1 /*
2 * Copyright (c) 1997, 2011, 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/assembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/exceptionHandlerTable.hpp"
29 #include "code/nmethod.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/block.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/callGenerator.hpp"
36 #include "opto/callnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/chaitin.hpp"
39 #include "opto/compile.hpp"
40 #include "opto/connode.hpp"
41 #include "opto/divnode.hpp"
42 #include "opto/escape.hpp"
43 #include "opto/idealGraphPrinter.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/macro.hpp"
47 #include "opto/matcher.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/node.hpp"
51 #include "opto/opcodes.hpp"
52 #include "opto/output.hpp"
53 #include "opto/parse.hpp"
54 #include "opto/phaseX.hpp"
55 #include "opto/rootnode.hpp"
56 #include "opto/runtime.hpp"
57 #include "opto/stringopts.hpp"
58 #include "opto/type.hpp"
59 #include "opto/vectornode.hpp"
60 #include "runtime/arguments.hpp"
61 #include "runtime/signature.hpp"
62 #include "runtime/stubRoutines.hpp"
63 #include "runtime/timer.hpp"
64 #include "utilities/copy.hpp"
65 #ifdef TARGET_ARCH_MODEL_x86_32
66 # include "adfiles/ad_x86_32.hpp"
67 #endif
68 #ifdef TARGET_ARCH_MODEL_x86_64
69 # include "adfiles/ad_x86_64.hpp"
70 #endif
71 #ifdef TARGET_ARCH_MODEL_sparc
72 # include "adfiles/ad_sparc.hpp"
73 #endif
74 #ifdef TARGET_ARCH_MODEL_zero
75 # include "adfiles/ad_zero.hpp"
76 #endif
77 #ifdef TARGET_ARCH_MODEL_arm
78 # include "adfiles/ad_arm.hpp"
79 #endif
80 #ifdef TARGET_ARCH_MODEL_ppc
81 # include "adfiles/ad_ppc.hpp"
82 #endif
83
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 #ifdef ASSERT
102 for (int i = 1; i < _intrinsics->length(); i++) {
103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 CallGenerator* cg2 = _intrinsics->at(i);
105 assert(cg1->method() != cg2->method()
106 ? cg1->method() < cg2->method()
107 : cg1->is_virtual() < cg2->is_virtual(),
108 "compiler intrinsics list must stay sorted");
109 }
110 #endif
111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 int lo = 0, hi = _intrinsics->length()-1;
113 while (lo <= hi) {
114 int mid = (uint)(hi + lo) / 2;
115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 if (m < mid_m) {
117 hi = mid-1;
118 } else if (m > mid_m) {
119 lo = mid+1;
120 } else {
121 // look at minor sort key
122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 if (is_virtual < mid_virt) {
124 hi = mid-1;
125 } else if (is_virtual > mid_virt) {
126 lo = mid+1;
127 } else {
128 return mid; // exact match
129 }
130 }
131 }
132 return lo; // inexact match
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new GrowableArray<CallGenerator*>(60);
138 }
139 // This code is stolen from ciObjectFactory::insert.
140 // Really, GrowableArray should have methods for
141 // insert_at, remove_at, and binary_search.
142 int len = _intrinsics->length();
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 if (index == len) {
145 _intrinsics->append(cg);
146 } else {
147 #ifdef ASSERT
148 CallGenerator* oldcg = _intrinsics->at(index);
149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 #endif
151 _intrinsics->append(_intrinsics->at(len-1));
152 int pos;
153 for (pos = len-2; pos >= index; pos--) {
154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 }
156 _intrinsics->at_put(index, cg);
157 }
158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 }
160
161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 if (_intrinsics != NULL) {
164 int index = intrinsic_insertion_index(m, is_virtual);
165 if (index < _intrinsics->length()
166 && _intrinsics->at(index)->method() == m
167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 return _intrinsics->at(index);
169 }
170 }
171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 if (cg != NULL) {
176 // Save it for next time:
177 register_intrinsic(cg);
178 return cg;
179 } else {
180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 }
182 }
183 return NULL;
184 }
185
186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 // in library_call.cpp.
188
189
190 #ifndef PRODUCT
191 // statistics gathering...
192
193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
195
196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 int oflags = _intrinsic_hist_flags[id];
199 assert(flags != 0, "what happened?");
200 if (is_virtual) {
201 flags |= _intrinsic_virtual;
202 }
203 bool changed = (flags != oflags);
204 if ((flags & _intrinsic_worked) != 0) {
205 juint count = (_intrinsic_hist_count[id] += 1);
206 if (count == 1) {
207 changed = true; // first time
208 }
209 // increment the overall count also:
210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 }
212 if (changed) {
213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 // Something changed about the intrinsic's virtuality.
215 if ((flags & _intrinsic_virtual) != 0) {
216 // This is the first use of this intrinsic as a virtual call.
217 if (oflags != 0) {
218 // We already saw it as a non-virtual, so note both cases.
219 flags |= _intrinsic_both;
220 }
221 } else if ((oflags & _intrinsic_both) == 0) {
222 // This is the first use of this intrinsic as a non-virtual
223 flags |= _intrinsic_both;
224 }
225 }
226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 }
228 // update the overall flags also:
229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 return changed;
231 }
232
233 static char* format_flags(int flags, char* buf) {
234 buf[0] = 0;
235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 if (buf[0] == 0) strcat(buf, ",");
241 assert(buf[0] == ',', "must be");
242 return &buf[1];
243 }
244
245 void Compile::print_intrinsic_statistics() {
246 char flagsbuf[100];
247 ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 tty->print_cr("Compiler intrinsic usage:");
250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 #define PRINT_STAT_LINE(name, c, f) \
253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 int flags = _intrinsic_hist_flags[id];
257 juint count = _intrinsic_hist_count[id];
258 if ((flags | count) != 0) {
259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 }
261 }
262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 if (xtty != NULL) xtty->tail("statistics");
264 }
265
266 void Compile::print_statistics() {
267 { ttyLocker ttyl;
268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 Parse::print_statistics();
270 PhaseCCP::print_statistics();
271 PhaseRegAlloc::print_statistics();
272 Scheduling::print_statistics();
273 PhasePeephole::print_statistics();
274 PhaseIdealLoop::print_statistics();
275 if (xtty != NULL) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
283
284 // Support for bundling info
285 Bundle* Compile::node_bundling(const Node *n) {
286 assert(valid_bundle_info(n), "oob");
287 return &_node_bundling_base[n->_idx];
288 }
289
290 bool Compile::valid_bundle_info(const Node *n) {
291 return (_node_bundling_limit > n->_idx);
292 }
293
294
295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 Node* use = n->last_out(i);
298 bool is_in_table = initial_gvn()->hash_delete(use);
299 uint uses_found = 0;
300 for (uint j = 0; j < use->len(); j++) {
301 if (use->in(j) == n) {
302 if (j < use->req())
303 use->set_req(j, nn);
304 else
305 use->set_prec(j, nn);
306 uses_found++;
307 }
308 }
309 if (is_in_table) {
310 // reinsert into table
311 initial_gvn()->hash_find_insert(use);
312 }
313 record_for_igvn(use);
314 i -= uses_found; // we deleted 1 or more copies of this edge
315 }
316 }
317
318
319
320
321 // Identify all nodes that are reachable from below, useful.
322 // Use breadth-first pass that records state in a Unique_Node_List,
323 // recursive traversal is slower.
324 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
325 int estimated_worklist_size = unique();
326 useful.map( estimated_worklist_size, NULL ); // preallocate space
327
328 // Initialize worklist
329 if (root() != NULL) { useful.push(root()); }
330 // If 'top' is cached, declare it useful to preserve cached node
331 if( cached_top_node() ) { useful.push(cached_top_node()); }
332
333 // Push all useful nodes onto the list, breadthfirst
334 for( uint next = 0; next < useful.size(); ++next ) {
335 assert( next < unique(), "Unique useful nodes < total nodes");
336 Node *n = useful.at(next);
337 uint max = n->len();
338 for( uint i = 0; i < max; ++i ) {
339 Node *m = n->in(i);
340 if( m == NULL ) continue;
341 useful.push(m);
342 }
343 }
344 }
345
346 // Disconnect all useless nodes by disconnecting those at the boundary.
347 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
348 uint next = 0;
349 while( next < useful.size() ) {
350 Node *n = useful.at(next++);
351 // Use raw traversal of out edges since this code removes out edges
352 int max = n->outcnt();
353 for (int j = 0; j < max; ++j ) {
354 Node* child = n->raw_out(j);
355 if( ! useful.member(child) ) {
356 assert( !child->is_top() || child != top(),
357 "If top is cached in Compile object it is in useful list");
358 // Only need to remove this out-edge to the useless node
359 n->raw_del_out(j);
360 --j;
361 --max;
362 }
363 }
364 if (n->outcnt() == 1 && n->has_special_unique_user()) {
365 record_for_igvn( n->unique_out() );
366 }
367 }
368 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
369 }
370
371 //------------------------------frame_size_in_words-----------------------------
372 // frame_slots in units of words
373 int Compile::frame_size_in_words() const {
374 // shift is 0 in LP32 and 1 in LP64
375 const int shift = (LogBytesPerWord - LogBytesPerInt);
376 int words = _frame_slots >> shift;
377 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
378 return words;
379 }
380
381 // ============================================================================
382 //------------------------------CompileWrapper---------------------------------
383 class CompileWrapper : public StackObj {
384 Compile *const _compile;
385 public:
386 CompileWrapper(Compile* compile);
387
388 ~CompileWrapper();
389 };
390
391 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
392 // the Compile* pointer is stored in the current ciEnv:
393 ciEnv* env = compile->env();
394 assert(env == ciEnv::current(), "must already be a ciEnv active");
395 assert(env->compiler_data() == NULL, "compile already active?");
396 env->set_compiler_data(compile);
397 assert(compile == Compile::current(), "sanity");
398
399 compile->set_type_dict(NULL);
400 compile->set_type_hwm(NULL);
401 compile->set_type_last_size(0);
402 compile->set_last_tf(NULL, NULL);
403 compile->set_indexSet_arena(NULL);
404 compile->set_indexSet_free_block_list(NULL);
405 compile->init_type_arena();
406 Type::Initialize(compile);
407 _compile->set_scratch_buffer_blob(NULL);
408 _compile->begin_method();
409 }
410 CompileWrapper::~CompileWrapper() {
411 _compile->end_method();
412 if (_compile->scratch_buffer_blob() != NULL)
413 BufferBlob::free(_compile->scratch_buffer_blob());
414 _compile->env()->set_compiler_data(NULL);
415 }
416
417
418 //----------------------------print_compile_messages---------------------------
419 void Compile::print_compile_messages() {
420 #ifndef PRODUCT
421 // Check if recompiling
422 if (_subsume_loads == false && PrintOpto) {
423 // Recompiling without allowing machine instructions to subsume loads
424 tty->print_cr("*********************************************************");
425 tty->print_cr("** Bailout: Recompile without subsuming loads **");
426 tty->print_cr("*********************************************************");
427 }
428 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
429 // Recompiling without escape analysis
430 tty->print_cr("*********************************************************");
431 tty->print_cr("** Bailout: Recompile without escape analysis **");
432 tty->print_cr("*********************************************************");
433 }
434 if (env()->break_at_compile()) {
435 // Open the debugger when compiling this method.
436 tty->print("### Breaking when compiling: ");
437 method()->print_short_name();
438 tty->cr();
439 BREAKPOINT;
440 }
441
442 if( PrintOpto ) {
443 if (is_osr_compilation()) {
444 tty->print("[OSR]%3d", _compile_id);
445 } else {
446 tty->print("%3d", _compile_id);
447 }
448 }
449 #endif
450 }
451
452
453 //-----------------------init_scratch_buffer_blob------------------------------
454 // Construct a temporary BufferBlob and cache it for this compile.
455 void Compile::init_scratch_buffer_blob(int const_size) {
456 // If there is already a scratch buffer blob allocated and the
457 // constant section is big enough, use it. Otherwise free the
458 // current and allocate a new one.
459 BufferBlob* blob = scratch_buffer_blob();
460 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
461 // Use the current blob.
462 } else {
463 if (blob != NULL) {
464 BufferBlob::free(blob);
465 }
466
467 ResourceMark rm;
468 _scratch_const_size = const_size;
469 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
470 blob = BufferBlob::create("Compile::scratch_buffer", size);
471 // Record the buffer blob for next time.
472 set_scratch_buffer_blob(blob);
473 // Have we run out of code space?
474 if (scratch_buffer_blob() == NULL) {
475 // Let CompilerBroker disable further compilations.
476 record_failure("Not enough space for scratch buffer in CodeCache");
477 return;
478 }
479 }
480
481 // Initialize the relocation buffers
482 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
483 set_scratch_locs_memory(locs_buf);
484 }
485
486
487 //-----------------------scratch_emit_size-------------------------------------
488 // Helper function that computes size by emitting code
489 uint Compile::scratch_emit_size(const Node* n) {
490 // Start scratch_emit_size section.
