1 /*
2 * Copyright (c) 1997, 2013, 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/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/node.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/output.hpp"
57 #include "opto/parse.hpp"
58 #include "opto/phaseX.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/runtime.hpp"
61 #include "opto/stringopts.hpp"
62 #include "opto/type.hpp"
63 #include "opto/vectornode.hpp"
64 #include "runtime/arguments.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/timer.hpp"
68 #include "trace/tracing.hpp"
69 #include "utilities/copy.hpp"
70 #ifdef TARGET_ARCH_MODEL_x86_32
71 # include "adfiles/ad_x86_32.hpp"
72 #endif
73 #ifdef TARGET_ARCH_MODEL_x86_64
74 # include "adfiles/ad_x86_64.hpp"
75 #endif
76 #ifdef TARGET_ARCH_MODEL_sparc
77 # include "adfiles/ad_sparc.hpp"
78 #endif
79 #ifdef TARGET_ARCH_MODEL_zero
80 # include "adfiles/ad_zero.hpp"
81 #endif
82 #ifdef TARGET_ARCH_MODEL_arm
83 # include "adfiles/ad_arm.hpp"
84 #endif
85 #ifdef TARGET_ARCH_MODEL_ppc
86 # include "adfiles/ad_ppc.hpp"
87 #endif
88
89
90 // -------------------- Compile::mach_constant_base_node -----------------------
91 // Constant table base node singleton.
92 MachConstantBaseNode* Compile::mach_constant_base_node() {
93 if (_mach_constant_base_node == NULL) {
94 _mach_constant_base_node = new (C) MachConstantBaseNode();
95 _mach_constant_base_node->add_req(C->root());
96 }
97 return _mach_constant_base_node;
98 }
99
100
101 /// Support for intrinsics.
102
103 // Return the index at which m must be inserted (or already exists).
104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
105 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
106 #ifdef ASSERT
107 for (int i = 1; i < _intrinsics->length(); i++) {
108 CallGenerator* cg1 = _intrinsics->at(i-1);
109 CallGenerator* cg2 = _intrinsics->at(i);
110 assert(cg1->method() != cg2->method()
111 ? cg1->method() < cg2->method()
112 : cg1->is_virtual() < cg2->is_virtual(),
113 "compiler intrinsics list must stay sorted");
114 }
115 #endif
116 // Binary search sorted list, in decreasing intervals [lo, hi].
117 int lo = 0, hi = _intrinsics->length()-1;
118 while (lo <= hi) {
119 int mid = (uint)(hi + lo) / 2;
120 ciMethod* mid_m = _intrinsics->at(mid)->method();
121 if (m < mid_m) {
122 hi = mid-1;
123 } else if (m > mid_m) {
124 lo = mid+1;
125 } else {
126 // look at minor sort key
127 bool mid_virt = _intrinsics->at(mid)->is_virtual();
128 if (is_virtual < mid_virt) {
129 hi = mid-1;
130 } else if (is_virtual > mid_virt) {
131 lo = mid+1;
132 } else {
133 return mid; // exact match
134 }
135 }
136 }
137 return lo; // inexact match
138 }
139
140 void Compile::register_intrinsic(CallGenerator* cg) {
141 if (_intrinsics == NULL) {
142 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
143 }
144 // This code is stolen from ciObjectFactory::insert.
145 // Really, GrowableArray should have methods for
146 // insert_at, remove_at, and binary_search.
147 int len = _intrinsics->length();
148 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
149 if (index == len) {
150 _intrinsics->append(cg);
151 } else {
152 #ifdef ASSERT
153 CallGenerator* oldcg = _intrinsics->at(index);
154 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
155 #endif
156 _intrinsics->append(_intrinsics->at(len-1));
157 int pos;
158 for (pos = len-2; pos >= index; pos--) {
159 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
160 }
161 _intrinsics->at_put(index, cg);
162 }
163 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
164 }
165
166 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
167 assert(m->is_loaded(), "don't try this on unloaded methods");
168 if (_intrinsics != NULL) {
169 int index = intrinsic_insertion_index(m, is_virtual);
170 if (index < _intrinsics->length()
171 && _intrinsics->at(index)->method() == m
172 && _intrinsics->at(index)->is_virtual() == is_virtual) {
173 return _intrinsics->at(index);
174 }
175 }
176 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
177 if (m->intrinsic_id() != vmIntrinsics::_none &&
178 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
179 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
180 if (cg != NULL) {
181 // Save it for next time:
182 register_intrinsic(cg);
183 return cg;
184 } else {
185 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
186 }
187 }
188 return NULL;
189 }
190
191 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
192 // in library_call.cpp.
193
194
195 #ifndef PRODUCT
196 // statistics gathering...
197
198 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
199 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
200
201 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
202 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
203 int oflags = _intrinsic_hist_flags[id];
204 assert(flags != 0, "what happened?");
205 if (is_virtual) {
206 flags |= _intrinsic_virtual;
207 }
208 bool changed = (flags != oflags);
209 if ((flags & _intrinsic_worked) != 0) {
210 juint count = (_intrinsic_hist_count[id] += 1);
211 if (count == 1) {
212 changed = true; // first time
213 }
214 // increment the overall count also:
215 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
216 }
217 if (changed) {
218 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
219 // Something changed about the intrinsic's virtuality.
220 if ((flags & _intrinsic_virtual) != 0) {
221 // This is the first use of this intrinsic as a virtual call.
222 if (oflags != 0) {
223 // We already saw it as a non-virtual, so note both cases.
224 flags |= _intrinsic_both;
225 }
226 } else if ((oflags & _intrinsic_both) == 0) {
227 // This is the first use of this intrinsic as a non-virtual
228 flags |= _intrinsic_both;
229 }
230 }
231 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
232 }
233 // update the overall flags also:
234 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
235 return changed;
236 }
237
238 static char* format_flags(int flags, char* buf) {
239 buf[0] = 0;
240 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
241 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
242 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
243 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
244 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
245 if (buf[0] == 0) strcat(buf, ",");
246 assert(buf[0] == ',', "must be");
247 return &buf[1];
248 }
249
250 void Compile::print_intrinsic_statistics() {
251 char flagsbuf[100];
252 ttyLocker ttyl;
253 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
254 tty->print_cr("Compiler intrinsic usage:");
255 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
256 if (total == 0) total = 1; // avoid div0 in case of no successes
257 #define PRINT_STAT_LINE(name, c, f) \
258 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
259 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
260 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
261 int flags = _intrinsic_hist_flags[id];
262 juint count = _intrinsic_hist_count[id];
263 if ((flags | count) != 0) {
264 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
265 }
266 }
267 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
268 if (xtty != NULL) xtty->tail("statistics");
269 }
270
271 void Compile::print_statistics() {
272 { ttyLocker ttyl;
273 if (xtty != NULL) xtty->head("statistics type='opto'");
274 Parse::print_statistics();
275 PhaseCCP::print_statistics();
276 PhaseRegAlloc::print_statistics();
277 Scheduling::print_statistics();
278 PhasePeephole::print_statistics();
279 PhaseIdealLoop::print_statistics();
280 if (xtty != NULL) xtty->tail("statistics");
281 }
282 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
283 // put this under its own <statistics> element.
284 print_intrinsic_statistics();
285 }
286 }
287 #endif //PRODUCT
288
289 // Support for bundling info
290 Bundle* Compile::node_bundling(const Node *n) {
291 assert(valid_bundle_info(n), "oob");
292 return &_node_bundling_base[n->_idx];
293 }
294
295 bool Compile::valid_bundle_info(const Node *n) {
296 return (_node_bundling_limit > n->_idx);
297 }
298
299
300 void Compile::gvn_replace_by(Node* n, Node* nn) {
301 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
302 Node* use = n->last_out(i);
303 bool is_in_table = initial_gvn()->hash_delete(use);
304 uint uses_found = 0;
305 for (uint j = 0; j < use->len(); j++) {
306 if (use->in(j) == n) {
307 if (j < use->req())
308 use->set_req(j, nn);
309 else
310 use->set_prec(j, nn);
311 uses_found++;
312 }
313 }
314 if (is_in_table) {
315 // reinsert into table
316 initial_gvn()->hash_find_insert(use);
317 }
318 record_for_igvn(use);
319 i -= uses_found; // we deleted 1 or more copies of this edge
320 }
321 }
322
323
324 static inline bool not_a_node(const Node* n) {
325 if (n == NULL) return true;
326 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
327 if (*(address*)n == badAddress) return true; // kill by Node::destruct
328 return false;
329 }
330
331 // Identify all nodes that are reachable from below, useful.
332 // Use breadth-first pass that records state in a Unique_Node_List,
333 // recursive traversal is slower.
334 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
335 int estimated_worklist_size = unique();
336 useful.map( estimated_worklist_size, NULL ); // preallocate space
337
338 // Initialize worklist
339 if (root() != NULL) { useful.push(root()); }
340 // If 'top' is cached, declare it useful to preserve cached node
341 if( cached_top_node() ) { useful.push(cached_top_node()); }
342
343 // Push all useful nodes onto the list, breadthfirst
344 for( uint next = 0; next < useful.size(); ++next ) {
345 assert( next < unique(), "Unique useful nodes < total nodes");
346 Node *n = useful.at(next);
347 uint max = n->len();
348 for( uint i = 0; i < max; ++i ) {
349 Node *m = n->in(i);
350 if (not_a_node(m)) continue;
351 useful.push(m);
352 }
353 }
354 }
355
356 // Update dead_node_list with any missing dead nodes using useful
357 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
358 void Compile::update_dead_node_list(Unique_Node_List &useful) {
359 uint max_idx = unique();
360 VectorSet& useful_node_set = useful.member_set();
361
362 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
363 // If node with index node_idx is not in useful set,
364 // mark it as dead in dead node list.
365 if (! useful_node_set.test(node_idx) ) {
366 record_dead_node(node_idx);
367 }
368 }
369 }
370
371 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
372 int shift = 0;
373 for (int i = 0; i < inlines->length(); i++) {
374 CallGenerator* cg = inlines->at(i);
375 CallNode* call = cg->call_node();
376 if (shift > 0) {
377 inlines->at_put(i-shift, cg);
378 }
379 if (!useful.member(call)) {
380 shift++;
381 }
382 }
383 inlines->trunc_to(inlines->length()-shift);
384 }
385
386 // Disconnect all useless nodes by disconnecting those at the boundary.
387 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
388 uint next = 0;
389 while (next < useful.size()) {
390 Node *n = useful.at(next++);
391 // Use raw traversal of out edges since this code removes out edges
392 int max = n->outcnt();
393 for (int j = 0; j < max; ++j) {
394 Node* child = n->raw_out(j);
395 if (! useful.member(child)) {
396 assert(!child->is_top() || child != top(),
397 "If top is cached in Compile object it is in useful list");
398 // Only need to remove this out-edge to the useless node
399 n->raw_del_out(j);
400 --j;
401 --max;
402 }
403 }
404 if (n->outcnt() == 1 && n->has_special_unique_user()) {
405 record_for_igvn(n->unique_out());
406 }
407 }
408 // Remove useless macro and predicate opaq nodes
409 for (int i = C->macro_count()-1; i >= 0; i--) {
410 Node* n = C->macro_node(i);
411 if (!useful.member(n)) {
412 remove_macro_node(n);
413 }
414 }
415 // Remove useless expensive node
416 for (int i = C->expensive_count()-1; i >= 0; i--) {
417 Node* n = C->expensive_node(i);
418 if (!useful.member(n)) {
419 remove_expensive_node(n);
420 }
421 }
422 // clean up the late inline lists
423 remove_useless_late_inlines(&_string_late_inlines, useful);
424 remove_useless_late_inlines(&_boxing_late_inlines, useful);
425 remove_useless_late_inlines(&_late_inlines, useful);
426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
427 }
428
429 //------------------------------frame_size_in_words-----------------------------
430 // frame_slots in units of words
431 int Compile::frame_size_in_words() const {
432 // shift is 0 in LP32 and 1 in LP64
433 const int shift = (LogBytesPerWord - LogBytesPerInt);
434 int words = _frame_slots >> shift;
435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
436 return words;
437 }
438
439 // ============================================================================
440 //------------------------------CompileWrapper---------------------------------
441 class CompileWrapper : public StackObj {
442 Compile *const _compile;
443 public:
444 CompileWrapper(Compile* compile);
445
446 ~CompileWrapper();
447 };
448
449 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
450 // the Compile* pointer is stored in the current ciEnv:
451 ciEnv* env = compile->env();
452 assert(env == ciEnv::current(), "must already be a ciEnv active");
453 assert(env->compiler_data() == NULL, "compile already active?");
454 env->set_compiler_data(compile);
455 assert(compile == Compile::current(), "sanity");
456
457 compile->set_type_dict(NULL);
458 compile->set_type_hwm(NULL);
459 compile->set_type_last_size(0);
460 compile->set_last_tf(NULL, NULL);
461 compile->set_indexSet_arena(NULL);
462 compile->set_indexSet_free_block_list(NULL);
463 compile->init_type_arena();
464 Type::Initialize(compile);
465 _compile->set_scratch_buffer_blob(NULL);
466 _compile->begin_method();
467 }
468 CompileWrapper::~CompileWrapper() {
469 _compile->end_method();
470 if (_compile->scratch_buffer_blob() != NULL)
471 BufferBlob::free(_compile->scratch_buffer_blob());
472 _compile->env()->set_compiler_data(NULL);
473 }
474
475
476 //----------------------------print_compile_messages---------------------------
477 void Compile::print_compile_messages() {
478 #ifndef PRODUCT
479 // Check if recompiling
480 if (_subsume_loads == false && PrintOpto) {
481 // Recompiling without allowing machine instructions to subsume loads
482 tty->print_cr("*********************************************************");
483 tty->print_cr("** Bailout: Recompile without subsuming loads **");
484 tty->print_cr("*********************************************************");
485 }
486 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
487 // Recompiling without escape analysis
488 tty->print_cr("*********************************************************");
489 tty->print_cr("** Bailout: Recompile without escape analysis **");
490 tty->print_cr("*********************************************************");
491 }
492 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
493 // Recompiling without boxing elimination
494 tty->print_cr("*********************************************************");
495 tty->print_cr("** Bailout: Recompile without boxing elimination **");
496 tty->print_cr("*********************************************************");
497 }
498 if (env()->break_at_compile()) {
499 // Open the debugger when compiling this method.
500 tty->print("### Breaking when compiling: ");
501 method()->print_short_name();
502 tty->cr();
503 BREAKPOINT;
504 }
505
506 if( PrintOpto ) {
507 if (is_osr_compilation()) {
508 tty->print("[OSR]%3d", _compile_id);
509 } else {
510 tty->print("%3d", _compile_id);
511 }
512 }
513 #endif
514 }
515
516
517 //-----------------------init_scratch_buffer_blob------------------------------
518 // Construct a temporary BufferBlob and cache it for this compile.
519 void Compile::init_scratch_buffer_blob(int const_size) {
520 // If there is already a scratch buffer blob allocated and the
521 // constant section is big enough, use it. Otherwise free the
522 // current and allocate a new one.
523 BufferBlob* blob = scratch_buffer_blob();
524 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
525 // Use the current blob.
526 } else {
527 if (blob != NULL) {
528 BufferBlob::free(blob);
529 }
530
531 ResourceMark rm;
532 _scratch_const_size = const_size;
533 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
534 blob = BufferBlob::create("Compile::scratch_buffer", size);
535 // Record the buffer blob for next time.
536 set_scratch_buffer_blob(blob);
537 // Have we run out of code space?
538 if (scratch_buffer_blob() == NULL) {
539 // Let CompilerBroker disable further compilations.
540 record_failure("Not enough space for scratch buffer in CodeCache");
541 return;
542 }
543 }
544
545 // Initialize the relocation buffers
546 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
547 set_scratch_locs_memory(locs_buf);
548 }
549
550
551 //-----------------------scratch_emit_size-------------------------------------
552 // Helper function that computes size by emitting code
553 uint Compile::scratch_emit_size(const Node* n) {
554 // Start scratch_emit_size section.
