1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/c2compiler.hpp"
38 #include "opto/callGenerator.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/cfgnode.hpp"
41 #include "opto/chaitin.hpp"
42 #include "opto/compile.hpp"
43 #include "opto/connode.hpp"
44 #include "opto/divnode.hpp"
45 #include "opto/escape.hpp"
46 #include "opto/idealGraphPrinter.hpp"
47 #include "opto/loopnode.hpp"
48 #include "opto/machnode.hpp"
49 #include "opto/macro.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/mathexactnode.hpp"
52 #include "opto/memnode.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/narrowptrnode.hpp"
55 #include "opto/node.hpp"
56 #include "opto/opcodes.hpp"
57 #include "opto/output.hpp"
58 #include "opto/parse.hpp"
59 #include "opto/phaseX.hpp"
60 #include "opto/rootnode.hpp"
61 #include "opto/runtime.hpp"
62 #include "opto/stringopts.hpp"
63 #include "opto/type.hpp"
64 #include "opto/vectornode.hpp"
65 #include "runtime/arguments.hpp"
66 #include "runtime/signature.hpp"
67 #include "runtime/stubRoutines.hpp"
68 #include "runtime/timer.hpp"
69 #include "trace/tracing.hpp"
70 #include "utilities/copy.hpp"
71 #ifdef TARGET_ARCH_MODEL_x86_32
72 # include "adfiles/ad_x86_32.hpp"
73 #endif
74 #ifdef TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #endif
77 #ifdef TARGET_ARCH_MODEL_sparc
78 # include "adfiles/ad_sparc.hpp"
79 #endif
80 #ifdef TARGET_ARCH_MODEL_zero
81 # include "adfiles/ad_zero.hpp"
82 #endif
83 #ifdef TARGET_ARCH_MODEL_arm
84 # include "adfiles/ad_arm.hpp"
85 #endif
86 #ifdef TARGET_ARCH_MODEL_ppc_32
87 # include "adfiles/ad_ppc_32.hpp"
88 #endif
89 #ifdef TARGET_ARCH_MODEL_ppc_64
90 # include "adfiles/ad_ppc_64.hpp"
91 #endif
92
93
94 // -------------------- Compile::mach_constant_base_node -----------------------
95 // Constant table base node singleton.
96 MachConstantBaseNode* Compile::mach_constant_base_node() {
97 if (_mach_constant_base_node == NULL) {
98 _mach_constant_base_node = new (C) MachConstantBaseNode();
99 _mach_constant_base_node->add_req(C->root());
100 }
101 return _mach_constant_base_node;
102 }
103
104
105 /// Support for intrinsics.
106
107 // Return the index at which m must be inserted (or already exists).
108 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
109 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
110 #ifdef ASSERT
111 for (int i = 1; i < _intrinsics->length(); i++) {
112 CallGenerator* cg1 = _intrinsics->at(i-1);
113 CallGenerator* cg2 = _intrinsics->at(i);
114 assert(cg1->method() != cg2->method()
115 ? cg1->method() < cg2->method()
116 : cg1->is_virtual() < cg2->is_virtual(),
117 "compiler intrinsics list must stay sorted");
118 }
119 #endif
120 // Binary search sorted list, in decreasing intervals [lo, hi].
121 int lo = 0, hi = _intrinsics->length()-1;
122 while (lo <= hi) {
123 int mid = (uint)(hi + lo) / 2;
124 ciMethod* mid_m = _intrinsics->at(mid)->method();
125 if (m < mid_m) {
126 hi = mid-1;
127 } else if (m > mid_m) {
128 lo = mid+1;
129 } else {
130 // look at minor sort key
131 bool mid_virt = _intrinsics->at(mid)->is_virtual();
132 if (is_virtual < mid_virt) {
133 hi = mid-1;
134 } else if (is_virtual > mid_virt) {
135 lo = mid+1;
136 } else {
137 return mid; // exact match
138 }
139 }
140 }
141 return lo; // inexact match
142 }
143
144 void Compile::register_intrinsic(CallGenerator* cg) {
145 if (_intrinsics == NULL) {
146 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
147 }
148 // This code is stolen from ciObjectFactory::insert.
149 // Really, GrowableArray should have methods for
150 // insert_at, remove_at, and binary_search.
151 int len = _intrinsics->length();
152 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
153 if (index == len) {
154 _intrinsics->append(cg);
155 } else {
156 #ifdef ASSERT
157 CallGenerator* oldcg = _intrinsics->at(index);
158 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
159 #endif
160 _intrinsics->append(_intrinsics->at(len-1));
161 int pos;
162 for (pos = len-2; pos >= index; pos--) {
163 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
164 }
165 _intrinsics->at_put(index, cg);
166 }
167 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
168 }
169
170 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
171 assert(m->is_loaded(), "don't try this on unloaded methods");
172 if (_intrinsics != NULL) {
173 int index = intrinsic_insertion_index(m, is_virtual);
174 if (index < _intrinsics->length()
175 && _intrinsics->at(index)->method() == m
176 && _intrinsics->at(index)->is_virtual() == is_virtual) {
177 return _intrinsics->at(index);
178 }
179 }
180 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
181 if (m->intrinsic_id() != vmIntrinsics::_none &&
182 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
183 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
184 if (cg != NULL) {
185 // Save it for next time:
186 register_intrinsic(cg);
187 return cg;
188 } else {
189 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
190 }
191 }
192 return NULL;
193 }
194
195 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
196 // in library_call.cpp.
197
198
199 #ifndef PRODUCT
200 // statistics gathering...
201
202 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
203 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
204
205 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
206 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
207 int oflags = _intrinsic_hist_flags[id];
208 assert(flags != 0, "what happened?");
209 if (is_virtual) {
210 flags |= _intrinsic_virtual;
211 }
212 bool changed = (flags != oflags);
213 if ((flags & _intrinsic_worked) != 0) {
214 juint count = (_intrinsic_hist_count[id] += 1);
215 if (count == 1) {
216 changed = true; // first time
217 }
218 // increment the overall count also:
219 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
220 }
221 if (changed) {
222 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
223 // Something changed about the intrinsic's virtuality.
224 if ((flags & _intrinsic_virtual) != 0) {
225 // This is the first use of this intrinsic as a virtual call.
226 if (oflags != 0) {
227 // We already saw it as a non-virtual, so note both cases.
228 flags |= _intrinsic_both;
229 }
230 } else if ((oflags & _intrinsic_both) == 0) {
231 // This is the first use of this intrinsic as a non-virtual
232 flags |= _intrinsic_both;
233 }
234 }
235 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
236 }
237 // update the overall flags also:
238 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
239 return changed;
240 }
241
242 static char* format_flags(int flags, char* buf) {
243 buf[0] = 0;
244 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
245 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
246 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
247 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
248 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
249 if (buf[0] == 0) strcat(buf, ",");
250 assert(buf[0] == ',', "must be");
251 return &buf[1];
252 }
253
254 void Compile::print_intrinsic_statistics() {
255 char flagsbuf[100];
256 ttyLocker ttyl;
257 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
258 tty->print_cr("Compiler intrinsic usage:");
259 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
260 if (total == 0) total = 1; // avoid div0 in case of no successes
261 #define PRINT_STAT_LINE(name, c, f) \
262 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
263 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
264 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
265 int flags = _intrinsic_hist_flags[id];
266 juint count = _intrinsic_hist_count[id];
267 if ((flags | count) != 0) {
268 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
269 }
270 }
271 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
272 if (xtty != NULL) xtty->tail("statistics");
273 }
274
275 void Compile::print_statistics() {
276 { ttyLocker ttyl;
277 if (xtty != NULL) xtty->head("statistics type='opto'");
278 Parse::print_statistics();
279 PhaseCCP::print_statistics();
280 PhaseRegAlloc::print_statistics();
281 Scheduling::print_statistics();
282 PhasePeephole::print_statistics();
283 PhaseIdealLoop::print_statistics();
284 if (xtty != NULL) xtty->tail("statistics");
285 }
286 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
287 // put this under its own <statistics> element.
