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
83
84 // -------------------- Compile::mach_constant_base_node -----------------------
85 // Constant table base node singleton.
86 MachConstantBaseNode* Compile::mach_constant_base_node() {
87 if (_mach_constant_base_node == NULL) {
88 _mach_constant_base_node = new MachConstantBaseNode();
89 _mach_constant_base_node->add_req(C->root());
90 }
91 return _mach_constant_base_node;
92 }
93
94
95 /// Support for intrinsics.
96
97 // Return the index at which m must be inserted (or already exists).
98 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
99 class IntrinsicDescPair {
100 private:
101 ciMethod* _m;
102 bool _is_virtual;
103 public:
104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
106 ciMethod* m= elt->method();
107 ciMethod* key_m = key->_m;
108 if (key_m < m) return -1;
109 else if (key_m > m) return 1;
110 else {
111 bool is_virtual = elt->is_virtual();
112 bool key_virtual = key->_is_virtual;
113 if (key_virtual < is_virtual) return -1;
114 else if (key_virtual > is_virtual) return 1;
115 else return 0;
116 }
117 }
118 };
119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
120 #ifdef ASSERT
121 for (int i = 1; i < _intrinsics->length(); i++) {
122 CallGenerator* cg1 = _intrinsics->at(i-1);
123 CallGenerator* cg2 = _intrinsics->at(i);
124 assert(cg1->method() != cg2->method()
125 ? cg1->method() < cg2->method()
126 : cg1->is_virtual() < cg2->is_virtual(),
127 "compiler intrinsics list must stay sorted");
128 }
129 #endif
130 IntrinsicDescPair pair(m, is_virtual);
131 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
132 }
133
134 void Compile::register_intrinsic(CallGenerator* cg) {
135 if (_intrinsics == NULL) {
136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
137 }
138 int len = _intrinsics->length();
139 bool found = false;
140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
141 assert(!found, "registering twice");
142 _intrinsics->insert_before(index, cg);
143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
144 }
145
146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
147 assert(m->is_loaded(), "don't try this on unloaded methods");
148 if (_intrinsics != NULL) {
149 bool found = false;
150 int index = intrinsic_insertion_index(m, is_virtual, found);
151 if (found) {
152 return _intrinsics->at(index);
153 }
154 }
155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
156 if (m->intrinsic_id() != vmIntrinsics::_none &&
157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
159 if (cg != NULL) {
160 // Save it for next time:
161 register_intrinsic(cg);
162 return cg;
163 } else {
164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
165 }
166 }
167 return NULL;
168 }
169
170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
171 // in library_call.cpp.
172
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
179
180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
182 int oflags = _intrinsic_hist_flags[id];
183 assert(flags != 0, "what happened?");
184 if (is_virtual) {
185 flags |= _intrinsic_virtual;
186 }
187 bool changed = (flags != oflags);
188 if ((flags & _intrinsic_worked) != 0) {
189 juint count = (_intrinsic_hist_count[id] += 1);
190 if (count == 1) {
191 changed = true; // first time
192 }
193 // increment the overall count also:
194 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
195 }
196 if (changed) {
197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
198 // Something changed about the intrinsic's virtuality.
199 if ((flags & _intrinsic_virtual) != 0) {
200 // This is the first use of this intrinsic as a virtual call.
201 if (oflags != 0) {
202 // We already saw it as a non-virtual, so note both cases.
203 flags |= _intrinsic_both;
204 }
205 } else if ((oflags & _intrinsic_both) == 0) {
206 // This is the first use of this intrinsic as a non-virtual
207 flags |= _intrinsic_both;
208 }
209 }
210 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
211 }
212 // update the overall flags also:
213 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
214 return changed;
215 }
216
217 static char* format_flags(int flags, char* buf) {
218 buf[0] = 0;
219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
224 if (buf[0] == 0) strcat(buf, ",");
225 assert(buf[0] == ',', "must be");
226 return &buf[1];
227 }
228
229 void Compile::print_intrinsic_statistics() {
230 char flagsbuf[100];
231 ttyLocker ttyl;
232 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
233 tty->print_cr("Compiler intrinsic usage:");
234 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
235 if (total == 0) total = 1; // avoid div0 in case of no successes
236 #define PRINT_STAT_LINE(name, c, f) \
237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
238 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
239 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
240 int flags = _intrinsic_hist_flags[id];
241 juint count = _intrinsic_hist_count[id];
242 if ((flags | count) != 0) {
243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
244 }
245 }
246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
247 if (xtty != NULL) xtty->tail("statistics");
248 }
249
250 void Compile::print_statistics() {
251 { ttyLocker ttyl;
252 if (xtty != NULL) xtty->head("statistics type='opto'");
253 Parse::print_statistics();
254 PhaseCCP::print_statistics();
255 PhaseRegAlloc::print_statistics();
256 Scheduling::print_statistics();
257 PhasePeephole::print_statistics();
258 PhaseIdealLoop::print_statistics();
259 if (xtty != NULL) xtty->tail("statistics");
260 }
261 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
262 // put this under its own <statistics> element.
263 print_intrinsic_statistics();
264 }
265 }
266 #endif //PRODUCT
267
268 // Support for bundling info
269 Bundle* Compile::node_bundling(const Node *n) {
270 assert(valid_bundle_info(n), "oob");
271 return &_node_bundling_base[n->_idx];
272 }
273
274 bool Compile::valid_bundle_info(const Node *n) {
275 return (_node_bundling_limit > n->_idx);
276 }
277
278
279 void Compile::gvn_replace_by(Node* n, Node* nn) {
280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
281 Node* use = n->last_out(i);
282 bool is_in_table = initial_gvn()->hash_delete(use);
283 uint uses_found = 0;
284 for (uint j = 0; j < use->len(); j++) {
285 if (use->in(j) == n) {
286 if (j < use->req())
287 use->set_req(j, nn);
288 else
289 use->set_prec(j, nn);
290 uses_found++;
291 }
292 }
293 if (is_in_table) {
294 // reinsert into table
295 initial_gvn()->hash_find_insert(use);
296 }
297 record_for_igvn(use);
298 i -= uses_found; // we deleted 1 or more copies of this edge
299 }
300 }
301
302
303 static inline bool not_a_node(const Node* n) {
304 if (n == NULL) return true;
305 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
306 if (*(address*)n == badAddress) return true; // kill by Node::destruct
307 return false;
308 }
309
310 // Identify all nodes that are reachable from below, useful.
311 // Use breadth-first pass that records state in a Unique_Node_List,
312 // recursive traversal is slower.
313 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
314 int estimated_worklist_size = live_nodes();
315 useful.map( estimated_worklist_size, NULL ); // preallocate space
316
317 // Initialize worklist
318 if (root() != NULL) { useful.push(root()); }
319 // If 'top' is cached, declare it useful to preserve cached node
320 if( cached_top_node() ) { useful.push(cached_top_node()); }
321
322 // Push all useful nodes onto the list, breadthfirst
323 for( uint next = 0; next < useful.size(); ++next ) {
324 assert( next < unique(), "Unique useful nodes < total nodes");
325 Node *n = useful.at(next);
326 uint max = n->len();
327 for( uint i = 0; i < max; ++i ) {
328 Node *m = n->in(i);
329 if (not_a_node(m)) continue;
330 useful.push(m);
331 }
332 }
333 }
334
335 // Update dead_node_list with any missing dead nodes using useful
336 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
337 void Compile::update_dead_node_list(Unique_Node_List &useful) {
338 uint max_idx = unique();
339 VectorSet& useful_node_set = useful.member_set();
340
341 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
342 // If node with index node_idx is not in useful set,
343 // mark it as dead in dead node list.
344 if (! useful_node_set.test(node_idx) ) {
345 record_dead_node(node_idx);
346 }
347 }
348 }
349
350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
351 int shift = 0;
352 for (int i = 0; i < inlines->length(); i++) {
353 CallGenerator* cg = inlines->at(i);
354 CallNode* call = cg->call_node();
355 if (shift > 0) {
356 inlines->at_put(i-shift, cg);
357 }
358 if (!useful.member(call)) {
359 shift++;
360 }
361 }
362 inlines->trunc_to(inlines->length()-shift);
363 }
364
365 // Disconnect all useless nodes by disconnecting those at the boundary.
366 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
367 uint next = 0;
368 while (next < useful.size()) {
369 Node *n = useful.at(next++);
370 if (n->is_SafePoint()) {
371 // We're done with a parsing phase. Replaced nodes are not valid
372 // beyond that point.
373 n->as_SafePoint()->delete_replaced_nodes();
374 }
375 // Use raw traversal of out edges since this code removes out edges
376 int max = n->outcnt();
377 for (int j = 0; j < max; ++j) {
378 Node* child = n->raw_out(j);
379 if (! useful.member(child)) {
380 assert(!child->is_top() || child != top(),
381 "If top is cached in Compile object it is in useful list");
382 // Only need to remove this out-edge to the useless node
383 n->raw_del_out(j);
384 --j;
385 --max;
386 }
387 }
388 if (n->outcnt() == 1 && n->has_special_unique_user()) {
389 record_for_igvn(n->unique_out());
390 }
391 }
392 // Remove useless macro and predicate opaq nodes
393 for (int i = C->macro_count()-1; i >= 0; i--) {
394 Node* n = C->macro_node(i);
395 if (!useful.member(n)) {
396 remove_macro_node(n);
397 }
398 }
399 // Remove useless CastII nodes with range check dependency
400 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
401 Node* cast = range_check_cast_node(i);
402 if (!useful.member(cast)) {
403 remove_range_check_cast(cast);
404 }
405 }
406 // Remove useless expensive nodes
407 for (int i = C->expensive_count()-1; i >= 0; i--) {
408 Node* n = C->expensive_node(i);
409 if (!useful.member(n)) {
410 remove_expensive_node(n);
411 }
412 }
413 // Remove useless Opaque4 nodes
414 for (int i = opaque4_count() - 1; i >= 0; i--) {
415 Node* opaq = opaque4_node(i);
416 if (!useful.member(opaq)) {
417 remove_opaque4_node(opaq);
418 }
419 }
420 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
421 bs->eliminate_useless_gc_barriers(useful, this);
422 // clean up the late inline lists
423 remove_useless_late_inlines(&_string_late_inlines, useful);
424 remove_useless_late_inlines(&_boxing_late_inlines, useful);
425 remove_useless_late_inlines(&_late_inlines, useful);
426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
427 }
428
429 //------------------------------frame_size_in_words-----------------------------
430 // frame_slots in units of words
431 int Compile::frame_size_in_words() const {
432 // shift is 0 in LP32 and 1 in LP64
433 const int shift = (LogBytesPerWord - LogBytesPerInt);
434 int words = _frame_slots >> shift;
435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
436 return words;
437 }
438
439 // To bang the stack of this compiled method we use the stack size
440 // that the interpreter would need in case of a deoptimization. This
441 // removes the need to bang the stack in the deoptimization blob which
442 // in turn simplifies stack overflow handling.
443 int Compile::bang_size_in_bytes() const {
444 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
445 }
446
447 // ============================================================================
448 //------------------------------CompileWrapper---------------------------------
449 class CompileWrapper : public StackObj {
450 Compile *const _compile;
451 public:
452 CompileWrapper(Compile* compile);
453
454 ~CompileWrapper();
455 };
456
457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
458 // the Compile* pointer is stored in the current ciEnv:
459 ciEnv* env = compile->env();
460 assert(env == ciEnv::current(), "must already be a ciEnv active");
461 assert(env->compiler_data() == NULL, "compile already active?");
462 env->set_compiler_data(compile);
463 assert(compile == Compile::current(), "sanity");
464
465 compile->set_type_dict(NULL);
466 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
467 compile->clone_map().set_clone_idx(0);
468 compile->set_type_hwm(NULL);
469 compile->set_type_last_size(0);
470 compile->set_last_tf(NULL, NULL);
471 compile->set_indexSet_arena(NULL);
472 compile->set_indexSet_free_block_list(NULL);
473 compile->init_type_arena();
474 Type::Initialize(compile);
475 _compile->set_scratch_buffer_blob(NULL);
476 _compile->begin_method();
477 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
478 }
479 CompileWrapper::~CompileWrapper() {
480 _compile->end_method();
481 if (_compile->scratch_buffer_blob() != NULL)
482 BufferBlob::free(_compile->scratch_buffer_blob());
483 _compile->env()->set_compiler_data(NULL);
484 }
485
486
487 //----------------------------print_compile_messages---------------------------
488 void Compile::print_compile_messages() {
489 #ifndef PRODUCT
490 // Check if recompiling
491 if (_subsume_loads == false && PrintOpto) {
492 // Recompiling without allowing machine instructions to subsume loads
493 tty->print_cr("*********************************************************");
494 tty->print_cr("** Bailout: Recompile without subsuming loads **");
495 tty->print_cr("*********************************************************");
496 }
497 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
498 // Recompiling without escape analysis
499 tty->print_cr("*********************************************************");
500 tty->print_cr("** Bailout: Recompile without escape analysis **");
501 tty->print_cr("*********************************************************");
502 }
503 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
504 // Recompiling without boxing elimination
505 tty->print_cr("*********************************************************");
506 tty->print_cr("** Bailout: Recompile without boxing elimination **");
507 tty->print_cr("*********************************************************");
508 }
509 if (C->directive()->BreakAtCompileOption) {
510 // Open the debugger when compiling this method.
511 tty->print("### Breaking when compiling: ");
512 method()->print_short_name();
513 tty->cr();
514 BREAKPOINT;
515 }
516
517 if( PrintOpto ) {
518 if (is_osr_compilation()) {
519 tty->print("[OSR]%3d", _compile_id);
520 } else {
521 tty->print("%3d", _compile_id);
522 }
523 }
524 #endif
525 }
526
527
528 //-----------------------init_scratch_buffer_blob------------------------------
529 // Construct a temporary BufferBlob and cache it for this compile.
530 void Compile::init_scratch_buffer_blob(int const_size) {
531 // If there is already a scratch buffer blob allocated and the
532 // constant section is big enough, use it. Otherwise free the
533 // current and allocate a new one.
534 BufferBlob* blob = scratch_buffer_blob();
535 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
536 // Use the current blob.
537 } else {
538 if (blob != NULL) {
539 BufferBlob::free(blob);
540 }
541
542 ResourceMark rm;
543 _scratch_const_size = const_size;
544 int size = C2Compiler::initial_code_buffer_size(const_size);
545 blob = BufferBlob::create("Compile::scratch_buffer", size);
546 // Record the buffer blob for next time.
547 set_scratch_buffer_blob(blob);
548 // Have we run out of code space?
549 if (scratch_buffer_blob() == NULL) {
550 // Let CompilerBroker disable further compilations.
551 record_failure("Not enough space for scratch buffer in CodeCache");
552 return;
553 }
554 }
555
556 // Initialize the relocation buffers
557 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
558 set_scratch_locs_memory(locs_buf);
559 }
560
561
562 //-----------------------scratch_emit_size-------------------------------------
563 // Helper function that computes size by emitting code
564 uint Compile::scratch_emit_size(const Node* n) {
565 // Start scratch_emit_size section.
566 set_in_scratch_emit_size(true);
567
568 // Emit into a trash buffer and count bytes emitted.
569 // This is a pretty expensive way to compute a size,
570 // but it works well enough if seldom used.
571 // All common fixed-size instructions are given a size
572 // method by the AD file.
573 // Note that the scratch buffer blob and locs memory are
574 // allocated at the beginning of the compile task, and
575 // may be shared by several calls to scratch_emit_size.
576 // The allocation of the scratch buffer blob is particularly
577 // expensive, since it has to grab the code cache lock.
578 BufferBlob* blob = this->scratch_buffer_blob();
579 assert(blob != NULL, "Initialize BufferBlob at start");
580 assert(blob->size() > MAX_inst_size, "sanity");
581 relocInfo* locs_buf = scratch_locs_memory();
582 address blob_begin = blob->content_begin();
583 address blob_end = (address)locs_buf;
584 assert(blob->contains(blob_end), "sanity");
585 CodeBuffer buf(blob_begin, blob_end - blob_begin);
586 buf.initialize_consts_size(_scratch_const_size);
587 buf.initialize_stubs_size(MAX_stubs_size);
588 assert(locs_buf != NULL, "sanity");
589 int lsize = MAX_locs_size / 3;
590 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
591 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
592 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
593 // Mark as scratch buffer.
594 buf.consts()->set_scratch_emit();
595 buf.insts()->set_scratch_emit();
596 buf.stubs()->set_scratch_emit();
597
598 // Do the emission.
599
600 Label fakeL; // Fake label for branch instructions.
601 Label* saveL = NULL;
602 uint save_bnum = 0;
603 bool is_branch = n->is_MachBranch();
604 if (is_branch) {
605 MacroAssembler masm(&buf);
606 masm.bind(fakeL);
607 n->as_MachBranch()->save_label(&saveL, &save_bnum);
608 n->as_MachBranch()->label_set(&fakeL, 0);
609 }
610 n->emit(buf, this->regalloc());
611
612 // Emitting into the scratch buffer should not fail
613 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
614
615 if (is_branch) // Restore label.
616 n->as_MachBranch()->label_set(saveL, save_bnum);
617
618 // End scratch_emit_size section.
619 set_in_scratch_emit_size(false);
620
621 return buf.insts_size();
622 }
623
624
625 // ============================================================================
626 //------------------------------Compile standard-------------------------------
627 debug_only( int Compile::_debug_idx = 100000; )
628
629 // Compile a method. entry_bci is -1 for normal compilations and indicates
630 // the continuation bci for on stack replacement.