491 set_in_scratch_emit_size(true);
492
493 // Emit into a trash buffer and count bytes emitted.
494 // This is a pretty expensive way to compute a size,
495 // but it works well enough if seldom used.
496 // All common fixed-size instructions are given a size
497 // method by the AD file.
498 // Note that the scratch buffer blob and locs memory are
499 // allocated at the beginning of the compile task, and
500 // may be shared by several calls to scratch_emit_size.
501 // The allocation of the scratch buffer blob is particularly
502 // expensive, since it has to grab the code cache lock.
503 BufferBlob* blob = this->scratch_buffer_blob();
504 assert(blob != NULL, "Initialize BufferBlob at start");
505 assert(blob->size() > MAX_inst_size, "sanity");
506 relocInfo* locs_buf = scratch_locs_memory();
507 address blob_begin = blob->content_begin();
508 address blob_end = (address)locs_buf;
509 assert(blob->content_contains(blob_end), "sanity");
510 CodeBuffer buf(blob_begin, blob_end - blob_begin);
511 buf.initialize_consts_size(_scratch_const_size);
512 buf.initialize_stubs_size(MAX_stubs_size);
513 assert(locs_buf != NULL, "sanity");
514 int lsize = MAX_locs_size / 3;
515 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
516 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
517 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
518
519 // Do the emission.
520
521 Label fakeL; // Fake label for branch instructions.
522 bool is_branch = n->is_Branch() && n->as_Mach()->ideal_Opcode() != Op_Jump;
523 if (is_branch) {
524 MacroAssembler masm(&buf);
525 masm.bind(fakeL);
526 n->as_Mach()->label_set(&fakeL, 0);
527 }
528 n->emit(buf, this->regalloc());
529 if (is_branch) // Clear the reference to fake label.
530 n->as_Mach()->label_set(NULL, 0);
531
532 // End scratch_emit_size section.
533 set_in_scratch_emit_size(false);
534
535 return buf.insts_size();
536 }
537
538
539 // ============================================================================
540 //------------------------------Compile standard-------------------------------
541 debug_only( int Compile::_debug_idx = 100000; )
542
543 // Compile a method. entry_bci is -1 for normal compilations and indicates
544 // the continuation bci for on stack replacement.
545
546
547 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
548 : Phase(Compiler),
549 _env(ci_env),
550 _log(ci_env->log()),
551 _compile_id(ci_env->compile_id()),
552 _save_argument_registers(false),
553 _stub_name(NULL),
554 _stub_function(NULL),
555 _stub_entry_point(NULL),
556 _method(target),
557 _entry_bci(osr_bci),
558 _initial_gvn(NULL),
559 _for_igvn(NULL),
560 _warm_calls(NULL),
561 _subsume_loads(subsume_loads),
562 _do_escape_analysis(do_escape_analysis),
563 _failure_reason(NULL),
564 _code_buffer("Compile::Fill_buffer"),
565 _orig_pc_slot(0),
566 _orig_pc_slot_offset_in_bytes(0),
567 _has_method_handle_invokes(false),
568 _mach_constant_base_node(NULL),
569 _node_bundling_limit(0),
570 _node_bundling_base(NULL),
571 _java_calls(0),
572 _inner_loops(0),
573 _scratch_const_size(-1),
574 _in_scratch_emit_size(false),
575 #ifndef PRODUCT
576 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
577 _printer(IdealGraphPrinter::printer()),
578 #endif
579 _congraph(NULL) {
580 C = this;
581
582 CompileWrapper cw(this);
583 #ifndef PRODUCT
584 if (TimeCompiler2) {
585 tty->print(" ");
586 target->holder()->name()->print();
587 tty->print(".");
588 target->print_short_name();
589 tty->print(" ");
590 }
591 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
592 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
593 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
594 if (!print_opto_assembly) {
595 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
596 if (print_assembly && !Disassembler::can_decode()) {
597 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
598 print_opto_assembly = true;
599 }
600 }
601 set_print_assembly(print_opto_assembly);
602 set_parsed_irreducible_loop(false);
603 #endif
604
605 if (ProfileTraps) {
606 // Make sure the method being compiled gets its own MDO,
607 // so we can at least track the decompile_count().
608 method()->ensure_method_data();
609 }
610
611 Init(::AliasLevel);
612
613
614 print_compile_messages();
615
616 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
617 _ilt = InlineTree::build_inline_tree_root();
618 else
619 _ilt = NULL;
620
621 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
622 assert(num_alias_types() >= AliasIdxRaw, "");
623
624 #define MINIMUM_NODE_HASH 1023
625 // Node list that Iterative GVN will start with
626 Unique_Node_List for_igvn(comp_arena());
627 set_for_igvn(&for_igvn);
628
629 // GVN that will be run immediately on new nodes
630 uint estimated_size = method()->code_size()*4+64;
631 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
632 PhaseGVN gvn(node_arena(), estimated_size);
633 set_initial_gvn(&gvn);
634
635 { // Scope for timing the parser
636 TracePhase t3("parse", &_t_parser, true);
637
638 // Put top into the hash table ASAP.
639 initial_gvn()->transform_no_reclaim(top());
640
641 // Set up tf(), start(), and find a CallGenerator.
642 CallGenerator* cg = NULL;
643 if (is_osr_compilation()) {
644 const TypeTuple *domain = StartOSRNode::osr_domain();
645 const TypeTuple *range = TypeTuple::make_range(method()->signature());
646 init_tf(TypeFunc::make(domain, range));
647 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
648 initial_gvn()->set_type_bottom(s);
649 init_start(s);
650 cg = CallGenerator::for_osr(method(), entry_bci());
651 } else {
652 // Normal case.
653 init_tf(TypeFunc::make(method()));
654 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
655 initial_gvn()->set_type_bottom(s);
656 init_start(s);
657 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
658 // With java.lang.ref.reference.get() we must go through the
659 // intrinsic when G1 is enabled - even when get() is the root
660 // method of the compile - so that, if necessary, the value in
661 // the referent field of the reference object gets recorded by
662 // the pre-barrier code.
663 // Specifically, if G1 is enabled, the value in the referent
664 // field is recorded by the G1 SATB pre barrier. This will
665 // result in the referent being marked live and the reference
666 // object removed from the list of discovered references during
667 // reference processing.
668 cg = find_intrinsic(method(), false);
669 }
670 if (cg == NULL) {
671 float past_uses = method()->interpreter_invocation_count();
672 float expected_uses = past_uses;
673 cg = CallGenerator::for_inline(method(), expected_uses);
674 }
675 }
676 if (failing()) return;
677 if (cg == NULL) {
678 record_method_not_compilable_all_tiers("cannot parse method");
679 return;
680 }
681 JVMState* jvms = build_start_state(start(), tf());
682 if ((jvms = cg->generate(jvms)) == NULL) {
683 record_method_not_compilable("method parse failed");
684 return;
685 }
686 GraphKit kit(jvms);
687
688 if (!kit.stopped()) {
689 // Accept return values, and transfer control we know not where.
690 // This is done by a special, unique ReturnNode bound to root.
691 return_values(kit.jvms());
692 }
693
694 if (kit.has_exceptions()) {
695 // Any exceptions that escape from this call must be rethrown
696 // to whatever caller is dynamically above us on the stack.
697 // This is done by a special, unique RethrowNode bound to root.
698 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
699 }
700
701 if (!failing() && has_stringbuilder()) {
702 {
703 // remove useless nodes to make the usage analysis simpler
704 ResourceMark rm;
705 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
706 }
707
708 {
709 ResourceMark rm;
710 print_method("Before StringOpts", 3);
711 PhaseStringOpts pso(initial_gvn(), &for_igvn);
712 print_method("After StringOpts", 3);
713 }
714
715 // now inline anything that we skipped the first time around
716 while (_late_inlines.length() > 0) {
717 CallGenerator* cg = _late_inlines.pop();
718 cg->do_late_inline();
719 }
720 }
721 assert(_late_inlines.length() == 0, "should have been processed");
722
723 print_method("Before RemoveUseless", 3);
724
725 // Remove clutter produced by parsing.
726 if (!failing()) {
727 ResourceMark rm;
728 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
729 }
730 }
731
732 // Note: Large methods are capped off in do_one_bytecode().
733 if (failing()) return;
734
735 // After parsing, node notes are no longer automagic.
736 // They must be propagated by register_new_node_with_optimizer(),
737 // clone(), or the like.
738 set_default_node_notes(NULL);
739
740 for (;;) {
741 int successes = Inline_Warm();
742 if (failing()) return;
743 if (successes == 0) break;
744 }
745
746 // Drain the list.
747 Finish_Warm();
748 #ifndef PRODUCT
749 if (_printer) {
750 _printer->print_inlining(this);
751 }
752 #endif
753
754 if (failing()) return;
755 NOT_PRODUCT( verify_graph_edges(); )
756
757 // Now optimize
758 Optimize();
759 if (failing()) return;
760 NOT_PRODUCT( verify_graph_edges(); )
761
762 #ifndef PRODUCT
763 if (PrintIdeal) {
764 ttyLocker ttyl; // keep the following output all in one block
765 // This output goes directly to the tty, not the compiler log.
766 // To enable tools to match it up with the compilation activity,
767 // be sure to tag this tty output with the compile ID.
768 if (xtty != NULL) {
769 xtty->head("ideal compile_id='%d'%s", compile_id(),
770 is_osr_compilation() ? " compile_kind='osr'" :
771 "");
772 }
773 root()->dump(9999);
774 if (xtty != NULL) {
775 xtty->tail("ideal");
776 }
777 }
778 #endif
779
780 // Now that we know the size of all the monitors we can add a fixed slot
781 // for the original deopt pc.
782
783 _orig_pc_slot = fixed_slots();
784 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
785 set_fixed_slots(next_slot);
786
787 // Now generate code
788 Code_Gen();
789 if (failing()) return;
790
791 // Check if we want to skip execution of all compiled code.
792 {
793 #ifndef PRODUCT
794 if (OptoNoExecute) {
795 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
796 return;
797 }
798 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
799 #endif
800
801 if (is_osr_compilation()) {
802 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
803 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
804 } else {
805 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
806 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
807 }
808
809 env()->register_method(_method, _entry_bci,
810 &_code_offsets,
811 _orig_pc_slot_offset_in_bytes,
812 code_buffer(),
813 frame_size_in_words(), _oop_map_set,
814 &_handler_table, &_inc_table,
815 compiler,
816 env()->comp_level(),
817 true, /*has_debug_info*/
818 has_unsafe_access()
819 );
820 }
821 }
822
823 //------------------------------Compile----------------------------------------
824 // Compile a runtime stub
825 Compile::Compile( ciEnv* ci_env,
826 TypeFunc_generator generator,
827 address stub_function,
828 const char *stub_name,
829 int is_fancy_jump,
830 bool pass_tls,
831 bool save_arg_registers,
832 bool return_pc )
833 : Phase(Compiler),
834 _env(ci_env),
835 _log(ci_env->log()),
836 _compile_id(-1),
837 _save_argument_registers(save_arg_registers),
838 _method(NULL),
839 _stub_name(stub_name),
840 _stub_function(stub_function),
841 _stub_entry_point(NULL),
842 _entry_bci(InvocationEntryBci),
843 _initial_gvn(NULL),
844 _for_igvn(NULL),
845 _warm_calls(NULL),
846 _orig_pc_slot(0),
847 _orig_pc_slot_offset_in_bytes(0),
848 _subsume_loads(true),
849 _do_escape_analysis(false),
850 _failure_reason(NULL),
851 _code_buffer("Compile::Fill_buffer"),
852 _has_method_handle_invokes(false),
853 _mach_constant_base_node(NULL),
854 _node_bundling_limit(0),
855 _node_bundling_base(NULL),
856 _java_calls(0),
857 _inner_loops(0),
858 #ifndef PRODUCT
859 _trace_opto_output(TraceOptoOutput),
860 _printer(NULL),
861 #endif
862 _congraph(NULL) {
863 C = this;
864
865 #ifndef PRODUCT
866 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
867 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
868 set_print_assembly(PrintFrameConverterAssembly);
869 set_parsed_irreducible_loop(false);
870 #endif
871 CompileWrapper cw(this);
872 Init(/*AliasLevel=*/ 0);
873 init_tf((*generator)());
874
875 {
876 // The following is a dummy for the sake of GraphKit::gen_stub
877 Unique_Node_List for_igvn(comp_arena());
878 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
879 PhaseGVN gvn(Thread::current()->resource_area(),255);
880 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
881 gvn.transform_no_reclaim(top());
882
883 GraphKit kit;
884 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
885 }
886
887 NOT_PRODUCT( verify_graph_edges(); )
888 Code_Gen();
889 if (failing()) return;
890
891
892 // Entry point will be accessed using compile->stub_entry_point();
893 if (code_buffer() == NULL) {
894 Matcher::soft_match_failure();
895 } else {
896 if (PrintAssembly && (WizardMode || Verbose))
897 tty->print_cr("### Stub::%s", stub_name);
898
899 if (!failing()) {
900 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
901
902 // Make the NMethod
903 // For now we mark the frame as never safe for profile stackwalking
904 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
905 code_buffer(),
906 CodeOffsets::frame_never_safe,
907 // _code_offsets.value(CodeOffsets::Frame_Complete),
908 frame_size_in_words(),
909 _oop_map_set,
910 save_arg_registers);
911 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
912
913 _stub_entry_point = rs->entry_point();
914 }
915 }
916 }
917
918 #ifndef PRODUCT
919 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
920 if(PrintOpto && Verbose) {
921 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
922 }
923 }
924 #endif
925
926 void Compile::print_codes() {
927 }
928
929 //------------------------------Init-------------------------------------------
930 // Prepare for a single compilation
931 void Compile::Init(int aliaslevel) {
932 _unique = 0;
933 _regalloc = NULL;
934
935 _tf = NULL; // filled in later
936 _top = NULL; // cached later
937 _matcher = NULL; // filled in later
938 _cfg = NULL; // filled in later
939
940 set_24_bit_selection_and_mode(Use24BitFP, false);
941
942 _node_note_array = NULL;
943 _default_node_notes = NULL;
944
945 _immutable_memory = NULL; // filled in at first inquiry
946
947 // Globally visible Nodes
948 // First set TOP to NULL to give safe behavior during creation of RootNode
949 set_cached_top_node(NULL);
950 set_root(new (this, 3) RootNode());
951 // Now that you have a Root to point to, create the real TOP
952 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
953 set_recent_alloc(NULL, NULL);
954
955 // Create Debug Information Recorder to record scopes, oopmaps, etc.
