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_ZGC
80 #include "gc/z/c2/zBarrierSetC2.hpp"
81 #endif
82 #if INCLUDE_SHENANDOAHGC
83 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.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 if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1472 ta = tj->isa_aryptr();
1473 } else { // Random constant offset into array body
1474 offset = Type::OffsetBot; // Flatten constant access into array body
1475 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1476 }
1477 }
1478 // Arrays of fixed size alias with arrays of unknown size.
1479 if (ta->size() != TypeInt::POS) {
1480 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1481 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1482 }
1483 // Arrays of known objects become arrays of unknown objects.
1484 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1485 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1486 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1487 }
1488 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1489 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1490 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1491 }
1492 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1493 // cannot be distinguished by bytecode alone.
1494 if (ta->elem() == TypeInt::BOOL) {
1495 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1496 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1497 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1498 }
1499 // During the 2nd round of IterGVN, NotNull castings are removed.
1500 // Make sure the Bottom and NotNull variants alias the same.
1501 // Also, make sure exact and non-exact variants alias the same.
1502 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1503 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1504 }
1505 }
1506
1507 // Oop pointers need some flattening
1508 const TypeInstPtr *to = tj->isa_instptr();
1509 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1510 ciInstanceKlass *k = to->klass()->as_instance_klass();
1511 if( ptr == TypePtr::Constant ) {
1512 if (to->klass() != ciEnv::current()->Class_klass() ||
1513 offset < k->size_helper() * wordSize) {
1514 // No constant oop pointers (such as Strings); they alias with
1515 // unknown strings.
1516 assert(!is_known_inst, "not scalarizable allocation");
1517 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1518 }
1519 } else if( is_known_inst ) {
1520 tj = to; // Keep NotNull and klass_is_exact for instance type
1521 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1522 // During the 2nd round of IterGVN, NotNull castings are removed.
1523 // Make sure the Bottom and NotNull variants alias the same.
1524 // Also, make sure exact and non-exact variants alias the same.
1525 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1526 }
1527 if (to->speculative() != NULL) {
1528 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1529 }
1530 // Canonicalize the holder of this field
1531 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1532 // First handle header references such as a LoadKlassNode, even if the
1533 // object's klass is unloaded at compile time (4965979).
1534 if (!is_known_inst) { // Do it only for non-instance types
1535 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1536 }
1537 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1538 to = tj->is_instptr();
1539 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1540 // Static fields are in the space above the normal instance
1541 // fields in the java.lang.Class instance.
1542 if (to->klass() != ciEnv::current()->Class_klass()) {
1543 to = NULL;
1544 tj = TypeOopPtr::BOTTOM;
1545 offset = tj->offset();
1546 }
1547 } else {
1548 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1549 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1550 if( is_known_inst ) {
1551 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1552 } else {
1553 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1554 }
1555 }
1556 }
1557 }
1558
1559 // Klass pointers to object array klasses need some flattening
1560 const TypeKlassPtr *tk = tj->isa_klassptr();
1561 if( tk ) {
1562 // If we are referencing a field within a Klass, we need
1563 // to assume the worst case of an Object. Both exact and
1564 // inexact types must flatten to the same alias class so
1565 // use NotNull as the PTR.
1566 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1567
1568 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1569 TypeKlassPtr::OBJECT->klass(),
1570 offset);
1571 }
1572
1573 ciKlass* klass = tk->klass();
1574 if( klass->is_obj_array_klass() ) {
1575 ciKlass* k = TypeAryPtr::OOPS->klass();
1576 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1577 k = TypeInstPtr::BOTTOM->klass();
1578 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1579 }
1580
1581 // Check for precise loads from the primary supertype array and force them
1582 // to the supertype cache alias index. Check for generic array loads from
1583 // the primary supertype array and also force them to the supertype cache
1584 // alias index. Since the same load can reach both, we need to merge
1585 // these 2 disparate memories into the same alias class. Since the
1586 // primary supertype array is read-only, there's no chance of confusion
1587 // where we bypass an array load and an array store.
1588 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1589 if (offset == Type::OffsetBot ||
1590 (offset >= primary_supers_offset &&
1591 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1592 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1593 offset = in_bytes(Klass::secondary_super_cache_offset());
1594 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1595 }
1596 }
1597
1598 // Flatten all Raw pointers together.
1599 if (tj->base() == Type::RawPtr)
1600 tj = TypeRawPtr::BOTTOM;
1601
1602 if (tj->base() == Type::AnyPtr)
1603 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1604
1605 // Flatten all to bottom for now
1606 switch( _AliasLevel ) {
1607 case 0:
1608 tj = TypePtr::BOTTOM;
1609 break;
1610 case 1: // Flatten to: oop, static, field or array
1611 switch (tj->base()) {
1612 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1613 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1614 case Type::AryPtr: // do not distinguish arrays at all
1615 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1616 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1617 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1618 default: ShouldNotReachHere();
1619 }
1620 break;
1621 case 2: // No collapsing at level 2; keep all splits
1622 case 3: // No collapsing at level 3; keep all splits
1623 break;
1624 default:
1625 Unimplemented();
1626 }
1627
1628 offset = tj->offset();
1629 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1630
1631 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1632 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1633 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1634 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1635 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1636 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1637 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1638 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1639 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1640 assert( tj->ptr() != TypePtr::TopPTR &&
1641 tj->ptr() != TypePtr::AnyNull &&
1642 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1643 // assert( tj->ptr() != TypePtr::Constant ||
1644 // tj->base() == Type::RawPtr ||
1645 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1646
1647 return tj;
1648 }
1649
1650 void Compile::AliasType::Init(int i, const TypePtr* at) {
1651 _index = i;
1652 _adr_type = at;
1653 _field = NULL;
1654 _element = NULL;
1655 _is_rewritable = true; // default
1656 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1657 if (atoop != NULL && atoop->is_known_instance()) {
1658 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1659 _general_index = Compile::current()->get_alias_index(gt);
1660 } else {
1661 _general_index = 0;
1662 }
1663 }
1664
1665 BasicType Compile::AliasType::basic_type() const {
1666 if (element() != NULL) {
1667 const Type* element = adr_type()->is_aryptr()->elem();
1668 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1669 } if (field() != NULL) {
1670 return field()->layout_type();
1671 } else {
1672 return T_ILLEGAL; // unknown
1673 }
1674 }
1675
1676 //---------------------------------print_on------------------------------------
1677 #ifndef PRODUCT
1678 void Compile::AliasType::print_on(outputStream* st) {
1679 if (index() < 10)
1680 st->print("@ <%d> ", index());
1681 else st->print("@ <%d>", index());
1682 st->print(is_rewritable() ? " " : " RO");
1683 int offset = adr_type()->offset();
1684 if (offset == Type::OffsetBot)
1685 st->print(" +any");
1686 else st->print(" +%-3d", offset);
1687 st->print(" in ");
1688 adr_type()->dump_on(st);
1689 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1690 if (field() != NULL && tjp) {
1691 if (tjp->klass() != field()->holder() ||
1692 tjp->offset() != field()->offset_in_bytes()) {
1693 st->print(" != ");
1694 field()->print();
1695 st->print(" ***");
1696 }
1697 }
1698 }
1699
1700 void print_alias_types() {
1701 Compile* C = Compile::current();
1702 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1703 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1704 C->alias_type(idx)->print_on(tty);
1705 tty->cr();
1706 }
1707 }
1708 #endif
1709
1710
1711 //----------------------------probe_alias_cache--------------------------------
1712 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1713 intptr_t key = (intptr_t) adr_type;
1714 key ^= key >> logAliasCacheSize;
1715 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1716 }
1717
1718
1719 //-----------------------------grow_alias_types--------------------------------
1720 void Compile::grow_alias_types() {
1721 const int old_ats = _max_alias_types; // how many before?
1722 const int new_ats = old_ats; // how many more?
1723 const int grow_ats = old_ats+new_ats; // how many now?
1724 _max_alias_types = grow_ats;
1725 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1726 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1727 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1728 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1729 }
1730
1731
1732 //--------------------------------find_alias_type------------------------------
1733 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1734 if (_AliasLevel == 0)
1735 return alias_type(AliasIdxBot);
1736
1737 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1738 if (ace->_adr_type == adr_type) {
1739 return alias_type(ace->_index);
1740 }
1741
1742 // Handle special cases.
1743 if (adr_type == NULL) return alias_type(AliasIdxTop);
1744 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1745
1746 // Do it the slow way.
1747 const TypePtr* flat = flatten_alias_type(adr_type);
1748
1749 #ifdef ASSERT
1750 {
1751 ResourceMark rm;
1752 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1753 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1754 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1755 Type::str(adr_type));
1756 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1757 const TypeOopPtr* foop = flat->is_oopptr();
1758 // Scalarizable allocations have exact klass always.
1759 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1760 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1761 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1762 Type::str(foop), Type::str(xoop));
1763 }
1764 }
1765 #endif
1766
1767 int idx = AliasIdxTop;
1768 for (int i = 0; i < num_alias_types(); i++) {
1769 if (alias_type(i)->adr_type() == flat) {
1770 idx = i;
1771 break;
1772 }
1773 }
1774
1775 if (idx == AliasIdxTop) {
1776 if (no_create) return NULL;
1777 // Grow the array if necessary.
1778 if (_num_alias_types == _max_alias_types) grow_alias_types();
1779 // Add a new alias type.
1780 idx = _num_alias_types++;
1781 _alias_types[idx]->Init(idx, flat);
1782 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1783 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1784 if (flat->isa_instptr()) {
1785 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1786 && flat->is_instptr()->klass() == env()->Class_klass())
1787 alias_type(idx)->set_rewritable(false);
1788 }
1789 if (flat->isa_aryptr()) {
1790 #ifdef ASSERT
1791 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1792 // (T_BYTE has the weakest alignment and size restrictions...)
1793 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1794 #endif
1795 if (flat->offset() == TypePtr::OffsetBot) {
1796 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1797 }
1798 }
1799 if (flat->isa_klassptr()) {
1800 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1801 alias_type(idx)->set_rewritable(false);
1802 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1803 alias_type(idx)->set_rewritable(false);
1804 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1805 alias_type(idx)->set_rewritable(false);
1806 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1807 alias_type(idx)->set_rewritable(false);
1808 }
1809 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1810 // but the base pointer type is not distinctive enough to identify
1811 // references into JavaThread.)
1812
1813 // Check for final fields.
1814 const TypeInstPtr* tinst = flat->isa_instptr();
1815 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1816 ciField* field;
1817 if (tinst->const_oop() != NULL &&
1818 tinst->klass() == ciEnv::current()->Class_klass() &&
1819 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1820 // static field
1821 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1822 field = k->get_field_by_offset(tinst->offset(), true);
1823 } else {
1824 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1825 field = k->get_field_by_offset(tinst->offset(), false);
1826 }
1827 assert(field == NULL ||
1828 original_field == NULL ||
1829 (field->holder() == original_field->holder() &&
1830 field->offset() == original_field->offset() &&
1831 field->is_static() == original_field->is_static()), "wrong field?");
1832 // Set field() and is_rewritable() attributes.
1833 if (field != NULL) alias_type(idx)->set_field(field);
1834 }
1835 }
1836
1837 // Fill the cache for next time.
1838 ace->_adr_type = adr_type;
1839 ace->_index = idx;
1840 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1841
1842 // Might as well try to fill the cache for the flattened version, too.
1843 AliasCacheEntry* face = probe_alias_cache(flat);
1844 if (face->_adr_type == NULL) {
1845 face->_adr_type = flat;
1846 face->_index = idx;
1847 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1848 }
1849
1850 return alias_type(idx);
1851 }
1852
1853
1854 Compile::AliasType* Compile::alias_type(ciField* field) {
1855 const TypeOopPtr* t;
1856 if (field->is_static())
1857 t = TypeInstPtr::make(field->holder()->java_mirror());
1858 else
1859 t = TypeOopPtr::make_from_klass_raw(field->holder());
1860 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1861 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1862 return atp;
1863 }
1864
1865
1866 //------------------------------have_alias_type--------------------------------
1867 bool Compile::have_alias_type(const TypePtr* adr_type) {
1868 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1869 if (ace->_adr_type == adr_type) {
1870 return true;
1871 }
1872
1873 // Handle special cases.
1874 if (adr_type == NULL) return true;
1875 if (adr_type == TypePtr::BOTTOM) return true;
1876
1877 return find_alias_type(adr_type, true, NULL) != NULL;
1878 }
1879
1880 //-----------------------------must_alias--------------------------------------
1881 // True if all values of the given address type are in the given alias category.
1882 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1883 if (alias_idx == AliasIdxBot) return true; // the universal category
1884 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1885 if (alias_idx == AliasIdxTop) return false; // the empty category
1886 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1887
1888 // the only remaining possible overlap is identity
1889 int adr_idx = get_alias_index(adr_type);
1890 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1891 assert(adr_idx == alias_idx ||
1892 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1893 && adr_type != TypeOopPtr::BOTTOM),
1894 "should not be testing for overlap with an unsafe pointer");
1895 return adr_idx == alias_idx;
1896 }
1897
1898 //------------------------------can_alias--------------------------------------
1899 // True if any values of the given address type are in the given alias category.
