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) alias_type(idx)->set_field(field);
1919 }
1920
1921 // Fill the cache for next time.
1922 if (!uncached) {
1923 ace->_adr_type = adr_type;
1924 ace->_index = idx;
1925 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1926
1927 // Might as well try to fill the cache for the flattened version, too.
1928 AliasCacheEntry* face = probe_alias_cache(flat);
1929 if (face->_adr_type == NULL) {
1930 face->_adr_type = flat;
1931 face->_index = idx;
1932 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1933 }
1934 }
1935
1936 return alias_type(idx);
1937 }
1938
1939
1940 Compile::AliasType* Compile::alias_type(ciField* field) {
1941 const TypeOopPtr* t;
1942 if (field->is_static())
1943 t = TypeInstPtr::make(field->holder()->java_mirror());
1944 else
1945 t = TypeOopPtr::make_from_klass_raw(field->holder());
1946 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1947 if(!((field->is_final() || field->is_stable()) == !atp->is_rewritable())) {
1948 ResourceMark rm;
1949 tty->print_cr("Problematic field: %s %s.%s",
1950 field->signature()->as_utf8(),
1951 field->holder()->name()->as_utf8(),
1952 field->name()->as_utf8());
1953 tty->print_cr("is_final = %d is_stable = %d is_rewritable = %d",
1954 field->is_final(), field->is_stable(), atp->is_rewritable());
1955 }
1956 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1957 return atp;
1958 }
1959
1960
1961 //------------------------------have_alias_type--------------------------------
1962 bool Compile::have_alias_type(const TypePtr* adr_type) {
1963 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1964 if (ace->_adr_type == adr_type) {
1965 return true;
1966 }
1967
1968 // Handle special cases.
1969 if (adr_type == NULL) return true;
1970 if (adr_type == TypePtr::BOTTOM) return true;
1971
1972 return find_alias_type(adr_type, true, NULL) != NULL;
1973 }
1974
1975 //-----------------------------must_alias--------------------------------------
1976 // True if all values of the given address type are in the given alias category.
1977 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1978 if (alias_idx == AliasIdxBot) return true; // the universal category
1979 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1980 if (alias_idx == AliasIdxTop) return false; // the empty category
1981 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1982
1983 // the only remaining possible overlap is identity
1984 int adr_idx = get_alias_index(adr_type);
1985 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1986 assert(adr_idx == alias_idx ||
1987 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1988 && adr_type != TypeOopPtr::BOTTOM),
1989 "should not be testing for overlap with an unsafe pointer");
1990 return adr_idx == alias_idx;
1991 }
1992
1993 //------------------------------can_alias--------------------------------------
1994 // True if any values of the given address type are in the given alias category.
1995 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1996 if (alias_idx == AliasIdxTop) return false; // the empty category
1997 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1998 if (alias_idx == AliasIdxBot) return true; // the universal category
1999 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
2000
2001 // the only remaining possible overlap is identity
2002 int adr_idx = get_alias_index(adr_type);
2003 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
2004 return adr_idx == alias_idx;
2005 }
2006
2007
2008
2009 //---------------------------pop_warm_call-------------------------------------
2010 WarmCallInfo* Compile::pop_warm_call() {
2011 WarmCallInfo* wci = _warm_calls;
2012 if (wci != NULL) _warm_calls = wci->remove_from(wci);
2013 return wci;
2014 }
2015
2016 //----------------------------Inline_Warm--------------------------------------
2017 int Compile::Inline_Warm() {
2018 // If there is room, try to inline some more warm call sites.
2019 // %%% Do a graph index compaction pass when we think we're out of space?
2020 if (!InlineWarmCalls) return 0;
2021
2022 int calls_made_hot = 0;
2023 int room_to_grow = NodeCountInliningCutoff - unique();
2024 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
2025 int amount_grown = 0;
2026 WarmCallInfo* call;
2027 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
2028 int est_size = (int)call->size();
2029 if (est_size > (room_to_grow - amount_grown)) {
2030 // This one won't fit anyway. Get rid of it.
2031 call->make_cold();
2032 continue;
2033 }
2034 call->make_hot();
2035 calls_made_hot++;
2036 amount_grown += est_size;
2037 amount_to_grow -= est_size;
2038 }
2039
2040 if (calls_made_hot > 0) set_major_progress();
2041 return calls_made_hot;
2042 }
2043
2044
2045 //----------------------------Finish_Warm--------------------------------------
2046 void Compile::Finish_Warm() {
2047 if (!InlineWarmCalls) return;
2048 if (failing()) return;
2049 if (warm_calls() == NULL) return;
2050
2051 // Clean up loose ends, if we are out of space for inlining.
2052 WarmCallInfo* call;
2053 while ((call = pop_warm_call()) != NULL) {
2054 call->make_cold();
2055 }
2056 }
2057
2058 //---------------------cleanup_loop_predicates-----------------------
2059 // Remove the opaque nodes that protect the predicates so that all unused
2060 // checks and uncommon_traps will be eliminated from the ideal graph
2061 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
2062 if (predicate_count()==0) return;
2063 for (int i = predicate_count(); i > 0; i--) {
2064 Node * n = predicate_opaque1_node(i-1);
2065 assert(n->Opcode() == Op_Opaque1, "must be");
2066 igvn.replace_node(n, n->in(1));
2067 }
2068 assert(predicate_count()==0, "should be clean!");
2069 }
2070
2071 void Compile::add_range_check_cast(Node* n) {
2072 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2073 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
2074 _range_check_casts->append(n);
2075 }
2076
2077 // Remove all range check dependent CastIINodes.
2078 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
2079 for (int i = range_check_cast_count(); i > 0; i--) {
2080 Node* cast = range_check_cast_node(i-1);
2081 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2082 igvn.replace_node(cast, cast->in(1));
2083 }
2084 assert(range_check_cast_count() == 0, "should be empty");
2085 }
2086
2087 void Compile::add_opaque4_node(Node* n) {
2088 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2089 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2090 _opaque4_nodes->append(n);
2091 }
2092
2093 // Remove all Opaque4 nodes.
2094 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2095 for (int i = opaque4_count(); i > 0; i--) {
2096 Node* opaq = opaque4_node(i-1);
2097 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2098 igvn.replace_node(opaq, opaq->in(2));
2099 }
2100 assert(opaque4_count() == 0, "should be empty");
2101 }
2102
2103 void Compile::add_value_type(Node* n) {
2104 assert(n->is_ValueTypeBase(), "unexpected node");
2105 if (_value_type_nodes != NULL) {
2106 _value_type_nodes->push(n);
2107 }
2108 }
2109
2110 void Compile::remove_value_type(Node* n) {
2111 assert(n->is_ValueTypeBase(), "unexpected node");
2112 if (_value_type_nodes != NULL) {
2113 _value_type_nodes->remove(n);
2114 }
2115 }
2116
2117 // Does the return value keep otherwise useless value type allocations
2118 // alive?
2119 static bool return_val_keeps_allocations_alive(Node* ret_val) {
2120 ResourceMark rm;
2121 Unique_Node_List wq;
2122 wq.push(ret_val);
2123 bool some_allocations = false;
2124 for (uint i = 0; i < wq.size(); i++) {
2125 Node* n = wq.at(i);
2126 assert(!n->is_ValueTypeBase(), "chain of value type nodes");
2127 if (n->outcnt() > 1) {
2128 // Some other use for the allocation
2129 return false;
2130 } else if (n->is_Phi()) {
2131 for (uint j = 1; j < n->req(); j++) {
2132 wq.push(n->in(j));
2133 }
2134 } else if (n->is_CheckCastPP() &&
2135 n->in(1)->is_Proj() &&
2136 n->in(1)->in(0)->is_Allocate()) {
2137 some_allocations = true;
2138 }
2139 }
2140 return some_allocations;
2141 }
2142
2143 void Compile::process_value_types(PhaseIterGVN &igvn) {
2144 // Make value types scalar in safepoints
2145 while (_value_type_nodes->size() != 0) {
2146 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase();
2147 vt->make_scalar_in_safepoints(&igvn);
2148 if (vt->is_ValueTypePtr()) {
2149 igvn.replace_node(vt, vt->get_oop());
2150 } else if (vt->outcnt() == 0) {
2151 igvn.remove_dead_node(vt);
2152 }
2153 }
2154 _value_type_nodes = NULL;
2155 if (tf()->returns_value_type_as_fields()) {
2156 Node* ret = NULL;
2157 for (uint i = 1; i < root()->req(); i++){
2158 Node* in = root()->in(i);
2159 if (in->Opcode() == Op_Return) {
2160 assert(ret == NULL, "only one return");
2161 ret = in;
2162 }
2163 }
2164 if (ret != NULL) {
2165 Node* ret_val = ret->in(TypeFunc::Parms);
2166 if (igvn.type(ret_val)->isa_oopptr() &&
2167 return_val_keeps_allocations_alive(ret_val)) {
2168 igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn));
2169 assert(ret_val->outcnt() == 0, "should be dead now");
2170 igvn.remove_dead_node(ret_val);
2171 }
2172 }
2173 }
2174 igvn.optimize();
2175 }
2176
2177 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2178 if (!_has_flattened_accesses) {
2179 return;
2180 }
2181 // Initially, all flattened array accesses share the same slice to
2182 // keep dependencies with Object[] array accesses (that could be
2183 // to a flattened array) correct. We're done with parsing so we
2184 // now know all flattened array accesses in this compile
2185 // unit. Let's move flattened array accesses to their own slice,
2186 // one per element field. This should help memory access
2187 // optimizations.
2188 ResourceMark rm;
2189 Unique_Node_List wq;
2190 wq.push(root());
2191
2192 Node_List mergememnodes;
2193 Node_List memnodes;
2194
2195 // Alias index currently shared by all flattened memory accesses
2196 int index = get_alias_index(TypeAryPtr::VALUES);
2197
2198 // Find MergeMem nodes and flattened array accesses
2199 for (uint i = 0; i < wq.size(); i++) {
2200 Node* n = wq.at(i);
2201 if (n->is_Mem()) {
2202 const TypePtr* adr_type = NULL;
2203 if (n->Opcode() == Op_StoreCM) {
2204 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type()));
2205 } else {
2206 adr_type = get_adr_type(get_alias_index(n->adr_type()));
2207 }
2208 if (adr_type == TypeAryPtr::VALUES) {
2209 memnodes.push(n);
2210 }
2211 } else if (n->is_MergeMem()) {
2212 MergeMemNode* mm = n->as_MergeMem();
2213 if (mm->memory_at(index) != mm->base_memory()) {
2214 mergememnodes.push(n);
2215 }
2216 }
2217 for (uint j = 0; j < n->req(); j++) {
2218 Node* m = n->in(j);
2219 if (m != NULL) {
2220 wq.push(m);
2221 }
2222 }
2223 }
2224
2225 if (memnodes.size() > 0) {
2226 _flattened_accesses_share_alias = false;
2227
2228 // We are going to change the slice for the flattened array
2229 // accesses so we need to clear the cache entries that refer to
2230 // them.
2231 for (uint i = 0; i < AliasCacheSize; i++) {
2232 AliasCacheEntry* ace = &_alias_cache[i];
2233 if (ace->_adr_type != NULL &&
2234 ace->_adr_type->isa_aryptr() &&
2235 ace->_adr_type->is_aryptr()->elem()->isa_valuetype()) {
2236 ace->_adr_type = NULL;
2237 ace->_index = 0;
2238 }
2239 }
2240
2241 // Find what aliases we are going to add
2242 int start_alias = num_alias_types()-1;
2243 int stop_alias = 0;
2244
2245 for (uint i = 0; i < memnodes.size(); i++) {
2246 Node* m = memnodes.at(i);
2247 const TypePtr* adr_type = NULL;
2248 if (m->Opcode() == Op_StoreCM) {
2249 adr_type = m->in(MemNode::OopStore)->adr_type();
2250 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address),
2251 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore),
2252 get_alias_index(adr_type));
2253 igvn.register_new_node_with_optimizer(clone);
2254 igvn.replace_node(m, clone);
2255 } else {
2256 adr_type = m->adr_type();
2257 #ifdef ASSERT
2258 m->as_Mem()->set_adr_type(adr_type);
2259 #endif
2260 }
2261 int idx = get_alias_index(adr_type);
2262 start_alias = MIN2(start_alias, idx);
2263 stop_alias = MAX2(stop_alias, idx);
2264 }
2265
2266 assert(stop_alias >= start_alias, "should have expanded aliases");
2267
2268 Node_Stack stack(0);
2269 #ifdef ASSERT
2270 VectorSet seen(Thread::current()->resource_area());
2271 #endif
2272 // Now let's fix the memory graph so each flattened array access
2273 // is moved to the right slice. Start from the MergeMem nodes.
2274 uint last = unique();
2275 for (uint i = 0; i < mergememnodes.size(); i++) {
2276 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2277 Node* n = current->memory_at(index);
2278 MergeMemNode* mm = NULL;
2279 do {
2280 // Follow memory edges through memory accesses, phis and
2281 // narrow membars and push nodes on the stack. Once we hit
2282 // bottom memory, we pop element off the stack one at a
2283 // time, in reverse order, and move them to the right slice
2284 // by changing their memory edges.
2285 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::VALUES) {
2286 assert(!seen.test_set(n->_idx), "");
2287 // Uses (a load for instance) will need to be moved to the
2288 // right slice as well and will get a new memory state
2289 // that we don't know yet. The use could also be the
2290 // backedge of a loop. We put a place holder node between
2291 // the memory node and its uses. We replace that place
2292 // holder with the correct memory state once we know it,
2293 // i.e. when nodes are popped off the stack. Using the
2294 // place holder make the logic work in the presence of
2295 // loops.
