1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vectornode.hpp"
71 #include "runtime/arguments.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stubRoutines.hpp"
75 #include "runtime/timer.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/copy.hpp"
78 #if INCLUDE_ALL_GCS
79 #include "gc/g1/g1ThreadLocalData.hpp"
80 #endif // INCLUDE_ALL_GCS
81
82
83 // -------------------- Compile::mach_constant_base_node -----------------------
84 // Constant table base node singleton.
85 MachConstantBaseNode* Compile::mach_constant_base_node() {
86 if (_mach_constant_base_node == NULL) {
87 _mach_constant_base_node = new MachConstantBaseNode();
88 _mach_constant_base_node->add_req(C->root());
89 }
90 return _mach_constant_base_node;
91 }
92
93
94 /// Support for intrinsics.
95
96 // Return the index at which m must be inserted (or already exists).
97 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
98 class IntrinsicDescPair {
99 private:
100 ciMethod* _m;
101 bool _is_virtual;
102 public:
103 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
104 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
105 ciMethod* m= elt->method();
106 ciMethod* key_m = key->_m;
107 if (key_m < m) return -1;
108 else if (key_m > m) return 1;
109 else {
110 bool is_virtual = elt->is_virtual();
111 bool key_virtual = key->_is_virtual;
112 if (key_virtual < is_virtual) return -1;
113 else if (key_virtual > is_virtual) return 1;
114 else return 0;
115 }
116 }
117 };
118 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
119 #ifdef ASSERT
120 for (int i = 1; i < _intrinsics->length(); i++) {
121 CallGenerator* cg1 = _intrinsics->at(i-1);
122 CallGenerator* cg2 = _intrinsics->at(i);
123 assert(cg1->method() != cg2->method()
124 ? cg1->method() < cg2->method()
125 : cg1->is_virtual() < cg2->is_virtual(),
126 "compiler intrinsics list must stay sorted");
127 }
128 #endif
129 IntrinsicDescPair pair(m, is_virtual);
130 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
131 }
132
133 void Compile::register_intrinsic(CallGenerator* cg) {
134 if (_intrinsics == NULL) {
135 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
136 }
137 int len = _intrinsics->length();
138 bool found = false;
139 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
140 assert(!found, "registering twice");
141 _intrinsics->insert_before(index, cg);
142 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
143 }
144
145 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
146 assert(m->is_loaded(), "don't try this on unloaded methods");
147 if (_intrinsics != NULL) {
148 bool found = false;
149 int index = intrinsic_insertion_index(m, is_virtual, found);
150 if (found) {
151 return _intrinsics->at(index);
152 }
153 }
154 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
155 if (m->intrinsic_id() != vmIntrinsics::_none &&
156 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
157 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
158 if (cg != NULL) {
159 // Save it for next time:
160 register_intrinsic(cg);
161 return cg;
162 } else {
163 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
164 }
165 }
166 return NULL;
167 }
168
169 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
170 // in library_call.cpp.
171
172
173 #ifndef PRODUCT
174 // statistics gathering...
175
176 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
178
179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
181 int oflags = _intrinsic_hist_flags[id];
182 assert(flags != 0, "what happened?");
183 if (is_virtual) {
184 flags |= _intrinsic_virtual;
185 }
186 bool changed = (flags != oflags);
187 if ((flags & _intrinsic_worked) != 0) {
188 juint count = (_intrinsic_hist_count[id] += 1);
189 if (count == 1) {
190 changed = true; // first time
191 }
192 // increment the overall count also:
193 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
194 }
195 if (changed) {
196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
197 // Something changed about the intrinsic's virtuality.
198 if ((flags & _intrinsic_virtual) != 0) {
199 // This is the first use of this intrinsic as a virtual call.
200 if (oflags != 0) {
201 // We already saw it as a non-virtual, so note both cases.
202 flags |= _intrinsic_both;
203 }
204 } else if ((oflags & _intrinsic_both) == 0) {
205 // This is the first use of this intrinsic as a non-virtual
206 flags |= _intrinsic_both;
207 }
208 }
209 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
210 }
211 // update the overall flags also:
212 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
213 return changed;
214 }
215
216 static char* format_flags(int flags, char* buf) {
217 buf[0] = 0;
218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
223 if (buf[0] == 0) strcat(buf, ",");
224 assert(buf[0] == ',', "must be");
225 return &buf[1];
226 }
227
228 void Compile::print_intrinsic_statistics() {
229 char flagsbuf[100];
230 ttyLocker ttyl;
231 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
232 tty->print_cr("Compiler intrinsic usage:");
233 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
234 if (total == 0) total = 1; // avoid div0 in case of no successes
235 #define PRINT_STAT_LINE(name, c, f) \
236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
237 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
238 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
239 int flags = _intrinsic_hist_flags[id];
240 juint count = _intrinsic_hist_count[id];
241 if ((flags | count) != 0) {
242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
243 }
244 }
245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
246 if (xtty != NULL) xtty->tail("statistics");
247 }
248
249 void Compile::print_statistics() {
250 { ttyLocker ttyl;
251 if (xtty != NULL) xtty->head("statistics type='opto'");
252 Parse::print_statistics();
253 PhaseCCP::print_statistics();
254 PhaseRegAlloc::print_statistics();
255 Scheduling::print_statistics();
256 PhasePeephole::print_statistics();
257 PhaseIdealLoop::print_statistics();
258 if (xtty != NULL) xtty->tail("statistics");
259 }
260 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
261 // put this under its own <statistics> element.
262 print_intrinsic_statistics();
263 }
264 }
265 #endif //PRODUCT
266
267 // Support for bundling info
268 Bundle* Compile::node_bundling(const Node *n) {
269 assert(valid_bundle_info(n), "oob");
270 return &_node_bundling_base[n->_idx];
271 }
272
273 bool Compile::valid_bundle_info(const Node *n) {
274 return (_node_bundling_limit > n->_idx);
275 }
276
277
278 void Compile::gvn_replace_by(Node* n, Node* nn) {
279 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
280 Node* use = n->last_out(i);
281 bool is_in_table = initial_gvn()->hash_delete(use);
282 uint uses_found = 0;
283 for (uint j = 0; j < use->len(); j++) {
284 if (use->in(j) == n) {
285 if (j < use->req())
286 use->set_req(j, nn);
287 else
288 use->set_prec(j, nn);
289 uses_found++;
290 }
291 }
292 if (is_in_table) {
293 // reinsert into table
294 initial_gvn()->hash_find_insert(use);
295 }
296 record_for_igvn(use);
297 i -= uses_found; // we deleted 1 or more copies of this edge
298 }
299 }
300
301
302 static inline bool not_a_node(const Node* n) {
303 if (n == NULL) return true;
304 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
305 if (*(address*)n == badAddress) return true; // kill by Node::destruct
306 return false;
307 }
308
309 // Identify all nodes that are reachable from below, useful.
310 // Use breadth-first pass that records state in a Unique_Node_List,
311 // recursive traversal is slower.
312 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
313 int estimated_worklist_size = live_nodes();
314 useful.map( estimated_worklist_size, NULL ); // preallocate space
315
316 // Initialize worklist
317 if (root() != NULL) { useful.push(root()); }
318 // If 'top' is cached, declare it useful to preserve cached node
319 if( cached_top_node() ) { useful.push(cached_top_node()); }
320
321 // Push all useful nodes onto the list, breadthfirst
322 for( uint next = 0; next < useful.size(); ++next ) {
323 assert( next < unique(), "Unique useful nodes < total nodes");
324 Node *n = useful.at(next);
325 uint max = n->len();
326 for( uint i = 0; i < max; ++i ) {
327 Node *m = n->in(i);
328 if (not_a_node(m)) continue;
329 useful.push(m);
330 }
331 }
332 }
333
334 // Update dead_node_list with any missing dead nodes using useful
335 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
336 void Compile::update_dead_node_list(Unique_Node_List &useful) {
337 uint max_idx = unique();
338 VectorSet& useful_node_set = useful.member_set();
339
340 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
341 // If node with index node_idx is not in useful set,
342 // mark it as dead in dead node list.
343 if (! useful_node_set.test(node_idx) ) {
344 record_dead_node(node_idx);
345 }
346 }
347 }
348
349 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
350 int shift = 0;
351 for (int i = 0; i < inlines->length(); i++) {
352 CallGenerator* cg = inlines->at(i);
353 CallNode* call = cg->call_node();
354 if (shift > 0) {
355 inlines->at_put(i-shift, cg);
356 }
357 if (!useful.member(call)) {
358 shift++;
359 }
360 }
361 inlines->trunc_to(inlines->length()-shift);
362 }
363
364 // Disconnect all useless nodes by disconnecting those at the boundary.
365 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
366 uint next = 0;
367 while (next < useful.size()) {
368 Node *n = useful.at(next++);
369 if (n->is_SafePoint()) {
370 // We're done with a parsing phase. Replaced nodes are not valid
371 // beyond that point.
372 n->as_SafePoint()->delete_replaced_nodes();
373 }
374 // Use raw traversal of out edges since this code removes out edges
375 int max = n->outcnt();
376 for (int j = 0; j < max; ++j) {
377 Node* child = n->raw_out(j);
378 if (! useful.member(child)) {
379 assert(!child->is_top() || child != top(),
380 "If top is cached in Compile object it is in useful list");
381 // Only need to remove this out-edge to the useless node
382 n->raw_del_out(j);
383 --j;
384 --max;
385 }
386 }
387 if (n->outcnt() == 1 && n->has_special_unique_user()) {
388 record_for_igvn(n->unique_out());
389 }
390 }
391 // Remove useless macro and predicate opaq nodes
392 for (int i = C->macro_count()-1; i >= 0; i--) {
393 Node* n = C->macro_node(i);
394 if (!useful.member(n)) {
395 remove_macro_node(n);
396 }
397 }
398 // Remove useless CastII nodes with range check dependency
399 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
400 Node* cast = range_check_cast_node(i);
401 if (!useful.member(cast)) {
402 remove_range_check_cast(cast);
403 }
404 }
405 // Remove useless expensive nodes
406 for (int i = C->expensive_count()-1; i >= 0; i--) {
407 Node* n = C->expensive_node(i);
408 if (!useful.member(n)) {
409 remove_expensive_node(n);
410 }
411 }
412 // Remove useless Opaque4 nodes
413 for (int i = opaque4_count() - 1; i >= 0; i--) {
414 Node* opaq = opaque4_node(i);
415 if (!useful.member(opaq)) {
416 remove_opaque4_node(opaq);
417 }
418 }
419 // clean up the late inline lists
420 remove_useless_late_inlines(&_string_late_inlines, useful);
421 remove_useless_late_inlines(&_boxing_late_inlines, useful);
422 remove_useless_late_inlines(&_late_inlines, useful);
423 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
424 }
425
426 //------------------------------frame_size_in_words-----------------------------
427 // frame_slots in units of words
428 int Compile::frame_size_in_words() const {
429 // shift is 0 in LP32 and 1 in LP64
430 const int shift = (LogBytesPerWord - LogBytesPerInt);
431 int words = _frame_slots >> shift;
432 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
433 return words;
434 }
435
436 // To bang the stack of this compiled method we use the stack size
437 // that the interpreter would need in case of a deoptimization. This
438 // removes the need to bang the stack in the deoptimization blob which
439 // in turn simplifies stack overflow handling.
440 int Compile::bang_size_in_bytes() const {
441 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
442 }
443
444 // ============================================================================
445 //------------------------------CompileWrapper---------------------------------
446 class CompileWrapper : public StackObj {
447 Compile *const _compile;
448 public:
449 CompileWrapper(Compile* compile);
450
451 ~CompileWrapper();
452 };
453
454 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
455 // the Compile* pointer is stored in the current ciEnv:
456 ciEnv* env = compile->env();
457 assert(env == ciEnv::current(), "must already be a ciEnv active");
458 assert(env->compiler_data() == NULL, "compile already active?");
459 env->set_compiler_data(compile);
460 assert(compile == Compile::current(), "sanity");
461
462 compile->set_type_dict(NULL);
463 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
464 compile->clone_map().set_clone_idx(0);
465 compile->set_type_hwm(NULL);
466 compile->set_type_last_size(0);
467 compile->set_last_tf(NULL, NULL);
468 compile->set_indexSet_arena(NULL);
469 compile->set_indexSet_free_block_list(NULL);
470 compile->init_type_arena();
471 Type::Initialize(compile);
472 _compile->set_scratch_buffer_blob(NULL);
473 _compile->begin_method();
474 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
475 }
476 CompileWrapper::~CompileWrapper() {
477 _compile->end_method();
478 if (_compile->scratch_buffer_blob() != NULL)
479 BufferBlob::free(_compile->scratch_buffer_blob());
480 _compile->env()->set_compiler_data(NULL);
481 }
482
483
484 //----------------------------print_compile_messages---------------------------
485 void Compile::print_compile_messages() {
486 #ifndef PRODUCT
487 // Check if recompiling
488 if (_subsume_loads == false && PrintOpto) {
489 // Recompiling without allowing machine instructions to subsume loads
490 tty->print_cr("*********************************************************");
491 tty->print_cr("** Bailout: Recompile without subsuming loads **");
492 tty->print_cr("*********************************************************");
493 }
494 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
495 // Recompiling without escape analysis
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without escape analysis **");
498 tty->print_cr("*********************************************************");
499 }
500 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
501 // Recompiling without boxing elimination
502 tty->print_cr("*********************************************************");
503 tty->print_cr("** Bailout: Recompile without boxing elimination **");
504 tty->print_cr("*********************************************************");
505 }
506 if (C->directive()->BreakAtCompileOption) {
507 // Open the debugger when compiling this method.
508 tty->print("### Breaking when compiling: ");
509 method()->print_short_name();
510 tty->cr();
511 BREAKPOINT;
512 }
513
514 if( PrintOpto ) {
515 if (is_osr_compilation()) {
516 tty->print("[OSR]%3d", _compile_id);
517 } else {
518 tty->print("%3d", _compile_id);
519 }
520 }
521 #endif
522 }
523
524
525 //-----------------------init_scratch_buffer_blob------------------------------
526 // Construct a temporary BufferBlob and cache it for this compile.
527 void Compile::init_scratch_buffer_blob(int const_size) {
528 // If there is already a scratch buffer blob allocated and the
529 // constant section is big enough, use it. Otherwise free the
530 // current and allocate a new one.
531 BufferBlob* blob = scratch_buffer_blob();
532 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
533 // Use the current blob.
534 } else {
535 if (blob != NULL) {
536 BufferBlob::free(blob);
537 }
538
539 ResourceMark rm;
540 _scratch_const_size = const_size;
541 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
542 blob = BufferBlob::create("Compile::scratch_buffer", size);
543 // Record the buffer blob for next time.
544 set_scratch_buffer_blob(blob);
545 // Have we run out of code space?
546 if (scratch_buffer_blob() == NULL) {
547 // Let CompilerBroker disable further compilations.
548 record_failure("Not enough space for scratch buffer in CodeCache");
549 return;
550 }
551 }
552
553 // Initialize the relocation buffers
554 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
555 set_scratch_locs_memory(locs_buf);
556 }
557
558
559 //-----------------------scratch_emit_size-------------------------------------
560 // Helper function that computes size by emitting code
561 uint Compile::scratch_emit_size(const Node* n) {
562 // Start scratch_emit_size section.
563 set_in_scratch_emit_size(true);
564
565 // Emit into a trash buffer and count bytes emitted.
566 // This is a pretty expensive way to compute a size,
567 // but it works well enough if seldom used.
568 // All common fixed-size instructions are given a size
569 // method by the AD file.
570 // Note that the scratch buffer blob and locs memory are
571 // allocated at the beginning of the compile task, and
572 // may be shared by several calls to scratch_emit_size.
573 // The allocation of the scratch buffer blob is particularly
574 // expensive, since it has to grab the code cache lock.
575 BufferBlob* blob = this->scratch_buffer_blob();
576 assert(blob != NULL, "Initialize BufferBlob at start");
577 assert(blob->size() > MAX_inst_size, "sanity");
578 relocInfo* locs_buf = scratch_locs_memory();
579 address blob_begin = blob->content_begin();
580 address blob_end = (address)locs_buf;
581 assert(blob->contains(blob_end), "sanity");
582 CodeBuffer buf(blob_begin, blob_end - blob_begin);
583 buf.initialize_consts_size(_scratch_const_size);
584 buf.initialize_stubs_size(MAX_stubs_size);
585 assert(locs_buf != NULL, "sanity");
586 int lsize = MAX_locs_size / 3;
587 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
588 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
589 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
590 // Mark as scratch buffer.
591 buf.consts()->set_scratch_emit();
592 buf.insts()->set_scratch_emit();
593 buf.stubs()->set_scratch_emit();
594
595 // Do the emission.
596
597 Label fakeL; // Fake label for branch instructions.
598 Label* saveL = NULL;
599 uint save_bnum = 0;
600 bool is_branch = n->is_MachBranch();
601 if (is_branch) {
602 MacroAssembler masm(&buf);
603 masm.bind(fakeL);
604 n->as_MachBranch()->save_label(&saveL, &save_bnum);
605 n->as_MachBranch()->label_set(&fakeL, 0);
606 }
607 n->emit(buf, this->regalloc());
608
609 // Emitting into the scratch buffer should not fail
610 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
611
612 if (is_branch) // Restore label.