956 env()->set_oop_recorder(new OopRecorder(comp_arena()));
957 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
958 env()->set_dependencies(new Dependencies(env()));
959
960 _fixed_slots = 0;
961 set_has_split_ifs(false);
962 set_has_loops(has_method() && method()->has_loops()); // first approximation
963 set_has_stringbuilder(false);
964 _trap_can_recompile = false; // no traps emitted yet
965 _major_progress = true; // start out assuming good things will happen
966 set_has_unsafe_access(false);
967 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
968 set_decompile_count(0);
969
970 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
971 set_num_loop_opts(LoopOptsCount);
972 set_do_inlining(Inline);
973 set_max_inline_size(MaxInlineSize);
974 set_freq_inline_size(FreqInlineSize);
975 set_do_scheduling(OptoScheduling);
976 set_do_count_invocations(false);
977 set_do_method_data_update(false);
978
979 if (debug_info()->recording_non_safepoints()) {
980 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
981 (comp_arena(), 8, 0, NULL));
982 set_default_node_notes(Node_Notes::make(this));
983 }
984
985 // // -- Initialize types before each compile --
986 // // Update cached type information
987 // if( _method && _method->constants() )
988 // Type::update_loaded_types(_method, _method->constants());
989
990 // Init alias_type map.
991 if (!_do_escape_analysis && aliaslevel == 3)
992 aliaslevel = 2; // No unique types without escape analysis
993 _AliasLevel = aliaslevel;
994 const int grow_ats = 16;
995 _max_alias_types = grow_ats;
996 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
997 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
998 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
999 {
1000 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1001 }
1002 // Initialize the first few types.
1003 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1004 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1005 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1006 _num_alias_types = AliasIdxRaw+1;
1007 // Zero out the alias type cache.
1008 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1009 // A NULL adr_type hits in the cache right away. Preload the right answer.
1010 probe_alias_cache(NULL)->_index = AliasIdxTop;
1011
1012 _intrinsics = NULL;
1013 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1014 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1015 register_library_intrinsics();
1016 }
1017
1018 //---------------------------init_start----------------------------------------
1019 // Install the StartNode on this compile object.
1020 void Compile::init_start(StartNode* s) {
1021 if (failing())
1022 return; // already failing
1023 assert(s == start(), "");
1024 }
1025
1026 StartNode* Compile::start() const {
1027 assert(!failing(), "");
1028 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1029 Node* start = root()->fast_out(i);
1030 if( start->is_Start() )
1031 return start->as_Start();
1032 }
1033 ShouldNotReachHere();
1034 return NULL;
1035 }
1036
1037 //-------------------------------immutable_memory-------------------------------------
1038 // Access immutable memory
1039 Node* Compile::immutable_memory() {
1040 if (_immutable_memory != NULL) {
1041 return _immutable_memory;
1042 }
1043 StartNode* s = start();
1044 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1045 Node *p = s->fast_out(i);
1046 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1047 _immutable_memory = p;
1048 return _immutable_memory;
1049 }
1050 }
1051 ShouldNotReachHere();
1052 return NULL;
1053 }
1054
1055 //----------------------set_cached_top_node------------------------------------
1056 // Install the cached top node, and make sure Node::is_top works correctly.
1057 void Compile::set_cached_top_node(Node* tn) {
1058 if (tn != NULL) verify_top(tn);
1059 Node* old_top = _top;
1060 _top = tn;
1061 // Calling Node::setup_is_top allows the nodes the chance to adjust
1062 // their _out arrays.
1063 if (_top != NULL) _top->setup_is_top();
1064 if (old_top != NULL) old_top->setup_is_top();
1065 assert(_top == NULL || top()->is_top(), "");
1066 }
1067
1068 #ifndef PRODUCT
1069 void Compile::verify_top(Node* tn) const {
1070 if (tn != NULL) {
1071 assert(tn->is_Con(), "top node must be a constant");
1072 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1073 assert(tn->in(0) != NULL, "must have live top node");
1074 }
1075 }
1076 #endif
1077
1078
1079 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1080
1081 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1082 guarantee(arr != NULL, "");
1083 int num_blocks = arr->length();
1084 if (grow_by < num_blocks) grow_by = num_blocks;
1085 int num_notes = grow_by * _node_notes_block_size;
1086 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1087 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1088 while (num_notes > 0) {
1089 arr->append(notes);
1090 notes += _node_notes_block_size;
1091 num_notes -= _node_notes_block_size;
1092 }
1093 assert(num_notes == 0, "exact multiple, please");
1094 }
1095
1096 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1097 if (source == NULL || dest == NULL) return false;
1098
1099 if (dest->is_Con())
1100 return false; // Do not push debug info onto constants.
1101
1102 #ifdef ASSERT
1103 // Leave a bread crumb trail pointing to the original node:
1104 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1105 dest->set_debug_orig(source);
1106 }
1107 #endif
1108
1109 if (node_note_array() == NULL)
1110 return false; // Not collecting any notes now.
1111
1112 // This is a copy onto a pre-existing node, which may already have notes.
1113 // If both nodes have notes, do not overwrite any pre-existing notes.
1114 Node_Notes* source_notes = node_notes_at(source->_idx);
1115 if (source_notes == NULL || source_notes->is_clear()) return false;
1116 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1117 if (dest_notes == NULL || dest_notes->is_clear()) {
1118 return set_node_notes_at(dest->_idx, source_notes);
1119 }
1120
1121 Node_Notes merged_notes = (*source_notes);
1122 // The order of operations here ensures that dest notes will win...
1123 merged_notes.update_from(dest_notes);
1124 return set_node_notes_at(dest->_idx, &merged_notes);
1125 }
1126
1127
1128 //--------------------------allow_range_check_smearing-------------------------
1129 // Gating condition for coalescing similar range checks.
1130 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1131 // single covering check that is at least as strong as any of them.
1132 // If the optimization succeeds, the simplified (strengthened) range check
1133 // will always succeed. If it fails, we will deopt, and then give up
1134 // on the optimization.
1135 bool Compile::allow_range_check_smearing() const {
1136 // If this method has already thrown a range-check,
1137 // assume it was because we already tried range smearing
1138 // and it failed.
1139 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1140 return !already_trapped;
1141 }
1142
1143
1144 //------------------------------flatten_alias_type-----------------------------
1145 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1146 int offset = tj->offset();
1147 TypePtr::PTR ptr = tj->ptr();
1148
1149 // Known instance (scalarizable allocation) alias only with itself.
1150 bool is_known_inst = tj->isa_oopptr() != NULL &&
1151 tj->is_oopptr()->is_known_instance();
1152
1153 // Process weird unsafe references.
1154 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1155 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1156 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1157 tj = TypeOopPtr::BOTTOM;
1158 ptr = tj->ptr();
1159 offset = tj->offset();
1160 }
1161
1162 // Array pointers need some flattening
1163 const TypeAryPtr *ta = tj->isa_aryptr();
1164 if( ta && is_known_inst ) {
1165 if ( offset != Type::OffsetBot &&
1166 offset > arrayOopDesc::length_offset_in_bytes() ) {
1167 offset = Type::OffsetBot; // Flatten constant access into array body only
1168 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1169 }
1170 } else if( ta && _AliasLevel >= 2 ) {
1171 // For arrays indexed by constant indices, we flatten the alias
1172 // space to include all of the array body. Only the header, klass
1173 // and array length can be accessed un-aliased.
1174 if( offset != Type::OffsetBot ) {
1175 if( ta->const_oop() ) { // methodDataOop or methodOop
1176 offset = Type::OffsetBot; // Flatten constant access into array body
1177 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1178 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1179 // range is OK as-is.
1180 tj = ta = TypeAryPtr::RANGE;
1181 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1182 tj = TypeInstPtr::KLASS; // all klass loads look alike
1183 ta = TypeAryPtr::RANGE; // generic ignored junk
1184 ptr = TypePtr::BotPTR;
1185 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1186 tj = TypeInstPtr::MARK;
1187 ta = TypeAryPtr::RANGE; // generic ignored junk
1188 ptr = TypePtr::BotPTR;
1189 } else { // Random constant offset into array body
1190 offset = Type::OffsetBot; // Flatten constant access into array body
1191 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1192 }
1193 }
1194 // Arrays of fixed size alias with arrays of unknown size.
1195 if (ta->size() != TypeInt::POS) {
1196 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1197 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1198 }
1199 // Arrays of known objects become arrays of unknown objects.
1200 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1201 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1202 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1203 }
1204 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1205 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1206 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1207 }
1208 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1209 // cannot be distinguished by bytecode alone.
1210 if (ta->elem() == TypeInt::BOOL) {
1211 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1212 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1213 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1214 }
1215 // During the 2nd round of IterGVN, NotNull castings are removed.
1216 // Make sure the Bottom and NotNull variants alias the same.
1217 // Also, make sure exact and non-exact variants alias the same.
1218 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1219 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1220 }
1221 }
1222
1223 // Oop pointers need some flattening
1224 const TypeInstPtr *to = tj->isa_instptr();
1225 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1226 ciInstanceKlass *k = to->klass()->as_instance_klass();
1227 if( ptr == TypePtr::Constant ) {
1228 if (to->klass() != ciEnv::current()->Class_klass() ||
1229 offset < k->size_helper() * wordSize) {
1230 // No constant oop pointers (such as Strings); they alias with
1231 // unknown strings.
1232 assert(!is_known_inst, "not scalarizable allocation");
1233 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1234 }
1235 } else if( is_known_inst ) {
1236 tj = to; // Keep NotNull and klass_is_exact for instance type
1237 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1238 // During the 2nd round of IterGVN, NotNull castings are removed.
1239 // Make sure the Bottom and NotNull variants alias the same.
1240 // Also, make sure exact and non-exact variants alias the same.
1241 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1242 }
1243 // Canonicalize the holder of this field
1244 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1245 // First handle header references such as a LoadKlassNode, even if the
1246 // object's klass is unloaded at compile time (4965979).
1247 if (!is_known_inst) { // Do it only for non-instance types
1248 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1249 }
1250 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1251 // Static fields are in the space above the normal instance
1252 // fields in the java.lang.Class instance.
1253 if (to->klass() != ciEnv::current()->Class_klass()) {
1254 to = NULL;
1255 tj = TypeOopPtr::BOTTOM;
1256 offset = tj->offset();
1257 }
1258 } else {
1259 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1260 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1261 if( is_known_inst ) {
1262 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1263 } else {
1264 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1265 }
1266 }
1267 }
1268 }
1269
1270 // Klass pointers to object array klasses need some flattening
1271 const TypeKlassPtr *tk = tj->isa_klassptr();
1272 if( tk ) {
1273 // If we are referencing a field within a Klass, we need
1274 // to assume the worst case of an Object. Both exact and
1275 // inexact types must flatten to the same alias class.
1276 // Since the flattened result for a klass is defined to be
1277 // precisely java.lang.Object, use a constant ptr.
1278 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1279
1280 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1281 TypeKlassPtr::OBJECT->klass(),
1282 offset);
1283 }
1284
1285 ciKlass* klass = tk->klass();
1286 if( klass->is_obj_array_klass() ) {
1287 ciKlass* k = TypeAryPtr::OOPS->klass();
1288 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1289 k = TypeInstPtr::BOTTOM->klass();
1290 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1291 }
1292
1293 // Check for precise loads from the primary supertype array and force them
1294 // to the supertype cache alias index. Check for generic array loads from
1295 // the primary supertype array and also force them to the supertype cache
1296 // alias index. Since the same load can reach both, we need to merge
1297 // these 2 disparate memories into the same alias class. Since the
1298 // primary supertype array is read-only, there's no chance of confusion
1299 // where we bypass an array load and an array store.
