1 /*
2 * Copyright (c) 1997, 2017, 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/valuetypenode.hpp"
69 #include "opto/vectornode.hpp"
70 #include "runtime/arguments.hpp"
71 #include "runtime/sharedRuntime.hpp"
72 #include "runtime/signature.hpp"
73 #include "runtime/stubRoutines.hpp"
74 #include "runtime/timer.hpp"
75 #include "utilities/align.hpp"
76 #include "utilities/copy.hpp"
77
78
79 // -------------------- Compile::mach_constant_base_node -----------------------
80 // Constant table base node singleton.
81 MachConstantBaseNode* Compile::mach_constant_base_node() {
82 if (_mach_constant_base_node == NULL) {
83 _mach_constant_base_node = new MachConstantBaseNode();
84 _mach_constant_base_node->add_req(C->root());
85 }
86 return _mach_constant_base_node;
87 }
88
89
90 /// Support for intrinsics.
91
92 // Return the index at which m must be inserted (or already exists).
93 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
94 class IntrinsicDescPair {
95 private:
96 ciMethod* _m;
97 bool _is_virtual;
98 public:
99 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
100 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
101 ciMethod* m= elt->method();
102 ciMethod* key_m = key->_m;
103 if (key_m < m) return -1;
104 else if (key_m > m) return 1;
105 else {
106 bool is_virtual = elt->is_virtual();
107 bool key_virtual = key->_is_virtual;
108 if (key_virtual < is_virtual) return -1;
109 else if (key_virtual > is_virtual) return 1;
110 else return 0;
111 }
112 }
113 };
114 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
115 #ifdef ASSERT
116 for (int i = 1; i < _intrinsics->length(); i++) {
117 CallGenerator* cg1 = _intrinsics->at(i-1);
118 CallGenerator* cg2 = _intrinsics->at(i);
119 assert(cg1->method() != cg2->method()
120 ? cg1->method() < cg2->method()
121 : cg1->is_virtual() < cg2->is_virtual(),
122 "compiler intrinsics list must stay sorted");
123 }
124 #endif
125 IntrinsicDescPair pair(m, is_virtual);
126 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
127 }
128
129 void Compile::register_intrinsic(CallGenerator* cg) {
130 if (_intrinsics == NULL) {
131 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
132 }
133 int len = _intrinsics->length();
134 bool found = false;
135 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
136 assert(!found, "registering twice");
137 _intrinsics->insert_before(index, cg);
138 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
139 }
140
141 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
142 assert(m->is_loaded(), "don't try this on unloaded methods");
143 if (_intrinsics != NULL) {
144 bool found = false;
145 int index = intrinsic_insertion_index(m, is_virtual, found);
146 if (found) {
147 return _intrinsics->at(index);
148 }
149 }
150 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
151 if (m->intrinsic_id() != vmIntrinsics::_none &&
152 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
153 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
154 if (cg != NULL) {
155 // Save it for next time:
156 register_intrinsic(cg);
157 return cg;
158 } else {
159 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
160 }
161 }
162 return NULL;
163 }
164
165 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
166 // in library_call.cpp.
167
168
169 #ifndef PRODUCT
170 // statistics gathering...
171
172 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
174
175 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
176 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
177 int oflags = _intrinsic_hist_flags[id];
178 assert(flags != 0, "what happened?");
179 if (is_virtual) {
180 flags |= _intrinsic_virtual;
181 }
182 bool changed = (flags != oflags);
183 if ((flags & _intrinsic_worked) != 0) {
184 juint count = (_intrinsic_hist_count[id] += 1);
185 if (count == 1) {
186 changed = true; // first time
187 }
188 // increment the overall count also:
189 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
190 }
191 if (changed) {
192 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
193 // Something changed about the intrinsic's virtuality.
194 if ((flags & _intrinsic_virtual) != 0) {
195 // This is the first use of this intrinsic as a virtual call.
196 if (oflags != 0) {
197 // We already saw it as a non-virtual, so note both cases.
198 flags |= _intrinsic_both;
199 }
200 } else if ((oflags & _intrinsic_both) == 0) {
201 // This is the first use of this intrinsic as a non-virtual
202 flags |= _intrinsic_both;
203 }
204 }
205 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
206 }
207 // update the overall flags also:
208 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
209 return changed;
210 }
211
212 static char* format_flags(int flags, char* buf) {
213 buf[0] = 0;
214 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
215 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
216 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
217 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
218 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
219 if (buf[0] == 0) strcat(buf, ",");
220 assert(buf[0] == ',', "must be");
221 return &buf[1];
222 }
223
224 void Compile::print_intrinsic_statistics() {
225 char flagsbuf[100];
226 ttyLocker ttyl;
227 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
228 tty->print_cr("Compiler intrinsic usage:");
229 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
230 if (total == 0) total = 1; // avoid div0 in case of no successes
231 #define PRINT_STAT_LINE(name, c, f) \
232 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
233 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
234 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
235 int flags = _intrinsic_hist_flags[id];
236 juint count = _intrinsic_hist_count[id];
237 if ((flags | count) != 0) {
238 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
239 }
240 }
241 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
242 if (xtty != NULL) xtty->tail("statistics");
243 }
244
245 void Compile::print_statistics() {
246 { ttyLocker ttyl;
247 if (xtty != NULL) xtty->head("statistics type='opto'");
248 Parse::print_statistics();
249 PhaseCCP::print_statistics();
250 PhaseRegAlloc::print_statistics();
251 Scheduling::print_statistics();
252 PhasePeephole::print_statistics();
253 PhaseIdealLoop::print_statistics();
254 if (xtty != NULL) xtty->tail("statistics");
255 }
256 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
257 // put this under its own <statistics> element.
258 print_intrinsic_statistics();
259 }
260 }
261 #endif //PRODUCT
262
263 // Support for bundling info
264 Bundle* Compile::node_bundling(const Node *n) {
265 assert(valid_bundle_info(n), "oob");
266 return &_node_bundling_base[n->_idx];
267 }
268
269 bool Compile::valid_bundle_info(const Node *n) {
270 return (_node_bundling_limit > n->_idx);
271 }
272
273
274 void Compile::gvn_replace_by(Node* n, Node* nn) {
275 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
276 Node* use = n->last_out(i);
277 bool is_in_table = initial_gvn()->hash_delete(use);
278 uint uses_found = 0;
279 for (uint j = 0; j < use->len(); j++) {
280 if (use->in(j) == n) {
281 if (j < use->req())
282 use->set_req(j, nn);
283 else
284 use->set_prec(j, nn);
285 uses_found++;
286 }
287 }
288 if (is_in_table) {
289 // reinsert into table
290 initial_gvn()->hash_find_insert(use);
291 }
292 record_for_igvn(use);
293 i -= uses_found; // we deleted 1 or more copies of this edge
294 }
295 }
296
297
298 static inline bool not_a_node(const Node* n) {
299 if (n == NULL) return true;
300 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
301 if (*(address*)n == badAddress) return true; // kill by Node::destruct
302 return false;
303 }
304
305 // Identify all nodes that are reachable from below, useful.
306 // Use breadth-first pass that records state in a Unique_Node_List,
307 // recursive traversal is slower.
308 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
309 int estimated_worklist_size = live_nodes();
310 useful.map( estimated_worklist_size, NULL ); // preallocate space
311
312 // Initialize worklist
313 if (root() != NULL) { useful.push(root()); }
314 // If 'top' is cached, declare it useful to preserve cached node
315 if( cached_top_node() ) { useful.push(cached_top_node()); }
316
317 // Push all useful nodes onto the list, breadthfirst
318 for( uint next = 0; next < useful.size(); ++next ) {
319 assert( next < unique(), "Unique useful nodes < total nodes");
320 Node *n = useful.at(next);
321 uint max = n->len();
322 for( uint i = 0; i < max; ++i ) {
323 Node *m = n->in(i);
324 if (not_a_node(m)) continue;
325 useful.push(m);
326 }
327 }
328 }
329
330 // Update dead_node_list with any missing dead nodes using useful
331 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
332 void Compile::update_dead_node_list(Unique_Node_List &useful) {
333 uint max_idx = unique();
334 VectorSet& useful_node_set = useful.member_set();
335
336 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
337 // If node with index node_idx is not in useful set,
338 // mark it as dead in dead node list.
339 if (! useful_node_set.test(node_idx) ) {
340 record_dead_node(node_idx);
341 }
342 }
343 }
344
345 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
346 int shift = 0;
347 for (int i = 0; i < inlines->length(); i++) {
348 CallGenerator* cg = inlines->at(i);
349 CallNode* call = cg->call_node();
350 if (shift > 0) {
351 inlines->at_put(i-shift, cg);
352 }
353 if (!useful.member(call)) {
354 shift++;
355 }
356 }
357 inlines->trunc_to(inlines->length()-shift);
358 }
359
360 // Disconnect all useless nodes by disconnecting those at the boundary.
361 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
362 uint next = 0;
363 while (next < useful.size()) {
364 Node *n = useful.at(next++);
365 if (n->is_SafePoint()) {
366 // We're done with a parsing phase. Replaced nodes are not valid
367 // beyond that point.
368 n->as_SafePoint()->delete_replaced_nodes();
369 }
370 // Use raw traversal of out edges since this code removes out edges
371 int max = n->outcnt();
372 for (int j = 0; j < max; ++j) {
373 Node* child = n->raw_out(j);
374 if (! useful.member(child)) {
375 assert(!child->is_top() || child != top(),
376 "If top is cached in Compile object it is in useful list");
377 // Only need to remove this out-edge to the useless node
378 n->raw_del_out(j);
379 --j;
380 --max;
381 }
382 }
383 if (n->outcnt() == 1 && n->has_special_unique_user()) {
384 record_for_igvn(n->unique_out());
385 }
386 }
387 // Remove useless macro and predicate opaq nodes
388 for (int i = C->macro_count()-1; i >= 0; i--) {
389 Node* n = C->macro_node(i);
390 if (!useful.member(n)) {
391 remove_macro_node(n);
392 }
393 }
394 // Remove useless CastII nodes with range check dependency
395 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
396 Node* cast = range_check_cast_node(i);
397 if (!useful.member(cast)) {
398 remove_range_check_cast(cast);
399 }
400 }
401 // Remove useless expensive node
402 for (int i = C->expensive_count()-1; i >= 0; i--) {
403 Node* n = C->expensive_node(i);
404 if (!useful.member(n)) {
405 remove_expensive_node(n);
406 }
407 }
408 for (int i = value_type_ptr_count() - 1; i >= 0; i--) {
409 ValueTypePtrNode* vtptr = value_type_ptr(i);
410 if (!useful.member(vtptr)) {
411 remove_value_type_ptr(vtptr);
412 }
413 }
414 // clean up the late inline lists
415 remove_useless_late_inlines(&_string_late_inlines, useful);
416 remove_useless_late_inlines(&_boxing_late_inlines, useful);
417 remove_useless_late_inlines(&_late_inlines, useful);
418 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
419 }
420
421 //------------------------------frame_size_in_words-----------------------------
422 // frame_slots in units of words
423 int Compile::frame_size_in_words() const {
424 // shift is 0 in LP32 and 1 in LP64
425 const int shift = (LogBytesPerWord - LogBytesPerInt);
426 int words = _frame_slots >> shift;
427 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
428 return words;
429 }
430
431 // To bang the stack of this compiled method we use the stack size
432 // that the interpreter would need in case of a deoptimization. This
433 // removes the need to bang the stack in the deoptimization blob which
434 // in turn simplifies stack overflow handling.
435 int Compile::bang_size_in_bytes() const {
436 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
437 }
438
439 // ============================================================================
440 //------------------------------CompileWrapper---------------------------------
441 class CompileWrapper : public StackObj {
442 Compile *const _compile;
443 public:
444 CompileWrapper(Compile* compile);
445
446 ~CompileWrapper();
447 };
448
449 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
450 // the Compile* pointer is stored in the current ciEnv:
451 ciEnv* env = compile->env();
452 assert(env == ciEnv::current(), "must already be a ciEnv active");
453 assert(env->compiler_data() == NULL, "compile already active?");
454 env->set_compiler_data(compile);
455 assert(compile == Compile::current(), "sanity");
456
457 compile->set_type_dict(NULL);
458 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
459 compile->clone_map().set_clone_idx(0);
460 compile->set_type_hwm(NULL);
461 compile->set_type_last_size(0);
462 compile->set_last_tf(NULL, NULL);
463 compile->set_indexSet_arena(NULL);
464 compile->set_indexSet_free_block_list(NULL);
465 compile->init_type_arena();
466 Type::Initialize(compile);
467 _compile->set_scratch_buffer_blob(NULL);
468 _compile->begin_method();
469 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
470 }
471 CompileWrapper::~CompileWrapper() {
472 _compile->end_method();
473 if (_compile->scratch_buffer_blob() != NULL)
474 BufferBlob::free(_compile->scratch_buffer_blob());
475 _compile->env()->set_compiler_data(NULL);
476 }
477
478
479 //----------------------------print_compile_messages---------------------------
480 void Compile::print_compile_messages() {
481 #ifndef PRODUCT
482 // Check if recompiling
483 if (_subsume_loads == false && PrintOpto) {
484 // Recompiling without allowing machine instructions to subsume loads
485 tty->print_cr("*********************************************************");
486 tty->print_cr("** Bailout: Recompile without subsuming loads **");
487 tty->print_cr("*********************************************************");
488 }
489 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
490 // Recompiling without escape analysis
491 tty->print_cr("*********************************************************");
492 tty->print_cr("** Bailout: Recompile without escape analysis **");
493 tty->print_cr("*********************************************************");
494 }
495 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
496 // Recompiling without boxing elimination
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without boxing elimination **");
499 tty->print_cr("*********************************************************");
500 }
501 if (C->directive()->BreakAtCompileOption) {
502 // Open the debugger when compiling this method.
503 tty->print("### Breaking when compiling: ");
504 method()->print_short_name();
505 tty->cr();
506 BREAKPOINT;
507 }
508
509 if( PrintOpto ) {
510 if (is_osr_compilation()) {
511 tty->print("[OSR]%3d", _compile_id);
512 } else {
513 tty->print("%3d", _compile_id);
514 }
515 }
516 #endif
517 }
518
519
520 //-----------------------init_scratch_buffer_blob------------------------------
521 // Construct a temporary BufferBlob and cache it for this compile.
522 void Compile::init_scratch_buffer_blob(int const_size) {
523 // If there is already a scratch buffer blob allocated and the
524 // constant section is big enough, use it. Otherwise free the
525 // current and allocate a new one.
526 BufferBlob* blob = scratch_buffer_blob();
527 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
528 // Use the current blob.
529 } else {
530 if (blob != NULL) {
531 BufferBlob::free(blob);
532 }
533
534 ResourceMark rm;
535 _scratch_const_size = const_size;
536 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
537 blob = BufferBlob::create("Compile::scratch_buffer", size);
538 // Record the buffer blob for next time.
539 set_scratch_buffer_blob(blob);
540 // Have we run out of code space?
541 if (scratch_buffer_blob() == NULL) {
542 // Let CompilerBroker disable further compilations.
543 record_failure("Not enough space for scratch buffer in CodeCache");
544 return;
545 }
546 }
547
548 // Initialize the relocation buffers
549 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
550 set_scratch_locs_memory(locs_buf);
551 }
552
553
554 //-----------------------scratch_emit_size-------------------------------------
555 // Helper function that computes size by emitting code
556 uint Compile::scratch_emit_size(const Node* n) {
557 // Start scratch_emit_size section.
558 set_in_scratch_emit_size(true);
559
560 // Emit into a trash buffer and count bytes emitted.
561 // This is a pretty expensive way to compute a size,
562 // but it works well enough if seldom used.
563 // All common fixed-size instructions are given a size
564 // method by the AD file.
565 // Note that the scratch buffer blob and locs memory are
566 // allocated at the beginning of the compile task, and
567 // may be shared by several calls to scratch_emit_size.
568 // The allocation of the scratch buffer blob is particularly
569 // expensive, since it has to grab the code cache lock.
570 BufferBlob* blob = this->scratch_buffer_blob();
571 assert(blob != NULL, "Initialize BufferBlob at start");
572 assert(blob->size() > MAX_inst_size, "sanity");
573 relocInfo* locs_buf = scratch_locs_memory();
574 address blob_begin = blob->content_begin();
575 address blob_end = (address)locs_buf;
576 assert(blob->contains(blob_end), "sanity");
577 CodeBuffer buf(blob_begin, blob_end - blob_begin);
578 buf.initialize_consts_size(_scratch_const_size);
579 buf.initialize_stubs_size(MAX_stubs_size);
580 assert(locs_buf != NULL, "sanity");
581 int lsize = MAX_locs_size / 3;
582 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
583 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
584 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
585 // Mark as scratch buffer.
586 buf.consts()->set_scratch_emit();
587 buf.insts()->set_scratch_emit();
588 buf.stubs()->set_scratch_emit();
589
590 // Do the emission.
591
592 Label fakeL; // Fake label for branch instructions.
593 Label* saveL = NULL;
594 uint save_bnum = 0;
595 bool is_branch = n->is_MachBranch();
596 if (is_branch) {
597 MacroAssembler masm(&buf);
598 masm.bind(fakeL);
599 n->as_MachBranch()->save_label(&saveL, &save_bnum);
600 n->as_MachBranch()->label_set(&fakeL, 0);
601 }
602 n->emit(buf, this->regalloc());
603
604 // Emitting into the scratch buffer should not fail
605 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
606
607 if (is_branch) // Restore label.
608 n->as_MachBranch()->label_set(saveL, save_bnum);
609
610 // End scratch_emit_size section.
611 set_in_scratch_emit_size(false);
612
613 return buf.insts_size();
614 }
615
616
617 // ============================================================================
618 //------------------------------Compile standard-------------------------------
619 debug_only( int Compile::_debug_idx = 100000; )
620
621 // Compile a method. entry_bci is -1 for normal compilations and indicates
622 // the continuation bci for on stack replacement.