288 print_intrinsic_statistics();
289 }
290 }
291 #endif //PRODUCT
292
293 // Support for bundling info
294 Bundle* Compile::node_bundling(const Node *n) {
295 assert(valid_bundle_info(n), "oob");
296 return &_node_bundling_base[n->_idx];
297 }
298
299 bool Compile::valid_bundle_info(const Node *n) {
300 return (_node_bundling_limit > n->_idx);
301 }
302
303
304 void Compile::gvn_replace_by(Node* n, Node* nn) {
305 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
306 Node* use = n->last_out(i);
307 bool is_in_table = initial_gvn()->hash_delete(use);
308 uint uses_found = 0;
309 for (uint j = 0; j < use->len(); j++) {
310 if (use->in(j) == n) {
311 if (j < use->req())
312 use->set_req(j, nn);
313 else
314 use->set_prec(j, nn);
315 uses_found++;
316 }
317 }
318 if (is_in_table) {
319 // reinsert into table
320 initial_gvn()->hash_find_insert(use);
321 }
322 record_for_igvn(use);
323 i -= uses_found; // we deleted 1 or more copies of this edge
324 }
325 }
326
327
328 static inline bool not_a_node(const Node* n) {
329 if (n == NULL) return true;
330 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
331 if (*(address*)n == badAddress) return true; // kill by Node::destruct
332 return false;
333 }
334
335 // Identify all nodes that are reachable from below, useful.
336 // Use breadth-first pass that records state in a Unique_Node_List,
337 // recursive traversal is slower.
338 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
339 int estimated_worklist_size = unique();
340 useful.map( estimated_worklist_size, NULL ); // preallocate space
341
342 // Initialize worklist
343 if (root() != NULL) { useful.push(root()); }
344 // If 'top' is cached, declare it useful to preserve cached node
345 if( cached_top_node() ) { useful.push(cached_top_node()); }
346
347 // Push all useful nodes onto the list, breadthfirst
348 for( uint next = 0; next < useful.size(); ++next ) {
349 assert( next < unique(), "Unique useful nodes < total nodes");
350 Node *n = useful.at(next);
351 uint max = n->len();
352 for( uint i = 0; i < max; ++i ) {
353 Node *m = n->in(i);
354 if (not_a_node(m)) continue;
355 useful.push(m);
356 }
357 }
358 }
359
360 // Update dead_node_list with any missing dead nodes using useful
361 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
362 void Compile::update_dead_node_list(Unique_Node_List &useful) {
363 uint max_idx = unique();
364 VectorSet& useful_node_set = useful.member_set();
365
366 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
367 // If node with index node_idx is not in useful set,
368 // mark it as dead in dead node list.
369 if (! useful_node_set.test(node_idx) ) {
370 record_dead_node(node_idx);
371 }
372 }
373 }
374
375 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
376 int shift = 0;
377 for (int i = 0; i < inlines->length(); i++) {
378 CallGenerator* cg = inlines->at(i);
379 CallNode* call = cg->call_node();
380 if (shift > 0) {
381 inlines->at_put(i-shift, cg);
382 }
383 if (!useful.member(call)) {
384 shift++;
385 }
386 }
387 inlines->trunc_to(inlines->length()-shift);
388 }
389
390 // Disconnect all useless nodes by disconnecting those at the boundary.
391 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
392 uint next = 0;
393 while (next < useful.size()) {
394 Node *n = useful.at(next++);
395 // Use raw traversal of out edges since this code removes out edges
396 int max = n->outcnt();
397 for (int j = 0; j < max; ++j) {
398 Node* child = n->raw_out(j);
399 if (! useful.member(child)) {
400 assert(!child->is_top() || child != top(),
401 "If top is cached in Compile object it is in useful list");
402 // Only need to remove this out-edge to the useless node
403 n->raw_del_out(j);
404 --j;
405 --max;
406 }
407 }
408 if (n->outcnt() == 1 && n->has_special_unique_user()) {
409 record_for_igvn(n->unique_out());
410 }
411 }
412 // Remove useless macro and predicate opaq nodes
413 for (int i = C->macro_count()-1; i >= 0; i--) {
414 Node* n = C->macro_node(i);
415 if (!useful.member(n)) {
416 remove_macro_node(n);
417 }
418 }
419 // Remove useless expensive node
420 for (int i = C->expensive_count()-1; i >= 0; i--) {
421 Node* n = C->expensive_node(i);
422 if (!useful.member(n)) {
423 remove_expensive_node(n);
424 }
425 }
426 // clean up the late inline lists
427 remove_useless_late_inlines(&_string_late_inlines, useful);
428 remove_useless_late_inlines(&_boxing_late_inlines, useful);
429 remove_useless_late_inlines(&_late_inlines, useful);
430 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
431 }
432
433 //------------------------------frame_size_in_words-----------------------------
434 // frame_slots in units of words
435 int Compile::frame_size_in_words() const {
436 // shift is 0 in LP32 and 1 in LP64
437 const int shift = (LogBytesPerWord - LogBytesPerInt);
438 int words = _frame_slots >> shift;
439 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
440 return words;
441 }
442
443 // ============================================================================
444 //------------------------------CompileWrapper---------------------------------
445 class CompileWrapper : public StackObj {
446 Compile *const _compile;
447 public:
448 CompileWrapper(Compile* compile);
449
450 ~CompileWrapper();
451 };
452
453 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
454 // the Compile* pointer is stored in the current ciEnv:
455 ciEnv* env = compile->env();
456 assert(env == ciEnv::current(), "must already be a ciEnv active");
457 assert(env->compiler_data() == NULL, "compile already active?");
458 env->set_compiler_data(compile);
459 assert(compile == Compile::current(), "sanity");
460
461 compile->set_type_dict(NULL);
462 compile->set_type_hwm(NULL);
463 compile->set_type_last_size(0);
464 compile->set_last_tf(NULL, NULL);
465 compile->set_indexSet_arena(NULL);
466 compile->set_indexSet_free_block_list(NULL);
467 compile->init_type_arena();
468 Type::Initialize(compile);
469 _compile->set_scratch_buffer_blob(NULL);
470 _compile->begin_method();
471 }
472 CompileWrapper::~CompileWrapper() {
473 _compile->end_method();
474 if (_compile->scratch_buffer_blob() != NULL)
475 BufferBlob::free(_compile->scratch_buffer_blob());
476 _compile->env()->set_compiler_data(NULL);
477 }
478
479
480 //----------------------------print_compile_messages---------------------------
481 void Compile::print_compile_messages() {
482 #ifndef PRODUCT
483 // Check if recompiling
484 if (_subsume_loads == false && PrintOpto) {
485 // Recompiling without allowing machine instructions to subsume loads
486 tty->print_cr("*********************************************************");
487 tty->print_cr("** Bailout: Recompile without subsuming loads **");
488 tty->print_cr("*********************************************************");
489 }
490 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
491 // Recompiling without escape analysis
492 tty->print_cr("*********************************************************");
493 tty->print_cr("** Bailout: Recompile without escape analysis **");
494 tty->print_cr("*********************************************************");
495 }
496 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
497 // Recompiling without boxing elimination
498 tty->print_cr("*********************************************************");
499 tty->print_cr("** Bailout: Recompile without boxing elimination **");
500 tty->print_cr("*********************************************************");
501 }
502 if (env()->break_at_compile()) {
503 // Open the debugger when compiling this method.
504 tty->print("### Breaking when compiling: ");
505 method()->print_short_name();
506 tty->cr();
507 BREAKPOINT;
508 }
509
510 if( PrintOpto ) {
511 if (is_osr_compilation()) {
512 tty->print("[OSR]%3d", _compile_id);
513 } else {
514 tty->print("%3d", _compile_id);
515 }
516 }
517 #endif
518 }
519
520
521 //-----------------------init_scratch_buffer_blob------------------------------
522 // Construct a temporary BufferBlob and cache it for this compile.
523 void Compile::init_scratch_buffer_blob(int const_size) {
524 // If there is already a scratch buffer blob allocated and the
525 // constant section is big enough, use it. Otherwise free the
526 // current and allocate a new one.
527 BufferBlob* blob = scratch_buffer_blob();
528 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
529 // Use the current blob.
530 } else {
531 if (blob != NULL) {
532 BufferBlob::free(blob);
533 }
534
535 ResourceMark rm;
536 _scratch_const_size = const_size;
537 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
538 blob = BufferBlob::create("Compile::scratch_buffer", size);
539 // Record the buffer blob for next time.
540 set_scratch_buffer_blob(blob);
541 // Have we run out of code space?
542 if (scratch_buffer_blob() == NULL) {
543 // Let CompilerBroker disable further compilations.
544 record_failure("Not enough space for scratch buffer in CodeCache");
545 return;
546 }
547 }
548
549 // Initialize the relocation buffers
550 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
551 set_scratch_locs_memory(locs_buf);
552 }
553
554
555 //-----------------------scratch_emit_size-------------------------------------
556 // Helper function that computes size by emitting code
557 uint Compile::scratch_emit_size(const Node* n) {
558 // Start scratch_emit_size section.
559 set_in_scratch_emit_size(true);
560
561 // Emit into a trash buffer and count bytes emitted.
