holder node between
2289 // the memory node and its uses. We replace that place
2290 // holder with the correct memory state once we know it,
2291 // i.e. when nodes are popped off the stack. Using the
2292 // place holder make the logic work in the presence of
2293 // loops.
2294 if (n->outcnt() > 1) {
2295 Node* place_holder = NULL;
2296 assert(!n->has_out_with(Op_Node), "");
2297 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2298 Node* u = n->out(k);
2299 if (u != current && u->_idx < last) {
2300 bool success = false;
2301 for (uint l = 0; l < u->req(); l++) {
2302 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2303 continue;
2304 }
2305 Node* in = u->in(l);
2306 if (in == n) {
2307 if (place_holder == NULL) {
2308 place_holder = new Node(1);
2309 place_holder->init_req(0, n);
2310 }
2311 igvn.replace_input_of(u, l, place_holder);
2312 success = true;
2313 }
2314 }
2315 if (success) {
2316 --k;
2317 }
2318 }
2319 }
2320 }
2321 if (n->is_Phi()) {
2322 stack.push(n, 1);
2323 n = n->in(1);
2324 } else if (n->is_Mem()) {
2325 stack.push(n, n->req());
2326 n = n->in(MemNode::Memory);
2327 } else {
2328 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2329 stack.push(n, n->req());
2330 n = n->in(0)->in(TypeFunc::Memory);
2331 }
2332 } else {
2333 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2334 // Build a new MergeMem node to carry the new memory state
2335 // as we build it. IGVN should fold extraneous MergeMem
2336 // nodes.
2337 mm = MergeMemNode::make(n);
2338 igvn.register_new_node_with_optimizer(mm);
2339 while (stack.size() > 0) {
2340 Node* m = stack.node();
2341 uint idx = stack.index();
2342 if (m->is_Mem()) {
2343 // Move memory node to its new slice
2344 const TypePtr* adr_type = m->adr_type();
2345 int alias = get_alias_index(adr_type);
2346 Node* prev = mm->memory_at(alias);
2347 igvn.replace_input_of(m, MemNode::Memory, prev);
2348 mm->set_memory_at(alias, m);
2349 } else if (m->is_Phi()) {
2350 // We need as many new phis as there are new aliases
2351 igvn.replace_input_of(m, idx, mm);
2352 if (idx == m->req()-1) {
2353 Node* r = m->in(0);
2354 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2355 const Type* adr_type = get_adr_type(j);
2356 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2357 continue;
2358 }
2359 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2360 igvn.register_new_node_with_optimizer(phi);
2361 for (uint k = 1; k < m->req(); k++) {
2362 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2363 }
2364 mm->set_memory_at(j, phi);
2365 }
2366 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2367 igvn.register_new_node_with_optimizer(base_phi);
2368 for (uint k = 1; k < m->req(); k++) {
2369 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2370 }
2371 mm->set_base_memory(base_phi);
2372 }
2373 } else {
2374 // This is a MemBarCPUOrder node from
2375 // Parse::array_load()/Parse::array_store(), in the
2376 // branch that handles flattened arrays hidden under
2377 // an Object[] array. We also need one new membar per
2378 // new alias to keep the unknown access that the
2379 // membars protect properly ordered with accesses to
2380 // known flattened array.
2381 assert(m->is_Proj(), "projection expected");
2382 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2383 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2384 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2385 const Type* adr_type = get_adr_type(j);
2386 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2387 continue;
2388 }
2389 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2390 igvn.register_new_node_with_optimizer(mb);
2391 Node* mem = mm->memory_at(j);
2392 mb->init_req(TypeFunc::Control, ctrl);
2393 mb->init_req(TypeFunc::Memory, mem);
2394 ctrl = new ProjNode(mb, TypeFunc::Control);
2395 igvn.register_new_node_with_optimizer(ctrl);
2396 mem = new ProjNode(mb, TypeFunc::Memory);
2397 igvn.register_new_node_with_optimizer(mem);
2398 mm->set_memory_at(j, mem);
2399 }
2400 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2401 }
2402 if (idx < m->req()-1) {
2403 idx += 1;
2404 stack.set_index(idx);
2405 n = m->in(idx);
2406 break;
2407 }
2408 // Take care of place holder nodes
2409 if (m->has_out_with(Op_Node)) {
2410 Node* place_holder = m->find_out_with(Op_Node);
2411 if (place_holder != NULL) {
2412 Node* mm_clone = mm->clone();
2413 igvn.register_new_node_with_optimizer(mm_clone);
2414 Node* hook = new Node(1);
2415 hook->init_req(0, mm);
2416 igvn.replace_node(place_holder, mm_clone);
2417 hook->destruct();
2418 }
2419 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2420 }
2421 stack.pop();
2422 }
2423 }
2424 } while(stack.size() > 0);
2425 // Fix the memory state at the MergeMem we started from
2426 igvn.rehash_node_delayed(current);
2427 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2428 const Type* adr_type = get_adr_type(j);
2429 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2430 continue;
2431 }
2432 current->set_memory_at(j, mm);
2433 }
2434 current->set_memory_at(index, current->base_memory());
2435 }
2436 igvn.optimize();
2437 }
2438 print_method(PHASE_SPLIT_VALUES_ARRAY, 2);
2439 }
2440
2441
2442 // StringOpts and late inlining of string methods
2443 void Compile::inline_string_calls(bool parse_time) {
2444 {
2445 // remove useless nodes to make the usage analysis simpler
2446 ResourceMark rm;
2447 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2448 }
2449
2450 {
2451 ResourceMark rm;
2452 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2453 PhaseStringOpts pso(initial_gvn(), for_igvn());
2454 print_method(PHASE_AFTER_STRINGOPTS, 3);
2455 }
2456
2457 // now inline anything that we skipped the first time around
2458 if (!parse_time) {
2459 _late_inlines_pos = _late_inlines.length();
2460 }
2461
2462 while (_string_late_inlines.length() > 0) {
2463 CallGenerator* cg = _string_late_inlines.pop();
2464 cg->do_late_inline();
2465 if (failing()) return;
2466 }
2467 _string_late_inlines.trunc_to(0);
2468 }
2469
2470 // Late inlining of boxing methods
2471 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2472 if (_boxing_late_inlines.length() > 0) {
2473 assert(has_boxed_value(), "inconsistent");
2474
2475 PhaseGVN* gvn = initial_gvn();
2476 set_inlining_incrementally(true);
2477
2478 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2479 for_igvn()->clear();
2480 gvn->replace_with(&igvn);
2481
2482 _late_inlines_pos = _late_inlines.length();
2483
2484 while (_boxing_late_inlines.length() > 0) {
2485 CallGenerator* cg = _boxing_late_inlines.pop();
2486 cg->do_late_inline();
2487 if (failing()) return;
2488 }
2489 _boxing_late_inlines.trunc_to(0);
2490
2491 inline_incrementally_cleanup(igvn);
2492
2493 set_inlining_incrementally(false);
2494 }
2495 }
2496
2497 bool Compile::inline_incrementally_one() {
2498 assert(IncrementalInline, "incremental inlining should be on");
2499
2500 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2501 set_inlining_progress(false);
2502 set_do_cleanup(false);
2503 int i = 0;
2504 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2505 CallGenerator* cg = _late_inlines.at(i);
2506 _late_inlines_pos = i+1;
2507 cg->do_late_inline();
2508 if (failing()) return false;
2509 }
2510 int j = 0;
2511 for (; i < _late_inlines.length(); i++, j++) {
2512 _late_inlines.at_put(j, _late_inlines.at(i));
2513 }
2514 _late_inlines.trunc_to(j);
2515 assert(inlining_progress() || _late_inlines.length() == 0, "");
2516
2517 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2518
2519 set_inlining_progress(false);
2520 set_do_cleanup(false);
2521 return (_late_inlines.length() > 0) && !needs_cleanup;
2522 }
2523
2524 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2525 {
2526 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2527 ResourceMark rm;
2528 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2529 }
2530 {
2531 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2532 igvn = PhaseIterGVN(initial_gvn());
2533 igvn.optimize();
2534 }
2535 }
2536
2537 // Perform incremental inlining until bound on number of live nodes is reached
2538 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2539 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2540
2541 set_inlining_incrementally(true);
2542 uint low_live_nodes = 0;
2543
2544 while (_late_inlines.length() > 0) {
2545 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2546 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2547 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2548 // PhaseIdealLoop is expensive so we only try it once we are
2549 // out of live nodes and we only try it again if the previous
2550 // helped got the number of nodes down significantly
2551 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2552 if (failing()) return;
2553 low_live_nodes = live_nodes();
2554 _major_progress = true;
2555 }
2556
2557 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2558 break; // finish
2559 }
2560 }
2561
2562 for_igvn()->clear();
2563 initial_gvn()->replace_with(&igvn);
2564
2565 while (inline_incrementally_one()) {
2566 assert(!failing(), "inconsistent");
2567 }
2568
2569 if (failing()) return;
2570
2571 inline_incrementally_cleanup(igvn);
2572
2573 if (failing()) return;
2574 }
2575 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2576
2577 if (_string_late_inlines.length() > 0) {
2578 assert(has_stringbuilder(), "inconsistent");
2579 for_igvn()->clear();
2580 initial_gvn()->replace_with(&igvn);
2581
2582 inline_string_calls(false);
2583
2584 if (failing()) return;
2585
2586 inline_incrementally_cleanup(igvn);
2587 }
2588
2589 set_inlining_incrementally(false);
2590 }
2591
2592
2593 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2594 if(_loop_opts_cnt > 0) {
2595 debug_only( int cnt = 0; );
2596 while(major_progress() && (_loop_opts_cnt > 0)) {
2597 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2598 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2599 PhaseIdealLoop::optimize(igvn, mode);
2600 _loop_opts_cnt--;
2601 if (failing()) return false;
2602 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2603 }
2604 }
2605 return true;
2606 }
2607
2608 // Remove edges from "root" to each SafePoint at a backward branch.
2609 // They were inserted during parsing (see add_safepoint()) to make
2610 // infinite loops without calls or exceptions visible to root, i.e.,
2611 // useful.
2612 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2613 Node *r = root();
2614 if (r != NULL) {
2615 for (uint i = r->req(); i < r->len(); ++i) {
2616 Node *n = r->in(i);
2617 if (n != NULL && n->is_SafePoint()) {
2618 r->rm_prec(i);
2619 if (n->outcnt() == 0) {
2620 igvn.remove_dead_node(n);
2621 }
2622 --i;
2623 }
2624 }
2625 }
2626 }
2627
2628 //------------------------------Optimize---------------------------------------
2629 // Given a graph, optimize it.
2630 void Compile::Optimize() {
2631 TracePhase tp("optimizer", &timers[_t_optimizer]);
2632
2633 #ifndef PRODUCT
2634 if (_directive->BreakAtCompileOption) {
2635 BREAKPOINT;
2636 }
2637
2638 #endif
2639
2640 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2641 #ifdef ASSERT
2642 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2643 #endif
2644
2645 ResourceMark rm;
2646
2647 print_inlining_reinit();
2648
2649 NOT_PRODUCT( verify_graph_edges(); )
2650
2651 print_method(PHASE_AFTER_PARSING);
2652
2653 {
2654 // Iterative Global Value Numbering, including ideal transforms
2655 // Initialize IterGVN with types and values from parse-time GVN
2656 PhaseIterGVN igvn(initial_gvn());
2657 #ifdef ASSERT
2658 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2659 #endif
2660 {
2661 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2662 igvn.optimize();
2663 }
2664
2665 if (failing()) return;
2666
2667 print_method(PHASE_ITER_GVN1, 2);
2668
2669 inline_incrementally(igvn);
2670
2671 print_method(PHASE_INCREMENTAL_INLINE, 2);
2672
2673 if (failing()) return;
2674
2675 if (eliminate_boxing()) {
2676 // Inline valueOf() methods now.
2677 inline_boxing_calls(igvn);
2678
2679 if (AlwaysIncrementalInline) {
2680 inline_incrementally(igvn);
2681 }
2682
2683 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2684
2685 if (failing()) return;
2686 }
2687
2688 // Now that all inlining is over, cut edge from root to loop
2689 // safepoints
2690 remove_root_to_sfpts_edges(igvn);
2691
2692 // Remove the speculative part of types and clean up the graph from
2693 // the extra CastPP nodes whose only purpose is to carry them. Do
2694 // that early so that optimizations are not disrupted by the extra
2695 // CastPP nodes.