555 set_in_scratch_emit_size(true);
556
557 // Emit into a trash buffer and count bytes emitted.
558 // This is a pretty expensive way to compute a size,
559 // but it works well enough if seldom used.
560 // All common fixed-size instructions are given a size
561 // method by the AD file.
562 // Note that the scratch buffer blob and locs memory are
563 // allocated at the beginning of the compile task, and
564 // may be shared by several calls to scratch_emit_size.
565 // The allocation of the scratch buffer blob is particularly
566 // expensive, since it has to grab the code cache lock.
567 BufferBlob* blob = this->scratch_buffer_blob();
568 assert(blob != NULL, "Initialize BufferBlob at start");
569 assert(blob->size() > MAX_inst_size, "sanity");
570 relocInfo* locs_buf = scratch_locs_memory();
571 address blob_begin = blob->content_begin();
572 address blob_end = (address)locs_buf;
573 assert(blob->content_contains(blob_end), "sanity");
574 CodeBuffer buf(blob_begin, blob_end - blob_begin);
575 buf.initialize_consts_size(_scratch_const_size);
576 buf.initialize_stubs_size(MAX_stubs_size);
577 assert(locs_buf != NULL, "sanity");
578 int lsize = MAX_locs_size / 3;
579 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
580 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
581 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
582
583 // Do the emission.
584
585 Label fakeL; // Fake label for branch instructions.
586 Label* saveL = NULL;
587 uint save_bnum = 0;
588 bool is_branch = n->is_MachBranch();
589 if (is_branch) {
590 MacroAssembler masm(&buf);
591 masm.bind(fakeL);
592 n->as_MachBranch()->save_label(&saveL, &save_bnum);
593 n->as_MachBranch()->label_set(&fakeL, 0);
594 }
595 n->emit(buf, this->regalloc());
596 if (is_branch) // Restore label.
597 n->as_MachBranch()->label_set(saveL, save_bnum);
598
599 // End scratch_emit_size section.
600 set_in_scratch_emit_size(false);
601
602 return buf.insts_size();
603 }
604
605
606 // ============================================================================
607 //------------------------------Compile standard-------------------------------
608 debug_only( int Compile::_debug_idx = 100000; )
609
610 // Compile a method. entry_bci is -1 for normal compilations and indicates
611 // the continuation bci for on stack replacement.
612
613
614 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
615 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
616 : Phase(Compiler),
617 _env(ci_env),
618 _log(ci_env->log()),
619 _compile_id(ci_env->compile_id()),
620 _save_argument_registers(false),
621 _stub_name(NULL),
622 _stub_function(NULL),
623 _stub_entry_point(NULL),
624 _method(target),
625 _entry_bci(osr_bci),
626 _initial_gvn(NULL),
627 _for_igvn(NULL),
628 _warm_calls(NULL),
629 _subsume_loads(subsume_loads),
630 _do_escape_analysis(do_escape_analysis),
631 _eliminate_boxing(eliminate_boxing),
632 _failure_reason(NULL),
633 _code_buffer("Compile::Fill_buffer"),
634 _orig_pc_slot(0),
635 _orig_pc_slot_offset_in_bytes(0),
636 _has_method_handle_invokes(false),
637 _mach_constant_base_node(NULL),
638 _node_bundling_limit(0),
639 _node_bundling_base(NULL),
640 _java_calls(0),
641 _inner_loops(0),
642 _scratch_const_size(-1),
643 _in_scratch_emit_size(false),
644 _dead_node_list(comp_arena()),
645 _dead_node_count(0),
646 #ifndef PRODUCT
647 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
648 _printer(IdealGraphPrinter::printer()),
649 #endif
650 _congraph(NULL),
651 _replay_inline_data(NULL),
652 _late_inlines(comp_arena(), 2, 0, NULL),
653 _string_late_inlines(comp_arena(), 2, 0, NULL),
654 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
655 _late_inlines_pos(0),
656 _number_of_mh_late_inlines(0),
657 _inlining_progress(false),
658 _inlining_incrementally(false),
659 _print_inlining_list(NULL),
660 _print_inlining_idx(0),
661 _preserve_jvm_state(0) {
662 C = this;
663
664 CompileWrapper cw(this);
665 #ifndef PRODUCT
666 if (TimeCompiler2) {
667 tty->print(" ");
668 target->holder()->name()->print();
669 tty->print(".");
670 target->print_short_name();
671 tty->print(" ");
672 }
673 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
674 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
675 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
676 if (!print_opto_assembly) {
677 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
678 if (print_assembly && !Disassembler::can_decode()) {
679 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
680 print_opto_assembly = true;
681 }
682 }
683 set_print_assembly(print_opto_assembly);
684 set_parsed_irreducible_loop(false);
685
686 if (method()->has_option("ReplayInline")) {
687 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
688 }
689 #endif
690 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
691 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
692
693 if (ProfileTraps) {
694 // Make sure the method being compiled gets its own MDO,
695 // so we can at least track the decompile_count().
696 method()->ensure_method_data();
697 }
698
699 Init(::AliasLevel);
700
701
702 print_compile_messages();
703
704 _ilt = InlineTree::build_inline_tree_root();
705
706 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
707 assert(num_alias_types() >= AliasIdxRaw, "");
708
709 #define MINIMUM_NODE_HASH 1023
710 // Node list that Iterative GVN will start with
711 Unique_Node_List for_igvn(comp_arena());
712 set_for_igvn(&for_igvn);
713
714 // GVN that will be run immediately on new nodes
715 uint estimated_size = method()->code_size()*4+64;
716 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
717 PhaseGVN gvn(node_arena(), estimated_size);
718 set_initial_gvn(&gvn);
719
720 if (print_inlining() || print_intrinsics()) {
721 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
722 }
723 { // Scope for timing the parser
724 TracePhase t3("parse", &_t_parser, true);
725
726 // Put top into the hash table ASAP.
727 initial_gvn()->transform_no_reclaim(top());
728
729 // Set up tf(), start(), and find a CallGenerator.
730 CallGenerator* cg = NULL;
731 if (is_osr_compilation()) {
732 const TypeTuple *domain = StartOSRNode::osr_domain();
733 const TypeTuple *range = TypeTuple::make_range(method()->signature());
734 init_tf(TypeFunc::make(domain, range));
735 StartNode* s = new (this) StartOSRNode(root(), domain);
736 initial_gvn()->set_type_bottom(s);
737 init_start(s);
738 cg = CallGenerator::for_osr(method(), entry_bci());
739 } else {
740 // Normal case.
741 init_tf(TypeFunc::make(method()));
742 StartNode* s = new (this) StartNode(root(), tf()->domain());
743 initial_gvn()->set_type_bottom(s);
744 init_start(s);
745 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
746 // With java.lang.ref.reference.get() we must go through the
747 // intrinsic when G1 is enabled - even when get() is the root
748 // method of the compile - so that, if necessary, the value in
749 // the referent field of the reference object gets recorded by
750 // the pre-barrier code.
751 // Specifically, if G1 is enabled, the value in the referent
752 // field is recorded by the G1 SATB pre barrier. This will
753 // result in the referent being marked live and the reference
754 // object removed from the list of discovered references during
755 // reference processing.
756 cg = find_intrinsic(method(), false);
757 }
758 if (cg == NULL) {
759 float past_uses = method()->interpreter_invocation_count();
760 float expected_uses = past_uses;
761 cg = CallGenerator::for_inline(method(), expected_uses);
762 }
763 }
764 if (failing()) return;
765 if (cg == NULL) {
766 record_method_not_compilable_all_tiers("cannot parse method");
767 return;
768 }
769 JVMState* jvms = build_start_state(start(), tf());
770 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
771 record_method_not_compilable("method parse failed");
772 return;
773 }
774 GraphKit kit(jvms);
775
776 if (!kit.stopped()) {
777 // Accept return values, and transfer control we know not where.
778 // This is done by a special, unique ReturnNode bound to root.
779 return_values(kit.jvms());
780 }
781
782 if (kit.has_exceptions()) {
783 // Any exceptions that escape from this call must be rethrown
784 // to whatever caller is dynamically above us on the stack.
785 // This is done by a special, unique RethrowNode bound to root.
786 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
787 }
788
789 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
790
791 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
792 inline_string_calls(true);
793 }
794
795 if (failing()) return;
796
797 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
798
799 // Remove clutter produced by parsing.
800 if (!failing()) {
801 ResourceMark rm;
802 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
803 }
804 }
805
806 // Note: Large methods are capped off in do_one_bytecode().
807 if (failing()) return;
808
809 // After parsing, node notes are no longer automagic.
810 // They must be propagated by register_new_node_with_optimizer(),
811 // clone(), or the like.
812 set_default_node_notes(NULL);
813
814 for (;;) {
815 int successes = Inline_Warm();
816 if (failing()) return;
817 if (successes == 0) break;
818 }
819
820 // Drain the list.
821 Finish_Warm();
822 #ifndef PRODUCT
823 if (_printer) {
824 _printer->print_inlining(this);
825 }
826 #endif
827
828 if (failing()) return;
829 NOT_PRODUCT( verify_graph_edges(); )
830
831 // Now optimize
832 Optimize();
833 if (failing()) return;
834 NOT_PRODUCT( verify_graph_edges(); )
835
836 #ifndef PRODUCT
837 if (PrintIdeal) {
838 ttyLocker ttyl; // keep the following output all in one block
839 // This output goes directly to the tty, not the compiler log.
840 // To enable tools to match it up with the compilation activity,
841 // be sure to tag this tty output with the compile ID.
842 if (xtty != NULL) {
843 xtty->head("ideal compile_id='%d'%s", compile_id(),
844 is_osr_compilation() ? " compile_kind='osr'" :
845 "");
846 }
847 root()->dump(9999);
848 if (xtty != NULL) {
849 xtty->tail("ideal");
850 }
851 }
852 #endif
853
854 NOT_PRODUCT( verify_barriers(); )
855
856 // Dump compilation data to replay it.
857 if (method()->has_option("DumpReplay")) {
858 env()->dump_replay_data(_compile_id);
859 }
860 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
861 env()->dump_inline_data(_compile_id);
862 }
863
864 // Now that we know the size of all the monitors we can add a fixed slot
865 // for the original deopt pc.
866
867 _orig_pc_slot = fixed_slots();
868 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
869 set_fixed_slots(next_slot);
870
871 // Now generate code
872 Code_Gen();
873 if (failing()) return;
874
875 // Check if we want to skip execution of all compiled code.
876 {
877 #ifndef PRODUCT
878 if (OptoNoExecute) {
879 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
880 return;
881 }
882 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
883 #endif
884
885 if (is_osr_compilation()) {
886 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
887 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
888 } else {
889 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
890 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
891 }
892
893 env()->register_method(_method, _entry_bci,
894 &_code_offsets,
895 _orig_pc_slot_offset_in_bytes,
896 code_buffer(),
897 frame_size_in_words(), _oop_map_set,
898 &_handler_table, &_inc_table,
899 compiler,
900 env()->comp_level(),
901 has_unsafe_access(),
902 SharedRuntime::is_wide_vector(max_vector_size())
903 );
904
905 if (log() != NULL) // Print code cache state into compiler log
906 log()->code_cache_state();
907 }
908 }
909
910 //------------------------------Compile----------------------------------------
911 // Compile a runtime stub
912 Compile::Compile( ciEnv* ci_env,
913 TypeFunc_generator generator,
914 address stub_function,
915 const char *stub_name,
916 int is_fancy_jump,
917 bool pass_tls,
918 bool save_arg_registers,
919 bool return_pc )
920 : Phase(Compiler),
921 _env(ci_env),
922 _log(ci_env->log()),
923 _compile_id(0),
924 _save_argument_registers(save_arg_registers),
925 _method(NULL),
926 _stub_name(stub_name),
927 _stub_function(stub_function),
928 _stub_entry_point(NULL),
929 _entry_bci(InvocationEntryBci),
930 _initial_gvn(NULL),
931 _for_igvn(NULL),
932 _warm_calls(NULL),
933 _orig_pc_slot(0),
934 _orig_pc_slot_offset_in_bytes(0),
935 _subsume_loads(true),
936 _do_escape_analysis(false),
937 _eliminate_boxing(false),
938 _failure_reason(NULL),
939 _code_buffer("Compile::Fill_buffer"),
940 _has_method_handle_invokes(false),
941 _mach_constant_base_node(NULL),
942 _node_bundling_limit(0),
943 _node_bundling_base(NULL),
944 _java_calls(0),
945 _inner_loops(0),
946 #ifndef PRODUCT
947 _trace_opto_output(TraceOptoOutput),
948 _printer(NULL),
949 #endif
950 _dead_node_list(comp_arena()),
951 _dead_node_count(0),
952 _congraph(NULL),
953 _replay_inline_data(NULL),
954 _number_of_mh_late_inlines(0),
955 _inlining_progress(false),
956 _inlining_incrementally(false),
957 _print_inlining_list(NULL),
958 _print_inlining_idx(0),
959 _preserve_jvm_state(0) {
960 C = this;
961
962 #ifndef PRODUCT
963 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
964 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
965 set_print_assembly(PrintFrameConverterAssembly);
966 set_parsed_irreducible_loop(false);
967 #endif
968 CompileWrapper cw(this);
969 Init(/*AliasLevel=*/ 0);
970 init_tf((*generator)());
971
972 {
973 // The following is a dummy for the sake of GraphKit::gen_stub
974 Unique_Node_List for_igvn(comp_arena());
975 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
976 PhaseGVN gvn(Thread::current()->resource_area(),255);
977 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
978 gvn.transform_no_reclaim(top());
979
980 GraphKit kit;
981 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
982 }
983
984 NOT_PRODUCT( verify_graph_edges(); )
985 Code_Gen();
986 if (failing()) return;
987
988
989 // Entry point will be accessed using compile->stub_entry_point();
990 if (code_buffer() == NULL) {
991 Matcher::soft_match_failure();
992 } else {
993 if (PrintAssembly && (WizardMode || Verbose))
994 tty->print_cr("### Stub::%s", stub_name);
995
996 if (!failing()) {
997 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
998
999 // Make the NMethod
1000 // For now we mark the frame as never safe for profile stackwalking
1001 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1002 code_buffer(),
1003 CodeOffsets::frame_never_safe,
1004 // _code_offsets.value(CodeOffsets::Frame_Complete),
1005 frame_size_in_words(),
1006 _oop_map_set,
1007 save_arg_registers);
1008 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1009
1010 _stub_entry_point = rs->entry_point();
1011 }
1012 }
1013 }
1014
1015 //------------------------------Init-------------------------------------------
1016 // Prepare for a single compilation
1017 void Compile::Init(int aliaslevel) {
1018 _unique = 0;
1019 _regalloc = NULL;
1020
1021 _tf = NULL; // filled in later
1022 _top = NULL; // cached later
1023 _matcher = NULL; // filled in later
1024 _cfg = NULL; // filled in later
1025
1026 set_24_bit_selection_and_mode(Use24BitFP, false);
1027
1028 _node_note_array = NULL;
1029 _default_node_notes = NULL;
1030
1031 _immutable_memory = NULL; // filled in at first inquiry
1032
1033 // Globally visible Nodes
1034 // First set TOP to NULL to give safe behavior during creation of RootNode
1035 set_cached_top_node(NULL);
1036 set_root(new (this) RootNode());
1037 // Now that you have a Root to point to, create the real TOP
1038 set_cached_top_node( new (this) ConNode(Type::TOP) );
1039 set_recent_alloc(NULL, NULL);
1040
1041 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1042 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1043 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1044 env()->set_dependencies(new Dependencies(env()));
1045
1046 _fixed_slots = 0;
1047 set_has_split_ifs(false);
1048 set_has_loops(has_method() && method()->has_loops()); // first approximation
1049 set_has_stringbuilder(false);
1050 set_has_boxed_value(false);
1051 _trap_can_recompile = false; // no traps emitted yet
1052 _major_progress = true; // start out assuming good things will happen
1053 set_has_unsafe_access(false);
1054 set_max_vector_size(0);
1055 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1056 set_decompile_count(0);
1057
1058 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1059 set_num_loop_opts(LoopOptsCount);
1060 set_do_inlining(Inline);
1061 set_max_inline_size(MaxInlineSize);
1062 set_freq_inline_size(FreqInlineSize);
1063 set_do_scheduling(OptoScheduling);
1064 set_do_count_invocations(false);
1065 set_do_method_data_update(false);
1066
1067 if (debug_info()->recording_non_safepoints()) {
1068 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1069 (comp_arena(), 8, 0, NULL));
1070 set_default_node_notes(Node_Notes::make(this));
1071 }
1072
1073 // // -- Initialize types before each compile --
1074 // // Update cached type information
1075 // if( _method && _method->constants() )
1076 // Type::update_loaded_types(_method, _method->constants());
1077
1078 // Init alias_type map.
