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/mathexactnode.hpp"
51 #include "opto/memnode.hpp"
52 #include "opto/mulnode.hpp"
53 #include "opto/node.hpp"
54 #include "opto/opcodes.hpp"
55 #include "opto/output.hpp"
56 #include "opto/parse.hpp"
57 #include "opto/phaseX.hpp"
58 #include "opto/rootnode.hpp"
59 #include "opto/runtime.hpp"
60 #include "opto/stringopts.hpp"
61 #include "opto/type.hpp"
62 #include "opto/vectornode.hpp"
63 #include "runtime/arguments.hpp"
64 #include "runtime/signature.hpp"
65 #include "runtime/stubRoutines.hpp"
66 #include "runtime/timer.hpp"
67 #include "trace/tracing.hpp"
68 #include "utilities/copy.hpp"
69 #ifdef TARGET_ARCH_MODEL_x86_32
70 # include "adfiles/ad_x86_32.hpp"
71 #endif
72 #ifdef TARGET_ARCH_MODEL_x86_64
73 # include "adfiles/ad_x86_64.hpp"
74 #endif
75 #ifdef TARGET_ARCH_MODEL_sparc
76 # include "adfiles/ad_sparc.hpp"
77 #endif
78 #ifdef TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #endif
81 #ifdef TARGET_ARCH_MODEL_arm
82 # include "adfiles/ad_arm.hpp"
83 #endif
84 #ifdef TARGET_ARCH_MODEL_ppc_32
85 # include "adfiles/ad_ppc_32.hpp"
86 #endif
87 #ifdef TARGET_ARCH_MODEL_ppc_64
88 # include "adfiles/ad_ppc_64.hpp"
89 #endif
90
91
92 // -------------------- Compile::mach_constant_base_node -----------------------
93 // Constant table base node singleton.
94 MachConstantBaseNode* Compile::mach_constant_base_node() {
95 if (_mach_constant_base_node == NULL) {
96 _mach_constant_base_node = new (C) MachConstantBaseNode();
97 _mach_constant_base_node->add_req(C->root());
98 }
99 return _mach_constant_base_node;
100 }
101
102
103 /// Support for intrinsics.
104
105 // Return the index at which m must be inserted (or already exists).
106 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
107 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
108 #ifdef ASSERT
109 for (int i = 1; i < _intrinsics->length(); i++) {
110 CallGenerator* cg1 = _intrinsics->at(i-1);
111 CallGenerator* cg2 = _intrinsics->at(i);
112 assert(cg1->method() != cg2->method()
113 ? cg1->method() < cg2->method()
114 : cg1->is_virtual() < cg2->is_virtual(),
115 "compiler intrinsics list must stay sorted");
116 }
117 #endif
118 // Binary search sorted list, in decreasing intervals [lo, hi].
119 int lo = 0, hi = _intrinsics->length()-1;
120 while (lo <= hi) {
121 int mid = (uint)(hi + lo) / 2;
122 ciMethod* mid_m = _intrinsics->at(mid)->method();
123 if (m < mid_m) {
124 hi = mid-1;
125 } else if (m > mid_m) {
126 lo = mid+1;
127 } else {
128 // look at minor sort key
129 bool mid_virt = _intrinsics->at(mid)->is_virtual();
130 if (is_virtual < mid_virt) {
131 hi = mid-1;
132 } else if (is_virtual > mid_virt) {
133 lo = mid+1;
134 } else {
135 return mid; // exact match
136 }
137 }
138 }
139 return lo; // inexact match
140 }
141
142 void Compile::register_intrinsic(CallGenerator* cg) {
143 if (_intrinsics == NULL) {
144 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
145 }
146 // This code is stolen from ciObjectFactory::insert.
147 // Really, GrowableArray should have methods for
148 // insert_at, remove_at, and binary_search.
149 int len = _intrinsics->length();
150 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
151 if (index == len) {
152 _intrinsics->append(cg);
153 } else {
154 #ifdef ASSERT
155 CallGenerator* oldcg = _intrinsics->at(index);
156 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
157 #endif
158 _intrinsics->append(_intrinsics->at(len-1));
159 int pos;
160 for (pos = len-2; pos >= index; pos--) {
161 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
162 }
163 _intrinsics->at_put(index, cg);
164 }
165 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
166 }
167
168 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
169 assert(m->is_loaded(), "don't try this on unloaded methods");
170 if (_intrinsics != NULL) {
171 int index = intrinsic_insertion_index(m, is_virtual);
172 if (index < _intrinsics->length()
173 && _intrinsics->at(index)->method() == m
174 && _intrinsics->at(index)->is_virtual() == is_virtual) {
175 return _intrinsics->at(index);
176 }
177 }
178 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
179 if (m->intrinsic_id() != vmIntrinsics::_none &&
180 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
181 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
182 if (cg != NULL) {
183 // Save it for next time:
184 register_intrinsic(cg);
185 return cg;
186 } else {
187 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
188 }
189 }
190 return NULL;
191 }
192
193 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
194 // in library_call.cpp.
195
196
197 #ifndef PRODUCT
198 // statistics gathering...
199
200 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
201 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
202
203 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
204 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
205 int oflags = _intrinsic_hist_flags[id];
206 assert(flags != 0, "what happened?");
207 if (is_virtual) {
208 flags |= _intrinsic_virtual;
209 }
210 bool changed = (flags != oflags);
211 if ((flags & _intrinsic_worked) != 0) {
212 juint count = (_intrinsic_hist_count[id] += 1);
213 if (count == 1) {
214 changed = true; // first time
215 }
216 // increment the overall count also:
217 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
218 }
219 if (changed) {
220 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
221 // Something changed about the intrinsic's virtuality.
222 if ((flags & _intrinsic_virtual) != 0) {
223 // This is the first use of this intrinsic as a virtual call.
224 if (oflags != 0) {
225 // We already saw it as a non-virtual, so note both cases.
226 flags |= _intrinsic_both;
227 }
228 } else if ((oflags & _intrinsic_both) == 0) {
229 // This is the first use of this intrinsic as a non-virtual
230 flags |= _intrinsic_both;
231 }
232 }
233 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
234 }
235 // update the overall flags also:
236 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
237 return changed;
238 }
239
240 static char* format_flags(int flags, char* buf) {
241 buf[0] = 0;
242 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
243 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
244 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
245 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
246 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
247 if (buf[0] == 0) strcat(buf, ",");
248 assert(buf[0] == ',', "must be");
249 return &buf[1];
250 }
251
252 void Compile::print_intrinsic_statistics() {
253 char flagsbuf[100];
254 ttyLocker ttyl;
255 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
256 tty->print_cr("Compiler intrinsic usage:");
257 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
258 if (total == 0) total = 1; // avoid div0 in case of no successes
259 #define PRINT_STAT_LINE(name, c, f) \
260 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
261 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
262 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
263 int flags = _intrinsic_hist_flags[id];
264 juint count = _intrinsic_hist_count[id];
265 if ((flags | count) != 0) {
266 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
267 }
268 }
269 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
270 if (xtty != NULL) xtty->tail("statistics");
271 }
272
273 void Compile::print_statistics() {
274 { ttyLocker ttyl;
275 if (xtty != NULL) xtty->head("statistics type='opto'");
276 Parse::print_statistics();
277 PhaseCCP::print_statistics();
278 PhaseRegAlloc::print_statistics();
279 Scheduling::print_statistics();
280 PhasePeephole::print_statistics();
281 PhaseIdealLoop::print_statistics();
282 if (xtty != NULL) xtty->tail("statistics");
283 }
284 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
285 // put this under its own <statistics> element.
286 print_intrinsic_statistics();
287 }
288 }
289 #endif //PRODUCT
290
291 // Support for bundling info
292 Bundle* Compile::node_bundling(const Node *n) {
293 assert(valid_bundle_info(n), "oob");
294 return &_node_bundling_base[n->_idx];
295 }
296
297 bool Compile::valid_bundle_info(const Node *n) {
298 return (_node_bundling_limit > n->_idx);
299 }
300
301
302 void Compile::gvn_replace_by(Node* n, Node* nn) {
303 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
304 Node* use = n->last_out(i);
305 bool is_in_table = initial_gvn()->hash_delete(use);
306 uint uses_found = 0;
307 for (uint j = 0; j < use->len(); j++) {
308 if (use->in(j) == n) {
309 if (j < use->req())
310 use->set_req(j, nn);
311 else
312 use->set_prec(j, nn);
313 uses_found++;
314 }
315 }
316 if (is_in_table) {
317 // reinsert into table
318 initial_gvn()->hash_find_insert(use);
319 }
320 record_for_igvn(use);
321 i -= uses_found; // we deleted 1 or more copies of this edge
322 }
323 }
324
325
326 static inline bool not_a_node(const Node* n) {
327 if (n == NULL) return true;
328 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
329 if (*(address*)n == badAddress) return true; // kill by Node::destruct
330 return false;
331 }
332
333 // Identify all nodes that are reachable from below, useful.
334 // Use breadth-first pass that records state in a Unique_Node_List,
335 // recursive traversal is slower.
336 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
337 int estimated_worklist_size = unique();
338 useful.map( estimated_worklist_size, NULL ); // preallocate space
339
340 // Initialize worklist
341 if (root() != NULL) { useful.push(root()); }
342 // If 'top' is cached, declare it useful to preserve cached node
343 if( cached_top_node() ) { useful.push(cached_top_node()); }
344
345 // Push all useful nodes onto the list, breadthfirst
346 for( uint next = 0; next < useful.size(); ++next ) {
347 assert( next < unique(), "Unique useful nodes < total nodes");
348 Node *n = useful.at(next);
349 uint max = n->len();
350 for( uint i = 0; i < max; ++i ) {
351 Node *m = n->in(i);
352 if (not_a_node(m)) continue;
353 useful.push(m);
354 }
355 }
356 }
357
358 // Update dead_node_list with any missing dead nodes using useful
359 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
360 void Compile::update_dead_node_list(Unique_Node_List &useful) {
361 uint max_idx = unique();
362 VectorSet& useful_node_set = useful.member_set();
363
364 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
365 // If node with index node_idx is not in useful set,
366 // mark it as dead in dead node list.
367 if (! useful_node_set.test(node_idx) ) {
368 record_dead_node(node_idx);
369 }
370 }
371 }
372
373 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
374 int shift = 0;
375 for (int i = 0; i < inlines->length(); i++) {
376 CallGenerator* cg = inlines->at(i);
377 CallNode* call = cg->call_node();
378 if (shift > 0) {
379 inlines->at_put(i-shift, cg);
380 }
381 if (!useful.member(call)) {
382 shift++;
383 }
384 }
385 inlines->trunc_to(inlines->length()-shift);
386 }
387
388 // Disconnect all useless nodes by disconnecting those at the boundary.
389 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
390 uint next = 0;
391 while (next < useful.size()) {
392 Node *n = useful.at(next++);
393 // Use raw traversal of out edges since this code removes out edges
394 int max = n->outcnt();
395 for (int j = 0; j < max; ++j) {
396 Node* child = n->raw_out(j);
397 if (! useful.member(child)) {
398 assert(!child->is_top() || child != top(),
399 "If top is cached in Compile object it is in useful list");
400 // Only need to remove this out-edge to the useless node
401 n->raw_del_out(j);
402 --j;
403 --max;
404 }
405 }
406 if (n->outcnt() == 1 && n->has_special_unique_user()) {
407 record_for_igvn(n->unique_out());
408 }
409 }
410 // Remove useless macro and predicate opaq nodes
411 for (int i = C->macro_count()-1; i >= 0; i--) {
412 Node* n = C->macro_node(i);
413 if (!useful.member(n)) {
414 remove_macro_node(n);
415 }
416 }
417 // Remove useless expensive node
418 for (int i = C->expensive_count()-1; i >= 0; i--) {
419 Node* n = C->expensive_node(i);
420 if (!useful.member(n)) {
421 remove_expensive_node(n);
422 }
423 }
424 // clean up the late inline lists
425 remove_useless_late_inlines(&_string_late_inlines, useful);
426 remove_useless_late_inlines(&_boxing_late_inlines, useful);
427 remove_useless_late_inlines(&_late_inlines, useful);
428 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
429 }
430
431 //------------------------------frame_size_in_words-----------------------------
432 // frame_slots in units of words
433 int Compile::frame_size_in_words() const {
434 // shift is 0 in LP32 and 1 in LP64
435 const int shift = (LogBytesPerWord - LogBytesPerInt);
436 int words = _frame_slots >> shift;
437 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
438 return words;
439 }
440
441 // ============================================================================
442 //------------------------------CompileWrapper---------------------------------
443 class CompileWrapper : public StackObj {
444 Compile *const _compile;
445 public:
446 CompileWrapper(Compile* compile);
447
448 ~CompileWrapper();
449 };
450
451 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
452 // the Compile* pointer is stored in the current ciEnv:
453 ciEnv* env = compile->env();
454 assert(env == ciEnv::current(), "must already be a ciEnv active");
455 assert(env->compiler_data() == NULL, "compile already active?");
456 env->set_compiler_data(compile);
457 assert(compile == Compile::current(), "sanity");
458
459 compile->set_type_dict(NULL);
460 compile->set_type_hwm(NULL);
461 compile->set_type_last_size(0);
462 compile->set_last_tf(NULL, NULL);
463 compile->set_indexSet_arena(NULL);
464 compile->set_indexSet_free_block_list(NULL);
465 compile->init_type_arena();
466 Type::Initialize(compile);
467 _compile->set_scratch_buffer_blob(NULL);
468 _compile->begin_method();
469 }
470 CompileWrapper::~CompileWrapper() {
471 _compile->end_method();
472 if (_compile->scratch_buffer_blob() != NULL)
473 BufferBlob::free(_compile->scratch_buffer_blob());
474 _compile->env()->set_compiler_data(NULL);
475 }
476
477
478 //----------------------------print_compile_messages---------------------------
479 void Compile::print_compile_messages() {
480 #ifndef PRODUCT
481 // Check if recompiling
482 if (_subsume_loads == false && PrintOpto) {
483 // Recompiling without allowing machine instructions to subsume loads
484 tty->print_cr("*********************************************************");
485 tty->print_cr("** Bailout: Recompile without subsuming loads **");
486 tty->print_cr("*********************************************************");
487 }
488 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
489 // Recompiling without escape analysis
490 tty->print_cr("*********************************************************");
491 tty->print_cr("** Bailout: Recompile without escape analysis **");
492 tty->print_cr("*********************************************************");
493 }
494 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
495 // Recompiling without boxing elimination
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without boxing elimination **");
498 tty->print_cr("*********************************************************");
499 }
500 if (env()->break_at_compile()) {
501 // Open the debugger when compiling this method.
502 tty->print("### Breaking when compiling: ");
503 method()->print_short_name();
504 tty->cr();
505 BREAKPOINT;
506 }
507
508 if( PrintOpto ) {
509 if (is_osr_compilation()) {
510 tty->print("[OSR]%3d", _compile_id);
511 } else {
512 tty->print("%3d", _compile_id);
513 }
514 }
515 #endif
516 }
517
518
519 //-----------------------init_scratch_buffer_blob------------------------------
520 // Construct a temporary BufferBlob and cache it for this compile.
521 void Compile::init_scratch_buffer_blob(int const_size) {
522 // If there is already a scratch buffer blob allocated and the
523 // constant section is big enough, use it. Otherwise free the
524 // current and allocate a new one.
525 BufferBlob* blob = scratch_buffer_blob();
526 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
527 // Use the current blob.
528 } else {
529 if (blob != NULL) {
530 BufferBlob::free(blob);
531 }
532
533 ResourceMark rm;
534 _scratch_const_size = const_size;
535 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
536 blob = BufferBlob::create("Compile::scratch_buffer", size);
537 // Record the buffer blob for next time.
538 set_scratch_buffer_blob(blob);
539 // Have we run out of code space?
540 if (scratch_buffer_blob() == NULL) {
541 // Let CompilerBroker disable further compilations.
