1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/exceptionHandlerTable.hpp"
30 #include "code/nmethod.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "compiler/disassembler.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/block.hpp"
36 #include "opto/c2compiler.hpp"
37 #include "opto/callGenerator.hpp"
38 #include "opto/callnode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/chaitin.hpp"
41 #include "opto/compile.hpp"
42 #include "opto/connode.hpp"
43 #include "opto/divnode.hpp"
44 #include "opto/escape.hpp"
45 #include "opto/idealGraphPrinter.hpp"
46 #include "opto/loopnode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/macro.hpp"
49 #include "opto/matcher.hpp"
50 #include "opto/memnode.hpp"
51 #include "opto/mulnode.hpp"
52 #include "opto/node.hpp"
53 #include "opto/opcodes.hpp"
54 #include "opto/output.hpp"
55 #include "opto/parse.hpp"
56 #include "opto/phaseX.hpp"
57 #include "opto/rootnode.hpp"
58 #include "opto/runtime.hpp"
59 #include "opto/stringopts.hpp"
60 #include "opto/type.hpp"
61 #include "opto/vectornode.hpp"
62 #include "runtime/arguments.hpp"
63 #include "runtime/signature.hpp"
64 #include "runtime/stubRoutines.hpp"
65 #include "runtime/timer.hpp"
66 #include "trace/tracing.hpp"
67 #include "utilities/copy.hpp"
68 #ifdef TARGET_ARCH_MODEL_x86_32
69 # include "adfiles/ad_x86_32.hpp"
70 #endif
71 #ifdef TARGET_ARCH_MODEL_x86_64
72 # include "adfiles/ad_x86_64.hpp"
73 #endif
74 #ifdef TARGET_ARCH_MODEL_sparc
75 # include "adfiles/ad_sparc.hpp"
76 #endif
77 #ifdef TARGET_ARCH_MODEL_zero
78 # include "adfiles/ad_zero.hpp"
79 #endif
80 #ifdef TARGET_ARCH_MODEL_arm
81 # include "adfiles/ad_arm.hpp"
82 #endif
83 #ifdef TARGET_ARCH_MODEL_ppc
84 # include "adfiles/ad_ppc.hpp"
85 #endif
86
87
88 // -------------------- Compile::mach_constant_base_node -----------------------
89 // Constant table base node singleton.
90 MachConstantBaseNode* Compile::mach_constant_base_node() {
91 if (_mach_constant_base_node == NULL) {
92 _mach_constant_base_node = new (C) MachConstantBaseNode();
93 _mach_constant_base_node->add_req(C->root());
94 }
95 return _mach_constant_base_node;
96 }
97
98
99 /// Support for intrinsics.
100
101 // Return the index at which m must be inserted (or already exists).
102 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
103 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
104 #ifdef ASSERT
105 for (int i = 1; i < _intrinsics->length(); i++) {
106 CallGenerator* cg1 = _intrinsics->at(i-1);
107 CallGenerator* cg2 = _intrinsics->at(i);
108 assert(cg1->method() != cg2->method()
109 ? cg1->method() < cg2->method()
110 : cg1->is_virtual() < cg2->is_virtual(),
111 "compiler intrinsics list must stay sorted");
112 }
113 #endif
114 // Binary search sorted list, in decreasing intervals [lo, hi].
115 int lo = 0, hi = _intrinsics->length()-1;
116 while (lo <= hi) {
117 int mid = (uint)(hi + lo) / 2;
118 ciMethod* mid_m = _intrinsics->at(mid)->method();
119 if (m < mid_m) {
120 hi = mid-1;
121 } else if (m > mid_m) {
122 lo = mid+1;
123 } else {
124 // look at minor sort key
125 bool mid_virt = _intrinsics->at(mid)->is_virtual();
126 if (is_virtual < mid_virt) {
127 hi = mid-1;
128 } else if (is_virtual > mid_virt) {
129 lo = mid+1;
130 } else {
131 return mid; // exact match
132 }
133 }
134 }
135 return lo; // inexact match
136 }
137
138 void Compile::register_intrinsic(CallGenerator* cg) {
139 if (_intrinsics == NULL) {
140 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
141 }
142 // This code is stolen from ciObjectFactory::insert.
143 // Really, GrowableArray should have methods for
144 // insert_at, remove_at, and binary_search.
145 int len = _intrinsics->length();
146 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
147 if (index == len) {
148 _intrinsics->append(cg);
149 } else {
150 #ifdef ASSERT
151 CallGenerator* oldcg = _intrinsics->at(index);
152 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
153 #endif
154 _intrinsics->append(_intrinsics->at(len-1));
155 int pos;
156 for (pos = len-2; pos >= index; pos--) {
157 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
158 }
159 _intrinsics->at_put(index, cg);
160 }
161 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
162 }
163
164 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
165 assert(m->is_loaded(), "don't try this on unloaded methods");
166 if (_intrinsics != NULL) {
167 int index = intrinsic_insertion_index(m, is_virtual);
168 if (index < _intrinsics->length()
169 && _intrinsics->at(index)->method() == m
170 && _intrinsics->at(index)->is_virtual() == is_virtual) {
171 return _intrinsics->at(index);
172 }
173 }
174 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
175 if (m->intrinsic_id() != vmIntrinsics::_none &&
176 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
177 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
178 if (cg != NULL) {
179 // Save it for next time:
180 register_intrinsic(cg);
181 return cg;
182 } else {
183 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
184 }
185 }
186 return NULL;
187 }
188
189 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
190 // in library_call.cpp.
191
192
193 #ifndef PRODUCT
194 // statistics gathering...
195
196 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
197 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
198
199 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
200 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
201 int oflags = _intrinsic_hist_flags[id];
202 assert(flags != 0, "what happened?");
203 if (is_virtual) {
204 flags |= _intrinsic_virtual;
205 }
206 bool changed = (flags != oflags);
207 if ((flags & _intrinsic_worked) != 0) {
208 juint count = (_intrinsic_hist_count[id] += 1);
209 if (count == 1) {
210 changed = true; // first time
211 }
212 // increment the overall count also:
213 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
214 }
215 if (changed) {
216 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
217 // Something changed about the intrinsic's virtuality.
218 if ((flags & _intrinsic_virtual) != 0) {
219 // This is the first use of this intrinsic as a virtual call.
220 if (oflags != 0) {
221 // We already saw it as a non-virtual, so note both cases.
222 flags |= _intrinsic_both;
223 }
224 } else if ((oflags & _intrinsic_both) == 0) {
225 // This is the first use of this intrinsic as a non-virtual
226 flags |= _intrinsic_both;
227 }
228 }
229 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
230 }
231 // update the overall flags also:
232 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
233 return changed;
234 }
235
236 static char* format_flags(int flags, char* buf) {
237 buf[0] = 0;
238 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
239 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
240 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
241 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
242 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
243 if (buf[0] == 0) strcat(buf, ",");
244 assert(buf[0] == ',', "must be");
245 return &buf[1];
246 }
247
248 void Compile::print_intrinsic_statistics() {
249 char flagsbuf[100];
250 ttyLocker ttyl;
251 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
252 tty->print_cr("Compiler intrinsic usage:");
253 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
254 if (total == 0) total = 1; // avoid div0 in case of no successes
255 #define PRINT_STAT_LINE(name, c, f) \
256 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
257 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
258 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
259 int flags = _intrinsic_hist_flags[id];
260 juint count = _intrinsic_hist_count[id];
261 if ((flags | count) != 0) {
262 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
263 }
264 }
265 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
266 if (xtty != NULL) xtty->tail("statistics");
267 }
268
269 void Compile::print_statistics() {
270 { ttyLocker ttyl;
271 if (xtty != NULL) xtty->head("statistics type='opto'");
272 Parse::print_statistics();
273 PhaseCCP::print_statistics();
274 PhaseRegAlloc::print_statistics();
275 Scheduling::print_statistics();
276 PhasePeephole::print_statistics();
277 PhaseIdealLoop::print_statistics();
278 if (xtty != NULL) xtty->tail("statistics");
279 }
280 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
281 // put this under its own <statistics> element.
282 print_intrinsic_statistics();
283 }
284 }
285 #endif //PRODUCT
286
287 // Support for bundling info
288 Bundle* Compile::node_bundling(const Node *n) {
289 assert(valid_bundle_info(n), "oob");
290 return &_node_bundling_base[n->_idx];
291 }
292
293 bool Compile::valid_bundle_info(const Node *n) {
294 return (_node_bundling_limit > n->_idx);
295 }
296
297
298 void Compile::gvn_replace_by(Node* n, Node* nn) {
299 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
300 Node* use = n->last_out(i);
301 bool is_in_table = initial_gvn()->hash_delete(use);
302 uint uses_found = 0;
303 for (uint j = 0; j < use->len(); j++) {
304 if (use->in(j) == n) {
305 if (j < use->req())
306 use->set_req(j, nn);
307 else
308 use->set_prec(j, nn);
309 uses_found++;
310 }
311 }
312 if (is_in_table) {
313 // reinsert into table
314 initial_gvn()->hash_find_insert(use);
315 }
316 record_for_igvn(use);
317 i -= uses_found; // we deleted 1 or more copies of this edge
318 }
319 }
320
321
322 static inline bool not_a_node(const Node* n) {
323 if (n == NULL) return true;
324 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
325 if (*(address*)n == badAddress) return true; // kill by Node::destruct
326 return false;
327 }
328
329 // Identify all nodes that are reachable from below, useful.
330 // Use breadth-first pass that records state in a Unique_Node_List,
331 // recursive traversal is slower.
332 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
333 int estimated_worklist_size = unique();
334 useful.map( estimated_worklist_size, NULL ); // preallocate space
335
336 // Initialize worklist
337 if (root() != NULL) { useful.push(root()); }
338 // If 'top' is cached, declare it useful to preserve cached node
339 if( cached_top_node() ) { useful.push(cached_top_node()); }
340
341 // Push all useful nodes onto the list, breadthfirst
342 for( uint next = 0; next < useful.size(); ++next ) {
343 assert( next < unique(), "Unique useful nodes < total nodes");
344 Node *n = useful.at(next);
345 uint max = n->len();
346 for( uint i = 0; i < max; ++i ) {
347 Node *m = n->in(i);
348 if (not_a_node(m)) continue;
349 useful.push(m);
350 }
351 }
352 }
353
354 // Update dead_node_list with any missing dead nodes using useful
355 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
356 void Compile::update_dead_node_list(Unique_Node_List &useful) {
357 uint max_idx = unique();
358 VectorSet& useful_node_set = useful.member_set();
359
360 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
361 // If node with index node_idx is not in useful set,
362 // mark it as dead in dead node list.
363 if (! useful_node_set.test(node_idx) ) {
364 record_dead_node(node_idx);
365 }
366 }
367 }
368
369 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
370 int shift = 0;
371 for (int i = 0; i < inlines->length(); i++) {
372 CallGenerator* cg = inlines->at(i);
373 CallNode* call = cg->call_node();
374 if (shift > 0) {
375 inlines->at_put(i-shift, cg);
376 }
377 if (!useful.member(call)) {
378 shift++;
379 }
380 }
381 inlines->trunc_to(inlines->length()-shift);
382 }
383
384 // Disconnect all useless nodes by disconnecting those at the boundary.
385 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
386 uint next = 0;
387 while (next < useful.size()) {
388 Node *n = useful.at(next++);
389 // Use raw traversal of out edges since this code removes out edges
390 int max = n->outcnt();
391 for (int j = 0; j < max; ++j) {
392 Node* child = n->raw_out(j);
393 if (! useful.member(child)) {
394 assert(!child->is_top() || child != top(),
395 "If top is cached in Compile object it is in useful list");
396 // Only need to remove this out-edge to the useless node
397 n->raw_del_out(j);
398 --j;
399 --max;
400 }
401 }
402 if (n->outcnt() == 1 && n->has_special_unique_user()) {
403 record_for_igvn(n->unique_out());
404 }
405 }
406 // Remove useless macro and predicate opaq nodes
407 for (int i = C->macro_count()-1; i >= 0; i--) {
408 Node* n = C->macro_node(i);
409 if (!useful.member(n)) {
410 remove_macro_node(n);
411 }
412 }
413 // Remove useless expensive node
414 for (int i = C->expensive_count()-1; i >= 0; i--) {
415 Node* n = C->expensive_node(i);
416 if (!useful.member(n)) {
417 remove_expensive_node(n);
418 }
419 }
420 // clean up the late inline lists
421 remove_useless_late_inlines(&_string_late_inlines, useful);
422 remove_useless_late_inlines(&_boxing_late_inlines, useful);
423 remove_useless_late_inlines(&_late_inlines, useful);
424 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
425 }
426
427 //------------------------------frame_size_in_words-----------------------------
428 // frame_slots in units of words
429 int Compile::frame_size_in_words() const {
430 // shift is 0 in LP32 and 1 in LP64
431 const int shift = (LogBytesPerWord - LogBytesPerInt);
432 int words = _frame_slots >> shift;
433 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
434 return words;
435 }
436
437 // ============================================================================
438 //------------------------------CompileWrapper---------------------------------
439 class CompileWrapper : public StackObj {
440 Compile *const _compile;
441 public:
442 CompileWrapper(Compile* compile);
443
444 ~CompileWrapper();
445 };
446
447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
448 // the Compile* pointer is stored in the current ciEnv:
449 ciEnv* env = compile->env();
450 assert(env == ciEnv::current(), "must already be a ciEnv active");
451 assert(env->compiler_data() == NULL, "compile already active?");
452 env->set_compiler_data(compile);
453 assert(compile == Compile::current(), "sanity");
454
455 compile->set_type_dict(NULL);
456 compile->set_type_hwm(NULL);
457 compile->set_type_last_size(0);
458 compile->set_last_tf(NULL, NULL);
459 compile->set_indexSet_arena(NULL);
460 compile->set_indexSet_free_block_list(NULL);
461 compile->init_type_arena();
462 Type::Initialize(compile);
463 _compile->set_scratch_buffer_blob(NULL);
464 _compile->begin_method();
465 }
466 CompileWrapper::~CompileWrapper() {
467 _compile->end_method();
468 if (_compile->scratch_buffer_blob() != NULL)
469 BufferBlob::free(_compile->scratch_buffer_blob());
470 _compile->env()->set_compiler_data(NULL);
471 }
472
473
474 //----------------------------print_compile_messages---------------------------
475 void Compile::print_compile_messages() {
476 #ifndef PRODUCT
477 // Check if recompiling
478 if (_subsume_loads == false && PrintOpto) {
479 // Recompiling without allowing machine instructions to subsume loads
480 tty->print_cr("*********************************************************");
481 tty->print_cr("** Bailout: Recompile without subsuming loads **");
482 tty->print_cr("*********************************************************");
483 }
484 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
485 // Recompiling without escape analysis
486 tty->print_cr("*********************************************************");
487 tty->print_cr("** Bailout: Recompile without escape analysis **");
488 tty->print_cr("*********************************************************");
489 }
490 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
491 // Recompiling without boxing elimination
492 tty->print_cr("*********************************************************");
493 tty->print_cr("** Bailout: Recompile without boxing elimination **");
494 tty->print_cr("*********************************************************");
495 }
496 if (env()->break_at_compile()) {
497 // Open the debugger when compiling this method.