1900 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1901 if (alias_idx == AliasIdxTop) return false; // the empty category
1902 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1903 if (alias_idx == AliasIdxBot) return true; // the universal category
1904 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1905
1906 // the only remaining possible overlap is identity
1907 int adr_idx = get_alias_index(adr_type);
1908 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1909 return adr_idx == alias_idx;
1910 }
1911
1912
1913
1914 //---------------------------pop_warm_call-------------------------------------
1915 WarmCallInfo* Compile::pop_warm_call() {
1916 WarmCallInfo* wci = _warm_calls;
1917 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1918 return wci;
1919 }
1920
1921 //----------------------------Inline_Warm--------------------------------------
1922 int Compile::Inline_Warm() {
1923 // If there is room, try to inline some more warm call sites.
1924 // %%% Do a graph index compaction pass when we think we're out of space?
1925 if (!InlineWarmCalls) return 0;
1926
1927 int calls_made_hot = 0;
1928 int room_to_grow = NodeCountInliningCutoff - unique();
1929 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1930 int amount_grown = 0;
1931 WarmCallInfo* call;
1932 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1933 int est_size = (int)call->size();
1934 if (est_size > (room_to_grow - amount_grown)) {
1935 // This one won't fit anyway. Get rid of it.
1936 call->make_cold();
1937 continue;
1938 }
1939 call->make_hot();
1940 calls_made_hot++;
1941 amount_grown += est_size;
1942 amount_to_grow -= est_size;
1943 }
1944
1945 if (calls_made_hot > 0) set_major_progress();
1946 return calls_made_hot;
1947 }
1948
1949
1950 //----------------------------Finish_Warm--------------------------------------
1951 void Compile::Finish_Warm() {
1952 if (!InlineWarmCalls) return;
1953 if (failing()) return;
1954 if (warm_calls() == NULL) return;
1955
1956 // Clean up loose ends, if we are out of space for inlining.
1957 WarmCallInfo* call;
1958 while ((call = pop_warm_call()) != NULL) {
1959 call->make_cold();
1960 }
1961 }
1962
1963 //---------------------cleanup_loop_predicates-----------------------
1964 // Remove the opaque nodes that protect the predicates so that all unused
1965 // checks and uncommon_traps will be eliminated from the ideal graph
1966 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1967 if (predicate_count()==0) return;
1968 for (int i = predicate_count(); i > 0; i--) {
1969 Node * n = predicate_opaque1_node(i-1);
1970 assert(n->Opcode() == Op_Opaque1, "must be");
1971 igvn.replace_node(n, n->in(1));
1972 }
1973 assert(predicate_count()==0, "should be clean!");
1974 }
1975
1976 void Compile::add_range_check_cast(Node* n) {
1977 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1978 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1979 _range_check_casts->append(n);
1980 }
1981
1982 // Remove all range check dependent CastIINodes.
1983 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1984 for (int i = range_check_cast_count(); i > 0; i--) {
1985 Node* cast = range_check_cast_node(i-1);
1986 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1987 igvn.replace_node(cast, cast->in(1));
1988 }
1989 assert(range_check_cast_count() == 0, "should be empty");
1990 }
1991
1992 void Compile::add_opaque4_node(Node* n) {
1993 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1994 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1995 _opaque4_nodes->append(n);
1996 }
1997
1998 // Remove all Opaque4 nodes.
1999 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2000 for (int i = opaque4_count(); i > 0; i--) {
2001 Node* opaq = opaque4_node(i-1);
2002 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2003 igvn.replace_node(opaq, opaq->in(2));
2004 }
2005 assert(opaque4_count() == 0, "should be empty");
2006 }
2007
2008 // StringOpts and late inlining of string methods
2009 void Compile::inline_string_calls(bool parse_time) {
2010 {
2011 // remove useless nodes to make the usage analysis simpler
2012 ResourceMark rm;
2013 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2014 }
2015
2016 {
2017 ResourceMark rm;
2018 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2019 PhaseStringOpts pso(initial_gvn(), for_igvn());
2020 print_method(PHASE_AFTER_STRINGOPTS, 3);
2021 }
2022
2023 // now inline anything that we skipped the first time around
2024 if (!parse_time) {
2025 _late_inlines_pos = _late_inlines.length();
2026 }
2027
2028 while (_string_late_inlines.length() > 0) {
2029 CallGenerator* cg = _string_late_inlines.pop();
2030 cg->do_late_inline();
2031 if (failing()) return;
2032 }
2033 _string_late_inlines.trunc_to(0);
2034 }
2035
2036 // Late inlining of boxing methods
2037 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2038 if (_boxing_late_inlines.length() > 0) {
2039 assert(has_boxed_value(), "inconsistent");
2040
2041 PhaseGVN* gvn = initial_gvn();
2042 set_inlining_incrementally(true);
2043
2044 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2045 for_igvn()->clear();
2046 gvn->replace_with(&igvn);
2047
2048 _late_inlines_pos = _late_inlines.length();
2049
2050 while (_boxing_late_inlines.length() > 0) {
2051 CallGenerator* cg = _boxing_late_inlines.pop();
2052 cg->do_late_inline();
2053 if (failing()) return;
2054 }
2055 _boxing_late_inlines.trunc_to(0);
2056
2057 {
2058 ResourceMark rm;
2059 PhaseRemoveUseless pru(gvn, for_igvn());
2060 }
2061
2062 igvn = PhaseIterGVN(gvn);
2063 igvn.optimize();
2064
2065 set_inlining_progress(false);
2066 set_inlining_incrementally(false);
2067 }
2068 }
2069
2070 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2071 assert(IncrementalInline, "incremental inlining should be on");
2072 PhaseGVN* gvn = initial_gvn();
2073
2074 set_inlining_progress(false);
2075 for_igvn()->clear();
2076 gvn->replace_with(&igvn);
2077
2078 {
2079 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2080 int i = 0;
2081 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2082 CallGenerator* cg = _late_inlines.at(i);
2083 _late_inlines_pos = i+1;
2084 cg->do_late_inline();
2085 if (failing()) return;
2086 }
2087 int j = 0;
2088 for (; i < _late_inlines.length(); i++, j++) {
2089 _late_inlines.at_put(j, _late_inlines.at(i));
2090 }
2091 _late_inlines.trunc_to(j);
2092 }
2093
2094 {
2095 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2096 ResourceMark rm;
2097 PhaseRemoveUseless pru(gvn, for_igvn());
2098 }
2099
2100 {
2101 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2102 igvn = PhaseIterGVN(gvn);
2103 }
2104 }
2105
2106 // Perform incremental inlining until bound on number of live nodes is reached
2107 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2108 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2109
2110 PhaseGVN* gvn = initial_gvn();
2111
2112 set_inlining_incrementally(true);
2113 set_inlining_progress(true);
2114 uint low_live_nodes = 0;
2115
2116 while(inlining_progress() && _late_inlines.length() > 0) {
2117
2118 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2119 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2120 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2121 // PhaseIdealLoop is expensive so we only try it once we are
2122 // out of live nodes and we only try it again if the previous
2123 // helped got the number of nodes down significantly
2124 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2125 if (failing()) return;
2126 low_live_nodes = live_nodes();
2127 _major_progress = true;
2128 }
2129
2130 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2131 break;
2132 }
2133 }
2134
2135 inline_incrementally_one(igvn);
2136
2137 if (failing()) return;
2138
2139 {
2140 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2141 igvn.optimize();
2142 }
2143
2144 if (failing()) return;
2145 }
2146
2147 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2148
2149 if (_string_late_inlines.length() > 0) {
2150 assert(has_stringbuilder(), "inconsistent");
2151 for_igvn()->clear();
2152 initial_gvn()->replace_with(&igvn);
2153
2154 inline_string_calls(false);
2155
2156 if (failing()) return;
2157
2158 {
2159 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2160 ResourceMark rm;
2161 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2162 }
2163
2164 {
2165 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2166 igvn = PhaseIterGVN(gvn);
2167 igvn.optimize();
2168 }
2169 }
2170
2171 set_inlining_incrementally(false);
2172 }
2173
2174
2175 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2176 if(loop_opts_cnt > 0) {
2177 debug_only( int cnt = 0; );
2178 while(major_progress() && (loop_opts_cnt > 0)) {
2179 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2180 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2181 PhaseIdealLoop ideal_loop(igvn, mode);
2182 loop_opts_cnt--;
2183 if (failing()) return false;
2184 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2185 }
2186 }
2187 return true;
2188 }
2189
2190 //------------------------------Optimize---------------------------------------
2191 // Given a graph, optimize it.
2192 void Compile::Optimize() {
2193 TracePhase tp("optimizer", &timers[_t_optimizer]);
2194
2195 #ifndef PRODUCT
2196 if (_directive->BreakAtCompileOption) {
2197 BREAKPOINT;
2198 }
2199
2200 #endif
2201
2202 #ifdef ASSERT
2203 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2204 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2205 #endif
2206
2207 ResourceMark rm;
2208 int loop_opts_cnt;
2209
2210 print_inlining_reinit();
2211
2212 NOT_PRODUCT( verify_graph_edges(); )
2213
2214 print_method(PHASE_AFTER_PARSING);
2215
2216 {
2217 // Iterative Global Value Numbering, including ideal transforms
2218 // Initialize IterGVN with types and values from parse-time GVN
2219 PhaseIterGVN igvn(initial_gvn());
2220 #ifdef ASSERT
2221 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2222 #endif
2223 {
2224 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2225 igvn.optimize();
2226 }
2227
2228 if (failing()) return;
2229
2230 print_method(PHASE_ITER_GVN1, 2);
2231
2232 inline_incrementally(igvn);
2233
2234 print_method(PHASE_INCREMENTAL_INLINE, 2);
2235
2236 if (failing()) return;
2237
2238 if (eliminate_boxing()) {
2239 // Inline valueOf() methods now.
2240 inline_boxing_calls(igvn);
2241
2242 if (AlwaysIncrementalInline) {
2243 inline_incrementally(igvn);
2244 }
2245
2246 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2247
2248 if (failing()) return;
2249 }
2250
2251 // Remove the speculative part of types and clean up the graph from
2252 // the extra CastPP nodes whose only purpose is to carry them. Do
2253 // that early so that optimizations are not disrupted by the extra
2254 // CastPP nodes.
2255 remove_speculative_types(igvn);
2256
2257 // No more new expensive nodes will be added to the list from here
2258 // so keep only the actual candidates for optimizations.
2259 cleanup_expensive_nodes(igvn);
2260
2261 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2262 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2263 initial_gvn()->replace_with(&igvn);
2264 for_igvn()->clear();
2265 Unique_Node_List new_worklist(C->comp_arena());
2266 {
2267 ResourceMark rm;
2268 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2269 }
2270 set_for_igvn(&new_worklist);
2271 igvn = PhaseIterGVN(initial_gvn());
2272 igvn.optimize();
2273 }
2274
2275 // Perform escape analysis
2276 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2277 if (has_loops()) {
2278 // Cleanup graph (remove dead nodes).
2279 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2280 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2281 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2282 if (failing()) return;
2283 }
2284 ConnectionGraph::do_analysis(this, &igvn);
2285
2286 if (failing()) return;
2287
2288 // Optimize out fields loads from scalar replaceable allocations.
2289 igvn.optimize();
2290 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2291
2292 if (failing()) return;
2293
2294 if (congraph() != NULL && macro_count() > 0) {
2295 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2296 PhaseMacroExpand mexp(igvn);
2297 mexp.eliminate_macro_nodes();
2298 igvn.set_delay_transform(false);
2299
2300 igvn.optimize();
2301 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2302
2303 if (failing()) return;
2304 }
2305 }
2306
2307 // Loop transforms on the ideal graph. Range Check Elimination,
2308 // peeling, unrolling, etc.
2309
2310 // Set loop opts counter
2311 loop_opts_cnt = num_loop_opts();
2312 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2313 {
2314 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2315 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2316 loop_opts_cnt--;
2317 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2318 if (failing()) return;
2319 }
2320 // Loop opts pass if partial peeling occurred in previous pass
2321 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2322 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2323 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2324 loop_opts_cnt--;
2325 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2326 if (failing()) return;
2327 }
2328 // Loop opts pass for loop-unrolling before CCP
2329 if(major_progress() && (loop_opts_cnt > 0)) {
2330 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2331 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2332 loop_opts_cnt--;
2333 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2334 }
2335 if (!failing()) {
2336 // Verify that last round of loop opts produced a valid graph
2337 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2338 PhaseIdealLoop::verify(igvn);
2339 }
2340 }
2341 if (failing()) return;
2342
2343 // Conditional Constant Propagation;
2344 PhaseCCP ccp( &igvn );
2345 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2346 {
2347 TracePhase tp("ccp", &timers[_t_ccp]);
2348 ccp.do_transform();
2349 }
2350 print_method(PHASE_CPP1, 2);
2351
2352 assert( true, "Break here to ccp.dump_old2new_map()");
2353
2354 // Iterative Global Value Numbering, including ideal transforms
2355 {
2356 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2357 igvn = ccp;
2358 igvn.optimize();
2359 }
2360
2361 print_method(PHASE_ITER_GVN2, 2);
2362
2363 if (failing()) return;
2364
2365 // Loop transforms on the ideal graph. Range Check Elimination,
2366 // peeling, unrolling, etc.
