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