2696 remove_speculative_types(igvn);
2697
2698 // No more new expensive nodes will be added to the list from here
2699 // so keep only the actual candidates for optimizations.
2700 cleanup_expensive_nodes(igvn);
2701
2702 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2703 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2704 initial_gvn()->replace_with(&igvn);
2705 for_igvn()->clear();
2706 Unique_Node_List new_worklist(C->comp_arena());
2707 {
2708 ResourceMark rm;
2709 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2710 }
2711 set_for_igvn(&new_worklist);
2712 igvn = PhaseIterGVN(initial_gvn());
2713 igvn.optimize();
2714 }
2715
2716 if (_value_type_nodes->size() > 0) {
2717 // Do this once all inlining is over to avoid getting inconsistent debug info
2718 process_value_types(igvn);
2719 }
2720
2721 adjust_flattened_array_access_aliases(igvn);
2722
2723 // Perform escape analysis
2724 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2725 if (has_loops()) {
2726 // Cleanup graph (remove dead nodes).
2727 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2728 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2729 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2730 if (failing()) return;
2731 }
2732 ConnectionGraph::do_analysis(this, &igvn);
2733
2734 if (failing()) return;
2735
2736 // Optimize out fields loads from scalar replaceable allocations.
2737 igvn.optimize();
2738 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2739
2740 if (failing()) return;
2741
2742 if (congraph() != NULL && macro_count() > 0) {
2743 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2744 PhaseMacroExpand mexp(igvn);
2745 mexp.eliminate_macro_nodes();
2746 igvn.set_delay_transform(false);
2747
2748 igvn.optimize();
2749 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2750
2751 if (failing()) return;
2752 }
2753 }
2754
2755 // Loop transforms on the ideal graph. Range Check Elimination,
2756 // peeling, unrolling, etc.
2757
2758 // Set loop opts counter
2759 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2760 {
2761 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2762 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2763 _loop_opts_cnt--;
2764 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2765 if (failing()) return;
2766 }
2767 // Loop opts pass if partial peeling occurred in previous pass
2768 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2769 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2770 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2771 _loop_opts_cnt--;
2772 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2773 if (failing()) return;
2774 }
2775 // Loop opts pass for loop-unrolling before CCP
2776 if(major_progress() && (_loop_opts_cnt > 0)) {
2777 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2778 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2779 _loop_opts_cnt--;
2780 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2781 }
2782 if (!failing()) {
2783 // Verify that last round of loop opts produced a valid graph
2784 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2785 PhaseIdealLoop::verify(igvn);
2786 }
2787 }
2788 if (failing()) return;
2789
2790 // Conditional Constant Propagation;
2791 PhaseCCP ccp( &igvn );
2792 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2793 {
2794 TracePhase tp("ccp", &timers[_t_ccp]);
2795 ccp.do_transform();
2796 }
2797 print_method(PHASE_CPP1, 2);
2798
2799 assert( true, "Break here to ccp.dump_old2new_map()");
2800
2801 // Iterative Global Value Numbering, including ideal transforms
2802 {
2803 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2804 igvn = ccp;
2805 igvn.optimize();
2806 }
2807 print_method(PHASE_ITER_GVN2, 2);
2808
2809 if (failing()) return;
2810
2811 // Loop transforms on the ideal graph. Range Check Elimination,
2812 // peeling, unrolling, etc.
2813 if (!optimize_loops(igvn, LoopOptsDefault)) {
2814 return;
2815 }
2816
2817 if (failing()) return;
2818
2819 // Ensure that major progress is now clear
2820 C->clear_major_progress();
2821
2822 {
2823 // Verify that all previous optimizations produced a valid graph
2824 // at least to this point, even if no loop optimizations were done.
2825 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2826 PhaseIdealLoop::verify(igvn);
2827 }
2828
2829 if (range_check_cast_count() > 0) {
2830 // No more loop optimizations. Remove all range check dependent CastIINodes.
2831 C->remove_range_check_casts(igvn);
2832 igvn.optimize();
2833 }
2834
2835 #ifdef ASSERT
2836 bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion);
2837 #endif
2838
2839 bs->barrier_insertion_phase(C, igvn);
2840 if (failing()) return;
2841
2842 #ifdef ASSERT
2843 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2844 #endif
2845
2846 {
2847 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2848 PhaseMacroExpand mex(igvn);
2849 if (mex.expand_macro_nodes()) {
2850 assert(failing(), "must bail out w/ explicit message");
2851 return;
2852 }
2853 print_method(PHASE_MACRO_EXPANSION, 2);
2854 }
2855
2856 {
2857 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2858 if (bs->expand_barriers(this, igvn)) {
2859 assert(failing(), "must bail out w/ explicit message");
2860 return;
2861 }
2862 print_method(PHASE_BARRIER_EXPANSION, 2);
2863 }
2864
2865 if (opaque4_count() > 0) {
2866 C->remove_opaque4_nodes(igvn);
2867 igvn.optimize();
2868 }
2869
2870 DEBUG_ONLY( _modified_nodes = NULL; )
2871 } // (End scope of igvn; run destructor if necessary for asserts.)
2872
2873 process_print_inlining();
2874 // A method with only infinite loops has no edges entering loops from root
2875 {
2876 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2877 if (final_graph_reshaping()) {
2878 assert(failing(), "must bail out w/ explicit message");
2879 return;
2880 }
2881 }
2882
2883 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2884 }
2885
2886 //------------------------------Code_Gen---------------------------------------
2887 // Given a graph, generate code for it
2888 void Compile::Code_Gen() {
2889 if (failing()) {
2890 return;
2891 }
2892
2893 // Perform instruction selection. You might think we could reclaim Matcher
2894 // memory PDQ, but actually the Matcher is used in generating spill code.
2895 // Internals of the Matcher (including some VectorSets) must remain live
2896 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2897 // set a bit in reclaimed memory.
2898
2899 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2900 // nodes. Mapping is only valid at the root of each matched subtree.
2901 NOT_PRODUCT( verify_graph_edges(); )
2902
2903 Matcher matcher;
2904 _matcher = &matcher;
2905 {
2906 TracePhase tp("matcher", &timers[_t_matcher]);
2907 matcher.match();
2908 }
2909 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2910 // nodes. Mapping is only valid at the root of each matched subtree.
2911 NOT_PRODUCT( verify_graph_edges(); )
2912
2913 // If you have too many nodes, or if matching has failed, bail out
2914 check_node_count(0, "out of nodes matching instructions");
2915 if (failing()) {
2916 return;
2917 }
2918
2919 print_method(PHASE_MATCHING, 2);
2920
2921 // Build a proper-looking CFG
2922 PhaseCFG cfg(node_arena(), root(), matcher);
2923 _cfg = &cfg;
2924 {
2925 TracePhase tp("scheduler", &timers[_t_scheduler]);
2926 bool success = cfg.do_global_code_motion();
2927 if (!success) {
2928 return;
2929 }
2930
2931 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2932 NOT_PRODUCT( verify_graph_edges(); )
2933 debug_only( cfg.verify(); )
2934 }
2935
2936 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2937 _regalloc = ®alloc;
2938 {
2939 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2940 // Perform register allocation. After Chaitin, use-def chains are
2941 // no longer accurate (at spill code) and so must be ignored.
2942 // Node->LRG->reg mappings are still accurate.
2943 _regalloc->Register_Allocate();
2944
2945 // Bail out if the allocator builds too many nodes
2946 if (failing()) {
2947 return;
2948 }
2949 }
2950
2951 // Prior to register allocation we kept empty basic blocks in case the
2952 // the allocator needed a place to spill. After register allocation we
2953 // are not adding any new instructions. If any basic block is empty, we
2954 // can now safely remove it.
2955 {
2956 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2957 cfg.remove_empty_blocks();
2958 if (do_freq_based_layout()) {
2959 PhaseBlockLayout layout(cfg);
2960 } else {
2961 cfg.set_loop_alignment();
2962 }
2963 cfg.fixup_flow();
2964 }
2965
2966 // Apply peephole optimizations
2967 if( OptoPeephole ) {
2968 TracePhase tp("peephole", &timers[_t_peephole]);
2969 PhasePeephole peep( _regalloc, cfg);
2970 peep.do_transform();
2971 }
2972
2973 // Do late expand if CPU requires this.
2974 if (Matcher::require_postalloc_expand) {
2975 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2976 cfg.postalloc_expand(_regalloc);
2977 }
2978
2979 // Convert Nodes to instruction bits in a buffer
2980 {
2981 TraceTime tp("output", &timers[_t_output], CITime);
2982 Output();
2983 }
2984
2985 print_method(PHASE_FINAL_CODE);
2986
2987 // He's dead, Jim.
2988 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2989 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2990 }
2991
2992
2993 //------------------------------dump_asm---------------------------------------
2994 // Dump formatted assembly
2995 #if defined(SUPPORT_OPTO_ASSEMBLY)
2996 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2997
2998 int pc_digits = 3; // #chars required for pc
2999 int sb_chars = 3; // #chars for "start bundle" indicator
3000 int tab_size = 8;
3001 if (pcs != NULL) {
3002 int max_pc = 0;
3003 for (uint i = 0; i < pc_limit; i++) {
3004 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3005 }
3006 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3007 }
3008 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3009
3010 bool cut_short = false;
3011 st->print_cr("#");
3012 st->print("# "); _tf->dump_on(st); st->cr();
3013 st->print_cr("#");
3014
3015 // For all blocks
3016 int pc = 0x0; // Program counter
3017 char starts_bundle = ' ';
3018 _regalloc->dump_frame();
3019
3020 Node *n = NULL;
3021 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
3022 if (VMThread::should_terminate()) {
3023 cut_short = true;
3024 break;
3025 }
3026 Block* block = _cfg->get_block(i);
3027 if (block->is_connector() && !Verbose) {
3028 continue;
3029 }
3030 n = block->head();
3031 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3032 pc = pcs[n->_idx];
3033 st->print("%*.*x", pc_digits, pc_digits, pc);
3034 }
3035 st->fill_to(prefix_len);
3036 block->dump_head(_cfg, st);
3037 if (block->is_connector()) {
3038 st->fill_to(prefix_len);
3039 st->print_cr("# Empty connector block");
3040 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3041 st->fill_to(prefix_len);
3042 st->print_cr("# Block is sole successor of call");
3043 }
3044
3045 // For all instructions
3046 Node *delay = NULL;
3047 for (uint j = 0; j < block->number_of_nodes(); j++) {
3048 if (VMThread::should_terminate()) {
3049 cut_short = true;
3050 break;
3051 }
3052 n = block->get_node(j);
3053 if (valid_bundle_info(n)) {
3054 Bundle* bundle = node_bundling(n);
3055 if (bundle->used_in_unconditional_delay()) {
3056 delay = n;
3057 continue;
3058 }
3059 if (bundle->starts_bundle()) {
3060 starts_bundle = '+';
3061 }
3062 }
3063
3064 if (WizardMode) {
3065 n->dump();
3066 }
3067
3068 if( !n->is_Region() && // Dont print in the Assembly
3069 !n->is_Phi() && // a few noisely useless nodes
3070 !n->is_Proj() &&
3071 !n->is_MachTemp() &&
3072 !n->is_SafePointScalarObject() &&
3073 !n->is_Catch() && // Would be nice to print exception table targets
3074 !n->is_MergeMem() && // Not very interesting
3075 !n->is_top() && // Debug info table constants
3076 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3077 ) {
3078 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3079 pc = pcs[n->_idx];
3080 st->print("%*.*x", pc_digits, pc_digits, pc);
3081 } else {
3082 st->fill_to(pc_digits);
3083 }
3084 st->print(" %c ", starts_bundle);
3085 starts_bundle = ' ';
3086 st->fill_to(prefix_len);
3087 n->format(_regalloc, st);
3088 st->cr();
3089 }
3090
3091 // If we have an instruction with a delay slot, and have seen a delay,
3092 // then back up and print it
3093 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3094 // Coverity finding - Explicit null dereferenced.