631
632
633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
634 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
635 : Phase(Compiler),
636 _compile_id(ci_env->compile_id()),
637 _save_argument_registers(false),
638 _subsume_loads(subsume_loads),
639 _do_escape_analysis(do_escape_analysis),
640 _eliminate_boxing(eliminate_boxing),
641 _method(target),
642 _entry_bci(osr_bci),
643 _stub_function(NULL),
644 _stub_name(NULL),
645 _stub_entry_point(NULL),
646 _max_node_limit(MaxNodeLimit),
647 _orig_pc_slot(0),
648 _orig_pc_slot_offset_in_bytes(0),
649 _inlining_progress(false),
650 _inlining_incrementally(false),
651 _has_reserved_stack_access(target->has_reserved_stack_access()),
652 #ifndef PRODUCT
653 _trace_opto_output(directive->TraceOptoOutputOption),
654 #endif
655 _has_method_handle_invokes(false),
656 _comp_arena(mtCompiler),
657 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
658 _env(ci_env),
659 _directive(directive),
660 _log(ci_env->log()),
661 _failure_reason(NULL),
662 _congraph(NULL),
663 #ifndef PRODUCT
664 _printer(IdealGraphPrinter::printer()),
665 #endif
666 _dead_node_list(comp_arena()),
667 _dead_node_count(0),
668 _node_arena(mtCompiler),
669 _old_arena(mtCompiler),
670 _mach_constant_base_node(NULL),
671 _Compile_types(mtCompiler),
672 _initial_gvn(NULL),
673 _for_igvn(NULL),
674 _warm_calls(NULL),
675 _late_inlines(comp_arena(), 2, 0, NULL),
676 _string_late_inlines(comp_arena(), 2, 0, NULL),
677 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
678 _late_inlines_pos(0),
679 _number_of_mh_late_inlines(0),
680 _print_inlining_stream(NULL),
681 _print_inlining_list(NULL),
682 _print_inlining_idx(0),
683 _print_inlining_output(NULL),
684 _replay_inline_data(NULL),
685 _java_calls(0),
686 _inner_loops(0),
687 _interpreter_frame_size(0),
688 _node_bundling_limit(0),
689 _node_bundling_base(NULL),
690 _code_buffer("Compile::Fill_buffer"),
691 _scratch_const_size(-1),
692 _in_scratch_emit_size(false)
693 #ifndef PRODUCT
694 , _in_dump_cnt(0)
695 #endif
696 {
697 C = this;
698 #ifndef PRODUCT
699 if (_printer != NULL) {
700 _printer->set_compile(this);
701 }
702 #endif
703 CompileWrapper cw(this);
704
705 if (CITimeVerbose) {
706 tty->print(" ");
707 target->holder()->name()->print();
708 tty->print(".");
709 target->print_short_name();
710 tty->print(" ");
711 }
712 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
713 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
714
715 #ifndef PRODUCT
716 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
717 if (!print_opto_assembly) {
718 bool print_assembly = directive->PrintAssemblyOption;
719 if (print_assembly && !Disassembler::can_decode()) {
720 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
721 print_opto_assembly = true;
722 }
723 }
724 set_print_assembly(print_opto_assembly);
725 set_parsed_irreducible_loop(false);
726
727 if (directive->ReplayInlineOption) {
728 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
729 }
730 #endif
731 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
732 set_print_intrinsics(directive->PrintIntrinsicsOption);
733 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
734
735 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
736 // Make sure the method being compiled gets its own MDO,
737 // so we can at least track the decompile_count().
738 // Need MDO to record RTM code generation state.
739 method()->ensure_method_data();
740 }
741
742 Init(::AliasLevel);
743
744
745 print_compile_messages();
746
747 _ilt = InlineTree::build_inline_tree_root();
748
749 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
750 assert(num_alias_types() >= AliasIdxRaw, "");
751
752 #define MINIMUM_NODE_HASH 1023
753 // Node list that Iterative GVN will start with
754 Unique_Node_List for_igvn(comp_arena());
755 set_for_igvn(&for_igvn);
756
757 // GVN that will be run immediately on new nodes
758 uint estimated_size = method()->code_size()*4+64;
759 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
760 PhaseGVN gvn(node_arena(), estimated_size);
761 set_initial_gvn(&gvn);
762
763 print_inlining_init();
764 { // Scope for timing the parser
765 TracePhase tp("parse", &timers[_t_parser]);
766
767 // Put top into the hash table ASAP.
768 initial_gvn()->transform_no_reclaim(top());
769
770 // Set up tf(), start(), and find a CallGenerator.
771 CallGenerator* cg = NULL;
772 if (is_osr_compilation()) {
773 const TypeTuple *domain = StartOSRNode::osr_domain();
774 const TypeTuple *range = TypeTuple::make_range(method()->signature());
775 init_tf(TypeFunc::make(domain, range));
776 StartNode* s = new StartOSRNode(root(), domain);
777 initial_gvn()->set_type_bottom(s);
778 init_start(s);
779 cg = CallGenerator::for_osr(method(), entry_bci());
780 } else {
781 // Normal case.
782 init_tf(TypeFunc::make(method()));
783 StartNode* s = new StartNode(root(), tf()->domain());
784 initial_gvn()->set_type_bottom(s);
785 init_start(s);
786 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
787 // With java.lang.ref.reference.get() we must go through the
788 // intrinsic - even when get() is the root
789 // method of the compile - so that, if necessary, the value in
790 // the referent field of the reference object gets recorded by
791 // the pre-barrier code.
792 cg = find_intrinsic(method(), false);
793 }
794 if (cg == NULL) {
795 float past_uses = method()->interpreter_invocation_count();
796 float expected_uses = past_uses;
797 cg = CallGenerator::for_inline(method(), expected_uses);
798 }
799 }
800 if (failing()) return;
801 if (cg == NULL) {
802 record_method_not_compilable("cannot parse method");
803 return;
804 }
805 JVMState* jvms = build_start_state(start(), tf());
806 if ((jvms = cg->generate(jvms)) == NULL) {
807 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
808 record_method_not_compilable("method parse failed");
809 }
810 return;
811 }
812 GraphKit kit(jvms);
813
814 if (!kit.stopped()) {
815 // Accept return values, and transfer control we know not where.
816 // This is done by a special, unique ReturnNode bound to root.
817 return_values(kit.jvms());
818 }
819
820 if (kit.has_exceptions()) {
821 // Any exceptions that escape from this call must be rethrown
822 // to whatever caller is dynamically above us on the stack.
823 // This is done by a special, unique RethrowNode bound to root.
824 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
825 }
826
827 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
828
829 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
830 inline_string_calls(true);
831 }
832
833 if (failing()) return;
834
835 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
836
837 // Remove clutter produced by parsing.
838 if (!failing()) {
839 ResourceMark rm;
840 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
841 }
842 }
843
844 // Note: Large methods are capped off in do_one_bytecode().
845 if (failing()) return;
846
847 // After parsing, node notes are no longer automagic.
848 // They must be propagated by register_new_node_with_optimizer(),
849 // clone(), or the like.
850 set_default_node_notes(NULL);
851
852 for (;;) {
853 int successes = Inline_Warm();
854 if (failing()) return;
855 if (successes == 0) break;
856 }
857
858 // Drain the list.
859 Finish_Warm();
860 #ifndef PRODUCT
861 if (_printer && _printer->should_print(1)) {
862 _printer->print_inlining();
863 }
864 #endif
865
866 if (failing()) return;
867 NOT_PRODUCT( verify_graph_edges(); )
868
869 // Now optimize
870 Optimize();
871 if (failing()) return;
872 NOT_PRODUCT( verify_graph_edges(); )
873
874 #ifndef PRODUCT
875 if (PrintIdeal) {
876 ttyLocker ttyl; // keep the following output all in one block
877 // This output goes directly to the tty, not the compiler log.
878 // To enable tools to match it up with the compilation activity,
879 // be sure to tag this tty output with the compile ID.
880 if (xtty != NULL) {
881 xtty->head("ideal compile_id='%d'%s", compile_id(),
882 is_osr_compilation() ? " compile_kind='osr'" :
883 "");
884 }
885 root()->dump(9999);
886 if (xtty != NULL) {
887 xtty->tail("ideal");
888 }
889 }
890 #endif
891
892 #ifdef ASSERT
893 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
894 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
895 #endif
896
897 // Dump compilation data to replay it.
898 if (directive->DumpReplayOption) {
899 env()->dump_replay_data(_compile_id);
900 }
901 if (directive->DumpInlineOption && (ilt() != NULL)) {
902 env()->dump_inline_data(_compile_id);
903 }
904
905 // Now that we know the size of all the monitors we can add a fixed slot
906 // for the original deopt pc.
907
908 _orig_pc_slot = fixed_slots();
909 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
910 set_fixed_slots(next_slot);
911
912 // Compute when to use implicit null checks. Used by matching trap based
913 // nodes and NullCheck optimization.
914 set_allowed_deopt_reasons();
915
916 // Now generate code
917 Code_Gen();
918 if (failing()) return;
919
920 // Check if we want to skip execution of all compiled code.
921 {
922 #ifndef PRODUCT
923 if (OptoNoExecute) {
924 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
925 return;
926 }
927 #endif
928 TracePhase tp("install_code", &timers[_t_registerMethod]);
929
930 if (is_osr_compilation()) {
931 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
932 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
933 } else {
934 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
935 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
936 }
937
938 env()->register_method(_method, _entry_bci,
939 &_code_offsets,
940 _orig_pc_slot_offset_in_bytes,
941 code_buffer(),
942 frame_size_in_words(), _oop_map_set,
943 &_handler_table, &_inc_table,
944 compiler,
945 has_unsafe_access(),
946 SharedRuntime::is_wide_vector(max_vector_size()),
947 rtm_state()
948 );
949
950 if (log() != NULL) // Print code cache state into compiler log
951 log()->code_cache_state();
952 }
953 }
954
955 //------------------------------Compile----------------------------------------
956 // Compile a runtime stub
957 Compile::Compile( ciEnv* ci_env,
958 TypeFunc_generator generator,
959 address stub_function,
960 const char *stub_name,
961 int is_fancy_jump,
962 bool pass_tls,
963 bool save_arg_registers,
964 bool return_pc,
965 DirectiveSet* directive)
966 : Phase(Compiler),
967 _compile_id(0),
968 _save_argument_registers(save_arg_registers),
969 _subsume_loads(true),
970 _do_escape_analysis(false),
971 _eliminate_boxing(false),
972 _method(NULL),
973 _entry_bci(InvocationEntryBci),
974 _stub_function(stub_function),
975 _stub_name(stub_name),
976 _stub_entry_point(NULL),
977 _max_node_limit(MaxNodeLimit),
978 _orig_pc_slot(0),
979 _orig_pc_slot_offset_in_bytes(0),
980 _inlining_progress(false),
981 _inlining_incrementally(false),
982 _has_reserved_stack_access(false),
983 #ifndef PRODUCT
984 _trace_opto_output(directive->TraceOptoOutputOption),
985 #endif
986 _has_method_handle_invokes(false),
987 _comp_arena(mtCompiler),
988 _env(ci_env),
989 _directive(directive),
990 _log(ci_env->log()),
991 _failure_reason(NULL),
992 _congraph(NULL),
993 #ifndef PRODUCT
994 _printer(NULL),
995 #endif
996 _dead_node_list(comp_arena()),
997 _dead_node_count(0),
998 _node_arena(mtCompiler),
999 _old_arena(mtCompiler),
1000 _mach_constant_base_node(NULL),
1001 _Compile_types(mtCompiler),
1002 _initial_gvn(NULL),
1003 _for_igvn(NULL),
1004 _warm_calls(NULL),
1005 _number_of_mh_late_inlines(0),
1006 _print_inlining_stream(NULL),
1007 _print_inlining_list(NULL),
1008 _print_inlining_idx(0),
1009 _print_inlining_output(NULL),
1010 _replay_inline_data(NULL),
1011 _java_calls(0),
1012 _inner_loops(0),
1013 _interpreter_frame_size(0),
1014 _node_bundling_limit(0),
1015 _node_bundling_base(NULL),
1016 _code_buffer("Compile::Fill_buffer"),
1017 #ifndef PRODUCT
1018 _in_dump_cnt(0),
1019 #endif
1020 _allowed_reasons(0) {
1021 C = this;
1022
1023 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1024 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1025
1026 #ifndef PRODUCT
1027 set_print_assembly(PrintFrameConverterAssembly);
1028 set_parsed_irreducible_loop(false);
1029 #endif
1030 set_has_irreducible_loop(false); // no loops
1031
1032 CompileWrapper cw(this);
1033 Init(/*AliasLevel=*/ 0);
1034 init_tf((*generator)());
1035
1036 {
1037 // The following is a dummy for the sake of GraphKit::gen_stub
1038 Unique_Node_List for_igvn(comp_arena());
1039 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1040 PhaseGVN gvn(Thread::current()->resource_area(),255);
1041 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1042 gvn.transform_no_reclaim(top());
1043
1044 GraphKit kit;
1045 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1046 }
1047
1048 NOT_PRODUCT( verify_graph_edges(); )
1049 Code_Gen();
1050 if (failing()) return;
1051
1052
1053 // Entry point will be accessed using compile->stub_entry_point();
1054 if (code_buffer() == NULL) {
1055 Matcher::soft_match_failure();
1056 } else {
1057 if (PrintAssembly && (WizardMode || Verbose))
1058 tty->print_cr("### Stub::%s", stub_name);
1059
1060 if (!failing()) {
1061 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1062
1063 // Make the NMethod
1064 // For now we mark the frame as never safe for profile stackwalking
1065 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1066 code_buffer(),
1067 CodeOffsets::frame_never_safe,
1068 // _code_offsets.value(CodeOffsets::Frame_Complete),
1069 frame_size_in_words(),
1070 _oop_map_set,
1071 save_arg_registers);
1072 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1073
1074 _stub_entry_point = rs->entry_point();
1075 }
1076 }
1077 }
1078
1079 //------------------------------Init-------------------------------------------
1080 // Prepare for a single compilation
1081 void Compile::Init(int aliaslevel) {
1082 _unique = 0;
1083 _regalloc = NULL;
1084
1085 _tf = NULL; // filled in later
1086 _top = NULL; // cached later
1087 _matcher = NULL; // filled in later
1088 _cfg = NULL; // filled in later
1089
1090 set_24_bit_selection_and_mode(Use24BitFP, false);
1091
1092 _node_note_array = NULL;
1093 _default_node_notes = NULL;
1094 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1095
1096 _immutable_memory = NULL; // filled in at first inquiry
1097
1098 // Globally visible Nodes
1099 // First set TOP to NULL to give safe behavior during creation of RootNode
1100 set_cached_top_node(NULL);
1101 set_root(new RootNode());
1102 // Now that you have a Root to point to, create the real TOP
1103 set_cached_top_node( new ConNode(Type::TOP) );
1104 set_recent_alloc(NULL, NULL);
1105
1106 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1107 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1108 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1109 env()->set_dependencies(new Dependencies(env()));
1110
1111 _fixed_slots = 0;
1112 set_has_split_ifs(false);
1113 set_has_loops(has_method() && method()->has_loops()); // first approximation
1114 set_has_stringbuilder(false);
1115 set_has_boxed_value(false);
1116 _trap_can_recompile = false; // no traps emitted yet
1117 _major_progress = true; // start out assuming good things will happen
1118 set_has_unsafe_access(false);
1119 set_max_vector_size(0);
1120 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1121 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1122 set_decompile_count(0);
1123
1124 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1125 set_num_loop_opts(LoopOptsCount);
1126 set_do_inlining(Inline);
1127 set_max_inline_size(MaxInlineSize);
1128 set_freq_inline_size(FreqInlineSize);
1129 set_do_scheduling(OptoScheduling);
1130 set_do_count_invocations(false);
1131 set_do_method_data_update(false);
1132
1133 set_do_vector_loop(false);
1134
1135 if (AllowVectorizeOnDemand) {
1136 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1137 set_do_vector_loop(true);
1138 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());})
1139 } else if (has_method() && method()->name() != 0 &&
1140 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1141 set_do_vector_loop(true);
1142 }
1143 }
1144 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1145 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());})
1146
1147 set_age_code(has_method() && method()->profile_aging());
1148 set_rtm_state(NoRTM); // No RTM lock eliding by default
1149 _max_node_limit = _directive->MaxNodeLimitOption;
1150
1151 #if INCLUDE_RTM_OPT
1152 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1153 int rtm_state = method()->method_data()->rtm_state();
1154 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1155 // Don't generate RTM lock eliding code.
1156 set_rtm_state(NoRTM);
1157 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1158 // Generate RTM lock eliding code without abort ratio calculation code.
1159 set_rtm_state(UseRTM);
1160 } else if (UseRTMDeopt) {
1161 // Generate RTM lock eliding code and include abort ratio calculation
1162 // code if UseRTMDeopt is on.
1163 set_rtm_state(ProfileRTM);
1164 }
1165 }
1166 #endif
1167 if (debug_info()->recording_non_safepoints()) {
1168 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1169 (comp_arena(), 8, 0, NULL));
1170 set_default_node_notes(Node_Notes::make(this));
1171 }
1172
1173 // // -- Initialize types before each compile --
1174 // // Update cached type information
1175 // if( _method && _method->constants() )
1176 // Type::update_loaded_types(_method, _method->constants());
1177
1178 // Init alias_type map.
1179 if (!_do_escape_analysis && aliaslevel == 3)
1180 aliaslevel = 2; // No unique types without escape analysis
1181 _AliasLevel = aliaslevel;
1182 const int grow_ats = 16;
1183 _max_alias_types = grow_ats;
1184 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1185 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1186 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1187 {
1188 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1189 }
1190 // Initialize the first few types.
1191 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1192 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1193 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1194 _num_alias_types = AliasIdxRaw+1;
1195 // Zero out the alias type cache.
1196 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1197 // A NULL adr_type hits in the cache right away. Preload the right answer.
1198 probe_alias_cache(NULL)->_index = AliasIdxTop;
1199
1200 _intrinsics = NULL;
1201 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1202 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1203 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1204 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1205 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1206 register_library_intrinsics();
1207 }
1208
1209 //---------------------------init_start----------------------------------------
1210 // Install the StartNode on this compile object.
1211 void Compile::init_start(StartNode* s) {
1212 if (failing())
1213 return; // already failing
1214 assert(s == start(), "");
1215 }
1216
1217 /**
1218 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1219 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1220 * the ideal graph.
1221 */
1222 StartNode* Compile::start() const {
1223 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1224 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1225 Node* start = root()->fast_out(i);
1226 if (start->is_Start()) {
1227 return start->as_Start();
1228 }
1229 }
1230 fatal("Did not find Start node!");
1231 return NULL;
1232 }
1233
1234 //-------------------------------immutable_memory-------------------------------------
1235 // Access immutable memory
1236 Node* Compile::immutable_memory() {
1237 if (_immutable_memory != NULL) {
1238 return _immutable_memory;
1239 }
1240 StartNode* s = start();
1241 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1242 Node *p = s->fast_out(i);
1243 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1244 _immutable_memory = p;
1245 return _immutable_memory;
1246 }
1247 }
1248 ShouldNotReachHere();
1249 return NULL;
1250 }
1251
1252 //----------------------set_cached_top_node------------------------------------
1253 // Install the cached top node, and make sure Node::is_top works correctly.