1300 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1301 if( offset == Type::OffsetBot ||
1302 off2 < Klass::primary_super_limit()*wordSize ) {
1303 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1304 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1305 }
1306 }
1307
1308 // Flatten all Raw pointers together.
1309 if (tj->base() == Type::RawPtr)
1310 tj = TypeRawPtr::BOTTOM;
1311
1312 if (tj->base() == Type::AnyPtr)
1313 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1314
1315 // Flatten all to bottom for now
1316 switch( _AliasLevel ) {
1317 case 0:
1318 tj = TypePtr::BOTTOM;
1319 break;
1320 case 1: // Flatten to: oop, static, field or array
1321 switch (tj->base()) {
1322 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1323 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1324 case Type::AryPtr: // do not distinguish arrays at all
1325 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1326 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1327 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1328 default: ShouldNotReachHere();
1329 }
1330 break;
1331 case 2: // No collapsing at level 2; keep all splits
1332 case 3: // No collapsing at level 3; keep all splits
1333 break;
1334 default:
1335 Unimplemented();
1336 }
1337
1338 offset = tj->offset();
1339 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1340
1341 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1342 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1343 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1344 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1345 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1346 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1347 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1348 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1349 assert( tj->ptr() != TypePtr::TopPTR &&
1350 tj->ptr() != TypePtr::AnyNull &&
1351 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1352 // assert( tj->ptr() != TypePtr::Constant ||
1353 // tj->base() == Type::RawPtr ||
1354 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1355
1356 return tj;
1357 }
1358
1359 void Compile::AliasType::Init(int i, const TypePtr* at) {
1360 _index = i;
1361 _adr_type = at;
1362 _field = NULL;
1363 _is_rewritable = true; // default
1364 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1365 if (atoop != NULL && atoop->is_known_instance()) {
1366 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1367 _general_index = Compile::current()->get_alias_index(gt);
1368 } else {
1369 _general_index = 0;
1370 }
1371 }
1372
1373 //---------------------------------print_on------------------------------------
1374 #ifndef PRODUCT
1375 void Compile::AliasType::print_on(outputStream* st) {
1376 if (index() < 10)
1377 st->print("@ <%d> ", index());
1378 else st->print("@ <%d>", index());
1379 st->print(is_rewritable() ? " " : " RO");
1380 int offset = adr_type()->offset();
1381 if (offset == Type::OffsetBot)
1382 st->print(" +any");
1383 else st->print(" +%-3d", offset);
1384 st->print(" in ");
1385 adr_type()->dump_on(st);
1386 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1387 if (field() != NULL && tjp) {
1388 if (tjp->klass() != field()->holder() ||
1389 tjp->offset() != field()->offset_in_bytes()) {
1390 st->print(" != ");
1391 field()->print();
1392 st->print(" ***");
1393 }
1394 }
1395 }
1396
1397 void print_alias_types() {
1398 Compile* C = Compile::current();
1399 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1400 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1401 C->alias_type(idx)->print_on(tty);
1402 tty->cr();
1403 }
1404 }
1405 #endif
1406
1407
1408 //----------------------------probe_alias_cache--------------------------------
1409 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1410 intptr_t key = (intptr_t) adr_type;
1411 key ^= key >> logAliasCacheSize;
1412 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1413 }
1414
1415
1416 //-----------------------------grow_alias_types--------------------------------
1417 void Compile::grow_alias_types() {
1418 const int old_ats = _max_alias_types; // how many before?
1419 const int new_ats = old_ats; // how many more?
1420 const int grow_ats = old_ats+new_ats; // how many now?
1421 _max_alias_types = grow_ats;
1422 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1423 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1424 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1425 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1426 }
1427
1428
1429 //--------------------------------find_alias_type------------------------------
1430 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1431 if (_AliasLevel == 0)
1432 return alias_type(AliasIdxBot);
1433
1434 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1435 if (ace->_adr_type == adr_type) {
1436 return alias_type(ace->_index);
1437 }
1438
1439 // Handle special cases.
1440 if (adr_type == NULL) return alias_type(AliasIdxTop);
1441 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1442
1443 // Do it the slow way.
1444 const TypePtr* flat = flatten_alias_type(adr_type);
1445
1446 #ifdef ASSERT
1447 assert(flat == flatten_alias_type(flat), "idempotent");
1448 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1449 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1450 const TypeOopPtr* foop = flat->is_oopptr();
1451 // Scalarizable allocations have exact klass always.
1452 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1453 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1454 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1455 }
1456 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1457 #endif
1458
1459 int idx = AliasIdxTop;
1460 for (int i = 0; i < num_alias_types(); i++) {
1461 if (alias_type(i)->adr_type() == flat) {
1462 idx = i;
1463 break;
1464 }
1465 }
1466
1467 if (idx == AliasIdxTop) {
1468 if (no_create) return NULL;
1469 // Grow the array if necessary.
1470 if (_num_alias_types == _max_alias_types) grow_alias_types();
1471 // Add a new alias type.
1472 idx = _num_alias_types++;
1473 _alias_types[idx]->Init(idx, flat);
1474 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1475 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1476 if (flat->isa_instptr()) {
1477 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1478 && flat->is_instptr()->klass() == env()->Class_klass())
1479 alias_type(idx)->set_rewritable(false);
1480 }
1481 if (flat->isa_klassptr()) {
1482 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1483 alias_type(idx)->set_rewritable(false);
1484 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1485 alias_type(idx)->set_rewritable(false);
1486 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1487 alias_type(idx)->set_rewritable(false);
1488 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1489 alias_type(idx)->set_rewritable(false);
1490 }
1491 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1492 // but the base pointer type is not distinctive enough to identify
1493 // references into JavaThread.)
1494
1495 // Check for final fields.
1496 const TypeInstPtr* tinst = flat->isa_instptr();
1497 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1498 ciField* field;
1499 if (tinst->const_oop() != NULL &&
1500 tinst->klass() == ciEnv::current()->Class_klass() &&
1501 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1502 // static field
1503 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1504 field = k->get_field_by_offset(tinst->offset(), true);
1505 } else {
1506 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1507 field = k->get_field_by_offset(tinst->offset(), false);
1508 }
1509 assert(field == NULL ||
1510 original_field == NULL ||
1511 (field->holder() == original_field->holder() &&
1512 field->offset() == original_field->offset() &&
1513 field->is_static() == original_field->is_static()), "wrong field?");
1514 // Set field() and is_rewritable() attributes.
1515 if (field != NULL) alias_type(idx)->set_field(field);
1516 }
1517 }
1518
1519 // Fill the cache for next time.
1520 ace->_adr_type = adr_type;
1521 ace->_index = idx;
1522 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1523
1524 // Might as well try to fill the cache for the flattened version, too.
1525 AliasCacheEntry* face = probe_alias_cache(flat);
1526 if (face->_adr_type == NULL) {
1527 face->_adr_type = flat;
1528 face->_index = idx;
1529 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1530 }
1531
1532 return alias_type(idx);
1533 }
1534
1535
1536 Compile::AliasType* Compile::alias_type(ciField* field) {
1537 const TypeOopPtr* t;
1538 if (field->is_static())
1539 t = TypeInstPtr::make(field->holder()->java_mirror());
1540 else
1541 t = TypeOopPtr::make_from_klass_raw(field->holder());
1542 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1543 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1544 return atp;
1545 }
1546
1547
1548 //------------------------------have_alias_type--------------------------------
1549 bool Compile::have_alias_type(const TypePtr* adr_type) {
1550 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1551 if (ace->_adr_type == adr_type) {
1552 return true;
1553 }
1554
1555 // Handle special cases.
1556 if (adr_type == NULL) return true;
1557 if (adr_type == TypePtr::BOTTOM) return true;
1558
1559 return find_alias_type(adr_type, true, NULL) != NULL;
1560 }
1561
1562 //-----------------------------must_alias--------------------------------------
1563 // True if all values of the given address type are in the given alias category.
1564 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1565 if (alias_idx == AliasIdxBot) return true; // the universal category
1566 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1567 if (alias_idx == AliasIdxTop) return false; // the empty category
1568 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1569
1570 // the only remaining possible overlap is identity
1571 int adr_idx = get_alias_index(adr_type);
1572 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1573 assert(adr_idx == alias_idx ||
1574 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1575 && adr_type != TypeOopPtr::BOTTOM),
1576 "should not be testing for overlap with an unsafe pointer");
1577 return adr_idx == alias_idx;
1578 }
1579
1580 //------------------------------can_alias--------------------------------------
1581 // True if any values of the given address type are in the given alias category.
1582 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1583 if (alias_idx == AliasIdxTop) return false; // the empty category
1584 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1585 if (alias_idx == AliasIdxBot) return true; // the universal category
1586 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1587
1588 // the only remaining possible overlap is identity
1589 int adr_idx = get_alias_index(adr_type);
1590 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1591 return adr_idx == alias_idx;
1592 }
1593
1594
1595
1596 //---------------------------pop_warm_call-------------------------------------
1597 WarmCallInfo* Compile::pop_warm_call() {
1598 WarmCallInfo* wci = _warm_calls;
1599 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1600 return wci;
1601 }
1602
1603 //----------------------------Inline_Warm--------------------------------------
1604 int Compile::Inline_Warm() {
1605 // If there is room, try to inline some more warm call sites.
1606 // %%% Do a graph index compaction pass when we think we're out of space?
1607 if (!InlineWarmCalls) return 0;
1608
1609 int calls_made_hot = 0;
1610 int room_to_grow = NodeCountInliningCutoff - unique();
1611 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1612 int amount_grown = 0;
1613 WarmCallInfo* call;
1614 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1615 int est_size = (int)call->size();
1616 if (est_size > (room_to_grow - amount_grown)) {
1617 // This one won't fit anyway. Get rid of it.
1618 call->make_cold();
1619 continue;
1620 }
1621 call->make_hot();
1622 calls_made_hot++;
1623 amount_grown += est_size;
1624 amount_to_grow -= est_size;
1625 }
1626
1627 if (calls_made_hot > 0) set_major_progress();
1628 return calls_made_hot;
1629 }
1630
1631
1632 //----------------------------Finish_Warm--------------------------------------
1633 void Compile::Finish_Warm() {
1634 if (!InlineWarmCalls) return;
1635 if (failing()) return;
1636 if (warm_calls() == NULL) return;
1637
1638 // Clean up loose ends, if we are out of space for inlining.
1639 WarmCallInfo* call;
1640 while ((call = pop_warm_call()) != NULL) {
1641 call->make_cold();
1642 }
1643 }
1644
1645 //---------------------cleanup_loop_predicates-----------------------
1646 // Remove the opaque nodes that protect the predicates so that all unused
1647 // checks and uncommon_traps will be eliminated from the ideal graph
1648 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1649 if (predicate_count()==0) return;
1650 for (int i = predicate_count(); i > 0; i--) {
1651 Node * n = predicate_opaque1_node(i-1);
1652 assert(n->Opcode() == Op_Opaque1, "must be");
1653 igvn.replace_node(n, n->in(1));
1654 }
1655 assert(predicate_count()==0, "should be clean!");
1656 }
1657
1658 //------------------------------Optimize---------------------------------------
1659 // Given a graph, optimize it.
1660 void Compile::Optimize() {
1661 TracePhase t1("optimizer", &_t_optimizer, true);
1662
1663 #ifndef PRODUCT
1664 if (env()->break_at_compile()) {
1665 BREAKPOINT;
1666 }
1667
1668 #endif
1669
1670 ResourceMark rm;
1671 int loop_opts_cnt;
1672
1673 NOT_PRODUCT( verify_graph_edges(); )
1674
1675 print_method("After Parsing");
1676
1677 {
1678 // Iterative Global Value Numbering, including ideal transforms
1679 // Initialize IterGVN with types and values from parse-time GVN
1680 PhaseIterGVN igvn(initial_gvn());
1681 {
1682 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1683 igvn.optimize();
1684 }
1685
1686 print_method("Iter GVN 1", 2);
1687
1688 if (failing()) return;
1689
1690 // Perform escape analysis
1691 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1692 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
1693 ConnectionGraph::do_analysis(this, &igvn);
1694
1695 if (failing()) return;
1696
1697 igvn.optimize();
1698 print_method("Iter GVN 3", 2);
1699
1700 if (failing()) return;
1701
1702 }
1703
1704 // Loop transforms on the ideal graph. Range Check Elimination,
1705 // peeling, unrolling, etc.
