1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vectornode.hpp"
71 #include "runtime/arguments.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stubRoutines.hpp"
75 #include "runtime/timer.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/copy.hpp"
78 #include "utilities/macros.hpp"
79 #if INCLUDE_G1GC
80 #include "gc/g1/g1ThreadLocalData.hpp"
81 #endif // INCLUDE_G1GC
82 #if INCLUDE_ZGC
83 #include "gc/z/c2/zBarrierSetC2.hpp"
84 #endif
85
86
87 // -------------------- Compile::mach_constant_base_node -----------------------
88 // Constant table base node singleton.
89 MachConstantBaseNode* Compile::mach_constant_base_node() {
90 if (_mach_constant_base_node == NULL) {
91 _mach_constant_base_node = new MachConstantBaseNode();
92 _mach_constant_base_node->add_req(C->root());
93 }
94 return _mach_constant_base_node;
95 }
96
97
98 /// Support for intrinsics.
99
100 // Return the index at which m must be inserted (or already exists).
101 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
102 class IntrinsicDescPair {
103 private:
104 ciMethod* _m;
105 bool _is_virtual;
106 public:
107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
109 ciMethod* m= elt->method();
110 ciMethod* key_m = key->_m;
111 if (key_m < m) return -1;
112 else if (key_m > m) return 1;
113 else {
114 bool is_virtual = elt->is_virtual();
115 bool key_virtual = key->_is_virtual;
116 if (key_virtual < is_virtual) return -1;
117 else if (key_virtual > is_virtual) return 1;
118 else return 0;
119 }
120 }
121 };
122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
123 #ifdef ASSERT
124 for (int i = 1; i < _intrinsics->length(); i++) {
125 CallGenerator* cg1 = _intrinsics->at(i-1);
126 CallGenerator* cg2 = _intrinsics->at(i);
127 assert(cg1->method() != cg2->method()
128 ? cg1->method() < cg2->method()
129 : cg1->is_virtual() < cg2->is_virtual(),
130 "compiler intrinsics list must stay sorted");
131 }
132 #endif
133 IntrinsicDescPair pair(m, is_virtual);
134 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
135 }
136
137 void Compile::register_intrinsic(CallGenerator* cg) {
138 if (_intrinsics == NULL) {
139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
140 }
141 int len = _intrinsics->length();
142 bool found = false;
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
144 assert(!found, "registering twice");
145 _intrinsics->insert_before(index, cg);
146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
147 }
148
149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
150 assert(m->is_loaded(), "don't try this on unloaded methods");
151 if (_intrinsics != NULL) {
152 bool found = false;
153 int index = intrinsic_insertion_index(m, is_virtual, found);
154 if (found) {
155 return _intrinsics->at(index);
156 }
157 }
158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
159 if (m->intrinsic_id() != vmIntrinsics::_none &&
160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
162 if (cg != NULL) {
163 // Save it for next time:
164 register_intrinsic(cg);
165 return cg;
166 } else {
167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
168 }
169 }
170 return NULL;
171 }
172
173 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
174 // in library_call.cpp.
175
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
182
183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
185 int oflags = _intrinsic_hist_flags[id];
186 assert(flags != 0, "what happened?");
187 if (is_virtual) {
188 flags |= _intrinsic_virtual;
189 }
190 bool changed = (flags != oflags);
191 if ((flags & _intrinsic_worked) != 0) {
192 juint count = (_intrinsic_hist_count[id] += 1);
193 if (count == 1) {
194 changed = true; // first time
195 }
196 // increment the overall count also:
197 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
198 }
199 if (changed) {
200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
201 // Something changed about the intrinsic's virtuality.
202 if ((flags & _intrinsic_virtual) != 0) {
203 // This is the first use of this intrinsic as a virtual call.
204 if (oflags != 0) {
205 // We already saw it as a non-virtual, so note both cases.
206 flags |= _intrinsic_both;
207 }
208 } else if ((oflags & _intrinsic_both) == 0) {
209 // This is the first use of this intrinsic as a non-virtual
210 flags |= _intrinsic_both;
211 }
212 }
213 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
214 }
215 // update the overall flags also:
216 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
217 return changed;
218 }
219
220 static char* format_flags(int flags, char* buf) {
221 buf[0] = 0;
222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
227 if (buf[0] == 0) strcat(buf, ",");
228 assert(buf[0] == ',', "must be");
229 return &buf[1];
230 }
231
232 void Compile::print_intrinsic_statistics() {
233 char flagsbuf[100];
234 ttyLocker ttyl;
235 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
236 tty->print_cr("Compiler intrinsic usage:");
237 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
238 if (total == 0) total = 1; // avoid div0 in case of no successes
239 #define PRINT_STAT_LINE(name, c, f) \
240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
241 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
242 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
243 int flags = _intrinsic_hist_flags[id];
244 juint count = _intrinsic_hist_count[id];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
250 if (xtty != NULL) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != NULL) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseCCP::print_statistics();
258 PhaseRegAlloc::print_statistics();
259 Scheduling::print_statistics();
260 PhasePeephole::print_statistics();
261 PhaseIdealLoop::print_statistics();
262 if (xtty != NULL) xtty->tail("statistics");
263 }
264 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
265 // put this under its own <statistics> element.
266 print_intrinsic_statistics();
267 }
268 }
269 #endif //PRODUCT
270
271 // Support for bundling info
272 Bundle* Compile::node_bundling(const Node *n) {
273 assert(valid_bundle_info(n), "oob");
274 return &_node_bundling_base[n->_idx];
275 }
276
277 bool Compile::valid_bundle_info(const Node *n) {
278 return (_node_bundling_limit > n->_idx);
279 }
280
281
282 void Compile::gvn_replace_by(Node* n, Node* nn) {
283 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
284 Node* use = n->last_out(i);
285 bool is_in_table = initial_gvn()->hash_delete(use);
286 uint uses_found = 0;
287 for (uint j = 0; j < use->len(); j++) {
288 if (use->in(j) == n) {
289 if (j < use->req())
290 use->set_req(j, nn);
291 else
292 use->set_prec(j, nn);
293 uses_found++;
294 }
295 }
296 if (is_in_table) {
297 // reinsert into table
298 initial_gvn()->hash_find_insert(use);
299 }
300 record_for_igvn(use);
301 i -= uses_found; // we deleted 1 or more copies of this edge
302 }
303 }
304
305
306 static inline bool not_a_node(const Node* n) {
307 if (n == NULL) return true;
308 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
309 if (*(address*)n == badAddress) return true; // kill by Node::destruct
310 return false;
311 }
312
313 // Identify all nodes that are reachable from below, useful.
314 // Use breadth-first pass that records state in a Unique_Node_List,
315 // recursive traversal is slower.
316 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
317 int estimated_worklist_size = live_nodes();
318 useful.map( estimated_worklist_size, NULL ); // preallocate space
319
320 // Initialize worklist
321 if (root() != NULL) { useful.push(root()); }
322 // If 'top' is cached, declare it useful to preserve cached node
323 if( cached_top_node() ) { useful.push(cached_top_node()); }
324
325 // Push all useful nodes onto the list, breadthfirst
326 for( uint next = 0; next < useful.size(); ++next ) {
327 assert( next < unique(), "Unique useful nodes < total nodes");
328 Node *n = useful.at(next);
329 uint max = n->len();
330 for( uint i = 0; i < max; ++i ) {
331 Node *m = n->in(i);
332 if (not_a_node(m)) continue;
333 useful.push(m);
334 }
335 }
336 }
337
338 // Update dead_node_list with any missing dead nodes using useful
339 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
340 void Compile::update_dead_node_list(Unique_Node_List &useful) {
341 uint max_idx = unique();
342 VectorSet& useful_node_set = useful.member_set();
343
344 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
345 // If node with index node_idx is not in useful set,
346 // mark it as dead in dead node list.
347 if (! useful_node_set.test(node_idx) ) {
348 record_dead_node(node_idx);
349 }
350 }
351 }
352
353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
354 int shift = 0;
355 for (int i = 0; i < inlines->length(); i++) {
356 CallGenerator* cg = inlines->at(i);
357 CallNode* call = cg->call_node();
358 if (shift > 0) {
359 inlines->at_put(i-shift, cg);
360 }
361 if (!useful.member(call)) {
362 shift++;
363 }
364 }
365 inlines->trunc_to(inlines->length()-shift);
366 }
367
368 // Disconnect all useless nodes by disconnecting those at the boundary.
369 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
370 uint next = 0;
371 while (next < useful.size()) {
372 Node *n = useful.at(next++);
373 if (n->is_SafePoint()) {
374 // We're done with a parsing phase. Replaced nodes are not valid
375 // beyond that point.
376 n->as_SafePoint()->delete_replaced_nodes();
377 }
378 // Use raw traversal of out edges since this code removes out edges
379 int max = n->outcnt();
380 for (int j = 0; j < max; ++j) {
381 Node* child = n->raw_out(j);
382 if (! useful.member(child)) {
383 assert(!child->is_top() || child != top(),
384 "If top is cached in Compile object it is in useful list");
385 // Only need to remove this out-edge to the useless node
386 n->raw_del_out(j);
387 --j;
388 --max;
389 }
390 }
391 if (n->outcnt() == 1 && n->has_special_unique_user()) {
392 record_for_igvn(n->unique_out());
393 }
394 }
395 // Remove useless macro and predicate opaq nodes
396 for (int i = C->macro_count()-1; i >= 0; i--) {
397 Node* n = C->macro_node(i);
398 if (!useful.member(n)) {
399 remove_macro_node(n);
400 }
401 }
402 // Remove useless CastII nodes with range check dependency
403 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
404 Node* cast = range_check_cast_node(i);
405 if (!useful.member(cast)) {
406 remove_range_check_cast(cast);
407 }
408 }
409 // Remove useless expensive nodes
410 for (int i = C->expensive_count()-1; i >= 0; i--) {
411 Node* n = C->expensive_node(i);
412 if (!useful.member(n)) {
413 remove_expensive_node(n);
414 }
415 }
416 // Remove useless Opaque4 nodes
417 for (int i = opaque4_count() - 1; i >= 0; i--) {
418 Node* opaq = opaque4_node(i);
419 if (!useful.member(opaq)) {
420 remove_opaque4_node(opaq);
421 }
422 }
423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
424 bs->eliminate_useless_gc_barriers(useful, this);
425 // clean up the late inline lists
426 remove_useless_late_inlines(&_string_late_inlines, useful);
427 remove_useless_late_inlines(&_boxing_late_inlines, useful);
428 remove_useless_late_inlines(&_late_inlines, useful);
429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
430 }
431
432 //------------------------------frame_size_in_words-----------------------------
433 // frame_slots in units of words
434 int Compile::frame_size_in_words() const {
435 // shift is 0 in LP32 and 1 in LP64
436 const int shift = (LogBytesPerWord - LogBytesPerInt);
437 int words = _frame_slots >> shift;
438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
439 return words;
440 }
441
442 // To bang the stack of this compiled method we use the stack size
443 // that the interpreter would need in case of a deoptimization. This
444 // removes the need to bang the stack in the deoptimization blob which
445 // in turn simplifies stack overflow handling.
446 int Compile::bang_size_in_bytes() const {
447 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
448 }
449
450 // ============================================================================
451 //------------------------------CompileWrapper---------------------------------
452 class CompileWrapper : public StackObj {
453 Compile *const _compile;
454 public:
455 CompileWrapper(Compile* compile);
456
457 ~CompileWrapper();
458 };
459
460 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
461 // the Compile* pointer is stored in the current ciEnv:
462 ciEnv* env = compile->env();
463 assert(env == ciEnv::current(), "must already be a ciEnv active");
464 assert(env->compiler_data() == NULL, "compile already active?");
465 env->set_compiler_data(compile);
466 assert(compile == Compile::current(), "sanity");
467
468 compile->set_type_dict(NULL);
469 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
470 compile->clone_map().set_clone_idx(0);
471 compile->set_type_hwm(NULL);
472 compile->set_type_last_size(0);
473 compile->set_last_tf(NULL, NULL);
474 compile->set_indexSet_arena(NULL);
475 compile->set_indexSet_free_block_list(NULL);
476 compile->init_type_arena();
477 Type::Initialize(compile);
478 _compile->set_scratch_buffer_blob(NULL);
479 _compile->begin_method();
480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
481 }
482 CompileWrapper::~CompileWrapper() {
483 _compile->end_method();
484 if (_compile->scratch_buffer_blob() != NULL)
485 BufferBlob::free(_compile->scratch_buffer_blob());
486 _compile->env()->set_compiler_data(NULL);
487 }
488
489
490 //----------------------------print_compile_messages---------------------------
491 void Compile::print_compile_messages() {
492 #ifndef PRODUCT
493 // Check if recompiling
494 if (_subsume_loads == false && PrintOpto) {
495 // Recompiling without allowing machine instructions to subsume loads
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without subsuming loads **");
498 tty->print_cr("*********************************************************");
499 }
500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
501 // Recompiling without escape analysis
502 tty->print_cr("*********************************************************");
503 tty->print_cr("** Bailout: Recompile without escape analysis **");
504 tty->print_cr("*********************************************************");
505 }
506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
507 // Recompiling without boxing elimination
508 tty->print_cr("*********************************************************");
509 tty->print_cr("** Bailout: Recompile without boxing elimination **");
510 tty->print_cr("*********************************************************");
511 }
512 if (C->directive()->BreakAtCompileOption) {
513 // Open the debugger when compiling this method.
514 tty->print("### Breaking when compiling: ");
515 method()->print_short_name();
516 tty->cr();
517 BREAKPOINT;
518 }
519
520 if( PrintOpto ) {
521 if (is_osr_compilation()) {
522 tty->print("[OSR]%3d", _compile_id);
523 } else {
524 tty->print("%3d", _compile_id);
525 }
526 }
527 #endif
528 }
529
530
531 //-----------------------init_scratch_buffer_blob------------------------------
532 // Construct a temporary BufferBlob and cache it for this compile.
533 void Compile::init_scratch_buffer_blob(int const_size) {
534 // If there is already a scratch buffer blob allocated and the
535 // constant section is big enough, use it. Otherwise free the
536 // current and allocate a new one.
537 BufferBlob* blob = scratch_buffer_blob();
538 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
539 // Use the current blob.
540 } else {
541 if (blob != NULL) {
542 BufferBlob::free(blob);
543 }
544
545 ResourceMark rm;
546 _scratch_const_size = const_size;
547 int size = C2Compiler::initial_code_buffer_size(const_size);
548 blob = BufferBlob::create("Compile::scratch_buffer", size);
549 // Record the buffer blob for next time.
550 set_scratch_buffer_blob(blob);
551 // Have we run out of code space?
552 if (scratch_buffer_blob() == NULL) {
553 // Let CompilerBroker disable further compilations.
554 record_failure("Not enough space for scratch buffer in CodeCache");
555 return;
556 }
557 }
558
559 // Initialize the relocation buffers
560 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
561 set_scratch_locs_memory(locs_buf);
562 }
563
564
565 //-----------------------scratch_emit_size-------------------------------------
566 // Helper function that computes size by emitting code
567 uint Compile::scratch_emit_size(const Node* n) {
568 // Start scratch_emit_size section.
569 set_in_scratch_emit_size(true);
570
571 // Emit into a trash buffer and count bytes emitted.
572 // This is a pretty expensive way to compute a size,
573 // but it works well enough if seldom used.
574 // All common fixed-size instructions are given a size
575 // method by the AD file.
576 // Note that the scratch buffer blob and locs memory are
577 // allocated at the beginning of the compile task, and
578 // may be shared by several calls to scratch_emit_size.
579 // The allocation of the scratch buffer blob is particularly
580 // expensive, since it has to grab the code cache lock.
581 BufferBlob* blob = this->scratch_buffer_blob();
582 assert(blob != NULL, "Initialize BufferBlob at start");
583 assert(blob->size() > MAX_inst_size, "sanity");
584 relocInfo* locs_buf = scratch_locs_memory();
585 address blob_begin = blob->content_begin();
586 address blob_end = (address)locs_buf;
587 assert(blob->contains(blob_end), "sanity");
588 CodeBuffer buf(blob_begin, blob_end - blob_begin);
589 buf.initialize_consts_size(_scratch_const_size);
590 buf.initialize_stubs_size(MAX_stubs_size);
591 assert(locs_buf != NULL, "sanity");
592 int lsize = MAX_locs_size / 3;
593 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
594 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
595 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
596 // Mark as scratch buffer.
597 buf.consts()->set_scratch_emit();
598 buf.insts()->set_scratch_emit();
599 buf.stubs()->set_scratch_emit();
600
601 // Do the emission.
602
603 Label fakeL; // Fake label for branch instructions.
604 Label* saveL = NULL;
605 uint save_bnum = 0;
606 bool is_branch = n->is_MachBranch();
607 if (is_branch) {
608 MacroAssembler masm(&buf);
609 masm.bind(fakeL);
610 n->as_MachBranch()->save_label(&saveL, &save_bnum);
611 n->as_MachBranch()->label_set(&fakeL, 0);
612 }
613 n->emit(buf, this->regalloc());
614
615 // Emitting into the scratch buffer should not fail
616 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
617
618 if (is_branch) // Restore label.
619 n->as_MachBranch()->label_set(saveL, save_bnum);
620
621 // End scratch_emit_size section.
622 set_in_scratch_emit_size(false);
623
624 return buf.insts_size();
625 }
626
627
628 // ============================================================================
629 //------------------------------Compile standard-------------------------------
630 debug_only( int Compile::_debug_idx = 100000; )
631
632 // Compile a method. entry_bci is -1 for normal compilations and indicates
633 // the continuation bci for on stack replacement.