1079 if (!_do_escape_analysis && aliaslevel == 3)
1080 aliaslevel = 2; // No unique types without escape analysis
1081 _AliasLevel = aliaslevel;
1082 const int grow_ats = 16;
1083 _max_alias_types = grow_ats;
1084 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1085 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1086 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1087 {
1088 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1089 }
1090 // Initialize the first few types.
1091 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1092 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1093 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1094 _num_alias_types = AliasIdxRaw+1;
1095 // Zero out the alias type cache.
1096 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1097 // A NULL adr_type hits in the cache right away. Preload the right answer.
1098 probe_alias_cache(NULL)->_index = AliasIdxTop;
1099
1100 _intrinsics = NULL;
1101 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1102 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1103 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1104 register_library_intrinsics();
1105 }
1106
1107 //---------------------------init_start----------------------------------------
1108 // Install the StartNode on this compile object.
1109 void Compile::init_start(StartNode* s) {
1110 if (failing())
1111 return; // already failing
1112 assert(s == start(), "");
1113 }
1114
1115 StartNode* Compile::start() const {
1116 assert(!failing(), "");
1117 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1118 Node* start = root()->fast_out(i);
1119 if( start->is_Start() )
1120 return start->as_Start();
1121 }
1122 ShouldNotReachHere();
1123 return NULL;
1124 }
1125
1126 //-------------------------------immutable_memory-------------------------------------
1127 // Access immutable memory
1128 Node* Compile::immutable_memory() {
1129 if (_immutable_memory != NULL) {
1130 return _immutable_memory;
1131 }
1132 StartNode* s = start();
1133 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1134 Node *p = s->fast_out(i);
1135 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1136 _immutable_memory = p;
1137 return _immutable_memory;
1138 }
1139 }
1140 ShouldNotReachHere();
1141 return NULL;
1142 }
1143
1144 //----------------------set_cached_top_node------------------------------------
1145 // Install the cached top node, and make sure Node::is_top works correctly.
1146 void Compile::set_cached_top_node(Node* tn) {
1147 if (tn != NULL) verify_top(tn);
1148 Node* old_top = _top;
1149 _top = tn;
1150 // Calling Node::setup_is_top allows the nodes the chance to adjust
1151 // their _out arrays.
1152 if (_top != NULL) _top->setup_is_top();
1153 if (old_top != NULL) old_top->setup_is_top();
1154 assert(_top == NULL || top()->is_top(), "");
1155 }
1156
1157 #ifdef ASSERT
1158 uint Compile::count_live_nodes_by_graph_walk() {
1159 Unique_Node_List useful(comp_arena());
1160 // Get useful node list by walking the graph.
1161 identify_useful_nodes(useful);
1162 return useful.size();
1163 }
1164
1165 void Compile::print_missing_nodes() {
1166
1167 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1168 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1169 return;
1170 }
1171
1172 // This is an expensive function. It is executed only when the user
1173 // specifies VerifyIdealNodeCount option or otherwise knows the
1174 // additional work that needs to be done to identify reachable nodes
1175 // by walking the flow graph and find the missing ones using
1176 // _dead_node_list.
1177
1178 Unique_Node_List useful(comp_arena());
1179 // Get useful node list by walking the graph.
1180 identify_useful_nodes(useful);
1181
1182 uint l_nodes = C->live_nodes();
1183 uint l_nodes_by_walk = useful.size();
1184
1185 if (l_nodes != l_nodes_by_walk) {
1186 if (_log != NULL) {
1187 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1188 _log->stamp();
1189 _log->end_head();
1190 }
1191 VectorSet& useful_member_set = useful.member_set();
1192 int last_idx = l_nodes_by_walk;
1193 for (int i = 0; i < last_idx; i++) {
1194 if (useful_member_set.test(i)) {
1195 if (_dead_node_list.test(i)) {
1196 if (_log != NULL) {
1197 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1198 }
1199 if (PrintIdealNodeCount) {
1200 // Print the log message to tty
1201 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1202 useful.at(i)->dump();
1203 }
1204 }
1205 }
1206 else if (! _dead_node_list.test(i)) {
1207 if (_log != NULL) {
1208 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1209 }
1210 if (PrintIdealNodeCount) {
1211 // Print the log message to tty
1212 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1213 }
1214 }
1215 }
1216 if (_log != NULL) {
1217 _log->tail("mismatched_nodes");
1218 }
1219 }
1220 }
1221 #endif
1222
1223 #ifndef PRODUCT
1224 void Compile::verify_top(Node* tn) const {
1225 if (tn != NULL) {
1226 assert(tn->is_Con(), "top node must be a constant");
1227 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1228 assert(tn->in(0) != NULL, "must have live top node");
1229 }
1230 }
1231 #endif
1232
1233
1234 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1235
1236 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1237 guarantee(arr != NULL, "");
1238 int num_blocks = arr->length();
1239 if (grow_by < num_blocks) grow_by = num_blocks;
1240 int num_notes = grow_by * _node_notes_block_size;
1241 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1242 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1243 while (num_notes > 0) {
1244 arr->append(notes);
1245 notes += _node_notes_block_size;
1246 num_notes -= _node_notes_block_size;
1247 }
1248 assert(num_notes == 0, "exact multiple, please");
1249 }
1250
1251 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1252 if (source == NULL || dest == NULL) return false;
1253
1254 if (dest->is_Con())
1255 return false; // Do not push debug info onto constants.
1256
1257 #ifdef ASSERT
1258 // Leave a bread crumb trail pointing to the original node:
1259 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1260 dest->set_debug_orig(source);
1261 }
1262 #endif
1263
1264 if (node_note_array() == NULL)
1265 return false; // Not collecting any notes now.
1266
1267 // This is a copy onto a pre-existing node, which may already have notes.
1268 // If both nodes have notes, do not overwrite any pre-existing notes.
1269 Node_Notes* source_notes = node_notes_at(source->_idx);
1270 if (source_notes == NULL || source_notes->is_clear()) return false;
1271 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1272 if (dest_notes == NULL || dest_notes->is_clear()) {
1273 return set_node_notes_at(dest->_idx, source_notes);
1274 }
1275
1276 Node_Notes merged_notes = (*source_notes);
1277 // The order of operations here ensures that dest notes will win...
1278 merged_notes.update_from(dest_notes);
1279 return set_node_notes_at(dest->_idx, &merged_notes);
1280 }
1281
1282
1283 //--------------------------allow_range_check_smearing-------------------------
1284 // Gating condition for coalescing similar range checks.
1285 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1286 // single covering check that is at least as strong as any of them.
1287 // If the optimization succeeds, the simplified (strengthened) range check
1288 // will always succeed. If it fails, we will deopt, and then give up
1289 // on the optimization.
1290 bool Compile::allow_range_check_smearing() const {
1291 // If this method has already thrown a range-check,
1292 // assume it was because we already tried range smearing
1293 // and it failed.
1294 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1295 return !already_trapped;
1296 }
1297
1298
1299 //------------------------------flatten_alias_type-----------------------------
1300 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1301 int offset = tj->offset();
1302 TypePtr::PTR ptr = tj->ptr();
1303
1304 // Known instance (scalarizable allocation) alias only with itself.
1305 bool is_known_inst = tj->isa_oopptr() != NULL &&
1306 tj->is_oopptr()->is_known_instance();
1307
1308 // Process weird unsafe references.
1309 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1310 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1311 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1312 tj = TypeOopPtr::BOTTOM;
1313 ptr = tj->ptr();
1314 offset = tj->offset();
1315 }
1316
1317 // Array pointers need some flattening
1318 const TypeAryPtr *ta = tj->isa_aryptr();
1319 if (ta && ta->is_stable()) {
1320 // Erase stability property for alias analysis.
1321 tj = ta = ta->cast_to_stable(false);
1322 }
1323 if( ta && is_known_inst ) {
1324 if ( offset != Type::OffsetBot &&
1325 offset > arrayOopDesc::length_offset_in_bytes() ) {
1326 offset = Type::OffsetBot; // Flatten constant access into array body only
1327 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1328 }
1329 } else if( ta && _AliasLevel >= 2 ) {
1330 // For arrays indexed by constant indices, we flatten the alias
1331 // space to include all of the array body. Only the header, klass
1332 // and array length can be accessed un-aliased.
1333 if( offset != Type::OffsetBot ) {
1334 if( ta->const_oop() ) { // MethodData* or Method*
1335 offset = Type::OffsetBot; // Flatten constant access into array body
1336 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1337 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1338 // range is OK as-is.
1339 tj = ta = TypeAryPtr::RANGE;
1340 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1341 tj = TypeInstPtr::KLASS; // all klass loads look alike
1342 ta = TypeAryPtr::RANGE; // generic ignored junk
1343 ptr = TypePtr::BotPTR;
1344 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1345 tj = TypeInstPtr::MARK;
1346 ta = TypeAryPtr::RANGE; // generic ignored junk
1347 ptr = TypePtr::BotPTR;
1348 } else { // Random constant offset into array body
1349 offset = Type::OffsetBot; // Flatten constant access into array body
1350 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1351 }
1352 }
1353 // Arrays of fixed size alias with arrays of unknown size.
1354 if (ta->size() != TypeInt::POS) {
1355 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1356 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1357 }
1358 // Arrays of known objects become arrays of unknown objects.
1359 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1360 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1361 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1362 }
1363 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1364 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1365 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1366 }
1367 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1368 // cannot be distinguished by bytecode alone.
1369 if (ta->elem() == TypeInt::BOOL) {
1370 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1371 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1372 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1373 }
1374 // During the 2nd round of IterGVN, NotNull castings are removed.
1375 // Make sure the Bottom and NotNull variants alias the same.
1376 // Also, make sure exact and non-exact variants alias the same.
1377 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1378 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1379 }
1380 }
1381
1382 // Oop pointers need some flattening
1383 const TypeInstPtr *to = tj->isa_instptr();
1384 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1385 ciInstanceKlass *k = to->klass()->as_instance_klass();
1386 if( ptr == TypePtr::Constant ) {
1387 if (to->klass() != ciEnv::current()->Class_klass() ||
1388 offset < k->size_helper() * wordSize) {
1389 // No constant oop pointers (such as Strings); they alias with
1390 // unknown strings.
1391 assert(!is_known_inst, "not scalarizable allocation");
1392 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1393 }
1394 } else if( is_known_inst ) {
1395 tj = to; // Keep NotNull and klass_is_exact for instance type
1396 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1397 // During the 2nd round of IterGVN, NotNull castings are removed.
1398 // Make sure the Bottom and NotNull variants alias the same.
1399 // Also, make sure exact and non-exact variants alias the same.
1400 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1401 }
1402 if (to->speculative() != NULL) {
1403 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1404 }
1405 // Canonicalize the holder of this field
1406 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1407 // First handle header references such as a LoadKlassNode, even if the
1408 // object's klass is unloaded at compile time (4965979).
1409 if (!is_known_inst) { // Do it only for non-instance types
1410 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1411 }
1412 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1413 // Static fields are in the space above the normal instance
1414 // fields in the java.lang.Class instance.
1415 if (to->klass() != ciEnv::current()->Class_klass()) {
1416 to = NULL;
1417 tj = TypeOopPtr::BOTTOM;
1418 offset = tj->offset();
1419 }
1420 } else {
1421 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1422 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1423 if( is_known_inst ) {
1424 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1425 } else {
1426 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1427 }
1428 }
1429 }
1430 }
1431
1432 // Klass pointers to object array klasses need some flattening
1433 const TypeKlassPtr *tk = tj->isa_klassptr();
1434 if( tk ) {
1435 // If we are referencing a field within a Klass, we need
1436 // to assume the worst case of an Object. Both exact and
1437 // inexact types must flatten to the same alias class so
1438 // use NotNull as the PTR.
1439 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1440
1441 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1442 TypeKlassPtr::OBJECT->klass(),
1443 offset);
1444 }
1445
1446 ciKlass* klass = tk->klass();
1447 if( klass->is_obj_array_klass() ) {
1448 ciKlass* k = TypeAryPtr::OOPS->klass();
1449 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1450 k = TypeInstPtr::BOTTOM->klass();
1451 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1452 }
1453
1454 // Check for precise loads from the primary supertype array and force them
1455 // to the supertype cache alias index. Check for generic array loads from
1456 // the primary supertype array and also force them to the supertype cache
1457 // alias index. Since the same load can reach both, we need to merge
1458 // these 2 disparate memories into the same alias class. Since the
1459 // primary supertype array is read-only, there's no chance of confusion
1460 // where we bypass an array load and an array store.
1461 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1462 if (offset == Type::OffsetBot ||
1463 (offset >= primary_supers_offset &&
1464 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1465 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1466 offset = in_bytes(Klass::secondary_super_cache_offset());
1467 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1468 }
1469 }
1470
1471 // Flatten all Raw pointers together.
1472 if (tj->base() == Type::RawPtr)
1473 tj = TypeRawPtr::BOTTOM;
1474
1475 if (tj->base() == Type::AnyPtr)
1476 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1477
1478 // Flatten all to bottom for now
1479 switch( _AliasLevel ) {
1480 case 0:
1481 tj = TypePtr::BOTTOM;
1482 break;
1483 case 1: // Flatten to: oop, static, field or array
1484 switch (tj->base()) {
1485 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1486 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1487 case Type::AryPtr: // do not distinguish arrays at all
1488 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1489 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1490 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1491 default: ShouldNotReachHere();
1492 }
1493 break;
1494 case 2: // No collapsing at level 2; keep all splits
1495 case 3: // No collapsing at level 3; keep all splits
1496 break;
1497 default:
1498 Unimplemented();
1499 }
1500
1501 offset = tj->offset();
1502 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1503
1504 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1505 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1506 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1507 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1508 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1509 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1510 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1511 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1512 assert( tj->ptr() != TypePtr::TopPTR &&
1513 tj->ptr() != TypePtr::AnyNull &&
1514 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1515 // assert( tj->ptr() != TypePtr::Constant ||
1516 // tj->base() == Type::RawPtr ||
1517 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1518
1519 return tj;
1520 }
1521
1522 void Compile::AliasType::Init(int i, const TypePtr* at) {
1523 _index = i;
1524 _adr_type = at;
1525 _field = NULL;
1526 _element = NULL;
1527 _is_rewritable = true; // default
1528 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1529 if (atoop != NULL && atoop->is_known_instance()) {
1530 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1531 _general_index = Compile::current()->get_alias_index(gt);
1532 } else {
1533 _general_index = 0;
1534 }
1535 }
1536
1537 //---------------------------------print_on------------------------------------
1538 #ifndef PRODUCT
1539 void Compile::AliasType::print_on(outputStream* st) {
1540 if (index() < 10)
1541 st->print("@ <%d> ", index());
1542 else st->print("@ <%d>", index());
1543 st->print(is_rewritable() ? " " : " RO");
1544 int offset = adr_type()->offset();
1545 if (offset == Type::OffsetBot)
1546 st->print(" +any");
1547 else st->print(" +%-3d", offset);
1548 st->print(" in ");
1549 adr_type()->dump_on(st);
1550 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1551 if (field() != NULL && tjp) {
1552 if (tjp->klass() != field()->holder() ||
1553 tjp->offset() != field()->offset_in_bytes()) {
1554 st->print(" != ");
1555 field()->print();
1556 st->print(" ***");
1557 }
1558 }
1559 }
1560
1561 void print_alias_types() {
1562 Compile* C = Compile::current();
1563 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1564 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1565 C->alias_type(idx)->print_on(tty);
1566 tty->cr();
1567 }
1568 }
1569 #endif
1570
1571
1572 //----------------------------probe_alias_cache--------------------------------
1573 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1574 intptr_t key = (intptr_t) adr_type;
1575 key ^= key >> logAliasCacheSize;
1576 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1577 }
1578
1579
1580 //-----------------------------grow_alias_types--------------------------------
1581 void Compile::grow_alias_types() {
1582 const int old_ats = _max_alias_types; // how many before?