542 record_failure("Not enough space for scratch buffer in CodeCache");
543 return;
544 }
545 }
546
547 // Initialize the relocation buffers
548 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
549 set_scratch_locs_memory(locs_buf);
550 }
551
552
553 //-----------------------scratch_emit_size-------------------------------------
554 // Helper function that computes size by emitting code
555 uint Compile::scratch_emit_size(const Node* n) {
556 // Start scratch_emit_size section.
557 set_in_scratch_emit_size(true);
558
559 // Emit into a trash buffer and count bytes emitted.
560 // This is a pretty expensive way to compute a size,
561 // but it works well enough if seldom used.
562 // All common fixed-size instructions are given a size
563 // method by the AD file.
564 // Note that the scratch buffer blob and locs memory are
565 // allocated at the beginning of the compile task, and
566 // may be shared by several calls to scratch_emit_size.
567 // The allocation of the scratch buffer blob is particularly
568 // expensive, since it has to grab the code cache lock.
569 BufferBlob* blob = this->scratch_buffer_blob();
570 assert(blob != NULL, "Initialize BufferBlob at start");
571 assert(blob->size() > MAX_inst_size, "sanity");
572 relocInfo* locs_buf = scratch_locs_memory();
573 address blob_begin = blob->content_begin();
574 address blob_end = (address)locs_buf;
575 assert(blob->content_contains(blob_end), "sanity");
576 CodeBuffer buf(blob_begin, blob_end - blob_begin);
577 buf.initialize_consts_size(_scratch_const_size);
578 buf.initialize_stubs_size(MAX_stubs_size);
579 assert(locs_buf != NULL, "sanity");
580 int lsize = MAX_locs_size / 3;
581 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
582 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
583 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
584
585 // Do the emission.
586
587 Label fakeL; // Fake label for branch instructions.
588 Label* saveL = NULL;
589 uint save_bnum = 0;
590 bool is_branch = n->is_MachBranch();
591 if (is_branch) {
592 MacroAssembler masm(&buf);
593 masm.bind(fakeL);
594 n->as_MachBranch()->save_label(&saveL, &save_bnum);
595 n->as_MachBranch()->label_set(&fakeL, 0);
596 }
597 n->emit(buf, this->regalloc());
598 if (is_branch) // Restore label.
599 n->as_MachBranch()->label_set(saveL, save_bnum);
600
601 // End scratch_emit_size section.
602 set_in_scratch_emit_size(false);
603
604 return buf.insts_size();
605 }
606
607
608 // ============================================================================
609 //------------------------------Compile standard-------------------------------
610 debug_only( int Compile::_debug_idx = 100000; )
611
612 // Compile a method. entry_bci is -1 for normal compilations and indicates
613 // the continuation bci for on stack replacement.
614
615
616 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
617 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
618 : Phase(Compiler),
619 _env(ci_env),
620 _log(ci_env->log()),
621 _compile_id(ci_env->compile_id()),
622 _save_argument_registers(false),
623 _stub_name(NULL),
624 _stub_function(NULL),
625 _stub_entry_point(NULL),
626 _method(target),
627 _entry_bci(osr_bci),
628 _initial_gvn(NULL),
629 _for_igvn(NULL),
630 _warm_calls(NULL),
631 _subsume_loads(subsume_loads),
632 _do_escape_analysis(do_escape_analysis),
633 _eliminate_boxing(eliminate_boxing),
634 _failure_reason(NULL),
635 _code_buffer("Compile::Fill_buffer"),
636 _orig_pc_slot(0),
637 _orig_pc_slot_offset_in_bytes(0),
638 _has_method_handle_invokes(false),
639 _mach_constant_base_node(NULL),
640 _node_bundling_limit(0),
641 _node_bundling_base(NULL),
642 _java_calls(0),
643 _inner_loops(0),
644 _scratch_const_size(-1),
645 _in_scratch_emit_size(false),
646 _dead_node_list(comp_arena()),
647 _dead_node_count(0),
648 #ifndef PRODUCT
649 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
650 _in_dump_cnt(0),
651 _printer(IdealGraphPrinter::printer()),
652 #endif
653 _congraph(NULL),
654 _late_inlines(comp_arena(), 2, 0, NULL),
655 _string_late_inlines(comp_arena(), 2, 0, NULL),
656 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
657 _late_inlines_pos(0),
658 _number_of_mh_late_inlines(0),
659 _inlining_progress(false),
660 _inlining_incrementally(false),
661 _print_inlining_list(NULL),
662 _print_inlining_idx(0),
663 _preserve_jvm_state(0) {
664 C = this;
665
666 CompileWrapper cw(this);
667 #ifndef PRODUCT
668 if (TimeCompiler2) {
669 tty->print(" ");
670 target->holder()->name()->print();
671 tty->print(".");
672 target->print_short_name();
673 tty->print(" ");
674 }
675 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
676 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
677 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
678 if (!print_opto_assembly) {
679 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
680 if (print_assembly && !Disassembler::can_decode()) {
681 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
682 print_opto_assembly = true;
683 }
684 }
685 set_print_assembly(print_opto_assembly);
686 set_parsed_irreducible_loop(false);
687 #endif
688 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
689 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
690
691 if (ProfileTraps) {
692 // Make sure the method being compiled gets its own MDO,
693 // so we can at least track the decompile_count().
694 method()->ensure_method_data();
695 }
696
697 Init(::AliasLevel);
698
699
700 print_compile_messages();
701
702 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
703 _ilt = InlineTree::build_inline_tree_root();
704 else
705 _ilt = NULL;
706
707 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
708 assert(num_alias_types() >= AliasIdxRaw, "");
709
710 #define MINIMUM_NODE_HASH 1023
711 // Node list that Iterative GVN will start with
712 Unique_Node_List for_igvn(comp_arena());
713 set_for_igvn(&for_igvn);
714
715 // GVN that will be run immediately on new nodes
716 uint estimated_size = method()->code_size()*4+64;
717 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
718 PhaseGVN gvn(node_arena(), estimated_size);
719 set_initial_gvn(&gvn);
720
721 if (print_inlining() || print_intrinsics()) {
722 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
723 }
724 { // Scope for timing the parser
725 TracePhase t3("parse", &_t_parser, true);
726
727 // Put top into the hash table ASAP.
728 initial_gvn()->transform_no_reclaim(top());
729
730 // Set up tf(), start(), and find a CallGenerator.
731 CallGenerator* cg = NULL;
732 if (is_osr_compilation()) {
733 const TypeTuple *domain = StartOSRNode::osr_domain();
734 const TypeTuple *range = TypeTuple::make_range(method()->signature());
735 init_tf(TypeFunc::make(domain, range));
736 StartNode* s = new (this) StartOSRNode(root(), domain);
737 initial_gvn()->set_type_bottom(s);
738 init_start(s);
739 cg = CallGenerator::for_osr(method(), entry_bci());
740 } else {
741 // Normal case.
742 init_tf(TypeFunc::make(method()));
743 StartNode* s = new (this) StartNode(root(), tf()->domain());
744 initial_gvn()->set_type_bottom(s);
745 init_start(s);
746 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
747 // With java.lang.ref.reference.get() we must go through the
748 // intrinsic when G1 is enabled - even when get() is the root
749 // method of the compile - so that, if necessary, the value in
750 // the referent field of the reference object gets recorded by
751 // the pre-barrier code.
752 // Specifically, if G1 is enabled, the value in the referent
753 // field is recorded by the G1 SATB pre barrier. This will
754 // result in the referent being marked live and the reference
755 // object removed from the list of discovered references during
756 // reference processing.
757 cg = find_intrinsic(method(), false);
758 }
759 if (cg == NULL) {
760 float past_uses = method()->interpreter_invocation_count();
761 float expected_uses = past_uses;
762 cg = CallGenerator::for_inline(method(), expected_uses);
763 }
764 }
765 if (failing()) return;
766 if (cg == NULL) {
767 record_method_not_compilable_all_tiers("cannot parse method");
768 return;
769 }
770 JVMState* jvms = build_start_state(start(), tf());
771 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
772 record_method_not_compilable("method parse failed");
773 return;
774 }
775 GraphKit kit(jvms);
776
777 if (!kit.stopped()) {
778 // Accept return values, and transfer control we know not where.
779 // This is done by a special, unique ReturnNode bound to root.
780 return_values(kit.jvms());
781 }
782
783 if (kit.has_exceptions()) {
784 // Any exceptions that escape from this call must be rethrown
785 // to whatever caller is dynamically above us on the stack.
786 // This is done by a special, unique RethrowNode bound to root.
787 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
788 }
789
790 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
791
792 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
793 inline_string_calls(true);
794 }
795
796 if (failing()) return;
797
798 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
799
800 // Remove clutter produced by parsing.
801 if (!failing()) {
802 ResourceMark rm;
803 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
804 }
805 }
806
807 // Note: Large methods are capped off in do_one_bytecode().
808 if (failing()) return;
809
810 // After parsing, node notes are no longer automagic.
811 // They must be propagated by register_new_node_with_optimizer(),
812 // clone(), or the like.
813 set_default_node_notes(NULL);
814
815 for (;;) {
816 int successes = Inline_Warm();
817 if (failing()) return;
818 if (successes == 0) break;
819 }
820
821 // Drain the list.
822 Finish_Warm();
823 #ifndef PRODUCT
824 if (_printer) {
825 _printer->print_inlining(this);
826 }
827 #endif
828
829 if (failing()) return;
830 NOT_PRODUCT( verify_graph_edges(); )
831
832 // Now optimize
833 Optimize();
834 if (failing()) return;
835 NOT_PRODUCT( verify_graph_edges(); )
836
837 #ifndef PRODUCT
838 if (PrintIdeal) {
839 ttyLocker ttyl; // keep the following output all in one block
840 // This output goes directly to the tty, not the compiler log.
841 // To enable tools to match it up with the compilation activity,
842 // be sure to tag this tty output with the compile ID.
843 if (xtty != NULL) {
844 xtty->head("ideal compile_id='%d'%s", compile_id(),
845 is_osr_compilation() ? " compile_kind='osr'" :
846 "");
847 }
848 root()->dump(9999);
849 if (xtty != NULL) {
850 xtty->tail("ideal");
851 }
852 }
853 #endif
854
855 NOT_PRODUCT( verify_barriers(); )
856 // Now that we know the size of all the monitors we can add a fixed slot
857 // for the original deopt pc.
858
859 _orig_pc_slot = fixed_slots();
860 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
861 set_fixed_slots(next_slot);
862
863 // Compute when to use implicit null checks. Used by matching trap based
864 // nodes and NullCheck optimization.
865 set_allowed_deopt_reasons();
866
867 // Now generate code
868 Code_Gen();
869 if (failing()) return;
870
871 // Check if we want to skip execution of all compiled code.
872 {
873 #ifndef PRODUCT
874 if (OptoNoExecute) {
875 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
876 return;
877 }
878 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
879 #endif
880
881 if (is_osr_compilation()) {
882 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
883 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
884 } else {
885 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
886 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
887 }
888
889 env()->register_method(_method, _entry_bci,
890 &_code_offsets,
891 _orig_pc_slot_offset_in_bytes,
892 code_buffer(),
893 frame_size_in_words(), _oop_map_set,
894 &_handler_table, &_inc_table,
895 compiler,
896 env()->comp_level(),
897 has_unsafe_access(),
898 SharedRuntime::is_wide_vector(max_vector_size())
899 );
900
901 if (log() != NULL) // Print code cache state into compiler log
902 log()->code_cache_state();
903 }
904 }
905
906 //------------------------------Compile----------------------------------------
907 // Compile a runtime stub
908 Compile::Compile( ciEnv* ci_env,
909 TypeFunc_generator generator,
910 address stub_function,
911 const char *stub_name,
912 int is_fancy_jump,
913 bool pass_tls,
914 bool save_arg_registers,
915 bool return_pc )
916 : Phase(Compiler),
917 _env(ci_env),
918 _log(ci_env->log()),
919 _compile_id(0),
920 _save_argument_registers(save_arg_registers),
921 _method(NULL),
922 _stub_name(stub_name),
923 _stub_function(stub_function),
924 _stub_entry_point(NULL),
925 _entry_bci(InvocationEntryBci),
926 _initial_gvn(NULL),
927 _for_igvn(NULL),
928 _warm_calls(NULL),
929 _orig_pc_slot(0),
930 _orig_pc_slot_offset_in_bytes(0),
931 _subsume_loads(true),
932 _do_escape_analysis(false),
933 _eliminate_boxing(false),
934 _failure_reason(NULL),
935 _code_buffer("Compile::Fill_buffer"),
936 _has_method_handle_invokes(false),
937 _mach_constant_base_node(NULL),
938 _node_bundling_limit(0),
939 _node_bundling_base(NULL),
940 _java_calls(0),
941 _inner_loops(0),
942 #ifndef PRODUCT
943 _trace_opto_output(TraceOptoOutput),
944 _in_dump_cnt(0),
945 _printer(NULL),
946 #endif
947 _dead_node_list(comp_arena()),
948 _dead_node_count(0),
949 _congraph(NULL),
950 _number_of_mh_late_inlines(0),
951 _inlining_progress(false),
952 _inlining_incrementally(false),
953 _print_inlining_list(NULL),
954 _print_inlining_idx(0),
955 _preserve_jvm_state(0),
956 _allowed_reasons(0) {
957 C = this;
958
959 #ifndef PRODUCT
960 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
961 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
962 set_print_assembly(PrintFrameConverterAssembly);
963 set_parsed_irreducible_loop(false);
964 #endif
965 CompileWrapper cw(this);
966 Init(/*AliasLevel=*/ 0);
967 init_tf((*generator)());
968
969 {
970 // The following is a dummy for the sake of GraphKit::gen_stub
971 Unique_Node_List for_igvn(comp_arena());
972 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
973 PhaseGVN gvn(Thread::current()->resource_area(),255);
974 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
975 gvn.transform_no_reclaim(top());
976
977 GraphKit kit;
978 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
979 }
980
981 NOT_PRODUCT( verify_graph_edges(); )
982 Code_Gen();
983 if (failing()) return;
984
985
986 // Entry point will be accessed using compile->stub_entry_point();
987 if (code_buffer() == NULL) {
988 Matcher::soft_match_failure();
989 } else {
990 if (PrintAssembly && (WizardMode || Verbose))
991 tty->print_cr("### Stub::%s", stub_name);
992
993 if (!failing()) {
994 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
995
996 // Make the NMethod
997 // For now we mark the frame as never safe for profile stackwalking
998 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
999 code_buffer(),
1000 CodeOffsets::frame_never_safe,
1001 // _code_offsets.value(CodeOffsets::Frame_Complete),
1002 frame_size_in_words(),
1003 _oop_map_set,
1004 save_arg_registers);
1005 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1006
1007 _stub_entry_point = rs->entry_point();
1008 }
1009 }
1010 }
1011
1012 //------------------------------Init-------------------------------------------
1013 // Prepare for a single compilation
1014 void Compile::Init(int aliaslevel) {
1015 _unique = 0;
1016 _regalloc = NULL;
1017
1018 _tf = NULL; // filled in later
1019 _top = NULL; // cached later
1020 _matcher = NULL; // filled in later
1021 _cfg = NULL; // filled in later
1022
1023 set_24_bit_selection_and_mode(Use24BitFP, false);
1024
1025 _node_note_array = NULL;
1026 _default_node_notes = NULL;
1027
1028 _immutable_memory = NULL; // filled in at first inquiry
1029
1030 // Globally visible Nodes
1031 // First set TOP to NULL to give safe behavior during creation of RootNode
1032 set_cached_top_node(NULL);
1033 set_root(new (this) RootNode());
1034 // Now that you have a Root to point to, create the real TOP
1035 set_cached_top_node( new (this) ConNode(Type::TOP) );
1036 set_recent_alloc(NULL, NULL);
1037
1038 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1039 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1040 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1041 env()->set_dependencies(new Dependencies(env()));
1042
1043 _fixed_slots = 0;
1044 set_has_split_ifs(false);
1045 set_has_loops(has_method() && method()->has_loops()); // first approximation
1046 set_has_stringbuilder(false);
1047 set_has_boxed_value(false);
1048 _trap_can_recompile = false; // no traps emitted yet
1049 _major_progress = true; // start out assuming good things will happen
1050 set_has_unsafe_access(false);
1051 set_max_vector_size(0);
1052 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1053 set_decompile_count(0);
1054
1055 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1056 set_num_loop_opts(LoopOptsCount);
1057 set_do_inlining(Inline);
1058 set_max_inline_size(MaxInlineSize);
1059 set_freq_inline_size(FreqInlineSize);
1060 set_do_scheduling(OptoScheduling);
1061 set_do_count_invocations(false);
1062 set_do_method_data_update(false);
1063
1064 if (debug_info()->recording_non_safepoints()) {
1065 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1066 (comp_arena(), 8, 0, NULL));
1067 set_default_node_notes(Node_Notes::make(this));
1068 }
1069
1070 // // -- Initialize types before each compile --
1071 // // Update cached type information
1072 // if( _method && _method->constants() )
1073 // Type::update_loaded_types(_method, _method->constants());
1074
1075 // Init alias_type map.