3095 guarantee(delay != NULL, "no unconditional delay instruction");
3096 if (WizardMode) delay->dump();
3097
3098 if (node_bundling(delay)->starts_bundle())
3099 starts_bundle = '+';
3100 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3101 pc = pcs[n->_idx];
3102 st->print("%*.*x", pc_digits, pc_digits, pc);
3103 } else {
3104 st->fill_to(pc_digits);
3105 }
3106 st->print(" %c ", starts_bundle);
3107 starts_bundle = ' ';
3108 st->fill_to(prefix_len);
3109 delay->format(_regalloc, st);
3110 st->cr();
3111 delay = NULL;
3112 }
3113
3114 // Dump the exception table as well
3115 if( n->is_Catch() && (Verbose || WizardMode) ) {
3116 // Print the exception table for this offset
3117 _handler_table.print_subtable_for(pc);
3118 }
3119 st->bol(); // Make sure we start on a new line
3120 }
3121 st->cr(); // one empty line between blocks
3122 assert(cut_short || delay == NULL, "no unconditional delay branch");
3123 } // End of per-block dump
3124
3125 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3126 }
3127 #endif
3128
3129 //------------------------------Final_Reshape_Counts---------------------------
3130 // This class defines counters to help identify when a method
3131 // may/must be executed using hardware with only 24-bit precision.
3132 struct Final_Reshape_Counts : public StackObj {
3133 int _call_count; // count non-inlined 'common' calls
3134 int _float_count; // count float ops requiring 24-bit precision
3135 int _double_count; // count double ops requiring more precision
3136 int _java_call_count; // count non-inlined 'java' calls
3137 int _inner_loop_count; // count loops which need alignment
3138 VectorSet _visited; // Visitation flags
3139 Node_List _tests; // Set of IfNodes & PCTableNodes
3140
3141 Final_Reshape_Counts() :
3142 _call_count(0), _float_count(0), _double_count(0),
3143 _java_call_count(0), _inner_loop_count(0),
3144 _visited( Thread::current()->resource_area() ) { }
3145
3146 void inc_call_count () { _call_count ++; }
3147 void inc_float_count () { _float_count ++; }
3148 void inc_double_count() { _double_count++; }
3149 void inc_java_call_count() { _java_call_count++; }
3150 void inc_inner_loop_count() { _inner_loop_count++; }
3151
3152 int get_call_count () const { return _call_count ; }
3153 int get_float_count () const { return _float_count ; }
3154 int get_double_count() const { return _double_count; }
3155 int get_java_call_count() const { return _java_call_count; }
3156 int get_inner_loop_count() const { return _inner_loop_count; }
3157 };
3158
3159 #ifdef ASSERT
3160 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
3161 ciInstanceKlass *k = tp->klass()->as_instance_klass();
3162 // Make sure the offset goes inside the instance layout.
3163 return k->contains_field_offset(tp->offset());
3164 // Note that OffsetBot and OffsetTop are very negative.
3165 }
3166 #endif
3167
3168 // Eliminate trivially redundant StoreCMs and accumulate their
3169 // precedence edges.
3170 void Compile::eliminate_redundant_card_marks(Node* n) {
3171 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3172 if (n->in(MemNode::Address)->outcnt() > 1) {
3173 // There are multiple users of the same address so it might be
3174 // possible to eliminate some of the StoreCMs
3175 Node* mem = n->in(MemNode::Memory);
3176 Node* adr = n->in(MemNode::Address);
3177 Node* val = n->in(MemNode::ValueIn);
3178 Node* prev = n;
3179 bool done = false;
3180 // Walk the chain of StoreCMs eliminating ones that match. As
3181 // long as it's a chain of single users then the optimization is
3182 // safe. Eliminating partially redundant StoreCMs would require
3183 // cloning copies down the other paths.
3184 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3185 if (adr == mem->in(MemNode::Address) &&
3186 val == mem->in(MemNode::ValueIn)) {
3187 // redundant StoreCM
3188 if (mem->req() > MemNode::OopStore) {
3189 // Hasn't been processed by this code yet.
3190 n->add_prec(mem->in(MemNode::OopStore));
3191 } else {
3192 // Already converted to precedence edge
3193 for (uint i = mem->req(); i < mem->len(); i++) {
3194 // Accumulate any precedence edges
3195 if (mem->in(i) != NULL) {
3196 n->add_prec(mem->in(i));
3197 }
3198 }
3199 // Everything above this point has been processed.
3200 done = true;
3201 }
3202 // Eliminate the previous StoreCM
3203 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3204 assert(mem->outcnt() == 0, "should be dead");
3205 mem->disconnect_inputs(NULL, this);
3206 } else {
3207 prev = mem;
3208 }
3209 mem = prev->in(MemNode::Memory);
3210 }
3211 }
3212 }
3213
3214
3215 //------------------------------final_graph_reshaping_impl----------------------
3216 // Implement items 1-5 from final_graph_reshaping below.
3217 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3218
3219 if ( n->outcnt() == 0 ) return; // dead node
3220 uint nop = n->Opcode();
3221
3222 // Check for 2-input instruction with "last use" on right input.
3223 // Swap to left input. Implements item (2).
3224 if( n->req() == 3 && // two-input instruction
3225 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3226 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3227 n->in(2)->outcnt() == 1 &&// right use IS a last use
3228 !n->in(2)->is_Con() ) { // right use is not a constant
3229 // Check for commutative opcode
3230 switch( nop ) {
3231 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3232 case Op_MaxI: case Op_MinI:
3233 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3234 case Op_AndL: case Op_XorL: case Op_OrL:
3235 case Op_AndI: case Op_XorI: case Op_OrI: {
3236 // Move "last use" input to left by swapping inputs
3237 n->swap_edges(1, 2);
3238 break;
3239 }
3240 default:
3241 break;
3242 }
3243 }
3244
3245 #ifdef ASSERT
3246 if( n->is_Mem() ) {
3247 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3248 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3249 // oop will be recorded in oop map if load crosses safepoint
3250 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3251 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3252 "raw memory operations should have control edge");
3253 }
3254 if (n->is_MemBar()) {
3255 MemBarNode* mb = n->as_MemBar();
3256 if (mb->trailing_store() || mb->trailing_load_store()) {
3257 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3258 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3259 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3260 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3261 } else if (mb->leading()) {
3262 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3263 }
3264 }
3265 #endif
3266 // Count FPU ops and common calls, implements item (3)
3267 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3268 if (!gc_handled) {
3269 final_graph_reshaping_main_switch(n, frc, nop);
3270 }
3271
3272 // Collect CFG split points
3273 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3274 frc._tests.push(n);
3275 }
3276 }
3277
3278 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3279 switch( nop ) {
3280 // Count all float operations that may use FPU
3281 case Op_AddF:
3282 case Op_SubF:
3283 case Op_MulF:
3284 case Op_DivF:
3285 case Op_NegF:
3286 case Op_ModF:
3287 case Op_ConvI2F:
3288 case Op_ConF:
3289 case Op_CmpF:
3290 case Op_CmpF3:
3291 // case Op_ConvL2F: // longs are split into 32-bit halves
3292 frc.inc_float_count();
3293 break;
3294
3295 case Op_ConvF2D:
3296 case Op_ConvD2F:
3297 frc.inc_float_count();
3298 frc.inc_double_count();
3299 break;
3300
3301 // Count all double operations that may use FPU
3302 case Op_AddD:
3303 case Op_SubD:
3304 case Op_MulD:
3305 case Op_DivD:
3306 case Op_NegD:
3307 case Op_ModD:
3308 case Op_ConvI2D:
3309 case Op_ConvD2I:
3310 // case Op_ConvL2D: // handled by leaf call
3311 // case Op_ConvD2L: // handled by leaf call
3312 case Op_ConD:
3313 case Op_CmpD:
3314 case Op_CmpD3:
3315 frc.inc_double_count();
3316 break;
3317 case Op_Opaque1: // Remove Opaque Nodes before matching
3318 case Op_Opaque2: // Remove Opaque Nodes before matching
3319 case Op_Opaque3:
3320 n->subsume_by(n->in(1), this);
3321 break;
3322 case Op_CallStaticJava:
3323 case Op_CallJava:
3324 case Op_CallDynamicJava:
3325 frc.inc_java_call_count(); // Count java call site;
3326 case Op_CallRuntime:
3327 case Op_CallLeaf:
3328 case Op_CallLeafNoFP: {
3329 assert (n->is_Call(), "");
3330 CallNode *call = n->as_Call();
3331 // Count call sites where the FP mode bit would have to be flipped.
3332 // Do not count uncommon runtime calls:
3333 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3334 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3335 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3336 frc.inc_call_count(); // Count the call site
3337 } else { // See if uncommon argument is shared
3338 Node *n = call->in(TypeFunc::Parms);
3339 int nop = n->Opcode();
3340 // Clone shared simple arguments to uncommon calls, item (1).
3341 if (n->outcnt() > 1 &&
3342 !n->is_Proj() &&
3343 nop != Op_CreateEx &&
3344 nop != Op_CheckCastPP &&
3345 nop != Op_DecodeN &&
3346 nop != Op_DecodeNKlass &&
3347 !n->is_Mem() &&
3348 !n->is_Phi()) {
3349 Node *x = n->clone();
3350 call->set_req(TypeFunc::Parms, x);
3351 }
3352 }
3353 break;
3354 }
3355
3356 case Op_StoreD:
3357 case Op_LoadD:
3358 case Op_LoadD_unaligned:
3359 frc.inc_double_count();
3360 goto handle_mem;
3361 case Op_StoreF:
3362 case Op_LoadF:
3363 frc.inc_float_count();
3364 goto handle_mem;
3365
3366 case Op_StoreCM:
3367 {
3368 // Convert OopStore dependence into precedence edge
3369 Node* prec = n->in(MemNode::OopStore);
3370 n->del_req(MemNode::OopStore);
3371 n->add_prec(prec);
3372 eliminate_redundant_card_marks(n);
3373 }
3374
3375 // fall through
3376
3377 case Op_StoreB:
3378 case Op_StoreC:
3379 case Op_StorePConditional:
3380 case Op_StoreI:
3381 case Op_StoreL:
3382 case Op_StoreIConditional:
3383 case Op_StoreLConditional:
3384 case Op_CompareAndSwapB:
3385 case Op_CompareAndSwapS:
3386 case Op_CompareAndSwapI:
3387 case Op_CompareAndSwapL:
3388 case Op_CompareAndSwapP:
3389 case Op_CompareAndSwapN:
3390 case Op_WeakCompareAndSwapB:
3391 case Op_WeakCompareAndSwapS:
3392 case Op_WeakCompareAndSwapI:
3393 case Op_WeakCompareAndSwapL:
3394 case Op_WeakCompareAndSwapP:
3395 case Op_WeakCompareAndSwapN:
3396 case Op_CompareAndExchangeB:
3397 case Op_CompareAndExchangeS:
3398 case Op_CompareAndExchangeI:
3399 case Op_CompareAndExchangeL:
3400 case Op_CompareAndExchangeP:
3401 case Op_CompareAndExchangeN:
3402 case Op_GetAndAddS:
3403 case Op_GetAndAddB:
3404 case Op_GetAndAddI:
3405 case Op_GetAndAddL:
3406 case Op_GetAndSetS:
3407 case Op_GetAndSetB:
3408 case Op_GetAndSetI:
3409 case Op_GetAndSetL:
3410 case Op_GetAndSetP:
3411 case Op_GetAndSetN:
3412 case Op_StoreP:
3413 case Op_StoreN:
3414 case Op_StoreNKlass:
3415 case Op_LoadB:
3416 case Op_LoadUB:
3417 case Op_LoadUS:
3418 case Op_LoadI:
3419 case Op_LoadKlass:
3420 case Op_LoadNKlass:
3421 case Op_LoadL:
3422 case Op_LoadL_unaligned:
3423 case Op_LoadPLocked:
3424 case Op_LoadP:
3425 case Op_LoadN:
3426 case Op_LoadRange:
3427 case Op_LoadS: {
3428 handle_mem:
3429 #ifdef ASSERT
3430 if( VerifyOptoOopOffsets ) {
3431 MemNode* mem = n->as_Mem();
3432 // Check to see if address types have grounded out somehow.
3433 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3434 assert( !tp || oop_offset_is_sane(tp), "" );
3435 }
3436 #endif
3437 if (EnableValhalla && (nop == Op_LoadKlass || nop == Op_LoadNKlass)) {
3438 const TypeKlassPtr* tk = n->bottom_type()->make_ptr()->is_klassptr();
3439 assert(!tk->klass_is_exact(), "should have been folded");
3440 if (tk->klass()->can_be_value_array_klass()) {
3441 // Array load klass needs to filter out property bits (but not
3442 // GetNullFreePropertyNode which needs to extract the null free bits)
3443 uint last = unique();
3444 Node* pointer = NULL;
3445 if (nop == Op_LoadKlass) {
3446 Node* cast = new CastP2XNode(NULL, n);
3447 Node* masked = new LShiftXNode(cast, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3448 masked = new RShiftXNode(masked, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3449 pointer = new CastX2PNode(masked);
3450 pointer = new CheckCastPPNode(NULL, pointer, n->bottom_type());
3451 } else {
3452 Node* cast = new CastN2INode(n);
3453 Node* masked = new AndINode(cast, new ConINode(TypeInt::make(oopDesc::compressed_klass_mask())));
3454 pointer = new CastI2NNode(masked, n->bottom_type());
3455 }
3456 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3457 Node* u = n->fast_out(i);
3458 if (u->_idx < last && u->Opcode() != Op_GetNullFreeProperty) {
3459 // If user is a comparison with a klass that can't be a value type
3460 // array klass, we don't need to clear the storage property bits.