2296 if (n->outcnt() > 1) {
2297 Node* place_holder = NULL;
2298 assert(!n->has_out_with(Op_Node), "");
2299 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2300 Node* u = n->out(k);
2301 if (u != current && u->_idx < last) {
2302 bool success = false;
2303 for (uint l = 0; l < u->req(); l++) {
2304 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2305 continue;
2306 }
2307 Node* in = u->in(l);
2308 if (in == n) {
2309 if (place_holder == NULL) {
2310 place_holder = new Node(1);
2311 place_holder->init_req(0, n);
2312 }
2313 igvn.replace_input_of(u, l, place_holder);
2314 success = true;
2315 }
2316 }
2317 if (success) {
2318 --k;
2319 }
2320 }
2321 }
2322 }
2323 if (n->is_Phi()) {
2324 stack.push(n, 1);
2325 n = n->in(1);
2326 } else if (n->is_Mem()) {
2327 stack.push(n, n->req());
2328 n = n->in(MemNode::Memory);
2329 } else {
2330 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2331 stack.push(n, n->req());
2332 n = n->in(0)->in(TypeFunc::Memory);
2333 }
2334 } else {
2335 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2336 // Build a new MergeMem node to carry the new memory state
2337 // as we build it. IGVN should fold extraneous MergeMem
2338 // nodes.
2339 mm = MergeMemNode::make(n);
2340 igvn.register_new_node_with_optimizer(mm);
2341 while (stack.size() > 0) {
2342 Node* m = stack.node();
2343 uint idx = stack.index();
2344 if (m->is_Mem()) {
2345 // Move memory node to its new slice
2346 const TypePtr* adr_type = m->adr_type();
2347 int alias = get_alias_index(adr_type);
2348 Node* prev = mm->memory_at(alias);
2349 igvn.replace_input_of(m, MemNode::Memory, prev);
2350 mm->set_memory_at(alias, m);
2351 } else if (m->is_Phi()) {
2352 // We need as many new phis as there are new aliases
2353 igvn.replace_input_of(m, idx, mm);
2354 if (idx == m->req()-1) {
2355 Node* r = m->in(0);
2356 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2357 const Type* adr_type = get_adr_type(j);
2358 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2359 continue;
2360 }
2361 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2362 igvn.register_new_node_with_optimizer(phi);
2363 for (uint k = 1; k < m->req(); k++) {
2364 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2365 }
2366 mm->set_memory_at(j, phi);
2367 }
2368 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2369 igvn.register_new_node_with_optimizer(base_phi);
2370 for (uint k = 1; k < m->req(); k++) {
2371 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2372 }
2373 mm->set_base_memory(base_phi);
2374 }
2375 } else {
2376 // This is a MemBarCPUOrder node from
2377 // Parse::array_load()/Parse::array_store(), in the
2378 // branch that handles flattened arrays hidden under
2379 // an Object[] array. We also need one new membar per
2380 // new alias to keep the unknown access that the
2381 // membars protect properly ordered with accesses to
2382 // known flattened array.
2383 assert(m->is_Proj(), "projection expected");
2384 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2385 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2386 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2387 const Type* adr_type = get_adr_type(j);
2388 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2389 continue;
2390 }
2391 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2392 igvn.register_new_node_with_optimizer(mb);
2393 Node* mem = mm->memory_at(j);
2394 mb->init_req(TypeFunc::Control, ctrl);
2395 mb->init_req(TypeFunc::Memory, mem);
2396 ctrl = new ProjNode(mb, TypeFunc::Control);
2397 igvn.register_new_node_with_optimizer(ctrl);
2398 mem = new ProjNode(mb, TypeFunc::Memory);
2399 igvn.register_new_node_with_optimizer(mem);
2400 mm->set_memory_at(j, mem);
2401 }
2402 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2403 }
2404 if (idx < m->req()-1) {
2405 idx += 1;
2406 stack.set_index(idx);
2407 n = m->in(idx);
2408 break;
2409 }
2410 // Take care of place holder nodes
2411 if (m->has_out_with(Op_Node)) {
2412 Node* place_holder = m->find_out_with(Op_Node);
2413 if (place_holder != NULL) {
2414 Node* mm_clone = mm->clone();
2415 igvn.register_new_node_with_optimizer(mm_clone);
2416 Node* hook = new Node(1);
2417 hook->init_req(0, mm);
2418 igvn.replace_node(place_holder, mm_clone);
2419 hook->destruct();
2420 }
2421 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2422 }
2423 stack.pop();
2424 }
2425 }
2426 } while(stack.size() > 0);
2427 // Fix the memory state at the MergeMem we started from
2428 igvn.rehash_node_delayed(current);
2429 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2430 const Type* adr_type = get_adr_type(j);
2431 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2432 continue;
2433 }
2434 current->set_memory_at(j, mm);
2435 }
2436 current->set_memory_at(index, current->base_memory());
2437 }
2438 igvn.optimize();
2439 }
2440 print_method(PHASE_SPLIT_VALUES_ARRAY, 2);
2441 }
2442
2443
2444 // StringOpts and late inlining of string methods
2445 void Compile::inline_string_calls(bool parse_time) {
2446 {
2447 // remove useless nodes to make the usage analysis simpler
2448 ResourceMark rm;
2449 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2450 }
2451
2452 {
2453 ResourceMark rm;
2454 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2455 PhaseStringOpts pso(initial_gvn(), for_igvn());
2456 print_method(PHASE_AFTER_STRINGOPTS, 3);
2457 }
2458
2459 // now inline anything that we skipped the first time around
2460 if (!parse_time) {
2461 _late_inlines_pos = _late_inlines.length();
2462 }
2463
2464 while (_string_late_inlines.length() > 0) {
2465 CallGenerator* cg = _string_late_inlines.pop();
2466 cg->do_late_inline();
2467 if (failing()) return;
2468 }
2469 _string_late_inlines.trunc_to(0);
2470 }
2471
2472 // Late inlining of boxing methods
2473 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2474 if (_boxing_late_inlines.length() > 0) {
2475 assert(has_boxed_value(), "inconsistent");
2476
2477 PhaseGVN* gvn = initial_gvn();
2478 set_inlining_incrementally(true);
2479
2480 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2481 for_igvn()->clear();
2482 gvn->replace_with(&igvn);
2483
2484 _late_inlines_pos = _late_inlines.length();
2485
2486 while (_boxing_late_inlines.length() > 0) {
2487 CallGenerator* cg = _boxing_late_inlines.pop();
2488 cg->do_late_inline();
2489 if (failing()) return;
2490 }
2491 _boxing_late_inlines.trunc_to(0);
2492
2493 inline_incrementally_cleanup(igvn);
2494
2495 set_inlining_incrementally(false);
2496 }
2497 }
2498
2499 bool Compile::inline_incrementally_one() {
2500 assert(IncrementalInline, "incremental inlining should be on");
2501
2502 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2503 set_inlining_progress(false);
2504 set_do_cleanup(false);
2505 int i = 0;
2506 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2507 CallGenerator* cg = _late_inlines.at(i);
2508 _late_inlines_pos = i+1;
2509 cg->do_late_inline();
2510 if (failing()) return false;
2511 }
2512 int j = 0;
2513 for (; i < _late_inlines.length(); i++, j++) {
2514 _late_inlines.at_put(j, _late_inlines.at(i));
2515 }
2516 _late_inlines.trunc_to(j);
2517 assert(inlining_progress() || _late_inlines.length() == 0, "");
2518
2519 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2520
2521 set_inlining_progress(false);
2522 set_do_cleanup(false);
2523 return (_late_inlines.length() > 0) && !needs_cleanup;
2524 }
2525
2526 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2527 {
2528 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2529 ResourceMark rm;
2530 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2531 }
2532 {
2533 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2534 igvn = PhaseIterGVN(initial_gvn());
2535 igvn.optimize();
2536 }
2537 }
2538
2539 // Perform incremental inlining until bound on number of live nodes is reached
2540 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2541 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2542
2543 set_inlining_incrementally(true);
2544 uint low_live_nodes = 0;
2545
2546 while (_late_inlines.length() > 0) {
2547 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2548 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2549 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2550 // PhaseIdealLoop is expensive so we only try it once we are
2551 // out of live nodes and we only try it again if the previous
2552 // helped got the number of nodes down significantly
2553 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2554 if (failing()) return;
2555 low_live_nodes = live_nodes();
2556 _major_progress = true;
2557 }
2558
2559 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2560 break; // finish
2561 }
2562 }
2563
2564 for_igvn()->clear();
2565 initial_gvn()->replace_with(&igvn);
2566
2567 while (inline_incrementally_one()) {
2568 assert(!failing(), "inconsistent");
2569 }
2570
2571 if (failing()) return;
2572
2573 inline_incrementally_cleanup(igvn);
2574
2575 if (failing()) return;
2576 }
2577 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2578
2579 if (_string_late_inlines.length() > 0) {
2580 assert(has_stringbuilder(), "inconsistent");
2581 for_igvn()->clear();
2582 initial_gvn()->replace_with(&igvn);
2583
2584 inline_string_calls(false);
2585
2586 if (failing()) return;
2587
2588 inline_incrementally_cleanup(igvn);
2589 }
2590
2591 set_inlining_incrementally(false);
2592 }
2593
2594
2595 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2596 if(_loop_opts_cnt > 0) {
2597 debug_only( int cnt = 0; );
2598 while(major_progress() && (_loop_opts_cnt > 0)) {
2599 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2600 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2601 PhaseIdealLoop::optimize(igvn, mode);
2602 _loop_opts_cnt--;
2603 if (failing()) return false;
2604 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2605 }
2606 }
2607 return true;
2608 }
2609
2610 // Remove edges from "root" to each SafePoint at a backward branch.
2611 // They were inserted during parsing (see add_safepoint()) to make
2612 // infinite loops without calls or exceptions visible to root, i.e.,
2613 // useful.
2614 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2615 Node *r = root();
2616 if (r != NULL) {
2617 for (uint i = r->req(); i < r->len(); ++i) {
2618 Node *n = r->in(i);
2619 if (n != NULL && n->is_SafePoint()) {
2620 r->rm_prec(i);
2621 if (n->outcnt() == 0) {
2622 igvn.remove_dead_node(n);
2623 }
2624 --i;
2625 }
2626 }
2627 }
2628 }
2629
2630 //------------------------------Optimize---------------------------------------
2631 // Given a graph, optimize it.
2632 void Compile::Optimize() {
2633 TracePhase tp("optimizer", &timers[_t_optimizer]);
2634
2635 #ifndef PRODUCT
2636 if (_directive->BreakAtCompileOption) {
2637 BREAKPOINT;
2638 }
2639
2640 #endif
2641
2642 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2643 #ifdef ASSERT
2644 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2645 #endif
2646
2647 ResourceMark rm;
2648
2649 print_inlining_reinit();
2650
2651 NOT_PRODUCT( verify_graph_edges(); )
2652
2653 print_method(PHASE_AFTER_PARSING);
2654
2655 {
2656 // Iterative Global Value Numbering, including ideal transforms
2657 // Initialize IterGVN with types and values from parse-time GVN
2658 PhaseIterGVN igvn(initial_gvn());
2659 #ifdef ASSERT
2660 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2661 #endif
2662 {
2663 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2664 igvn.optimize();
2665 }
2666
2667 if (failing()) return;
2668
2669 print_method(PHASE_ITER_GVN1, 2);
2670
2671 inline_incrementally(igvn);
2672
2673 print_method(PHASE_INCREMENTAL_INLINE, 2);
2674
2675 if (failing()) return;
2676
2677 if (eliminate_boxing()) {
2678 // Inline valueOf() methods now.
2679 inline_boxing_calls(igvn);
2680
2681 if (AlwaysIncrementalInline) {
2682 inline_incrementally(igvn);
2683 }
2684
2685 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2686
2687 if (failing()) return;
2688 }
2689
2690 // Now that all inlining is over, cut edge from root to loop
2691 // safepoints
2692 remove_root_to_sfpts_edges(igvn);
2693
2694 // Remove the speculative part of types and clean up the graph from
2695 // the extra CastPP nodes whose only purpose is to carry them. Do
2696 // that early so that optimizations are not disrupted by the extra
2697 // CastPP nodes.
2698 remove_speculative_types(igvn);
2699
2700 // No more new expensive nodes will be added to the list from here
2701 // so keep only the actual candidates for optimizations.
2702 cleanup_expensive_nodes(igvn);
2703
2704 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2705 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2706 initial_gvn()->replace_with(&igvn);
2707 for_igvn()->clear();
2708 Unique_Node_List new_worklist(C->comp_arena());
2709 {
2710 ResourceMark rm;
2711 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2712 }
2713 set_for_igvn(&new_worklist);
2714 igvn = PhaseIterGVN(initial_gvn());
2715 igvn.optimize();
2716 }
2717
2718 if (_value_type_nodes->size() > 0) {
2719 // Do this once all inlining is over to avoid getting inconsistent debug info
2720 process_value_types(igvn);
2721 }
2722
2723 adjust_flattened_array_access_aliases(igvn);
2724
2725 // Perform escape analysis
2726 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2727 if (has_loops()) {
2728 // Cleanup graph (remove dead nodes).
2729 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2730 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2731 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2732 if (failing()) return;
2733 }
2734 ConnectionGraph::do_analysis(this, &igvn);
2735
2736 if (failing()) return;
2737
2738 // Optimize out fields loads from scalar replaceable allocations.
2739 igvn.optimize();
2740 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2741
2742 if (failing()) return;
2743
2744 if (congraph() != NULL && macro_count() > 0) {
2745 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2746 PhaseMacroExpand mexp(igvn);
2747 mexp.eliminate_macro_nodes();
2748 igvn.set_delay_transform(false);
2749
2750 igvn.optimize();
2751 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2752
2753 if (failing()) return;
2754 }
2755 }
2756
2757 // Loop transforms on the ideal graph. Range Check Elimination,
2758 // peeling, unrolling, etc.