498 tty->print("### Breaking when compiling: ");
499 method()->print_short_name();
500 tty->cr();
501 BREAKPOINT;
502 }
503
504 if( PrintOpto ) {
505 if (is_osr_compilation()) {
506 tty->print("[OSR]%3d", _compile_id);
507 } else {
508 tty->print("%3d", _compile_id);
509 }
510 }
511 #endif
512 }
513
514
515 //-----------------------init_scratch_buffer_blob------------------------------
516 // Construct a temporary BufferBlob and cache it for this compile.
517 void Compile::init_scratch_buffer_blob(int const_size) {
518 // If there is already a scratch buffer blob allocated and the
519 // constant section is big enough, use it. Otherwise free the
520 // current and allocate a new one.
521 BufferBlob* blob = scratch_buffer_blob();
522 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
523 // Use the current blob.
524 } else {
525 if (blob != NULL) {
526 BufferBlob::free(blob);
527 }
528
529 ResourceMark rm;
530 _scratch_const_size = const_size;
531 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
532 blob = BufferBlob::create("Compile::scratch_buffer", size);
533 // Record the buffer blob for next time.
534 set_scratch_buffer_blob(blob);
535 // Have we run out of code space?
536 if (scratch_buffer_blob() == NULL) {
537 // Let CompilerBroker disable further compilations.
538 record_failure("Not enough space for scratch buffer in CodeCache");
539 return;
540 }
541 }
542
543 // Initialize the relocation buffers
544 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
545 set_scratch_locs_memory(locs_buf);
546 }
547
548
549 //-----------------------scratch_emit_size-------------------------------------
550 // Helper function that computes size by emitting code
551 uint Compile::scratch_emit_size(const Node* n) {
552 // Start scratch_emit_size section.
553 set_in_scratch_emit_size(true);
554
555 // Emit into a trash buffer and count bytes emitted.
556 // This is a pretty expensive way to compute a size,
557 // but it works well enough if seldom used.
558 // All common fixed-size instructions are given a size
559 // method by the AD file.
560 // Note that the scratch buffer blob and locs memory are
561 // allocated at the beginning of the compile task, and
562 // may be shared by several calls to scratch_emit_size.
563 // The allocation of the scratch buffer blob is particularly
564 // expensive, since it has to grab the code cache lock.
565 BufferBlob* blob = this->scratch_buffer_blob();
566 assert(blob != NULL, "Initialize BufferBlob at start");
567 assert(blob->size() > MAX_inst_size, "sanity");
568 relocInfo* locs_buf = scratch_locs_memory();
569 address blob_begin = blob->content_begin();
570 address blob_end = (address)locs_buf;
571 assert(blob->content_contains(blob_end), "sanity");
572 CodeBuffer buf(blob_begin, blob_end - blob_begin);
573 buf.initialize_consts_size(_scratch_const_size);
574 buf.initialize_stubs_size(MAX_stubs_size);
575 assert(locs_buf != NULL, "sanity");
576 int lsize = MAX_locs_size / 3;
577 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
578 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
579 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
580
581 // Do the emission.
582
583 Label fakeL; // Fake label for branch instructions.
584 Label* saveL = NULL;
585 uint save_bnum = 0;
586 bool is_branch = n->is_MachBranch();
587 if (is_branch) {
588 MacroAssembler masm(&buf);
589 masm.bind(fakeL);
590 n->as_MachBranch()->save_label(&saveL, &save_bnum);
591 n->as_MachBranch()->label_set(&fakeL, 0);
592 }
593 n->emit(buf, this->regalloc());
594 if (is_branch) // Restore label.
595 n->as_MachBranch()->label_set(saveL, save_bnum);
596
597 // End scratch_emit_size section.
598 set_in_scratch_emit_size(false);
599
600 return buf.insts_size();
601 }
602
603
604 // ============================================================================
605 //------------------------------Compile standard-------------------------------
606 debug_only( int Compile::_debug_idx = 100000; )
607
608 // Compile a method. entry_bci is -1 for normal compilations and indicates
609 // the continuation bci for on stack replacement.
610
611
612 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
613 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
614 : Phase(Compiler),
615 _env(ci_env),
616 _log(ci_env->log()),
617 _compile_id(ci_env->compile_id()),
618 _save_argument_registers(false),
619 _stub_name(NULL),
620 _stub_function(NULL),
621 _stub_entry_point(NULL),
622 _method(target),
623 _entry_bci(osr_bci),
624 _initial_gvn(NULL),
625 _for_igvn(NULL),
626 _warm_calls(NULL),
627 _subsume_loads(subsume_loads),
628 _do_escape_analysis(do_escape_analysis),
629 _eliminate_boxing(eliminate_boxing),
630 _failure_reason(NULL),
631 _code_buffer("Compile::Fill_buffer"),
632 _orig_pc_slot(0),
633 _orig_pc_slot_offset_in_bytes(0),
634 _has_method_handle_invokes(false),
635 _mach_constant_base_node(NULL),
636 _node_bundling_limit(0),
637 _node_bundling_base(NULL),
638 _java_calls(0),
639 _inner_loops(0),
640 _scratch_const_size(-1),
641 _in_scratch_emit_size(false),
642 _dead_node_list(comp_arena()),
643 _dead_node_count(0),
644 #ifndef PRODUCT
645 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
646 _printer(IdealGraphPrinter::printer()),
647 #endif
648 _congraph(NULL),
649 _late_inlines(comp_arena(), 2, 0, NULL),
650 _string_late_inlines(comp_arena(), 2, 0, NULL),
651 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
652 _late_inlines_pos(0),
653 _number_of_mh_late_inlines(0),
654 _inlining_progress(false),
655 _inlining_incrementally(false),
656 _print_inlining_list(NULL),
657 _print_inlining_idx(0) {
658 C = this;
659
660 CompileWrapper cw(this);
661 #ifndef PRODUCT
662 if (TimeCompiler2) {
663 tty->print(" ");
664 target->holder()->name()->print();
665 tty->print(".");
666 target->print_short_name();
667 tty->print(" ");
668 }
669 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
670 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
671 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
672 if (!print_opto_assembly) {
673 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
674 if (print_assembly && !Disassembler::can_decode()) {
675 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
676 print_opto_assembly = true;
677 }
678 }
679 set_print_assembly(print_opto_assembly);
680 set_parsed_irreducible_loop(false);
681 #endif
682 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
683 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
684
685 if (ProfileTraps) {
686 // Make sure the method being compiled gets its own MDO,
687 // so we can at least track the decompile_count().
688 method()->ensure_method_data();
689 }
690
691 Init(::AliasLevel);
692
693
694 print_compile_messages();
695
696 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
697 _ilt = InlineTree::build_inline_tree_root();
698 else
699 _ilt = NULL;
700
701 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
702 assert(num_alias_types() >= AliasIdxRaw, "");
703
704 #define MINIMUM_NODE_HASH 1023
705 // Node list that Iterative GVN will start with
706 Unique_Node_List for_igvn(comp_arena());
707 set_for_igvn(&for_igvn);
708
709 // GVN that will be run immediately on new nodes
710 uint estimated_size = method()->code_size()*4+64;
711 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
712 PhaseGVN gvn(node_arena(), estimated_size);
713 set_initial_gvn(&gvn);
714
715 if (print_inlining() || print_intrinsics()) {
716 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
717 }
718 { // Scope for timing the parser
719 TracePhase t3("parse", &_t_parser, true);
720
721 // Put top into the hash table ASAP.
722 initial_gvn()->transform_no_reclaim(top());
723
724 // Set up tf(), start(), and find a CallGenerator.
725 CallGenerator* cg = NULL;
726 if (is_osr_compilation()) {
727 const TypeTuple *domain = StartOSRNode::osr_domain();
728 const TypeTuple *range = TypeTuple::make_range(method()->signature());
729 init_tf(TypeFunc::make(domain, range));
730 StartNode* s = new (this) StartOSRNode(root(), domain);
731 initial_gvn()->set_type_bottom(s);
732 init_start(s);
733 cg = CallGenerator::for_osr(method(), entry_bci());
734 } else {
735 // Normal case.
736 init_tf(TypeFunc::make(method()));
737 StartNode* s = new (this) StartNode(root(), tf()->domain());
738 initial_gvn()->set_type_bottom(s);
739 init_start(s);
740 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
741 // With java.lang.ref.reference.get() we must go through the
742 // intrinsic when G1 is enabled - even when get() is the root
743 // method of the compile - so that, if necessary, the value in
744 // the referent field of the reference object gets recorded by
745 // the pre-barrier code.
746 // Specifically, if G1 is enabled, the value in the referent
747 // field is recorded by the G1 SATB pre barrier. This will
748 // result in the referent being marked live and the reference
749 // object removed from the list of discovered references during
750 // reference processing.
751 cg = find_intrinsic(method(), false);
752 }
753 if (cg == NULL) {
754 float past_uses = method()->interpreter_invocation_count();
755 float expected_uses = past_uses;
756 cg = CallGenerator::for_inline(method(), expected_uses);
757 }
758 }
759 if (failing()) return;
760 if (cg == NULL) {
761 record_method_not_compilable_all_tiers("cannot parse method");
762 return;
763 }
764 JVMState* jvms = build_start_state(start(), tf());
765 if ((jvms = cg->generate(jvms)) == NULL) {
766 record_method_not_compilable("method parse failed");
767 return;
768 }
769 GraphKit kit(jvms);
770
771 if (!kit.stopped()) {
772 // Accept return values, and transfer control we know not where.
773 // This is done by a special, unique ReturnNode bound to root.
774 return_values(kit.jvms());
775 }
776
777 if (kit.has_exceptions()) {
778 // Any exceptions that escape from this call must be rethrown
779 // to whatever caller is dynamically above us on the stack.
780 // This is done by a special, unique RethrowNode bound to root.
781 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
782 }
783
784 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
785
786 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
787 inline_string_calls(true);
788 }
789
790 if (failing()) return;
791
792 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
793
794 // Remove clutter produced by parsing.
795 if (!failing()) {
796 ResourceMark rm;
797 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
798 }
799 }
800
801 // Note: Large methods are capped off in do_one_bytecode().
802 if (failing()) return;
803
804 // After parsing, node notes are no longer automagic.
805 // They must be propagated by register_new_node_with_optimizer(),
806 // clone(), or the like.
807 set_default_node_notes(NULL);
808
809 for (;;) {
810 int successes = Inline_Warm();
811 if (failing()) return;
812 if (successes == 0) break;
813 }
814
815 // Drain the list.
816 Finish_Warm();
817 #ifndef PRODUCT
818 if (_printer) {
819 _printer->print_inlining(this);
820 }
821 #endif
822
823 if (failing()) return;
824 NOT_PRODUCT( verify_graph_edges(); )
825
826 // Now optimize
827 Optimize();
828 if (failing()) return;
829 NOT_PRODUCT( verify_graph_edges(); )
830
831 #ifndef PRODUCT
832 if (PrintIdeal) {
833 ttyLocker ttyl; // keep the following output all in one block
834 // This output goes directly to the tty, not the compiler log.
835 // To enable tools to match it up with the compilation activity,
836 // be sure to tag this tty output with the compile ID.
837 if (xtty != NULL) {
838 xtty->head("ideal compile_id='%d'%s", compile_id(),
839 is_osr_compilation() ? " compile_kind='osr'" :
840 "");
841 }
842 root()->dump(9999);
843 if (xtty != NULL) {
844 xtty->tail("ideal");
845 }
846 }
847 #endif
848
849 // Now that we know the size of all the monitors we can add a fixed slot
850 // for the original deopt pc.
851
852 _orig_pc_slot = fixed_slots();
853 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
854 set_fixed_slots(next_slot);
855
856 // Now generate code
857 Code_Gen();
858 if (failing()) return;
859
860 // Check if we want to skip execution of all compiled code.
861 {
862 #ifndef PRODUCT
863 if (OptoNoExecute) {
864 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
865 return;
866 }
867 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
868 #endif
869
870 if (is_osr_compilation()) {
871 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
872 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
873 } else {
874 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
875 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
876 }
877
878 env()->register_method(_method, _entry_bci,
879 &_code_offsets,
880 _orig_pc_slot_offset_in_bytes,
881 code_buffer(),
882 frame_size_in_words(), _oop_map_set,
883 &_handler_table, &_inc_table,
884 compiler,
885 env()->comp_level(),
886 has_unsafe_access(),
887 SharedRuntime::is_wide_vector(max_vector_size())
888 );
889
890 if (log() != NULL) // Print code cache state into compiler log
891 log()->code_cache_state();
892 }
893 }
894
895 //------------------------------Compile----------------------------------------
896 // Compile a runtime stub
897 Compile::Compile( ciEnv* ci_env,
898 TypeFunc_generator generator,
899 address stub_function,
900 const char *stub_name,
901 int is_fancy_jump,
902 bool pass_tls,
903 bool save_arg_registers,
904 bool return_pc )
905 : Phase(Compiler),
906 _env(ci_env),
907 _log(ci_env->log()),
908 _compile_id(0),
909 _save_argument_registers(save_arg_registers),
910 _method(NULL),
911 _stub_name(stub_name),
912 _stub_function(stub_function),
913 _stub_entry_point(NULL),
914 _entry_bci(InvocationEntryBci),
915 _initial_gvn(NULL),
916 _for_igvn(NULL),
917 _warm_calls(NULL),
918 _orig_pc_slot(0),
919 _orig_pc_slot_offset_in_bytes(0),
920 _subsume_loads(true),
921 _do_escape_analysis(false),
922 _eliminate_boxing(false),
923 _failure_reason(NULL),
924 _code_buffer("Compile::Fill_buffer"),
925 _has_method_handle_invokes(false),
926 _mach_constant_base_node(NULL),
927 _node_bundling_limit(0),
928 _node_bundling_base(NULL),
929 _java_calls(0),
930 _inner_loops(0),
931 #ifndef PRODUCT
932 _trace_opto_output(TraceOptoOutput),
933 _printer(NULL),
934 #endif
935 _dead_node_list(comp_arena()),
936 _dead_node_count(0),
937 _congraph(NULL),
938 _number_of_mh_late_inlines(0),
939 _inlining_progress(false),
940 _inlining_incrementally(false),
941 _print_inlining_list(NULL),
942 _print_inlining_idx(0) {
943 C = this;
944
945 #ifndef PRODUCT
946 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
947 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
948 set_print_assembly(PrintFrameConverterAssembly);
949 set_parsed_irreducible_loop(false);
950 #endif
951 CompileWrapper cw(this);
952 Init(/*AliasLevel=*/ 0);
953 init_tf((*generator)());
954
955 {
956 // The following is a dummy for the sake of GraphKit::gen_stub
957 Unique_Node_List for_igvn(comp_arena());
958 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
959 PhaseGVN gvn(Thread::current()->resource_area(),255);
960 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
961 gvn.transform_no_reclaim(top());
962
963 GraphKit kit;
964 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
965 }
966
967 NOT_PRODUCT( verify_graph_edges(); )
968 Code_Gen();
969 if (failing()) return;
970
971
972 // Entry point will be accessed using compile->stub_entry_point();
973 if (code_buffer() == NULL) {
974 Matcher::soft_match_failure();
975 } else {
976 if (PrintAssembly && (WizardMode || Verbose))
977 tty->print_cr("### Stub::%s", stub_name);
978
979 if (!failing()) {
980 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
981
982 // Make the NMethod
983 // For now we mark the frame as never safe for profile stackwalking
984 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
985 code_buffer(),
986 CodeOffsets::frame_never_safe,
987 // _code_offsets.value(CodeOffsets::Frame_Complete),
988 frame_size_in_words(),
989 _oop_map_set,
990 save_arg_registers);
991 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
992
993 _stub_entry_point = rs->entry_point();
994 }
995 }
996 }
997
998 //------------------------------Init-------------------------------------------
999 // Prepare for a single compilation
1000 void Compile::Init(int aliaslevel) {
1001 _unique = 0;
1002 _regalloc = NULL;
1003
1004 _tf = NULL; // filled in later
1005 _top = NULL; // cached later
1006 _matcher = NULL; // filled in later
1007 _cfg = NULL; // filled in later
1008
1009 set_24_bit_selection_and_mode(Use24BitFP, false);
1010
1011 _node_note_array = NULL;
1012 _default_node_notes = NULL;
1013
1014 _immutable_memory = NULL; // filled in at first inquiry
1015
1016 // Globally visible Nodes
1017 // First set TOP to NULL to give safe behavior during creation of RootNode
1018 set_cached_top_node(NULL);
1019 set_root(new (this) RootNode());
1020 // Now that you have a Root to point to, create the real TOP
1021 set_cached_top_node( new (this) ConNode(Type::TOP) );
1022 set_recent_alloc(NULL, NULL);
1023
1024 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1025 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1026 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1027 env()->set_dependencies(new Dependencies(env()));
1028
1029 _fixed_slots = 0;
1030 set_has_split_ifs(false);
1031 set_has_loops(has_method() && method()->has_loops()); // first approximation
1032 set_has_stringbuilder(false);
1033 set_has_boxed_value(false);
1034 _trap_can_recompile = false; // no traps emitted yet
1035 _major_progress = true; // start out assuming good things will happen
1036 set_has_unsafe_access(false);
1037 set_max_vector_size(0);
1038 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1039 set_decompile_count(0);
1040
1041 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1042 set_num_loop_opts(LoopOptsCount);
1043 set_do_inlining(Inline);
1044 set_max_inline_size(MaxInlineSize);
1045 set_freq_inline_size(FreqInlineSize);
1046 set_do_scheduling(OptoScheduling);
1047 set_do_count_invocations(false);
1048 set_do_method_data_update(false);
1049
1050 if (debug_info()->recording_non_safepoints()) {
1051 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1052 (comp_arena(), 8, 0, NULL));
1053 set_default_node_notes(Node_Notes::make(this));
1054 }
1055
1056 // // -- Initialize types before each compile --
1057 // // Update cached type information
1058 // if( _method && _method->constants() )
1059 // Type::update_loaded_types(_method, _method->constants());
1060
1061 // Init alias_type map.