562 // This is a pretty expensive way to compute a size,
563 // but it works well enough if seldom used.
564 // All common fixed-size instructions are given a size
565 // method by the AD file.
566 // Note that the scratch buffer blob and locs memory are
567 // allocated at the beginning of the compile task, and
568 // may be shared by several calls to scratch_emit_size.
569 // The allocation of the scratch buffer blob is particularly
570 // expensive, since it has to grab the code cache lock.
571 BufferBlob* blob = this->scratch_buffer_blob();
572 assert(blob != NULL, "Initialize BufferBlob at start");
573 assert(blob->size() > MAX_inst_size, "sanity");
574 relocInfo* locs_buf = scratch_locs_memory();
575 address blob_begin = blob->content_begin();
576 address blob_end = (address)locs_buf;
577 assert(blob->content_contains(blob_end), "sanity");
578 CodeBuffer buf(blob_begin, blob_end - blob_begin);
579 buf.initialize_consts_size(_scratch_const_size);
580 buf.initialize_stubs_size(MAX_stubs_size);
581 assert(locs_buf != NULL, "sanity");
582 int lsize = MAX_locs_size / 3;
583 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
584 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
585 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
586
587 // Do the emission.
588
589 Label fakeL; // Fake label for branch instructions.
590 Label* saveL = NULL;
591 uint save_bnum = 0;
592 bool is_branch = n->is_MachBranch();
593 if (is_branch) {
594 MacroAssembler masm(&buf);
595 masm.bind(fakeL);
596 n->as_MachBranch()->save_label(&saveL, &save_bnum);
597 n->as_MachBranch()->label_set(&fakeL, 0);
598 }
599 n->emit(buf, this->regalloc());
600 if (is_branch) // Restore label.
601 n->as_MachBranch()->label_set(saveL, save_bnum);
602
603 // End scratch_emit_size section.
604 set_in_scratch_emit_size(false);
605
606 return buf.insts_size();
607 }
608
609
610 // ============================================================================
611 //------------------------------Compile standard-------------------------------
612 debug_only( int Compile::_debug_idx = 100000; )
613
614 // Compile a method. entry_bci is -1 for normal compilations and indicates
615 // the continuation bci for on stack replacement.
616
617
618 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
619 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
620 : Phase(Compiler),
621 _env(ci_env),
622 _log(ci_env->log()),
623 _compile_id(ci_env->compile_id()),
624 _save_argument_registers(false),
625 _stub_name(NULL),
626 _stub_function(NULL),
627 _stub_entry_point(NULL),
628 _method(target),
629 _entry_bci(osr_bci),
630 _initial_gvn(NULL),
631 _for_igvn(NULL),
632 _warm_calls(NULL),
633 _subsume_loads(subsume_loads),
634 _do_escape_analysis(do_escape_analysis),
635 _eliminate_boxing(eliminate_boxing),
636 _failure_reason(NULL),
637 _code_buffer("Compile::Fill_buffer"),
638 _orig_pc_slot(0),
639 _orig_pc_slot_offset_in_bytes(0),
640 _has_method_handle_invokes(false),
641 _mach_constant_base_node(NULL),
642 _node_bundling_limit(0),
643 _node_bundling_base(NULL),
644 _java_calls(0),
645 _inner_loops(0),
646 _scratch_const_size(-1),
647 _in_scratch_emit_size(false),
648 _dead_node_list(comp_arena()),
649 _dead_node_count(0),
650 #ifndef PRODUCT
651 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
652 _in_dump_cnt(0),
653 _printer(IdealGraphPrinter::printer()),
654 #endif
655 _congraph(NULL),
656 _replay_inline_data(NULL),
657 _late_inlines(comp_arena(), 2, 0, NULL),
658 _string_late_inlines(comp_arena(), 2, 0, NULL),
659 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
660 _late_inlines_pos(0),
661 _number_of_mh_late_inlines(0),
662 _inlining_progress(false),
663 _inlining_incrementally(false),
664 _print_inlining_list(NULL),
665 _print_inlining_idx(0),
666 _preserve_jvm_state(0) {
667 C = this;
668
669 CompileWrapper cw(this);
670 #ifndef PRODUCT
671 if (TimeCompiler2) {
672 tty->print(" ");
673 target->holder()->name()->print();
674 tty->print(".");
675 target->print_short_name();
676 tty->print(" ");
677 }
678 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
679 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
680 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
681 if (!print_opto_assembly) {
682 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
683 if (print_assembly && !Disassembler::can_decode()) {
684 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
685 print_opto_assembly = true;
686 }
687 }
688 set_print_assembly(print_opto_assembly);
689 set_parsed_irreducible_loop(false);
690
691 if (method()->has_option("ReplayInline")) {
692 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
693 }
694 #endif
695 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
696 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
697
698 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
699 // Make sure the method being compiled gets its own MDO,
700 // so we can at least track the decompile_count().
701 // Need MDO to record RTM code generation state.
702 method()->ensure_method_data();
703 }
704
705 Init(::AliasLevel);
706
707
708 print_compile_messages();
709
710 _ilt = InlineTree::build_inline_tree_root();
711
712 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
713 assert(num_alias_types() >= AliasIdxRaw, "");
714
715 #define MINIMUM_NODE_HASH 1023
716 // Node list that Iterative GVN will start with
717 Unique_Node_List for_igvn(comp_arena());
718 set_for_igvn(&for_igvn);
719
720 // GVN that will be run immediately on new nodes
721 uint estimated_size = method()->code_size()*4+64;
722 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
723 PhaseGVN gvn(node_arena(), estimated_size);
724 set_initial_gvn(&gvn);
725
726 if (print_inlining() || print_intrinsics()) {
727 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
728 }
729 { // Scope for timing the parser
730 TracePhase t3("parse", &_t_parser, true);
731
732 // Put top into the hash table ASAP.
733 initial_gvn()->transform_no_reclaim(top());
734
735 // Set up tf(), start(), and find a CallGenerator.
736 CallGenerator* cg = NULL;
737 if (is_osr_compilation()) {
738 const TypeTuple *domain = StartOSRNode::osr_domain();
739 const TypeTuple *range = TypeTuple::make_range(method()->signature());
740 init_tf(TypeFunc::make(domain, range));
741 StartNode* s = new (this) StartOSRNode(root(), domain);
742 initial_gvn()->set_type_bottom(s);
743 init_start(s);
744 cg = CallGenerator::for_osr(method(), entry_bci());
745 } else {
746 // Normal case.
747 init_tf(TypeFunc::make(method()));
748 StartNode* s = new (this) StartNode(root(), tf()->domain());
749 initial_gvn()->set_type_bottom(s);
750 init_start(s);
751 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
752 // With java.lang.ref.reference.get() we must go through the
753 // intrinsic when G1 is enabled - even when get() is the root
754 // method of the compile - so that, if necessary, the value in
755 // the referent field of the reference object gets recorded by
756 // the pre-barrier code.
757 // Specifically, if G1 is enabled, the value in the referent
758 // field is recorded by the G1 SATB pre barrier. This will
759 // result in the referent being marked live and the reference
760 // object removed from the list of discovered references during
761 // reference processing.
762 cg = find_intrinsic(method(), false);
763 }
764 if (cg == NULL) {
765 float past_uses = method()->interpreter_invocation_count();
766 float expected_uses = past_uses;
767 cg = CallGenerator::for_inline(method(), expected_uses);
768 }
769 }
770 if (failing()) return;
771 if (cg == NULL) {
772 record_method_not_compilable_all_tiers("cannot parse method");
773 return;
774 }
775 JVMState* jvms = build_start_state(start(), tf());
776 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
777 record_method_not_compilable("method parse failed");
778 return;
779 }
780 GraphKit kit(jvms);
781
782 if (!kit.stopped()) {
783 // Accept return values, and transfer control we know not where.
784 // This is done by a special, unique ReturnNode bound to root.
785 return_values(kit.jvms());
786 }
787
788 if (kit.has_exceptions()) {
789 // Any exceptions that escape from this call must be rethrown
790 // to whatever caller is dynamically above us on the stack.
791 // This is done by a special, unique RethrowNode bound to root.
792 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
793 }
794
795 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
796
797 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
798 inline_string_calls(true);
799 }
800
801 if (failing()) return;
802
803 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
804
805 // Remove clutter produced by parsing.
806 if (!failing()) {
807 ResourceMark rm;
808 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
809 }
810 }
811
812 // Note: Large methods are capped off in do_one_bytecode().
813 if (failing()) return;
814
815 // After parsing, node notes are no longer automagic.
816 // They must be propagated by register_new_node_with_optimizer(),
817 // clone(), or the like.