634
635
636 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
637 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
638 : Phase(Compiler),
639 _compile_id(ci_env->compile_id()),
640 _save_argument_registers(false),
641 _subsume_loads(subsume_loads),
642 _do_escape_analysis(do_escape_analysis),
643 _eliminate_boxing(eliminate_boxing),
644 _method(target),
645 _entry_bci(osr_bci),
646 _stub_function(NULL),
647 _stub_name(NULL),
648 _stub_entry_point(NULL),
649 _max_node_limit(MaxNodeLimit),
650 _orig_pc_slot(0),
651 _orig_pc_slot_offset_in_bytes(0),
652 _inlining_progress(false),
653 _inlining_incrementally(false),
654 _has_reserved_stack_access(target->has_reserved_stack_access()),
655 #ifndef PRODUCT
656 _trace_opto_output(directive->TraceOptoOutputOption),
657 #endif
658 _has_method_handle_invokes(false),
659 _comp_arena(mtCompiler),
660 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
661 _env(ci_env),
662 _directive(directive),
663 _log(ci_env->log()),
664 _failure_reason(NULL),
665 _congraph(NULL),
666 #ifndef PRODUCT
667 _printer(IdealGraphPrinter::printer()),
668 #endif
669 _dead_node_list(comp_arena()),
670 _dead_node_count(0),
671 _node_arena(mtCompiler),
672 _old_arena(mtCompiler),
673 _mach_constant_base_node(NULL),
674 _Compile_types(mtCompiler),
675 _initial_gvn(NULL),
676 _for_igvn(NULL),
677 _warm_calls(NULL),
678 _late_inlines(comp_arena(), 2, 0, NULL),
679 _string_late_inlines(comp_arena(), 2, 0, NULL),
680 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
681 _late_inlines_pos(0),
682 _number_of_mh_late_inlines(0),
683 _print_inlining_stream(NULL),
684 _print_inlining_list(NULL),
685 _print_inlining_idx(0),
686 _print_inlining_output(NULL),
687 _replay_inline_data(NULL),
688 _java_calls(0),
689 _inner_loops(0),
690 _interpreter_frame_size(0),
691 _node_bundling_limit(0),
692 _node_bundling_base(NULL),
693 _code_buffer("Compile::Fill_buffer"),
694 _scratch_const_size(-1),
695 _in_scratch_emit_size(false)
696 #ifndef PRODUCT
697 , _in_dump_cnt(0)
698 #endif
699 {
700 C = this;
701 #ifndef PRODUCT
702 if (_printer != NULL) {
703 _printer->set_compile(this);
704 }
705 #endif
706 CompileWrapper cw(this);
707
708 if (CITimeVerbose) {
709 tty->print(" ");
710 target->holder()->name()->print();
711 tty->print(".");
712 target->print_short_name();
713 tty->print(" ");
714 }
715 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
716 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
717
718 #ifndef PRODUCT
719 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
720 if (!print_opto_assembly) {
721 bool print_assembly = directive->PrintAssemblyOption;
722 if (print_assembly && !Disassembler::can_decode()) {
723 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
724 print_opto_assembly = true;
725 }
726 }
727 set_print_assembly(print_opto_assembly);
728 set_parsed_irreducible_loop(false);
729
730 if (directive->ReplayInlineOption) {
731 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
732 }
733 #endif
734 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
735 set_print_intrinsics(directive->PrintIntrinsicsOption);
736 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
737
738 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
739 // Make sure the method being compiled gets its own MDO,
740 // so we can at least track the decompile_count().
741 // Need MDO to record RTM code generation state.
742 method()->ensure_method_data();
743 }
744
745 Init(::AliasLevel);
746
747
748 print_compile_messages();
749
750 _ilt = InlineTree::build_inline_tree_root();
751
752 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
753 assert(num_alias_types() >= AliasIdxRaw, "");
754
755 #define MINIMUM_NODE_HASH 1023
756 // Node list that Iterative GVN will start with
757 Unique_Node_List for_igvn(comp_arena());
758 set_for_igvn(&for_igvn);
759
760 // GVN that will be run immediately on new nodes
761 uint estimated_size = method()->code_size()*4+64;
762 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
763 PhaseGVN gvn(node_arena(), estimated_size);
764 set_initial_gvn(&gvn);
765
766 print_inlining_init();
767 { // Scope for timing the parser
768 TracePhase tp("parse", &timers[_t_parser]);
769
770 // Put top into the hash table ASAP.
771 initial_gvn()->transform_no_reclaim(top());
772
773 // Set up tf(), start(), and find a CallGenerator.
774 CallGenerator* cg = NULL;
775 if (is_osr_compilation()) {
776 const TypeTuple *domain = StartOSRNode::osr_domain();
777 const TypeTuple *range = TypeTuple::make_range(method()->signature());
778 init_tf(TypeFunc::make(domain, range));
779 StartNode* s = new StartOSRNode(root(), domain);
780 initial_gvn()->set_type_bottom(s);
781 init_start(s);
782 cg = CallGenerator::for_osr(method(), entry_bci());
783 } else {
784 // Normal case.
785 init_tf(TypeFunc::make(method()));
786 StartNode* s = new StartNode(root(), tf()->domain());
787 initial_gvn()->set_type_bottom(s);
788 init_start(s);
789 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
790 // With java.lang.ref.reference.get() we must go through the
791 // intrinsic - even when get() is the root
792 // method of the compile - so that, if necessary, the value in
793 // the referent field of the reference object gets recorded by
794 // the pre-barrier code.
795 cg = find_intrinsic(method(), false);
796 }
797 if (cg == NULL) {
798 float past_uses = method()->interpreter_invocation_count();
799 float expected_uses = past_uses;
800 cg = CallGenerator::for_inline(method(), expected_uses);
801 }
802 }
803 if (failing()) return;
804 if (cg == NULL) {
805 record_method_not_compilable("cannot parse method");
806 return;
807 }
808 JVMState* jvms = build_start_state(start(), tf());
809 if ((jvms = cg->generate(jvms)) == NULL) {
810 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
811 record_method_not_compilable("method parse failed");
812 }
813 return;
814 }
815 GraphKit kit(jvms);
816
817 if (!kit.stopped()) {
818 // Accept return values, and transfer control we know not where.
819 // This is done by a special, unique ReturnNode bound to root.
820 return_values(kit.jvms());
821 }
822
823 if (kit.has_exceptions()) {
824 // Any exceptions that escape from this call must be rethrown
825 // to whatever caller is dynamically above us on the stack.
826 // This is done by a special, unique RethrowNode bound to root.
827 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
828 }
829
830 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
831
832 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
833 inline_string_calls(true);
834 }
835
836 if (failing()) return;
837
838 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
839
840 // Remove clutter produced by parsing.
841 if (!failing()) {
842 ResourceMark rm;
843 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
844 }
845 }
846
847 // Note: Large methods are capped off in do_one_bytecode().
848 if (failing()) return;
849
850 // After parsing, node notes are no longer automagic.
851 // They must be propagated by register_new_node_with_optimizer(),
852 // clone(), or the like.
853 set_default_node_notes(NULL);
854
855 for (;;) {
856 int successes = Inline_Warm();
857 if (failing()) return;
858 if (successes == 0) break;
859 }
860
861 // Drain the list.
862 Finish_Warm();
863 #ifndef PRODUCT
864 if (_printer && _printer->should_print(1)) {
865 _printer->print_inlining();
866 }
867 #endif
868
869 if (failing()) return;
870 NOT_PRODUCT( verify_graph_edges(); )
871
872 // Now optimize
873 Optimize();
874 if (failing()) return;
875 NOT_PRODUCT( verify_graph_edges(); )
876
877 #ifndef PRODUCT
878 if (PrintIdeal) {
879 ttyLocker ttyl; // keep the following output all in one block
880 // This output goes directly to the tty, not the compiler log.
881 // To enable tools to match it up with the compilation activity,
882 // be sure to tag this tty output with the compile ID.
883 if (xtty != NULL) {
884 xtty->head("ideal compile_id='%d'%s", compile_id(),
885 is_osr_compilation() ? " compile_kind='osr'" :
886 "");
887 }
888 root()->dump(9999);
889 if (xtty != NULL) {
890 xtty->tail("ideal");
891 }
892 }
893 #endif
894
895 #ifdef ASSERT
896 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
897 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
898 #endif
899
900 // Dump compilation data to replay it.
901 if (directive->DumpReplayOption) {
902 env()->dump_replay_data(_compile_id);
903 }
904 if (directive->DumpInlineOption && (ilt() != NULL)) {
905 env()->dump_inline_data(_compile_id);
906 }
907
908 // Now that we know the size of all the monitors we can add a fixed slot
909 // for the original deopt pc.
910
911 _orig_pc_slot = fixed_slots();
912 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
913 set_fixed_slots(next_slot);
914
915 // Compute when to use implicit null checks. Used by matching trap based
916 // nodes and NullCheck optimization.
917 set_allowed_deopt_reasons();
918
919 // Now generate code
920 Code_Gen();
921 if (failing()) return;
922
923 // Check if we want to skip execution of all compiled code.
924 {
925 #ifndef PRODUCT
926 if (OptoNoExecute) {
927 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
928 return;
929 }
930 #endif
931 TracePhase tp("install_code", &timers[_t_registerMethod]);
932
933 if (is_osr_compilation()) {
934 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
935 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
936 } else {
937 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
938 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
939 }
940
941 env()->register_method(_method, _entry_bci,
942 &_code_offsets,
943 _orig_pc_slot_offset_in_bytes,
944 code_buffer(),
945 frame_size_in_words(), _oop_map_set,
946 &_handler_table, &_inc_table,
947 compiler,
948 has_unsafe_access(),
949 SharedRuntime::is_wide_vector(max_vector_size()),
950 rtm_state()
951 );
952
953 if (log() != NULL) // Print code cache state into compiler log
954 log()->code_cache_state();
955 }
956 }
957
958 //------------------------------Compile----------------------------------------
959 // Compile a runtime stub
960 Compile::Compile( ciEnv* ci_env,
961 TypeFunc_generator generator,
962 address stub_function,
963 const char *stub_name,
964 int is_fancy_jump,
965 bool pass_tls,
966 bool save_arg_registers,
967 bool return_pc,
968 DirectiveSet* directive)
969 : Phase(Compiler),
970 _compile_id(0),
971 _save_argument_registers(save_arg_registers),
972 _subsume_loads(true),
973 _do_escape_analysis(false),
974 _eliminate_boxing(false),
975 _method(NULL),
976 _entry_bci(InvocationEntryBci),
977 _stub_function(stub_function),
978 _stub_name(stub_name),
979 _stub_entry_point(NULL),
980 _max_node_limit(MaxNodeLimit),
981 _orig_pc_slot(0),
982 _orig_pc_slot_offset_in_bytes(0),
983 _inlining_progress(false),
984 _inlining_incrementally(false),
985 _has_reserved_stack_access(false),
986 #ifndef PRODUCT
987 _trace_opto_output(directive->TraceOptoOutputOption),
988 #endif
989 _has_method_handle_invokes(false),
990 _comp_arena(mtCompiler),
991 _env(ci_env),
992 _directive(directive),
993 _log(ci_env->log()),
994 _failure_reason(NULL),
995 _congraph(NULL),
996 #ifndef PRODUCT
997 _printer(NULL),
998 #endif
999 _dead_node_list(comp_arena()),
1000 _dead_node_count(0),
1001 _node_arena(mtCompiler),
1002 _old_arena(mtCompiler),
1003 _mach_constant_base_node(NULL),
1004 _Compile_types(mtCompiler),
1005 _initial_gvn(NULL),
1006 _for_igvn(NULL),
1007 _warm_calls(NULL),
1008 _number_of_mh_late_inlines(0),
1009 _print_inlining_stream(NULL),
1010 _print_inlining_list(NULL),
1011 _print_inlining_idx(0),
1012 _print_inlining_output(NULL),
1013 _replay_inline_data(NULL),
1014 _java_calls(0),
1015 _inner_loops(0),
1016 _interpreter_frame_size(0),
1017 _node_bundling_limit(0),
1018 _node_bundling_base(NULL),
1019 _code_buffer("Compile::Fill_buffer"),
1020 #ifndef PRODUCT
1021 _in_dump_cnt(0),
1022 #endif
1023 _allowed_reasons(0) {
1024 C = this;
1025
1026 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1027 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1028
1029 #ifndef PRODUCT
1030 set_print_assembly(PrintFrameConverterAssembly);
1031 set_parsed_irreducible_loop(false);
1032 #endif
1033 set_has_irreducible_loop(false); // no loops
1034
1035 CompileWrapper cw(this);
1036 Init(/*AliasLevel=*/ 0);
1037 init_tf((*generator)());
1038
1039 {
1040 // The following is a dummy for the sake of GraphKit::gen_stub
1041 Unique_Node_List for_igvn(comp_arena());
1042 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1043 PhaseGVN gvn(Thread::current()->resource_area(),255);
1044 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1045 gvn.transform_no_reclaim(top());
1046
1047 GraphKit kit;
1048 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1049 }
1050
1051 NOT_PRODUCT( verify_graph_edges(); )
1052 Code_Gen();
1053 if (failing()) return;
1054
1055
1056 // Entry point will be accessed using compile->stub_entry_point();
1057 if (code_buffer() == NULL) {
1058 Matcher::soft_match_failure();
1059 } else {
1060 if (PrintAssembly && (WizardMode || Verbose))
1061 tty->print_cr("### Stub::%s", stub_name);
1062
1063 if (!failing()) {
1064 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1065
1066 // Make the NMethod
1067 // For now we mark the frame as never safe for profile stackwalking
1068 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1069 code_buffer(),
1070 CodeOffsets::frame_never_safe,
1071 // _code_offsets.value(CodeOffsets::Frame_Complete),
1072 frame_size_in_words(),
1073 _oop_map_set,
1074 save_arg_registers);
1075 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1076
1077 _stub_entry_point = rs->entry_point();
1078 }
1079 }
1080 }
1081
1082 //------------------------------Init-------------------------------------------
1083 // Prepare for a single compilation
1084 void Compile::Init(int aliaslevel) {
1085 _unique = 0;
1086 _regalloc = NULL;
1087
1088 _tf = NULL; // filled in later
1089 _top = NULL; // cached later
1090 _matcher = NULL; // filled in later
1091 _cfg = NULL; // filled in later
1092
1093 set_24_bit_selection_and_mode(Use24BitFP, false);
1094
1095 _node_note_array = NULL;
1096 _default_node_notes = NULL;
1097 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1098
1099 _immutable_memory = NULL; // filled in at first inquiry
1100
1101 // Globally visible Nodes
1102 // First set TOP to NULL to give safe behavior during creation of RootNode
1103 set_cached_top_node(NULL);
1104 set_root(new RootNode());
1105 // Now that you have a Root to point to, create the real TOP
1106 set_cached_top_node( new ConNode(Type::TOP) );
1107 set_recent_alloc(NULL, NULL);
1108
1109 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1110 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1111 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1112 env()->set_dependencies(new Dependencies(env()));
1113
1114 _fixed_slots = 0;
1115 set_has_split_ifs(false);
1116 set_has_loops(has_method() && method()->has_loops()); // first approximation
1117 set_has_stringbuilder(false);
1118 set_has_boxed_value(false);
1119 _trap_can_recompile = false; // no traps emitted yet
1120 _major_progress = true; // start out assuming good things will happen
1121 set_has_unsafe_access(false);
1122 set_max_vector_size(0);
1123 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1124 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1125 set_decompile_count(0);
1126
1127 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1128 set_num_loop_opts(LoopOptsCount);
1129 set_do_inlining(Inline);
1130 set_max_inline_size(MaxInlineSize);
1131 set_freq_inline_size(FreqInlineSize);
1132 set_do_scheduling(OptoScheduling);
1133 set_do_count_invocations(false);
1134 set_do_method_data_update(false);
1135
1136 set_do_vector_loop(false);
1137
1138 if (AllowVectorizeOnDemand) {
1139 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1140 set_do_vector_loop(true);
1141 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1142 } else if (has_method() && method()->name() != 0 &&
1143 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1144 set_do_vector_loop(true);
1145 }
1146 }
1147 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1148 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1149
1150 set_age_code(has_method() && method()->profile_aging());
1151 set_rtm_state(NoRTM); // No RTM lock eliding by default
1152 _max_node_limit = _directive->MaxNodeLimitOption;
1153
1154 #if INCLUDE_RTM_OPT
1155 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1156 int rtm_state = method()->method_data()->rtm_state();
1157 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1158 // Don't generate RTM lock eliding code.
1159 set_rtm_state(NoRTM);
1160 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1161 // Generate RTM lock eliding code without abort ratio calculation code.
1162 set_rtm_state(UseRTM);
1163 } else if (UseRTMDeopt) {
1164 // Generate RTM lock eliding code and include abort ratio calculation
1165 // code if UseRTMDeopt is on.
1166 set_rtm_state(ProfileRTM);
1167 }
1168 }
1169 #endif
1170 if (debug_info()->recording_non_safepoints()) {
1171 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1172 (comp_arena(), 8, 0, NULL));
1173 set_default_node_notes(Node_Notes::make(this));
1174 }
1175
1176 // // -- Initialize types before each compile --
1177 // // Update cached type information
1178 // if( _method && _method->constants() )
1179 // Type::update_loaded_types(_method, _method->constants());
1180
1181 // Init alias_type map.
1182 if (!_do_escape_analysis && aliaslevel == 3)
1183 aliaslevel = 2; // No unique types without escape analysis
1184 _AliasLevel = aliaslevel;
1185 const int grow_ats = 16;
1186 _max_alias_types = grow_ats;
1187 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1188 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1189 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1190 {
1191 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1192 }
1193 // Initialize the first few types.
1194 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1195 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1196 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1197 _num_alias_types = AliasIdxRaw+1;
1198 // Zero out the alias type cache.
1199 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1200 // A NULL adr_type hits in the cache right away. Preload the right answer.
1201 probe_alias_cache(NULL)->_index = AliasIdxTop;
1202
1203 _intrinsics = NULL;
1204 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1205 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1206 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1207 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1208 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1209 register_library_intrinsics();
1210 }
1211
1212 //---------------------------init_start----------------------------------------
1213 // Install the StartNode on this compile object.
1214 void Compile::init_start(StartNode* s) {
1215 if (failing())
1216 return; // already failing
1217 assert(s == start(), "");
1218 }
1219
1220 /**
1221 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1222 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1223 * the ideal graph.
1224 */
1225 StartNode* Compile::start() const {
1226 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1227 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1228 Node* start = root()->fast_out(i);
1229 if (start->is_Start()) {
1230 return start->as_Start();
1231 }
1232 }
1233 fatal("Did not find Start node!");
1234 return NULL;
1235 }
1236
1237 //-------------------------------immutable_memory-------------------------------------
1238 // Access immutable memory
1239 Node* Compile::immutable_memory() {
1240 if (_immutable_memory != NULL) {
1241 return _immutable_memory;
1242 }
1243 StartNode* s = start();
1244 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1245 Node *p = s->fast_out(i);
1246 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1247 _immutable_memory = p;
1248 return _immutable_memory;
1249 }
1250 }
1251 ShouldNotReachHere();
1252 return NULL;
1253 }
1254
1255 //----------------------set_cached_top_node------------------------------------
1256 // Install the cached top node, and make sure Node::is_top works correctly.