1583 const int new_ats = old_ats; // how many more?
1584 const int grow_ats = old_ats+new_ats; // how many now?
1585 _max_alias_types = grow_ats;
1586 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1587 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1588 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1589 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1590 }
1591
1592
1593 //--------------------------------find_alias_type------------------------------
1594 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1595 if (_AliasLevel == 0)
1596 return alias_type(AliasIdxBot);
1597
1598 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1599 if (ace->_adr_type == adr_type) {
1600 return alias_type(ace->_index);
1601 }
1602
1603 // Handle special cases.
1604 if (adr_type == NULL) return alias_type(AliasIdxTop);
1605 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1606
1607 // Do it the slow way.
1608 const TypePtr* flat = flatten_alias_type(adr_type);
1609
1610 #ifdef ASSERT
1611 assert(flat == flatten_alias_type(flat), "idempotent");
1612 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1613 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1614 const TypeOopPtr* foop = flat->is_oopptr();
1615 // Scalarizable allocations have exact klass always.
1616 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1617 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1618 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1619 }
1620 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1621 #endif
1622
1623 int idx = AliasIdxTop;
1624 for (int i = 0; i < num_alias_types(); i++) {
1625 if (alias_type(i)->adr_type() == flat) {
1626 idx = i;
1627 break;
1628 }
1629 }
1630
1631 if (idx == AliasIdxTop) {
1632 if (no_create) return NULL;
1633 // Grow the array if necessary.
1634 if (_num_alias_types == _max_alias_types) grow_alias_types();
1635 // Add a new alias type.
1636 idx = _num_alias_types++;
1637 _alias_types[idx]->Init(idx, flat);
1638 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1639 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1640 if (flat->isa_instptr()) {
1641 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1642 && flat->is_instptr()->klass() == env()->Class_klass())
1643 alias_type(idx)->set_rewritable(false);
1644 }
1645 if (flat->isa_aryptr()) {
1646 #ifdef ASSERT
1647 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1648 // (T_BYTE has the weakest alignment and size restrictions...)
1649 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1650 #endif
1651 if (flat->offset() == TypePtr::OffsetBot) {
1652 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1653 }
1654 }
1655 if (flat->isa_klassptr()) {
1656 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1657 alias_type(idx)->set_rewritable(false);
1658 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1659 alias_type(idx)->set_rewritable(false);
1660 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1661 alias_type(idx)->set_rewritable(false);
1662 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1663 alias_type(idx)->set_rewritable(false);
1664 }
1665 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1666 // but the base pointer type is not distinctive enough to identify
1667 // references into JavaThread.)
1668
1669 // Check for final fields.
1670 const TypeInstPtr* tinst = flat->isa_instptr();
1671 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1672 ciField* field;
1673 if (tinst->const_oop() != NULL &&
1674 tinst->klass() == ciEnv::current()->Class_klass() &&
1675 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1676 // static field
1677 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1678 field = k->get_field_by_offset(tinst->offset(), true);
1679 } else {
1680 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1681 field = k->get_field_by_offset(tinst->offset(), false);
1682 }
1683 assert(field == NULL ||
1684 original_field == NULL ||
1685 (field->holder() == original_field->holder() &&
1686 field->offset() == original_field->offset() &&
1687 field->is_static() == original_field->is_static()), "wrong field?");
1688 // Set field() and is_rewritable() attributes.
1689 if (field != NULL) alias_type(idx)->set_field(field);
1690 }
1691 }
1692
1693 // Fill the cache for next time.
1694 ace->_adr_type = adr_type;
1695 ace->_index = idx;
1696 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1697
1698 // Might as well try to fill the cache for the flattened version, too.
1699 AliasCacheEntry* face = probe_alias_cache(flat);
1700 if (face->_adr_type == NULL) {
1701 face->_adr_type = flat;
1702 face->_index = idx;
1703 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1704 }
1705
1706 return alias_type(idx);
1707 }
1708
1709
1710 Compile::AliasType* Compile::alias_type(ciField* field) {
1711 const TypeOopPtr* t;
1712 if (field->is_static())
1713 t = TypeInstPtr::make(field->holder()->java_mirror());
1714 else
1715 t = TypeOopPtr::make_from_klass_raw(field->holder());
1716 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1717 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1718 return atp;
1719 }
1720
1721
1722 //------------------------------have_alias_type--------------------------------
1723 bool Compile::have_alias_type(const TypePtr* adr_type) {
1724 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1725 if (ace->_adr_type == adr_type) {
1726 return true;
1727 }
1728
1729 // Handle special cases.
1730 if (adr_type == NULL) return true;
1731 if (adr_type == TypePtr::BOTTOM) return true;
1732
1733 return find_alias_type(adr_type, true, NULL) != NULL;
1734 }
1735
1736 //-----------------------------must_alias--------------------------------------
1737 // True if all values of the given address type are in the given alias category.
1738 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1739 if (alias_idx == AliasIdxBot) return true; // the universal category
1740 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1741 if (alias_idx == AliasIdxTop) return false; // the empty category
1742 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1743
1744 // the only remaining possible overlap is identity
1745 int adr_idx = get_alias_index(adr_type);
1746 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1747 assert(adr_idx == alias_idx ||
1748 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1749 && adr_type != TypeOopPtr::BOTTOM),
1750 "should not be testing for overlap with an unsafe pointer");
1751 return adr_idx == alias_idx;
1752 }
1753
1754 //------------------------------can_alias--------------------------------------
1755 // True if any values of the given address type are in the given alias category.
1756 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1757 if (alias_idx == AliasIdxTop) return false; // the empty category
1758 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1759 if (alias_idx == AliasIdxBot) return true; // the universal category
1760 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1761
1762 // the only remaining possible overlap is identity
1763 int adr_idx = get_alias_index(adr_type);
1764 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1765 return adr_idx == alias_idx;
1766 }
1767
1768
1769
1770 //---------------------------pop_warm_call-------------------------------------
1771 WarmCallInfo* Compile::pop_warm_call() {
1772 WarmCallInfo* wci = _warm_calls;
1773 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1774 return wci;
1775 }
1776
1777 //----------------------------Inline_Warm--------------------------------------
1778 int Compile::Inline_Warm() {
1779 // If there is room, try to inline some more warm call sites.
1780 // %%% Do a graph index compaction pass when we think we're out of space?
1781 if (!InlineWarmCalls) return 0;
1782
1783 int calls_made_hot = 0;
1784 int room_to_grow = NodeCountInliningCutoff - unique();
1785 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1786 int amount_grown = 0;
1787 WarmCallInfo* call;
1788 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1789 int est_size = (int)call->size();
1790 if (est_size > (room_to_grow - amount_grown)) {
1791 // This one won't fit anyway. Get rid of it.
1792 call->make_cold();
1793 continue;
1794 }
1795 call->make_hot();
1796 calls_made_hot++;
1797 amount_grown += est_size;
1798 amount_to_grow -= est_size;
1799 }
1800
1801 if (calls_made_hot > 0) set_major_progress();
1802 return calls_made_hot;
1803 }
1804
1805
1806 //----------------------------Finish_Warm--------------------------------------
1807 void Compile::Finish_Warm() {
1808 if (!InlineWarmCalls) return;
1809 if (failing()) return;
1810 if (warm_calls() == NULL) return;
1811
1812 // Clean up loose ends, if we are out of space for inlining.
1813 WarmCallInfo* call;
1814 while ((call = pop_warm_call()) != NULL) {
1815 call->make_cold();
1816 }
1817 }
1818
1819 //---------------------cleanup_loop_predicates-----------------------
1820 // Remove the opaque nodes that protect the predicates so that all unused
1821 // checks and uncommon_traps will be eliminated from the ideal graph
1822 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1823 if (predicate_count()==0) return;
1824 for (int i = predicate_count(); i > 0; i--) {
1825 Node * n = predicate_opaque1_node(i-1);
1826 assert(n->Opcode() == Op_Opaque1, "must be");
1827 igvn.replace_node(n, n->in(1));
1828 }
1829 assert(predicate_count()==0, "should be clean!");
1830 }
1831
1832 // StringOpts and late inlining of string methods
1833 void Compile::inline_string_calls(bool parse_time) {
1834 {
1835 // remove useless nodes to make the usage analysis simpler
1836 ResourceMark rm;
1837 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1838 }
1839
1840 {
1841 ResourceMark rm;
1842 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1843 PhaseStringOpts pso(initial_gvn(), for_igvn());
1844 print_method(PHASE_AFTER_STRINGOPTS, 3);
1845 }
1846
1847 // now inline anything that we skipped the first time around
1848 if (!parse_time) {
1849 _late_inlines_pos = _late_inlines.length();
1850 }
1851
1852 while (_string_late_inlines.length() > 0) {
1853 CallGenerator* cg = _string_late_inlines.pop();
1854 cg->do_late_inline();
1855 if (failing()) return;
1856 }
1857 _string_late_inlines.trunc_to(0);
1858 }
1859
1860 // Late inlining of boxing methods
1861 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1862 if (_boxing_late_inlines.length() > 0) {
1863 assert(has_boxed_value(), "inconsistent");
1864
1865 PhaseGVN* gvn = initial_gvn();
1866 set_inlining_incrementally(true);
1867
1868 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1869 for_igvn()->clear();
1870 gvn->replace_with(&igvn);
1871
1872 while (_boxing_late_inlines.length() > 0) {
1873 CallGenerator* cg = _boxing_late_inlines.pop();
1874 cg->do_late_inline();
1875 if (failing()) return;
1876 }
1877 _boxing_late_inlines.trunc_to(0);
1878
1879 {
1880 ResourceMark rm;
1881 PhaseRemoveUseless pru(gvn, for_igvn());
1882 }
1883
1884 igvn = PhaseIterGVN(gvn);
1885 igvn.optimize();
1886
1887 set_inlining_progress(false);
1888 set_inlining_incrementally(false);
1889 }
1890 }
1891
1892 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1893 assert(IncrementalInline, "incremental inlining should be on");
1894 PhaseGVN* gvn = initial_gvn();
1895
1896 set_inlining_progress(false);
1897 for_igvn()->clear();
1898 gvn->replace_with(&igvn);
1899
1900 int i = 0;
1901
1902 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1903 CallGenerator* cg = _late_inlines.at(i);
1904 _late_inlines_pos = i+1;
1905 cg->do_late_inline();
1906 if (failing()) return;
1907 }
1908 int j = 0;
1909 for (; i < _late_inlines.length(); i++, j++) {
1910 _late_inlines.at_put(j, _late_inlines.at(i));
1911 }
1912 _late_inlines.trunc_to(j);
1913
1914 {
1915 ResourceMark rm;
1916 PhaseRemoveUseless pru(gvn, for_igvn());
1917 }
1918
1919 igvn = PhaseIterGVN(gvn);
1920 }
1921
1922 // Perform incremental inlining until bound on number of live nodes is reached
1923 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1924 PhaseGVN* gvn = initial_gvn();
1925
1926 set_inlining_incrementally(true);
1927 set_inlining_progress(true);
1928 uint low_live_nodes = 0;
1929
1930 while(inlining_progress() && _late_inlines.length() > 0) {
1931
1932 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1933 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1934 // PhaseIdealLoop is expensive so we only try it once we are
1935 // out of loop and we only try it again if the previous helped
1936 // got the number of nodes down significantly
1937 PhaseIdealLoop ideal_loop( igvn, false, true );
1938 if (failing()) return;
1939 low_live_nodes = live_nodes();
1940 _major_progress = true;
1941 }
1942
1943 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1944 break;
1945 }
1946 }
1947
1948 inline_incrementally_one(igvn);
1949
1950 if (failing()) return;
1951
1952 igvn.optimize();
1953
1954 if (failing()) return;
1955 }
1956
1957 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1958
1959 if (_string_late_inlines.length() > 0) {
1960 assert(has_stringbuilder(), "inconsistent");
1961 for_igvn()->clear();
1962 initial_gvn()->replace_with(&igvn);
1963
1964 inline_string_calls(false);
1965
1966 if (failing()) return;
1967
1968 {
1969 ResourceMark rm;
1970 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1971 }
1972
1973 igvn = PhaseIterGVN(gvn);
1974
1975 igvn.optimize();
1976 }
1977
1978 set_inlining_incrementally(false);
1979 }
1980
1981
1982 //------------------------------Optimize---------------------------------------
1983 // Given a graph, optimize it.
1984 void Compile::Optimize() {
1985 TracePhase t1("optimizer", &_t_optimizer, true);
1986
1987 #ifndef PRODUCT
1988 if (env()->break_at_compile()) {
1989 BREAKPOINT;
1990 }
1991
1992 #endif
1993
1994 ResourceMark rm;
1995 int loop_opts_cnt;
1996
1997 NOT_PRODUCT( verify_graph_edges(); )
1998
1999 print_method(PHASE_AFTER_PARSING);
2000
2001 {
2002 // Iterative Global Value Numbering, including ideal transforms
2003 // Initialize IterGVN with types and values from parse-time GVN
2004 PhaseIterGVN igvn(initial_gvn());
2005 {
2006 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2007 igvn.optimize();
2008 }
2009
2010 print_method(PHASE_ITER_GVN1, 2);
2011
2012 if (failing()) return;
2013
2014 {
2015 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2016 inline_incrementally(igvn);
2017 }
2018
2019 print_method(PHASE_INCREMENTAL_INLINE, 2);
2020
2021 if (failing()) return;
2022
2023 if (eliminate_boxing()) {
2024 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2025 // Inline valueOf() methods now.
2026 inline_boxing_calls(igvn);
2027
2028 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2029
2030 if (failing()) return;
2031 }
2032
2033 // Remove the speculative part of types and clean up the graph from
2034 // the extra CastPP nodes whose only purpose is to carry them. Do
2035 // that early so that optimizations are not disrupted by the extra
2036 // CastPP nodes.
2037 remove_speculative_types(igvn);
2038
2039 // No more new expensive nodes will be added to the list from here
2040 // so keep only the actual candidates for optimizations.
2041 cleanup_expensive_nodes(igvn);
2042
2043 // Perform escape analysis
2044 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2045 if (has_loops()) {
2046 // Cleanup graph (remove dead nodes).
2047 TracePhase t2("idealLoop", &_t_idealLoop, true);
2048 PhaseIdealLoop ideal_loop( igvn, false, true );
2049 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2050 if (failing()) return;
2051 }
2052 ConnectionGraph::do_analysis(this, &igvn);
2053
2054 if (failing()) return;
2055
2056 // Optimize out fields loads from scalar replaceable allocations.
2057 igvn.optimize();
2058 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2059
2060 if (failing()) return;
2061
2062 if (congraph() != NULL && macro_count() > 0) {
2063 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2064 PhaseMacroExpand mexp(igvn);
2065 mexp.eliminate_macro_nodes();
2066 igvn.set_delay_transform(false);
2067
2068 igvn.optimize();
2069 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2070
2071 if (failing()) return;
2072 }
2073 }
2074
2075 // Loop transforms on the ideal graph. Range Check Elimination,
2076 // peeling, unrolling, etc.