818 set_default_node_notes(NULL);
819
820 for (;;) {
821 int successes = Inline_Warm();
822 if (failing()) return;
823 if (successes == 0) break;
824 }
825
826 // Drain the list.
827 Finish_Warm();
828 #ifndef PRODUCT
829 if (_printer) {
830 _printer->print_inlining(this);
831 }
832 #endif
833
834 if (failing()) return;
835 NOT_PRODUCT( verify_graph_edges(); )
836
837 // Now optimize
838 Optimize();
839 if (failing()) return;
840 NOT_PRODUCT( verify_graph_edges(); )
841
842 #ifndef PRODUCT
843 if (PrintIdeal) {
844 ttyLocker ttyl; // keep the following output all in one block
845 // This output goes directly to the tty, not the compiler log.
846 // To enable tools to match it up with the compilation activity,
847 // be sure to tag this tty output with the compile ID.
848 if (xtty != NULL) {
849 xtty->head("ideal compile_id='%d'%s", compile_id(),
850 is_osr_compilation() ? " compile_kind='osr'" :
851 "");
852 }
853 root()->dump(9999);
854 if (xtty != NULL) {
855 xtty->tail("ideal");
856 }
857 }
858 #endif
859
860 NOT_PRODUCT( verify_barriers(); )
861
862 // Dump compilation data to replay it.
863 if (method()->has_option("DumpReplay")) {
864 env()->dump_replay_data(_compile_id);
865 }
866 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
867 env()->dump_inline_data(_compile_id);
868 }
869
870 // Now that we know the size of all the monitors we can add a fixed slot
871 // for the original deopt pc.
872
873 _orig_pc_slot = fixed_slots();
874 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
875 set_fixed_slots(next_slot);
876
877 // Compute when to use implicit null checks. Used by matching trap based
878 // nodes and NullCheck optimization.
879 set_allowed_deopt_reasons();
880
881 // Now generate code
882 Code_Gen();
883 if (failing()) return;
884
885 // Check if we want to skip execution of all compiled code.
886 {
887 #ifndef PRODUCT
888 if (OptoNoExecute) {
889 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
890 return;
891 }
892 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
893 #endif
894
895 if (is_osr_compilation()) {
896 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
897 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
898 } else {
899 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
900 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
901 }
902
903 env()->register_method(_method, _entry_bci,
904 &_code_offsets,
905 _orig_pc_slot_offset_in_bytes,
906 code_buffer(),
907 frame_size_in_words(), _oop_map_set,
908 &_handler_table, &_inc_table,
909 compiler,
910 env()->comp_level(),
911 has_unsafe_access(),
912 SharedRuntime::is_wide_vector(max_vector_size()),
913 rtm_state()
914 );
915
916 if (log() != NULL) // Print code cache state into compiler log
917 log()->code_cache_state();
918 }
919 }
920
921 //------------------------------Compile----------------------------------------
922 // Compile a runtime stub
923 Compile::Compile( ciEnv* ci_env,
924 TypeFunc_generator generator,
925 address stub_function,
926 const char *stub_name,
927 int is_fancy_jump,
928 bool pass_tls,
929 bool save_arg_registers,
930 bool return_pc )
931 : Phase(Compiler),
932 _env(ci_env),
933 _log(ci_env->log()),
934 _compile_id(0),
935 _save_argument_registers(save_arg_registers),
936 _method(NULL),
937 _stub_name(stub_name),
938 _stub_function(stub_function),
939 _stub_entry_point(NULL),
940 _entry_bci(InvocationEntryBci),
941 _initial_gvn(NULL),
942 _for_igvn(NULL),
943 _warm_calls(NULL),
944 _orig_pc_slot(0),
945 _orig_pc_slot_offset_in_bytes(0),
946 _subsume_loads(true),
947 _do_escape_analysis(false),
948 _eliminate_boxing(false),
949 _failure_reason(NULL),
950 _code_buffer("Compile::Fill_buffer"),
951 _has_method_handle_invokes(false),
952 _mach_constant_base_node(NULL),
953 _node_bundling_limit(0),
954 _node_bundling_base(NULL),
955 _java_calls(0),
956 _inner_loops(0),
957 #ifndef PRODUCT
958 _trace_opto_output(TraceOptoOutput),
959 _in_dump_cnt(0),
960 _printer(NULL),
961 #endif
962 _dead_node_list(comp_arena()),
963 _dead_node_count(0),
964 _congraph(NULL),
965 _replay_inline_data(NULL),
966 _number_of_mh_late_inlines(0),
967 _inlining_progress(false),
968 _inlining_incrementally(false),
969 _print_inlining_list(NULL),
970 _print_inlining_idx(0),
971 _preserve_jvm_state(0),
972 _allowed_reasons(0) {
973 C = this;
974
975 #ifndef PRODUCT
976 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
977 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
978 set_print_assembly(PrintFrameConverterAssembly);
979 set_parsed_irreducible_loop(false);
980 #endif
981 CompileWrapper cw(this);
982 Init(/*AliasLevel=*/ 0);
983 init_tf((*generator)());
984
985 {
986 // The following is a dummy for the sake of GraphKit::gen_stub
987 Unique_Node_List for_igvn(comp_arena());
988 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
989 PhaseGVN gvn(Thread::current()->resource_area(),255);
990 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
991 gvn.transform_no_reclaim(top());
992
993 GraphKit kit;
994 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
995 }
996
997 NOT_PRODUCT( verify_graph_edges(); )
998 Code_Gen();
999 if (failing()) return;
1000
1001
1002 // Entry point will be accessed using compile->stub_entry_point();
1003 if (code_buffer() == NULL) {
1004 Matcher::soft_match_failure();
1005 } else {
1006 if (PrintAssembly && (WizardMode || Verbose))
1007 tty->print_cr("### Stub::%s", stub_name);
1008
1009 if (!failing()) {
1010 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1011
1012 // Make the NMethod
1013 // For now we mark the frame as never safe for profile stackwalking
1014 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1015 code_buffer(),
1016 CodeOffsets::frame_never_safe,
1017 // _code_offsets.value(CodeOffsets::Frame_Complete),
1018 frame_size_in_words(),
1019 _oop_map_set,
1020 save_arg_registers);
1021 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1022
1023 _stub_entry_point = rs->entry_point();
1024 }
1025 }
1026 }
1027
1028 //------------------------------Init-------------------------------------------
1029 // Prepare for a single compilation
1030 void Compile::Init(int aliaslevel) {
1031 _unique = 0;
1032 _regalloc = NULL;
1033
1034 _tf = NULL; // filled in later
1035 _top = NULL; // cached later
1036 _matcher = NULL; // filled in later
1037 _cfg = NULL; // filled in later
1038
1039 set_24_bit_selection_and_mode(Use24BitFP, false);
1040
1041 _node_note_array = NULL;
1042 _default_node_notes = NULL;
1043
1044 _immutable_memory = NULL; // filled in at first inquiry
1045
1046 // Globally visible Nodes
1047 // First set TOP to NULL to give safe behavior during creation of RootNode
1048 set_cached_top_node(NULL);
1049 set_root(new (this) RootNode());
1050 // Now that you have a Root to point to, create the real TOP
1051 set_cached_top_node( new (this) ConNode(Type::TOP) );
1052 set_recent_alloc(NULL, NULL);
1053
1054 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1055 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1056 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1057 env()->set_dependencies(new Dependencies(env()));
1058
1059 _fixed_slots = 0;
1060 set_has_split_ifs(false);
1061 set_has_loops(has_method() && method()->has_loops()); // first approximation
1062 set_has_stringbuilder(false);
1063 set_has_boxed_value(false);
1064 _trap_can_recompile = false; // no traps emitted yet
1065 _major_progress = true; // start out assuming good things will happen
1066 set_has_unsafe_access(false);
1067 set_max_vector_size(0);
1068 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1069 set_decompile_count(0);
1070
1071 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1072 set_num_loop_opts(LoopOptsCount);
1073 set_do_inlining(Inline);
1074 set_max_inline_size(MaxInlineSize);
1075 set_freq_inline_size(FreqInlineSize);
1076 set_do_scheduling(OptoScheduling);
1077 set_do_count_invocations(false);
1078 set_do_method_data_update(false);
1079 set_rtm_state(NoRTM); // No RTM lock eliding by default
1080 #if INCLUDE_RTM_OPT
1081 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1082 int rtm_state = method()->method_data()->rtm_state();
1083 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1084 // Don't generate RTM lock eliding code.
1085 set_rtm_state(NoRTM);
1086 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1087 // Generate RTM lock eliding code without abort ratio calculation code.