623
624
625 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
626 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
627 : Phase(Compiler),
628 _env(ci_env),
629 _directive(directive),
630 _log(ci_env->log()),
631 _compile_id(ci_env->compile_id()),
632 _save_argument_registers(false),
633 _stub_name(NULL),
634 _stub_function(NULL),
635 _stub_entry_point(NULL),
636 _method(target),
637 _entry_bci(osr_bci),
638 _initial_gvn(NULL),
639 _for_igvn(NULL),
640 _warm_calls(NULL),
641 _subsume_loads(subsume_loads),
642 _do_escape_analysis(do_escape_analysis),
643 _eliminate_boxing(eliminate_boxing),
644 _failure_reason(NULL),
645 _code_buffer("Compile::Fill_buffer"),
646 _orig_pc_slot(0),
647 _orig_pc_slot_offset_in_bytes(0),
648 _has_method_handle_invokes(false),
649 _mach_constant_base_node(NULL),
650 _node_bundling_limit(0),
651 _node_bundling_base(NULL),
652 _java_calls(0),
653 _inner_loops(0),
654 _scratch_const_size(-1),
655 _in_scratch_emit_size(false),
656 _dead_node_list(comp_arena()),
657 _dead_node_count(0),
658 #ifndef PRODUCT
659 _trace_opto_output(directive->TraceOptoOutputOption),
660 _in_dump_cnt(0),
661 _printer(IdealGraphPrinter::printer()),
662 #endif
663 _congraph(NULL),
664 _comp_arena(mtCompiler),
665 _node_arena(mtCompiler),
666 _old_arena(mtCompiler),
667 _Compile_types(mtCompiler),
668 _replay_inline_data(NULL),
669 _late_inlines(comp_arena(), 2, 0, NULL),
670 _string_late_inlines(comp_arena(), 2, 0, NULL),
671 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
672 _late_inlines_pos(0),
673 _number_of_mh_late_inlines(0),
674 _inlining_progress(false),
675 _inlining_incrementally(false),
676 _print_inlining_list(NULL),
677 _print_inlining_stream(NULL),
678 _print_inlining_idx(0),
679 _print_inlining_output(NULL),
680 _interpreter_frame_size(0),
681 _max_node_limit(MaxNodeLimit),
682 _has_reserved_stack_access(target->has_reserved_stack_access()) {
683 C = this;
684 #ifndef PRODUCT
685 if (_printer != NULL) {
686 _printer->set_compile(this);
687 }
688 #endif
689 CompileWrapper cw(this);
690
691 if (CITimeVerbose) {
692 tty->print(" ");
693 target->holder()->name()->print();
694 tty->print(".");
695 target->print_short_name();
696 tty->print(" ");
697 }
698 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
699 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
700
701 #ifndef PRODUCT
702 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
703 if (!print_opto_assembly) {
704 bool print_assembly = directive->PrintAssemblyOption;
705 if (print_assembly && !Disassembler::can_decode()) {
706 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
707 print_opto_assembly = true;
708 }
709 }
710 set_print_assembly(print_opto_assembly);
711 set_parsed_irreducible_loop(false);
712
713 if (directive->ReplayInlineOption) {
714 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
715 }
716 #endif
717 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
718 set_print_intrinsics(directive->PrintIntrinsicsOption);
719 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
720
721 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
722 // Make sure the method being compiled gets its own MDO,
723 // so we can at least track the decompile_count().
724 // Need MDO to record RTM code generation state.
725 method()->ensure_method_data();
726 }
727
728 Init(::AliasLevel);
729
730
731 print_compile_messages();
732
733 _ilt = InlineTree::build_inline_tree_root();
734
735 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
736 assert(num_alias_types() >= AliasIdxRaw, "");
737
738 #define MINIMUM_NODE_HASH 1023
739 // Node list that Iterative GVN will start with
740 Unique_Node_List for_igvn(comp_arena());
741 set_for_igvn(&for_igvn);
742
743 // GVN that will be run immediately on new nodes
744 uint estimated_size = method()->code_size()*4+64;
745 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
746 PhaseGVN gvn(node_arena(), estimated_size);
747 set_initial_gvn(&gvn);
748
749 print_inlining_init();
750 { // Scope for timing the parser
751 TracePhase tp("parse", &timers[_t_parser]);
752
753 // Put top into the hash table ASAP.
754 initial_gvn()->transform_no_reclaim(top());
755
756 // Set up tf(), start(), and find a CallGenerator.
757 CallGenerator* cg = NULL;
758 if (is_osr_compilation()) {
759 const TypeTuple *domain = StartOSRNode::osr_domain();
760 const TypeTuple *range = TypeTuple::make_range(method()->signature());
761 init_tf(TypeFunc::make(domain, range));
762 StartNode* s = new StartOSRNode(root(), domain);
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 cg = CallGenerator::for_osr(method(), entry_bci());
766 } else {
767 // Normal case.
768 init_tf(TypeFunc::make(method()));
769 StartNode* s = new StartNode(root(), tf()->domain_cc());
770 initial_gvn()->set_type_bottom(s);
771 init_start(s);
772 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
773 // With java.lang.ref.reference.get() we must go through the
774 // intrinsic when G1 is enabled - even when get() is the root
775 // method of the compile - so that, if necessary, the value in
776 // the referent field of the reference object gets recorded by
777 // the pre-barrier code.
778 // Specifically, if G1 is enabled, the value in the referent
779 // field is recorded by the G1 SATB pre barrier. This will
780 // result in the referent being marked live and the reference
781 // object removed from the list of discovered references during
782 // reference processing.
783 cg = find_intrinsic(method(), false);
784 }
785 if (cg == NULL) {
786 float past_uses = method()->interpreter_invocation_count();
787 float expected_uses = past_uses;
788 cg = CallGenerator::for_inline(method(), expected_uses);
789 }
790 }
791 if (failing()) return;
792 if (cg == NULL) {
793 record_method_not_compilable("cannot parse method");
794 return;
795 }
796 JVMState* jvms = build_start_state(start(), tf());
797 if ((jvms = cg->generate(jvms)) == NULL) {
798 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
799 record_method_not_compilable("method parse failed");
800 }
801 return;
802 }
803 GraphKit kit(jvms);
804
805 if (!kit.stopped()) {
806 // Accept return values, and transfer control we know not where.
807 // This is done by a special, unique ReturnNode bound to root.
808 return_values(kit.jvms());
809 }
810
811 if (kit.has_exceptions()) {
812 // Any exceptions that escape from this call must be rethrown
813 // to whatever caller is dynamically above us on the stack.
814 // This is done by a special, unique RethrowNode bound to root.
815 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
816 }
817
818 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
819
820 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
821 inline_string_calls(true);
822 }
823
824 if (failing()) return;
825
826 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
827
828 // Remove clutter produced by parsing.
829 if (!failing()) {
830 ResourceMark rm;
831 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
832 }
833 }
834
835 // Note: Large methods are capped off in do_one_bytecode().
836 if (failing()) return;
837
838 // After parsing, node notes are no longer automagic.
839 // They must be propagated by register_new_node_with_optimizer(),
840 // clone(), or the like.
841 set_default_node_notes(NULL);
842
843 for (;;) {
844 int successes = Inline_Warm();
845 if (failing()) return;
846 if (successes == 0) break;
847 }
848
849 // Drain the list.
850 Finish_Warm();
851 #ifndef PRODUCT
852 if (_printer && _printer->should_print(1)) {
853 _printer->print_inlining();
854 }
855 #endif
856
857 if (failing()) return;
858 NOT_PRODUCT( verify_graph_edges(); )
859
860 // Now optimize
861 Optimize();
862 if (failing()) return;
863 NOT_PRODUCT( verify_graph_edges(); )
864
865 #ifndef PRODUCT
866 if (PrintIdeal) {
867 ttyLocker ttyl; // keep the following output all in one block
868 // This output goes directly to the tty, not the compiler log.
869 // To enable tools to match it up with the compilation activity,
870 // be sure to tag this tty output with the compile ID.
871 if (xtty != NULL) {
872 xtty->head("ideal compile_id='%d'%s", compile_id(),
873 is_osr_compilation() ? " compile_kind='osr'" :
874 "");
875 }
876 root()->dump(9999);
877 if (xtty != NULL) {
878 xtty->tail("ideal");
879 }
880 }
881 #endif
882
883 NOT_PRODUCT( verify_barriers(); )
884
885 // Dump compilation data to replay it.
886 if (directive->DumpReplayOption) {
887 env()->dump_replay_data(_compile_id);
888 }
889 if (directive->DumpInlineOption && (ilt() != NULL)) {
890 env()->dump_inline_data(_compile_id);
891 }
892
893 // Now that we know the size of all the monitors we can add a fixed slot
894 // for the original deopt pc.
895
896 _orig_pc_slot = fixed_slots();
897 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
898 set_fixed_slots(next_slot);
899
900 // Compute when to use implicit null checks. Used by matching trap based
901 // nodes and NullCheck optimization.
902 set_allowed_deopt_reasons();
903
904 // Now generate code
905 Code_Gen();
906 if (failing()) return;
907
908 // Check if we want to skip execution of all compiled code.
909 {
910 #ifndef PRODUCT
911 if (OptoNoExecute) {
912 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
913 return;
914 }
915 #endif
916 TracePhase tp("install_code", &timers[_t_registerMethod]);
917
918 if (is_osr_compilation()) {
919 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
920 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
921 } else {
922 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
923 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
924 }
925
926 env()->register_method(_method, _entry_bci,
927 &_code_offsets,
928 _orig_pc_slot_offset_in_bytes,
929 code_buffer(),
930 frame_size_in_words(), _oop_map_set,
931 &_handler_table, &_inc_table,
932 compiler,
933 has_unsafe_access(),
934 SharedRuntime::is_wide_vector(max_vector_size()),
935 rtm_state()
936 );
937
938 if (log() != NULL) // Print code cache state into compiler log
939 log()->code_cache_state();
940 }
941 }
942
943 //------------------------------Compile----------------------------------------
944 // Compile a runtime stub
945 Compile::Compile( ciEnv* ci_env,
946 TypeFunc_generator generator,
947 address stub_function,
948 const char *stub_name,
949 int is_fancy_jump,
950 bool pass_tls,
951 bool save_arg_registers,
952 bool return_pc,
953 DirectiveSet* directive)
954 : Phase(Compiler),
955 _env(ci_env),
956 _directive(directive),
957 _log(ci_env->log()),
958 _compile_id(0),
959 _save_argument_registers(save_arg_registers),
960 _method(NULL),
961 _stub_name(stub_name),
962 _stub_function(stub_function),
963 _stub_entry_point(NULL),
964 _entry_bci(InvocationEntryBci),
965 _initial_gvn(NULL),
966 _for_igvn(NULL),
967 _warm_calls(NULL),
968 _orig_pc_slot(0),
969 _orig_pc_slot_offset_in_bytes(0),
970 _subsume_loads(true),
971 _do_escape_analysis(false),
972 _eliminate_boxing(false),
973 _failure_reason(NULL),
974 _code_buffer("Compile::Fill_buffer"),
975 _has_method_handle_invokes(false),
976 _mach_constant_base_node(NULL),
977 _node_bundling_limit(0),
978 _node_bundling_base(NULL),
979 _java_calls(0),
980 _inner_loops(0),
981 #ifndef PRODUCT
982 _trace_opto_output(directive->TraceOptoOutputOption),
983 _in_dump_cnt(0),
984 _printer(NULL),
985 #endif
986 _comp_arena(mtCompiler),
987 _node_arena(mtCompiler),
988 _old_arena(mtCompiler),
989 _Compile_types(mtCompiler),
990 _dead_node_list(comp_arena()),
991 _dead_node_count(0),
992 _congraph(NULL),
993 _replay_inline_data(NULL),
994 _number_of_mh_late_inlines(0),
995 _inlining_progress(false),
996 _inlining_incrementally(false),
997 _print_inlining_list(NULL),
998 _print_inlining_stream(NULL),
999 _print_inlining_idx(0),
1000 _print_inlining_output(NULL),
1001 _allowed_reasons(0),
1002 _interpreter_frame_size(0),
1003 _max_node_limit(MaxNodeLimit),
1004 _has_reserved_stack_access(false) {
1005 C = this;
1006
1007 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1008 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1009
1010 #ifndef PRODUCT
1011 set_print_assembly(PrintFrameConverterAssembly);
1012 set_parsed_irreducible_loop(false);
1013 #endif
1014 set_has_irreducible_loop(false); // no loops
1015
1016 CompileWrapper cw(this);
1017 Init(/*AliasLevel=*/ 0);
1018 init_tf((*generator)());
1019
1020 {
1021 // The following is a dummy for the sake of GraphKit::gen_stub
1022 Unique_Node_List for_igvn(comp_arena());
1023 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1024 PhaseGVN gvn(Thread::current()->resource_area(),255);
1025 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1026 gvn.transform_no_reclaim(top());
1027
1028 GraphKit kit;
1029 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1030 }
1031
1032 NOT_PRODUCT( verify_graph_edges(); )
1033 Code_Gen();
1034 if (failing()) return;
1035
1036
1037 // Entry point will be accessed using compile->stub_entry_point();
1038 if (code_buffer() == NULL) {
1039 Matcher::soft_match_failure();
1040 } else {
1041 if (PrintAssembly && (WizardMode || Verbose))
1042 tty->print_cr("### Stub::%s", stub_name);
1043
1044 if (!failing()) {
1045 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1046
1047 // Make the NMethod
1048 // For now we mark the frame as never safe for profile stackwalking
1049 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1050 code_buffer(),
1051 CodeOffsets::frame_never_safe,
1052 // _code_offsets.value(CodeOffsets::Frame_Complete),
1053 frame_size_in_words(),
1054 _oop_map_set,
1055 save_arg_registers);
1056 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1057
1058 _stub_entry_point = rs->entry_point();
1059 }
1060 }
1061 }
1062
1063 //------------------------------Init-------------------------------------------
1064 // Prepare for a single compilation
1065 void Compile::Init(int aliaslevel) {
1066 _unique = 0;
1067 _regalloc = NULL;
1068
1069 _tf = NULL; // filled in later
1070 _top = NULL; // cached later
1071 _matcher = NULL; // filled in later
1072 _cfg = NULL; // filled in later
1073
1074 set_24_bit_selection_and_mode(Use24BitFP, false);
1075
1076 _node_note_array = NULL;
1077 _default_node_notes = NULL;
1078 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1079
1080 _immutable_memory = NULL; // filled in at first inquiry
1081
1082 // Globally visible Nodes
1083 // First set TOP to NULL to give safe behavior during creation of RootNode
1084 set_cached_top_node(NULL);
1085 set_root(new RootNode());
1086 // Now that you have a Root to point to, create the real TOP
1087 set_cached_top_node( new ConNode(Type::TOP) );
1088 set_recent_alloc(NULL, NULL);
1089
1090 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1091 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1092 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1093 env()->set_dependencies(new Dependencies(env()));
1094
1095 _fixed_slots = 0;
1096 set_has_split_ifs(false);
1097 set_has_loops(has_method() && method()->has_loops()); // first approximation
1098 set_has_stringbuilder(false);
1099 set_has_boxed_value(false);
1100 _trap_can_recompile = false; // no traps emitted yet
1101 _major_progress = true; // start out assuming good things will happen
1102 set_has_unsafe_access(false);
1103 set_max_vector_size(0);
1104 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1105 set_decompile_count(0);
1106
1107 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1108 set_num_loop_opts(LoopOptsCount);
1109 set_do_inlining(Inline);
1110 set_max_inline_size(MaxInlineSize);
1111 set_freq_inline_size(FreqInlineSize);
1112 set_do_scheduling(OptoScheduling);
1113 set_do_count_invocations(false);
1114 set_do_method_data_update(false);
1115
1116 set_do_vector_loop(false);
1117
1118 if (AllowVectorizeOnDemand) {
1119 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1120 set_do_vector_loop(true);
1121 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());})
1122 } else if (has_method() && method()->name() != 0 &&
1123 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1124 set_do_vector_loop(true);
1125 }
1126 }
1127 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1128 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());})
1129
1130 set_age_code(has_method() && method()->profile_aging());
1131 set_rtm_state(NoRTM); // No RTM lock eliding by default
1132 _max_node_limit = _directive->MaxNodeLimitOption;
1133
1134 #if INCLUDE_RTM_OPT
1135 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1136 int rtm_state = method()->method_data()->rtm_state();
1137 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1138 // Don't generate RTM lock eliding code.
1139 set_rtm_state(NoRTM);
1140 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1141 // Generate RTM lock eliding code without abort ratio calculation code.
1142 set_rtm_state(UseRTM);
1143 } else if (UseRTMDeopt) {
1144 // Generate RTM lock eliding code and include abort ratio calculation
1145 // code if UseRTMDeopt is on.
1146 set_rtm_state(ProfileRTM);
1147 }
1148 }
1149 #endif
1150 if (debug_info()->recording_non_safepoints()) {
1151 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1152 (comp_arena(), 8, 0, NULL));
1153 set_default_node_notes(Node_Notes::make(this));
1154 }
1155
1156 // // -- Initialize types before each compile --
1157 // // Update cached type information
1158 // if( _method && _method->constants() )
1159 // Type::update_loaded_types(_method, _method->constants());
1160
1161 // Init alias_type map.
1162 if (!_do_escape_analysis && aliaslevel == 3)
1163 aliaslevel = 2; // No unique types without escape analysis
1164 _AliasLevel = aliaslevel;
1165 const int grow_ats = 16;
1166 _max_alias_types = grow_ats;
1167 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1168 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1169 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1170 {
1171 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1172 }
1173 // Initialize the first few types.
1174 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1175 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1176 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1177 _num_alias_types = AliasIdxRaw+1;
1178 // Zero out the alias type cache.
1179 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1180 // A NULL adr_type hits in the cache right away. Preload the right answer.
1181 probe_alias_cache(NULL)->_index = AliasIdxTop;
1182
1183 _intrinsics = NULL;
1184 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1185 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1186 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1187 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1188 _value_type_ptr_nodes = new(comp_arena()) GrowableArray<ValueTypePtrNode*>(comp_arena(), 8, 0, NULL);
1189 register_library_intrinsics();
1190 }
1191
1192 //---------------------------init_start----------------------------------------
1193 // Install the StartNode on this compile object.
1194 void Compile::init_start(StartNode* s) {
1195 if (failing())
1196 return; // already failing
1197 assert(s == start(), "");
1198 }
1199
1200 /**
1201 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1202 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1203 * the ideal graph.
1204 */
1205 StartNode* Compile::start() const {
1206 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1207 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1208 Node* start = root()->fast_out(i);
1209 if (start->is_Start()) {
1210 return start->as_Start();
1211 }
1212 }
1213 fatal("Did not find Start node!");
1214 return NULL;
1215 }
1216
1217 //-------------------------------immutable_memory-------------------------------------
1218 // Access immutable memory
1219 Node* Compile::immutable_memory() {
1220 if (_immutable_memory != NULL) {
1221 return _immutable_memory;
1222 }
1223 StartNode* s = start();
1224 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1225 Node *p = s->fast_out(i);
1226 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1227 _immutable_memory = p;
1228 return _immutable_memory;
1229 }
1230 }
1231 ShouldNotReachHere();
1232 return NULL;
1233 }
1234
1235 //----------------------set_cached_top_node------------------------------------
1236 // Install the cached top node, and make sure Node::is_top works correctly.
1237 void Compile::set_cached_top_node(Node* tn) {
1238 if (tn != NULL) verify_top(tn);
1239 Node* old_top = _top;
1240 _top = tn;
1241 // Calling Node::setup_is_top allows the nodes the chance to adjust
1242 // their _out arrays.
1243 if (_top != NULL) _top->setup_is_top();
1244 if (old_top != NULL) old_top->setup_is_top();
1245 assert(_top == NULL || top()->is_top(), "");
1246 }
1247
1248 #ifdef ASSERT
1249 uint Compile::count_live_nodes_by_graph_walk() {
1250 Unique_Node_List useful(comp_arena());
1251 // Get useful node list by walking the graph.
1252 identify_useful_nodes(useful);
1253 return useful.size();
1254 }
1255
1256 void Compile::print_missing_nodes() {
1257
1258 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1259 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1260 return;
1261 }
1262
1263 // This is an expensive function. It is executed only when the user
1264 // specifies VerifyIdealNodeCount option or otherwise knows the
1265 // additional work that needs to be done to identify reachable nodes
1266 // by walking the flow graph and find the missing ones using
1267 // _dead_node_list.
1268
1269 Unique_Node_List useful(comp_arena());
1270 // Get useful node list by walking the graph.