2759
2760 // Set loop opts counter
2761 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2762 {
2763 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2764 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2765 _loop_opts_cnt--;
2766 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2767 if (failing()) return;
2768 }
2769 // Loop opts pass if partial peeling occurred in previous pass
2770 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2771 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2772 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2773 _loop_opts_cnt--;
2774 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2775 if (failing()) return;
2776 }
2777 // Loop opts pass for loop-unrolling before CCP
2778 if(major_progress() && (_loop_opts_cnt > 0)) {
2779 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2780 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2781 _loop_opts_cnt--;
2782 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2783 }
2784 if (!failing()) {
2785 // Verify that last round of loop opts produced a valid graph
2786 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2787 PhaseIdealLoop::verify(igvn);
2788 }
2789 }
2790 if (failing()) return;
2791
2792 // Conditional Constant Propagation;
2793 PhaseCCP ccp( &igvn );
2794 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2795 {
2796 TracePhase tp("ccp", &timers[_t_ccp]);
2797 ccp.do_transform();
2798 }
2799 print_method(PHASE_CPP1, 2);
2800
2801 assert( true, "Break here to ccp.dump_old2new_map()");
2802
2803 // Iterative Global Value Numbering, including ideal transforms
2804 {
2805 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2806 igvn = ccp;
2807 igvn.optimize();
2808 }
2809 print_method(PHASE_ITER_GVN2, 2);
2810
2811 if (failing()) return;
2812
2813 // Loop transforms on the ideal graph. Range Check Elimination,
2814 // peeling, unrolling, etc.
2815 if (!optimize_loops(igvn, LoopOptsDefault)) {
2816 return;
2817 }
2818
2819 if (failing()) return;
2820
2821 // Ensure that major progress is now clear
2822 C->clear_major_progress();
2823
2824 {
2825 // Verify that all previous optimizations produced a valid graph
2826 // at least to this point, even if no loop optimizations were done.
2827 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2828 PhaseIdealLoop::verify(igvn);
2829 }
2830
2831 if (range_check_cast_count() > 0) {
2832 // No more loop optimizations. Remove all range check dependent CastIINodes.
2833 C->remove_range_check_casts(igvn);
2834 igvn.optimize();
2835 }
2836
2837 #ifdef ASSERT
2838 bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion);
2839 #endif
2840
2841 bs->barrier_insertion_phase(C, igvn);
2842 if (failing()) return;
2843
2844 #ifdef ASSERT
2845 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2846 #endif
2847
2848 {
2849 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2850 PhaseMacroExpand mex(igvn);
2851 if (mex.expand_macro_nodes()) {
2852 assert(failing(), "must bail out w/ explicit message");
2853 return;
2854 }
2855 print_method(PHASE_MACRO_EXPANSION, 2);
2856 }
2857
2858 {
2859 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2860 if (bs->expand_barriers(this, igvn)) {
2861 assert(failing(), "must bail out w/ explicit message");
2862 return;
2863 }
2864 print_method(PHASE_BARRIER_EXPANSION, 2);
2865 }
2866
2867 if (opaque4_count() > 0) {
2868 C->remove_opaque4_nodes(igvn);
2869 igvn.optimize();
2870 }
2871
2872 DEBUG_ONLY( _modified_nodes = NULL; )
2873 } // (End scope of igvn; run destructor if necessary for asserts.)
2874
2875 process_print_inlining();
2876 // A method with only infinite loops has no edges entering loops from root
2877 {
2878 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2879 if (final_graph_reshaping()) {
2880 assert(failing(), "must bail out w/ explicit message");
2881 return;
2882 }
2883 }
2884
2885 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2886 }
2887
2888 //------------------------------Code_Gen---------------------------------------
2889 // Given a graph, generate code for it
2890 void Compile::Code_Gen() {
2891 if (failing()) {
2892 return;
2893 }
2894
2895 // Perform instruction selection. You might think we could reclaim Matcher
2896 // memory PDQ, but actually the Matcher is used in generating spill code.
2897 // Internals of the Matcher (including some VectorSets) must remain live
2898 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2899 // set a bit in reclaimed memory.
2900
2901 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2902 // nodes. Mapping is only valid at the root of each matched subtree.
2903 NOT_PRODUCT( verify_graph_edges(); )
2904
2905 Matcher matcher;
2906 _matcher = &matcher;
2907 {
2908 TracePhase tp("matcher", &timers[_t_matcher]);
2909 matcher.match();
2910 }
2911 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2912 // nodes. Mapping is only valid at the root of each matched subtree.
2913 NOT_PRODUCT( verify_graph_edges(); )
2914
2915 // If you have too many nodes, or if matching has failed, bail out
2916 check_node_count(0, "out of nodes matching instructions");
2917 if (failing()) {
2918 return;
2919 }
2920
2921 print_method(PHASE_MATCHING, 2);
2922
2923 // Build a proper-looking CFG
2924 PhaseCFG cfg(node_arena(), root(), matcher);
2925 _cfg = &cfg;
2926 {
2927 TracePhase tp("scheduler", &timers[_t_scheduler]);
2928 bool success = cfg.do_global_code_motion();
2929 if (!success) {
2930 return;
2931 }
2932
2933 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2934 NOT_PRODUCT( verify_graph_edges(); )
2935 debug_only( cfg.verify(); )
2936 }
2937
2938 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2939 _regalloc = ®alloc;
2940 {
2941 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2942 // Perform register allocation. After Chaitin, use-def chains are
2943 // no longer accurate (at spill code) and so must be ignored.
2944 // Node->LRG->reg mappings are still accurate.
2945 _regalloc->Register_Allocate();
2946
2947 // Bail out if the allocator builds too many nodes
2948 if (failing()) {
2949 return;
2950 }
2951 }
2952
2953 // Prior to register allocation we kept empty basic blocks in case the
2954 // the allocator needed a place to spill. After register allocation we
2955 // are not adding any new instructions. If any basic block is empty, we
2956 // can now safely remove it.
2957 {
2958 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2959 cfg.remove_empty_blocks();
2960 if (do_freq_based_layout()) {
2961 PhaseBlockLayout layout(cfg);
2962 } else {
2963 cfg.set_loop_alignment();
2964 }
2965 cfg.fixup_flow();
2966 }
2967
2968 // Apply peephole optimizations
2969 if( OptoPeephole ) {
2970 TracePhase tp("peephole", &timers[_t_peephole]);
2971 PhasePeephole peep( _regalloc, cfg);
2972 peep.do_transform();
2973 }
2974
2975 // Do late expand if CPU requires this.
2976 if (Matcher::require_postalloc_expand) {
2977 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2978 cfg.postalloc_expand(_regalloc);
2979 }
2980
2981 // Convert Nodes to instruction bits in a buffer
2982 {
2983 TraceTime tp("output", &timers[_t_output], CITime);
2984 Output();
2985 }
2986
2987 print_method(PHASE_FINAL_CODE);
2988
2989 // He's dead, Jim.
2990 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2991 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2992 }
2993
2994
2995 //------------------------------dump_asm---------------------------------------
2996 // Dump formatted assembly
2997 #if defined(SUPPORT_OPTO_ASSEMBLY)
2998 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2999
3000 int pc_digits = 3; // #chars required for pc
3001 int sb_chars = 3; // #chars for "start bundle" indicator
3002 int tab_size = 8;
3003 if (pcs != NULL) {
3004 int max_pc = 0;
3005 for (uint i = 0; i < pc_limit; i++) {
3006 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
3007 }
3008 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
3009 }
3010 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
3011
3012 bool cut_short = false;
3013 st->print_cr("#");
3014 st->print("# "); _tf->dump_on(st); st->cr();
3015 st->print_cr("#");
3016
3017 // For all blocks
3018 int pc = 0x0; // Program counter
3019 char starts_bundle = ' ';
3020 _regalloc->dump_frame();
3021
3022 Node *n = NULL;
3023 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
3024 if (VMThread::should_terminate()) {
3025 cut_short = true;
3026 break;
3027 }
3028 Block* block = _cfg->get_block(i);
3029 if (block->is_connector() && !Verbose) {
3030 continue;
3031 }
3032 n = block->head();
3033 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3034 pc = pcs[n->_idx];
3035 st->print("%*.*x", pc_digits, pc_digits, pc);
3036 }
3037 st->fill_to(prefix_len);
3038 block->dump_head(_cfg, st);
3039 if (block->is_connector()) {
3040 st->fill_to(prefix_len);
3041 st->print_cr("# Empty connector block");
3042 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3043 st->fill_to(prefix_len);
3044 st->print_cr("# Block is sole successor of call");
3045 }
3046
3047 // For all instructions
3048 Node *delay = NULL;
3049 for (uint j = 0; j < block->number_of_nodes(); j++) {
3050 if (VMThread::should_terminate()) {
3051 cut_short = true;
3052 break;
3053 }
3054 n = block->get_node(j);
3055 if (valid_bundle_info(n)) {
3056 Bundle* bundle = node_bundling(n);
3057 if (bundle->used_in_unconditional_delay()) {
3058 delay = n;
3059 continue;
3060 }
3061 if (bundle->starts_bundle()) {
3062 starts_bundle = '+';
3063 }
3064 }
3065
3066 if (WizardMode) {
3067 n->dump();
3068 }
3069
3070 if( !n->is_Region() && // Dont print in the Assembly
3071 !n->is_Phi() && // a few noisely useless nodes
3072 !n->is_Proj() &&
3073 !n->is_MachTemp() &&
3074 !n->is_SafePointScalarObject() &&
3075 !n->is_Catch() && // Would be nice to print exception table targets
3076 !n->is_MergeMem() && // Not very interesting
3077 !n->is_top() && // Debug info table constants
3078 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3079 ) {
3080 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3081 pc = pcs[n->_idx];
3082 st->print("%*.*x", pc_digits, pc_digits, pc);
3083 } else {
3084 st->fill_to(pc_digits);
3085 }
3086 st->print(" %c ", starts_bundle);
3087 starts_bundle = ' ';
3088 st->fill_to(prefix_len);
3089 n->format(_regalloc, st);
3090 st->cr();
3091 }
3092
3093 // If we have an instruction with a delay slot, and have seen a delay,
3094 // then back up and print it
3095 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3096 // Coverity finding - Explicit null dereferenced.
3097 guarantee(delay != NULL, "no unconditional delay instruction");
3098 if (WizardMode) delay->dump();
3099
3100 if (node_bundling(delay)->starts_bundle())
3101 starts_bundle = '+';
3102 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3103 pc = pcs[n->_idx];
3104 st->print("%*.*x", pc_digits, pc_digits, pc);
3105 } else {
3106 st->fill_to(pc_digits);
3107 }
3108 st->print(" %c ", starts_bundle);
3109 starts_bundle = ' ';
3110 st->fill_to(prefix_len);
3111 delay->format(_regalloc, st);
3112 st->cr();
3113 delay = NULL;
3114 }
3115
3116 // Dump the exception table as well
3117 if( n->is_Catch() && (Verbose || WizardMode) ) {
3118 // Print the exception table for this offset
3119 _handler_table.print_subtable_for(pc);
3120 }
3121 st->bol(); // Make sure we start on a new line
3122 }
3123 st->cr(); // one empty line between blocks
3124 assert(cut_short || delay == NULL, "no unconditional delay branch");
3125 } // End of per-block dump
3126
3127 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3128 }
3129 #endif
3130
3131 //------------------------------Final_Reshape_Counts---------------------------
3132 // This class defines counters to help identify when a method
3133 // may/must be executed using hardware with only 24-bit precision.
3134 struct Final_Reshape_Counts : public StackObj {
3135 int _call_count; // count non-inlined 'common' calls
3136 int _float_count; // count float ops requiring 24-bit precision
3137 int _double_count; // count double ops requiring more precision
3138 int _java_call_count; // count non-inlined 'java' calls
3139 int _inner_loop_count; // count loops which need alignment
3140 VectorSet _visited; // Visitation flags
3141 Node_List _tests; // Set of IfNodes & PCTableNodes
3142
3143 Final_Reshape_Counts() :
3144 _call_count(0), _float_count(0), _double_count(0),
3145 _java_call_count(0), _inner_loop_count(0),
3146 _visited( Thread::current()->resource_area() ) { }
3147
3148 void inc_call_count () { _call_count ++; }
3149 void inc_float_count () { _float_count ++; }
3150 void inc_double_count() { _double_count++; }
3151 void inc_java_call_count() { _java_call_count++; }
3152 void inc_inner_loop_count() { _inner_loop_count++; }
3153
3154 int get_call_count () const { return _call_count ; }
3155 int get_float_count () const { return _float_count ; }
3156 int get_double_count() const { return _double_count; }
3157 int get_java_call_count() const { return _java_call_count; }
3158 int get_inner_loop_count() const { return _inner_loop_count; }
3159 };
3160
3161 #ifdef ASSERT
3162 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
3163 ciInstanceKlass *k = tp->klass()->as_instance_klass();
3164 // Make sure the offset goes inside the instance layout.
3165 return k->contains_field_offset(tp->offset());
3166 // Note that OffsetBot and OffsetTop are very negative.
3167 }
3168 #endif
3169
3170 // Eliminate trivially redundant StoreCMs and accumulate their
3171 // precedence edges.
3172 void Compile::eliminate_redundant_card_marks(Node* n) {
3173 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3174 if (n->in(MemNode::Address)->outcnt() > 1) {
3175 // There are multiple users of the same address so it might be
3176 // possible to eliminate some of the StoreCMs
3177 Node* mem = n->in(MemNode::Memory);
3178 Node* adr = n->in(MemNode::Address);
3179 Node* val = n->in(MemNode::ValueIn);
3180 Node* prev = n;
3181 bool done = false;
3182 // Walk the chain of StoreCMs eliminating ones that match. As
3183 // long as it's a chain of single users then the optimization is
3184 // safe. Eliminating partially redundant StoreCMs would require
3185 // cloning copies down the other paths.
3186 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3187 if (adr == mem->in(MemNode::Address) &&
3188 val == mem->in(MemNode::ValueIn)) {
3189 // redundant StoreCM
3190 if (mem->req() > MemNode::OopStore) {
3191 // Hasn't been processed by this code yet.
3192 n->add_prec(mem->in(MemNode::OopStore));
3193 } else {
3194 // Already converted to precedence edge
3195 for (uint i = mem->req(); i < mem->len(); i++) {
3196 // Accumulate any precedence edges
3197 if (mem->in(i) != NULL) {
3198 n->add_prec(mem->in(i));
3199 }
3200 }
3201 // Everything above this point has been processed.