2077
2078 // Set loop opts counter
2079 loop_opts_cnt = num_loop_opts();
2080 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2081 {
2082 TracePhase t2("idealLoop", &_t_idealLoop, true);
2083 PhaseIdealLoop ideal_loop( igvn, true );
2084 loop_opts_cnt--;
2085 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2086 if (failing()) return;
2087 }
2088 // Loop opts pass if partial peeling occurred in previous pass
2089 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2090 TracePhase t3("idealLoop", &_t_idealLoop, true);
2091 PhaseIdealLoop ideal_loop( igvn, false );
2092 loop_opts_cnt--;
2093 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2094 if (failing()) return;
2095 }
2096 // Loop opts pass for loop-unrolling before CCP
2097 if(major_progress() && (loop_opts_cnt > 0)) {
2098 TracePhase t4("idealLoop", &_t_idealLoop, true);
2099 PhaseIdealLoop ideal_loop( igvn, false );
2100 loop_opts_cnt--;
2101 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2102 }
2103 if (!failing()) {
2104 // Verify that last round of loop opts produced a valid graph
2105 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2106 PhaseIdealLoop::verify(igvn);
2107 }
2108 }
2109 if (failing()) return;
2110
2111 // Conditional Constant Propagation;
2112 PhaseCCP ccp( &igvn );
2113 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2114 {
2115 TracePhase t2("ccp", &_t_ccp, true);
2116 ccp.do_transform();
2117 }
2118 print_method(PHASE_CPP1, 2);
2119
2120 assert( true, "Break here to ccp.dump_old2new_map()");
2121
2122 // Iterative Global Value Numbering, including ideal transforms
2123 {
2124 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2125 igvn = ccp;
2126 igvn.optimize();
2127 }
2128
2129 print_method(PHASE_ITER_GVN2, 2);
2130
2131 if (failing()) return;
2132
2133 // Loop transforms on the ideal graph. Range Check Elimination,
2134 // peeling, unrolling, etc.
2135 if(loop_opts_cnt > 0) {
2136 debug_only( int cnt = 0; );
2137 while(major_progress() && (loop_opts_cnt > 0)) {
2138 TracePhase t2("idealLoop", &_t_idealLoop, true);
2139 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2140 PhaseIdealLoop ideal_loop( igvn, true);
2141 loop_opts_cnt--;
2142 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2143 if (failing()) return;
2144 }
2145 }
2146
2147 {
2148 // Verify that all previous optimizations produced a valid graph
2149 // at least to this point, even if no loop optimizations were done.
2150 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2151 PhaseIdealLoop::verify(igvn);
2152 }
2153
2154 {
2155 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2156 PhaseMacroExpand mex(igvn);
2157 if (mex.expand_macro_nodes()) {
2158 assert(failing(), "must bail out w/ explicit message");
2159 return;
2160 }
2161 }
2162
2163 } // (End scope of igvn; run destructor if necessary for asserts.)
2164
2165 dump_inlining();
2166 // A method with only infinite loops has no edges entering loops from root
2167 {
2168 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2169 if (final_graph_reshaping()) {
2170 assert(failing(), "must bail out w/ explicit message");
2171 return;
2172 }
2173 }
2174
2175 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2176 }
2177
2178
2179 //------------------------------Code_Gen---------------------------------------
2180 // Given a graph, generate code for it
2181 void Compile::Code_Gen() {
2182 if (failing()) {
2183 return;
2184 }
2185
2186 // Perform instruction selection. You might think we could reclaim Matcher
2187 // memory PDQ, but actually the Matcher is used in generating spill code.
2188 // Internals of the Matcher (including some VectorSets) must remain live
2189 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2190 // set a bit in reclaimed memory.
2191
2192 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2193 // nodes. Mapping is only valid at the root of each matched subtree.
2194 NOT_PRODUCT( verify_graph_edges(); )
2195
2196 Matcher matcher;
2197 _matcher = &matcher;
2198 {
2199 TracePhase t2("matcher", &_t_matcher, true);
2200 matcher.match();
2201 }
2202 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2203 // nodes. Mapping is only valid at the root of each matched subtree.
2204 NOT_PRODUCT( verify_graph_edges(); )
2205
2206 // If you have too many nodes, or if matching has failed, bail out
2207 check_node_count(0, "out of nodes matching instructions");
2208 if (failing()) {
2209 return;
2210 }
2211
2212 // Build a proper-looking CFG
2213 PhaseCFG cfg(node_arena(), root(), matcher);
2214 _cfg = &cfg;
2215 {
2216 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2217 bool success = cfg.do_global_code_motion();
2218 if (!success) {
2219 return;
2220 }
2221
2222 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2223 NOT_PRODUCT( verify_graph_edges(); )
2224 debug_only( cfg.verify(); )
2225 }
2226
2227 PhaseChaitin regalloc(unique(), cfg, matcher);
2228 _regalloc = ®alloc;
2229 {
2230 TracePhase t2("regalloc", &_t_registerAllocation, true);
2231 // Perform register allocation. After Chaitin, use-def chains are
2232 // no longer accurate (at spill code) and so must be ignored.
2233 // Node->LRG->reg mappings are still accurate.
2234 _regalloc->Register_Allocate();
2235
2236 // Bail out if the allocator builds too many nodes
2237 if (failing()) {
2238 return;
2239 }
2240 }
2241
2242 // Prior to register allocation we kept empty basic blocks in case the
2243 // the allocator needed a place to spill. After register allocation we
2244 // are not adding any new instructions. If any basic block is empty, we
2245 // can now safely remove it.
2246 {
2247 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2248 cfg.remove_empty_blocks();
2249 if (do_freq_based_layout()) {
2250 PhaseBlockLayout layout(cfg);
2251 } else {
2252 cfg.set_loop_alignment();
2253 }
2254 cfg.fixup_flow();
2255 }
2256
2257 // Apply peephole optimizations
2258 if( OptoPeephole ) {
2259 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2260 PhasePeephole peep( _regalloc, cfg);
2261 peep.do_transform();
2262 }
2263
2264 // Convert Nodes to instruction bits in a buffer
2265 {
2266 // %%%% workspace merge brought two timers together for one job
2267 TracePhase t2a("output", &_t_output, true);
2268 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2269 Output();
2270 }
2271
2272 print_method(PHASE_FINAL_CODE);
2273
2274 // He's dead, Jim.
2275 _cfg = (PhaseCFG*)0xdeadbeef;
2276 _regalloc = (PhaseChaitin*)0xdeadbeef;
2277 }
2278
2279
2280 //------------------------------dump_asm---------------------------------------
2281 // Dump formatted assembly
2282 #ifndef PRODUCT
2283 void Compile::dump_asm(int *pcs, uint pc_limit) {
2284 bool cut_short = false;
2285 tty->print_cr("#");
2286 tty->print("# "); _tf->dump(); tty->cr();
2287 tty->print_cr("#");
2288
2289 // For all blocks
2290 int pc = 0x0; // Program counter
2291 char starts_bundle = ' ';
2292 _regalloc->dump_frame();
2293
2294 Node *n = NULL;
2295 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2296 if (VMThread::should_terminate()) {
2297 cut_short = true;
2298 break;
2299 }
2300 Block* block = _cfg->get_block(i);
2301 if (block->is_connector() && !Verbose) {
2302 continue;
2303 }
2304 n = block->head();
2305 if (pcs && n->_idx < pc_limit) {
2306 tty->print("%3.3x ", pcs[n->_idx]);
2307 } else {
2308 tty->print(" ");
2309 }
2310 block->dump_head(_cfg);
2311 if (block->is_connector()) {
2312 tty->print_cr(" # Empty connector block");
2313 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2314 tty->print_cr(" # Block is sole successor of call");
2315 }
2316
2317 // For all instructions
2318 Node *delay = NULL;
2319 for (uint j = 0; j < block->number_of_nodes(); j++) {
2320 if (VMThread::should_terminate()) {
2321 cut_short = true;
2322 break;
2323 }
2324 n = block->get_node(j);
2325 if (valid_bundle_info(n)) {
2326 Bundle* bundle = node_bundling(n);
2327 if (bundle->used_in_unconditional_delay()) {
2328 delay = n;
2329 continue;
2330 }
2331 if (bundle->starts_bundle()) {
2332 starts_bundle = '+';
2333 }
2334 }
2335
2336 if (WizardMode) {
2337 n->dump();
2338 }
2339
2340 if( !n->is_Region() && // Dont print in the Assembly
2341 !n->is_Phi() && // a few noisely useless nodes
2342 !n->is_Proj() &&
2343 !n->is_MachTemp() &&
2344 !n->is_SafePointScalarObject() &&
2345 !n->is_Catch() && // Would be nice to print exception table targets
2346 !n->is_MergeMem() && // Not very interesting
2347 !n->is_top() && // Debug info table constants
2348 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2349 ) {
2350 if (pcs && n->_idx < pc_limit)
2351 tty->print("%3.3x", pcs[n->_idx]);
2352 else
2353 tty->print(" ");
2354 tty->print(" %c ", starts_bundle);
2355 starts_bundle = ' ';
2356 tty->print("\t");
2357 n->format(_regalloc, tty);
2358 tty->cr();
2359 }
2360
2361 // If we have an instruction with a delay slot, and have seen a delay,
2362 // then back up and print it
2363 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2364 assert(delay != NULL, "no unconditional delay instruction");
2365 if (WizardMode) delay->dump();
2366
2367 if (node_bundling(delay)->starts_bundle())
2368 starts_bundle = '+';
2369 if (pcs && n->_idx < pc_limit)
2370 tty->print("%3.3x", pcs[n->_idx]);
2371 else
2372 tty->print(" ");
2373 tty->print(" %c ", starts_bundle);
2374 starts_bundle = ' ';
2375 tty->print("\t");
2376 delay->format(_regalloc, tty);
2377 tty->print_cr("");
2378 delay = NULL;
2379 }
2380
2381 // Dump the exception table as well
2382 if( n->is_Catch() && (Verbose || WizardMode) ) {
2383 // Print the exception table for this offset
2384 _handler_table.print_subtable_for(pc);
2385 }
2386 }
2387
2388 if (pcs && n->_idx < pc_limit)
2389 tty->print_cr("%3.3x", pcs[n->_idx]);
2390 else
2391 tty->print_cr("");
2392
2393 assert(cut_short || delay == NULL, "no unconditional delay branch");
2394
2395 } // End of per-block dump
2396 tty->print_cr("");
2397
2398 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2399 }
2400 #endif
2401
2402 //------------------------------Final_Reshape_Counts---------------------------
2403 // This class defines counters to help identify when a method
2404 // may/must be executed using hardware with only 24-bit precision.
2405 struct Final_Reshape_Counts : public StackObj {
2406 int _call_count; // count non-inlined 'common' calls
2407 int _float_count; // count float ops requiring 24-bit precision
2408 int _double_count; // count double ops requiring more precision
2409 int _java_call_count; // count non-inlined 'java' calls
2410 int _inner_loop_count; // count loops which need alignment
2411 VectorSet _visited; // Visitation flags
2412 Node_List _tests; // Set of IfNodes & PCTableNodes
2413
2414 Final_Reshape_Counts() :
2415 _call_count(0), _float_count(0), _double_count(0),
2416 _java_call_count(0), _inner_loop_count(0),
2417 _visited( Thread::current()->resource_area() ) { }
2418
2419 void inc_call_count () { _call_count ++; }
2420 void inc_float_count () { _float_count ++; }
2421 void inc_double_count() { _double_count++; }
2422 void inc_java_call_count() { _java_call_count++; }
2423 void inc_inner_loop_count() { _inner_loop_count++; }
2424
2425 int get_call_count () const { return _call_count ; }
2426 int get_float_count () const { return _float_count ; }
2427 int get_double_count() const { return _double_count; }
2428 int get_java_call_count() const { return _java_call_count; }
2429 int get_inner_loop_count() const { return _inner_loop_count; }
2430 };
2431
2432 #ifdef ASSERT
2433 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2434 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2435 // Make sure the offset goes inside the instance layout.
2436 return k->contains_field_offset(tp->offset());
2437 // Note that OffsetBot and OffsetTop are very negative.
2438 }
2439 #endif
2440
2441 // Eliminate trivially redundant StoreCMs and accumulate their
2442 // precedence edges.
2443 void Compile::eliminate_redundant_card_marks(Node* n) {
2444 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2445 if (n->in(MemNode::Address)->outcnt() > 1) {
2446 // There are multiple users of the same address so it might be
2447 // possible to eliminate some of the StoreCMs
2448 Node* mem = n->in(MemNode::Memory);
2449 Node* adr = n->in(MemNode::Address);
2450 Node* val = n->in(MemNode::ValueIn);
2451 Node* prev = n;
2452 bool done = false;
2453 // Walk the chain of StoreCMs eliminating ones that match. As
2454 // long as it's a chain of single users then the optimization is
2455 // safe. Eliminating partially redundant StoreCMs would require
2456 // cloning copies down the other paths.
2457 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2458 if (adr == mem->in(MemNode::Address) &&
2459 val == mem->in(MemNode::ValueIn)) {
2460 // redundant StoreCM
2461 if (mem->req() > MemNode::OopStore) {
2462 // Hasn't been processed by this code yet.
2463 n->add_prec(mem->in(MemNode::OopStore));
2464 } else {
2465 // Already converted to precedence edge
2466 for (uint i = mem->req(); i < mem->len(); i++) {
2467 // Accumulate any precedence edges
2468 if (mem->in(i) != NULL) {
2469 n->add_prec(mem->in(i));
2470 }
2471 }
2472 // Everything above this point has been processed.
2473 done = true;
2474 }
2475 // Eliminate the previous StoreCM
2476 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2477 assert(mem->outcnt() == 0, "should be dead");
2478 mem->disconnect_inputs(NULL, this);
2479 } else {
2480 prev = mem;
2481 }
2482 mem = prev->in(MemNode::Memory);
2483 }
2484 }
2485 }
2486
2487 //------------------------------final_graph_reshaping_impl----------------------
2488 // Implement items 1-5 from final_graph_reshaping below.
2489 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2490
2491 if ( n->outcnt() == 0 ) return; // dead node
2492 uint nop = n->Opcode();
2493
2494 // Check for 2-input instruction with "last use" on right input.
2495 // Swap to left input. Implements item (2).
2496 if( n->req() == 3 && // two-input instruction
2497 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2498 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2499 n->in(2)->outcnt() == 1 &&// right use IS a last use
2500 !n->in(2)->is_Con() ) { // right use is not a constant
2501 // Check for commutative opcode
2502 switch( nop ) {
2503 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2504 case Op_MaxI: case Op_MinI:
2505 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2506 case Op_AndL: case Op_XorL: case Op_OrL:
2507 case Op_AndI: case Op_XorI: case Op_OrI: {
2508 // Move "last use" input to left by swapping inputs
2509 n->swap_edges(1, 2);
2510 break;
2511 }
2512 default:
2513 break;
2514 }
2515 }
2516
2517 #ifdef ASSERT
2518 if( n->is_Mem() ) {
2519 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2520 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2521 // oop will be recorded in oop map if load crosses safepoint
2522 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2523 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2524 "raw memory operations should have control edge");
2525 }
2526 #endif
2527 // Count FPU ops and common calls, implements item (3)
2528 switch( nop ) {
2529 // Count all float operations that may use FPU
2530 case Op_AddF:
2531 case Op_SubF:
2532 case Op_MulF:
2533 case Op_DivF:
2534 case Op_NegF:
2535 case Op_ModF:
2536 case Op_ConvI2F:
2537 case Op_ConF:
2538 case Op_CmpF:
2539 case Op_CmpF3:
2540 // case Op_ConvL2F: // longs are split into 32-bit halves
2541 frc.inc_float_count();
2542 break;
2543
2544 case Op_ConvF2D:
2545 case Op_ConvD2F:
2546 frc.inc_float_count();
2547 frc.inc_double_count();
2548 break;
2549
2550 // Count all double operations that may use FPU
2551 case Op_AddD:
2552 case Op_SubD:
2553 case Op_MulD:
2554 case Op_DivD:
2555 case Op_NegD:
2556 case Op_ModD:
2557 case Op_ConvI2D:
2558 case Op_ConvD2I:
2559 // case Op_ConvL2D: // handled by leaf call
2560 // case Op_ConvD2L: // handled by leaf call
2561 case Op_ConD:
2562 case Op_CmpD:
2563 case Op_CmpD3:
2564 frc.inc_double_count();
2565 break;
2566 case Op_Opaque1: // Remove Opaque Nodes before matching
2567 case Op_Opaque2: // Remove Opaque Nodes before matching
2568 n->subsume_by(n->in(1), this);
2569 break;
2570 case Op_CallStaticJava:
2571 case Op_CallJava:
2572 case Op_CallDynamicJava:
2573 frc.inc_java_call_count(); // Count java call site;
2574 case Op_CallRuntime:
2575 case Op_CallLeaf:
2576 case Op_CallLeafNoFP: {
2577 assert( n->is_Call(), "" );
2578 CallNode *call = n->as_Call();
2579 // Count call sites where the FP mode bit would have to be flipped.