1257 void Compile::set_cached_top_node(Node* tn) {
1258 if (tn != NULL) verify_top(tn);
1259 Node* old_top = _top;
1260 _top = tn;
1261 // Calling Node::setup_is_top allows the nodes the chance to adjust
1262 // their _out arrays.
1263 if (_top != NULL) _top->setup_is_top();
1264 if (old_top != NULL) old_top->setup_is_top();
1265 assert(_top == NULL || top()->is_top(), "");
1266 }
1267
1268 #ifdef ASSERT
1269 uint Compile::count_live_nodes_by_graph_walk() {
1270 Unique_Node_List useful(comp_arena());
1271 // Get useful node list by walking the graph.
1272 identify_useful_nodes(useful);
1273 return useful.size();
1274 }
1275
1276 void Compile::print_missing_nodes() {
1277
1278 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1279 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1280 return;
1281 }
1282
1283 // This is an expensive function. It is executed only when the user
1284 // specifies VerifyIdealNodeCount option or otherwise knows the
1285 // additional work that needs to be done to identify reachable nodes
1286 // by walking the flow graph and find the missing ones using
1287 // _dead_node_list.
1288
1289 Unique_Node_List useful(comp_arena());
1290 // Get useful node list by walking the graph.
1291 identify_useful_nodes(useful);
1292
1293 uint l_nodes = C->live_nodes();
1294 uint l_nodes_by_walk = useful.size();
1295
1296 if (l_nodes != l_nodes_by_walk) {
1297 if (_log != NULL) {
1298 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1299 _log->stamp();
1300 _log->end_head();
1301 }
1302 VectorSet& useful_member_set = useful.member_set();
1303 int last_idx = l_nodes_by_walk;
1304 for (int i = 0; i < last_idx; i++) {
1305 if (useful_member_set.test(i)) {
1306 if (_dead_node_list.test(i)) {
1307 if (_log != NULL) {
1308 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1309 }
1310 if (PrintIdealNodeCount) {
1311 // Print the log message to tty
1312 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1313 useful.at(i)->dump();
1314 }
1315 }
1316 }
1317 else if (! _dead_node_list.test(i)) {
1318 if (_log != NULL) {
1319 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1320 }
1321 if (PrintIdealNodeCount) {
1322 // Print the log message to tty
1323 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1324 }
1325 }
1326 }
1327 if (_log != NULL) {
1328 _log->tail("mismatched_nodes");
1329 }
1330 }
1331 }
1332 void Compile::record_modified_node(Node* n) {
1333 if (_modified_nodes != NULL && !_inlining_incrementally &&
1334 n->outcnt() != 0 && !n->is_Con()) {
1335 _modified_nodes->push(n);
1336 }
1337 }
1338
1339 void Compile::remove_modified_node(Node* n) {
1340 if (_modified_nodes != NULL) {
1341 _modified_nodes->remove(n);
1342 }
1343 }
1344 #endif
1345
1346 #ifndef PRODUCT
1347 void Compile::verify_top(Node* tn) const {
1348 if (tn != NULL) {
1349 assert(tn->is_Con(), "top node must be a constant");
1350 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1351 assert(tn->in(0) != NULL, "must have live top node");
1352 }
1353 }
1354 #endif
1355
1356
1357 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1358
1359 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1360 guarantee(arr != NULL, "");
1361 int num_blocks = arr->length();
1362 if (grow_by < num_blocks) grow_by = num_blocks;
1363 int num_notes = grow_by * _node_notes_block_size;
1364 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1365 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1366 while (num_notes > 0) {
1367 arr->append(notes);
1368 notes += _node_notes_block_size;
1369 num_notes -= _node_notes_block_size;
1370 }
1371 assert(num_notes == 0, "exact multiple, please");
1372 }
1373
1374 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1375 if (source == NULL || dest == NULL) return false;
1376
1377 if (dest->is_Con())
1378 return false; // Do not push debug info onto constants.
1379
1380 #ifdef ASSERT
1381 // Leave a bread crumb trail pointing to the original node:
1382 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1383 dest->set_debug_orig(source);
1384 }
1385 #endif
1386
1387 if (node_note_array() == NULL)
1388 return false; // Not collecting any notes now.
1389
1390 // This is a copy onto a pre-existing node, which may already have notes.
1391 // If both nodes have notes, do not overwrite any pre-existing notes.
1392 Node_Notes* source_notes = node_notes_at(source->_idx);
1393 if (source_notes == NULL || source_notes->is_clear()) return false;
1394 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1395 if (dest_notes == NULL || dest_notes->is_clear()) {
1396 return set_node_notes_at(dest->_idx, source_notes);
1397 }
1398
1399 Node_Notes merged_notes = (*source_notes);
1400 // The order of operations here ensures that dest notes will win...
1401 merged_notes.update_from(dest_notes);
1402 return set_node_notes_at(dest->_idx, &merged_notes);
1403 }
1404
1405
1406 //--------------------------allow_range_check_smearing-------------------------
1407 // Gating condition for coalescing similar range checks.
1408 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1409 // single covering check that is at least as strong as any of them.
1410 // If the optimization succeeds, the simplified (strengthened) range check
1411 // will always succeed. If it fails, we will deopt, and then give up
1412 // on the optimization.
1413 bool Compile::allow_range_check_smearing() const {
1414 // If this method has already thrown a range-check,
1415 // assume it was because we already tried range smearing
1416 // and it failed.
1417 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1418 return !already_trapped;
1419 }
1420
1421
1422 //------------------------------flatten_alias_type-----------------------------
1423 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1424 int offset = tj->offset();
1425 TypePtr::PTR ptr = tj->ptr();
1426
1427 // Known instance (scalarizable allocation) alias only with itself.
1428 bool is_known_inst = tj->isa_oopptr() != NULL &&
1429 tj->is_oopptr()->is_known_instance();
1430
1431 // Process weird unsafe references.
1432 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1433 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1434 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1435 tj = TypeOopPtr::BOTTOM;
1436 ptr = tj->ptr();
1437 offset = tj->offset();
1438 }
1439
1440 // Array pointers need some flattening
1441 const TypeAryPtr *ta = tj->isa_aryptr();
1442 if (ta && ta->is_stable()) {
1443 // Erase stability property for alias analysis.
1444 tj = ta = ta->cast_to_stable(false);
1445 }
1446 if( ta && is_known_inst ) {
1447 if ( offset != Type::OffsetBot &&
1448 offset > arrayOopDesc::length_offset_in_bytes() ) {
1449 offset = Type::OffsetBot; // Flatten constant access into array body only
1450 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1451 }
1452 } else if( ta && _AliasLevel >= 2 ) {
1453 // For arrays indexed by constant indices, we flatten the alias
1454 // space to include all of the array body. Only the header, klass
1455 // and array length can be accessed un-aliased.
1456 if( offset != Type::OffsetBot ) {
1457 if( ta->const_oop() ) { // MethodData* or Method*
1458 offset = Type::OffsetBot; // Flatten constant access into array body
1459 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1460 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1461 // range is OK as-is.
1462 tj = ta = TypeAryPtr::RANGE;
1463 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1464 tj = TypeInstPtr::KLASS; // all klass loads look alike
1465 ta = TypeAryPtr::RANGE; // generic ignored junk
1466 ptr = TypePtr::BotPTR;
1467 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1468 tj = TypeInstPtr::MARK;
1469 ta = TypeAryPtr::RANGE; // generic ignored junk
1470 ptr = TypePtr::BotPTR;
1471 } else { // Random constant offset into array body
1472 offset = Type::OffsetBot; // Flatten constant access into array body
1473 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1474 }
1475 }
1476 // Arrays of fixed size alias with arrays of unknown size.
1477 if (ta->size() != TypeInt::POS) {
1478 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1479 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1480 }
1481 // Arrays of known objects become arrays of unknown objects.
1482 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1483 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1484 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1485 }
1486 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1487 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1488 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1489 }
1490 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1491 // cannot be distinguished by bytecode alone.
1492 if (ta->elem() == TypeInt::BOOL) {
1493 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1494 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1495 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1496 }
1497 // During the 2nd round of IterGVN, NotNull castings are removed.
1498 // Make sure the Bottom and NotNull variants alias the same.
1499 // Also, make sure exact and non-exact variants alias the same.
1500 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1501 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1502 }
1503 }
1504
1505 // Oop pointers need some flattening
1506 const TypeInstPtr *to = tj->isa_instptr();
1507 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1508 ciInstanceKlass *k = to->klass()->as_instance_klass();
1509 if( ptr == TypePtr::Constant ) {
1510 if (to->klass() != ciEnv::current()->Class_klass() ||
1511 offset < k->size_helper() * wordSize) {
1512 // No constant oop pointers (such as Strings); they alias with
1513 // unknown strings.
1514 assert(!is_known_inst, "not scalarizable allocation");
1515 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1516 }
1517 } else if( is_known_inst ) {
1518 tj = to; // Keep NotNull and klass_is_exact for instance type
1519 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1520 // During the 2nd round of IterGVN, NotNull castings are removed.
1521 // Make sure the Bottom and NotNull variants alias the same.
1522 // Also, make sure exact and non-exact variants alias the same.
1523 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1524 }
1525 if (to->speculative() != NULL) {
1526 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1527 }
1528 // Canonicalize the holder of this field
1529 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1530 // First handle header references such as a LoadKlassNode, even if the
1531 // object's klass is unloaded at compile time (4965979).
1532 if (!is_known_inst) { // Do it only for non-instance types
1533 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1534 }
1535 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1536 // Static fields are in the space above the normal instance
1537 // fields in the java.lang.Class instance.
1538 if (to->klass() != ciEnv::current()->Class_klass()) {
1539 to = NULL;
1540 tj = TypeOopPtr::BOTTOM;
1541 offset = tj->offset();
1542 }
1543 } else {
1544 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1545 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1546 if( is_known_inst ) {
1547 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1548 } else {
1549 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1550 }
1551 }
1552 }
1553 }
1554
1555 // Klass pointers to object array klasses need some flattening
1556 const TypeKlassPtr *tk = tj->isa_klassptr();
1557 if( tk ) {
1558 // If we are referencing a field within a Klass, we need
1559 // to assume the worst case of an Object. Both exact and
1560 // inexact types must flatten to the same alias class so
1561 // use NotNull as the PTR.
1562 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1563
1564 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1565 TypeKlassPtr::OBJECT->klass(),
1566 offset);
1567 }
1568
1569 ciKlass* klass = tk->klass();
1570 if( klass->is_obj_array_klass() ) {
1571 ciKlass* k = TypeAryPtr::OOPS->klass();
1572 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1573 k = TypeInstPtr::BOTTOM->klass();
1574 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1575 }
1576
1577 // Check for precise loads from the primary supertype array and force them
1578 // to the supertype cache alias index. Check for generic array loads from
1579 // the primary supertype array and also force them to the supertype cache
1580 // alias index. Since the same load can reach both, we need to merge
1581 // these 2 disparate memories into the same alias class. Since the
1582 // primary supertype array is read-only, there's no chance of confusion
1583 // where we bypass an array load and an array store.
1584 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1585 if (offset == Type::OffsetBot ||
1586 (offset >= primary_supers_offset &&
1587 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1588 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1589 offset = in_bytes(Klass::secondary_super_cache_offset());
1590 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1591 }
1592 }
1593
1594 // Flatten all Raw pointers together.
1595 if (tj->base() == Type::RawPtr)
1596 tj = TypeRawPtr::BOTTOM;
1597
1598 if (tj->base() == Type::AnyPtr)
1599 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1600
1601 // Flatten all to bottom for now
1602 switch( _AliasLevel ) {
1603 case 0:
1604 tj = TypePtr::BOTTOM;
1605 break;
1606 case 1: // Flatten to: oop, static, field or array
1607 switch (tj->base()) {
1608 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1609 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1610 case Type::AryPtr: // do not distinguish arrays at all
1611 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1612 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1613 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1614 default: ShouldNotReachHere();
1615 }
1616 break;
1617 case 2: // No collapsing at level 2; keep all splits
1618 case 3: // No collapsing at level 3; keep all splits
1619 break;
1620 default:
1621 Unimplemented();
1622 }
1623
1624 offset = tj->offset();
1625 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1626
1627 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1628 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1629 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1630 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1631 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1632 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1633 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1634 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1635 assert( tj->ptr() != TypePtr::TopPTR &&
1636 tj->ptr() != TypePtr::AnyNull &&
1637 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1638 // assert( tj->ptr() != TypePtr::Constant ||
1639 // tj->base() == Type::RawPtr ||
1640 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1641
1642 return tj;
1643 }
1644
1645 void Compile::AliasType::Init(int i, const TypePtr* at) {
1646 _index = i;
1647 _adr_type = at;
1648 _field = NULL;
1649 _element = NULL;
1650 _is_rewritable = true; // default
1651 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1652 if (atoop != NULL && atoop->is_known_instance()) {
1653 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1654 _general_index = Compile::current()->get_alias_index(gt);
1655 } else {
1656 _general_index = 0;
1657 }
1658 }
1659
1660 BasicType Compile::AliasType::basic_type() const {
1661 if (element() != NULL) {
1662 const Type* element = adr_type()->is_aryptr()->elem();
1663 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1664 } if (field() != NULL) {
1665 return field()->layout_type();
1666 } else {
1667 return T_ILLEGAL; // unknown
1668 }
1669 }
1670
1671 //---------------------------------print_on------------------------------------
1672 #ifndef PRODUCT
1673 void Compile::AliasType::print_on(outputStream* st) {
1674 if (index() < 10)
1675 st->print("@ <%d> ", index());
1676 else st->print("@ <%d>", index());
1677 st->print(is_rewritable() ? " " : " RO");
1678 int offset = adr_type()->offset();
1679 if (offset == Type::OffsetBot)
1680 st->print(" +any");
1681 else st->print(" +%-3d", offset);
1682 st->print(" in ");
1683 adr_type()->dump_on(st);
1684 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1685 if (field() != NULL && tjp) {
1686 if (tjp->klass() != field()->holder() ||
1687 tjp->offset() != field()->offset_in_bytes()) {
1688 st->print(" != ");
1689 field()->print();
1690 st->print(" ***");
1691 }
1692 }
1693 }
1694
1695 void print_alias_types() {
1696 Compile* C = Compile::current();
1697 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1698 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1699 C->alias_type(idx)->print_on(tty);
1700 tty->cr();
1701 }
1702 }
1703 #endif
1704
1705
1706 //----------------------------probe_alias_cache--------------------------------
1707 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1708 intptr_t key = (intptr_t) adr_type;
1709 key ^= key >> logAliasCacheSize;
1710 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1711 }
1712
1713
1714 //-----------------------------grow_alias_types--------------------------------
1715 void Compile::grow_alias_types() {
1716 const int old_ats = _max_alias_types; // how many before?
1717 const int new_ats = old_ats; // how many more?
1718 const int grow_ats = old_ats+new_ats; // how many now?
1719 _max_alias_types = grow_ats;
1720 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1721 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1722 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1723 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1724 }
1725
1726
1727 //--------------------------------find_alias_type------------------------------
1728 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1729 if (_AliasLevel == 0)
1730 return alias_type(AliasIdxBot);
1731
1732 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1733 if (ace->_adr_type == adr_type) {
1734 return alias_type(ace->_index);
1735 }
1736
1737 // Handle special cases.
1738 if (adr_type == NULL) return alias_type(AliasIdxTop);
1739 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1740
1741 // Do it the slow way.
1742 const TypePtr* flat = flatten_alias_type(adr_type);
1743
1744 #ifdef ASSERT
1745 {
1746 ResourceMark rm;
1747 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1748 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1749 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1750 Type::str(adr_type));
1751 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1752 const TypeOopPtr* foop = flat->is_oopptr();
1753 // Scalarizable allocations have exact klass always.
1754 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1755 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1756 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1757 Type::str(foop), Type::str(xoop));
1758 }
1759 }
1760 #endif
1761
1762 int idx = AliasIdxTop;
1763 for (int i = 0; i < num_alias_types(); i++) {
1764 if (alias_type(i)->adr_type() == flat) {
1765 idx = i;
1766 break;
1767 }
1768 }
1769
1770 if (idx == AliasIdxTop) {
1771 if (no_create) return NULL;
1772 // Grow the array if necessary.
1773 if (_num_alias_types == _max_alias_types) grow_alias_types();
1774 // Add a new alias type.
1775 idx = _num_alias_types++;
1776 _alias_types[idx]->Init(idx, flat);
1777 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1778 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1779 if (flat->isa_instptr()) {
1780 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1781 && flat->is_instptr()->klass() == env()->Class_klass())
1782 alias_type(idx)->set_rewritable(false);
1783 }
1784 if (flat->isa_aryptr()) {
1785 #ifdef ASSERT
1786 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1787 // (T_BYTE has the weakest alignment and size restrictions...)
1788 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1789 #endif
1790 if (flat->offset() == TypePtr::OffsetBot) {
1791 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1792 }
1793 }
1794 if (flat->isa_klassptr()) {
1795 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1796 alias_type(idx)->set_rewritable(false);
1797 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1798 alias_type(idx)->set_rewritable(false);
1799 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1800 alias_type(idx)->set_rewritable(false);
1801 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1802 alias_type(idx)->set_rewritable(false);
1803 }
1804 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1805 // but the base pointer type is not distinctive enough to identify
1806 // references into JavaThread.)
1807
1808 // Check for final fields.
1809 const TypeInstPtr* tinst = flat->isa_instptr();
1810 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1811 ciField* field;
1812 if (tinst->const_oop() != NULL &&
1813 tinst->klass() == ciEnv::current()->Class_klass() &&
1814 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1815 // static field
1816 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1817 field = k->get_field_by_offset(tinst->offset(), true);
1818 } else {
1819 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1820 field = k->get_field_by_offset(tinst->offset(), false);
1821 }
1822 assert(field == NULL ||
1823 original_field == NULL ||
1824 (field->holder() == original_field->holder() &&
1825 field->offset() == original_field->offset() &&
1826 field->is_static() == original_field->is_static()), "wrong field?");
1827 // Set field() and is_rewritable() attributes.
1828 if (field != NULL) alias_type(idx)->set_field(field);
1829 }
1830 }
1831
1832 // Fill the cache for next time.
1833 ace->_adr_type = adr_type;
1834 ace->_index = idx;
1835 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1836
1837 // Might as well try to fill the cache for the flattened version, too.