3202 done = true;
3203 }
3204 // Eliminate the previous StoreCM
3205 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3206 assert(mem->outcnt() == 0, "should be dead");
3207 mem->disconnect_inputs(NULL, this);
3208 } else {
3209 prev = mem;
3210 }
3211 mem = prev->in(MemNode::Memory);
3212 }
3213 }
3214 }
3215
3216
3217 //------------------------------final_graph_reshaping_impl----------------------
3218 // Implement items 1-5 from final_graph_reshaping below.
3219 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3220
3221 if ( n->outcnt() == 0 ) return; // dead node
3222 uint nop = n->Opcode();
3223
3224 // Check for 2-input instruction with "last use" on right input.
3225 // Swap to left input. Implements item (2).
3226 if( n->req() == 3 && // two-input instruction
3227 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3228 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3229 n->in(2)->outcnt() == 1 &&// right use IS a last use
3230 !n->in(2)->is_Con() ) { // right use is not a constant
3231 // Check for commutative opcode
3232 switch( nop ) {
3233 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3234 case Op_MaxI: case Op_MinI:
3235 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3236 case Op_AndL: case Op_XorL: case Op_OrL:
3237 case Op_AndI: case Op_XorI: case Op_OrI: {
3238 // Move "last use" input to left by swapping inputs
3239 n->swap_edges(1, 2);
3240 break;
3241 }
3242 default:
3243 break;
3244 }
3245 }
3246
3247 #ifdef ASSERT
3248 if( n->is_Mem() ) {
3249 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3250 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3251 // oop will be recorded in oop map if load crosses safepoint
3252 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3253 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3254 "raw memory operations should have control edge");
3255 }
3256 if (n->is_MemBar()) {
3257 MemBarNode* mb = n->as_MemBar();
3258 if (mb->trailing_store() || mb->trailing_load_store()) {
3259 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3260 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3261 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3262 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3263 } else if (mb->leading()) {
3264 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3265 }
3266 }
3267 #endif
3268 // Count FPU ops and common calls, implements item (3)
3269 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3270 if (!gc_handled) {
3271 final_graph_reshaping_main_switch(n, frc, nop);
3272 }
3273
3274 // Collect CFG split points
3275 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3276 frc._tests.push(n);
3277 }
3278 }
3279
3280 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3281 switch( nop ) {
3282 // Count all float operations that may use FPU
3283 case Op_AddF:
3284 case Op_SubF:
3285 case Op_MulF:
3286 case Op_DivF:
3287 case Op_NegF:
3288 case Op_ModF:
3289 case Op_ConvI2F:
3290 case Op_ConF:
3291 case Op_CmpF:
3292 case Op_CmpF3:
3293 // case Op_ConvL2F: // longs are split into 32-bit halves
3294 frc.inc_float_count();
3295 break;
3296
3297 case Op_ConvF2D:
3298 case Op_ConvD2F:
3299 frc.inc_float_count();
3300 frc.inc_double_count();
3301 break;
3302
3303 // Count all double operations that may use FPU
3304 case Op_AddD:
3305 case Op_SubD:
3306 case Op_MulD:
3307 case Op_DivD:
3308 case Op_NegD:
3309 case Op_ModD:
3310 case Op_ConvI2D:
3311 case Op_ConvD2I:
3312 // case Op_ConvL2D: // handled by leaf call
3313 // case Op_ConvD2L: // handled by leaf call
3314 case Op_ConD:
3315 case Op_CmpD:
3316 case Op_CmpD3:
3317 frc.inc_double_count();
3318 break;
3319 case Op_Opaque1: // Remove Opaque Nodes before matching
3320 case Op_Opaque2: // Remove Opaque Nodes before matching
3321 case Op_Opaque3:
3322 n->subsume_by(n->in(1), this);
3323 break;
3324 case Op_CallStaticJava:
3325 case Op_CallJava:
3326 case Op_CallDynamicJava:
3327 frc.inc_java_call_count(); // Count java call site;
3328 case Op_CallRuntime:
3329 case Op_CallLeaf:
3330 case Op_CallLeafNoFP: {
3331 assert (n->is_Call(), "");
3332 CallNode *call = n->as_Call();
3333 // Count call sites where the FP mode bit would have to be flipped.
3334 // Do not count uncommon runtime calls:
3335 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3336 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3337 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3338 frc.inc_call_count(); // Count the call site
3339 } else { // See if uncommon argument is shared
3340 Node *n = call->in(TypeFunc::Parms);
3341 int nop = n->Opcode();
3342 // Clone shared simple arguments to uncommon calls, item (1).
3343 if (n->outcnt() > 1 &&
3344 !n->is_Proj() &&
3345 nop != Op_CreateEx &&
3346 nop != Op_CheckCastPP &&
3347 nop != Op_DecodeN &&
3348 nop != Op_DecodeNKlass &&
3349 !n->is_Mem() &&
3350 !n->is_Phi()) {
3351 Node *x = n->clone();
3352 call->set_req(TypeFunc::Parms, x);
3353 }
3354 }
3355 break;
3356 }
3357
3358 case Op_StoreD:
3359 case Op_LoadD:
3360 case Op_LoadD_unaligned:
3361 frc.inc_double_count();
3362 goto handle_mem;
3363 case Op_StoreF:
3364 case Op_LoadF:
3365 frc.inc_float_count();
3366 goto handle_mem;
3367
3368 case Op_StoreCM:
3369 {
3370 // Convert OopStore dependence into precedence edge
3371 Node* prec = n->in(MemNode::OopStore);
3372 n->del_req(MemNode::OopStore);
3373 n->add_prec(prec);
3374 eliminate_redundant_card_marks(n);
3375 }
3376
3377 // fall through
3378
3379 case Op_StoreB:
3380 case Op_StoreC:
3381 case Op_StorePConditional:
3382 case Op_StoreI:
3383 case Op_StoreL:
3384 case Op_StoreIConditional:
3385 case Op_StoreLConditional:
3386 case Op_CompareAndSwapB:
3387 case Op_CompareAndSwapS:
3388 case Op_CompareAndSwapI:
3389 case Op_CompareAndSwapL:
3390 case Op_CompareAndSwapP:
3391 case Op_CompareAndSwapN:
3392 case Op_WeakCompareAndSwapB:
3393 case Op_WeakCompareAndSwapS:
3394 case Op_WeakCompareAndSwapI:
3395 case Op_WeakCompareAndSwapL:
3396 case Op_WeakCompareAndSwapP:
3397 case Op_WeakCompareAndSwapN:
3398 case Op_CompareAndExchangeB:
3399 case Op_CompareAndExchangeS:
3400 case Op_CompareAndExchangeI:
3401 case Op_CompareAndExchangeL:
3402 case Op_CompareAndExchangeP:
3403 case Op_CompareAndExchangeN:
3404 case Op_GetAndAddS:
3405 case Op_GetAndAddB:
3406 case Op_GetAndAddI:
3407 case Op_GetAndAddL:
3408 case Op_GetAndSetS:
3409 case Op_GetAndSetB:
3410 case Op_GetAndSetI:
3411 case Op_GetAndSetL:
3412 case Op_GetAndSetP:
3413 case Op_GetAndSetN:
3414 case Op_StoreP:
3415 case Op_StoreN:
3416 case Op_StoreNKlass:
3417 case Op_LoadB:
3418 case Op_LoadUB:
3419 case Op_LoadUS:
3420 case Op_LoadI:
3421 case Op_LoadKlass:
3422 case Op_LoadNKlass:
3423 case Op_LoadL:
3424 case Op_LoadL_unaligned:
3425 case Op_LoadPLocked:
3426 case Op_LoadP:
3427 case Op_LoadN:
3428 case Op_LoadRange:
3429 case Op_LoadS: {
3430 handle_mem:
3431 #ifdef ASSERT
3432 if( VerifyOptoOopOffsets ) {
3433 MemNode* mem = n->as_Mem();
3434 // Check to see if address types have grounded out somehow.
3435 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3436 assert( !tp || oop_offset_is_sane(tp), "" );
3437 }
3438 #endif
3439 if (EnableValhalla && (nop == Op_LoadKlass || nop == Op_LoadNKlass)) {
3440 const TypeKlassPtr* tk = n->bottom_type()->make_ptr()->is_klassptr();
3441 assert(!tk->klass_is_exact(), "should have been folded");
3442 ciKlass* klass = tk->klass();
3443 bool maybe_value_array = klass->is_java_lang_Object();
3444 if (!maybe_value_array && klass->is_obj_array_klass()) {
3445 klass = klass->as_array_klass()->element_klass();
3446 maybe_value_array = klass->is_java_lang_Object() || klass->is_interface() || klass->is_valuetype();
3447 }
3448 if (maybe_value_array) {
3449 // Array load klass needs to filter out property bits (but not
3450 // GetNullFreePropertyNode which needs to extract the null free bits)
3451 uint last = unique();
3452 Node* pointer = NULL;
3453 if (nop == Op_LoadKlass) {
3454 Node* cast = new CastP2XNode(NULL, n);
3455 Node* masked = new LShiftXNode(cast, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3456 masked = new RShiftXNode(masked, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3457 pointer = new CastX2PNode(masked);
3458 pointer = new CheckCastPPNode(NULL, pointer, n->bottom_type());
3459 } else {
3460 Node* cast = new CastN2INode(n);
3461 Node* masked = new AndINode(cast, new ConINode(TypeInt::make(oopDesc::compressed_klass_mask())));
3462 pointer = new CastI2NNode(masked, n->bottom_type());
3463 }
3464 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3465 Node* u = n->fast_out(i);
3466 if (u->_idx < last && u->Opcode() != Op_GetNullFreeProperty) {
3467 int nb = u->replace_edge(n, pointer);
3468 --i, imax -= nb;
3469 }
3470 }
3471 }
3472 }
3473 break;
3474 }
3475
3476 case Op_AddP: { // Assert sane base pointers
3477 Node *addp = n->in(AddPNode::Address);
3478 assert( !addp->is_AddP() ||
3479 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3480 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3481 "Base pointers must match (addp %u)", addp->_idx );
3482 #ifdef _LP64
3483 if ((UseCompressedOops || UseCompressedClassPointers) &&
3484 addp->Opcode() == Op_ConP &&
3485 addp == n->in(AddPNode::Base) &&
3486 n->in(AddPNode::Offset)->is_Con()) {
3487 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3488 // on the platform and on the compressed oops mode.
3489 // Use addressing with narrow klass to load with offset on x86.
3490 // Some platforms can use the constant pool to load ConP.
3491 // Do this transformation here since IGVN will convert ConN back to ConP.
3492 const Type* t = addp->bottom_type();
3493 bool is_oop = t->isa_oopptr() != NULL;
3494 bool is_klass = t->isa_klassptr() != NULL;
3495
3496 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3497 (is_klass && Matcher::const_klass_prefer_decode())) {
3498 Node* nn = NULL;
3499
3500 int op = is_oop ? Op_ConN : Op_ConNKlass;
3501
3502 // Look for existing ConN node of the same exact type.
3503 Node* r = root();
3504 uint cnt = r->outcnt();
3505 for (uint i = 0; i < cnt; i++) {
3506 Node* m = r->raw_out(i);
3507 if (m!= NULL && m->Opcode() == op &&
3508 m->bottom_type()->make_ptr() == t) {
3509 nn = m;
3510 break;
3511 }
3512 }
3513 if (nn != NULL) {
3514 // Decode a narrow oop to match address
3515 // [R12 + narrow_oop_reg<<3 + offset]
3516 if (is_oop) {
3517 nn = new DecodeNNode(nn, t);
3518 } else {
3519 nn = new DecodeNKlassNode(nn, t);
3520 }
3521 // Check for succeeding AddP which uses the same Base.
3522 // Otherwise we will run into the assertion above when visiting that guy.
3523 for (uint i = 0; i < n->outcnt(); ++i) {
3524 Node *out_i = n->raw_out(i);
3525 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3526 out_i->set_req(AddPNode::Base, nn);
3527 #ifdef ASSERT
3528 for (uint j = 0; j < out_i->outcnt(); ++j) {
3529 Node *out_j = out_i->raw_out(j);
3530 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3531 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3532 }
3533 #endif
3534 }
3535 }
3536 n->set_req(AddPNode::Base, nn);
3537 n->set_req(AddPNode::Address, nn);
3538 if (addp->outcnt() == 0) {
3539 addp->disconnect_inputs(NULL, this);
3540 }
3541 }
3542 }
3543 }
3544 #endif
3545 // platform dependent reshaping of the address expression
3546 reshape_address(n->as_AddP());
3547 break;
3548 }
3549
3550 case Op_CastPP: {
3551 // Remove CastPP nodes to gain more freedom during scheduling but
3552 // keep the dependency they encode as control or precedence edges
3553 // (if control is set already) on memory operations. Some CastPP
3554 // nodes don't have a control (don't carry a dependency): skip
3555 // those.
3556 if (n->in(0) != NULL) {
3557 ResourceMark rm;
3558 Unique_Node_List wq;
3559 wq.push(n);
3560 for (uint next = 0; next < wq.size(); ++next) {
3561 Node *m = wq.at(next);
3562 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3563 Node* use = m->fast_out(i);
3564 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3565 use->ensure_control_or_add_prec(n->in(0));
3566 } else {
3567 switch(use->Opcode()) {
3568 case Op_AddP:
3569 case Op_DecodeN:
3570 case Op_DecodeNKlass:
3571 case Op_CheckCastPP:
3572 case Op_CastPP:
3573 wq.push(use);
3574 break;
3575 }
3576 }
3577 }
3578 }
3579 }
3580 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3581 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3582 Node* in1 = n->in(1);
3583 const Type* t = n->bottom_type();
3584 Node* new_in1 = in1->clone();
3585 new_in1->as_DecodeN()->set_type(t);
3586
3587 if (!Matcher::narrow_oop_use_complex_address()) {
3588 //
3589 // x86, ARM and friends can handle 2 adds in addressing mode
3590 // and Matcher can fold a DecodeN node into address by using
3591 // a narrow oop directly and do implicit NULL check in address:
3592 //
3593 // [R12 + narrow_oop_reg<<3 + offset]
3594 // NullCheck narrow_oop_reg
3595 //
3596 // On other platforms (Sparc) we have to keep new DecodeN node and
3597 // use it to do implicit NULL check in address:
3598 //
3599 // decode_not_null narrow_oop_reg, base_reg
3600 // [base_reg + offset]
3601 // NullCheck base_reg
3602 //
3603 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3604 // to keep the information to which NULL check the new DecodeN node
3605 // corresponds to use it as value in implicit_null_check().