1076 if (!_do_escape_analysis && aliaslevel == 3)
1077 aliaslevel = 2; // No unique types without escape analysis
1078 _AliasLevel = aliaslevel;
1079 const int grow_ats = 16;
1080 _max_alias_types = grow_ats;
1081 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1082 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1083 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1084 {
1085 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1086 }
1087 // Initialize the first few types.
1088 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1089 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1090 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1091 _num_alias_types = AliasIdxRaw+1;
1092 // Zero out the alias type cache.
1093 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1094 // A NULL adr_type hits in the cache right away. Preload the right answer.
1095 probe_alias_cache(NULL)->_index = AliasIdxTop;
1096
1097 _intrinsics = NULL;
1098 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1099 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1100 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1101 register_library_intrinsics();
1102 }
1103
1104 //---------------------------init_start----------------------------------------
1105 // Install the StartNode on this compile object.
1106 void Compile::init_start(StartNode* s) {
1107 if (failing())
1108 return; // already failing
1109 assert(s == start(), "");
1110 }
1111
1112 StartNode* Compile::start() const {
1113 assert(!failing(), "");
1114 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1115 Node* start = root()->fast_out(i);
1116 if( start->is_Start() )
1117 return start->as_Start();
1118 }
1119 ShouldNotReachHere();
1120 return NULL;
1121 }
1122
1123 //-------------------------------immutable_memory-------------------------------------
1124 // Access immutable memory
1125 Node* Compile::immutable_memory() {
1126 if (_immutable_memory != NULL) {
1127 return _immutable_memory;
1128 }
1129 StartNode* s = start();
1130 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1131 Node *p = s->fast_out(i);
1132 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1133 _immutable_memory = p;
1134 return _immutable_memory;
1135 }
1136 }
1137 ShouldNotReachHere();
1138 return NULL;
1139 }
1140
1141 //----------------------set_cached_top_node------------------------------------
1142 // Install the cached top node, and make sure Node::is_top works correctly.
1143 void Compile::set_cached_top_node(Node* tn) {
1144 if (tn != NULL) verify_top(tn);
1145 Node* old_top = _top;
1146 _top = tn;
1147 // Calling Node::setup_is_top allows the nodes the chance to adjust
1148 // their _out arrays.
1149 if (_top != NULL) _top->setup_is_top();
1150 if (old_top != NULL) old_top->setup_is_top();
1151 assert(_top == NULL || top()->is_top(), "");
1152 }
1153
1154 #ifdef ASSERT
1155 uint Compile::count_live_nodes_by_graph_walk() {
1156 Unique_Node_List useful(comp_arena());
1157 // Get useful node list by walking the graph.
1158 identify_useful_nodes(useful);
1159 return useful.size();
1160 }
1161
1162 void Compile::print_missing_nodes() {
1163
1164 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1165 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1166 return;
1167 }
1168
1169 // This is an expensive function. It is executed only when the user
1170 // specifies VerifyIdealNodeCount option or otherwise knows the
1171 // additional work that needs to be done to identify reachable nodes
1172 // by walking the flow graph and find the missing ones using
1173 // _dead_node_list.
1174
1175 Unique_Node_List useful(comp_arena());
1176 // Get useful node list by walking the graph.
1177 identify_useful_nodes(useful);
1178
1179 uint l_nodes = C->live_nodes();
1180 uint l_nodes_by_walk = useful.size();
1181
1182 if (l_nodes != l_nodes_by_walk) {
1183 if (_log != NULL) {
1184 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1185 _log->stamp();
1186 _log->end_head();
1187 }
1188 VectorSet& useful_member_set = useful.member_set();
1189 int last_idx = l_nodes_by_walk;
1190 for (int i = 0; i < last_idx; i++) {
1191 if (useful_member_set.test(i)) {
1192 if (_dead_node_list.test(i)) {
1193 if (_log != NULL) {
1194 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1195 }
1196 if (PrintIdealNodeCount) {
1197 // Print the log message to tty
1198 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1199 useful.at(i)->dump();
1200 }
1201 }
1202 }
1203 else if (! _dead_node_list.test(i)) {
1204 if (_log != NULL) {
1205 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1206 }
1207 if (PrintIdealNodeCount) {
1208 // Print the log message to tty
1209 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1210 }
1211 }
1212 }
1213 if (_log != NULL) {
1214 _log->tail("mismatched_nodes");
1215 }
1216 }
1217 }
1218 #endif
1219
1220 #ifndef PRODUCT
1221 void Compile::verify_top(Node* tn) const {
1222 if (tn != NULL) {
1223 assert(tn->is_Con(), "top node must be a constant");
1224 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1225 assert(tn->in(0) != NULL, "must have live top node");
1226 }
1227 }
1228 #endif
1229
1230
1231 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1232
1233 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1234 guarantee(arr != NULL, "");
1235 int num_blocks = arr->length();
1236 if (grow_by < num_blocks) grow_by = num_blocks;
1237 int num_notes = grow_by * _node_notes_block_size;
1238 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1239 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1240 while (num_notes > 0) {
1241 arr->append(notes);
1242 notes += _node_notes_block_size;
1243 num_notes -= _node_notes_block_size;
1244 }
1245 assert(num_notes == 0, "exact multiple, please");
1246 }
1247
1248 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1249 if (source == NULL || dest == NULL) return false;
1250
1251 if (dest->is_Con())
1252 return false; // Do not push debug info onto constants.
1253
1254 #ifdef ASSERT
1255 // Leave a bread crumb trail pointing to the original node:
1256 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1257 dest->set_debug_orig(source);
1258 }
1259 #endif
1260
1261 if (node_note_array() == NULL)
1262 return false; // Not collecting any notes now.
1263
1264 // This is a copy onto a pre-existing node, which may already have notes.
1265 // If both nodes have notes, do not overwrite any pre-existing notes.
1266 Node_Notes* source_notes = node_notes_at(source->_idx);
1267 if (source_notes == NULL || source_notes->is_clear()) return false;
1268 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1269 if (dest_notes == NULL || dest_notes->is_clear()) {
1270 return set_node_notes_at(dest->_idx, source_notes);
1271 }
1272
1273 Node_Notes merged_notes = (*source_notes);
1274 // The order of operations here ensures that dest notes will win...
1275 merged_notes.update_from(dest_notes);
1276 return set_node_notes_at(dest->_idx, &merged_notes);
1277 }
1278
1279
1280 //--------------------------allow_range_check_smearing-------------------------
1281 // Gating condition for coalescing similar range checks.
1282 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1283 // single covering check that is at least as strong as any of them.
1284 // If the optimization succeeds, the simplified (strengthened) range check
1285 // will always succeed. If it fails, we will deopt, and then give up
1286 // on the optimization.
1287 bool Compile::allow_range_check_smearing() const {
1288 // If this method has already thrown a range-check,
1289 // assume it was because we already tried range smearing
1290 // and it failed.
1291 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1292 return !already_trapped;
1293 }
1294
1295
1296 //------------------------------flatten_alias_type-----------------------------
1297 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1298 int offset = tj->offset();
1299 TypePtr::PTR ptr = tj->ptr();
1300
1301 // Known instance (scalarizable allocation) alias only with itself.
1302 bool is_known_inst = tj->isa_oopptr() != NULL &&
1303 tj->is_oopptr()->is_known_instance();
1304
1305 // Process weird unsafe references.
1306 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1307 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1308 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1309 tj = TypeOopPtr::BOTTOM;
1310 ptr = tj->ptr();
1311 offset = tj->offset();
1312 }
1313
1314 // Array pointers need some flattening
1315 const TypeAryPtr *ta = tj->isa_aryptr();
1316 if (ta && ta->is_stable()) {
1317 // Erase stability property for alias analysis.
1318 tj = ta = ta->cast_to_stable(false);
1319 }
1320 if( ta && is_known_inst ) {
1321 if ( offset != Type::OffsetBot &&
1322 offset > arrayOopDesc::length_offset_in_bytes() ) {
1323 offset = Type::OffsetBot; // Flatten constant access into array body only
1324 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1325 }
1326 } else if( ta && _AliasLevel >= 2 ) {
1327 // For arrays indexed by constant indices, we flatten the alias
1328 // space to include all of the array body. Only the header, klass
1329 // and array length can be accessed un-aliased.
1330 if( offset != Type::OffsetBot ) {
1331 if( ta->const_oop() ) { // MethodData* or Method*
1332 offset = Type::OffsetBot; // Flatten constant access into array body
1333 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1334 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1335 // range is OK as-is.
1336 tj = ta = TypeAryPtr::RANGE;
1337 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1338 tj = TypeInstPtr::KLASS; // all klass loads look alike
1339 ta = TypeAryPtr::RANGE; // generic ignored junk
1340 ptr = TypePtr::BotPTR;
1341 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1342 tj = TypeInstPtr::MARK;
1343 ta = TypeAryPtr::RANGE; // generic ignored junk
1344 ptr = TypePtr::BotPTR;
1345 } else { // Random constant offset into array body
1346 offset = Type::OffsetBot; // Flatten constant access into array body
1347 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1348 }
1349 }
1350 // Arrays of fixed size alias with arrays of unknown size.
1351 if (ta->size() != TypeInt::POS) {
1352 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1353 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1354 }
1355 // Arrays of known objects become arrays of unknown objects.
1356 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1357 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1358 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1359 }
1360 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1361 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1362 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1363 }
1364 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1365 // cannot be distinguished by bytecode alone.
1366 if (ta->elem() == TypeInt::BOOL) {
1367 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1368 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1369 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1370 }
1371 // During the 2nd round of IterGVN, NotNull castings are removed.
1372 // Make sure the Bottom and NotNull variants alias the same.
1373 // Also, make sure exact and non-exact variants alias the same.
1374 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1375 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1376 }
1377 }
1378
1379 // Oop pointers need some flattening
1380 const TypeInstPtr *to = tj->isa_instptr();
1381 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1382 ciInstanceKlass *k = to->klass()->as_instance_klass();
1383 if( ptr == TypePtr::Constant ) {
1384 if (to->klass() != ciEnv::current()->Class_klass() ||
1385 offset < k->size_helper() * wordSize) {
1386 // No constant oop pointers (such as Strings); they alias with
1387 // unknown strings.
1388 assert(!is_known_inst, "not scalarizable allocation");
1389 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1390 }
1391 } else if( is_known_inst ) {
1392 tj = to; // Keep NotNull and klass_is_exact for instance type
1393 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1394 // During the 2nd round of IterGVN, NotNull castings are removed.
1395 // Make sure the Bottom and NotNull variants alias the same.
1396 // Also, make sure exact and non-exact variants alias the same.
1397 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1398 }
1399 if (to->speculative() != NULL) {
1400 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1401 }
1402 // Canonicalize the holder of this field
1403 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1404 // First handle header references such as a LoadKlassNode, even if the
1405 // object's klass is unloaded at compile time (4965979).
1406 if (!is_known_inst) { // Do it only for non-instance types
1407 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1408 }
1409 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1410 // Static fields are in the space above the normal instance
1411 // fields in the java.lang.Class instance.
1412 if (to->klass() != ciEnv::current()->Class_klass()) {
1413 to = NULL;
1414 tj = TypeOopPtr::BOTTOM;
1415 offset = tj->offset();
1416 }
1417 } else {
1418 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1419 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1420 if( is_known_inst ) {
1421 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1422 } else {
1423 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1424 }
1425 }
1426 }
1427 }
1428
1429 // Klass pointers to object array klasses need some flattening
1430 const TypeKlassPtr *tk = tj->isa_klassptr();
1431 if( tk ) {
1432 // If we are referencing a field within a Klass, we need
1433 // to assume the worst case of an Object. Both exact and
1434 // inexact types must flatten to the same alias class so
1435 // use NotNull as the PTR.
1436 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1437
1438 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1439 TypeKlassPtr::OBJECT->klass(),
1440 offset);
1441 }
1442
1443 ciKlass* klass = tk->klass();
1444 if( klass->is_obj_array_klass() ) {
1445 ciKlass* k = TypeAryPtr::OOPS->klass();
1446 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1447 k = TypeInstPtr::BOTTOM->klass();
1448 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1449 }
1450
1451 // Check for precise loads from the primary supertype array and force them
1452 // to the supertype cache alias index. Check for generic array loads from
1453 // the primary supertype array and also force them to the supertype cache
1454 // alias index. Since the same load can reach both, we need to merge
1455 // these 2 disparate memories into the same alias class. Since the
1456 // primary supertype array is read-only, there's no chance of confusion
1457 // where we bypass an array load and an array store.
1458 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1459 if (offset == Type::OffsetBot ||
1460 (offset >= primary_supers_offset &&
1461 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1462 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1463 offset = in_bytes(Klass::secondary_super_cache_offset());
1464 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1465 }
1466 }
1467
1468 // Flatten all Raw pointers together.
1469 if (tj->base() == Type::RawPtr)
1470 tj = TypeRawPtr::BOTTOM;
1471
1472 if (tj->base() == Type::AnyPtr)
1473 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1474
1475 // Flatten all to bottom for now
1476 switch( _AliasLevel ) {
1477 case 0:
1478 tj = TypePtr::BOTTOM;
1479 break;
1480 case 1: // Flatten to: oop, static, field or array
1481 switch (tj->base()) {
1482 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1483 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1484 case Type::AryPtr: // do not distinguish arrays at all
1485 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1486 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1487 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1488 default: ShouldNotReachHere();
1489 }
1490 break;
1491 case 2: // No collapsing at level 2; keep all splits
1492 case 3: // No collapsing at level 3; keep all splits
1493 break;
1494 default:
1495 Unimplemented();
1496 }
1497
1498 offset = tj->offset();
1499 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1500
1501 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1502 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1503 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1504 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1505 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1506 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1507 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1508 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1509 assert( tj->ptr() != TypePtr::TopPTR &&
1510 tj->ptr() != TypePtr::AnyNull &&
1511 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1512 // assert( tj->ptr() != TypePtr::Constant ||
1513 // tj->base() == Type::RawPtr ||
1514 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1515
1516 return tj;
1517 }
1518
1519 void Compile::AliasType::Init(int i, const TypePtr* at) {
1520 _index = i;
1521 _adr_type = at;
1522 _field = NULL;
1523 _element = NULL;
1524 _is_rewritable = true; // default
1525 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1526 if (atoop != NULL && atoop->is_known_instance()) {
1527 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1528 _general_index = Compile::current()->get_alias_index(gt);
1529 } else {
1530 _general_index = 0;
1531 }
1532 }
1533
1534 //---------------------------------print_on------------------------------------
1535 #ifndef PRODUCT
1536 void Compile::AliasType::print_on(outputStream* st) {
1537 if (index() < 10)
1538 st->print("@ <%d> ", index());
1539 else st->print("@ <%d>", index());
1540 st->print(is_rewritable() ? " " : " RO");
1541 int offset = adr_type()->offset();
1542 if (offset == Type::OffsetBot)
1543 st->print(" +any");
1544 else st->print(" +%-3d", offset);
1545 st->print(" in ");
1546 adr_type()->dump_on(st);
1547 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1548 if (field() != NULL && tjp) {
1549 if (tjp->klass() != field()->holder() ||
1550 tjp->offset() != field()->offset_in_bytes()) {
1551 st->print(" != ");
1552 field()->print();
1553 st->print(" ***");
1554 }
1555 }
1556 }
1557
1558 void print_alias_types() {
1559 Compile* C = Compile::current();
1560 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1561 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1562 C->alias_type(idx)->print_on(tty);
1563 tty->cr();
1564 }
1565 }
1566 #endif
1567
1568
1569 //----------------------------probe_alias_cache--------------------------------
1570 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1571 intptr_t key = (intptr_t) adr_type;
1572 key ^= key >> logAliasCacheSize;
1573 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1574 }
1575
1576
1577 //-----------------------------grow_alias_types--------------------------------
1578 void Compile::grow_alias_types() {
1579 const int old_ats = _max_alias_types; // how many before?