1706
1707 // Set loop opts counter
1708 loop_opts_cnt = num_loop_opts();
1709 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1710 {
1711 TracePhase t2("idealLoop", &_t_idealLoop, true);
1712 PhaseIdealLoop ideal_loop( igvn, true );
1713 loop_opts_cnt--;
1714 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1715 if (failing()) return;
1716 }
1717 // Loop opts pass if partial peeling occurred in previous pass
1718 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1719 TracePhase t3("idealLoop", &_t_idealLoop, true);
1720 PhaseIdealLoop ideal_loop( igvn, false );
1721 loop_opts_cnt--;
1722 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1723 if (failing()) return;
1724 }
1725 // Loop opts pass for loop-unrolling before CCP
1726 if(major_progress() && (loop_opts_cnt > 0)) {
1727 TracePhase t4("idealLoop", &_t_idealLoop, true);
1728 PhaseIdealLoop ideal_loop( igvn, false );
1729 loop_opts_cnt--;
1730 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1731 }
1732 if (!failing()) {
1733 // Verify that last round of loop opts produced a valid graph
1734 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1735 PhaseIdealLoop::verify(igvn);
1736 }
1737 }
1738 if (failing()) return;
1739
1740 // Conditional Constant Propagation;
1741 PhaseCCP ccp( &igvn );
1742 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1743 {
1744 TracePhase t2("ccp", &_t_ccp, true);
1745 ccp.do_transform();
1746 }
1747 print_method("PhaseCPP 1", 2);
1748
1749 assert( true, "Break here to ccp.dump_old2new_map()");
1750
1751 // Iterative Global Value Numbering, including ideal transforms
1752 {
1753 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1754 igvn = ccp;
1755 igvn.optimize();
1756 }
1757
1758 print_method("Iter GVN 2", 2);
1759
1760 if (failing()) return;
1761
1762 // Loop transforms on the ideal graph. Range Check Elimination,
1763 // peeling, unrolling, etc.
1764 if(loop_opts_cnt > 0) {
1765 debug_only( int cnt = 0; );
1766 while(major_progress() && (loop_opts_cnt > 0)) {
1767 TracePhase t2("idealLoop", &_t_idealLoop, true);
1768 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1769 PhaseIdealLoop ideal_loop( igvn, true);
1770 loop_opts_cnt--;
1771 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1772 if (failing()) return;
1773 }
1774 }
1775
1776 {
1777 // Verify that all previous optimizations produced a valid graph
1778 // at least to this point, even if no loop optimizations were done.
1779 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1780 PhaseIdealLoop::verify(igvn);
1781 }
1782
1783 {
1784 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1785 PhaseMacroExpand mex(igvn);
1786 if (mex.expand_macro_nodes()) {
1787 assert(failing(), "must bail out w/ explicit message");
1788 return;
1789 }
1790 }
1791
1792 } // (End scope of igvn; run destructor if necessary for asserts.)
1793
1794 // A method with only infinite loops has no edges entering loops from root
1795 {
1796 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1797 if (final_graph_reshaping()) {
1798 assert(failing(), "must bail out w/ explicit message");
1799 return;
1800 }
1801 }
1802
1803 print_method("Optimize finished", 2);
1804 }
1805
1806
1807 //------------------------------Code_Gen---------------------------------------
1808 // Given a graph, generate code for it
1809 void Compile::Code_Gen() {
1810 if (failing()) return;
1811
1812 // Perform instruction selection. You might think we could reclaim Matcher
1813 // memory PDQ, but actually the Matcher is used in generating spill code.
1814 // Internals of the Matcher (including some VectorSets) must remain live
1815 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1816 // set a bit in reclaimed memory.
1817
1818 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1819 // nodes. Mapping is only valid at the root of each matched subtree.
1820 NOT_PRODUCT( verify_graph_edges(); )
1821
1822 Node_List proj_list;
1823 Matcher m(proj_list);
1824 _matcher = &m;
1825 {
1826 TracePhase t2("matcher", &_t_matcher, true);
1827 m.match();
1828 }
1829 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1830 // nodes. Mapping is only valid at the root of each matched subtree.
1831 NOT_PRODUCT( verify_graph_edges(); )
1832
1833 // If you have too many nodes, or if matching has failed, bail out
1834 check_node_count(0, "out of nodes matching instructions");
1835 if (failing()) return;
1836
1837 // Build a proper-looking CFG
1838 PhaseCFG cfg(node_arena(), root(), m);
1839 _cfg = &cfg;
1840 {
1841 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1842 cfg.Dominators();
1843 if (failing()) return;
1844
1845 NOT_PRODUCT( verify_graph_edges(); )
1846
1847 cfg.Estimate_Block_Frequency();
1848 cfg.GlobalCodeMotion(m,unique(),proj_list);
1849
1850 print_method("Global code motion", 2);
1851
1852 if (failing()) return;
1853 NOT_PRODUCT( verify_graph_edges(); )
1854
1855 debug_only( cfg.verify(); )
1856 }
1857 NOT_PRODUCT( verify_graph_edges(); )
1858
1859 PhaseChaitin regalloc(unique(),cfg,m);
1860 _regalloc = ®alloc;
1861 {
1862 TracePhase t2("regalloc", &_t_registerAllocation, true);
1863 // Perform any platform dependent preallocation actions. This is used,
1864 // for example, to avoid taking an implicit null pointer exception
1865 // using the frame pointer on win95.
1866 _regalloc->pd_preallocate_hook();
1867
1868 // Perform register allocation. After Chaitin, use-def chains are
1869 // no longer accurate (at spill code) and so must be ignored.
1870 // Node->LRG->reg mappings are still accurate.
1871 _regalloc->Register_Allocate();
1872
1873 // Bail out if the allocator builds too many nodes
1874 if (failing()) return;
1875 }
1876
1877 // Prior to register allocation we kept empty basic blocks in case the
1878 // the allocator needed a place to spill. After register allocation we
1879 // are not adding any new instructions. If any basic block is empty, we
1880 // can now safely remove it.
1881 {
1882 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1883 cfg.remove_empty();
1884 if (do_freq_based_layout()) {
1885 PhaseBlockLayout layout(cfg);
1886 } else {
1887 cfg.set_loop_alignment();
1888 }
1889 cfg.fixup_flow();
1890 }
1891
1892 // Perform any platform dependent postallocation verifications.
1893 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1894
1895 // Apply peephole optimizations
1896 if( OptoPeephole ) {
1897 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1898 PhasePeephole peep( _regalloc, cfg);
1899 peep.do_transform();
1900 }
1901
1902 // Convert Nodes to instruction bits in a buffer
1903 {
1904 // %%%% workspace merge brought two timers together for one job
1905 TracePhase t2a("output", &_t_output, true);
1906 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1907 Output();
1908 }
1909
1910 print_method("Final Code");
1911
1912 // He's dead, Jim.
1913 _cfg = (PhaseCFG*)0xdeadbeef;
1914 _regalloc = (PhaseChaitin*)0xdeadbeef;
1915 }
1916
1917
1918 //------------------------------dump_asm---------------------------------------
1919 // Dump formatted assembly
1920 #ifndef PRODUCT
1921 void Compile::dump_asm(int *pcs, uint pc_limit) {
1922 bool cut_short = false;
1923 tty->print_cr("#");
1924 tty->print("# "); _tf->dump(); tty->cr();
1925 tty->print_cr("#");
1926
1927 // For all blocks
1928 int pc = 0x0; // Program counter
1929 char starts_bundle = ' ';
1930 _regalloc->dump_frame();
1931
1932 Node *n = NULL;
1933 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1934 if (VMThread::should_terminate()) { cut_short = true; break; }
1935 Block *b = _cfg->_blocks[i];
1936 if (b->is_connector() && !Verbose) continue;
1937 n = b->_nodes[0];
1938 if (pcs && n->_idx < pc_limit)
1939 tty->print("%3.3x ", pcs[n->_idx]);
1940 else
1941 tty->print(" ");
1942 b->dump_head( &_cfg->_bbs );
1943 if (b->is_connector()) {
1944 tty->print_cr(" # Empty connector block");
1945 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1946 tty->print_cr(" # Block is sole successor of call");
1947 }
1948
1949 // For all instructions
1950 Node *delay = NULL;
1951 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1952 if (VMThread::should_terminate()) { cut_short = true; break; }
1953 n = b->_nodes[j];
1954 if (valid_bundle_info(n)) {
1955 Bundle *bundle = node_bundling(n);
1956 if (bundle->used_in_unconditional_delay()) {
1957 delay = n;
1958 continue;
1959 }
1960 if (bundle->starts_bundle())
1961 starts_bundle = '+';
1962 }
1963
1964 if (WizardMode) n->dump();
1965
1966 if( !n->is_Region() && // Dont print in the Assembly
1967 !n->is_Phi() && // a few noisely useless nodes
1968 !n->is_Proj() &&
1969 !n->is_MachTemp() &&
1970 !n->is_SafePointScalarObject() &&
1971 !n->is_Catch() && // Would be nice to print exception table targets
1972 !n->is_MergeMem() && // Not very interesting
1973 !n->is_top() && // Debug info table constants
1974 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1975 ) {
1976 if (pcs && n->_idx < pc_limit)
1977 tty->print("%3.3x", pcs[n->_idx]);
1978 else
1979 tty->print(" ");
1980 tty->print(" %c ", starts_bundle);
1981 starts_bundle = ' ';
1982 tty->print("\t");
1983 n->format(_regalloc, tty);
1984 tty->cr();
1985 }
1986
1987 // If we have an instruction with a delay slot, and have seen a delay,
1988 // then back up and print it
1989 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1990 assert(delay != NULL, "no unconditional delay instruction");
1991 if (WizardMode) delay->dump();
1992
1993 if (node_bundling(delay)->starts_bundle())
1994 starts_bundle = '+';
1995 if (pcs && n->_idx < pc_limit)
1996 tty->print("%3.3x", pcs[n->_idx]);
1997 else
1998 tty->print(" ");
1999 tty->print(" %c ", starts_bundle);
2000 starts_bundle = ' ';
2001 tty->print("\t");
2002 delay->format(_regalloc, tty);
2003 tty->print_cr("");
2004 delay = NULL;
2005 }
2006
2007 // Dump the exception table as well
2008 if( n->is_Catch() && (Verbose || WizardMode) ) {
2009 // Print the exception table for this offset
2010 _handler_table.print_subtable_for(pc);
2011 }
2012 }
2013
2014 if (pcs && n->_idx < pc_limit)
2015 tty->print_cr("%3.3x", pcs[n->_idx]);
2016 else
2017 tty->print_cr("");
2018
2019 assert(cut_short || delay == NULL, "no unconditional delay branch");
2020
2021 } // End of per-block dump
2022 tty->print_cr("");
2023
2024 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2025 }
2026 #endif
2027
2028 //------------------------------Final_Reshape_Counts---------------------------
2029 // This class defines counters to help identify when a method
2030 // may/must be executed using hardware with only 24-bit precision.
2031 struct Final_Reshape_Counts : public StackObj {
2032 int _call_count; // count non-inlined 'common' calls
2033 int _float_count; // count float ops requiring 24-bit precision
2034 int _double_count; // count double ops requiring more precision
2035 int _java_call_count; // count non-inlined 'java' calls
2036 int _inner_loop_count; // count loops which need alignment
2037 VectorSet _visited; // Visitation flags
2038 Node_List _tests; // Set of IfNodes & PCTableNodes
2039
2040 Final_Reshape_Counts() :
2041 _call_count(0), _float_count(0), _double_count(0),
2042 _java_call_count(0), _inner_loop_count(0),
2043 _visited( Thread::current()->resource_area() ) { }
2044
2045 void inc_call_count () { _call_count ++; }
2046 void inc_float_count () { _float_count ++; }
2047 void inc_double_count() { _double_count++; }
2048 void inc_java_call_count() { _java_call_count++; }
2049 void inc_inner_loop_count() { _inner_loop_count++; }
2050
2051 int get_call_count () const { return _call_count ; }
2052 int get_float_count () const { return _float_count ; }
2053 int get_double_count() const { return _double_count; }
2054 int get_java_call_count() const { return _java_call_count; }
2055 int get_inner_loop_count() const { return _inner_loop_count; }
2056 };
2057
2058 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2059 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2060 // Make sure the offset goes inside the instance layout.
2061 return k->contains_field_offset(tp->offset());
2062 // Note that OffsetBot and OffsetTop are very negative.
2063 }
2064
2065 // Eliminate trivially redundant StoreCMs and accumulate their
2066 // precedence edges.
2067 static void eliminate_redundant_card_marks(Node* n) {
2068 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2069 if (n->in(MemNode::Address)->outcnt() > 1) {
2070 // There are multiple users of the same address so it might be
2071 // possible to eliminate some of the StoreCMs
2072 Node* mem = n->in(MemNode::Memory);
2073 Node* adr = n->in(MemNode::Address);
2074 Node* val = n->in(MemNode::ValueIn);
2075 Node* prev = n;
2076 bool done = false;
2077 // Walk the chain of StoreCMs eliminating ones that match. As
2078 // long as it's a chain of single users then the optimization is
2079 // safe. Eliminating partially redundant StoreCMs would require
2080 // cloning copies down the other paths.