3606 //
3607 new_in1->set_req(0, n->in(0));
3608 }
3609
3610 n->subsume_by(new_in1, this);
3611 if (in1->outcnt() == 0) {
3612 in1->disconnect_inputs(NULL, this);
3613 }
3614 } else {
3615 n->subsume_by(n->in(1), this);
3616 if (n->outcnt() == 0) {
3617 n->disconnect_inputs(NULL, this);
3618 }
3619 }
3620 break;
3621 }
3622 #ifdef _LP64
3623 case Op_CmpP:
3624 // Do this transformation here to preserve CmpPNode::sub() and
3625 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3626 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3627 Node* in1 = n->in(1);
3628 Node* in2 = n->in(2);
3629 if (!in1->is_DecodeNarrowPtr()) {
3630 in2 = in1;
3631 in1 = n->in(2);
3632 }
3633 assert(in1->is_DecodeNarrowPtr(), "sanity");
3634
3635 Node* new_in2 = NULL;
3636 if (in2->is_DecodeNarrowPtr()) {
3637 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3638 new_in2 = in2->in(1);
3639 } else if (in2->Opcode() == Op_ConP) {
3640 const Type* t = in2->bottom_type();
3641 if (t == TypePtr::NULL_PTR) {
3642 assert(in1->is_DecodeN(), "compare klass to null?");
3643 // Don't convert CmpP null check into CmpN if compressed
3644 // oops implicit null check is not generated.
3645 // This will allow to generate normal oop implicit null check.
3646 if (Matcher::gen_narrow_oop_implicit_null_checks())
3647 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3648 //
3649 // This transformation together with CastPP transformation above
3650 // will generated code for implicit NULL checks for compressed oops.
3651 //
3652 // The original code after Optimize()
3653 //
3654 // LoadN memory, narrow_oop_reg
3655 // decode narrow_oop_reg, base_reg
3656 // CmpP base_reg, NULL
3657 // CastPP base_reg // NotNull
3658 // Load [base_reg + offset], val_reg
3659 //
3660 // after these transformations will be
3661 //
3662 // LoadN memory, narrow_oop_reg
3663 // CmpN narrow_oop_reg, NULL
3664 // decode_not_null narrow_oop_reg, base_reg
3665 // Load [base_reg + offset], val_reg
3666 //
3667 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3668 // since narrow oops can be used in debug info now (see the code in
3669 // final_graph_reshaping_walk()).
3670 //
3671 // At the end the code will be matched to
3672 // on x86:
3673 //
3674 // Load_narrow_oop memory, narrow_oop_reg
3675 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3676 // NullCheck narrow_oop_reg
3677 //
3678 // and on sparc:
3679 //
3680 // Load_narrow_oop memory, narrow_oop_reg
3681 // decode_not_null narrow_oop_reg, base_reg
3682 // Load [base_reg + offset], val_reg
3683 // NullCheck base_reg
3684 //
3685 } else if (t->isa_oopptr()) {
3686 new_in2 = ConNode::make(t->make_narrowoop());
3687 } else if (t->isa_klassptr()) {
3688 new_in2 = ConNode::make(t->make_narrowklass());
3689 }
3690 }
3691 if (new_in2 != NULL) {
3692 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3693 n->subsume_by(cmpN, this);
3694 if (in1->outcnt() == 0) {
3695 in1->disconnect_inputs(NULL, this);
3696 }
3697 if (in2->outcnt() == 0) {
3698 in2->disconnect_inputs(NULL, this);
3699 }
3700 }
3701 }
3702 break;
3703
3704 case Op_DecodeN:
3705 case Op_DecodeNKlass:
3706 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3707 // DecodeN could be pinned when it can't be fold into
3708 // an address expression, see the code for Op_CastPP above.
3709 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3710 break;
3711
3712 case Op_EncodeP:
3713 case Op_EncodePKlass: {
3714 Node* in1 = n->in(1);
3715 if (in1->is_DecodeNarrowPtr()) {
3716 n->subsume_by(in1->in(1), this);
3717 } else if (in1->Opcode() == Op_ConP) {
3718 const Type* t = in1->bottom_type();
3719 if (t == TypePtr::NULL_PTR) {
3720 assert(t->isa_oopptr(), "null klass?");
3721 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3722 } else if (t->isa_oopptr()) {
3723 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3724 } else if (t->isa_klassptr()) {
3725 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3726 }
3727 }
3728 if (in1->outcnt() == 0) {
3729 in1->disconnect_inputs(NULL, this);
3730 }
3731 break;
3732 }
3733
3734 case Op_Proj: {
3735 if (OptimizeStringConcat) {
3736 ProjNode* p = n->as_Proj();
3737 if (p->_is_io_use) {
3738 // Separate projections were used for the exception path which
3739 // are normally removed by a late inline. If it wasn't inlined
3740 // then they will hang around and should just be replaced with
3741 // the original one.
3742 Node* proj = NULL;
3743 // Replace with just one
3744 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3745 Node *use = i.get();
3746 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3747 proj = use;
3748 break;
3749 }
3750 }
3751 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3752 if (proj != NULL) {
3753 p->subsume_by(proj, this);
3754 }
3755 }
3756 }
3757 break;
3758 }
3759
3760 case Op_Phi:
3761 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3762 // The EncodeP optimization may create Phi with the same edges
3763 // for all paths. It is not handled well by Register Allocator.
3764 Node* unique_in = n->in(1);
3765 assert(unique_in != NULL, "");
3766 uint cnt = n->req();
3767 for (uint i = 2; i < cnt; i++) {
3768 Node* m = n->in(i);
3769 assert(m != NULL, "");
3770 if (unique_in != m)
3771 unique_in = NULL;
3772 }
3773 if (unique_in != NULL) {
3774 n->subsume_by(unique_in, this);
3775 }
3776 }
3777 break;
3778
3779 #endif
3780
3781 #ifdef ASSERT
3782 case Op_CastII:
3783 // Verify that all range check dependent CastII nodes were removed.
3784 if (n->isa_CastII()->has_range_check()) {
3785 n->dump(3);
3786 assert(false, "Range check dependent CastII node was not removed");
3787 }
3788 break;
3789 #endif
3790
3791 case Op_ModI:
3792 if (UseDivMod) {
3793 // Check if a%b and a/b both exist
3794 Node* d = n->find_similar(Op_DivI);
3795 if (d) {
3796 // Replace them with a fused divmod if supported
3797 if (Matcher::has_match_rule(Op_DivModI)) {
3798 DivModINode* divmod = DivModINode::make(n);
3799 d->subsume_by(divmod->div_proj(), this);
3800 n->subsume_by(divmod->mod_proj(), this);
3801 } else {
3802 // replace a%b with a-((a/b)*b)
3803 Node* mult = new MulINode(d, d->in(2));
3804 Node* sub = new SubINode(d->in(1), mult);
3805 n->subsume_by(sub, this);
3806 }
3807 }
3808 }
3809 break;
3810
3811 case Op_ModL:
3812 if (UseDivMod) {
3813 // Check if a%b and a/b both exist
3814 Node* d = n->find_similar(Op_DivL);
3815 if (d) {
3816 // Replace them with a fused divmod if supported
3817 if (Matcher::has_match_rule(Op_DivModL)) {
3818 DivModLNode* divmod = DivModLNode::make(n);
3819 d->subsume_by(divmod->div_proj(), this);
3820 n->subsume_by(divmod->mod_proj(), this);
3821 } else {
3822 // replace a%b with a-((a/b)*b)
3823 Node* mult = new MulLNode(d, d->in(2));
3824 Node* sub = new SubLNode(d->in(1), mult);
3825 n->subsume_by(sub, this);
3826 }
3827 }
3828 }
3829 break;
3830
3831 case Op_LoadVector:
3832 case Op_StoreVector:
3833 break;
3834
3835 case Op_AddReductionVI:
3836 case Op_AddReductionVL:
3837 case Op_AddReductionVF:
3838 case Op_AddReductionVD:
3839 case Op_MulReductionVI:
3840 case Op_MulReductionVL:
3841 case Op_MulReductionVF:
3842 case Op_MulReductionVD:
3843 case Op_MinReductionV:
3844 case Op_MaxReductionV:
3845 break;
3846
3847 case Op_PackB:
3848 case Op_PackS:
3849 case Op_PackI:
3850 case Op_PackF:
3851 case Op_PackL:
3852 case Op_PackD:
3853 if (n->req()-1 > 2) {
3854 // Replace many operand PackNodes with a binary tree for matching
3855 PackNode* p = (PackNode*) n;
3856 Node* btp = p->binary_tree_pack(1, n->req());
3857 n->subsume_by(btp, this);
3858 }
3859 break;
3860 case Op_Loop:
3861 case Op_CountedLoop:
3862 case Op_OuterStripMinedLoop:
3863 if (n->as_Loop()->is_inner_loop()) {
3864 frc.inc_inner_loop_count();
3865 }
3866 n->as_Loop()->verify_strip_mined(0);
3867 break;
3868 case Op_LShiftI:
3869 case Op_RShiftI:
3870 case Op_URShiftI:
3871 case Op_LShiftL:
3872 case Op_RShiftL:
3873 case Op_URShiftL:
3874 if (Matcher::need_masked_shift_count) {
3875 // The cpu's shift instructions don't restrict the count to the
3876 // lower 5/6 bits. We need to do the masking ourselves.
3877 Node* in2 = n->in(2);
3878 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3879 const TypeInt* t = in2->find_int_type();
3880 if (t != NULL && t->is_con()) {
3881 juint shift = t->get_con();
3882 if (shift > mask) { // Unsigned cmp
3883 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3884 }
3885 } else {
3886 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3887 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3888 n->set_req(2, shift);
3889 }
3890 }
3891 if (in2->outcnt() == 0) { // Remove dead node
3892 in2->disconnect_inputs(NULL, this);
3893 }
3894 }
3895 break;
3896 case Op_MemBarStoreStore:
3897 case Op_MemBarRelease:
3898 // Break the link with AllocateNode: it is no longer useful and
3899 // confuses register allocation.
3900 if (n->req() > MemBarNode::Precedent) {
3901 n->set_req(MemBarNode::Precedent, top());
3902 }
3903 break;
3904 case Op_MemBarAcquire: {
3905 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3906 // At parse time, the trailing MemBarAcquire for a volatile load
3907 // is created with an edge to the load. After optimizations,
3908 // that input may be a chain of Phis. If those phis have no
3909 // other use, then the MemBarAcquire keeps them alive and
3910 // register allocation can be confused.
3911 ResourceMark rm;
3912 Unique_Node_List wq;
3913 wq.push(n->in(MemBarNode::Precedent));
3914 n->set_req(MemBarNode::Precedent, top());
3915 while (wq.size() > 0) {
3916 Node* m = wq.pop();
3917 if (m->outcnt() == 0) {
3918 for (uint j = 0; j < m->req(); j++) {
3919 Node* in = m->in(j);
3920 if (in != NULL) {
3921 wq.push(in);
3922 }
3923 }
3924 m->disconnect_inputs(NULL, this);
3925 }
3926 }
3927 }
3928 break;
3929 }
3930 case Op_RangeCheck: {
3931 RangeCheckNode* rc = n->as_RangeCheck();
3932 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3933 n->subsume_by(iff, this);
3934 frc._tests.push(iff);
3935 break;
3936 }
3937 case Op_ConvI2L: {
3938 if (!Matcher::convi2l_type_required) {
3939 // Code generation on some platforms doesn't need accurate
3940 // ConvI2L types. Widening the type can help remove redundant
3941 // address computations.
3942 n->as_Type()->set_type(TypeLong::INT);
3943 ResourceMark rm;
3944 Node_List wq;
3945 wq.push(n);
3946 for (uint next = 0; next < wq.size(); next++) {
3947 Node *m = wq.at(next);
3948
3949 for(;;) {
3950 // Loop over all nodes with identical inputs edges as m
3951 Node* k = m->find_similar(m->Opcode());
3952 if (k == NULL) {
3953 break;
3954 }
3955 // Push their uses so we get a chance to remove node made
3956 // redundant
3957 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3958 Node* u = k->fast_out(i);
3959 assert(!wq.contains(u), "shouldn't process one node several times");
3960 if (u->Opcode() == Op_LShiftL ||
3961 u->Opcode() == Op_AddL ||
3962 u->Opcode() == Op_SubL ||
3963 u->Opcode() == Op_AddP) {
3964 wq.push(u);
3965 }
3966 }
3967 // Replace all nodes with identical edges as m with m
3968 k->subsume_by(m, this);
3969 }
3970 }
3971 }
3972 break;
3973 }
3974 case Op_CmpUL: {
3975 if (!Matcher::has_match_rule(Op_CmpUL)) {
3976 // No support for unsigned long comparisons
3977 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3978 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3979 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3980 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3981 Node* andl = new AndLNode(orl, remove_sign_mask);
3982 Node* cmp = new CmpLNode(andl, n->in(2));
3983 n->subsume_by(cmp, this);
3984 }
3985 break;
3986 }
3987 #ifdef ASSERT
3988 case Op_ValueTypePtr:
3989 case Op_ValueType: {
3990 n->dump(-1);
3991 assert(false, "value type node was not removed");
3992 break;
3993 }
3994 #endif
3995 case Op_GetNullFreeProperty: {
3996 // Extract the null free bits
3997 uint last = unique();
3998 Node* null_free = NULL;
3999 if (n->in(1)->Opcode() == Op_LoadKlass) {
4000 Node* cast = new CastP2XNode(NULL, n->in(1));
4001 null_free = new AndLNode(cast, new ConLNode(TypeLong::make(((jlong)1)<<(oopDesc::wide_storage_props_shift + ArrayStorageProperties::null_free_bit))));
4002 } else {
4003 assert(n->in(1)->Opcode() == Op_LoadNKlass, "not a compressed klass?");
4004 Node* cast = new CastN2INode(n->in(1));
4005 null_free = new AndINode(cast, new ConINode(TypeInt::make(1<<(oopDesc::narrow_storage_props_shift + ArrayStorageProperties::null_free_bit))));
4006 }
4007 n->replace_by(null_free);
4008 break;
4009 }
4010 default:
4011 assert(!n->is_Call(), "");
4012 assert(!n->is_Mem(), "");
4013 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4014 break;
4015 }
4016 }
4017
4018 //------------------------------final_graph_reshaping_walk---------------------
4019 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4020 // requires that the walk visits a node's inputs before visiting the node.