613 n->as_MachBranch()->label_set(saveL, save_bnum);
614
615 // End scratch_emit_size section.
616 set_in_scratch_emit_size(false);
617
618 return buf.insts_size();
619 }
620
621
622 // ============================================================================
623 //------------------------------Compile standard-------------------------------
624 debug_only( int Compile::_debug_idx = 100000; )
625
626 // Compile a method. entry_bci is -1 for normal compilations and indicates
627 // the continuation bci for on stack replacement.
628
629
630 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
631 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
632 : Phase(Compiler),
633 _env(ci_env),
634 _directive(directive),
635 _log(ci_env->log()),
636 _compile_id(ci_env->compile_id()),
637 _save_argument_registers(false),
638 _stub_name(NULL),
639 _stub_function(NULL),
640 _stub_entry_point(NULL),
641 _method(target),
642 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state()),
643 _entry_bci(osr_bci),
644 _initial_gvn(NULL),
645 _for_igvn(NULL),
646 _warm_calls(NULL),
647 _subsume_loads(subsume_loads),
648 _do_escape_analysis(do_escape_analysis),
649 _eliminate_boxing(eliminate_boxing),
650 _failure_reason(NULL),
651 _code_buffer("Compile::Fill_buffer"),
652 _orig_pc_slot(0),
653 _orig_pc_slot_offset_in_bytes(0),
654 _has_method_handle_invokes(false),
655 _mach_constant_base_node(NULL),
656 _node_bundling_limit(0),
657 _node_bundling_base(NULL),
658 _java_calls(0),
659 _inner_loops(0),
660 _scratch_const_size(-1),
661 _in_scratch_emit_size(false),
662 _dead_node_list(comp_arena()),
663 _dead_node_count(0),
664 #ifndef PRODUCT
665 _trace_opto_output(directive->TraceOptoOutputOption),
666 _in_dump_cnt(0),
667 _printer(IdealGraphPrinter::printer()),
668 #endif
669 _congraph(NULL),
670 _comp_arena(mtCompiler),
671 _node_arena(mtCompiler),
672 _old_arena(mtCompiler),
673 _Compile_types(mtCompiler),
674 _replay_inline_data(NULL),
675 _late_inlines(comp_arena(), 2, 0, NULL),
676 _string_late_inlines(comp_arena(), 2, 0, NULL),
677 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
678 _late_inlines_pos(0),
679 _number_of_mh_late_inlines(0),
680 _inlining_progress(false),
681 _inlining_incrementally(false),
682 _print_inlining_list(NULL),
683 _print_inlining_stream(NULL),
684 _print_inlining_idx(0),
685 _print_inlining_output(NULL),
686 _interpreter_frame_size(0),
687 _max_node_limit(MaxNodeLimit),
688 _has_reserved_stack_access(target->has_reserved_stack_access()) {
689 C = this;
690 #ifndef PRODUCT
691 if (_printer != NULL) {
692 _printer->set_compile(this);
693 }
694 #endif
695 CompileWrapper cw(this);
696
697 if (CITimeVerbose) {
698 tty->print(" ");
699 target->holder()->name()->print();
700 tty->print(".");
701 target->print_short_name();
702 tty->print(" ");
703 }
704 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
705 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
706
707 #ifndef PRODUCT
708 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
709 if (!print_opto_assembly) {
710 bool print_assembly = directive->PrintAssemblyOption;
711 if (print_assembly && !Disassembler::can_decode()) {
712 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
713 print_opto_assembly = true;
714 }
715 }
716 set_print_assembly(print_opto_assembly);
717 set_parsed_irreducible_loop(false);
718
719 if (directive->ReplayInlineOption) {
720 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
721 }
722 #endif
723 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
724 set_print_intrinsics(directive->PrintIntrinsicsOption);
725 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
726
727 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
728 // Make sure the method being compiled gets its own MDO,
729 // so we can at least track the decompile_count().
730 // Need MDO to record RTM code generation state.
731 method()->ensure_method_data();
732 }
733
734 Init(::AliasLevel);
735
736
737 print_compile_messages();
738
739 _ilt = InlineTree::build_inline_tree_root();
740
741 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
742 assert(num_alias_types() >= AliasIdxRaw, "");
743
744 #define MINIMUM_NODE_HASH 1023
745 // Node list that Iterative GVN will start with
746 Unique_Node_List for_igvn(comp_arena());
747 set_for_igvn(&for_igvn);
748
749 // GVN that will be run immediately on new nodes
750 uint estimated_size = method()->code_size()*4+64;
751 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
752 PhaseGVN gvn(node_arena(), estimated_size);
753 set_initial_gvn(&gvn);
754
755 print_inlining_init();
756 { // Scope for timing the parser
757 TracePhase tp("parse", &timers[_t_parser]);
758
759 // Put top into the hash table ASAP.
760 initial_gvn()->transform_no_reclaim(top());
761
762 // Set up tf(), start(), and find a CallGenerator.
763 CallGenerator* cg = NULL;
764 if (is_osr_compilation()) {
765 const TypeTuple *domain = StartOSRNode::osr_domain();
766 const TypeTuple *range = TypeTuple::make_range(method()->signature());
767 init_tf(TypeFunc::make(domain, range));
768 StartNode* s = new StartOSRNode(root(), domain);
769 initial_gvn()->set_type_bottom(s);
770 init_start(s);
771 cg = CallGenerator::for_osr(method(), entry_bci());
772 } else {
773 // Normal case.
774 init_tf(TypeFunc::make(method()));
775 StartNode* s = new StartNode(root(), tf()->domain());
776 initial_gvn()->set_type_bottom(s);
777 init_start(s);
778 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
779 // With java.lang.ref.reference.get() we must go through the
780 // intrinsic - even when get() is the root
781 // method of the compile - so that, if necessary, the value in
782 // the referent field of the reference object gets recorded by
783 // the pre-barrier code.
784 cg = find_intrinsic(method(), false);
785 }
786 if (cg == NULL) {
787 float past_uses = method()->interpreter_invocation_count();
788 float expected_uses = past_uses;
789 cg = CallGenerator::for_inline(method(), expected_uses);
790 }
791 }
792 if (failing()) return;
793 if (cg == NULL) {
794 record_method_not_compilable("cannot parse method");
795 return;
796 }
797 JVMState* jvms = build_start_state(start(), tf());
798 if ((jvms = cg->generate(jvms)) == NULL) {
799 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
800 record_method_not_compilable("method parse failed");
801 }
802 return;
803 }
804 GraphKit kit(jvms);
805
806 if (!kit.stopped()) {
807 // Accept return values, and transfer control we know not where.
808 // This is done by a special, unique ReturnNode bound to root.
809 return_values(kit.jvms());
810 }
811
812 if (kit.has_exceptions()) {
813 // Any exceptions that escape from this call must be rethrown
814 // to whatever caller is dynamically above us on the stack.
815 // This is done by a special, unique RethrowNode bound to root.
816 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
817 }
818
819 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
820
821 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
822 inline_string_calls(true);
823 }
824
825 if (failing()) return;
826
827 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
828
829 // Remove clutter produced by parsing.
830 if (!failing()) {
831 ResourceMark rm;
832 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
833 }
834 }
835
836 // Note: Large methods are capped off in do_one_bytecode().
837 if (failing()) return;
838
839 // After parsing, node notes are no longer automagic.
840 // They must be propagated by register_new_node_with_optimizer(),
841 // clone(), or the like.
842 set_default_node_notes(NULL);
843
844 for (;;) {
845 int successes = Inline_Warm();
846 if (failing()) return;
847 if (successes == 0) break;
848 }
849
850 // Drain the list.
851 Finish_Warm();
852 #ifndef PRODUCT
853 if (_printer && _printer->should_print(1)) {
854 _printer->print_inlining();
855 }
856 #endif
857
858 if (failing()) return;
859 NOT_PRODUCT( verify_graph_edges(); )
860
861 // Now optimize
862 Optimize();
863 if (failing()) return;
864 NOT_PRODUCT( verify_graph_edges(); )
865
866 #ifndef PRODUCT
867 if (PrintIdeal) {
868 ttyLocker ttyl; // keep the following output all in one block
869 // This output goes directly to the tty, not the compiler log.
870 // To enable tools to match it up with the compilation activity,
871 // be sure to tag this tty output with the compile ID.
872 if (xtty != NULL) {
873 xtty->head("ideal compile_id='%d'%s", compile_id(),
874 is_osr_compilation() ? " compile_kind='osr'" :
875 "");
876 }
877 root()->dump(9999);
878 if (xtty != NULL) {
879 xtty->tail("ideal");
880 }
881 }
882 #endif
883
884 NOT_PRODUCT( verify_barriers(); )
885
886 // Dump compilation data to replay it.
887 if (directive->DumpReplayOption) {
888 env()->dump_replay_data(_compile_id);
889 }
890 if (directive->DumpInlineOption && (ilt() != NULL)) {
891 env()->dump_inline_data(_compile_id);
892 }
893
894 // Now that we know the size of all the monitors we can add a fixed slot
895 // for the original deopt pc.
896
897 _orig_pc_slot = fixed_slots();
898 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
899 set_fixed_slots(next_slot);
900
901 // Compute when to use implicit null checks. Used by matching trap based
902 // nodes and NullCheck optimization.
903 set_allowed_deopt_reasons();
904
905 // Now generate code
906 Code_Gen();
907 if (failing()) return;
908
909 // Check if we want to skip execution of all compiled code.
910 {
911 #ifndef PRODUCT
912 if (OptoNoExecute) {
913 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
914 return;
915 }
916 #endif
917 TracePhase tp("install_code", &timers[_t_registerMethod]);
918
919 if (is_osr_compilation()) {
920 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
921 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
922 } else {
923 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
924 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
925 }
926
927 env()->register_method(_method, _entry_bci,
928 &_code_offsets,
929 _orig_pc_slot_offset_in_bytes,
930 code_buffer(),
931 frame_size_in_words(), _oop_map_set,
932 &_handler_table, &_inc_table,
933 compiler,
934 has_unsafe_access(),
935 SharedRuntime::is_wide_vector(max_vector_size()),
936 rtm_state()
937 );
938
939 if (log() != NULL) // Print code cache state into compiler log
940 log()->code_cache_state();
941 }
942 }
943
944 //------------------------------Compile----------------------------------------
945 // Compile a runtime stub
946 Compile::Compile( ciEnv* ci_env,
947 TypeFunc_generator generator,
948 address stub_function,
949 const char *stub_name,
950 int is_fancy_jump,
951 bool pass_tls,
952 bool save_arg_registers,
953 bool return_pc,
954 DirectiveSet* directive)
955 : Phase(Compiler),
956 _env(ci_env),
957 _directive(directive),
958 _log(ci_env->log()),
959 _compile_id(0),
960 _save_argument_registers(save_arg_registers),
961 _method(NULL),
962 _stub_name(stub_name),
963 _stub_function(stub_function),
964 _stub_entry_point(NULL),
965 _entry_bci(InvocationEntryBci),
966 _initial_gvn(NULL),
967 _for_igvn(NULL),
968 _warm_calls(NULL),
969 _orig_pc_slot(0),
970 _orig_pc_slot_offset_in_bytes(0),
971 _subsume_loads(true),
972 _do_escape_analysis(false),
973 _eliminate_boxing(false),
974 _failure_reason(NULL),
975 _code_buffer("Compile::Fill_buffer"),
976 _has_method_handle_invokes(false),
977 _mach_constant_base_node(NULL),
978 _node_bundling_limit(0),
979 _node_bundling_base(NULL),
980 _java_calls(0),
981 _inner_loops(0),
982 #ifndef PRODUCT
983 _trace_opto_output(directive->TraceOptoOutputOption),
984 _in_dump_cnt(0),
985 _printer(NULL),
986 #endif
987 _comp_arena(mtCompiler),
988 _node_arena(mtCompiler),
989 _old_arena(mtCompiler),
990 _Compile_types(mtCompiler),
991 _dead_node_list(comp_arena()),
992 _dead_node_count(0),
993 _congraph(NULL),
994 _replay_inline_data(NULL),
995 _number_of_mh_late_inlines(0),
996 _inlining_progress(false),
997 _inlining_incrementally(false),
998 _print_inlining_list(NULL),
999 _print_inlining_stream(NULL),
1000 _print_inlining_idx(0),
1001 _print_inlining_output(NULL),
1002 _allowed_reasons(0),
1003 _interpreter_frame_size(0),
1004 _max_node_limit(MaxNodeLimit),
1005 _has_reserved_stack_access(false) {
1006 C = this;
1007
1008 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1009 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1010
1011 #ifndef PRODUCT
1012 set_print_assembly(PrintFrameConverterAssembly);
1013 set_parsed_irreducible_loop(false);
1014 #endif
1015 set_has_irreducible_loop(false); // no loops
1016
1017 CompileWrapper cw(this);
1018 Init(/*AliasLevel=*/ 0);
1019 init_tf((*generator)());
1020
1021 {
1022 // The following is a dummy for the sake of GraphKit::gen_stub
1023 Unique_Node_List for_igvn(comp_arena());
1024 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1025 PhaseGVN gvn(Thread::current()->resource_area(),255);
1026 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1027 gvn.transform_no_reclaim(top());
1028
1029 GraphKit kit;
1030 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1031 }
1032
1033 NOT_PRODUCT( verify_graph_edges(); )
1034 Code_Gen();
1035 if (failing()) return;
1036
1037
1038 // Entry point will be accessed using compile->stub_entry_point();
1039 if (code_buffer() == NULL) {
1040 Matcher::soft_match_failure();
1041 } else {
1042 if (PrintAssembly && (WizardMode || Verbose))
1043 tty->print_cr("### Stub::%s", stub_name);
1044
1045 if (!failing()) {
1046 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1047
1048 // Make the NMethod
1049 // For now we mark the frame as never safe for profile stackwalking
1050 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1051 code_buffer(),
1052 CodeOffsets::frame_never_safe,
1053 // _code_offsets.value(CodeOffsets::Frame_Complete),
1054 frame_size_in_words(),
1055 _oop_map_set,
1056 save_arg_registers);
1057 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1058
1059 _stub_entry_point = rs->entry_point();
1060 }
1061 }
1062 }
1063
1064 //------------------------------Init-------------------------------------------
1065 // Prepare for a single compilation
1066 void Compile::Init(int aliaslevel) {
1067 _unique = 0;
1068 _regalloc = NULL;
1069
1070 _tf = NULL; // filled in later
1071 _top = NULL; // cached later
1072 _matcher = NULL; // filled in later
1073 _cfg = NULL; // filled in later
1074
1075 set_24_bit_selection_and_mode(Use24BitFP, false);
1076
1077 _node_note_array = NULL;
1078 _default_node_notes = NULL;
1079 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1080
1081 _immutable_memory = NULL; // filled in at first inquiry
1082
1083 // Globally visible Nodes
1084 // First set TOP to NULL to give safe behavior during creation of RootNode
1085 set_cached_top_node(NULL);
1086 set_root(new RootNode());
1087 // Now that you have a Root to point to, create the real TOP
1088 set_cached_top_node( new ConNode(Type::TOP) );
1089 set_recent_alloc(NULL, NULL);
1090
1091 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1092 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1093 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1094 env()->set_dependencies(new Dependencies(env()));
1095
1096 _fixed_slots = 0;
1097 set_has_split_ifs(false);
1098 set_has_loops(has_method() && method()->has_loops()); // first approximation
1099 set_has_stringbuilder(false);
1100 set_has_boxed_value(false);
1101 _trap_can_recompile = false; // no traps emitted yet
1102 _major_progress = true; // start out assuming good things will happen
1103 set_has_unsafe_access(false);
1104 set_max_vector_size(0);
1105 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1106 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1107 set_decompile_count(0);
1108
1109 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1110 set_num_loop_opts(LoopOptsCount);
1111 set_do_inlining(Inline);
1112 set_max_inline_size(MaxInlineSize);
1113 set_freq_inline_size(FreqInlineSize);
1114 set_do_scheduling(OptoScheduling);
1115 set_do_count_invocations(false);
1116 set_do_method_data_update(false);
1117
1118 set_do_vector_loop(false);
1119
1120 if (AllowVectorizeOnDemand) {
1121 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1122 set_do_vector_loop(true);
1123 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());})
1124 } else if (has_method() && method()->name() != 0 &&
1125 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1126 set_do_vector_loop(true);
1127 }
1128 }
1129 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1130 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());})
1131
1132 set_age_code(has_method() && method()->profile_aging());
1133 set_rtm_state(NoRTM); // No RTM lock eliding by default
1134 _max_node_limit = _directive->MaxNodeLimitOption;
1135
1136 #if INCLUDE_RTM_OPT
1137 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1138 int rtm_state = method()->method_data()->rtm_state();
1139 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1140 // Don't generate RTM lock eliding code.
1141 set_rtm_state(NoRTM);
1142 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1143 // Generate RTM lock eliding code without abort ratio calculation code.
1144 set_rtm_state(UseRTM);
1145 } else if (UseRTMDeopt) {
1146 // Generate RTM lock eliding code and include abort ratio calculation
1147 // code if UseRTMDeopt is on.