2081 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2082 if (adr == mem->in(MemNode::Address) &&
2083 val == mem->in(MemNode::ValueIn)) {
2084 // redundant StoreCM
2085 if (mem->req() > MemNode::OopStore) {
2086 // Hasn't been processed by this code yet.
2087 n->add_prec(mem->in(MemNode::OopStore));
2088 } else {
2089 // Already converted to precedence edge
2090 for (uint i = mem->req(); i < mem->len(); i++) {
2091 // Accumulate any precedence edges
2092 if (mem->in(i) != NULL) {
2093 n->add_prec(mem->in(i));
2094 }
2095 }
2096 // Everything above this point has been processed.
2097 done = true;
2098 }
2099 // Eliminate the previous StoreCM
2100 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2101 assert(mem->outcnt() == 0, "should be dead");
2102 mem->disconnect_inputs(NULL);
2103 } else {
2104 prev = mem;
2105 }
2106 mem = prev->in(MemNode::Memory);
2107 }
2108 }
2109 }
2110
2111 //------------------------------final_graph_reshaping_impl----------------------
2112 // Implement items 1-5 from final_graph_reshaping below.
2113 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
2114
2115 if ( n->outcnt() == 0 ) return; // dead node
2116 uint nop = n->Opcode();
2117
2118 // Check for 2-input instruction with "last use" on right input.
2119 // Swap to left input. Implements item (2).
2120 if( n->req() == 3 && // two-input instruction
2121 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2122 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2123 n->in(2)->outcnt() == 1 &&// right use IS a last use
2124 !n->in(2)->is_Con() ) { // right use is not a constant
2125 // Check for commutative opcode
2126 switch( nop ) {
2127 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2128 case Op_MaxI: case Op_MinI:
2129 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2130 case Op_AndL: case Op_XorL: case Op_OrL:
2131 case Op_AndI: case Op_XorI: case Op_OrI: {
2132 // Move "last use" input to left by swapping inputs
2133 n->swap_edges(1, 2);
2134 break;
2135 }
2136 default:
2137 break;
2138 }
2139 }
2140
2141 #ifdef ASSERT
2142 if( n->is_Mem() ) {
2143 Compile* C = Compile::current();
2144 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
2145 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2146 // oop will be recorded in oop map if load crosses safepoint
2147 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2148 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2149 "raw memory operations should have control edge");
2150 }
2151 #endif
2152 // Count FPU ops and common calls, implements item (3)
2153 switch( nop ) {
2154 // Count all float operations that may use FPU
2155 case Op_AddF:
2156 case Op_SubF:
2157 case Op_MulF:
2158 case Op_DivF:
2159 case Op_NegF:
2160 case Op_ModF:
2161 case Op_ConvI2F:
2162 case Op_ConF:
2163 case Op_CmpF:
2164 case Op_CmpF3:
2165 // case Op_ConvL2F: // longs are split into 32-bit halves
2166 frc.inc_float_count();
2167 break;
2168
2169 case Op_ConvF2D:
2170 case Op_ConvD2F:
2171 frc.inc_float_count();
2172 frc.inc_double_count();
2173 break;
2174
2175 // Count all double operations that may use FPU
2176 case Op_AddD:
2177 case Op_SubD:
2178 case Op_MulD:
2179 case Op_DivD:
2180 case Op_NegD:
2181 case Op_ModD:
2182 case Op_ConvI2D:
2183 case Op_ConvD2I:
2184 // case Op_ConvL2D: // handled by leaf call
2185 // case Op_ConvD2L: // handled by leaf call
2186 case Op_ConD:
2187 case Op_CmpD:
2188 case Op_CmpD3:
2189 frc.inc_double_count();
2190 break;
2191 case Op_Opaque1: // Remove Opaque Nodes before matching
2192 case Op_Opaque2: // Remove Opaque Nodes before matching
2193 n->subsume_by(n->in(1));
2194 break;
2195 case Op_CallStaticJava:
2196 case Op_CallJava:
2197 case Op_CallDynamicJava:
2198 frc.inc_java_call_count(); // Count java call site;
2199 case Op_CallRuntime:
2200 case Op_CallLeaf:
2201 case Op_CallLeafNoFP: {
2202 assert( n->is_Call(), "" );
2203 CallNode *call = n->as_Call();
2204 // Count call sites where the FP mode bit would have to be flipped.
2205 // Do not count uncommon runtime calls:
2206 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2207 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2208 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2209 frc.inc_call_count(); // Count the call site
2210 } else { // See if uncommon argument is shared
2211 Node *n = call->in(TypeFunc::Parms);
2212 int nop = n->Opcode();
2213 // Clone shared simple arguments to uncommon calls, item (1).
2214 if( n->outcnt() > 1 &&
2215 !n->is_Proj() &&
2216 nop != Op_CreateEx &&
2217 nop != Op_CheckCastPP &&
2218 nop != Op_DecodeN &&
2219 !n->is_Mem() ) {
2220 Node *x = n->clone();
2221 call->set_req( TypeFunc::Parms, x );
2222 }
2223 }
2224 break;
2225 }
2226
2227 case Op_StoreD:
2228 case Op_LoadD:
2229 case Op_LoadD_unaligned:
2230 frc.inc_double_count();
2231 goto handle_mem;
2232 case Op_StoreF:
2233 case Op_LoadF:
2234 frc.inc_float_count();
2235 goto handle_mem;
2236
2237 case Op_StoreCM:
2238 {
2239 // Convert OopStore dependence into precedence edge
2240 Node* prec = n->in(MemNode::OopStore);
2241 n->del_req(MemNode::OopStore);
2242 n->add_prec(prec);
2243 eliminate_redundant_card_marks(n);
2244 }
2245
2246 // fall through
2247
2248 case Op_StoreB:
2249 case Op_StoreC:
2250 case Op_StorePConditional:
2251 case Op_StoreI:
2252 case Op_StoreL:
2253 case Op_StoreIConditional:
2254 case Op_StoreLConditional:
2255 case Op_CompareAndSwapI:
2256 case Op_CompareAndSwapL:
2257 case Op_CompareAndSwapP:
2258 case Op_CompareAndSwapN:
2259 case Op_StoreP:
2260 case Op_StoreN:
2261 case Op_LoadB:
2262 case Op_LoadUB:
2263 case Op_LoadUS:
2264 case Op_LoadI:
2265 case Op_LoadUI2L:
2266 case Op_LoadKlass:
2267 case Op_LoadNKlass:
2268 case Op_LoadL:
2269 case Op_LoadL_unaligned:
2270 case Op_LoadPLocked:
2271 case Op_LoadLLocked:
2272 case Op_LoadP:
2273 case Op_LoadN:
2274 case Op_LoadRange:
2275 case Op_LoadS: {
2276 handle_mem:
2277 #ifdef ASSERT
2278 if( VerifyOptoOopOffsets ) {
2279 assert( n->is_Mem(), "" );
2280 MemNode *mem = (MemNode*)n;
2281 // Check to see if address types have grounded out somehow.
2282 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2283 assert( !tp || oop_offset_is_sane(tp), "" );
2284 }
2285 #endif
2286 break;
2287 }
2288
2289 case Op_AddP: { // Assert sane base pointers
2290 Node *addp = n->in(AddPNode::Address);
2291 assert( !addp->is_AddP() ||
2292 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2293 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2294 "Base pointers must match" );
2295 #ifdef _LP64
2296 if (UseCompressedOops &&
2297 addp->Opcode() == Op_ConP &&
2298 addp == n->in(AddPNode::Base) &&
2299 n->in(AddPNode::Offset)->is_Con()) {
2300 // Use addressing with narrow klass to load with offset on x86.
2301 // On sparc loading 32-bits constant and decoding it have less
2302 // instructions (4) then load 64-bits constant (7).
2303 // Do this transformation here since IGVN will convert ConN back to ConP.
2304 const Type* t = addp->bottom_type();
2305 if (t->isa_oopptr()) {
2306 Node* nn = NULL;
2307
2308 // Look for existing ConN node of the same exact type.
2309 Compile* C = Compile::current();
2310 Node* r = C->root();
2311 uint cnt = r->outcnt();
2312 for (uint i = 0; i < cnt; i++) {
2313 Node* m = r->raw_out(i);
2314 if (m!= NULL && m->Opcode() == Op_ConN &&
2315 m->bottom_type()->make_ptr() == t) {
2316 nn = m;
2317 break;
2318 }
2319 }
2320 if (nn != NULL) {
2321 // Decode a narrow oop to match address
2322 // [R12 + narrow_oop_reg<<3 + offset]
2323 nn = new (C, 2) DecodeNNode(nn, t);
2324 n->set_req(AddPNode::Base, nn);
2325 n->set_req(AddPNode::Address, nn);
2326 if (addp->outcnt() == 0) {
2327 addp->disconnect_inputs(NULL);
2328 }
2329 }
2330 }
2331 }
2332 #endif
2333 break;
2334 }
2335
2336 #ifdef _LP64
2337 case Op_CastPP:
2338 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2339 Compile* C = Compile::current();
2340 Node* in1 = n->in(1);
2341 const Type* t = n->bottom_type();
2342 Node* new_in1 = in1->clone();
2343 new_in1->as_DecodeN()->set_type(t);
2344
2345 if (!Matcher::narrow_oop_use_complex_address()) {
2346 //
2347 // x86, ARM and friends can handle 2 adds in addressing mode
2348 // and Matcher can fold a DecodeN node into address by using
2349 // a narrow oop directly and do implicit NULL check in address:
2350 //
2351 // [R12 + narrow_oop_reg<<3 + offset]
2352 // NullCheck narrow_oop_reg
2353 //
2354 // On other platforms (Sparc) we have to keep new DecodeN node and
2355 // use it to do implicit NULL check in address:
2356 //
2357 // decode_not_null narrow_oop_reg, base_reg
2358 // [base_reg + offset]
2359 // NullCheck base_reg
2360 //
2361 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2362 // to keep the information to which NULL check the new DecodeN node
2363 // corresponds to use it as value in implicit_null_check().
2364 //
2365 new_in1->set_req(0, n->in(0));
2366 }
2367
2368 n->subsume_by(new_in1);
2369 if (in1->outcnt() == 0) {
2370 in1->disconnect_inputs(NULL);
2371 }
2372 }
2373 break;
2374
2375 case Op_CmpP:
2376 // Do this transformation here to preserve CmpPNode::sub() and
2377 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2378 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2379 Node* in1 = n->in(1);
2380 Node* in2 = n->in(2);
2381 if (!in1->is_DecodeN()) {
2382 in2 = in1;
2383 in1 = n->in(2);
2384 }
2385 assert(in1->is_DecodeN(), "sanity");
2386
2387 Compile* C = Compile::current();
2388 Node* new_in2 = NULL;
2389 if (in2->is_DecodeN()) {
2390 new_in2 = in2->in(1);
2391 } else if (in2->Opcode() == Op_ConP) {
2392 const Type* t = in2->bottom_type();
2393 if (t == TypePtr::NULL_PTR) {
2394 // Don't convert CmpP null check into CmpN if compressed
2395 // oops implicit null check is not generated.
2396 // This will allow to generate normal oop implicit null check.
2397 if (Matcher::gen_narrow_oop_implicit_null_checks())
2398 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2399 //
2400 // This transformation together with CastPP transformation above
2401 // will generated code for implicit NULL checks for compressed oops.
2402 //
2403 // The original code after Optimize()
2404 //
2405 // LoadN memory, narrow_oop_reg
2406 // decode narrow_oop_reg, base_reg
2407 // CmpP base_reg, NULL
2408 // CastPP base_reg // NotNull
2409 // Load [base_reg + offset], val_reg
2410 //
2411 // after these transformations will be
2412 //
2413 // LoadN memory, narrow_oop_reg
2414 // CmpN narrow_oop_reg, NULL
2415 // decode_not_null narrow_oop_reg, base_reg
2416 // Load [base_reg + offset], val_reg
2417 //
2418 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2419 // since narrow oops can be used in debug info now (see the code in
2420 // final_graph_reshaping_walk()).
2421 //
2422 // At the end the code will be matched to
2423 // on x86:
2424 //
2425 // Load_narrow_oop memory, narrow_oop_reg
2426 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2427 // NullCheck narrow_oop_reg
2428 //
2429 // and on sparc:
2430 //
2431 // Load_narrow_oop memory, narrow_oop_reg
2432 // decode_not_null narrow_oop_reg, base_reg
2433 // Load [base_reg + offset], val_reg
2434 // NullCheck base_reg
2435 //
2436 } else if (t->isa_oopptr()) {
2437 new_in2 = ConNode::make(C, t->make_narrowoop());
2438 }
2439 }
2440 if (new_in2 != NULL) {
2441 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2442 n->subsume_by( cmpN );
2443 if (in1->outcnt() == 0) {
2444 in1->disconnect_inputs(NULL);
2445 }
2446 if (in2->outcnt() == 0) {
2447 in2->disconnect_inputs(NULL);
2448 }
2449 }
2450 }
2451 break;
2452
2453 case Op_DecodeN:
2454 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2455 // DecodeN could be pinned when it can't be fold into
2456 // an address expression, see the code for Op_CastPP above.