1271 identify_useful_nodes(useful);
1272
1273 uint l_nodes = C->live_nodes();
1274 uint l_nodes_by_walk = useful.size();
1275
1276 if (l_nodes != l_nodes_by_walk) {
1277 if (_log != NULL) {
1278 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1279 _log->stamp();
1280 _log->end_head();
1281 }
1282 VectorSet& useful_member_set = useful.member_set();
1283 int last_idx = l_nodes_by_walk;
1284 for (int i = 0; i < last_idx; i++) {
1285 if (useful_member_set.test(i)) {
1286 if (_dead_node_list.test(i)) {
1287 if (_log != NULL) {
1288 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1289 }
1290 if (PrintIdealNodeCount) {
1291 // Print the log message to tty
1292 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1293 useful.at(i)->dump();
1294 }
1295 }
1296 }
1297 else if (! _dead_node_list.test(i)) {
1298 if (_log != NULL) {
1299 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1300 }
1301 if (PrintIdealNodeCount) {
1302 // Print the log message to tty
1303 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1304 }
1305 }
1306 }
1307 if (_log != NULL) {
1308 _log->tail("mismatched_nodes");
1309 }
1310 }
1311 }
1312 void Compile::record_modified_node(Node* n) {
1313 if (_modified_nodes != NULL && !_inlining_incrementally &&
1314 n->outcnt() != 0 && !n->is_Con()) {
1315 _modified_nodes->push(n);
1316 }
1317 }
1318
1319 void Compile::remove_modified_node(Node* n) {
1320 if (_modified_nodes != NULL) {
1321 _modified_nodes->remove(n);
1322 }
1323 }
1324 #endif
1325
1326 #ifndef PRODUCT
1327 void Compile::verify_top(Node* tn) const {
1328 if (tn != NULL) {
1329 assert(tn->is_Con(), "top node must be a constant");
1330 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1331 assert(tn->in(0) != NULL, "must have live top node");
1332 }
1333 }
1334 #endif
1335
1336
1337 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1338
1339 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1340 guarantee(arr != NULL, "");
1341 int num_blocks = arr->length();
1342 if (grow_by < num_blocks) grow_by = num_blocks;
1343 int num_notes = grow_by * _node_notes_block_size;
1344 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1345 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1346 while (num_notes > 0) {
1347 arr->append(notes);
1348 notes += _node_notes_block_size;
1349 num_notes -= _node_notes_block_size;
1350 }
1351 assert(num_notes == 0, "exact multiple, please");
1352 }
1353
1354 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1355 if (source == NULL || dest == NULL) return false;
1356
1357 if (dest->is_Con())
1358 return false; // Do not push debug info onto constants.
1359
1360 #ifdef ASSERT
1361 // Leave a bread crumb trail pointing to the original node:
1362 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1363 dest->set_debug_orig(source);
1364 }
1365 #endif
1366
1367 if (node_note_array() == NULL)
1368 return false; // Not collecting any notes now.
1369
1370 // This is a copy onto a pre-existing node, which may already have notes.
1371 // If both nodes have notes, do not overwrite any pre-existing notes.
1372 Node_Notes* source_notes = node_notes_at(source->_idx);
1373 if (source_notes == NULL || source_notes->is_clear()) return false;
1374 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1375 if (dest_notes == NULL || dest_notes->is_clear()) {
1376 return set_node_notes_at(dest->_idx, source_notes);
1377 }
1378
1379 Node_Notes merged_notes = (*source_notes);
1380 // The order of operations here ensures that dest notes will win...
1381 merged_notes.update_from(dest_notes);
1382 return set_node_notes_at(dest->_idx, &merged_notes);
1383 }
1384
1385
1386 //--------------------------allow_range_check_smearing-------------------------
1387 // Gating condition for coalescing similar range checks.
1388 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1389 // single covering check that is at least as strong as any of them.
1390 // If the optimization succeeds, the simplified (strengthened) range check
1391 // will always succeed. If it fails, we will deopt, and then give up
1392 // on the optimization.
1393 bool Compile::allow_range_check_smearing() const {
1394 // If this method has already thrown a range-check,
1395 // assume it was because we already tried range smearing
1396 // and it failed.
1397 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1398 return !already_trapped;
1399 }
1400
1401
1402 //------------------------------flatten_alias_type-----------------------------
1403 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1404 int offset = tj->offset();
1405 TypePtr::PTR ptr = tj->ptr();
1406
1407 // Known instance (scalarizable allocation) alias only with itself.
1408 bool is_known_inst = tj->isa_oopptr() != NULL &&
1409 tj->is_oopptr()->is_known_instance();
1410
1411 // Process weird unsafe references.
1412 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1413 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1414 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1415 tj = TypeOopPtr::BOTTOM;
1416 ptr = tj->ptr();
1417 offset = tj->offset();
1418 }
1419
1420 // Array pointers need some flattening
1421 const TypeAryPtr *ta = tj->isa_aryptr();
1422 if (ta && ta->is_stable()) {
1423 // Erase stability property for alias analysis.
1424 tj = ta = ta->cast_to_stable(false);
1425 }
1426 if( ta && is_known_inst ) {
1427 if ( offset != Type::OffsetBot &&
1428 offset > arrayOopDesc::length_offset_in_bytes() ) {
1429 offset = Type::OffsetBot; // Flatten constant access into array body only
1430 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1431 }
1432 } else if( ta && _AliasLevel >= 2 ) {
1433 // For arrays indexed by constant indices, we flatten the alias
1434 // space to include all of the array body. Only the header, klass
1435 // and array length can be accessed un-aliased.
1436 // For flattened value type array, each field has its own slice so
1437 // we must include the field offset.
1438 if( offset != Type::OffsetBot ) {
1439 if( ta->const_oop() ) { // MethodData* or Method*
1440 offset = Type::OffsetBot; // Flatten constant access into array body
1441 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1442 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1443 // range is OK as-is.
1444 tj = ta = TypeAryPtr::RANGE;
1445 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1446 tj = TypeInstPtr::KLASS; // all klass loads look alike
1447 ta = TypeAryPtr::RANGE; // generic ignored junk
1448 ptr = TypePtr::BotPTR;
1449 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1450 tj = TypeInstPtr::MARK;
1451 ta = TypeAryPtr::RANGE; // generic ignored junk
1452 ptr = TypePtr::BotPTR;
1453 } else { // Random constant offset into array body
1454 offset = Type::OffsetBot; // Flatten constant access into array body
1455 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1456 }
1457 }
1458 // Arrays of fixed size alias with arrays of unknown size.
1459 if (ta->size() != TypeInt::POS) {
1460 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1461 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1462 }
1463 // Arrays of known objects become arrays of unknown objects.
1464 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1465 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1466 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1467 }
1468 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1469 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1470 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1471 }
1472 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1473 // cannot be distinguished by bytecode alone.
1474 if (ta->elem() == TypeInt::BOOL) {
1475 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1476 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1477 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1478 }
1479 // During the 2nd round of IterGVN, NotNull castings are removed.
1480 // Make sure the Bottom and NotNull variants alias the same.
1481 // Also, make sure exact and non-exact variants alias the same.
1482 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1483 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1484 }
1485 }
1486
1487 // Oop pointers need some flattening
1488 const TypeInstPtr *to = tj->isa_instptr();
1489 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1490 ciInstanceKlass *k = to->klass()->as_instance_klass();
1491 if( ptr == TypePtr::Constant ) {
1492 if (to->klass() != ciEnv::current()->Class_klass() ||
1493 offset < k->size_helper() * wordSize) {
1494 // No constant oop pointers (such as Strings); they alias with
1495 // unknown strings.
1496 assert(!is_known_inst, "not scalarizable allocation");
1497 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1498 }
1499 } else if( is_known_inst ) {
1500 tj = to; // Keep NotNull and klass_is_exact for instance type
1501 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1502 // During the 2nd round of IterGVN, NotNull castings are removed.
1503 // Make sure the Bottom and NotNull variants alias the same.
1504 // Also, make sure exact and non-exact variants alias the same.
1505 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1506 }
1507 if (to->speculative() != NULL) {
1508 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id());
1509 }
1510 // Canonicalize the holder of this field
1511 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1512 // First handle header references such as a LoadKlassNode, even if the
1513 // object's klass is unloaded at compile time (4965979).
1514 if (!is_known_inst) { // Do it only for non-instance types
1515 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1516 }
1517 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1518 // Static fields are in the space above the normal instance
1519 // fields in the java.lang.Class instance.
1520 if (to->klass() != ciEnv::current()->Class_klass()) {
1521 to = NULL;
1522 tj = TypeOopPtr::BOTTOM;
1523 offset = tj->offset();
1524 }
1525 } else {
1526 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1527 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1528 if( is_known_inst ) {
1529 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id());
1530 } else {
1531 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1532 }
1533 }
1534 }
1535 }
1536
1537 // Klass pointers to object array klasses need some flattening
1538 const TypeKlassPtr *tk = tj->isa_klassptr();
1539 if( tk ) {
1540 // If we are referencing a field within a Klass, we need
1541 // to assume the worst case of an Object. Both exact and
1542 // inexact types must flatten to the same alias class so
1543 // use NotNull as the PTR.
1544 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1545
1546 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1547 TypeKlassPtr::OBJECT->klass(),
1548 Type::Offset(offset));
1549 }
1550
1551 ciKlass* klass = tk->klass();
1552 if (klass != NULL && klass->is_obj_array_klass()) {
1553 ciKlass* k = TypeAryPtr::OOPS->klass();
1554 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1555 k = TypeInstPtr::BOTTOM->klass();
1556 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1557 }
1558
1559 // Check for precise loads from the primary supertype array and force them
1560 // to the supertype cache alias index. Check for generic array loads from
1561 // the primary supertype array and also force them to the supertype cache
1562 // alias index. Since the same load can reach both, we need to merge
1563 // these 2 disparate memories into the same alias class. Since the
1564 // primary supertype array is read-only, there's no chance of confusion
1565 // where we bypass an array load and an array store.
1566 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1567 if (offset == Type::OffsetBot ||
1568 (offset >= primary_supers_offset &&
1569 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1570 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1571 offset = in_bytes(Klass::secondary_super_cache_offset());
1572 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1573 }
1574 }
1575
1576 // Flatten all Raw pointers together.
1577 if (tj->base() == Type::RawPtr)
1578 tj = TypeRawPtr::BOTTOM;
1579
1580 if (tj->base() == Type::AnyPtr)
1581 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1582
1583 // Flatten all to bottom for now
1584 switch( _AliasLevel ) {
1585 case 0:
1586 tj = TypePtr::BOTTOM;
1587 break;
1588 case 1: // Flatten to: oop, static, field or array
1589 switch (tj->base()) {
1590 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1591 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1592 case Type::AryPtr: // do not distinguish arrays at all
1593 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1594 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1595 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1596 default: ShouldNotReachHere();
1597 }
1598 break;
1599 case 2: // No collapsing at level 2; keep all splits
1600 case 3: // No collapsing at level 3; keep all splits
1601 break;
1602 default:
1603 Unimplemented();
1604 }
1605
1606 offset = tj->offset();
1607 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1608
1609 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1610 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1611 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1612 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1613 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1614 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1615 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1616 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1617 assert( tj->ptr() != TypePtr::TopPTR &&
1618 tj->ptr() != TypePtr::AnyNull &&
1619 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1620 // assert( tj->ptr() != TypePtr::Constant ||
1621 // tj->base() == Type::RawPtr ||
1622 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1623
1624 return tj;
1625 }
1626
1627 void Compile::AliasType::Init(int i, const TypePtr* at) {
1628 _index = i;
1629 _adr_type = at;
1630 _field = NULL;
1631 _element = NULL;
1632 _is_rewritable = true; // default
1633 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1634 if (atoop != NULL && atoop->is_known_instance()) {
1635 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1636 _general_index = Compile::current()->get_alias_index(gt);
1637 } else {
1638 _general_index = 0;
1639 }
1640 }
1641
1642 BasicType Compile::AliasType::basic_type() const {
1643 if (element() != NULL) {
1644 const Type* element = adr_type()->is_aryptr()->elem();
1645 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1646 } if (field() != NULL) {
1647 return field()->layout_type();
1648 } else {
1649 return T_ILLEGAL; // unknown
1650 }
1651 }
1652
1653 //---------------------------------print_on------------------------------------
1654 #ifndef PRODUCT
1655 void Compile::AliasType::print_on(outputStream* st) {
1656 if (index() < 10)
1657 st->print("@ <%d> ", index());
1658 else st->print("@ <%d>", index());
1659 st->print(is_rewritable() ? " " : " RO");
1660 int offset = adr_type()->offset();
1661 if (offset == Type::OffsetBot)
1662 st->print(" +any");
1663 else st->print(" +%-3d", offset);
1664 st->print(" in ");
1665 adr_type()->dump_on(st);
1666 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1667 if (field() != NULL && tjp) {
1668 if (tjp->klass() != field()->holder() ||
1669 tjp->offset() != field()->offset_in_bytes()) {
1670 st->print(" != ");
1671 field()->print();
1672 st->print(" ***");
1673 }
1674 }
1675 }
1676
1677 void print_alias_types() {
1678 Compile* C = Compile::current();
1679 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1680 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1681 C->alias_type(idx)->print_on(tty);
1682 tty->cr();
1683 }
1684 }
1685 #endif
1686
1687
1688 //----------------------------probe_alias_cache--------------------------------
1689 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1690 intptr_t key = (intptr_t) adr_type;
1691 key ^= key >> logAliasCacheSize;
1692 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1693 }
1694
1695
1696 //-----------------------------grow_alias_types--------------------------------
1697 void Compile::grow_alias_types() {
1698 const int old_ats = _max_alias_types; // how many before?
1699 const int new_ats = old_ats; // how many more?
1700 const int grow_ats = old_ats+new_ats; // how many now?
1701 _max_alias_types = grow_ats;
1702 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1703 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1704 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1705 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1706 }
1707
1708
1709 //--------------------------------find_alias_type------------------------------
1710 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1711 if (_AliasLevel == 0)
1712 return alias_type(AliasIdxBot);
1713
1714 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1715 if (ace->_adr_type == adr_type) {
1716 return alias_type(ace->_index);
1717 }
1718
1719 // Handle special cases.
1720 if (adr_type == NULL) return alias_type(AliasIdxTop);
1721 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1722
1723 // Do it the slow way.
1724 const TypePtr* flat = flatten_alias_type(adr_type);
1725
1726 #ifdef ASSERT
1727 {
1728 ResourceMark rm;
1729 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1730 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1731 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1732 Type::str(adr_type));
1733 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1734 const TypeOopPtr* foop = flat->is_oopptr();
1735 // Scalarizable allocations have exact klass always.
1736 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1737 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1738 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1739 Type::str(foop), Type::str(xoop));
1740 }
1741 }
1742 #endif
1743
1744 int idx = AliasIdxTop;
1745 for (int i = 0; i < num_alias_types(); i++) {
1746 if (alias_type(i)->adr_type() == flat) {
1747 idx = i;
1748 break;
1749 }
1750 }
1751
1752 if (idx == AliasIdxTop) {
1753 if (no_create) return NULL;
1754 // Grow the array if necessary.
1755 if (_num_alias_types == _max_alias_types) grow_alias_types();
1756 // Add a new alias type.
1757 idx = _num_alias_types++;
1758 _alias_types[idx]->Init(idx, flat);
1759 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1760 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1761 if (flat->isa_instptr()) {
1762 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1763 && flat->is_instptr()->klass() == env()->Class_klass())
1764 alias_type(idx)->set_rewritable(false);
1765 }
1766 if (flat->isa_aryptr()) {
1767 #ifdef ASSERT
1768 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1769 // (T_BYTE has the weakest alignment and size restrictions...)
1770 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1771 #endif
1772 if (flat->offset() == TypePtr::OffsetBot) {
1773 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1774 }
1775 }
1776 if (flat->isa_klassptr()) {
1777 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1778 alias_type(idx)->set_rewritable(false);
1779 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1780 alias_type(idx)->set_rewritable(false);
1781 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1782 alias_type(idx)->set_rewritable(false);
1783 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1784 alias_type(idx)->set_rewritable(false);
1785 }
1786 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1787 // but the base pointer type is not distinctive enough to identify
1788 // references into JavaThread.)
1789
1790 // Check for final fields.
1791 const TypeInstPtr* tinst = flat->isa_instptr();
1792 const TypeValueTypePtr* vtptr = flat->isa_valuetypeptr();
1793 ciField* field = NULL;
1794 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1795 if (tinst->const_oop() != NULL &&
1796 tinst->klass() == ciEnv::current()->Class_klass() &&
1797 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1798 // static field
1799 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1800 field = k->get_field_by_offset(tinst->offset(), true);
1801 } else {
1802 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1803 field = k->get_field_by_offset(tinst->offset(), false);
1804 }
1805 } else if (vtptr) {
1806 // Value type field
1807 ciValueKlass* vk = vtptr->klass()->as_value_klass();
1808 field = vk->get_field_by_offset(vtptr->offset(), false);
1809 }
1810 assert(field == NULL ||
1811 original_field == NULL ||
1812 (field->holder() == original_field->holder() &&
1813 field->offset() == original_field->offset() &&
1814 field->is_static() == original_field->is_static()), "wrong field?");
1815 // Set field() and is_rewritable() attributes.
1816 if (field != NULL) alias_type(idx)->set_field(field);
1817 }
1818
1819 // Fill the cache for next time.
1820 ace->_adr_type = adr_type;
1821 ace->_index = idx;
1822 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1823
1824 // Might as well try to fill the cache for the flattened version, too.
1825 AliasCacheEntry* face = probe_alias_cache(flat);
1826 if (face->_adr_type == NULL) {
1827 face->_adr_type = flat;
1828 face->_index = idx;
1829 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1830 }
1831
1832 return alias_type(idx);
1833 }
1834
1835
1836 Compile::AliasType* Compile::alias_type(ciField* field) {
1837 const TypeOopPtr* t;
1838 if (field->is_static())
1839 t = TypeInstPtr::make(field->holder()->java_mirror());
1840 else
1841 t = TypeOopPtr::make_from_klass_raw(field->holder());
1842 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1843 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1844 return atp;
1845 }
1846
1847
1848 //------------------------------have_alias_type--------------------------------
1849 bool Compile::have_alias_type(const TypePtr* adr_type) {
1850 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1851 if (ace->_adr_type == adr_type) {
1852 return true;
1853 }
1854
1855 // Handle special cases.
1856 if (adr_type == NULL) return true;
1857 if (adr_type == TypePtr::BOTTOM) return true;
1858
1859 return find_alias_type(adr_type, true, NULL) != NULL;
1860 }
1861
1862 //-----------------------------must_alias--------------------------------------
1863 // True if all values of the given address type are in the given alias category.
1864 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1865 if (alias_idx == AliasIdxBot) return true; // the universal category
1866 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1867 if (alias_idx == AliasIdxTop) return false; // the empty category
1868 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1869
1870 // the only remaining possible overlap is identity
1871 int adr_idx = get_alias_index(adr_type);
1872 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1873 assert(adr_idx == alias_idx ||
1874 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1875 && adr_type != TypeOopPtr::BOTTOM),
1876 "should not be testing for overlap with an unsafe pointer");
1877 return adr_idx == alias_idx;
1878 }
1879
1880 //------------------------------can_alias--------------------------------------
1881 // True if any values of the given address type are in the given alias category.