1062 if (!_do_escape_analysis && aliaslevel == 3)
1063 aliaslevel = 2; // No unique types without escape analysis
1064 _AliasLevel = aliaslevel;
1065 const int grow_ats = 16;
1066 _max_alias_types = grow_ats;
1067 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1068 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1069 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1070 {
1071 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1072 }
1073 // Initialize the first few types.
1074 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1075 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1076 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1077 _num_alias_types = AliasIdxRaw+1;
1078 // Zero out the alias type cache.
1079 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1080 // A NULL adr_type hits in the cache right away. Preload the right answer.
1081 probe_alias_cache(NULL)->_index = AliasIdxTop;
1082
1083 _intrinsics = NULL;
1084 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1085 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1086 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1087 register_library_intrinsics();
1088 }
1089
1090 //---------------------------init_start----------------------------------------
1091 // Install the StartNode on this compile object.
1092 void Compile::init_start(StartNode* s) {
1093 if (failing())
1094 return; // already failing
1095 assert(s == start(), "");
1096 }
1097
1098 StartNode* Compile::start() const {
1099 assert(!failing(), "");
1100 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1101 Node* start = root()->fast_out(i);
1102 if( start->is_Start() )
1103 return start->as_Start();
1104 }
1105 ShouldNotReachHere();
1106 return NULL;
1107 }
1108
1109 //-------------------------------immutable_memory-------------------------------------
1110 // Access immutable memory
1111 Node* Compile::immutable_memory() {
1112 if (_immutable_memory != NULL) {
1113 return _immutable_memory;
1114 }
1115 StartNode* s = start();
1116 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1117 Node *p = s->fast_out(i);
1118 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1119 _immutable_memory = p;
1120 return _immutable_memory;
1121 }
1122 }
1123 ShouldNotReachHere();
1124 return NULL;
1125 }
1126
1127 //----------------------set_cached_top_node------------------------------------
1128 // Install the cached top node, and make sure Node::is_top works correctly.
1129 void Compile::set_cached_top_node(Node* tn) {
1130 if (tn != NULL) verify_top(tn);
1131 Node* old_top = _top;
1132 _top = tn;
1133 // Calling Node::setup_is_top allows the nodes the chance to adjust
1134 // their _out arrays.
1135 if (_top != NULL) _top->setup_is_top();
1136 if (old_top != NULL) old_top->setup_is_top();
1137 assert(_top == NULL || top()->is_top(), "");
1138 }
1139
1140 #ifdef ASSERT
1141 uint Compile::count_live_nodes_by_graph_walk() {
1142 Unique_Node_List useful(comp_arena());
1143 // Get useful node list by walking the graph.
1144 identify_useful_nodes(useful);
1145 return useful.size();
1146 }
1147
1148 void Compile::print_missing_nodes() {
1149
1150 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1151 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1152 return;
1153 }
1154
1155 // This is an expensive function. It is executed only when the user
1156 // specifies VerifyIdealNodeCount option or otherwise knows the
1157 // additional work that needs to be done to identify reachable nodes
1158 // by walking the flow graph and find the missing ones using
1159 // _dead_node_list.
1160
1161 Unique_Node_List useful(comp_arena());
1162 // Get useful node list by walking the graph.
1163 identify_useful_nodes(useful);
1164
1165 uint l_nodes = C->live_nodes();
1166 uint l_nodes_by_walk = useful.size();
1167
1168 if (l_nodes != l_nodes_by_walk) {
1169 if (_log != NULL) {
1170 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1171 _log->stamp();
1172 _log->end_head();
1173 }
1174 VectorSet& useful_member_set = useful.member_set();
1175 int last_idx = l_nodes_by_walk;
1176 for (int i = 0; i < last_idx; i++) {
1177 if (useful_member_set.test(i)) {
1178 if (_dead_node_list.test(i)) {
1179 if (_log != NULL) {
1180 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1181 }
1182 if (PrintIdealNodeCount) {
1183 // Print the log message to tty
1184 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1185 useful.at(i)->dump();
1186 }
1187 }
1188 }
1189 else if (! _dead_node_list.test(i)) {
1190 if (_log != NULL) {
1191 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1192 }
1193 if (PrintIdealNodeCount) {
1194 // Print the log message to tty
1195 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1196 }
1197 }
1198 }
1199 if (_log != NULL) {
1200 _log->tail("mismatched_nodes");
1201 }
1202 }
1203 }
1204 #endif
1205
1206 #ifndef PRODUCT
1207 void Compile::verify_top(Node* tn) const {
1208 if (tn != NULL) {
1209 assert(tn->is_Con(), "top node must be a constant");
1210 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1211 assert(tn->in(0) != NULL, "must have live top node");
1212 }
1213 }
1214 #endif
1215
1216
1217 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1218
1219 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1220 guarantee(arr != NULL, "");
1221 int num_blocks = arr->length();
1222 if (grow_by < num_blocks) grow_by = num_blocks;
1223 int num_notes = grow_by * _node_notes_block_size;
1224 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1225 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1226 while (num_notes > 0) {
1227 arr->append(notes);
1228 notes += _node_notes_block_size;
1229 num_notes -= _node_notes_block_size;
1230 }
1231 assert(num_notes == 0, "exact multiple, please");
1232 }
1233
1234 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1235 if (source == NULL || dest == NULL) return false;
1236
1237 if (dest->is_Con())
1238 return false; // Do not push debug info onto constants.
1239
1240 #ifdef ASSERT
1241 // Leave a bread crumb trail pointing to the original node:
1242 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1243 dest->set_debug_orig(source);
1244 }
1245 #endif
1246
1247 if (node_note_array() == NULL)
1248 return false; // Not collecting any notes now.
1249
1250 // This is a copy onto a pre-existing node, which may already have notes.
1251 // If both nodes have notes, do not overwrite any pre-existing notes.
1252 Node_Notes* source_notes = node_notes_at(source->_idx);
1253 if (source_notes == NULL || source_notes->is_clear()) return false;
1254 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1255 if (dest_notes == NULL || dest_notes->is_clear()) {
1256 return set_node_notes_at(dest->_idx, source_notes);
1257 }
1258
1259 Node_Notes merged_notes = (*source_notes);
1260 // The order of operations here ensures that dest notes will win...
1261 merged_notes.update_from(dest_notes);
1262 return set_node_notes_at(dest->_idx, &merged_notes);
1263 }
1264
1265
1266 //--------------------------allow_range_check_smearing-------------------------
1267 // Gating condition for coalescing similar range checks.
1268 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1269 // single covering check that is at least as strong as any of them.
1270 // If the optimization succeeds, the simplified (strengthened) range check
1271 // will always succeed. If it fails, we will deopt, and then give up
1272 // on the optimization.
1273 bool Compile::allow_range_check_smearing() const {
1274 // If this method has already thrown a range-check,
1275 // assume it was because we already tried range smearing
1276 // and it failed.
1277 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1278 return !already_trapped;
1279 }
1280
1281
1282 //------------------------------flatten_alias_type-----------------------------
1283 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1284 int offset = tj->offset();
1285 TypePtr::PTR ptr = tj->ptr();
1286
1287 // Known instance (scalarizable allocation) alias only with itself.
1288 bool is_known_inst = tj->isa_oopptr() != NULL &&
1289 tj->is_oopptr()->is_known_instance();
1290
1291 // Process weird unsafe references.
1292 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1293 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1294 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1295 tj = TypeOopPtr::BOTTOM;
1296 ptr = tj->ptr();
1297 offset = tj->offset();
1298 }
1299
1300 // Array pointers need some flattening
1301 const TypeAryPtr *ta = tj->isa_aryptr();
1302 if (ta && ta->is_stable()) {
1303 // Erase stability property for alias analysis.
1304 tj = ta = ta->cast_to_stable(false);
1305 }
1306 if( ta && is_known_inst ) {
1307 if ( offset != Type::OffsetBot &&
1308 offset > arrayOopDesc::length_offset_in_bytes() ) {
1309 offset = Type::OffsetBot; // Flatten constant access into array body only
1310 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1311 }
1312 } else if( ta && _AliasLevel >= 2 ) {
1313 // For arrays indexed by constant indices, we flatten the alias
1314 // space to include all of the array body. Only the header, klass
1315 // and array length can be accessed un-aliased.
1316 if( offset != Type::OffsetBot ) {
1317 if( ta->const_oop() ) { // MethodData* or Method*
1318 offset = Type::OffsetBot; // Flatten constant access into array body
1319 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1320 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1321 // range is OK as-is.
1322 tj = ta = TypeAryPtr::RANGE;
1323 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1324 tj = TypeInstPtr::KLASS; // all klass loads look alike
1325 ta = TypeAryPtr::RANGE; // generic ignored junk
1326 ptr = TypePtr::BotPTR;
1327 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1328 tj = TypeInstPtr::MARK;
1329 ta = TypeAryPtr::RANGE; // generic ignored junk
1330 ptr = TypePtr::BotPTR;
1331 } else { // Random constant offset into array body
1332 offset = Type::OffsetBot; // Flatten constant access into array body
1333 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1334 }
1335 }
1336 // Arrays of fixed size alias with arrays of unknown size.
1337 if (ta->size() != TypeInt::POS) {
1338 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1339 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1340 }
1341 // Arrays of known objects become arrays of unknown objects.
1342 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1343 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1344 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1345 }
1346 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1347 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1348 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1349 }
1350 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1351 // cannot be distinguished by bytecode alone.
1352 if (ta->elem() == TypeInt::BOOL) {
1353 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1354 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1355 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1356 }
1357 // During the 2nd round of IterGVN, NotNull castings are removed.
1358 // Make sure the Bottom and NotNull variants alias the same.
1359 // Also, make sure exact and non-exact variants alias the same.
1360 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1361 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1362 }
1363 }
1364
1365 // Oop pointers need some flattening
1366 const TypeInstPtr *to = tj->isa_instptr();
1367 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1368 ciInstanceKlass *k = to->klass()->as_instance_klass();
1369 if( ptr == TypePtr::Constant ) {
1370 if (to->klass() != ciEnv::current()->Class_klass() ||
1371 offset < k->size_helper() * wordSize) {
1372 // No constant oop pointers (such as Strings); they alias with
1373 // unknown strings.
1374 assert(!is_known_inst, "not scalarizable allocation");
1375 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1376 }
1377 } else if( is_known_inst ) {
1378 tj = to; // Keep NotNull and klass_is_exact for instance type
1379 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1380 // During the 2nd round of IterGVN, NotNull castings are removed.
1381 // Make sure the Bottom and NotNull variants alias the same.
1382 // Also, make sure exact and non-exact variants alias the same.
1383 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1384 }
1385 // Canonicalize the holder of this field
1386 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1387 // First handle header references such as a LoadKlassNode, even if the
1388 // object's klass is unloaded at compile time (4965979).
1389 if (!is_known_inst) { // Do it only for non-instance types
1390 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1391 }
1392 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1393 // Static fields are in the space above the normal instance
1394 // fields in the java.lang.Class instance.
1395 if (to->klass() != ciEnv::current()->Class_klass()) {
1396 to = NULL;
1397 tj = TypeOopPtr::BOTTOM;
1398 offset = tj->offset();
1399 }
1400 } else {
1401 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1402 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1403 if( is_known_inst ) {
1404 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1405 } else {
1406 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1407 }
1408 }
1409 }
1410 }
1411
1412 // Klass pointers to object array klasses need some flattening
1413 const TypeKlassPtr *tk = tj->isa_klassptr();
1414 if( tk ) {
1415 // If we are referencing a field within a Klass, we need
1416 // to assume the worst case of an Object. Both exact and
1417 // inexact types must flatten to the same alias class so
1418 // use NotNull as the PTR.
1419 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1420
1421 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1422 TypeKlassPtr::OBJECT->klass(),
1423 offset);
1424 }
1425
1426 ciKlass* klass = tk->klass();
1427 if( klass->is_obj_array_klass() ) {
1428 ciKlass* k = TypeAryPtr::OOPS->klass();
1429 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1430 k = TypeInstPtr::BOTTOM->klass();
1431 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1432 }
1433
1434 // Check for precise loads from the primary supertype array and force them
1435 // to the supertype cache alias index. Check for generic array loads from
1436 // the primary supertype array and also force them to the supertype cache
1437 // alias index. Since the same load can reach both, we need to merge
1438 // these 2 disparate memories into the same alias class. Since the
1439 // primary supertype array is read-only, there's no chance of confusion
1440 // where we bypass an array load and an array store.
1441 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1442 if (offset == Type::OffsetBot ||
1443 (offset >= primary_supers_offset &&
1444 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1445 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1446 offset = in_bytes(Klass::secondary_super_cache_offset());
1447 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1448 }
1449 }
1450
1451 // Flatten all Raw pointers together.