1148 set_rtm_state(ProfileRTM);
1149 }
1150 }
1151 #endif
1152 if (debug_info()->recording_non_safepoints()) {
1153 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1154 (comp_arena(), 8, 0, NULL));
1155 set_default_node_notes(Node_Notes::make(this));
1156 }
1157
1158 // // -- Initialize types before each compile --
1159 // // Update cached type information
1160 // if( _method && _method->constants() )
1161 // Type::update_loaded_types(_method, _method->constants());
1162
1163 // Init alias_type map.
1164 if (!_do_escape_analysis && aliaslevel == 3)
1165 aliaslevel = 2; // No unique types without escape analysis
1166 _AliasLevel = aliaslevel;
1167 const int grow_ats = 16;
1168 _max_alias_types = grow_ats;
1169 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1170 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1171 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1172 {
1173 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1174 }
1175 // Initialize the first few types.
1176 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1177 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1178 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1179 _num_alias_types = AliasIdxRaw+1;
1180 // Zero out the alias type cache.
1181 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1182 // A NULL adr_type hits in the cache right away. Preload the right answer.
1183 probe_alias_cache(NULL)->_index = AliasIdxTop;
1184
1185 _intrinsics = NULL;
1186 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1187 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1188 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1189 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1190 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1191 register_library_intrinsics();
1192 }
1193
1194 //---------------------------init_start----------------------------------------
1195 // Install the StartNode on this compile object.
1196 void Compile::init_start(StartNode* s) {
1197 if (failing())
1198 return; // already failing
1199 assert(s == start(), "");
1200 }
1201
1202 /**
1203 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1204 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1205 * the ideal graph.
1206 */
1207 StartNode* Compile::start() const {
1208 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1209 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1210 Node* start = root()->fast_out(i);
1211 if (start->is_Start()) {
1212 return start->as_Start();
1213 }
1214 }
1215 fatal("Did not find Start node!");
1216 return NULL;
1217 }
1218
1219 //-------------------------------immutable_memory-------------------------------------
1220 // Access immutable memory
1221 Node* Compile::immutable_memory() {
1222 if (_immutable_memory != NULL) {
1223 return _immutable_memory;
1224 }
1225 StartNode* s = start();
1226 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1227 Node *p = s->fast_out(i);
1228 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1229 _immutable_memory = p;
1230 return _immutable_memory;
1231 }
1232 }
1233 ShouldNotReachHere();
1234 return NULL;
1235 }
1236
1237 //----------------------set_cached_top_node------------------------------------
1238 // Install the cached top node, and make sure Node::is_top works correctly.
1239 void Compile::set_cached_top_node(Node* tn) {
1240 if (tn != NULL) verify_top(tn);
1241 Node* old_top = _top;
1242 _top = tn;
1243 // Calling Node::setup_is_top allows the nodes the chance to adjust
1244 // their _out arrays.
1245 if (_top != NULL) _top->setup_is_top();
1246 if (old_top != NULL) old_top->setup_is_top();
1247 assert(_top == NULL || top()->is_top(), "");
1248 }
1249
1250 #ifdef ASSERT
1251 uint Compile::count_live_nodes_by_graph_walk() {
1252 Unique_Node_List useful(comp_arena());
1253 // Get useful node list by walking the graph.
1254 identify_useful_nodes(useful);
1255 return useful.size();
1256 }
1257
1258 void Compile::print_missing_nodes() {
1259
1260 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1261 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1262 return;
1263 }
1264
1265 // This is an expensive function. It is executed only when the user
1266 // specifies VerifyIdealNodeCount option or otherwise knows the
1267 // additional work that needs to be done to identify reachable nodes
1268 // by walking the flow graph and find the missing ones using
1269 // _dead_node_list.
1270
1271 Unique_Node_List useful(comp_arena());
1272 // Get useful node list by walking the graph.
1273 identify_useful_nodes(useful);
1274
1275 uint l_nodes = C->live_nodes();
1276 uint l_nodes_by_walk = useful.size();
1277
1278 if (l_nodes != l_nodes_by_walk) {
1279 if (_log != NULL) {
1280 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1281 _log->stamp();
1282 _log->end_head();
1283 }
1284 VectorSet& useful_member_set = useful.member_set();
1285 int last_idx = l_nodes_by_walk;
1286 for (int i = 0; i < last_idx; i++) {
1287 if (useful_member_set.test(i)) {
1288 if (_dead_node_list.test(i)) {
1289 if (_log != NULL) {
1290 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1291 }
1292 if (PrintIdealNodeCount) {
1293 // Print the log message to tty
1294 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1295 useful.at(i)->dump();
1296 }
1297 }
1298 }
1299 else if (! _dead_node_list.test(i)) {
1300 if (_log != NULL) {
1301 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1302 }
1303 if (PrintIdealNodeCount) {
1304 // Print the log message to tty
1305 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1306 }
1307 }
1308 }
1309 if (_log != NULL) {
1310 _log->tail("mismatched_nodes");
1311 }
1312 }
1313 }
1314 void Compile::record_modified_node(Node* n) {
1315 if (_modified_nodes != NULL && !_inlining_incrementally &&
1316 n->outcnt() != 0 && !n->is_Con()) {
1317 _modified_nodes->push(n);
1318 }
1319 }
1320
1321 void Compile::remove_modified_node(Node* n) {
1322 if (_modified_nodes != NULL) {
1323 _modified_nodes->remove(n);
1324 }
1325 }
1326 #endif
1327
1328 #ifndef PRODUCT
1329 void Compile::verify_top(Node* tn) const {
1330 if (tn != NULL) {
1331 assert(tn->is_Con(), "top node must be a constant");
1332 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1333 assert(tn->in(0) != NULL, "must have live top node");
1334 }
1335 }
1336 #endif
1337
1338
1339 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1340
1341 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1342 guarantee(arr != NULL, "");
1343 int num_blocks = arr->length();
1344 if (grow_by < num_blocks) grow_by = num_blocks;
1345 int num_notes = grow_by * _node_notes_block_size;
1346 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1347 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1348 while (num_notes > 0) {
1349 arr->append(notes);
1350 notes += _node_notes_block_size;
1351 num_notes -= _node_notes_block_size;
1352 }
1353 assert(num_notes == 0, "exact multiple, please");
1354 }
1355
1356 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1357 if (source == NULL || dest == NULL) return false;
1358
1359 if (dest->is_Con())
1360 return false; // Do not push debug info onto constants.
1361
1362 #ifdef ASSERT
1363 // Leave a bread crumb trail pointing to the original node:
1364 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1365 dest->set_debug_orig(source);
1366 }
1367 #endif
1368
1369 if (node_note_array() == NULL)
1370 return false; // Not collecting any notes now.
1371
1372 // This is a copy onto a pre-existing node, which may already have notes.
1373 // If both nodes have notes, do not overwrite any pre-existing notes.
1374 Node_Notes* source_notes = node_notes_at(source->_idx);
1375 if (source_notes == NULL || source_notes->is_clear()) return false;
1376 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1377 if (dest_notes == NULL || dest_notes->is_clear()) {
1378 return set_node_notes_at(dest->_idx, source_notes);
1379 }
1380
1381 Node_Notes merged_notes = (*source_notes);
1382 // The order of operations here ensures that dest notes will win...
1383 merged_notes.update_from(dest_notes);
1384 return set_node_notes_at(dest->_idx, &merged_notes);
1385 }
1386
1387
1388 //--------------------------allow_range_check_smearing-------------------------
1389 // Gating condition for coalescing similar range checks.
1390 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1391 // single covering check that is at least as strong as any of them.
1392 // If the optimization succeeds, the simplified (strengthened) range check
1393 // will always succeed. If it fails, we will deopt, and then give up
1394 // on the optimization.
1395 bool Compile::allow_range_check_smearing() const {
1396 // If this method has already thrown a range-check,
1397 // assume it was because we already tried range smearing
1398 // and it failed.
1399 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1400 return !already_trapped;
1401 }
1402
1403
1404 //------------------------------flatten_alias_type-----------------------------
1405 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1406 int offset = tj->offset();
1407 TypePtr::PTR ptr = tj->ptr();
1408
1409 // Known instance (scalarizable allocation) alias only with itself.
1410 bool is_known_inst = tj->isa_oopptr() != NULL &&
1411 tj->is_oopptr()->is_known_instance();
1412
1413 // Process weird unsafe references.
1414 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1415 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1416 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1417 tj = TypeOopPtr::BOTTOM;
1418 ptr = tj->ptr();
1419 offset = tj->offset();
1420 }
1421
1422 // Array pointers need some flattening
1423 const TypeAryPtr *ta = tj->isa_aryptr();
1424 if (ta && ta->is_stable()) {
1425 // Erase stability property for alias analysis.
1426 tj = ta = ta->cast_to_stable(false);
1427 }
1428 if( ta && is_known_inst ) {
1429 if ( offset != Type::OffsetBot &&
1430 offset > arrayOopDesc::length_offset_in_bytes() ) {
1431 offset = Type::OffsetBot; // Flatten constant access into array body only
1432 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1433 }
1434 } else if( ta && _AliasLevel >= 2 ) {
1435 // For arrays indexed by constant indices, we flatten the alias
1436 // space to include all of the array body. Only the header, klass
1437 // and array length can be accessed un-aliased.
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,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,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,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,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,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,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,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,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,offset);
1506 }
1507 if (to->speculative() != NULL) {
1508 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),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, 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, offset, to->instance_id());
1530 } else {
1531 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, 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 offset);
1549 }
1550
1551 ciKlass* klass = tk->klass();
1552 if( 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, 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(), 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 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1793 ciField* field;
1794 if (tinst->const_oop() != NULL &&
1795 tinst->klass() == ciEnv::current()->Class_klass() &&
1796 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1797 // static field
1798 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1799 field = k->get_field_by_offset(tinst->offset(), true);
1800 } else {
1801 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1802 field = k->get_field_by_offset(tinst->offset(), false);
1803 }
1804 assert(field == NULL ||
1805 original_field == NULL ||
1806 (field->holder() == original_field->holder() &&
1807 field->offset() == original_field->offset() &&
1808 field->is_static() == original_field->is_static()), "wrong field?");
1809 // Set field() and is_rewritable() attributes.
1810 if (field != NULL) alias_type(idx)->set_field(field);
1811 }
1812 }
1813
1814 // Fill the cache for next time.
1815 ace->_adr_type = adr_type;
1816 ace->_index = idx;
1817 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1818
1819 // Might as well try to fill the cache for the flattened version, too.
1820 AliasCacheEntry* face = probe_alias_cache(flat);
1821 if (face->_adr_type == NULL) {
1822 face->_adr_type = flat;
1823 face->_index = idx;
1824 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1825 }
1826
1827 return alias_type(idx);
1828 }
1829
1830
1831 Compile::AliasType* Compile::alias_type(ciField* field) {
1832 const TypeOopPtr* t;
1833 if (field->is_static())
1834 t = TypeInstPtr::make(field->holder()->java_mirror());
1835 else
1836 t = TypeOopPtr::make_from_klass_raw(field->holder());
1837 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1838 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1839 return atp;
1840 }
1841
1842
1843 //------------------------------have_alias_type--------------------------------
1844 bool Compile::have_alias_type(const TypePtr* adr_type) {
1845 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1846 if (ace->_adr_type == adr_type) {
1847 return true;
1848 }
1849
1850 // Handle special cases.
1851 if (adr_type == NULL) return true;
1852 if (adr_type == TypePtr::BOTTOM) return true;
1853
1854 return find_alias_type(adr_type, true, NULL) != NULL;
1855 }
1856
1857 //-----------------------------must_alias--------------------------------------
1858 // True if all values of the given address type are in the given alias category.
1859 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1860 if (alias_idx == AliasIdxBot) return true; // the universal category
1861 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1862 if (alias_idx == AliasIdxTop) return false; // the empty category
1863 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1864
1865 // the only remaining possible overlap is identity
1866 int adr_idx = get_alias_index(adr_type);
1867 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1868 assert(adr_idx == alias_idx ||
1869 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1870 && adr_type != TypeOopPtr::BOTTOM),
1871 "should not be testing for overlap with an unsafe pointer");
1872 return adr_idx == alias_idx;
1873 }
1874
1875 //------------------------------can_alias--------------------------------------
1876 // True if any values of the given address type are in the given alias category.
1877 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1878 if (alias_idx == AliasIdxTop) return false; // the empty category
1879 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1880 if (alias_idx == AliasIdxBot) return true; // the universal category
1881 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1882
1883 // the only remaining possible overlap is identity
1884 int adr_idx = get_alias_index(adr_type);
1885 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1886 return adr_idx == alias_idx;
1887 }
1888
1889
1890
1891 //---------------------------pop_warm_call-------------------------------------
1892 WarmCallInfo* Compile::pop_warm_call() {
1893 WarmCallInfo* wci = _warm_calls;
1894 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1895 return wci;
1896 }
1897
1898 //----------------------------Inline_Warm--------------------------------------
1899 int Compile::Inline_Warm() {
1900 // If there is room, try to inline some more warm call sites.
1901 // %%% Do a graph index compaction pass when we think we're out of space?
1902 if (!InlineWarmCalls) return 0;
1903
1904 int calls_made_hot = 0;
1905 int room_to_grow = NodeCountInliningCutoff - unique();
1906 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1907 int amount_grown = 0;
1908 WarmCallInfo* call;
1909 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1910 int est_size = (int)call->size();
1911 if (est_size > (room_to_grow - amount_grown)) {
1912 // This one won't fit anyway. Get rid of it.
1913 call->make_cold();
1914 continue;
1915 }
1916 call->make_hot();
1917 calls_made_hot++;
1918 amount_grown += est_size;
1919 amount_to_grow -= est_size;
1920 }
1921
1922 if (calls_made_hot > 0) set_major_progress();
1923 return calls_made_hot;
1924 }
1925
1926
1927 //----------------------------Finish_Warm--------------------------------------
1928 void Compile::Finish_Warm() {
1929 if (!InlineWarmCalls) return;
1930 if (failing()) return;
1931 if (warm_calls() == NULL) return;
1932
1933 // Clean up loose ends, if we are out of space for inlining.
1934 WarmCallInfo* call;
1935 while ((call = pop_warm_call()) != NULL) {
1936 call->make_cold();
1937 }
1938 }
1939
1940 //---------------------cleanup_loop_predicates-----------------------
1941 // Remove the opaque nodes that protect the predicates so that all unused
1942 // checks and uncommon_traps will be eliminated from the ideal graph
1943 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1944 if (predicate_count()==0) return;
1945 for (int i = predicate_count(); i > 0; i--) {
1946 Node * n = predicate_opaque1_node(i-1);
1947 assert(n->Opcode() == Op_Opaque1, "must be");
1948 igvn.replace_node(n, n->in(1));
1949 }
1950 assert(predicate_count()==0, "should be clean!");
1951 }
1952
1953 void Compile::add_range_check_cast(Node* n) {
1954 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1955 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1956 _range_check_casts->append(n);
1957 }
1958
1959 // Remove all range check dependent CastIINodes.
1960 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1961 for (int i = range_check_cast_count(); i > 0; i--) {
1962 Node* cast = range_check_cast_node(i-1);
1963 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1964 igvn.replace_node(cast, cast->in(1));
1965 }
1966 assert(range_check_cast_count() == 0, "should be empty");
1967 }
1968
1969 void Compile::add_opaque4_node(Node* n) {
1970 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1971 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1972 _opaque4_nodes->append(n);
1973 }
1974
1975 // Remove all Opaque4 nodes.