2367 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2368 return;
2369 }
2370
2371 #if INCLUDE_ZGC
2372 if (UseZGC) {
2373 ZBarrierSetC2::find_dominating_barriers(igvn);
2374 }
2375 #endif
2376
2377 if (failing()) return;
2378
2379 // Ensure that major progress is now clear
2380 C->clear_major_progress();
2381
2382 {
2383 // Verify that all previous optimizations produced a valid graph
2384 // at least to this point, even if no loop optimizations were done.
2385 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2386 PhaseIdealLoop::verify(igvn);
2387 }
2388
2389 if (range_check_cast_count() > 0) {
2390 // No more loop optimizations. Remove all range check dependent CastIINodes.
2391 C->remove_range_check_casts(igvn);
2392 igvn.optimize();
2393 }
2394
2395 #ifdef ASSERT
2396 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2397 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand);
2398 #endif
2399
2400 {
2401 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2402 PhaseMacroExpand mex(igvn);
2403 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2404 if (mex.expand_macro_nodes()) {
2405 assert(failing(), "must bail out w/ explicit message");
2406 return;
2407 }
2408 }
2409
2410 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2);
2411
2412 #if INCLUDE_SHENANDOAHGC
2413 if (!ShenandoahWriteBarrierNode::expand(this, igvn, loop_opts_cnt)) {
2414 assert(failing(), "must bail out w/ explicit message");
2415 return;
2416 }
2417 #endif
2418
2419 if (opaque4_count() > 0) {
2420 C->remove_opaque4_nodes(igvn);
2421 igvn.optimize();
2422 }
2423
2424 DEBUG_ONLY( _modified_nodes = NULL; )
2425 } // (End scope of igvn; run destructor if necessary for asserts.)
2426
2427 process_print_inlining();
2428 // A method with only infinite loops has no edges entering loops from root
2429 {
2430 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2431 if (final_graph_reshaping()) {
2432 assert(failing(), "must bail out w/ explicit message");
2433 return;
2434 }
2435 }
2436
2437 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2438 }
2439
2440
2441 //------------------------------Code_Gen---------------------------------------
2442 // Given a graph, generate code for it
2443 void Compile::Code_Gen() {
2444 if (failing()) {
2445 return;
2446 }
2447
2448 // Perform instruction selection. You might think we could reclaim Matcher
2449 // memory PDQ, but actually the Matcher is used in generating spill code.
2450 // Internals of the Matcher (including some VectorSets) must remain live
2451 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2452 // set a bit in reclaimed memory.
2453
2454 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2455 // nodes. Mapping is only valid at the root of each matched subtree.
2456 NOT_PRODUCT( verify_graph_edges(); )
2457
2458 Matcher matcher;
2459 _matcher = &matcher;
2460 {
2461 TracePhase tp("matcher", &timers[_t_matcher]);
2462 matcher.match();
2463 }
2464 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2465 // nodes. Mapping is only valid at the root of each matched subtree.
2466 NOT_PRODUCT( verify_graph_edges(); )
2467
2468 // If you have too many nodes, or if matching has failed, bail out
2469 check_node_count(0, "out of nodes matching instructions");
2470 if (failing()) {
2471 return;
2472 }
2473
2474 print_method(PHASE_MATCHING, 2);
2475
2476 // Build a proper-looking CFG
2477 PhaseCFG cfg(node_arena(), root(), matcher);
2478 _cfg = &cfg;
2479 {
2480 TracePhase tp("scheduler", &timers[_t_scheduler]);
2481 bool success = cfg.do_global_code_motion();
2482 if (!success) {
2483 return;
2484 }
2485
2486 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2487 NOT_PRODUCT( verify_graph_edges(); )
2488 debug_only( cfg.verify(); )
2489 }
2490
2491 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2492 _regalloc = ®alloc;
2493 {
2494 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2495 // Perform register allocation. After Chaitin, use-def chains are
2496 // no longer accurate (at spill code) and so must be ignored.
2497 // Node->LRG->reg mappings are still accurate.
2498 _regalloc->Register_Allocate();
2499
2500 // Bail out if the allocator builds too many nodes
2501 if (failing()) {
2502 return;
2503 }
2504 }
2505
2506 // Prior to register allocation we kept empty basic blocks in case the
2507 // the allocator needed a place to spill. After register allocation we
2508 // are not adding any new instructions. If any basic block is empty, we
2509 // can now safely remove it.
2510 {
2511 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2512 cfg.remove_empty_blocks();
2513 if (do_freq_based_layout()) {
2514 PhaseBlockLayout layout(cfg);
2515 } else {
2516 cfg.set_loop_alignment();
2517 }
2518 cfg.fixup_flow();
2519 }
2520
2521 // Apply peephole optimizations
2522 if( OptoPeephole ) {
2523 TracePhase tp("peephole", &timers[_t_peephole]);
2524 PhasePeephole peep( _regalloc, cfg);
2525 peep.do_transform();
2526 }
2527
2528 // Do late expand if CPU requires this.
2529 if (Matcher::require_postalloc_expand) {
2530 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2531 cfg.postalloc_expand(_regalloc);
2532 }
2533
2534 // Convert Nodes to instruction bits in a buffer
2535 {
2536 TraceTime tp("output", &timers[_t_output], CITime);
2537 Output();
2538 }
2539
2540 print_method(PHASE_FINAL_CODE);
2541
2542 // He's dead, Jim.
2543 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2544 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2545 }
2546
2547
2548 //------------------------------dump_asm---------------------------------------
2549 // Dump formatted assembly
2550 #ifndef PRODUCT
2551 void Compile::dump_asm(int *pcs, uint pc_limit) {
2552 bool cut_short = false;
2553 tty->print_cr("#");
2554 tty->print("# "); _tf->dump(); tty->cr();
2555 tty->print_cr("#");
2556
2557 // For all blocks
2558 int pc = 0x0; // Program counter
2559 char starts_bundle = ' ';
2560 _regalloc->dump_frame();
2561
2562 Node *n = NULL;
2563 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2564 if (VMThread::should_terminate()) {
2565 cut_short = true;
2566 break;
2567 }
2568 Block* block = _cfg->get_block(i);
2569 if (block->is_connector() && !Verbose) {
2570 continue;
2571 }
2572 n = block->head();
2573 if (pcs && n->_idx < pc_limit) {
2574 tty->print("%3.3x ", pcs[n->_idx]);
2575 } else {
2576 tty->print(" ");
2577 }
2578 block->dump_head(_cfg);
2579 if (block->is_connector()) {
2580 tty->print_cr(" # Empty connector block");
2581 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2582 tty->print_cr(" # Block is sole successor of call");
2583 }
2584
2585 // For all instructions
2586 Node *delay = NULL;
2587 for (uint j = 0; j < block->number_of_nodes(); j++) {
2588 if (VMThread::should_terminate()) {
2589 cut_short = true;
2590 break;
2591 }
2592 n = block->get_node(j);
2593 if (valid_bundle_info(n)) {
2594 Bundle* bundle = node_bundling(n);
2595 if (bundle->used_in_unconditional_delay()) {
2596 delay = n;
2597 continue;
2598 }
2599 if (bundle->starts_bundle()) {
2600 starts_bundle = '+';
2601 }
2602 }
2603
2604 if (WizardMode) {
2605 n->dump();
2606 }
2607
2608 if( !n->is_Region() && // Dont print in the Assembly
2609 !n->is_Phi() && // a few noisely useless nodes
2610 !n->is_Proj() &&
2611 !n->is_MachTemp() &&
2612 !n->is_SafePointScalarObject() &&
2613 !n->is_Catch() && // Would be nice to print exception table targets
2614 !n->is_MergeMem() && // Not very interesting
2615 !n->is_top() && // Debug info table constants
2616 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2617 ) {
2618 if (pcs && n->_idx < pc_limit)
2619 tty->print("%3.3x", pcs[n->_idx]);
2620 else
2621 tty->print(" ");
2622 tty->print(" %c ", starts_bundle);
2623 starts_bundle = ' ';
2624 tty->print("\t");
2625 n->format(_regalloc, tty);
2626 tty->cr();
2627 }
2628
2629 // If we have an instruction with a delay slot, and have seen a delay,
2630 // then back up and print it
2631 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2632 assert(delay != NULL, "no unconditional delay instruction");
2633 if (WizardMode) delay->dump();
2634
2635 if (node_bundling(delay)->starts_bundle())
2636 starts_bundle = '+';
2637 if (pcs && n->_idx < pc_limit)
2638 tty->print("%3.3x", pcs[n->_idx]);
2639 else
2640 tty->print(" ");
2641 tty->print(" %c ", starts_bundle);
2642 starts_bundle = ' ';
2643 tty->print("\t");
2644 delay->format(_regalloc, tty);
2645 tty->cr();
2646 delay = NULL;
2647 }
2648
2649 // Dump the exception table as well
2650 if( n->is_Catch() && (Verbose || WizardMode) ) {
2651 // Print the exception table for this offset
2652 _handler_table.print_subtable_for(pc);
2653 }
2654 }
2655
2656 if (pcs && n->_idx < pc_limit)
2657 tty->print_cr("%3.3x", pcs[n->_idx]);
2658 else
2659 tty->cr();
2660
2661 assert(cut_short || delay == NULL, "no unconditional delay branch");
2662
2663 } // End of per-block dump
2664 tty->cr();
2665
2666 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2667 }
2668 #endif
2669
2670 //------------------------------Final_Reshape_Counts---------------------------
2671 // This class defines counters to help identify when a method
2672 // may/must be executed using hardware with only 24-bit precision.
2673 struct Final_Reshape_Counts : public StackObj {
2674 int _call_count; // count non-inlined 'common' calls
2675 int _float_count; // count float ops requiring 24-bit precision
2676 int _double_count; // count double ops requiring more precision
2677 int _java_call_count; // count non-inlined 'java' calls
2678 int _inner_loop_count; // count loops which need alignment
2679 VectorSet _visited; // Visitation flags
2680 Node_List _tests; // Set of IfNodes & PCTableNodes
2681
2682 Final_Reshape_Counts() :
2683 _call_count(0), _float_count(0), _double_count(0),
2684 _java_call_count(0), _inner_loop_count(0),
2685 _visited( Thread::current()->resource_area() ) { }
2686
2687 void inc_call_count () { _call_count ++; }
2688 void inc_float_count () { _float_count ++; }
2689 void inc_double_count() { _double_count++; }
2690 void inc_java_call_count() { _java_call_count++; }
2691 void inc_inner_loop_count() { _inner_loop_count++; }
2692
2693 int get_call_count () const { return _call_count ; }
2694 int get_float_count () const { return _float_count ; }
2695 int get_double_count() const { return _double_count; }
2696 int get_java_call_count() const { return _java_call_count; }
2697 int get_inner_loop_count() const { return _inner_loop_count; }
2698 };
2699
2700 #ifdef ASSERT
2701 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2702 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2703 // Make sure the offset goes inside the instance layout.
2704 return k->contains_field_offset(tp->offset());
2705 // Note that OffsetBot and OffsetTop are very negative.
2706 }
2707 #endif
2708
2709 // Eliminate trivially redundant StoreCMs and accumulate their
2710 // precedence edges.
2711 void Compile::eliminate_redundant_card_marks(Node* n) {
2712 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2713 if (n->in(MemNode::Address)->outcnt() > 1) {
2714 // There are multiple users of the same address so it might be
2715 // possible to eliminate some of the StoreCMs
2716 Node* mem = n->in(MemNode::Memory);
2717 Node* adr = n->in(MemNode::Address);
2718 Node* val = n->in(MemNode::ValueIn);
2719 Node* prev = n;
2720 bool done = false;
2721 // Walk the chain of StoreCMs eliminating ones that match. As
2722 // long as it's a chain of single users then the optimization is
2723 // safe. Eliminating partially redundant StoreCMs would require
2724 // cloning copies down the other paths.
2725 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2726 if (adr == mem->in(MemNode::Address) &&
2727 val == mem->in(MemNode::ValueIn)) {
2728 // redundant StoreCM
2729 if (mem->req() > MemNode::OopStore) {
2730 // Hasn't been processed by this code yet.
2731 n->add_prec(mem->in(MemNode::OopStore));
2732 } else {
2733 // Already converted to precedence edge
2734 for (uint i = mem->req(); i < mem->len(); i++) {
2735 // Accumulate any precedence edges
2736 if (mem->in(i) != NULL) {
2737 n->add_prec(mem->in(i));
2738 }
2739 }
2740 // Everything above this point has been processed.