3461 Node* cmp = (u->is_DecodeNKlass() && u->outcnt() == 1) ? u->unique_out() : u;
3462 if (cmp->is_Cmp()) {
3463 const TypeKlassPtr* kp1 = cmp->in(1)->bottom_type()->make_ptr()->isa_klassptr();
3464 const TypeKlassPtr* kp2 = cmp->in(2)->bottom_type()->make_ptr()->isa_klassptr();
3465 if ((kp1 != NULL && !kp1->klass()->can_be_value_array_klass()) ||
3466 (kp2 != NULL && !kp2->klass()->can_be_value_array_klass())) {
3467 continue;
3468 }
3469 }
3470 int nb = u->replace_edge(n, pointer);
3471 --i, imax -= nb;
3472 }
3473 }
3474 }
3475 }
3476 break;
3477 }
3478
3479 case Op_AddP: { // Assert sane base pointers
3480 Node *addp = n->in(AddPNode::Address);
3481 assert( !addp->is_AddP() ||
3482 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3483 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3484 "Base pointers must match (addp %u)", addp->_idx );
3485 #ifdef _LP64
3486 if ((UseCompressedOops || UseCompressedClassPointers) &&
3487 addp->Opcode() == Op_ConP &&
3488 addp == n->in(AddPNode::Base) &&
3489 n->in(AddPNode::Offset)->is_Con()) {
3490 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3491 // on the platform and on the compressed oops mode.
3492 // Use addressing with narrow klass to load with offset on x86.
3493 // Some platforms can use the constant pool to load ConP.
3494 // Do this transformation here since IGVN will convert ConN back to ConP.
3495 const Type* t = addp->bottom_type();
3496 bool is_oop = t->isa_oopptr() != NULL;
3497 bool is_klass = t->isa_klassptr() != NULL;
3498
3499 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3500 (is_klass && Matcher::const_klass_prefer_decode())) {
3501 Node* nn = NULL;
3502
3503 int op = is_oop ? Op_ConN : Op_ConNKlass;
3504
3505 // Look for existing ConN node of the same exact type.
3506 Node* r = root();
3507 uint cnt = r->outcnt();
3508 for (uint i = 0; i < cnt; i++) {
3509 Node* m = r->raw_out(i);
3510 if (m!= NULL && m->Opcode() == op &&
3511 m->bottom_type()->make_ptr() == t) {
3512 nn = m;
3513 break;
3514 }
3515 }
3516 if (nn != NULL) {
3517 // Decode a narrow oop to match address
3518 // [R12 + narrow_oop_reg<<3 + offset]
3519 if (is_oop) {
3520 nn = new DecodeNNode(nn, t);
3521 } else {
3522 nn = new DecodeNKlassNode(nn, t);
3523 }
3524 // Check for succeeding AddP which uses the same Base.
3525 // Otherwise we will run into the assertion above when visiting that guy.
3526 for (uint i = 0; i < n->outcnt(); ++i) {
3527 Node *out_i = n->raw_out(i);
3528 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3529 out_i->set_req(AddPNode::Base, nn);
3530 #ifdef ASSERT
3531 for (uint j = 0; j < out_i->outcnt(); ++j) {
3532 Node *out_j = out_i->raw_out(j);
3533 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3534 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3535 }
3536 #endif
3537 }
3538 }
3539 n->set_req(AddPNode::Base, nn);
3540 n->set_req(AddPNode::Address, nn);
3541 if (addp->outcnt() == 0) {
3542 addp->disconnect_inputs(NULL, this);
3543 }
3544 }
3545 }
3546 }
3547 #endif
3548 // platform dependent reshaping of the address expression
3549 reshape_address(n->as_AddP());
3550 break;
3551 }
3552
3553 case Op_CastPP: {
3554 // Remove CastPP nodes to gain more freedom during scheduling but
3555 // keep the dependency they encode as control or precedence edges
3556 // (if control is set already) on memory operations. Some CastPP
3557 // nodes don't have a control (don't carry a dependency): skip
3558 // those.
3559 if (n->in(0) != NULL) {
3560 ResourceMark rm;
3561 Unique_Node_List wq;
3562 wq.push(n);
3563 for (uint next = 0; next < wq.size(); ++next) {
3564 Node *m = wq.at(next);
3565 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3566 Node* use = m->fast_out(i);
3567 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3568 use->ensure_control_or_add_prec(n->in(0));
3569 } else {
3570 switch(use->Opcode()) {
3571 case Op_AddP:
3572 case Op_DecodeN:
3573 case Op_DecodeNKlass:
3574 case Op_CheckCastPP:
3575 case Op_CastPP:
3576 wq.push(use);
3577 break;
3578 }
3579 }
3580 }
3581 }
3582 }
3583 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3584 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3585 Node* in1 = n->in(1);
3586 const Type* t = n->bottom_type();
3587 Node* new_in1 = in1->clone();
3588 new_in1->as_DecodeN()->set_type(t);
3589
3590 if (!Matcher::narrow_oop_use_complex_address()) {
3591 //
3592 // x86, ARM and friends can handle 2 adds in addressing mode
3593 // and Matcher can fold a DecodeN node into address by using
3594 // a narrow oop directly and do implicit NULL check in address:
3595 //
3596 // [R12 + narrow_oop_reg<<3 + offset]
3597 // NullCheck narrow_oop_reg
3598 //
3599 // On other platforms (Sparc) we have to keep new DecodeN node and
3600 // use it to do implicit NULL check in address:
3601 //
3602 // decode_not_null narrow_oop_reg, base_reg
3603 // [base_reg + offset]
3604 // NullCheck base_reg
3605 //
3606 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3607 // to keep the information to which NULL check the new DecodeN node
3608 // corresponds to use it as value in implicit_null_check().
3609 //
3610 new_in1->set_req(0, n->in(0));
3611 }
3612
3613 n->subsume_by(new_in1, this);
3614 if (in1->outcnt() == 0) {
3615 in1->disconnect_inputs(NULL, this);
3616 }
3617 } else {
3618 n->subsume_by(n->in(1), this);
3619 if (n->outcnt() == 0) {
3620 n->disconnect_inputs(NULL, this);
3621 }
3622 }
3623 break;
3624 }
3625 #ifdef _LP64
3626 case Op_CmpP:
3627 // Do this transformation here to preserve CmpPNode::sub() and
3628 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3629 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3630 Node* in1 = n->in(1);
3631 Node* in2 = n->in(2);
3632 if (!in1->is_DecodeNarrowPtr()) {
3633 in2 = in1;
3634 in1 = n->in(2);
3635 }
3636 assert(in1->is_DecodeNarrowPtr(), "sanity");
3637
3638 Node* new_in2 = NULL;
3639 if (in2->is_DecodeNarrowPtr()) {
3640 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3641 new_in2 = in2->in(1);
3642 } else if (in2->Opcode() == Op_ConP) {
3643 const Type* t = in2->bottom_type();
3644 if (t == TypePtr::NULL_PTR) {
3645 assert(in1->is_DecodeN(), "compare klass to null?");
3646 // Don't convert CmpP null check into CmpN if compressed
3647 // oops implicit null check is not generated.
3648 // This will allow to generate normal oop implicit null check.
3649 if (Matcher::gen_narrow_oop_implicit_null_checks())
3650 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3651 //
3652 // This transformation together with CastPP transformation above
3653 // will generated code for implicit NULL checks for compressed oops.
3654 //
3655 // The original code after Optimize()
3656 //
3657 // LoadN memory, narrow_oop_reg
3658 // decode narrow_oop_reg, base_reg
3659 // CmpP base_reg, NULL
3660 // CastPP base_reg // NotNull
3661 // Load [base_reg + offset], val_reg
3662 //
3663 // after these transformations will be
3664 //
3665 // LoadN memory, narrow_oop_reg
3666 // CmpN narrow_oop_reg, NULL
3667 // decode_not_null narrow_oop_reg, base_reg
3668 // Load [base_reg + offset], val_reg
3669 //
3670 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3671 // since narrow oops can be used in debug info now (see the code in
3672 // final_graph_reshaping_walk()).
3673 //
3674 // At the end the code will be matched to
3675 // on x86:
3676 //
3677 // Load_narrow_oop memory, narrow_oop_reg
3678 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3679 // NullCheck narrow_oop_reg
3680 //
3681 // and on sparc:
3682 //
3683 // Load_narrow_oop memory, narrow_oop_reg
3684 // decode_not_null narrow_oop_reg, base_reg
3685 // Load [base_reg + offset], val_reg
3686 // NullCheck base_reg
3687 //
3688 } else if (t->isa_oopptr()) {
3689 new_in2 = ConNode::make(t->make_narrowoop());
3690 } else if (t->isa_klassptr()) {
3691 new_in2 = ConNode::make(t->make_narrowklass());
3692 }
3693 }
3694 if (new_in2 != NULL) {
3695 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3696 n->subsume_by(cmpN, this);
3697 if (in1->outcnt() == 0) {
3698 in1->disconnect_inputs(NULL, this);
3699 }
3700 if (in2->outcnt() == 0) {
3701 in2->disconnect_inputs(NULL, this);
3702 }
3703 }
3704 }
3705 break;
3706
3707 case Op_DecodeN:
3708 case Op_DecodeNKlass:
3709 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3710 // DecodeN could be pinned when it can't be fold into
3711 // an address expression, see the code for Op_CastPP above.
3712 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3713 break;
3714
3715 case Op_EncodeP:
3716 case Op_EncodePKlass: {
3717 Node* in1 = n->in(1);
3718 if (in1->is_DecodeNarrowPtr()) {
3719 n->subsume_by(in1->in(1), this);
3720 } else if (in1->Opcode() == Op_ConP) {
3721 const Type* t = in1->bottom_type();
3722 if (t == TypePtr::NULL_PTR) {
3723 assert(t->isa_oopptr(), "null klass?");
3724 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3725 } else if (t->isa_oopptr()) {
3726 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3727 } else if (t->isa_klassptr()) {
3728 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3729 }
3730 }
3731 if (in1->outcnt() == 0) {
3732 in1->disconnect_inputs(NULL, this);
3733 }
3734 break;
3735 }
3736
3737 case Op_Proj: {
3738 if (OptimizeStringConcat) {
3739 ProjNode* p = n->as_Proj();
3740 if (p->_is_io_use) {
3741 // Separate projections were used for the exception path which
3742 // are normally removed by a late inline. If it wasn't inlined
3743 // then they will hang around and should just be replaced with
3744 // the original one.
3745 Node* proj = NULL;
3746 // Replace with just one
3747 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3748 Node *use = i.get();
3749 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3750 proj = use;
3751 break;
3752 }
3753 }
3754 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3755 if (proj != NULL) {
3756 p->subsume_by(proj, this);
3757 }
3758 }
3759 }
3760 break;
3761 }
3762
3763 case Op_Phi:
3764 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3765 // The EncodeP optimization may create Phi with the same edges
3766 // for all paths. It is not handled well by Register Allocator.
3767 Node* unique_in = n->in(1);
3768 assert(unique_in != NULL, "");
3769 uint cnt = n->req();
3770 for (uint i = 2; i < cnt; i++) {
3771 Node* m = n->in(i);
3772 assert(m != NULL, "");
3773 if (unique_in != m)
3774 unique_in = NULL;
3775 }
3776 if (unique_in != NULL) {
3777 n->subsume_by(unique_in, this);
3778 }
3779 }
3780 break;
3781
3782 #endif
3783
3784 #ifdef ASSERT
3785 case Op_CastII:
3786 // Verify that all range check dependent CastII nodes were removed.