2580 // Do not count uncommon runtime calls:
2581 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2582 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2583 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2584 frc.inc_call_count(); // Count the call site
2585 } else { // See if uncommon argument is shared
2586 Node *n = call->in(TypeFunc::Parms);
2587 int nop = n->Opcode();
2588 // Clone shared simple arguments to uncommon calls, item (1).
2589 if( n->outcnt() > 1 &&
2590 !n->is_Proj() &&
2591 nop != Op_CreateEx &&
2592 nop != Op_CheckCastPP &&
2593 nop != Op_DecodeN &&
2594 nop != Op_DecodeNKlass &&
2595 !n->is_Mem() ) {
2596 Node *x = n->clone();
2597 call->set_req( TypeFunc::Parms, x );
2598 }
2599 }
2600 break;
2601 }
2602
2603 case Op_StoreD:
2604 case Op_LoadD:
2605 case Op_LoadD_unaligned:
2606 frc.inc_double_count();
2607 goto handle_mem;
2608 case Op_StoreF:
2609 case Op_LoadF:
2610 frc.inc_float_count();
2611 goto handle_mem;
2612
2613 case Op_StoreCM:
2614 {
2615 // Convert OopStore dependence into precedence edge
2616 Node* prec = n->in(MemNode::OopStore);
2617 n->del_req(MemNode::OopStore);
2618 n->add_prec(prec);
2619 eliminate_redundant_card_marks(n);
2620 }
2621
2622 // fall through
2623
2624 case Op_StoreB:
2625 case Op_StoreC:
2626 case Op_StorePConditional:
2627 case Op_StoreI:
2628 case Op_StoreL:
2629 case Op_StoreIConditional:
2630 case Op_StoreLConditional:
2631 case Op_CompareAndSwapI:
2632 case Op_CompareAndSwapL:
2633 case Op_CompareAndSwapP:
2634 case Op_CompareAndSwapN:
2635 case Op_GetAndAddI:
2636 case Op_GetAndAddL:
2637 case Op_GetAndSetI:
2638 case Op_GetAndSetL:
2639 case Op_GetAndSetP:
2640 case Op_GetAndSetN:
2641 case Op_StoreP:
2642 case Op_StoreN:
2643 case Op_StoreNKlass:
2644 case Op_LoadB:
2645 case Op_LoadUB:
2646 case Op_LoadUS:
2647 case Op_LoadI:
2648 case Op_LoadKlass:
2649 case Op_LoadNKlass:
2650 case Op_LoadL:
2651 case Op_LoadL_unaligned:
2652 case Op_LoadPLocked:
2653 case Op_LoadP:
2654 case Op_LoadN:
2655 case Op_LoadRange:
2656 case Op_LoadS: {
2657 handle_mem:
2658 #ifdef ASSERT
2659 if( VerifyOptoOopOffsets ) {
2660 assert( n->is_Mem(), "" );
2661 MemNode *mem = (MemNode*)n;
2662 // Check to see if address types have grounded out somehow.
2663 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2664 assert( !tp || oop_offset_is_sane(tp), "" );
2665 }
2666 #endif
2667 break;
2668 }
2669
2670 case Op_AddP: { // Assert sane base pointers
2671 Node *addp = n->in(AddPNode::Address);
2672 assert( !addp->is_AddP() ||
2673 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2674 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2675 "Base pointers must match" );
2676 #ifdef _LP64
2677 if ((UseCompressedOops || UseCompressedClassPointers) &&
2678 addp->Opcode() == Op_ConP &&
2679 addp == n->in(AddPNode::Base) &&
2680 n->in(AddPNode::Offset)->is_Con()) {
2681 // Use addressing with narrow klass to load with offset on x86.
2682 // On sparc loading 32-bits constant and decoding it have less
2683 // instructions (4) then load 64-bits constant (7).
2684 // Do this transformation here since IGVN will convert ConN back to ConP.
2685 const Type* t = addp->bottom_type();
2686 if (t->isa_oopptr() || t->isa_klassptr()) {
2687 Node* nn = NULL;
2688
2689 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2690
2691 // Look for existing ConN node of the same exact type.
2692 Node* r = root();
2693 uint cnt = r->outcnt();
2694 for (uint i = 0; i < cnt; i++) {
2695 Node* m = r->raw_out(i);
2696 if (m!= NULL && m->Opcode() == op &&
2697 m->bottom_type()->make_ptr() == t) {
2698 nn = m;
2699 break;
2700 }
2701 }
2702 if (nn != NULL) {
2703 // Decode a narrow oop to match address
2704 // [R12 + narrow_oop_reg<<3 + offset]
2705 if (t->isa_oopptr()) {
2706 nn = new (this) DecodeNNode(nn, t);
2707 } else {
2708 nn = new (this) DecodeNKlassNode(nn, t);
2709 }
2710 n->set_req(AddPNode::Base, nn);
2711 n->set_req(AddPNode::Address, nn);
2712 if (addp->outcnt() == 0) {
2713 addp->disconnect_inputs(NULL, this);
2714 }
2715 }
2716 }
2717 }
2718 #endif
2719 break;
2720 }
2721
2722 #ifdef _LP64
2723 case Op_CastPP:
2724 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2725 Node* in1 = n->in(1);
2726 const Type* t = n->bottom_type();
2727 Node* new_in1 = in1->clone();
2728 new_in1->as_DecodeN()->set_type(t);
2729
2730 if (!Matcher::narrow_oop_use_complex_address()) {
2731 //
2732 // x86, ARM and friends can handle 2 adds in addressing mode
2733 // and Matcher can fold a DecodeN node into address by using
2734 // a narrow oop directly and do implicit NULL check in address:
2735 //
2736 // [R12 + narrow_oop_reg<<3 + offset]
2737 // NullCheck narrow_oop_reg
2738 //
2739 // On other platforms (Sparc) we have to keep new DecodeN node and
2740 // use it to do implicit NULL check in address:
2741 //
2742 // decode_not_null narrow_oop_reg, base_reg
2743 // [base_reg + offset]
2744 // NullCheck base_reg
2745 //
2746 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2747 // to keep the information to which NULL check the new DecodeN node
2748 // corresponds to use it as value in implicit_null_check().
2749 //
2750 new_in1->set_req(0, n->in(0));
2751 }
2752
2753 n->subsume_by(new_in1, this);
2754 if (in1->outcnt() == 0) {
2755 in1->disconnect_inputs(NULL, this);
2756 }
2757 }
2758 break;
2759
2760 case Op_CmpP:
2761 // Do this transformation here to preserve CmpPNode::sub() and
2762 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2763 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2764 Node* in1 = n->in(1);
2765 Node* in2 = n->in(2);
2766 if (!in1->is_DecodeNarrowPtr()) {
2767 in2 = in1;
2768 in1 = n->in(2);
2769 }
2770 assert(in1->is_DecodeNarrowPtr(), "sanity");
2771
2772 Node* new_in2 = NULL;
2773 if (in2->is_DecodeNarrowPtr()) {
2774 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2775 new_in2 = in2->in(1);
2776 } else if (in2->Opcode() == Op_ConP) {
2777 const Type* t = in2->bottom_type();
2778 if (t == TypePtr::NULL_PTR) {
2779 assert(in1->is_DecodeN(), "compare klass to null?");
2780 // Don't convert CmpP null check into CmpN if compressed
2781 // oops implicit null check is not generated.
2782 // This will allow to generate normal oop implicit null check.
2783 if (Matcher::gen_narrow_oop_implicit_null_checks())
2784 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2785 //
2786 // This transformation together with CastPP transformation above
2787 // will generated code for implicit NULL checks for compressed oops.
2788 //
2789 // The original code after Optimize()
2790 //
2791 // LoadN memory, narrow_oop_reg
2792 // decode narrow_oop_reg, base_reg
2793 // CmpP base_reg, NULL
2794 // CastPP base_reg // NotNull
2795 // Load [base_reg + offset], val_reg
2796 //
2797 // after these transformations will be
2798 //
2799 // LoadN memory, narrow_oop_reg
2800 // CmpN narrow_oop_reg, NULL
2801 // decode_not_null narrow_oop_reg, base_reg
2802 // Load [base_reg + offset], val_reg
2803 //
2804 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2805 // since narrow oops can be used in debug info now (see the code in
2806 // final_graph_reshaping_walk()).
2807 //
2808 // At the end the code will be matched to
2809 // on x86:
2810 //
2811 // Load_narrow_oop memory, narrow_oop_reg
2812 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2813 // NullCheck narrow_oop_reg
2814 //
2815 // and on sparc:
2816 //
2817 // Load_narrow_oop memory, narrow_oop_reg
2818 // decode_not_null narrow_oop_reg, base_reg
2819 // Load [base_reg + offset], val_reg
2820 // NullCheck base_reg
2821 //
2822 } else if (t->isa_oopptr()) {
2823 new_in2 = ConNode::make(this, t->make_narrowoop());
2824 } else if (t->isa_klassptr()) {
2825 new_in2 = ConNode::make(this, t->make_narrowklass());
2826 }
2827 }
2828 if (new_in2 != NULL) {
2829 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2830 n->subsume_by(cmpN, this);
2831 if (in1->outcnt() == 0) {
2832 in1->disconnect_inputs(NULL, this);
2833 }
2834 if (in2->outcnt() == 0) {
2835 in2->disconnect_inputs(NULL, this);
2836 }
2837 }
2838 }
2839 break;
2840
2841 case Op_DecodeN:
2842 case Op_DecodeNKlass:
2843 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2844 // DecodeN could be pinned when it can't be fold into
2845 // an address expression, see the code for Op_CastPP above.
2846 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2847 break;
2848
2849 case Op_EncodeP:
2850 case Op_EncodePKlass: {
2851 Node* in1 = n->in(1);
2852 if (in1->is_DecodeNarrowPtr()) {
2853 n->subsume_by(in1->in(1), this);
2854 } else if (in1->Opcode() == Op_ConP) {
2855 const Type* t = in1->bottom_type();
2856 if (t == TypePtr::NULL_PTR) {
2857 assert(t->isa_oopptr(), "null klass?");
2858 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2859 } else if (t->isa_oopptr()) {
2860 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2861 } else if (t->isa_klassptr()) {
2862 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2863 }
2864 }
2865 if (in1->outcnt() == 0) {
2866 in1->disconnect_inputs(NULL, this);
2867 }
2868 break;
2869 }
2870
2871 case Op_Proj: {
2872 if (OptimizeStringConcat) {
2873 ProjNode* p = n->as_Proj();
2874 if (p->_is_io_use) {
2875 // Separate projections were used for the exception path which
2876 // are normally removed by a late inline. If it wasn't inlined
2877 // then they will hang around and should just be replaced with
2878 // the original one.
2879 Node* proj = NULL;
2880 // Replace with just one
2881 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2882 Node *use = i.get();
2883 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2884 proj = use;
2885 break;
2886 }
2887 }
2888 assert(proj != NULL, "must be found");
2889 p->subsume_by(proj, this);
2890 }
2891 }
2892 break;
2893 }
2894
2895 case Op_Phi:
2896 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2897 // The EncodeP optimization may create Phi with the same edges
2898 // for all paths. It is not handled well by Register Allocator.
2899 Node* unique_in = n->in(1);
2900 assert(unique_in != NULL, "");
2901 uint cnt = n->req();
2902 for (uint i = 2; i < cnt; i++) {
2903 Node* m = n->in(i);
2904 assert(m != NULL, "");
2905 if (unique_in != m)
2906 unique_in = NULL;
2907 }
2908 if (unique_in != NULL) {
2909 n->subsume_by(unique_in, this);
2910 }
2911 }
2912 break;
2913
2914 #endif
2915
2916 case Op_ModI:
2917 if (UseDivMod) {
2918 // Check if a%b and a/b both exist
2919 Node* d = n->find_similar(Op_DivI);
2920 if (d) {
2921 // Replace them with a fused divmod if supported
2922 if (Matcher::has_match_rule(Op_DivModI)) {
2923 DivModINode* divmod = DivModINode::make(this, n);
2924 d->subsume_by(divmod->div_proj(), this);
2925 n->subsume_by(divmod->mod_proj(), this);
2926 } else {
2927 // replace a%b with a-((a/b)*b)
2928 Node* mult = new (this) MulINode(d, d->in(2));
2929 Node* sub = new (this) SubINode(d->in(1), mult);
2930 n->subsume_by(sub, this);
2931 }
2932 }
2933 }
2934 break;
2935
2936 case Op_ModL:
2937 if (UseDivMod) {
2938 // Check if a%b and a/b both exist
2939 Node* d = n->find_similar(Op_DivL);
2940 if (d) {
2941 // Replace them with a fused divmod if supported
2942 if (Matcher::has_match_rule(Op_DivModL)) {
2943 DivModLNode* divmod = DivModLNode::make(this, n);
2944 d->subsume_by(divmod->div_proj(), this);
2945 n->subsume_by(divmod->mod_proj(), this);
2946 } else {
2947 // replace a%b with a-((a/b)*b)
2948 Node* mult = new (this) MulLNode(d, d->in(2));
2949 Node* sub = new (this) SubLNode(d->in(1), mult);
2950 n->subsume_by(sub, this);
2951 }
2952 }
2953 }
2954 break;
2955
2956 case Op_LoadVector:
2957 case Op_StoreVector:
2958 break;
2959
2960 case Op_PackB:
2961 case Op_PackS:
2962 case Op_PackI:
2963 case Op_PackF:
2964 case Op_PackL:
2965 case Op_PackD:
2966 if (n->req()-1 > 2) {
2967 // Replace many operand PackNodes with a binary tree for matching
2968 PackNode* p = (PackNode*) n;
2969 Node* btp = p->binary_tree_pack(this, 1, n->req());
2970 n->subsume_by(btp, this);
2971 }
2972 break;
2973 case Op_Loop:
2974 case Op_CountedLoop:
2975 if (n->as_Loop()->is_inner_loop()) {
2976 frc.inc_inner_loop_count();
2977 }
2978 break;
2979 case Op_LShiftI:
2980 case Op_RShiftI:
2981 case Op_URShiftI:
2982 case Op_LShiftL:
2983 case Op_RShiftL:
2984 case Op_URShiftL:
2985 if (Matcher::need_masked_shift_count) {
2986 // The cpu's shift instructions don't restrict the count to the
2987 // lower 5/6 bits. We need to do the masking ourselves.
2988 Node* in2 = n->in(2);
2989 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2990 const TypeInt* t = in2->find_int_type();
2991 if (t != NULL && t->is_con()) {
2992 juint shift = t->get_con();
2993 if (shift > mask) { // Unsigned cmp
2994 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2995 }
2996 } else {
2997 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2998 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2999 n->set_req(2, shift);
3000 }
3001 }
3002 if (in2->outcnt() == 0) { // Remove dead node
3003 in2->disconnect_inputs(NULL, this);
3004 }
3005 }
3006 break;
3007 case Op_MemBarStoreStore:
3008 case Op_MemBarRelease:
3009 // Break the link with AllocateNode: it is no longer useful and
3010 // confuses register allocation.