1452 if (tj->base() == Type::RawPtr)
1453 tj = TypeRawPtr::BOTTOM;
1454
1455 if (tj->base() == Type::AnyPtr)
1456 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1457
1458 // Flatten all to bottom for now
1459 switch( _AliasLevel ) {
1460 case 0:
1461 tj = TypePtr::BOTTOM;
1462 break;
1463 case 1: // Flatten to: oop, static, field or array
1464 switch (tj->base()) {
1465 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1466 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1467 case Type::AryPtr: // do not distinguish arrays at all
1468 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1469 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1470 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1471 default: ShouldNotReachHere();
1472 }
1473 break;
1474 case 2: // No collapsing at level 2; keep all splits
1475 case 3: // No collapsing at level 3; keep all splits
1476 break;
1477 default:
1478 Unimplemented();
1479 }
1480
1481 offset = tj->offset();
1482 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1483
1484 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1485 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1486 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1487 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1488 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1489 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1490 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1491 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1492 assert( tj->ptr() != TypePtr::TopPTR &&
1493 tj->ptr() != TypePtr::AnyNull &&
1494 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1495 // assert( tj->ptr() != TypePtr::Constant ||
1496 // tj->base() == Type::RawPtr ||
1497 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1498
1499 return tj;
1500 }
1501
1502 void Compile::AliasType::Init(int i, const TypePtr* at) {
1503 _index = i;
1504 _adr_type = at;
1505 _field = NULL;
1506 _element = NULL;
1507 _is_rewritable = true; // default
1508 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1509 if (atoop != NULL && atoop->is_known_instance()) {
1510 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1511 _general_index = Compile::current()->get_alias_index(gt);
1512 } else {
1513 _general_index = 0;
1514 }
1515 }
1516
1517 //---------------------------------print_on------------------------------------
1518 #ifndef PRODUCT
1519 void Compile::AliasType::print_on(outputStream* st) {
1520 if (index() < 10)
1521 st->print("@ <%d> ", index());
1522 else st->print("@ <%d>", index());
1523 st->print(is_rewritable() ? " " : " RO");
1524 int offset = adr_type()->offset();
1525 if (offset == Type::OffsetBot)
1526 st->print(" +any");
1527 else st->print(" +%-3d", offset);
1528 st->print(" in ");
1529 adr_type()->dump_on(st);
1530 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1531 if (field() != NULL && tjp) {
1532 if (tjp->klass() != field()->holder() ||
1533 tjp->offset() != field()->offset_in_bytes()) {
1534 st->print(" != ");
1535 field()->print();
1536 st->print(" ***");
1537 }
1538 }
1539 }
1540
1541 void print_alias_types() {
1542 Compile* C = Compile::current();
1543 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1544 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1545 C->alias_type(idx)->print_on(tty);
1546 tty->cr();
1547 }
1548 }
1549 #endif
1550
1551
1552 //----------------------------probe_alias_cache--------------------------------
1553 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1554 intptr_t key = (intptr_t) adr_type;
1555 key ^= key >> logAliasCacheSize;
1556 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1557 }
1558
1559
1560 //-----------------------------grow_alias_types--------------------------------
1561 void Compile::grow_alias_types() {
1562 const int old_ats = _max_alias_types; // how many before?
1563 const int new_ats = old_ats; // how many more?
1564 const int grow_ats = old_ats+new_ats; // how many now?
1565 _max_alias_types = grow_ats;
1566 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1567 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1568 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1569 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1570 }
1571
1572
1573 //--------------------------------find_alias_type------------------------------
1574 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1575 if (_AliasLevel == 0)
1576 return alias_type(AliasIdxBot);
1577
1578 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1579 if (ace->_adr_type == adr_type) {
1580 return alias_type(ace->_index);
1581 }
1582
1583 // Handle special cases.
1584 if (adr_type == NULL) return alias_type(AliasIdxTop);
1585 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1586
1587 // Do it the slow way.
1588 const TypePtr* flat = flatten_alias_type(adr_type);
1589
1590 #ifdef ASSERT
1591 assert(flat == flatten_alias_type(flat), "idempotent");
1592 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1593 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1594 const TypeOopPtr* foop = flat->is_oopptr();
1595 // Scalarizable allocations have exact klass always.
1596 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1597 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1598 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1599 }
1600 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1601 #endif
1602
1603 int idx = AliasIdxTop;
1604 for (int i = 0; i < num_alias_types(); i++) {
1605 if (alias_type(i)->adr_type() == flat) {
1606 idx = i;
1607 break;
1608 }
1609 }
1610
1611 if (idx == AliasIdxTop) {
1612 if (no_create) return NULL;
1613 // Grow the array if necessary.
1614 if (_num_alias_types == _max_alias_types) grow_alias_types();
1615 // Add a new alias type.
1616 idx = _num_alias_types++;
1617 _alias_types[idx]->Init(idx, flat);
1618 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1619 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1620 if (flat->isa_instptr()) {
1621 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1622 && flat->is_instptr()->klass() == env()->Class_klass())
1623 alias_type(idx)->set_rewritable(false);
1624 }
1625 if (flat->isa_aryptr()) {
1626 #ifdef ASSERT
1627 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1628 // (T_BYTE has the weakest alignment and size restrictions...)
1629 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1630 #endif
1631 if (flat->offset() == TypePtr::OffsetBot) {
1632 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1633 }
1634 }
1635 if (flat->isa_klassptr()) {
1636 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1637 alias_type(idx)->set_rewritable(false);
1638 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1639 alias_type(idx)->set_rewritable(false);
1640 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1641 alias_type(idx)->set_rewritable(false);
1642 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1643 alias_type(idx)->set_rewritable(false);
1644 }
1645 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1646 // but the base pointer type is not distinctive enough to identify
1647 // references into JavaThread.)
1648
1649 // Check for final fields.
1650 const TypeInstPtr* tinst = flat->isa_instptr();
1651 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1652 ciField* field;
1653 if (tinst->const_oop() != NULL &&
1654 tinst->klass() == ciEnv::current()->Class_klass() &&
1655 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1656 // static field
1657 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1658 field = k->get_field_by_offset(tinst->offset(), true);
1659 } else {
1660 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1661 field = k->get_field_by_offset(tinst->offset(), false);
1662 }
1663 assert(field == NULL ||
1664 original_field == NULL ||
1665 (field->holder() == original_field->holder() &&
1666 field->offset() == original_field->offset() &&
1667 field->is_static() == original_field->is_static()), "wrong field?");
1668 // Set field() and is_rewritable() attributes.
1669 if (field != NULL) alias_type(idx)->set_field(field);
1670 }
1671 }
1672
1673 // Fill the cache for next time.
1674 ace->_adr_type = adr_type;
1675 ace->_index = idx;
1676 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1677
1678 // Might as well try to fill the cache for the flattened version, too.
1679 AliasCacheEntry* face = probe_alias_cache(flat);
1680 if (face->_adr_type == NULL) {
1681 face->_adr_type = flat;
1682 face->_index = idx;
1683 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1684 }
1685
1686 return alias_type(idx);
1687 }
1688
1689
1690 Compile::AliasType* Compile::alias_type(ciField* field) {
1691 const TypeOopPtr* t;
1692 if (field->is_static())
1693 t = TypeInstPtr::make(field->holder()->java_mirror());
1694 else
1695 t = TypeOopPtr::make_from_klass_raw(field->holder());
1696 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1697 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1698 return atp;
1699 }
1700
1701
1702 //------------------------------have_alias_type--------------------------------
1703 bool Compile::have_alias_type(const TypePtr* adr_type) {
1704 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1705 if (ace->_adr_type == adr_type) {
1706 return true;
1707 }
1708
1709 // Handle special cases.
1710 if (adr_type == NULL) return true;
1711 if (adr_type == TypePtr::BOTTOM) return true;
1712
1713 return find_alias_type(adr_type, true, NULL) != NULL;
1714 }
1715
1716 //-----------------------------must_alias--------------------------------------
1717 // True if all values of the given address type are in the given alias category.
1718 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1719 if (alias_idx == AliasIdxBot) return true; // the universal category
1720 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1721 if (alias_idx == AliasIdxTop) return false; // the empty category
1722 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1723
1724 // the only remaining possible overlap is identity
1725 int adr_idx = get_alias_index(adr_type);
1726 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1727 assert(adr_idx == alias_idx ||
1728 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1729 && adr_type != TypeOopPtr::BOTTOM),
1730 "should not be testing for overlap with an unsafe pointer");
1731 return adr_idx == alias_idx;
1732 }
1733
1734 //------------------------------can_alias--------------------------------------
1735 // True if any values of the given address type are in the given alias category.
1736 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1737 if (alias_idx == AliasIdxTop) return false; // the empty category
1738 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1739 if (alias_idx == AliasIdxBot) return true; // the universal category
1740 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1741
1742 // the only remaining possible overlap is identity
1743 int adr_idx = get_alias_index(adr_type);
1744 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1745 return adr_idx == alias_idx;
1746 }
1747
1748
1749
1750 //---------------------------pop_warm_call-------------------------------------
1751 WarmCallInfo* Compile::pop_warm_call() {
1752 WarmCallInfo* wci = _warm_calls;
1753 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1754 return wci;
1755 }
1756
1757 //----------------------------Inline_Warm--------------------------------------
1758 int Compile::Inline_Warm() {
1759 // If there is room, try to inline some more warm call sites.
1760 // %%% Do a graph index compaction pass when we think we're out of space?
1761 if (!InlineWarmCalls) return 0;
1762
1763 int calls_made_hot = 0;
1764 int room_to_grow = NodeCountInliningCutoff - unique();
1765 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1766 int amount_grown = 0;
1767 WarmCallInfo* call;
1768 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1769 int est_size = (int)call->size();
1770 if (est_size > (room_to_grow - amount_grown)) {
1771 // This one won't fit anyway. Get rid of it.
1772 call->make_cold();
1773 continue;
1774 }
1775 call->make_hot();
1776 calls_made_hot++;
1777 amount_grown += est_size;
1778 amount_to_grow -= est_size;
1779 }
1780
1781 if (calls_made_hot > 0) set_major_progress();
1782 return calls_made_hot;
1783 }
1784
1785
1786 //----------------------------Finish_Warm--------------------------------------
1787 void Compile::Finish_Warm() {
1788 if (!InlineWarmCalls) return;
1789 if (failing()) return;
1790 if (warm_calls() == NULL) return;
1791
1792 // Clean up loose ends, if we are out of space for inlining.
1793 WarmCallInfo* call;
1794 while ((call = pop_warm_call()) != NULL) {
1795 call->make_cold();
1796 }
1797 }
1798
1799 //---------------------cleanup_loop_predicates-----------------------
1800 // Remove the opaque nodes that protect the predicates so that all unused
1801 // checks and uncommon_traps will be eliminated from the ideal graph
1802 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1803 if (predicate_count()==0) return;
1804 for (int i = predicate_count(); i > 0; i--) {
1805 Node * n = predicate_opaque1_node(i-1);
1806 assert(n->Opcode() == Op_Opaque1, "must be");
1807 igvn.replace_node(n, n->in(1));
1808 }
1809 assert(predicate_count()==0, "should be clean!");
1810 }
1811
1812 // StringOpts and late inlining of string methods
1813 void Compile::inline_string_calls(bool parse_time) {
1814 {
1815 // remove useless nodes to make the usage analysis simpler
1816 ResourceMark rm;
1817 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1818 }
1819
1820 {
1821 ResourceMark rm;
1822 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1823 PhaseStringOpts pso(initial_gvn(), for_igvn());
1824 print_method(PHASE_AFTER_STRINGOPTS, 3);
1825 }
1826
1827 // now inline anything that we skipped the first time around
1828 if (!parse_time) {
1829 _late_inlines_pos = _late_inlines.length();
1830 }
1831
1832 while (_string_late_inlines.length() > 0) {
1833 CallGenerator* cg = _string_late_inlines.pop();
1834 cg->do_late_inline();
1835 if (failing()) return;
1836 }
1837 _string_late_inlines.trunc_to(0);
1838 }
1839
1840 // Late inlining of boxing methods
1841 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1842 if (_boxing_late_inlines.length() > 0) {
1843 assert(has_boxed_value(), "inconsistent");
1844
1845 PhaseGVN* gvn = initial_gvn();
1846 set_inlining_incrementally(true);
1847
1848 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1849 for_igvn()->clear();
1850 gvn->replace_with(&igvn);
1851
1852 while (_boxing_late_inlines.length() > 0) {
1853 CallGenerator* cg = _boxing_late_inlines.pop();
1854 cg->do_late_inline();
1855 if (failing()) return;
1856 }
1857 _boxing_late_inlines.trunc_to(0);
1858
1859 {
1860 ResourceMark rm;
1861 PhaseRemoveUseless pru(gvn, for_igvn());
1862 }
1863
1864 igvn = PhaseIterGVN(gvn);
1865 igvn.optimize();
1866
1867 set_inlining_progress(false);
1868 set_inlining_incrementally(false);
1869 }
1870 }
1871
1872 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1873 assert(IncrementalInline, "incremental inlining should be on");
1874 PhaseGVN* gvn = initial_gvn();
1875
1876 set_inlining_progress(false);
1877 for_igvn()->clear();
1878 gvn->replace_with(&igvn);
1879
1880 int i = 0;
1881
1882 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1883 CallGenerator* cg = _late_inlines.at(i);
1884 _late_inlines_pos = i+1;
1885 cg->do_late_inline();
1886 if (failing()) return;
1887 }
1888 int j = 0;
1889 for (; i < _late_inlines.length(); i++, j++) {
1890 _late_inlines.at_put(j, _late_inlines.at(i));
1891 }
1892 _late_inlines.trunc_to(j);
1893
1894 {
1895 ResourceMark rm;
1896 PhaseRemoveUseless pru(gvn, for_igvn());
1897 }
1898
1899 igvn = PhaseIterGVN(gvn);
1900 }
1901
1902 // Perform incremental inlining until bound on number of live nodes is reached
1903 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1904 PhaseGVN* gvn = initial_gvn();
1905
1906 set_inlining_incrementally(true);
1907 set_inlining_progress(true);
1908 uint low_live_nodes = 0;
1909
1910 while(inlining_progress() && _late_inlines.length() > 0) {
1911
1912 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1913 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1914 // PhaseIdealLoop is expensive so we only try it once we are
1915 // out of loop and we only try it again if the previous helped
1916 // got the number of nodes down significantly
1917 PhaseIdealLoop ideal_loop( igvn, false, true );
1918 if (failing()) return;
1919 low_live_nodes = live_nodes();
1920 _major_progress = true;
1921 }
1922
1923 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1924 break;
1925 }
1926 }
1927
1928 inline_incrementally_one(igvn);
1929
1930 if (failing()) return;
1931
1932 igvn.optimize();
1933
1934 if (failing()) return;
1935 }
1936
1937 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1938
1939 if (_string_late_inlines.length() > 0) {
1940 assert(has_stringbuilder(), "inconsistent");
1941 for_igvn()->clear();
1942 initial_gvn()->replace_with(&igvn);
1943
1944 inline_string_calls(false);
1945
1946 if (failing()) return;
1947
1948 {
1949 ResourceMark rm;
1950 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1951 }
1952
1953 igvn = PhaseIterGVN(gvn);
1954
1955 igvn.optimize();
1956 }
1957
1958 set_inlining_incrementally(false);
1959 }
1960
1961
1962 //------------------------------Optimize---------------------------------------
1963 // Given a graph, optimize it.