3787 if (n->isa_CastII()->has_range_check()) {
3788 n->dump(3);
3789 assert(false, "Range check dependent CastII node was not removed");
3790 }
3791 break;
3792 #endif
3793
3794 case Op_ModI:
3795 if (UseDivMod) {
3796 // Check if a%b and a/b both exist
3797 Node* d = n->find_similar(Op_DivI);
3798 if (d) {
3799 // Replace them with a fused divmod if supported
3800 if (Matcher::has_match_rule(Op_DivModI)) {
3801 DivModINode* divmod = DivModINode::make(n);
3802 d->subsume_by(divmod->div_proj(), this);
3803 n->subsume_by(divmod->mod_proj(), this);
3804 } else {
3805 // replace a%b with a-((a/b)*b)
3806 Node* mult = new MulINode(d, d->in(2));
3807 Node* sub = new SubINode(d->in(1), mult);
3808 n->subsume_by(sub, this);
3809 }
3810 }
3811 }
3812 break;
3813
3814 case Op_ModL:
3815 if (UseDivMod) {
3816 // Check if a%b and a/b both exist
3817 Node* d = n->find_similar(Op_DivL);
3818 if (d) {
3819 // Replace them with a fused divmod if supported
3820 if (Matcher::has_match_rule(Op_DivModL)) {
3821 DivModLNode* divmod = DivModLNode::make(n);
3822 d->subsume_by(divmod->div_proj(), this);
3823 n->subsume_by(divmod->mod_proj(), this);
3824 } else {
3825 // replace a%b with a-((a/b)*b)
3826 Node* mult = new MulLNode(d, d->in(2));
3827 Node* sub = new SubLNode(d->in(1), mult);
3828 n->subsume_by(sub, this);
3829 }
3830 }
3831 }
3832 break;
3833
3834 case Op_LoadVector:
3835 case Op_StoreVector:
3836 break;
3837
3838 case Op_AddReductionVI:
3839 case Op_AddReductionVL:
3840 case Op_AddReductionVF:
3841 case Op_AddReductionVD:
3842 case Op_MulReductionVI:
3843 case Op_MulReductionVL:
3844 case Op_MulReductionVF:
3845 case Op_MulReductionVD:
3846 case Op_MinReductionV:
3847 case Op_MaxReductionV:
3848 break;
3849
3850 case Op_PackB:
3851 case Op_PackS:
3852 case Op_PackI:
3853 case Op_PackF:
3854 case Op_PackL:
3855 case Op_PackD:
3856 if (n->req()-1 > 2) {
3857 // Replace many operand PackNodes with a binary tree for matching
3858 PackNode* p = (PackNode*) n;
3859 Node* btp = p->binary_tree_pack(1, n->req());
3860 n->subsume_by(btp, this);
3861 }
3862 break;
3863 case Op_Loop:
3864 case Op_CountedLoop:
3865 case Op_OuterStripMinedLoop:
3866 if (n->as_Loop()->is_inner_loop()) {
3867 frc.inc_inner_loop_count();
3868 }
3869 n->as_Loop()->verify_strip_mined(0);
3870 break;
3871 case Op_LShiftI:
3872 case Op_RShiftI:
3873 case Op_URShiftI:
3874 case Op_LShiftL:
3875 case Op_RShiftL:
3876 case Op_URShiftL:
3877 if (Matcher::need_masked_shift_count) {
3878 // The cpu's shift instructions don't restrict the count to the
3879 // lower 5/6 bits. We need to do the masking ourselves.
3880 Node* in2 = n->in(2);
3881 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3882 const TypeInt* t = in2->find_int_type();
3883 if (t != NULL && t->is_con()) {
3884 juint shift = t->get_con();
3885 if (shift > mask) { // Unsigned cmp
3886 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3887 }
3888 } else {
3889 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3890 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3891 n->set_req(2, shift);
3892 }
3893 }
3894 if (in2->outcnt() == 0) { // Remove dead node
3895 in2->disconnect_inputs(NULL, this);
3896 }
3897 }
3898 break;
3899 case Op_MemBarStoreStore:
3900 case Op_MemBarRelease:
3901 // Break the link with AllocateNode: it is no longer useful and
3902 // confuses register allocation.
3903 if (n->req() > MemBarNode::Precedent) {
3904 n->set_req(MemBarNode::Precedent, top());
3905 }
3906 break;
3907 case Op_MemBarAcquire: {
3908 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3909 // At parse time, the trailing MemBarAcquire for a volatile load
3910 // is created with an edge to the load. After optimizations,
3911 // that input may be a chain of Phis. If those phis have no
3912 // other use, then the MemBarAcquire keeps them alive and
3913 // register allocation can be confused.
3914 ResourceMark rm;
3915 Unique_Node_List wq;
3916 wq.push(n->in(MemBarNode::Precedent));
3917 n->set_req(MemBarNode::Precedent, top());
3918 while (wq.size() > 0) {
3919 Node* m = wq.pop();
3920 if (m->outcnt() == 0) {
3921 for (uint j = 0; j < m->req(); j++) {
3922 Node* in = m->in(j);
3923 if (in != NULL) {
3924 wq.push(in);
3925 }
3926 }
3927 m->disconnect_inputs(NULL, this);
3928 }
3929 }
3930 }
3931 break;
3932 }
3933 case Op_RangeCheck: {
3934 RangeCheckNode* rc = n->as_RangeCheck();
3935 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3936 n->subsume_by(iff, this);
3937 frc._tests.push(iff);
3938 break;
3939 }
3940 case Op_ConvI2L: {
3941 if (!Matcher::convi2l_type_required) {
3942 // Code generation on some platforms doesn't need accurate
3943 // ConvI2L types. Widening the type can help remove redundant
3944 // address computations.
3945 n->as_Type()->set_type(TypeLong::INT);
3946 ResourceMark rm;
3947 Node_List wq;
3948 wq.push(n);
3949 for (uint next = 0; next < wq.size(); next++) {
3950 Node *m = wq.at(next);
3951
3952 for(;;) {
3953 // Loop over all nodes with identical inputs edges as m
3954 Node* k = m->find_similar(m->Opcode());
3955 if (k == NULL) {
3956 break;
3957 }
3958 // Push their uses so we get a chance to remove node made
3959 // redundant
3960 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3961 Node* u = k->fast_out(i);
3962 assert(!wq.contains(u), "shouldn't process one node several times");
3963 if (u->Opcode() == Op_LShiftL ||
3964 u->Opcode() == Op_AddL ||
3965 u->Opcode() == Op_SubL ||
3966 u->Opcode() == Op_AddP) {
3967 wq.push(u);
3968 }
3969 }
3970 // Replace all nodes with identical edges as m with m
3971 k->subsume_by(m, this);
3972 }
3973 }
3974 }
3975 break;
3976 }
3977 case Op_CmpUL: {
3978 if (!Matcher::has_match_rule(Op_CmpUL)) {
3979 // No support for unsigned long comparisons
3980 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3981 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3982 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3983 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3984 Node* andl = new AndLNode(orl, remove_sign_mask);
3985 Node* cmp = new CmpLNode(andl, n->in(2));
3986 n->subsume_by(cmp, this);
3987 }
3988 break;
3989 }
3990 #ifdef ASSERT
3991 case Op_ValueTypePtr:
3992 case Op_ValueType: {
3993 n->dump(-1);
3994 assert(false, "value type node was not removed");
3995 break;
3996 }
3997 #endif
3998 case Op_GetNullFreeProperty: {
3999 // Extract the null free bits
4000 uint last = unique();
4001 Node* null_free = NULL;
4002 if (n->in(1)->Opcode() == Op_LoadKlass) {
4003 Node* cast = new CastP2XNode(NULL, n->in(1));
4004 null_free = new AndLNode(cast, new ConLNode(TypeLong::make(((jlong)1)<<(oopDesc::wide_storage_props_shift + ArrayStorageProperties::null_free_bit))));
4005 } else {
4006 assert(n->in(1)->Opcode() == Op_LoadNKlass, "not a compressed klass?");
4007 Node* cast = new CastN2INode(n->in(1));
4008 null_free = new AndINode(cast, new ConINode(TypeInt::make(1<<(oopDesc::narrow_storage_props_shift + ArrayStorageProperties::null_free_bit))));
4009 }
4010 n->replace_by(null_free);
4011 break;
4012 }
4013 default:
4014 assert(!n->is_Call(), "");
4015 assert(!n->is_Mem(), "");
4016 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4017 break;
4018 }
4019 }
4020
4021 //------------------------------final_graph_reshaping_walk---------------------
4022 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4023 // requires that the walk visits a node's inputs before visiting the node.
4024 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4025 ResourceArea *area = Thread::current()->resource_area();
4026 Unique_Node_List sfpt(area);
4027
4028 frc._visited.set(root->_idx); // first, mark node as visited
4029 uint cnt = root->req();
4030 Node *n = root;
4031 uint i = 0;
4032 while (true) {
4033 if (i < cnt) {
4034 // Place all non-visited non-null inputs onto stack
4035 Node* m = n->in(i);
4036 ++i;
4037 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4038 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4039 // compute worst case interpreter size in case of a deoptimization
4040 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4041
4042 sfpt.push(m);
4043 }
4044 cnt = m->req();
4045 nstack.push(n, i); // put on stack parent and next input's index
4046 n = m;
4047 i = 0;
4048 }
4049 } else {
4050 // Now do post-visit work
4051 final_graph_reshaping_impl( n, frc );
4052 if (nstack.is_empty())
4053 break; // finished
4054 n = nstack.node(); // Get node from stack
4055 cnt = n->req();
4056 i = nstack.index();
4057 nstack.pop(); // Shift to the next node on stack
4058 }
4059 }
4060
4061 // Skip next transformation if compressed oops are not used.
4062 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4063 (!UseCompressedOops && !UseCompressedClassPointers))
4064 return;
4065
4066 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4067 // It could be done for an uncommon traps or any safepoints/calls
4068 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4069 while (sfpt.size() > 0) {
4070 n = sfpt.pop();
4071 JVMState *jvms = n->as_SafePoint()->jvms();
4072 assert(jvms != NULL, "sanity");
4073 int start = jvms->debug_start();
4074 int end = n->req();
4075 bool is_uncommon = (n->is_CallStaticJava() &&
4076 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4077 for (int j = start; j < end; j++) {
4078 Node* in = n->in(j);
4079 if (in->is_DecodeNarrowPtr()) {
4080 bool safe_to_skip = true;
4081 if (!is_uncommon ) {
4082 // Is it safe to skip?
4083 for (uint i = 0; i < in->outcnt(); i++) {
4084 Node* u = in->raw_out(i);
4085 if (!u->is_SafePoint() ||
4086 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4087 safe_to_skip = false;
4088 }
4089 }
4090 }
4091 if (safe_to_skip) {
4092 n->set_req(j, in->in(1));
4093 }
4094 if (in->outcnt() == 0) {
4095 in->disconnect_inputs(NULL, this);
4096 }
4097 }
4098 }
4099 }
4100 }
4101
4102 //------------------------------final_graph_reshaping--------------------------
4103 // Final Graph Reshaping.
4104 //
4105 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4106 // and not commoned up and forced early. Must come after regular
4107 // optimizations to avoid GVN undoing the cloning. Clone constant
4108 // inputs to Loop Phis; these will be split by the allocator anyways.
4109 // Remove Opaque nodes.
4110 // (2) Move last-uses by commutative operations to the left input to encourage
4111 // Intel update-in-place two-address operations and better register usage
4112 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4113 // calls canonicalizing them back.
4114 // (3) Count the number of double-precision FP ops, single-precision FP ops
4115 // and call sites. On Intel, we can get correct rounding either by
4116 // forcing singles to memory (requires extra stores and loads after each
4117 // FP bytecode) or we can set a rounding mode bit (requires setting and
4118 // clearing the mode bit around call sites). The mode bit is only used
4119 // if the relative frequency of single FP ops to calls is low enough.
4120 // This is a key transform for SPEC mpeg_audio.
4121 // (4) Detect infinite loops; blobs of code reachable from above but not
4122 // below. Several of the Code_Gen algorithms fail on such code shapes,
4123 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4124 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4125 // Detection is by looking for IfNodes where only 1 projection is
4126 // reachable from below or CatchNodes missing some targets.
4127 // (5) Assert for insane oop offsets in debug mode.
4128
4129 bool Compile::final_graph_reshaping() {
4130 // an infinite loop may have been eliminated by the optimizer,
4131 // in which case the graph will be empty.
4132 if (root()->req() == 1) {
4133 record_method_not_compilable("trivial infinite loop");
4134 return true;
4135 }
4136
4137 // Expensive nodes have their control input set to prevent the GVN
4138 // from freely commoning them. There's no GVN beyond this point so
4139 // no need to keep the control input. We want the expensive nodes to
4140 // be freely moved to the least frequent code path by gcm.