1254 void Compile::set_cached_top_node(Node* tn) {
1255 if (tn != NULL) verify_top(tn);
1256 Node* old_top = _top;
1257 _top = tn;
1258 // Calling Node::setup_is_top allows the nodes the chance to adjust
1259 // their _out arrays.
1260 if (_top != NULL) _top->setup_is_top();
1261 if (old_top != NULL) old_top->setup_is_top();
1262 assert(_top == NULL || top()->is_top(), "");
1263 }
1264
1265 #ifdef ASSERT
1266 uint Compile::count_live_nodes_by_graph_walk() {
1267 Unique_Node_List useful(comp_arena());
1268 // Get useful node list by walking the graph.
1269 identify_useful_nodes(useful);
1270 return useful.size();
1271 }
1272
1273 void Compile::print_missing_nodes() {
1274
1275 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1276 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1277 return;
1278 }
1279
1280 // This is an expensive function. It is executed only when the user
1281 // specifies VerifyIdealNodeCount option or otherwise knows the
1282 // additional work that needs to be done to identify reachable nodes
1283 // by walking the flow graph and find the missing ones using
1284 // _dead_node_list.
1285
1286 Unique_Node_List useful(comp_arena());
1287 // Get useful node list by walking the graph.
1288 identify_useful_nodes(useful);
1289
1290 uint l_nodes = C->live_nodes();
1291 uint l_nodes_by_walk = useful.size();
1292
1293 if (l_nodes != l_nodes_by_walk) {
1294 if (_log != NULL) {
1295 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1296 _log->stamp();
1297 _log->end_head();
1298 }
1299 VectorSet& useful_member_set = useful.member_set();
1300 int last_idx = l_nodes_by_walk;
1301 for (int i = 0; i < last_idx; i++) {
1302 if (useful_member_set.test(i)) {
1303 if (_dead_node_list.test(i)) {
1304 if (_log != NULL) {
1305 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1306 }
1307 if (PrintIdealNodeCount) {
1308 // Print the log message to tty
1309 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1310 useful.at(i)->dump();
1311 }
1312 }
1313 }
1314 else if (! _dead_node_list.test(i)) {
1315 if (_log != NULL) {
1316 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1317 }
1318 if (PrintIdealNodeCount) {
1319 // Print the log message to tty
1320 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1321 }
1322 }
1323 }
1324 if (_log != NULL) {
1325 _log->tail("mismatched_nodes");
1326 }
1327 }
1328 }
1329 void Compile::record_modified_node(Node* n) {
1330 if (_modified_nodes != NULL && !_inlining_incrementally &&
1331 n->outcnt() != 0 && !n->is_Con()) {
1332 _modified_nodes->push(n);
1333 }
1334 }
1335
1336 void Compile::remove_modified_node(Node* n) {
1337 if (_modified_nodes != NULL) {
1338 _modified_nodes->remove(n);
1339 }
1340 }
1341 #endif
1342
1343 #ifndef PRODUCT
1344 void Compile::verify_top(Node* tn) const {
1345 if (tn != NULL) {
1346 assert(tn->is_Con(), "top node must be a constant");
1347 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1348 assert(tn->in(0) != NULL, "must have live top node");
1349 }
1350 }
1351 #endif
1352
1353
1354 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1355
1356 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1357 guarantee(arr != NULL, "");
1358 int num_blocks = arr->length();
1359 if (grow_by < num_blocks) grow_by = num_blocks;
1360 int num_notes = grow_by * _node_notes_block_size;
1361 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1362 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1363 while (num_notes > 0) {
1364 arr->append(notes);
1365 notes += _node_notes_block_size;
1366 num_notes -= _node_notes_block_size;
1367 }
1368 assert(num_notes == 0, "exact multiple, please");
1369 }
1370
1371 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1372 if (source == NULL || dest == NULL) return false;
1373
1374 if (dest->is_Con())
1375 return false; // Do not push debug info onto constants.
1376
1377 #ifdef ASSERT
1378 // Leave a bread crumb trail pointing to the original node:
1379 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1380 dest->set_debug_orig(source);
1381 }
1382 #endif
1383
1384 if (node_note_array() == NULL)
1385 return false; // Not collecting any notes now.
1386
1387 // This is a copy onto a pre-existing node, which may already have notes.
1388 // If both nodes have notes, do not overwrite any pre-existing notes.
1389 Node_Notes* source_notes = node_notes_at(source->_idx);
1390 if (source_notes == NULL || source_notes->is_clear()) return false;
1391 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1392 if (dest_notes == NULL || dest_notes->is_clear()) {
1393 return set_node_notes_at(dest->_idx, source_notes);
1394 }
1395
1396 Node_Notes merged_notes = (*source_notes);
1397 // The order of operations here ensures that dest notes will win...
1398 merged_notes.update_from(dest_notes);
1399 return set_node_notes_at(dest->_idx, &merged_notes);
1400 }
1401
1402
1403 //--------------------------allow_range_check_smearing-------------------------
1404 // Gating condition for coalescing similar range checks.
1405 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1406 // single covering check that is at least as strong as any of them.
1407 // If the optimization succeeds, the simplified (strengthened) range check
1408 // will always succeed. If it fails, we will deopt, and then give up
1409 // on the optimization.
1410 bool Compile::allow_range_check_smearing() const {
1411 // If this method has already thrown a range-check,
1412 // assume it was because we already tried range smearing
1413 // and it failed.
1414 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1415 return !already_trapped;
1416 }
1417
1418
1419 //------------------------------flatten_alias_type-----------------------------
1420 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1421 int offset = tj->offset();
1422 TypePtr::PTR ptr = tj->ptr();
1423
1424 // Known instance (scalarizable allocation) alias only with itself.
1425 bool is_known_inst = tj->isa_oopptr() != NULL &&
1426 tj->is_oopptr()->is_known_instance();
1427
1428 // Process weird unsafe references.
1429 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1430 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1431 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1432 tj = TypeOopPtr::BOTTOM;
1433 ptr = tj->ptr();
1434 offset = tj->offset();
1435 }
1436
1437 // Array pointers need some flattening
1438 const TypeAryPtr *ta = tj->isa_aryptr();
1439 if (ta && ta->is_stable()) {
1440 // Erase stability property for alias analysis.
1441 tj = ta = ta->cast_to_stable(false);
1442 }
1443 if( ta && is_known_inst ) {
1444 if ( offset != Type::OffsetBot &&
1445 offset > arrayOopDesc::length_offset_in_bytes() ) {
1446 offset = Type::OffsetBot; // Flatten constant access into array body only
1447 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1448 }
1449 } else if( ta && _AliasLevel >= 2 ) {
1450 // For arrays indexed by constant indices, we flatten the alias
1451 // space to include all of the array body. Only the header, klass
1452 // and array length can be accessed un-aliased.
1453 if( offset != Type::OffsetBot ) {
1454 if( ta->const_oop() ) { // MethodData* or Method*
1455 offset = Type::OffsetBot; // Flatten constant access into array body
1456 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1457 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1458 // range is OK as-is.
1459 tj = ta = TypeAryPtr::RANGE;
1460 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1461 tj = TypeInstPtr::KLASS; // all klass loads look alike
1462 ta = TypeAryPtr::RANGE; // generic ignored junk
1463 ptr = TypePtr::BotPTR;
1464 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1465 tj = TypeInstPtr::MARK;
1466 ta = TypeAryPtr::RANGE; // generic ignored junk
1467 ptr = TypePtr::BotPTR;
1468 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1469 ta = tj->isa_aryptr();
1470 } else { // Random constant offset into array body
1471 offset = Type::OffsetBot; // Flatten constant access into array body
1472 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1473 }
1474 }
1475 // Arrays of fixed size alias with arrays of unknown size.
1476 if (ta->size() != TypeInt::POS) {
1477 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1478 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1479 }
1480 // Arrays of known objects become arrays of unknown objects.
1481 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1482 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1483 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1484 }
1485 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1486 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1487 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1488 }
1489 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1490 // cannot be distinguished by bytecode alone.
1491 if (ta->elem() == TypeInt::BOOL) {
1492 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1493 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1494 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1495 }
1496 // During the 2nd round of IterGVN, NotNull castings are removed.
1497 // Make sure the Bottom and NotNull variants alias the same.
1498 // Also, make sure exact and non-exact variants alias the same.
1499 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1500 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1501 }
1502 }
1503
1504 // Oop pointers need some flattening
1505 const TypeInstPtr *to = tj->isa_instptr();
1506 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1507 ciInstanceKlass *k = to->klass()->as_instance_klass();
1508 if( ptr == TypePtr::Constant ) {
1509 if (to->klass() != ciEnv::current()->Class_klass() ||
1510 offset < k->size_helper() * wordSize) {
1511 // No constant oop pointers (such as Strings); they alias with
1512 // unknown strings.
1513 assert(!is_known_inst, "not scalarizable allocation");
1514 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1515 }
1516 } else if( is_known_inst ) {
1517 tj = to; // Keep NotNull and klass_is_exact for instance type
1518 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1519 // During the 2nd round of IterGVN, NotNull castings are removed.
1520 // Make sure the Bottom and NotNull variants alias the same.
1521 // Also, make sure exact and non-exact variants alias the same.
1522 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1523 }
1524 if (to->speculative() != NULL) {
1525 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1526 }
1527 // Canonicalize the holder of this field
1528 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1529 // First handle header references such as a LoadKlassNode, even if the
1530 // object's klass is unloaded at compile time (4965979).
1531 if (!is_known_inst) { // Do it only for non-instance types
1532 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1533 }
1534 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1535 to = tj->is_instptr();
1536 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1537 // Static fields are in the space above the normal instance
1538 // fields in the java.lang.Class instance.
1539 if (to->klass() != ciEnv::current()->Class_klass()) {
1540 to = NULL;
1541 tj = TypeOopPtr::BOTTOM;
1542 offset = tj->offset();
1543 }
1544 } else {
1545 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1546 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1547 if( is_known_inst ) {
1548 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1549 } else {
1550 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1551 }
1552 }
1553 }
1554 }
1555
1556 // Klass pointers to object array klasses need some flattening
1557 const TypeKlassPtr *tk = tj->isa_klassptr();
1558 if( tk ) {
1559 // If we are referencing a field within a Klass, we need
1560 // to assume the worst case of an Object. Both exact and
1561 // inexact types must flatten to the same alias class so
1562 // use NotNull as the PTR.
1563 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1564
1565 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1566 TypeKlassPtr::OBJECT->klass(),
1567 offset);
1568 }
1569
1570 ciKlass* klass = tk->klass();
1571 if( klass->is_obj_array_klass() ) {
1572 ciKlass* k = TypeAryPtr::OOPS->klass();
1573 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1574 k = TypeInstPtr::BOTTOM->klass();
1575 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1576 }
1577
1578 // Check for precise loads from the primary supertype array and force them
1579 // to the supertype cache alias index. Check for generic array loads from
1580 // the primary supertype array and also force them to the supertype cache
1581 // alias index. Since the same load can reach both, we need to merge
1582 // these 2 disparate memories into the same alias class. Since the
1583 // primary supertype array is read-only, there's no chance of confusion
1584 // where we bypass an array load and an array store.
1585 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1586 if (offset == Type::OffsetBot ||
1587 (offset >= primary_supers_offset &&
1588 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1589 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1590 offset = in_bytes(Klass::secondary_super_cache_offset());
1591 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1592 }
1593 }
1594
1595 // Flatten all Raw pointers together.
1596 if (tj->base() == Type::RawPtr)
1597 tj = TypeRawPtr::BOTTOM;
1598
1599 if (tj->base() == Type::AnyPtr)
1600 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1601
1602 // Flatten all to bottom for now
1603 switch( _AliasLevel ) {
1604 case 0:
1605 tj = TypePtr::BOTTOM;
1606 break;
1607 case 1: // Flatten to: oop, static, field or array
1608 switch (tj->base()) {
1609 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1610 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1611 case Type::AryPtr: // do not distinguish arrays at all
1612 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1613 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1614 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1615 default: ShouldNotReachHere();
1616 }
1617 break;
1618 case 2: // No collapsing at level 2; keep all splits
1619 case 3: // No collapsing at level 3; keep all splits
1620 break;
1621 default:
1622 Unimplemented();
1623 }
1624
1625 offset = tj->offset();
1626 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1627
1628 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1629 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1630 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1631 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1632 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1633 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1634 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1635 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1636 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1637 assert( tj->ptr() != TypePtr::TopPTR &&
1638 tj->ptr() != TypePtr::AnyNull &&
1639 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1640 // assert( tj->ptr() != TypePtr::Constant ||
1641 // tj->base() == Type::RawPtr ||
1642 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1643
1644 return tj;
1645 }
1646
1647 void Compile::AliasType::Init(int i, const TypePtr* at) {
1648 _index = i;
1649 _adr_type = at;
1650 _field = NULL;
1651 _element = NULL;
1652 _is_rewritable = true; // default
1653 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1654 if (atoop != NULL && atoop->is_known_instance()) {
1655 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1656 _general_index = Compile::current()->get_alias_index(gt);
1657 } else {
1658 _general_index = 0;
1659 }
1660 }
1661
1662 BasicType Compile::AliasType::basic_type() const {
1663 if (element() != NULL) {
1664 const Type* element = adr_type()->is_aryptr()->elem();
1665 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1666 } if (field() != NULL) {
1667 return field()->layout_type();
1668 } else {
1669 return T_ILLEGAL; // unknown
1670 }
1671 }
1672
1673 //---------------------------------print_on------------------------------------
1674 #ifndef PRODUCT
1675 void Compile::AliasType::print_on(outputStream* st) {
1676 if (index() < 10)
1677 st->print("@ <%d> ", index());
1678 else st->print("@ <%d>", index());
1679 st->print(is_rewritable() ? " " : " RO");
1680 int offset = adr_type()->offset();
1681 if (offset == Type::OffsetBot)
1682 st->print(" +any");
1683 else st->print(" +%-3d", offset);
1684 st->print(" in ");
1685 adr_type()->dump_on(st);
1686 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1687 if (field() != NULL && tjp) {
1688 if (tjp->klass() != field()->holder() ||
1689 tjp->offset() != field()->offset_in_bytes()) {
1690 st->print(" != ");
1691 field()->print();
1692 st->print(" ***");
1693 }
1694 }
1695 }
1696
1697 void print_alias_types() {
1698 Compile* C = Compile::current();
1699 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1700 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1701 C->alias_type(idx)->print_on(tty);
1702 tty->cr();
1703 }
1704 }
1705 #endif
1706
1707
1708 //----------------------------probe_alias_cache--------------------------------
1709 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1710 intptr_t key = (intptr_t) adr_type;
1711 key ^= key >> logAliasCacheSize;
1712 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1713 }
1714
1715
1716 //-----------------------------grow_alias_types--------------------------------
1717 void Compile::grow_alias_types() {
1718 const int old_ats = _max_alias_types; // how many before?
1719 const int new_ats = old_ats; // how many more?
1720 const int grow_ats = old_ats+new_ats; // how many now?
1721 _max_alias_types = grow_ats;
1722 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1723 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1724 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1725 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1726 }
1727
1728
1729 //--------------------------------find_alias_type------------------------------
1730 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1731 if (_AliasLevel == 0)
1732 return alias_type(AliasIdxBot);
1733
1734 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1735 if (ace->_adr_type == adr_type) {
1736 return alias_type(ace->_index);
1737 }
1738
1739 // Handle special cases.
1740 if (adr_type == NULL) return alias_type(AliasIdxTop);
1741 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1742
1743 // Do it the slow way.
1744 const TypePtr* flat = flatten_alias_type(adr_type);
1745
1746 #ifdef ASSERT
1747 {
1748 ResourceMark rm;
1749 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1750 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1751 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1752 Type::str(adr_type));
1753 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1754 const TypeOopPtr* foop = flat->is_oopptr();
1755 // Scalarizable allocations have exact klass always.
1756 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1757 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1758 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1759 Type::str(foop), Type::str(xoop));
1760 }
1761 }
1762 #endif
1763
1764 int idx = AliasIdxTop;
1765 for (int i = 0; i < num_alias_types(); i++) {
1766 if (alias_type(i)->adr_type() == flat) {
1767 idx = i;
1768 break;
1769 }
1770 }
1771
1772 if (idx == AliasIdxTop) {
1773 if (no_create) return NULL;
1774 // Grow the array if necessary.
1775 if (_num_alias_types == _max_alias_types) grow_alias_types();
1776 // Add a new alias type.
1777 idx = _num_alias_types++;
1778 _alias_types[idx]->Init(idx, flat);
1779 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1780 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1781 if (flat->isa_instptr()) {
1782 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1783 && flat->is_instptr()->klass() == env()->Class_klass())
1784 alias_type(idx)->set_rewritable(false);
1785 }
1786 if (flat->isa_aryptr()) {
1787 #ifdef ASSERT
1788 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1789 // (T_BYTE has the weakest alignment and size restrictions...)
1790 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1791 #endif
1792 if (flat->offset() == TypePtr::OffsetBot) {
1793 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1794 }
1795 }
1796 if (flat->isa_klassptr()) {
1797 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1798 alias_type(idx)->set_rewritable(false);
1799 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1800 alias_type(idx)->set_rewritable(false);
1801 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1802 alias_type(idx)->set_rewritable(false);
1803 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1804 alias_type(idx)->set_rewritable(false);
1805 }
1806 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1807 // but the base pointer type is not distinctive enough to identify
1808 // references into JavaThread.)
1809
1810 // Check for final fields.
1811 const TypeInstPtr* tinst = flat->isa_instptr();
1812 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1813 ciField* field;
1814 if (tinst->const_oop() != NULL &&
1815 tinst->klass() == ciEnv::current()->Class_klass() &&
1816 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1817 // static field
1818 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1819 field = k->get_field_by_offset(tinst->offset(), true);
1820 } else {
1821 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1822 field = k->get_field_by_offset(tinst->offset(), false);
1823 }
1824 assert(field == NULL ||
1825 original_field == NULL ||
1826 (field->holder() == original_field->holder() &&
1827 field->offset() == original_field->offset() &&
1828 field->is_static() == original_field->is_static()), "wrong field?");
1829 // Set field() and is_rewritable() attributes.
1830 if (field != NULL) alias_type(idx)->set_field(field);
1831 }
1832 }
1833
1834 // Fill the cache for next time.
1835 ace->_adr_type = adr_type;
1836 ace->_index = idx;
1837 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1838
1839 // Might as well try to fill the cache for the flattened version, too.