1838 AliasCacheEntry* face = probe_alias_cache(flat);
1839 if (face->_adr_type == NULL) {
1840 face->_adr_type = flat;
1841 face->_index = idx;
1842 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1843 }
1844
1845 return alias_type(idx);
1846 }
1847
1848
1849 Compile::AliasType* Compile::alias_type(ciField* field) {
1850 const TypeOopPtr* t;
1851 if (field->is_static())
1852 t = TypeInstPtr::make(field->holder()->java_mirror());
1853 else
1854 t = TypeOopPtr::make_from_klass_raw(field->holder());
1855 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1856 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1857 return atp;
1858 }
1859
1860
1861 //------------------------------have_alias_type--------------------------------
1862 bool Compile::have_alias_type(const TypePtr* adr_type) {
1863 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1864 if (ace->_adr_type == adr_type) {
1865 return true;
1866 }
1867
1868 // Handle special cases.
1869 if (adr_type == NULL) return true;
1870 if (adr_type == TypePtr::BOTTOM) return true;
1871
1872 return find_alias_type(adr_type, true, NULL) != NULL;
1873 }
1874
1875 //-----------------------------must_alias--------------------------------------
1876 // True if all values of the given address type are in the given alias category.
1877 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1878 if (alias_idx == AliasIdxBot) return true; // the universal category
1879 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1880 if (alias_idx == AliasIdxTop) return false; // the empty category
1881 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1882
1883 // the only remaining possible overlap is identity
1884 int adr_idx = get_alias_index(adr_type);
1885 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1886 assert(adr_idx == alias_idx ||
1887 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1888 && adr_type != TypeOopPtr::BOTTOM),
1889 "should not be testing for overlap with an unsafe pointer");
1890 return adr_idx == alias_idx;
1891 }
1892
1893 //------------------------------can_alias--------------------------------------
1894 // True if any values of the given address type are in the given alias category.
1895 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1896 if (alias_idx == AliasIdxTop) return false; // the empty category
1897 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1898 if (alias_idx == AliasIdxBot) return true; // the universal category
1899 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1900
1901 // the only remaining possible overlap is identity
1902 int adr_idx = get_alias_index(adr_type);
1903 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1904 return adr_idx == alias_idx;
1905 }
1906
1907
1908
1909 //---------------------------pop_warm_call-------------------------------------
1910 WarmCallInfo* Compile::pop_warm_call() {
1911 WarmCallInfo* wci = _warm_calls;
1912 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1913 return wci;
1914 }
1915
1916 //----------------------------Inline_Warm--------------------------------------
1917 int Compile::Inline_Warm() {
1918 // If there is room, try to inline some more warm call sites.
1919 // %%% Do a graph index compaction pass when we think we're out of space?
1920 if (!InlineWarmCalls) return 0;
1921
1922 int calls_made_hot = 0;
1923 int room_to_grow = NodeCountInliningCutoff - unique();
1924 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1925 int amount_grown = 0;
1926 WarmCallInfo* call;
1927 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1928 int est_size = (int)call->size();
1929 if (est_size > (room_to_grow - amount_grown)) {
1930 // This one won't fit anyway. Get rid of it.
1931 call->make_cold();
1932 continue;
1933 }
1934 call->make_hot();
1935 calls_made_hot++;
1936 amount_grown += est_size;
1937 amount_to_grow -= est_size;
1938 }
1939
1940 if (calls_made_hot > 0) set_major_progress();
1941 return calls_made_hot;
1942 }
1943
1944
1945 //----------------------------Finish_Warm--------------------------------------
1946 void Compile::Finish_Warm() {
1947 if (!InlineWarmCalls) return;
1948 if (failing()) return;
1949 if (warm_calls() == NULL) return;
1950
1951 // Clean up loose ends, if we are out of space for inlining.
1952 WarmCallInfo* call;
1953 while ((call = pop_warm_call()) != NULL) {
1954 call->make_cold();
1955 }
1956 }
1957
1958 //---------------------cleanup_loop_predicates-----------------------
1959 // Remove the opaque nodes that protect the predicates so that all unused
1960 // checks and uncommon_traps will be eliminated from the ideal graph
1961 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1962 if (predicate_count()==0) return;
1963 for (int i = predicate_count(); i > 0; i--) {
1964 Node * n = predicate_opaque1_node(i-1);
1965 assert(n->Opcode() == Op_Opaque1, "must be");
1966 igvn.replace_node(n, n->in(1));
1967 }
1968 assert(predicate_count()==0, "should be clean!");
1969 }
1970
1971 void Compile::add_range_check_cast(Node* n) {
1972 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1973 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1974 _range_check_casts->append(n);
1975 }
1976
1977 // Remove all range check dependent CastIINodes.
1978 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1979 for (int i = range_check_cast_count(); i > 0; i--) {
1980 Node* cast = range_check_cast_node(i-1);
1981 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1982 igvn.replace_node(cast, cast->in(1));
1983 }
1984 assert(range_check_cast_count() == 0, "should be empty");
1985 }
1986
1987 void Compile::add_opaque4_node(Node* n) {
1988 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1989 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1990 _opaque4_nodes->append(n);
1991 }
1992
1993 // Remove all Opaque4 nodes.
1994 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1995 for (int i = opaque4_count(); i > 0; i--) {
1996 Node* opaq = opaque4_node(i-1);
1997 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
1998 igvn.replace_node(opaq, opaq->in(2));
1999 }
2000 assert(opaque4_count() == 0, "should be empty");
2001 }
2002
2003 // StringOpts and late inlining of string methods
2004 void Compile::inline_string_calls(bool parse_time) {
2005 {
2006 // remove useless nodes to make the usage analysis simpler
2007 ResourceMark rm;
2008 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2009 }
2010
2011 {
2012 ResourceMark rm;
2013 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2014 PhaseStringOpts pso(initial_gvn(), for_igvn());
2015 print_method(PHASE_AFTER_STRINGOPTS, 3);
2016 }
2017
2018 // now inline anything that we skipped the first time around
2019 if (!parse_time) {
2020 _late_inlines_pos = _late_inlines.length();
2021 }
2022
2023 while (_string_late_inlines.length() > 0) {
2024 CallGenerator* cg = _string_late_inlines.pop();
2025 cg->do_late_inline();
2026 if (failing()) return;
2027 }
2028 _string_late_inlines.trunc_to(0);
2029 }
2030
2031 // Late inlining of boxing methods
2032 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2033 if (_boxing_late_inlines.length() > 0) {
2034 assert(has_boxed_value(), "inconsistent");
2035
2036 PhaseGVN* gvn = initial_gvn();
2037 set_inlining_incrementally(true);
2038
2039 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2040 for_igvn()->clear();
2041 gvn->replace_with(&igvn);
2042
2043 _late_inlines_pos = _late_inlines.length();
2044
2045 while (_boxing_late_inlines.length() > 0) {
2046 CallGenerator* cg = _boxing_late_inlines.pop();
2047 cg->do_late_inline();
2048 if (failing()) return;
2049 }
2050 _boxing_late_inlines.trunc_to(0);
2051
2052 {
2053 ResourceMark rm;
2054 PhaseRemoveUseless pru(gvn, for_igvn());
2055 }
2056
2057 igvn = PhaseIterGVN(gvn);
2058 igvn.optimize();
2059
2060 set_inlining_progress(false);
2061 set_inlining_incrementally(false);
2062 }
2063 }
2064
2065 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2066 assert(IncrementalInline, "incremental inlining should be on");
2067 PhaseGVN* gvn = initial_gvn();
2068
2069 set_inlining_progress(false);
2070 for_igvn()->clear();
2071 gvn->replace_with(&igvn);
2072
2073 {
2074 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2075 int i = 0;
2076 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2077 CallGenerator* cg = _late_inlines.at(i);
2078 _late_inlines_pos = i+1;
2079 cg->do_late_inline();
2080 if (failing()) return;
2081 }
2082 int j = 0;
2083 for (; i < _late_inlines.length(); i++, j++) {
2084 _late_inlines.at_put(j, _late_inlines.at(i));
2085 }
2086 _late_inlines.trunc_to(j);
2087 }
2088
2089 {
2090 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2091 ResourceMark rm;
2092 PhaseRemoveUseless pru(gvn, for_igvn());
2093 }
2094
2095 {
2096 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2097 igvn = PhaseIterGVN(gvn);
2098 }
2099 }
2100
2101 // Perform incremental inlining until bound on number of live nodes is reached
2102 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2103 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2104
2105 PhaseGVN* gvn = initial_gvn();
2106
2107 set_inlining_incrementally(true);
2108 set_inlining_progress(true);
2109 uint low_live_nodes = 0;
2110
2111 while(inlining_progress() && _late_inlines.length() > 0) {
2112
2113 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2114 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2115 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2116 // PhaseIdealLoop is expensive so we only try it once we are
2117 // out of live nodes and we only try it again if the previous
2118 // helped got the number of nodes down significantly
2119 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2120 if (failing()) return;
2121 low_live_nodes = live_nodes();
2122 _major_progress = true;
2123 }
2124
2125 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2126 break;
2127 }
2128 }
2129
2130 inline_incrementally_one(igvn);
2131
2132 if (failing()) return;
2133
2134 {
2135 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2136 igvn.optimize();
2137 }
2138
2139 if (failing()) return;
2140 }
2141
2142 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2143
2144 if (_string_late_inlines.length() > 0) {
2145 assert(has_stringbuilder(), "inconsistent");
2146 for_igvn()->clear();
2147 initial_gvn()->replace_with(&igvn);
2148
2149 inline_string_calls(false);
2150
2151 if (failing()) return;
2152
2153 {
2154 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2155 ResourceMark rm;
2156 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2157 }
2158
2159 {
2160 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2161 igvn = PhaseIterGVN(gvn);
2162 igvn.optimize();
2163 }
2164 }
2165
2166 set_inlining_incrementally(false);
2167 }
2168
2169
2170 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2171 if(loop_opts_cnt > 0) {
2172 debug_only( int cnt = 0; );
2173 while(major_progress() && (loop_opts_cnt > 0)) {
2174 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2175 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2176 PhaseIdealLoop ideal_loop(igvn, mode);
2177 loop_opts_cnt--;
2178 if (failing()) return false;
2179 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2180 }
2181 }
2182 return true;
2183 }
2184
2185 //------------------------------Optimize---------------------------------------
2186 // Given a graph, optimize it.
2187 void Compile::Optimize() {
2188 TracePhase tp("optimizer", &timers[_t_optimizer]);
2189
2190 #ifndef PRODUCT
2191 if (_directive->BreakAtCompileOption) {
2192 BREAKPOINT;
2193 }
2194
2195 #endif
2196
2197 #ifdef ASSERT
2198 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2199 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2200 #endif
2201
2202 ResourceMark rm;
2203 int loop_opts_cnt;
2204
2205 print_inlining_reinit();
2206
2207 NOT_PRODUCT( verify_graph_edges(); )
2208
2209 print_method(PHASE_AFTER_PARSING);
2210
2211 {
2212 // Iterative Global Value Numbering, including ideal transforms
2213 // Initialize IterGVN with types and values from parse-time GVN
2214 PhaseIterGVN igvn(initial_gvn());
2215 #ifdef ASSERT
2216 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2217 #endif
2218 {
2219 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2220 igvn.optimize();
2221 }
2222
2223 if (failing()) return;
2224
2225 print_method(PHASE_ITER_GVN1, 2);
2226
2227 inline_incrementally(igvn);
2228
2229 print_method(PHASE_INCREMENTAL_INLINE, 2);
2230
2231 if (failing()) return;
2232
2233 if (eliminate_boxing()) {
2234 // Inline valueOf() methods now.
2235 inline_boxing_calls(igvn);
2236
2237 if (AlwaysIncrementalInline) {
2238 inline_incrementally(igvn);
2239 }
2240
2241 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2242
2243 if (failing()) return;
2244 }
2245
2246 // Remove the speculative part of types and clean up the graph from
2247 // the extra CastPP nodes whose only purpose is to carry them. Do
2248 // that early so that optimizations are not disrupted by the extra
2249 // CastPP nodes.
2250 remove_speculative_types(igvn);
2251
2252 // No more new expensive nodes will be added to the list from here
2253 // so keep only the actual candidates for optimizations.
2254 cleanup_expensive_nodes(igvn);
2255
2256 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2257 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2258 initial_gvn()->replace_with(&igvn);
2259 for_igvn()->clear();
2260 Unique_Node_List new_worklist(C->comp_arena());
2261 {
2262 ResourceMark rm;
2263 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2264 }
2265 set_for_igvn(&new_worklist);
2266 igvn = PhaseIterGVN(initial_gvn());
2267 igvn.optimize();
2268 }
2269
2270 // Perform escape analysis
2271 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2272 if (has_loops()) {
2273 // Cleanup graph (remove dead nodes).
2274 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2275 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2276 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2277 if (failing()) return;
2278 }
2279 ConnectionGraph::do_analysis(this, &igvn);
2280
2281 if (failing()) return;
2282
2283 // Optimize out fields loads from scalar replaceable allocations.
2284 igvn.optimize();
2285 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2286
2287 if (failing()) return;
2288
2289 if (congraph() != NULL && macro_count() > 0) {
2290 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2291 PhaseMacroExpand mexp(igvn);
2292 mexp.eliminate_macro_nodes();
2293 igvn.set_delay_transform(false);
2294
2295 igvn.optimize();
2296 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2297
2298 if (failing()) return;
2299 }
2300 }
2301
2302 // Loop transforms on the ideal graph. Range Check Elimination,
2303 // peeling, unrolling, etc.
2304
2305 // Set loop opts counter
2306 loop_opts_cnt = num_loop_opts();
2307 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2308 {
2309 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2310 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2311 loop_opts_cnt--;
2312 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2313 if (failing()) return;
2314 }
2315 // Loop opts pass if partial peeling occurred in previous pass
2316 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2317 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2318 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2319 loop_opts_cnt--;
2320 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2321 if (failing()) return;
2322 }
2323 // Loop opts pass for loop-unrolling before CCP
2324 if(major_progress() && (loop_opts_cnt > 0)) {
2325 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2326 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2327 loop_opts_cnt--;
2328 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2329 }
2330 if (!failing()) {
2331 // Verify that last round of loop opts produced a valid graph
2332 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2333 PhaseIdealLoop::verify(igvn);
2334 }
2335 }
2336 if (failing()) return;
2337
2338 // Conditional Constant Propagation;
2339 PhaseCCP ccp( &igvn );
2340 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2341 {
2342 TracePhase tp("ccp", &timers[_t_ccp]);
2343 ccp.do_transform();
2344 }
2345 print_method(PHASE_CPP1, 2);
2346
2347 assert( true, "Break here to ccp.dump_old2new_map()");
2348
2349 // Iterative Global Value Numbering, including ideal transforms
2350 {
2351 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2352 igvn = ccp;
2353 igvn.optimize();
2354 }
2355
2356 print_method(PHASE_ITER_GVN2, 2);
2357
2358 if (failing()) return;
2359
2360 // Loop transforms on the ideal graph. Range Check Elimination,
2361 // peeling, unrolling, etc.
2362 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2363 return;
2364 }
2365
2366 #if INCLUDE_ZGC
2367 if (UseZGC) {
2368 ZBarrierSetC2::find_dominating_barriers(igvn);
2369 }
2370 #endif
2371
2372 if (failing()) return;
2373
2374 // Ensure that major progress is now clear
2375 C->clear_major_progress();
2376
2377 {
2378 // Verify that all previous optimizations produced a valid graph
2379 // at least to this point, even if no loop optimizations were done.
2380 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2381 PhaseIdealLoop::verify(igvn);
2382 }
2383
2384 if (range_check_cast_count() > 0) {
2385 // No more loop optimizations. Remove all range check dependent CastIINodes.
2386 C->remove_range_check_casts(igvn);
2387 igvn.optimize();
2388 }
2389
2390 #ifdef ASSERT
2391 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2392 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand);
2393 #endif
2394
2395 {
2396 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2397 PhaseMacroExpand mex(igvn);
2398 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2399 if (mex.expand_macro_nodes()) {
2400 assert(failing(), "must bail out w/ explicit message");
2401 return;
2402 }
2403 }
2404
2405 if (opaque4_count() > 0) {
2406 C->remove_opaque4_nodes(igvn);
2407 igvn.optimize();
2408 }
2409
2410 DEBUG_ONLY( _modified_nodes = NULL; )
2411 } // (End scope of igvn; run destructor if necessary for asserts.)
2412
2413 process_print_inlining();
2414 // A method with only infinite loops has no edges entering loops from root
2415 {
2416 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2417 if (final_graph_reshaping()) {
2418 assert(failing(), "must bail out w/ explicit message");
2419 return;
2420 }
2421 }
2422
2423 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2424 }
2425
2426
2427 //------------------------------Code_Gen---------------------------------------
2428 // Given a graph, generate code for it
2429 void Compile::Code_Gen() {
2430 if (failing()) {
2431 return;
2432 }
2433
2434 // Perform instruction selection. You might think we could reclaim Matcher
2435 // memory PDQ, but actually the Matcher is used in generating spill code.
2436 // Internals of the Matcher (including some VectorSets) must remain live
2437 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2438 // set a bit in reclaimed memory.
2439
2440 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2441 // nodes. Mapping is only valid at the root of each matched subtree.
2442 NOT_PRODUCT( verify_graph_edges(); )
2443
2444 Matcher matcher;
2445 _matcher = &matcher;
2446 {
2447 TracePhase tp("matcher", &timers[_t_matcher]);
2448 matcher.match();
2449 }
2450 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2451 // nodes. Mapping is only valid at the root of each matched subtree.
2452 NOT_PRODUCT( verify_graph_edges(); )
2453
2454 // If you have too many nodes, or if matching has failed, bail out
2455 check_node_count(0, "out of nodes matching instructions");
2456 if (failing()) {
2457 return;
2458 }
2459
2460 print_method(PHASE_MATCHING, 2);
2461
2462 // Build a proper-looking CFG
2463 PhaseCFG cfg(node_arena(), root(), matcher);
2464 _cfg = &cfg;
2465 {
2466 TracePhase tp("scheduler", &timers[_t_scheduler]);
2467 bool success = cfg.do_global_code_motion();
2468 if (!success) {
2469 return;
2470 }
2471
2472 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2473 NOT_PRODUCT( verify_graph_edges(); )
2474 debug_only( cfg.verify(); )
2475 }
2476
2477 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2478 _regalloc = ®alloc;
2479 {
2480 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2481 // Perform register allocation. After Chaitin, use-def chains are
2482 // no longer accurate (at spill code) and so must be ignored.