4141 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4142 for (int i = 0; i < expensive_count(); i++) {
4143 _expensive_nodes->at(i)->set_req(0, NULL);
4144 }
4145
4146 Final_Reshape_Counts frc;
4147
4148 // Visit everybody reachable!
4149 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4150 Node_Stack nstack(live_nodes() >> 1);
4151 final_graph_reshaping_walk(nstack, root(), frc);
4152
4153 // Check for unreachable (from below) code (i.e., infinite loops).
4154 for( uint i = 0; i < frc._tests.size(); i++ ) {
4155 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4156 // Get number of CFG targets.
4157 // Note that PCTables include exception targets after calls.
4158 uint required_outcnt = n->required_outcnt();
4159 if (n->outcnt() != required_outcnt) {
4160 // Check for a few special cases. Rethrow Nodes never take the
4161 // 'fall-thru' path, so expected kids is 1 less.
4162 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4163 if (n->in(0)->in(0)->is_Call()) {
4164 CallNode *call = n->in(0)->in(0)->as_Call();
4165 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4166 required_outcnt--; // Rethrow always has 1 less kid
4167 } else if (call->req() > TypeFunc::Parms &&
4168 call->is_CallDynamicJava()) {
4169 // Check for null receiver. In such case, the optimizer has
4170 // detected that the virtual call will always result in a null
4171 // pointer exception. The fall-through projection of this CatchNode
4172 // will not be populated.
4173 Node *arg0 = call->in(TypeFunc::Parms);
4174 if (arg0->is_Type() &&
4175 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4176 required_outcnt--;
4177 }
4178 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4179 call->req() > TypeFunc::Parms+1 &&
4180 call->is_CallStaticJava()) {
4181 // Check for negative array length. In such case, the optimizer has
4182 // detected that the allocation attempt will always result in an
4183 // exception. There is no fall-through projection of this CatchNode .
4184 Node *arg1 = call->in(TypeFunc::Parms+1);
4185 if (arg1->is_Type() &&
4186 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4187 required_outcnt--;
4188 }
4189 }
4190 }
4191 }
4192 // Recheck with a better notion of 'required_outcnt'
4193 if (n->outcnt() != required_outcnt) {
4194 record_method_not_compilable("malformed control flow");
4195 return true; // Not all targets reachable!
4196 }
4197 }
4198 // Check that I actually visited all kids. Unreached kids
4199 // must be infinite loops.
4200 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4201 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4202 record_method_not_compilable("infinite loop");
4203 return true; // Found unvisited kid; must be unreach
4204 }
4205
4206 // Here so verification code in final_graph_reshaping_walk()
4207 // always see an OuterStripMinedLoopEnd
4208 if (n->is_OuterStripMinedLoopEnd()) {
4209 IfNode* init_iff = n->as_If();
4210 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4211 n->subsume_by(iff, this);
4212 }
4213 }
4214
4215 // If original bytecodes contained a mixture of floats and doubles
4216 // check if the optimizer has made it homogenous, item (3).
4217 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
4218 frc.get_float_count() > 32 &&
4219 frc.get_double_count() == 0 &&
4220 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4221 set_24_bit_selection_and_mode( false, true );
4222 }
4223
4224 set_java_calls(frc.get_java_call_count());
4225 set_inner_loops(frc.get_inner_loop_count());
4226
4227 // No infinite loops, no reason to bail out.
4228 return false;
4229 }
4230
4231 //-----------------------------too_many_traps----------------------------------
4232 // Report if there are too many traps at the current method and bci.
4233 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4234 bool Compile::too_many_traps(ciMethod* method,
4235 int bci,
4236 Deoptimization::DeoptReason reason) {
4237 ciMethodData* md = method->method_data();
4238 if (md->is_empty()) {
4239 // Assume the trap has not occurred, or that it occurred only
4240 // because of a transient condition during start-up in the interpreter.
4241 return false;
4242 }
4243 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4244 if (md->has_trap_at(bci, m, reason) != 0) {
4245 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4246 // Also, if there are multiple reasons, or if there is no per-BCI record,
4247 // assume the worst.
4248 if (log())
4249 log()->elem("observe trap='%s' count='%d'",
4250 Deoptimization::trap_reason_name(reason),
4251 md->trap_count(reason));
4252 return true;
4253 } else {
4254 // Ignore method/bci and see if there have been too many globally.
4255 return too_many_traps(reason, md);
4256 }
4257 }
4258
4259 // Less-accurate variant which does not require a method and bci.
4260 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4261 ciMethodData* logmd) {
4262 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4263 // Too many traps globally.
4264 // Note that we use cumulative trap_count, not just md->trap_count.
4265 if (log()) {
4266 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4267 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4268 Deoptimization::trap_reason_name(reason),
4269 mcount, trap_count(reason));
4270 }
4271 return true;
4272 } else {
4273 // The coast is clear.
4274 return false;
4275 }
4276 }
4277
4278 //--------------------------too_many_recompiles--------------------------------
4279 // Report if there are too many recompiles at the current method and bci.
4280 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4281 // Is not eager to return true, since this will cause the compiler to use
4282 // Action_none for a trap point, to avoid too many recompilations.
4283 bool Compile::too_many_recompiles(ciMethod* method,
4284 int bci,
4285 Deoptimization::DeoptReason reason) {
4286 ciMethodData* md = method->method_data();
4287 if (md->is_empty()) {
4288 // Assume the trap has not occurred, or that it occurred only
4289 // because of a transient condition during start-up in the interpreter.
4290 return false;
4291 }
4292 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4293 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4294 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4295 Deoptimization::DeoptReason per_bc_reason
4296 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4297 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4298 if ((per_bc_reason == Deoptimization::Reason_none
4299 || md->has_trap_at(bci, m, reason) != 0)
4300 // The trap frequency measure we care about is the recompile count:
4301 && md->trap_recompiled_at(bci, m)
4302 && md->overflow_recompile_count() >= bc_cutoff) {
4303 // Do not emit a trap here if it has already caused recompilations.
4304 // Also, if there are multiple reasons, or if there is no per-BCI record,
4305 // assume the worst.
4306 if (log())
4307 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4308 Deoptimization::trap_reason_name(reason),
4309 md->trap_count(reason),
4310 md->overflow_recompile_count());
4311 return true;
4312 } else if (trap_count(reason) != 0
4313 && decompile_count() >= m_cutoff) {
4314 // Too many recompiles globally, and we have seen this sort of trap.
4315 // Use cumulative decompile_count, not just md->decompile_count.
4316 if (log())
4317 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4318 Deoptimization::trap_reason_name(reason),
4319 md->trap_count(reason), trap_count(reason),
4320 md->decompile_count(), decompile_count());
4321 return true;
4322 } else {
4323 // The coast is clear.
4324 return false;
4325 }
4326 }
4327
4328 // Compute when not to trap. Used by matching trap based nodes and
4329 // NullCheck optimization.
4330 void Compile::set_allowed_deopt_reasons() {
4331 _allowed_reasons = 0;
4332 if (is_method_compilation()) {
4333 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4334 assert(rs < BitsPerInt, "recode bit map");
4335 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4336 _allowed_reasons |= nth_bit(rs);
4337 }
4338 }
4339 }
4340 }
4341
4342 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4343 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4344 }
4345
4346 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4347 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4348 }
4349
4350 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4351 if (holder->is_initialized()) {
4352 return false;
4353 }
4354 if (holder->is_being_initialized()) {
4355 if (accessing_method->holder() == holder) {
4356 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4357 // <init>, or a static method. In all those cases, there was an initialization
4358 // barrier on the holder klass passed.
4359 if (accessing_method->is_class_initializer() ||
4360 accessing_method->is_object_constructor() ||
4361 accessing_method->is_static()) {
4362 return false;
4363 }
4364 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4365 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4366 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4367 // child class can become fully initialized while its parent class is still being initialized.
4368 if (accessing_method->is_class_initializer()) {
4369 return false;
4370 }
4371 }
4372 ciMethod* root = method(); // the root method of compilation
4373 if (root != accessing_method) {
4374 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4375 }
4376 }
4377 return true;
4378 }
4379
4380 #ifndef PRODUCT
4381 //------------------------------verify_graph_edges---------------------------
4382 // Walk the Graph and verify that there is a one-to-one correspondence
4383 // between Use-Def edges and Def-Use edges in the graph.
4384 void Compile::verify_graph_edges(bool no_dead_code) {
4385 if (VerifyGraphEdges) {
4386 ResourceArea *area = Thread::current()->resource_area();
4387 Unique_Node_List visited(area);
4388 // Call recursive graph walk to check edges
4389 _root->verify_edges(visited);
4390 if (no_dead_code) {
4391 // Now make sure that no visited node is used by an unvisited node.
4392 bool dead_nodes = false;
4393 Unique_Node_List checked(area);
4394 while (visited.size() > 0) {
4395 Node* n = visited.pop();
4396 checked.push(n);
4397 for (uint i = 0; i < n->outcnt(); i++) {
4398 Node* use = n->raw_out(i);
4399 if (checked.member(use)) continue; // already checked
4400 if (visited.member(use)) continue; // already in the graph
4401 if (use->is_Con()) continue; // a dead ConNode is OK
4402 // At this point, we have found a dead node which is DU-reachable.
4403 if (!dead_nodes) {
4404 tty->print_cr("*** Dead nodes reachable via DU edges:");
4405 dead_nodes = true;
4406 }
4407 use->dump(2);
4408 tty->print_cr("---");
4409 checked.push(use); // No repeats; pretend it is now checked.
4410 }
4411 }
4412 assert(!dead_nodes, "using nodes must be reachable from root");
4413 }
4414 }
4415 }
4416 #endif
4417
4418 // The Compile object keeps track of failure reasons separately from the ciEnv.
4419 // This is required because there is not quite a 1-1 relation between the
4420 // ciEnv and its compilation task and the Compile object. Note that one
4421 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4422 // to backtrack and retry without subsuming loads. Other than this backtracking
4423 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4424 // by the logic in C2Compiler.
4425 void Compile::record_failure(const char* reason) {
4426 if (log() != NULL) {
4427 log()->elem("failure reason='%s' phase='compile'", reason);
4428 }
4429 if (_failure_reason == NULL) {
4430 // Record the first failure reason.
4431 _failure_reason = reason;
4432 }
4433
4434 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4435 C->print_method(PHASE_FAILURE);
4436 }
4437 _root = NULL; // flush the graph, too
4438 }
4439
4440 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4441 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4442 _phase_name(name), _dolog(CITimeVerbose)
4443 {
4444 if (_dolog) {
4445 C = Compile::current();
4446 _log = C->log();
4447 } else {
4448 C = NULL;
4449 _log = NULL;
4450 }
4451 if (_log != NULL) {
4452 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4453 _log->stamp();
4454 _log->end_head();
4455 }
4456 }
4457
4458 Compile::TracePhase::~TracePhase() {
4459
4460 C = Compile::current();
4461 if (_dolog) {
4462 _log = C->log();
4463 } else {
4464 _log = NULL;
4465 }
4466
4467 #ifdef ASSERT
4468 if (PrintIdealNodeCount) {
4469 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4470 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4471 }
4472
4473 if (VerifyIdealNodeCount) {
4474 Compile::current()->print_missing_nodes();
4475 }
4476 #endif
4477
4478 if (_log != NULL) {
4479 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4480 }
4481 }
4482
4483 //=============================================================================
4484 // Two Constant's are equal when the type and the value are equal.
4485 bool Compile::Constant::operator==(const Constant& other) {
4486 if (type() != other.type() ) return false;
4487 if (can_be_reused() != other.can_be_reused()) return false;
4488 // For floating point values we compare the bit pattern.
4489 switch (type()) {
4490 case T_INT:
4491 case T_FLOAT: return (_v._value.i == other._v._value.i);
4492 case T_LONG:
4493 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4494 case T_OBJECT:
4495 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4496 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4497 case T_METADATA: return (_v._metadata == other._v._metadata);
4498 default: ShouldNotReachHere(); return false;
4499 }
4500 }
4501
4502 static int type_to_size_in_bytes(BasicType t) {
4503 switch (t) {
4504 case T_INT: return sizeof(jint );
4505 case T_LONG: return sizeof(jlong );
4506 case T_FLOAT: return sizeof(jfloat );
4507 case T_DOUBLE: return sizeof(jdouble);
4508 case T_METADATA: return sizeof(Metadata*);
4509 // We use T_VOID as marker for jump-table entries (labels) which
4510 // need an internal word relocation.