1088 set_rtm_state(UseRTM);
1089 } else if (UseRTMDeopt) {
1090 // Generate RTM lock eliding code and include abort ratio calculation
1091 // code if UseRTMDeopt is on.
1092 set_rtm_state(ProfileRTM);
1093 }
1094 }
1095 #endif
1096 if (debug_info()->recording_non_safepoints()) {
1097 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1098 (comp_arena(), 8, 0, NULL));
1099 set_default_node_notes(Node_Notes::make(this));
1100 }
1101
1102 // // -- Initialize types before each compile --
1103 // // Update cached type information
1104 // if( _method && _method->constants() )
1105 // Type::update_loaded_types(_method, _method->constants());
1106
1107 // Init alias_type map.
1108 if (!_do_escape_analysis && aliaslevel == 3)
1109 aliaslevel = 2; // No unique types without escape analysis
1110 _AliasLevel = aliaslevel;
1111 const int grow_ats = 16;
1112 _max_alias_types = grow_ats;
1113 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1114 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1115 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1116 {
1117 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1118 }
1119 // Initialize the first few types.
1120 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1121 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1122 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1123 _num_alias_types = AliasIdxRaw+1;
1124 // Zero out the alias type cache.
1125 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1126 // A NULL adr_type hits in the cache right away. Preload the right answer.
1127 probe_alias_cache(NULL)->_index = AliasIdxTop;
1128
1129 _intrinsics = NULL;
1130 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1131 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1132 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1133 register_library_intrinsics();
1134 }
1135
1136 //---------------------------init_start----------------------------------------
1137 // Install the StartNode on this compile object.
1138 void Compile::init_start(StartNode* s) {
1139 if (failing())
1140 return; // already failing
1141 assert(s == start(), "");
1142 }
1143
1144 StartNode* Compile::start() const {
1145 assert(!failing(), "");
1146 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1147 Node* start = root()->fast_out(i);
1148 if( start->is_Start() )
1149 return start->as_Start();
1150 }
1151 ShouldNotReachHere();
1152 return NULL;
1153 }
1154
1155 //-------------------------------immutable_memory-------------------------------------
1156 // Access immutable memory
1157 Node* Compile::immutable_memory() {
1158 if (_immutable_memory != NULL) {
1159 return _immutable_memory;
1160 }
1161 StartNode* s = start();
1162 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1163 Node *p = s->fast_out(i);
1164 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1165 _immutable_memory = p;
1166 return _immutable_memory;
1167 }
1168 }
1169 ShouldNotReachHere();
1170 return NULL;
1171 }
1172
1173 //----------------------set_cached_top_node------------------------------------
1174 // Install the cached top node, and make sure Node::is_top works correctly.
1175 void Compile::set_cached_top_node(Node* tn) {
1176 if (tn != NULL) verify_top(tn);
1177 Node* old_top = _top;
1178 _top = tn;
1179 // Calling Node::setup_is_top allows the nodes the chance to adjust
1180 // their _out arrays.
1181 if (_top != NULL) _top->setup_is_top();
1182 if (old_top != NULL) old_top->setup_is_top();
1183 assert(_top == NULL || top()->is_top(), "");
1184 }
1185
1186 #ifdef ASSERT
1187 uint Compile::count_live_nodes_by_graph_walk() {
1188 Unique_Node_List useful(comp_arena());
1189 // Get useful node list by walking the graph.
1190 identify_useful_nodes(useful);
1191 return useful.size();
1192 }
1193
1194 void Compile::print_missing_nodes() {
1195
1196 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1197 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1198 return;
1199 }
1200
1201 // This is an expensive function. It is executed only when the user
1202 // specifies VerifyIdealNodeCount option or otherwise knows the
1203 // additional work that needs to be done to identify reachable nodes
1204 // by walking the flow graph and find the missing ones using
1205 // _dead_node_list.
1206
1207 Unique_Node_List useful(comp_arena());
1208 // Get useful node list by walking the graph.
1209 identify_useful_nodes(useful);
1210
1211 uint l_nodes = C->live_nodes();
1212 uint l_nodes_by_walk = useful.size();
1213
1214 if (l_nodes != l_nodes_by_walk) {
1215 if (_log != NULL) {
1216 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1217 _log->stamp();
1218 _log->end_head();
1219 }
1220 VectorSet& useful_member_set = useful.member_set();
1221 int last_idx = l_nodes_by_walk;
1222 for (int i = 0; i < last_idx; i++) {
1223 if (useful_member_set.test(i)) {
1224 if (_dead_node_list.test(i)) {
1225 if (_log != NULL) {
1226 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1227 }
1228 if (PrintIdealNodeCount) {
1229 // Print the log message to tty
1230 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1231 useful.at(i)->dump();
1232 }
1233 }
1234 }
1235 else if (! _dead_node_list.test(i)) {
1236 if (_log != NULL) {
1237 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1238 }
1239 if (PrintIdealNodeCount) {
1240 // Print the log message to tty
1241 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1242 }
1243 }
1244 }
1245 if (_log != NULL) {
1246 _log->tail("mismatched_nodes");
1247 }
1248 }
1249 }
1250 #endif
1251
1252 #ifndef PRODUCT
1253 void Compile::verify_top(Node* tn) const {
1254 if (tn != NULL) {
1255 assert(tn->is_Con(), "top node must be a constant");
1256 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1257 assert(tn->in(0) != NULL, "must have live top node");
1258 }
1259 }
1260 #endif
1261
1262
1263 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1264
1265 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1266 guarantee(arr != NULL, "");
1267 int num_blocks = arr->length();
1268 if (grow_by < num_blocks) grow_by = num_blocks;
1269 int num_notes = grow_by * _node_notes_block_size;
1270 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1271 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1272 while (num_notes > 0) {
1273 arr->append(notes);
1274 notes += _node_notes_block_size;
1275 num_notes -= _node_notes_block_size;
1276 }
1277 assert(num_notes == 0, "exact multiple, please");
1278 }
1279
1280 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1281 if (source == NULL || dest == NULL) return false;
1282
1283 if (dest->is_Con())
1284 return false; // Do not push debug info onto constants.
1285
1286 #ifdef ASSERT
1287 // Leave a bread crumb trail pointing to the original node:
1288 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1289 dest->set_debug_orig(source);
1290 }
1291 #endif
1292
1293 if (node_note_array() == NULL)
1294 return false; // Not collecting any notes now.
1295
1296 // This is a copy onto a pre-existing node, which may already have notes.
1297 // If both nodes have notes, do not overwrite any pre-existing notes.
1298 Node_Notes* source_notes = node_notes_at(source->_idx);
1299 if (source_notes == NULL || source_notes->is_clear()) return false;
1300 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1301 if (dest_notes == NULL || dest_notes->is_clear()) {
1302 return set_node_notes_at(dest->_idx, source_notes);
1303 }
1304
1305 Node_Notes merged_notes = (*source_notes);
1306 // The order of operations here ensures that dest notes will win...
1307 merged_notes.update_from(dest_notes);
1308 return set_node_notes_at(dest->_idx, &merged_notes);
1309 }
1310
1311
1312 //--------------------------allow_range_check_smearing-------------------------
1313 // Gating condition for coalescing similar range checks.
1314 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1315 // single covering check that is at least as strong as any of them.
1316 // If the optimization succeeds, the simplified (strengthened) range check
1317 // will always succeed. If it fails, we will deopt, and then give up
1318 // on the optimization.
1319 bool Compile::allow_range_check_smearing() const {
1320 // If this method has already thrown a range-check,
1321 // assume it was because we already tried range smearing
1322 // and it failed.
1323 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1324 return !already_trapped;
1325 }
1326
1327
1328 //------------------------------flatten_alias_type-----------------------------
1329 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1330 int offset = tj->offset();
1331 TypePtr::PTR ptr = tj->ptr();
1332
1333 // Known instance (scalarizable allocation) alias only with itself.
1334 bool is_known_inst = tj->isa_oopptr() != NULL &&
1335 tj->is_oopptr()->is_known_instance();
1336
1337 // Process weird unsafe references.
1338 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1339 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1340 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1341 tj = TypeOopPtr::BOTTOM;
1342 ptr = tj->ptr();
1343 offset = tj->offset();
1344 }
1345
1346 // Array pointers need some flattening
1347 const TypeAryPtr *ta = tj->isa_aryptr();
1348 if (ta && ta->is_stable()) {
1349 // Erase stability property for alias analysis.