1964 void Compile::Optimize() {
1965 TracePhase t1("optimizer", &_t_optimizer, true);
1966
1967 #ifndef PRODUCT
1968 if (env()->break_at_compile()) {
1969 BREAKPOINT;
1970 }
1971
1972 #endif
1973
1974 ResourceMark rm;
1975 int loop_opts_cnt;
1976
1977 NOT_PRODUCT( verify_graph_edges(); )
1978
1979 print_method(PHASE_AFTER_PARSING);
1980
1981 {
1982 // Iterative Global Value Numbering, including ideal transforms
1983 // Initialize IterGVN with types and values from parse-time GVN
1984 PhaseIterGVN igvn(initial_gvn());
1985 {
1986 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1987 igvn.optimize();
1988 }
1989
1990 print_method(PHASE_ITER_GVN1, 2);
1991
1992 if (failing()) return;
1993
1994 {
1995 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
1996 inline_incrementally(igvn);
1997 }
1998
1999 print_method(PHASE_INCREMENTAL_INLINE, 2);
2000
2001 if (failing()) return;
2002
2003 if (eliminate_boxing()) {
2004 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2005 // Inline valueOf() methods now.
2006 inline_boxing_calls(igvn);
2007
2008 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2009
2010 if (failing()) return;
2011 }
2012
2013 // No more new expensive nodes will be added to the list from here
2014 // so keep only the actual candidates for optimizations.
2015 cleanup_expensive_nodes(igvn);
2016
2017 // Perform escape analysis
2018 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2019 if (has_loops()) {
2020 // Cleanup graph (remove dead nodes).
2021 TracePhase t2("idealLoop", &_t_idealLoop, true);
2022 PhaseIdealLoop ideal_loop( igvn, false, true );
2023 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2024 if (failing()) return;
2025 }
2026 ConnectionGraph::do_analysis(this, &igvn);
2027
2028 if (failing()) return;
2029
2030 // Optimize out fields loads from scalar replaceable allocations.
2031 igvn.optimize();
2032 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2033
2034 if (failing()) return;
2035
2036 if (congraph() != NULL && macro_count() > 0) {
2037 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2038 PhaseMacroExpand mexp(igvn);
2039 mexp.eliminate_macro_nodes();
2040 igvn.set_delay_transform(false);
2041
2042 igvn.optimize();
2043 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2044
2045 if (failing()) return;
2046 }
2047 }
2048
2049 // Loop transforms on the ideal graph. Range Check Elimination,
2050 // peeling, unrolling, etc.
2051
2052 // Set loop opts counter
2053 loop_opts_cnt = num_loop_opts();
2054 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2055 {
2056 TracePhase t2("idealLoop", &_t_idealLoop, true);
2057 PhaseIdealLoop ideal_loop( igvn, true );
2058 loop_opts_cnt--;
2059 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2060 if (failing()) return;
2061 }
2062 // Loop opts pass if partial peeling occurred in previous pass
2063 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2064 TracePhase t3("idealLoop", &_t_idealLoop, true);
2065 PhaseIdealLoop ideal_loop( igvn, false );
2066 loop_opts_cnt--;
2067 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2068 if (failing()) return;
2069 }
2070 // Loop opts pass for loop-unrolling before CCP
2071 if(major_progress() && (loop_opts_cnt > 0)) {
2072 TracePhase t4("idealLoop", &_t_idealLoop, true);
2073 PhaseIdealLoop ideal_loop( igvn, false );
2074 loop_opts_cnt--;
2075 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2076 }
2077 if (!failing()) {
2078 // Verify that last round of loop opts produced a valid graph
2079 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2080 PhaseIdealLoop::verify(igvn);
2081 }
2082 }
2083 if (failing()) return;
2084
2085 // Conditional Constant Propagation;
2086 PhaseCCP ccp( &igvn );
2087 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2088 {
2089 TracePhase t2("ccp", &_t_ccp, true);
2090 ccp.do_transform();
2091 }
2092 print_method(PHASE_CPP1, 2);
2093
2094 assert( true, "Break here to ccp.dump_old2new_map()");
2095
2096 // Iterative Global Value Numbering, including ideal transforms
2097 {
2098 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2099 igvn = ccp;
2100 igvn.optimize();
2101 }
2102
2103 print_method(PHASE_ITER_GVN2, 2);
2104
2105 if (failing()) return;
2106
2107 // Loop transforms on the ideal graph. Range Check Elimination,
2108 // peeling, unrolling, etc.
2109 if(loop_opts_cnt > 0) {
2110 debug_only( int cnt = 0; );
2111 while(major_progress() && (loop_opts_cnt > 0)) {
2112 TracePhase t2("idealLoop", &_t_idealLoop, true);
2113 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2114 PhaseIdealLoop ideal_loop( igvn, true);
2115 loop_opts_cnt--;
2116 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2117 if (failing()) return;
2118 }
2119 }
2120
2121 {
2122 // Verify that all previous optimizations produced a valid graph
2123 // at least to this point, even if no loop optimizations were done.
2124 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2125 PhaseIdealLoop::verify(igvn);
2126 }
2127
2128 {
2129 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2130 PhaseMacroExpand mex(igvn);
2131 if (mex.expand_macro_nodes()) {
2132 assert(failing(), "must bail out w/ explicit message");
2133 return;
2134 }
2135 }
2136
2137 } // (End scope of igvn; run destructor if necessary for asserts.)
2138
2139 dump_inlining();
2140 // A method with only infinite loops has no edges entering loops from root
2141 {
2142 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2143 if (final_graph_reshaping()) {
2144 assert(failing(), "must bail out w/ explicit message");
2145 return;
2146 }
2147 }
2148
2149 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2150 }
2151
2152
2153 //------------------------------Code_Gen---------------------------------------
2154 // Given a graph, generate code for it
2155 void Compile::Code_Gen() {
2156 if (failing()) {
2157 return;
2158 }
2159
2160 // Perform instruction selection. You might think we could reclaim Matcher
2161 // memory PDQ, but actually the Matcher is used in generating spill code.
2162 // Internals of the Matcher (including some VectorSets) must remain live
2163 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2164 // set a bit in reclaimed memory.
2165
2166 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2167 // nodes. Mapping is only valid at the root of each matched subtree.
2168 NOT_PRODUCT( verify_graph_edges(); )
2169
2170 Matcher matcher;
2171 _matcher = &matcher;
2172 {
2173 TracePhase t2("matcher", &_t_matcher, true);
2174 matcher.match();
2175 }
2176 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2177 // nodes. Mapping is only valid at the root of each matched subtree.
2178 NOT_PRODUCT( verify_graph_edges(); )
2179
2180 // If you have too many nodes, or if matching has failed, bail out
2181 check_node_count(0, "out of nodes matching instructions");
2182 if (failing()) {
2183 return;
2184 }
2185
2186 // Build a proper-looking CFG
2187 PhaseCFG cfg(node_arena(), root(), matcher);
2188 _cfg = &cfg;
2189 {
2190 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2191 bool success = cfg.do_global_code_motion();
2192 if (!success) {
2193 return;
2194 }
2195
2196 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2197 NOT_PRODUCT( verify_graph_edges(); )
2198 debug_only( cfg.verify(); )
2199 }
2200
2201 PhaseChaitin regalloc(unique(), cfg, matcher);
2202 _regalloc = ®alloc;
2203 {
2204 TracePhase t2("regalloc", &_t_registerAllocation, true);
2205 // Perform register allocation. After Chaitin, use-def chains are
2206 // no longer accurate (at spill code) and so must be ignored.
2207 // Node->LRG->reg mappings are still accurate.
2208 _regalloc->Register_Allocate();
2209
2210 // Bail out if the allocator builds too many nodes
2211 if (failing()) {
2212 return;
2213 }
2214 }
2215
2216 // Prior to register allocation we kept empty basic blocks in case the
2217 // the allocator needed a place to spill. After register allocation we
2218 // are not adding any new instructions. If any basic block is empty, we
2219 // can now safely remove it.
2220 {
2221 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2222 cfg.remove_empty_blocks();
2223 if (do_freq_based_layout()) {
2224 PhaseBlockLayout layout(cfg);
2225 } else {
2226 cfg.set_loop_alignment();
2227 }
2228 cfg.fixup_flow();
2229 }
2230
2231 // Apply peephole optimizations
2232 if( OptoPeephole ) {
2233 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2234 PhasePeephole peep( _regalloc, cfg);
2235 peep.do_transform();
2236 }
2237
2238 // Convert Nodes to instruction bits in a buffer
2239 {
2240 // %%%% workspace merge brought two timers together for one job
2241 TracePhase t2a("output", &_t_output, true);
2242 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2243 Output();
2244 }
2245
2246 print_method(PHASE_FINAL_CODE);
2247
2248 // He's dead, Jim.
2249 _cfg = (PhaseCFG*)0xdeadbeef;
2250 _regalloc = (PhaseChaitin*)0xdeadbeef;
2251 }
2252
2253
2254 //------------------------------dump_asm---------------------------------------
2255 // Dump formatted assembly
2256 #ifndef PRODUCT
2257 void Compile::dump_asm(int *pcs, uint pc_limit) {
2258 bool cut_short = false;
2259 tty->print_cr("#");
2260 tty->print("# "); _tf->dump(); tty->cr();
2261 tty->print_cr("#");
2262
2263 // For all blocks
2264 int pc = 0x0; // Program counter
2265 char starts_bundle = ' ';
2266 _regalloc->dump_frame();
2267
2268 Node *n = NULL;
2269 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2270 if (VMThread::should_terminate()) {
2271 cut_short = true;
2272 break;
2273 }
2274 Block* block = _cfg->get_block(i);
2275 if (block->is_connector() && !Verbose) {
2276 continue;
2277 }
2278 n = block->head();
2279 if (pcs && n->_idx < pc_limit) {
2280 tty->print("%3.3x ", pcs[n->_idx]);
2281 } else {
2282 tty->print(" ");
2283 }
2284 block->dump_head(_cfg);
2285 if (block->is_connector()) {
2286 tty->print_cr(" # Empty connector block");
2287 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2288 tty->print_cr(" # Block is sole successor of call");
2289 }
2290
2291 // For all instructions
2292 Node *delay = NULL;
2293 for (uint j = 0; j < block->number_of_nodes(); j++) {
2294 if (VMThread::should_terminate()) {
2295 cut_short = true;
2296 break;
2297 }
2298 n = block->get_node(j);
2299 if (valid_bundle_info(n)) {
2300 Bundle* bundle = node_bundling(n);
2301 if (bundle->used_in_unconditional_delay()) {
2302 delay = n;
2303 continue;
2304 }
2305 if (bundle->starts_bundle()) {
2306 starts_bundle = '+';
2307 }
2308 }
2309
2310 if (WizardMode) {
2311 n->dump();
2312 }
2313
2314 if( !n->is_Region() && // Dont print in the Assembly
2315 !n->is_Phi() && // a few noisely useless nodes
2316 !n->is_Proj() &&
2317 !n->is_MachTemp() &&
2318 !n->is_SafePointScalarObject() &&
2319 !n->is_Catch() && // Would be nice to print exception table targets
2320 !n->is_MergeMem() && // Not very interesting
2321 !n->is_top() && // Debug info table constants
2322 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2323 ) {
2324 if (pcs && n->_idx < pc_limit)
2325 tty->print("%3.3x", pcs[n->_idx]);
2326 else
2327 tty->print(" ");
2328 tty->print(" %c ", starts_bundle);
2329 starts_bundle = ' ';
2330 tty->print("\t");
2331 n->format(_regalloc, tty);
2332 tty->cr();
2333 }
2334
2335 // If we have an instruction with a delay slot, and have seen a delay,
2336 // then back up and print it
2337 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2338 assert(delay != NULL, "no unconditional delay instruction");
2339 if (WizardMode) delay->dump();
2340
2341 if (node_bundling(delay)->starts_bundle())
2342 starts_bundle = '+';
2343 if (pcs && n->_idx < pc_limit)
2344 tty->print("%3.3x", pcs[n->_idx]);
2345 else
2346 tty->print(" ");
2347 tty->print(" %c ", starts_bundle);
2348 starts_bundle = ' ';
2349 tty->print("\t");
2350 delay->format(_regalloc, tty);
2351 tty->print_cr("");
2352 delay = NULL;
2353 }
2354
2355 // Dump the exception table as well
2356 if( n->is_Catch() && (Verbose || WizardMode) ) {
2357 // Print the exception table for this offset
2358 _handler_table.print_subtable_for(pc);
2359 }
2360 }
2361
2362 if (pcs && n->_idx < pc_limit)
2363 tty->print_cr("%3.3x", pcs[n->_idx]);
2364 else
2365 tty->print_cr("");
2366
2367 assert(cut_short || delay == NULL, "no unconditional delay branch");
2368
2369 } // End of per-block dump
2370 tty->print_cr("");
2371
2372 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2373 }
2374 #endif
2375
2376 //------------------------------Final_Reshape_Counts---------------------------
2377 // This class defines counters to help identify when a method
2378 // may/must be executed using hardware with only 24-bit precision.
2379 struct Final_Reshape_Counts : public StackObj {
2380 int _call_count; // count non-inlined 'common' calls
2381 int _float_count; // count float ops requiring 24-bit precision
2382 int _double_count; // count double ops requiring more precision
2383 int _java_call_count; // count non-inlined 'java' calls
2384 int _inner_loop_count; // count loops which need alignment
2385 VectorSet _visited; // Visitation flags
2386 Node_List _tests; // Set of IfNodes & PCTableNodes
2387
2388 Final_Reshape_Counts() :
2389 _call_count(0), _float_count(0), _double_count(0),
2390 _java_call_count(0), _inner_loop_count(0),
2391 _visited( Thread::current()->resource_area() ) { }
2392
2393 void inc_call_count () { _call_count ++; }
2394 void inc_float_count () { _float_count ++; }
2395 void inc_double_count() { _double_count++; }
2396 void inc_java_call_count() { _java_call_count++; }
2397 void inc_inner_loop_count() { _inner_loop_count++; }
2398
2399 int get_call_count () const { return _call_count ; }
2400 int get_float_count () const { return _float_count ; }
2401 int get_double_count() const { return _double_count; }
2402 int get_java_call_count() const { return _java_call_count; }
2403 int get_inner_loop_count() const { return _inner_loop_count; }
2404 };
2405
2406 #ifdef ASSERT
2407 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2408 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2409 // Make sure the offset goes inside the instance layout.
2410 return k->contains_field_offset(tp->offset());
2411 // Note that OffsetBot and OffsetTop are very negative.
2412 }
2413 #endif
2414
2415 // Eliminate trivially redundant StoreCMs and accumulate their
2416 // precedence edges.
2417 void Compile::eliminate_redundant_card_marks(Node* n) {
2418 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2419 if (n->in(MemNode::Address)->outcnt() > 1) {
2420 // There are multiple users of the same address so it might be
2421 // possible to eliminate some of the StoreCMs
2422 Node* mem = n->in(MemNode::Memory);
2423 Node* adr = n->in(MemNode::Address);
2424 Node* val = n->in(MemNode::ValueIn);
2425 Node* prev = n;
2426 bool done = false;
2427 // Walk the chain of StoreCMs eliminating ones that match. As
2428 // long as it's a chain of single users then the optimization is
2429 // safe. Eliminating partially redundant StoreCMs would require
2430 // cloning copies down the other paths.
2431 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2432 if (adr == mem->in(MemNode::Address) &&
2433 val == mem->in(MemNode::ValueIn)) {
2434 // redundant StoreCM
2435 if (mem->req() > MemNode::OopStore) {
2436 // Hasn't been processed by this code yet.