1840 AliasCacheEntry* face = probe_alias_cache(flat);
1841 if (face->_adr_type == NULL) {
1842 face->_adr_type = flat;
1843 face->_index = idx;
1844 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1845 }
1846
1847 return alias_type(idx);
1848 }
1849
1850
1851 Compile::AliasType* Compile::alias_type(ciField* field) {
1852 const TypeOopPtr* t;
1853 if (field->is_static())
1854 t = TypeInstPtr::make(field->holder()->java_mirror());
1855 else
1856 t = TypeOopPtr::make_from_klass_raw(field->holder());
1857 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1858 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1859 return atp;
1860 }
1861
1862
1863 //------------------------------have_alias_type--------------------------------
1864 bool Compile::have_alias_type(const TypePtr* adr_type) {
1865 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1866 if (ace->_adr_type == adr_type) {
1867 return true;
1868 }
1869
1870 // Handle special cases.
1871 if (adr_type == NULL) return true;
1872 if (adr_type == TypePtr::BOTTOM) return true;
1873
1874 return find_alias_type(adr_type, true, NULL) != NULL;
1875 }
1876
1877 //-----------------------------must_alias--------------------------------------
1878 // True if all values of the given address type are in the given alias category.
1879 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1880 if (alias_idx == AliasIdxBot) return true; // the universal category
1881 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1882 if (alias_idx == AliasIdxTop) return false; // the empty category
1883 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1884
1885 // the only remaining possible overlap is identity
1886 int adr_idx = get_alias_index(adr_type);
1887 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1888 assert(adr_idx == alias_idx ||
1889 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1890 && adr_type != TypeOopPtr::BOTTOM),
1891 "should not be testing for overlap with an unsafe pointer");
1892 return adr_idx == alias_idx;
1893 }
1894
1895 //------------------------------can_alias--------------------------------------
1896 // True if any values of the given address type are in the given alias category.
1897 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1898 if (alias_idx == AliasIdxTop) return false; // the empty category
1899 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1900 if (alias_idx == AliasIdxBot) return true; // the universal category
1901 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1902
1903 // the only remaining possible overlap is identity
1904 int adr_idx = get_alias_index(adr_type);
1905 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1906 return adr_idx == alias_idx;
1907 }
1908
1909
1910
1911 //---------------------------pop_warm_call-------------------------------------
1912 WarmCallInfo* Compile::pop_warm_call() {
1913 WarmCallInfo* wci = _warm_calls;
1914 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1915 return wci;
1916 }
1917
1918 //----------------------------Inline_Warm--------------------------------------
1919 int Compile::Inline_Warm() {
1920 // If there is room, try to inline some more warm call sites.
1921 // %%% Do a graph index compaction pass when we think we're out of space?
1922 if (!InlineWarmCalls) return 0;
1923
1924 int calls_made_hot = 0;
1925 int room_to_grow = NodeCountInliningCutoff - unique();
1926 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1927 int amount_grown = 0;
1928 WarmCallInfo* call;
1929 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1930 int est_size = (int)call->size();
1931 if (est_size > (room_to_grow - amount_grown)) {
1932 // This one won't fit anyway. Get rid of it.
1933 call->make_cold();
1934 continue;
1935 }
1936 call->make_hot();
1937 calls_made_hot++;
1938 amount_grown += est_size;
1939 amount_to_grow -= est_size;
1940 }
1941
1942 if (calls_made_hot > 0) set_major_progress();
1943 return calls_made_hot;
1944 }
1945
1946
1947 //----------------------------Finish_Warm--------------------------------------
1948 void Compile::Finish_Warm() {
1949 if (!InlineWarmCalls) return;
1950 if (failing()) return;
1951 if (warm_calls() == NULL) return;
1952
1953 // Clean up loose ends, if we are out of space for inlining.
1954 WarmCallInfo* call;
1955 while ((call = pop_warm_call()) != NULL) {
1956 call->make_cold();
1957 }
1958 }
1959
1960 //---------------------cleanup_loop_predicates-----------------------
1961 // Remove the opaque nodes that protect the predicates so that all unused
1962 // checks and uncommon_traps will be eliminated from the ideal graph
1963 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1964 if (predicate_count()==0) return;
1965 for (int i = predicate_count(); i > 0; i--) {
1966 Node * n = predicate_opaque1_node(i-1);
1967 assert(n->Opcode() == Op_Opaque1, "must be");
1968 igvn.replace_node(n, n->in(1));
1969 }
1970 assert(predicate_count()==0, "should be clean!");
1971 }
1972
1973 void Compile::add_range_check_cast(Node* n) {
1974 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1975 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1976 _range_check_casts->append(n);
1977 }
1978
1979 // Remove all range check dependent CastIINodes.
1980 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1981 for (int i = range_check_cast_count(); i > 0; i--) {
1982 Node* cast = range_check_cast_node(i-1);
1983 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1984 igvn.replace_node(cast, cast->in(1));
1985 }
1986 assert(range_check_cast_count() == 0, "should be empty");
1987 }
1988
1989 void Compile::add_opaque4_node(Node* n) {
1990 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1991 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1992 _opaque4_nodes->append(n);
1993 }
1994
1995 // Remove all Opaque4 nodes.
1996 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1997 for (int i = opaque4_count(); i > 0; i--) {
1998 Node* opaq = opaque4_node(i-1);
1999 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2000 igvn.replace_node(opaq, opaq->in(2));
2001 }
2002 assert(opaque4_count() == 0, "should be empty");
2003 }
2004
2005 // StringOpts and late inlining of string methods
2006 void Compile::inline_string_calls(bool parse_time) {
2007 {
2008 // remove useless nodes to make the usage analysis simpler
2009 ResourceMark rm;
2010 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2011 }
2012
2013 {
2014 ResourceMark rm;
2015 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2016 PhaseStringOpts pso(initial_gvn(), for_igvn());
2017 print_method(PHASE_AFTER_STRINGOPTS, 3);
2018 }
2019
2020 // now inline anything that we skipped the first time around
2021 if (!parse_time) {
2022 _late_inlines_pos = _late_inlines.length();
2023 }
2024
2025 while (_string_late_inlines.length() > 0) {
2026 CallGenerator* cg = _string_late_inlines.pop();
2027 cg->do_late_inline();
2028 if (failing()) return;
2029 }
2030 _string_late_inlines.trunc_to(0);
2031 }
2032
2033 // Late inlining of boxing methods
2034 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2035 if (_boxing_late_inlines.length() > 0) {
2036 assert(has_boxed_value(), "inconsistent");
2037
2038 PhaseGVN* gvn = initial_gvn();
2039 set_inlining_incrementally(true);
2040
2041 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2042 for_igvn()->clear();
2043 gvn->replace_with(&igvn);
2044
2045 _late_inlines_pos = _late_inlines.length();
2046
2047 while (_boxing_late_inlines.length() > 0) {
2048 CallGenerator* cg = _boxing_late_inlines.pop();
2049 cg->do_late_inline();
2050 if (failing()) return;
2051 }
2052 _boxing_late_inlines.trunc_to(0);
2053
2054 {
2055 ResourceMark rm;
2056 PhaseRemoveUseless pru(gvn, for_igvn());
2057 }
2058
2059 igvn = PhaseIterGVN(gvn);
2060 igvn.optimize();
2061
2062 set_inlining_progress(false);
2063 set_inlining_incrementally(false);
2064 }
2065 }
2066
2067 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2068 assert(IncrementalInline, "incremental inlining should be on");
2069 PhaseGVN* gvn = initial_gvn();
2070
2071 set_inlining_progress(false);
2072 for_igvn()->clear();
2073 gvn->replace_with(&igvn);
2074
2075 {
2076 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2077 int i = 0;
2078 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2079 CallGenerator* cg = _late_inlines.at(i);
2080 _late_inlines_pos = i+1;
2081 cg->do_late_inline();
2082 if (failing()) return;
2083 }
2084 int j = 0;
2085 for (; i < _late_inlines.length(); i++, j++) {
2086 _late_inlines.at_put(j, _late_inlines.at(i));
2087 }
2088 _late_inlines.trunc_to(j);
2089 }
2090
2091 {
2092 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2093 ResourceMark rm;
2094 PhaseRemoveUseless pru(gvn, for_igvn());
2095 }
2096
2097 {
2098 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2099 igvn = PhaseIterGVN(gvn);
2100 }
2101 }
2102
2103 // Perform incremental inlining until bound on number of live nodes is reached
2104 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2105 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2106
2107 PhaseGVN* gvn = initial_gvn();
2108
2109 set_inlining_incrementally(true);
2110 set_inlining_progress(true);
2111 uint low_live_nodes = 0;
2112
2113 while(inlining_progress() && _late_inlines.length() > 0) {
2114
2115 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2116 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2117 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2118 // PhaseIdealLoop is expensive so we only try it once we are
2119 // out of live nodes and we only try it again if the previous
2120 // helped got the number of nodes down significantly
2121 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2122 if (failing()) return;
2123 low_live_nodes = live_nodes();
2124 _major_progress = true;
2125 }
2126
2127 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2128 break;
2129 }
2130 }
2131
2132 inline_incrementally_one(igvn);
2133
2134 if (failing()) return;
2135
2136 {
2137 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2138 igvn.optimize();
2139 }
2140
2141 if (failing()) return;
2142 }
2143
2144 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2145
2146 if (_string_late_inlines.length() > 0) {
2147 assert(has_stringbuilder(), "inconsistent");
2148 for_igvn()->clear();
2149 initial_gvn()->replace_with(&igvn);
2150
2151 inline_string_calls(false);
2152
2153 if (failing()) return;
2154
2155 {
2156 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2157 ResourceMark rm;
2158 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2159 }
2160
2161 {
2162 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2163 igvn = PhaseIterGVN(gvn);
2164 igvn.optimize();
2165 }
2166 }
2167
2168 set_inlining_incrementally(false);
2169 }
2170
2171
2172 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2173 if(loop_opts_cnt > 0) {
2174 debug_only( int cnt = 0; );
2175 while(major_progress() && (loop_opts_cnt > 0)) {
2176 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2177 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2178 PhaseIdealLoop ideal_loop(igvn, mode);
2179 loop_opts_cnt--;
2180 if (failing()) return false;
2181 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2182 }
2183 }
2184 return true;
2185 }
2186
2187 //------------------------------Optimize---------------------------------------
2188 // Given a graph, optimize it.
2189 void Compile::Optimize() {
2190 TracePhase tp("optimizer", &timers[_t_optimizer]);
2191
2192 #ifndef PRODUCT
2193 if (_directive->BreakAtCompileOption) {
2194 BREAKPOINT;
2195 }
2196
2197 #endif
2198
2199 #ifdef ASSERT
2200 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2201 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2202 #endif
2203
2204 ResourceMark rm;
2205 int loop_opts_cnt;
2206
2207 print_inlining_reinit();
2208
2209 NOT_PRODUCT( verify_graph_edges(); )
2210
2211 print_method(PHASE_AFTER_PARSING);
2212
2213 {
2214 // Iterative Global Value Numbering, including ideal transforms
2215 // Initialize IterGVN with types and values from parse-time GVN
2216 PhaseIterGVN igvn(initial_gvn());
2217 #ifdef ASSERT
2218 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2219 #endif
2220 {
2221 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2222 igvn.optimize();
2223 }
2224
2225 if (failing()) return;
2226
2227 print_method(PHASE_ITER_GVN1, 2);
2228
2229 inline_incrementally(igvn);
2230
2231 print_method(PHASE_INCREMENTAL_INLINE, 2);
2232
2233 if (failing()) return;
2234
2235 if (eliminate_boxing()) {
2236 // Inline valueOf() methods now.
2237 inline_boxing_calls(igvn);
2238
2239 if (AlwaysIncrementalInline) {
2240 inline_incrementally(igvn);
2241 }
2242
2243 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2244
2245 if (failing()) return;
2246 }
2247
2248 // Remove the speculative part of types and clean up the graph from
2249 // the extra CastPP nodes whose only purpose is to carry them. Do
2250 // that early so that optimizations are not disrupted by the extra
2251 // CastPP nodes.
2252 remove_speculative_types(igvn);
2253
2254 // No more new expensive nodes will be added to the list from here
2255 // so keep only the actual candidates for optimizations.
2256 cleanup_expensive_nodes(igvn);
2257
2258 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2259 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2260 initial_gvn()->replace_with(&igvn);
2261 for_igvn()->clear();
2262 Unique_Node_List new_worklist(C->comp_arena());
2263 {
2264 ResourceMark rm;
2265 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2266 }
2267 set_for_igvn(&new_worklist);
2268 igvn = PhaseIterGVN(initial_gvn());
2269 igvn.optimize();
2270 }
2271
2272 // Perform escape analysis
2273 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2274 if (has_loops()) {
2275 // Cleanup graph (remove dead nodes).
2276 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2277 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2278 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2279 if (failing()) return;
2280 }
2281 ConnectionGraph::do_analysis(this, &igvn);
2282
2283 if (failing()) return;
2284
2285 // Optimize out fields loads from scalar replaceable allocations.
2286 igvn.optimize();
2287 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2288
2289 if (failing()) return;
2290
2291 if (congraph() != NULL && macro_count() > 0) {
2292 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2293 PhaseMacroExpand mexp(igvn);
2294 mexp.eliminate_macro_nodes();
2295 igvn.set_delay_transform(false);
2296
2297 igvn.optimize();
2298 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2299
2300 if (failing()) return;
2301 }
2302 }
2303
2304 // Loop transforms on the ideal graph. Range Check Elimination,
2305 // peeling, unrolling, etc.
2306
2307 // Set loop opts counter
2308 loop_opts_cnt = num_loop_opts();
2309 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2310 {
2311 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2312 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2313 loop_opts_cnt--;
2314 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2315 if (failing()) return;
2316 }
2317 // Loop opts pass if partial peeling occurred in previous pass
2318 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2319 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2320 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2321 loop_opts_cnt--;
2322 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2323 if (failing()) return;
2324 }
2325 // Loop opts pass for loop-unrolling before CCP
2326 if(major_progress() && (loop_opts_cnt > 0)) {
2327 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2328 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2329 loop_opts_cnt--;
2330 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2331 }
2332 if (!failing()) {
2333 // Verify that last round of loop opts produced a valid graph
2334 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2335 PhaseIdealLoop::verify(igvn);
2336 }
2337 }
2338 if (failing()) return;
2339
2340 // Conditional Constant Propagation;
2341 PhaseCCP ccp( &igvn );
2342 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2343 {
2344 TracePhase tp("ccp", &timers[_t_ccp]);
2345 ccp.do_transform();
2346 }
2347 print_method(PHASE_CPP1, 2);
2348
2349 assert( true, "Break here to ccp.dump_old2new_map()");
2350
2351 // Iterative Global Value Numbering, including ideal transforms
2352 {
2353 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2354 igvn = ccp;
2355 igvn.optimize();
2356 }
2357
2358 print_method(PHASE_ITER_GVN2, 2);
2359
2360 if (failing()) return;
2361
2362 // Loop transforms on the ideal graph. Range Check Elimination,
2363 // peeling, unrolling, etc.
2364 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2365 return;
2366 }
2367
2368 #if INCLUDE_ZGC
2369 if (UseZGC) {
2370 ZBarrierSetC2::find_dominating_barriers(igvn);
2371 }
2372 #endif
2373
2374 if (failing()) return;
2375
2376 // Ensure that major progress is now clear
2377 C->clear_major_progress();
2378
2379 {
2380 // Verify that all previous optimizations produced a valid graph
2381 // at least to this point, even if no loop optimizations were done.
2382 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2383 PhaseIdealLoop::verify(igvn);
2384 }
2385
2386 if (range_check_cast_count() > 0) {
2387 // No more loop optimizations. Remove all range check dependent CastIINodes.
2388 C->remove_range_check_casts(igvn);
2389 igvn.optimize();
2390 }
2391
2392 #ifdef ASSERT
2393 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2394 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand);
2395 #endif
2396
2397 {
2398 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2399 PhaseMacroExpand mex(igvn);
2400 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2401 if (mex.expand_macro_nodes()) {
2402 assert(failing(), "must bail out w/ explicit message");
2403 return;
2404 }
2405 }
2406
2407 if (opaque4_count() > 0) {
2408 C->remove_opaque4_nodes(igvn);
2409 igvn.optimize();
2410 }
2411
2412 DEBUG_ONLY( _modified_nodes = NULL; )
2413 } // (End scope of igvn; run destructor if necessary for asserts.)
2414
2415 process_print_inlining();
2416 // A method with only infinite loops has no edges entering loops from root
2417 {
2418 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2419 if (final_graph_reshaping()) {
2420 assert(failing(), "must bail out w/ explicit message");
2421 return;
2422 }
2423 }
2424
2425 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2426 }
2427
2428
2429 //------------------------------Code_Gen---------------------------------------
2430 // Given a graph, generate code for it
2431 void Compile::Code_Gen() {
2432 if (failing()) {
2433 return;
2434 }
2435
2436 // Perform instruction selection. You might think we could reclaim Matcher
2437 // memory PDQ, but actually the Matcher is used in generating spill code.
2438 // Internals of the Matcher (including some VectorSets) must remain live
2439 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2440 // set a bit in reclaimed memory.
2441
2442 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2443 // nodes. Mapping is only valid at the root of each matched subtree.
2444 NOT_PRODUCT( verify_graph_edges(); )
2445
2446 Matcher matcher;
2447 _matcher = &matcher;
2448 {
2449 TracePhase tp("matcher", &timers[_t_matcher]);
2450 matcher.match();
2451 }
2452 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2453 // nodes. Mapping is only valid at the root of each matched subtree.
2454 NOT_PRODUCT( verify_graph_edges(); )
2455
2456 // If you have too many nodes, or if matching has failed, bail out
2457 check_node_count(0, "out of nodes matching instructions");
2458 if (failing()) {
2459 return;
2460 }
2461
2462 print_method(PHASE_MATCHING, 2);
2463
2464 // Build a proper-looking CFG
2465 PhaseCFG cfg(node_arena(), root(), matcher);
2466 _cfg = &cfg;
2467 {
2468 TracePhase tp("scheduler", &timers[_t_scheduler]);
2469 bool success = cfg.do_global_code_motion();
2470 if (!success) {
2471 return;
2472 }
2473
2474 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2475 NOT_PRODUCT( verify_graph_edges(); )
2476 debug_only( cfg.verify(); )
2477 }
2478
2479 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2480 _regalloc = ®alloc;
2481 {
2482 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2483 // Perform register allocation. After Chaitin, use-def chains are
2484 // no longer accurate (at spill code) and so must be ignored.