2457 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2458 break;
2459
2460 case Op_EncodeP: {
2461 Node* in1 = n->in(1);
2462 if (in1->is_DecodeN()) {
2463 n->subsume_by(in1->in(1));
2464 } else if (in1->Opcode() == Op_ConP) {
2465 Compile* C = Compile::current();
2466 const Type* t = in1->bottom_type();
2467 if (t == TypePtr::NULL_PTR) {
2468 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2469 } else if (t->isa_oopptr()) {
2470 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2471 }
2472 }
2473 if (in1->outcnt() == 0) {
2474 in1->disconnect_inputs(NULL);
2475 }
2476 break;
2477 }
2478
2479 case Op_Proj: {
2480 if (OptimizeStringConcat) {
2481 ProjNode* p = n->as_Proj();
2482 if (p->_is_io_use) {
2483 // Separate projections were used for the exception path which
2484 // are normally removed by a late inline. If it wasn't inlined
2485 // then they will hang around and should just be replaced with
2486 // the original one.
2487 Node* proj = NULL;
2488 // Replace with just one
2489 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2490 Node *use = i.get();
2491 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2492 proj = use;
2493 break;
2494 }
2495 }
2496 assert(p != NULL, "must be found");
2497 p->subsume_by(proj);
2498 }
2499 }
2500 break;
2501 }
2502
2503 case Op_Phi:
2504 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2505 // The EncodeP optimization may create Phi with the same edges
2506 // for all paths. It is not handled well by Register Allocator.
2507 Node* unique_in = n->in(1);
2508 assert(unique_in != NULL, "");
2509 uint cnt = n->req();
2510 for (uint i = 2; i < cnt; i++) {
2511 Node* m = n->in(i);
2512 assert(m != NULL, "");
2513 if (unique_in != m)
2514 unique_in = NULL;
2515 }
2516 if (unique_in != NULL) {
2517 n->subsume_by(unique_in);
2518 }
2519 }
2520 break;
2521
2522 #endif
2523
2524 case Op_ModI:
2525 if (UseDivMod) {
2526 // Check if a%b and a/b both exist
2527 Node* d = n->find_similar(Op_DivI);
2528 if (d) {
2529 // Replace them with a fused divmod if supported
2530 Compile* C = Compile::current();
2531 if (Matcher::has_match_rule(Op_DivModI)) {
2532 DivModINode* divmod = DivModINode::make(C, n);
2533 d->subsume_by(divmod->div_proj());
2534 n->subsume_by(divmod->mod_proj());
2535 } else {
2536 // replace a%b with a-((a/b)*b)
2537 Node* mult = new (C, 3) MulINode(d, d->in(2));
2538 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2539 n->subsume_by( sub );
2540 }
2541 }
2542 }
2543 break;
2544
2545 case Op_ModL:
2546 if (UseDivMod) {
2547 // Check if a%b and a/b both exist
2548 Node* d = n->find_similar(Op_DivL);
2549 if (d) {
2550 // Replace them with a fused divmod if supported
2551 Compile* C = Compile::current();
2552 if (Matcher::has_match_rule(Op_DivModL)) {
2553 DivModLNode* divmod = DivModLNode::make(C, n);
2554 d->subsume_by(divmod->div_proj());
2555 n->subsume_by(divmod->mod_proj());
2556 } else {
2557 // replace a%b with a-((a/b)*b)
2558 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2559 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2560 n->subsume_by( sub );
2561 }
2562 }
2563 }
2564 break;
2565
2566 case Op_Load16B:
2567 case Op_Load8B:
2568 case Op_Load4B:
2569 case Op_Load8S:
2570 case Op_Load4S:
2571 case Op_Load2S:
2572 case Op_Load8C:
2573 case Op_Load4C:
2574 case Op_Load2C:
2575 case Op_Load4I:
2576 case Op_Load2I:
2577 case Op_Load2L:
2578 case Op_Load4F:
2579 case Op_Load2F:
2580 case Op_Load2D:
2581 case Op_Store16B:
2582 case Op_Store8B:
2583 case Op_Store4B:
2584 case Op_Store8C:
2585 case Op_Store4C:
2586 case Op_Store2C:
2587 case Op_Store4I:
2588 case Op_Store2I:
2589 case Op_Store2L:
2590 case Op_Store4F:
2591 case Op_Store2F:
2592 case Op_Store2D:
2593 break;
2594
2595 case Op_PackB:
2596 case Op_PackS:
2597 case Op_PackC:
2598 case Op_PackI:
2599 case Op_PackF:
2600 case Op_PackL:
2601 case Op_PackD:
2602 if (n->req()-1 > 2) {
2603 // Replace many operand PackNodes with a binary tree for matching
2604 PackNode* p = (PackNode*) n;
2605 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2606 n->subsume_by(btp);
2607 }
2608 break;
2609 case Op_Loop:
2610 case Op_CountedLoop:
2611 if (n->as_Loop()->is_inner_loop()) {
2612 frc.inc_inner_loop_count();
2613 }
2614 break;
2615 case Op_LShiftI:
2616 case Op_RShiftI:
2617 case Op_URShiftI:
2618 case Op_LShiftL:
2619 case Op_RShiftL:
2620 case Op_URShiftL:
2621 if (Matcher::need_masked_shift_count) {
2622 // The cpu's shift instructions don't restrict the count to the
2623 // lower 5/6 bits. We need to do the masking ourselves.
2624 Node* in2 = n->in(2);
2625 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2626 const TypeInt* t = in2->find_int_type();
2627 if (t != NULL && t->is_con()) {
2628 juint shift = t->get_con();
2629 if (shift > mask) { // Unsigned cmp
2630 Compile* C = Compile::current();
2631 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask)));
2632 }
2633 } else {
2634 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2635 Compile* C = Compile::current();
2636 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask)));
2637 n->set_req(2, shift);
2638 }
2639 }
2640 if (in2->outcnt() == 0) { // Remove dead node
2641 in2->disconnect_inputs(NULL);
2642 }
2643 }
2644 break;
2645 default:
2646 assert( !n->is_Call(), "" );
2647 assert( !n->is_Mem(), "" );
2648 break;
2649 }
2650
2651 // Collect CFG split points
2652 if (n->is_MultiBranch())
2653 frc._tests.push(n);
2654 }
2655
2656 //------------------------------final_graph_reshaping_walk---------------------
2657 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2658 // requires that the walk visits a node's inputs before visiting the node.
2659 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2660 ResourceArea *area = Thread::current()->resource_area();
2661 Unique_Node_List sfpt(area);
2662
2663 frc._visited.set(root->_idx); // first, mark node as visited
2664 uint cnt = root->req();
2665 Node *n = root;
2666 uint i = 0;
2667 while (true) {
2668 if (i < cnt) {
2669 // Place all non-visited non-null inputs onto stack
2670 Node* m = n->in(i);
2671 ++i;
2672 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2673 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2674 sfpt.push(m);
2675 cnt = m->req();
2676 nstack.push(n, i); // put on stack parent and next input's index
2677 n = m;
2678 i = 0;
2679 }
2680 } else {
2681 // Now do post-visit work
2682 final_graph_reshaping_impl( n, frc );
2683 if (nstack.is_empty())
2684 break; // finished
2685 n = nstack.node(); // Get node from stack
2686 cnt = n->req();
2687 i = nstack.index();
2688 nstack.pop(); // Shift to the next node on stack
2689 }
2690 }
2691
2692 // Skip next transformation if compressed oops are not used.
2693 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2694 return;
2695
2696 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2697 // It could be done for an uncommon traps or any safepoints/calls
2698 // if the DecodeN node is referenced only in a debug info.
2699 while (sfpt.size() > 0) {
2700 n = sfpt.pop();
2701 JVMState *jvms = n->as_SafePoint()->jvms();
2702 assert(jvms != NULL, "sanity");
2703 int start = jvms->debug_start();
2704 int end = n->req();
2705 bool is_uncommon = (n->is_CallStaticJava() &&
2706 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2707 for (int j = start; j < end; j++) {
2708 Node* in = n->in(j);
2709 if (in->is_DecodeN()) {
2710 bool safe_to_skip = true;
2711 if (!is_uncommon ) {
2712 // Is it safe to skip?
2713 for (uint i = 0; i < in->outcnt(); i++) {
2714 Node* u = in->raw_out(i);
2715 if (!u->is_SafePoint() ||
2716 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2717 safe_to_skip = false;
2718 }
2719 }
2720 }
2721 if (safe_to_skip) {
2722 n->set_req(j, in->in(1));
2723 }
2724 if (in->outcnt() == 0) {
2725 in->disconnect_inputs(NULL);
2726 }
2727 }
2728 }
2729 }
2730 }
2731
2732 //------------------------------final_graph_reshaping--------------------------
2733 // Final Graph Reshaping.
2734 //
2735 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2736 // and not commoned up and forced early. Must come after regular
2737 // optimizations to avoid GVN undoing the cloning. Clone constant
2738 // inputs to Loop Phis; these will be split by the allocator anyways.
2739 // Remove Opaque nodes.
2740 // (2) Move last-uses by commutative operations to the left input to encourage
2741 // Intel update-in-place two-address operations and better register usage
2742 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2743 // calls canonicalizing them back.
2744 // (3) Count the number of double-precision FP ops, single-precision FP ops
2745 // and call sites. On Intel, we can get correct rounding either by
2746 // forcing singles to memory (requires extra stores and loads after each
2747 // FP bytecode) or we can set a rounding mode bit (requires setting and
2748 // clearing the mode bit around call sites). The mode bit is only used
2749 // if the relative frequency of single FP ops to calls is low enough.
2750 // This is a key transform for SPEC mpeg_audio.
2751 // (4) Detect infinite loops; blobs of code reachable from above but not
2752 // below. Several of the Code_Gen algorithms fail on such code shapes,
2753 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2754 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2755 // Detection is by looking for IfNodes where only 1 projection is
2756 // reachable from below or CatchNodes missing some targets.
2757 // (5) Assert for insane oop offsets in debug mode.
2758
2759 bool Compile::final_graph_reshaping() {
2760 // an infinite loop may have been eliminated by the optimizer,
2761 // in which case the graph will be empty.
2762 if (root()->req() == 1) {
2763 record_method_not_compilable("trivial infinite loop");
2764 return true;
2765 }
2766
2767 Final_Reshape_Counts frc;
2768
2769 // Visit everybody reachable!
2770 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2771 Node_Stack nstack(unique() >> 1);
2772 final_graph_reshaping_walk(nstack, root(), frc);
2773
2774 // Check for unreachable (from below) code (i.e., infinite loops).
2775 for( uint i = 0; i < frc._tests.size(); i++ ) {
2776 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2777 // Get number of CFG targets.
2778 // Note that PCTables include exception targets after calls.
2779 uint required_outcnt = n->required_outcnt();
2780 if (n->outcnt() != required_outcnt) {
2781 // Check for a few special cases. Rethrow Nodes never take the
2782 // 'fall-thru' path, so expected kids is 1 less.
2783 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2784 if (n->in(0)->in(0)->is_Call()) {
2785 CallNode *call = n->in(0)->in(0)->as_Call();
2786 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2787 required_outcnt--; // Rethrow always has 1 less kid
2788 } else if (call->req() > TypeFunc::Parms &&
2789 call->is_CallDynamicJava()) {
2790 // Check for null receiver. In such case, the optimizer has
2791 // detected that the virtual call will always result in a null
2792 // pointer exception. The fall-through projection of this CatchNode
2793 // will not be populated.
2794 Node *arg0 = call->in(TypeFunc::Parms);
2795 if (arg0->is_Type() &&
2796 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2797 required_outcnt--;
2798 }
2799 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2800 call->req() > TypeFunc::Parms+1 &&
2801 call->is_CallStaticJava()) {
2802 // Check for negative array length. In such case, the optimizer has
2803 // detected that the allocation attempt will always result in an
2804 // exception. There is no fall-through projection of this CatchNode .
2805 Node *arg1 = call->in(TypeFunc::Parms+1);
2806 if (arg1->is_Type() &&
2807 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2808 required_outcnt--;
2809 }
2810 }
2811 }
2812 }
2813 // Recheck with a better notion of 'required_outcnt'
2814 if (n->outcnt() != required_outcnt) {
2815 record_method_not_compilable("malformed control flow");
2816 return true; // Not all targets reachable!
2817 }
2818 }
2819 // Check that I actually visited all kids. Unreached kids
2820 // must be infinite loops.
2821 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2822 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2823 record_method_not_compilable("infinite loop");
2824 return true; // Found unvisited kid; must be unreach
2825 }
2826 }
2827
2828 // If original bytecodes contained a mixture of floats and doubles
2829 // check if the optimizer has made it homogenous, item (3).
2830 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2831 frc.get_float_count() > 32 &&
2832 frc.get_double_count() == 0 &&
2833 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2834 set_24_bit_selection_and_mode( false, true );
2835 }
2836
2837 set_java_calls(frc.get_java_call_count());
2838 set_inner_loops(frc.get_inner_loop_count());
2839
2840 // No infinite loops, no reason to bail out.