2741 done = true;
2742 }
2743 // Eliminate the previous StoreCM
2744 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2745 assert(mem->outcnt() == 0, "should be dead");
2746 mem->disconnect_inputs(NULL, this);
2747 } else {
2748 prev = mem;
2749 }
2750 mem = prev->in(MemNode::Memory);
2751 }
2752 }
2753 }
2754
2755 //------------------------------final_graph_reshaping_impl----------------------
2756 // Implement items 1-5 from final_graph_reshaping below.
2757 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2758
2759 if ( n->outcnt() == 0 ) return; // dead node
2760 uint nop = n->Opcode();
2761
2762 // Check for 2-input instruction with "last use" on right input.
2763 // Swap to left input. Implements item (2).
2764 if( n->req() == 3 && // two-input instruction
2765 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2766 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2767 n->in(2)->outcnt() == 1 &&// right use IS a last use
2768 !n->in(2)->is_Con() ) { // right use is not a constant
2769 // Check for commutative opcode
2770 switch( nop ) {
2771 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2772 case Op_MaxI: case Op_MinI:
2773 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2774 case Op_AndL: case Op_XorL: case Op_OrL:
2775 case Op_AndI: case Op_XorI: case Op_OrI: {
2776 // Move "last use" input to left by swapping inputs
2777 n->swap_edges(1, 2);
2778 break;
2779 }
2780 default:
2781 break;
2782 }
2783 }
2784
2785 #ifdef ASSERT
2786 if( n->is_Mem() ) {
2787 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2788 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2789 // oop will be recorded in oop map if load crosses safepoint
2790 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2791 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2792 "raw memory operations should have control edge");
2793 }
2794 if (n->is_MemBar()) {
2795 MemBarNode* mb = n->as_MemBar();
2796 if (mb->trailing_store() || mb->trailing_load_store()) {
2797 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2798 Node* mem = mb->in(MemBarNode::Precedent);
2799 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2800 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2801 } else if (mb->leading()) {
2802 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2803 }
2804 }
2805 #endif
2806 // Count FPU ops and common calls, implements item (3)
2807 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2808 if (!gc_handled) {
2809 final_graph_reshaping_main_switch(n, frc, nop);
2810 }
2811
2812 // Collect CFG split points
2813 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2814 frc._tests.push(n);
2815 }
2816 }
2817
2818 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2819 switch( nop ) {
2820 // Count all float operations that may use FPU
2821 case Op_AddF:
2822 case Op_SubF:
2823 case Op_MulF:
2824 case Op_DivF:
2825 case Op_NegF:
2826 case Op_ModF:
2827 case Op_ConvI2F:
2828 case Op_ConF:
2829 case Op_CmpF:
2830 case Op_CmpF3:
2831 // case Op_ConvL2F: // longs are split into 32-bit halves
2832 frc.inc_float_count();
2833 break;
2834
2835 case Op_ConvF2D:
2836 case Op_ConvD2F:
2837 frc.inc_float_count();
2838 frc.inc_double_count();
2839 break;
2840
2841 // Count all double operations that may use FPU
2842 case Op_AddD:
2843 case Op_SubD:
2844 case Op_MulD:
2845 case Op_DivD:
2846 case Op_NegD:
2847 case Op_ModD:
2848 case Op_ConvI2D:
2849 case Op_ConvD2I:
2850 // case Op_ConvL2D: // handled by leaf call
2851 // case Op_ConvD2L: // handled by leaf call
2852 case Op_ConD:
2853 case Op_CmpD:
2854 case Op_CmpD3:
2855 frc.inc_double_count();
2856 break;
2857 case Op_Opaque1: // Remove Opaque Nodes before matching
2858 case Op_Opaque2: // Remove Opaque Nodes before matching
2859 case Op_Opaque3:
2860 n->subsume_by(n->in(1), this);
2861 break;
2862 case Op_CallStaticJava:
2863 case Op_CallJava:
2864 case Op_CallDynamicJava:
2865 frc.inc_java_call_count(); // Count java call site;
2866 case Op_CallRuntime:
2867 case Op_CallLeaf:
2868 case Op_CallLeafNoFP: {
2869 assert (n->is_Call(), "");
2870 CallNode *call = n->as_Call();
2871 // Count call sites where the FP mode bit would have to be flipped.
2872 // Do not count uncommon runtime calls:
2873 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2874 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2875 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2876 frc.inc_call_count(); // Count the call site
2877 } else { // See if uncommon argument is shared
2878 Node *n = call->in(TypeFunc::Parms);
2879 int nop = n->Opcode();
2880 // Clone shared simple arguments to uncommon calls, item (1).
2881 if (n->outcnt() > 1 &&
2882 !n->is_Proj() &&
2883 nop != Op_CreateEx &&
2884 nop != Op_CheckCastPP &&
2885 nop != Op_DecodeN &&
2886 nop != Op_DecodeNKlass &&
2887 !n->is_Mem() &&
2888 !n->is_Phi()) {
2889 Node *x = n->clone();
2890 call->set_req(TypeFunc::Parms, x);
2891 }
2892 }
2893 break;
2894 }
2895
2896 case Op_StoreD:
2897 case Op_LoadD:
2898 case Op_LoadD_unaligned:
2899 frc.inc_double_count();
2900 goto handle_mem;
2901 case Op_StoreF:
2902 case Op_LoadF:
2903 frc.inc_float_count();
2904 goto handle_mem;
2905
2906 case Op_StoreCM:
2907 {
2908 // Convert OopStore dependence into precedence edge
2909 Node* prec = n->in(MemNode::OopStore);
2910 n->del_req(MemNode::OopStore);
2911 n->add_prec(prec);
2912 eliminate_redundant_card_marks(n);
2913 }
2914
2915 // fall through
2916
2917 case Op_StoreB:
2918 case Op_StoreC:
2919 case Op_StorePConditional:
2920 case Op_StoreI:
2921 case Op_StoreL:
2922 case Op_StoreIConditional:
2923 case Op_StoreLConditional:
2924 case Op_CompareAndSwapB:
2925 case Op_CompareAndSwapS:
2926 case Op_CompareAndSwapI:
2927 case Op_CompareAndSwapL:
2928 case Op_CompareAndSwapP:
2929 case Op_CompareAndSwapN:
2930 case Op_WeakCompareAndSwapB:
2931 case Op_WeakCompareAndSwapS:
2932 case Op_WeakCompareAndSwapI:
2933 case Op_WeakCompareAndSwapL:
2934 case Op_WeakCompareAndSwapP:
2935 case Op_WeakCompareAndSwapN:
2936 case Op_CompareAndExchangeB:
2937 case Op_CompareAndExchangeS:
2938 case Op_CompareAndExchangeI:
2939 case Op_CompareAndExchangeL:
2940 case Op_CompareAndExchangeP:
2941 case Op_CompareAndExchangeN:
2942 case Op_GetAndAddS:
2943 case Op_GetAndAddB:
2944 case Op_GetAndAddI:
2945 case Op_GetAndAddL:
2946 case Op_GetAndSetS:
2947 case Op_GetAndSetB:
2948 case Op_GetAndSetI:
2949 case Op_GetAndSetL:
2950 case Op_GetAndSetP:
2951 case Op_GetAndSetN:
2952 case Op_StoreP:
2953 case Op_StoreN:
2954 case Op_StoreNKlass:
2955 case Op_LoadB:
2956 case Op_LoadUB:
2957 case Op_LoadUS:
2958 case Op_LoadI:
2959 case Op_LoadKlass:
2960 case Op_LoadNKlass:
2961 case Op_LoadL:
2962 case Op_LoadL_unaligned:
2963 case Op_LoadPLocked:
2964 case Op_LoadP:
2965 case Op_LoadN:
2966 case Op_LoadRange:
2967 case Op_LoadS: {
2968 handle_mem:
2969 #ifdef ASSERT
2970 if( VerifyOptoOopOffsets ) {
2971 MemNode* mem = n->as_Mem();
2972 // Check to see if address types have grounded out somehow.
2973 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2974 assert( !tp || oop_offset_is_sane(tp), "" );
2975 }
2976 #endif
2977 break;
2978 }
2979
2980 case Op_AddP: { // Assert sane base pointers
2981 Node *addp = n->in(AddPNode::Address);
2982 assert( !addp->is_AddP() ||
2983 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2984 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2985 "Base pointers must match (addp %u)", addp->_idx );
2986 #ifdef _LP64
2987 if ((UseCompressedOops || UseCompressedClassPointers) &&
2988 addp->Opcode() == Op_ConP &&
2989 addp == n->in(AddPNode::Base) &&
2990 n->in(AddPNode::Offset)->is_Con()) {
2991 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2992 // on the platform and on the compressed oops mode.
2993 // Use addressing with narrow klass to load with offset on x86.
2994 // Some platforms can use the constant pool to load ConP.
2995 // Do this transformation here since IGVN will convert ConN back to ConP.
2996 const Type* t = addp->bottom_type();
2997 bool is_oop = t->isa_oopptr() != NULL;
2998 bool is_klass = t->isa_klassptr() != NULL;
2999
3000 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3001 (is_klass && Matcher::const_klass_prefer_decode())) {
3002 Node* nn = NULL;
3003
3004 int op = is_oop ? Op_ConN : Op_ConNKlass;
3005
3006 // Look for existing ConN node of the same exact type.
3007 Node* r = root();
3008 uint cnt = r->outcnt();
3009 for (uint i = 0; i < cnt; i++) {
3010 Node* m = r->raw_out(i);
3011 if (m!= NULL && m->Opcode() == op &&
3012 m->bottom_type()->make_ptr() == t) {
3013 nn = m;
3014 break;
3015 }
3016 }
3017 if (nn != NULL) {
3018 // Decode a narrow oop to match address
3019 // [R12 + narrow_oop_reg<<3 + offset]
3020 if (is_oop) {
3021 nn = new DecodeNNode(nn, t);
3022 } else {
3023 nn = new DecodeNKlassNode(nn, t);
3024 }
3025 // Check for succeeding AddP which uses the same Base.
3026 // Otherwise we will run into the assertion above when visiting that guy.
3027 for (uint i = 0; i < n->outcnt(); ++i) {
3028 Node *out_i = n->raw_out(i);
3029 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3030 out_i->set_req(AddPNode::Base, nn);
3031 #ifdef ASSERT
3032 for (uint j = 0; j < out_i->outcnt(); ++j) {
3033 Node *out_j = out_i->raw_out(j);
3034 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3035 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3036 }
3037 #endif
3038 }
3039 }
3040 n->set_req(AddPNode::Base, nn);
3041 n->set_req(AddPNode::Address, nn);
3042 if (addp->outcnt() == 0) {
3043 addp->disconnect_inputs(NULL, this);
3044 }
3045 }
3046 }
3047 }
3048 #endif
3049 // platform dependent reshaping of the address expression
3050 reshape_address(n->as_AddP());
3051 break;
3052 }
3053
3054 case Op_CastPP: {
3055 // Remove CastPP nodes to gain more freedom during scheduling but
3056 // keep the dependency they encode as control or precedence edges
3057 // (if control is set already) on memory operations. Some CastPP
3058 // nodes don't have a control (don't carry a dependency): skip
3059 // those.
3060 if (n->in(0) != NULL) {
3061 ResourceMark rm;
3062 Unique_Node_List wq;
3063 wq.push(n);
3064 for (uint next = 0; next < wq.size(); ++next) {
3065 Node *m = wq.at(next);
3066 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3067 Node* use = m->fast_out(i);
3068 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) {
3069 use->ensure_control_or_add_prec(n->in(0));
3070 } else {
3071 switch(use->Opcode()) {
3072 case Op_AddP:
3073 case Op_DecodeN:
3074 case Op_DecodeNKlass:
3075 case Op_CheckCastPP:
3076 case Op_CastPP:
3077 wq.push(use);
3078 break;
3079 }
3080 }
3081 }
3082 }
3083 }
3084 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3085 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3086 Node* in1 = n->in(1);
3087 const Type* t = n->bottom_type();
3088 Node* new_in1 = in1->clone();
3089 new_in1->as_DecodeN()->set_type(t);
3090
3091 if (!Matcher::narrow_oop_use_complex_address()) {
3092 //
3093 // x86, ARM and friends can handle 2 adds in addressing mode
3094 // and Matcher can fold a DecodeN node into address by using
3095 // a narrow oop directly and do implicit NULL check in address:
3096 //
3097 // [R12 + narrow_oop_reg<<3 + offset]
3098 // NullCheck narrow_oop_reg
3099 //
3100 // On other platforms (Sparc) we have to keep new DecodeN node and
3101 // use it to do implicit NULL check in address:
3102 //
3103 // decode_not_null narrow_oop_reg, base_reg
3104 // [base_reg + offset]
3105 // NullCheck base_reg
3106 //
3107 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3108 // to keep the information to which NULL check the new DecodeN node
3109 // corresponds to use it as value in implicit_null_check().