3011 if (n->req() > MemBarNode::Precedent) {
3012 n->set_req(MemBarNode::Precedent, top());
3013 }
3014 break;
3015 default:
3016 assert( !n->is_Call(), "" );
3017 assert( !n->is_Mem(), "" );
3018 break;
3019 }
3020
3021 // Collect CFG split points
3022 if (n->is_MultiBranch())
3023 frc._tests.push(n);
3024 }
3025
3026 //------------------------------final_graph_reshaping_walk---------------------
3027 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3028 // requires that the walk visits a node's inputs before visiting the node.
3029 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3030 ResourceArea *area = Thread::current()->resource_area();
3031 Unique_Node_List sfpt(area);
3032
3033 frc._visited.set(root->_idx); // first, mark node as visited
3034 uint cnt = root->req();
3035 Node *n = root;
3036 uint i = 0;
3037 while (true) {
3038 if (i < cnt) {
3039 // Place all non-visited non-null inputs onto stack
3040 Node* m = n->in(i);
3041 ++i;
3042 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3043 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3044 sfpt.push(m);
3045 cnt = m->req();
3046 nstack.push(n, i); // put on stack parent and next input's index
3047 n = m;
3048 i = 0;
3049 }
3050 } else {
3051 // Now do post-visit work
3052 final_graph_reshaping_impl( n, frc );
3053 if (nstack.is_empty())
3054 break; // finished
3055 n = nstack.node(); // Get node from stack
3056 cnt = n->req();
3057 i = nstack.index();
3058 nstack.pop(); // Shift to the next node on stack
3059 }
3060 }
3061
3062 // Skip next transformation if compressed oops are not used.
3063 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3064 (!UseCompressedOops && !UseCompressedClassPointers))
3065 return;
3066
3067 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3068 // It could be done for an uncommon traps or any safepoints/calls
3069 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3070 while (sfpt.size() > 0) {
3071 n = sfpt.pop();
3072 JVMState *jvms = n->as_SafePoint()->jvms();
3073 assert(jvms != NULL, "sanity");
3074 int start = jvms->debug_start();
3075 int end = n->req();
3076 bool is_uncommon = (n->is_CallStaticJava() &&
3077 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3078 for (int j = start; j < end; j++) {
3079 Node* in = n->in(j);
3080 if (in->is_DecodeNarrowPtr()) {
3081 bool safe_to_skip = true;
3082 if (!is_uncommon ) {
3083 // Is it safe to skip?
3084 for (uint i = 0; i < in->outcnt(); i++) {
3085 Node* u = in->raw_out(i);
3086 if (!u->is_SafePoint() ||
3087 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3088 safe_to_skip = false;
3089 }
3090 }
3091 }
3092 if (safe_to_skip) {
3093 n->set_req(j, in->in(1));
3094 }
3095 if (in->outcnt() == 0) {
3096 in->disconnect_inputs(NULL, this);
3097 }
3098 }
3099 }
3100 }
3101 }
3102
3103 //------------------------------final_graph_reshaping--------------------------
3104 // Final Graph Reshaping.
3105 //
3106 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3107 // and not commoned up and forced early. Must come after regular
3108 // optimizations to avoid GVN undoing the cloning. Clone constant
3109 // inputs to Loop Phis; these will be split by the allocator anyways.
3110 // Remove Opaque nodes.
3111 // (2) Move last-uses by commutative operations to the left input to encourage
3112 // Intel update-in-place two-address operations and better register usage
3113 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3114 // calls canonicalizing them back.
3115 // (3) Count the number of double-precision FP ops, single-precision FP ops
3116 // and call sites. On Intel, we can get correct rounding either by
3117 // forcing singles to memory (requires extra stores and loads after each
3118 // FP bytecode) or we can set a rounding mode bit (requires setting and
3119 // clearing the mode bit around call sites). The mode bit is only used
3120 // if the relative frequency of single FP ops to calls is low enough.
3121 // This is a key transform for SPEC mpeg_audio.
3122 // (4) Detect infinite loops; blobs of code reachable from above but not
3123 // below. Several of the Code_Gen algorithms fail on such code shapes,
3124 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3125 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3126 // Detection is by looking for IfNodes where only 1 projection is
3127 // reachable from below or CatchNodes missing some targets.
3128 // (5) Assert for insane oop offsets in debug mode.
3129
3130 bool Compile::final_graph_reshaping() {
3131 // an infinite loop may have been eliminated by the optimizer,
3132 // in which case the graph will be empty.
3133 if (root()->req() == 1) {
3134 record_method_not_compilable("trivial infinite loop");
3135 return true;
3136 }
3137
3138 // Expensive nodes have their control input set to prevent the GVN
3139 // from freely commoning them. There's no GVN beyond this point so
3140 // no need to keep the control input. We want the expensive nodes to
3141 // be freely moved to the least frequent code path by gcm.
3142 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3143 for (int i = 0; i < expensive_count(); i++) {
3144 _expensive_nodes->at(i)->set_req(0, NULL);
3145 }
3146
3147 Final_Reshape_Counts frc;
3148
3149 // Visit everybody reachable!
3150 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3151 Node_Stack nstack(unique() >> 1);
3152 final_graph_reshaping_walk(nstack, root(), frc);
3153
3154 // Check for unreachable (from below) code (i.e., infinite loops).
3155 for( uint i = 0; i < frc._tests.size(); i++ ) {
3156 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3157 // Get number of CFG targets.
3158 // Note that PCTables include exception targets after calls.
3159 uint required_outcnt = n->required_outcnt();
3160 if (n->outcnt() != required_outcnt) {
3161 // Check for a few special cases. Rethrow Nodes never take the
3162 // 'fall-thru' path, so expected kids is 1 less.
3163 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3164 if (n->in(0)->in(0)->is_Call()) {
3165 CallNode *call = n->in(0)->in(0)->as_Call();
3166 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3167 required_outcnt--; // Rethrow always has 1 less kid
3168 } else if (call->req() > TypeFunc::Parms &&
3169 call->is_CallDynamicJava()) {
3170 // Check for null receiver. In such case, the optimizer has
3171 // detected that the virtual call will always result in a null
3172 // pointer exception. The fall-through projection of this CatchNode
3173 // will not be populated.
3174 Node *arg0 = call->in(TypeFunc::Parms);
3175 if (arg0->is_Type() &&
3176 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3177 required_outcnt--;
3178 }
3179 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3180 call->req() > TypeFunc::Parms+1 &&
3181 call->is_CallStaticJava()) {
3182 // Check for negative array length. In such case, the optimizer has
3183 // detected that the allocation attempt will always result in an
3184 // exception. There is no fall-through projection of this CatchNode .
3185 Node *arg1 = call->in(TypeFunc::Parms+1);
3186 if (arg1->is_Type() &&
3187 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3188 required_outcnt--;
3189 }
3190 }
3191 }
3192 }
3193 // Recheck with a better notion of 'required_outcnt'
3194 if (n->outcnt() != required_outcnt) {
3195 record_method_not_compilable("malformed control flow");
3196 return true; // Not all targets reachable!
3197 }
3198 }
3199 // Check that I actually visited all kids. Unreached kids
3200 // must be infinite loops.
3201 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3202 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3203 record_method_not_compilable("infinite loop");
3204 return true; // Found unvisited kid; must be unreach
3205 }
3206 }
3207
3208 // If original bytecodes contained a mixture of floats and doubles
3209 // check if the optimizer has made it homogenous, item (3).
3210 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3211 frc.get_float_count() > 32 &&
3212 frc.get_double_count() == 0 &&
3213 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3214 set_24_bit_selection_and_mode( false, true );
3215 }
3216
3217 set_java_calls(frc.get_java_call_count());
3218 set_inner_loops(frc.get_inner_loop_count());
3219
3220 // No infinite loops, no reason to bail out.
3221 return false;
3222 }
3223
3224 //-----------------------------too_many_traps----------------------------------
3225 // Report if there are too many traps at the current method and bci.
3226 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3227 bool Compile::too_many_traps(ciMethod* method,
3228 int bci,
3229 Deoptimization::DeoptReason reason) {
3230 ciMethodData* md = method->method_data();
3231 if (md->is_empty()) {
3232 // Assume the trap has not occurred, or that it occurred only
3233 // because of a transient condition during start-up in the interpreter.
3234 return false;
3235 }
3236 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3237 if (md->has_trap_at(bci, m, reason) != 0) {
3238 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3239 // Also, if there are multiple reasons, or if there is no per-BCI record,
3240 // assume the worst.
3241 if (log())
3242 log()->elem("observe trap='%s' count='%d'",
3243 Deoptimization::trap_reason_name(reason),
3244 md->trap_count(reason));
3245 return true;
3246 } else {
3247 // Ignore method/bci and see if there have been too many globally.
3248 return too_many_traps(reason, md);
3249 }
3250 }
3251
3252 // Less-accurate variant which does not require a method and bci.
3253 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3254 ciMethodData* logmd) {
3255 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3256 // Too many traps globally.
3257 // Note that we use cumulative trap_count, not just md->trap_count.
3258 if (log()) {
3259 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3260 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3261 Deoptimization::trap_reason_name(reason),
3262 mcount, trap_count(reason));
3263 }
3264 return true;
3265 } else {
3266 // The coast is clear.
3267 return false;
3268 }
3269 }
3270
3271 //--------------------------too_many_recompiles--------------------------------
3272 // Report if there are too many recompiles at the current method and bci.
3273 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3274 // Is not eager to return true, since this will cause the compiler to use
3275 // Action_none for a trap point, to avoid too many recompilations.
3276 bool Compile::too_many_recompiles(ciMethod* method,
3277 int bci,
3278 Deoptimization::DeoptReason reason) {
3279 ciMethodData* md = method->method_data();
3280 if (md->is_empty()) {
3281 // Assume the trap has not occurred, or that it occurred only
3282 // because of a transient condition during start-up in the interpreter.
3283 return false;
3284 }
3285 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3286 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3287 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3288 Deoptimization::DeoptReason per_bc_reason
3289 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3290 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3291 if ((per_bc_reason == Deoptimization::Reason_none
3292 || md->has_trap_at(bci, m, reason) != 0)
3293 // The trap frequency measure we care about is the recompile count:
3294 && md->trap_recompiled_at(bci, m)
3295 && md->overflow_recompile_count() >= bc_cutoff) {
3296 // Do not emit a trap here if it has already caused recompilations.
3297 // Also, if there are multiple reasons, or if there is no per-BCI record,
3298 // assume the worst.
3299 if (log())
3300 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3301 Deoptimization::trap_reason_name(reason),
3302 md->trap_count(reason),
3303 md->overflow_recompile_count());
3304 return true;
3305 } else if (trap_count(reason) != 0
3306 && decompile_count() >= m_cutoff) {
3307 // Too many recompiles globally, and we have seen this sort of trap.
3308 // Use cumulative decompile_count, not just md->decompile_count.
3309 if (log())
3310 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3311 Deoptimization::trap_reason_name(reason),
3312 md->trap_count(reason), trap_count(reason),
3313 md->decompile_count(), decompile_count());
3314 return true;
3315 } else {
3316 // The coast is clear.
3317 return false;
3318 }
3319 }
3320
3321
3322 #ifndef PRODUCT
3323 //------------------------------verify_graph_edges---------------------------
3324 // Walk the Graph and verify that there is a one-to-one correspondence
3325 // between Use-Def edges and Def-Use edges in the graph.
3326 void Compile::verify_graph_edges(bool no_dead_code) {
3327 if (VerifyGraphEdges) {
3328 ResourceArea *area = Thread::current()->resource_area();
3329 Unique_Node_List visited(area);
3330 // Call recursive graph walk to check edges
3331 _root->verify_edges(visited);
3332 if (no_dead_code) {
3333 // Now make sure that no visited node is used by an unvisited node.
3334 bool dead_nodes = 0;
3335 Unique_Node_List checked(area);
3336 while (visited.size() > 0) {
3337 Node* n = visited.pop();
3338 checked.push(n);
3339 for (uint i = 0; i < n->outcnt(); i++) {
3340 Node* use = n->raw_out(i);
3341 if (checked.member(use)) continue; // already checked
3342 if (visited.member(use)) continue; // already in the graph
3343 if (use->is_Con()) continue; // a dead ConNode is OK
3344 // At this point, we have found a dead node which is DU-reachable.
3345 if (dead_nodes++ == 0)
3346 tty->print_cr("*** Dead nodes reachable via DU edges:");
3347 use->dump(2);
3348 tty->print_cr("---");
3349 checked.push(use); // No repeats; pretend it is now checked.
3350 }
3351 }
3352 assert(dead_nodes == 0, "using nodes must be reachable from root");
3353 }
3354 }
3355 }
3356
3357 // Verify GC barriers consistency
3358 // Currently supported:
3359 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3360 void Compile::verify_barriers() {
3361 if (UseG1GC) {
3362 // Verify G1 pre-barriers
3363 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3364
3365 ResourceArea *area = Thread::current()->resource_area();
3366 Unique_Node_List visited(area);
3367 Node_List worklist(area);
3368 // We're going to walk control flow backwards starting from the Root
3369 worklist.push(_root);
3370 while (worklist.size() > 0) {
3371 Node* x = worklist.pop();
3372 if (x == NULL || x == top()) continue;
3373 if (visited.member(x)) {
3374 continue;
3375 } else {
3376 visited.push(x);
3377 }
3378
3379 if (x->is_Region()) {
3380 for (uint i = 1; i < x->req(); i++) {
3381 worklist.push(x->in(i));
3382 }
3383 } else {
3384 worklist.push(x->in(0));
3385 // We are looking for the pattern:
3386 // /->ThreadLocal
3387 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3388 // \->ConI(0)
3389 // We want to verify that the If and the LoadB have the same control
3390 // See GraphKit::g1_write_barrier_pre()
3391 if (x->is_If()) {
3392 IfNode *iff = x->as_If();
3393 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3394 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3395 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3396 && cmp->in(1)->is_Load()) {
3397 LoadNode* load = cmp->in(1)->as_Load();
3398 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3399 && load->in(2)->in(3)->is_Con()
3400 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3401
3402 Node* if_ctrl = iff->in(0);
3403 Node* load_ctrl = load->in(0);
3404
3405 if (if_ctrl != load_ctrl) {
3406 // Skip possible CProj->NeverBranch in infinite loops
3407 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3408 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3409 if_ctrl = if_ctrl->in(0)->in(0);
3410 }
3411 }
3412 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3413 }
3414 }
3415 }
3416 }
3417 }
3418 }
3419 }
3420 }
3421
3422 #endif
3423
3424 // The Compile object keeps track of failure reasons separately from the ciEnv.
3425 // This is required because there is not quite a 1-1 relation between the
3426 // ciEnv and its compilation task and the Compile object. Note that one
3427 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3428 // to backtrack and retry without subsuming loads. Other than this backtracking
3429 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3430 // by the logic in C2Compiler.
3431 void Compile::record_failure(const char* reason) {
3432 if (log() != NULL) {
3433 log()->elem("failure reason='%s' phase='compile'", reason);
3434 }
3435 if (_failure_reason == NULL) {
3436 // Record the first failure reason.