2437 n->add_prec(mem->in(MemNode::OopStore));
2438 } else {
2439 // Already converted to precedence edge
2440 for (uint i = mem->req(); i < mem->len(); i++) {
2441 // Accumulate any precedence edges
2442 if (mem->in(i) != NULL) {
2443 n->add_prec(mem->in(i));
2444 }
2445 }
2446 // Everything above this point has been processed.
2447 done = true;
2448 }
2449 // Eliminate the previous StoreCM
2450 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2451 assert(mem->outcnt() == 0, "should be dead");
2452 mem->disconnect_inputs(NULL, this);
2453 } else {
2454 prev = mem;
2455 }
2456 mem = prev->in(MemNode::Memory);
2457 }
2458 }
2459 }
2460
2461 //------------------------------final_graph_reshaping_impl----------------------
2462 // Implement items 1-5 from final_graph_reshaping below.
2463 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2464
2465 if ( n->outcnt() == 0 ) return; // dead node
2466 uint nop = n->Opcode();
2467
2468 // Check for 2-input instruction with "last use" on right input.
2469 // Swap to left input. Implements item (2).
2470 if( n->req() == 3 && // two-input instruction
2471 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2472 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2473 n->in(2)->outcnt() == 1 &&// right use IS a last use
2474 !n->in(2)->is_Con() ) { // right use is not a constant
2475 // Check for commutative opcode
2476 switch( nop ) {
2477 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2478 case Op_MaxI: case Op_MinI:
2479 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2480 case Op_AndL: case Op_XorL: case Op_OrL:
2481 case Op_AndI: case Op_XorI: case Op_OrI: {
2482 // Move "last use" input to left by swapping inputs
2483 n->swap_edges(1, 2);
2484 break;
2485 }
2486 default:
2487 break;
2488 }
2489 }
2490
2491 #ifdef ASSERT
2492 if( n->is_Mem() ) {
2493 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2494 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2495 // oop will be recorded in oop map if load crosses safepoint
2496 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2497 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2498 "raw memory operations should have control edge");
2499 }
2500 #endif
2501 // Count FPU ops and common calls, implements item (3)
2502 switch( nop ) {
2503 // Count all float operations that may use FPU
2504 case Op_AddF:
2505 case Op_SubF:
2506 case Op_MulF:
2507 case Op_DivF:
2508 case Op_NegF:
2509 case Op_ModF:
2510 case Op_ConvI2F:
2511 case Op_ConF:
2512 case Op_CmpF:
2513 case Op_CmpF3:
2514 // case Op_ConvL2F: // longs are split into 32-bit halves
2515 frc.inc_float_count();
2516 break;
2517
2518 case Op_ConvF2D:
2519 case Op_ConvD2F:
2520 frc.inc_float_count();
2521 frc.inc_double_count();
2522 break;
2523
2524 // Count all double operations that may use FPU
2525 case Op_AddD:
2526 case Op_SubD:
2527 case Op_MulD:
2528 case Op_DivD:
2529 case Op_NegD:
2530 case Op_ModD:
2531 case Op_ConvI2D:
2532 case Op_ConvD2I:
2533 // case Op_ConvL2D: // handled by leaf call
2534 // case Op_ConvD2L: // handled by leaf call
2535 case Op_ConD:
2536 case Op_CmpD:
2537 case Op_CmpD3:
2538 frc.inc_double_count();
2539 break;
2540 case Op_Opaque1: // Remove Opaque Nodes before matching
2541 case Op_Opaque2: // Remove Opaque Nodes before matching
2542 n->subsume_by(n->in(1), this);
2543 break;
2544 case Op_CallStaticJava:
2545 case Op_CallJava:
2546 case Op_CallDynamicJava:
2547 frc.inc_java_call_count(); // Count java call site;
2548 case Op_CallRuntime:
2549 case Op_CallLeaf:
2550 case Op_CallLeafNoFP: {
2551 assert( n->is_Call(), "" );
2552 CallNode *call = n->as_Call();
2553 // Count call sites where the FP mode bit would have to be flipped.
2554 // Do not count uncommon runtime calls:
2555 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2556 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2557 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2558 frc.inc_call_count(); // Count the call site
2559 } else { // See if uncommon argument is shared
2560 Node *n = call->in(TypeFunc::Parms);
2561 int nop = n->Opcode();
2562 // Clone shared simple arguments to uncommon calls, item (1).
2563 if( n->outcnt() > 1 &&
2564 !n->is_Proj() &&
2565 nop != Op_CreateEx &&
2566 nop != Op_CheckCastPP &&
2567 nop != Op_DecodeN &&
2568 nop != Op_DecodeNKlass &&
2569 !n->is_Mem() ) {
2570 Node *x = n->clone();
2571 call->set_req( TypeFunc::Parms, x );
2572 }
2573 }
2574 break;
2575 }
2576
2577 case Op_StoreD:
2578 case Op_LoadD:
2579 case Op_LoadD_unaligned:
2580 frc.inc_double_count();
2581 goto handle_mem;
2582 case Op_StoreF:
2583 case Op_LoadF:
2584 frc.inc_float_count();
2585 goto handle_mem;
2586
2587 case Op_StoreCM:
2588 {
2589 // Convert OopStore dependence into precedence edge
2590 Node* prec = n->in(MemNode::OopStore);
2591 n->del_req(MemNode::OopStore);
2592 n->add_prec(prec);
2593 eliminate_redundant_card_marks(n);
2594 }
2595
2596 // fall through
2597
2598 case Op_StoreB:
2599 case Op_StoreC:
2600 case Op_StorePConditional:
2601 case Op_StoreI:
2602 case Op_StoreL:
2603 case Op_StoreIConditional:
2604 case Op_StoreLConditional:
2605 case Op_CompareAndSwapI:
2606 case Op_CompareAndSwapL:
2607 case Op_CompareAndSwapP:
2608 case Op_CompareAndSwapN:
2609 case Op_GetAndAddI:
2610 case Op_GetAndAddL:
2611 case Op_GetAndSetI:
2612 case Op_GetAndSetL:
2613 case Op_GetAndSetP:
2614 case Op_GetAndSetN:
2615 case Op_StoreP:
2616 case Op_StoreN:
2617 case Op_StoreNKlass:
2618 case Op_LoadB:
2619 case Op_LoadUB:
2620 case Op_LoadUS:
2621 case Op_LoadI:
2622 case Op_LoadKlass:
2623 case Op_LoadNKlass:
2624 case Op_LoadL:
2625 case Op_LoadL_unaligned:
2626 case Op_LoadPLocked:
2627 case Op_LoadP:
2628 case Op_LoadN:
2629 case Op_LoadRange:
2630 case Op_LoadS: {
2631 handle_mem:
2632 #ifdef ASSERT
2633 if( VerifyOptoOopOffsets ) {
2634 assert( n->is_Mem(), "" );
2635 MemNode *mem = (MemNode*)n;
2636 // Check to see if address types have grounded out somehow.
2637 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2638 assert( !tp || oop_offset_is_sane(tp), "" );
2639 }
2640 #endif
2641 break;
2642 }
2643
2644 case Op_AddP: { // Assert sane base pointers
2645 Node *addp = n->in(AddPNode::Address);
2646 assert( !addp->is_AddP() ||
2647 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2648 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2649 "Base pointers must match" );
2650 #ifdef _LP64
2651 if ((UseCompressedOops || UseCompressedClassPointers) &&
2652 addp->Opcode() == Op_ConP &&
2653 addp == n->in(AddPNode::Base) &&
2654 n->in(AddPNode::Offset)->is_Con()) {
2655 // Use addressing with narrow klass to load with offset on x86.
2656 // On sparc loading 32-bits constant and decoding it have less
2657 // instructions (4) then load 64-bits constant (7).
2658 // Do this transformation here since IGVN will convert ConN back to ConP.
2659 const Type* t = addp->bottom_type();
2660 if (t->isa_oopptr() || t->isa_klassptr()) {
2661 Node* nn = NULL;
2662
2663 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2664
2665 // Look for existing ConN node of the same exact type.
2666 Node* r = root();
2667 uint cnt = r->outcnt();
2668 for (uint i = 0; i < cnt; i++) {
2669 Node* m = r->raw_out(i);
2670 if (m!= NULL && m->Opcode() == op &&
2671 m->bottom_type()->make_ptr() == t) {
2672 nn = m;
2673 break;
2674 }
2675 }
2676 if (nn != NULL) {
2677 // Decode a narrow oop to match address
2678 // [R12 + narrow_oop_reg<<3 + offset]
2679 if (t->isa_oopptr()) {
2680 nn = new (this) DecodeNNode(nn, t);
2681 } else {
2682 nn = new (this) DecodeNKlassNode(nn, t);
2683 }
2684 n->set_req(AddPNode::Base, nn);
2685 n->set_req(AddPNode::Address, nn);
2686 if (addp->outcnt() == 0) {
2687 addp->disconnect_inputs(NULL, this);
2688 }
2689 }
2690 }
2691 }
2692 #endif
2693 break;
2694 }
2695
2696 #ifdef _LP64
2697 case Op_CastPP:
2698 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2699 Node* in1 = n->in(1);
2700 const Type* t = n->bottom_type();
2701 Node* new_in1 = in1->clone();
2702 new_in1->as_DecodeN()->set_type(t);
2703
2704 if (!Matcher::narrow_oop_use_complex_address()) {
2705 //
2706 // x86, ARM and friends can handle 2 adds in addressing mode
2707 // and Matcher can fold a DecodeN node into address by using
2708 // a narrow oop directly and do implicit NULL check in address:
2709 //
2710 // [R12 + narrow_oop_reg<<3 + offset]
2711 // NullCheck narrow_oop_reg
2712 //
2713 // On other platforms (Sparc) we have to keep new DecodeN node and
2714 // use it to do implicit NULL check in address:
2715 //
2716 // decode_not_null narrow_oop_reg, base_reg
2717 // [base_reg + offset]
2718 // NullCheck base_reg
2719 //
2720 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2721 // to keep the information to which NULL check the new DecodeN node
2722 // corresponds to use it as value in implicit_null_check().
2723 //
2724 new_in1->set_req(0, n->in(0));
2725 }
2726
2727 n->subsume_by(new_in1, this);
2728 if (in1->outcnt() == 0) {
2729 in1->disconnect_inputs(NULL, this);
2730 }
2731 }
2732 break;
2733
2734 case Op_CmpP:
2735 // Do this transformation here to preserve CmpPNode::sub() and
2736 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2737 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2738 Node* in1 = n->in(1);
2739 Node* in2 = n->in(2);
2740 if (!in1->is_DecodeNarrowPtr()) {
2741 in2 = in1;
2742 in1 = n->in(2);
2743 }
2744 assert(in1->is_DecodeNarrowPtr(), "sanity");
2745
2746 Node* new_in2 = NULL;
2747 if (in2->is_DecodeNarrowPtr()) {
2748 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2749 new_in2 = in2->in(1);
2750 } else if (in2->Opcode() == Op_ConP) {
2751 const Type* t = in2->bottom_type();
2752 if (t == TypePtr::NULL_PTR) {
2753 assert(in1->is_DecodeN(), "compare klass to null?");
2754 // Don't convert CmpP null check into CmpN if compressed
2755 // oops implicit null check is not generated.
2756 // This will allow to generate normal oop implicit null check.
2757 if (Matcher::gen_narrow_oop_implicit_null_checks())
2758 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2759 //
2760 // This transformation together with CastPP transformation above
2761 // will generated code for implicit NULL checks for compressed oops.
2762 //
2763 // The original code after Optimize()
2764 //
2765 // LoadN memory, narrow_oop_reg
2766 // decode narrow_oop_reg, base_reg
2767 // CmpP base_reg, NULL
2768 // CastPP base_reg // NotNull
2769 // Load [base_reg + offset], val_reg
2770 //
2771 // after these transformations will be
2772 //
2773 // LoadN memory, narrow_oop_reg
2774 // CmpN narrow_oop_reg, NULL
2775 // decode_not_null narrow_oop_reg, base_reg
2776 // Load [base_reg + offset], val_reg
2777 //
2778 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2779 // since narrow oops can be used in debug info now (see the code in
2780 // final_graph_reshaping_walk()).
2781 //
2782 // At the end the code will be matched to
2783 // on x86:
2784 //
2785 // Load_narrow_oop memory, narrow_oop_reg
2786 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2787 // NullCheck narrow_oop_reg
2788 //
2789 // and on sparc:
2790 //
2791 // Load_narrow_oop memory, narrow_oop_reg
2792 // decode_not_null narrow_oop_reg, base_reg
2793 // Load [base_reg + offset], val_reg
2794 // NullCheck base_reg
2795 //
2796 } else if (t->isa_oopptr()) {
2797 new_in2 = ConNode::make(this, t->make_narrowoop());
2798 } else if (t->isa_klassptr()) {
2799 new_in2 = ConNode::make(this, t->make_narrowklass());
2800 }
2801 }
2802 if (new_in2 != NULL) {
2803 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2804 n->subsume_by(cmpN, this);
2805 if (in1->outcnt() == 0) {
2806 in1->disconnect_inputs(NULL, this);
2807 }
2808 if (in2->outcnt() == 0) {
2809 in2->disconnect_inputs(NULL, this);
2810 }
2811 }
2812 }
2813 break;
2814
2815 case Op_DecodeN:
2816 case Op_DecodeNKlass:
2817 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2818 // DecodeN could be pinned when it can't be fold into
2819 // an address expression, see the code for Op_CastPP above.
2820 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2821 break;
2822
2823 case Op_EncodeP:
2824 case Op_EncodePKlass: {
2825 Node* in1 = n->in(1);
2826 if (in1->is_DecodeNarrowPtr()) {
2827 n->subsume_by(in1->in(1), this);
2828 } else if (in1->Opcode() == Op_ConP) {
2829 const Type* t = in1->bottom_type();
2830 if (t == TypePtr::NULL_PTR) {
2831 assert(t->isa_oopptr(), "null klass?");
2832 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2833 } else if (t->isa_oopptr()) {
2834 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2835 } else if (t->isa_klassptr()) {
2836 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2837 }
2838 }
2839 if (in1->outcnt() == 0) {
2840 in1->disconnect_inputs(NULL, this);
2841 }
2842 break;
2843 }
2844
2845 case Op_Proj: {
2846 if (OptimizeStringConcat) {
2847 ProjNode* p = n->as_Proj();
2848 if (p->_is_io_use) {
2849 // Separate projections were used for the exception path which
2850 // are normally removed by a late inline. If it wasn't inlined
2851 // then they will hang around and should just be replaced with
2852 // the original one.
2853 Node* proj = NULL;
2854 // Replace with just one
2855 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2856 Node *use = i.get();
2857 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2858 proj = use;
2859 break;
2860 }
2861 }
2862 assert(proj != NULL, "must be found");
2863 p->subsume_by(proj, this);
2864 }
2865 }
2866 break;
2867 }
2868
2869 case Op_Phi:
2870 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2871 // The EncodeP optimization may create Phi with the same edges
2872 // for all paths. It is not handled well by Register Allocator.