4021 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4022 ResourceArea *area = Thread::current()->resource_area();
4023 Unique_Node_List sfpt(area);
4024
4025 frc._visited.set(root->_idx); // first, mark node as visited
4026 uint cnt = root->req();
4027 Node *n = root;
4028 uint i = 0;
4029 while (true) {
4030 if (i < cnt) {
4031 // Place all non-visited non-null inputs onto stack
4032 Node* m = n->in(i);
4033 ++i;
4034 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4035 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4036 // compute worst case interpreter size in case of a deoptimization
4037 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4038
4039 sfpt.push(m);
4040 }
4041 cnt = m->req();
4042 nstack.push(n, i); // put on stack parent and next input's index
4043 n = m;
4044 i = 0;
4045 }
4046 } else {
4047 // Now do post-visit work
4048 final_graph_reshaping_impl( n, frc );
4049 if (nstack.is_empty())
4050 break; // finished
4051 n = nstack.node(); // Get node from stack
4052 cnt = n->req();
4053 i = nstack.index();
4054 nstack.pop(); // Shift to the next node on stack
4055 }
4056 }
4057
4058 // Skip next transformation if compressed oops are not used.
4059 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4060 (!UseCompressedOops && !UseCompressedClassPointers))
4061 return;
4062
4063 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4064 // It could be done for an uncommon traps or any safepoints/calls
4065 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4066 while (sfpt.size() > 0) {
4067 n = sfpt.pop();
4068 JVMState *jvms = n->as_SafePoint()->jvms();
4069 assert(jvms != NULL, "sanity");
4070 int start = jvms->debug_start();
4071 int end = n->req();
4072 bool is_uncommon = (n->is_CallStaticJava() &&
4073 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4074 for (int j = start; j < end; j++) {
4075 Node* in = n->in(j);
4076 if (in->is_DecodeNarrowPtr()) {
4077 bool safe_to_skip = true;
4078 if (!is_uncommon ) {
4079 // Is it safe to skip?
4080 for (uint i = 0; i < in->outcnt(); i++) {
4081 Node* u = in->raw_out(i);
4082 if (!u->is_SafePoint() ||
4083 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4084 safe_to_skip = false;
4085 }
4086 }
4087 }
4088 if (safe_to_skip) {
4089 n->set_req(j, in->in(1));
4090 }
4091 if (in->outcnt() == 0) {
4092 in->disconnect_inputs(NULL, this);
4093 }
4094 }
4095 }
4096 }
4097 }
4098
4099 //------------------------------final_graph_reshaping--------------------------
4100 // Final Graph Reshaping.
4101 //
4102 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4103 // and not commoned up and forced early. Must come after regular
4104 // optimizations to avoid GVN undoing the cloning. Clone constant
4105 // inputs to Loop Phis; these will be split by the allocator anyways.
4106 // Remove Opaque nodes.
4107 // (2) Move last-uses by commutative operations to the left input to encourage
4108 // Intel update-in-place two-address operations and better register usage
4109 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4110 // calls canonicalizing them back.
4111 // (3) Count the number of double-precision FP ops, single-precision FP ops
4112 // and call sites. On Intel, we can get correct rounding either by
4113 // forcing singles to memory (requires extra stores and loads after each
4114 // FP bytecode) or we can set a rounding mode bit (requires setting and
4115 // clearing the mode bit around call sites). The mode bit is only used
4116 // if the relative frequency of single FP ops to calls is low enough.
4117 // This is a key transform for SPEC mpeg_audio.
4118 // (4) Detect infinite loops; blobs of code reachable from above but not
4119 // below. Several of the Code_Gen algorithms fail on such code shapes,
4120 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4121 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4122 // Detection is by looking for IfNodes where only 1 projection is
4123 // reachable from below or CatchNodes missing some targets.
4124 // (5) Assert for insane oop offsets in debug mode.
4125
4126 bool Compile::final_graph_reshaping() {
4127 // an infinite loop may have been eliminated by the optimizer,
4128 // in which case the graph will be empty.
4129 if (root()->req() == 1) {
4130 record_method_not_compilable("trivial infinite loop");
4131 return true;
4132 }
4133
4134 // Expensive nodes have their control input set to prevent the GVN
4135 // from freely commoning them. There's no GVN beyond this point so
4136 // no need to keep the control input. We want the expensive nodes to
4137 // be freely moved to the least frequent code path by gcm.
4138 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4139 for (int i = 0; i < expensive_count(); i++) {
4140 _expensive_nodes->at(i)->set_req(0, NULL);
4141 }
4142
4143 Final_Reshape_Counts frc;
4144
4145 // Visit everybody reachable!
4146 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4147 Node_Stack nstack(live_nodes() >> 1);
4148 final_graph_reshaping_walk(nstack, root(), frc);
4149
4150 // Check for unreachable (from below) code (i.e., infinite loops).
4151 for( uint i = 0; i < frc._tests.size(); i++ ) {
4152 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4153 // Get number of CFG targets.
4154 // Note that PCTables include exception targets after calls.
4155 uint required_outcnt = n->required_outcnt();
4156 if (n->outcnt() != required_outcnt) {
4157 // Check for a few special cases. Rethrow Nodes never take the
4158 // 'fall-thru' path, so expected kids is 1 less.
4159 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4160 if (n->in(0)->in(0)->is_Call()) {
4161 CallNode *call = n->in(0)->in(0)->as_Call();
4162 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4163 required_outcnt--; // Rethrow always has 1 less kid
4164 } else if (call->req() > TypeFunc::Parms &&
4165 call->is_CallDynamicJava()) {
4166 // Check for null receiver. In such case, the optimizer has
4167 // detected that the virtual call will always result in a null
4168 // pointer exception. The fall-through projection of this CatchNode
4169 // will not be populated.
4170 Node *arg0 = call->in(TypeFunc::Parms);
4171 if (arg0->is_Type() &&
4172 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4173 required_outcnt--;
4174 }
4175 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4176 call->req() > TypeFunc::Parms+1 &&
4177 call->is_CallStaticJava()) {
4178 // Check for negative array length. In such case, the optimizer has
4179 // detected that the allocation attempt will always result in an
4180 // exception. There is no fall-through projection of this CatchNode .
4181 Node *arg1 = call->in(TypeFunc::Parms+1);
4182 if (arg1->is_Type() &&
4183 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4184 required_outcnt--;
4185 }
4186 }
4187 }
4188 }
4189 // Recheck with a better notion of 'required_outcnt'
4190 if (n->outcnt() != required_outcnt) {
4191 record_method_not_compilable("malformed control flow");
4192 return true; // Not all targets reachable!
4193 }
4194 }
4195 // Check that I actually visited all kids. Unreached kids
4196 // must be infinite loops.
4197 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4198 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4199 record_method_not_compilable("infinite loop");
4200 return true; // Found unvisited kid; must be unreach
4201 }
4202
4203 // Here so verification code in final_graph_reshaping_walk()
4204 // always see an OuterStripMinedLoopEnd
4205 if (n->is_OuterStripMinedLoopEnd()) {
4206 IfNode* init_iff = n->as_If();
4207 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4208 n->subsume_by(iff, this);
4209 }
4210 }
4211
4212 // If original bytecodes contained a mixture of floats and doubles
4213 // check if the optimizer has made it homogenous, item (3).
4214 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
4215 frc.get_float_count() > 32 &&
4216 frc.get_double_count() == 0 &&
4217 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4218 set_24_bit_selection_and_mode( false, true );
4219 }
4220
4221 set_java_calls(frc.get_java_call_count());
4222 set_inner_loops(frc.get_inner_loop_count());
4223
4224 // No infinite loops, no reason to bail out.
4225 return false;
4226 }
4227
4228 //-----------------------------too_many_traps----------------------------------
4229 // Report if there are too many traps at the current method and bci.
4230 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4231 bool Compile::too_many_traps(ciMethod* method,
4232 int bci,
4233 Deoptimization::DeoptReason reason) {
4234 ciMethodData* md = method->method_data();
4235 if (md->is_empty()) {
4236 // Assume the trap has not occurred, or that it occurred only
4237 // because of a transient condition during start-up in the interpreter.
4238 return false;
4239 }
4240 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4241 if (md->has_trap_at(bci, m, reason) != 0) {
4242 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4243 // Also, if there are multiple reasons, or if there is no per-BCI record,
4244 // assume the worst.
4245 if (log())
4246 log()->elem("observe trap='%s' count='%d'",
4247 Deoptimization::trap_reason_name(reason),
4248 md->trap_count(reason));
4249 return true;
4250 } else {
4251 // Ignore method/bci and see if there have been too many globally.
4252 return too_many_traps(reason, md);
4253 }
4254 }
4255
4256 // Less-accurate variant which does not require a method and bci.
4257 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4258 ciMethodData* logmd) {
4259 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4260 // Too many traps globally.
4261 // Note that we use cumulative trap_count, not just md->trap_count.
4262 if (log()) {
4263 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4264 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4265 Deoptimization::trap_reason_name(reason),
4266 mcount, trap_count(reason));
4267 }
4268 return true;
4269 } else {
4270 // The coast is clear.
4271 return false;
4272 }
4273 }
4274
4275 //--------------------------too_many_recompiles--------------------------------
4276 // Report if there are too many recompiles at the current method and bci.
4277 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4278 // Is not eager to return true, since this will cause the compiler to use
4279 // Action_none for a trap point, to avoid too many recompilations.
4280 bool Compile::too_many_recompiles(ciMethod* method,
4281 int bci,
4282 Deoptimization::DeoptReason reason) {
4283 ciMethodData* md = method->method_data();
4284 if (md->is_empty()) {
4285 // Assume the trap has not occurred, or that it occurred only
4286 // because of a transient condition during start-up in the interpreter.
4287 return false;
4288 }
4289 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4290 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4291 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4292 Deoptimization::DeoptReason per_bc_reason
4293 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4294 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4295 if ((per_bc_reason == Deoptimization::Reason_none
4296 || md->has_trap_at(bci, m, reason) != 0)
4297 // The trap frequency measure we care about is the recompile count:
4298 && md->trap_recompiled_at(bci, m)
4299 && md->overflow_recompile_count() >= bc_cutoff) {
4300 // Do not emit a trap here if it has already caused recompilations.
4301 // Also, if there are multiple reasons, or if there is no per-BCI record,
4302 // assume the worst.
4303 if (log())
4304 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4305 Deoptimization::trap_reason_name(reason),
4306 md->trap_count(reason),
4307 md->overflow_recompile_count());
4308 return true;
4309 } else if (trap_count(reason) != 0
4310 && decompile_count() >= m_cutoff) {
4311 // Too many recompiles globally, and we have seen this sort of trap.
4312 // Use cumulative decompile_count, not just md->decompile_count.
4313 if (log())
4314 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4315 Deoptimization::trap_reason_name(reason),
4316 md->trap_count(reason), trap_count(reason),
4317 md->decompile_count(), decompile_count());
4318 return true;
4319 } else {
4320 // The coast is clear.
4321 return false;
4322 }
4323 }
4324
4325 // Compute when not to trap. Used by matching trap based nodes and
4326 // NullCheck optimization.
4327 void Compile::set_allowed_deopt_reasons() {
4328 _allowed_reasons = 0;
4329 if (is_method_compilation()) {
4330 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4331 assert(rs < BitsPerInt, "recode bit map");
4332 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4333 _allowed_reasons |= nth_bit(rs);
4334 }
4335 }
4336 }
4337 }
4338
4339 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4340 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4341 }
4342
4343 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4344 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4345 }
4346
4347 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4348 if (holder->is_initialized()) {
4349 return false;
4350 }
4351 if (holder->is_being_initialized()) {
4352 if (accessing_method->holder() == holder) {
4353 // Access inside a class. The barrier can be elided when access happens in <clinit>,
4354 // <init>, or a static method. In all those cases, there was an initialization
4355 // barrier on the holder klass passed.
4356 if (accessing_method->is_class_initializer() ||
4357 accessing_method->is_object_constructor() ||
4358 accessing_method->is_static()) {
4359 return false;
4360 }
4361 } else if (accessing_method->holder()->is_subclass_of(holder)) {
4362 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4363 // In case of <init> or a static method, the barrier is on the subclass is not enough:
4364 // child class can become fully initialized while its parent class is still being initialized.
4365 if (accessing_method->is_class_initializer()) {
4366 return false;
4367 }
4368 }
4369 ciMethod* root = method(); // the root method of compilation
4370 if (root != accessing_method) {
4371 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4372 }
4373 }
4374 return true;
4375 }
4376
4377 #ifndef PRODUCT
4378 //------------------------------verify_graph_edges---------------------------
4379 // Walk the Graph and verify that there is a one-to-one correspondence
4380 // between Use-Def edges and Def-Use edges in the graph.
4381 void Compile::verify_graph_edges(bool no_dead_code) {
4382 if (VerifyGraphEdges) {
4383 ResourceArea *area = Thread::current()->resource_area();
4384 Unique_Node_List visited(area);
4385 // Call recursive graph walk to check edges
4386 _root->verify_edges(visited);
4387 if (no_dead_code) {
4388 // Now make sure that no visited node is used by an unvisited node.
4389 bool dead_nodes = false;
4390 Unique_Node_List checked(area);
4391 while (visited.size() > 0) {
4392 Node* n = visited.pop();
4393 checked.push(n);
4394 for (uint i = 0; i < n->outcnt(); i++) {
4395 Node* use = n->raw_out(i);
4396 if (checked.member(use)) continue; // already checked
4397 if (visited.member(use)) continue; // already in the graph
4398 if (use->is_Con()) continue; // a dead ConNode is OK
4399 // At this point, we have found a dead node which is DU-reachable.