1882 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1883 if (alias_idx == AliasIdxTop) return false; // the empty category
1884 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1885 if (alias_idx == AliasIdxBot) return true; // the universal category
1886 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1887
1888 // the only remaining possible overlap is identity
1889 int adr_idx = get_alias_index(adr_type);
1890 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1891 return adr_idx == alias_idx;
1892 }
1893
1894
1895
1896 //---------------------------pop_warm_call-------------------------------------
1897 WarmCallInfo* Compile::pop_warm_call() {
1898 WarmCallInfo* wci = _warm_calls;
1899 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1900 return wci;
1901 }
1902
1903 //----------------------------Inline_Warm--------------------------------------
1904 int Compile::Inline_Warm() {
1905 // If there is room, try to inline some more warm call sites.
1906 // %%% Do a graph index compaction pass when we think we're out of space?
1907 if (!InlineWarmCalls) return 0;
1908
1909 int calls_made_hot = 0;
1910 int room_to_grow = NodeCountInliningCutoff - unique();
1911 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1912 int amount_grown = 0;
1913 WarmCallInfo* call;
1914 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1915 int est_size = (int)call->size();
1916 if (est_size > (room_to_grow - amount_grown)) {
1917 // This one won't fit anyway. Get rid of it.
1918 call->make_cold();
1919 continue;
1920 }
1921 call->make_hot();
1922 calls_made_hot++;
1923 amount_grown += est_size;
1924 amount_to_grow -= est_size;
1925 }
1926
1927 if (calls_made_hot > 0) set_major_progress();
1928 return calls_made_hot;
1929 }
1930
1931
1932 //----------------------------Finish_Warm--------------------------------------
1933 void Compile::Finish_Warm() {
1934 if (!InlineWarmCalls) return;
1935 if (failing()) return;
1936 if (warm_calls() == NULL) return;
1937
1938 // Clean up loose ends, if we are out of space for inlining.
1939 WarmCallInfo* call;
1940 while ((call = pop_warm_call()) != NULL) {
1941 call->make_cold();
1942 }
1943 }
1944
1945 //---------------------cleanup_loop_predicates-----------------------
1946 // Remove the opaque nodes that protect the predicates so that all unused
1947 // checks and uncommon_traps will be eliminated from the ideal graph
1948 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1949 if (predicate_count()==0) return;
1950 for (int i = predicate_count(); i > 0; i--) {
1951 Node * n = predicate_opaque1_node(i-1);
1952 assert(n->Opcode() == Op_Opaque1, "must be");
1953 igvn.replace_node(n, n->in(1));
1954 }
1955 assert(predicate_count()==0, "should be clean!");
1956 }
1957
1958 void Compile::add_range_check_cast(Node* n) {
1959 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1960 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1961 _range_check_casts->append(n);
1962 }
1963
1964 // Remove all range check dependent CastIINodes.
1965 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1966 for (int i = range_check_cast_count(); i > 0; i--) {
1967 Node* cast = range_check_cast_node(i-1);
1968 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1969 igvn.replace_node(cast, cast->in(1));
1970 }
1971 assert(range_check_cast_count() == 0, "should be empty");
1972 }
1973
1974 void Compile::add_value_type_ptr(ValueTypePtrNode* n) {
1975 assert(can_add_value_type_ptr(), "too late");
1976 assert(!_value_type_ptr_nodes->contains(n), "duplicate entry");
1977 _value_type_ptr_nodes->append(n);
1978 }
1979
1980 void Compile::process_value_type_ptr_nodes(PhaseIterGVN &igvn) {
1981 for (int i = value_type_ptr_count(); i > 0; i--) {
1982 ValueTypePtrNode* vtptr = value_type_ptr(i-1);
1983 // once all inlining is over otherwise debug info can get
1984 // inconsistent
1985 vtptr->make_scalar_in_safepoints(igvn.C->root(), &igvn);
1986 igvn.replace_node(vtptr, vtptr->get_oop());
1987 }
1988 assert(value_type_ptr_count() == 0, "should be empty");
1989 _value_type_ptr_nodes = NULL;
1990 igvn.optimize();
1991 }
1992
1993 // StringOpts and late inlining of string methods
1994 void Compile::inline_string_calls(bool parse_time) {
1995 {
1996 // remove useless nodes to make the usage analysis simpler
1997 ResourceMark rm;
1998 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1999 }
2000
2001 {
2002 ResourceMark rm;
2003 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2004 PhaseStringOpts pso(initial_gvn(), for_igvn());
2005 print_method(PHASE_AFTER_STRINGOPTS, 3);
2006 }
2007
2008 // now inline anything that we skipped the first time around
2009 if (!parse_time) {
2010 _late_inlines_pos = _late_inlines.length();
2011 }
2012
2013 while (_string_late_inlines.length() > 0) {
2014 CallGenerator* cg = _string_late_inlines.pop();
2015 cg->do_late_inline();
2016 if (failing()) return;
2017 }
2018 _string_late_inlines.trunc_to(0);
2019 }
2020
2021 // Late inlining of boxing methods
2022 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2023 if (_boxing_late_inlines.length() > 0) {
2024 assert(has_boxed_value(), "inconsistent");
2025
2026 PhaseGVN* gvn = initial_gvn();
2027 set_inlining_incrementally(true);
2028
2029 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2030 for_igvn()->clear();
2031 gvn->replace_with(&igvn);
2032
2033 _late_inlines_pos = _late_inlines.length();
2034
2035 while (_boxing_late_inlines.length() > 0) {
2036 CallGenerator* cg = _boxing_late_inlines.pop();
2037 cg->do_late_inline();
2038 if (failing()) return;
2039 }
2040 _boxing_late_inlines.trunc_to(0);
2041
2042 {
2043 ResourceMark rm;
2044 PhaseRemoveUseless pru(gvn, for_igvn());
2045 }
2046
2047 igvn = PhaseIterGVN(gvn);
2048 igvn.optimize();
2049
2050 set_inlining_progress(false);
2051 set_inlining_incrementally(false);
2052 }
2053 }
2054
2055 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2056 assert(IncrementalInline, "incremental inlining should be on");
2057 PhaseGVN* gvn = initial_gvn();
2058
2059 set_inlining_progress(false);
2060 for_igvn()->clear();
2061 gvn->replace_with(&igvn);
2062
2063 {
2064 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2065 int i = 0;
2066 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2067 CallGenerator* cg = _late_inlines.at(i);
2068 _late_inlines_pos = i+1;
2069 cg->do_late_inline();
2070 if (failing()) return;
2071 }
2072 int j = 0;
2073 for (; i < _late_inlines.length(); i++, j++) {
2074 _late_inlines.at_put(j, _late_inlines.at(i));
2075 }
2076 _late_inlines.trunc_to(j);
2077 }
2078
2079 {
2080 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2081 ResourceMark rm;
2082 PhaseRemoveUseless pru(gvn, for_igvn());
2083 }
2084
2085 {
2086 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2087 igvn = PhaseIterGVN(gvn);
2088 }
2089 }
2090
2091 // Perform incremental inlining until bound on number of live nodes is reached
2092 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2093 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2094
2095 PhaseGVN* gvn = initial_gvn();
2096
2097 set_inlining_incrementally(true);
2098 set_inlining_progress(true);
2099 uint low_live_nodes = 0;
2100
2101 while(inlining_progress() && _late_inlines.length() > 0) {
2102
2103 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2104 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2105 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2106 // PhaseIdealLoop is expensive so we only try it once we are
2107 // out of live nodes and we only try it again if the previous
2108 // helped got the number of nodes down significantly
2109 PhaseIdealLoop ideal_loop( igvn, false, true );
2110 if (failing()) return;
2111 low_live_nodes = live_nodes();
2112 _major_progress = true;
2113 }
2114
2115 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2116 break;
2117 }
2118 }
2119
2120 inline_incrementally_one(igvn);
2121
2122 if (failing()) return;
2123
2124 {
2125 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2126 igvn.optimize();
2127 }
2128
2129 if (failing()) return;
2130 }
2131
2132 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2133
2134 if (_string_late_inlines.length() > 0) {
2135 assert(has_stringbuilder(), "inconsistent");
2136 for_igvn()->clear();
2137 initial_gvn()->replace_with(&igvn);
2138
2139 inline_string_calls(false);
2140
2141 if (failing()) return;
2142
2143 {
2144 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2145 ResourceMark rm;
2146 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2147 }
2148
2149 {
2150 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2151 igvn = PhaseIterGVN(gvn);
2152 igvn.optimize();
2153 }
2154 }
2155
2156 set_inlining_incrementally(false);
2157 }
2158
2159
2160 //------------------------------Optimize---------------------------------------
2161 // Given a graph, optimize it.
2162 void Compile::Optimize() {
2163 TracePhase tp("optimizer", &timers[_t_optimizer]);
2164
2165 #ifndef PRODUCT
2166 if (_directive->BreakAtCompileOption) {
2167 BREAKPOINT;
2168 }
2169
2170 #endif
2171
2172 ResourceMark rm;
2173 int loop_opts_cnt;
2174
2175 print_inlining_reinit();
2176
2177 NOT_PRODUCT( verify_graph_edges(); )
2178
2179 print_method(PHASE_AFTER_PARSING);
2180
2181 {
2182 // Iterative Global Value Numbering, including ideal transforms
2183 // Initialize IterGVN with types and values from parse-time GVN
2184 PhaseIterGVN igvn(initial_gvn());
2185 #ifdef ASSERT
2186 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2187 #endif
2188 {
2189 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2190 igvn.optimize();
2191 }
2192
2193 print_method(PHASE_ITER_GVN1, 2);
2194
2195 if (failing()) return;
2196
2197 inline_incrementally(igvn);
2198
2199 print_method(PHASE_INCREMENTAL_INLINE, 2);
2200
2201 if (failing()) return;
2202
2203 if (eliminate_boxing()) {
2204 // Inline valueOf() methods now.
2205 inline_boxing_calls(igvn);
2206
2207 if (AlwaysIncrementalInline) {
2208 inline_incrementally(igvn);
2209 }
2210
2211 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2212
2213 if (failing()) return;
2214 }
2215
2216 // Remove the speculative part of types and clean up the graph from
2217 // the extra CastPP nodes whose only purpose is to carry them. Do
2218 // that early so that optimizations are not disrupted by the extra
2219 // CastPP nodes.
2220 remove_speculative_types(igvn);
2221
2222 // No more new expensive nodes will be added to the list from here
2223 // so keep only the actual candidates for optimizations.
2224 cleanup_expensive_nodes(igvn);
2225
2226 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2227 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2228 initial_gvn()->replace_with(&igvn);
2229 for_igvn()->clear();
2230 Unique_Node_List new_worklist(C->comp_arena());
2231 {
2232 ResourceMark rm;
2233 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2234 }
2235 set_for_igvn(&new_worklist);
2236 igvn = PhaseIterGVN(initial_gvn());
2237 igvn.optimize();
2238 }
2239
2240 if (value_type_ptr_count() > 0) {
2241 process_value_type_ptr_nodes(igvn);
2242 }
2243
2244 // Perform escape analysis
2245 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2246 if (has_loops()) {
2247 // Cleanup graph (remove dead nodes).
2248 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2249 PhaseIdealLoop ideal_loop( igvn, false, true );
2250 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2251 if (failing()) return;
2252 }
2253 ConnectionGraph::do_analysis(this, &igvn);
2254
2255 if (failing()) return;
2256
2257 // Optimize out fields loads from scalar replaceable allocations.
2258 igvn.optimize();
2259 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2260
2261 if (failing()) return;
2262
2263 if (congraph() != NULL && macro_count() > 0) {
2264 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2265 PhaseMacroExpand mexp(igvn);
2266 mexp.eliminate_macro_nodes();
2267 igvn.set_delay_transform(false);
2268
2269 igvn.optimize();
2270 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2271
2272 if (failing()) return;
2273 }
2274 }
2275
2276 // Loop transforms on the ideal graph. Range Check Elimination,
2277 // peeling, unrolling, etc.
2278
2279 // Set loop opts counter
2280 loop_opts_cnt = num_loop_opts();
2281 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2282 {
2283 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2284 PhaseIdealLoop ideal_loop( igvn, true );
2285 loop_opts_cnt--;
2286 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2287 if (failing()) return;
2288 }
2289 // Loop opts pass if partial peeling occurred in previous pass
2290 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2291 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2292 PhaseIdealLoop ideal_loop( igvn, false );
2293 loop_opts_cnt--;
2294 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2295 if (failing()) return;
2296 }
2297 // Loop opts pass for loop-unrolling before CCP
2298 if(major_progress() && (loop_opts_cnt > 0)) {
2299 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2300 PhaseIdealLoop ideal_loop( igvn, false );
2301 loop_opts_cnt--;
2302 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2303 }
2304 if (!failing()) {
2305 // Verify that last round of loop opts produced a valid graph
2306 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2307 PhaseIdealLoop::verify(igvn);
2308 }
2309 }
2310 if (failing()) return;
2311
2312 // Conditional Constant Propagation;
2313 PhaseCCP ccp( &igvn );
2314 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2315 {
2316 TracePhase tp("ccp", &timers[_t_ccp]);
2317 ccp.do_transform();
2318 }
2319 print_method(PHASE_CPP1, 2);
2320
2321 assert( true, "Break here to ccp.dump_old2new_map()");
2322
2323 // Iterative Global Value Numbering, including ideal transforms
2324 {
2325 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2326 igvn = ccp;
2327 igvn.optimize();
2328 }
2329
2330 print_method(PHASE_ITER_GVN2, 2);
2331
2332 if (failing()) return;
2333
2334 // Loop transforms on the ideal graph. Range Check Elimination,
2335 // peeling, unrolling, etc.
2336 if(loop_opts_cnt > 0) {
2337 debug_only( int cnt = 0; );
2338 while(major_progress() && (loop_opts_cnt > 0)) {
2339 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2340 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2341 PhaseIdealLoop ideal_loop( igvn, true);
2342 loop_opts_cnt--;
2343 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2344 if (failing()) return;
2345 }
2346 }
2347 // Ensure that major progress is now clear
2348 C->clear_major_progress();
2349
2350 {
2351 // Verify that all previous optimizations produced a valid graph
2352 // at least to this point, even if no loop optimizations were done.
2353 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2354 PhaseIdealLoop::verify(igvn);
2355 }
2356
2357 if (range_check_cast_count() > 0) {
2358 // No more loop optimizations. Remove all range check dependent CastIINodes.
2359 C->remove_range_check_casts(igvn);
2360 igvn.optimize();
2361 }
2362
2363 {
2364 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2365 PhaseMacroExpand mex(igvn);
2366 if (mex.expand_macro_nodes()) {
2367 assert(failing(), "must bail out w/ explicit message");
2368 return;
2369 }
2370 }
2371
2372 DEBUG_ONLY( _modified_nodes = NULL; )
2373 } // (End scope of igvn; run destructor if necessary for asserts.)
2374
2375 process_print_inlining();
2376 // A method with only infinite loops has no edges entering loops from root
2377 {
2378 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2379 if (final_graph_reshaping()) {
2380 assert(failing(), "must bail out w/ explicit message");
2381 return;
2382 }
2383 }
2384
2385 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2386 }
2387
2388 // Fixme remove
2389 static void check_for_value_node(Node &n, void* C) {
2390 if (n.is_ValueType()) {
2391 #ifdef ASSERT
2392 ((Compile*)C)->method()->print_short_name();
2393 tty->print_cr("");
2394 n.dump(-1);
2395 assert(false, "Unable to match ValueTypeNode");
2396 #endif
2397 ((Compile*)C)->record_failure("Unable to match ValueTypeNode");
2398 }
2399 }
2400
2401 //------------------------------Code_Gen---------------------------------------
2402 // Given a graph, generate code for it
2403 void Compile::Code_Gen() {
2404 // FIXME remove
2405 root()->walk(Node::nop, check_for_value_node, this);
2406
2407 if (failing()) {
2408 return;
2409 }
2410
2411 // Perform instruction selection. You might think we could reclaim Matcher
2412 // memory PDQ, but actually the Matcher is used in generating spill code.
2413 // Internals of the Matcher (including some VectorSets) must remain live
2414 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2415 // set a bit in reclaimed memory.
2416
2417 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2418 // nodes. Mapping is only valid at the root of each matched subtree.
2419 NOT_PRODUCT( verify_graph_edges(); )
2420
2421 Matcher matcher;
2422 _matcher = &matcher;
2423 {
2424 TracePhase tp("matcher", &timers[_t_matcher]);
2425 matcher.match();
2426 }
2427 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2428 // nodes. Mapping is only valid at the root of each matched subtree.
2429 NOT_PRODUCT( verify_graph_edges(); )
2430
2431 // If you have too many nodes, or if matching has failed, bail out
2432 check_node_count(0, "out of nodes matching instructions");
2433 if (failing()) {
2434 return;
2435 }
2436
2437 // Build a proper-looking CFG
2438 PhaseCFG cfg(node_arena(), root(), matcher);
2439 _cfg = &cfg;
2440 {
2441 TracePhase tp("scheduler", &timers[_t_scheduler]);
2442 bool success = cfg.do_global_code_motion();
2443 if (!success) {
2444 return;
2445 }
2446
2447 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2448 NOT_PRODUCT( verify_graph_edges(); )
2449 debug_only( cfg.verify(); )
2450 }
2451
2452 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2453 _regalloc = ®alloc;
2454 {
2455 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2456 // Perform register allocation. After Chaitin, use-def chains are
2457 // no longer accurate (at spill code) and so must be ignored.
2458 // Node->LRG->reg mappings are still accurate.
2459 _regalloc->Register_Allocate();
2460
2461 // Bail out if the allocator builds too many nodes
2462 if (failing()) {
2463 return;
2464 }
2465 }
2466
2467 // Prior to register allocation we kept empty basic blocks in case the
2468 // the allocator needed a place to spill. After register allocation we
2469 // are not adding any new instructions. If any basic block is empty, we
2470 // can now safely remove it.
2471 {
2472 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2473 cfg.remove_empty_blocks();
2474 if (do_freq_based_layout()) {
2475 PhaseBlockLayout layout(cfg);
2476 } else {
2477 cfg.set_loop_alignment();
2478 }
2479 cfg.fixup_flow();
2480 }
2481
2482 // Apply peephole optimizations
2483 if( OptoPeephole ) {
2484 TracePhase tp("peephole", &timers[_t_peephole]);
2485 PhasePeephole peep( _regalloc, cfg);
2486 peep.do_transform();
2487 }
2488
2489 // Do late expand if CPU requires this.
2490 if (Matcher::require_postalloc_expand) {
2491 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2492 cfg.postalloc_expand(_regalloc);
2493 }
2494
2495 // Convert Nodes to instruction bits in a buffer
2496 {
2497 TraceTime tp("output", &timers[_t_output], CITime);
2498 Output();
2499 }
2500
2501 print_method(PHASE_FINAL_CODE);
2502
2503 // He's dead, Jim.