1976 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1977 for (int i = opaque4_count(); i > 0; i--) {
1978 Node* opaq = opaque4_node(i-1);
1979 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
1980 igvn.replace_node(opaq, opaq->in(2));
1981 }
1982 assert(opaque4_count() == 0, "should be empty");
1983 }
1984
1985 // StringOpts and late inlining of string methods
1986 void Compile::inline_string_calls(bool parse_time) {
1987 {
1988 // remove useless nodes to make the usage analysis simpler
1989 ResourceMark rm;
1990 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1991 }
1992
1993 {
1994 ResourceMark rm;
1995 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1996 PhaseStringOpts pso(initial_gvn(), for_igvn());
1997 print_method(PHASE_AFTER_STRINGOPTS, 3);
1998 }
1999
2000 // now inline anything that we skipped the first time around
2001 if (!parse_time) {
2002 _late_inlines_pos = _late_inlines.length();
2003 }
2004
2005 while (_string_late_inlines.length() > 0) {
2006 CallGenerator* cg = _string_late_inlines.pop();
2007 cg->do_late_inline();
2008 if (failing()) return;
2009 }
2010 _string_late_inlines.trunc_to(0);
2011 }
2012
2013 // Late inlining of boxing methods
2014 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2015 if (_boxing_late_inlines.length() > 0) {
2016 assert(has_boxed_value(), "inconsistent");
2017
2018 PhaseGVN* gvn = initial_gvn();
2019 set_inlining_incrementally(true);
2020
2021 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2022 for_igvn()->clear();
2023 gvn->replace_with(&igvn);
2024
2025 _late_inlines_pos = _late_inlines.length();
2026
2027 while (_boxing_late_inlines.length() > 0) {
2028 CallGenerator* cg = _boxing_late_inlines.pop();
2029 cg->do_late_inline();
2030 if (failing()) return;
2031 }
2032 _boxing_late_inlines.trunc_to(0);
2033
2034 {
2035 ResourceMark rm;
2036 PhaseRemoveUseless pru(gvn, for_igvn());
2037 }
2038
2039 igvn = PhaseIterGVN(gvn);
2040 igvn.optimize();
2041
2042 set_inlining_progress(false);
2043 set_inlining_incrementally(false);
2044 }
2045 }
2046
2047 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2048 assert(IncrementalInline, "incremental inlining should be on");
2049 PhaseGVN* gvn = initial_gvn();
2050
2051 set_inlining_progress(false);
2052 for_igvn()->clear();
2053 gvn->replace_with(&igvn);
2054
2055 {
2056 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2057 int i = 0;
2058 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2059 CallGenerator* cg = _late_inlines.at(i);
2060 _late_inlines_pos = i+1;
2061 cg->do_late_inline();
2062 if (failing()) return;
2063 }
2064 int j = 0;
2065 for (; i < _late_inlines.length(); i++, j++) {
2066 _late_inlines.at_put(j, _late_inlines.at(i));
2067 }
2068 _late_inlines.trunc_to(j);
2069 }
2070
2071 {
2072 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2073 ResourceMark rm;
2074 PhaseRemoveUseless pru(gvn, for_igvn());
2075 }
2076
2077 {
2078 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2079 igvn = PhaseIterGVN(gvn);
2080 }
2081 }
2082
2083 // Perform incremental inlining until bound on number of live nodes is reached
2084 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2085 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2086
2087 PhaseGVN* gvn = initial_gvn();
2088
2089 set_inlining_incrementally(true);
2090 set_inlining_progress(true);
2091 uint low_live_nodes = 0;
2092
2093 while(inlining_progress() && _late_inlines.length() > 0) {
2094
2095 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2096 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2097 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2098 // PhaseIdealLoop is expensive so we only try it once we are
2099 // out of live nodes and we only try it again if the previous
2100 // helped got the number of nodes down significantly
2101 PhaseIdealLoop ideal_loop( igvn, false, true );
2102 if (failing()) return;
2103 low_live_nodes = live_nodes();
2104 _major_progress = true;
2105 }
2106
2107 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2108 break;
2109 }
2110 }
2111
2112 inline_incrementally_one(igvn);
2113
2114 if (failing()) return;
2115
2116 {
2117 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2118 igvn.optimize();
2119 }
2120
2121 if (failing()) return;
2122 }
2123
2124 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2125
2126 if (_string_late_inlines.length() > 0) {
2127 assert(has_stringbuilder(), "inconsistent");
2128 for_igvn()->clear();
2129 initial_gvn()->replace_with(&igvn);
2130
2131 inline_string_calls(false);
2132
2133 if (failing()) return;
2134
2135 {
2136 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2137 ResourceMark rm;
2138 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2139 }
2140
2141 {
2142 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2143 igvn = PhaseIterGVN(gvn);
2144 igvn.optimize();
2145 }
2146 }
2147
2148 set_inlining_incrementally(false);
2149 }
2150
2151
2152 //------------------------------Optimize---------------------------------------
2153 // Given a graph, optimize it.
2154 void Compile::Optimize() {
2155 TracePhase tp("optimizer", &timers[_t_optimizer]);
2156
2157 #ifndef PRODUCT
2158 if (_directive->BreakAtCompileOption) {
2159 BREAKPOINT;
2160 }
2161
2162 #endif
2163
2164 ResourceMark rm;
2165 int loop_opts_cnt;
2166
2167 print_inlining_reinit();
2168
2169 NOT_PRODUCT( verify_graph_edges(); )
2170
2171 print_method(PHASE_AFTER_PARSING);
2172
2173 {
2174 // Iterative Global Value Numbering, including ideal transforms
2175 // Initialize IterGVN with types and values from parse-time GVN
2176 PhaseIterGVN igvn(initial_gvn());
2177 #ifdef ASSERT
2178 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2179 #endif
2180 {
2181 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2182 igvn.optimize();
2183 }
2184
2185 print_method(PHASE_ITER_GVN1, 2);
2186
2187 if (failing()) return;
2188
2189 inline_incrementally(igvn);
2190
2191 print_method(PHASE_INCREMENTAL_INLINE, 2);
2192
2193 if (failing()) return;
2194
2195 if (eliminate_boxing()) {
2196 // Inline valueOf() methods now.
2197 inline_boxing_calls(igvn);
2198
2199 if (AlwaysIncrementalInline) {
2200 inline_incrementally(igvn);
2201 }
2202
2203 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2204
2205 if (failing()) return;
2206 }
2207
2208 // Remove the speculative part of types and clean up the graph from
2209 // the extra CastPP nodes whose only purpose is to carry them. Do
2210 // that early so that optimizations are not disrupted by the extra
2211 // CastPP nodes.
2212 remove_speculative_types(igvn);
2213
2214 // No more new expensive nodes will be added to the list from here
2215 // so keep only the actual candidates for optimizations.
2216 cleanup_expensive_nodes(igvn);
2217
2218 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2219 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2220 initial_gvn()->replace_with(&igvn);
2221 for_igvn()->clear();
2222 Unique_Node_List new_worklist(C->comp_arena());
2223 {
2224 ResourceMark rm;
2225 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2226 }
2227 set_for_igvn(&new_worklist);
2228 igvn = PhaseIterGVN(initial_gvn());
2229 igvn.optimize();
2230 }
2231
2232 // Perform escape analysis
2233 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2234 if (has_loops()) {
2235 // Cleanup graph (remove dead nodes).
2236 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2237 PhaseIdealLoop ideal_loop( igvn, false, true );
2238 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2239 if (failing()) return;
2240 }
2241 ConnectionGraph::do_analysis(this, &igvn);
2242
2243 if (failing()) return;
2244
2245 // Optimize out fields loads from scalar replaceable allocations.
2246 igvn.optimize();
2247 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2248
2249 if (failing()) return;
2250
2251 if (congraph() != NULL && macro_count() > 0) {
2252 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2253 PhaseMacroExpand mexp(igvn);
2254 mexp.eliminate_macro_nodes();
2255 igvn.set_delay_transform(false);
2256
2257 igvn.optimize();
2258 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2259
2260 if (failing()) return;
2261 }
2262 }
2263
2264 // Loop transforms on the ideal graph. Range Check Elimination,
2265 // peeling, unrolling, etc.
2266
2267 // Set loop opts counter
2268 loop_opts_cnt = num_loop_opts();
2269 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2270 {
2271 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2272 PhaseIdealLoop ideal_loop( igvn, true );
2273 loop_opts_cnt--;
2274 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2275 if (failing()) return;
2276 }
2277 // Loop opts pass if partial peeling occurred in previous pass
2278 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2279 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2280 PhaseIdealLoop ideal_loop( igvn, false );
2281 loop_opts_cnt--;
2282 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2283 if (failing()) return;
2284 }
2285 // Loop opts pass for loop-unrolling before CCP
2286 if(major_progress() && (loop_opts_cnt > 0)) {
2287 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2288 PhaseIdealLoop ideal_loop( igvn, false );
2289 loop_opts_cnt--;
2290 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2291 }
2292 if (!failing()) {
2293 // Verify that last round of loop opts produced a valid graph
2294 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2295 PhaseIdealLoop::verify(igvn);
2296 }
2297 }
2298 if (failing()) return;
2299
2300 // Conditional Constant Propagation;
2301 PhaseCCP ccp( &igvn );
2302 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2303 {
2304 TracePhase tp("ccp", &timers[_t_ccp]);
2305 ccp.do_transform();
2306 }
2307 print_method(PHASE_CPP1, 2);
2308
2309 assert( true, "Break here to ccp.dump_old2new_map()");
2310
2311 // Iterative Global Value Numbering, including ideal transforms
2312 {
2313 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2314 igvn = ccp;
2315 igvn.optimize();
2316 }
2317
2318 print_method(PHASE_ITER_GVN2, 2);
2319
2320 if (failing()) return;
2321
2322 // Loop transforms on the ideal graph. Range Check Elimination,
2323 // peeling, unrolling, etc.
2324 if(loop_opts_cnt > 0) {
2325 debug_only( int cnt = 0; );
2326 while(major_progress() && (loop_opts_cnt > 0)) {
2327 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2328 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2329 PhaseIdealLoop ideal_loop( igvn, true);
2330 loop_opts_cnt--;
2331 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2332 if (failing()) return;
2333 }
2334 }
2335 // Ensure that major progress is now clear
2336 C->clear_major_progress();
2337
2338 {
2339 // Verify that all previous optimizations produced a valid graph
2340 // at least to this point, even if no loop optimizations were done.
2341 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2342 PhaseIdealLoop::verify(igvn);
2343 }
2344
2345 if (range_check_cast_count() > 0) {
2346 // No more loop optimizations. Remove all range check dependent CastIINodes.
2347 C->remove_range_check_casts(igvn);
2348 igvn.optimize();
2349 }
2350
2351 {
2352 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2353 PhaseMacroExpand mex(igvn);
2354 if (mex.expand_macro_nodes()) {
2355 assert(failing(), "must bail out w/ explicit message");
2356 return;
2357 }
2358 }
2359
2360 if (opaque4_count() > 0) {
2361 C->remove_opaque4_nodes(igvn);
2362 igvn.optimize();
2363 }
2364
2365 DEBUG_ONLY( _modified_nodes = NULL; )
2366 } // (End scope of igvn; run destructor if necessary for asserts.)
2367
2368 process_print_inlining();
2369 // A method with only infinite loops has no edges entering loops from root
2370 {
2371 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2372 if (final_graph_reshaping()) {
2373 assert(failing(), "must bail out w/ explicit message");
2374 return;
2375 }
2376 }
2377
2378 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2379 }
2380
2381
2382 //------------------------------Code_Gen---------------------------------------
2383 // Given a graph, generate code for it
2384 void Compile::Code_Gen() {
2385 if (failing()) {
2386 return;
2387 }
2388
2389 // Perform instruction selection. You might think we could reclaim Matcher
2390 // memory PDQ, but actually the Matcher is used in generating spill code.
2391 // Internals of the Matcher (including some VectorSets) must remain live
2392 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2393 // set a bit in reclaimed memory.
2394
2395 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2396 // nodes. Mapping is only valid at the root of each matched subtree.
2397 NOT_PRODUCT( verify_graph_edges(); )
2398
2399 Matcher matcher;
2400 _matcher = &matcher;
2401 {
2402 TracePhase tp("matcher", &timers[_t_matcher]);
2403 matcher.match();
2404 }
2405 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2406 // nodes. Mapping is only valid at the root of each matched subtree.
2407 NOT_PRODUCT( verify_graph_edges(); )
2408
2409 // If you have too many nodes, or if matching has failed, bail out
2410 check_node_count(0, "out of nodes matching instructions");
2411 if (failing()) {
2412 return;
2413 }
2414
2415 // Build a proper-looking CFG
2416 PhaseCFG cfg(node_arena(), root(), matcher);
2417 _cfg = &cfg;
2418 {
2419 TracePhase tp("scheduler", &timers[_t_scheduler]);
2420 bool success = cfg.do_global_code_motion();
2421 if (!success) {
2422 return;
2423 }
2424
2425 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2426 NOT_PRODUCT( verify_graph_edges(); )
2427 debug_only( cfg.verify(); )
2428 }
2429
2430 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2431 _regalloc = ®alloc;
2432 {
2433 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2434 // Perform register allocation. After Chaitin, use-def chains are
2435 // no longer accurate (at spill code) and so must be ignored.
2436 // Node->LRG->reg mappings are still accurate.
2437 _regalloc->Register_Allocate();
2438
2439 // Bail out if the allocator builds too many nodes
2440 if (failing()) {
2441 return;
2442 }
2443 }
2444
2445 // Prior to register allocation we kept empty basic blocks in case the
2446 // the allocator needed a place to spill. After register allocation we
2447 // are not adding any new instructions. If any basic block is empty, we
2448 // can now safely remove it.
2449 {
2450 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2451 cfg.remove_empty_blocks();
2452 if (do_freq_based_layout()) {
2453 PhaseBlockLayout layout(cfg);
2454 } else {
2455 cfg.set_loop_alignment();
2456 }
2457 cfg.fixup_flow();
2458 }
2459
2460 // Apply peephole optimizations
2461 if( OptoPeephole ) {
2462 TracePhase tp("peephole", &timers[_t_peephole]);
2463 PhasePeephole peep( _regalloc, cfg);
2464 peep.do_transform();
2465 }
2466
2467 // Do late expand if CPU requires this.
2468 if (Matcher::require_postalloc_expand) {
2469 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2470 cfg.postalloc_expand(_regalloc);
2471 }
2472
2473 // Convert Nodes to instruction bits in a buffer
2474 {
2475 TraceTime tp("output", &timers[_t_output], CITime);
2476 Output();
2477 }
2478
2479 print_method(PHASE_FINAL_CODE);
2480
2481 // He's dead, Jim.
2482 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2483 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2484 }
2485
2486
2487 //------------------------------dump_asm---------------------------------------
2488 // Dump formatted assembly
2489 #ifndef PRODUCT
2490 void Compile::dump_asm(int *pcs, uint pc_limit) {
2491 bool cut_short = false;
2492 tty->print_cr("#");
2493 tty->print("# "); _tf->dump(); tty->cr();
2494 tty->print_cr("#");
2495
2496 // For all blocks
2497 int pc = 0x0; // Program counter
2498 char starts_bundle = ' ';
2499 _regalloc->dump_frame();
2500
2501 Node *n = NULL;
2502 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2503 if (VMThread::should_terminate()) {
2504 cut_short = true;
2505 break;
2506 }
2507 Block* block = _cfg->get_block(i);
2508 if (block->is_connector() && !Verbose) {
2509 continue;
2510 }
2511 n = block->head();
2512 if (pcs && n->_idx < pc_limit) {
2513 tty->print("%3.3x ", pcs[n->_idx]);
2514 } else {
2515 tty->print(" ");
2516 }
2517 block->dump_head(_cfg);
2518 if (block->is_connector()) {
2519 tty->print_cr(" # Empty connector block");
2520 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2521 tty->print_cr(" # Block is sole successor of call");
2522 }
2523
2524 // For all instructions
2525 Node *delay = NULL;
2526 for (uint j = 0; j < block->number_of_nodes(); j++) {
2527 if (VMThread::should_terminate()) {
2528 cut_short = true;
2529 break;
2530 }
2531 n = block->get_node(j);
2532 if (valid_bundle_info(n)) {
2533 Bundle* bundle = node_bundling(n);
2534 if (bundle->used_in_unconditional_delay()) {
2535 delay = n;
2536 continue;
2537 }
2538 if (bundle->starts_bundle()) {
2539 starts_bundle = '+';
2540 }
2541 }
2542
2543 if (WizardMode) {
2544 n->dump();
2545 }
2546
2547 if( !n->is_Region() && // Dont print in the Assembly
2548 !n->is_Phi() && // a few noisely useless nodes
2549 !n->is_Proj() &&
2550 !n->is_MachTemp() &&
2551 !n->is_SafePointScalarObject() &&
2552 !n->is_Catch() && // Would be nice to print exception table targets
2553 !n->is_MergeMem() && // Not very interesting
2554 !n->is_top() && // Debug info table constants
2555 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2556 ) {
2557 if (pcs && n->_idx < pc_limit)
2558 tty->print("%3.3x", pcs[n->_idx]);
2559 else
2560 tty->print(" ");
2561 tty->print(" %c ", starts_bundle);
2562 starts_bundle = ' ';
2563 tty->print("\t");
2564 n->format(_regalloc, tty);
2565 tty->cr();
2566 }
2567
2568 // If we have an instruction with a delay slot, and have seen a delay,
2569 // then back up and print it
2570 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2571 assert(delay != NULL, "no unconditional delay instruction");
2572 if (WizardMode) delay->dump();
2573
2574 if (node_bundling(delay)->starts_bundle())
2575 starts_bundle = '+';
2576 if (pcs && n->_idx < pc_limit)
2577 tty->print("%3.3x", pcs[n->_idx]);
2578 else
2579 tty->print(" ");
2580 tty->print(" %c ", starts_bundle);
2581 starts_bundle = ' ';
2582 tty->print("\t");
2583 delay->format(_regalloc, tty);
2584 tty->cr();
2585 delay = NULL;
2586 }
2587
2588 // Dump the exception table as well
2589 if( n->is_Catch() && (Verbose || WizardMode) ) {
2590 // Print the exception table for this offset
2591 _handler_table.print_subtable_for(pc);
2592 }
2593 }
2594
2595 if (pcs && n->_idx < pc_limit)
2596 tty->print_cr("%3.3x", pcs[n->_idx]);
2597 else
2598 tty->cr();
2599
2600 assert(cut_short || delay == NULL, "no unconditional delay branch");
2601
2602 } // End of per-block dump
2603 tty->cr();
2604
2605 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2606 }
2607 #endif
2608
2609 //------------------------------Final_Reshape_Counts---------------------------
2610 // This class defines counters to help identify when a method
2611 // may/must be executed using hardware with only 24-bit precision.