1580 const int new_ats = old_ats; // how many more?
1581 const int grow_ats = old_ats+new_ats; // how many now?
1582 _max_alias_types = grow_ats;
1583 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1584 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1585 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1586 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1587 }
1588
1589
1590 //--------------------------------find_alias_type------------------------------
1591 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1592 if (_AliasLevel == 0)
1593 return alias_type(AliasIdxBot);
1594
1595 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1596 if (ace->_adr_type == adr_type) {
1597 return alias_type(ace->_index);
1598 }
1599
1600 // Handle special cases.
1601 if (adr_type == NULL) return alias_type(AliasIdxTop);
1602 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1603
1604 // Do it the slow way.
1605 const TypePtr* flat = flatten_alias_type(adr_type);
1606
1607 #ifdef ASSERT
1608 assert(flat == flatten_alias_type(flat), "idempotent");
1609 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1610 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1611 const TypeOopPtr* foop = flat->is_oopptr();
1612 // Scalarizable allocations have exact klass always.
1613 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1614 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1615 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1616 }
1617 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1618 #endif
1619
1620 int idx = AliasIdxTop;
1621 for (int i = 0; i < num_alias_types(); i++) {
1622 if (alias_type(i)->adr_type() == flat) {
1623 idx = i;
1624 break;
1625 }
1626 }
1627
1628 if (idx == AliasIdxTop) {
1629 if (no_create) return NULL;
1630 // Grow the array if necessary.
1631 if (_num_alias_types == _max_alias_types) grow_alias_types();
1632 // Add a new alias type.
1633 idx = _num_alias_types++;
1634 _alias_types[idx]->Init(idx, flat);
1635 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1636 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1637 if (flat->isa_instptr()) {
1638 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1639 && flat->is_instptr()->klass() == env()->Class_klass())
1640 alias_type(idx)->set_rewritable(false);
1641 }
1642 if (flat->isa_aryptr()) {
1643 #ifdef ASSERT
1644 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1645 // (T_BYTE has the weakest alignment and size restrictions...)
1646 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1647 #endif
1648 if (flat->offset() == TypePtr::OffsetBot) {
1649 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1650 }
1651 }
1652 if (flat->isa_klassptr()) {
1653 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1654 alias_type(idx)->set_rewritable(false);
1655 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1656 alias_type(idx)->set_rewritable(false);
1657 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1658 alias_type(idx)->set_rewritable(false);
1659 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1660 alias_type(idx)->set_rewritable(false);
1661 }
1662 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1663 // but the base pointer type is not distinctive enough to identify
1664 // references into JavaThread.)
1665
1666 // Check for final fields.
1667 const TypeInstPtr* tinst = flat->isa_instptr();
1668 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1669 ciField* field;
1670 if (tinst->const_oop() != NULL &&
1671 tinst->klass() == ciEnv::current()->Class_klass() &&
1672 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1673 // static field
1674 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1675 field = k->get_field_by_offset(tinst->offset(), true);
1676 } else {
1677 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1678 field = k->get_field_by_offset(tinst->offset(), false);
1679 }
1680 assert(field == NULL ||
1681 original_field == NULL ||
1682 (field->holder() == original_field->holder() &&
1683 field->offset() == original_field->offset() &&
1684 field->is_static() == original_field->is_static()), "wrong field?");
1685 // Set field() and is_rewritable() attributes.
1686 if (field != NULL) alias_type(idx)->set_field(field);
1687 }
1688 }
1689
1690 // Fill the cache for next time.
1691 ace->_adr_type = adr_type;
1692 ace->_index = idx;
1693 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1694
1695 // Might as well try to fill the cache for the flattened version, too.
1696 AliasCacheEntry* face = probe_alias_cache(flat);
1697 if (face->_adr_type == NULL) {
1698 face->_adr_type = flat;
1699 face->_index = idx;
1700 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1701 }
1702
1703 return alias_type(idx);
1704 }
1705
1706
1707 Compile::AliasType* Compile::alias_type(ciField* field) {
1708 const TypeOopPtr* t;
1709 if (field->is_static())
1710 t = TypeInstPtr::make(field->holder()->java_mirror());
1711 else
1712 t = TypeOopPtr::make_from_klass_raw(field->holder());
1713 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1714 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1715 return atp;
1716 }
1717
1718
1719 //------------------------------have_alias_type--------------------------------
1720 bool Compile::have_alias_type(const TypePtr* adr_type) {
1721 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1722 if (ace->_adr_type == adr_type) {
1723 return true;
1724 }
1725
1726 // Handle special cases.
1727 if (adr_type == NULL) return true;
1728 if (adr_type == TypePtr::BOTTOM) return true;
1729
1730 return find_alias_type(adr_type, true, NULL) != NULL;
1731 }
1732
1733 //-----------------------------must_alias--------------------------------------
1734 // True if all values of the given address type are in the given alias category.
1735 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1736 if (alias_idx == AliasIdxBot) return true; // the universal category
1737 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1738 if (alias_idx == AliasIdxTop) return false; // the empty category
1739 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1740
1741 // the only remaining possible overlap is identity
1742 int adr_idx = get_alias_index(adr_type);
1743 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1744 assert(adr_idx == alias_idx ||
1745 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1746 && adr_type != TypeOopPtr::BOTTOM),
1747 "should not be testing for overlap with an unsafe pointer");
1748 return adr_idx == alias_idx;
1749 }
1750
1751 //------------------------------can_alias--------------------------------------
1752 // True if any values of the given address type are in the given alias category.
1753 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1754 if (alias_idx == AliasIdxTop) return false; // the empty category
1755 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1756 if (alias_idx == AliasIdxBot) return true; // the universal category
1757 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1758
1759 // the only remaining possible overlap is identity
1760 int adr_idx = get_alias_index(adr_type);
1761 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1762 return adr_idx == alias_idx;
1763 }
1764
1765
1766
1767 //---------------------------pop_warm_call-------------------------------------
1768 WarmCallInfo* Compile::pop_warm_call() {
1769 WarmCallInfo* wci = _warm_calls;
1770 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1771 return wci;
1772 }
1773
1774 //----------------------------Inline_Warm--------------------------------------
1775 int Compile::Inline_Warm() {
1776 // If there is room, try to inline some more warm call sites.
1777 // %%% Do a graph index compaction pass when we think we're out of space?
1778 if (!InlineWarmCalls) return 0;
1779
1780 int calls_made_hot = 0;
1781 int room_to_grow = NodeCountInliningCutoff - unique();
1782 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1783 int amount_grown = 0;
1784 WarmCallInfo* call;
1785 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1786 int est_size = (int)call->size();
1787 if (est_size > (room_to_grow - amount_grown)) {
1788 // This one won't fit anyway. Get rid of it.
1789 call->make_cold();
1790 continue;
1791 }
1792 call->make_hot();
1793 calls_made_hot++;
1794 amount_grown += est_size;
1795 amount_to_grow -= est_size;
1796 }
1797
1798 if (calls_made_hot > 0) set_major_progress();
1799 return calls_made_hot;
1800 }
1801
1802
1803 //----------------------------Finish_Warm--------------------------------------
1804 void Compile::Finish_Warm() {
1805 if (!InlineWarmCalls) return;
1806 if (failing()) return;
1807 if (warm_calls() == NULL) return;
1808
1809 // Clean up loose ends, if we are out of space for inlining.
1810 WarmCallInfo* call;
1811 while ((call = pop_warm_call()) != NULL) {
1812 call->make_cold();
1813 }
1814 }
1815
1816 //---------------------cleanup_loop_predicates-----------------------
1817 // Remove the opaque nodes that protect the predicates so that all unused
1818 // checks and uncommon_traps will be eliminated from the ideal graph
1819 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1820 if (predicate_count()==0) return;
1821 for (int i = predicate_count(); i > 0; i--) {
1822 Node * n = predicate_opaque1_node(i-1);
1823 assert(n->Opcode() == Op_Opaque1, "must be");
1824 igvn.replace_node(n, n->in(1));
1825 }
1826 assert(predicate_count()==0, "should be clean!");
1827 }
1828
1829 // StringOpts and late inlining of string methods
1830 void Compile::inline_string_calls(bool parse_time) {
1831 {
1832 // remove useless nodes to make the usage analysis simpler
1833 ResourceMark rm;
1834 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1835 }
1836
1837 {
1838 ResourceMark rm;
1839 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1840 PhaseStringOpts pso(initial_gvn(), for_igvn());
1841 print_method(PHASE_AFTER_STRINGOPTS, 3);
1842 }
1843
1844 // now inline anything that we skipped the first time around
1845 if (!parse_time) {
1846 _late_inlines_pos = _late_inlines.length();
1847 }
1848
1849 while (_string_late_inlines.length() > 0) {
1850 CallGenerator* cg = _string_late_inlines.pop();
1851 cg->do_late_inline();
1852 if (failing()) return;
1853 }
1854 _string_late_inlines.trunc_to(0);
1855 }
1856
1857 // Late inlining of boxing methods
1858 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1859 if (_boxing_late_inlines.length() > 0) {
1860 assert(has_boxed_value(), "inconsistent");
1861
1862 PhaseGVN* gvn = initial_gvn();
1863 set_inlining_incrementally(true);
1864
1865 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1866 for_igvn()->clear();
1867 gvn->replace_with(&igvn);
1868
1869 while (_boxing_late_inlines.length() > 0) {
1870 CallGenerator* cg = _boxing_late_inlines.pop();
1871 cg->do_late_inline();
1872 if (failing()) return;
1873 }
1874 _boxing_late_inlines.trunc_to(0);
1875
1876 {
1877 ResourceMark rm;
1878 PhaseRemoveUseless pru(gvn, for_igvn());
1879 }
1880
1881 igvn = PhaseIterGVN(gvn);
1882 igvn.optimize();
1883
1884 set_inlining_progress(false);
1885 set_inlining_incrementally(false);
1886 }
1887 }
1888
1889 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1890 assert(IncrementalInline, "incremental inlining should be on");
1891 PhaseGVN* gvn = initial_gvn();
1892
1893 set_inlining_progress(false);
1894 for_igvn()->clear();
1895 gvn->replace_with(&igvn);
1896
1897 int i = 0;
1898
1899 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1900 CallGenerator* cg = _late_inlines.at(i);
1901 _late_inlines_pos = i+1;
1902 cg->do_late_inline();
1903 if (failing()) return;
1904 }
1905 int j = 0;
1906 for (; i < _late_inlines.length(); i++, j++) {
1907 _late_inlines.at_put(j, _late_inlines.at(i));
1908 }
1909 _late_inlines.trunc_to(j);
1910
1911 {
1912 ResourceMark rm;
1913 PhaseRemoveUseless pru(gvn, for_igvn());
1914 }
1915
1916 igvn = PhaseIterGVN(gvn);
1917 }
1918
1919 // Perform incremental inlining until bound on number of live nodes is reached
1920 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1921 PhaseGVN* gvn = initial_gvn();
1922
1923 set_inlining_incrementally(true);
1924 set_inlining_progress(true);
1925 uint low_live_nodes = 0;
1926
1927 while(inlining_progress() && _late_inlines.length() > 0) {
1928
1929 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1930 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1931 // PhaseIdealLoop is expensive so we only try it once we are
1932 // out of loop and we only try it again if the previous helped
1933 // got the number of nodes down significantly
1934 PhaseIdealLoop ideal_loop( igvn, false, true );
1935 if (failing()) return;
1936 low_live_nodes = live_nodes();
1937 _major_progress = true;
1938 }
1939
1940 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1941 break;
1942 }
1943 }
1944
1945 inline_incrementally_one(igvn);
1946
1947 if (failing()) return;
1948
1949 igvn.optimize();
1950
1951 if (failing()) return;
1952 }
1953
1954 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1955
1956 if (_string_late_inlines.length() > 0) {
1957 assert(has_stringbuilder(), "inconsistent");
1958 for_igvn()->clear();
1959 initial_gvn()->replace_with(&igvn);
1960
1961 inline_string_calls(false);
1962
1963 if (failing()) return;
1964
1965 {
1966 ResourceMark rm;
1967 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1968 }
1969
1970 igvn = PhaseIterGVN(gvn);
1971
1972 igvn.optimize();
1973 }
1974
1975 set_inlining_incrementally(false);
1976 }
1977
1978
1979 //------------------------------Optimize---------------------------------------
1980 // Given a graph, optimize it.
1981 void Compile::Optimize() {
1982 TracePhase t1("optimizer", &_t_optimizer, true);
1983
1984 #ifndef PRODUCT
1985 if (env()->break_at_compile()) {
1986 BREAKPOINT;
1987 }
1988
1989 #endif
1990
1991 ResourceMark rm;
1992 int loop_opts_cnt;
1993
1994 NOT_PRODUCT( verify_graph_edges(); )
1995
1996 print_method(PHASE_AFTER_PARSING);
1997
1998 {
1999 // Iterative Global Value Numbering, including ideal transforms
2000 // Initialize IterGVN with types and values from parse-time GVN
2001 PhaseIterGVN igvn(initial_gvn());
2002 {
2003 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2004 igvn.optimize();
2005 }
2006
2007 print_method(PHASE_ITER_GVN1, 2);
2008
2009 if (failing()) return;
2010
2011 {
2012 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2013 inline_incrementally(igvn);
2014 }
2015
2016 print_method(PHASE_INCREMENTAL_INLINE, 2);
2017
2018 if (failing()) return;
2019
2020 if (eliminate_boxing()) {
2021 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2022 // Inline valueOf() methods now.
2023 inline_boxing_calls(igvn);
2024
2025 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2026
2027 if (failing()) return;
2028 }
2029
2030 // Remove the speculative part of types and clean up the graph from
2031 // the extra CastPP nodes whose only purpose is to carry them. Do
2032 // that early so that optimizations are not disrupted by the extra
2033 // CastPP nodes.
2034 remove_speculative_types(igvn);
2035
2036 // No more new expensive nodes will be added to the list from here
2037 // so keep only the actual candidates for optimizations.
2038 cleanup_expensive_nodes(igvn);
2039
2040 // Perform escape analysis
2041 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2042 if (has_loops()) {
2043 // Cleanup graph (remove dead nodes).
2044 TracePhase t2("idealLoop", &_t_idealLoop, true);
2045 PhaseIdealLoop ideal_loop( igvn, false, true );
2046 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2047 if (failing()) return;
2048 }
2049 ConnectionGraph::do_analysis(this, &igvn);
2050
2051 if (failing()) return;
2052
2053 // Optimize out fields loads from scalar replaceable allocations.
2054 igvn.optimize();
2055 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2056
2057 if (failing()) return;
2058
2059 if (congraph() != NULL && macro_count() > 0) {
2060 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2061 PhaseMacroExpand mexp(igvn);
2062 mexp.eliminate_macro_nodes();
2063 igvn.set_delay_transform(false);
2064
2065 igvn.optimize();
2066 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2067
2068 if (failing()) return;
2069 }
2070 }
2071
2072 // Loop transforms on the ideal graph. Range Check Elimination,
2073 // peeling, unrolling, etc.