2485 // Node->LRG->reg mappings are still accurate.
2486 _regalloc->Register_Allocate();
2487
2488 // Bail out if the allocator builds too many nodes
2489 if (failing()) {
2490 return;
2491 }
2492 }
2493
2494 // Prior to register allocation we kept empty basic blocks in case the
2495 // the allocator needed a place to spill. After register allocation we
2496 // are not adding any new instructions. If any basic block is empty, we
2497 // can now safely remove it.
2498 {
2499 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2500 cfg.remove_empty_blocks();
2501 if (do_freq_based_layout()) {
2502 PhaseBlockLayout layout(cfg);
2503 } else {
2504 cfg.set_loop_alignment();
2505 }
2506 cfg.fixup_flow();
2507 }
2508
2509 // Apply peephole optimizations
2510 if( OptoPeephole ) {
2511 TracePhase tp("peephole", &timers[_t_peephole]);
2512 PhasePeephole peep( _regalloc, cfg);
2513 peep.do_transform();
2514 }
2515
2516 // Do late expand if CPU requires this.
2517 if (Matcher::require_postalloc_expand) {
2518 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2519 cfg.postalloc_expand(_regalloc);
2520 }
2521
2522 // Convert Nodes to instruction bits in a buffer
2523 {
2524 TraceTime tp("output", &timers[_t_output], CITime);
2525 Output();
2526 }
2527
2528 print_method(PHASE_FINAL_CODE);
2529
2530 // He's dead, Jim.
2531 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2532 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2533 }
2534
2535
2536 //------------------------------dump_asm---------------------------------------
2537 // Dump formatted assembly
2538 #ifndef PRODUCT
2539 void Compile::dump_asm(int *pcs, uint pc_limit) {
2540 bool cut_short = false;
2541 tty->print_cr("#");
2542 tty->print("# "); _tf->dump(); tty->cr();
2543 tty->print_cr("#");
2544
2545 // For all blocks
2546 int pc = 0x0; // Program counter
2547 char starts_bundle = ' ';
2548 _regalloc->dump_frame();
2549
2550 Node *n = NULL;
2551 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2552 if (VMThread::should_terminate()) {
2553 cut_short = true;
2554 break;
2555 }
2556 Block* block = _cfg->get_block(i);
2557 if (block->is_connector() && !Verbose) {
2558 continue;
2559 }
2560 n = block->head();
2561 if (pcs && n->_idx < pc_limit) {
2562 tty->print("%3.3x ", pcs[n->_idx]);
2563 } else {
2564 tty->print(" ");
2565 }
2566 block->dump_head(_cfg);
2567 if (block->is_connector()) {
2568 tty->print_cr(" # Empty connector block");
2569 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2570 tty->print_cr(" # Block is sole successor of call");
2571 }
2572
2573 // For all instructions
2574 Node *delay = NULL;
2575 for (uint j = 0; j < block->number_of_nodes(); j++) {
2576 if (VMThread::should_terminate()) {
2577 cut_short = true;
2578 break;
2579 }
2580 n = block->get_node(j);
2581 if (valid_bundle_info(n)) {
2582 Bundle* bundle = node_bundling(n);
2583 if (bundle->used_in_unconditional_delay()) {
2584 delay = n;
2585 continue;
2586 }
2587 if (bundle->starts_bundle()) {
2588 starts_bundle = '+';
2589 }
2590 }
2591
2592 if (WizardMode) {
2593 n->dump();
2594 }
2595
2596 if( !n->is_Region() && // Dont print in the Assembly
2597 !n->is_Phi() && // a few noisely useless nodes
2598 !n->is_Proj() &&
2599 !n->is_MachTemp() &&
2600 !n->is_SafePointScalarObject() &&
2601 !n->is_Catch() && // Would be nice to print exception table targets
2602 !n->is_MergeMem() && // Not very interesting
2603 !n->is_top() && // Debug info table constants
2604 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2605 ) {
2606 if (pcs && n->_idx < pc_limit)
2607 tty->print("%3.3x", pcs[n->_idx]);
2608 else
2609 tty->print(" ");
2610 tty->print(" %c ", starts_bundle);
2611 starts_bundle = ' ';
2612 tty->print("\t");
2613 n->format(_regalloc, tty);
2614 tty->cr();
2615 }
2616
2617 // If we have an instruction with a delay slot, and have seen a delay,
2618 // then back up and print it
2619 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2620 assert(delay != NULL, "no unconditional delay instruction");
2621 if (WizardMode) delay->dump();
2622
2623 if (node_bundling(delay)->starts_bundle())
2624 starts_bundle = '+';
2625 if (pcs && n->_idx < pc_limit)
2626 tty->print("%3.3x", pcs[n->_idx]);
2627 else
2628 tty->print(" ");
2629 tty->print(" %c ", starts_bundle);
2630 starts_bundle = ' ';
2631 tty->print("\t");
2632 delay->format(_regalloc, tty);
2633 tty->cr();
2634 delay = NULL;
2635 }
2636
2637 // Dump the exception table as well
2638 if( n->is_Catch() && (Verbose || WizardMode) ) {
2639 // Print the exception table for this offset
2640 _handler_table.print_subtable_for(pc);
2641 }
2642 }
2643
2644 if (pcs && n->_idx < pc_limit)
2645 tty->print_cr("%3.3x", pcs[n->_idx]);
2646 else
2647 tty->cr();
2648
2649 assert(cut_short || delay == NULL, "no unconditional delay branch");
2650
2651 } // End of per-block dump
2652 tty->cr();
2653
2654 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2655 }
2656 #endif
2657
2658 //------------------------------Final_Reshape_Counts---------------------------
2659 // This class defines counters to help identify when a method
2660 // may/must be executed using hardware with only 24-bit precision.
2661 struct Final_Reshape_Counts : public StackObj {
2662 int _call_count; // count non-inlined 'common' calls
2663 int _float_count; // count float ops requiring 24-bit precision
2664 int _double_count; // count double ops requiring more precision
2665 int _java_call_count; // count non-inlined 'java' calls
2666 int _inner_loop_count; // count loops which need alignment
2667 VectorSet _visited; // Visitation flags
2668 Node_List _tests; // Set of IfNodes & PCTableNodes
2669
2670 Final_Reshape_Counts() :
2671 _call_count(0), _float_count(0), _double_count(0),
2672 _java_call_count(0), _inner_loop_count(0),
2673 _visited( Thread::current()->resource_area() ) { }
2674
2675 void inc_call_count () { _call_count ++; }
2676 void inc_float_count () { _float_count ++; }
2677 void inc_double_count() { _double_count++; }
2678 void inc_java_call_count() { _java_call_count++; }
2679 void inc_inner_loop_count() { _inner_loop_count++; }
2680
2681 int get_call_count () const { return _call_count ; }
2682 int get_float_count () const { return _float_count ; }
2683 int get_double_count() const { return _double_count; }
2684 int get_java_call_count() const { return _java_call_count; }
2685 int get_inner_loop_count() const { return _inner_loop_count; }
2686 };
2687
2688 #ifdef ASSERT
2689 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2690 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2691 // Make sure the offset goes inside the instance layout.
2692 return k->contains_field_offset(tp->offset());
2693 // Note that OffsetBot and OffsetTop are very negative.
2694 }
2695 #endif
2696
2697 // Eliminate trivially redundant StoreCMs and accumulate their
2698 // precedence edges.
2699 void Compile::eliminate_redundant_card_marks(Node* n) {
2700 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2701 if (n->in(MemNode::Address)->outcnt() > 1) {
2702 // There are multiple users of the same address so it might be
2703 // possible to eliminate some of the StoreCMs
2704 Node* mem = n->in(MemNode::Memory);
2705 Node* adr = n->in(MemNode::Address);
2706 Node* val = n->in(MemNode::ValueIn);
2707 Node* prev = n;
2708 bool done = false;
2709 // Walk the chain of StoreCMs eliminating ones that match. As
2710 // long as it's a chain of single users then the optimization is
2711 // safe. Eliminating partially redundant StoreCMs would require
2712 // cloning copies down the other paths.
2713 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2714 if (adr == mem->in(MemNode::Address) &&
2715 val == mem->in(MemNode::ValueIn)) {
2716 // redundant StoreCM
2717 if (mem->req() > MemNode::OopStore) {
2718 // Hasn't been processed by this code yet.
2719 n->add_prec(mem->in(MemNode::OopStore));
2720 } else {
2721 // Already converted to precedence edge
2722 for (uint i = mem->req(); i < mem->len(); i++) {
2723 // Accumulate any precedence edges
2724 if (mem->in(i) != NULL) {
2725 n->add_prec(mem->in(i));
2726 }
2727 }
2728 // Everything above this point has been processed.
2729 done = true;
2730 }
2731 // Eliminate the previous StoreCM
2732 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2733 assert(mem->outcnt() == 0, "should be dead");
2734 mem->disconnect_inputs(NULL, this);
2735 } else {
2736 prev = mem;
2737 }
2738 mem = prev->in(MemNode::Memory);
2739 }
2740 }
2741 }
2742
2743 //------------------------------final_graph_reshaping_impl----------------------
2744 // Implement items 1-5 from final_graph_reshaping below.
2745 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2746
2747 if ( n->outcnt() == 0 ) return; // dead node
2748 uint nop = n->Opcode();
2749
2750 // Check for 2-input instruction with "last use" on right input.
2751 // Swap to left input. Implements item (2).
2752 if( n->req() == 3 && // two-input instruction
2753 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2754 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2755 n->in(2)->outcnt() == 1 &&// right use IS a last use
2756 !n->in(2)->is_Con() ) { // right use is not a constant
2757 // Check for commutative opcode
2758 switch( nop ) {
2759 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2760 case Op_MaxI: case Op_MinI:
2761 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2762 case Op_AndL: case Op_XorL: case Op_OrL:
2763 case Op_AndI: case Op_XorI: case Op_OrI: {
2764 // Move "last use" input to left by swapping inputs
2765 n->swap_edges(1, 2);
2766 break;
2767 }
2768 default:
2769 break;
2770 }
2771 }
2772
2773 #ifdef ASSERT
2774 if( n->is_Mem() ) {
2775 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2776 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2777 // oop will be recorded in oop map if load crosses safepoint
2778 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2779 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2780 "raw memory operations should have control edge");
2781 }
2782 if (n->is_MemBar()) {
2783 MemBarNode* mb = n->as_MemBar();
2784 if (mb->trailing_store() || mb->trailing_load_store()) {
2785 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2786 Node* mem = mb->in(MemBarNode::Precedent);
2787 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2788 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2789 } else if (mb->leading()) {
2790 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2791 }
2792 }
2793 #endif
2794 // Count FPU ops and common calls, implements item (3)
2795 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2796 if (!gc_handled) {
2797 final_graph_reshaping_main_switch(n, frc, nop);
2798 }
2799
2800 // Collect CFG split points
2801 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2802 frc._tests.push(n);
2803 }
2804 }
2805
2806 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2807 switch( nop ) {
2808 // Count all float operations that may use FPU
2809 case Op_AddF:
2810 case Op_SubF:
2811 case Op_MulF:
2812 case Op_DivF:
2813 case Op_NegF:
2814 case Op_ModF:
2815 case Op_ConvI2F:
2816 case Op_ConF:
2817 case Op_CmpF:
2818 case Op_CmpF3:
2819 // case Op_ConvL2F: // longs are split into 32-bit halves
2820 frc.inc_float_count();
2821 break;
2822
2823 case Op_ConvF2D:
2824 case Op_ConvD2F:
2825 frc.inc_float_count();
2826 frc.inc_double_count();
2827 break;
2828
2829 // Count all double operations that may use FPU
2830 case Op_AddD:
2831 case Op_SubD:
2832 case Op_MulD:
2833 case Op_DivD:
2834 case Op_NegD:
2835 case Op_ModD:
2836 case Op_ConvI2D:
2837 case Op_ConvD2I:
2838 // case Op_ConvL2D: // handled by leaf call
2839 // case Op_ConvD2L: // handled by leaf call
2840 case Op_ConD:
2841 case Op_CmpD:
2842 case Op_CmpD3:
2843 frc.inc_double_count();
2844 break;
2845 case Op_Opaque1: // Remove Opaque Nodes before matching
2846 case Op_Opaque2: // Remove Opaque Nodes before matching
2847 case Op_Opaque3:
2848 n->subsume_by(n->in(1), this);
2849 break;
2850 case Op_CallStaticJava:
2851 case Op_CallJava:
2852 case Op_CallDynamicJava:
2853 frc.inc_java_call_count(); // Count java call site;
2854 case Op_CallRuntime:
2855 case Op_CallLeaf:
2856 case Op_CallLeafNoFP: {
2857 assert (n->is_Call(), "");
2858 CallNode *call = n->as_Call();
2859 // Count call sites where the FP mode bit would have to be flipped.
2860 // Do not count uncommon runtime calls:
2861 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2862 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2863 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2864 frc.inc_call_count(); // Count the call site
2865 } else { // See if uncommon argument is shared
2866 Node *n = call->in(TypeFunc::Parms);
2867 int nop = n->Opcode();
2868 // Clone shared simple arguments to uncommon calls, item (1).
2869 if (n->outcnt() > 1 &&
2870 !n->is_Proj() &&
2871 nop != Op_CreateEx &&
2872 nop != Op_CheckCastPP &&
2873 nop != Op_DecodeN &&
2874 nop != Op_DecodeNKlass &&
2875 !n->is_Mem() &&
2876 !n->is_Phi()) {
2877 Node *x = n->clone();
2878 call->set_req(TypeFunc::Parms, x);
2879 }
2880 }
2881 break;
2882 }
2883
2884 case Op_StoreD:
2885 case Op_LoadD:
2886 case Op_LoadD_unaligned:
2887 frc.inc_double_count();
2888 goto handle_mem;
2889 case Op_StoreF:
2890 case Op_LoadF:
2891 frc.inc_float_count();
2892 goto handle_mem;
2893
2894 case Op_StoreCM:
2895 {
2896 // Convert OopStore dependence into precedence edge
2897 Node* prec = n->in(MemNode::OopStore);
2898 n->del_req(MemNode::OopStore);
2899 n->add_prec(prec);
2900 eliminate_redundant_card_marks(n);
2901 }
2902
2903 // fall through
2904
2905 case Op_StoreB:
2906 case Op_StoreC:
2907 case Op_StorePConditional:
2908 case Op_StoreI:
2909 case Op_StoreL:
2910 case Op_StoreIConditional:
2911 case Op_StoreLConditional:
2912 case Op_CompareAndSwapB:
2913 case Op_CompareAndSwapS:
2914 case Op_CompareAndSwapI:
2915 case Op_CompareAndSwapL:
2916 case Op_CompareAndSwapP:
2917 case Op_CompareAndSwapN:
2918 case Op_WeakCompareAndSwapB:
2919 case Op_WeakCompareAndSwapS:
2920 case Op_WeakCompareAndSwapI:
2921 case Op_WeakCompareAndSwapL:
2922 case Op_WeakCompareAndSwapP:
2923 case Op_WeakCompareAndSwapN:
2924 case Op_CompareAndExchangeB:
2925 case Op_CompareAndExchangeS:
2926 case Op_CompareAndExchangeI:
2927 case Op_CompareAndExchangeL:
2928 case Op_CompareAndExchangeP:
2929 case Op_CompareAndExchangeN:
2930 case Op_GetAndAddS:
2931 case Op_GetAndAddB:
2932 case Op_GetAndAddI:
2933 case Op_GetAndAddL:
2934 case Op_GetAndSetS:
2935 case Op_GetAndSetB:
2936 case Op_GetAndSetI:
2937 case Op_GetAndSetL:
2938 case Op_GetAndSetP:
2939 case Op_GetAndSetN:
2940 case Op_StoreP:
2941 case Op_StoreN:
2942 case Op_StoreNKlass:
2943 case Op_LoadB:
2944 case Op_LoadUB:
2945 case Op_LoadUS:
2946 case Op_LoadI:
2947 case Op_LoadKlass:
2948 case Op_LoadNKlass:
2949 case Op_LoadL:
2950 case Op_LoadL_unaligned:
2951 case Op_LoadPLocked:
2952 case Op_LoadP:
2953 case Op_LoadN:
2954 case Op_LoadRange:
2955 case Op_LoadS: {
2956 handle_mem:
2957 #ifdef ASSERT
2958 if( VerifyOptoOopOffsets ) {
2959 MemNode* mem = n->as_Mem();
2960 // Check to see if address types have grounded out somehow.
2961 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2962 assert( !tp || oop_offset_is_sane(tp), "" );
2963 }
2964 #endif
2965 break;
2966 }
2967
2968 case Op_AddP: { // Assert sane base pointers
2969 Node *addp = n->in(AddPNode::Address);
2970 assert( !addp->is_AddP() ||
2971 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2972 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2973 "Base pointers must match (addp %u)", addp->_idx );
2974 #ifdef _LP64
2975 if ((UseCompressedOops || UseCompressedClassPointers) &&
2976 addp->Opcode() == Op_ConP &&
2977 addp == n->in(AddPNode::Base) &&
2978 n->in(AddPNode::Offset)->is_Con()) {
2979 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2980 // on the platform and on the compressed oops mode.
2981 // Use addressing with narrow klass to load with offset on x86.
2982 // Some platforms can use the constant pool to load ConP.
2983 // Do this transformation here since IGVN will convert ConN back to ConP.
2984 const Type* t = addp->bottom_type();
2985 bool is_oop = t->isa_oopptr() != NULL;
2986 bool is_klass = t->isa_klassptr() != NULL;
2987
2988 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2989 (is_klass && Matcher::const_klass_prefer_decode())) {
2990 Node* nn = NULL;
2991
2992 int op = is_oop ? Op_ConN : Op_ConNKlass;
2993
2994 // Look for existing ConN node of the same exact type.
2995 Node* r = root();
2996 uint cnt = r->outcnt();
2997 for (uint i = 0; i < cnt; i++) {
2998 Node* m = r->raw_out(i);
2999 if (m!= NULL && m->Opcode() == op &&
3000 m->bottom_type()->make_ptr() == t) {
3001 nn = m;
3002 break;
3003 }
3004 }
3005 if (nn != NULL) {
3006 // Decode a narrow oop to match address
3007 // [R12 + narrow_oop_reg<<3 + offset]
3008 if (is_oop) {
3009 nn = new DecodeNNode(nn, t);
3010 } else {
3011 nn = new DecodeNKlassNode(nn, t);
3012 }
3013 // Check for succeeding AddP which uses the same Base.