2483 // Node->LRG->reg mappings are still accurate.
2484 _regalloc->Register_Allocate();
2485
2486 // Bail out if the allocator builds too many nodes
2487 if (failing()) {
2488 return;
2489 }
2490 }
2491
2492 // Prior to register allocation we kept empty basic blocks in case the
2493 // the allocator needed a place to spill. After register allocation we
2494 // are not adding any new instructions. If any basic block is empty, we
2495 // can now safely remove it.
2496 {
2497 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2498 cfg.remove_empty_blocks();
2499 if (do_freq_based_layout()) {
2500 PhaseBlockLayout layout(cfg);
2501 } else {
2502 cfg.set_loop_alignment();
2503 }
2504 cfg.fixup_flow();
2505 }
2506
2507 // Apply peephole optimizations
2508 if( OptoPeephole ) {
2509 TracePhase tp("peephole", &timers[_t_peephole]);
2510 PhasePeephole peep( _regalloc, cfg);
2511 peep.do_transform();
2512 }
2513
2514 // Do late expand if CPU requires this.
2515 if (Matcher::require_postalloc_expand) {
2516 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2517 cfg.postalloc_expand(_regalloc);
2518 }
2519
2520 // Convert Nodes to instruction bits in a buffer
2521 {
2522 TraceTime tp("output", &timers[_t_output], CITime);
2523 Output();
2524 }
2525
2526 print_method(PHASE_FINAL_CODE);
2527
2528 // He's dead, Jim.
2529 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2530 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2531 }
2532
2533
2534 //------------------------------dump_asm---------------------------------------
2535 // Dump formatted assembly
2536 #ifndef PRODUCT
2537 void Compile::dump_asm(int *pcs, uint pc_limit) {
2538 bool cut_short = false;
2539 tty->print_cr("#");
2540 tty->print("# "); _tf->dump(); tty->cr();
2541 tty->print_cr("#");
2542
2543 // For all blocks
2544 int pc = 0x0; // Program counter
2545 char starts_bundle = ' ';
2546 _regalloc->dump_frame();
2547
2548 Node *n = NULL;
2549 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2550 if (VMThread::should_terminate()) {
2551 cut_short = true;
2552 break;
2553 }
2554 Block* block = _cfg->get_block(i);
2555 if (block->is_connector() && !Verbose) {
2556 continue;
2557 }
2558 n = block->head();
2559 if (pcs && n->_idx < pc_limit) {
2560 tty->print("%3.3x ", pcs[n->_idx]);
2561 } else {
2562 tty->print(" ");
2563 }
2564 block->dump_head(_cfg);
2565 if (block->is_connector()) {
2566 tty->print_cr(" # Empty connector block");
2567 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2568 tty->print_cr(" # Block is sole successor of call");
2569 }
2570
2571 // For all instructions
2572 Node *delay = NULL;
2573 for (uint j = 0; j < block->number_of_nodes(); j++) {
2574 if (VMThread::should_terminate()) {
2575 cut_short = true;
2576 break;
2577 }
2578 n = block->get_node(j);
2579 if (valid_bundle_info(n)) {
2580 Bundle* bundle = node_bundling(n);
2581 if (bundle->used_in_unconditional_delay()) {
2582 delay = n;
2583 continue;
2584 }
2585 if (bundle->starts_bundle()) {
2586 starts_bundle = '+';
2587 }
2588 }
2589
2590 if (WizardMode) {
2591 n->dump();
2592 }
2593
2594 if( !n->is_Region() && // Dont print in the Assembly
2595 !n->is_Phi() && // a few noisely useless nodes
2596 !n->is_Proj() &&
2597 !n->is_MachTemp() &&
2598 !n->is_SafePointScalarObject() &&
2599 !n->is_Catch() && // Would be nice to print exception table targets
2600 !n->is_MergeMem() && // Not very interesting
2601 !n->is_top() && // Debug info table constants
2602 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2603 ) {
2604 if (pcs && n->_idx < pc_limit)
2605 tty->print("%3.3x", pcs[n->_idx]);
2606 else
2607 tty->print(" ");
2608 tty->print(" %c ", starts_bundle);
2609 starts_bundle = ' ';
2610 tty->print("\t");
2611 n->format(_regalloc, tty);
2612 tty->cr();
2613 }
2614
2615 // If we have an instruction with a delay slot, and have seen a delay,
2616 // then back up and print it
2617 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2618 assert(delay != NULL, "no unconditional delay instruction");
2619 if (WizardMode) delay->dump();
2620
2621 if (node_bundling(delay)->starts_bundle())
2622 starts_bundle = '+';
2623 if (pcs && n->_idx < pc_limit)
2624 tty->print("%3.3x", pcs[n->_idx]);
2625 else
2626 tty->print(" ");
2627 tty->print(" %c ", starts_bundle);
2628 starts_bundle = ' ';
2629 tty->print("\t");
2630 delay->format(_regalloc, tty);
2631 tty->cr();
2632 delay = NULL;
2633 }
2634
2635 // Dump the exception table as well
2636 if( n->is_Catch() && (Verbose || WizardMode) ) {
2637 // Print the exception table for this offset
2638 _handler_table.print_subtable_for(pc);
2639 }
2640 }
2641
2642 if (pcs && n->_idx < pc_limit)
2643 tty->print_cr("%3.3x", pcs[n->_idx]);
2644 else
2645 tty->cr();
2646
2647 assert(cut_short || delay == NULL, "no unconditional delay branch");
2648
2649 } // End of per-block dump
2650 tty->cr();
2651
2652 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2653 }
2654 #endif
2655
2656 //------------------------------Final_Reshape_Counts---------------------------
2657 // This class defines counters to help identify when a method
2658 // may/must be executed using hardware with only 24-bit precision.
2659 struct Final_Reshape_Counts : public StackObj {
2660 int _call_count; // count non-inlined 'common' calls
2661 int _float_count; // count float ops requiring 24-bit precision
2662 int _double_count; // count double ops requiring more precision
2663 int _java_call_count; // count non-inlined 'java' calls
2664 int _inner_loop_count; // count loops which need alignment
2665 VectorSet _visited; // Visitation flags
2666 Node_List _tests; // Set of IfNodes & PCTableNodes
2667
2668 Final_Reshape_Counts() :
2669 _call_count(0), _float_count(0), _double_count(0),
2670 _java_call_count(0), _inner_loop_count(0),
2671 _visited( Thread::current()->resource_area() ) { }
2672
2673 void inc_call_count () { _call_count ++; }
2674 void inc_float_count () { _float_count ++; }
2675 void inc_double_count() { _double_count++; }
2676 void inc_java_call_count() { _java_call_count++; }
2677 void inc_inner_loop_count() { _inner_loop_count++; }
2678
2679 int get_call_count () const { return _call_count ; }
2680 int get_float_count () const { return _float_count ; }
2681 int get_double_count() const { return _double_count; }
2682 int get_java_call_count() const { return _java_call_count; }
2683 int get_inner_loop_count() const { return _inner_loop_count; }
2684 };
2685
2686 #ifdef ASSERT
2687 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2688 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2689 // Make sure the offset goes inside the instance layout.
2690 return k->contains_field_offset(tp->offset());
2691 // Note that OffsetBot and OffsetTop are very negative.
2692 }
2693 #endif
2694
2695 // Eliminate trivially redundant StoreCMs and accumulate their
2696 // precedence edges.
2697 void Compile::eliminate_redundant_card_marks(Node* n) {
2698 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2699 if (n->in(MemNode::Address)->outcnt() > 1) {
2700 // There are multiple users of the same address so it might be
2701 // possible to eliminate some of the StoreCMs
2702 Node* mem = n->in(MemNode::Memory);
2703 Node* adr = n->in(MemNode::Address);
2704 Node* val = n->in(MemNode::ValueIn);
2705 Node* prev = n;
2706 bool done = false;
2707 // Walk the chain of StoreCMs eliminating ones that match. As
2708 // long as it's a chain of single users then the optimization is
2709 // safe. Eliminating partially redundant StoreCMs would require
2710 // cloning copies down the other paths.
2711 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2712 if (adr == mem->in(MemNode::Address) &&
2713 val == mem->in(MemNode::ValueIn)) {
2714 // redundant StoreCM
2715 if (mem->req() > MemNode::OopStore) {
2716 // Hasn't been processed by this code yet.
2717 n->add_prec(mem->in(MemNode::OopStore));
2718 } else {
2719 // Already converted to precedence edge
2720 for (uint i = mem->req(); i < mem->len(); i++) {
2721 // Accumulate any precedence edges
2722 if (mem->in(i) != NULL) {
2723 n->add_prec(mem->in(i));
2724 }
2725 }
2726 // Everything above this point has been processed.
2727 done = true;
2728 }
2729 // Eliminate the previous StoreCM
2730 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2731 assert(mem->outcnt() == 0, "should be dead");
2732 mem->disconnect_inputs(NULL, this);
2733 } else {
2734 prev = mem;
2735 }
2736 mem = prev->in(MemNode::Memory);
2737 }
2738 }
2739 }
2740
2741 //------------------------------final_graph_reshaping_impl----------------------
2742 // Implement items 1-5 from final_graph_reshaping below.
2743 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2744
2745 if ( n->outcnt() == 0 ) return; // dead node
2746 uint nop = n->Opcode();
2747
2748 // Check for 2-input instruction with "last use" on right input.
2749 // Swap to left input. Implements item (2).
2750 if( n->req() == 3 && // two-input instruction
2751 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2752 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2753 n->in(2)->outcnt() == 1 &&// right use IS a last use
2754 !n->in(2)->is_Con() ) { // right use is not a constant
2755 // Check for commutative opcode
2756 switch( nop ) {
2757 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2758 case Op_MaxI: case Op_MinI:
2759 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2760 case Op_AndL: case Op_XorL: case Op_OrL:
2761 case Op_AndI: case Op_XorI: case Op_OrI: {
2762 // Move "last use" input to left by swapping inputs
2763 n->swap_edges(1, 2);
2764 break;
2765 }
2766 default:
2767 break;
2768 }
2769 }
2770
2771 #ifdef ASSERT
2772 if( n->is_Mem() ) {
2773 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2774 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2775 // oop will be recorded in oop map if load crosses safepoint
2776 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2777 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2778 "raw memory operations should have control edge");
2779 }
2780 if (n->is_MemBar()) {
2781 MemBarNode* mb = n->as_MemBar();
2782 if (mb->trailing_store() || mb->trailing_load_store()) {
2783 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2784 Node* mem = mb->in(MemBarNode::Precedent);
2785 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2786 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2787 } else if (mb->leading()) {
2788 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2789 }
2790 }
2791 #endif
2792 // Count FPU ops and common calls, implements item (3)
2793 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n);
2794 switch( nop ) {
2795 // Count all float operations that may use FPU
2796 case Op_AddF:
2797 case Op_SubF:
2798 case Op_MulF:
2799 case Op_DivF:
2800 case Op_NegF:
2801 case Op_ModF:
2802 case Op_ConvI2F:
2803 case Op_ConF:
2804 case Op_CmpF:
2805 case Op_CmpF3:
2806 // case Op_ConvL2F: // longs are split into 32-bit halves
2807 frc.inc_float_count();
2808 break;
2809
2810 case Op_ConvF2D:
2811 case Op_ConvD2F:
2812 frc.inc_float_count();
2813 frc.inc_double_count();
2814 break;
2815
2816 // Count all double operations that may use FPU
2817 case Op_AddD:
2818 case Op_SubD:
2819 case Op_MulD:
2820 case Op_DivD:
2821 case Op_NegD:
2822 case Op_ModD:
2823 case Op_ConvI2D:
2824 case Op_ConvD2I:
2825 // case Op_ConvL2D: // handled by leaf call
2826 // case Op_ConvD2L: // handled by leaf call
2827 case Op_ConD:
2828 case Op_CmpD:
2829 case Op_CmpD3:
2830 frc.inc_double_count();
2831 break;
2832 case Op_Opaque1: // Remove Opaque Nodes before matching
2833 case Op_Opaque2: // Remove Opaque Nodes before matching
2834 case Op_Opaque3:
2835 n->subsume_by(n->in(1), this);
2836 break;
2837 case Op_CallStaticJava:
2838 case Op_CallJava:
2839 case Op_CallDynamicJava:
2840 frc.inc_java_call_count(); // Count java call site;
2841 case Op_CallRuntime:
2842 case Op_CallLeaf:
2843 case Op_CallLeafNoFP: {
2844 assert (n->is_Call(), "");
2845 CallNode *call = n->as_Call();
2846 // Count call sites where the FP mode bit would have to be flipped.
2847 // Do not count uncommon runtime calls:
2848 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2849 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2850 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2851 frc.inc_call_count(); // Count the call site
2852 } else { // See if uncommon argument is shared
2853 Node *n = call->in(TypeFunc::Parms);
2854 int nop = n->Opcode();
2855 // Clone shared simple arguments to uncommon calls, item (1).
2856 if (n->outcnt() > 1 &&
2857 !n->is_Proj() &&
2858 nop != Op_CreateEx &&
2859 nop != Op_CheckCastPP &&
2860 nop != Op_DecodeN &&
2861 nop != Op_DecodeNKlass &&
2862 !n->is_Mem() &&
2863 !n->is_Phi()) {
2864 Node *x = n->clone();
2865 call->set_req(TypeFunc::Parms, x);
2866 }
2867 }
2868 break;
2869 }
2870
2871 case Op_StoreD:
2872 case Op_LoadD:
2873 case Op_LoadD_unaligned:
2874 frc.inc_double_count();
2875 goto handle_mem;
2876 case Op_StoreF:
2877 case Op_LoadF:
2878 frc.inc_float_count();
2879 goto handle_mem;
2880
2881 case Op_StoreCM:
2882 {
2883 // Convert OopStore dependence into precedence edge
2884 Node* prec = n->in(MemNode::OopStore);
2885 n->del_req(MemNode::OopStore);
2886 n->add_prec(prec);
2887 eliminate_redundant_card_marks(n);
2888 }
2889
2890 // fall through
2891
2892 case Op_StoreB:
2893 case Op_StoreC:
2894 case Op_StorePConditional:
2895 case Op_StoreI:
2896 case Op_StoreL:
2897 case Op_StoreIConditional:
2898 case Op_StoreLConditional:
2899 case Op_CompareAndSwapB:
2900 case Op_CompareAndSwapS:
2901 case Op_CompareAndSwapI:
2902 case Op_CompareAndSwapL:
2903 case Op_CompareAndSwapP:
2904 case Op_CompareAndSwapN:
2905 case Op_WeakCompareAndSwapB:
2906 case Op_WeakCompareAndSwapS:
2907 case Op_WeakCompareAndSwapI:
2908 case Op_WeakCompareAndSwapL:
2909 case Op_WeakCompareAndSwapP:
2910 case Op_WeakCompareAndSwapN:
2911 case Op_CompareAndExchangeB:
2912 case Op_CompareAndExchangeS:
2913 case Op_CompareAndExchangeI:
2914 case Op_CompareAndExchangeL:
2915 case Op_CompareAndExchangeP:
2916 case Op_CompareAndExchangeN:
2917 case Op_GetAndAddS:
2918 case Op_GetAndAddB:
2919 case Op_GetAndAddI:
2920 case Op_GetAndAddL:
2921 case Op_GetAndSetS:
2922 case Op_GetAndSetB:
2923 case Op_GetAndSetI:
2924 case Op_GetAndSetL:
2925 case Op_GetAndSetP:
2926 case Op_GetAndSetN:
2927 case Op_StoreP:
2928 case Op_StoreN:
2929 case Op_StoreNKlass:
2930 case Op_LoadB:
2931 case Op_LoadUB:
2932 case Op_LoadUS:
2933 case Op_LoadI:
2934 case Op_LoadKlass:
2935 case Op_LoadNKlass:
2936 case Op_LoadL:
2937 case Op_LoadL_unaligned:
2938 case Op_LoadPLocked:
2939 case Op_LoadP:
2940 #if INCLUDE_ZGC
2941 case Op_LoadBarrierSlowReg:
2942 case Op_LoadBarrierWeakSlowReg:
2943 #endif
2944 case Op_LoadN:
2945 case Op_LoadRange:
2946 case Op_LoadS: {
2947 handle_mem:
2948 #ifdef ASSERT
2949 if( VerifyOptoOopOffsets ) {
2950 assert( n->is_Mem(), "" );
2951 MemNode *mem = (MemNode*)n;
2952 // Check to see if address types have grounded out somehow.
2953 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2954 assert( !tp || oop_offset_is_sane(tp), "" );
2955 }
2956 #endif
2957 break;
2958 }
2959
2960 case Op_AddP: { // Assert sane base pointers
2961 Node *addp = n->in(AddPNode::Address);
2962 assert( !addp->is_AddP() ||
2963 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2964 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2965 "Base pointers must match (addp %u)", addp->_idx );
2966 #ifdef _LP64
2967 if ((UseCompressedOops || UseCompressedClassPointers) &&
2968 addp->Opcode() == Op_ConP &&
2969 addp == n->in(AddPNode::Base) &&
2970 n->in(AddPNode::Offset)->is_Con()) {
2971 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2972 // on the platform and on the compressed oops mode.
2973 // Use addressing with narrow klass to load with offset on x86.
2974 // Some platforms can use the constant pool to load ConP.
2975 // Do this transformation here since IGVN will convert ConN back to ConP.
2976 const Type* t = addp->bottom_type();
2977 bool is_oop = t->isa_oopptr() != NULL;
2978 bool is_klass = t->isa_klassptr() != NULL;
2979
2980 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2981 (is_klass && Matcher::const_klass_prefer_decode())) {
2982 Node* nn = NULL;
2983
2984 int op = is_oop ? Op_ConN : Op_ConNKlass;
2985
2986 // Look for existing ConN node of the same exact type.
2987 Node* r = root();
2988 uint cnt = r->outcnt();
2989 for (uint i = 0; i < cnt; i++) {
2990 Node* m = r->raw_out(i);
2991 if (m!= NULL && m->Opcode() == op &&
2992 m->bottom_type()->make_ptr() == t) {
2993 nn = m;
2994 break;
2995 }
2996 }
2997 if (nn != NULL) {
2998 // Decode a narrow oop to match address
2999 // [R12 + narrow_oop_reg<<3 + offset]
3000 if (is_oop) {
3001 nn = new DecodeNNode(nn, t);
3002 } else {
3003 nn = new DecodeNKlassNode(nn, t);
3004 }
3005 // Check for succeeding AddP which uses the same Base.