2074
2075 // Set loop opts counter
2076 loop_opts_cnt = num_loop_opts();
2077 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2078 {
2079 TracePhase t2("idealLoop", &_t_idealLoop, true);
2080 PhaseIdealLoop ideal_loop( igvn, true );
2081 loop_opts_cnt--;
2082 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2083 if (failing()) return;
2084 }
2085 // Loop opts pass if partial peeling occurred in previous pass
2086 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2087 TracePhase t3("idealLoop", &_t_idealLoop, true);
2088 PhaseIdealLoop ideal_loop( igvn, false );
2089 loop_opts_cnt--;
2090 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2091 if (failing()) return;
2092 }
2093 // Loop opts pass for loop-unrolling before CCP
2094 if(major_progress() && (loop_opts_cnt > 0)) {
2095 TracePhase t4("idealLoop", &_t_idealLoop, true);
2096 PhaseIdealLoop ideal_loop( igvn, false );
2097 loop_opts_cnt--;
2098 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2099 }
2100 if (!failing()) {
2101 // Verify that last round of loop opts produced a valid graph
2102 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2103 PhaseIdealLoop::verify(igvn);
2104 }
2105 }
2106 if (failing()) return;
2107
2108 // Conditional Constant Propagation;
2109 PhaseCCP ccp( &igvn );
2110 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2111 {
2112 TracePhase t2("ccp", &_t_ccp, true);
2113 ccp.do_transform();
2114 }
2115 print_method(PHASE_CPP1, 2);
2116
2117 assert( true, "Break here to ccp.dump_old2new_map()");
2118
2119 // Iterative Global Value Numbering, including ideal transforms
2120 {
2121 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2122 igvn = ccp;
2123 igvn.optimize();
2124 }
2125
2126 print_method(PHASE_ITER_GVN2, 2);
2127
2128 if (failing()) return;
2129
2130 // Loop transforms on the ideal graph. Range Check Elimination,
2131 // peeling, unrolling, etc.
2132 if(loop_opts_cnt > 0) {
2133 debug_only( int cnt = 0; );
2134 while(major_progress() && (loop_opts_cnt > 0)) {
2135 TracePhase t2("idealLoop", &_t_idealLoop, true);
2136 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2137 PhaseIdealLoop ideal_loop( igvn, true);
2138 loop_opts_cnt--;
2139 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2140 if (failing()) return;
2141 }
2142 }
2143
2144 {
2145 // Verify that all previous optimizations produced a valid graph
2146 // at least to this point, even if no loop optimizations were done.
2147 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2148 PhaseIdealLoop::verify(igvn);
2149 }
2150
2151 {
2152 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2153 PhaseMacroExpand mex(igvn);
2154 if (mex.expand_macro_nodes()) {
2155 assert(failing(), "must bail out w/ explicit message");
2156 return;
2157 }
2158 }
2159
2160 } // (End scope of igvn; run destructor if necessary for asserts.)
2161
2162 dump_inlining();
2163 // A method with only infinite loops has no edges entering loops from root
2164 {
2165 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2166 if (final_graph_reshaping()) {
2167 assert(failing(), "must bail out w/ explicit message");
2168 return;
2169 }
2170 }
2171
2172 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2173 }
2174
2175
2176 //------------------------------Code_Gen---------------------------------------
2177 // Given a graph, generate code for it
2178 void Compile::Code_Gen() {
2179 if (failing()) {
2180 return;
2181 }
2182
2183 // Perform instruction selection. You might think we could reclaim Matcher
2184 // memory PDQ, but actually the Matcher is used in generating spill code.
2185 // Internals of the Matcher (including some VectorSets) must remain live
2186 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2187 // set a bit in reclaimed memory.
2188
2189 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2190 // nodes. Mapping is only valid at the root of each matched subtree.
2191 NOT_PRODUCT( verify_graph_edges(); )
2192
2193 Matcher matcher;
2194 _matcher = &matcher;
2195 {
2196 TracePhase t2("matcher", &_t_matcher, true);
2197 matcher.match();
2198 }
2199 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2200 // nodes. Mapping is only valid at the root of each matched subtree.
2201 NOT_PRODUCT( verify_graph_edges(); )
2202
2203 // If you have too many nodes, or if matching has failed, bail out
2204 check_node_count(0, "out of nodes matching instructions");
2205 if (failing()) {
2206 return;
2207 }
2208
2209 // Build a proper-looking CFG
2210 PhaseCFG cfg(node_arena(), root(), matcher);
2211 _cfg = &cfg;
2212 {
2213 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2214 bool success = cfg.do_global_code_motion();
2215 if (!success) {
2216 return;
2217 }
2218
2219 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2220 NOT_PRODUCT( verify_graph_edges(); )
2221 debug_only( cfg.verify(); )
2222 }
2223
2224 PhaseChaitin regalloc(unique(), cfg, matcher);
2225 _regalloc = ®alloc;
2226 {
2227 TracePhase t2("regalloc", &_t_registerAllocation, true);
2228 // Perform register allocation. After Chaitin, use-def chains are
2229 // no longer accurate (at spill code) and so must be ignored.
2230 // Node->LRG->reg mappings are still accurate.
2231 _regalloc->Register_Allocate();
2232
2233 // Bail out if the allocator builds too many nodes
2234 if (failing()) {
2235 return;
2236 }
2237 }
2238
2239 // Prior to register allocation we kept empty basic blocks in case the
2240 // the allocator needed a place to spill. After register allocation we
2241 // are not adding any new instructions. If any basic block is empty, we
2242 // can now safely remove it.
2243 {
2244 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2245 cfg.remove_empty_blocks();
2246 if (do_freq_based_layout()) {
2247 PhaseBlockLayout layout(cfg);
2248 } else {
2249 cfg.set_loop_alignment();
2250 }
2251 cfg.fixup_flow();
2252 }
2253
2254 // Apply peephole optimizations
2255 if( OptoPeephole ) {
2256 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2257 PhasePeephole peep( _regalloc, cfg);
2258 peep.do_transform();
2259 }
2260
2261 // Do late expand if CPU requires this.
2262 if (Matcher::require_postalloc_expand) {
2263 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2264 cfg.postalloc_expand(_regalloc);
2265 }
2266
2267 // Convert Nodes to instruction bits in a buffer
2268 {
2269 // %%%% workspace merge brought two timers together for one job
2270 TracePhase t2a("output", &_t_output, true);
2271 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2272 Output();
2273 }
2274
2275 print_method(PHASE_FINAL_CODE);
2276
2277 // He's dead, Jim.
2278 _cfg = (PhaseCFG*)0xdeadbeef;
2279 _regalloc = (PhaseChaitin*)0xdeadbeef;
2280 }
2281
2282
2283 //------------------------------dump_asm---------------------------------------
2284 // Dump formatted assembly
2285 #ifndef PRODUCT
2286 void Compile::dump_asm(int *pcs, uint pc_limit) {
2287 bool cut_short = false;
2288 tty->print_cr("#");
2289 tty->print("# "); _tf->dump(); tty->cr();
2290 tty->print_cr("#");
2291
2292 // For all blocks
2293 int pc = 0x0; // Program counter
2294 char starts_bundle = ' ';
2295 _regalloc->dump_frame();
2296
2297 Node *n = NULL;
2298 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2299 if (VMThread::should_terminate()) {
2300 cut_short = true;
2301 break;
2302 }
2303 Block* block = _cfg->get_block(i);
2304 if (block->is_connector() && !Verbose) {
2305 continue;
2306 }
2307 n = block->head();
2308 if (pcs && n->_idx < pc_limit) {
2309 tty->print("%3.3x ", pcs[n->_idx]);
2310 } else {
2311 tty->print(" ");
2312 }
2313 block->dump_head(_cfg);
2314 if (block->is_connector()) {
2315 tty->print_cr(" # Empty connector block");
2316 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2317 tty->print_cr(" # Block is sole successor of call");
2318 }
2319
2320 // For all instructions
2321 Node *delay = NULL;
2322 for (uint j = 0; j < block->number_of_nodes(); j++) {
2323 if (VMThread::should_terminate()) {
2324 cut_short = true;
2325 break;
2326 }
2327 n = block->get_node(j);
2328 if (valid_bundle_info(n)) {
2329 Bundle* bundle = node_bundling(n);
2330 if (bundle->used_in_unconditional_delay()) {
2331 delay = n;
2332 continue;
2333 }
2334 if (bundle->starts_bundle()) {
2335 starts_bundle = '+';
2336 }
2337 }
2338
2339 if (WizardMode) {
2340 n->dump();
2341 }
2342
2343 if( !n->is_Region() && // Dont print in the Assembly
2344 !n->is_Phi() && // a few noisely useless nodes
2345 !n->is_Proj() &&
2346 !n->is_MachTemp() &&
2347 !n->is_SafePointScalarObject() &&
2348 !n->is_Catch() && // Would be nice to print exception table targets
2349 !n->is_MergeMem() && // Not very interesting
2350 !n->is_top() && // Debug info table constants
2351 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2352 ) {
2353 if (pcs && n->_idx < pc_limit)
2354 tty->print("%3.3x", pcs[n->_idx]);
2355 else
2356 tty->print(" ");
2357 tty->print(" %c ", starts_bundle);
2358 starts_bundle = ' ';
2359 tty->print("\t");
2360 n->format(_regalloc, tty);
2361 tty->cr();
2362 }
2363
2364 // If we have an instruction with a delay slot, and have seen a delay,
2365 // then back up and print it
2366 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2367 assert(delay != NULL, "no unconditional delay instruction");
2368 if (WizardMode) delay->dump();
2369
2370 if (node_bundling(delay)->starts_bundle())
2371 starts_bundle = '+';
2372 if (pcs && n->_idx < pc_limit)
2373 tty->print("%3.3x", pcs[n->_idx]);
2374 else
2375 tty->print(" ");
2376 tty->print(" %c ", starts_bundle);
2377 starts_bundle = ' ';
2378 tty->print("\t");
2379 delay->format(_regalloc, tty);
2380 tty->print_cr("");
2381 delay = NULL;
2382 }
2383
2384 // Dump the exception table as well
2385 if( n->is_Catch() && (Verbose || WizardMode) ) {
2386 // Print the exception table for this offset
2387 _handler_table.print_subtable_for(pc);
2388 }
2389 }
2390
2391 if (pcs && n->_idx < pc_limit)
2392 tty->print_cr("%3.3x", pcs[n->_idx]);
2393 else
2394 tty->print_cr("");
2395
2396 assert(cut_short || delay == NULL, "no unconditional delay branch");
2397
2398 } // End of per-block dump
2399 tty->print_cr("");
2400
2401 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2402 }
2403 #endif
2404
2405 //------------------------------Final_Reshape_Counts---------------------------
2406 // This class defines counters to help identify when a method
2407 // may/must be executed using hardware with only 24-bit precision.
2408 struct Final_Reshape_Counts : public StackObj {
2409 int _call_count; // count non-inlined 'common' calls
2410 int _float_count; // count float ops requiring 24-bit precision
2411 int _double_count; // count double ops requiring more precision
2412 int _java_call_count; // count non-inlined 'java' calls
2413 int _inner_loop_count; // count loops which need alignment
2414 VectorSet _visited; // Visitation flags
2415 Node_List _tests; // Set of IfNodes & PCTableNodes
2416
2417 Final_Reshape_Counts() :
2418 _call_count(0), _float_count(0), _double_count(0),
2419 _java_call_count(0), _inner_loop_count(0),
2420 _visited( Thread::current()->resource_area() ) { }
2421
2422 void inc_call_count () { _call_count ++; }
2423 void inc_float_count () { _float_count ++; }
2424 void inc_double_count() { _double_count++; }
2425 void inc_java_call_count() { _java_call_count++; }
2426 void inc_inner_loop_count() { _inner_loop_count++; }
2427
2428 int get_call_count () const { return _call_count ; }
2429 int get_float_count () const { return _float_count ; }
2430 int get_double_count() const { return _double_count; }
2431 int get_java_call_count() const { return _java_call_count; }
2432 int get_inner_loop_count() const { return _inner_loop_count; }
2433 };
2434
2435 #ifdef ASSERT
2436 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2437 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2438 // Make sure the offset goes inside the instance layout.
2439 return k->contains_field_offset(tp->offset());
2440 // Note that OffsetBot and OffsetTop are very negative.
2441 }
2442 #endif
2443
2444 // Eliminate trivially redundant StoreCMs and accumulate their
2445 // precedence edges.
2446 void Compile::eliminate_redundant_card_marks(Node* n) {
2447 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2448 if (n->in(MemNode::Address)->outcnt() > 1) {
2449 // There are multiple users of the same address so it might be
2450 // possible to eliminate some of the StoreCMs
2451 Node* mem = n->in(MemNode::Memory);
2452 Node* adr = n->in(MemNode::Address);
2453 Node* val = n->in(MemNode::ValueIn);
2454 Node* prev = n;
2455 bool done = false;
2456 // Walk the chain of StoreCMs eliminating ones that match. As
2457 // long as it's a chain of single users then the optimization is
2458 // safe. Eliminating partially redundant StoreCMs would require
2459 // cloning copies down the other paths.
2460 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2461 if (adr == mem->in(MemNode::Address) &&
2462 val == mem->in(MemNode::ValueIn)) {
2463 // redundant StoreCM
2464 if (mem->req() > MemNode::OopStore) {
2465 // Hasn't been processed by this code yet.
2466 n->add_prec(mem->in(MemNode::OopStore));
2467 } else {
2468 // Already converted to precedence edge
2469 for (uint i = mem->req(); i < mem->len(); i++) {
2470 // Accumulate any precedence edges
2471 if (mem->in(i) != NULL) {
2472 n->add_prec(mem->in(i));
2473 }
2474 }
2475 // Everything above this point has been processed.
2476 done = true;
2477 }
2478 // Eliminate the previous StoreCM
2479 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2480 assert(mem->outcnt() == 0, "should be dead");
2481 mem->disconnect_inputs(NULL, this);
2482 } else {
2483 prev = mem;
2484 }
2485 mem = prev->in(MemNode::Memory);
2486 }
2487 }
2488 }
2489
2490 //------------------------------final_graph_reshaping_impl----------------------
2491 // Implement items 1-5 from final_graph_reshaping below.
2492 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2493
2494 if ( n->outcnt() == 0 ) return; // dead node
2495 uint nop = n->Opcode();
2496
2497 // Check for 2-input instruction with "last use" on right input.
2498 // Swap to left input. Implements item (2).
2499 if( n->req() == 3 && // two-input instruction
2500 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2501 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2502 n->in(2)->outcnt() == 1 &&// right use IS a last use
2503 !n->in(2)->is_Con() ) { // right use is not a constant
2504 // Check for commutative opcode
2505 switch( nop ) {
2506 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2507 case Op_MaxI: case Op_MinI:
2508 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2509 case Op_AndL: case Op_XorL: case Op_OrL:
2510 case Op_AndI: case Op_XorI: case Op_OrI: {
2511 // Move "last use" input to left by swapping inputs
2512 n->swap_edges(1, 2);
2513 break;
2514 }
2515 default:
2516 break;
2517 }
2518 }
2519
2520 #ifdef ASSERT
2521 if( n->is_Mem() ) {
2522 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2523 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2524 // oop will be recorded in oop map if load crosses safepoint
2525 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2526 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2527 "raw memory operations should have control edge");
2528 }
2529 #endif
2530 // Count FPU ops and common calls, implements item (3)
2531 switch( nop ) {
2532 // Count all float operations that may use FPU
2533 case Op_AddF:
2534 case Op_SubF:
2535 case Op_MulF:
2536 case Op_DivF:
2537 case Op_NegF:
2538 case Op_ModF:
2539 case Op_ConvI2F:
2540 case Op_ConF:
2541 case Op_CmpF:
2542 case Op_CmpF3:
2543 // case Op_ConvL2F: // longs are split into 32-bit halves
2544 frc.inc_float_count();
2545 break;
2546
2547 case Op_ConvF2D:
2548 case Op_ConvD2F:
2549 frc.inc_float_count();
2550 frc.inc_double_count();
2551 break;
2552
2553 // Count all double operations that may use FPU
2554 case Op_AddD:
2555 case Op_SubD:
2556 case Op_MulD:
2557 case Op_DivD:
2558 case Op_NegD:
2559 case Op_ModD:
2560 case Op_ConvI2D:
2561 case Op_ConvD2I:
2562 // case Op_ConvL2D: // handled by leaf call
2563 // case Op_ConvD2L: // handled by leaf call
2564 case Op_ConD:
2565 case Op_CmpD:
2566 case Op_CmpD3:
2567 frc.inc_double_count();
2568 break;
2569 case Op_Opaque1: // Remove Opaque Nodes before matching
2570 case Op_Opaque2: // Remove Opaque Nodes before matching
2571 n->subsume_by(n->in(1), this);
2572 break;
2573 case Op_CallStaticJava:
2574 case Op_CallJava:
2575 case Op_CallDynamicJava:
2576 frc.inc_java_call_count(); // Count java call site;
2577 case Op_CallRuntime:
2578 case Op_CallLeaf:
2579 case Op_CallLeafNoFP: {
2580 assert( n->is_Call(), "" );
2581 CallNode *call = n->as_Call();
2582 // Count call sites where the FP mode bit would have to be flipped.