3014 // Otherwise we will run into the assertion above when visiting that guy.
3015 for (uint i = 0; i < n->outcnt(); ++i) {
3016 Node *out_i = n->raw_out(i);
3017 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3018 out_i->set_req(AddPNode::Base, nn);
3019 #ifdef ASSERT
3020 for (uint j = 0; j < out_i->outcnt(); ++j) {
3021 Node *out_j = out_i->raw_out(j);
3022 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3023 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3024 }
3025 #endif
3026 }
3027 }
3028 n->set_req(AddPNode::Base, nn);
3029 n->set_req(AddPNode::Address, nn);
3030 if (addp->outcnt() == 0) {
3031 addp->disconnect_inputs(NULL, this);
3032 }
3033 }
3034 }
3035 }
3036 #endif
3037 // platform dependent reshaping of the address expression
3038 reshape_address(n->as_AddP());
3039 break;
3040 }
3041
3042 case Op_CastPP: {
3043 // Remove CastPP nodes to gain more freedom during scheduling but
3044 // keep the dependency they encode as control or precedence edges
3045 // (if control is set already) on memory operations. Some CastPP
3046 // nodes don't have a control (don't carry a dependency): skip
3047 // those.
3048 if (n->in(0) != NULL) {
3049 ResourceMark rm;
3050 Unique_Node_List wq;
3051 wq.push(n);
3052 for (uint next = 0; next < wq.size(); ++next) {
3053 Node *m = wq.at(next);
3054 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3055 Node* use = m->fast_out(i);
3056 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3057 use->ensure_control_or_add_prec(n->in(0));
3058 } else {
3059 switch(use->Opcode()) {
3060 case Op_AddP:
3061 case Op_DecodeN:
3062 case Op_DecodeNKlass:
3063 case Op_CheckCastPP:
3064 case Op_CastPP:
3065 wq.push(use);
3066 break;
3067 }
3068 }
3069 }
3070 }
3071 }
3072 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3073 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3074 Node* in1 = n->in(1);
3075 const Type* t = n->bottom_type();
3076 Node* new_in1 = in1->clone();
3077 new_in1->as_DecodeN()->set_type(t);
3078
3079 if (!Matcher::narrow_oop_use_complex_address()) {
3080 //
3081 // x86, ARM and friends can handle 2 adds in addressing mode
3082 // and Matcher can fold a DecodeN node into address by using
3083 // a narrow oop directly and do implicit NULL check in address:
3084 //
3085 // [R12 + narrow_oop_reg<<3 + offset]
3086 // NullCheck narrow_oop_reg
3087 //
3088 // On other platforms (Sparc) we have to keep new DecodeN node and
3089 // use it to do implicit NULL check in address:
3090 //
3091 // decode_not_null narrow_oop_reg, base_reg
3092 // [base_reg + offset]
3093 // NullCheck base_reg
3094 //
3095 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3096 // to keep the information to which NULL check the new DecodeN node
3097 // corresponds to use it as value in implicit_null_check().
3098 //
3099 new_in1->set_req(0, n->in(0));
3100 }
3101
3102 n->subsume_by(new_in1, this);
3103 if (in1->outcnt() == 0) {
3104 in1->disconnect_inputs(NULL, this);
3105 }
3106 } else {
3107 n->subsume_by(n->in(1), this);
3108 if (n->outcnt() == 0) {
3109 n->disconnect_inputs(NULL, this);
3110 }
3111 }
3112 break;
3113 }
3114 #ifdef _LP64
3115 case Op_CmpP:
3116 // Do this transformation here to preserve CmpPNode::sub() and
3117 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3118 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3119 Node* in1 = n->in(1);
3120 Node* in2 = n->in(2);
3121 if (!in1->is_DecodeNarrowPtr()) {
3122 in2 = in1;
3123 in1 = n->in(2);
3124 }
3125 assert(in1->is_DecodeNarrowPtr(), "sanity");
3126
3127 Node* new_in2 = NULL;
3128 if (in2->is_DecodeNarrowPtr()) {
3129 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3130 new_in2 = in2->in(1);
3131 } else if (in2->Opcode() == Op_ConP) {
3132 const Type* t = in2->bottom_type();
3133 if (t == TypePtr::NULL_PTR) {
3134 assert(in1->is_DecodeN(), "compare klass to null?");
3135 // Don't convert CmpP null check into CmpN if compressed
3136 // oops implicit null check is not generated.
3137 // This will allow to generate normal oop implicit null check.
3138 if (Matcher::gen_narrow_oop_implicit_null_checks())
3139 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3140 //
3141 // This transformation together with CastPP transformation above
3142 // will generated code for implicit NULL checks for compressed oops.
3143 //
3144 // The original code after Optimize()
3145 //
3146 // LoadN memory, narrow_oop_reg
3147 // decode narrow_oop_reg, base_reg
3148 // CmpP base_reg, NULL
3149 // CastPP base_reg // NotNull
3150 // Load [base_reg + offset], val_reg
3151 //
3152 // after these transformations will be
3153 //
3154 // LoadN memory, narrow_oop_reg
3155 // CmpN narrow_oop_reg, NULL
3156 // decode_not_null narrow_oop_reg, base_reg
3157 // Load [base_reg + offset], val_reg
3158 //
3159 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3160 // since narrow oops can be used in debug info now (see the code in
3161 // final_graph_reshaping_walk()).
3162 //
3163 // At the end the code will be matched to
3164 // on x86:
3165 //
3166 // Load_narrow_oop memory, narrow_oop_reg
3167 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3168 // NullCheck narrow_oop_reg
3169 //
3170 // and on sparc:
3171 //
3172 // Load_narrow_oop memory, narrow_oop_reg
3173 // decode_not_null narrow_oop_reg, base_reg
3174 // Load [base_reg + offset], val_reg
3175 // NullCheck base_reg
3176 //
3177 } else if (t->isa_oopptr()) {
3178 new_in2 = ConNode::make(t->make_narrowoop());
3179 } else if (t->isa_klassptr()) {
3180 new_in2 = ConNode::make(t->make_narrowklass());
3181 }
3182 }
3183 if (new_in2 != NULL) {
3184 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3185 n->subsume_by(cmpN, this);
3186 if (in1->outcnt() == 0) {
3187 in1->disconnect_inputs(NULL, this);
3188 }
3189 if (in2->outcnt() == 0) {
3190 in2->disconnect_inputs(NULL, this);
3191 }
3192 }
3193 }
3194 break;
3195
3196 case Op_DecodeN:
3197 case Op_DecodeNKlass:
3198 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3199 // DecodeN could be pinned when it can't be fold into
3200 // an address expression, see the code for Op_CastPP above.
3201 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3202 break;
3203
3204 case Op_EncodeP:
3205 case Op_EncodePKlass: {
3206 Node* in1 = n->in(1);
3207 if (in1->is_DecodeNarrowPtr()) {
3208 n->subsume_by(in1->in(1), this);
3209 } else if (in1->Opcode() == Op_ConP) {
3210 const Type* t = in1->bottom_type();
3211 if (t == TypePtr::NULL_PTR) {
3212 assert(t->isa_oopptr(), "null klass?");
3213 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3214 } else if (t->isa_oopptr()) {
3215 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3216 } else if (t->isa_klassptr()) {
3217 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3218 }
3219 }
3220 if (in1->outcnt() == 0) {
3221 in1->disconnect_inputs(NULL, this);
3222 }
3223 break;
3224 }
3225
3226 case Op_Proj: {
3227 if (OptimizeStringConcat) {
3228 ProjNode* p = n->as_Proj();
3229 if (p->_is_io_use) {
3230 // Separate projections were used for the exception path which
3231 // are normally removed by a late inline. If it wasn't inlined
3232 // then they will hang around and should just be replaced with
3233 // the original one.
3234 Node* proj = NULL;
3235 // Replace with just one
3236 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3237 Node *use = i.get();
3238 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3239 proj = use;
3240 break;
3241 }
3242 }
3243 assert(proj != NULL, "must be found");
3244 p->subsume_by(proj, this);
3245 }
3246 }
3247 break;
3248 }
3249
3250 case Op_Phi:
3251 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3252 // The EncodeP optimization may create Phi with the same edges
3253 // for all paths. It is not handled well by Register Allocator.
3254 Node* unique_in = n->in(1);
3255 assert(unique_in != NULL, "");
3256 uint cnt = n->req();
3257 for (uint i = 2; i < cnt; i++) {
3258 Node* m = n->in(i);
3259 assert(m != NULL, "");
3260 if (unique_in != m)
3261 unique_in = NULL;
3262 }
3263 if (unique_in != NULL) {
3264 n->subsume_by(unique_in, this);
3265 }
3266 }
3267 break;
3268
3269 #endif
3270
3271 #ifdef ASSERT
3272 case Op_CastII:
3273 // Verify that all range check dependent CastII nodes were removed.
3274 if (n->isa_CastII()->has_range_check()) {
3275 n->dump(3);
3276 assert(false, "Range check dependent CastII node was not removed");
3277 }
3278 break;
3279 #endif
3280
3281 case Op_ModI:
3282 if (UseDivMod) {
3283 // Check if a%b and a/b both exist
3284 Node* d = n->find_similar(Op_DivI);
3285 if (d) {
3286 // Replace them with a fused divmod if supported
3287 if (Matcher::has_match_rule(Op_DivModI)) {
3288 DivModINode* divmod = DivModINode::make(n);
3289 d->subsume_by(divmod->div_proj(), this);
3290 n->subsume_by(divmod->mod_proj(), this);
3291 } else {
3292 // replace a%b with a-((a/b)*b)
3293 Node* mult = new MulINode(d, d->in(2));
3294 Node* sub = new SubINode(d->in(1), mult);
3295 n->subsume_by(sub, this);
3296 }
3297 }
3298 }
3299 break;
3300
3301 case Op_ModL:
3302 if (UseDivMod) {
3303 // Check if a%b and a/b both exist
3304 Node* d = n->find_similar(Op_DivL);
3305 if (d) {
3306 // Replace them with a fused divmod if supported
3307 if (Matcher::has_match_rule(Op_DivModL)) {
3308 DivModLNode* divmod = DivModLNode::make(n);
3309 d->subsume_by(divmod->div_proj(), this);
3310 n->subsume_by(divmod->mod_proj(), this);
3311 } else {
3312 // replace a%b with a-((a/b)*b)
3313 Node* mult = new MulLNode(d, d->in(2));
3314 Node* sub = new SubLNode(d->in(1), mult);
3315 n->subsume_by(sub, this);
3316 }
3317 }
3318 }
3319 break;
3320
3321 case Op_LoadVector:
3322 case Op_StoreVector:
3323 break;
3324
3325 case Op_AddReductionVI:
3326 case Op_AddReductionVL:
3327 case Op_AddReductionVF:
3328 case Op_AddReductionVD:
3329 case Op_MulReductionVI:
3330 case Op_MulReductionVL:
3331 case Op_MulReductionVF:
3332 case Op_MulReductionVD:
3333 break;
3334
3335 case Op_PackB:
3336 case Op_PackS:
3337 case Op_PackI:
3338 case Op_PackF:
3339 case Op_PackL:
3340 case Op_PackD:
3341 if (n->req()-1 > 2) {
3342 // Replace many operand PackNodes with a binary tree for matching
3343 PackNode* p = (PackNode*) n;
3344 Node* btp = p->binary_tree_pack(1, n->req());
3345 n->subsume_by(btp, this);
3346 }
3347 break;
3348 case Op_Loop:
3349 case Op_CountedLoop:
3350 case Op_OuterStripMinedLoop:
3351 if (n->as_Loop()->is_inner_loop()) {
3352 frc.inc_inner_loop_count();
3353 }
3354 n->as_Loop()->verify_strip_mined(0);
3355 break;
3356 case Op_LShiftI:
3357 case Op_RShiftI:
3358 case Op_URShiftI:
3359 case Op_LShiftL:
3360 case Op_RShiftL:
3361 case Op_URShiftL:
3362 if (Matcher::need_masked_shift_count) {
3363 // The cpu's shift instructions don't restrict the count to the
3364 // lower 5/6 bits. We need to do the masking ourselves.
3365 Node* in2 = n->in(2);
3366 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3367 const TypeInt* t = in2->find_int_type();
3368 if (t != NULL && t->is_con()) {
3369 juint shift = t->get_con();
3370 if (shift > mask) { // Unsigned cmp
3371 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3372 }
3373 } else {
3374 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3375 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3376 n->set_req(2, shift);
3377 }
3378 }
3379 if (in2->outcnt() == 0) { // Remove dead node
3380 in2->disconnect_inputs(NULL, this);
3381 }
3382 }
3383 break;
3384 case Op_MemBarStoreStore:
3385 case Op_MemBarRelease:
3386 // Break the link with AllocateNode: it is no longer useful and
3387 // confuses register allocation.
3388 if (n->req() > MemBarNode::Precedent) {
3389 n->set_req(MemBarNode::Precedent, top());
3390 }
3391 break;
3392 case Op_MemBarAcquire: {
3393 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3394 // At parse time, the trailing MemBarAcquire for a volatile load
3395 // is created with an edge to the load. After optimizations,
3396 // that input may be a chain of Phis. If those phis have no
3397 // other use, then the MemBarAcquire keeps them alive and
3398 // register allocation can be confused.
3399 ResourceMark rm;
3400 Unique_Node_List wq;
3401 wq.push(n->in(MemBarNode::Precedent));
3402 n->set_req(MemBarNode::Precedent, top());
3403 while (wq.size() > 0) {
3404 Node* m = wq.pop();
3405 if (m->outcnt() == 0) {
3406 for (uint j = 0; j < m->req(); j++) {
3407 Node* in = m->in(j);
3408 if (in != NULL) {
3409 wq.push(in);
3410 }
3411 }
3412 m->disconnect_inputs(NULL, this);
3413 }
3414 }
3415 }
3416 break;
3417 }
3418 case Op_RangeCheck: {
3419 RangeCheckNode* rc = n->as_RangeCheck();
3420 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3421 n->subsume_by(iff, this);
3422 frc._tests.push(iff);
3423 break;
3424 }
3425 case Op_ConvI2L: {
3426 if (!Matcher::convi2l_type_required) {
3427 // Code generation on some platforms doesn't need accurate
3428 // ConvI2L types. Widening the type can help remove redundant
3429 // address computations.
3430 n->as_Type()->set_type(TypeLong::INT);
3431 ResourceMark rm;
3432 Node_List wq;
3433 wq.push(n);
3434 for (uint next = 0; next < wq.size(); next++) {
3435 Node *m = wq.at(next);
3436
3437 for(;;) {
3438 // Loop over all nodes with identical inputs edges as m
3439 Node* k = m->find_similar(m->Opcode());
3440 if (k == NULL) {
3441 break;
3442 }
3443 // Push their uses so we get a chance to remove node made
3444 // redundant
3445 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3446 Node* u = k->fast_out(i);
3447 assert(!wq.contains(u), "shouldn't process one node several times");
3448 if (u->Opcode() == Op_LShiftL ||
3449 u->Opcode() == Op_AddL ||
3450 u->Opcode() == Op_SubL ||
3451 u->Opcode() == Op_AddP) {
3452 wq.push(u);
3453 }
3454 }
3455 // Replace all nodes with identical edges as m with m
3456 k->subsume_by(m, this);
3457 }
3458 }
3459 }
3460 break;
3461 }
3462 case Op_CmpUL: {
3463 if (!Matcher::has_match_rule(Op_CmpUL)) {
3464 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3465 // to max_julong if < 0 to make the signed comparison fail.
3466 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3467 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3468 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3469 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3470 Node* andl = new AndLNode(orl, remove_sign_mask);
3471 Node* cmp = new CmpLNode(andl, n->in(2));
3472 n->subsume_by(cmp, this);
3473 }
3474 break;
3475 }
3476 default:
3477 assert(!n->is_Call(), "");
3478 assert(!n->is_Mem(), "");
3479 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3480 break;
3481 }
3482 }
3483
3484 //------------------------------final_graph_reshaping_walk---------------------
3485 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3486 // requires that the walk visits a node's inputs before visiting the node.
3487 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3488 ResourceArea *area = Thread::current()->resource_area();
3489 Unique_Node_List sfpt(area);
3490
3491 frc._visited.set(root->_idx); // first, mark node as visited
3492 uint cnt = root->req();
3493 Node *n = root;
3494 uint i = 0;
3495 while (true) {
3496 if (i < cnt) {
3497 // Place all non-visited non-null inputs onto stack
3498 Node* m = n->in(i);
3499 ++i;
3500 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3501 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3502 // compute worst case interpreter size in case of a deoptimization
3503 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3504
3505 sfpt.push(m);
3506 }
3507 cnt = m->req();
3508 nstack.push(n, i); // put on stack parent and next input's index
3509 n = m;
3510 i = 0;
3511 }
3512 } else {
3513 // Now do post-visit work
3514 final_graph_reshaping_impl( n, frc );
3515 if (nstack.is_empty())
3516 break; // finished
3517 n = nstack.node(); // Get node from stack
3518 cnt = n->req();
3519 i = nstack.index();
3520 nstack.pop(); // Shift to the next node on stack
3521 }
3522 }
3523
3524 // Skip next transformation if compressed oops are not used.
3525 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3526 (!UseCompressedOops && !UseCompressedClassPointers))
3527 return;
3528
3529 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3530 // It could be done for an uncommon traps or any safepoints/calls
3531 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3532 while (sfpt.size() > 0) {
3533 n = sfpt.pop();
3534 JVMState *jvms = n->as_SafePoint()->jvms();
3535 assert(jvms != NULL, "sanity");
3536 int start = jvms->debug_start();
3537 int end = n->req();
3538 bool is_uncommon = (n->is_CallStaticJava() &&
3539 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3540 for (int j = start; j < end; j++) {
3541 Node* in = n->in(j);
3542 if (in->is_DecodeNarrowPtr()) {
3543 bool safe_to_skip = true;
3544 if (!is_uncommon ) {
3545 // Is it safe to skip?