3110 //
3111 new_in1->set_req(0, n->in(0));
3112 }
3113
3114 n->subsume_by(new_in1, this);
3115 if (in1->outcnt() == 0) {
3116 in1->disconnect_inputs(NULL, this);
3117 }
3118 } else {
3119 n->subsume_by(n->in(1), this);
3120 if (n->outcnt() == 0) {
3121 n->disconnect_inputs(NULL, this);
3122 }
3123 }
3124 break;
3125 }
3126 #ifdef _LP64
3127 case Op_CmpP:
3128 // Do this transformation here to preserve CmpPNode::sub() and
3129 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3130 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3131 Node* in1 = n->in(1);
3132 Node* in2 = n->in(2);
3133 if (!in1->is_DecodeNarrowPtr()) {
3134 in2 = in1;
3135 in1 = n->in(2);
3136 }
3137 assert(in1->is_DecodeNarrowPtr(), "sanity");
3138
3139 Node* new_in2 = NULL;
3140 if (in2->is_DecodeNarrowPtr()) {
3141 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3142 new_in2 = in2->in(1);
3143 } else if (in2->Opcode() == Op_ConP) {
3144 const Type* t = in2->bottom_type();
3145 if (t == TypePtr::NULL_PTR) {
3146 assert(in1->is_DecodeN(), "compare klass to null?");
3147 // Don't convert CmpP null check into CmpN if compressed
3148 // oops implicit null check is not generated.
3149 // This will allow to generate normal oop implicit null check.
3150 if (Matcher::gen_narrow_oop_implicit_null_checks())
3151 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3152 //
3153 // This transformation together with CastPP transformation above
3154 // will generated code for implicit NULL checks for compressed oops.
3155 //
3156 // The original code after Optimize()
3157 //
3158 // LoadN memory, narrow_oop_reg
3159 // decode narrow_oop_reg, base_reg
3160 // CmpP base_reg, NULL
3161 // CastPP base_reg // NotNull
3162 // Load [base_reg + offset], val_reg
3163 //
3164 // after these transformations will be
3165 //
3166 // LoadN memory, narrow_oop_reg
3167 // CmpN narrow_oop_reg, NULL
3168 // decode_not_null narrow_oop_reg, base_reg
3169 // Load [base_reg + offset], val_reg
3170 //
3171 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3172 // since narrow oops can be used in debug info now (see the code in
3173 // final_graph_reshaping_walk()).
3174 //
3175 // At the end the code will be matched to
3176 // on x86:
3177 //
3178 // Load_narrow_oop memory, narrow_oop_reg
3179 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3180 // NullCheck narrow_oop_reg
3181 //
3182 // and on sparc:
3183 //
3184 // Load_narrow_oop memory, narrow_oop_reg
3185 // decode_not_null narrow_oop_reg, base_reg
3186 // Load [base_reg + offset], val_reg
3187 // NullCheck base_reg
3188 //
3189 } else if (t->isa_oopptr()) {
3190 new_in2 = ConNode::make(t->make_narrowoop());
3191 } else if (t->isa_klassptr()) {
3192 new_in2 = ConNode::make(t->make_narrowklass());
3193 }
3194 }
3195 if (new_in2 != NULL) {
3196 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3197 n->subsume_by(cmpN, this);
3198 if (in1->outcnt() == 0) {
3199 in1->disconnect_inputs(NULL, this);
3200 }
3201 if (in2->outcnt() == 0) {
3202 in2->disconnect_inputs(NULL, this);
3203 }
3204 }
3205 }
3206 break;
3207
3208 case Op_DecodeN:
3209 case Op_DecodeNKlass:
3210 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3211 // DecodeN could be pinned when it can't be fold into
3212 // an address expression, see the code for Op_CastPP above.
3213 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3214 break;
3215
3216 case Op_EncodeP:
3217 case Op_EncodePKlass: {
3218 Node* in1 = n->in(1);
3219 if (in1->is_DecodeNarrowPtr()) {
3220 n->subsume_by(in1->in(1), this);
3221 } else if (in1->Opcode() == Op_ConP) {
3222 const Type* t = in1->bottom_type();
3223 if (t == TypePtr::NULL_PTR) {
3224 assert(t->isa_oopptr(), "null klass?");
3225 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3226 } else if (t->isa_oopptr()) {
3227 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3228 } else if (t->isa_klassptr()) {
3229 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3230 }
3231 }
3232 if (in1->outcnt() == 0) {
3233 in1->disconnect_inputs(NULL, this);
3234 }
3235 break;
3236 }
3237
3238 case Op_Proj: {
3239 if (OptimizeStringConcat) {
3240 ProjNode* p = n->as_Proj();
3241 if (p->_is_io_use) {
3242 // Separate projections were used for the exception path which
3243 // are normally removed by a late inline. If it wasn't inlined
3244 // then they will hang around and should just be replaced with
3245 // the original one.
3246 Node* proj = NULL;
3247 // Replace with just one
3248 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3249 Node *use = i.get();
3250 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3251 proj = use;
3252 break;
3253 }
3254 }
3255 assert(proj != NULL, "must be found");
3256 p->subsume_by(proj, this);
3257 }
3258 }
3259 break;
3260 }
3261
3262 case Op_Phi:
3263 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3264 // The EncodeP optimization may create Phi with the same edges
3265 // for all paths. It is not handled well by Register Allocator.
3266 Node* unique_in = n->in(1);
3267 assert(unique_in != NULL, "");
3268 uint cnt = n->req();
3269 for (uint i = 2; i < cnt; i++) {
3270 Node* m = n->in(i);
3271 assert(m != NULL, "");
3272 if (unique_in != m)
3273 unique_in = NULL;
3274 }
3275 if (unique_in != NULL) {
3276 n->subsume_by(unique_in, this);
3277 }
3278 }
3279 break;
3280
3281 #endif
3282
3283 #ifdef ASSERT
3284 case Op_CastII:
3285 // Verify that all range check dependent CastII nodes were removed.
3286 if (n->isa_CastII()->has_range_check()) {
3287 n->dump(3);
3288 assert(false, "Range check dependent CastII node was not removed");
3289 }
3290 break;
3291 #endif
3292
3293 case Op_ModI:
3294 if (UseDivMod) {
3295 // Check if a%b and a/b both exist
3296 Node* d = n->find_similar(Op_DivI);
3297 if (d) {
3298 // Replace them with a fused divmod if supported
3299 if (Matcher::has_match_rule(Op_DivModI)) {
3300 DivModINode* divmod = DivModINode::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 MulINode(d, d->in(2));
3306 Node* sub = new SubINode(d->in(1), mult);
3307 n->subsume_by(sub, this);
3308 }
3309 }
3310 }
3311 break;
3312
3313 case Op_ModL:
3314 if (UseDivMod) {
3315 // Check if a%b and a/b both exist
3316 Node* d = n->find_similar(Op_DivL);
3317 if (d) {
3318 // Replace them with a fused divmod if supported
3319 if (Matcher::has_match_rule(Op_DivModL)) {
3320 DivModLNode* divmod = DivModLNode::make(n);
3321 d->subsume_by(divmod->div_proj(), this);
3322 n->subsume_by(divmod->mod_proj(), this);
3323 } else {
3324 // replace a%b with a-((a/b)*b)
3325 Node* mult = new MulLNode(d, d->in(2));
3326 Node* sub = new SubLNode(d->in(1), mult);
3327 n->subsume_by(sub, this);
3328 }
3329 }
3330 }
3331 break;
3332
3333 case Op_LoadVector:
3334 case Op_StoreVector:
3335 break;
3336
3337 case Op_AddReductionVI:
3338 case Op_AddReductionVL:
3339 case Op_AddReductionVF:
3340 case Op_AddReductionVD:
3341 case Op_MulReductionVI:
3342 case Op_MulReductionVL:
3343 case Op_MulReductionVF:
3344 case Op_MulReductionVD:
3345 break;
3346
3347 case Op_PackB:
3348 case Op_PackS:
3349 case Op_PackI:
3350 case Op_PackF:
3351 case Op_PackL:
3352 case Op_PackD:
3353 if (n->req()-1 > 2) {
3354 // Replace many operand PackNodes with a binary tree for matching
3355 PackNode* p = (PackNode*) n;
3356 Node* btp = p->binary_tree_pack(1, n->req());
3357 n->subsume_by(btp, this);
3358 }
3359 break;
3360 case Op_Loop:
3361 case Op_CountedLoop:
3362 case Op_OuterStripMinedLoop:
3363 if (n->as_Loop()->is_inner_loop()) {
3364 frc.inc_inner_loop_count();
3365 }
3366 n->as_Loop()->verify_strip_mined(0);
3367 break;
3368 case Op_LShiftI:
3369 case Op_RShiftI:
3370 case Op_URShiftI:
3371 case Op_LShiftL:
3372 case Op_RShiftL:
3373 case Op_URShiftL:
3374 if (Matcher::need_masked_shift_count) {
3375 // The cpu's shift instructions don't restrict the count to the
3376 // lower 5/6 bits. We need to do the masking ourselves.
3377 Node* in2 = n->in(2);
3378 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3379 const TypeInt* t = in2->find_int_type();
3380 if (t != NULL && t->is_con()) {
3381 juint shift = t->get_con();
3382 if (shift > mask) { // Unsigned cmp
3383 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3384 }
3385 } else {
3386 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3387 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3388 n->set_req(2, shift);
3389 }
3390 }
3391 if (in2->outcnt() == 0) { // Remove dead node
3392 in2->disconnect_inputs(NULL, this);
3393 }
3394 }
3395 break;
3396 case Op_MemBarStoreStore:
3397 case Op_MemBarRelease:
3398 // Break the link with AllocateNode: it is no longer useful and
3399 // confuses register allocation.
3400 if (n->req() > MemBarNode::Precedent) {
3401 n->set_req(MemBarNode::Precedent, top());
3402 }
3403 break;
3404 case Op_MemBarAcquire: {
3405 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3406 // At parse time, the trailing MemBarAcquire for a volatile load
3407 // is created with an edge to the load. After optimizations,
3408 // that input may be a chain of Phis. If those phis have no
3409 // other use, then the MemBarAcquire keeps them alive and
3410 // register allocation can be confused.
3411 ResourceMark rm;
3412 Unique_Node_List wq;
3413 wq.push(n->in(MemBarNode::Precedent));
3414 n->set_req(MemBarNode::Precedent, top());
3415 while (wq.size() > 0) {
3416 Node* m = wq.pop();
3417 if (m->outcnt() == 0) {
3418 for (uint j = 0; j < m->req(); j++) {
3419 Node* in = m->in(j);
3420 if (in != NULL) {
3421 wq.push(in);
3422 }
3423 }
3424 m->disconnect_inputs(NULL, this);
3425 }
3426 }
3427 }
3428 break;
3429 }
3430 case Op_RangeCheck: {
3431 RangeCheckNode* rc = n->as_RangeCheck();
3432 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3433 n->subsume_by(iff, this);
3434 frc._tests.push(iff);
3435 break;
3436 }
3437 case Op_ConvI2L: {
3438 if (!Matcher::convi2l_type_required) {
3439 // Code generation on some platforms doesn't need accurate
3440 // ConvI2L types. Widening the type can help remove redundant
3441 // address computations.
3442 n->as_Type()->set_type(TypeLong::INT);
3443 ResourceMark rm;
3444 Node_List wq;
3445 wq.push(n);
3446 for (uint next = 0; next < wq.size(); next++) {
3447 Node *m = wq.at(next);
3448
3449 for(;;) {
3450 // Loop over all nodes with identical inputs edges as m
3451 Node* k = m->find_similar(m->Opcode());
3452 if (k == NULL) {
3453 break;
3454 }
3455 // Push their uses so we get a chance to remove node made
3456 // redundant
3457 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3458 Node* u = k->fast_out(i);
3459 assert(!wq.contains(u), "shouldn't process one node several times");
3460 if (u->Opcode() == Op_LShiftL ||
3461 u->Opcode() == Op_AddL ||
3462 u->Opcode() == Op_SubL ||
3463 u->Opcode() == Op_AddP) {
3464 wq.push(u);
3465 }
3466 }
3467 // Replace all nodes with identical edges as m with m
3468 k->subsume_by(m, this);
3469 }
3470 }
3471 }
3472 break;
3473 }
3474 case Op_CmpUL: {
3475 if (!Matcher::has_match_rule(Op_CmpUL)) {
3476 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3477 // to max_julong if < 0 to make the signed comparison fail.
3478 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3479 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3480 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3481 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3482 Node* andl = new AndLNode(orl, remove_sign_mask);
3483 Node* cmp = new CmpLNode(andl, n->in(2));
3484 n->subsume_by(cmp, this);
3485 }
3486 break;
3487 }
3488 default:
3489 assert(!n->is_Call(), "");
3490 assert(!n->is_Mem(), "");
3491 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3492 break;
3493 }
3494 }
3495
3496 //------------------------------final_graph_reshaping_walk---------------------
3497 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3498 // requires that the walk visits a node's inputs before visiting the node.