2583 // Do not count uncommon runtime calls:
2584 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2585 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2586 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2587 frc.inc_call_count(); // Count the call site
2588 } else { // See if uncommon argument is shared
2589 Node *n = call->in(TypeFunc::Parms);
2590 int nop = n->Opcode();
2591 // Clone shared simple arguments to uncommon calls, item (1).
2592 if( n->outcnt() > 1 &&
2593 !n->is_Proj() &&
2594 nop != Op_CreateEx &&
2595 nop != Op_CheckCastPP &&
2596 nop != Op_DecodeN &&
2597 nop != Op_DecodeNKlass &&
2598 !n->is_Mem() ) {
2599 Node *x = n->clone();
2600 call->set_req( TypeFunc::Parms, x );
2601 }
2602 }
2603 break;
2604 }
2605
2606 case Op_StoreD:
2607 case Op_LoadD:
2608 case Op_LoadD_unaligned:
2609 frc.inc_double_count();
2610 goto handle_mem;
2611 case Op_StoreF:
2612 case Op_LoadF:
2613 frc.inc_float_count();
2614 goto handle_mem;
2615
2616 case Op_StoreCM:
2617 {
2618 // Convert OopStore dependence into precedence edge
2619 Node* prec = n->in(MemNode::OopStore);
2620 n->del_req(MemNode::OopStore);
2621 n->add_prec(prec);
2622 eliminate_redundant_card_marks(n);
2623 }
2624
2625 // fall through
2626
2627 case Op_StoreB:
2628 case Op_StoreC:
2629 case Op_StorePConditional:
2630 case Op_StoreI:
2631 case Op_StoreL:
2632 case Op_StoreIConditional:
2633 case Op_StoreLConditional:
2634 case Op_CompareAndSwapI:
2635 case Op_CompareAndSwapL:
2636 case Op_CompareAndSwapP:
2637 case Op_CompareAndSwapN:
2638 case Op_GetAndAddI:
2639 case Op_GetAndAddL:
2640 case Op_GetAndSetI:
2641 case Op_GetAndSetL:
2642 case Op_GetAndSetP:
2643 case Op_GetAndSetN:
2644 case Op_StoreP:
2645 case Op_StoreN:
2646 case Op_StoreNKlass:
2647 case Op_LoadB:
2648 case Op_LoadUB:
2649 case Op_LoadUS:
2650 case Op_LoadI:
2651 case Op_LoadKlass:
2652 case Op_LoadNKlass:
2653 case Op_LoadL:
2654 case Op_LoadL_unaligned:
2655 case Op_LoadPLocked:
2656 case Op_LoadP:
2657 case Op_LoadN:
2658 case Op_LoadRange:
2659 case Op_LoadS: {
2660 handle_mem:
2661 #ifdef ASSERT
2662 if( VerifyOptoOopOffsets ) {
2663 assert( n->is_Mem(), "" );
2664 MemNode *mem = (MemNode*)n;
2665 // Check to see if address types have grounded out somehow.
2666 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2667 assert( !tp || oop_offset_is_sane(tp), "" );
2668 }
2669 #endif
2670 break;
2671 }
2672
2673 case Op_AddP: { // Assert sane base pointers
2674 Node *addp = n->in(AddPNode::Address);
2675 assert( !addp->is_AddP() ||
2676 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2677 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2678 "Base pointers must match" );
2679 #ifdef _LP64
2680 if ((UseCompressedOops || UseCompressedClassPointers) &&
2681 addp->Opcode() == Op_ConP &&
2682 addp == n->in(AddPNode::Base) &&
2683 n->in(AddPNode::Offset)->is_Con()) {
2684 // Use addressing with narrow klass to load with offset on x86.
2685 // On sparc loading 32-bits constant and decoding it have less
2686 // instructions (4) then load 64-bits constant (7).
2687 // Do this transformation here since IGVN will convert ConN back to ConP.
2688 const Type* t = addp->bottom_type();
2689 if (t->isa_oopptr() || t->isa_klassptr()) {
2690 Node* nn = NULL;
2691
2692 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2693
2694 // Look for existing ConN node of the same exact type.
2695 Node* r = root();
2696 uint cnt = r->outcnt();
2697 for (uint i = 0; i < cnt; i++) {
2698 Node* m = r->raw_out(i);
2699 if (m!= NULL && m->Opcode() == op &&
2700 m->bottom_type()->make_ptr() == t) {
2701 nn = m;
2702 break;
2703 }
2704 }
2705 if (nn != NULL) {
2706 // Decode a narrow oop to match address
2707 // [R12 + narrow_oop_reg<<3 + offset]
2708 if (t->isa_oopptr()) {
2709 nn = new (this) DecodeNNode(nn, t);
2710 } else {
2711 nn = new (this) DecodeNKlassNode(nn, t);
2712 }
2713 n->set_req(AddPNode::Base, nn);
2714 n->set_req(AddPNode::Address, nn);
2715 if (addp->outcnt() == 0) {
2716 addp->disconnect_inputs(NULL, this);
2717 }
2718 }
2719 }
2720 }
2721 #endif
2722 break;
2723 }
2724
2725 #ifdef _LP64
2726 case Op_CastPP:
2727 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2728 Node* in1 = n->in(1);
2729 const Type* t = n->bottom_type();
2730 Node* new_in1 = in1->clone();
2731 new_in1->as_DecodeN()->set_type(t);
2732
2733 if (!Matcher::narrow_oop_use_complex_address()) {
2734 //
2735 // x86, ARM and friends can handle 2 adds in addressing mode
2736 // and Matcher can fold a DecodeN node into address by using
2737 // a narrow oop directly and do implicit NULL check in address:
2738 //
2739 // [R12 + narrow_oop_reg<<3 + offset]
2740 // NullCheck narrow_oop_reg
2741 //
2742 // On other platforms (Sparc) we have to keep new DecodeN node and
2743 // use it to do implicit NULL check in address:
2744 //
2745 // decode_not_null narrow_oop_reg, base_reg
2746 // [base_reg + offset]
2747 // NullCheck base_reg
2748 //
2749 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2750 // to keep the information to which NULL check the new DecodeN node
2751 // corresponds to use it as value in implicit_null_check().
2752 //
2753 new_in1->set_req(0, n->in(0));
2754 }
2755
2756 n->subsume_by(new_in1, this);
2757 if (in1->outcnt() == 0) {
2758 in1->disconnect_inputs(NULL, this);
2759 }
2760 }
2761 break;
2762
2763 case Op_CmpP:
2764 // Do this transformation here to preserve CmpPNode::sub() and
2765 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2766 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2767 Node* in1 = n->in(1);
2768 Node* in2 = n->in(2);
2769 if (!in1->is_DecodeNarrowPtr()) {
2770 in2 = in1;
2771 in1 = n->in(2);
2772 }
2773 assert(in1->is_DecodeNarrowPtr(), "sanity");
2774
2775 Node* new_in2 = NULL;
2776 if (in2->is_DecodeNarrowPtr()) {
2777 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2778 new_in2 = in2->in(1);
2779 } else if (in2->Opcode() == Op_ConP) {
2780 const Type* t = in2->bottom_type();
2781 if (t == TypePtr::NULL_PTR) {
2782 assert(in1->is_DecodeN(), "compare klass to null?");
2783 // Don't convert CmpP null check into CmpN if compressed
2784 // oops implicit null check is not generated.
2785 // This will allow to generate normal oop implicit null check.
2786 if (Matcher::gen_narrow_oop_implicit_null_checks())
2787 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2788 //
2789 // This transformation together with CastPP transformation above
2790 // will generated code for implicit NULL checks for compressed oops.
2791 //
2792 // The original code after Optimize()
2793 //
2794 // LoadN memory, narrow_oop_reg
2795 // decode narrow_oop_reg, base_reg
2796 // CmpP base_reg, NULL
2797 // CastPP base_reg // NotNull
2798 // Load [base_reg + offset], val_reg
2799 //
2800 // after these transformations will be
2801 //
2802 // LoadN memory, narrow_oop_reg
2803 // CmpN narrow_oop_reg, NULL
2804 // decode_not_null narrow_oop_reg, base_reg
2805 // Load [base_reg + offset], val_reg
2806 //
2807 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2808 // since narrow oops can be used in debug info now (see the code in
2809 // final_graph_reshaping_walk()).
2810 //
2811 // At the end the code will be matched to
2812 // on x86:
2813 //
2814 // Load_narrow_oop memory, narrow_oop_reg
2815 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2816 // NullCheck narrow_oop_reg
2817 //
2818 // and on sparc:
2819 //
2820 // Load_narrow_oop memory, narrow_oop_reg
2821 // decode_not_null narrow_oop_reg, base_reg
2822 // Load [base_reg + offset], val_reg
2823 // NullCheck base_reg
2824 //
2825 } else if (t->isa_oopptr()) {
2826 new_in2 = ConNode::make(this, t->make_narrowoop());
2827 } else if (t->isa_klassptr()) {
2828 new_in2 = ConNode::make(this, t->make_narrowklass());
2829 }
2830 }
2831 if (new_in2 != NULL) {
2832 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2833 n->subsume_by(cmpN, this);
2834 if (in1->outcnt() == 0) {
2835 in1->disconnect_inputs(NULL, this);
2836 }
2837 if (in2->outcnt() == 0) {
2838 in2->disconnect_inputs(NULL, this);
2839 }
2840 }
2841 }
2842 break;
2843
2844 case Op_DecodeN:
2845 case Op_DecodeNKlass:
2846 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2847 // DecodeN could be pinned when it can't be fold into
2848 // an address expression, see the code for Op_CastPP above.
2849 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2850 break;
2851
2852 case Op_EncodeP:
2853 case Op_EncodePKlass: {
2854 Node* in1 = n->in(1);
2855 if (in1->is_DecodeNarrowPtr()) {
2856 n->subsume_by(in1->in(1), this);
2857 } else if (in1->Opcode() == Op_ConP) {
2858 const Type* t = in1->bottom_type();
2859 if (t == TypePtr::NULL_PTR) {
2860 assert(t->isa_oopptr(), "null klass?");
2861 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2862 } else if (t->isa_oopptr()) {
2863 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2864 } else if (t->isa_klassptr()) {
2865 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2866 }
2867 }
2868 if (in1->outcnt() == 0) {
2869 in1->disconnect_inputs(NULL, this);
2870 }
2871 break;
2872 }
2873
2874 case Op_Proj: {
2875 if (OptimizeStringConcat) {
2876 ProjNode* p = n->as_Proj();
2877 if (p->_is_io_use) {
2878 // Separate projections were used for the exception path which
2879 // are normally removed by a late inline. If it wasn't inlined
2880 // then they will hang around and should just be replaced with
2881 // the original one.
2882 Node* proj = NULL;
2883 // Replace with just one
2884 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2885 Node *use = i.get();
2886 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2887 proj = use;
2888 break;
2889 }
2890 }
2891 assert(proj != NULL, "must be found");
2892 p->subsume_by(proj, this);
2893 }
2894 }
2895 break;
2896 }
2897
2898 case Op_Phi:
2899 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2900 // The EncodeP optimization may create Phi with the same edges
2901 // for all paths. It is not handled well by Register Allocator.
2902 Node* unique_in = n->in(1);
2903 assert(unique_in != NULL, "");
2904 uint cnt = n->req();
2905 for (uint i = 2; i < cnt; i++) {
2906 Node* m = n->in(i);
2907 assert(m != NULL, "");
2908 if (unique_in != m)
2909 unique_in = NULL;
2910 }
2911 if (unique_in != NULL) {
2912 n->subsume_by(unique_in, this);
2913 }
2914 }
2915 break;
2916
2917 #endif
2918
2919 case Op_ModI:
2920 if (UseDivMod) {
2921 // Check if a%b and a/b both exist
2922 Node* d = n->find_similar(Op_DivI);
2923 if (d) {
2924 // Replace them with a fused divmod if supported
2925 if (Matcher::has_match_rule(Op_DivModI)) {
2926 DivModINode* divmod = DivModINode::make(this, n);
2927 d->subsume_by(divmod->div_proj(), this);
2928 n->subsume_by(divmod->mod_proj(), this);
2929 } else {
2930 // replace a%b with a-((a/b)*b)
2931 Node* mult = new (this) MulINode(d, d->in(2));
2932 Node* sub = new (this) SubINode(d->in(1), mult);
2933 n->subsume_by(sub, this);
2934 }
2935 }
2936 }
2937 break;
2938
2939 case Op_ModL:
2940 if (UseDivMod) {
2941 // Check if a%b and a/b both exist
2942 Node* d = n->find_similar(Op_DivL);
2943 if (d) {
2944 // Replace them with a fused divmod if supported
2945 if (Matcher::has_match_rule(Op_DivModL)) {
2946 DivModLNode* divmod = DivModLNode::make(this, n);
2947 d->subsume_by(divmod->div_proj(), this);
2948 n->subsume_by(divmod->mod_proj(), this);
2949 } else {
2950 // replace a%b with a-((a/b)*b)
2951 Node* mult = new (this) MulLNode(d, d->in(2));
2952 Node* sub = new (this) SubLNode(d->in(1), mult);
2953 n->subsume_by(sub, this);
2954 }
2955 }
2956 }
2957 break;
2958
2959 case Op_LoadVector:
2960 case Op_StoreVector:
2961 break;
2962
2963 case Op_PackB:
2964 case Op_PackS:
2965 case Op_PackI:
2966 case Op_PackF:
2967 case Op_PackL:
2968 case Op_PackD:
2969 if (n->req()-1 > 2) {
2970 // Replace many operand PackNodes with a binary tree for matching
2971 PackNode* p = (PackNode*) n;
2972 Node* btp = p->binary_tree_pack(this, 1, n->req());
2973 n->subsume_by(btp, this);
2974 }
2975 break;
2976 case Op_Loop:
2977 case Op_CountedLoop:
2978 if (n->as_Loop()->is_inner_loop()) {
2979 frc.inc_inner_loop_count();
2980 }
2981 break;
2982 case Op_LShiftI:
2983 case Op_RShiftI:
2984 case Op_URShiftI:
2985 case Op_LShiftL:
2986 case Op_RShiftL:
2987 case Op_URShiftL:
2988 if (Matcher::need_masked_shift_count) {
2989 // The cpu's shift instructions don't restrict the count to the
2990 // lower 5/6 bits. We need to do the masking ourselves.
2991 Node* in2 = n->in(2);
2992 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2993 const TypeInt* t = in2->find_int_type();
2994 if (t != NULL && t->is_con()) {
2995 juint shift = t->get_con();
2996 if (shift > mask) { // Unsigned cmp
2997 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2998 }
2999 } else {
3000 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3001 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3002 n->set_req(2, shift);
3003 }
3004 }
3005 if (in2->outcnt() == 0) { // Remove dead node
3006 in2->disconnect_inputs(NULL, this);
3007 }
3008 }
3009 break;
3010 case Op_MemBarStoreStore:
3011 case Op_MemBarRelease:
3012 // Break the link with AllocateNode: it is no longer useful and
3013 // confuses register allocation.