1350 tj = ta = ta->cast_to_stable(false);
1351 }
1352 if( ta && is_known_inst ) {
1353 if ( offset != Type::OffsetBot &&
1354 offset > arrayOopDesc::length_offset_in_bytes() ) {
1355 offset = Type::OffsetBot; // Flatten constant access into array body only
1356 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1357 }
1358 } else if( ta && _AliasLevel >= 2 ) {
1359 // For arrays indexed by constant indices, we flatten the alias
1360 // space to include all of the array body. Only the header, klass
1361 // and array length can be accessed un-aliased.
1362 if( offset != Type::OffsetBot ) {
1363 if( ta->const_oop() ) { // MethodData* or Method*
1364 offset = Type::OffsetBot; // Flatten constant access into array body
1365 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1366 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1367 // range is OK as-is.
1368 tj = ta = TypeAryPtr::RANGE;
1369 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1370 tj = TypeInstPtr::KLASS; // all klass loads look alike
1371 ta = TypeAryPtr::RANGE; // generic ignored junk
1372 ptr = TypePtr::BotPTR;
1373 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1374 tj = TypeInstPtr::MARK;
1375 ta = TypeAryPtr::RANGE; // generic ignored junk
1376 ptr = TypePtr::BotPTR;
1377 } else { // Random constant offset into array body
1378 offset = Type::OffsetBot; // Flatten constant access into array body
1379 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1380 }
1381 }
1382 // Arrays of fixed size alias with arrays of unknown size.
1383 if (ta->size() != TypeInt::POS) {
1384 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1385 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1386 }
1387 // Arrays of known objects become arrays of unknown objects.
1388 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1389 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1390 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1391 }
1392 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1393 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1394 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1395 }
1396 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1397 // cannot be distinguished by bytecode alone.
1398 if (ta->elem() == TypeInt::BOOL) {
1399 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1400 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1402 }
1403 // During the 2nd round of IterGVN, NotNull castings are removed.
1404 // Make sure the Bottom and NotNull variants alias the same.
1405 // Also, make sure exact and non-exact variants alias the same.
1406 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1407 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1408 }
1409 }
1410
1411 // Oop pointers need some flattening
1412 const TypeInstPtr *to = tj->isa_instptr();
1413 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1414 ciInstanceKlass *k = to->klass()->as_instance_klass();
1415 if( ptr == TypePtr::Constant ) {
1416 if (to->klass() != ciEnv::current()->Class_klass() ||
1417 offset < k->size_helper() * wordSize) {
1418 // No constant oop pointers (such as Strings); they alias with
1419 // unknown strings.
1420 assert(!is_known_inst, "not scalarizable allocation");
1421 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1422 }
1423 } else if( is_known_inst ) {
1424 tj = to; // Keep NotNull and klass_is_exact for instance type
1425 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1426 // During the 2nd round of IterGVN, NotNull castings are removed.
1427 // Make sure the Bottom and NotNull variants alias the same.
1428 // Also, make sure exact and non-exact variants alias the same.
1429 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1430 }
1431 if (to->speculative() != NULL) {
1432 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1433 }
1434 // Canonicalize the holder of this field
1435 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1436 // First handle header references such as a LoadKlassNode, even if the
1437 // object's klass is unloaded at compile time (4965979).
1438 if (!is_known_inst) { // Do it only for non-instance types
1439 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1440 }
1441 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1442 // Static fields are in the space above the normal instance
1443 // fields in the java.lang.Class instance.
1444 if (to->klass() != ciEnv::current()->Class_klass()) {
1445 to = NULL;
1446 tj = TypeOopPtr::BOTTOM;
1447 offset = tj->offset();
1448 }
1449 } else {
1450 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1451 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1452 if( is_known_inst ) {
1453 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1454 } else {
1455 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1456 }
1457 }
1458 }
1459 }
1460
1461 // Klass pointers to object array klasses need some flattening
1462 const TypeKlassPtr *tk = tj->isa_klassptr();
1463 if( tk ) {
1464 // If we are referencing a field within a Klass, we need
1465 // to assume the worst case of an Object. Both exact and
1466 // inexact types must flatten to the same alias class so
1467 // use NotNull as the PTR.
1468 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1469
1470 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1471 TypeKlassPtr::OBJECT->klass(),
1472 offset);
1473 }
1474
1475 ciKlass* klass = tk->klass();
1476 if( klass->is_obj_array_klass() ) {
1477 ciKlass* k = TypeAryPtr::OOPS->klass();
1478 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1479 k = TypeInstPtr::BOTTOM->klass();
1480 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1481 }
1482
1483 // Check for precise loads from the primary supertype array and force them
1484 // to the supertype cache alias index. Check for generic array loads from
1485 // the primary supertype array and also force them to the supertype cache
1486 // alias index. Since the same load can reach both, we need to merge
1487 // these 2 disparate memories into the same alias class. Since the
1488 // primary supertype array is read-only, there's no chance of confusion
1489 // where we bypass an array load and an array store.
1490 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1491 if (offset == Type::OffsetBot ||
1492 (offset >= primary_supers_offset &&
1493 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1494 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1495 offset = in_bytes(Klass::secondary_super_cache_offset());
1496 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1497 }
1498 }
1499
1500 // Flatten all Raw pointers together.
1501 if (tj->base() == Type::RawPtr)
1502 tj = TypeRawPtr::BOTTOM;
1503
1504 if (tj->base() == Type::AnyPtr)
1505 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1506
1507 // Flatten all to bottom for now
1508 switch( _AliasLevel ) {
1509 case 0:
1510 tj = TypePtr::BOTTOM;
1511 break;
1512 case 1: // Flatten to: oop, static, field or array
1513 switch (tj->base()) {
1514 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1515 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1516 case Type::AryPtr: // do not distinguish arrays at all
1517 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1518 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1519 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1520 default: ShouldNotReachHere();
1521 }
1522 break;
1523 case 2: // No collapsing at level 2; keep all splits
1524 case 3: // No collapsing at level 3; keep all splits
1525 break;
1526 default:
1527 Unimplemented();
1528 }
1529
1530 offset = tj->offset();
1531 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1532
1533 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1534 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1535 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1536 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1537 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1538 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1539 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1540 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1541 assert( tj->ptr() != TypePtr::TopPTR &&
1542 tj->ptr() != TypePtr::AnyNull &&
1543 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1544 // assert( tj->ptr() != TypePtr::Constant ||
1545 // tj->base() == Type::RawPtr ||
1546 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1547
1548 return tj;
1549 }
1550
1551 void Compile::AliasType::Init(int i, const TypePtr* at) {
1552 _index = i;
1553 _adr_type = at;
1554 _field = NULL;
1555 _element = NULL;
1556 _is_rewritable = true; // default
1557 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1558 if (atoop != NULL && atoop->is_known_instance()) {
1559 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1560 _general_index = Compile::current()->get_alias_index(gt);
1561 } else {
1562 _general_index = 0;
1563 }
1564 }
1565
1566 //---------------------------------print_on------------------------------------
1567 #ifndef PRODUCT
1568 void Compile::AliasType::print_on(outputStream* st) {
1569 if (index() < 10)
1570 st->print("@ <%d> ", index());
1571 else st->print("@ <%d>", index());
1572 st->print(is_rewritable() ? " " : " RO");
1573 int offset = adr_type()->offset();
1574 if (offset == Type::OffsetBot)
1575 st->print(" +any");
1576 else st->print(" +%-3d", offset);
1577 st->print(" in ");
1578 adr_type()->dump_on(st);
1579 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1580 if (field() != NULL && tjp) {
1581 if (tjp->klass() != field()->holder() ||
1582 tjp->offset() != field()->offset_in_bytes()) {
1583 st->print(" != ");
1584 field()->print();
1585 st->print(" ***");
1586 }
1587 }
1588 }
1589
1590 void print_alias_types() {
1591 Compile* C = Compile::current();
1592 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1593 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1594 C->alias_type(idx)->print_on(tty);
1595 tty->cr();
1596 }
1597 }
1598 #endif
1599
1600
1601 //----------------------------probe_alias_cache--------------------------------
1602 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1603 intptr_t key = (intptr_t) adr_type;
1604 key ^= key >> logAliasCacheSize;
1605 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1606 }
1607
1608
1609 //-----------------------------grow_alias_types--------------------------------
1610 void Compile::grow_alias_types() {
1611 const int old_ats = _max_alias_types; // how many before?
1612 const int new_ats = old_ats; // how many more?
1613 const int grow_ats = old_ats+new_ats; // how many now?
1614 _max_alias_types = grow_ats;
1615 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1616 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1617 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1618 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1619 }
1620
1621
1622 //--------------------------------find_alias_type------------------------------
1623 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1624 if (_AliasLevel == 0)
1625 return alias_type(AliasIdxBot);
1626
1627 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1628 if (ace->_adr_type == adr_type) {
1629 return alias_type(ace->_index);
1630 }
1631
1632 // Handle special cases.