3499 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3500 ResourceArea *area = Thread::current()->resource_area();
3501 Unique_Node_List sfpt(area);
3502
3503 frc._visited.set(root->_idx); // first, mark node as visited
3504 uint cnt = root->req();
3505 Node *n = root;
3506 uint i = 0;
3507 while (true) {
3508 if (i < cnt) {
3509 // Place all non-visited non-null inputs onto stack
3510 Node* m = n->in(i);
3511 ++i;
3512 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3513 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3514 // compute worst case interpreter size in case of a deoptimization
3515 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3516
3517 sfpt.push(m);
3518 }
3519 cnt = m->req();
3520 nstack.push(n, i); // put on stack parent and next input's index
3521 n = m;
3522 i = 0;
3523 }
3524 } else {
3525 // Now do post-visit work
3526 final_graph_reshaping_impl( n, frc );
3527 if (nstack.is_empty())
3528 break; // finished
3529 n = nstack.node(); // Get node from stack
3530 cnt = n->req();
3531 i = nstack.index();
3532 nstack.pop(); // Shift to the next node on stack
3533 }
3534 }
3535
3536 // Skip next transformation if compressed oops are not used.
3537 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3538 (!UseCompressedOops && !UseCompressedClassPointers))
3539 return;
3540
3541 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3542 // It could be done for an uncommon traps or any safepoints/calls
3543 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3544 while (sfpt.size() > 0) {
3545 n = sfpt.pop();
3546 JVMState *jvms = n->as_SafePoint()->jvms();
3547 assert(jvms != NULL, "sanity");
3548 int start = jvms->debug_start();
3549 int end = n->req();
3550 bool is_uncommon = (n->is_CallStaticJava() &&
3551 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3552 for (int j = start; j < end; j++) {
3553 Node* in = n->in(j);
3554 if (in->is_DecodeNarrowPtr()) {
3555 bool safe_to_skip = true;
3556 if (!is_uncommon ) {
3557 // Is it safe to skip?
3558 for (uint i = 0; i < in->outcnt(); i++) {
3559 Node* u = in->raw_out(i);
3560 if (!u->is_SafePoint() ||
3561 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3562 safe_to_skip = false;
3563 }
3564 }
3565 }
3566 if (safe_to_skip) {
3567 n->set_req(j, in->in(1));
3568 }
3569 if (in->outcnt() == 0) {
3570 in->disconnect_inputs(NULL, this);
3571 }
3572 }
3573 }
3574 }
3575 }
3576
3577 //------------------------------final_graph_reshaping--------------------------
3578 // Final Graph Reshaping.
3579 //
3580 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3581 // and not commoned up and forced early. Must come after regular
3582 // optimizations to avoid GVN undoing the cloning. Clone constant
3583 // inputs to Loop Phis; these will be split by the allocator anyways.
3584 // Remove Opaque nodes.
3585 // (2) Move last-uses by commutative operations to the left input to encourage
3586 // Intel update-in-place two-address operations and better register usage
3587 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3588 // calls canonicalizing them back.
3589 // (3) Count the number of double-precision FP ops, single-precision FP ops
3590 // and call sites. On Intel, we can get correct rounding either by
3591 // forcing singles to memory (requires extra stores and loads after each
3592 // FP bytecode) or we can set a rounding mode bit (requires setting and
3593 // clearing the mode bit around call sites). The mode bit is only used
3594 // if the relative frequency of single FP ops to calls is low enough.
3595 // This is a key transform for SPEC mpeg_audio.
3596 // (4) Detect infinite loops; blobs of code reachable from above but not
3597 // below. Several of the Code_Gen algorithms fail on such code shapes,
3598 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3599 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3600 // Detection is by looking for IfNodes where only 1 projection is
3601 // reachable from below or CatchNodes missing some targets.
3602 // (5) Assert for insane oop offsets in debug mode.
3603
3604 bool Compile::final_graph_reshaping() {
3605 // an infinite loop may have been eliminated by the optimizer,
3606 // in which case the graph will be empty.
3607 if (root()->req() == 1) {
3608 record_method_not_compilable("trivial infinite loop");
3609 return true;
3610 }
3611
3612 // Expensive nodes have their control input set to prevent the GVN
3613 // from freely commoning them. There's no GVN beyond this point so
3614 // no need to keep the control input. We want the expensive nodes to
3615 // be freely moved to the least frequent code path by gcm.
3616 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3617 for (int i = 0; i < expensive_count(); i++) {
3618 _expensive_nodes->at(i)->set_req(0, NULL);
3619 }
3620
3621 Final_Reshape_Counts frc;
3622
3623 // Visit everybody reachable!
3624 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3625 Node_Stack nstack(live_nodes() >> 1);
3626 final_graph_reshaping_walk(nstack, root(), frc);
3627
3628 // Check for unreachable (from below) code (i.e., infinite loops).
3629 for( uint i = 0; i < frc._tests.size(); i++ ) {
3630 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3631 // Get number of CFG targets.
3632 // Note that PCTables include exception targets after calls.
3633 uint required_outcnt = n->required_outcnt();
3634 if (n->outcnt() != required_outcnt) {
3635 // Check for a few special cases. Rethrow Nodes never take the
3636 // 'fall-thru' path, so expected kids is 1 less.
3637 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3638 if (n->in(0)->in(0)->is_Call()) {
3639 CallNode *call = n->in(0)->in(0)->as_Call();
3640 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3641 required_outcnt--; // Rethrow always has 1 less kid
3642 } else if (call->req() > TypeFunc::Parms &&
3643 call->is_CallDynamicJava()) {
3644 // Check for null receiver. In such case, the optimizer has
3645 // detected that the virtual call will always result in a null
3646 // pointer exception. The fall-through projection of this CatchNode
3647 // will not be populated.
3648 Node *arg0 = call->in(TypeFunc::Parms);
3649 if (arg0->is_Type() &&
3650 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3651 required_outcnt--;
3652 }
3653 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3654 call->req() > TypeFunc::Parms+1 &&
3655 call->is_CallStaticJava()) {
3656 // Check for negative array length. In such case, the optimizer has
3657 // detected that the allocation attempt will always result in an
3658 // exception. There is no fall-through projection of this CatchNode .
3659 Node *arg1 = call->in(TypeFunc::Parms+1);
3660 if (arg1->is_Type() &&
3661 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3662 required_outcnt--;
3663 }
3664 }
3665 }
3666 }
3667 // Recheck with a better notion of 'required_outcnt'
3668 if (n->outcnt() != required_outcnt) {
3669 record_method_not_compilable("malformed control flow");
3670 return true; // Not all targets reachable!
3671 }
3672 }
3673 // Check that I actually visited all kids. Unreached kids
3674 // must be infinite loops.
3675 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3676 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3677 record_method_not_compilable("infinite loop");
3678 return true; // Found unvisited kid; must be unreach
3679 }
3680
3681 // Here so verification code in final_graph_reshaping_walk()
3682 // always see an OuterStripMinedLoopEnd
3683 if (n->is_OuterStripMinedLoopEnd()) {
3684 IfNode* init_iff = n->as_If();
3685 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3686 n->subsume_by(iff, this);
3687 }
3688 }
3689
3690 // If original bytecodes contained a mixture of floats and doubles
3691 // check if the optimizer has made it homogenous, item (3).
3692 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3693 frc.get_float_count() > 32 &&
3694 frc.get_double_count() == 0 &&
3695 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3696 set_24_bit_selection_and_mode( false, true );
3697 }
3698
3699 set_java_calls(frc.get_java_call_count());
3700 set_inner_loops(frc.get_inner_loop_count());
3701
3702 // No infinite loops, no reason to bail out.
3703 return false;
3704 }
3705
3706 //-----------------------------too_many_traps----------------------------------
3707 // Report if there are too many traps at the current method and bci.
3708 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3709 bool Compile::too_many_traps(ciMethod* method,
3710 int bci,
3711 Deoptimization::DeoptReason reason) {
3712 ciMethodData* md = method->method_data();
3713 if (md->is_empty()) {
3714 // Assume the trap has not occurred, or that it occurred only
3715 // because of a transient condition during start-up in the interpreter.
3716 return false;
3717 }
3718 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3719 if (md->has_trap_at(bci, m, reason) != 0) {
3720 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3721 // Also, if there are multiple reasons, or if there is no per-BCI record,
3722 // assume the worst.
3723 if (log())
3724 log()->elem("observe trap='%s' count='%d'",
3725 Deoptimization::trap_reason_name(reason),
3726 md->trap_count(reason));
3727 return true;
3728 } else {
3729 // Ignore method/bci and see if there have been too many globally.
3730 return too_many_traps(reason, md);
3731 }
3732 }
3733
3734 // Less-accurate variant which does not require a method and bci.
3735 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3736 ciMethodData* logmd) {
3737 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3738 // Too many traps globally.
3739 // Note that we use cumulative trap_count, not just md->trap_count.
3740 if (log()) {
3741 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3742 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3743 Deoptimization::trap_reason_name(reason),
3744 mcount, trap_count(reason));
3745 }
3746 return true;
3747 } else {
3748 // The coast is clear.
3749 return false;
3750 }
3751 }
3752
3753 //--------------------------too_many_recompiles--------------------------------
3754 // Report if there are too many recompiles at the current method and bci.
3755 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3756 // Is not eager to return true, since this will cause the compiler to use
3757 // Action_none for a trap point, to avoid too many recompilations.
3758 bool Compile::too_many_recompiles(ciMethod* method,
3759 int bci,
3760 Deoptimization::DeoptReason reason) {
3761 ciMethodData* md = method->method_data();
3762 if (md->is_empty()) {
3763 // Assume the trap has not occurred, or that it occurred only
3764 // because of a transient condition during start-up in the interpreter.
3765 return false;
3766 }
3767 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3768 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3769 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3770 Deoptimization::DeoptReason per_bc_reason
3771 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3772 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3773 if ((per_bc_reason == Deoptimization::Reason_none
3774 || md->has_trap_at(bci, m, reason) != 0)
3775 // The trap frequency measure we care about is the recompile count:
3776 && md->trap_recompiled_at(bci, m)
3777 && md->overflow_recompile_count() >= bc_cutoff) {
3778 // Do not emit a trap here if it has already caused recompilations.
3779 // Also, if there are multiple reasons, or if there is no per-BCI record,
3780 // assume the worst.
3781 if (log())
3782 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3783 Deoptimization::trap_reason_name(reason),
3784 md->trap_count(reason),
3785 md->overflow_recompile_count());
3786 return true;
3787 } else if (trap_count(reason) != 0
3788 && decompile_count() >= m_cutoff) {
3789 // Too many recompiles globally, and we have seen this sort of trap.
3790 // Use cumulative decompile_count, not just md->decompile_count.
3791 if (log())
3792 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3793 Deoptimization::trap_reason_name(reason),
3794 md->trap_count(reason), trap_count(reason),
3795 md->decompile_count(), decompile_count());
3796 return true;
3797 } else {
3798 // The coast is clear.
3799 return false;
3800 }
3801 }
3802
3803 // Compute when not to trap. Used by matching trap based nodes and
3804 // NullCheck optimization.
3805 void Compile::set_allowed_deopt_reasons() {
3806 _allowed_reasons = 0;
3807 if (is_method_compilation()) {
3808 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3809 assert(rs < BitsPerInt, "recode bit map");
3810 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3811 _allowed_reasons |= nth_bit(rs);
3812 }
3813 }
3814 }
3815 }
3816
3817 #ifndef PRODUCT
3818 //------------------------------verify_graph_edges---------------------------
3819 // Walk the Graph and verify that there is a one-to-one correspondence
3820 // between Use-Def edges and Def-Use edges in the graph.
3821 void Compile::verify_graph_edges(bool no_dead_code) {
3822 if (VerifyGraphEdges) {
3823 ResourceArea *area = Thread::current()->resource_area();
3824 Unique_Node_List visited(area);
3825 // Call recursive graph walk to check edges
3826 _root->verify_edges(visited);
3827 if (no_dead_code) {
3828 // Now make sure that no visited node is used by an unvisited node.
3829 bool dead_nodes = false;
3830 Unique_Node_List checked(area);
3831 while (visited.size() > 0) {
3832 Node* n = visited.pop();
3833 checked.push(n);
3834 for (uint i = 0; i < n->outcnt(); i++) {
3835 Node* use = n->raw_out(i);
3836 if (checked.member(use)) continue; // already checked
3837 if (visited.member(use)) continue; // already in the graph
3838 if (use->is_Con()) continue; // a dead ConNode is OK
3839 // At this point, we have found a dead node which is DU-reachable.
3840 if (!dead_nodes) {
3841 tty->print_cr("*** Dead nodes reachable via DU edges:");
3842 dead_nodes = true;
3843 }
3844 use->dump(2);
3845 tty->print_cr("---");
3846 checked.push(use); // No repeats; pretend it is now checked.
3847 }
3848 }
3849 assert(!dead_nodes, "using nodes must be reachable from root");
3850 }
3851 }
3852 }
3853 #endif
3854
3855 // The Compile object keeps track of failure reasons separately from the ciEnv.
3856 // This is required because there is not quite a 1-1 relation between the
3857 // ciEnv and its compilation task and the Compile object. Note that one
3858 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3859 // to backtrack and retry without subsuming loads. Other than this backtracking
3860 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3861 // by the logic in C2Compiler.