3014 if (n->req() > MemBarNode::Precedent) {
3015 n->set_req(MemBarNode::Precedent, top());
3016 }
3017 break;
3018 // Must set a control edge on all nodes that produce a FlagsProj
3019 // so they can't escape the block that consumes the flags.
3020 // Must also set the non throwing branch as the control
3021 // for all nodes that depends on the result. Unless the node
3022 // already have a control that isn't the control of the
3023 // flag producer
3024 case Op_FlagsProj:
3025 {
3026 MathExactNode* math = (MathExactNode*) n->in(0);
3027 Node* ctrl = math->control_node();
3028 Node* non_throwing = math->non_throwing_branch();
3029 math->set_req(0, ctrl);
3030
3031 Node* result = math->result_node();
3032 if (result != NULL) {
3033 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) {
3034 Node* out = result->fast_out(j);
3035 // Phi nodes shouldn't be moved. They would only match below if they
3036 // had the same control as the MathExactNode. The only time that
3037 // would happen is if the Phi is also an input to the MathExact
3038 //
3039 // Cmp nodes shouldn't have control set at all.
3040 if (out->is_Phi() ||
3041 out->is_Cmp()) {
3042 continue;
3043 }
3044
3045 if (out->in(0) == NULL) {
3046 out->set_req(0, non_throwing);
3047 } else if (out->in(0) == ctrl) {
3048 out->set_req(0, non_throwing);
3049 }
3050 }
3051 }
3052 }
3053 break;
3054 default:
3055 assert( !n->is_Call(), "" );
3056 assert( !n->is_Mem(), "" );
3057 break;
3058 }
3059
3060 // Collect CFG split points
3061 if (n->is_MultiBranch())
3062 frc._tests.push(n);
3063 }
3064
3065 //------------------------------final_graph_reshaping_walk---------------------
3066 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3067 // requires that the walk visits a node's inputs before visiting the node.
3068 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3069 ResourceArea *area = Thread::current()->resource_area();
3070 Unique_Node_List sfpt(area);
3071
3072 frc._visited.set(root->_idx); // first, mark node as visited
3073 uint cnt = root->req();
3074 Node *n = root;
3075 uint i = 0;
3076 while (true) {
3077 if (i < cnt) {
3078 // Place all non-visited non-null inputs onto stack
3079 Node* m = n->in(i);
3080 ++i;
3081 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3082 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3083 sfpt.push(m);
3084 cnt = m->req();
3085 nstack.push(n, i); // put on stack parent and next input's index
3086 n = m;
3087 i = 0;
3088 }
3089 } else {
3090 // Now do post-visit work
3091 final_graph_reshaping_impl( n, frc );
3092 if (nstack.is_empty())
3093 break; // finished
3094 n = nstack.node(); // Get node from stack
3095 cnt = n->req();
3096 i = nstack.index();
3097 nstack.pop(); // Shift to the next node on stack
3098 }
3099 }
3100
3101 // Skip next transformation if compressed oops are not used.
3102 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3103 (!UseCompressedOops && !UseCompressedClassPointers))
3104 return;
3105
3106 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3107 // It could be done for an uncommon traps or any safepoints/calls
3108 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3109 while (sfpt.size() > 0) {
3110 n = sfpt.pop();
3111 JVMState *jvms = n->as_SafePoint()->jvms();
3112 assert(jvms != NULL, "sanity");
3113 int start = jvms->debug_start();
3114 int end = n->req();
3115 bool is_uncommon = (n->is_CallStaticJava() &&
3116 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3117 for (int j = start; j < end; j++) {
3118 Node* in = n->in(j);
3119 if (in->is_DecodeNarrowPtr()) {
3120 bool safe_to_skip = true;
3121 if (!is_uncommon ) {
3122 // Is it safe to skip?
3123 for (uint i = 0; i < in->outcnt(); i++) {
3124 Node* u = in->raw_out(i);
3125 if (!u->is_SafePoint() ||
3126 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3127 safe_to_skip = false;
3128 }
3129 }
3130 }
3131 if (safe_to_skip) {
3132 n->set_req(j, in->in(1));
3133 }
3134 if (in->outcnt() == 0) {
3135 in->disconnect_inputs(NULL, this);
3136 }
3137 }
3138 }
3139 }
3140 }
3141
3142 //------------------------------final_graph_reshaping--------------------------
3143 // Final Graph Reshaping.
3144 //
3145 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3146 // and not commoned up and forced early. Must come after regular
3147 // optimizations to avoid GVN undoing the cloning. Clone constant
3148 // inputs to Loop Phis; these will be split by the allocator anyways.
3149 // Remove Opaque nodes.
3150 // (2) Move last-uses by commutative operations to the left input to encourage
3151 // Intel update-in-place two-address operations and better register usage
3152 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3153 // calls canonicalizing them back.
3154 // (3) Count the number of double-precision FP ops, single-precision FP ops
3155 // and call sites. On Intel, we can get correct rounding either by
3156 // forcing singles to memory (requires extra stores and loads after each
3157 // FP bytecode) or we can set a rounding mode bit (requires setting and
3158 // clearing the mode bit around call sites). The mode bit is only used
3159 // if the relative frequency of single FP ops to calls is low enough.
3160 // This is a key transform for SPEC mpeg_audio.
3161 // (4) Detect infinite loops; blobs of code reachable from above but not
3162 // below. Several of the Code_Gen algorithms fail on such code shapes,
3163 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3164 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3165 // Detection is by looking for IfNodes where only 1 projection is
3166 // reachable from below or CatchNodes missing some targets.
3167 // (5) Assert for insane oop offsets in debug mode.
3168
3169 bool Compile::final_graph_reshaping() {
3170 // an infinite loop may have been eliminated by the optimizer,
3171 // in which case the graph will be empty.
3172 if (root()->req() == 1) {
3173 record_method_not_compilable("trivial infinite loop");
3174 return true;
3175 }
3176
3177 // Expensive nodes have their control input set to prevent the GVN
3178 // from freely commoning them. There's no GVN beyond this point so
3179 // no need to keep the control input. We want the expensive nodes to
3180 // be freely moved to the least frequent code path by gcm.
3181 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3182 for (int i = 0; i < expensive_count(); i++) {
3183 _expensive_nodes->at(i)->set_req(0, NULL);
3184 }
3185
3186 Final_Reshape_Counts frc;
3187
3188 // Visit everybody reachable!
3189 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3190 Node_Stack nstack(unique() >> 1);
3191 final_graph_reshaping_walk(nstack, root(), frc);
3192
3193 // Check for unreachable (from below) code (i.e., infinite loops).
3194 for( uint i = 0; i < frc._tests.size(); i++ ) {
3195 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3196 // Get number of CFG targets.
3197 // Note that PCTables include exception targets after calls.
3198 uint required_outcnt = n->required_outcnt();
3199 if (n->outcnt() != required_outcnt) {
3200 // Check for a few special cases. Rethrow Nodes never take the
3201 // 'fall-thru' path, so expected kids is 1 less.
3202 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3203 if (n->in(0)->in(0)->is_Call()) {
3204 CallNode *call = n->in(0)->in(0)->as_Call();
3205 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3206 required_outcnt--; // Rethrow always has 1 less kid
3207 } else if (call->req() > TypeFunc::Parms &&
3208 call->is_CallDynamicJava()) {
3209 // Check for null receiver. In such case, the optimizer has
3210 // detected that the virtual call will always result in a null
3211 // pointer exception. The fall-through projection of this CatchNode
3212 // will not be populated.
3213 Node *arg0 = call->in(TypeFunc::Parms);
3214 if (arg0->is_Type() &&
3215 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3216 required_outcnt--;
3217 }
3218 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3219 call->req() > TypeFunc::Parms+1 &&
3220 call->is_CallStaticJava()) {
3221 // Check for negative array length. In such case, the optimizer has
3222 // detected that the allocation attempt will always result in an
3223 // exception. There is no fall-through projection of this CatchNode .
3224 Node *arg1 = call->in(TypeFunc::Parms+1);
3225 if (arg1->is_Type() &&
3226 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3227 required_outcnt--;
3228 }
3229 }
3230 }
3231 }
3232 // Recheck with a better notion of 'required_outcnt'
3233 if (n->outcnt() != required_outcnt) {
3234 record_method_not_compilable("malformed control flow");
3235 return true; // Not all targets reachable!
3236 }
3237 }
3238 // Check that I actually visited all kids. Unreached kids
3239 // must be infinite loops.
3240 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3241 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3242 record_method_not_compilable("infinite loop");
3243 return true; // Found unvisited kid; must be unreach
3244 }
3245 }
3246
3247 // If original bytecodes contained a mixture of floats and doubles
3248 // check if the optimizer has made it homogenous, item (3).
3249 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3250 frc.get_float_count() > 32 &&
3251 frc.get_double_count() == 0 &&
3252 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3253 set_24_bit_selection_and_mode( false, true );
3254 }
3255
3256 set_java_calls(frc.get_java_call_count());
3257 set_inner_loops(frc.get_inner_loop_count());
3258
3259 // No infinite loops, no reason to bail out.
3260 return false;
3261 }
3262
3263 //-----------------------------too_many_traps----------------------------------
3264 // Report if there are too many traps at the current method and bci.
3265 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3266 bool Compile::too_many_traps(ciMethod* method,
3267 int bci,
3268 Deoptimization::DeoptReason reason) {
3269 ciMethodData* md = method->method_data();
3270 if (md->is_empty()) {
3271 // Assume the trap has not occurred, or that it occurred only
3272 // because of a transient condition during start-up in the interpreter.
3273 return false;
3274 }
3275 if (md->has_trap_at(bci, reason) != 0) {
3276 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3277 // Also, if there are multiple reasons, or if there is no per-BCI record,
3278 // assume the worst.
3279 if (log())
3280 log()->elem("observe trap='%s' count='%d'",
3281 Deoptimization::trap_reason_name(reason),
3282 md->trap_count(reason));
3283 return true;
3284 } else {
3285 // Ignore method/bci and see if there have been too many globally.
3286 return too_many_traps(reason, md);
3287 }
3288 }
3289
3290 // Less-accurate variant which does not require a method and bci.
3291 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3292 ciMethodData* logmd) {
3293 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3294 // Too many traps globally.
3295 // Note that we use cumulative trap_count, not just md->trap_count.
3296 if (log()) {
3297 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3298 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3299 Deoptimization::trap_reason_name(reason),
3300 mcount, trap_count(reason));
3301 }
3302 return true;
3303 } else {
3304 // The coast is clear.
3305 return false;
3306 }
3307 }
3308
3309 //--------------------------too_many_recompiles--------------------------------
3310 // Report if there are too many recompiles at the current method and bci.
3311 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3312 // Is not eager to return true, since this will cause the compiler to use
3313 // Action_none for a trap point, to avoid too many recompilations.
3314 bool Compile::too_many_recompiles(ciMethod* method,
3315 int bci,
3316 Deoptimization::DeoptReason reason) {
3317 ciMethodData* md = method->method_data();
3318 if (md->is_empty()) {
3319 // Assume the trap has not occurred, or that it occurred only
3320 // because of a transient condition during start-up in the interpreter.
3321 return false;
3322 }
3323 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3324 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3325 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3326 Deoptimization::DeoptReason per_bc_reason
3327 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3328 if ((per_bc_reason == Deoptimization::Reason_none
3329 || md->has_trap_at(bci, reason) != 0)
3330 // The trap frequency measure we care about is the recompile count:
3331 && md->trap_recompiled_at(bci)
3332 && md->overflow_recompile_count() >= bc_cutoff) {
3333 // Do not emit a trap here if it has already caused recompilations.
3334 // Also, if there are multiple reasons, or if there is no per-BCI record,
3335 // assume the worst.
3336 if (log())
3337 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3338 Deoptimization::trap_reason_name(reason),
3339 md->trap_count(reason),
3340 md->overflow_recompile_count());
3341 return true;
3342 } else if (trap_count(reason) != 0
3343 && decompile_count() >= m_cutoff) {
3344 // Too many recompiles globally, and we have seen this sort of trap.
3345 // Use cumulative decompile_count, not just md->decompile_count.
3346 if (log())
3347 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3348 Deoptimization::trap_reason_name(reason),
3349 md->trap_count(reason), trap_count(reason),
3350 md->decompile_count(), decompile_count());
3351 return true;
3352 } else {
3353 // The coast is clear.
3354 return false;
3355 }
3356 }
3357
3358 // Compute when not to trap. Used by matching trap based nodes and
3359 // NullCheck optimization.
3360 void Compile::set_allowed_deopt_reasons() {
3361 _allowed_reasons = 0;
3362 if (is_method_compilation()) {
3363 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3364 assert(rs < BitsPerInt, "recode bit map");
3365 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3366 _allowed_reasons |= nth_bit(rs);
3367 }
3368 }
3369 }
3370 }
3371
3372 #ifndef PRODUCT
3373 //------------------------------verify_graph_edges---------------------------
3374 // Walk the Graph and verify that there is a one-to-one correspondence
3375 // between Use-Def edges and Def-Use edges in the graph.
3376 void Compile::verify_graph_edges(bool no_dead_code) {
3377 if (VerifyGraphEdges) {
3378 ResourceArea *area = Thread::current()->resource_area();
3379 Unique_Node_List visited(area);
3380 // Call recursive graph walk to check edges
3381 _root->verify_edges(visited);
3382 if (no_dead_code) {
3383 // Now make sure that no visited node is used by an unvisited node.
3384 bool dead_nodes = 0;
3385 Unique_Node_List checked(area);
3386 while (visited.size() > 0) {
3387 Node* n = visited.pop();
3388 checked.push(n);
3389 for (uint i = 0; i < n->outcnt(); i++) {
3390 Node* use = n->raw_out(i);
3391 if (checked.member(use)) continue; // already checked
3392 if (visited.member(use)) continue; // already in the graph
3393 if (use->is_Con()) continue; // a dead ConNode is OK
3394 // At this point, we have found a dead node which is DU-reachable.
3395 if (dead_nodes++ == 0)
3396 tty->print_cr("*** Dead nodes reachable via DU edges:");
3397 use->dump(2);
3398 tty->print_cr("---");
3399 checked.push(use); // No repeats; pretend it is now checked.
3400 }
3401 }
3402 assert(dead_nodes == 0, "using nodes must be reachable from root");
3403 }
3404 }
3405 }
3406
3407 // Verify GC barriers consistency
3408 // Currently supported:
3409 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3410 void Compile::verify_barriers() {
3411 if (UseG1GC) {
3412 // Verify G1 pre-barriers
3413 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3414
3415 ResourceArea *area = Thread::current()->resource_area();
3416 Unique_Node_List visited(area);
3417 Node_List worklist(area);
3418 // We're going to walk control flow backwards starting from the Root
3419 worklist.push(_root);
3420 while (worklist.size() > 0) {
3421 Node* x = worklist.pop();
3422 if (x == NULL || x == top()) continue;
3423 if (visited.member(x)) {
3424 continue;
3425 } else {
3426 visited.push(x);
3427 }
3428
3429 if (x->is_Region()) {
3430 for (uint i = 1; i < x->req(); i++) {
3431 worklist.push(x->in(i));
3432 }
3433 } else {
3434 worklist.push(x->in(0));
3435 // We are looking for the pattern:
3436 // /->ThreadLocal
3437 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3438 // \->ConI(0)
3439 // We want to verify that the If and the LoadB have the same control
3440 // See GraphKit::g1_write_barrier_pre()
3441 if (x->is_If()) {
3442 IfNode *iff = x->as_If();
3443 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3444 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3445 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3446 && cmp->in(1)->is_Load()) {
3447 LoadNode* load = cmp->in(1)->as_Load();
3448 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3449 && load->in(2)->in(3)->is_Con()
3450 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3451
3452 Node* if_ctrl = iff->in(0);
3453 Node* load_ctrl = load->in(0);
3454
3455 if (if_ctrl != load_ctrl) {
3456 // Skip possible CProj->NeverBranch in infinite loops
3457 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3458 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3459 if_ctrl = if_ctrl->in(0)->in(0);
3460 }
3461 }
3462 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3463 }
3464 }
3465 }
3466 }
3467 }
3468 }
3469 }
3470 }
3471
3472 #endif
3473
3474 // The Compile object keeps track of failure reasons separately from the ciEnv.