3006 // Otherwise we will run into the assertion above when visiting that guy.
3007 for (uint i = 0; i < n->outcnt(); ++i) {
3008 Node *out_i = n->raw_out(i);
3009 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3010 out_i->set_req(AddPNode::Base, nn);
3011 #ifdef ASSERT
3012 for (uint j = 0; j < out_i->outcnt(); ++j) {
3013 Node *out_j = out_i->raw_out(j);
3014 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3015 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3016 }
3017 #endif
3018 }
3019 }
3020 n->set_req(AddPNode::Base, nn);
3021 n->set_req(AddPNode::Address, nn);
3022 if (addp->outcnt() == 0) {
3023 addp->disconnect_inputs(NULL, this);
3024 }
3025 }
3026 }
3027 }
3028 #endif
3029 // platform dependent reshaping of the address expression
3030 reshape_address(n->as_AddP());
3031 break;
3032 }
3033
3034 case Op_CastPP: {
3035 // Remove CastPP nodes to gain more freedom during scheduling but
3036 // keep the dependency they encode as control or precedence edges
3037 // (if control is set already) on memory operations. Some CastPP
3038 // nodes don't have a control (don't carry a dependency): skip
3039 // those.
3040 if (n->in(0) != NULL) {
3041 ResourceMark rm;
3042 Unique_Node_List wq;
3043 wq.push(n);
3044 for (uint next = 0; next < wq.size(); ++next) {
3045 Node *m = wq.at(next);
3046 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3047 Node* use = m->fast_out(i);
3048 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3049 use->ensure_control_or_add_prec(n->in(0));
3050 } else {
3051 switch(use->Opcode()) {
3052 case Op_AddP:
3053 case Op_DecodeN:
3054 case Op_DecodeNKlass:
3055 case Op_CheckCastPP:
3056 case Op_CastPP:
3057 wq.push(use);
3058 break;
3059 }
3060 }
3061 }
3062 }
3063 }
3064 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3065 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3066 Node* in1 = n->in(1);
3067 const Type* t = n->bottom_type();
3068 Node* new_in1 = in1->clone();
3069 new_in1->as_DecodeN()->set_type(t);
3070
3071 if (!Matcher::narrow_oop_use_complex_address()) {
3072 //
3073 // x86, ARM and friends can handle 2 adds in addressing mode
3074 // and Matcher can fold a DecodeN node into address by using
3075 // a narrow oop directly and do implicit NULL check in address:
3076 //
3077 // [R12 + narrow_oop_reg<<3 + offset]
3078 // NullCheck narrow_oop_reg
3079 //
3080 // On other platforms (Sparc) we have to keep new DecodeN node and
3081 // use it to do implicit NULL check in address:
3082 //
3083 // decode_not_null narrow_oop_reg, base_reg
3084 // [base_reg + offset]
3085 // NullCheck base_reg
3086 //
3087 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3088 // to keep the information to which NULL check the new DecodeN node
3089 // corresponds to use it as value in implicit_null_check().
3090 //
3091 new_in1->set_req(0, n->in(0));
3092 }
3093
3094 n->subsume_by(new_in1, this);
3095 if (in1->outcnt() == 0) {
3096 in1->disconnect_inputs(NULL, this);
3097 }
3098 } else {
3099 n->subsume_by(n->in(1), this);
3100 if (n->outcnt() == 0) {
3101 n->disconnect_inputs(NULL, this);
3102 }
3103 }
3104 break;
3105 }
3106 #ifdef _LP64
3107 case Op_CmpP:
3108 // Do this transformation here to preserve CmpPNode::sub() and
3109 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3110 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3111 Node* in1 = n->in(1);
3112 Node* in2 = n->in(2);
3113 if (!in1->is_DecodeNarrowPtr()) {
3114 in2 = in1;
3115 in1 = n->in(2);
3116 }
3117 assert(in1->is_DecodeNarrowPtr(), "sanity");
3118
3119 Node* new_in2 = NULL;
3120 if (in2->is_DecodeNarrowPtr()) {
3121 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3122 new_in2 = in2->in(1);
3123 } else if (in2->Opcode() == Op_ConP) {
3124 const Type* t = in2->bottom_type();
3125 if (t == TypePtr::NULL_PTR) {
3126 assert(in1->is_DecodeN(), "compare klass to null?");
3127 // Don't convert CmpP null check into CmpN if compressed
3128 // oops implicit null check is not generated.
3129 // This will allow to generate normal oop implicit null check.
3130 if (Matcher::gen_narrow_oop_implicit_null_checks())
3131 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3132 //
3133 // This transformation together with CastPP transformation above
3134 // will generated code for implicit NULL checks for compressed oops.
3135 //
3136 // The original code after Optimize()
3137 //
3138 // LoadN memory, narrow_oop_reg
3139 // decode narrow_oop_reg, base_reg
3140 // CmpP base_reg, NULL
3141 // CastPP base_reg // NotNull
3142 // Load [base_reg + offset], val_reg
3143 //
3144 // after these transformations will be
3145 //
3146 // LoadN memory, narrow_oop_reg
3147 // CmpN narrow_oop_reg, NULL
3148 // decode_not_null narrow_oop_reg, base_reg
3149 // Load [base_reg + offset], val_reg
3150 //
3151 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3152 // since narrow oops can be used in debug info now (see the code in
3153 // final_graph_reshaping_walk()).
3154 //
3155 // At the end the code will be matched to
3156 // on x86:
3157 //
3158 // Load_narrow_oop memory, narrow_oop_reg
3159 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3160 // NullCheck narrow_oop_reg
3161 //
3162 // and on sparc:
3163 //
3164 // Load_narrow_oop memory, narrow_oop_reg
3165 // decode_not_null narrow_oop_reg, base_reg
3166 // Load [base_reg + offset], val_reg
3167 // NullCheck base_reg
3168 //
3169 } else if (t->isa_oopptr()) {
3170 new_in2 = ConNode::make(t->make_narrowoop());
3171 } else if (t->isa_klassptr()) {
3172 new_in2 = ConNode::make(t->make_narrowklass());
3173 }
3174 }
3175 if (new_in2 != NULL) {
3176 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3177 n->subsume_by(cmpN, this);
3178 if (in1->outcnt() == 0) {
3179 in1->disconnect_inputs(NULL, this);
3180 }
3181 if (in2->outcnt() == 0) {
3182 in2->disconnect_inputs(NULL, this);
3183 }
3184 }
3185 }
3186 break;
3187
3188 case Op_DecodeN:
3189 case Op_DecodeNKlass:
3190 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3191 // DecodeN could be pinned when it can't be fold into
3192 // an address expression, see the code for Op_CastPP above.
3193 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3194 break;
3195
3196 case Op_EncodeP:
3197 case Op_EncodePKlass: {
3198 Node* in1 = n->in(1);
3199 if (in1->is_DecodeNarrowPtr()) {
3200 n->subsume_by(in1->in(1), this);
3201 } else if (in1->Opcode() == Op_ConP) {
3202 const Type* t = in1->bottom_type();
3203 if (t == TypePtr::NULL_PTR) {
3204 assert(t->isa_oopptr(), "null klass?");
3205 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3206 } else if (t->isa_oopptr()) {
3207 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3208 } else if (t->isa_klassptr()) {
3209 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3210 }
3211 }
3212 if (in1->outcnt() == 0) {
3213 in1->disconnect_inputs(NULL, this);
3214 }
3215 break;
3216 }
3217
3218 case Op_Proj: {
3219 if (OptimizeStringConcat) {
3220 ProjNode* p = n->as_Proj();
3221 if (p->_is_io_use) {
3222 // Separate projections were used for the exception path which
3223 // are normally removed by a late inline. If it wasn't inlined
3224 // then they will hang around and should just be replaced with
3225 // the original one.
3226 Node* proj = NULL;
3227 // Replace with just one
3228 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3229 Node *use = i.get();
3230 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3231 proj = use;
3232 break;
3233 }
3234 }
3235 assert(proj != NULL, "must be found");
3236 p->subsume_by(proj, this);
3237 }
3238 }
3239 break;
3240 }
3241
3242 case Op_Phi:
3243 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3244 // The EncodeP optimization may create Phi with the same edges
3245 // for all paths. It is not handled well by Register Allocator.
3246 Node* unique_in = n->in(1);
3247 assert(unique_in != NULL, "");
3248 uint cnt = n->req();
3249 for (uint i = 2; i < cnt; i++) {
3250 Node* m = n->in(i);
3251 assert(m != NULL, "");
3252 if (unique_in != m)
3253 unique_in = NULL;
3254 }
3255 if (unique_in != NULL) {
3256 n->subsume_by(unique_in, this);
3257 }
3258 }
3259 break;
3260
3261 #endif
3262
3263 #ifdef ASSERT
3264 case Op_CastII:
3265 // Verify that all range check dependent CastII nodes were removed.
3266 if (n->isa_CastII()->has_range_check()) {
3267 n->dump(3);
3268 assert(false, "Range check dependent CastII node was not removed");
3269 }
3270 break;
3271 #endif
3272
3273 case Op_ModI:
3274 if (UseDivMod) {
3275 // Check if a%b and a/b both exist
3276 Node* d = n->find_similar(Op_DivI);
3277 if (d) {
3278 // Replace them with a fused divmod if supported
3279 if (Matcher::has_match_rule(Op_DivModI)) {
3280 DivModINode* divmod = DivModINode::make(n);
3281 d->subsume_by(divmod->div_proj(), this);
3282 n->subsume_by(divmod->mod_proj(), this);
3283 } else {
3284 // replace a%b with a-((a/b)*b)
3285 Node* mult = new MulINode(d, d->in(2));
3286 Node* sub = new SubINode(d->in(1), mult);
3287 n->subsume_by(sub, this);
3288 }
3289 }
3290 }
3291 break;
3292
3293 case Op_ModL:
3294 if (UseDivMod) {
3295 // Check if a%b and a/b both exist
3296 Node* d = n->find_similar(Op_DivL);
3297 if (d) {
3298 // Replace them with a fused divmod if supported
3299 if (Matcher::has_match_rule(Op_DivModL)) {
3300 DivModLNode* divmod = DivModLNode::make(n);
3301 d->subsume_by(divmod->div_proj(), this);
3302 n->subsume_by(divmod->mod_proj(), this);
3303 } else {
3304 // replace a%b with a-((a/b)*b)
3305 Node* mult = new MulLNode(d, d->in(2));
3306 Node* sub = new SubLNode(d->in(1), mult);
3307 n->subsume_by(sub, this);
3308 }
3309 }
3310 }
3311 break;
3312
3313 case Op_LoadVector:
3314 case Op_StoreVector:
3315 break;
3316
3317 case Op_AddReductionVI:
3318 case Op_AddReductionVL:
3319 case Op_AddReductionVF:
3320 case Op_AddReductionVD:
3321 case Op_MulReductionVI:
3322 case Op_MulReductionVL:
3323 case Op_MulReductionVF:
3324 case Op_MulReductionVD:
3325 break;
3326
3327 case Op_PackB:
3328 case Op_PackS:
3329 case Op_PackI:
3330 case Op_PackF:
3331 case Op_PackL:
3332 case Op_PackD:
3333 if (n->req()-1 > 2) {
3334 // Replace many operand PackNodes with a binary tree for matching
3335 PackNode* p = (PackNode*) n;
3336 Node* btp = p->binary_tree_pack(1, n->req());
3337 n->subsume_by(btp, this);
3338 }
3339 break;
3340 case Op_Loop:
3341 case Op_CountedLoop:
3342 case Op_OuterStripMinedLoop:
3343 if (n->as_Loop()->is_inner_loop()) {
3344 frc.inc_inner_loop_count();
3345 }
3346 n->as_Loop()->verify_strip_mined(0);
3347 break;
3348 case Op_LShiftI:
3349 case Op_RShiftI:
3350 case Op_URShiftI:
3351 case Op_LShiftL:
3352 case Op_RShiftL:
3353 case Op_URShiftL:
3354 if (Matcher::need_masked_shift_count) {
3355 // The cpu's shift instructions don't restrict the count to the
3356 // lower 5/6 bits. We need to do the masking ourselves.
3357 Node* in2 = n->in(2);
3358 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3359 const TypeInt* t = in2->find_int_type();
3360 if (t != NULL && t->is_con()) {
3361 juint shift = t->get_con();
3362 if (shift > mask) { // Unsigned cmp
3363 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3364 }
3365 } else {
3366 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3367 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3368 n->set_req(2, shift);
3369 }
3370 }
3371 if (in2->outcnt() == 0) { // Remove dead node
3372 in2->disconnect_inputs(NULL, this);
3373 }
3374 }
3375 break;
3376 case Op_MemBarStoreStore:
3377 case Op_MemBarRelease:
3378 // Break the link with AllocateNode: it is no longer useful and
3379 // confuses register allocation.
3380 if (n->req() > MemBarNode::Precedent) {
3381 n->set_req(MemBarNode::Precedent, top());
3382 }
3383 break;
3384 case Op_MemBarAcquire: {
3385 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3386 // At parse time, the trailing MemBarAcquire for a volatile load
3387 // is created with an edge to the load. After optimizations,
3388 // that input may be a chain of Phis. If those phis have no
3389 // other use, then the MemBarAcquire keeps them alive and
3390 // register allocation can be confused.
3391 ResourceMark rm;
3392 Unique_Node_List wq;
3393 wq.push(n->in(MemBarNode::Precedent));
3394 n->set_req(MemBarNode::Precedent, top());
3395 while (wq.size() > 0) {
3396 Node* m = wq.pop();
3397 if (m->outcnt() == 0) {
3398 for (uint j = 0; j < m->req(); j++) {
3399 Node* in = m->in(j);
3400 if (in != NULL) {
3401 wq.push(in);
3402 }
3403 }
3404 m->disconnect_inputs(NULL, this);
3405 }
3406 }
3407 }
3408 break;
3409 }
3410 case Op_RangeCheck: {
3411 RangeCheckNode* rc = n->as_RangeCheck();
3412 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3413 n->subsume_by(iff, this);
3414 frc._tests.push(iff);
3415 break;
3416 }
3417 case Op_ConvI2L: {
3418 if (!Matcher::convi2l_type_required) {
3419 // Code generation on some platforms doesn't need accurate
3420 // ConvI2L types. Widening the type can help remove redundant
3421 // address computations.
3422 n->as_Type()->set_type(TypeLong::INT);
3423 ResourceMark rm;
3424 Node_List wq;
3425 wq.push(n);
3426 for (uint next = 0; next < wq.size(); next++) {
3427 Node *m = wq.at(next);
3428
3429 for(;;) {
3430 // Loop over all nodes with identical inputs edges as m
3431 Node* k = m->find_similar(m->Opcode());
3432 if (k == NULL) {
3433 break;
3434 }
3435 // Push their uses so we get a chance to remove node made
3436 // redundant
3437 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3438 Node* u = k->fast_out(i);
3439 assert(!wq.contains(u), "shouldn't process one node several times");
3440 if (u->Opcode() == Op_LShiftL ||
3441 u->Opcode() == Op_AddL ||
3442 u->Opcode() == Op_SubL ||
3443 u->Opcode() == Op_AddP) {
3444 wq.push(u);
3445 }
3446 }
3447 // Replace all nodes with identical edges as m with m
3448 k->subsume_by(m, this);
3449 }
3450 }
3451 }
3452 break;
3453 }
3454 case Op_CmpUL: {
3455 if (!Matcher::has_match_rule(Op_CmpUL)) {
3456 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3457 // to max_julong if < 0 to make the signed comparison fail.
3458 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3459 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3460 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3461 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3462 Node* andl = new AndLNode(orl, remove_sign_mask);
3463 Node* cmp = new CmpLNode(andl, n->in(2));
3464 n->subsume_by(cmp, this);
3465 }
3466 break;
3467 }
3468 default:
3469 if (!gc_handled) {
3470 assert(!n->is_Call(), "");
3471 assert(!n->is_Mem(), "");
3472 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3473 }
3474 break;
3475 }
3476
3477 // Collect CFG split points
3478 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3479 frc._tests.push(n);
3480 }
3481 }
3482
3483 //------------------------------final_graph_reshaping_walk---------------------
3484 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3485 // requires that the walk visits a node's inputs before visiting the node.
3486 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3487 ResourceArea *area = Thread::current()->resource_area();
3488 Unique_Node_List sfpt(area);
3489
3490 frc._visited.set(root->_idx); // first, mark node as visited
3491 uint cnt = root->req();
3492 Node *n = root;
3493 uint i = 0;
3494 while (true) {
3495 if (i < cnt) {
3496 // Place all non-visited non-null inputs onto stack
3497 Node* m = n->in(i);
3498 ++i;
3499 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3500 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3501 // compute worst case interpreter size in case of a deoptimization
3502 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3503
3504 sfpt.push(m);
3505 }
3506 cnt = m->req();
3507 nstack.push(n, i); // put on stack parent and next input's index
3508 n = m;
3509 i = 0;
3510 }
3511 } else {
3512 // Now do post-visit work
3513 final_graph_reshaping_impl( n, frc );
3514 if (nstack.is_empty())
3515 break; // finished
3516 n = nstack.node(); // Get node from stack
3517 cnt = n->req();
3518 i = nstack.index();
3519 nstack.pop(); // Shift to the next node on stack
3520 }
3521 }
3522
3523 // Skip next transformation if compressed oops are not used.
3524 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3525 (!UseCompressedOops && !UseCompressedClassPointers))
3526 return;
3527
3528 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3529 // It could be done for an uncommon traps or any safepoints/calls
3530 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3531 while (sfpt.size() > 0) {
3532 n = sfpt.pop();
3533 JVMState *jvms = n->as_SafePoint()->jvms();
3534 assert(jvms != NULL, "sanity");
3535 int start = jvms->debug_start();
3536 int end = n->req();
3537 bool is_uncommon = (n->is_CallStaticJava() &&
3538 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3539 for (int j = start; j < end; j++) {
3540 Node* in = n->in(j);
3541 if (in->is_DecodeNarrowPtr()) {
3542 bool safe_to_skip = true;
3543 if (!is_uncommon ) {
3544 // Is it safe to skip?