3862 void Compile::record_failure(const char* reason) {
3863 if (log() != NULL) {
3864 log()->elem("failure reason='%s' phase='compile'", reason);
3865 }
3866 if (_failure_reason == NULL) {
3867 // Record the first failure reason.
3868 _failure_reason = reason;
3869 }
3870
3871 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3872 C->print_method(PHASE_FAILURE);
3873 }
3874 _root = NULL; // flush the graph, too
3875 }
3876
3877 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3878 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3879 _phase_name(name), _dolog(CITimeVerbose)
3880 {
3881 if (_dolog) {
3882 C = Compile::current();
3883 _log = C->log();
3884 } else {
3885 C = NULL;
3886 _log = NULL;
3887 }
3888 if (_log != NULL) {
3889 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3890 _log->stamp();
3891 _log->end_head();
3892 }
3893 }
3894
3895 Compile::TracePhase::~TracePhase() {
3896
3897 C = Compile::current();
3898 if (_dolog) {
3899 _log = C->log();
3900 } else {
3901 _log = NULL;
3902 }
3903
3904 #ifdef ASSERT
3905 if (PrintIdealNodeCount) {
3906 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3907 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3908 }
3909
3910 if (VerifyIdealNodeCount) {
3911 Compile::current()->print_missing_nodes();
3912 }
3913 #endif
3914
3915 if (_log != NULL) {
3916 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3917 }
3918 }
3919
3920 //=============================================================================
3921 // Two Constant's are equal when the type and the value are equal.
3922 bool Compile::Constant::operator==(const Constant& other) {
3923 if (type() != other.type() ) return false;
3924 if (can_be_reused() != other.can_be_reused()) return false;
3925 // For floating point values we compare the bit pattern.
3926 switch (type()) {
3927 case T_INT:
3928 case T_FLOAT: return (_v._value.i == other._v._value.i);
3929 case T_LONG:
3930 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3931 case T_OBJECT:
3932 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3933 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3934 case T_METADATA: return (_v._metadata == other._v._metadata);
3935 default: ShouldNotReachHere(); return false;
3936 }
3937 }
3938
3939 static int type_to_size_in_bytes(BasicType t) {
3940 switch (t) {
3941 case T_INT: return sizeof(jint );
3942 case T_LONG: return sizeof(jlong );
3943 case T_FLOAT: return sizeof(jfloat );
3944 case T_DOUBLE: return sizeof(jdouble);
3945 case T_METADATA: return sizeof(Metadata*);
3946 // We use T_VOID as marker for jump-table entries (labels) which
3947 // need an internal word relocation.
3948 case T_VOID:
3949 case T_ADDRESS:
3950 case T_OBJECT: return sizeof(jobject);
3951 default:
3952 ShouldNotReachHere();
3953 return -1;
3954 }
3955 }
3956
3957 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3958 // sort descending
3959 if (a->freq() > b->freq()) return -1;
3960 if (a->freq() < b->freq()) return 1;
3961 return 0;
3962 }
3963
3964 void Compile::ConstantTable::calculate_offsets_and_size() {
3965 // First, sort the array by frequencies.
3966 _constants.sort(qsort_comparator);
3967
3968 #ifdef ASSERT
3969 // Make sure all jump-table entries were sorted to the end of the
3970 // array (they have a negative frequency).
3971 bool found_void = false;
3972 for (int i = 0; i < _constants.length(); i++) {
3973 Constant con = _constants.at(i);
3974 if (con.type() == T_VOID)
3975 found_void = true; // jump-tables
3976 else
3977 assert(!found_void, "wrong sorting");
3978 }
3979 #endif
3980
3981 int offset = 0;
3982 for (int i = 0; i < _constants.length(); i++) {
3983 Constant* con = _constants.adr_at(i);
3984
3985 // Align offset for type.
3986 int typesize = type_to_size_in_bytes(con->type());
3987 offset = align_up(offset, typesize);
3988 con->set_offset(offset); // set constant's offset
3989
3990 if (con->type() == T_VOID) {
3991 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3992 offset = offset + typesize * n->outcnt(); // expand jump-table
3993 } else {
3994 offset = offset + typesize;
3995 }
3996 }
3997
3998 // Align size up to the next section start (which is insts; see
3999 // CodeBuffer::align_at_start).
4000 assert(_size == -1, "already set?");
4001 _size = align_up(offset, (int)CodeEntryAlignment);
4002 }
4003
4004 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4005 MacroAssembler _masm(&cb);
4006 for (int i = 0; i < _constants.length(); i++) {
4007 Constant con = _constants.at(i);
4008 address constant_addr = NULL;
4009 switch (con.type()) {
4010 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4011 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4012 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4013 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4014 case T_OBJECT: {
4015 jobject obj = con.get_jobject();
4016 int oop_index = _masm.oop_recorder()->find_index(obj);
4017 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4018 break;
4019 }
4020 case T_ADDRESS: {
4021 address addr = (address) con.get_jobject();
4022 constant_addr = _masm.address_constant(addr);
4023 break;
4024 }
4025 // We use T_VOID as marker for jump-table entries (labels) which
4026 // need an internal word relocation.
4027 case T_VOID: {
4028 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4029 // Fill the jump-table with a dummy word. The real value is
4030 // filled in later in fill_jump_table.
4031 address dummy = (address) n;
4032 constant_addr = _masm.address_constant(dummy);
4033 // Expand jump-table
4034 for (uint i = 1; i < n->outcnt(); i++) {
4035 address temp_addr = _masm.address_constant(dummy + i);
4036 assert(temp_addr, "consts section too small");
4037 }
4038 break;
4039 }
4040 case T_METADATA: {
4041 Metadata* obj = con.get_metadata();
4042 int metadata_index = _masm.oop_recorder()->find_index(obj);
4043 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4044 break;
4045 }
4046 default: ShouldNotReachHere();
4047 }
4048 assert(constant_addr, "consts section too small");
4049 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4050 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4051 }
4052 }
4053
4054 int Compile::ConstantTable::find_offset(Constant& con) const {
4055 int idx = _constants.find(con);
4056 guarantee(idx != -1, "constant must be in constant table");
4057 int offset = _constants.at(idx).offset();
4058 guarantee(offset != -1, "constant table not emitted yet?");
4059 return offset;
4060 }
4061
4062 void Compile::ConstantTable::add(Constant& con) {
4063 if (con.can_be_reused()) {
4064 int idx = _constants.find(con);
4065 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4066 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4067 return;
4068 }
4069 }
4070 (void) _constants.append(con);
4071 }
4072
4073 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4074 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4075 Constant con(type, value, b->_freq);
4076 add(con);
4077 return con;
4078 }
4079
4080 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4081 Constant con(metadata);
4082 add(con);
4083 return con;
4084 }
4085
4086 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4087 jvalue value;
4088 BasicType type = oper->type()->basic_type();
4089 switch (type) {
4090 case T_LONG: value.j = oper->constantL(); break;
4091 case T_FLOAT: value.f = oper->constantF(); break;
4092 case T_DOUBLE: value.d = oper->constantD(); break;
4093 case T_OBJECT:
4094 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4095 case T_METADATA: return add((Metadata*)oper->constant()); break;
4096 default: guarantee(false, "unhandled type: %s", type2name(type));
4097 }
4098 return add(n, type, value);
4099 }
4100
4101 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4102 jvalue value;
4103 // We can use the node pointer here to identify the right jump-table
4104 // as this method is called from Compile::Fill_buffer right before
4105 // the MachNodes are emitted and the jump-table is filled (means the
4106 // MachNode pointers do not change anymore).
4107 value.l = (jobject) n;
4108 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4109 add(con);
4110 return con;
4111 }
4112
4113 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4114 // If called from Compile::scratch_emit_size do nothing.
4115 if (Compile::current()->in_scratch_emit_size()) return;
4116
4117 assert(labels.is_nonempty(), "must be");
4118 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4119
4120 // Since MachConstantNode::constant_offset() also contains
4121 // table_base_offset() we need to subtract the table_base_offset()
4122 // to get the plain offset into the constant table.
4123 int offset = n->constant_offset() - table_base_offset();
4124
4125 MacroAssembler _masm(&cb);
4126 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4127
4128 for (uint i = 0; i < n->outcnt(); i++) {
4129 address* constant_addr = &jump_table_base[i];
4130 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));
4131 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4132 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4133 }
4134 }
4135
4136 //----------------------------static_subtype_check-----------------------------
4137 // Shortcut important common cases when superklass is exact:
4138 // (0) superklass is java.lang.Object (can occur in reflective code)
4139 // (1) subklass is already limited to a subtype of superklass => always ok
4140 // (2) subklass does not overlap with superklass => always fail
4141 // (3) superklass has NO subtypes and we can check with a simple compare.
4142 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4143 if (StressReflectiveCode) {
4144 return SSC_full_test; // Let caller generate the general case.
4145 }
4146
4147 if (superk == env()->Object_klass()) {
4148 return SSC_always_true; // (0) this test cannot fail
4149 }
4150
4151 ciType* superelem = superk;
4152 if (superelem->is_array_klass())
4153 superelem = superelem->as_array_klass()->base_element_type();
4154
4155 if (!subk->is_interface()) { // cannot trust static interface types yet
4156 if (subk->is_subtype_of(superk)) {
4157 return SSC_always_true; // (1) false path dead; no dynamic test needed
4158 }
4159 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4160 !superk->is_subtype_of(subk)) {
4161 return SSC_always_false;
4162 }
4163 }
4164
4165 // If casting to an instance klass, it must have no subtypes
4166 if (superk->is_interface()) {
4167 // Cannot trust interfaces yet.
4168 // %%% S.B. superk->nof_implementors() == 1
4169 } else if (superelem->is_instance_klass()) {
4170 ciInstanceKlass* ik = superelem->as_instance_klass();
4171 if (!ik->has_subklass() && !ik->is_interface()) {
4172 if (!ik->is_final()) {
4173 // Add a dependency if there is a chance of a later subclass.
4174 dependencies()->assert_leaf_type(ik);
4175 }
4176 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4177 }
4178 } else {
4179 // A primitive array type has no subtypes.
4180 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4181 }
4182
4183 return SSC_full_test;
4184 }
4185
4186 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4187 #ifdef _LP64
4188 // The scaled index operand to AddP must be a clean 64-bit value.
4189 // Java allows a 32-bit int to be incremented to a negative
4190 // value, which appears in a 64-bit register as a large
4191 // positive number. Using that large positive number as an
4192 // operand in pointer arithmetic has bad consequences.
4193 // On the other hand, 32-bit overflow is rare, and the possibility
4194 // can often be excluded, if we annotate the ConvI2L node with
4195 // a type assertion that its value is known to be a small positive
4196 // number. (The prior range check has ensured this.)
4197 // This assertion is used by ConvI2LNode::Ideal.
4198 int index_max = max_jint - 1; // array size is max_jint, index is one less
4199 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4200 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4201 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4202 #endif
4203 return idx;
4204 }
4205
4206 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4207 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4208 if (ctrl != NULL) {
4209 // Express control dependency by a CastII node with a narrow type.
4210 value = new CastIINode(value, itype, false, true /* range check dependency */);
4211 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4212 // node from floating above the range check during loop optimizations. Otherwise, the
4213 // ConvI2L node may be eliminated independently of the range check, causing the data path
4214 // to become TOP while the control path is still there (although it's unreachable).
4215 value->set_req(0, ctrl);
4216 // Save CastII node to remove it after loop optimizations.
4217 phase->C->add_range_check_cast(value);
4218 value = phase->transform(value);
4219 }
4220 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4221 return phase->transform(new ConvI2LNode(value, ltype));
4222 }
4223
4224 // The message about the current inlining is accumulated in
4225 // _print_inlining_stream and transfered into the _print_inlining_list
4226 // once we know whether inlining succeeds or not. For regular
4227 // inlining, messages are appended to the buffer pointed by
4228 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4229 // a new buffer is added after _print_inlining_idx in the list. This
4230 // way we can update the inlining message for late inlining call site
4231 // when the inlining is attempted again.
4232 void Compile::print_inlining_init() {
4233 if (print_inlining() || print_intrinsics()) {
4234 _print_inlining_stream = new stringStream();
4235 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4236 }
4237 }
4238
4239 void Compile::print_inlining_reinit() {
4240 if (print_inlining() || print_intrinsics()) {
4241 // Re allocate buffer when we change ResourceMark
4242 _print_inlining_stream = new stringStream();
4243 }
4244 }
4245
4246 void Compile::print_inlining_reset() {
4247 _print_inlining_stream->reset();
4248 }
4249
4250 void Compile::print_inlining_commit() {
4251 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4252 // Transfer the message from _print_inlining_stream to the current
4253 // _print_inlining_list buffer and clear _print_inlining_stream.