2841 return false;
2842 }
2843
2844 //-----------------------------too_many_traps----------------------------------
2845 // Report if there are too many traps at the current method and bci.
2846 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2847 bool Compile::too_many_traps(ciMethod* method,
2848 int bci,
2849 Deoptimization::DeoptReason reason) {
2850 ciMethodData* md = method->method_data();
2851 if (md->is_empty()) {
2852 // Assume the trap has not occurred, or that it occurred only
2853 // because of a transient condition during start-up in the interpreter.
2854 return false;
2855 }
2856 if (md->has_trap_at(bci, reason) != 0) {
2857 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2858 // Also, if there are multiple reasons, or if there is no per-BCI record,
2859 // assume the worst.
2860 if (log())
2861 log()->elem("observe trap='%s' count='%d'",
2862 Deoptimization::trap_reason_name(reason),
2863 md->trap_count(reason));
2864 return true;
2865 } else {
2866 // Ignore method/bci and see if there have been too many globally.
2867 return too_many_traps(reason, md);
2868 }
2869 }
2870
2871 // Less-accurate variant which does not require a method and bci.
2872 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2873 ciMethodData* logmd) {
2874 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2875 // Too many traps globally.
2876 // Note that we use cumulative trap_count, not just md->trap_count.
2877 if (log()) {
2878 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2879 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2880 Deoptimization::trap_reason_name(reason),
2881 mcount, trap_count(reason));
2882 }
2883 return true;
2884 } else {
2885 // The coast is clear.
2886 return false;
2887 }
2888 }
2889
2890 //--------------------------too_many_recompiles--------------------------------
2891 // Report if there are too many recompiles at the current method and bci.
2892 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2893 // Is not eager to return true, since this will cause the compiler to use
2894 // Action_none for a trap point, to avoid too many recompilations.
2895 bool Compile::too_many_recompiles(ciMethod* method,
2896 int bci,
2897 Deoptimization::DeoptReason reason) {
2898 ciMethodData* md = method->method_data();
2899 if (md->is_empty()) {
2900 // Assume the trap has not occurred, or that it occurred only
2901 // because of a transient condition during start-up in the interpreter.
2902 return false;
2903 }
2904 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2905 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2906 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2907 Deoptimization::DeoptReason per_bc_reason
2908 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2909 if ((per_bc_reason == Deoptimization::Reason_none
2910 || md->has_trap_at(bci, reason) != 0)
2911 // The trap frequency measure we care about is the recompile count:
2912 && md->trap_recompiled_at(bci)
2913 && md->overflow_recompile_count() >= bc_cutoff) {
2914 // Do not emit a trap here if it has already caused recompilations.
2915 // Also, if there are multiple reasons, or if there is no per-BCI record,
2916 // assume the worst.
2917 if (log())
2918 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2919 Deoptimization::trap_reason_name(reason),
2920 md->trap_count(reason),
2921 md->overflow_recompile_count());
2922 return true;
2923 } else if (trap_count(reason) != 0
2924 && decompile_count() >= m_cutoff) {
2925 // Too many recompiles globally, and we have seen this sort of trap.
2926 // Use cumulative decompile_count, not just md->decompile_count.
2927 if (log())
2928 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2929 Deoptimization::trap_reason_name(reason),
2930 md->trap_count(reason), trap_count(reason),
2931 md->decompile_count(), decompile_count());
2932 return true;
2933 } else {
2934 // The coast is clear.
2935 return false;
2936 }
2937 }
2938
2939
2940 #ifndef PRODUCT
2941 //------------------------------verify_graph_edges---------------------------
2942 // Walk the Graph and verify that there is a one-to-one correspondence
2943 // between Use-Def edges and Def-Use edges in the graph.
2944 void Compile::verify_graph_edges(bool no_dead_code) {
2945 if (VerifyGraphEdges) {
2946 ResourceArea *area = Thread::current()->resource_area();
2947 Unique_Node_List visited(area);
2948 // Call recursive graph walk to check edges
2949 _root->verify_edges(visited);
2950 if (no_dead_code) {
2951 // Now make sure that no visited node is used by an unvisited node.
2952 bool dead_nodes = 0;
2953 Unique_Node_List checked(area);
2954 while (visited.size() > 0) {
2955 Node* n = visited.pop();
2956 checked.push(n);
2957 for (uint i = 0; i < n->outcnt(); i++) {
2958 Node* use = n->raw_out(i);
2959 if (checked.member(use)) continue; // already checked
2960 if (visited.member(use)) continue; // already in the graph
2961 if (use->is_Con()) continue; // a dead ConNode is OK
2962 // At this point, we have found a dead node which is DU-reachable.
2963 if (dead_nodes++ == 0)
2964 tty->print_cr("*** Dead nodes reachable via DU edges:");
2965 use->dump(2);
2966 tty->print_cr("---");
2967 checked.push(use); // No repeats; pretend it is now checked.
2968 }
2969 }
2970 assert(dead_nodes == 0, "using nodes must be reachable from root");
2971 }
2972 }
2973 }
2974 #endif
2975
2976 // The Compile object keeps track of failure reasons separately from the ciEnv.
2977 // This is required because there is not quite a 1-1 relation between the
2978 // ciEnv and its compilation task and the Compile object. Note that one
2979 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2980 // to backtrack and retry without subsuming loads. Other than this backtracking
2981 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2982 // by the logic in C2Compiler.
2983 void Compile::record_failure(const char* reason) {
2984 if (log() != NULL) {
2985 log()->elem("failure reason='%s' phase='compile'", reason);
2986 }
2987 if (_failure_reason == NULL) {
2988 // Record the first failure reason.
2989 _failure_reason = reason;
2990 }
2991 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2992 C->print_method(_failure_reason);
2993 }
2994 _root = NULL; // flush the graph, too
2995 }
2996
2997 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2998 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2999 {
3000 if (dolog) {
3001 C = Compile::current();
3002 _log = C->log();
3003 } else {
3004 C = NULL;
3005 _log = NULL;
3006 }
3007 if (_log != NULL) {
3008 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
3009 _log->stamp();
3010 _log->end_head();
3011 }
3012 }
3013
3014 Compile::TracePhase::~TracePhase() {
3015 if (_log != NULL) {
3016 _log->done("phase nodes='%d'", C->unique());
3017 }
3018 }
3019
3020 //=============================================================================
3021 // Two Constant's are equal when the type and the value are equal.
3022 bool Compile::Constant::operator==(const Constant& other) {
3023 if (type() != other.type() ) return false;
3024 if (can_be_reused() != other.can_be_reused()) return false;
3025 // For floating point values we compare the bit pattern.
3026 switch (type()) {
3027 case T_FLOAT: return (_value.i == other._value.i);
3028 case T_LONG:
3029 case T_DOUBLE: return (_value.j == other._value.j);
3030 case T_OBJECT:
3031 case T_ADDRESS: return (_value.l == other._value.l);
3032 case T_VOID: return (_value.l == other._value.l); // jump-table entries
3033 default: ShouldNotReachHere();
3034 }
3035 return false;
3036 }
3037
3038 // Emit constants grouped in the following order:
3039 static BasicType type_order[] = {
3040 T_FLOAT, // 32-bit
3041 T_OBJECT, // 32 or 64-bit
3042 T_ADDRESS, // 32 or 64-bit
3043 T_DOUBLE, // 64-bit
3044 T_LONG, // 64-bit
3045 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
3046 T_ILLEGAL
3047 };
3048
3049 static int type_to_size_in_bytes(BasicType t) {
3050 switch (t) {
3051 case T_LONG: return sizeof(jlong );
3052 case T_FLOAT: return sizeof(jfloat );
3053 case T_DOUBLE: return sizeof(jdouble);
3054 // We use T_VOID as marker for jump-table entries (labels) which
3055 // need an interal word relocation.
3056 case T_VOID:
3057 case T_ADDRESS:
3058 case T_OBJECT: return sizeof(jobject);
3059 }
3060
3061 ShouldNotReachHere();
3062 return -1;
3063 }
3064
3065 void Compile::ConstantTable::calculate_offsets_and_size() {
3066 int size = 0;
3067 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3068 BasicType type = type_order[t];
3069
3070 for (int i = 0; i < _constants.length(); i++) {
3071 Constant con = _constants.at(i);
3072 if (con.type() != type) continue; // Skip other types.
3073
3074 // Align size for type.
3075 int typesize = type_to_size_in_bytes(con.type());
3076 size = align_size_up(size, typesize);
3077
3078 // Set offset.
3079 con.set_offset(size);
3080 _constants.at_put(i, con);
3081
3082 // Add type size.
3083 size = size + typesize;
3084 }
3085 }
3086
3087 // Align size up to the next section start (which is insts; see
3088 // CodeBuffer::align_at_start).
3089 assert(_size == -1, "already set?");
3090 _size = align_size_up(size, CodeEntryAlignment);
3091
3092 if (Matcher::constant_table_absolute_addressing) {
3093 set_table_base_offset(0); // No table base offset required
3094 } else {
3095 if (UseRDPCForConstantTableBase) {
3096 // table base offset is set in MachConstantBaseNode::emit
3097 } else {
3098 // When RDPC is not used, the table base is set into the middle of
3099 // the constant table.
3100 int half_size = _size / 2;
3101 assert(half_size * 2 == _size, "sanity");
3102 set_table_base_offset(-half_size);
3103 }
3104 }
3105 }
3106
3107 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3108 MacroAssembler _masm(&cb);
3109 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3110 BasicType type = type_order[t];
3111
3112 for (int i = 0; i < _constants.length(); i++) {
3113 Constant con = _constants.at(i);
3114 if (con.type() != type) continue; // Skip other types.
3115
3116 address constant_addr;
3117 switch (con.type()) {
3118 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3119 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3120 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3121 case T_OBJECT: {
3122 jobject obj = con.get_jobject();
3123 int oop_index = _masm.oop_recorder()->find_index(obj);
3124 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3125 break;
3126 }
3127 case T_ADDRESS: {
3128 address addr = (address) con.get_jobject();
3129 constant_addr = _masm.address_constant(addr);
3130 break;
3131 }
3132 // We use T_VOID as marker for jump-table entries (labels) which
3133 // need an interal word relocation.
3134 case T_VOID: {
3135 // Write a dummy word. The real value is filled in later
3136 // in fill_jump_table_in_constant_table.
3137 address addr = (address) con.get_jobject();
3138 constant_addr = _masm.address_constant(addr);
3139 break;
3140 }
3141 default: ShouldNotReachHere();
3142 }
3143 assert(constant_addr != NULL, "consts section too small");
3144 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3145 }
3146 }
3147 }
3148
3149 int Compile::ConstantTable::find_offset(Constant& con) const {
3150 int idx = _constants.find(con);
3151 assert(idx != -1, "constant must be in constant table");
3152 int offset = _constants.at(idx).offset();
3153 assert(offset != -1, "constant table not emitted yet?");
3154 return offset;
3155 }
3156
3157 void Compile::ConstantTable::add(Constant& con) {
3158 if (con.can_be_reused()) {
3159 int idx = _constants.find(con);
3160 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3161 return;
3162 }
3163 }
3164 (void) _constants.append(con);
3165 }
3166
3167 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
3168 Constant con(type, value);
3169 add(con);
3170 return con;
3171 }
3172
3173 Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
3174 jvalue value;
3175 BasicType type = oper->type()->basic_type();
3176 switch (type) {
3177 case T_LONG: value.j = oper->constantL(); break;
3178 case T_FLOAT: value.f = oper->constantF(); break;
3179 case T_DOUBLE: value.d = oper->constantD(); break;
3180 case T_OBJECT:
3181 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3182 default: ShouldNotReachHere();
3183 }
3184 return add(type, value);
3185 }
3186
3187 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
3188 jvalue value;
3189 // We can use the node pointer here to identify the right jump-table
3190 // as this method is called from Compile::Fill_buffer right before
3191 // the MachNodes are emitted and the jump-table is filled (means the
3192 // MachNode pointers do not change anymore).
3193 value.l = (jobject) n;
3194 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused.
3195 for (uint i = 0; i < n->outcnt(); i++) {
3196 add(con);
3197 }
3198 return con;
3199 }
3200
3201 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3202 // If called from Compile::scratch_emit_size do nothing.
3203 if (Compile::current()->in_scratch_emit_size()) return;
3204
3205 assert(labels.is_nonempty(), "must be");
3206 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
3207
3208 // Since MachConstantNode::constant_offset() also contains
3209 // table_base_offset() we need to subtract the table_base_offset()
3210 // to get the plain offset into the constant table.
3211 int offset = n->constant_offset() - table_base_offset();
3212
3213 MacroAssembler _masm(&cb);
3214 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3215
3216 for (int i = 0; i < labels.length(); i++) {
3217 address* constant_addr = &jump_table_base[i];
3218 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
3219 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3220 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3221 }
3222 }