3546 for (uint i = 0; i < in->outcnt(); i++) {
3547 Node* u = in->raw_out(i);
3548 if (!u->is_SafePoint() ||
3549 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3550 safe_to_skip = false;
3551 }
3552 }
3553 }
3554 if (safe_to_skip) {
3555 n->set_req(j, in->in(1));
3556 }
3557 if (in->outcnt() == 0) {
3558 in->disconnect_inputs(NULL, this);
3559 }
3560 }
3561 }
3562 }
3563 }
3564
3565 //------------------------------final_graph_reshaping--------------------------
3566 // Final Graph Reshaping.
3567 //
3568 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3569 // and not commoned up and forced early. Must come after regular
3570 // optimizations to avoid GVN undoing the cloning. Clone constant
3571 // inputs to Loop Phis; these will be split by the allocator anyways.
3572 // Remove Opaque nodes.
3573 // (2) Move last-uses by commutative operations to the left input to encourage
3574 // Intel update-in-place two-address operations and better register usage
3575 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3576 // calls canonicalizing them back.
3577 // (3) Count the number of double-precision FP ops, single-precision FP ops
3578 // and call sites. On Intel, we can get correct rounding either by
3579 // forcing singles to memory (requires extra stores and loads after each
3580 // FP bytecode) or we can set a rounding mode bit (requires setting and
3581 // clearing the mode bit around call sites). The mode bit is only used
3582 // if the relative frequency of single FP ops to calls is low enough.
3583 // This is a key transform for SPEC mpeg_audio.
3584 // (4) Detect infinite loops; blobs of code reachable from above but not
3585 // below. Several of the Code_Gen algorithms fail on such code shapes,
3586 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3587 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3588 // Detection is by looking for IfNodes where only 1 projection is
3589 // reachable from below or CatchNodes missing some targets.
3590 // (5) Assert for insane oop offsets in debug mode.
3591
3592 bool Compile::final_graph_reshaping() {
3593 // an infinite loop may have been eliminated by the optimizer,
3594 // in which case the graph will be empty.
3595 if (root()->req() == 1) {
3596 record_method_not_compilable("trivial infinite loop");
3597 return true;
3598 }
3599
3600 // Expensive nodes have their control input set to prevent the GVN
3601 // from freely commoning them. There's no GVN beyond this point so
3602 // no need to keep the control input. We want the expensive nodes to
3603 // be freely moved to the least frequent code path by gcm.
3604 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3605 for (int i = 0; i < expensive_count(); i++) {
3606 _expensive_nodes->at(i)->set_req(0, NULL);
3607 }
3608
3609 Final_Reshape_Counts frc;
3610
3611 // Visit everybody reachable!
3612 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3613 Node_Stack nstack(live_nodes() >> 1);
3614 final_graph_reshaping_walk(nstack, root(), frc);
3615
3616 // Check for unreachable (from below) code (i.e., infinite loops).
3617 for( uint i = 0; i < frc._tests.size(); i++ ) {
3618 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3619 // Get number of CFG targets.
3620 // Note that PCTables include exception targets after calls.
3621 uint required_outcnt = n->required_outcnt();
3622 if (n->outcnt() != required_outcnt) {
3623 // Check for a few special cases. Rethrow Nodes never take the
3624 // 'fall-thru' path, so expected kids is 1 less.
3625 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3626 if (n->in(0)->in(0)->is_Call()) {
3627 CallNode *call = n->in(0)->in(0)->as_Call();
3628 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3629 required_outcnt--; // Rethrow always has 1 less kid
3630 } else if (call->req() > TypeFunc::Parms &&
3631 call->is_CallDynamicJava()) {
3632 // Check for null receiver. In such case, the optimizer has
3633 // detected that the virtual call will always result in a null
3634 // pointer exception. The fall-through projection of this CatchNode
3635 // will not be populated.
3636 Node *arg0 = call->in(TypeFunc::Parms);
3637 if (arg0->is_Type() &&
3638 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3639 required_outcnt--;
3640 }
3641 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3642 call->req() > TypeFunc::Parms+1 &&
3643 call->is_CallStaticJava()) {
3644 // Check for negative array length. In such case, the optimizer has
3645 // detected that the allocation attempt will always result in an
3646 // exception. There is no fall-through projection of this CatchNode .
3647 Node *arg1 = call->in(TypeFunc::Parms+1);
3648 if (arg1->is_Type() &&
3649 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3650 required_outcnt--;
3651 }
3652 }
3653 }
3654 }
3655 // Recheck with a better notion of 'required_outcnt'
3656 if (n->outcnt() != required_outcnt) {
3657 record_method_not_compilable("malformed control flow");
3658 return true; // Not all targets reachable!
3659 }
3660 }
3661 // Check that I actually visited all kids. Unreached kids
3662 // must be infinite loops.
3663 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3664 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3665 record_method_not_compilable("infinite loop");
3666 return true; // Found unvisited kid; must be unreach
3667 }
3668
3669 // Here so verification code in final_graph_reshaping_walk()
3670 // always see an OuterStripMinedLoopEnd
3671 if (n->is_OuterStripMinedLoopEnd()) {
3672 IfNode* init_iff = n->as_If();
3673 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3674 n->subsume_by(iff, this);
3675 }
3676 }
3677
3678 // If original bytecodes contained a mixture of floats and doubles
3679 // check if the optimizer has made it homogenous, item (3).
3680 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3681 frc.get_float_count() > 32 &&
3682 frc.get_double_count() == 0 &&
3683 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3684 set_24_bit_selection_and_mode( false, true );
3685 }
3686
3687 set_java_calls(frc.get_java_call_count());
3688 set_inner_loops(frc.get_inner_loop_count());
3689
3690 // No infinite loops, no reason to bail out.
3691 return false;
3692 }
3693
3694 //-----------------------------too_many_traps----------------------------------
3695 // Report if there are too many traps at the current method and bci.
3696 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3697 bool Compile::too_many_traps(ciMethod* method,
3698 int bci,
3699 Deoptimization::DeoptReason reason) {
3700 ciMethodData* md = method->method_data();
3701 if (md->is_empty()) {
3702 // Assume the trap has not occurred, or that it occurred only
3703 // because of a transient condition during start-up in the interpreter.
3704 return false;
3705 }
3706 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3707 if (md->has_trap_at(bci, m, reason) != 0) {
3708 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3709 // Also, if there are multiple reasons, or if there is no per-BCI record,
3710 // assume the worst.
3711 if (log())
3712 log()->elem("observe trap='%s' count='%d'",
3713 Deoptimization::trap_reason_name(reason),
3714 md->trap_count(reason));
3715 return true;
3716 } else {
3717 // Ignore method/bci and see if there have been too many globally.
3718 return too_many_traps(reason, md);
3719 }
3720 }
3721
3722 // Less-accurate variant which does not require a method and bci.
3723 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3724 ciMethodData* logmd) {
3725 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3726 // Too many traps globally.
3727 // Note that we use cumulative trap_count, not just md->trap_count.
3728 if (log()) {
3729 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3730 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3731 Deoptimization::trap_reason_name(reason),
3732 mcount, trap_count(reason));
3733 }
3734 return true;
3735 } else {
3736 // The coast is clear.
3737 return false;
3738 }
3739 }
3740
3741 //--------------------------too_many_recompiles--------------------------------
3742 // Report if there are too many recompiles at the current method and bci.
3743 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3744 // Is not eager to return true, since this will cause the compiler to use
3745 // Action_none for a trap point, to avoid too many recompilations.
3746 bool Compile::too_many_recompiles(ciMethod* method,
3747 int bci,
3748 Deoptimization::DeoptReason reason) {
3749 ciMethodData* md = method->method_data();
3750 if (md->is_empty()) {
3751 // Assume the trap has not occurred, or that it occurred only
3752 // because of a transient condition during start-up in the interpreter.
3753 return false;
3754 }
3755 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3756 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3757 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3758 Deoptimization::DeoptReason per_bc_reason
3759 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3760 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3761 if ((per_bc_reason == Deoptimization::Reason_none
3762 || md->has_trap_at(bci, m, reason) != 0)
3763 // The trap frequency measure we care about is the recompile count:
3764 && md->trap_recompiled_at(bci, m)
3765 && md->overflow_recompile_count() >= bc_cutoff) {
3766 // Do not emit a trap here if it has already caused recompilations.
3767 // Also, if there are multiple reasons, or if there is no per-BCI record,
3768 // assume the worst.
3769 if (log())
3770 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3771 Deoptimization::trap_reason_name(reason),
3772 md->trap_count(reason),
3773 md->overflow_recompile_count());
3774 return true;
3775 } else if (trap_count(reason) != 0
3776 && decompile_count() >= m_cutoff) {
3777 // Too many recompiles globally, and we have seen this sort of trap.
3778 // Use cumulative decompile_count, not just md->decompile_count.
3779 if (log())
3780 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3781 Deoptimization::trap_reason_name(reason),
3782 md->trap_count(reason), trap_count(reason),
3783 md->decompile_count(), decompile_count());
3784 return true;
3785 } else {
3786 // The coast is clear.
3787 return false;
3788 }
3789 }
3790
3791 // Compute when not to trap. Used by matching trap based nodes and
3792 // NullCheck optimization.
3793 void Compile::set_allowed_deopt_reasons() {
3794 _allowed_reasons = 0;
3795 if (is_method_compilation()) {
3796 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3797 assert(rs < BitsPerInt, "recode bit map");
3798 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3799 _allowed_reasons |= nth_bit(rs);
3800 }
3801 }
3802 }
3803 }
3804
3805 #ifndef PRODUCT
3806 //------------------------------verify_graph_edges---------------------------
3807 // Walk the Graph and verify that there is a one-to-one correspondence
3808 // between Use-Def edges and Def-Use edges in the graph.
3809 void Compile::verify_graph_edges(bool no_dead_code) {
3810 if (VerifyGraphEdges) {
3811 ResourceArea *area = Thread::current()->resource_area();
3812 Unique_Node_List visited(area);
3813 // Call recursive graph walk to check edges
3814 _root->verify_edges(visited);
3815 if (no_dead_code) {
3816 // Now make sure that no visited node is used by an unvisited node.
3817 bool dead_nodes = false;
3818 Unique_Node_List checked(area);
3819 while (visited.size() > 0) {
3820 Node* n = visited.pop();
3821 checked.push(n);
3822 for (uint i = 0; i < n->outcnt(); i++) {
3823 Node* use = n->raw_out(i);
3824 if (checked.member(use)) continue; // already checked
3825 if (visited.member(use)) continue; // already in the graph
3826 if (use->is_Con()) continue; // a dead ConNode is OK
3827 // At this point, we have found a dead node which is DU-reachable.
3828 if (!dead_nodes) {
3829 tty->print_cr("*** Dead nodes reachable via DU edges:");
3830 dead_nodes = true;
3831 }
3832 use->dump(2);
3833 tty->print_cr("---");
3834 checked.push(use); // No repeats; pretend it is now checked.
3835 }
3836 }
3837 assert(!dead_nodes, "using nodes must be reachable from root");
3838 }
3839 }
3840 }
3841 #endif
3842
3843 // The Compile object keeps track of failure reasons separately from the ciEnv.
3844 // This is required because there is not quite a 1-1 relation between the
3845 // ciEnv and its compilation task and the Compile object. Note that one
3846 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3847 // to backtrack and retry without subsuming loads. Other than this backtracking
3848 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3849 // by the logic in C2Compiler.
3850 void Compile::record_failure(const char* reason) {
3851 if (log() != NULL) {
3852 log()->elem("failure reason='%s' phase='compile'", reason);
3853 }
3854 if (_failure_reason == NULL) {
3855 // Record the first failure reason.
3856 _failure_reason = reason;
3857 }
3858
3859 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3860 C->print_method(PHASE_FAILURE);
3861 }
3862 _root = NULL; // flush the graph, too
3863 }
3864
3865 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3866 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3867 _phase_name(name), _dolog(CITimeVerbose)
3868 {
3869 if (_dolog) {
3870 C = Compile::current();
3871 _log = C->log();
3872 } else {
3873 C = NULL;
3874 _log = NULL;
3875 }
3876 if (_log != NULL) {
3877 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3878 _log->stamp();
3879 _log->end_head();
3880 }
3881 }
3882
3883 Compile::TracePhase::~TracePhase() {
3884
3885 C = Compile::current();
3886 if (_dolog) {
3887 _log = C->log();
3888 } else {
3889 _log = NULL;
3890 }
3891
3892 #ifdef ASSERT
3893 if (PrintIdealNodeCount) {
3894 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3895 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3896 }
3897
3898 if (VerifyIdealNodeCount) {
3899 Compile::current()->print_missing_nodes();
3900 }
3901 #endif
3902
3903 if (_log != NULL) {
3904 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3905 }
3906 }
3907
3908 //=============================================================================
3909 // Two Constant's are equal when the type and the value are equal.
3910 bool Compile::Constant::operator==(const Constant& other) {
3911 if (type() != other.type() ) return false;
3912 if (can_be_reused() != other.can_be_reused()) return false;
3913 // For floating point values we compare the bit pattern.
3914 switch (type()) {
3915 case T_INT:
3916 case T_FLOAT: return (_v._value.i == other._v._value.i);
3917 case T_LONG:
3918 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3919 case T_OBJECT:
3920 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3921 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3922 case T_METADATA: return (_v._metadata == other._v._metadata);
3923 default: ShouldNotReachHere(); return false;
3924 }
3925 }
3926
3927 static int type_to_size_in_bytes(BasicType t) {
3928 switch (t) {
3929 case T_INT: return sizeof(jint );
3930 case T_LONG: return sizeof(jlong );
3931 case T_FLOAT: return sizeof(jfloat );
3932 case T_DOUBLE: return sizeof(jdouble);
3933 case T_METADATA: return sizeof(Metadata*);
3934 // We use T_VOID as marker for jump-table entries (labels) which
3935 // need an internal word relocation.
3936 case T_VOID:
3937 case T_ADDRESS:
3938 case T_OBJECT: return sizeof(jobject);
3939 default:
3940 ShouldNotReachHere();
3941 return -1;
3942 }
3943 }
3944
3945 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3946 // sort descending
3947 if (a->freq() > b->freq()) return -1;
3948 if (a->freq() < b->freq()) return 1;
3949 return 0;
3950 }
3951
3952 void Compile::ConstantTable::calculate_offsets_and_size() {
3953 // First, sort the array by frequencies.
3954 _constants.sort(qsort_comparator);
3955
3956 #ifdef ASSERT
3957 // Make sure all jump-table entries were sorted to the end of the
3958 // array (they have a negative frequency).
3959 bool found_void = false;
3960 for (int i = 0; i < _constants.length(); i++) {
3961 Constant con = _constants.at(i);
3962 if (con.type() == T_VOID)
3963 found_void = true; // jump-tables
3964 else
3965 assert(!found_void, "wrong sorting");
3966 }
3967 #endif
3968
3969 int offset = 0;
3970 for (int i = 0; i < _constants.length(); i++) {
3971 Constant* con = _constants.adr_at(i);
3972
3973 // Align offset for type.
3974 int typesize = type_to_size_in_bytes(con->type());
3975 offset = align_up(offset, typesize);
3976 con->set_offset(offset); // set constant's offset
3977
3978 if (con->type() == T_VOID) {
3979 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3980 offset = offset + typesize * n->outcnt(); // expand jump-table
3981 } else {
3982 offset = offset + typesize;
3983 }
3984 }
3985
3986 // Align size up to the next section start (which is insts; see
3987 // CodeBuffer::align_at_start).
3988 assert(_size == -1, "already set?");
3989 _size = align_up(offset, (int)CodeEntryAlignment);
3990 }
3991
3992 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3993 MacroAssembler _masm(&cb);
3994 for (int i = 0; i < _constants.length(); i++) {
3995 Constant con = _constants.at(i);
3996 address constant_addr = NULL;
3997 switch (con.type()) {
3998 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
3999 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4000 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4001 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4002 case T_OBJECT: {
4003 jobject obj = con.get_jobject();
4004 int oop_index = _masm.oop_recorder()->find_index(obj);
4005 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4006 break;
4007 }
4008 case T_ADDRESS: {
4009 address addr = (address) con.get_jobject();
4010 constant_addr = _masm.address_constant(addr);
4011 break;
4012 }
4013 // We use T_VOID as marker for jump-table entries (labels) which
4014 // need an internal word relocation.
4015 case T_VOID: {
4016 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4017 // Fill the jump-table with a dummy word. The real value is
4018 // filled in later in fill_jump_table.
4019 address dummy = (address) n;
4020 constant_addr = _masm.address_constant(dummy);
4021 // Expand jump-table
4022 for (uint i = 1; i < n->outcnt(); i++) {
4023 address temp_addr = _masm.address_constant(dummy + i);
4024 assert(temp_addr, "consts section too small");
4025 }
4026 break;
4027 }
4028 case T_METADATA: {
4029 Metadata* obj = con.get_metadata();
4030 int metadata_index = _masm.oop_recorder()->find_index(obj);
4031 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4032 break;
4033 }
4034 default: ShouldNotReachHere();
4035 }
4036 assert(constant_addr, "consts section too small");
4037 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4038 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4039 }
4040 }
4041
4042 int Compile::ConstantTable::find_offset(Constant& con) const {
4043 int idx = _constants.find(con);
4044 guarantee(idx != -1, "constant must be in constant table");
4045 int offset = _constants.at(idx).offset();
4046 guarantee(offset != -1, "constant table not emitted yet?");
4047 return offset;
4048 }
4049
4050 void Compile::ConstantTable::add(Constant& con) {
4051 if (con.can_be_reused()) {
4052 int idx = _constants.find(con);
4053 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4054 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4055 return;
4056 }
4057 }
4058 (void) _constants.append(con);
4059 }
4060
4061 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4062 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4063 Constant con(type, value, b->_freq);
4064 add(con);
4065 return con;
4066 }
4067
4068 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4069 Constant con(metadata);
4070 add(con);
4071 return con;
4072 }
4073
4074 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4075 jvalue value;
4076 BasicType type = oper->type()->basic_type();
4077 switch (type) {
4078 case T_LONG: value.j = oper->constantL(); break;
4079 case T_FLOAT: value.f = oper->constantF(); break;
4080 case T_DOUBLE: value.d = oper->constantD(); break;
4081 case T_OBJECT:
4082 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4083 case T_METADATA: return add((Metadata*)oper->constant()); break;
4084 default: guarantee(false, "unhandled type: %s", type2name(type));
4085 }
4086 return add(n, type, value);
4087 }
4088
4089 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4090 jvalue value;
4091 // We can use the node pointer here to identify the right jump-table
4092 // as this method is called from Compile::Fill_buffer right before
4093 // the MachNodes are emitted and the jump-table is filled (means the
4094 // MachNode pointers do not change anymore).