2873 Node* unique_in = n->in(1);
2874 assert(unique_in != NULL, "");
2875 uint cnt = n->req();
2876 for (uint i = 2; i < cnt; i++) {
2877 Node* m = n->in(i);
2878 assert(m != NULL, "");
2879 if (unique_in != m)
2880 unique_in = NULL;
2881 }
2882 if (unique_in != NULL) {
2883 n->subsume_by(unique_in, this);
2884 }
2885 }
2886 break;
2887
2888 #endif
2889
2890 case Op_ModI:
2891 if (UseDivMod) {
2892 // Check if a%b and a/b both exist
2893 Node* d = n->find_similar(Op_DivI);
2894 if (d) {
2895 // Replace them with a fused divmod if supported
2896 if (Matcher::has_match_rule(Op_DivModI)) {
2897 DivModINode* divmod = DivModINode::make(this, n);
2898 d->subsume_by(divmod->div_proj(), this);
2899 n->subsume_by(divmod->mod_proj(), this);
2900 } else {
2901 // replace a%b with a-((a/b)*b)
2902 Node* mult = new (this) MulINode(d, d->in(2));
2903 Node* sub = new (this) SubINode(d->in(1), mult);
2904 n->subsume_by(sub, this);
2905 }
2906 }
2907 }
2908 break;
2909
2910 case Op_ModL:
2911 if (UseDivMod) {
2912 // Check if a%b and a/b both exist
2913 Node* d = n->find_similar(Op_DivL);
2914 if (d) {
2915 // Replace them with a fused divmod if supported
2916 if (Matcher::has_match_rule(Op_DivModL)) {
2917 DivModLNode* divmod = DivModLNode::make(this, n);
2918 d->subsume_by(divmod->div_proj(), this);
2919 n->subsume_by(divmod->mod_proj(), this);
2920 } else {
2921 // replace a%b with a-((a/b)*b)
2922 Node* mult = new (this) MulLNode(d, d->in(2));
2923 Node* sub = new (this) SubLNode(d->in(1), mult);
2924 n->subsume_by(sub, this);
2925 }
2926 }
2927 }
2928 break;
2929
2930 case Op_LoadVector:
2931 case Op_StoreVector:
2932 break;
2933
2934 case Op_PackB:
2935 case Op_PackS:
2936 case Op_PackI:
2937 case Op_PackF:
2938 case Op_PackL:
2939 case Op_PackD:
2940 if (n->req()-1 > 2) {
2941 // Replace many operand PackNodes with a binary tree for matching
2942 PackNode* p = (PackNode*) n;
2943 Node* btp = p->binary_tree_pack(this, 1, n->req());
2944 n->subsume_by(btp, this);
2945 }
2946 break;
2947 case Op_Loop:
2948 case Op_CountedLoop:
2949 if (n->as_Loop()->is_inner_loop()) {
2950 frc.inc_inner_loop_count();
2951 }
2952 break;
2953 case Op_LShiftI:
2954 case Op_RShiftI:
2955 case Op_URShiftI:
2956 case Op_LShiftL:
2957 case Op_RShiftL:
2958 case Op_URShiftL:
2959 if (Matcher::need_masked_shift_count) {
2960 // The cpu's shift instructions don't restrict the count to the
2961 // lower 5/6 bits. We need to do the masking ourselves.
2962 Node* in2 = n->in(2);
2963 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2964 const TypeInt* t = in2->find_int_type();
2965 if (t != NULL && t->is_con()) {
2966 juint shift = t->get_con();
2967 if (shift > mask) { // Unsigned cmp
2968 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2969 }
2970 } else {
2971 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2972 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2973 n->set_req(2, shift);
2974 }
2975 }
2976 if (in2->outcnt() == 0) { // Remove dead node
2977 in2->disconnect_inputs(NULL, this);
2978 }
2979 }
2980 break;
2981 case Op_MemBarStoreStore:
2982 case Op_MemBarRelease:
2983 // Break the link with AllocateNode: it is no longer useful and
2984 // confuses register allocation.
2985 if (n->req() > MemBarNode::Precedent) {
2986 n->set_req(MemBarNode::Precedent, top());
2987 }
2988 break;
2989 default:
2990 assert( !n->is_Call(), "" );
2991 assert( !n->is_Mem(), "" );
2992 break;
2993 }
2994
2995 // Collect CFG split points
2996 if (n->is_MultiBranch())
2997 frc._tests.push(n);
2998 }
2999
3000 //------------------------------final_graph_reshaping_walk---------------------
3001 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3002 // requires that the walk visits a node's inputs before visiting the node.
3003 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3004 ResourceArea *area = Thread::current()->resource_area();
3005 Unique_Node_List sfpt(area);
3006
3007 frc._visited.set(root->_idx); // first, mark node as visited
3008 uint cnt = root->req();
3009 Node *n = root;
3010 uint i = 0;
3011 while (true) {
3012 if (i < cnt) {
3013 // Place all non-visited non-null inputs onto stack
3014 Node* m = n->in(i);
3015 ++i;
3016 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3017 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3018 sfpt.push(m);
3019 cnt = m->req();
3020 nstack.push(n, i); // put on stack parent and next input's index
3021 n = m;
3022 i = 0;
3023 }
3024 } else {
3025 // Now do post-visit work
3026 final_graph_reshaping_impl( n, frc );
3027 if (nstack.is_empty())
3028 break; // finished
3029 n = nstack.node(); // Get node from stack
3030 cnt = n->req();
3031 i = nstack.index();
3032 nstack.pop(); // Shift to the next node on stack
3033 }
3034 }
3035
3036 // Skip next transformation if compressed oops are not used.
3037 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3038 (!UseCompressedOops && !UseCompressedClassPointers))
3039 return;
3040
3041 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3042 // It could be done for an uncommon traps or any safepoints/calls
3043 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3044 while (sfpt.size() > 0) {
3045 n = sfpt.pop();
3046 JVMState *jvms = n->as_SafePoint()->jvms();
3047 assert(jvms != NULL, "sanity");
3048 int start = jvms->debug_start();
3049 int end = n->req();
3050 bool is_uncommon = (n->is_CallStaticJava() &&
3051 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3052 for (int j = start; j < end; j++) {
3053 Node* in = n->in(j);
3054 if (in->is_DecodeNarrowPtr()) {
3055 bool safe_to_skip = true;
3056 if (!is_uncommon ) {
3057 // Is it safe to skip?
3058 for (uint i = 0; i < in->outcnt(); i++) {
3059 Node* u = in->raw_out(i);
3060 if (!u->is_SafePoint() ||
3061 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3062 safe_to_skip = false;
3063 }
3064 }
3065 }
3066 if (safe_to_skip) {
3067 n->set_req(j, in->in(1));
3068 }
3069 if (in->outcnt() == 0) {
3070 in->disconnect_inputs(NULL, this);
3071 }
3072 }
3073 }
3074 }
3075 }
3076
3077 //------------------------------final_graph_reshaping--------------------------
3078 // Final Graph Reshaping.
3079 //
3080 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3081 // and not commoned up and forced early. Must come after regular
3082 // optimizations to avoid GVN undoing the cloning. Clone constant
3083 // inputs to Loop Phis; these will be split by the allocator anyways.
3084 // Remove Opaque nodes.
3085 // (2) Move last-uses by commutative operations to the left input to encourage
3086 // Intel update-in-place two-address operations and better register usage
3087 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3088 // calls canonicalizing them back.
3089 // (3) Count the number of double-precision FP ops, single-precision FP ops
3090 // and call sites. On Intel, we can get correct rounding either by
3091 // forcing singles to memory (requires extra stores and loads after each
3092 // FP bytecode) or we can set a rounding mode bit (requires setting and
3093 // clearing the mode bit around call sites). The mode bit is only used
3094 // if the relative frequency of single FP ops to calls is low enough.
3095 // This is a key transform for SPEC mpeg_audio.
3096 // (4) Detect infinite loops; blobs of code reachable from above but not
3097 // below. Several of the Code_Gen algorithms fail on such code shapes,
3098 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3099 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3100 // Detection is by looking for IfNodes where only 1 projection is
3101 // reachable from below or CatchNodes missing some targets.
3102 // (5) Assert for insane oop offsets in debug mode.
3103
3104 bool Compile::final_graph_reshaping() {
3105 // an infinite loop may have been eliminated by the optimizer,
3106 // in which case the graph will be empty.
3107 if (root()->req() == 1) {
3108 record_method_not_compilable("trivial infinite loop");
3109 return true;
3110 }
3111
3112 // Expensive nodes have their control input set to prevent the GVN
3113 // from freely commoning them. There's no GVN beyond this point so
3114 // no need to keep the control input. We want the expensive nodes to
3115 // be freely moved to the least frequent code path by gcm.
3116 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3117 for (int i = 0; i < expensive_count(); i++) {
3118 _expensive_nodes->at(i)->set_req(0, NULL);
3119 }
3120
3121 Final_Reshape_Counts frc;
3122
3123 // Visit everybody reachable!
3124 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3125 Node_Stack nstack(unique() >> 1);
3126 final_graph_reshaping_walk(nstack, root(), frc);
3127
3128 // Check for unreachable (from below) code (i.e., infinite loops).
3129 for( uint i = 0; i < frc._tests.size(); i++ ) {
3130 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3131 // Get number of CFG targets.
3132 // Note that PCTables include exception targets after calls.
3133 uint required_outcnt = n->required_outcnt();
3134 if (n->outcnt() != required_outcnt) {
3135 // Check for a few special cases. Rethrow Nodes never take the
3136 // 'fall-thru' path, so expected kids is 1 less.
3137 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3138 if (n->in(0)->in(0)->is_Call()) {
3139 CallNode *call = n->in(0)->in(0)->as_Call();
3140 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3141 required_outcnt--; // Rethrow always has 1 less kid
3142 } else if (call->req() > TypeFunc::Parms &&
3143 call->is_CallDynamicJava()) {
3144 // Check for null receiver. In such case, the optimizer has
3145 // detected that the virtual call will always result in a null
3146 // pointer exception. The fall-through projection of this CatchNode
3147 // will not be populated.
3148 Node *arg0 = call->in(TypeFunc::Parms);
3149 if (arg0->is_Type() &&
3150 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3151 required_outcnt--;
3152 }
3153 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3154 call->req() > TypeFunc::Parms+1 &&
3155 call->is_CallStaticJava()) {
3156 // Check for negative array length. In such case, the optimizer has
3157 // detected that the allocation attempt will always result in an
3158 // exception. There is no fall-through projection of this CatchNode .
3159 Node *arg1 = call->in(TypeFunc::Parms+1);
3160 if (arg1->is_Type() &&
3161 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3162 required_outcnt--;
3163 }
3164 }
3165 }
3166 }
3167 // Recheck with a better notion of 'required_outcnt'
3168 if (n->outcnt() != required_outcnt) {
3169 record_method_not_compilable("malformed control flow");
3170 return true; // Not all targets reachable!
3171 }
3172 }
3173 // Check that I actually visited all kids. Unreached kids
3174 // must be infinite loops.
3175 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3176 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3177 record_method_not_compilable("infinite loop");
3178 return true; // Found unvisited kid; must be unreach
3179 }
3180 }
3181
3182 // If original bytecodes contained a mixture of floats and doubles
3183 // check if the optimizer has made it homogenous, item (3).
3184 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3185 frc.get_float_count() > 32 &&
3186 frc.get_double_count() == 0 &&
3187 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3188 set_24_bit_selection_and_mode( false, true );
3189 }
3190
3191 set_java_calls(frc.get_java_call_count());
3192 set_inner_loops(frc.get_inner_loop_count());
3193
3194 // No infinite loops, no reason to bail out.
3195 return false;
3196 }
3197
3198 //-----------------------------too_many_traps----------------------------------
3199 // Report if there are too many traps at the current method and bci.
3200 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3201 bool Compile::too_many_traps(ciMethod* method,
3202 int bci,
3203 Deoptimization::DeoptReason reason) {
3204 ciMethodData* md = method->method_data();
3205 if (md->is_empty()) {
3206 // Assume the trap has not occurred, or that it occurred only
3207 // because of a transient condition during start-up in the interpreter.
3208 return false;
3209 }
3210 if (md->has_trap_at(bci, reason) != 0) {
3211 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3212 // Also, if there are multiple reasons, or if there is no per-BCI record,
3213 // assume the worst.
3214 if (log())
3215 log()->elem("observe trap='%s' count='%d'",
3216 Deoptimization::trap_reason_name(reason),
3217 md->trap_count(reason));
3218 return true;
3219 } else {
3220 // Ignore method/bci and see if there have been too many globally.
3221 return too_many_traps(reason, md);
3222 }
3223 }
3224
3225 // Less-accurate variant which does not require a method and bci.
3226 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3227 ciMethodData* logmd) {
3228 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3229 // Too many traps globally.
3230 // Note that we use cumulative trap_count, not just md->trap_count.
3231 if (log()) {
3232 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3233 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3234 Deoptimization::trap_reason_name(reason),
3235 mcount, trap_count(reason));
3236 }
3237 return true;
3238 } else {
3239 // The coast is clear.
3240 return false;
3241 }
3242 }
3243
3244 //--------------------------too_many_recompiles--------------------------------
3245 // Report if there are too many recompiles at the current method and bci.
3246 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3247 // Is not eager to return true, since this will cause the compiler to use
3248 // Action_none for a trap point, to avoid too many recompilations.
3249 bool Compile::too_many_recompiles(ciMethod* method,
3250 int bci,
3251 Deoptimization::DeoptReason reason) {
3252 ciMethodData* md = method->method_data();
3253 if (md->is_empty()) {
3254 // Assume the trap has not occurred, or that it occurred only
3255 // because of a transient condition during start-up in the interpreter.
3256 return false;
3257 }
3258 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3259 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3260 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3261 Deoptimization::DeoptReason per_bc_reason
3262 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3263 if ((per_bc_reason == Deoptimization::Reason_none
3264 || md->has_trap_at(bci, reason) != 0)
3265 // The trap frequency measure we care about is the recompile count:
3266 && md->trap_recompiled_at(bci)
3267 && md->overflow_recompile_count() >= bc_cutoff) {
3268 // Do not emit a trap here if it has already caused recompilations.
3269 // Also, if there are multiple reasons, or if there is no per-BCI record,
3270 // assume the worst.
3271 if (log())
3272 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3273 Deoptimization::trap_reason_name(reason),
3274 md->trap_count(reason),
3275 md->overflow_recompile_count());
3276 return true;
3277 } else if (trap_count(reason) != 0
3278 && decompile_count() >= m_cutoff) {
3279 // Too many recompiles globally, and we have seen this sort of trap.
3280 // Use cumulative decompile_count, not just md->decompile_count.
3281 if (log())
3282 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3283 Deoptimization::trap_reason_name(reason),
3284 md->trap_count(reason), trap_count(reason),
3285 md->decompile_count(), decompile_count());
3286 return true;
3287 } else {
3288 // The coast is clear.
3289 return false;
3290 }
3291 }
3292
3293
3294 #ifndef PRODUCT
3295 //------------------------------verify_graph_edges---------------------------
3296 // Walk the Graph and verify that there is a one-to-one correspondence
3297 // between Use-Def edges and Def-Use edges in the graph.
3298 void Compile::verify_graph_edges(bool no_dead_code) {
3299 if (VerifyGraphEdges) {
3300 ResourceArea *area = Thread::current()->resource_area();
3301 Unique_Node_List visited(area);
3302 // Call recursive graph walk to check edges
3303 _root->verify_edges(visited);
3304 if (no_dead_code) {
3305 // Now make sure that no visited node is used by an unvisited node.