3437 _failure_reason = reason;
3438 }
3439
3440 EventCompilerFailure event;
3441 if (event.should_commit()) {
3442 event.set_compileID(Compile::compile_id());
3443 event.set_failure(reason);
3444 event.commit();
3445 }
3446
3447 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3448 C->print_method(PHASE_FAILURE);
3449 }
3450 _root = NULL; // flush the graph, too
3451 }
3452
3453 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3454 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3455 _phase_name(name), _dolog(dolog)
3456 {
3457 if (dolog) {
3458 C = Compile::current();
3459 _log = C->log();
3460 } else {
3461 C = NULL;
3462 _log = NULL;
3463 }
3464 if (_log != NULL) {
3465 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3466 _log->stamp();
3467 _log->end_head();
3468 }
3469 }
3470
3471 Compile::TracePhase::~TracePhase() {
3472
3473 C = Compile::current();
3474 if (_dolog) {
3475 _log = C->log();
3476 } else {
3477 _log = NULL;
3478 }
3479
3480 #ifdef ASSERT
3481 if (PrintIdealNodeCount) {
3482 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3483 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3484 }
3485
3486 if (VerifyIdealNodeCount) {
3487 Compile::current()->print_missing_nodes();
3488 }
3489 #endif
3490
3491 if (_log != NULL) {
3492 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3493 }
3494 }
3495
3496 //=============================================================================
3497 // Two Constant's are equal when the type and the value are equal.
3498 bool Compile::Constant::operator==(const Constant& other) {
3499 if (type() != other.type() ) return false;
3500 if (can_be_reused() != other.can_be_reused()) return false;
3501 // For floating point values we compare the bit pattern.
3502 switch (type()) {
3503 case T_FLOAT: return (_v._value.i == other._v._value.i);
3504 case T_LONG:
3505 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3506 case T_OBJECT:
3507 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3508 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3509 case T_METADATA: return (_v._metadata == other._v._metadata);
3510 default: ShouldNotReachHere();
3511 }
3512 return false;
3513 }
3514
3515 static int type_to_size_in_bytes(BasicType t) {
3516 switch (t) {
3517 case T_LONG: return sizeof(jlong );
3518 case T_FLOAT: return sizeof(jfloat );
3519 case T_DOUBLE: return sizeof(jdouble);
3520 case T_METADATA: return sizeof(Metadata*);
3521 // We use T_VOID as marker for jump-table entries (labels) which
3522 // need an internal word relocation.
3523 case T_VOID:
3524 case T_ADDRESS:
3525 case T_OBJECT: return sizeof(jobject);
3526 }
3527
3528 ShouldNotReachHere();
3529 return -1;
3530 }
3531
3532 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3533 // sort descending
3534 if (a->freq() > b->freq()) return -1;
3535 if (a->freq() < b->freq()) return 1;
3536 return 0;
3537 }
3538
3539 void Compile::ConstantTable::calculate_offsets_and_size() {
3540 // First, sort the array by frequencies.
3541 _constants.sort(qsort_comparator);
3542
3543 #ifdef ASSERT
3544 // Make sure all jump-table entries were sorted to the end of the
3545 // array (they have a negative frequency).
3546 bool found_void = false;
3547 for (int i = 0; i < _constants.length(); i++) {
3548 Constant con = _constants.at(i);
3549 if (con.type() == T_VOID)
3550 found_void = true; // jump-tables
3551 else
3552 assert(!found_void, "wrong sorting");
3553 }
3554 #endif
3555
3556 int offset = 0;
3557 for (int i = 0; i < _constants.length(); i++) {
3558 Constant* con = _constants.adr_at(i);
3559
3560 // Align offset for type.
3561 int typesize = type_to_size_in_bytes(con->type());
3562 offset = align_size_up(offset, typesize);
3563 con->set_offset(offset); // set constant's offset
3564
3565 if (con->type() == T_VOID) {
3566 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3567 offset = offset + typesize * n->outcnt(); // expand jump-table
3568 } else {
3569 offset = offset + typesize;
3570 }
3571 }
3572
3573 // Align size up to the next section start (which is insts; see
3574 // CodeBuffer::align_at_start).
3575 assert(_size == -1, "already set?");
3576 _size = align_size_up(offset, CodeEntryAlignment);
3577 }
3578
3579 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3580 MacroAssembler _masm(&cb);
3581 for (int i = 0; i < _constants.length(); i++) {
3582 Constant con = _constants.at(i);
3583 address constant_addr;
3584 switch (con.type()) {
3585 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3586 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3587 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3588 case T_OBJECT: {
3589 jobject obj = con.get_jobject();
3590 int oop_index = _masm.oop_recorder()->find_index(obj);
3591 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3592 break;
3593 }
3594 case T_ADDRESS: {
3595 address addr = (address) con.get_jobject();
3596 constant_addr = _masm.address_constant(addr);
3597 break;
3598 }
3599 // We use T_VOID as marker for jump-table entries (labels) which
3600 // need an internal word relocation.
3601 case T_VOID: {
3602 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3603 // Fill the jump-table with a dummy word. The real value is
3604 // filled in later in fill_jump_table.
3605 address dummy = (address) n;
3606 constant_addr = _masm.address_constant(dummy);
3607 // Expand jump-table
3608 for (uint i = 1; i < n->outcnt(); i++) {
3609 address temp_addr = _masm.address_constant(dummy + i);
3610 assert(temp_addr, "consts section too small");
3611 }
3612 break;
3613 }
3614 case T_METADATA: {
3615 Metadata* obj = con.get_metadata();
3616 int metadata_index = _masm.oop_recorder()->find_index(obj);
3617 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3618 break;
3619 }
3620 default: ShouldNotReachHere();
3621 }
3622 assert(constant_addr, "consts section too small");
3623 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3624 }
3625 }
3626
3627 int Compile::ConstantTable::find_offset(Constant& con) const {
3628 int idx = _constants.find(con);
3629 assert(idx != -1, "constant must be in constant table");
3630 int offset = _constants.at(idx).offset();
3631 assert(offset != -1, "constant table not emitted yet?");
3632 return offset;
3633 }
3634
3635 void Compile::ConstantTable::add(Constant& con) {
3636 if (con.can_be_reused()) {
3637 int idx = _constants.find(con);
3638 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3639 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3640 return;
3641 }
3642 }
3643 (void) _constants.append(con);
3644 }
3645
3646 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3647 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3648 Constant con(type, value, b->_freq);
3649 add(con);
3650 return con;
3651 }
3652
3653 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3654 Constant con(metadata);
3655 add(con);
3656 return con;
3657 }
3658
3659 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3660 jvalue value;
3661 BasicType type = oper->type()->basic_type();
3662 switch (type) {
3663 case T_LONG: value.j = oper->constantL(); break;
3664 case T_FLOAT: value.f = oper->constantF(); break;
3665 case T_DOUBLE: value.d = oper->constantD(); break;
3666 case T_OBJECT:
3667 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3668 case T_METADATA: return add((Metadata*)oper->constant()); break;
3669 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3670 }
3671 return add(n, type, value);
3672 }
3673
3674 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3675 jvalue value;
3676 // We can use the node pointer here to identify the right jump-table
3677 // as this method is called from Compile::Fill_buffer right before
3678 // the MachNodes are emitted and the jump-table is filled (means the
3679 // MachNode pointers do not change anymore).
3680 value.l = (jobject) n;
3681 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3682 add(con);
3683 return con;
3684 }
3685
3686 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3687 // If called from Compile::scratch_emit_size do nothing.
3688 if (Compile::current()->in_scratch_emit_size()) return;
3689
3690 assert(labels.is_nonempty(), "must be");
3691 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3692
3693 // Since MachConstantNode::constant_offset() also contains
3694 // table_base_offset() we need to subtract the table_base_offset()
3695 // to get the plain offset into the constant table.
3696 int offset = n->constant_offset() - table_base_offset();
3697
3698 MacroAssembler _masm(&cb);
3699 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3700
3701 for (uint i = 0; i < n->outcnt(); i++) {
3702 address* constant_addr = &jump_table_base[i];
3703 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3704 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3705 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3706 }
3707 }
3708
3709 void Compile::dump_inlining() {
3710 if (print_inlining() || print_intrinsics()) {
3711 // Print inlining message for candidates that we couldn't inline
3712 // for lack of space or non constant receiver
3713 for (int i = 0; i < _late_inlines.length(); i++) {
3714 CallGenerator* cg = _late_inlines.at(i);
3715 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3716 }
3717 Unique_Node_List useful;
3718 useful.push(root());
3719 for (uint next = 0; next < useful.size(); ++next) {
3720 Node* n = useful.at(next);
3721 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3722 CallNode* call = n->as_Call();
3723 CallGenerator* cg = call->generator();
3724 cg->print_inlining_late("receiver not constant");
3725 }
3726 uint max = n->len();
3727 for ( uint i = 0; i < max; ++i ) {
3728 Node *m = n->in(i);
3729 if ( m == NULL ) continue;
3730 useful.push(m);
3731 }
3732 }
3733 for (int i = 0; i < _print_inlining_list->length(); i++) {
3734 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3735 }
3736 }
3737 }
3738
3739 // Dump inlining replay data to the stream.
3740 // Don't change thread state and acquire any locks.
3741 void Compile::dump_inline_data(outputStream* out) {
3742 InlineTree* inl_tree = ilt();
3743 if (inl_tree != NULL) {
3744 out->print(" inline %d", inl_tree->count());
3745 inl_tree->dump_replay_data(out);
3746 }
3747 }
3748
3749 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3750 if (n1->Opcode() < n2->Opcode()) return -1;
3751 else if (n1->Opcode() > n2->Opcode()) return 1;
3752
3753 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
3754 for (uint i = 1; i < n1->req(); i++) {
3755 if (n1->in(i) < n2->in(i)) return -1;
3756 else if (n1->in(i) > n2->in(i)) return 1;
3757 }
3758
3759 return 0;
3760 }
3761
3762 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3763 Node* n1 = *n1p;
3764 Node* n2 = *n2p;
3765
3766 return cmp_expensive_nodes(n1, n2);
3767 }
3768
3769 void Compile::sort_expensive_nodes() {
3770 if (!expensive_nodes_sorted()) {
3771 _expensive_nodes->sort(cmp_expensive_nodes);
3772 }
3773 }
3774
3775 bool Compile::expensive_nodes_sorted() const {
3776 for (int i = 1; i < _expensive_nodes->length(); i++) {
3777 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3778 return false;
3779 }
3780 }
3781 return true;
3782 }
3783
3784 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3785 if (_expensive_nodes->length() == 0) {
3786 return false;
3787 }
3788
3789 assert(OptimizeExpensiveOps, "optimization off?");
3790
3791 // Take this opportunity to remove dead nodes from the list
3792 int j = 0;
3793 for (int i = 0; i < _expensive_nodes->length(); i++) {
3794 Node* n = _expensive_nodes->at(i);
3795 if (!n->is_unreachable(igvn)) {
3796 assert(n->is_expensive(), "should be expensive");
3797 _expensive_nodes->at_put(j, n);
3798 j++;
3799 }
3800 }
3801 _expensive_nodes->trunc_to(j);
3802
3803 // Then sort the list so that similar nodes are next to each other
3804 // and check for at least two nodes of identical kind with same data
3805 // inputs.
3806 sort_expensive_nodes();
3807
3808 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3809 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3810 return true;
3811 }
3812 }
3813
3814 return false;
3815 }
3816
3817 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3818 if (_expensive_nodes->length() == 0) {
3819 return;
3820 }
3821
3822 assert(OptimizeExpensiveOps, "optimization off?");
3823
3824 // Sort to bring similar nodes next to each other and clear the
3825 // control input of nodes for which there's only a single copy.
3826 sort_expensive_nodes();
3827
3828 int j = 0;
3829 int identical = 0;
3830 int i = 0;
3831 for (; i < _expensive_nodes->length()-1; i++) {
3832 assert(j <= i, "can't write beyond current index");
3833 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3834 identical++;
3835 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3836 continue;
3837 }
3838 if (identical > 0) {
3839 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3840 identical = 0;
3841 } else {
3842 Node* n = _expensive_nodes->at(i);
3843 igvn.hash_delete(n);
3844 n->set_req(0, NULL);
3845 igvn.hash_insert(n);
3846 }
3847 }
3848 if (identical > 0) {
3849 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3850 } else if (_expensive_nodes->length() >= 1) {
3851 Node* n = _expensive_nodes->at(i);
3852 igvn.hash_delete(n);
3853 n->set_req(0, NULL);
3854 igvn.hash_insert(n);
3855 }
3856 _expensive_nodes->trunc_to(j);
3857 }
3858
3859 void Compile::add_expensive_node(Node * n) {
3860 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3861 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3862 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3863 if (OptimizeExpensiveOps) {
3864 _expensive_nodes->append(n);
3865 } else {
3866 // Clear control input and let IGVN optimize expensive nodes if
3867 // OptimizeExpensiveOps is off.
3868 n->set_req(0, NULL);
3869 }
3870 }
3871
3872 /**
3873 * Remove the speculative part of types and clean up the graph
3874 */
3875 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3876 if (UseTypeSpeculation) {
3877 Unique_Node_List worklist;
3878 worklist.push(root());
3879 int modified = 0;
3880 // Go over all type nodes that carry a speculative type, drop the
3881 // speculative part of the type and enqueue the node for an igvn
3882 // which may optimize it out.
3883 for (uint next = 0; next < worklist.size(); ++next) {
3884 Node *n = worklist.at(next);
3885 if (n->is_Type()) {
3886 TypeNode* tn = n->as_Type();
3887 const Type* t = tn->type();
3888 const Type* t_no_spec = t->remove_speculative();
3889 if (t_no_spec != t) {
3890 bool in_hash = igvn.hash_delete(n);
3891 assert(in_hash, "node should be in igvn hash table");
3892 tn->set_type(t_no_spec);
3893 igvn.hash_insert(n);
3894 igvn._worklist.push(n); // give it a chance to go away
3895 modified++;
3896 }
3897 }
3898 uint max = n->len();
3899 for( uint i = 0; i < max; ++i ) {
3900 Node *m = n->in(i);
3901 if (not_a_node(m)) continue;
3902 worklist.push(m);
3903 }
3904 }
3905 // Drop the speculative part of all types in the igvn's type table
3906 igvn.remove_speculative_types();
3907 if (modified > 0) {
3908 igvn.optimize();
3909 }
3910 #ifdef ASSERT
3911 // Verify that after the IGVN is over no speculative type has resurfaced
3912 worklist.clear();
3913 worklist.push(root());
3914 for (uint next = 0; next < worklist.size(); ++next) {
3915 Node *n = worklist.at(next);
3916 const Type* t = igvn.type(n);
3917 assert(t == t->remove_speculative(), "no more speculative types");
3918 if (n->is_Type()) {
3919 t = n->as_Type()->type();
3920 assert(t == t->remove_speculative(), "no more speculative types");
3921 }
3922 uint max = n->len();
3923 for( uint i = 0; i < max; ++i ) {
3924 Node *m = n->in(i);
3925 if (not_a_node(m)) continue;
3926 worklist.push(m);
3927 }
3928 }
3929 igvn.check_no_speculative_types();
3930 #endif
3931 }
3932 }
3933
3934 // Auxiliary method to support randomized stressing/fuzzing.
3935 //
3936 // This method can be called the arbitrary number of times, with current count
3937 // as the argument. The logic allows selecting a single candidate from the
3938 // running list of candidates as follows:
3939 // int count = 0;
3940 // Cand* selected = null;
3941 // while(cand = cand->next()) {
3942 // if (randomized_select(++count)) {
3943 // selected = cand;
3944 // }
3945 // }
3946 //
3947 // Including count equalizes the chances any candidate is "selected".
3948 // This is useful when we don't have the complete list of candidates to choose
3949 // from uniformly. In this case, we need to adjust the randomicity of the
3950 // selection, or else we will end up biasing the selection towards the latter
3951 // candidates.
3952 //
3953 // Quick back-envelope calculation shows that for the list of n candidates
3954 // the equal probability for the candidate to persist as "best" can be
3955 // achieved by replacing it with "next" k-th candidate with the probability
3956 // of 1/k. It can be easily shown that by the end of the run, the
3957 // probability for any candidate is converged to 1/n, thus giving the
3958 // uniform distribution among all the candidates.
3959 //
3960 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3961 #define RANDOMIZED_DOMAIN_POW 29
3962 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3963 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3964 bool Compile::randomized_select(int count) {
3965 assert(count > 0, "only positive");
3966 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3967 }