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