1633 if (adr_type == NULL) return alias_type(AliasIdxTop);
1634 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1635
1636 // Do it the slow way.
1637 const TypePtr* flat = flatten_alias_type(adr_type);
1638
1639 #ifdef ASSERT
1640 assert(flat == flatten_alias_type(flat), "idempotent");
1641 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1642 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1643 const TypeOopPtr* foop = flat->is_oopptr();
1644 // Scalarizable allocations have exact klass always.
1645 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1646 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1647 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1648 }
1649 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1650 #endif
1651
1652 int idx = AliasIdxTop;
1653 for (int i = 0; i < num_alias_types(); i++) {
1654 if (alias_type(i)->adr_type() == flat) {
1655 idx = i;
1656 break;
1657 }
1658 }
1659
1660 if (idx == AliasIdxTop) {
1661 if (no_create) return NULL;
1662 // Grow the array if necessary.
1663 if (_num_alias_types == _max_alias_types) grow_alias_types();
1664 // Add a new alias type.
1665 idx = _num_alias_types++;
1666 _alias_types[idx]->Init(idx, flat);
1667 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1668 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1669 if (flat->isa_instptr()) {
1670 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1671 && flat->is_instptr()->klass() == env()->Class_klass())
1672 alias_type(idx)->set_rewritable(false);
1673 }
1674 if (flat->isa_aryptr()) {
1675 #ifdef ASSERT
1676 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1677 // (T_BYTE has the weakest alignment and size restrictions...)
1678 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1679 #endif
1680 if (flat->offset() == TypePtr::OffsetBot) {
1681 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1682 }
1683 }
1684 if (flat->isa_klassptr()) {
1685 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1686 alias_type(idx)->set_rewritable(false);
1687 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1688 alias_type(idx)->set_rewritable(false);
1689 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1690 alias_type(idx)->set_rewritable(false);
1691 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1692 alias_type(idx)->set_rewritable(false);
1693 }
1694 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1695 // but the base pointer type is not distinctive enough to identify
1696 // references into JavaThread.)
1697
1698 // Check for final fields.
1699 const TypeInstPtr* tinst = flat->isa_instptr();
1700 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1701 ciField* field;
1702 if (tinst->const_oop() != NULL &&
1703 tinst->klass() == ciEnv::current()->Class_klass() &&
1704 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1705 // static field
1706 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1707 field = k->get_field_by_offset(tinst->offset(), true);
1708 } else {
1709 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1710 field = k->get_field_by_offset(tinst->offset(), false);
1711 }
1712 assert(field == NULL ||
1713 original_field == NULL ||
1714 (field->holder() == original_field->holder() &&
1715 field->offset() == original_field->offset() &&
1716 field->is_static() == original_field->is_static()), "wrong field?");
1717 // Set field() and is_rewritable() attributes.
1718 if (field != NULL) alias_type(idx)->set_field(field);
1719 }
1720 }
1721
1722 // Fill the cache for next time.
1723 ace->_adr_type = adr_type;
1724 ace->_index = idx;
1725 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1726
1727 // Might as well try to fill the cache for the flattened version, too.
1728 AliasCacheEntry* face = probe_alias_cache(flat);
1729 if (face->_adr_type == NULL) {
1730 face->_adr_type = flat;
1731 face->_index = idx;
1732 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1733 }
1734
1735 return alias_type(idx);
1736 }
1737
1738
1739 Compile::AliasType* Compile::alias_type(ciField* field) {
1740 const TypeOopPtr* t;
1741 if (field->is_static())
1742 t = TypeInstPtr::make(field->holder()->java_mirror());
1743 else
1744 t = TypeOopPtr::make_from_klass_raw(field->holder());
1745 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1746 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1747 return atp;
1748 }
1749
1750
1751 //------------------------------have_alias_type--------------------------------
1752 bool Compile::have_alias_type(const TypePtr* adr_type) {
1753 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1754 if (ace->_adr_type == adr_type) {
1755 return true;
1756 }
1757
1758 // Handle special cases.
1759 if (adr_type == NULL) return true;
1760 if (adr_type == TypePtr::BOTTOM) return true;
1761
1762 return find_alias_type(adr_type, true, NULL) != NULL;
1763 }
1764
1765 //-----------------------------must_alias--------------------------------------
1766 // True if all values of the given address type are in the given alias category.
1767 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1768 if (alias_idx == AliasIdxBot) return true; // the universal category
1769 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1770 if (alias_idx == AliasIdxTop) return false; // the empty category
1771 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1772
1773 // the only remaining possible overlap is identity
1774 int adr_idx = get_alias_index(adr_type);
1775 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1776 assert(adr_idx == alias_idx ||
1777 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1778 && adr_type != TypeOopPtr::BOTTOM),
1779 "should not be testing for overlap with an unsafe pointer");
1780 return adr_idx == alias_idx;
1781 }
1782
1783 //------------------------------can_alias--------------------------------------
1784 // True if any values of the given address type are in the given alias category.
1785 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1786 if (alias_idx == AliasIdxTop) return false; // the empty category
1787 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1788 if (alias_idx == AliasIdxBot) return true; // the universal category
1789 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1790
1791 // the only remaining possible overlap is identity
1792 int adr_idx = get_alias_index(adr_type);
1793 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1794 return adr_idx == alias_idx;
1795 }
1796
1797
1798
1799 //---------------------------pop_warm_call-------------------------------------
1800 WarmCallInfo* Compile::pop_warm_call() {
1801 WarmCallInfo* wci = _warm_calls;
1802 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1803 return wci;
1804 }
1805
1806 //----------------------------Inline_Warm--------------------------------------
1807 int Compile::Inline_Warm() {
1808 // If there is room, try to inline some more warm call sites.
1809 // %%% Do a graph index compaction pass when we think we're out of space?
1810 if (!InlineWarmCalls) return 0;
1811
1812 int calls_made_hot = 0;
1813 int room_to_grow = NodeCountInliningCutoff - unique();
1814 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1815 int amount_grown = 0;
1816 WarmCallInfo* call;
1817 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1818 int est_size = (int)call->size();
1819 if (est_size > (room_to_grow - amount_grown)) {
1820 // This one won't fit anyway. Get rid of it.
1821 call->make_cold();
1822 continue;
1823 }
1824 call->make_hot();
1825 calls_made_hot++;
1826 amount_grown += est_size;
1827 amount_to_grow -= est_size;
1828 }
1829
1830 if (calls_made_hot > 0) set_major_progress();
1831 return calls_made_hot;
1832 }
1833
1834
1835 //----------------------------Finish_Warm--------------------------------------
1836 void Compile::Finish_Warm() {
1837 if (!InlineWarmCalls) return;
1838 if (failing()) return;
1839 if (warm_calls() == NULL) return;
1840
1841 // Clean up loose ends, if we are out of space for inlining.