4254 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4255 print_inlining_reset();
4256 }
4257
4258 void Compile::print_inlining_push() {
4259 // Add new buffer to the _print_inlining_list at current position
4260 _print_inlining_idx++;
4261 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4262 }
4263
4264 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4265 return _print_inlining_list->at(_print_inlining_idx);
4266 }
4267
4268 void Compile::print_inlining_update(CallGenerator* cg) {
4269 if (print_inlining() || print_intrinsics()) {
4270 if (!cg->is_late_inline()) {
4271 if (print_inlining_current().cg() != NULL) {
4272 print_inlining_push();
4273 }
4274 print_inlining_commit();
4275 } else {
4276 if (print_inlining_current().cg() != cg &&
4277 (print_inlining_current().cg() != NULL ||
4278 print_inlining_current().ss()->size() != 0)) {
4279 print_inlining_push();
4280 }
4281 print_inlining_commit();
4282 print_inlining_current().set_cg(cg);
4283 }
4284 }
4285 }
4286
4287 void Compile::print_inlining_move_to(CallGenerator* cg) {
4288 // We resume inlining at a late inlining call site. Locate the
4289 // corresponding inlining buffer so that we can update it.
4290 if (print_inlining()) {
4291 for (int i = 0; i < _print_inlining_list->length(); i++) {
4292 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4293 _print_inlining_idx = i;
4294 return;
4295 }
4296 }
4297 ShouldNotReachHere();
4298 }
4299 }
4300
4301 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4302 if (print_inlining()) {
4303 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4304 assert(print_inlining_current().cg() == cg, "wrong entry");
4305 // replace message with new message
4306 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4307 print_inlining_commit();
4308 print_inlining_current().set_cg(cg);
4309 }
4310 }
4311
4312 void Compile::print_inlining_assert_ready() {
4313 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4314 }
4315
4316 void Compile::process_print_inlining() {
4317 bool do_print_inlining = print_inlining() || print_intrinsics();
4318 if (do_print_inlining || log() != NULL) {
4319 // Print inlining message for candidates that we couldn't inline
4320 // for lack of space
4321 for (int i = 0; i < _late_inlines.length(); i++) {
4322 CallGenerator* cg = _late_inlines.at(i);
4323 if (!cg->is_mh_late_inline()) {
4324 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4325 if (do_print_inlining) {
4326 cg->print_inlining_late(msg);
4327 }
4328 log_late_inline_failure(cg, msg);
4329 }
4330 }
4331 }
4332 if (do_print_inlining) {
4333 ResourceMark rm;
4334 stringStream ss;
4335 for (int i = 0; i < _print_inlining_list->length(); i++) {
4336 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4337 }
4338 size_t end = ss.size();
4339 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4340 strncpy(_print_inlining_output, ss.base(), end+1);
4341 _print_inlining_output[end] = 0;
4342 }
4343 }
4344
4345 void Compile::dump_print_inlining() {
4346 if (_print_inlining_output != NULL) {
4347 tty->print_raw(_print_inlining_output);
4348 }
4349 }
4350
4351 void Compile::log_late_inline(CallGenerator* cg) {
4352 if (log() != NULL) {
4353 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4354 cg->unique_id());
4355 JVMState* p = cg->call_node()->jvms();
4356 while (p != NULL) {
4357 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4358 p = p->caller();
4359 }
4360 log()->tail("late_inline");
4361 }
4362 }
4363
4364 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4365 log_late_inline(cg);
4366 if (log() != NULL) {
4367 log()->inline_fail(msg);
4368 }
4369 }
4370
4371 void Compile::log_inline_id(CallGenerator* cg) {
4372 if (log() != NULL) {
4373 // The LogCompilation tool needs a unique way to identify late
4374 // inline call sites. This id must be unique for this call site in
4375 // this compilation. Try to have it unique across compilations as
4376 // well because it can be convenient when grepping through the log
4377 // file.
4378 // Distinguish OSR compilations from others in case CICountOSR is
4379 // on.
4380 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4381 cg->set_unique_id(id);
4382 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4383 }
4384 }
4385
4386 void Compile::log_inline_failure(const char* msg) {
4387 if (C->log() != NULL) {
4388 C->log()->inline_fail(msg);
4389 }
4390 }
4391
4392
4393 // Dump inlining replay data to the stream.
4394 // Don't change thread state and acquire any locks.
4395 void Compile::dump_inline_data(outputStream* out) {
4396 InlineTree* inl_tree = ilt();
4397 if (inl_tree != NULL) {
4398 out->print(" inline %d", inl_tree->count());
4399 inl_tree->dump_replay_data(out);
4400 }
4401 }
4402
4403 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4404 if (n1->Opcode() < n2->Opcode()) return -1;
4405 else if (n1->Opcode() > n2->Opcode()) return 1;
4406
4407 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4408 for (uint i = 1; i < n1->req(); i++) {
4409 if (n1->in(i) < n2->in(i)) return -1;
4410 else if (n1->in(i) > n2->in(i)) return 1;
4411 }
4412
4413 return 0;
4414 }
4415
4416 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4417 Node* n1 = *n1p;
4418 Node* n2 = *n2p;
4419
4420 return cmp_expensive_nodes(n1, n2);
4421 }
4422
4423 void Compile::sort_expensive_nodes() {
4424 if (!expensive_nodes_sorted()) {
4425 _expensive_nodes->sort(cmp_expensive_nodes);
4426 }
4427 }
4428
4429 bool Compile::expensive_nodes_sorted() const {
4430 for (int i = 1; i < _expensive_nodes->length(); i++) {
4431 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4432 return false;
4433 }
4434 }
4435 return true;
4436 }
4437
4438 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4439 if (_expensive_nodes->length() == 0) {
4440 return false;
4441 }
4442
4443 assert(OptimizeExpensiveOps, "optimization off?");
4444
4445 // Take this opportunity to remove dead nodes from the list
4446 int j = 0;
4447 for (int i = 0; i < _expensive_nodes->length(); i++) {
4448 Node* n = _expensive_nodes->at(i);
4449 if (!n->is_unreachable(igvn)) {
4450 assert(n->is_expensive(), "should be expensive");
4451 _expensive_nodes->at_put(j, n);
4452 j++;
4453 }
4454 }
4455 _expensive_nodes->trunc_to(j);
4456
4457 // Then sort the list so that similar nodes are next to each other
4458 // and check for at least two nodes of identical kind with same data
4459 // inputs.
4460 sort_expensive_nodes();
4461
4462 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4463 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4464 return true;
4465 }
4466 }
4467
4468 return false;
4469 }
4470
4471 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4472 if (_expensive_nodes->length() == 0) {
4473 return;
4474 }
4475
4476 assert(OptimizeExpensiveOps, "optimization off?");
4477
4478 // Sort to bring similar nodes next to each other and clear the
4479 // control input of nodes for which there's only a single copy.
4480 sort_expensive_nodes();
4481
4482 int j = 0;
4483 int identical = 0;
4484 int i = 0;
4485 bool modified = false;
4486 for (; i < _expensive_nodes->length()-1; i++) {
4487 assert(j <= i, "can't write beyond current index");
4488 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4489 identical++;
4490 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4491 continue;
4492 }
4493 if (identical > 0) {
4494 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4495 identical = 0;
4496 } else {
4497 Node* n = _expensive_nodes->at(i);
4498 igvn.replace_input_of(n, 0, NULL);
4499 igvn.hash_insert(n);
4500 modified = true;
4501 }
4502 }
4503 if (identical > 0) {
4504 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4505 } else if (_expensive_nodes->length() >= 1) {
4506 Node* n = _expensive_nodes->at(i);
4507 igvn.replace_input_of(n, 0, NULL);
4508 igvn.hash_insert(n);
4509 modified = true;
4510 }
4511 _expensive_nodes->trunc_to(j);
4512 if (modified) {
4513 igvn.optimize();
4514 }
4515 }
4516
4517 void Compile::add_expensive_node(Node * n) {
4518 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4519 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4520 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4521 if (OptimizeExpensiveOps) {
4522 _expensive_nodes->append(n);
4523 } else {
4524 // Clear control input and let IGVN optimize expensive nodes if
4525 // OptimizeExpensiveOps is off.
4526 n->set_req(0, NULL);
4527 }
4528 }
4529
4530 /**
4531 * Remove the speculative part of types and clean up the graph
4532 */
4533 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4534 if (UseTypeSpeculation) {
4535 Unique_Node_List worklist;
4536 worklist.push(root());
4537 int modified = 0;
4538 // Go over all type nodes that carry a speculative type, drop the
4539 // speculative part of the type and enqueue the node for an igvn
4540 // which may optimize it out.
4541 for (uint next = 0; next < worklist.size(); ++next) {
4542 Node *n = worklist.at(next);
4543 if (n->is_Type()) {
4544 TypeNode* tn = n->as_Type();
4545 const Type* t = tn->type();
4546 const Type* t_no_spec = t->remove_speculative();
4547 if (t_no_spec != t) {
4548 bool in_hash = igvn.hash_delete(n);
4549 assert(in_hash, "node should be in igvn hash table");
4550 tn->set_type(t_no_spec);
4551 igvn.hash_insert(n);
4552 igvn._worklist.push(n); // give it a chance to go away
4553 modified++;
4554 }
4555 }
4556 uint max = n->len();
4557 for( uint i = 0; i < max; ++i ) {
4558 Node *m = n->in(i);
4559 if (not_a_node(m)) continue;
4560 worklist.push(m);
4561 }
4562 }
4563 // Drop the speculative part of all types in the igvn's type table
4564 igvn.remove_speculative_types();
4565 if (modified > 0) {
4566 igvn.optimize();
4567 }
4568 #ifdef ASSERT
4569 // Verify that after the IGVN is over no speculative type has resurfaced
4570 worklist.clear();
4571 worklist.push(root());
4572 for (uint next = 0; next < worklist.size(); ++next) {
4573 Node *n = worklist.at(next);
4574 const Type* t = igvn.type_or_null(n);
4575 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4576 if (n->is_Type()) {
4577 t = n->as_Type()->type();
4578 assert(t == t->remove_speculative(), "no more speculative types");
4579 }
4580 uint max = n->len();
4581 for( uint i = 0; i < max; ++i ) {
4582 Node *m = n->in(i);
4583 if (not_a_node(m)) continue;
4584 worklist.push(m);
4585 }
4586 }
4587 igvn.check_no_speculative_types();
4588 #endif
4589 }
4590 }
4591
4592 // Auxiliary method to support randomized stressing/fuzzing.
4593 //
4594 // This method can be called the arbitrary number of times, with current count
4595 // as the argument. The logic allows selecting a single candidate from the
4596 // running list of candidates as follows:
4597 // int count = 0;
4598 // Cand* selected = null;
4599 // while(cand = cand->next()) {
4600 // if (randomized_select(++count)) {
4601 // selected = cand;
4602 // }
4603 // }
4604 //
4605 // Including count equalizes the chances any candidate is "selected".
4606 // This is useful when we don't have the complete list of candidates to choose
4607 // from uniformly. In this case, we need to adjust the randomicity of the
4608 // selection, or else we will end up biasing the selection towards the latter
4609 // candidates.
4610 //
4611 // Quick back-envelope calculation shows that for the list of n candidates
4612 // the equal probability for the candidate to persist as "best" can be
4613 // achieved by replacing it with "next" k-th candidate with the probability
4614 // of 1/k. It can be easily shown that by the end of the run, the
4615 // probability for any candidate is converged to 1/n, thus giving the
4616 // uniform distribution among all the candidates.
4617 //
4618 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4619 #define RANDOMIZED_DOMAIN_POW 29
4620 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4621 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4622 bool Compile::randomized_select(int count) {
4623 assert(count > 0, "only positive");
4624 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4625 }
4626
4627 CloneMap& Compile::clone_map() { return _clone_map; }
4628 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4629
4630 void NodeCloneInfo::dump() const {
4631 tty->print(" {%d:%d} ", idx(), gen());
4632 }
4633
4634 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4635 uint64_t val = value(old->_idx);
4636 NodeCloneInfo cio(val);
4637 assert(val != 0, "old node should be in the map");
4638 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4639 insert(nnn->_idx, cin.get());
4640 #ifndef PRODUCT
4641 if (is_debug()) {
4642 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4643 }
4644 #endif
4645 }
4646
4647 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4648 NodeCloneInfo cio(value(old->_idx));
4649 if (cio.get() == 0) {
4650 cio.set(old->_idx, 0);
4651 insert(old->_idx, cio.get());
4652 #ifndef PRODUCT
4653 if (is_debug()) {
4654 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4655 }
4656 #endif
4657 }
4658 clone(old, nnn, gen);
4659 }
4660
4661 int CloneMap::max_gen() const {
4662 int g = 0;
4663 DictI di(_dict);
4664 for(; di.test(); ++di) {
4665 int t = gen(di._key);
4666 if (g < t) {
4667 g = t;
4668 #ifndef PRODUCT
4669 if (is_debug()) {
4670 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4671 }
4672 #endif
4673 }
4674 }
4675 return g;
4676 }
4677
4678 void CloneMap::dump(node_idx_t key) const {
4679 uint64_t val = value(key);
4680 if (val != 0) {
4681 NodeCloneInfo ni(val);
4682 ni.dump();
4683 }
4684 }