2504 _cfg = (PhaseCFG*)0xdeadbeef;
2505 _regalloc = (PhaseChaitin*)0xdeadbeef;
2506 }
2507
2508
2509 //------------------------------dump_asm---------------------------------------
2510 // Dump formatted assembly
2511 #ifndef PRODUCT
2512 void Compile::dump_asm(int *pcs, uint pc_limit) {
2513 bool cut_short = false;
2514 tty->print_cr("#");
2515 tty->print("# "); _tf->dump(); tty->cr();
2516 tty->print_cr("#");
2517
2518 // For all blocks
2519 int pc = 0x0; // Program counter
2520 char starts_bundle = ' ';
2521 _regalloc->dump_frame();
2522
2523 Node *n = NULL;
2524 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2525 if (VMThread::should_terminate()) {
2526 cut_short = true;
2527 break;
2528 }
2529 Block* block = _cfg->get_block(i);
2530 if (block->is_connector() && !Verbose) {
2531 continue;
2532 }
2533 n = block->head();
2534 if (pcs && n->_idx < pc_limit) {
2535 tty->print("%3.3x ", pcs[n->_idx]);
2536 } else {
2537 tty->print(" ");
2538 }
2539 block->dump_head(_cfg);
2540 if (block->is_connector()) {
2541 tty->print_cr(" # Empty connector block");
2542 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2543 tty->print_cr(" # Block is sole successor of call");
2544 }
2545
2546 // For all instructions
2547 Node *delay = NULL;
2548 for (uint j = 0; j < block->number_of_nodes(); j++) {
2549 if (VMThread::should_terminate()) {
2550 cut_short = true;
2551 break;
2552 }
2553 n = block->get_node(j);
2554 if (valid_bundle_info(n)) {
2555 Bundle* bundle = node_bundling(n);
2556 if (bundle->used_in_unconditional_delay()) {
2557 delay = n;
2558 continue;
2559 }
2560 if (bundle->starts_bundle()) {
2561 starts_bundle = '+';
2562 }
2563 }
2564
2565 if (WizardMode) {
2566 n->dump();
2567 }
2568
2569 if( !n->is_Region() && // Dont print in the Assembly
2570 !n->is_Phi() && // a few noisely useless nodes
2571 !n->is_Proj() &&
2572 !n->is_MachTemp() &&
2573 !n->is_SafePointScalarObject() &&
2574 !n->is_Catch() && // Would be nice to print exception table targets
2575 !n->is_MergeMem() && // Not very interesting
2576 !n->is_top() && // Debug info table constants
2577 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2578 ) {
2579 if (pcs && n->_idx < pc_limit)
2580 tty->print("%3.3x", pcs[n->_idx]);
2581 else
2582 tty->print(" ");
2583 tty->print(" %c ", starts_bundle);
2584 starts_bundle = ' ';
2585 tty->print("\t");
2586 n->format(_regalloc, tty);
2587 tty->cr();
2588 }
2589
2590 // If we have an instruction with a delay slot, and have seen a delay,
2591 // then back up and print it
2592 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2593 assert(delay != NULL, "no unconditional delay instruction");
2594 if (WizardMode) delay->dump();
2595
2596 if (node_bundling(delay)->starts_bundle())
2597 starts_bundle = '+';
2598 if (pcs && n->_idx < pc_limit)
2599 tty->print("%3.3x", pcs[n->_idx]);
2600 else
2601 tty->print(" ");
2602 tty->print(" %c ", starts_bundle);
2603 starts_bundle = ' ';
2604 tty->print("\t");
2605 delay->format(_regalloc, tty);
2606 tty->cr();
2607 delay = NULL;
2608 }
2609
2610 // Dump the exception table as well
2611 if( n->is_Catch() && (Verbose || WizardMode) ) {
2612 // Print the exception table for this offset
2613 _handler_table.print_subtable_for(pc);
2614 }
2615 }
2616
2617 if (pcs && n->_idx < pc_limit)
2618 tty->print_cr("%3.3x", pcs[n->_idx]);
2619 else
2620 tty->cr();
2621
2622 assert(cut_short || delay == NULL, "no unconditional delay branch");
2623
2624 } // End of per-block dump
2625 tty->cr();
2626
2627 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2628 }
2629 #endif
2630
2631 //------------------------------Final_Reshape_Counts---------------------------
2632 // This class defines counters to help identify when a method
2633 // may/must be executed using hardware with only 24-bit precision.
2634 struct Final_Reshape_Counts : public StackObj {
2635 int _call_count; // count non-inlined 'common' calls
2636 int _float_count; // count float ops requiring 24-bit precision
2637 int _double_count; // count double ops requiring more precision
2638 int _java_call_count; // count non-inlined 'java' calls
2639 int _inner_loop_count; // count loops which need alignment
2640 VectorSet _visited; // Visitation flags
2641 Node_List _tests; // Set of IfNodes & PCTableNodes
2642
2643 Final_Reshape_Counts() :
2644 _call_count(0), _float_count(0), _double_count(0),
2645 _java_call_count(0), _inner_loop_count(0),
2646 _visited( Thread::current()->resource_area() ) { }
2647
2648 void inc_call_count () { _call_count ++; }
2649 void inc_float_count () { _float_count ++; }
2650 void inc_double_count() { _double_count++; }
2651 void inc_java_call_count() { _java_call_count++; }
2652 void inc_inner_loop_count() { _inner_loop_count++; }
2653
2654 int get_call_count () const { return _call_count ; }
2655 int get_float_count () const { return _float_count ; }
2656 int get_double_count() const { return _double_count; }
2657 int get_java_call_count() const { return _java_call_count; }
2658 int get_inner_loop_count() const { return _inner_loop_count; }
2659 };
2660
2661 #ifdef ASSERT
2662 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2663 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2664 // Make sure the offset goes inside the instance layout.
2665 return k->contains_field_offset(tp->offset());
2666 // Note that OffsetBot and OffsetTop are very negative.
2667 }
2668 #endif
2669
2670 // Eliminate trivially redundant StoreCMs and accumulate their
2671 // precedence edges.
2672 void Compile::eliminate_redundant_card_marks(Node* n) {
2673 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2674 if (n->in(MemNode::Address)->outcnt() > 1) {
2675 // There are multiple users of the same address so it might be
2676 // possible to eliminate some of the StoreCMs
2677 Node* mem = n->in(MemNode::Memory);
2678 Node* adr = n->in(MemNode::Address);
2679 Node* val = n->in(MemNode::ValueIn);
2680 Node* prev = n;
2681 bool done = false;
2682 // Walk the chain of StoreCMs eliminating ones that match. As
2683 // long as it's a chain of single users then the optimization is
2684 // safe. Eliminating partially redundant StoreCMs would require
2685 // cloning copies down the other paths.
2686 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2687 if (adr == mem->in(MemNode::Address) &&
2688 val == mem->in(MemNode::ValueIn)) {
2689 // redundant StoreCM
2690 if (mem->req() > MemNode::OopStore) {
2691 // Hasn't been processed by this code yet.
2692 n->add_prec(mem->in(MemNode::OopStore));
2693 } else {
2694 // Already converted to precedence edge
2695 for (uint i = mem->req(); i < mem->len(); i++) {
2696 // Accumulate any precedence edges
2697 if (mem->in(i) != NULL) {
2698 n->add_prec(mem->in(i));
2699 }
2700 }
2701 // Everything above this point has been processed.
2702 done = true;
2703 }
2704 // Eliminate the previous StoreCM
2705 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2706 assert(mem->outcnt() == 0, "should be dead");
2707 mem->disconnect_inputs(NULL, this);
2708 } else {
2709 prev = mem;
2710 }
2711 mem = prev->in(MemNode::Memory);
2712 }
2713 }
2714 }
2715
2716
2717 //------------------------------final_graph_reshaping_impl----------------------
2718 // Implement items 1-5 from final_graph_reshaping below.
2719 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2720
2721 if ( n->outcnt() == 0 ) return; // dead node
2722 uint nop = n->Opcode();
2723
2724 // Check for 2-input instruction with "last use" on right input.
2725 // Swap to left input. Implements item (2).
2726 if( n->req() == 3 && // two-input instruction
2727 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2728 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2729 n->in(2)->outcnt() == 1 &&// right use IS a last use
2730 !n->in(2)->is_Con() ) { // right use is not a constant
2731 // Check for commutative opcode
2732 switch( nop ) {
2733 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2734 case Op_MaxI: case Op_MinI:
2735 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2736 case Op_AndL: case Op_XorL: case Op_OrL:
2737 case Op_AndI: case Op_XorI: case Op_OrI: {
2738 // Move "last use" input to left by swapping inputs
2739 n->swap_edges(1, 2);
2740 break;
2741 }
2742 default:
2743 break;
2744 }
2745 }
2746
2747 #ifdef ASSERT
2748 if( n->is_Mem() ) {
2749 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2750 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2751 // oop will be recorded in oop map if load crosses safepoint
2752 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2753 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2754 "raw memory operations should have control edge");
2755 }
2756 #endif
2757 // Count FPU ops and common calls, implements item (3)
2758 switch( nop ) {
2759 // Count all float operations that may use FPU
2760 case Op_AddF:
2761 case Op_SubF:
2762 case Op_MulF:
2763 case Op_DivF:
2764 case Op_NegF:
2765 case Op_ModF:
2766 case Op_ConvI2F:
2767 case Op_ConF:
2768 case Op_CmpF:
2769 case Op_CmpF3:
2770 // case Op_ConvL2F: // longs are split into 32-bit halves
2771 frc.inc_float_count();
2772 break;
2773
2774 case Op_ConvF2D:
2775 case Op_ConvD2F:
2776 frc.inc_float_count();
2777 frc.inc_double_count();
2778 break;
2779
2780 // Count all double operations that may use FPU
2781 case Op_AddD:
2782 case Op_SubD:
2783 case Op_MulD:
2784 case Op_DivD:
2785 case Op_NegD:
2786 case Op_ModD:
2787 case Op_ConvI2D:
2788 case Op_ConvD2I:
2789 // case Op_ConvL2D: // handled by leaf call
2790 // case Op_ConvD2L: // handled by leaf call
2791 case Op_ConD:
2792 case Op_CmpD:
2793 case Op_CmpD3:
2794 frc.inc_double_count();
2795 break;
2796 case Op_Opaque1: // Remove Opaque Nodes before matching
2797 case Op_Opaque2: // Remove Opaque Nodes before matching
2798 case Op_Opaque3:
2799 n->subsume_by(n->in(1), this);
2800 break;
2801 case Op_CallStaticJava:
2802 case Op_CallJava:
2803 case Op_CallDynamicJava:
2804 frc.inc_java_call_count(); // Count java call site;
2805 case Op_CallRuntime:
2806 case Op_CallLeaf:
2807 case Op_CallLeafNoFP: {
2808 assert( n->is_Call(), "" );
2809 CallNode *call = n->as_Call();
2810 // Count call sites where the FP mode bit would have to be flipped.
2811 // Do not count uncommon runtime calls:
2812 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2813 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2814 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2815 frc.inc_call_count(); // Count the call site
2816 } else { // See if uncommon argument is shared
2817 Node *n = call->in(TypeFunc::Parms);
2818 int nop = n->Opcode();
2819 // Clone shared simple arguments to uncommon calls, item (1).
2820 if( n->outcnt() > 1 &&
2821 !n->is_Proj() &&
2822 nop != Op_CreateEx &&
2823 nop != Op_CheckCastPP &&
2824 nop != Op_DecodeN &&
2825 nop != Op_DecodeNKlass &&
2826 !n->is_Mem() ) {
2827 Node *x = n->clone();
2828 call->set_req( TypeFunc::Parms, x );
2829 }
2830 }
2831 break;
2832 }
2833
2834 case Op_StoreD:
2835 case Op_LoadD:
2836 case Op_LoadD_unaligned:
2837 frc.inc_double_count();
2838 goto handle_mem;
2839 case Op_StoreF:
2840 case Op_LoadF:
2841 frc.inc_float_count();
2842 goto handle_mem;
2843
2844 case Op_StoreCM:
2845 {
2846 // Convert OopStore dependence into precedence edge
2847 Node* prec = n->in(MemNode::OopStore);
2848 n->del_req(MemNode::OopStore);
2849 n->add_prec(prec);
2850 eliminate_redundant_card_marks(n);
2851 }
2852
2853 // fall through
2854
2855 case Op_StoreB:
2856 case Op_StoreC:
2857 case Op_StorePConditional:
2858 case Op_StoreI:
2859 case Op_StoreL:
2860 case Op_StoreIConditional:
2861 case Op_StoreLConditional:
2862 case Op_CompareAndSwapB:
2863 case Op_CompareAndSwapS:
2864 case Op_CompareAndSwapI:
2865 case Op_CompareAndSwapL:
2866 case Op_CompareAndSwapP:
2867 case Op_CompareAndSwapN:
2868 case Op_WeakCompareAndSwapB:
2869 case Op_WeakCompareAndSwapS:
2870 case Op_WeakCompareAndSwapI:
2871 case Op_WeakCompareAndSwapL:
2872 case Op_WeakCompareAndSwapP:
2873 case Op_WeakCompareAndSwapN:
2874 case Op_CompareAndExchangeB:
2875 case Op_CompareAndExchangeS:
2876 case Op_CompareAndExchangeI:
2877 case Op_CompareAndExchangeL:
2878 case Op_CompareAndExchangeP:
2879 case Op_CompareAndExchangeN:
2880 case Op_GetAndAddS:
2881 case Op_GetAndAddB:
2882 case Op_GetAndAddI:
2883 case Op_GetAndAddL:
2884 case Op_GetAndSetS:
2885 case Op_GetAndSetB:
2886 case Op_GetAndSetI:
2887 case Op_GetAndSetL:
2888 case Op_GetAndSetP:
2889 case Op_GetAndSetN:
2890 case Op_StoreP:
2891 case Op_StoreN:
2892 case Op_StoreNKlass:
2893 case Op_LoadB:
2894 case Op_LoadUB:
2895 case Op_LoadUS:
2896 case Op_LoadI:
2897 case Op_LoadKlass:
2898 case Op_LoadNKlass:
2899 case Op_LoadL:
2900 case Op_LoadL_unaligned:
2901 case Op_LoadPLocked:
2902 case Op_LoadP:
2903 case Op_LoadN:
2904 case Op_LoadRange:
2905 case Op_LoadS: {
2906 handle_mem:
2907 #ifdef ASSERT
2908 if( VerifyOptoOopOffsets ) {
2909 assert( n->is_Mem(), "" );
2910 MemNode *mem = (MemNode*)n;
2911 // Check to see if address types have grounded out somehow.
2912 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2913 assert( !tp || oop_offset_is_sane(tp), "" );
2914 }
2915 #endif
2916 break;
2917 }
2918
2919 case Op_AddP: { // Assert sane base pointers
2920 Node *addp = n->in(AddPNode::Address);
2921 assert( !addp->is_AddP() ||
2922 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2923 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2924 "Base pointers must match (addp %u)", addp->_idx );
2925 #ifdef _LP64
2926 if ((UseCompressedOops || UseCompressedClassPointers) &&
2927 addp->Opcode() == Op_ConP &&
2928 addp == n->in(AddPNode::Base) &&
2929 n->in(AddPNode::Offset)->is_Con()) {
2930 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2931 // on the platform and on the compressed oops mode.
2932 // Use addressing with narrow klass to load with offset on x86.
2933 // Some platforms can use the constant pool to load ConP.
2934 // Do this transformation here since IGVN will convert ConN back to ConP.
2935 const Type* t = addp->bottom_type();
2936 bool is_oop = t->isa_oopptr() != NULL;
2937 bool is_klass = t->isa_klassptr() != NULL;
2938
2939 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2940 (is_klass && Matcher::const_klass_prefer_decode())) {
2941 Node* nn = NULL;
2942
2943 int op = is_oop ? Op_ConN : Op_ConNKlass;
2944
2945 // Look for existing ConN node of the same exact type.
2946 Node* r = root();
2947 uint cnt = r->outcnt();
2948 for (uint i = 0; i < cnt; i++) {
2949 Node* m = r->raw_out(i);
2950 if (m!= NULL && m->Opcode() == op &&
2951 m->bottom_type()->make_ptr() == t) {
2952 nn = m;
2953 break;
2954 }
2955 }
2956 if (nn != NULL) {
2957 // Decode a narrow oop to match address
2958 // [R12 + narrow_oop_reg<<3 + offset]
2959 if (is_oop) {
2960 nn = new DecodeNNode(nn, t);
2961 } else {
2962 nn = new DecodeNKlassNode(nn, t);
2963 }
2964 // Check for succeeding AddP which uses the same Base.
2965 // Otherwise we will run into the assertion above when visiting that guy.
2966 for (uint i = 0; i < n->outcnt(); ++i) {
2967 Node *out_i = n->raw_out(i);
2968 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
2969 out_i->set_req(AddPNode::Base, nn);
2970 #ifdef ASSERT
2971 for (uint j = 0; j < out_i->outcnt(); ++j) {
2972 Node *out_j = out_i->raw_out(j);
2973 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
2974 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
2975 }
2976 #endif
2977 }
2978 }
2979 n->set_req(AddPNode::Base, nn);
2980 n->set_req(AddPNode::Address, nn);
2981 if (addp->outcnt() == 0) {
2982 addp->disconnect_inputs(NULL, this);
2983 }
2984 }
2985 }
2986 }
2987 #endif
2988 // platform dependent reshaping of the address expression
2989 reshape_address(n->as_AddP());
2990 break;
2991 }
2992
2993 case Op_CastPP: {
2994 // Remove CastPP nodes to gain more freedom during scheduling but
2995 // keep the dependency they encode as control or precedence edges
2996 // (if control is set already) on memory operations. Some CastPP
2997 // nodes don't have a control (don't carry a dependency): skip
2998 // those.
2999 if (n->in(0) != NULL) {
3000 ResourceMark rm;
3001 Unique_Node_List wq;
3002 wq.push(n);
3003 for (uint next = 0; next < wq.size(); ++next) {
3004 Node *m = wq.at(next);
3005 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3006 Node* use = m->fast_out(i);
3007 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3008 use->ensure_control_or_add_prec(n->in(0));
3009 } else {
3010 switch(use->Opcode()) {
3011 case Op_AddP:
3012 case Op_DecodeN:
3013 case Op_DecodeNKlass:
3014 case Op_CheckCastPP:
3015 case Op_CastPP:
3016 wq.push(use);
3017 break;
3018 }
3019 }
3020 }
3021 }
3022 }
3023 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3024 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3025 Node* in1 = n->in(1);
3026 const Type* t = n->bottom_type();
3027 Node* new_in1 = in1->clone();
3028 new_in1->as_DecodeN()->set_type(t);
3029
3030 if (!Matcher::narrow_oop_use_complex_address()) {
3031 //
3032 // x86, ARM and friends can handle 2 adds in addressing mode
3033 // and Matcher can fold a DecodeN node into address by using
3034 // a narrow oop directly and do implicit NULL check in address:
3035 //
3036 // [R12 + narrow_oop_reg<<3 + offset]
3037 // NullCheck narrow_oop_reg
3038 //
3039 // On other platforms (Sparc) we have to keep new DecodeN node and
3040 // use it to do implicit NULL check in address:
3041 //
3042 // decode_not_null narrow_oop_reg, base_reg
3043 // [base_reg + offset]
3044 // NullCheck base_reg
3045 //
3046 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3047 // to keep the information to which NULL check the new DecodeN node
3048 // corresponds to use it as value in implicit_null_check().