4400 if (!dead_nodes) {
4401 tty->print_cr("*** Dead nodes reachable via DU edges:");
4402 dead_nodes = true;
4403 }
4404 use->dump(2);
4405 tty->print_cr("---");
4406 checked.push(use); // No repeats; pretend it is now checked.
4407 }
4408 }
4409 assert(!dead_nodes, "using nodes must be reachable from root");
4410 }
4411 }
4412 }
4413 #endif
4414
4415 // The Compile object keeps track of failure reasons separately from the ciEnv.
4416 // This is required because there is not quite a 1-1 relation between the
4417 // ciEnv and its compilation task and the Compile object. Note that one
4418 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4419 // to backtrack and retry without subsuming loads. Other than this backtracking
4420 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4421 // by the logic in C2Compiler.
4422 void Compile::record_failure(const char* reason) {
4423 if (log() != NULL) {
4424 log()->elem("failure reason='%s' phase='compile'", reason);
4425 }
4426 if (_failure_reason == NULL) {
4427 // Record the first failure reason.
4428 _failure_reason = reason;
4429 }
4430
4431 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4432 C->print_method(PHASE_FAILURE);
4433 }
4434 _root = NULL; // flush the graph, too
4435 }
4436
4437 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4438 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4439 _phase_name(name), _dolog(CITimeVerbose)
4440 {
4441 if (_dolog) {
4442 C = Compile::current();
4443 _log = C->log();
4444 } else {
4445 C = NULL;
4446 _log = NULL;
4447 }
4448 if (_log != NULL) {
4449 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4450 _log->stamp();
4451 _log->end_head();
4452 }
4453 }
4454
4455 Compile::TracePhase::~TracePhase() {
4456
4457 C = Compile::current();
4458 if (_dolog) {
4459 _log = C->log();
4460 } else {
4461 _log = NULL;
4462 }
4463
4464 #ifdef ASSERT
4465 if (PrintIdealNodeCount) {
4466 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4467 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4468 }
4469
4470 if (VerifyIdealNodeCount) {
4471 Compile::current()->print_missing_nodes();
4472 }
4473 #endif
4474
4475 if (_log != NULL) {
4476 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4477 }
4478 }
4479
4480 //=============================================================================
4481 // Two Constant's are equal when the type and the value are equal.
4482 bool Compile::Constant::operator==(const Constant& other) {
4483 if (type() != other.type() ) return false;
4484 if (can_be_reused() != other.can_be_reused()) return false;
4485 // For floating point values we compare the bit pattern.
4486 switch (type()) {
4487 case T_INT:
4488 case T_FLOAT: return (_v._value.i == other._v._value.i);
4489 case T_LONG:
4490 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4491 case T_OBJECT:
4492 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4493 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4494 case T_METADATA: return (_v._metadata == other._v._metadata);
4495 default: ShouldNotReachHere(); return false;
4496 }
4497 }
4498
4499 static int type_to_size_in_bytes(BasicType t) {
4500 switch (t) {
4501 case T_INT: return sizeof(jint );
4502 case T_LONG: return sizeof(jlong );
4503 case T_FLOAT: return sizeof(jfloat );
4504 case T_DOUBLE: return sizeof(jdouble);
4505 case T_METADATA: return sizeof(Metadata*);
4506 // We use T_VOID as marker for jump-table entries (labels) which
4507 // need an internal word relocation.
4508 case T_VOID:
4509 case T_ADDRESS:
4510 case T_OBJECT: return sizeof(jobject);
4511 default:
4512 ShouldNotReachHere();
4513 return -1;
4514 }
4515 }
4516
4517 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4518 // sort descending
4519 if (a->freq() > b->freq()) return -1;
4520 if (a->freq() < b->freq()) return 1;
4521 return 0;
4522 }
4523
4524 void Compile::ConstantTable::calculate_offsets_and_size() {
4525 // First, sort the array by frequencies.
4526 _constants.sort(qsort_comparator);
4527
4528 #ifdef ASSERT
4529 // Make sure all jump-table entries were sorted to the end of the
4530 // array (they have a negative frequency).
4531 bool found_void = false;
4532 for (int i = 0; i < _constants.length(); i++) {
4533 Constant con = _constants.at(i);
4534 if (con.type() == T_VOID)
4535 found_void = true; // jump-tables
4536 else
4537 assert(!found_void, "wrong sorting");
4538 }
4539 #endif
4540
4541 int offset = 0;
4542 for (int i = 0; i < _constants.length(); i++) {
4543 Constant* con = _constants.adr_at(i);
4544
4545 // Align offset for type.
4546 int typesize = type_to_size_in_bytes(con->type());
4547 offset = align_up(offset, typesize);
4548 con->set_offset(offset); // set constant's offset
4549
4550 if (con->type() == T_VOID) {
4551 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4552 offset = offset + typesize * n->outcnt(); // expand jump-table
4553 } else {
4554 offset = offset + typesize;
4555 }
4556 }
4557
4558 // Align size up to the next section start (which is insts; see
4559 // CodeBuffer::align_at_start).
4560 assert(_size == -1, "already set?");
4561 _size = align_up(offset, (int)CodeEntryAlignment);
4562 }
4563
4564 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4565 MacroAssembler _masm(&cb);
4566 for (int i = 0; i < _constants.length(); i++) {
4567 Constant con = _constants.at(i);
4568 address constant_addr = NULL;
4569 switch (con.type()) {
4570 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4571 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4572 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4573 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4574 case T_OBJECT: {
4575 jobject obj = con.get_jobject();
4576 int oop_index = _masm.oop_recorder()->find_index(obj);
4577 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4578 break;
4579 }
4580 case T_ADDRESS: {
4581 address addr = (address) con.get_jobject();
4582 constant_addr = _masm.address_constant(addr);
4583 break;
4584 }
4585 // We use T_VOID as marker for jump-table entries (labels) which
4586 // need an internal word relocation.
4587 case T_VOID: {
4588 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4589 // Fill the jump-table with a dummy word. The real value is
4590 // filled in later in fill_jump_table.
4591 address dummy = (address) n;
4592 constant_addr = _masm.address_constant(dummy);
4593 // Expand jump-table
4594 for (uint i = 1; i < n->outcnt(); i++) {
4595 address temp_addr = _masm.address_constant(dummy + i);
4596 assert(temp_addr, "consts section too small");
4597 }
4598 break;
4599 }
4600 case T_METADATA: {
4601 Metadata* obj = con.get_metadata();
4602 int metadata_index = _masm.oop_recorder()->find_index(obj);
4603 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4604 break;
4605 }
4606 default: ShouldNotReachHere();
4607 }
4608 assert(constant_addr, "consts section too small");
4609 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4610 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4611 }
4612 }
4613
4614 int Compile::ConstantTable::find_offset(Constant& con) const {
4615 int idx = _constants.find(con);
4616 guarantee(idx != -1, "constant must be in constant table");
4617 int offset = _constants.at(idx).offset();
4618 guarantee(offset != -1, "constant table not emitted yet?");
4619 return offset;
4620 }
4621
4622 void Compile::ConstantTable::add(Constant& con) {
4623 if (con.can_be_reused()) {
4624 int idx = _constants.find(con);
4625 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4626 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4627 return;
4628 }
4629 }
4630 (void) _constants.append(con);
4631 }
4632
4633 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4634 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4635 Constant con(type, value, b->_freq);
4636 add(con);
4637 return con;
4638 }
4639
4640 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4641 Constant con(metadata);
4642 add(con);
4643 return con;
4644 }
4645
4646 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4647 jvalue value;
4648 BasicType type = oper->type()->basic_type();
4649 switch (type) {
4650 case T_LONG: value.j = oper->constantL(); break;
4651 case T_FLOAT: value.f = oper->constantF(); break;
4652 case T_DOUBLE: value.d = oper->constantD(); break;
4653 case T_OBJECT:
4654 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4655 case T_METADATA: return add((Metadata*)oper->constant()); break;
4656 default: guarantee(false, "unhandled type: %s", type2name(type));
4657 }
4658 return add(n, type, value);
4659 }
4660
4661 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4662 jvalue value;
4663 // We can use the node pointer here to identify the right jump-table
4664 // as this method is called from Compile::Fill_buffer right before
4665 // the MachNodes are emitted and the jump-table is filled (means the
4666 // MachNode pointers do not change anymore).
4667 value.l = (jobject) n;
4668 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4669 add(con);
4670 return con;
4671 }
4672
4673 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4674 // If called from Compile::scratch_emit_size do nothing.
4675 if (Compile::current()->in_scratch_emit_size()) return;
4676
4677 assert(labels.is_nonempty(), "must be");
4678 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4679
4680 // Since MachConstantNode::constant_offset() also contains
4681 // table_base_offset() we need to subtract the table_base_offset()
4682 // to get the plain offset into the constant table.
4683 int offset = n->constant_offset() - table_base_offset();
4684
4685 MacroAssembler _masm(&cb);
4686 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4687
4688 for (uint i = 0; i < n->outcnt(); i++) {
4689 address* constant_addr = &jump_table_base[i];
4690 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));
4691 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4692 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4693 }
4694 }
4695
4696 //----------------------------static_subtype_check-----------------------------
4697 // Shortcut important common cases when superklass is exact:
4698 // (0) superklass is java.lang.Object (can occur in reflective code)
4699 // (1) subklass is already limited to a subtype of superklass => always ok
4700 // (2) subklass does not overlap with superklass => always fail
4701 // (3) superklass has NO subtypes and we can check with a simple compare.
4702 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4703 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4704 return SSC_full_test; // Let caller generate the general case.
4705 }
4706
4707 if (superk == env()->Object_klass()) {
4708 return SSC_always_true; // (0) this test cannot fail
4709 }
4710
4711 ciType* superelem = superk;
4712 if (superelem->is_array_klass()) {
4713 ciArrayKlass* ak = superelem->as_array_klass();
4714 superelem = superelem->as_array_klass()->base_element_type();
4715 }
4716
4717 if (!subk->is_interface()) { // cannot trust static interface types yet
4718 if (subk->is_subtype_of(superk)) {
4719 return SSC_always_true; // (1) false path dead; no dynamic test needed
4720 }
4721 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4722 !superk->is_subtype_of(subk)) {
4723 return SSC_always_false;
4724 }
4725 }
4726
4727 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4728 // [V is a subtype of [V? but the klass for [V is not equal to the klass for [V?. Perform a full test.
4729 if (superk->is_obj_array_klass() && !superk->as_array_klass()->storage_properties().is_null_free() && superk->as_array_klass()->element_klass()->is_valuetype()) {
4730 return SSC_full_test;
4731 }
4732 // If casting to an instance klass, it must have no subtypes
4733 if (superk->is_interface()) {
4734 // Cannot trust interfaces yet.
4735 // %%% S.B. superk->nof_implementors() == 1
4736 } else if (superelem->is_instance_klass()) {
4737 ciInstanceKlass* ik = superelem->as_instance_klass();
4738 if (!ik->has_subklass() && !ik->is_interface()) {
4739 if (!ik->is_final()) {
4740 // Add a dependency if there is a chance of a later subclass.
4741 dependencies()->assert_leaf_type(ik);
4742 }
4743 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4744 }
4745 } else {
4746 // A primitive array type has no subtypes.
4747 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4748 }
4749
4750 return SSC_full_test;
4751 }
4752
4753 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4754 #ifdef _LP64
4755 // The scaled index operand to AddP must be a clean 64-bit value.
4756 // Java allows a 32-bit int to be incremented to a negative
4757 // value, which appears in a 64-bit register as a large
4758 // positive number. Using that large positive number as an
4759 // operand in pointer arithmetic has bad consequences.
4760 // On the other hand, 32-bit overflow is rare, and the possibility
4761 // can often be excluded, if we annotate the ConvI2L node with
4762 // a type assertion that its value is known to be a small positive
4763 // number. (The prior range check has ensured this.)
4764 // This assertion is used by ConvI2LNode::Ideal.
4765 int index_max = max_jint - 1; // array size is max_jint, index is one less
4766 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4767 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4768 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4769 #endif
4770 return idx;
4771 }
4772
4773 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4774 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4775 if (ctrl != NULL) {
4776 // Express control dependency by a CastII node with a narrow type.
4777 value = new CastIINode(value, itype, false, true /* range check dependency */);
4778 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4779 // node from floating above the range check during loop optimizations. Otherwise, the
4780 // ConvI2L node may be eliminated independently of the range check, causing the data path
4781 // to become TOP while the control path is still there (although it's unreachable).
4782 value->set_req(0, ctrl);
4783 // Save CastII node to remove it after loop optimizations.
4784 phase->C->add_range_check_cast(value);
4785 value = phase->transform(value);
4786 }
4787 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4788 return phase->transform(new ConvI2LNode(value, ltype));
4789 }
4790
4791 // The message about the current inlining is accumulated in
4792 // _print_inlining_stream and transfered into the _print_inlining_list
4793 // once we know whether inlining succeeds or not. For regular
4794 // inlining, messages are appended to the buffer pointed by
4795 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4796 // a new buffer is added after _print_inlining_idx in the list. This
4797 // way we can update the inlining message for late inlining call site
4798 // when the inlining is attempted again.
4799 void Compile::print_inlining_init() {
4800 if (print_inlining() || print_intrinsics()) {
4801 _print_inlining_stream = new stringStream();
4802 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4803 }
4804 }
4805
4806 void Compile::print_inlining_reinit() {
4807 if (print_inlining() || print_intrinsics()) {
4808 // Re allocate buffer when we change ResourceMark
4809 _print_inlining_stream = new stringStream();
4810 }
4811 }
4812
4813 void Compile::print_inlining_reset() {
4814 _print_inlining_stream->reset();
4815 }
4816
4817 void Compile::print_inlining_commit() {
4818 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4819 // Transfer the message from _print_inlining_stream to the current
4820 // _print_inlining_list buffer and clear _print_inlining_stream.