2612 struct Final_Reshape_Counts : public StackObj {
2613 int _call_count; // count non-inlined 'common' calls
2614 int _float_count; // count float ops requiring 24-bit precision
2615 int _double_count; // count double ops requiring more precision
2616 int _java_call_count; // count non-inlined 'java' calls
2617 int _inner_loop_count; // count loops which need alignment
2618 VectorSet _visited; // Visitation flags
2619 Node_List _tests; // Set of IfNodes & PCTableNodes
2620
2621 Final_Reshape_Counts() :
2622 _call_count(0), _float_count(0), _double_count(0),
2623 _java_call_count(0), _inner_loop_count(0),
2624 _visited( Thread::current()->resource_area() ) { }
2625
2626 void inc_call_count () { _call_count ++; }
2627 void inc_float_count () { _float_count ++; }
2628 void inc_double_count() { _double_count++; }
2629 void inc_java_call_count() { _java_call_count++; }
2630 void inc_inner_loop_count() { _inner_loop_count++; }
2631
2632 int get_call_count () const { return _call_count ; }
2633 int get_float_count () const { return _float_count ; }
2634 int get_double_count() const { return _double_count; }
2635 int get_java_call_count() const { return _java_call_count; }
2636 int get_inner_loop_count() const { return _inner_loop_count; }
2637 };
2638
2639 #ifdef ASSERT
2640 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2641 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2642 // Make sure the offset goes inside the instance layout.
2643 return k->contains_field_offset(tp->offset());
2644 // Note that OffsetBot and OffsetTop are very negative.
2645 }
2646 #endif
2647
2648 // Eliminate trivially redundant StoreCMs and accumulate their
2649 // precedence edges.
2650 void Compile::eliminate_redundant_card_marks(Node* n) {
2651 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2652 if (n->in(MemNode::Address)->outcnt() > 1) {
2653 // There are multiple users of the same address so it might be
2654 // possible to eliminate some of the StoreCMs
2655 Node* mem = n->in(MemNode::Memory);
2656 Node* adr = n->in(MemNode::Address);
2657 Node* val = n->in(MemNode::ValueIn);
2658 Node* prev = n;
2659 bool done = false;
2660 // Walk the chain of StoreCMs eliminating ones that match. As
2661 // long as it's a chain of single users then the optimization is
2662 // safe. Eliminating partially redundant StoreCMs would require
2663 // cloning copies down the other paths.
2664 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2665 if (adr == mem->in(MemNode::Address) &&
2666 val == mem->in(MemNode::ValueIn)) {
2667 // redundant StoreCM
2668 if (mem->req() > MemNode::OopStore) {
2669 // Hasn't been processed by this code yet.
2670 n->add_prec(mem->in(MemNode::OopStore));
2671 } else {
2672 // Already converted to precedence edge
2673 for (uint i = mem->req(); i < mem->len(); i++) {
2674 // Accumulate any precedence edges
2675 if (mem->in(i) != NULL) {
2676 n->add_prec(mem->in(i));
2677 }
2678 }
2679 // Everything above this point has been processed.
2680 done = true;
2681 }
2682 // Eliminate the previous StoreCM
2683 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2684 assert(mem->outcnt() == 0, "should be dead");
2685 mem->disconnect_inputs(NULL, this);
2686 } else {
2687 prev = mem;
2688 }
2689 mem = prev->in(MemNode::Memory);
2690 }
2691 }
2692 }
2693
2694 //------------------------------final_graph_reshaping_impl----------------------
2695 // Implement items 1-5 from final_graph_reshaping below.
2696 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2697
2698 if ( n->outcnt() == 0 ) return; // dead node
2699 uint nop = n->Opcode();
2700
2701 // Check for 2-input instruction with "last use" on right input.
2702 // Swap to left input. Implements item (2).
2703 if( n->req() == 3 && // two-input instruction
2704 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2705 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2706 n->in(2)->outcnt() == 1 &&// right use IS a last use
2707 !n->in(2)->is_Con() ) { // right use is not a constant
2708 // Check for commutative opcode
2709 switch( nop ) {
2710 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2711 case Op_MaxI: case Op_MinI:
2712 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2713 case Op_AndL: case Op_XorL: case Op_OrL:
2714 case Op_AndI: case Op_XorI: case Op_OrI: {
2715 // Move "last use" input to left by swapping inputs
2716 n->swap_edges(1, 2);
2717 break;
2718 }
2719 default:
2720 break;
2721 }
2722 }
2723
2724 #ifdef ASSERT
2725 if( n->is_Mem() ) {
2726 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2727 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2728 // oop will be recorded in oop map if load crosses safepoint
2729 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2730 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2731 "raw memory operations should have control edge");
2732 }
2733 #endif
2734 // Count FPU ops and common calls, implements item (3)
2735 switch( nop ) {
2736 // Count all float operations that may use FPU
2737 case Op_AddF:
2738 case Op_SubF:
2739 case Op_MulF:
2740 case Op_DivF:
2741 case Op_NegF:
2742 case Op_ModF:
2743 case Op_ConvI2F:
2744 case Op_ConF:
2745 case Op_CmpF:
2746 case Op_CmpF3:
2747 // case Op_ConvL2F: // longs are split into 32-bit halves
2748 frc.inc_float_count();
2749 break;
2750
2751 case Op_ConvF2D:
2752 case Op_ConvD2F:
2753 frc.inc_float_count();
2754 frc.inc_double_count();
2755 break;
2756
2757 // Count all double operations that may use FPU
2758 case Op_AddD:
2759 case Op_SubD:
2760 case Op_MulD:
2761 case Op_DivD:
2762 case Op_NegD:
2763 case Op_ModD:
2764 case Op_ConvI2D:
2765 case Op_ConvD2I:
2766 // case Op_ConvL2D: // handled by leaf call
2767 // case Op_ConvD2L: // handled by leaf call
2768 case Op_ConD:
2769 case Op_CmpD:
2770 case Op_CmpD3:
2771 frc.inc_double_count();
2772 break;
2773 case Op_Opaque1: // Remove Opaque Nodes before matching
2774 case Op_Opaque2: // Remove Opaque Nodes before matching
2775 case Op_Opaque3:
2776 n->subsume_by(n->in(1), this);
2777 break;
2778 case Op_CallStaticJava:
2779 case Op_CallJava:
2780 case Op_CallDynamicJava:
2781 frc.inc_java_call_count(); // Count java call site;
2782 case Op_CallRuntime:
2783 case Op_CallLeaf:
2784 case Op_CallLeafNoFP: {
2785 assert (n->is_Call(), "");
2786 CallNode *call = n->as_Call();
2787 // Count call sites where the FP mode bit would have to be flipped.
2788 // Do not count uncommon runtime calls:
2789 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2790 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2791 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2792 frc.inc_call_count(); // Count the call site
2793 } else { // See if uncommon argument is shared
2794 Node *n = call->in(TypeFunc::Parms);
2795 int nop = n->Opcode();
2796 // Clone shared simple arguments to uncommon calls, item (1).
2797 if (n->outcnt() > 1 &&
2798 !n->is_Proj() &&
2799 nop != Op_CreateEx &&
2800 nop != Op_CheckCastPP &&
2801 nop != Op_DecodeN &&
2802 nop != Op_DecodeNKlass &&
2803 !n->is_Mem() &&
2804 !n->is_Phi()) {
2805 Node *x = n->clone();
2806 call->set_req(TypeFunc::Parms, x);
2807 }
2808 }
2809 break;
2810 }
2811
2812 case Op_StoreD:
2813 case Op_LoadD:
2814 case Op_LoadD_unaligned:
2815 frc.inc_double_count();
2816 goto handle_mem;
2817 case Op_StoreF:
2818 case Op_LoadF:
2819 frc.inc_float_count();
2820 goto handle_mem;
2821
2822 case Op_StoreCM:
2823 {
2824 // Convert OopStore dependence into precedence edge
2825 Node* prec = n->in(MemNode::OopStore);
2826 n->del_req(MemNode::OopStore);
2827 n->add_prec(prec);
2828 eliminate_redundant_card_marks(n);
2829 }
2830
2831 // fall through
2832
2833 case Op_StoreB:
2834 case Op_StoreC:
2835 case Op_StorePConditional:
2836 case Op_StoreI:
2837 case Op_StoreL:
2838 case Op_StoreIConditional:
2839 case Op_StoreLConditional:
2840 case Op_CompareAndSwapB:
2841 case Op_CompareAndSwapS:
2842 case Op_CompareAndSwapI:
2843 case Op_CompareAndSwapL:
2844 case Op_CompareAndSwapP:
2845 case Op_CompareAndSwapN:
2846 case Op_WeakCompareAndSwapB:
2847 case Op_WeakCompareAndSwapS:
2848 case Op_WeakCompareAndSwapI:
2849 case Op_WeakCompareAndSwapL:
2850 case Op_WeakCompareAndSwapP:
2851 case Op_WeakCompareAndSwapN:
2852 case Op_CompareAndExchangeB:
2853 case Op_CompareAndExchangeS:
2854 case Op_CompareAndExchangeI:
2855 case Op_CompareAndExchangeL:
2856 case Op_CompareAndExchangeP:
2857 case Op_CompareAndExchangeN:
2858 case Op_GetAndAddS:
2859 case Op_GetAndAddB:
2860 case Op_GetAndAddI:
2861 case Op_GetAndAddL:
2862 case Op_GetAndSetS:
2863 case Op_GetAndSetB:
2864 case Op_GetAndSetI:
2865 case Op_GetAndSetL:
2866 case Op_GetAndSetP:
2867 case Op_GetAndSetN:
2868 case Op_StoreP:
2869 case Op_StoreN:
2870 case Op_StoreNKlass:
2871 case Op_LoadB:
2872 case Op_LoadUB:
2873 case Op_LoadUS:
2874 case Op_LoadI:
2875 case Op_LoadKlass:
2876 case Op_LoadNKlass:
2877 case Op_LoadL:
2878 case Op_LoadL_unaligned:
2879 case Op_LoadPLocked:
2880 case Op_LoadP:
2881 case Op_LoadN:
2882 case Op_LoadRange:
2883 case Op_LoadS: {
2884 handle_mem:
2885 #ifdef ASSERT
2886 if( VerifyOptoOopOffsets ) {
2887 assert( n->is_Mem(), "" );
2888 MemNode *mem = (MemNode*)n;
2889 // Check to see if address types have grounded out somehow.
2890 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2891 assert( !tp || oop_offset_is_sane(tp), "" );
2892 }
2893 #endif
2894 break;
2895 }
2896
2897 case Op_AddP: { // Assert sane base pointers
2898 Node *addp = n->in(AddPNode::Address);
2899 assert( !addp->is_AddP() ||
2900 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2901 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2902 "Base pointers must match (addp %u)", addp->_idx );
2903 #ifdef _LP64
2904 if ((UseCompressedOops || UseCompressedClassPointers) &&
2905 addp->Opcode() == Op_ConP &&
2906 addp == n->in(AddPNode::Base) &&
2907 n->in(AddPNode::Offset)->is_Con()) {
2908 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2909 // on the platform and on the compressed oops mode.
2910 // Use addressing with narrow klass to load with offset on x86.
2911 // Some platforms can use the constant pool to load ConP.
2912 // Do this transformation here since IGVN will convert ConN back to ConP.
2913 const Type* t = addp->bottom_type();
2914 bool is_oop = t->isa_oopptr() != NULL;
2915 bool is_klass = t->isa_klassptr() != NULL;
2916
2917 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2918 (is_klass && Matcher::const_klass_prefer_decode())) {
2919 Node* nn = NULL;
2920
2921 int op = is_oop ? Op_ConN : Op_ConNKlass;
2922
2923 // Look for existing ConN node of the same exact type.
2924 Node* r = root();
2925 uint cnt = r->outcnt();
2926 for (uint i = 0; i < cnt; i++) {
2927 Node* m = r->raw_out(i);
2928 if (m!= NULL && m->Opcode() == op &&
2929 m->bottom_type()->make_ptr() == t) {
2930 nn = m;
2931 break;
2932 }
2933 }
2934 if (nn != NULL) {
2935 // Decode a narrow oop to match address
2936 // [R12 + narrow_oop_reg<<3 + offset]
2937 if (is_oop) {
2938 nn = new DecodeNNode(nn, t);
2939 } else {
2940 nn = new DecodeNKlassNode(nn, t);
2941 }
2942 // Check for succeeding AddP which uses the same Base.
2943 // Otherwise we will run into the assertion above when visiting that guy.
2944 for (uint i = 0; i < n->outcnt(); ++i) {
2945 Node *out_i = n->raw_out(i);
2946 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
2947 out_i->set_req(AddPNode::Base, nn);
2948 #ifdef ASSERT
2949 for (uint j = 0; j < out_i->outcnt(); ++j) {
2950 Node *out_j = out_i->raw_out(j);
2951 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
2952 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
2953 }
2954 #endif
2955 }
2956 }
2957 n->set_req(AddPNode::Base, nn);
2958 n->set_req(AddPNode::Address, nn);
2959 if (addp->outcnt() == 0) {
2960 addp->disconnect_inputs(NULL, this);
2961 }
2962 }
2963 }
2964 }
2965 #endif
2966 // platform dependent reshaping of the address expression
2967 reshape_address(n->as_AddP());
2968 break;
2969 }
2970
2971 case Op_CastPP: {
2972 // Remove CastPP nodes to gain more freedom during scheduling but
2973 // keep the dependency they encode as control or precedence edges
2974 // (if control is set already) on memory operations. Some CastPP
2975 // nodes don't have a control (don't carry a dependency): skip
2976 // those.
2977 if (n->in(0) != NULL) {
2978 ResourceMark rm;
2979 Unique_Node_List wq;
2980 wq.push(n);
2981 for (uint next = 0; next < wq.size(); ++next) {
2982 Node *m = wq.at(next);
2983 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2984 Node* use = m->fast_out(i);
2985 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2986 use->ensure_control_or_add_prec(n->in(0));
2987 } else {
2988 switch(use->Opcode()) {
2989 case Op_AddP:
2990 case Op_DecodeN:
2991 case Op_DecodeNKlass:
2992 case Op_CheckCastPP:
2993 case Op_CastPP:
2994 wq.push(use);
2995 break;
2996 }
2997 }
2998 }
2999 }
3000 }
3001 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3002 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3003 Node* in1 = n->in(1);
3004 const Type* t = n->bottom_type();
3005 Node* new_in1 = in1->clone();
3006 new_in1->as_DecodeN()->set_type(t);
3007
3008 if (!Matcher::narrow_oop_use_complex_address()) {
3009 //
3010 // x86, ARM and friends can handle 2 adds in addressing mode
3011 // and Matcher can fold a DecodeN node into address by using
3012 // a narrow oop directly and do implicit NULL check in address:
3013 //
3014 // [R12 + narrow_oop_reg<<3 + offset]
3015 // NullCheck narrow_oop_reg
3016 //
3017 // On other platforms (Sparc) we have to keep new DecodeN node and
3018 // use it to do implicit NULL check in address:
3019 //
3020 // decode_not_null narrow_oop_reg, base_reg
3021 // [base_reg + offset]
3022 // NullCheck base_reg
3023 //
3024 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3025 // to keep the information to which NULL check the new DecodeN node
3026 // corresponds to use it as value in implicit_null_check().
3027 //
3028 new_in1->set_req(0, n->in(0));
3029 }
3030
3031 n->subsume_by(new_in1, this);
3032 if (in1->outcnt() == 0) {
3033 in1->disconnect_inputs(NULL, this);
3034 }
3035 } else {
3036 n->subsume_by(n->in(1), this);
3037 if (n->outcnt() == 0) {
3038 n->disconnect_inputs(NULL, this);
3039 }
3040 }
3041 break;
3042 }
3043 #ifdef _LP64
3044 case Op_CmpP:
3045 // Do this transformation here to preserve CmpPNode::sub() and
3046 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3047 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3048 Node* in1 = n->in(1);
3049 Node* in2 = n->in(2);
3050 if (!in1->is_DecodeNarrowPtr()) {
3051 in2 = in1;
3052 in1 = n->in(2);
3053 }
3054 assert(in1->is_DecodeNarrowPtr(), "sanity");
3055
3056 Node* new_in2 = NULL;
3057 if (in2->is_DecodeNarrowPtr()) {
3058 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3059 new_in2 = in2->in(1);
3060 } else if (in2->Opcode() == Op_ConP) {
3061 const Type* t = in2->bottom_type();
3062 if (t == TypePtr::NULL_PTR) {
3063 assert(in1->is_DecodeN(), "compare klass to null?");
3064 // Don't convert CmpP null check into CmpN if compressed
3065 // oops implicit null check is not generated.
3066 // This will allow to generate normal oop implicit null check.
3067 if (Matcher::gen_narrow_oop_implicit_null_checks())
3068 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3069 //
3070 // This transformation together with CastPP transformation above
3071 // will generated code for implicit NULL checks for compressed oops.