3545 for (uint i = 0; i < in->outcnt(); i++) {
3546 Node* u = in->raw_out(i);
3547 if (!u->is_SafePoint() ||
3548 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3549 safe_to_skip = false;
3550 }
3551 }
3552 }
3553 if (safe_to_skip) {
3554 n->set_req(j, in->in(1));
3555 }
3556 if (in->outcnt() == 0) {
3557 in->disconnect_inputs(NULL, this);
3558 }
3559 }
3560 }
3561 }
3562 }
3563
3564 //------------------------------final_graph_reshaping--------------------------
3565 // Final Graph Reshaping.
3566 //
3567 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3568 // and not commoned up and forced early. Must come after regular
3569 // optimizations to avoid GVN undoing the cloning. Clone constant
3570 // inputs to Loop Phis; these will be split by the allocator anyways.
3571 // Remove Opaque nodes.
3572 // (2) Move last-uses by commutative operations to the left input to encourage
3573 // Intel update-in-place two-address operations and better register usage
3574 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3575 // calls canonicalizing them back.
3576 // (3) Count the number of double-precision FP ops, single-precision FP ops
3577 // and call sites. On Intel, we can get correct rounding either by
3578 // forcing singles to memory (requires extra stores and loads after each
3579 // FP bytecode) or we can set a rounding mode bit (requires setting and
3580 // clearing the mode bit around call sites). The mode bit is only used
3581 // if the relative frequency of single FP ops to calls is low enough.
3582 // This is a key transform for SPEC mpeg_audio.
3583 // (4) Detect infinite loops; blobs of code reachable from above but not
3584 // below. Several of the Code_Gen algorithms fail on such code shapes,
3585 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3586 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3587 // Detection is by looking for IfNodes where only 1 projection is
3588 // reachable from below or CatchNodes missing some targets.
3589 // (5) Assert for insane oop offsets in debug mode.
3590
3591 bool Compile::final_graph_reshaping() {
3592 // an infinite loop may have been eliminated by the optimizer,
3593 // in which case the graph will be empty.
3594 if (root()->req() == 1) {
3595 record_method_not_compilable("trivial infinite loop");
3596 return true;
3597 }
3598
3599 // Expensive nodes have their control input set to prevent the GVN
3600 // from freely commoning them. There's no GVN beyond this point so
3601 // no need to keep the control input. We want the expensive nodes to
3602 // be freely moved to the least frequent code path by gcm.
3603 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3604 for (int i = 0; i < expensive_count(); i++) {
3605 _expensive_nodes->at(i)->set_req(0, NULL);
3606 }
3607
3608 Final_Reshape_Counts frc;
3609
3610 // Visit everybody reachable!
3611 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3612 Node_Stack nstack(live_nodes() >> 1);
3613 final_graph_reshaping_walk(nstack, root(), frc);
3614
3615 // Check for unreachable (from below) code (i.e., infinite loops).
3616 for( uint i = 0; i < frc._tests.size(); i++ ) {
3617 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3618 // Get number of CFG targets.
3619 // Note that PCTables include exception targets after calls.
3620 uint required_outcnt = n->required_outcnt();
3621 if (n->outcnt() != required_outcnt) {
3622 // Check for a few special cases. Rethrow Nodes never take the
3623 // 'fall-thru' path, so expected kids is 1 less.
3624 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3625 if (n->in(0)->in(0)->is_Call()) {
3626 CallNode *call = n->in(0)->in(0)->as_Call();
3627 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3628 required_outcnt--; // Rethrow always has 1 less kid
3629 } else if (call->req() > TypeFunc::Parms &&
3630 call->is_CallDynamicJava()) {
3631 // Check for null receiver. In such case, the optimizer has
3632 // detected that the virtual call will always result in a null
3633 // pointer exception. The fall-through projection of this CatchNode
3634 // will not be populated.
3635 Node *arg0 = call->in(TypeFunc::Parms);
3636 if (arg0->is_Type() &&
3637 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3638 required_outcnt--;
3639 }
3640 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3641 call->req() > TypeFunc::Parms+1 &&
3642 call->is_CallStaticJava()) {
3643 // Check for negative array length. In such case, the optimizer has
3644 // detected that the allocation attempt will always result in an
3645 // exception. There is no fall-through projection of this CatchNode .
3646 Node *arg1 = call->in(TypeFunc::Parms+1);
3647 if (arg1->is_Type() &&
3648 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3649 required_outcnt--;
3650 }
3651 }
3652 }
3653 }
3654 // Recheck with a better notion of 'required_outcnt'
3655 if (n->outcnt() != required_outcnt) {
3656 record_method_not_compilable("malformed control flow");
3657 return true; // Not all targets reachable!
3658 }
3659 }
3660 // Check that I actually visited all kids. Unreached kids
3661 // must be infinite loops.
3662 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3663 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3664 record_method_not_compilable("infinite loop");
3665 return true; // Found unvisited kid; must be unreach
3666 }
3667
3668 // Here so verification code in final_graph_reshaping_walk()
3669 // always see an OuterStripMinedLoopEnd
3670 if (n->is_OuterStripMinedLoopEnd()) {
3671 IfNode* init_iff = n->as_If();
3672 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3673 n->subsume_by(iff, this);
3674 }
3675 }
3676
3677 // If original bytecodes contained a mixture of floats and doubles
3678 // check if the optimizer has made it homogenous, item (3).
3679 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3680 frc.get_float_count() > 32 &&
3681 frc.get_double_count() == 0 &&
3682 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3683 set_24_bit_selection_and_mode( false, true );
3684 }
3685
3686 set_java_calls(frc.get_java_call_count());
3687 set_inner_loops(frc.get_inner_loop_count());
3688
3689 // No infinite loops, no reason to bail out.
3690 return false;
3691 }
3692
3693 //-----------------------------too_many_traps----------------------------------
3694 // Report if there are too many traps at the current method and bci.
3695 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3696 bool Compile::too_many_traps(ciMethod* method,
3697 int bci,
3698 Deoptimization::DeoptReason reason) {
3699 ciMethodData* md = method->method_data();
3700 if (md->is_empty()) {
3701 // Assume the trap has not occurred, or that it occurred only
3702 // because of a transient condition during start-up in the interpreter.
3703 return false;
3704 }
3705 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3706 if (md->has_trap_at(bci, m, reason) != 0) {
3707 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3708 // Also, if there are multiple reasons, or if there is no per-BCI record,
3709 // assume the worst.
3710 if (log())
3711 log()->elem("observe trap='%s' count='%d'",
3712 Deoptimization::trap_reason_name(reason),
3713 md->trap_count(reason));
3714 return true;
3715 } else {
3716 // Ignore method/bci and see if there have been too many globally.
3717 return too_many_traps(reason, md);
3718 }
3719 }
3720
3721 // Less-accurate variant which does not require a method and bci.
3722 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3723 ciMethodData* logmd) {
3724 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3725 // Too many traps globally.
3726 // Note that we use cumulative trap_count, not just md->trap_count.
3727 if (log()) {
3728 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3729 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3730 Deoptimization::trap_reason_name(reason),
3731 mcount, trap_count(reason));
3732 }
3733 return true;
3734 } else {
3735 // The coast is clear.
3736 return false;
3737 }
3738 }
3739
3740 //--------------------------too_many_recompiles--------------------------------
3741 // Report if there are too many recompiles at the current method and bci.
3742 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3743 // Is not eager to return true, since this will cause the compiler to use
3744 // Action_none for a trap point, to avoid too many recompilations.
3745 bool Compile::too_many_recompiles(ciMethod* method,
3746 int bci,
3747 Deoptimization::DeoptReason reason) {
3748 ciMethodData* md = method->method_data();
3749 if (md->is_empty()) {
3750 // Assume the trap has not occurred, or that it occurred only
3751 // because of a transient condition during start-up in the interpreter.
3752 return false;
3753 }
3754 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3755 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3756 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3757 Deoptimization::DeoptReason per_bc_reason
3758 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3759 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3760 if ((per_bc_reason == Deoptimization::Reason_none
3761 || md->has_trap_at(bci, m, reason) != 0)
3762 // The trap frequency measure we care about is the recompile count:
3763 && md->trap_recompiled_at(bci, m)
3764 && md->overflow_recompile_count() >= bc_cutoff) {
3765 // Do not emit a trap here if it has already caused recompilations.
3766 // Also, if there are multiple reasons, or if there is no per-BCI record,
3767 // assume the worst.
3768 if (log())
3769 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3770 Deoptimization::trap_reason_name(reason),
3771 md->trap_count(reason),
3772 md->overflow_recompile_count());
3773 return true;
3774 } else if (trap_count(reason) != 0
3775 && decompile_count() >= m_cutoff) {
3776 // Too many recompiles globally, and we have seen this sort of trap.
3777 // Use cumulative decompile_count, not just md->decompile_count.
3778 if (log())
3779 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3780 Deoptimization::trap_reason_name(reason),
3781 md->trap_count(reason), trap_count(reason),
3782 md->decompile_count(), decompile_count());
3783 return true;
3784 } else {
3785 // The coast is clear.
3786 return false;
3787 }
3788 }
3789
3790 // Compute when not to trap. Used by matching trap based nodes and
3791 // NullCheck optimization.
3792 void Compile::set_allowed_deopt_reasons() {
3793 _allowed_reasons = 0;
3794 if (is_method_compilation()) {
3795 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3796 assert(rs < BitsPerInt, "recode bit map");
3797 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3798 _allowed_reasons |= nth_bit(rs);
3799 }
3800 }
3801 }
3802 }
3803
3804 #ifndef PRODUCT
3805 //------------------------------verify_graph_edges---------------------------
3806 // Walk the Graph and verify that there is a one-to-one correspondence
3807 // between Use-Def edges and Def-Use edges in the graph.
3808 void Compile::verify_graph_edges(bool no_dead_code) {
3809 if (VerifyGraphEdges) {
3810 ResourceArea *area = Thread::current()->resource_area();
3811 Unique_Node_List visited(area);
3812 // Call recursive graph walk to check edges
3813 _root->verify_edges(visited);
3814 if (no_dead_code) {
3815 // Now make sure that no visited node is used by an unvisited node.
3816 bool dead_nodes = false;
3817 Unique_Node_List checked(area);
3818 while (visited.size() > 0) {
3819 Node* n = visited.pop();
3820 checked.push(n);
3821 for (uint i = 0; i < n->outcnt(); i++) {
3822 Node* use = n->raw_out(i);
3823 if (checked.member(use)) continue; // already checked
3824 if (visited.member(use)) continue; // already in the graph
3825 if (use->is_Con()) continue; // a dead ConNode is OK
3826 // At this point, we have found a dead node which is DU-reachable.
3827 if (!dead_nodes) {
3828 tty->print_cr("*** Dead nodes reachable via DU edges:");
3829 dead_nodes = true;
3830 }
3831 use->dump(2);
3832 tty->print_cr("---");
3833 checked.push(use); // No repeats; pretend it is now checked.
3834 }
3835 }
3836 assert(!dead_nodes, "using nodes must be reachable from root");
3837 }
3838 }
3839 }
3840 #endif
3841
3842 // The Compile object keeps track of failure reasons separately from the ciEnv.
3843 // This is required because there is not quite a 1-1 relation between the
3844 // ciEnv and its compilation task and the Compile object. Note that one
3845 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3846 // to backtrack and retry without subsuming loads. Other than this backtracking
3847 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3848 // by the logic in C2Compiler.
3849 void Compile::record_failure(const char* reason) {
3850 if (log() != NULL) {
3851 log()->elem("failure reason='%s' phase='compile'", reason);
3852 }
3853 if (_failure_reason == NULL) {
3854 // Record the first failure reason.
3855 _failure_reason = reason;
3856 }
3857
3858 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3859 C->print_method(PHASE_FAILURE);
3860 }
3861 _root = NULL; // flush the graph, too
3862 }
3863
3864 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3865 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3866 _phase_name(name), _dolog(CITimeVerbose)
3867 {
3868 if (_dolog) {
3869 C = Compile::current();
3870 _log = C->log();
3871 } else {
3872 C = NULL;
3873 _log = NULL;
3874 }
3875 if (_log != NULL) {
3876 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3877 _log->stamp();
3878 _log->end_head();
3879 }
3880 }
3881
3882 Compile::TracePhase::~TracePhase() {
3883
3884 C = Compile::current();
3885 if (_dolog) {
3886 _log = C->log();
3887 } else {
3888 _log = NULL;
3889 }
3890
3891 #ifdef ASSERT
3892 if (PrintIdealNodeCount) {
3893 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3894 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3895 }
3896
3897 if (VerifyIdealNodeCount) {
3898 Compile::current()->print_missing_nodes();
3899 }
3900 #endif
3901
3902 if (_log != NULL) {
3903 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3904 }
3905 }
3906
3907 //=============================================================================
3908 // Two Constant's are equal when the type and the value are equal.
3909 bool Compile::Constant::operator==(const Constant& other) {
3910 if (type() != other.type() ) return false;
3911 if (can_be_reused() != other.can_be_reused()) return false;
3912 // For floating point values we compare the bit pattern.
3913 switch (type()) {
3914 case T_INT:
3915 case T_FLOAT: return (_v._value.i == other._v._value.i);
3916 case T_LONG:
3917 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3918 case T_OBJECT:
3919 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3920 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3921 case T_METADATA: return (_v._metadata == other._v._metadata);
3922 default: ShouldNotReachHere(); return false;
3923 }
3924 }
3925
3926 static int type_to_size_in_bytes(BasicType t) {
3927 switch (t) {
3928 case T_INT: return sizeof(jint );
3929 case T_LONG: return sizeof(jlong );
3930 case T_FLOAT: return sizeof(jfloat );
3931 case T_DOUBLE: return sizeof(jdouble);
3932 case T_METADATA: return sizeof(Metadata*);
3933 // We use T_VOID as marker for jump-table entries (labels) which
3934 // need an internal word relocation.
3935 case T_VOID:
3936 case T_ADDRESS:
3937 case T_OBJECT: return sizeof(jobject);
3938 default:
3939 ShouldNotReachHere();
3940 return -1;
3941 }
3942 }
3943
3944 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3945 // sort descending
3946 if (a->freq() > b->freq()) return -1;
3947 if (a->freq() < b->freq()) return 1;
3948 return 0;
3949 }
3950
3951 void Compile::ConstantTable::calculate_offsets_and_size() {
3952 // First, sort the array by frequencies.
3953 _constants.sort(qsort_comparator);
3954
3955 #ifdef ASSERT
3956 // Make sure all jump-table entries were sorted to the end of the
3957 // array (they have a negative frequency).
3958 bool found_void = false;
3959 for (int i = 0; i < _constants.length(); i++) {
3960 Constant con = _constants.at(i);
3961 if (con.type() == T_VOID)
3962 found_void = true; // jump-tables
3963 else
3964 assert(!found_void, "wrong sorting");
3965 }
3966 #endif
3967
3968 int offset = 0;
3969 for (int i = 0; i < _constants.length(); i++) {
3970 Constant* con = _constants.adr_at(i);
3971
3972 // Align offset for type.
3973 int typesize = type_to_size_in_bytes(con->type());
3974 offset = align_up(offset, typesize);
3975 con->set_offset(offset); // set constant's offset
3976
3977 if (con->type() == T_VOID) {
3978 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3979 offset = offset + typesize * n->outcnt(); // expand jump-table
3980 } else {
3981 offset = offset + typesize;
3982 }
3983 }
3984
3985 // Align size up to the next section start (which is insts; see
3986 // CodeBuffer::align_at_start).
3987 assert(_size == -1, "already set?");
3988 _size = align_up(offset, (int)CodeEntryAlignment);
3989 }
3990
3991 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3992 MacroAssembler _masm(&cb);
3993 for (int i = 0; i < _constants.length(); i++) {
3994 Constant con = _constants.at(i);
3995 address constant_addr = NULL;
3996 switch (con.type()) {
3997 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
3998 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3999 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4000 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4001 case T_OBJECT: {
4002 jobject obj = con.get_jobject();
4003 int oop_index = _masm.oop_recorder()->find_index(obj);
4004 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4005 break;
4006 }
4007 case T_ADDRESS: {
4008 address addr = (address) con.get_jobject();
4009 constant_addr = _masm.address_constant(addr);
4010 break;
4011 }
4012 // We use T_VOID as marker for jump-table entries (labels) which
4013 // need an internal word relocation.
4014 case T_VOID: {
4015 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4016 // Fill the jump-table with a dummy word. The real value is
4017 // filled in later in fill_jump_table.
4018 address dummy = (address) n;
4019 constant_addr = _masm.address_constant(dummy);
4020 // Expand jump-table
4021 for (uint i = 1; i < n->outcnt(); i++) {
4022 address temp_addr = _masm.address_constant(dummy + i);
4023 assert(temp_addr, "consts section too small");
4024 }
4025 break;
4026 }
4027 case T_METADATA: {
4028 Metadata* obj = con.get_metadata();
4029 int metadata_index = _masm.oop_recorder()->find_index(obj);
4030 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4031 break;
4032 }
4033 default: ShouldNotReachHere();
4034 }
4035 assert(constant_addr, "consts section too small");
4036 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4037 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4038 }
4039 }
4040
4041 int Compile::ConstantTable::find_offset(Constant& con) const {
4042 int idx = _constants.find(con);
4043 guarantee(idx != -1, "constant must be in constant table");
4044 int offset = _constants.at(idx).offset();
4045 guarantee(offset != -1, "constant table not emitted yet?");
4046 return offset;
4047 }
4048
4049 void Compile::ConstantTable::add(Constant& con) {
4050 if (con.can_be_reused()) {
4051 int idx = _constants.find(con);
4052 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4053 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4054 return;
4055 }
4056 }
4057 (void) _constants.append(con);
4058 }
4059
4060 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4061 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4062 Constant con(type, value, b->_freq);
4063 add(con);
4064 return con;
4065 }
4066
4067 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4068 Constant con(metadata);
4069 add(con);
4070 return con;
4071 }
4072
4073 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4074 jvalue value;
4075 BasicType type = oper->type()->basic_type();
4076 switch (type) {
4077 case T_LONG: value.j = oper->constantL(); break;
4078 case T_FLOAT: value.f = oper->constantF(); break;
4079 case T_DOUBLE: value.d = oper->constantD(); break;
4080 case T_OBJECT:
4081 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4082 case T_METADATA: return add((Metadata*)oper->constant()); break;
4083 default: guarantee(false, "unhandled type: %s", type2name(type));
4084 }
4085 return add(n, type, value);
4086 }
4087
4088 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4089 jvalue value;
4090 // We can use the node pointer here to identify the right jump-table
4091 // as this method is called from Compile::Fill_buffer right before
4092 // the MachNodes are emitted and the jump-table is filled (means the
4093 // MachNode pointers do not change anymore).