4095 value.l = (jobject) n;
4096 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4097 add(con);
4098 return con;
4099 }
4100
4101 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4102 // If called from Compile::scratch_emit_size do nothing.
4103 if (Compile::current()->in_scratch_emit_size()) return;
4104
4105 assert(labels.is_nonempty(), "must be");
4106 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4107
4108 // Since MachConstantNode::constant_offset() also contains
4109 // table_base_offset() we need to subtract the table_base_offset()
4110 // to get the plain offset into the constant table.
4111 int offset = n->constant_offset() - table_base_offset();
4112
4113 MacroAssembler _masm(&cb);
4114 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4115
4116 for (uint i = 0; i < n->outcnt(); i++) {
4117 address* constant_addr = &jump_table_base[i];
4118 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));
4119 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4120 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4121 }
4122 }
4123
4124 //----------------------------static_subtype_check-----------------------------
4125 // Shortcut important common cases when superklass is exact:
4126 // (0) superklass is java.lang.Object (can occur in reflective code)
4127 // (1) subklass is already limited to a subtype of superklass => always ok
4128 // (2) subklass does not overlap with superklass => always fail
4129 // (3) superklass has NO subtypes and we can check with a simple compare.
4130 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4131 if (StressReflectiveCode) {
4132 return SSC_full_test; // Let caller generate the general case.
4133 }
4134
4135 if (superk == env()->Object_klass()) {
4136 return SSC_always_true; // (0) this test cannot fail
4137 }
4138
4139 ciType* superelem = superk;
4140 if (superelem->is_array_klass())
4141 superelem = superelem->as_array_klass()->base_element_type();
4142
4143 if (!subk->is_interface()) { // cannot trust static interface types yet
4144 if (subk->is_subtype_of(superk)) {
4145 return SSC_always_true; // (1) false path dead; no dynamic test needed
4146 }
4147 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4148 !superk->is_subtype_of(subk)) {
4149 return SSC_always_false;
4150 }
4151 }
4152
4153 // If casting to an instance klass, it must have no subtypes
4154 if (superk->is_interface()) {
4155 // Cannot trust interfaces yet.
4156 // %%% S.B. superk->nof_implementors() == 1
4157 } else if (superelem->is_instance_klass()) {
4158 ciInstanceKlass* ik = superelem->as_instance_klass();
4159 if (!ik->has_subklass() && !ik->is_interface()) {
4160 if (!ik->is_final()) {
4161 // Add a dependency if there is a chance of a later subclass.
4162 dependencies()->assert_leaf_type(ik);
4163 }
4164 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4165 }
4166 } else {
4167 // A primitive array type has no subtypes.
4168 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4169 }
4170
4171 return SSC_full_test;
4172 }
4173
4174 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4175 #ifdef _LP64
4176 // The scaled index operand to AddP must be a clean 64-bit value.
4177 // Java allows a 32-bit int to be incremented to a negative
4178 // value, which appears in a 64-bit register as a large
4179 // positive number. Using that large positive number as an
4180 // operand in pointer arithmetic has bad consequences.
4181 // On the other hand, 32-bit overflow is rare, and the possibility
4182 // can often be excluded, if we annotate the ConvI2L node with
4183 // a type assertion that its value is known to be a small positive
4184 // number. (The prior range check has ensured this.)
4185 // This assertion is used by ConvI2LNode::Ideal.
4186 int index_max = max_jint - 1; // array size is max_jint, index is one less
4187 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4188 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4189 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4190 #endif
4191 return idx;
4192 }
4193
4194 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4195 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4196 if (ctrl != NULL) {
4197 // Express control dependency by a CastII node with a narrow type.
4198 value = new CastIINode(value, itype, false, true /* range check dependency */);
4199 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4200 // node from floating above the range check during loop optimizations. Otherwise, the
4201 // ConvI2L node may be eliminated independently of the range check, causing the data path
4202 // to become TOP while the control path is still there (although it's unreachable).
4203 value->set_req(0, ctrl);
4204 // Save CastII node to remove it after loop optimizations.
4205 phase->C->add_range_check_cast(value);
4206 value = phase->transform(value);
4207 }
4208 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4209 return phase->transform(new ConvI2LNode(value, ltype));
4210 }
4211
4212 // The message about the current inlining is accumulated in
4213 // _print_inlining_stream and transfered into the _print_inlining_list
4214 // once we know whether inlining succeeds or not. For regular
4215 // inlining, messages are appended to the buffer pointed by
4216 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4217 // a new buffer is added after _print_inlining_idx in the list. This
4218 // way we can update the inlining message for late inlining call site
4219 // when the inlining is attempted again.
4220 void Compile::print_inlining_init() {
4221 if (print_inlining() || print_intrinsics()) {
4222 _print_inlining_stream = new stringStream();
4223 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4224 }
4225 }
4226
4227 void Compile::print_inlining_reinit() {
4228 if (print_inlining() || print_intrinsics()) {
4229 // Re allocate buffer when we change ResourceMark
4230 _print_inlining_stream = new stringStream();
4231 }
4232 }
4233
4234 void Compile::print_inlining_reset() {
4235 _print_inlining_stream->reset();
4236 }
4237
4238 void Compile::print_inlining_commit() {
4239 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4240 // Transfer the message from _print_inlining_stream to the current
4241 // _print_inlining_list buffer and clear _print_inlining_stream.
4242 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4243 print_inlining_reset();
4244 }
4245
4246 void Compile::print_inlining_push() {
4247 // Add new buffer to the _print_inlining_list at current position
4248 _print_inlining_idx++;
4249 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4250 }
4251
4252 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4253 return _print_inlining_list->at(_print_inlining_idx);
4254 }
4255
4256 void Compile::print_inlining_update(CallGenerator* cg) {
4257 if (print_inlining() || print_intrinsics()) {
4258 if (!cg->is_late_inline()) {
4259 if (print_inlining_current().cg() != NULL) {
4260 print_inlining_push();
4261 }
4262 print_inlining_commit();
4263 } else {
4264 if (print_inlining_current().cg() != cg &&
4265 (print_inlining_current().cg() != NULL ||
4266 print_inlining_current().ss()->size() != 0)) {
4267 print_inlining_push();
4268 }
4269 print_inlining_commit();
4270 print_inlining_current().set_cg(cg);
4271 }
4272 }
4273 }
4274
4275 void Compile::print_inlining_move_to(CallGenerator* cg) {
4276 // We resume inlining at a late inlining call site. Locate the
4277 // corresponding inlining buffer so that we can update it.
4278 if (print_inlining()) {
4279 for (int i = 0; i < _print_inlining_list->length(); i++) {
4280 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4281 _print_inlining_idx = i;
4282 return;
4283 }
4284 }
4285 ShouldNotReachHere();
4286 }
4287 }
4288
4289 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4290 if (print_inlining()) {
4291 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4292 assert(print_inlining_current().cg() == cg, "wrong entry");
4293 // replace message with new message
4294 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4295 print_inlining_commit();
4296 print_inlining_current().set_cg(cg);
4297 }
4298 }
4299
4300 void Compile::print_inlining_assert_ready() {
4301 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4302 }
4303
4304 void Compile::process_print_inlining() {
4305 bool do_print_inlining = print_inlining() || print_intrinsics();
4306 if (do_print_inlining || log() != NULL) {
4307 // Print inlining message for candidates that we couldn't inline
4308 // for lack of space
4309 for (int i = 0; i < _late_inlines.length(); i++) {
4310 CallGenerator* cg = _late_inlines.at(i);
4311 if (!cg->is_mh_late_inline()) {
4312 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4313 if (do_print_inlining) {
4314 cg->print_inlining_late(msg);
4315 }
4316 log_late_inline_failure(cg, msg);
4317 }
4318 }
4319 }
4320 if (do_print_inlining) {
4321 ResourceMark rm;
4322 stringStream ss;
4323 for (int i = 0; i < _print_inlining_list->length(); i++) {
4324 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4325 }
4326 size_t end = ss.size();
4327 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4328 strncpy(_print_inlining_output, ss.base(), end+1);
4329 _print_inlining_output[end] = 0;
4330 }
4331 }
4332
4333 void Compile::dump_print_inlining() {
4334 if (_print_inlining_output != NULL) {
4335 tty->print_raw(_print_inlining_output);
4336 }
4337 }
4338
4339 void Compile::log_late_inline(CallGenerator* cg) {
4340 if (log() != NULL) {
4341 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4342 cg->unique_id());
4343 JVMState* p = cg->call_node()->jvms();
4344 while (p != NULL) {
4345 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4346 p = p->caller();
4347 }
4348 log()->tail("late_inline");
4349 }
4350 }
4351
4352 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4353 log_late_inline(cg);
4354 if (log() != NULL) {
4355 log()->inline_fail(msg);
4356 }
4357 }
4358
4359 void Compile::log_inline_id(CallGenerator* cg) {
4360 if (log() != NULL) {
4361 // The LogCompilation tool needs a unique way to identify late
4362 // inline call sites. This id must be unique for this call site in
4363 // this compilation. Try to have it unique across compilations as
4364 // well because it can be convenient when grepping through the log
4365 // file.
4366 // Distinguish OSR compilations from others in case CICountOSR is
4367 // on.
4368 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4369 cg->set_unique_id(id);
4370 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4371 }
4372 }
4373
4374 void Compile::log_inline_failure(const char* msg) {
4375 if (C->log() != NULL) {
4376 C->log()->inline_fail(msg);
4377 }
4378 }
4379
4380
4381 // Dump inlining replay data to the stream.
4382 // Don't change thread state and acquire any locks.
4383 void Compile::dump_inline_data(outputStream* out) {
4384 InlineTree* inl_tree = ilt();
4385 if (inl_tree != NULL) {
4386 out->print(" inline %d", inl_tree->count());
4387 inl_tree->dump_replay_data(out);
4388 }
4389 }
4390
4391 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4392 if (n1->Opcode() < n2->Opcode()) return -1;
4393 else if (n1->Opcode() > n2->Opcode()) return 1;
4394
4395 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4396 for (uint i = 1; i < n1->req(); i++) {
4397 if (n1->in(i) < n2->in(i)) return -1;
4398 else if (n1->in(i) > n2->in(i)) return 1;
4399 }
4400
4401 return 0;
4402 }
4403
4404 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4405 Node* n1 = *n1p;
4406 Node* n2 = *n2p;
4407
4408 return cmp_expensive_nodes(n1, n2);
4409 }
4410
4411 void Compile::sort_expensive_nodes() {
4412 if (!expensive_nodes_sorted()) {
4413 _expensive_nodes->sort(cmp_expensive_nodes);
4414 }
4415 }
4416
4417 bool Compile::expensive_nodes_sorted() const {
4418 for (int i = 1; i < _expensive_nodes->length(); i++) {
4419 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4420 return false;
4421 }
4422 }
4423 return true;
4424 }
4425
4426 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4427 if (_expensive_nodes->length() == 0) {
4428 return false;
4429 }
4430
4431 assert(OptimizeExpensiveOps, "optimization off?");
4432
4433 // Take this opportunity to remove dead nodes from the list
4434 int j = 0;
4435 for (int i = 0; i < _expensive_nodes->length(); i++) {
4436 Node* n = _expensive_nodes->at(i);
4437 if (!n->is_unreachable(igvn)) {
4438 assert(n->is_expensive(), "should be expensive");
4439 _expensive_nodes->at_put(j, n);
4440 j++;
4441 }
4442 }
4443 _expensive_nodes->trunc_to(j);
4444
4445 // Then sort the list so that similar nodes are next to each other
4446 // and check for at least two nodes of identical kind with same data
4447 // inputs.
4448 sort_expensive_nodes();
4449
4450 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4451 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4452 return true;
4453 }
4454 }
4455
4456 return false;
4457 }
4458
4459 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4460 if (_expensive_nodes->length() == 0) {
4461 return;
4462 }
4463
4464 assert(OptimizeExpensiveOps, "optimization off?");
4465
4466 // Sort to bring similar nodes next to each other and clear the
4467 // control input of nodes for which there's only a single copy.
4468 sort_expensive_nodes();
4469
4470 int j = 0;
4471 int identical = 0;
4472 int i = 0;
4473 bool modified = false;
4474 for (; i < _expensive_nodes->length()-1; i++) {
4475 assert(j <= i, "can't write beyond current index");
4476 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4477 identical++;
4478 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4479 continue;
4480 }
4481 if (identical > 0) {
4482 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4483 identical = 0;
4484 } else {
4485 Node* n = _expensive_nodes->at(i);
4486 igvn.replace_input_of(n, 0, NULL);
4487 igvn.hash_insert(n);
4488 modified = true;
4489 }
4490 }
4491 if (identical > 0) {
4492 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4493 } else if (_expensive_nodes->length() >= 1) {
4494 Node* n = _expensive_nodes->at(i);
4495 igvn.replace_input_of(n, 0, NULL);
4496 igvn.hash_insert(n);
4497 modified = true;
4498 }
4499 _expensive_nodes->trunc_to(j);
4500 if (modified) {
4501 igvn.optimize();
4502 }
4503 }
4504
4505 void Compile::add_expensive_node(Node * n) {
4506 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4507 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4508 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4509 if (OptimizeExpensiveOps) {
4510 _expensive_nodes->append(n);
4511 } else {
4512 // Clear control input and let IGVN optimize expensive nodes if
4513 // OptimizeExpensiveOps is off.
4514 n->set_req(0, NULL);
4515 }
4516 }
4517
4518 /**
4519 * Remove the speculative part of types and clean up the graph
4520 */
4521 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4522 if (UseTypeSpeculation) {
4523 Unique_Node_List worklist;
4524 worklist.push(root());
4525 int modified = 0;
4526 // Go over all type nodes that carry a speculative type, drop the
4527 // speculative part of the type and enqueue the node for an igvn
4528 // which may optimize it out.
4529 for (uint next = 0; next < worklist.size(); ++next) {
4530 Node *n = worklist.at(next);
4531 if (n->is_Type()) {
4532 TypeNode* tn = n->as_Type();
4533 const Type* t = tn->type();
4534 const Type* t_no_spec = t->remove_speculative();
4535 if (t_no_spec != t) {
4536 bool in_hash = igvn.hash_delete(n);
4537 assert(in_hash, "node should be in igvn hash table");
4538 tn->set_type(t_no_spec);
4539 igvn.hash_insert(n);
4540 igvn._worklist.push(n); // give it a chance to go away
4541 modified++;
4542 }
4543 }
4544 uint max = n->len();
4545 for( uint i = 0; i < max; ++i ) {
4546 Node *m = n->in(i);
4547 if (not_a_node(m)) continue;
4548 worklist.push(m);
4549 }
4550 }
4551 // Drop the speculative part of all types in the igvn's type table
4552 igvn.remove_speculative_types();
4553 if (modified > 0) {
4554 igvn.optimize();
4555 }
4556 #ifdef ASSERT
4557 // Verify that after the IGVN is over no speculative type has resurfaced
4558 worklist.clear();
4559 worklist.push(root());
4560 for (uint next = 0; next < worklist.size(); ++next) {
4561 Node *n = worklist.at(next);
4562 const Type* t = igvn.type_or_null(n);
4563 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4564 if (n->is_Type()) {
4565 t = n->as_Type()->type();
4566 assert(t == t->remove_speculative(), "no more speculative types");
4567 }
4568 uint max = n->len();
4569 for( uint i = 0; i < max; ++i ) {
4570 Node *m = n->in(i);
4571 if (not_a_node(m)) continue;
4572 worklist.push(m);
4573 }
4574 }
4575 igvn.check_no_speculative_types();
4576 #endif
4577 }
4578 }
4579
4580 // Auxiliary method to support randomized stressing/fuzzing.
4581 //
4582 // This method can be called the arbitrary number of times, with current count
4583 // as the argument. The logic allows selecting a single candidate from the
4584 // running list of candidates as follows:
4585 // int count = 0;
4586 // Cand* selected = null;
4587 // while(cand = cand->next()) {
4588 // if (randomized_select(++count)) {
4589 // selected = cand;
4590 // }
4591 // }
4592 //
4593 // Including count equalizes the chances any candidate is "selected".
4594 // This is useful when we don't have the complete list of candidates to choose
4595 // from uniformly. In this case, we need to adjust the randomicity of the
4596 // selection, or else we will end up biasing the selection towards the latter
4597 // candidates.
4598 //
4599 // Quick back-envelope calculation shows that for the list of n candidates
4600 // the equal probability for the candidate to persist as "best" can be
4601 // achieved by replacing it with "next" k-th candidate with the probability
4602 // of 1/k. It can be easily shown that by the end of the run, the
4603 // probability for any candidate is converged to 1/n, thus giving the
4604 // uniform distribution among all the candidates.
4605 //
4606 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4607 #define RANDOMIZED_DOMAIN_POW 29
4608 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4609 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4610 bool Compile::randomized_select(int count) {
4611 assert(count > 0, "only positive");
4612 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4613 }
4614
4615 CloneMap& Compile::clone_map() { return _clone_map; }
4616 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4617
4618 void NodeCloneInfo::dump() const {
4619 tty->print(" {%d:%d} ", idx(), gen());
4620 }
4621
4622 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4623 uint64_t val = value(old->_idx);
4624 NodeCloneInfo cio(val);
4625 assert(val != 0, "old node should be in the map");
4626 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4627 insert(nnn->_idx, cin.get());
4628 #ifndef PRODUCT
4629 if (is_debug()) {
4630 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4631 }
4632 #endif
4633 }
4634
4635 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4636 NodeCloneInfo cio(value(old->_idx));
4637 if (cio.get() == 0) {
4638 cio.set(old->_idx, 0);
4639 insert(old->_idx, cio.get());
4640 #ifndef PRODUCT
4641 if (is_debug()) {
4642 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4643 }
4644 #endif
4645 }
4646 clone(old, nnn, gen);
4647 }
4648
4649 int CloneMap::max_gen() const {
4650 int g = 0;
4651 DictI di(_dict);
4652 for(; di.test(); ++di) {
4653 int t = gen(di._key);
4654 if (g < t) {
4655 g = t;
4656 #ifndef PRODUCT
4657 if (is_debug()) {
4658 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4659 }
4660 #endif
4661 }
4662 }
4663 return g;
4664 }
4665
4666 void CloneMap::dump(node_idx_t key) const {
4667 uint64_t val = value(key);
4668 if (val != 0) {
4669 NodeCloneInfo ni(val);
4670 ni.dump();
4671 }
4672 }