3072 //
3073 // The original code after Optimize()
3074 //
3075 // LoadN memory, narrow_oop_reg
3076 // decode narrow_oop_reg, base_reg
3077 // CmpP base_reg, NULL
3078 // CastPP base_reg // NotNull
3079 // Load [base_reg + offset], val_reg
3080 //
3081 // after these transformations will be
3082 //
3083 // LoadN memory, narrow_oop_reg
3084 // CmpN narrow_oop_reg, NULL
3085 // decode_not_null narrow_oop_reg, base_reg
3086 // Load [base_reg + offset], val_reg
3087 //
3088 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3089 // since narrow oops can be used in debug info now (see the code in
3090 // final_graph_reshaping_walk()).
3091 //
3092 // At the end the code will be matched to
3093 // on x86:
3094 //
3095 // Load_narrow_oop memory, narrow_oop_reg
3096 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3097 // NullCheck narrow_oop_reg
3098 //
3099 // and on sparc:
3100 //
3101 // Load_narrow_oop memory, narrow_oop_reg
3102 // decode_not_null narrow_oop_reg, base_reg
3103 // Load [base_reg + offset], val_reg
3104 // NullCheck base_reg
3105 //
3106 } else if (t->isa_oopptr()) {
3107 new_in2 = ConNode::make(t->make_narrowoop());
3108 } else if (t->isa_klassptr()) {
3109 new_in2 = ConNode::make(t->make_narrowklass());
3110 }
3111 }
3112 if (new_in2 != NULL) {
3113 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3114 n->subsume_by(cmpN, this);
3115 if (in1->outcnt() == 0) {
3116 in1->disconnect_inputs(NULL, this);
3117 }
3118 if (in2->outcnt() == 0) {
3119 in2->disconnect_inputs(NULL, this);
3120 }
3121 }
3122 }
3123 break;
3124
3125 case Op_DecodeN:
3126 case Op_DecodeNKlass:
3127 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3128 // DecodeN could be pinned when it can't be fold into
3129 // an address expression, see the code for Op_CastPP above.
3130 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3131 break;
3132
3133 case Op_EncodeP:
3134 case Op_EncodePKlass: {
3135 Node* in1 = n->in(1);
3136 if (in1->is_DecodeNarrowPtr()) {
3137 n->subsume_by(in1->in(1), this);
3138 } else if (in1->Opcode() == Op_ConP) {
3139 const Type* t = in1->bottom_type();
3140 if (t == TypePtr::NULL_PTR) {
3141 assert(t->isa_oopptr(), "null klass?");
3142 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3143 } else if (t->isa_oopptr()) {
3144 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3145 } else if (t->isa_klassptr()) {
3146 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3147 }
3148 }
3149 if (in1->outcnt() == 0) {
3150 in1->disconnect_inputs(NULL, this);
3151 }
3152 break;
3153 }
3154
3155 case Op_Proj: {
3156 if (OptimizeStringConcat) {
3157 ProjNode* p = n->as_Proj();
3158 if (p->_is_io_use) {
3159 // Separate projections were used for the exception path which
3160 // are normally removed by a late inline. If it wasn't inlined
3161 // then they will hang around and should just be replaced with
3162 // the original one.
3163 Node* proj = NULL;
3164 // Replace with just one
3165 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3166 Node *use = i.get();
3167 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3168 proj = use;
3169 break;
3170 }
3171 }
3172 assert(proj != NULL, "must be found");
3173 p->subsume_by(proj, this);
3174 }
3175 }
3176 break;
3177 }
3178
3179 case Op_Phi:
3180 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3181 // The EncodeP optimization may create Phi with the same edges
3182 // for all paths. It is not handled well by Register Allocator.
3183 Node* unique_in = n->in(1);
3184 assert(unique_in != NULL, "");
3185 uint cnt = n->req();
3186 for (uint i = 2; i < cnt; i++) {
3187 Node* m = n->in(i);
3188 assert(m != NULL, "");
3189 if (unique_in != m)
3190 unique_in = NULL;
3191 }
3192 if (unique_in != NULL) {
3193 n->subsume_by(unique_in, this);
3194 }
3195 }
3196 break;
3197
3198 #endif
3199
3200 #ifdef ASSERT
3201 case Op_CastII:
3202 // Verify that all range check dependent CastII nodes were removed.
3203 if (n->isa_CastII()->has_range_check()) {
3204 n->dump(3);
3205 assert(false, "Range check dependent CastII node was not removed");
3206 }
3207 break;
3208 #endif
3209
3210 case Op_ModI:
3211 if (UseDivMod) {
3212 // Check if a%b and a/b both exist
3213 Node* d = n->find_similar(Op_DivI);
3214 if (d) {
3215 // Replace them with a fused divmod if supported
3216 if (Matcher::has_match_rule(Op_DivModI)) {
3217 DivModINode* divmod = DivModINode::make(n);
3218 d->subsume_by(divmod->div_proj(), this);
3219 n->subsume_by(divmod->mod_proj(), this);
3220 } else {
3221 // replace a%b with a-((a/b)*b)
3222 Node* mult = new MulINode(d, d->in(2));
3223 Node* sub = new SubINode(d->in(1), mult);
3224 n->subsume_by(sub, this);
3225 }
3226 }
3227 }
3228 break;
3229
3230 case Op_ModL:
3231 if (UseDivMod) {
3232 // Check if a%b and a/b both exist
3233 Node* d = n->find_similar(Op_DivL);
3234 if (d) {
3235 // Replace them with a fused divmod if supported
3236 if (Matcher::has_match_rule(Op_DivModL)) {
3237 DivModLNode* divmod = DivModLNode::make(n);
3238 d->subsume_by(divmod->div_proj(), this);
3239 n->subsume_by(divmod->mod_proj(), this);
3240 } else {
3241 // replace a%b with a-((a/b)*b)
3242 Node* mult = new MulLNode(d, d->in(2));
3243 Node* sub = new SubLNode(d->in(1), mult);
3244 n->subsume_by(sub, this);
3245 }
3246 }
3247 }
3248 break;
3249
3250 case Op_LoadVector:
3251 case Op_StoreVector:
3252 break;
3253
3254 case Op_AddReductionVI:
3255 case Op_AddReductionVL:
3256 case Op_AddReductionVF:
3257 case Op_AddReductionVD:
3258 case Op_MulReductionVI:
3259 case Op_MulReductionVL:
3260 case Op_MulReductionVF:
3261 case Op_MulReductionVD:
3262 break;
3263
3264 case Op_PackB:
3265 case Op_PackS:
3266 case Op_PackI:
3267 case Op_PackF:
3268 case Op_PackL:
3269 case Op_PackD:
3270 if (n->req()-1 > 2) {
3271 // Replace many operand PackNodes with a binary tree for matching
3272 PackNode* p = (PackNode*) n;
3273 Node* btp = p->binary_tree_pack(1, n->req());
3274 n->subsume_by(btp, this);
3275 }
3276 break;
3277 case Op_Loop:
3278 case Op_CountedLoop:
3279 case Op_OuterStripMinedLoop:
3280 if (n->as_Loop()->is_inner_loop()) {
3281 frc.inc_inner_loop_count();
3282 }
3283 n->as_Loop()->verify_strip_mined(0);
3284 break;
3285 case Op_LShiftI:
3286 case Op_RShiftI:
3287 case Op_URShiftI:
3288 case Op_LShiftL:
3289 case Op_RShiftL:
3290 case Op_URShiftL:
3291 if (Matcher::need_masked_shift_count) {
3292 // The cpu's shift instructions don't restrict the count to the
3293 // lower 5/6 bits. We need to do the masking ourselves.
3294 Node* in2 = n->in(2);
3295 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3296 const TypeInt* t = in2->find_int_type();
3297 if (t != NULL && t->is_con()) {
3298 juint shift = t->get_con();
3299 if (shift > mask) { // Unsigned cmp
3300 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3301 }
3302 } else {
3303 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3304 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3305 n->set_req(2, shift);
3306 }
3307 }
3308 if (in2->outcnt() == 0) { // Remove dead node
3309 in2->disconnect_inputs(NULL, this);
3310 }
3311 }
3312 break;
3313 case Op_MemBarStoreStore:
3314 case Op_MemBarRelease:
3315 // Break the link with AllocateNode: it is no longer useful and
3316 // confuses register allocation.
3317 if (n->req() > MemBarNode::Precedent) {
3318 n->set_req(MemBarNode::Precedent, top());
3319 }
3320 break;
3321 case Op_RangeCheck: {
3322 RangeCheckNode* rc = n->as_RangeCheck();
3323 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3324 n->subsume_by(iff, this);
3325 frc._tests.push(iff);
3326 break;
3327 }
3328 case Op_ConvI2L: {
3329 if (!Matcher::convi2l_type_required) {
3330 // Code generation on some platforms doesn't need accurate
3331 // ConvI2L types. Widening the type can help remove redundant
3332 // address computations.
3333 n->as_Type()->set_type(TypeLong::INT);
3334 ResourceMark rm;
3335 Node_List wq;
3336 wq.push(n);
3337 for (uint next = 0; next < wq.size(); next++) {
3338 Node *m = wq.at(next);
3339
3340 for(;;) {
3341 // Loop over all nodes with identical inputs edges as m
3342 Node* k = m->find_similar(m->Opcode());
3343 if (k == NULL) {
3344 break;
3345 }
3346 // Push their uses so we get a chance to remove node made
3347 // redundant
3348 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3349 Node* u = k->fast_out(i);
3350 assert(!wq.contains(u), "shouldn't process one node several times");
3351 if (u->Opcode() == Op_LShiftL ||
3352 u->Opcode() == Op_AddL ||
3353 u->Opcode() == Op_SubL ||
3354 u->Opcode() == Op_AddP) {
3355 wq.push(u);
3356 }
3357 }
3358 // Replace all nodes with identical edges as m with m
3359 k->subsume_by(m, this);
3360 }
3361 }
3362 }
3363 break;
3364 }
3365 case Op_CmpUL: {
3366 if (!Matcher::has_match_rule(Op_CmpUL)) {
3367 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3368 // to max_julong if < 0 to make the signed comparison fail.
3369 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3370 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3371 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3372 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3373 Node* andl = new AndLNode(orl, remove_sign_mask);
3374 Node* cmp = new CmpLNode(andl, n->in(2));
3375 n->subsume_by(cmp, this);
3376 }
3377 break;
3378 }
3379 default:
3380 assert( !n->is_Call(), "" );
3381 assert( !n->is_Mem(), "" );
3382 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3383 break;
3384 }
3385
3386 // Collect CFG split points
3387 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3388 frc._tests.push(n);
3389 }
3390 }
3391
3392 //------------------------------final_graph_reshaping_walk---------------------
3393 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3394 // requires that the walk visits a node's inputs before visiting the node.
3395 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3396 ResourceArea *area = Thread::current()->resource_area();
3397 Unique_Node_List sfpt(area);
3398
3399 frc._visited.set(root->_idx); // first, mark node as visited
3400 uint cnt = root->req();
3401 Node *n = root;
3402 uint i = 0;
3403 while (true) {
3404 if (i < cnt) {
3405 // Place all non-visited non-null inputs onto stack
3406 Node* m = n->in(i);
3407 ++i;
3408 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3409 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3410 // compute worst case interpreter size in case of a deoptimization
3411 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3412
3413 sfpt.push(m);
3414 }
3415 cnt = m->req();
3416 nstack.push(n, i); // put on stack parent and next input's index
3417 n = m;
3418 i = 0;
3419 }
3420 } else {
3421 // Now do post-visit work
3422 final_graph_reshaping_impl( n, frc );
3423 if (nstack.is_empty())
3424 break; // finished
3425 n = nstack.node(); // Get node from stack
3426 cnt = n->req();
3427 i = nstack.index();
3428 nstack.pop(); // Shift to the next node on stack
3429 }
3430 }
3431
3432 // Skip next transformation if compressed oops are not used.
3433 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3434 (!UseCompressedOops && !UseCompressedClassPointers))
3435 return;
3436
3437 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3438 // It could be done for an uncommon traps or any safepoints/calls
3439 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3440 while (sfpt.size() > 0) {
3441 n = sfpt.pop();
3442 JVMState *jvms = n->as_SafePoint()->jvms();
3443 assert(jvms != NULL, "sanity");
3444 int start = jvms->debug_start();
3445 int end = n->req();
3446 bool is_uncommon = (n->is_CallStaticJava() &&
3447 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3448 for (int j = start; j < end; j++) {
3449 Node* in = n->in(j);
3450 if (in->is_DecodeNarrowPtr()) {
3451 bool safe_to_skip = true;
3452 if (!is_uncommon ) {
3453 // Is it safe to skip?
3454 for (uint i = 0; i < in->outcnt(); i++) {
3455 Node* u = in->raw_out(i);
3456 if (!u->is_SafePoint() ||
3457 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3458 safe_to_skip = false;
3459 }
3460 }
3461 }
3462 if (safe_to_skip) {
3463 n->set_req(j, in->in(1));
3464 }
3465 if (in->outcnt() == 0) {
3466 in->disconnect_inputs(NULL, this);
3467 }
3468 }
3469 }
3470 }
3471 }
3472
3473 //------------------------------final_graph_reshaping--------------------------
3474 // Final Graph Reshaping.
3475 //
3476 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3477 // and not commoned up and forced early. Must come after regular
3478 // optimizations to avoid GVN undoing the cloning. Clone constant
3479 // inputs to Loop Phis; these will be split by the allocator anyways.
3480 // Remove Opaque nodes.
3481 // (2) Move last-uses by commutative operations to the left input to encourage
3482 // Intel update-in-place two-address operations and better register usage
3483 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3484 // calls canonicalizing them back.
3485 // (3) Count the number of double-precision FP ops, single-precision FP ops
3486 // and call sites. On Intel, we can get correct rounding either by
3487 // forcing singles to memory (requires extra stores and loads after each
3488 // FP bytecode) or we can set a rounding mode bit (requires setting and
3489 // clearing the mode bit around call sites). The mode bit is only used
3490 // if the relative frequency of single FP ops to calls is low enough.
3491 // This is a key transform for SPEC mpeg_audio.
3492 // (4) Detect infinite loops; blobs of code reachable from above but not
3493 // below. Several of the Code_Gen algorithms fail on such code shapes,
3494 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3495 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3496 // Detection is by looking for IfNodes where only 1 projection is
3497 // reachable from below or CatchNodes missing some targets.
3498 // (5) Assert for insane oop offsets in debug mode.
3499
3500 bool Compile::final_graph_reshaping() {
3501 // an infinite loop may have been eliminated by the optimizer,
3502 // in which case the graph will be empty.
3503 if (root()->req() == 1) {
3504 record_method_not_compilable("trivial infinite loop");
3505 return true;
3506 }
3507
3508 // Expensive nodes have their control input set to prevent the GVN
3509 // from freely commoning them. There's no GVN beyond this point so
3510 // no need to keep the control input. We want the expensive nodes to
3511 // be freely moved to the least frequent code path by gcm.
3512 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3513 for (int i = 0; i < expensive_count(); i++) {
3514 _expensive_nodes->at(i)->set_req(0, NULL);
3515 }
3516
3517 Final_Reshape_Counts frc;
3518
3519 // Visit everybody reachable!
3520 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3521 Node_Stack nstack(live_nodes() >> 1);
3522 final_graph_reshaping_walk(nstack, root(), frc);
3523
3524 // Check for unreachable (from below) code (i.e., infinite loops).
3525 for( uint i = 0; i < frc._tests.size(); i++ ) {
3526 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3527 // Get number of CFG targets.
3528 // Note that PCTables include exception targets after calls.
3529 uint required_outcnt = n->required_outcnt();
3530 if (n->outcnt() != required_outcnt) {
3531 // Check for a few special cases. Rethrow Nodes never take the
3532 // 'fall-thru' path, so expected kids is 1 less.
3533 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3534 if (n->in(0)->in(0)->is_Call()) {
3535 CallNode *call = n->in(0)->in(0)->as_Call();
3536 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3537 required_outcnt--; // Rethrow always has 1 less kid
3538 } else if (call->req() > TypeFunc::Parms &&
3539 call->is_CallDynamicJava()) {
3540 // Check for null receiver. In such case, the optimizer has
3541 // detected that the virtual call will always result in a null
3542 // pointer exception. The fall-through projection of this CatchNode
3543 // will not be populated.
3544 Node *arg0 = call->in(TypeFunc::Parms);
3545 if (arg0->is_Type() &&
3546 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3547 required_outcnt--;
3548 }
3549 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3550 call->req() > TypeFunc::Parms+1 &&
3551 call->is_CallStaticJava()) {
3552 // Check for negative array length. In such case, the optimizer has
3553 // detected that the allocation attempt will always result in an
3554 // exception. There is no fall-through projection of this CatchNode .
3555 Node *arg1 = call->in(TypeFunc::Parms+1);
3556 if (arg1->is_Type() &&
3557 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3558 required_outcnt--;
3559 }
3560 }
3561 }
3562 }
3563 // Recheck with a better notion of 'required_outcnt'
3564 if (n->outcnt() != required_outcnt) {
3565 record_method_not_compilable("malformed control flow");
3566 return true; // Not all targets reachable!
3567 }
3568 }
3569 // Check that I actually visited all kids. Unreached kids
3570 // must be infinite loops.