3306 bool dead_nodes = 0;
3307 Unique_Node_List checked(area);
3308 while (visited.size() > 0) {
3309 Node* n = visited.pop();
3310 checked.push(n);
3311 for (uint i = 0; i < n->outcnt(); i++) {
3312 Node* use = n->raw_out(i);
3313 if (checked.member(use)) continue; // already checked
3314 if (visited.member(use)) continue; // already in the graph
3315 if (use->is_Con()) continue; // a dead ConNode is OK
3316 // At this point, we have found a dead node which is DU-reachable.
3317 if (dead_nodes++ == 0)
3318 tty->print_cr("*** Dead nodes reachable via DU edges:");
3319 use->dump(2);
3320 tty->print_cr("---");
3321 checked.push(use); // No repeats; pretend it is now checked.
3322 }
3323 }
3324 assert(dead_nodes == 0, "using nodes must be reachable from root");
3325 }
3326 }
3327 }
3328 #endif
3329
3330 // The Compile object keeps track of failure reasons separately from the ciEnv.
3331 // This is required because there is not quite a 1-1 relation between the
3332 // ciEnv and its compilation task and the Compile object. Note that one
3333 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3334 // to backtrack and retry without subsuming loads. Other than this backtracking
3335 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3336 // by the logic in C2Compiler.
3337 void Compile::record_failure(const char* reason) {
3338 if (log() != NULL) {
3339 log()->elem("failure reason='%s' phase='compile'", reason);
3340 }
3341 if (_failure_reason == NULL) {
3342 // Record the first failure reason.
3343 _failure_reason = reason;
3344 }
3345
3346 EventCompilerFailure event;
3347 if (event.should_commit()) {
3348 event.set_compileID(Compile::compile_id());
3349 event.set_failure(reason);
3350 event.commit();
3351 }
3352
3353 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3354 C->print_method(PHASE_FAILURE);
3355 }
3356 _root = NULL; // flush the graph, too
3357 }
3358
3359 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3360 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3361 _phase_name(name), _dolog(dolog)
3362 {
3363 if (dolog) {
3364 C = Compile::current();
3365 _log = C->log();
3366 } else {
3367 C = NULL;
3368 _log = NULL;
3369 }
3370 if (_log != NULL) {
3371 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3372 _log->stamp();
3373 _log->end_head();
3374 }
3375 }
3376
3377 Compile::TracePhase::~TracePhase() {
3378
3379 C = Compile::current();
3380 if (_dolog) {
3381 _log = C->log();
3382 } else {
3383 _log = NULL;
3384 }
3385
3386 #ifdef ASSERT
3387 if (PrintIdealNodeCount) {
3388 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3389 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3390 }
3391
3392 if (VerifyIdealNodeCount) {
3393 Compile::current()->print_missing_nodes();
3394 }
3395 #endif
3396
3397 if (_log != NULL) {
3398 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3399 }
3400 }
3401
3402 //=============================================================================
3403 // Two Constant's are equal when the type and the value are equal.
3404 bool Compile::Constant::operator==(const Constant& other) {
3405 if (type() != other.type() ) return false;
3406 if (can_be_reused() != other.can_be_reused()) return false;
3407 // For floating point values we compare the bit pattern.
3408 switch (type()) {
3409 case T_FLOAT: return (_v._value.i == other._v._value.i);
3410 case T_LONG:
3411 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3412 case T_OBJECT:
3413 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3414 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3415 case T_METADATA: return (_v._metadata == other._v._metadata);
3416 default: ShouldNotReachHere();
3417 }
3418 return false;
3419 }
3420
3421 static int type_to_size_in_bytes(BasicType t) {
3422 switch (t) {
3423 case T_LONG: return sizeof(jlong );
3424 case T_FLOAT: return sizeof(jfloat );
3425 case T_DOUBLE: return sizeof(jdouble);
3426 case T_METADATA: return sizeof(Metadata*);
3427 // We use T_VOID as marker for jump-table entries (labels) which
3428 // need an internal word relocation.
3429 case T_VOID:
3430 case T_ADDRESS:
3431 case T_OBJECT: return sizeof(jobject);
3432 }
3433
3434 ShouldNotReachHere();
3435 return -1;
3436 }
3437
3438 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3439 // sort descending
3440 if (a->freq() > b->freq()) return -1;
3441 if (a->freq() < b->freq()) return 1;
3442 return 0;
3443 }
3444
3445 void Compile::ConstantTable::calculate_offsets_and_size() {
3446 // First, sort the array by frequencies.
3447 _constants.sort(qsort_comparator);
3448
3449 #ifdef ASSERT
3450 // Make sure all jump-table entries were sorted to the end of the
3451 // array (they have a negative frequency).
3452 bool found_void = false;
3453 for (int i = 0; i < _constants.length(); i++) {
3454 Constant con = _constants.at(i);
3455 if (con.type() == T_VOID)
3456 found_void = true; // jump-tables
3457 else
3458 assert(!found_void, "wrong sorting");
3459 }
3460 #endif
3461
3462 int offset = 0;
3463 for (int i = 0; i < _constants.length(); i++) {
3464 Constant* con = _constants.adr_at(i);
3465
3466 // Align offset for type.
3467 int typesize = type_to_size_in_bytes(con->type());
3468 offset = align_size_up(offset, typesize);
3469 con->set_offset(offset); // set constant's offset
3470
3471 if (con->type() == T_VOID) {
3472 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3473 offset = offset + typesize * n->outcnt(); // expand jump-table
3474 } else {
3475 offset = offset + typesize;
3476 }
3477 }
3478
3479 // Align size up to the next section start (which is insts; see
3480 // CodeBuffer::align_at_start).
3481 assert(_size == -1, "already set?");
3482 _size = align_size_up(offset, CodeEntryAlignment);
3483 }
3484
3485 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3486 MacroAssembler _masm(&cb);
3487 for (int i = 0; i < _constants.length(); i++) {
3488 Constant con = _constants.at(i);
3489 address constant_addr;
3490 switch (con.type()) {
3491 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3492 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3493 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3494 case T_OBJECT: {
3495 jobject obj = con.get_jobject();
3496 int oop_index = _masm.oop_recorder()->find_index(obj);
3497 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3498 break;
3499 }
3500 case T_ADDRESS: {
3501 address addr = (address) con.get_jobject();
3502 constant_addr = _masm.address_constant(addr);
3503 break;
3504 }
3505 // We use T_VOID as marker for jump-table entries (labels) which
3506 // need an internal word relocation.
3507 case T_VOID: {
3508 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3509 // Fill the jump-table with a dummy word. The real value is
3510 // filled in later in fill_jump_table.
3511 address dummy = (address) n;
3512 constant_addr = _masm.address_constant(dummy);
3513 // Expand jump-table
3514 for (uint i = 1; i < n->outcnt(); i++) {
3515 address temp_addr = _masm.address_constant(dummy + i);
3516 assert(temp_addr, "consts section too small");
3517 }
3518 break;
3519 }
3520 case T_METADATA: {
3521 Metadata* obj = con.get_metadata();
3522 int metadata_index = _masm.oop_recorder()->find_index(obj);
3523 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3524 break;
3525 }
3526 default: ShouldNotReachHere();
3527 }
3528 assert(constant_addr, "consts section too small");
3529 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3530 }
3531 }
3532
3533 int Compile::ConstantTable::find_offset(Constant& con) const {
3534 int idx = _constants.find(con);
3535 assert(idx != -1, "constant must be in constant table");
3536 int offset = _constants.at(idx).offset();
3537 assert(offset != -1, "constant table not emitted yet?");
3538 return offset;
3539 }
3540
3541 void Compile::ConstantTable::add(Constant& con) {
3542 if (con.can_be_reused()) {
3543 int idx = _constants.find(con);
3544 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3545 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3546 return;
3547 }
3548 }
3549 (void) _constants.append(con);
3550 }
3551
3552 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3553 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3554 Constant con(type, value, b->_freq);
3555 add(con);
3556 return con;
3557 }
3558
3559 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3560 Constant con(metadata);
3561 add(con);
3562 return con;
3563 }
3564
3565 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3566 jvalue value;
3567 BasicType type = oper->type()->basic_type();
3568 switch (type) {
3569 case T_LONG: value.j = oper->constantL(); break;
3570 case T_FLOAT: value.f = oper->constantF(); break;
3571 case T_DOUBLE: value.d = oper->constantD(); break;
3572 case T_OBJECT:
3573 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3574 case T_METADATA: return add((Metadata*)oper->constant()); break;
3575 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3576 }
3577 return add(n, type, value);
3578 }
3579
3580 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3581 jvalue value;
3582 // We can use the node pointer here to identify the right jump-table
3583 // as this method is called from Compile::Fill_buffer right before
3584 // the MachNodes are emitted and the jump-table is filled (means the
3585 // MachNode pointers do not change anymore).
3586 value.l = (jobject) n;
3587 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3588 add(con);
3589 return con;
3590 }
3591
3592 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3593 // If called from Compile::scratch_emit_size do nothing.
3594 if (Compile::current()->in_scratch_emit_size()) return;
3595
3596 assert(labels.is_nonempty(), "must be");
3597 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3598
3599 // Since MachConstantNode::constant_offset() also contains
3600 // table_base_offset() we need to subtract the table_base_offset()
3601 // to get the plain offset into the constant table.
3602 int offset = n->constant_offset() - table_base_offset();
3603
3604 MacroAssembler _masm(&cb);
3605 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3606
3607 for (uint i = 0; i < n->outcnt(); i++) {
3608 address* constant_addr = &jump_table_base[i];
3609 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3610 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3611 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3612 }
3613 }
3614
3615 void Compile::dump_inlining() {
3616 if (print_inlining() || print_intrinsics()) {
3617 // Print inlining message for candidates that we couldn't inline
3618 // for lack of space or non constant receiver
3619 for (int i = 0; i < _late_inlines.length(); i++) {
3620 CallGenerator* cg = _late_inlines.at(i);
3621 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3622 }
3623 Unique_Node_List useful;
3624 useful.push(root());
3625 for (uint next = 0; next < useful.size(); ++next) {
3626 Node* n = useful.at(next);
3627 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3628 CallNode* call = n->as_Call();
3629 CallGenerator* cg = call->generator();
3630 cg->print_inlining_late("receiver not constant");
3631 }
3632 uint max = n->len();
3633 for ( uint i = 0; i < max; ++i ) {
3634 Node *m = n->in(i);
3635 if ( m == NULL ) continue;
3636 useful.push(m);
3637 }
3638 }
3639 for (int i = 0; i < _print_inlining_list->length(); i++) {
3640 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3641 }
3642 }
3643 }
3644
3645 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3646 if (n1->Opcode() < n2->Opcode()) return -1;
3647 else if (n1->Opcode() > n2->Opcode()) return 1;
3648
3649 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
3650 for (uint i = 1; i < n1->req(); i++) {
3651 if (n1->in(i) < n2->in(i)) return -1;
3652 else if (n1->in(i) > n2->in(i)) return 1;
3653 }
3654
3655 return 0;
3656 }
3657
3658 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3659 Node* n1 = *n1p;
3660 Node* n2 = *n2p;
3661
3662 return cmp_expensive_nodes(n1, n2);
3663 }
3664
3665 void Compile::sort_expensive_nodes() {
3666 if (!expensive_nodes_sorted()) {
3667 _expensive_nodes->sort(cmp_expensive_nodes);
3668 }
3669 }
3670
3671 bool Compile::expensive_nodes_sorted() const {
3672 for (int i = 1; i < _expensive_nodes->length(); i++) {
3673 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3674 return false;
3675 }
3676 }
3677 return true;
3678 }
3679
3680 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3681 if (_expensive_nodes->length() == 0) {
3682 return false;
3683 }
3684
3685 assert(OptimizeExpensiveOps, "optimization off?");
3686
3687 // Take this opportunity to remove dead nodes from the list
3688 int j = 0;
3689 for (int i = 0; i < _expensive_nodes->length(); i++) {
3690 Node* n = _expensive_nodes->at(i);
3691 if (!n->is_unreachable(igvn)) {
3692 assert(n->is_expensive(), "should be expensive");
3693 _expensive_nodes->at_put(j, n);
3694 j++;
3695 }
3696 }
3697 _expensive_nodes->trunc_to(j);
3698
3699 // Then sort the list so that similar nodes are next to each other
3700 // and check for at least two nodes of identical kind with same data
3701 // inputs.
3702 sort_expensive_nodes();
3703
3704 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3705 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3706 return true;
3707 }
3708 }
3709
3710 return false;
3711 }
3712
3713 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3714 if (_expensive_nodes->length() == 0) {
3715 return;
3716 }
3717
3718 assert(OptimizeExpensiveOps, "optimization off?");
3719
3720 // Sort to bring similar nodes next to each other and clear the
3721 // control input of nodes for which there's only a single copy.
3722 sort_expensive_nodes();
3723
3724 int j = 0;
3725 int identical = 0;
3726 int i = 0;
3727 for (; i < _expensive_nodes->length()-1; i++) {
3728 assert(j <= i, "can't write beyond current index");
3729 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3730 identical++;
3731 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3732 continue;
3733 }
3734 if (identical > 0) {
3735 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3736 identical = 0;
3737 } else {
3738 Node* n = _expensive_nodes->at(i);
3739 igvn.hash_delete(n);
3740 n->set_req(0, NULL);
3741 igvn.hash_insert(n);
3742 }
3743 }
3744 if (identical > 0) {
3745 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3746 } else if (_expensive_nodes->length() >= 1) {
3747 Node* n = _expensive_nodes->at(i);
3748 igvn.hash_delete(n);
3749 n->set_req(0, NULL);
3750 igvn.hash_insert(n);
3751 }
3752 _expensive_nodes->trunc_to(j);
3753 }
3754
3755 void Compile::add_expensive_node(Node * n) {
3756 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3757 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3758 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3759 if (OptimizeExpensiveOps) {
3760 _expensive_nodes->append(n);
3761 } else {
3762 // Clear control input and let IGVN optimize expensive nodes if
3763 // OptimizeExpensiveOps is off.
3764 n->set_req(0, NULL);
3765 }
3766 }
3767
3768 // Auxiliary method to support randomized stressing/fuzzing.
3769 //
3770 // This method can be called the arbitrary number of times, with current count
3771 // as the argument. The logic allows selecting a single candidate from the
3772 // running list of candidates as follows:
3773 // int count = 0;
3774 // Cand* selected = null;
3775 // while(cand = cand->next()) {
3776 // if (randomized_select(++count)) {
3777 // selected = cand;
3778 // }
3779 // }
3780 //
3781 // Including count equalizes the chances any candidate is "selected".
3782 // This is useful when we don't have the complete list of candidates to choose
3783 // from uniformly. In this case, we need to adjust the randomicity of the
3784 // selection, or else we will end up biasing the selection towards the latter
3785 // candidates.
3786 //
3787 // Quick back-envelope calculation shows that for the list of n candidates
3788 // the equal probability for the candidate to persist as "best" can be
3789 // achieved by replacing it with "next" k-th candidate with the probability
3790 // of 1/k. It can be easily shown that by the end of the run, the
3791 // probability for any candidate is converged to 1/n, thus giving the
3792 // uniform distribution among all the candidates.
3793 //
3794 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3795 #define RANDOMIZED_DOMAIN_POW 29
3796 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3797 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3798 bool Compile::randomized_select(int count) {
3799 assert(count > 0, "only positive");
3800 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3801 }