4821 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4822 print_inlining_reset();
4823 }
4824
4825 void Compile::print_inlining_push() {
4826 // Add new buffer to the _print_inlining_list at current position
4827 _print_inlining_idx++;
4828 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4829 }
4830
4831 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4832 return _print_inlining_list->at(_print_inlining_idx);
4833 }
4834
4835 void Compile::print_inlining_update(CallGenerator* cg) {
4836 if (print_inlining() || print_intrinsics()) {
4837 if (!cg->is_late_inline()) {
4838 if (print_inlining_current().cg() != NULL) {
4839 print_inlining_push();
4840 }
4841 print_inlining_commit();
4842 } else {
4843 if (print_inlining_current().cg() != cg &&
4844 (print_inlining_current().cg() != NULL ||
4845 print_inlining_current().ss()->size() != 0)) {
4846 print_inlining_push();
4847 }
4848 print_inlining_commit();
4849 print_inlining_current().set_cg(cg);
4850 }
4851 }
4852 }
4853
4854 void Compile::print_inlining_move_to(CallGenerator* cg) {
4855 // We resume inlining at a late inlining call site. Locate the
4856 // corresponding inlining buffer so that we can update it.
4857 if (print_inlining()) {
4858 for (int i = 0; i < _print_inlining_list->length(); i++) {
4859 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4860 _print_inlining_idx = i;
4861 return;
4862 }
4863 }
4864 ShouldNotReachHere();
4865 }
4866 }
4867
4868 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4869 if (print_inlining()) {
4870 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4871 assert(print_inlining_current().cg() == cg, "wrong entry");
4872 // replace message with new message
4873 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4874 print_inlining_commit();
4875 print_inlining_current().set_cg(cg);
4876 }
4877 }
4878
4879 void Compile::print_inlining_assert_ready() {
4880 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4881 }
4882
4883 void Compile::process_print_inlining() {
4884 bool do_print_inlining = print_inlining() || print_intrinsics();
4885 if (do_print_inlining || log() != NULL) {
4886 // Print inlining message for candidates that we couldn't inline
4887 // for lack of space
4888 for (int i = 0; i < _late_inlines.length(); i++) {
4889 CallGenerator* cg = _late_inlines.at(i);
4890 if (!cg->is_mh_late_inline()) {
4891 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4892 if (do_print_inlining) {
4893 cg->print_inlining_late(msg);
4894 }
4895 log_late_inline_failure(cg, msg);
4896 }
4897 }
4898 }
4899 if (do_print_inlining) {
4900 ResourceMark rm;
4901 stringStream ss;
4902 for (int i = 0; i < _print_inlining_list->length(); i++) {
4903 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4904 }
4905 size_t end = ss.size();
4906 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4907 strncpy(_print_inlining_output, ss.base(), end+1);
4908 _print_inlining_output[end] = 0;
4909 }
4910 }
4911
4912 void Compile::dump_print_inlining() {
4913 if (_print_inlining_output != NULL) {
4914 tty->print_raw(_print_inlining_output);
4915 }
4916 }
4917
4918 void Compile::log_late_inline(CallGenerator* cg) {
4919 if (log() != NULL) {
4920 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4921 cg->unique_id());
4922 JVMState* p = cg->call_node()->jvms();
4923 while (p != NULL) {
4924 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4925 p = p->caller();
4926 }
4927 log()->tail("late_inline");
4928 }
4929 }
4930
4931 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4932 log_late_inline(cg);
4933 if (log() != NULL) {
4934 log()->inline_fail(msg);
4935 }
4936 }
4937
4938 void Compile::log_inline_id(CallGenerator* cg) {
4939 if (log() != NULL) {
4940 // The LogCompilation tool needs a unique way to identify late
4941 // inline call sites. This id must be unique for this call site in
4942 // this compilation. Try to have it unique across compilations as
4943 // well because it can be convenient when grepping through the log
4944 // file.
4945 // Distinguish OSR compilations from others in case CICountOSR is
4946 // on.
4947 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4948 cg->set_unique_id(id);
4949 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4950 }
4951 }
4952
4953 void Compile::log_inline_failure(const char* msg) {
4954 if (C->log() != NULL) {
4955 C->log()->inline_fail(msg);
4956 }
4957 }
4958
4959
4960 // Dump inlining replay data to the stream.
4961 // Don't change thread state and acquire any locks.
4962 void Compile::dump_inline_data(outputStream* out) {
4963 InlineTree* inl_tree = ilt();
4964 if (inl_tree != NULL) {
4965 out->print(" inline %d", inl_tree->count());
4966 inl_tree->dump_replay_data(out);
4967 }
4968 }
4969
4970 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4971 if (n1->Opcode() < n2->Opcode()) return -1;
4972 else if (n1->Opcode() > n2->Opcode()) return 1;
4973
4974 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4975 for (uint i = 1; i < n1->req(); i++) {
4976 if (n1->in(i) < n2->in(i)) return -1;
4977 else if (n1->in(i) > n2->in(i)) return 1;
4978 }
4979
4980 return 0;
4981 }
4982
4983 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4984 Node* n1 = *n1p;
4985 Node* n2 = *n2p;
4986
4987 return cmp_expensive_nodes(n1, n2);
4988 }
4989
4990 void Compile::sort_expensive_nodes() {
4991 if (!expensive_nodes_sorted()) {
4992 _expensive_nodes->sort(cmp_expensive_nodes);
4993 }
4994 }
4995
4996 bool Compile::expensive_nodes_sorted() const {
4997 for (int i = 1; i < _expensive_nodes->length(); i++) {
4998 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4999 return false;
5000 }
5001 }
5002 return true;
5003 }
5004
5005 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
5006 if (_expensive_nodes->length() == 0) {
5007 return false;
5008 }
5009
5010 assert(OptimizeExpensiveOps, "optimization off?");
5011
5012 // Take this opportunity to remove dead nodes from the list
5013 int j = 0;
5014 for (int i = 0; i < _expensive_nodes->length(); i++) {
5015 Node* n = _expensive_nodes->at(i);
5016 if (!n->is_unreachable(igvn)) {
5017 assert(n->is_expensive(), "should be expensive");
5018 _expensive_nodes->at_put(j, n);
5019 j++;
5020 }
5021 }
5022 _expensive_nodes->trunc_to(j);
5023
5024 // Then sort the list so that similar nodes are next to each other
5025 // and check for at least two nodes of identical kind with same data
5026 // inputs.
5027 sort_expensive_nodes();
5028
5029 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
5030 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
5031 return true;
5032 }
5033 }
5034
5035 return false;
5036 }
5037
5038 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
5039 if (_expensive_nodes->length() == 0) {
5040 return;
5041 }
5042
5043 assert(OptimizeExpensiveOps, "optimization off?");
5044
5045 // Sort to bring similar nodes next to each other and clear the
5046 // control input of nodes for which there's only a single copy.
5047 sort_expensive_nodes();
5048
5049 int j = 0;
5050 int identical = 0;
5051 int i = 0;
5052 bool modified = false;
5053 for (; i < _expensive_nodes->length()-1; i++) {
5054 assert(j <= i, "can't write beyond current index");
5055 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
5056 identical++;
5057 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5058 continue;
5059 }
5060 if (identical > 0) {
5061 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5062 identical = 0;
5063 } else {
5064 Node* n = _expensive_nodes->at(i);
5065 igvn.replace_input_of(n, 0, NULL);
5066 igvn.hash_insert(n);
5067 modified = true;
5068 }
5069 }
5070 if (identical > 0) {
5071 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5072 } else if (_expensive_nodes->length() >= 1) {
5073 Node* n = _expensive_nodes->at(i);
5074 igvn.replace_input_of(n, 0, NULL);
5075 igvn.hash_insert(n);
5076 modified = true;
5077 }
5078 _expensive_nodes->trunc_to(j);
5079 if (modified) {
5080 igvn.optimize();
5081 }
5082 }
5083
5084 void Compile::add_expensive_node(Node * n) {
5085 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
5086 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5087 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5088 if (OptimizeExpensiveOps) {
5089 _expensive_nodes->append(n);
5090 } else {
5091 // Clear control input and let IGVN optimize expensive nodes if
5092 // OptimizeExpensiveOps is off.
5093 n->set_req(0, NULL);
5094 }
5095 }
5096
5097 /**
5098 * Remove the speculative part of types and clean up the graph
5099 */
5100 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5101 if (UseTypeSpeculation) {
5102 Unique_Node_List worklist;
5103 worklist.push(root());
5104 int modified = 0;
5105 // Go over all type nodes that carry a speculative type, drop the
5106 // speculative part of the type and enqueue the node for an igvn
5107 // which may optimize it out.
5108 for (uint next = 0; next < worklist.size(); ++next) {
5109 Node *n = worklist.at(next);
5110 if (n->is_Type()) {
5111 TypeNode* tn = n->as_Type();
5112 const Type* t = tn->type();
5113 const Type* t_no_spec = t->remove_speculative();
5114 if (t_no_spec != t) {
5115 bool in_hash = igvn.hash_delete(n);
5116 assert(in_hash, "node should be in igvn hash table");
5117 tn->set_type(t_no_spec);
5118 igvn.hash_insert(n);
5119 igvn._worklist.push(n); // give it a chance to go away
5120 modified++;
5121 }
5122 }
5123 uint max = n->len();
5124 for( uint i = 0; i < max; ++i ) {
5125 Node *m = n->in(i);
5126 if (not_a_node(m)) continue;
5127 worklist.push(m);
5128 }
5129 }
5130 // Drop the speculative part of all types in the igvn's type table
5131 igvn.remove_speculative_types();
5132 if (modified > 0) {
5133 igvn.optimize();
5134 }
5135 #ifdef ASSERT
5136 // Verify that after the IGVN is over no speculative type has resurfaced
5137 worklist.clear();
5138 worklist.push(root());
5139 for (uint next = 0; next < worklist.size(); ++next) {
5140 Node *n = worklist.at(next);
5141 const Type* t = igvn.type_or_null(n);
5142 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5143 if (n->is_Type()) {
5144 t = n->as_Type()->type();
5145 assert(t == t->remove_speculative(), "no more speculative types");
5146 }
5147 uint max = n->len();
5148 for( uint i = 0; i < max; ++i ) {
5149 Node *m = n->in(i);
5150 if (not_a_node(m)) continue;
5151 worklist.push(m);
5152 }
5153 }
5154 igvn.check_no_speculative_types();
5155 #endif
5156 }
5157 }
5158
5159 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5160 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5161 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5162 if (!EnableValhalla || ta == NULL || tb == NULL ||
5163 ta->is_zero_type() || tb->is_zero_type() ||
5164 !ta->can_be_value_type() || !tb->can_be_value_type()) {
5165 // Use old acmp if one operand is null or not a value type
5166 return new CmpPNode(a, b);
5167 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) {
5168 // We know that one operand is a value type. Therefore,
5169 // new acmp will only return true if both operands are NULL.
5170 // Check if both operands are null by or'ing the oops.
5171 a = phase->transform(new CastP2XNode(NULL, a));
5172 b = phase->transform(new CastP2XNode(NULL, b));
5173 a = phase->transform(new OrXNode(a, b));
5174 return new CmpXNode(a, phase->MakeConX(0));
5175 }
5176 // Use new acmp
5177 return NULL;
5178 }
5179
5180 // Auxiliary method to support randomized stressing/fuzzing.
5181 //
5182 // This method can be called the arbitrary number of times, with current count
5183 // as the argument. The logic allows selecting a single candidate from the
5184 // running list of candidates as follows:
5185 // int count = 0;
5186 // Cand* selected = null;
5187 // while(cand = cand->next()) {
5188 // if (randomized_select(++count)) {
5189 // selected = cand;
5190 // }
5191 // }
5192 //
5193 // Including count equalizes the chances any candidate is "selected".
5194 // This is useful when we don't have the complete list of candidates to choose
5195 // from uniformly. In this case, we need to adjust the randomicity of the
5196 // selection, or else we will end up biasing the selection towards the latter
5197 // candidates.
5198 //
5199 // Quick back-envelope calculation shows that for the list of n candidates
5200 // the equal probability for the candidate to persist as "best" can be
5201 // achieved by replacing it with "next" k-th candidate with the probability
5202 // of 1/k. It can be easily shown that by the end of the run, the
5203 // probability for any candidate is converged to 1/n, thus giving the
5204 // uniform distribution among all the candidates.
5205 //
5206 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5207 #define RANDOMIZED_DOMAIN_POW 29
5208 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5209 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5210 bool Compile::randomized_select(int count) {
5211 assert(count > 0, "only positive");
5212 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5213 }
5214
5215 CloneMap& Compile::clone_map() { return _clone_map; }
5216 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5217
5218 void NodeCloneInfo::dump() const {
5219 tty->print(" {%d:%d} ", idx(), gen());
5220 }
5221
5222 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5223 uint64_t val = value(old->_idx);
5224 NodeCloneInfo cio(val);
5225 assert(val != 0, "old node should be in the map");
5226 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5227 insert(nnn->_idx, cin.get());
5228 #ifndef PRODUCT
5229 if (is_debug()) {
5230 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5231 }
5232 #endif
5233 }
5234
5235 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5236 NodeCloneInfo cio(value(old->_idx));
5237 if (cio.get() == 0) {
5238 cio.set(old->_idx, 0);
5239 insert(old->_idx, cio.get());
5240 #ifndef PRODUCT
5241 if (is_debug()) {
5242 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5243 }
5244 #endif
5245 }
5246 clone(old, nnn, gen);
5247 }
5248
5249 int CloneMap::max_gen() const {
5250 int g = 0;
5251 DictI di(_dict);
5252 for(; di.test(); ++di) {
5253 int t = gen(di._key);
5254 if (g < t) {
5255 g = t;
5256 #ifndef PRODUCT
5257 if (is_debug()) {
5258 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5259 }
5260 #endif
5261 }
5262 }
5263 return g;
5264 }
5265
5266 void CloneMap::dump(node_idx_t key) const {
5267 uint64_t val = value(key);
5268 if (val != 0) {
5269 NodeCloneInfo ni(val);
5270 ni.dump();
5271 }
5272 }