3571 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3572 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3573 record_method_not_compilable("infinite loop");
3574 return true; // Found unvisited kid; must be unreach
3575 }
3576
3577 // Here so verification code in final_graph_reshaping_walk()
3578 // always see an OuterStripMinedLoopEnd
3579 if (n->is_OuterStripMinedLoopEnd()) {
3580 IfNode* init_iff = n->as_If();
3581 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3582 n->subsume_by(iff, this);
3583 }
3584 }
3585
3586 // If original bytecodes contained a mixture of floats and doubles
3587 // check if the optimizer has made it homogenous, item (3).
3588 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3589 frc.get_float_count() > 32 &&
3590 frc.get_double_count() == 0 &&
3591 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3592 set_24_bit_selection_and_mode( false, true );
3593 }
3594
3595 set_java_calls(frc.get_java_call_count());
3596 set_inner_loops(frc.get_inner_loop_count());
3597
3598 // No infinite loops, no reason to bail out.
3599 return false;
3600 }
3601
3602 //-----------------------------too_many_traps----------------------------------
3603 // Report if there are too many traps at the current method and bci.
3604 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3605 bool Compile::too_many_traps(ciMethod* method,
3606 int bci,
3607 Deoptimization::DeoptReason reason) {
3608 ciMethodData* md = method->method_data();
3609 if (md->is_empty()) {
3610 // Assume the trap has not occurred, or that it occurred only
3611 // because of a transient condition during start-up in the interpreter.
3612 return false;
3613 }
3614 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3615 if (md->has_trap_at(bci, m, reason) != 0) {
3616 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3617 // Also, if there are multiple reasons, or if there is no per-BCI record,
3618 // assume the worst.
3619 if (log())
3620 log()->elem("observe trap='%s' count='%d'",
3621 Deoptimization::trap_reason_name(reason),
3622 md->trap_count(reason));
3623 return true;
3624 } else {
3625 // Ignore method/bci and see if there have been too many globally.
3626 return too_many_traps(reason, md);
3627 }
3628 }
3629
3630 // Less-accurate variant which does not require a method and bci.
3631 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3632 ciMethodData* logmd) {
3633 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3634 // Too many traps globally.
3635 // Note that we use cumulative trap_count, not just md->trap_count.
3636 if (log()) {
3637 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3638 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3639 Deoptimization::trap_reason_name(reason),
3640 mcount, trap_count(reason));
3641 }
3642 return true;
3643 } else {
3644 // The coast is clear.
3645 return false;
3646 }
3647 }
3648
3649 //--------------------------too_many_recompiles--------------------------------
3650 // Report if there are too many recompiles at the current method and bci.
3651 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3652 // Is not eager to return true, since this will cause the compiler to use
3653 // Action_none for a trap point, to avoid too many recompilations.
3654 bool Compile::too_many_recompiles(ciMethod* method,
3655 int bci,
3656 Deoptimization::DeoptReason reason) {
3657 ciMethodData* md = method->method_data();
3658 if (md->is_empty()) {
3659 // Assume the trap has not occurred, or that it occurred only
3660 // because of a transient condition during start-up in the interpreter.
3661 return false;
3662 }
3663 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3664 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3665 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3666 Deoptimization::DeoptReason per_bc_reason
3667 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3668 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3669 if ((per_bc_reason == Deoptimization::Reason_none
3670 || md->has_trap_at(bci, m, reason) != 0)
3671 // The trap frequency measure we care about is the recompile count:
3672 && md->trap_recompiled_at(bci, m)
3673 && md->overflow_recompile_count() >= bc_cutoff) {
3674 // Do not emit a trap here if it has already caused recompilations.
3675 // Also, if there are multiple reasons, or if there is no per-BCI record,
3676 // assume the worst.
3677 if (log())
3678 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3679 Deoptimization::trap_reason_name(reason),
3680 md->trap_count(reason),
3681 md->overflow_recompile_count());
3682 return true;
3683 } else if (trap_count(reason) != 0
3684 && decompile_count() >= m_cutoff) {
3685 // Too many recompiles globally, and we have seen this sort of trap.
3686 // Use cumulative decompile_count, not just md->decompile_count.
3687 if (log())
3688 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3689 Deoptimization::trap_reason_name(reason),
3690 md->trap_count(reason), trap_count(reason),
3691 md->decompile_count(), decompile_count());
3692 return true;
3693 } else {
3694 // The coast is clear.
3695 return false;
3696 }
3697 }
3698
3699 // Compute when not to trap. Used by matching trap based nodes and
3700 // NullCheck optimization.
3701 void Compile::set_allowed_deopt_reasons() {
3702 _allowed_reasons = 0;
3703 if (is_method_compilation()) {
3704 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3705 assert(rs < BitsPerInt, "recode bit map");
3706 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3707 _allowed_reasons |= nth_bit(rs);
3708 }
3709 }
3710 }
3711 }
3712
3713 #ifndef PRODUCT
3714 //------------------------------verify_graph_edges---------------------------
3715 // Walk the Graph and verify that there is a one-to-one correspondence
3716 // between Use-Def edges and Def-Use edges in the graph.
3717 void Compile::verify_graph_edges(bool no_dead_code) {
3718 if (VerifyGraphEdges) {
3719 ResourceArea *area = Thread::current()->resource_area();
3720 Unique_Node_List visited(area);
3721 // Call recursive graph walk to check edges
3722 _root->verify_edges(visited);
3723 if (no_dead_code) {
3724 // Now make sure that no visited node is used by an unvisited node.
3725 bool dead_nodes = false;
3726 Unique_Node_List checked(area);
3727 while (visited.size() > 0) {
3728 Node* n = visited.pop();
3729 checked.push(n);
3730 for (uint i = 0; i < n->outcnt(); i++) {
3731 Node* use = n->raw_out(i);
3732 if (checked.member(use)) continue; // already checked
3733 if (visited.member(use)) continue; // already in the graph
3734 if (use->is_Con()) continue; // a dead ConNode is OK
3735 // At this point, we have found a dead node which is DU-reachable.
3736 if (!dead_nodes) {
3737 tty->print_cr("*** Dead nodes reachable via DU edges:");
3738 dead_nodes = true;
3739 }
3740 use->dump(2);
3741 tty->print_cr("---");
3742 checked.push(use); // No repeats; pretend it is now checked.
3743 }
3744 }
3745 assert(!dead_nodes, "using nodes must be reachable from root");
3746 }
3747 }
3748 }
3749
3750 // Verify GC barriers consistency
3751 // Currently supported:
3752 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3753 void Compile::verify_barriers() {
3754 if (UseG1GC) {
3755 // Verify G1 pre-barriers
3756 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
3757
3758 ResourceArea *area = Thread::current()->resource_area();
3759 Unique_Node_List visited(area);
3760 Node_List worklist(area);
3761 // We're going to walk control flow backwards starting from the Root
3762 worklist.push(_root);
3763 while (worklist.size() > 0) {
3764 Node* x = worklist.pop();
3765 if (x == NULL || x == top()) continue;
3766 if (visited.member(x)) {
3767 continue;
3768 } else {
3769 visited.push(x);
3770 }
3771
3772 if (x->is_Region()) {
3773 for (uint i = 1; i < x->req(); i++) {
3774 worklist.push(x->in(i));
3775 }
3776 } else {
3777 worklist.push(x->in(0));
3778 // We are looking for the pattern:
3779 // /->ThreadLocal
3780 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3781 // \->ConI(0)
3782 // We want to verify that the If and the LoadB have the same control
3783 // See GraphKit::g1_write_barrier_pre()
3784 if (x->is_If()) {
3785 IfNode *iff = x->as_If();
3786 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3787 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3788 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3789 && cmp->in(1)->is_Load()) {
3790 LoadNode* load = cmp->in(1)->as_Load();
3791 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3792 && load->in(2)->in(3)->is_Con()
3793 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3794
3795 Node* if_ctrl = iff->in(0);
3796 Node* load_ctrl = load->in(0);
3797
3798 if (if_ctrl != load_ctrl) {
3799 // Skip possible CProj->NeverBranch in infinite loops
3800 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3801 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3802 if_ctrl = if_ctrl->in(0)->in(0);
3803 }
3804 }
3805 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3806 }
3807 }
3808 }
3809 }
3810 }
3811 }
3812 }
3813 }
3814
3815 #endif
3816
3817 // The Compile object keeps track of failure reasons separately from the ciEnv.
3818 // This is required because there is not quite a 1-1 relation between the
3819 // ciEnv and its compilation task and the Compile object. Note that one
3820 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3821 // to backtrack and retry without subsuming loads. Other than this backtracking
3822 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3823 // by the logic in C2Compiler.
3824 void Compile::record_failure(const char* reason) {
3825 if (log() != NULL) {
3826 log()->elem("failure reason='%s' phase='compile'", reason);
3827 }
3828 if (_failure_reason == NULL) {
3829 // Record the first failure reason.
3830 _failure_reason = reason;
3831 }
3832
3833 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3834 C->print_method(PHASE_FAILURE);
3835 }
3836 _root = NULL; // flush the graph, too
3837 }
3838
3839 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3840 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3841 _phase_name(name), _dolog(CITimeVerbose)
3842 {
3843 if (_dolog) {
3844 C = Compile::current();
3845 _log = C->log();
3846 } else {
3847 C = NULL;
3848 _log = NULL;
3849 }
3850 if (_log != NULL) {
3851 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3852 _log->stamp();
3853 _log->end_head();
3854 }
3855 }
3856
3857 Compile::TracePhase::~TracePhase() {
3858
3859 C = Compile::current();
3860 if (_dolog) {
3861 _log = C->log();
3862 } else {
3863 _log = NULL;
3864 }
3865
3866 #ifdef ASSERT
3867 if (PrintIdealNodeCount) {
3868 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3869 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3870 }
3871
3872 if (VerifyIdealNodeCount) {
3873 Compile::current()->print_missing_nodes();
3874 }
3875 #endif
3876
3877 if (_log != NULL) {
3878 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3879 }
3880 }
3881
3882 //=============================================================================
3883 // Two Constant's are equal when the type and the value are equal.
3884 bool Compile::Constant::operator==(const Constant& other) {
3885 if (type() != other.type() ) return false;
3886 if (can_be_reused() != other.can_be_reused()) return false;
3887 // For floating point values we compare the bit pattern.
3888 switch (type()) {
3889 case T_INT:
3890 case T_FLOAT: return (_v._value.i == other._v._value.i);
3891 case T_LONG:
3892 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3893 case T_OBJECT:
3894 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3895 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3896 case T_METADATA: return (_v._metadata == other._v._metadata);
3897 default: ShouldNotReachHere(); return false;
3898 }
3899 }
3900
3901 static int type_to_size_in_bytes(BasicType t) {
3902 switch (t) {
3903 case T_INT: return sizeof(jint );
3904 case T_LONG: return sizeof(jlong );
3905 case T_FLOAT: return sizeof(jfloat );
3906 case T_DOUBLE: return sizeof(jdouble);
3907 case T_METADATA: return sizeof(Metadata*);
3908 // We use T_VOID as marker for jump-table entries (labels) which
3909 // need an internal word relocation.
3910 case T_VOID:
3911 case T_ADDRESS:
3912 case T_OBJECT: return sizeof(jobject);
3913 default:
3914 ShouldNotReachHere();
3915 return -1;
3916 }
3917 }
3918
3919 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3920 // sort descending
3921 if (a->freq() > b->freq()) return -1;
3922 if (a->freq() < b->freq()) return 1;
3923 return 0;
3924 }
3925
3926 void Compile::ConstantTable::calculate_offsets_and_size() {
3927 // First, sort the array by frequencies.
3928 _constants.sort(qsort_comparator);
3929
3930 #ifdef ASSERT
3931 // Make sure all jump-table entries were sorted to the end of the
3932 // array (they have a negative frequency).
3933 bool found_void = false;
3934 for (int i = 0; i < _constants.length(); i++) {
3935 Constant con = _constants.at(i);
3936 if (con.type() == T_VOID)
3937 found_void = true; // jump-tables
3938 else
3939 assert(!found_void, "wrong sorting");
3940 }
3941 #endif
3942
3943 int offset = 0;
3944 for (int i = 0; i < _constants.length(); i++) {
3945 Constant* con = _constants.adr_at(i);
3946
3947 // Align offset for type.
3948 int typesize = type_to_size_in_bytes(con->type());
3949 offset = align_up(offset, typesize);
3950 con->set_offset(offset); // set constant's offset
3951
3952 if (con->type() == T_VOID) {
3953 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3954 offset = offset + typesize * n->outcnt(); // expand jump-table
3955 } else {
3956 offset = offset + typesize;
3957 }
3958 }
3959
3960 // Align size up to the next section start (which is insts; see
3961 // CodeBuffer::align_at_start).
3962 assert(_size == -1, "already set?");
3963 _size = align_up(offset, (int)CodeEntryAlignment);
3964 }
3965
3966 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3967 MacroAssembler _masm(&cb);
3968 for (int i = 0; i < _constants.length(); i++) {
3969 Constant con = _constants.at(i);
3970 address constant_addr = NULL;
3971 switch (con.type()) {
3972 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
3973 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3974 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3975 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3976 case T_OBJECT: {
3977 jobject obj = con.get_jobject();
3978 int oop_index = _masm.oop_recorder()->find_index(obj);
3979 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3980 break;
3981 }
3982 case T_ADDRESS: {
3983 address addr = (address) con.get_jobject();
3984 constant_addr = _masm.address_constant(addr);
3985 break;
3986 }
3987 // We use T_VOID as marker for jump-table entries (labels) which
3988 // need an internal word relocation.
3989 case T_VOID: {
3990 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3991 // Fill the jump-table with a dummy word. The real value is
3992 // filled in later in fill_jump_table.
3993 address dummy = (address) n;
3994 constant_addr = _masm.address_constant(dummy);
3995 // Expand jump-table
3996 for (uint i = 1; i < n->outcnt(); i++) {
3997 address temp_addr = _masm.address_constant(dummy + i);
3998 assert(temp_addr, "consts section too small");
3999 }
4000 break;
4001 }
4002 case T_METADATA: {
4003 Metadata* obj = con.get_metadata();
4004 int metadata_index = _masm.oop_recorder()->find_index(obj);
4005 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4006 break;
4007 }
4008 default: ShouldNotReachHere();
4009 }
4010 assert(constant_addr, "consts section too small");
4011 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4012 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4013 }
4014 }
4015
4016 int Compile::ConstantTable::find_offset(Constant& con) const {
4017 int idx = _constants.find(con);
4018 assert(idx != -1, "constant must be in constant table");
4019 int offset = _constants.at(idx).offset();
4020 assert(offset != -1, "constant table not emitted yet?");
4021 return offset;
4022 }
4023
4024 void Compile::ConstantTable::add(Constant& con) {
4025 if (con.can_be_reused()) {
4026 int idx = _constants.find(con);
4027 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4028 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4029 return;
4030 }
4031 }
4032 (void) _constants.append(con);
4033 }
4034
4035 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4036 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4037 Constant con(type, value, b->_freq);
4038 add(con);
4039 return con;
4040 }
4041
4042 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4043 Constant con(metadata);
4044 add(con);
4045 return con;
4046 }
4047
4048 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4049 jvalue value;
4050 BasicType type = oper->type()->basic_type();
4051 switch (type) {
4052 case T_LONG: value.j = oper->constantL(); break;
4053 case T_FLOAT: value.f = oper->constantF(); break;
4054 case T_DOUBLE: value.d = oper->constantD(); break;
4055 case T_OBJECT:
4056 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4057 case T_METADATA: return add((Metadata*)oper->constant()); break;
4058 default: guarantee(false, "unhandled type: %s", type2name(type));
4059 }
4060 return add(n, type, value);
4061 }
4062
4063 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4064 jvalue value;
4065 // We can use the node pointer here to identify the right jump-table
4066 // as this method is called from Compile::Fill_buffer right before
4067 // the MachNodes are emitted and the jump-table is filled (means the
4068 // MachNode pointers do not change anymore).
4069 value.l = (jobject) n;
4070 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4071 add(con);
4072 return con;
4073 }
4074
4075 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4076 // If called from Compile::scratch_emit_size do nothing.
4077 if (Compile::current()->in_scratch_emit_size()) return;
4078
4079 assert(labels.is_nonempty(), "must be");
4080 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4081
4082 // Since MachConstantNode::constant_offset() also contains
4083 // table_base_offset() we need to subtract the table_base_offset()
4084 // to get the plain offset into the constant table.
4085 int offset = n->constant_offset() - table_base_offset();
4086
4087 MacroAssembler _masm(&cb);
4088 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4089
4090 for (uint i = 0; i < n->outcnt(); i++) {
4091 address* constant_addr = &jump_table_base[i];
4092 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));
4093 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4094 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4095 }
4096 }
4097
4098 //----------------------------static_subtype_check-----------------------------
4099 // Shortcut important common cases when superklass is exact:
4100 // (0) superklass is java.lang.Object (can occur in reflective code)
4101 // (1) subklass is already limited to a subtype of superklass => always ok
4102 // (2) subklass does not overlap with superklass => always fail
4103 // (3) superklass has NO subtypes and we can check with a simple compare.
4104 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4105 if (StressReflectiveCode) {
4106 return SSC_full_test; // Let caller generate the general case.