4094 value.l = (jobject) n;
4095 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4096 add(con);
4097 return con;
4098 }
4099
4100 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4101 // If called from Compile::scratch_emit_size do nothing.
4102 if (Compile::current()->in_scratch_emit_size()) return;
4103
4104 assert(labels.is_nonempty(), "must be");
4105 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4106
4107 // Since MachConstantNode::constant_offset() also contains
4108 // table_base_offset() we need to subtract the table_base_offset()
4109 // to get the plain offset into the constant table.
4110 int offset = n->constant_offset() - table_base_offset();
4111
4112 MacroAssembler _masm(&cb);
4113 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4114
4115 for (uint i = 0; i < n->outcnt(); i++) {
4116 address* constant_addr = &jump_table_base[i];
4117 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));
4118 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4119 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4120 }
4121 }
4122
4123 //----------------------------static_subtype_check-----------------------------
4124 // Shortcut important common cases when superklass is exact:
4125 // (0) superklass is java.lang.Object (can occur in reflective code)
4126 // (1) subklass is already limited to a subtype of superklass => always ok
4127 // (2) subklass does not overlap with superklass => always fail
4128 // (3) superklass has NO subtypes and we can check with a simple compare.
4129 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4130 if (StressReflectiveCode) {
4131 return SSC_full_test; // Let caller generate the general case.
4132 }
4133
4134 if (superk == env()->Object_klass()) {
4135 return SSC_always_true; // (0) this test cannot fail
4136 }
4137
4138 ciType* superelem = superk;
4139 if (superelem->is_array_klass())
4140 superelem = superelem->as_array_klass()->base_element_type();
4141
4142 if (!subk->is_interface()) { // cannot trust static interface types yet
4143 if (subk->is_subtype_of(superk)) {
4144 return SSC_always_true; // (1) false path dead; no dynamic test needed
4145 }
4146 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4147 !superk->is_subtype_of(subk)) {
4148 return SSC_always_false;
4149 }
4150 }
4151
4152 // If casting to an instance klass, it must have no subtypes
4153 if (superk->is_interface()) {
4154 // Cannot trust interfaces yet.
4155 // %%% S.B. superk->nof_implementors() == 1
4156 } else if (superelem->is_instance_klass()) {
4157 ciInstanceKlass* ik = superelem->as_instance_klass();
4158 if (!ik->has_subklass() && !ik->is_interface()) {
4159 if (!ik->is_final()) {
4160 // Add a dependency if there is a chance of a later subclass.
4161 dependencies()->assert_leaf_type(ik);
4162 }
4163 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4164 }
4165 } else {
4166 // A primitive array type has no subtypes.
4167 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4168 }
4169
4170 return SSC_full_test;
4171 }
4172
4173 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4174 #ifdef _LP64
4175 // The scaled index operand to AddP must be a clean 64-bit value.
4176 // Java allows a 32-bit int to be incremented to a negative
4177 // value, which appears in a 64-bit register as a large
4178 // positive number. Using that large positive number as an
4179 // operand in pointer arithmetic has bad consequences.
4180 // On the other hand, 32-bit overflow is rare, and the possibility
4181 // can often be excluded, if we annotate the ConvI2L node with
4182 // a type assertion that its value is known to be a small positive
4183 // number. (The prior range check has ensured this.)
4184 // This assertion is used by ConvI2LNode::Ideal.
4185 int index_max = max_jint - 1; // array size is max_jint, index is one less
4186 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4187 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4188 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4189 #endif
4190 return idx;
4191 }
4192
4193 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4194 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4195 if (ctrl != NULL) {
4196 // Express control dependency by a CastII node with a narrow type.
4197 value = new CastIINode(value, itype, false, true /* range check dependency */);
4198 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4199 // node from floating above the range check during loop optimizations. Otherwise, the
4200 // ConvI2L node may be eliminated independently of the range check, causing the data path
4201 // to become TOP while the control path is still there (although it's unreachable).
4202 value->set_req(0, ctrl);
4203 // Save CastII node to remove it after loop optimizations.
4204 phase->C->add_range_check_cast(value);
4205 value = phase->transform(value);
4206 }
4207 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4208 return phase->transform(new ConvI2LNode(value, ltype));
4209 }
4210
4211 // The message about the current inlining is accumulated in
4212 // _print_inlining_stream and transfered into the _print_inlining_list
4213 // once we know whether inlining succeeds or not. For regular
4214 // inlining, messages are appended to the buffer pointed by
4215 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4216 // a new buffer is added after _print_inlining_idx in the list. This
4217 // way we can update the inlining message for late inlining call site
4218 // when the inlining is attempted again.
4219 void Compile::print_inlining_init() {
4220 if (print_inlining() || print_intrinsics()) {
4221 _print_inlining_stream = new stringStream();
4222 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4223 }
4224 }
4225
4226 void Compile::print_inlining_reinit() {
4227 if (print_inlining() || print_intrinsics()) {
4228 // Re allocate buffer when we change ResourceMark
4229 _print_inlining_stream = new stringStream();
4230 }
4231 }
4232
4233 void Compile::print_inlining_reset() {
4234 _print_inlining_stream->reset();
4235 }
4236
4237 void Compile::print_inlining_commit() {
4238 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4239 // Transfer the message from _print_inlining_stream to the current
4240 // _print_inlining_list buffer and clear _print_inlining_stream.
4241 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4242 print_inlining_reset();
4243 }
4244
4245 void Compile::print_inlining_push() {
4246 // Add new buffer to the _print_inlining_list at current position
4247 _print_inlining_idx++;
4248 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4249 }
4250
4251 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4252 return _print_inlining_list->at(_print_inlining_idx);
4253 }
4254
4255 void Compile::print_inlining_update(CallGenerator* cg) {
4256 if (print_inlining() || print_intrinsics()) {
4257 if (!cg->is_late_inline()) {
4258 if (print_inlining_current().cg() != NULL) {
4259 print_inlining_push();
4260 }
4261 print_inlining_commit();
4262 } else {
4263 if (print_inlining_current().cg() != cg &&
4264 (print_inlining_current().cg() != NULL ||
4265 print_inlining_current().ss()->size() != 0)) {
4266 print_inlining_push();
4267 }
4268 print_inlining_commit();
4269 print_inlining_current().set_cg(cg);
4270 }
4271 }
4272 }
4273
4274 void Compile::print_inlining_move_to(CallGenerator* cg) {
4275 // We resume inlining at a late inlining call site. Locate the
4276 // corresponding inlining buffer so that we can update it.
4277 if (print_inlining()) {
4278 for (int i = 0; i < _print_inlining_list->length(); i++) {
4279 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4280 _print_inlining_idx = i;
4281 return;
4282 }
4283 }
4284 ShouldNotReachHere();
4285 }
4286 }
4287
4288 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4289 if (print_inlining()) {
4290 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4291 assert(print_inlining_current().cg() == cg, "wrong entry");
4292 // replace message with new message
4293 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4294 print_inlining_commit();
4295 print_inlining_current().set_cg(cg);
4296 }
4297 }
4298
4299 void Compile::print_inlining_assert_ready() {
4300 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4301 }
4302
4303 void Compile::process_print_inlining() {
4304 bool do_print_inlining = print_inlining() || print_intrinsics();
4305 if (do_print_inlining || log() != NULL) {
4306 // Print inlining message for candidates that we couldn't inline
4307 // for lack of space
4308 for (int i = 0; i < _late_inlines.length(); i++) {
4309 CallGenerator* cg = _late_inlines.at(i);
4310 if (!cg->is_mh_late_inline()) {
4311 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4312 if (do_print_inlining) {
4313 cg->print_inlining_late(msg);
4314 }
4315 log_late_inline_failure(cg, msg);
4316 }
4317 }
4318 }
4319 if (do_print_inlining) {
4320 ResourceMark rm;
4321 stringStream ss;
4322 for (int i = 0; i < _print_inlining_list->length(); i++) {
4323 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4324 }
4325 size_t end = ss.size();
4326 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4327 strncpy(_print_inlining_output, ss.base(), end+1);
4328 _print_inlining_output[end] = 0;
4329 }
4330 }
4331
4332 void Compile::dump_print_inlining() {
4333 if (_print_inlining_output != NULL) {
4334 tty->print_raw(_print_inlining_output);
4335 }
4336 }
4337
4338 void Compile::log_late_inline(CallGenerator* cg) {
4339 if (log() != NULL) {
4340 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4341 cg->unique_id());
4342 JVMState* p = cg->call_node()->jvms();
4343 while (p != NULL) {
4344 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4345 p = p->caller();
4346 }
4347 log()->tail("late_inline");
4348 }
4349 }
4350
4351 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4352 log_late_inline(cg);
4353 if (log() != NULL) {
4354 log()->inline_fail(msg);
4355 }
4356 }
4357
4358 void Compile::log_inline_id(CallGenerator* cg) {
4359 if (log() != NULL) {
4360 // The LogCompilation tool needs a unique way to identify late
4361 // inline call sites. This id must be unique for this call site in
4362 // this compilation. Try to have it unique across compilations as
4363 // well because it can be convenient when grepping through the log
4364 // file.
4365 // Distinguish OSR compilations from others in case CICountOSR is
4366 // on.
4367 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4368 cg->set_unique_id(id);
4369 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4370 }
4371 }
4372
4373 void Compile::log_inline_failure(const char* msg) {
4374 if (C->log() != NULL) {
4375 C->log()->inline_fail(msg);
4376 }
4377 }
4378
4379
4380 // Dump inlining replay data to the stream.
4381 // Don't change thread state and acquire any locks.
4382 void Compile::dump_inline_data(outputStream* out) {
4383 InlineTree* inl_tree = ilt();
4384 if (inl_tree != NULL) {
4385 out->print(" inline %d", inl_tree->count());
4386 inl_tree->dump_replay_data(out);
4387 }
4388 }
4389
4390 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4391 if (n1->Opcode() < n2->Opcode()) return -1;
4392 else if (n1->Opcode() > n2->Opcode()) return 1;
4393
4394 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4395 for (uint i = 1; i < n1->req(); i++) {
4396 if (n1->in(i) < n2->in(i)) return -1;
4397 else if (n1->in(i) > n2->in(i)) return 1;
4398 }
4399
4400 return 0;
4401 }
4402
4403 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4404 Node* n1 = *n1p;
4405 Node* n2 = *n2p;
4406
4407 return cmp_expensive_nodes(n1, n2);
4408 }
4409
4410 void Compile::sort_expensive_nodes() {
4411 if (!expensive_nodes_sorted()) {
4412 _expensive_nodes->sort(cmp_expensive_nodes);
4413 }
4414 }
4415
4416 bool Compile::expensive_nodes_sorted() const {
4417 for (int i = 1; i < _expensive_nodes->length(); i++) {
4418 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4419 return false;
4420 }
4421 }
4422 return true;
4423 }
4424
4425 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4426 if (_expensive_nodes->length() == 0) {
4427 return false;
4428 }
4429
4430 assert(OptimizeExpensiveOps, "optimization off?");
4431
4432 // Take this opportunity to remove dead nodes from the list
4433 int j = 0;
4434 for (int i = 0; i < _expensive_nodes->length(); i++) {
4435 Node* n = _expensive_nodes->at(i);
4436 if (!n->is_unreachable(igvn)) {
4437 assert(n->is_expensive(), "should be expensive");
4438 _expensive_nodes->at_put(j, n);
4439 j++;
4440 }
4441 }
4442 _expensive_nodes->trunc_to(j);
4443
4444 // Then sort the list so that similar nodes are next to each other
4445 // and check for at least two nodes of identical kind with same data
4446 // inputs.
4447 sort_expensive_nodes();
4448
4449 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4450 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4451 return true;
4452 }
4453 }
4454
4455 return false;
4456 }
4457
4458 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4459 if (_expensive_nodes->length() == 0) {
4460 return;
4461 }
4462
4463 assert(OptimizeExpensiveOps, "optimization off?");
4464
4465 // Sort to bring similar nodes next to each other and clear the
4466 // control input of nodes for which there's only a single copy.
4467 sort_expensive_nodes();
4468
4469 int j = 0;
4470 int identical = 0;
4471 int i = 0;
4472 bool modified = false;
4473 for (; i < _expensive_nodes->length()-1; i++) {
4474 assert(j <= i, "can't write beyond current index");
4475 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4476 identical++;
4477 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4478 continue;
4479 }
4480 if (identical > 0) {
4481 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4482 identical = 0;
4483 } else {
4484 Node* n = _expensive_nodes->at(i);
4485 igvn.replace_input_of(n, 0, NULL);
4486 igvn.hash_insert(n);
4487 modified = true;
4488 }
4489 }
4490 if (identical > 0) {
4491 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4492 } else if (_expensive_nodes->length() >= 1) {
4493 Node* n = _expensive_nodes->at(i);
4494 igvn.replace_input_of(n, 0, NULL);
4495 igvn.hash_insert(n);
4496 modified = true;
4497 }
4498 _expensive_nodes->trunc_to(j);
4499 if (modified) {
4500 igvn.optimize();
4501 }
4502 }
4503
4504 void Compile::add_expensive_node(Node * n) {
4505 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4506 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4507 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4508 if (OptimizeExpensiveOps) {
4509 _expensive_nodes->append(n);
4510 } else {
4511 // Clear control input and let IGVN optimize expensive nodes if
4512 // OptimizeExpensiveOps is off.
4513 n->set_req(0, NULL);
4514 }
4515 }
4516
4517 /**
4518 * Remove the speculative part of types and clean up the graph
4519 */
4520 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4521 if (UseTypeSpeculation) {
4522 Unique_Node_List worklist;
4523 worklist.push(root());
4524 int modified = 0;
4525 // Go over all type nodes that carry a speculative type, drop the
4526 // speculative part of the type and enqueue the node for an igvn
4527 // which may optimize it out.
4528 for (uint next = 0; next < worklist.size(); ++next) {
4529 Node *n = worklist.at(next);
4530 if (n->is_Type()) {
4531 TypeNode* tn = n->as_Type();
4532 const Type* t = tn->type();
4533 const Type* t_no_spec = t->remove_speculative();
4534 if (t_no_spec != t) {
4535 bool in_hash = igvn.hash_delete(n);
4536 assert(in_hash, "node should be in igvn hash table");
4537 tn->set_type(t_no_spec);
4538 igvn.hash_insert(n);
4539 igvn._worklist.push(n); // give it a chance to go away
4540 modified++;
4541 }
4542 }
4543 uint max = n->len();
4544 for( uint i = 0; i < max; ++i ) {
4545 Node *m = n->in(i);
4546 if (not_a_node(m)) continue;
4547 worklist.push(m);
4548 }
4549 }
4550 // Drop the speculative part of all types in the igvn's type table
4551 igvn.remove_speculative_types();
4552 if (modified > 0) {
4553 igvn.optimize();
4554 }
4555 #ifdef ASSERT
4556 // Verify that after the IGVN is over no speculative type has resurfaced
4557 worklist.clear();
4558 worklist.push(root());
4559 for (uint next = 0; next < worklist.size(); ++next) {
4560 Node *n = worklist.at(next);
4561 const Type* t = igvn.type_or_null(n);
4562 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4563 if (n->is_Type()) {
4564 t = n->as_Type()->type();
4565 assert(t == t->remove_speculative(), "no more speculative types");
4566 }
4567 uint max = n->len();
4568 for( uint i = 0; i < max; ++i ) {
4569 Node *m = n->in(i);
4570 if (not_a_node(m)) continue;
4571 worklist.push(m);
4572 }
4573 }
4574 igvn.check_no_speculative_types();
4575 #endif
4576 }
4577 }
4578
4579 // Auxiliary method to support randomized stressing/fuzzing.
4580 //
4581 // This method can be called the arbitrary number of times, with current count
4582 // as the argument. The logic allows selecting a single candidate from the
4583 // running list of candidates as follows:
4584 // int count = 0;
4585 // Cand* selected = null;
4586 // while(cand = cand->next()) {
4587 // if (randomized_select(++count)) {
4588 // selected = cand;
4589 // }
4590 // }
4591 //
4592 // Including count equalizes the chances any candidate is "selected".
4593 // This is useful when we don't have the complete list of candidates to choose
4594 // from uniformly. In this case, we need to adjust the randomicity of the
4595 // selection, or else we will end up biasing the selection towards the latter
4596 // candidates.
4597 //
4598 // Quick back-envelope calculation shows that for the list of n candidates
4599 // the equal probability for the candidate to persist as "best" can be
4600 // achieved by replacing it with "next" k-th candidate with the probability
4601 // of 1/k. It can be easily shown that by the end of the run, the
4602 // probability for any candidate is converged to 1/n, thus giving the
4603 // uniform distribution among all the candidates.
4604 //
4605 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4606 #define RANDOMIZED_DOMAIN_POW 29
4607 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4608 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4609 bool Compile::randomized_select(int count) {
4610 assert(count > 0, "only positive");
4611 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4612 }
4613
4614 CloneMap& Compile::clone_map() { return _clone_map; }
4615 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4616
4617 void NodeCloneInfo::dump() const {
4618 tty->print(" {%d:%d} ", idx(), gen());
4619 }
4620
4621 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4622 uint64_t val = value(old->_idx);
4623 NodeCloneInfo cio(val);
4624 assert(val != 0, "old node should be in the map");
4625 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4626 insert(nnn->_idx, cin.get());
4627 #ifndef PRODUCT
4628 if (is_debug()) {
4629 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4630 }
4631 #endif
4632 }
4633
4634 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4635 NodeCloneInfo cio(value(old->_idx));
4636 if (cio.get() == 0) {
4637 cio.set(old->_idx, 0);
4638 insert(old->_idx, cio.get());
4639 #ifndef PRODUCT
4640 if (is_debug()) {
4641 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4642 }
4643 #endif
4644 }
4645 clone(old, nnn, gen);
4646 }
4647
4648 int CloneMap::max_gen() const {
4649 int g = 0;
4650 DictI di(_dict);
4651 for(; di.test(); ++di) {
4652 int t = gen(di._key);
4653 if (g < t) {
4654 g = t;
4655 #ifndef PRODUCT
4656 if (is_debug()) {
4657 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4658 }
4659 #endif
4660 }
4661 }
4662 return g;
4663 }
4664
4665 void CloneMap::dump(node_idx_t key) const {
4666 uint64_t val = value(key);
4667 if (val != 0) {
4668 NodeCloneInfo ni(val);
4669 ni.dump();
4670 }
4671 }