3049 //
3050 new_in1->set_req(0, n->in(0));
3051 }
3052
3053 n->subsume_by(new_in1, this);
3054 if (in1->outcnt() == 0) {
3055 in1->disconnect_inputs(NULL, this);
3056 }
3057 } else {
3058 n->subsume_by(n->in(1), this);
3059 if (n->outcnt() == 0) {
3060 n->disconnect_inputs(NULL, this);
3061 }
3062 }
3063 break;
3064 }
3065 #ifdef _LP64
3066 case Op_CmpP:
3067 // Do this transformation here to preserve CmpPNode::sub() and
3068 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3069 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3070 Node* in1 = n->in(1);
3071 Node* in2 = n->in(2);
3072 if (!in1->is_DecodeNarrowPtr()) {
3073 in2 = in1;
3074 in1 = n->in(2);
3075 }
3076 assert(in1->is_DecodeNarrowPtr(), "sanity");
3077
3078 Node* new_in2 = NULL;
3079 if (in2->is_DecodeNarrowPtr()) {
3080 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3081 new_in2 = in2->in(1);
3082 } else if (in2->Opcode() == Op_ConP) {
3083 const Type* t = in2->bottom_type();
3084 if (t == TypePtr::NULL_PTR) {
3085 assert(in1->is_DecodeN(), "compare klass to null?");
3086 // Don't convert CmpP null check into CmpN if compressed
3087 // oops implicit null check is not generated.
3088 // This will allow to generate normal oop implicit null check.
3089 if (Matcher::gen_narrow_oop_implicit_null_checks())
3090 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3091 //
3092 // This transformation together with CastPP transformation above
3093 // will generated code for implicit NULL checks for compressed oops.
3094 //
3095 // The original code after Optimize()
3096 //
3097 // LoadN memory, narrow_oop_reg
3098 // decode narrow_oop_reg, base_reg
3099 // CmpP base_reg, NULL
3100 // CastPP base_reg // NotNull
3101 // Load [base_reg + offset], val_reg
3102 //
3103 // after these transformations will be
3104 //
3105 // LoadN memory, narrow_oop_reg
3106 // CmpN narrow_oop_reg, NULL
3107 // decode_not_null narrow_oop_reg, base_reg
3108 // Load [base_reg + offset], val_reg
3109 //
3110 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3111 // since narrow oops can be used in debug info now (see the code in
3112 // final_graph_reshaping_walk()).
3113 //
3114 // At the end the code will be matched to
3115 // on x86:
3116 //
3117 // Load_narrow_oop memory, narrow_oop_reg
3118 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3119 // NullCheck narrow_oop_reg
3120 //
3121 // and on sparc:
3122 //
3123 // Load_narrow_oop memory, narrow_oop_reg
3124 // decode_not_null narrow_oop_reg, base_reg
3125 // Load [base_reg + offset], val_reg
3126 // NullCheck base_reg
3127 //
3128 } else if (t->isa_oopptr()) {
3129 new_in2 = ConNode::make(t->make_narrowoop());
3130 } else if (t->isa_klassptr()) {
3131 new_in2 = ConNode::make(t->make_narrowklass());
3132 }
3133 }
3134 if (new_in2 != NULL) {
3135 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3136 n->subsume_by(cmpN, this);
3137 if (in1->outcnt() == 0) {
3138 in1->disconnect_inputs(NULL, this);
3139 }
3140 if (in2->outcnt() == 0) {
3141 in2->disconnect_inputs(NULL, this);
3142 }
3143 }
3144 }
3145 break;
3146
3147 case Op_DecodeN:
3148 case Op_DecodeNKlass:
3149 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3150 // DecodeN could be pinned when it can't be fold into
3151 // an address expression, see the code for Op_CastPP above.
3152 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3153 break;
3154
3155 case Op_EncodeP:
3156 case Op_EncodePKlass: {
3157 Node* in1 = n->in(1);
3158 if (in1->is_DecodeNarrowPtr()) {
3159 n->subsume_by(in1->in(1), this);
3160 } else if (in1->Opcode() == Op_ConP) {
3161 const Type* t = in1->bottom_type();
3162 if (t == TypePtr::NULL_PTR) {
3163 assert(t->isa_oopptr(), "null klass?");
3164 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3165 } else if (t->isa_oopptr()) {
3166 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3167 } else if (t->isa_klassptr()) {
3168 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3169 }
3170 }
3171 if (in1->outcnt() == 0) {
3172 in1->disconnect_inputs(NULL, this);
3173 }
3174 break;
3175 }
3176
3177 case Op_Proj: {
3178 if (OptimizeStringConcat) {
3179 ProjNode* p = n->as_Proj();
3180 if (p->_is_io_use) {
3181 // Separate projections were used for the exception path which
3182 // are normally removed by a late inline. If it wasn't inlined
3183 // then they will hang around and should just be replaced with
3184 // the original one.
3185 Node* proj = NULL;
3186 // Replace with just one
3187 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3188 Node *use = i.get();
3189 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3190 proj = use;
3191 break;
3192 }
3193 }
3194 assert(proj != NULL, "must be found");
3195 p->subsume_by(proj, this);
3196 }
3197 }
3198 break;
3199 }
3200
3201 case Op_Phi:
3202 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3203 // The EncodeP optimization may create Phi with the same edges
3204 // for all paths. It is not handled well by Register Allocator.
3205 Node* unique_in = n->in(1);
3206 assert(unique_in != NULL, "");
3207 uint cnt = n->req();
3208 for (uint i = 2; i < cnt; i++) {
3209 Node* m = n->in(i);
3210 assert(m != NULL, "");
3211 if (unique_in != m)
3212 unique_in = NULL;
3213 }
3214 if (unique_in != NULL) {
3215 n->subsume_by(unique_in, this);
3216 }
3217 }
3218 break;
3219
3220 #endif
3221
3222 #ifdef ASSERT
3223 case Op_CastII:
3224 // Verify that all range check dependent CastII nodes were removed.
3225 if (n->isa_CastII()->has_range_check()) {
3226 n->dump(3);
3227 assert(false, "Range check dependent CastII node was not removed");
3228 }
3229 break;
3230 #endif
3231
3232 case Op_ModI:
3233 if (UseDivMod) {
3234 // Check if a%b and a/b both exist
3235 Node* d = n->find_similar(Op_DivI);
3236 if (d) {
3237 // Replace them with a fused divmod if supported
3238 if (Matcher::has_match_rule(Op_DivModI)) {
3239 DivModINode* divmod = DivModINode::make(n);
3240 d->subsume_by(divmod->div_proj(), this);
3241 n->subsume_by(divmod->mod_proj(), this);
3242 } else {
3243 // replace a%b with a-((a/b)*b)
3244 Node* mult = new MulINode(d, d->in(2));
3245 Node* sub = new SubINode(d->in(1), mult);
3246 n->subsume_by(sub, this);
3247 }
3248 }
3249 }
3250 break;
3251
3252 case Op_ModL:
3253 if (UseDivMod) {
3254 // Check if a%b and a/b both exist
3255 Node* d = n->find_similar(Op_DivL);
3256 if (d) {
3257 // Replace them with a fused divmod if supported
3258 if (Matcher::has_match_rule(Op_DivModL)) {
3259 DivModLNode* divmod = DivModLNode::make(n);
3260 d->subsume_by(divmod->div_proj(), this);
3261 n->subsume_by(divmod->mod_proj(), this);
3262 } else {
3263 // replace a%b with a-((a/b)*b)
3264 Node* mult = new MulLNode(d, d->in(2));
3265 Node* sub = new SubLNode(d->in(1), mult);
3266 n->subsume_by(sub, this);
3267 }
3268 }
3269 }
3270 break;
3271
3272 case Op_LoadVector:
3273 case Op_StoreVector:
3274 break;
3275
3276 case Op_AddReductionVI:
3277 case Op_AddReductionVL:
3278 case Op_AddReductionVF:
3279 case Op_AddReductionVD:
3280 case Op_MulReductionVI:
3281 case Op_MulReductionVL:
3282 case Op_MulReductionVF:
3283 case Op_MulReductionVD:
3284 break;
3285
3286 case Op_PackB:
3287 case Op_PackS:
3288 case Op_PackI:
3289 case Op_PackF:
3290 case Op_PackL:
3291 case Op_PackD:
3292 if (n->req()-1 > 2) {
3293 // Replace many operand PackNodes with a binary tree for matching
3294 PackNode* p = (PackNode*) n;
3295 Node* btp = p->binary_tree_pack(1, n->req());
3296 n->subsume_by(btp, this);
3297 }
3298 break;
3299 case Op_Loop:
3300 case Op_CountedLoop:
3301 if (n->as_Loop()->is_inner_loop()) {
3302 frc.inc_inner_loop_count();
3303 }
3304 break;
3305 case Op_LShiftI:
3306 case Op_RShiftI:
3307 case Op_URShiftI:
3308 case Op_LShiftL:
3309 case Op_RShiftL:
3310 case Op_URShiftL:
3311 if (Matcher::need_masked_shift_count) {
3312 // The cpu's shift instructions don't restrict the count to the
3313 // lower 5/6 bits. We need to do the masking ourselves.
3314 Node* in2 = n->in(2);
3315 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3316 const TypeInt* t = in2->find_int_type();
3317 if (t != NULL && t->is_con()) {
3318 juint shift = t->get_con();
3319 if (shift > mask) { // Unsigned cmp
3320 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3321 }
3322 } else {
3323 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3324 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3325 n->set_req(2, shift);
3326 }
3327 }
3328 if (in2->outcnt() == 0) { // Remove dead node
3329 in2->disconnect_inputs(NULL, this);
3330 }
3331 }
3332 break;
3333 case Op_MemBarStoreStore:
3334 case Op_MemBarRelease:
3335 // Break the link with AllocateNode: it is no longer useful and
3336 // confuses register allocation.
3337 if (n->req() > MemBarNode::Precedent) {
3338 n->set_req(MemBarNode::Precedent, top());
3339 }
3340 break;
3341 case Op_RangeCheck: {
3342 RangeCheckNode* rc = n->as_RangeCheck();
3343 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3344 n->subsume_by(iff, this);
3345 frc._tests.push(iff);
3346 break;
3347 }
3348 case Op_ConvI2L: {
3349 if (!Matcher::convi2l_type_required) {
3350 // Code generation on some platforms doesn't need accurate
3351 // ConvI2L types. Widening the type can help remove redundant
3352 // address computations.
3353 n->as_Type()->set_type(TypeLong::INT);
3354 ResourceMark rm;
3355 Node_List wq;
3356 wq.push(n);
3357 for (uint next = 0; next < wq.size(); next++) {
3358 Node *m = wq.at(next);
3359
3360 for(;;) {
3361 // Loop over all nodes with identical inputs edges as m
3362 Node* k = m->find_similar(m->Opcode());
3363 if (k == NULL) {
3364 break;
3365 }
3366 // Push their uses so we get a chance to remove node made
3367 // redundant
3368 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3369 Node* u = k->fast_out(i);
3370 assert(!wq.contains(u), "shouldn't process one node several times");
3371 if (u->Opcode() == Op_LShiftL ||
3372 u->Opcode() == Op_AddL ||
3373 u->Opcode() == Op_SubL ||
3374 u->Opcode() == Op_AddP) {
3375 wq.push(u);
3376 }
3377 }
3378 // Replace all nodes with identical edges as m with m
3379 k->subsume_by(m, this);
3380 }
3381 }
3382 }
3383 break;
3384 }
3385 case Op_ValueType: {
3386 ValueTypeNode* vt = n->as_ValueType();
3387 vt->make_scalar_in_safepoints(root(), NULL);
3388 if (vt->outcnt() == 0) {
3389 vt->disconnect_inputs(NULL, this);
3390 }
3391 break;
3392 }
3393 case Op_ValueTypePtr: {
3394 ShouldNotReachHere();
3395 break;
3396 }
3397 default:
3398 assert( !n->is_Call(), "" );
3399 assert( !n->is_Mem(), "" );
3400 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3401 break;
3402 }
3403
3404 // Collect CFG split points
3405 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3406 frc._tests.push(n);
3407 }
3408 }
3409
3410 //------------------------------final_graph_reshaping_walk---------------------
3411 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3412 // requires that the walk visits a node's inputs before visiting the node.
3413 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3414 ResourceArea *area = Thread::current()->resource_area();
3415 Unique_Node_List sfpt(area);
3416
3417 frc._visited.set(root->_idx); // first, mark node as visited
3418 uint cnt = root->req();
3419 Node *n = root;
3420 uint i = 0;
3421 while (true) {
3422 if (i < cnt) {
3423 // Place all non-visited non-null inputs onto stack
3424 Node* m = n->in(i);
3425 ++i;
3426 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3427 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3428 // compute worst case interpreter size in case of a deoptimization
3429 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3430
3431 sfpt.push(m);
3432 }
3433 cnt = m->req();
3434 nstack.push(n, i); // put on stack parent and next input's index
3435 n = m;
3436 i = 0;
3437 }
3438 } else {
3439 // Now do post-visit work
3440 final_graph_reshaping_impl( n, frc );
3441 if (nstack.is_empty())
3442 break; // finished
3443 n = nstack.node(); // Get node from stack
3444 cnt = n->req();
3445 i = nstack.index();
3446 nstack.pop(); // Shift to the next node on stack
3447 }
3448 }
3449
3450 // Skip next transformation if compressed oops are not used.
3451 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3452 (!UseCompressedOops && !UseCompressedClassPointers))
3453 return;
3454
3455 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3456 // It could be done for an uncommon traps or any safepoints/calls
3457 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3458 while (sfpt.size() > 0) {
3459 n = sfpt.pop();
3460 JVMState *jvms = n->as_SafePoint()->jvms();
3461 assert(jvms != NULL, "sanity");
3462 int start = jvms->debug_start();
3463 int end = n->req();
3464 bool is_uncommon = (n->is_CallStaticJava() &&
3465 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3466 for (int j = start; j < end; j++) {
3467 Node* in = n->in(j);
3468 if (in->is_DecodeNarrowPtr()) {
3469 bool safe_to_skip = true;
3470 if (!is_uncommon ) {
3471 // Is it safe to skip?
3472 for (uint i = 0; i < in->outcnt(); i++) {
3473 Node* u = in->raw_out(i);
3474 if (!u->is_SafePoint() ||
3475 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3476 safe_to_skip = false;
3477 }
3478 }
3479 }
3480 if (safe_to_skip) {
3481 n->set_req(j, in->in(1));
3482 }
3483 if (in->outcnt() == 0) {
3484 in->disconnect_inputs(NULL, this);
3485 }
3486 }
3487 }
3488 }
3489 }
3490
3491 //------------------------------final_graph_reshaping--------------------------
3492 // Final Graph Reshaping.
3493 //
3494 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3495 // and not commoned up and forced early. Must come after regular
3496 // optimizations to avoid GVN undoing the cloning. Clone constant
3497 // inputs to Loop Phis; these will be split by the allocator anyways.
3498 // Remove Opaque nodes.
3499 // (2) Move last-uses by commutative operations to the left input to encourage
3500 // Intel update-in-place two-address operations and better register usage
3501 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3502 // calls canonicalizing them back.
3503 // (3) Count the number of double-precision FP ops, single-precision FP ops
3504 // and call sites. On Intel, we can get correct rounding either by
3505 // forcing singles to memory (requires extra stores and loads after each
3506 // FP bytecode) or we can set a rounding mode bit (requires setting and
3507 // clearing the mode bit around call sites). The mode bit is only used
3508 // if the relative frequency of single FP ops to calls is low enough.
3509 // This is a key transform for SPEC mpeg_audio.
3510 // (4) Detect infinite loops; blobs of code reachable from above but not
3511 // below. Several of the Code_Gen algorithms fail on such code shapes,
3512 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3513 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3514 // Detection is by looking for IfNodes where only 1 projection is
3515 // reachable from below or CatchNodes missing some targets.
3516 // (5) Assert for insane oop offsets in debug mode.
3517
3518 bool Compile::final_graph_reshaping() {
3519 // an infinite loop may have been eliminated by the optimizer,
3520 // in which case the graph will be empty.
3521 if (root()->req() == 1) {
3522 record_method_not_compilable("trivial infinite loop");
3523 return true;
3524 }
3525
3526 // Expensive nodes have their control input set to prevent the GVN
3527 // from freely commoning them. There's no GVN beyond this point so
3528 // no need to keep the control input. We want the expensive nodes to
3529 // be freely moved to the least frequent code path by gcm.
3530 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3531 for (int i = 0; i < expensive_count(); i++) {
3532 _expensive_nodes->at(i)->set_req(0, NULL);
3533 }
3534
3535 Final_Reshape_Counts frc;
3536
3537 // Visit everybody reachable!
3538 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3539 Node_Stack nstack(live_nodes() >> 1);
3540 final_graph_reshaping_walk(nstack, root(), frc);
3541
3542 // Check for unreachable (from below) code (i.e., infinite loops).
3543 for( uint i = 0; i < frc._tests.size(); i++ ) {
3544 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3545 // Get number of CFG targets.
3546 // Note that PCTables include exception targets after calls.
3547 uint required_outcnt = n->required_outcnt();
3548 if (n->outcnt() != required_outcnt) {
3549 // Check for a few special cases. Rethrow Nodes never take the
3550 // 'fall-thru' path, so expected kids is 1 less.
3551 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3552 if (n->in(0)->in(0)->is_Call()) {
3553 CallNode *call = n->in(0)->in(0)->as_Call();
3554 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3555 required_outcnt--; // Rethrow always has 1 less kid
3556 } else if (call->req() > TypeFunc::Parms &&
3557 call->is_CallDynamicJava()) {
3558 // Check for null receiver. In such case, the optimizer has
3559 // detected that the virtual call will always result in a null
3560 // pointer exception. The fall-through projection of this CatchNode
3561 // will not be populated.
3562 Node *arg0 = call->in(TypeFunc::Parms);
3563 if (arg0->is_Type() &&
3564 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3565 required_outcnt--;
3566 }
3567 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3568 call->req() > TypeFunc::Parms+1 &&
3569 call->is_CallStaticJava()) {
3570 // Check for negative array length. In such case, the optimizer has
3571 // detected that the allocation attempt will always result in an
3572 // exception. There is no fall-through projection of this CatchNode .
3573 Node *arg1 = call->in(TypeFunc::Parms+1);
3574 if (arg1->is_Type() &&
3575 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3576 required_outcnt--;
3577 }
3578 }
3579 }
3580 }
3581 // Recheck with a better notion of 'required_outcnt'
3582 if (n->outcnt() != required_outcnt) {
3583 assert(false, "malformed control flow");
3584 record_method_not_compilable("malformed control flow");
3585 return true; // Not all targets reachable!
3586 }
3587 }
3588 // Check that I actually visited all kids. Unreached kids
3589 // must be infinite loops.
3590 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3591 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3592 record_method_not_compilable("infinite loop");
3593 assert(false, "infinite loop");
3594 return true; // Found unvisited kid; must be unreach
3595 }
3596 }
3597
3598 // If original bytecodes contained a mixture of floats and doubles
3599 // check if the optimizer has made it homogenous, item (3).
3600 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3601 frc.get_float_count() > 32 &&
3602 frc.get_double_count() == 0 &&
3603 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3604 set_24_bit_selection_and_mode( false, true );
3605 }
3606
3607 set_java_calls(frc.get_java_call_count());
3608 set_inner_loops(frc.get_inner_loop_count());
3609
3610 // No infinite loops, no reason to bail out.