3475 // This is required because there is not quite a 1-1 relation between the
3476 // ciEnv and its compilation task and the Compile object. Note that one
3477 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3478 // to backtrack and retry without subsuming loads. Other than this backtracking
3479 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3480 // by the logic in C2Compiler.
3481 void Compile::record_failure(const char* reason) {
3482 if (log() != NULL) {
3483 log()->elem("failure reason='%s' phase='compile'", reason);
3484 }
3485 if (_failure_reason == NULL) {
3486 // Record the first failure reason.
3487 _failure_reason = reason;
3488 }
3489
3490 EventCompilerFailure event;
3491 if (event.should_commit()) {
3492 event.set_compileID(Compile::compile_id());
3493 event.set_failure(reason);
3494 event.commit();
3495 }
3496
3497 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3498 C->print_method(PHASE_FAILURE);
3499 }
3500 _root = NULL; // flush the graph, too
3501 }
3502
3503 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3504 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3505 _phase_name(name), _dolog(dolog)
3506 {
3507 if (dolog) {
3508 C = Compile::current();
3509 _log = C->log();
3510 } else {
3511 C = NULL;
3512 _log = NULL;
3513 }
3514 if (_log != NULL) {
3515 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3516 _log->stamp();
3517 _log->end_head();
3518 }
3519 }
3520
3521 Compile::TracePhase::~TracePhase() {
3522
3523 C = Compile::current();
3524 if (_dolog) {
3525 _log = C->log();
3526 } else {
3527 _log = NULL;
3528 }
3529
3530 #ifdef ASSERT
3531 if (PrintIdealNodeCount) {
3532 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3533 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3534 }
3535
3536 if (VerifyIdealNodeCount) {
3537 Compile::current()->print_missing_nodes();
3538 }
3539 #endif
3540
3541 if (_log != NULL) {
3542 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3543 }
3544 }
3545
3546 //=============================================================================
3547 // Two Constant's are equal when the type and the value are equal.
3548 bool Compile::Constant::operator==(const Constant& other) {
3549 if (type() != other.type() ) return false;
3550 if (can_be_reused() != other.can_be_reused()) return false;
3551 // For floating point values we compare the bit pattern.
3552 switch (type()) {
3553 case T_FLOAT: return (_v._value.i == other._v._value.i);
3554 case T_LONG:
3555 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3556 case T_OBJECT:
3557 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3558 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3559 case T_METADATA: return (_v._metadata == other._v._metadata);
3560 default: ShouldNotReachHere();
3561 }
3562 return false;
3563 }
3564
3565 static int type_to_size_in_bytes(BasicType t) {
3566 switch (t) {
3567 case T_LONG: return sizeof(jlong );
3568 case T_FLOAT: return sizeof(jfloat );
3569 case T_DOUBLE: return sizeof(jdouble);
3570 case T_METADATA: return sizeof(Metadata*);
3571 // We use T_VOID as marker for jump-table entries (labels) which
3572 // need an internal word relocation.
3573 case T_VOID:
3574 case T_ADDRESS:
3575 case T_OBJECT: return sizeof(jobject);
3576 }
3577
3578 ShouldNotReachHere();
3579 return -1;
3580 }
3581
3582 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3583 // sort descending
3584 if (a->freq() > b->freq()) return -1;
3585 if (a->freq() < b->freq()) return 1;
3586 return 0;
3587 }
3588
3589 void Compile::ConstantTable::calculate_offsets_and_size() {
3590 // First, sort the array by frequencies.
3591 _constants.sort(qsort_comparator);
3592
3593 #ifdef ASSERT
3594 // Make sure all jump-table entries were sorted to the end of the
3595 // array (they have a negative frequency).
3596 bool found_void = false;
3597 for (int i = 0; i < _constants.length(); i++) {
3598 Constant con = _constants.at(i);
3599 if (con.type() == T_VOID)
3600 found_void = true; // jump-tables
3601 else
3602 assert(!found_void, "wrong sorting");
3603 }
3604 #endif
3605
3606 int offset = 0;
3607 for (int i = 0; i < _constants.length(); i++) {
3608 Constant* con = _constants.adr_at(i);
3609
3610 // Align offset for type.
3611 int typesize = type_to_size_in_bytes(con->type());
3612 offset = align_size_up(offset, typesize);
3613 con->set_offset(offset); // set constant's offset
3614
3615 if (con->type() == T_VOID) {
3616 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3617 offset = offset + typesize * n->outcnt(); // expand jump-table
3618 } else {
3619 offset = offset + typesize;
3620 }
3621 }
3622
3623 // Align size up to the next section start (which is insts; see
3624 // CodeBuffer::align_at_start).
3625 assert(_size == -1, "already set?");
3626 _size = align_size_up(offset, CodeEntryAlignment);
3627 }
3628
3629 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3630 MacroAssembler _masm(&cb);
3631 for (int i = 0; i < _constants.length(); i++) {
3632 Constant con = _constants.at(i);
3633 address constant_addr;
3634 switch (con.type()) {
3635 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3636 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3637 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3638 case T_OBJECT: {
3639 jobject obj = con.get_jobject();
3640 int oop_index = _masm.oop_recorder()->find_index(obj);
3641 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3642 break;
3643 }
3644 case T_ADDRESS: {
3645 address addr = (address) con.get_jobject();
3646 constant_addr = _masm.address_constant(addr);
3647 break;
3648 }
3649 // We use T_VOID as marker for jump-table entries (labels) which
3650 // need an internal word relocation.
3651 case T_VOID: {
3652 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3653 // Fill the jump-table with a dummy word. The real value is
3654 // filled in later in fill_jump_table.
3655 address dummy = (address) n;
3656 constant_addr = _masm.address_constant(dummy);
3657 // Expand jump-table
3658 for (uint i = 1; i < n->outcnt(); i++) {
3659 address temp_addr = _masm.address_constant(dummy + i);
3660 assert(temp_addr, "consts section too small");
3661 }
3662 break;
3663 }
3664 case T_METADATA: {
3665 Metadata* obj = con.get_metadata();
3666 int metadata_index = _masm.oop_recorder()->find_index(obj);
3667 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3668 break;
3669 }
3670 default: ShouldNotReachHere();
3671 }
3672 assert(constant_addr, "consts section too small");
3673 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3674 }
3675 }
3676
3677 int Compile::ConstantTable::find_offset(Constant& con) const {
3678 int idx = _constants.find(con);
3679 assert(idx != -1, "constant must be in constant table");
3680 int offset = _constants.at(idx).offset();
3681 assert(offset != -1, "constant table not emitted yet?");
3682 return offset;
3683 }
3684
3685 void Compile::ConstantTable::add(Constant& con) {
3686 if (con.can_be_reused()) {
3687 int idx = _constants.find(con);
3688 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3689 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3690 return;
3691 }
3692 }
3693 (void) _constants.append(con);
3694 }
3695
3696 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3697 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3698 Constant con(type, value, b->_freq);
3699 add(con);
3700 return con;
3701 }
3702
3703 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3704 Constant con(metadata);
3705 add(con);
3706 return con;
3707 }
3708
3709 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3710 jvalue value;
3711 BasicType type = oper->type()->basic_type();
3712 switch (type) {
3713 case T_LONG: value.j = oper->constantL(); break;
3714 case T_FLOAT: value.f = oper->constantF(); break;
3715 case T_DOUBLE: value.d = oper->constantD(); break;
3716 case T_OBJECT:
3717 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3718 case T_METADATA: return add((Metadata*)oper->constant()); break;
3719 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3720 }
3721 return add(n, type, value);
3722 }
3723
3724 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3725 jvalue value;
3726 // We can use the node pointer here to identify the right jump-table
3727 // as this method is called from Compile::Fill_buffer right before
3728 // the MachNodes are emitted and the jump-table is filled (means the
3729 // MachNode pointers do not change anymore).
3730 value.l = (jobject) n;
3731 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3732 add(con);
3733 return con;
3734 }
3735
3736 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3737 // If called from Compile::scratch_emit_size do nothing.
3738 if (Compile::current()->in_scratch_emit_size()) return;
3739
3740 assert(labels.is_nonempty(), "must be");
3741 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3742
3743 // Since MachConstantNode::constant_offset() also contains
3744 // table_base_offset() we need to subtract the table_base_offset()
3745 // to get the plain offset into the constant table.
3746 int offset = n->constant_offset() - table_base_offset();
3747
3748 MacroAssembler _masm(&cb);
3749 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3750
3751 for (uint i = 0; i < n->outcnt(); i++) {
3752 address* constant_addr = &jump_table_base[i];
3753 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)));
3754 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3755 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3756 }
3757 }
3758
3759 void Compile::dump_inlining() {
3760 if (print_inlining() || print_intrinsics()) {
3761 // Print inlining message for candidates that we couldn't inline
3762 // for lack of space or non constant receiver
3763 for (int i = 0; i < _late_inlines.length(); i++) {
3764 CallGenerator* cg = _late_inlines.at(i);
3765 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3766 }
3767 Unique_Node_List useful;
3768 useful.push(root());
3769 for (uint next = 0; next < useful.size(); ++next) {
3770 Node* n = useful.at(next);
3771 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3772 CallNode* call = n->as_Call();
3773 CallGenerator* cg = call->generator();
3774 cg->print_inlining_late("receiver not constant");
3775 }
3776 uint max = n->len();
3777 for ( uint i = 0; i < max; ++i ) {
3778 Node *m = n->in(i);
3779 if ( m == NULL ) continue;
3780 useful.push(m);
3781 }
3782 }
3783 for (int i = 0; i < _print_inlining_list->length(); i++) {
3784 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3785 }
3786 }
3787 }
3788
3789 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3790 if (n1->Opcode() < n2->Opcode()) return -1;
3791 else if (n1->Opcode() > n2->Opcode()) return 1;
3792
3793 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()));
3794 for (uint i = 1; i < n1->req(); i++) {
3795 if (n1->in(i) < n2->in(i)) return -1;
3796 else if (n1->in(i) > n2->in(i)) return 1;
3797 }
3798
3799 return 0;
3800 }
3801
3802 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3803 Node* n1 = *n1p;
3804 Node* n2 = *n2p;
3805
3806 return cmp_expensive_nodes(n1, n2);
3807 }
3808
3809 void Compile::sort_expensive_nodes() {
3810 if (!expensive_nodes_sorted()) {
3811 _expensive_nodes->sort(cmp_expensive_nodes);
3812 }
3813 }
3814
3815 bool Compile::expensive_nodes_sorted() const {
3816 for (int i = 1; i < _expensive_nodes->length(); i++) {
3817 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3818 return false;
3819 }
3820 }
3821 return true;
3822 }
3823
3824 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3825 if (_expensive_nodes->length() == 0) {
3826 return false;
3827 }
3828
3829 assert(OptimizeExpensiveOps, "optimization off?");
3830
3831 // Take this opportunity to remove dead nodes from the list
3832 int j = 0;
3833 for (int i = 0; i < _expensive_nodes->length(); i++) {
3834 Node* n = _expensive_nodes->at(i);
3835 if (!n->is_unreachable(igvn)) {
3836 assert(n->is_expensive(), "should be expensive");
3837 _expensive_nodes->at_put(j, n);
3838 j++;
3839 }
3840 }
3841 _expensive_nodes->trunc_to(j);
3842
3843 // Then sort the list so that similar nodes are next to each other
3844 // and check for at least two nodes of identical kind with same data
3845 // inputs.
3846 sort_expensive_nodes();
3847
3848 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3849 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3850 return true;
3851 }
3852 }
3853
3854 return false;
3855 }
3856
3857 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3858 if (_expensive_nodes->length() == 0) {
3859 return;
3860 }
3861
3862 assert(OptimizeExpensiveOps, "optimization off?");
3863
3864 // Sort to bring similar nodes next to each other and clear the
3865 // control input of nodes for which there's only a single copy.
3866 sort_expensive_nodes();
3867
3868 int j = 0;
3869 int identical = 0;
3870 int i = 0;
3871 for (; i < _expensive_nodes->length()-1; i++) {
3872 assert(j <= i, "can't write beyond current index");
3873 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3874 identical++;
3875 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3876 continue;
3877 }
3878 if (identical > 0) {
3879 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3880 identical = 0;
3881 } else {
3882 Node* n = _expensive_nodes->at(i);
3883 igvn.hash_delete(n);
3884 n->set_req(0, NULL);
3885 igvn.hash_insert(n);
3886 }
3887 }
3888 if (identical > 0) {
3889 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3890 } else if (_expensive_nodes->length() >= 1) {
3891 Node* n = _expensive_nodes->at(i);
3892 igvn.hash_delete(n);
3893 n->set_req(0, NULL);
3894 igvn.hash_insert(n);
3895 }
3896 _expensive_nodes->trunc_to(j);
3897 }
3898
3899 void Compile::add_expensive_node(Node * n) {
3900 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3901 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3902 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3903 if (OptimizeExpensiveOps) {
3904 _expensive_nodes->append(n);
3905 } else {
3906 // Clear control input and let IGVN optimize expensive nodes if
3907 // OptimizeExpensiveOps is off.
3908 n->set_req(0, NULL);
3909 }
3910 }
3911
3912 /**
3913 * Remove the speculative part of types and clean up the graph
3914 */
3915 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3916 if (UseTypeSpeculation) {
3917 Unique_Node_List worklist;
3918 worklist.push(root());
3919 int modified = 0;
3920 // Go over all type nodes that carry a speculative type, drop the
3921 // speculative part of the type and enqueue the node for an igvn
3922 // which may optimize it out.
3923 for (uint next = 0; next < worklist.size(); ++next) {
3924 Node *n = worklist.at(next);
3925 if (n->is_Type() && n->as_Type()->type()->isa_oopptr() != NULL &&
3926 n->as_Type()->type()->is_oopptr()->speculative() != NULL) {
3927 TypeNode* tn = n->as_Type();
3928 const TypeOopPtr* t = tn->type()->is_oopptr();
3929 bool in_hash = igvn.hash_delete(n);
3930 assert(in_hash, "node should be in igvn hash table");
3931 tn->set_type(t->remove_speculative());
3932 igvn.hash_insert(n);
3933 igvn._worklist.push(n); // give it a chance to go away
3934 modified++;
3935 }
3936 uint max = n->len();
3937 for( uint i = 0; i < max; ++i ) {
3938 Node *m = n->in(i);
3939 if (not_a_node(m)) continue;
3940 worklist.push(m);
3941 }
3942 }
3943 // Drop the speculative part of all types in the igvn's type table
3944 igvn.remove_speculative_types();
3945 if (modified > 0) {
3946 igvn.optimize();
3947 }
3948 }
3949 }
3950
3951 // Auxiliary method to support randomized stressing/fuzzing.
3952 //
3953 // This method can be called the arbitrary number of times, with current count
3954 // as the argument. The logic allows selecting a single candidate from the
3955 // running list of candidates as follows:
3956 // int count = 0;
3957 // Cand* selected = null;
3958 // while(cand = cand->next()) {
3959 // if (randomized_select(++count)) {
3960 // selected = cand;
3961 // }
3962 // }
3963 //
3964 // Including count equalizes the chances any candidate is "selected".
3965 // This is useful when we don't have the complete list of candidates to choose
3966 // from uniformly. In this case, we need to adjust the randomicity of the
3967 // selection, or else we will end up biasing the selection towards the latter
3968 // candidates.
3969 //
3970 // Quick back-envelope calculation shows that for the list of n candidates
3971 // the equal probability for the candidate to persist as "best" can be
3972 // achieved by replacing it with "next" k-th candidate with the probability
3973 // of 1/k. It can be easily shown that by the end of the run, the
3974 // probability for any candidate is converged to 1/n, thus giving the
3975 // uniform distribution among all the candidates.
3976 //
3977 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
3978 #define RANDOMIZED_DOMAIN_POW 29
3979 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
3980 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
3981 bool Compile::randomized_select(int count) {
3982 assert(count > 0, "only positive");
3983 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
3984 }