4511 case T_VOID:
4512 case T_ADDRESS:
4513 case T_OBJECT: return sizeof(jobject);
4514 default:
4515 ShouldNotReachHere();
4516 return -1;
4517 }
4518 }
4519
4520 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4521 // sort descending
4522 if (a->freq() > b->freq()) return -1;
4523 if (a->freq() < b->freq()) return 1;
4524 return 0;
4525 }
4526
4527 void Compile::ConstantTable::calculate_offsets_and_size() {
4528 // First, sort the array by frequencies.
4529 _constants.sort(qsort_comparator);
4530
4531 #ifdef ASSERT
4532 // Make sure all jump-table entries were sorted to the end of the
4533 // array (they have a negative frequency).
4534 bool found_void = false;
4535 for (int i = 0; i < _constants.length(); i++) {
4536 Constant con = _constants.at(i);
4537 if (con.type() == T_VOID)
4538 found_void = true; // jump-tables
4539 else
4540 assert(!found_void, "wrong sorting");
4541 }
4542 #endif
4543
4544 int offset = 0;
4545 for (int i = 0; i < _constants.length(); i++) {
4546 Constant* con = _constants.adr_at(i);
4547
4548 // Align offset for type.
4549 int typesize = type_to_size_in_bytes(con->type());
4550 offset = align_up(offset, typesize);
4551 con->set_offset(offset); // set constant's offset
4552
4553 if (con->type() == T_VOID) {
4554 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4555 offset = offset + typesize * n->outcnt(); // expand jump-table
4556 } else {
4557 offset = offset + typesize;
4558 }
4559 }
4560
4561 // Align size up to the next section start (which is insts; see
4562 // CodeBuffer::align_at_start).
4563 assert(_size == -1, "already set?");
4564 _size = align_up(offset, (int)CodeEntryAlignment);
4565 }
4566
4567 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4568 MacroAssembler _masm(&cb);
4569 for (int i = 0; i < _constants.length(); i++) {
4570 Constant con = _constants.at(i);
4571 address constant_addr = NULL;
4572 switch (con.type()) {
4573 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4574 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4575 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4576 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4577 case T_OBJECT: {
4578 jobject obj = con.get_jobject();
4579 int oop_index = _masm.oop_recorder()->find_index(obj);
4580 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4581 break;
4582 }
4583 case T_ADDRESS: {
4584 address addr = (address) con.get_jobject();
4585 constant_addr = _masm.address_constant(addr);
4586 break;
4587 }
4588 // We use T_VOID as marker for jump-table entries (labels) which
4589 // need an internal word relocation.
4590 case T_VOID: {
4591 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4592 // Fill the jump-table with a dummy word. The real value is
4593 // filled in later in fill_jump_table.
4594 address dummy = (address) n;
4595 constant_addr = _masm.address_constant(dummy);
4596 // Expand jump-table
4597 for (uint i = 1; i < n->outcnt(); i++) {
4598 address temp_addr = _masm.address_constant(dummy + i);
4599 assert(temp_addr, "consts section too small");
4600 }
4601 break;
4602 }
4603 case T_METADATA: {
4604 Metadata* obj = con.get_metadata();
4605 int metadata_index = _masm.oop_recorder()->find_index(obj);
4606 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4607 break;
4608 }
4609 default: ShouldNotReachHere();
4610 }
4611 assert(constant_addr, "consts section too small");
4612 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4613 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4614 }
4615 }
4616
4617 int Compile::ConstantTable::find_offset(Constant& con) const {
4618 int idx = _constants.find(con);
4619 guarantee(idx != -1, "constant must be in constant table");
4620 int offset = _constants.at(idx).offset();
4621 guarantee(offset != -1, "constant table not emitted yet?");
4622 return offset;
4623 }
4624
4625 void Compile::ConstantTable::add(Constant& con) {
4626 if (con.can_be_reused()) {
4627 int idx = _constants.find(con);
4628 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4629 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4630 return;
4631 }
4632 }
4633 (void) _constants.append(con);
4634 }
4635
4636 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4637 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4638 Constant con(type, value, b->_freq);
4639 add(con);
4640 return con;
4641 }
4642
4643 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4644 Constant con(metadata);
4645 add(con);
4646 return con;
4647 }
4648
4649 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4650 jvalue value;
4651 BasicType type = oper->type()->basic_type();
4652 switch (type) {
4653 case T_LONG: value.j = oper->constantL(); break;
4654 case T_FLOAT: value.f = oper->constantF(); break;
4655 case T_DOUBLE: value.d = oper->constantD(); break;
4656 case T_OBJECT:
4657 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4658 case T_METADATA: return add((Metadata*)oper->constant()); break;
4659 default: guarantee(false, "unhandled type: %s", type2name(type));
4660 }
4661 return add(n, type, value);
4662 }
4663
4664 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4665 jvalue value;
4666 // We can use the node pointer here to identify the right jump-table
4667 // as this method is called from Compile::Fill_buffer right before
4668 // the MachNodes are emitted and the jump-table is filled (means the
4669 // MachNode pointers do not change anymore).
4670 value.l = (jobject) n;
4671 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4672 add(con);
4673 return con;
4674 }
4675
4676 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4677 // If called from Compile::scratch_emit_size do nothing.
4678 if (Compile::current()->in_scratch_emit_size()) return;
4679
4680 assert(labels.is_nonempty(), "must be");
4681 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4682
4683 // Since MachConstantNode::constant_offset() also contains
4684 // table_base_offset() we need to subtract the table_base_offset()
4685 // to get the plain offset into the constant table.
4686 int offset = n->constant_offset() - table_base_offset();
4687
4688 MacroAssembler _masm(&cb);
4689 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4690
4691 for (uint i = 0; i < n->outcnt(); i++) {
4692 address* constant_addr = &jump_table_base[i];
4693 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));
4694 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4695 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4696 }
4697 }
4698
4699 //----------------------------static_subtype_check-----------------------------
4700 // Shortcut important common cases when superklass is exact:
4701 // (0) superklass is java.lang.Object (can occur in reflective code)
4702 // (1) subklass is already limited to a subtype of superklass => always ok
4703 // (2) subklass does not overlap with superklass => always fail
4704 // (3) superklass has NO subtypes and we can check with a simple compare.
4705 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4706 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4707 return SSC_full_test; // Let caller generate the general case.
4708 }
4709
4710 if (superk == env()->Object_klass()) {
4711 return SSC_always_true; // (0) this test cannot fail
4712 }
4713
4714 ciType* superelem = superk;
4715 if (superelem->is_array_klass()) {
4716 ciArrayKlass* ak = superelem->as_array_klass();
4717 superelem = superelem->as_array_klass()->base_element_type();
4718 }
4719
4720 if (!subk->is_interface()) { // cannot trust static interface types yet
4721 if (subk->is_subtype_of(superk)) {
4722 return SSC_always_true; // (1) false path dead; no dynamic test needed
4723 }
4724 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4725 !superk->is_subtype_of(subk)) {
4726 return SSC_always_false;
4727 }
4728 }
4729
4730 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4731 // [V is a subtype of [V? but the klass for [V is not equal to the klass for [V?. Perform a full test.
4732 if (superk->is_obj_array_klass() && !superk->as_array_klass()->storage_properties().is_null_free() && superk->as_array_klass()->element_klass()->is_valuetype()) {
4733 return SSC_full_test;
4734 }
4735 // If casting to an instance klass, it must have no subtypes
4736 if (superk->is_interface()) {
4737 // Cannot trust interfaces yet.
4738 // %%% S.B. superk->nof_implementors() == 1
4739 } else if (superelem->is_instance_klass()) {
4740 ciInstanceKlass* ik = superelem->as_instance_klass();
4741 if (!ik->has_subklass() && !ik->is_interface()) {
4742 if (!ik->is_final()) {
4743 // Add a dependency if there is a chance of a later subclass.
4744 dependencies()->assert_leaf_type(ik);
4745 }
4746 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4747 }
4748 } else {
4749 // A primitive array type has no subtypes.
4750 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4751 }
4752
4753 return SSC_full_test;
4754 }
4755
4756 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4757 #ifdef _LP64
4758 // The scaled index operand to AddP must be a clean 64-bit value.
4759 // Java allows a 32-bit int to be incremented to a negative
4760 // value, which appears in a 64-bit register as a large
4761 // positive number. Using that large positive number as an
4762 // operand in pointer arithmetic has bad consequences.
4763 // On the other hand, 32-bit overflow is rare, and the possibility
4764 // can often be excluded, if we annotate the ConvI2L node with
4765 // a type assertion that its value is known to be a small positive
4766 // number. (The prior range check has ensured this.)
4767 // This assertion is used by ConvI2LNode::Ideal.
4768 int index_max = max_jint - 1; // array size is max_jint, index is one less
4769 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4770 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4771 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4772 #endif
4773 return idx;
4774 }
4775
4776 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4777 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4778 if (ctrl != NULL) {
4779 // Express control dependency by a CastII node with a narrow type.
4780 value = new CastIINode(value, itype, false, true /* range check dependency */);
4781 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4782 // node from floating above the range check during loop optimizations. Otherwise, the
4783 // ConvI2L node may be eliminated independently of the range check, causing the data path
4784 // to become TOP while the control path is still there (although it's unreachable).
4785 value->set_req(0, ctrl);
4786 // Save CastII node to remove it after loop optimizations.
4787 phase->C->add_range_check_cast(value);
4788 value = phase->transform(value);
4789 }
4790 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4791 return phase->transform(new ConvI2LNode(value, ltype));
4792 }
4793
4794 // The message about the current inlining is accumulated in
4795 // _print_inlining_stream and transfered into the _print_inlining_list
4796 // once we know whether inlining succeeds or not. For regular
4797 // inlining, messages are appended to the buffer pointed by
4798 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4799 // a new buffer is added after _print_inlining_idx in the list. This
4800 // way we can update the inlining message for late inlining call site
4801 // when the inlining is attempted again.
4802 void Compile::print_inlining_init() {
4803 if (print_inlining() || print_intrinsics()) {
4804 _print_inlining_stream = new stringStream();
4805 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4806 }
4807 }
4808
4809 void Compile::print_inlining_reinit() {
4810 if (print_inlining() || print_intrinsics()) {
4811 // Re allocate buffer when we change ResourceMark
4812 _print_inlining_stream = new stringStream();
4813 }
4814 }
4815
4816 void Compile::print_inlining_reset() {
4817 _print_inlining_stream->reset();
4818 }
4819
4820 void Compile::print_inlining_commit() {
4821 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4822 // Transfer the message from _print_inlining_stream to the current
4823 // _print_inlining_list buffer and clear _print_inlining_stream.
4824 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4825 print_inlining_reset();
4826 }
4827
4828 void Compile::print_inlining_push() {
4829 // Add new buffer to the _print_inlining_list at current position
4830 _print_inlining_idx++;
4831 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4832 }
4833
4834 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4835 return _print_inlining_list->at(_print_inlining_idx);
4836 }
4837
4838 void Compile::print_inlining_update(CallGenerator* cg) {
4839 if (print_inlining() || print_intrinsics()) {
4840 if (!cg->is_late_inline()) {
4841 if (print_inlining_current().cg() != NULL) {
4842 print_inlining_push();
4843 }
4844 print_inlining_commit();
4845 } else {
4846 if (print_inlining_current().cg() != cg &&
4847 (print_inlining_current().cg() != NULL ||
4848 print_inlining_current().ss()->size() != 0)) {
4849 print_inlining_push();
4850 }
4851 print_inlining_commit();
4852 print_inlining_current().set_cg(cg);
4853 }
4854 }
4855 }
4856
4857 void Compile::print_inlining_move_to(CallGenerator* cg) {
4858 // We resume inlining at a late inlining call site. Locate the
4859 // corresponding inlining buffer so that we can update it.