3611 return false;
3612 }
3613
3614 //-----------------------------too_many_traps----------------------------------
3615 // Report if there are too many traps at the current method and bci.
3616 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3617 bool Compile::too_many_traps(ciMethod* method,
3618 int bci,
3619 Deoptimization::DeoptReason reason) {
3620 ciMethodData* md = method->method_data();
3621 if (md->is_empty()) {
3622 // Assume the trap has not occurred, or that it occurred only
3623 // because of a transient condition during start-up in the interpreter.
3624 return false;
3625 }
3626 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3627 if (md->has_trap_at(bci, m, reason) != 0) {
3628 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3629 // Also, if there are multiple reasons, or if there is no per-BCI record,
3630 // assume the worst.
3631 if (log())
3632 log()->elem("observe trap='%s' count='%d'",
3633 Deoptimization::trap_reason_name(reason),
3634 md->trap_count(reason));
3635 return true;
3636 } else {
3637 // Ignore method/bci and see if there have been too many globally.
3638 return too_many_traps(reason, md);
3639 }
3640 }
3641
3642 // Less-accurate variant which does not require a method and bci.
3643 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3644 ciMethodData* logmd) {
3645 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3646 // Too many traps globally.
3647 // Note that we use cumulative trap_count, not just md->trap_count.
3648 if (log()) {
3649 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3650 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3651 Deoptimization::trap_reason_name(reason),
3652 mcount, trap_count(reason));
3653 }
3654 return true;
3655 } else {
3656 // The coast is clear.
3657 return false;
3658 }
3659 }
3660
3661 //--------------------------too_many_recompiles--------------------------------
3662 // Report if there are too many recompiles at the current method and bci.
3663 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3664 // Is not eager to return true, since this will cause the compiler to use
3665 // Action_none for a trap point, to avoid too many recompilations.
3666 bool Compile::too_many_recompiles(ciMethod* method,
3667 int bci,
3668 Deoptimization::DeoptReason reason) {
3669 ciMethodData* md = method->method_data();
3670 if (md->is_empty()) {
3671 // Assume the trap has not occurred, or that it occurred only
3672 // because of a transient condition during start-up in the interpreter.
3673 return false;
3674 }
3675 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3676 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3677 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3678 Deoptimization::DeoptReason per_bc_reason
3679 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3680 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3681 if ((per_bc_reason == Deoptimization::Reason_none
3682 || md->has_trap_at(bci, m, reason) != 0)
3683 // The trap frequency measure we care about is the recompile count:
3684 && md->trap_recompiled_at(bci, m)
3685 && md->overflow_recompile_count() >= bc_cutoff) {
3686 // Do not emit a trap here if it has already caused recompilations.
3687 // Also, if there are multiple reasons, or if there is no per-BCI record,
3688 // assume the worst.
3689 if (log())
3690 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3691 Deoptimization::trap_reason_name(reason),
3692 md->trap_count(reason),
3693 md->overflow_recompile_count());
3694 return true;
3695 } else if (trap_count(reason) != 0
3696 && decompile_count() >= m_cutoff) {
3697 // Too many recompiles globally, and we have seen this sort of trap.
3698 // Use cumulative decompile_count, not just md->decompile_count.
3699 if (log())
3700 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3701 Deoptimization::trap_reason_name(reason),
3702 md->trap_count(reason), trap_count(reason),
3703 md->decompile_count(), decompile_count());
3704 return true;
3705 } else {
3706 // The coast is clear.
3707 return false;
3708 }
3709 }
3710
3711 // Compute when not to trap. Used by matching trap based nodes and
3712 // NullCheck optimization.
3713 void Compile::set_allowed_deopt_reasons() {
3714 _allowed_reasons = 0;
3715 if (is_method_compilation()) {
3716 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3717 assert(rs < BitsPerInt, "recode bit map");
3718 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3719 _allowed_reasons |= nth_bit(rs);
3720 }
3721 }
3722 }
3723 }
3724
3725 #ifndef PRODUCT
3726 //------------------------------verify_graph_edges---------------------------
3727 // Walk the Graph and verify that there is a one-to-one correspondence
3728 // between Use-Def edges and Def-Use edges in the graph.
3729 void Compile::verify_graph_edges(bool no_dead_code) {
3730 if (VerifyGraphEdges) {
3731 ResourceArea *area = Thread::current()->resource_area();
3732 Unique_Node_List visited(area);
3733 // Call recursive graph walk to check edges
3734 _root->verify_edges(visited);
3735 if (no_dead_code) {
3736 // Now make sure that no visited node is used by an unvisited node.
3737 bool dead_nodes = false;
3738 Unique_Node_List checked(area);
3739 while (visited.size() > 0) {
3740 Node* n = visited.pop();
3741 checked.push(n);
3742 for (uint i = 0; i < n->outcnt(); i++) {
3743 Node* use = n->raw_out(i);
3744 if (checked.member(use)) continue; // already checked
3745 if (visited.member(use)) continue; // already in the graph
3746 if (use->is_Con()) continue; // a dead ConNode is OK
3747 // At this point, we have found a dead node which is DU-reachable.
3748 if (!dead_nodes) {
3749 tty->print_cr("*** Dead nodes reachable via DU edges:");
3750 dead_nodes = true;
3751 }
3752 use->dump(2);
3753 tty->print_cr("---");
3754 checked.push(use); // No repeats; pretend it is now checked.
3755 }
3756 }
3757 assert(!dead_nodes, "using nodes must be reachable from root");
3758 }
3759 }
3760 }
3761
3762 // Verify GC barriers consistency
3763 // Currently supported:
3764 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3765 void Compile::verify_barriers() {
3766 if (UseG1GC) {
3767 // Verify G1 pre-barriers
3768 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3769
3770 ResourceArea *area = Thread::current()->resource_area();
3771 Unique_Node_List visited(area);
3772 Node_List worklist(area);
3773 // We're going to walk control flow backwards starting from the Root
3774 worklist.push(_root);
3775 while (worklist.size() > 0) {
3776 Node* x = worklist.pop();
3777 if (x == NULL || x == top()) continue;
3778 if (visited.member(x)) {
3779 continue;
3780 } else {
3781 visited.push(x);
3782 }
3783
3784 if (x->is_Region()) {
3785 for (uint i = 1; i < x->req(); i++) {
3786 worklist.push(x->in(i));
3787 }
3788 } else {
3789 worklist.push(x->in(0));
3790 // We are looking for the pattern:
3791 // /->ThreadLocal
3792 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3793 // \->ConI(0)
3794 // We want to verify that the If and the LoadB have the same control
3795 // See GraphKit::g1_write_barrier_pre()
3796 if (x->is_If()) {
3797 IfNode *iff = x->as_If();
3798 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3799 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3800 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3801 && cmp->in(1)->is_Load()) {
3802 LoadNode* load = cmp->in(1)->as_Load();
3803 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3804 && load->in(2)->in(3)->is_Con()
3805 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3806
3807 Node* if_ctrl = iff->in(0);
3808 Node* load_ctrl = load->in(0);
3809
3810 if (if_ctrl != load_ctrl) {
3811 // Skip possible CProj->NeverBranch in infinite loops
3812 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3813 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3814 if_ctrl = if_ctrl->in(0)->in(0);
3815 }
3816 }
3817 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3818 }
3819 }
3820 }
3821 }
3822 }
3823 }
3824 }
3825 }
3826
3827 #endif
3828
3829 // The Compile object keeps track of failure reasons separately from the ciEnv.
3830 // This is required because there is not quite a 1-1 relation between the
3831 // ciEnv and its compilation task and the Compile object. Note that one
3832 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3833 // to backtrack and retry without subsuming loads. Other than this backtracking
3834 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3835 // by the logic in C2Compiler.
3836 void Compile::record_failure(const char* reason) {
3837 if (log() != NULL) {
3838 log()->elem("failure reason='%s' phase='compile'", reason);
3839 }
3840 if (_failure_reason == NULL) {
3841 // Record the first failure reason.
3842 _failure_reason = reason;
3843 }
3844
3845 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3846 C->print_method(PHASE_FAILURE);
3847 }
3848 _root = NULL; // flush the graph, too
3849 }
3850
3851 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3852 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3853 _phase_name(name), _dolog(CITimeVerbose)
3854 {
3855 if (_dolog) {
3856 C = Compile::current();
3857 _log = C->log();
3858 } else {
3859 C = NULL;
3860 _log = NULL;
3861 }
3862 if (_log != NULL) {
3863 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3864 _log->stamp();
3865 _log->end_head();
3866 }
3867 }
3868
3869 Compile::TracePhase::~TracePhase() {
3870
3871 C = Compile::current();
3872 if (_dolog) {
3873 _log = C->log();
3874 } else {
3875 _log = NULL;
3876 }
3877
3878 #ifdef ASSERT
3879 if (PrintIdealNodeCount) {
3880 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3881 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3882 }
3883
3884 if (VerifyIdealNodeCount) {
3885 Compile::current()->print_missing_nodes();
3886 }
3887 #endif
3888
3889 if (_log != NULL) {
3890 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3891 }
3892 }
3893
3894 //=============================================================================
3895 // Two Constant's are equal when the type and the value are equal.
3896 bool Compile::Constant::operator==(const Constant& other) {
3897 if (type() != other.type() ) return false;
3898 if (can_be_reused() != other.can_be_reused()) return false;
3899 // For floating point values we compare the bit pattern.
3900 switch (type()) {
3901 case T_INT:
3902 case T_FLOAT: return (_v._value.i == other._v._value.i);
3903 case T_LONG:
3904 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3905 case T_OBJECT:
3906 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3907 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3908 case T_METADATA: return (_v._metadata == other._v._metadata);
3909 default: ShouldNotReachHere(); return false;
3910 }
3911 }
3912
3913 static int type_to_size_in_bytes(BasicType t) {
3914 switch (t) {
3915 case T_INT: return sizeof(jint );
3916 case T_LONG: return sizeof(jlong );
3917 case T_FLOAT: return sizeof(jfloat );
3918 case T_DOUBLE: return sizeof(jdouble);
3919 case T_METADATA: return sizeof(Metadata*);
3920 // We use T_VOID as marker for jump-table entries (labels) which
3921 // need an internal word relocation.
3922 case T_VOID:
3923 case T_ADDRESS:
3924 case T_OBJECT: return sizeof(jobject);
3925 default:
3926 ShouldNotReachHere();
3927 return -1;
3928 }
3929 }
3930
3931 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3932 // sort descending
3933 if (a->freq() > b->freq()) return -1;
3934 if (a->freq() < b->freq()) return 1;
3935 return 0;
3936 }
3937
3938 void Compile::ConstantTable::calculate_offsets_and_size() {
3939 // First, sort the array by frequencies.
3940 _constants.sort(qsort_comparator);
3941
3942 #ifdef ASSERT
3943 // Make sure all jump-table entries were sorted to the end of the
3944 // array (they have a negative frequency).
3945 bool found_void = false;
3946 for (int i = 0; i < _constants.length(); i++) {
3947 Constant con = _constants.at(i);
3948 if (con.type() == T_VOID)
3949 found_void = true; // jump-tables
3950 else
3951 assert(!found_void, "wrong sorting");
3952 }
3953 #endif
3954
3955 int offset = 0;
3956 for (int i = 0; i < _constants.length(); i++) {
3957 Constant* con = _constants.adr_at(i);
3958
3959 // Align offset for type.
3960 int typesize = type_to_size_in_bytes(con->type());
3961 offset = align_up(offset, typesize);
3962 con->set_offset(offset); // set constant's offset
3963
3964 if (con->type() == T_VOID) {
3965 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3966 offset = offset + typesize * n->outcnt(); // expand jump-table
3967 } else {
3968 offset = offset + typesize;
3969 }
3970 }
3971
3972 // Align size up to the next section start (which is insts; see
3973 // CodeBuffer::align_at_start).
3974 assert(_size == -1, "already set?");
3975 _size = align_up(offset, (int)CodeEntryAlignment);
3976 }
3977
3978 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3979 MacroAssembler _masm(&cb);
3980 for (int i = 0; i < _constants.length(); i++) {
3981 Constant con = _constants.at(i);
3982 address constant_addr = NULL;
3983 switch (con.type()) {
3984 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
3985 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3986 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3987 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3988 case T_OBJECT: {
3989 jobject obj = con.get_jobject();
3990 int oop_index = _masm.oop_recorder()->find_index(obj);
3991 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3992 break;
3993 }
3994 case T_ADDRESS: {
3995 address addr = (address) con.get_jobject();
3996 constant_addr = _masm.address_constant(addr);
3997 break;
3998 }
3999 // We use T_VOID as marker for jump-table entries (labels) which
4000 // need an internal word relocation.
4001 case T_VOID: {
4002 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4003 // Fill the jump-table with a dummy word. The real value is
4004 // filled in later in fill_jump_table.
4005 address dummy = (address) n;
4006 constant_addr = _masm.address_constant(dummy);
4007 // Expand jump-table
4008 for (uint i = 1; i < n->outcnt(); i++) {
4009 address temp_addr = _masm.address_constant(dummy + i);
4010 assert(temp_addr, "consts section too small");
4011 }
4012 break;
4013 }
4014 case T_METADATA: {
4015 Metadata* obj = con.get_metadata();
4016 int metadata_index = _masm.oop_recorder()->find_index(obj);
4017 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4018 break;
4019 }
4020 default: ShouldNotReachHere();
4021 }
4022 assert(constant_addr, "consts section too small");
4023 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4024 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4025 }
4026 }
4027
4028 int Compile::ConstantTable::find_offset(Constant& con) const {
4029 int idx = _constants.find(con);
4030 assert(idx != -1, "constant must be in constant table");
4031 int offset = _constants.at(idx).offset();
4032 assert(offset != -1, "constant table not emitted yet?");
4033 return offset;
4034 }
4035
4036 void Compile::ConstantTable::add(Constant& con) {
4037 if (con.can_be_reused()) {
4038 int idx = _constants.find(con);
4039 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4040 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4041 return;
4042 }
4043 }
4044 (void) _constants.append(con);
4045 }
4046
4047 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4048 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4049 Constant con(type, value, b->_freq);
4050 add(con);
4051 return con;
4052 }
4053
4054 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4055 Constant con(metadata);
4056 add(con);
4057 return con;
4058 }
4059
4060 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4061 jvalue value;
4062 BasicType type = oper->type()->basic_type();
4063 switch (type) {
4064 case T_LONG: value.j = oper->constantL(); break;
4065 case T_FLOAT: value.f = oper->constantF(); break;
4066 case T_DOUBLE: value.d = oper->constantD(); break;
4067 case T_OBJECT:
4068 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4069 case T_METADATA: return add((Metadata*)oper->constant()); break;
4070 default: guarantee(false, "unhandled type: %s", type2name(type));
4071 }
4072 return add(n, type, value);
4073 }
4074
4075 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4076 jvalue value;
4077 // We can use the node pointer here to identify the right jump-table
4078 // as this method is called from Compile::Fill_buffer right before
4079 // the MachNodes are emitted and the jump-table is filled (means the
4080 // MachNode pointers do not change anymore).
4081 value.l = (jobject) n;
4082 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4083 add(con);
4084 return con;
4085 }
4086
4087 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4088 // If called from Compile::scratch_emit_size do nothing.
4089 if (Compile::current()->in_scratch_emit_size()) return;
4090
4091 assert(labels.is_nonempty(), "must be");
4092 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4093
4094 // Since MachConstantNode::constant_offset() also contains
4095 // table_base_offset() we need to subtract the table_base_offset()
4096 // to get the plain offset into the constant table.
4097 int offset = n->constant_offset() - table_base_offset();
4098
4099 MacroAssembler _masm(&cb);
4100 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4101
4102 for (uint i = 0; i < n->outcnt(); i++) {
4103 address* constant_addr = &jump_table_base[i];
4104 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));
4105 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4106 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4107 }
4108 }
4109
4110 //----------------------------static_subtype_check-----------------------------
4111 // Shortcut important common cases when superklass is exact:
4112 // (0) superklass is java.lang.Object (can occur in reflective code)
4113 // (1) subklass is already limited to a subtype of superklass => always ok
4114 // (2) subklass does not overlap with superklass => always fail
4115 // (3) superklass has NO subtypes and we can check with a simple compare.
4116 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4117 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4118 return SSC_full_test; // Let caller generate the general case.
4119 }
4120
4121 if (!EnableMVT && !EnableValhalla && superk == env()->Object_klass()) {
4122 return SSC_always_true; // (0) this test cannot fail
4123 }
4124
4125 ciType* superelem = superk;
4126 if (superelem->is_array_klass())
4127 superelem = superelem->as_array_klass()->base_element_type();
4128
4129 if (!subk->is_interface()) { // cannot trust static interface types yet
4130 if (subk->is_subtype_of(superk)) {
4131 return SSC_always_true; // (1) false path dead; no dynamic test needed
4132 }
4133 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4134 !superk->is_subtype_of(subk)) {
4135 return SSC_always_false;
4136 }
4137 }
4138
4139 // If casting to an instance klass, it must have no subtypes
4140 if (superk->is_interface()) {
4141 // Cannot trust interfaces yet.
4142 // %%% S.B. superk->nof_implementors() == 1
4143 } else if (superelem->is_instance_klass()) {
4144 ciInstanceKlass* ik = superelem->as_instance_klass();
4145 if (!ik->has_subklass() && !ik->is_interface()) {
4146 if (!ik->is_final()) {
4147 // Add a dependency if there is a chance of a later subclass.
4148 dependencies()->assert_leaf_type(ik);
4149 }
4150 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4151 }
4152 } else {
4153 // A primitive array type has no subtypes.
4154 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4155 }
4156
4157 return SSC_full_test;
4158 }
4159
4160 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4161 #ifdef _LP64
4162 // The scaled index operand to AddP must be a clean 64-bit value.
4163 // Java allows a 32-bit int to be incremented to a negative
4164 // value, which appears in a 64-bit register as a large
4165 // positive number. Using that large positive number as an
4166 // operand in pointer arithmetic has bad consequences.
4167 // On the other hand, 32-bit overflow is rare, and the possibility
4168 // can often be excluded, if we annotate the ConvI2L node with
4169 // a type assertion that its value is known to be a small positive
4170 // number. (The prior range check has ensured this.)
4171 // This assertion is used by ConvI2LNode::Ideal.
4172 int index_max = max_jint - 1; // array size is max_jint, index is one less
4173 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4174 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4175 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4176 #endif
4177 return idx;
4178 }
4179
4180 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4181 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4182 if (ctrl != NULL) {
4183 // Express control dependency by a CastII node with a narrow type.
4184 value = new CastIINode(value, itype, false, true /* range check dependency */);
4185 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4186 // node from floating above the range check during loop optimizations. Otherwise, the
4187 // ConvI2L node may be eliminated independently of the range check, causing the data path
4188 // to become TOP while the control path is still there (although it's unreachable).
4189 value->set_req(0, ctrl);
4190 // Save CastII node to remove it after loop optimizations.