4107 }
4108
4109 if (superk == env()->Object_klass()) {
4110 return SSC_always_true; // (0) this test cannot fail
4111 }
4112
4113 ciType* superelem = superk;
4114 if (superelem->is_array_klass())
4115 superelem = superelem->as_array_klass()->base_element_type();
4116
4117 if (!subk->is_interface()) { // cannot trust static interface types yet
4118 if (subk->is_subtype_of(superk)) {
4119 return SSC_always_true; // (1) false path dead; no dynamic test needed
4120 }
4121 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4122 !superk->is_subtype_of(subk)) {
4123 return SSC_always_false;
4124 }
4125 }
4126
4127 // If casting to an instance klass, it must have no subtypes
4128 if (superk->is_interface()) {
4129 // Cannot trust interfaces yet.
4130 // %%% S.B. superk->nof_implementors() == 1
4131 } else if (superelem->is_instance_klass()) {
4132 ciInstanceKlass* ik = superelem->as_instance_klass();
4133 if (!ik->has_subklass() && !ik->is_interface()) {
4134 if (!ik->is_final()) {
4135 // Add a dependency if there is a chance of a later subclass.
4136 dependencies()->assert_leaf_type(ik);
4137 }
4138 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4139 }
4140 } else {
4141 // A primitive array type has no subtypes.
4142 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4143 }
4144
4145 return SSC_full_test;
4146 }
4147
4148 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4149 #ifdef _LP64
4150 // The scaled index operand to AddP must be a clean 64-bit value.
4151 // Java allows a 32-bit int to be incremented to a negative
4152 // value, which appears in a 64-bit register as a large
4153 // positive number. Using that large positive number as an
4154 // operand in pointer arithmetic has bad consequences.
4155 // On the other hand, 32-bit overflow is rare, and the possibility
4156 // can often be excluded, if we annotate the ConvI2L node with
4157 // a type assertion that its value is known to be a small positive
4158 // number. (The prior range check has ensured this.)
4159 // This assertion is used by ConvI2LNode::Ideal.
4160 int index_max = max_jint - 1; // array size is max_jint, index is one less
4161 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4162 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4163 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4164 #endif
4165 return idx;
4166 }
4167
4168 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4169 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4170 if (ctrl != NULL) {
4171 // Express control dependency by a CastII node with a narrow type.
4172 value = new CastIINode(value, itype, false, true /* range check dependency */);
4173 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4174 // node from floating above the range check during loop optimizations. Otherwise, the
4175 // ConvI2L node may be eliminated independently of the range check, causing the data path
4176 // to become TOP while the control path is still there (although it's unreachable).
4177 value->set_req(0, ctrl);
4178 // Save CastII node to remove it after loop optimizations.
4179 phase->C->add_range_check_cast(value);
4180 value = phase->transform(value);
4181 }
4182 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4183 return phase->transform(new ConvI2LNode(value, ltype));
4184 }
4185
4186 // The message about the current inlining is accumulated in
4187 // _print_inlining_stream and transfered into the _print_inlining_list
4188 // once we know whether inlining succeeds or not. For regular
4189 // inlining, messages are appended to the buffer pointed by
4190 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4191 // a new buffer is added after _print_inlining_idx in the list. This
4192 // way we can update the inlining message for late inlining call site
4193 // when the inlining is attempted again.
4194 void Compile::print_inlining_init() {
4195 if (print_inlining() || print_intrinsics()) {
4196 _print_inlining_stream = new stringStream();
4197 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4198 }
4199 }
4200
4201 void Compile::print_inlining_reinit() {
4202 if (print_inlining() || print_intrinsics()) {
4203 // Re allocate buffer when we change ResourceMark
4204 _print_inlining_stream = new stringStream();
4205 }
4206 }
4207
4208 void Compile::print_inlining_reset() {
4209 _print_inlining_stream->reset();
4210 }
4211
4212 void Compile::print_inlining_commit() {
4213 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4214 // Transfer the message from _print_inlining_stream to the current
4215 // _print_inlining_list buffer and clear _print_inlining_stream.
4216 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4217 print_inlining_reset();
4218 }
4219
4220 void Compile::print_inlining_push() {
4221 // Add new buffer to the _print_inlining_list at current position
4222 _print_inlining_idx++;
4223 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4224 }
4225
4226 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4227 return _print_inlining_list->at(_print_inlining_idx);
4228 }
4229
4230 void Compile::print_inlining_update(CallGenerator* cg) {
4231 if (print_inlining() || print_intrinsics()) {
4232 if (!cg->is_late_inline()) {
4233 if (print_inlining_current().cg() != NULL) {
4234 print_inlining_push();
4235 }
4236 print_inlining_commit();
4237 } else {
4238 if (print_inlining_current().cg() != cg &&
4239 (print_inlining_current().cg() != NULL ||
4240 print_inlining_current().ss()->size() != 0)) {
4241 print_inlining_push();
4242 }
4243 print_inlining_commit();
4244 print_inlining_current().set_cg(cg);
4245 }
4246 }
4247 }
4248
4249 void Compile::print_inlining_move_to(CallGenerator* cg) {
4250 // We resume inlining at a late inlining call site. Locate the
4251 // corresponding inlining buffer so that we can update it.
4252 if (print_inlining()) {
4253 for (int i = 0; i < _print_inlining_list->length(); i++) {
4254 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4255 _print_inlining_idx = i;
4256 return;
4257 }
4258 }
4259 ShouldNotReachHere();
4260 }
4261 }
4262
4263 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4264 if (print_inlining()) {
4265 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4266 assert(print_inlining_current().cg() == cg, "wrong entry");
4267 // replace message with new message
4268 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4269 print_inlining_commit();
4270 print_inlining_current().set_cg(cg);
4271 }
4272 }
4273
4274 void Compile::print_inlining_assert_ready() {
4275 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4276 }
4277
4278 void Compile::process_print_inlining() {
4279 bool do_print_inlining = print_inlining() || print_intrinsics();
4280 if (do_print_inlining || log() != NULL) {
4281 // Print inlining message for candidates that we couldn't inline
4282 // for lack of space
4283 for (int i = 0; i < _late_inlines.length(); i++) {
4284 CallGenerator* cg = _late_inlines.at(i);
4285 if (!cg->is_mh_late_inline()) {
4286 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4287 if (do_print_inlining) {
4288 cg->print_inlining_late(msg);
4289 }
4290 log_late_inline_failure(cg, msg);
4291 }
4292 }
4293 }
4294 if (do_print_inlining) {
4295 ResourceMark rm;
4296 stringStream ss;
4297 for (int i = 0; i < _print_inlining_list->length(); i++) {
4298 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4299 }
4300 size_t end = ss.size();
4301 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4302 strncpy(_print_inlining_output, ss.base(), end+1);
4303 _print_inlining_output[end] = 0;
4304 }
4305 }
4306
4307 void Compile::dump_print_inlining() {
4308 if (_print_inlining_output != NULL) {
4309 tty->print_raw(_print_inlining_output);
4310 }
4311 }
4312
4313 void Compile::log_late_inline(CallGenerator* cg) {
4314 if (log() != NULL) {
4315 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4316 cg->unique_id());
4317 JVMState* p = cg->call_node()->jvms();
4318 while (p != NULL) {
4319 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4320 p = p->caller();
4321 }
4322 log()->tail("late_inline");
4323 }
4324 }
4325
4326 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4327 log_late_inline(cg);
4328 if (log() != NULL) {
4329 log()->inline_fail(msg);
4330 }
4331 }
4332
4333 void Compile::log_inline_id(CallGenerator* cg) {
4334 if (log() != NULL) {
4335 // The LogCompilation tool needs a unique way to identify late
4336 // inline call sites. This id must be unique for this call site in
4337 // this compilation. Try to have it unique across compilations as
4338 // well because it can be convenient when grepping through the log
4339 // file.
4340 // Distinguish OSR compilations from others in case CICountOSR is
4341 // on.
4342 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4343 cg->set_unique_id(id);
4344 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4345 }
4346 }
4347
4348 void Compile::log_inline_failure(const char* msg) {
4349 if (C->log() != NULL) {
4350 C->log()->inline_fail(msg);
4351 }
4352 }
4353
4354
4355 // Dump inlining replay data to the stream.
4356 // Don't change thread state and acquire any locks.
4357 void Compile::dump_inline_data(outputStream* out) {
4358 InlineTree* inl_tree = ilt();
4359 if (inl_tree != NULL) {
4360 out->print(" inline %d", inl_tree->count());
4361 inl_tree->dump_replay_data(out);
4362 }
4363 }
4364
4365 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4366 if (n1->Opcode() < n2->Opcode()) return -1;
4367 else if (n1->Opcode() > n2->Opcode()) return 1;
4368
4369 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4370 for (uint i = 1; i < n1->req(); i++) {
4371 if (n1->in(i) < n2->in(i)) return -1;
4372 else if (n1->in(i) > n2->in(i)) return 1;
4373 }
4374
4375 return 0;
4376 }
4377
4378 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4379 Node* n1 = *n1p;
4380 Node* n2 = *n2p;
4381
4382 return cmp_expensive_nodes(n1, n2);
4383 }
4384
4385 void Compile::sort_expensive_nodes() {
4386 if (!expensive_nodes_sorted()) {
4387 _expensive_nodes->sort(cmp_expensive_nodes);
4388 }
4389 }
4390
4391 bool Compile::expensive_nodes_sorted() const {
4392 for (int i = 1; i < _expensive_nodes->length(); i++) {
4393 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4394 return false;
4395 }
4396 }
4397 return true;
4398 }
4399
4400 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4401 if (_expensive_nodes->length() == 0) {
4402 return false;
4403 }
4404
4405 assert(OptimizeExpensiveOps, "optimization off?");
4406
4407 // Take this opportunity to remove dead nodes from the list
4408 int j = 0;
4409 for (int i = 0; i < _expensive_nodes->length(); i++) {
4410 Node* n = _expensive_nodes->at(i);
4411 if (!n->is_unreachable(igvn)) {
4412 assert(n->is_expensive(), "should be expensive");
4413 _expensive_nodes->at_put(j, n);
4414 j++;
4415 }
4416 }
4417 _expensive_nodes->trunc_to(j);
4418
4419 // Then sort the list so that similar nodes are next to each other
4420 // and check for at least two nodes of identical kind with same data
4421 // inputs.
4422 sort_expensive_nodes();
4423
4424 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4425 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4426 return true;
4427 }
4428 }
4429
4430 return false;
4431 }
4432
4433 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4434 if (_expensive_nodes->length() == 0) {
4435 return;
4436 }
4437
4438 assert(OptimizeExpensiveOps, "optimization off?");
4439
4440 // Sort to bring similar nodes next to each other and clear the
4441 // control input of nodes for which there's only a single copy.
4442 sort_expensive_nodes();
4443
4444 int j = 0;
4445 int identical = 0;
4446 int i = 0;
4447 bool modified = false;
4448 for (; i < _expensive_nodes->length()-1; i++) {
4449 assert(j <= i, "can't write beyond current index");
4450 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4451 identical++;
4452 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4453 continue;
4454 }
4455 if (identical > 0) {
4456 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4457 identical = 0;
4458 } else {
4459 Node* n = _expensive_nodes->at(i);
4460 igvn.replace_input_of(n, 0, NULL);
4461 igvn.hash_insert(n);
4462 modified = true;
4463 }
4464 }
4465 if (identical > 0) {
4466 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4467 } else if (_expensive_nodes->length() >= 1) {
4468 Node* n = _expensive_nodes->at(i);
4469 igvn.replace_input_of(n, 0, NULL);
4470 igvn.hash_insert(n);
4471 modified = true;
4472 }
4473 _expensive_nodes->trunc_to(j);
4474 if (modified) {
4475 igvn.optimize();
4476 }
4477 }
4478
4479 void Compile::add_expensive_node(Node * n) {
4480 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4481 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4482 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4483 if (OptimizeExpensiveOps) {
4484 _expensive_nodes->append(n);
4485 } else {
4486 // Clear control input and let IGVN optimize expensive nodes if
4487 // OptimizeExpensiveOps is off.
4488 n->set_req(0, NULL);
4489 }
4490 }
4491
4492 /**
4493 * Remove the speculative part of types and clean up the graph
4494 */
4495 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4496 if (UseTypeSpeculation) {
4497 Unique_Node_List worklist;
4498 worklist.push(root());
4499 int modified = 0;
4500 // Go over all type nodes that carry a speculative type, drop the
4501 // speculative part of the type and enqueue the node for an igvn
4502 // which may optimize it out.
4503 for (uint next = 0; next < worklist.size(); ++next) {
4504 Node *n = worklist.at(next);
4505 if (n->is_Type()) {
4506 TypeNode* tn = n->as_Type();
4507 const Type* t = tn->type();
4508 const Type* t_no_spec = t->remove_speculative();
4509 if (t_no_spec != t) {
4510 bool in_hash = igvn.hash_delete(n);
4511 assert(in_hash, "node should be in igvn hash table");
4512 tn->set_type(t_no_spec);
4513 igvn.hash_insert(n);
4514 igvn._worklist.push(n); // give it a chance to go away
4515 modified++;
4516 }
4517 }
4518 uint max = n->len();
4519 for( uint i = 0; i < max; ++i ) {
4520 Node *m = n->in(i);
4521 if (not_a_node(m)) continue;
4522 worklist.push(m);
4523 }
4524 }
4525 // Drop the speculative part of all types in the igvn's type table
4526 igvn.remove_speculative_types();
4527 if (modified > 0) {
4528 igvn.optimize();
4529 }
4530 #ifdef ASSERT
4531 // Verify that after the IGVN is over no speculative type has resurfaced
4532 worklist.clear();
4533 worklist.push(root());
4534 for (uint next = 0; next < worklist.size(); ++next) {
4535 Node *n = worklist.at(next);
4536 const Type* t = igvn.type_or_null(n);
4537 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4538 if (n->is_Type()) {
4539 t = n->as_Type()->type();
4540 assert(t == t->remove_speculative(), "no more speculative types");
4541 }
4542 uint max = n->len();
4543 for( uint i = 0; i < max; ++i ) {
4544 Node *m = n->in(i);
4545 if (not_a_node(m)) continue;
4546 worklist.push(m);
4547 }
4548 }
4549 igvn.check_no_speculative_types();
4550 #endif
4551 }
4552 }
4553
4554 // Auxiliary method to support randomized stressing/fuzzing.
4555 //
4556 // This method can be called the arbitrary number of times, with current count
4557 // as the argument. The logic allows selecting a single candidate from the
4558 // running list of candidates as follows:
4559 // int count = 0;
4560 // Cand* selected = null;
4561 // while(cand = cand->next()) {
4562 // if (randomized_select(++count)) {
4563 // selected = cand;
4564 // }
4565 // }
4566 //
4567 // Including count equalizes the chances any candidate is "selected".
4568 // This is useful when we don't have the complete list of candidates to choose
4569 // from uniformly. In this case, we need to adjust the randomicity of the
4570 // selection, or else we will end up biasing the selection towards the latter
4571 // candidates.
4572 //
4573 // Quick back-envelope calculation shows that for the list of n candidates
4574 // the equal probability for the candidate to persist as "best" can be
4575 // achieved by replacing it with "next" k-th candidate with the probability
4576 // of 1/k. It can be easily shown that by the end of the run, the
4577 // probability for any candidate is converged to 1/n, thus giving the
4578 // uniform distribution among all the candidates.
4579 //
4580 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4581 #define RANDOMIZED_DOMAIN_POW 29
4582 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4583 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4584 bool Compile::randomized_select(int count) {
4585 assert(count > 0, "only positive");
4586 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4587 }
4588
4589 CloneMap& Compile::clone_map() { return _clone_map; }
4590 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4591
4592 void NodeCloneInfo::dump() const {
4593 tty->print(" {%d:%d} ", idx(), gen());
4594 }
4595
4596 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4597 uint64_t val = value(old->_idx);
4598 NodeCloneInfo cio(val);
4599 assert(val != 0, "old node should be in the map");
4600 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4601 insert(nnn->_idx, cin.get());
4602 #ifndef PRODUCT
4603 if (is_debug()) {
4604 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4605 }
4606 #endif
4607 }
4608
4609 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4610 NodeCloneInfo cio(value(old->_idx));
4611 if (cio.get() == 0) {
4612 cio.set(old->_idx, 0);
4613 insert(old->_idx, cio.get());
4614 #ifndef PRODUCT
4615 if (is_debug()) {
4616 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4617 }
4618 #endif
4619 }
4620 clone(old, nnn, gen);
4621 }
4622
4623 int CloneMap::max_gen() const {
4624 int g = 0;
4625 DictI di(_dict);
4626 for(; di.test(); ++di) {
4627 int t = gen(di._key);
4628 if (g < t) {
4629 g = t;
4630 #ifndef PRODUCT
4631 if (is_debug()) {
4632 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4633 }
4634 #endif
4635 }
4636 }
4637 return g;
4638 }
4639
4640 void CloneMap::dump(node_idx_t key) const {
4641 uint64_t val = value(key);
4642 if (val != 0) {
4643 NodeCloneInfo ni(val);
4644 ni.dump();
4645 }
4646 }