4860 if (print_inlining()) {
4861 for (int i = 0; i < _print_inlining_list->length(); i++) {
4862 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4863 _print_inlining_idx = i;
4864 return;
4865 }
4866 }
4867 ShouldNotReachHere();
4868 }
4869 }
4870
4871 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4872 if (print_inlining()) {
4873 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4874 assert(print_inlining_current().cg() == cg, "wrong entry");
4875 // replace message with new message
4876 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4877 print_inlining_commit();
4878 print_inlining_current().set_cg(cg);
4879 }
4880 }
4881
4882 void Compile::print_inlining_assert_ready() {
4883 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4884 }
4885
4886 void Compile::process_print_inlining() {
4887 bool do_print_inlining = print_inlining() || print_intrinsics();
4888 if (do_print_inlining || log() != NULL) {
4889 // Print inlining message for candidates that we couldn't inline
4890 // for lack of space
4891 for (int i = 0; i < _late_inlines.length(); i++) {
4892 CallGenerator* cg = _late_inlines.at(i);
4893 if (!cg->is_mh_late_inline()) {
4894 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4895 if (do_print_inlining) {
4896 cg->print_inlining_late(msg);
4897 }
4898 log_late_inline_failure(cg, msg);
4899 }
4900 }
4901 }
4902 if (do_print_inlining) {
4903 ResourceMark rm;
4904 stringStream ss;
4905 for (int i = 0; i < _print_inlining_list->length(); i++) {
4906 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4907 }
4908 size_t end = ss.size();
4909 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4910 strncpy(_print_inlining_output, ss.base(), end+1);
4911 _print_inlining_output[end] = 0;
4912 }
4913 }
4914
4915 void Compile::dump_print_inlining() {
4916 if (_print_inlining_output != NULL) {
4917 tty->print_raw(_print_inlining_output);
4918 }
4919 }
4920
4921 void Compile::log_late_inline(CallGenerator* cg) {
4922 if (log() != NULL) {
4923 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4924 cg->unique_id());
4925 JVMState* p = cg->call_node()->jvms();
4926 while (p != NULL) {
4927 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4928 p = p->caller();
4929 }
4930 log()->tail("late_inline");
4931 }
4932 }
4933
4934 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4935 log_late_inline(cg);
4936 if (log() != NULL) {
4937 log()->inline_fail(msg);
4938 }
4939 }
4940
4941 void Compile::log_inline_id(CallGenerator* cg) {
4942 if (log() != NULL) {
4943 // The LogCompilation tool needs a unique way to identify late
4944 // inline call sites. This id must be unique for this call site in
4945 // this compilation. Try to have it unique across compilations as
4946 // well because it can be convenient when grepping through the log
4947 // file.
4948 // Distinguish OSR compilations from others in case CICountOSR is
4949 // on.
4950 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4951 cg->set_unique_id(id);
4952 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4953 }
4954 }
4955
4956 void Compile::log_inline_failure(const char* msg) {
4957 if (C->log() != NULL) {
4958 C->log()->inline_fail(msg);
4959 }
4960 }
4961
4962
4963 // Dump inlining replay data to the stream.
4964 // Don't change thread state and acquire any locks.
4965 void Compile::dump_inline_data(outputStream* out) {
4966 InlineTree* inl_tree = ilt();
4967 if (inl_tree != NULL) {
4968 out->print(" inline %d", inl_tree->count());
4969 inl_tree->dump_replay_data(out);
4970 }
4971 }
4972
4973 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4974 if (n1->Opcode() < n2->Opcode()) return -1;
4975 else if (n1->Opcode() > n2->Opcode()) return 1;
4976
4977 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4978 for (uint i = 1; i < n1->req(); i++) {
4979 if (n1->in(i) < n2->in(i)) return -1;
4980 else if (n1->in(i) > n2->in(i)) return 1;
4981 }
4982
4983 return 0;
4984 }
4985
4986 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4987 Node* n1 = *n1p;
4988 Node* n2 = *n2p;
4989
4990 return cmp_expensive_nodes(n1, n2);
4991 }
4992
4993 void Compile::sort_expensive_nodes() {
4994 if (!expensive_nodes_sorted()) {
4995 _expensive_nodes->sort(cmp_expensive_nodes);
4996 }
4997 }
4998
4999 bool Compile::expensive_nodes_sorted() const {
5000 for (int i = 1; i < _expensive_nodes->length(); i++) {
5001 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
5002 return false;
5003 }
5004 }
5005 return true;
5006 }
5007
5008 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5009 if (_expensive_nodes->length() == 0) {
5010 return false;
5011 }
5012
5013 assert(OptimizeExpensiveOps, "optimization off?");
5014
5015 // Take this opportunity to remove dead nodes from the list
5016 int j = 0;
5017 for (int i = 0; i < _expensive_nodes->length(); i++) {
5018 Node* n = _expensive_nodes->at(i);
5019 if (!n->is_unreachable(igvn)) {
5020 assert(n->is_expensive(), "should be expensive");
5021 _expensive_nodes->at_put(j, n);
5022 j++;
5023 }
5024 }
5025 _expensive_nodes->trunc_to(j);
5026
5027 // Then sort the list so that similar nodes are next to each other
5028 // and check for at least two nodes of identical kind with same data
5029 // inputs.
5030 sort_expensive_nodes();
5031
5032 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
5033 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
5034 return true;
5035 }
5036 }
5037
5038 return false;
5039 }
5040
5041 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5042 if (_expensive_nodes->length() == 0) {
5043 return;
5044 }
5045
5046 assert(OptimizeExpensiveOps, "optimization off?");
5047
5048 // Sort to bring similar nodes next to each other and clear the
5049 // control input of nodes for which there's only a single copy.
5050 sort_expensive_nodes();
5051
5052 int j = 0;
5053 int identical = 0;
5054 int i = 0;
5055 bool modified = false;
5056 for (; i < _expensive_nodes->length()-1; i++) {
5057 assert(j <= i, "can't write beyond current index");
5058 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
5059 identical++;
5060 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5061 continue;
5062 }
5063 if (identical > 0) {
5064 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5065 identical = 0;
5066 } else {
5067 Node* n = _expensive_nodes->at(i);
5068 igvn.replace_input_of(n, 0, NULL);
5069 igvn.hash_insert(n);
5070 modified = true;
5071 }
5072 }
5073 if (identical > 0) {
5074 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5075 } else if (_expensive_nodes->length() >= 1) {
5076 Node* n = _expensive_nodes->at(i);
5077 igvn.replace_input_of(n, 0, NULL);
5078 igvn.hash_insert(n);
5079 modified = true;
5080 }
5081 _expensive_nodes->trunc_to(j);
5082 if (modified) {
5083 igvn.optimize();
5084 }
5085 }
5086
5087 void Compile::add_expensive_node(Node * n) {
5088 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
5089 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5090 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5091 if (OptimizeExpensiveOps) {
5092 _expensive_nodes->append(n);
5093 } else {
5094 // Clear control input and let IGVN optimize expensive nodes if
5095 // OptimizeExpensiveOps is off.
5096 n->set_req(0, NULL);
5097 }
5098 }
5099
5100 /**
5101 * Remove the speculative part of types and clean up the graph
5102 */
5103 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5104 if (UseTypeSpeculation) {
5105 Unique_Node_List worklist;
5106 worklist.push(root());
5107 int modified = 0;
5108 // Go over all type nodes that carry a speculative type, drop the
5109 // speculative part of the type and enqueue the node for an igvn
5110 // which may optimize it out.
5111 for (uint next = 0; next < worklist.size(); ++next) {
5112 Node *n = worklist.at(next);
5113 if (n->is_Type()) {
5114 TypeNode* tn = n->as_Type();
5115 const Type* t = tn->type();
5116 const Type* t_no_spec = t->remove_speculative();
5117 if (t_no_spec != t) {
5118 bool in_hash = igvn.hash_delete(n);
5119 assert(in_hash, "node should be in igvn hash table");
5120 tn->set_type(t_no_spec);
5121 igvn.hash_insert(n);
5122 igvn._worklist.push(n); // give it a chance to go away
5123 modified++;
5124 }
5125 }
5126 uint max = n->len();
5127 for( uint i = 0; i < max; ++i ) {
5128 Node *m = n->in(i);
5129 if (not_a_node(m)) continue;
5130 worklist.push(m);
5131 }
5132 }
5133 // Drop the speculative part of all types in the igvn's type table
5134 igvn.remove_speculative_types();
5135 if (modified > 0) {
5136 igvn.optimize();
5137 }
5138 #ifdef ASSERT
5139 // Verify that after the IGVN is over no speculative type has resurfaced
5140 worklist.clear();
5141 worklist.push(root());
5142 for (uint next = 0; next < worklist.size(); ++next) {
5143 Node *n = worklist.at(next);
5144 const Type* t = igvn.type_or_null(n);
5145 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5146 if (n->is_Type()) {
5147 t = n->as_Type()->type();
5148 assert(t == t->remove_speculative(), "no more speculative types");
5149 }
5150 uint max = n->len();
5151 for( uint i = 0; i < max; ++i ) {
5152 Node *m = n->in(i);
5153 if (not_a_node(m)) continue;
5154 worklist.push(m);
5155 }
5156 }
5157 igvn.check_no_speculative_types();
5158 #endif
5159 }
5160 }
5161
5162 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5163 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5164 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5165 if (!EnableValhalla || ta == NULL || tb == NULL ||
5166 ta->is_zero_type() || tb->is_zero_type() ||
5167 !ta->can_be_value_type() || !tb->can_be_value_type()) {
5168 // Use old acmp if one operand is null or not a value type
5169 return new CmpPNode(a, b);
5170 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) {
5171 // We know that one operand is a value type. Therefore,
5172 // new acmp will only return true if both operands are NULL.
5173 // Check if both operands are null by or'ing the oops.
5174 a = phase->transform(new CastP2XNode(NULL, a));
5175 b = phase->transform(new CastP2XNode(NULL, b));
5176 a = phase->transform(new OrXNode(a, b));
5177 return new CmpXNode(a, phase->MakeConX(0));
5178 }
5179 // Use new acmp
5180 return NULL;
5181 }
5182
5183 // Auxiliary method to support randomized stressing/fuzzing.
5184 //
5185 // This method can be called the arbitrary number of times, with current count
5186 // as the argument. The logic allows selecting a single candidate from the
5187 // running list of candidates as follows:
5188 // int count = 0;
5189 // Cand* selected = null;
5190 // while(cand = cand->next()) {
5191 // if (randomized_select(++count)) {
5192 // selected = cand;
5193 // }
5194 // }
5195 //
5196 // Including count equalizes the chances any candidate is "selected".
5197 // This is useful when we don't have the complete list of candidates to choose
5198 // from uniformly. In this case, we need to adjust the randomicity of the
5199 // selection, or else we will end up biasing the selection towards the latter
5200 // candidates.
5201 //
5202 // Quick back-envelope calculation shows that for the list of n candidates
5203 // the equal probability for the candidate to persist as "best" can be
5204 // achieved by replacing it with "next" k-th candidate with the probability
5205 // of 1/k. It can be easily shown that by the end of the run, the
5206 // probability for any candidate is converged to 1/n, thus giving the
5207 // uniform distribution among all the candidates.
5208 //
5209 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5210 #define RANDOMIZED_DOMAIN_POW 29
5211 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5212 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5213 bool Compile::randomized_select(int count) {
5214 assert(count > 0, "only positive");
5215 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5216 }
5217
5218 CloneMap& Compile::clone_map() { return _clone_map; }
5219 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5220
5221 void NodeCloneInfo::dump() const {
5222 tty->print(" {%d:%d} ", idx(), gen());
5223 }
5224
5225 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5226 uint64_t val = value(old->_idx);
5227 NodeCloneInfo cio(val);
5228 assert(val != 0, "old node should be in the map");
5229 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5230 insert(nnn->_idx, cin.get());
5231 #ifndef PRODUCT
5232 if (is_debug()) {
5233 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5234 }
5235 #endif
5236 }
5237
5238 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5239 NodeCloneInfo cio(value(old->_idx));
5240 if (cio.get() == 0) {
5241 cio.set(old->_idx, 0);
5242 insert(old->_idx, cio.get());
5243 #ifndef PRODUCT
5244 if (is_debug()) {
5245 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5246 }
5247 #endif
5248 }
5249 clone(old, nnn, gen);
5250 }
5251
5252 int CloneMap::max_gen() const {
5253 int g = 0;
5254 DictI di(_dict);
5255 for(; di.test(); ++di) {
5256 int t = gen(di._key);
5257 if (g < t) {
5258 g = t;
5259 #ifndef PRODUCT
5260 if (is_debug()) {
5261 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5262 }
5263 #endif
5264 }
5265 }
5266 return g;
5267 }
5268
5269 void CloneMap::dump(node_idx_t key) const {
5270 uint64_t val = value(key);
5271 if (val != 0) {
5272 NodeCloneInfo ni(val);
5273 ni.dump();
5274 }
5275 }