4191 phase->C->add_range_check_cast(value);
4192 value = phase->transform(value);
4193 }
4194 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4195 return phase->transform(new ConvI2LNode(value, ltype));
4196 }
4197
4198 // The message about the current inlining is accumulated in
4199 // _print_inlining_stream and transfered into the _print_inlining_list
4200 // once we know whether inlining succeeds or not. For regular
4201 // inlining, messages are appended to the buffer pointed by
4202 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4203 // a new buffer is added after _print_inlining_idx in the list. This
4204 // way we can update the inlining message for late inlining call site
4205 // when the inlining is attempted again.
4206 void Compile::print_inlining_init() {
4207 if (print_inlining() || print_intrinsics()) {
4208 _print_inlining_stream = new stringStream();
4209 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4210 }
4211 }
4212
4213 void Compile::print_inlining_reinit() {
4214 if (print_inlining() || print_intrinsics()) {
4215 // Re allocate buffer when we change ResourceMark
4216 _print_inlining_stream = new stringStream();
4217 }
4218 }
4219
4220 void Compile::print_inlining_reset() {
4221 _print_inlining_stream->reset();
4222 }
4223
4224 void Compile::print_inlining_commit() {
4225 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4226 // Transfer the message from _print_inlining_stream to the current
4227 // _print_inlining_list buffer and clear _print_inlining_stream.
4228 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4229 print_inlining_reset();
4230 }
4231
4232 void Compile::print_inlining_push() {
4233 // Add new buffer to the _print_inlining_list at current position
4234 _print_inlining_idx++;
4235 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4236 }
4237
4238 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4239 return _print_inlining_list->at(_print_inlining_idx);
4240 }
4241
4242 void Compile::print_inlining_update(CallGenerator* cg) {
4243 if (print_inlining() || print_intrinsics()) {
4244 if (!cg->is_late_inline()) {
4245 if (print_inlining_current().cg() != NULL) {
4246 print_inlining_push();
4247 }
4248 print_inlining_commit();
4249 } else {
4250 if (print_inlining_current().cg() != cg &&
4251 (print_inlining_current().cg() != NULL ||
4252 print_inlining_current().ss()->size() != 0)) {
4253 print_inlining_push();
4254 }
4255 print_inlining_commit();
4256 print_inlining_current().set_cg(cg);
4257 }
4258 }
4259 }
4260
4261 void Compile::print_inlining_move_to(CallGenerator* cg) {
4262 // We resume inlining at a late inlining call site. Locate the
4263 // corresponding inlining buffer so that we can update it.
4264 if (print_inlining()) {
4265 for (int i = 0; i < _print_inlining_list->length(); i++) {
4266 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4267 _print_inlining_idx = i;
4268 return;
4269 }
4270 }
4271 ShouldNotReachHere();
4272 }
4273 }
4274
4275 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4276 if (print_inlining()) {
4277 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4278 assert(print_inlining_current().cg() == cg, "wrong entry");
4279 // replace message with new message
4280 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4281 print_inlining_commit();
4282 print_inlining_current().set_cg(cg);
4283 }
4284 }
4285
4286 void Compile::print_inlining_assert_ready() {
4287 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4288 }
4289
4290 void Compile::process_print_inlining() {
4291 bool do_print_inlining = print_inlining() || print_intrinsics();
4292 if (do_print_inlining || log() != NULL) {
4293 // Print inlining message for candidates that we couldn't inline
4294 // for lack of space
4295 for (int i = 0; i < _late_inlines.length(); i++) {
4296 CallGenerator* cg = _late_inlines.at(i);
4297 if (!cg->is_mh_late_inline()) {
4298 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4299 if (do_print_inlining) {
4300 cg->print_inlining_late(msg);
4301 }
4302 log_late_inline_failure(cg, msg);
4303 }
4304 }
4305 }
4306 if (do_print_inlining) {
4307 ResourceMark rm;
4308 stringStream ss;
4309 for (int i = 0; i < _print_inlining_list->length(); i++) {
4310 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4311 }
4312 size_t end = ss.size();
4313 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4314 strncpy(_print_inlining_output, ss.base(), end+1);
4315 _print_inlining_output[end] = 0;
4316 }
4317 }
4318
4319 void Compile::dump_print_inlining() {
4320 if (_print_inlining_output != NULL) {
4321 tty->print_raw(_print_inlining_output);
4322 }
4323 }
4324
4325 void Compile::log_late_inline(CallGenerator* cg) {
4326 if (log() != NULL) {
4327 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4328 cg->unique_id());
4329 JVMState* p = cg->call_node()->jvms();
4330 while (p != NULL) {
4331 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4332 p = p->caller();
4333 }
4334 log()->tail("late_inline");
4335 }
4336 }
4337
4338 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4339 log_late_inline(cg);
4340 if (log() != NULL) {
4341 log()->inline_fail(msg);
4342 }
4343 }
4344
4345 void Compile::log_inline_id(CallGenerator* cg) {
4346 if (log() != NULL) {
4347 // The LogCompilation tool needs a unique way to identify late
4348 // inline call sites. This id must be unique for this call site in
4349 // this compilation. Try to have it unique across compilations as
4350 // well because it can be convenient when grepping through the log
4351 // file.
4352 // Distinguish OSR compilations from others in case CICountOSR is
4353 // on.
4354 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4355 cg->set_unique_id(id);
4356 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4357 }
4358 }
4359
4360 void Compile::log_inline_failure(const char* msg) {
4361 if (C->log() != NULL) {
4362 C->log()->inline_fail(msg);
4363 }
4364 }
4365
4366
4367 // Dump inlining replay data to the stream.
4368 // Don't change thread state and acquire any locks.
4369 void Compile::dump_inline_data(outputStream* out) {
4370 InlineTree* inl_tree = ilt();
4371 if (inl_tree != NULL) {
4372 out->print(" inline %d", inl_tree->count());
4373 inl_tree->dump_replay_data(out);
4374 }
4375 }
4376
4377 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4378 if (n1->Opcode() < n2->Opcode()) return -1;
4379 else if (n1->Opcode() > n2->Opcode()) return 1;
4380
4381 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4382 for (uint i = 1; i < n1->req(); i++) {
4383 if (n1->in(i) < n2->in(i)) return -1;
4384 else if (n1->in(i) > n2->in(i)) return 1;
4385 }
4386
4387 return 0;
4388 }
4389
4390 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4391 Node* n1 = *n1p;
4392 Node* n2 = *n2p;
4393
4394 return cmp_expensive_nodes(n1, n2);
4395 }
4396
4397 void Compile::sort_expensive_nodes() {
4398 if (!expensive_nodes_sorted()) {
4399 _expensive_nodes->sort(cmp_expensive_nodes);
4400 }
4401 }
4402
4403 bool Compile::expensive_nodes_sorted() const {
4404 for (int i = 1; i < _expensive_nodes->length(); i++) {
4405 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4406 return false;
4407 }
4408 }
4409 return true;
4410 }
4411
4412 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4413 if (_expensive_nodes->length() == 0) {
4414 return false;
4415 }
4416
4417 assert(OptimizeExpensiveOps, "optimization off?");
4418
4419 // Take this opportunity to remove dead nodes from the list
4420 int j = 0;
4421 for (int i = 0; i < _expensive_nodes->length(); i++) {
4422 Node* n = _expensive_nodes->at(i);
4423 if (!n->is_unreachable(igvn)) {
4424 assert(n->is_expensive(), "should be expensive");
4425 _expensive_nodes->at_put(j, n);
4426 j++;
4427 }
4428 }
4429 _expensive_nodes->trunc_to(j);
4430
4431 // Then sort the list so that similar nodes are next to each other
4432 // and check for at least two nodes of identical kind with same data
4433 // inputs.
4434 sort_expensive_nodes();
4435
4436 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4437 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4438 return true;
4439 }
4440 }
4441
4442 return false;
4443 }
4444
4445 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4446 if (_expensive_nodes->length() == 0) {
4447 return;
4448 }
4449
4450 assert(OptimizeExpensiveOps, "optimization off?");
4451
4452 // Sort to bring similar nodes next to each other and clear the
4453 // control input of nodes for which there's only a single copy.
4454 sort_expensive_nodes();
4455
4456 int j = 0;
4457 int identical = 0;
4458 int i = 0;
4459 bool modified = false;
4460 for (; i < _expensive_nodes->length()-1; i++) {
4461 assert(j <= i, "can't write beyond current index");
4462 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4463 identical++;
4464 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4465 continue;
4466 }
4467 if (identical > 0) {
4468 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4469 identical = 0;
4470 } else {
4471 Node* n = _expensive_nodes->at(i);
4472 igvn.replace_input_of(n, 0, NULL);
4473 igvn.hash_insert(n);
4474 modified = true;
4475 }
4476 }
4477 if (identical > 0) {
4478 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4479 } else if (_expensive_nodes->length() >= 1) {
4480 Node* n = _expensive_nodes->at(i);
4481 igvn.replace_input_of(n, 0, NULL);
4482 igvn.hash_insert(n);
4483 modified = true;
4484 }
4485 _expensive_nodes->trunc_to(j);
4486 if (modified) {
4487 igvn.optimize();
4488 }
4489 }
4490
4491 void Compile::add_expensive_node(Node * n) {
4492 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4493 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4494 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4495 if (OptimizeExpensiveOps) {
4496 _expensive_nodes->append(n);
4497 } else {
4498 // Clear control input and let IGVN optimize expensive nodes if
4499 // OptimizeExpensiveOps is off.
4500 n->set_req(0, NULL);
4501 }
4502 }
4503
4504 /**
4505 * Remove the speculative part of types and clean up the graph
4506 */
4507 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4508 if (UseTypeSpeculation) {
4509 Unique_Node_List worklist;
4510 worklist.push(root());
4511 int modified = 0;
4512 // Go over all type nodes that carry a speculative type, drop the
4513 // speculative part of the type and enqueue the node for an igvn
4514 // which may optimize it out.
4515 for (uint next = 0; next < worklist.size(); ++next) {
4516 Node *n = worklist.at(next);
4517 if (n->is_Type()) {
4518 TypeNode* tn = n->as_Type();
4519 const Type* t = tn->type();
4520 const Type* t_no_spec = t->remove_speculative();
4521 if (t_no_spec != t) {
4522 bool in_hash = igvn.hash_delete(n);
4523 assert(in_hash, "node should be in igvn hash table");
4524 tn->set_type(t_no_spec);
4525 igvn.hash_insert(n);
4526 igvn._worklist.push(n); // give it a chance to go away
4527 modified++;
4528 }
4529 }
4530 uint max = n->len();
4531 for( uint i = 0; i < max; ++i ) {
4532 Node *m = n->in(i);
4533 if (not_a_node(m)) continue;
4534 worklist.push(m);
4535 }
4536 }
4537 // Drop the speculative part of all types in the igvn's type table
4538 igvn.remove_speculative_types();
4539 if (modified > 0) {
4540 igvn.optimize();
4541 }
4542 #ifdef ASSERT
4543 // Verify that after the IGVN is over no speculative type has resurfaced
4544 worklist.clear();
4545 worklist.push(root());
4546 for (uint next = 0; next < worklist.size(); ++next) {
4547 Node *n = worklist.at(next);
4548 const Type* t = igvn.type_or_null(n);
4549 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4550 if (n->is_Type()) {
4551 t = n->as_Type()->type();
4552 assert(t == t->remove_speculative(), "no more speculative types");
4553 }
4554 uint max = n->len();
4555 for( uint i = 0; i < max; ++i ) {
4556 Node *m = n->in(i);
4557 if (not_a_node(m)) continue;
4558 worklist.push(m);
4559 }
4560 }
4561 igvn.check_no_speculative_types();
4562 #endif
4563 }
4564 }
4565
4566 // Auxiliary method to support randomized stressing/fuzzing.
4567 //
4568 // This method can be called the arbitrary number of times, with current count
4569 // as the argument. The logic allows selecting a single candidate from the
4570 // running list of candidates as follows:
4571 // int count = 0;
4572 // Cand* selected = null;
4573 // while(cand = cand->next()) {
4574 // if (randomized_select(++count)) {
4575 // selected = cand;
4576 // }
4577 // }
4578 //
4579 // Including count equalizes the chances any candidate is "selected".
4580 // This is useful when we don't have the complete list of candidates to choose
4581 // from uniformly. In this case, we need to adjust the randomicity of the
4582 // selection, or else we will end up biasing the selection towards the latter
4583 // candidates.
4584 //
4585 // Quick back-envelope calculation shows that for the list of n candidates
4586 // the equal probability for the candidate to persist as "best" can be
4587 // achieved by replacing it with "next" k-th candidate with the probability
4588 // of 1/k. It can be easily shown that by the end of the run, the
4589 // probability for any candidate is converged to 1/n, thus giving the
4590 // uniform distribution among all the candidates.
4591 //
4592 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4593 #define RANDOMIZED_DOMAIN_POW 29
4594 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4595 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4596 bool Compile::randomized_select(int count) {
4597 assert(count > 0, "only positive");
4598 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4599 }
4600
4601 CloneMap& Compile::clone_map() { return _clone_map; }
4602 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4603
4604 void NodeCloneInfo::dump() const {
4605 tty->print(" {%d:%d} ", idx(), gen());
4606 }
4607
4608 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4609 uint64_t val = value(old->_idx);
4610 NodeCloneInfo cio(val);
4611 assert(val != 0, "old node should be in the map");
4612 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4613 insert(nnn->_idx, cin.get());
4614 #ifndef PRODUCT
4615 if (is_debug()) {
4616 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4617 }
4618 #endif
4619 }
4620
4621 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4622 NodeCloneInfo cio(value(old->_idx));
4623 if (cio.get() == 0) {
4624 cio.set(old->_idx, 0);
4625 insert(old->_idx, cio.get());
4626 #ifndef PRODUCT
4627 if (is_debug()) {
4628 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4629 }
4630 #endif
4631 }
4632 clone(old, nnn, gen);
4633 }
4634
4635 int CloneMap::max_gen() const {
4636 int g = 0;
4637 DictI di(_dict);
4638 for(; di.test(); ++di) {
4639 int t = gen(di._key);
4640 if (g < t) {
4641 g = t;
4642 #ifndef PRODUCT
4643 if (is_debug()) {
4644 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4645 }
4646 #endif
4647 }
4648 }
4649 return g;
4650 }
4651
4652 void CloneMap::dump(node_idx_t key) const {
4653 uint64_t val = value(key);
4654 if (val != 0) {
4655 NodeCloneInfo ni(val);
4656 ni.dump();
4657 }
4658 }
4659
4660 // Helper function for enforcing certain bytecodes to reexecute if
4661 // deoptimization happens
4662 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
4663 ciMethod* cur_method = jvms->method();
4664 int cur_bci = jvms->bci();
4665 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
4666 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
4667 return Interpreter::bytecode_should_reexecute(code) ||
4668 (is_anewarray && (code == Bytecodes::_multianewarray));
4669 // Reexecute _multianewarray bytecode which was replaced with
4670 // sequence of [a]newarray. See Parse::do_multianewarray().
4671 //
4672 // Note: interpreter should not have it set since this optimization
4673 // is limited by dimensions and guarded by flag so in some cases
4674 // multianewarray() runtime calls will be generated and
4675 // the bytecode should not be reexecutes (stack will not be reset).
4676 } else
4677 return false;
4678 }
4679
4680 void Compile::add_safepoint_edges(SafePointNode* call, JVMState* youngest_jvms, bool can_prune_locals, uint stack_slots_not_pruned) {
4681 // do not scribble on the input jvms
4682 JVMState* out_jvms = youngest_jvms->clone_deep(C);
4683 call->set_jvms(out_jvms); // Start jvms list for call node
4684
4685 // For a known set of bytecodes, the interpreter should reexecute them if
4686 // deoptimization happens. We set the reexecute state for them here
4687 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
4688 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
4689 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
4690 }
4691
4692 // Presize the call:
4693 DEBUG_ONLY(uint non_debug_edges = call->req());
4694 call->add_req_batch(top(), youngest_jvms->debug_depth());
4695 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
4696
4697 // Set up edges so that the call looks like this:
4698 // Call [state:] ctl io mem fptr retadr
4699 // [parms:] parm0 ... parmN
4700 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
4701 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
4702 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
4703 // Note that caller debug info precedes callee debug info.
4704
4705 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
4706 uint debug_ptr = call->req();
4707
4708 // Loop over the map input edges associated with jvms, add them
4709 // to the call node, & reset all offsets to match call node array.
4710 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
4711 uint debug_end = debug_ptr;
4712 uint debug_start = debug_ptr - in_jvms->debug_size();
4713 debug_ptr = debug_start; // back up the ptr
4714
4715 uint p = debug_start; // walks forward in [debug_start, debug_end)
4716 uint j, k, l;
4717 SafePointNode* in_map = in_jvms->map();
4718 out_jvms->set_map(call);
4719
4720 if (can_prune_locals) {
4721 assert(in_jvms->method() == out_jvms->method(), "sanity");
4722 // If the current throw can reach an exception handler in this JVMS,
4723 // then we must keep everything live that can reach that handler.
4724 // As a quick and dirty approximation, we look for any handlers at all.
4725 if (in_jvms->method()->has_exception_handlers()) {
4726 can_prune_locals = false;
4727 }
4728 }
4729
4730 // Add the Locals
4731 k = in_jvms->locoff();
4732 l = in_jvms->loc_size();
4733 out_jvms->set_locoff(p);
4734 if (!can_prune_locals) {
4735 for (j = 0; j < l; j++)
4736 call->set_req(p++, in_map->in(k+j));
4737 } else {
4738 p += l; // already set to top above by add_req_batch
4739 }
4740
4741 // Add the Expression Stack
4742 k = in_jvms->stkoff();
4743 l = in_jvms->sp();
4744 out_jvms->set_stkoff(p);
4745 if (!can_prune_locals) {
4746 for (j = 0; j < l; j++)
4747 call->set_req(p++, in_map->in(k+j));
4748 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
4749 // Divide stack into {S0,...,S1}, where S0 is set to top.
4750 uint s1 = stack_slots_not_pruned;
4751 stack_slots_not_pruned = 0; // for next iteration
4752 if (s1 > l) s1 = l;
4753 uint s0 = l - s1;
4754 p += s0; // skip the tops preinstalled by add_req_batch
4755 for (j = s0; j < l; j++)
4756 call->set_req(p++, in_map->in(k+j));
4757 } else {
4758 p += l; // already set to top above by add_req_batch
4759 }
4760
4761 // Add the Monitors
4762 k = in_jvms->monoff();
4763 l = in_jvms->mon_size();
4764 out_jvms->set_monoff(p);
4765 for (j = 0; j < l; j++)
4766 call->set_req(p++, in_map->in(k+j));
4767
4768 // Copy any scalar object fields.
4769 k = in_jvms->scloff();
4770 l = in_jvms->scl_size();
4771 out_jvms->set_scloff(p);
4772 for (j = 0; j < l; j++)
4773 call->set_req(p++, in_map->in(k+j));
4774
4775 // Finish the new jvms.
4776 out_jvms->set_endoff(p);
4777
4778 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
4779 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
4780 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
4781 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
4782 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
4783 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
4784
4785 // Update the two tail pointers in parallel.
4786 out_jvms = out_jvms->caller();
4787 in_jvms = in_jvms->caller();
4788 }
4789
4790 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
4791
4792 // Test the correctness of JVMState::debug_xxx accessors:
4793 assert(call->jvms()->debug_start() == non_debug_edges, "");
4794 assert(call->jvms()->debug_end() == call->req(), "");
4795 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
4796 }