1842 WarmCallInfo* call;
1843 while ((call = pop_warm_call()) != NULL) {
1844 call->make_cold();
1845 }
1846 }
1847
1848 //---------------------cleanup_loop_predicates-----------------------
1849 // Remove the opaque nodes that protect the predicates so that all unused
1850 // checks and uncommon_traps will be eliminated from the ideal graph
1851 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1852 if (predicate_count()==0) return;
1853 for (int i = predicate_count(); i > 0; i--) {
1854 Node * n = predicate_opaque1_node(i-1);
1855 assert(n->Opcode() == Op_Opaque1, "must be");
1856 igvn.replace_node(n, n->in(1));
1857 }
1858 assert(predicate_count()==0, "should be clean!");
1859 }
1860
1861 // StringOpts and late inlining of string methods
1862 void Compile::inline_string_calls(bool parse_time) {
1863 {
1864 // remove useless nodes to make the usage analysis simpler
1865 ResourceMark rm;
1866 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1867 }
1868
1869 {
1870 ResourceMark rm;
1871 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1872 PhaseStringOpts pso(initial_gvn(), for_igvn());
1873 print_method(PHASE_AFTER_STRINGOPTS, 3);
1874 }
1875
1876 // now inline anything that we skipped the first time around
1877 if (!parse_time) {
1878 _late_inlines_pos = _late_inlines.length();
1879 }
1880
1881 while (_string_late_inlines.length() > 0) {
1882 CallGenerator* cg = _string_late_inlines.pop();
1883 cg->do_late_inline();
1884 if (failing()) return;
1885 }
1886 _string_late_inlines.trunc_to(0);
1887 }
1888
1889 // Late inlining of boxing methods
1890 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1891 if (_boxing_late_inlines.length() > 0) {
1892 assert(has_boxed_value(), "inconsistent");
1893
1894 PhaseGVN* gvn = initial_gvn();
1895 set_inlining_incrementally(true);
1896
1897 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1898 for_igvn()->clear();
1899 gvn->replace_with(&igvn);
1900
1901 while (_boxing_late_inlines.length() > 0) {
1902 CallGenerator* cg = _boxing_late_inlines.pop();
1903 cg->do_late_inline();
1904 if (failing()) return;
1905 }
1906 _boxing_late_inlines.trunc_to(0);
1907
1908 {
1909 ResourceMark rm;
1910 PhaseRemoveUseless pru(gvn, for_igvn());
1911 }
1912
1913 igvn = PhaseIterGVN(gvn);
1914 igvn.optimize();
1915
1916 set_inlining_progress(false);
1917 set_inlining_incrementally(false);
1918 }
1919 }
1920
1921 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1922 assert(IncrementalInline, "incremental inlining should be on");
1923 PhaseGVN* gvn = initial_gvn();
1924
1925 set_inlining_progress(false);
1926 for_igvn()->clear();
1927 gvn->replace_with(&igvn);
1928
1929 int i = 0;
1930
1931 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1932 CallGenerator* cg = _late_inlines.at(i);
1933 _late_inlines_pos = i+1;
1934 cg->do_late_inline();
1935 if (failing()) return;
1936 }
1937 int j = 0;
1938 for (; i < _late_inlines.length(); i++, j++) {
1939 _late_inlines.at_put(j, _late_inlines.at(i));
1940 }
1941 _late_inlines.trunc_to(j);
1942
1943 {
1944 ResourceMark rm;
1945 PhaseRemoveUseless pru(gvn, for_igvn());
1946 }
1947
1948 igvn = PhaseIterGVN(gvn);
1949 }
1950
1951 // Perform incremental inlining until bound on number of live nodes is reached
1952 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1953 PhaseGVN* gvn = initial_gvn();
1954
1955 set_inlining_incrementally(true);
1956 set_inlining_progress(true);
1957 uint low_live_nodes = 0;
1958
1959 while(inlining_progress() && _late_inlines.length() > 0) {
1960
1961 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1962 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1963 // PhaseIdealLoop is expensive so we only try it once we are
1964 // out of loop and we only try it again if the previous helped
1965 // got the number of nodes down significantly
1966 PhaseIdealLoop ideal_loop( igvn, false, true );
1967 if (failing()) return;
1968 low_live_nodes = live_nodes();
1969 _major_progress = true;
1970 }
1971
1972 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1973 break;
1974 }
1975 }
1976
1977 inline_incrementally_one(igvn);
1978
1979 if (failing()) return;
1980
1981 igvn.optimize();
1982
1983 if (failing()) return;
1984 }
1985
1986 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1987
1988 if (_string_late_inlines.length() > 0) {
1989 assert(has_stringbuilder(), "inconsistent");
1990 for_igvn()->clear();
1991 initial_gvn()->replace_with(&igvn);
1992
1993 inline_string_calls(false);
1994
1995 if (failing()) return;
1996
1997 {
1998 ResourceMark rm;
1999 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2000 }
2001
2002 igvn = PhaseIterGVN(gvn);
2003
2004 igvn.optimize();
2005 }
2006
2007 set_inlining_incrementally(false);
2008 }
2009
2010
2011 //------------------------------Optimize---------------------------------------
2012 // Given a graph, optimize it.
2013 void Compile::Optimize() {
2014 TracePhase t1("optimizer", &_t_optimizer, true);
2015
2016 #ifndef PRODUCT
2017 if (env()->break_at_compile()) {
2018 BREAKPOINT;
2019 }
2020
2021 #endif
2022
2023 ResourceMark rm;
2024 int loop_opts_cnt;
2025
2026 NOT_PRODUCT( verify_graph_edges(); )
2027
2028 print_method(PHASE_AFTER_PARSING);
2029
2030 {
2031 // Iterative Global Value Numbering, including ideal transforms
2032 // Initialize IterGVN with types and values from parse-time GVN
2033 PhaseIterGVN igvn(initial_gvn());
2034 {
2035 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2036 igvn.optimize();
2037 }
2038
2039 print_method(PHASE_ITER_GVN1, 2);
2040
2041 if (failing()) return;
2042
2043 {
2044 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2045 inline_incrementally(igvn);
2046 }
2047
2048 print_method(PHASE_INCREMENTAL_INLINE, 2);
2049
2050 if (failing()) return;
2051
2052 if (eliminate_boxing()) {
2053 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2054 // Inline valueOf() methods now.
2055 inline_boxing_calls(igvn);
2056
2057 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2058
2059 if (failing()) return;
2060 }
2061
2062 // Remove the speculative part of types and clean up the graph from
2063 // the extra CastPP nodes whose only purpose is to carry them. Do
2064 // that early so that optimizations are not disrupted by the extra
2065 // CastPP nodes.
2066 remove_speculative_types(igvn);
2067
2068 // No more new expensive nodes will be added to the list from here
2069 // so keep only the actual candidates for optimizations.
2070 cleanup_expensive_nodes(igvn);
2071
2072 // Perform escape analysis
2073 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2074 if (has_loops()) {
2075 // Cleanup graph (remove dead nodes).
2076 TracePhase t2("idealLoop", &_t_idealLoop, true);
2077 PhaseIdealLoop ideal_loop( igvn, false, true );
2078 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2079 if (failing()) return;
2080 }
2081 ConnectionGraph::do_analysis(this, &igvn);
2082
2083 if (failing()) return;
2084
2085 // Optimize out fields loads from scalar replaceable allocations.
2086 igvn.optimize();
2087 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2088
2089 if (failing()) return;
2090
2091 if (congraph() != NULL && macro_count() > 0) {
2092 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2093 PhaseMacroExpand mexp(igvn);
2094 mexp.eliminate_macro_nodes();
2095 igvn.set_delay_transform(false);
2096
2097 igvn.optimize();
2098 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2099
2100 if (failing()) return;
2101 }
2102 }
2103
2104 // Loop transforms on the ideal graph. Range Check Elimination,
2105 // peeling, unrolling, etc.
2106
2107 // Set loop opts counter
2108 loop_opts_cnt = num_loop_opts();
2109 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2110 {
2111 TracePhase t2("idealLoop", &_t_idealLoop, true);
2112 PhaseIdealLoop ideal_loop( igvn, true );
2113 loop_opts_cnt--;
2114 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2115 if (failing()) return;
2116 }
2117 // Loop opts pass if partial peeling occurred in previous pass
2118 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2119 TracePhase t3("idealLoop", &_t_idealLoop, true);
2120 PhaseIdealLoop ideal_loop( igvn, false );
2121 loop_opts_cnt--;
2122 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2123 if (failing()) return;
2124 }
2125 // Loop opts pass for loop-unrolling before CCP
2126 if(major_progress() && (loop_opts_cnt > 0)) {
2127 TracePhase t4("idealLoop", &_t_idealLoop, true);
2128 PhaseIdealLoop ideal_loop( igvn, false );
2129 loop_opts_cnt--;
2130 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2131 }
2132 if (!failing()) {
2133 // Verify that last round of loop opts produced a valid graph
2134 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2135 PhaseIdealLoop::verify(igvn);
2136 }
2137 }
2138 if (failing()) return;
2139
2140 // Conditional Constant Propagation;
2141 PhaseCCP ccp( &igvn );
2142 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2143 {
2144 TracePhase t2("ccp", &_t_ccp, true);
2145 ccp.do_transform();
2146 }
2147 print_method(PHASE_CPP1, 2);
2148
2149 assert( true, "Break here to ccp.dump_old2new_map()");
2150
2151 // Iterative Global Value Numbering, including ideal transforms
2152 {
2153 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2154 igvn = ccp;
2155 igvn.optimize();
2156 }
2157
2158 print_method(PHASE_ITER_GVN2, 2);
2159
2160 if (failing()) return;
2161
2162 // Loop transforms on the ideal graph. Range Check Elimination,
2163 // peeling, unrolling, etc.
2164 if(loop_opts_cnt > 0) {
2165 debug_only( int cnt = 0; );
2166 while(major_progress() && (loop_opts_cnt > 0)) {
2167 TracePhase t2("idealLoop", &_t_idealLoop, true);
2168 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2169 PhaseIdealLoop ideal_loop( igvn, true);
2170 loop_opts_cnt--;
2171 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2172 if (failing()) return;
2173 }
2174 }
2175
2176 {
2177 // Verify that all previous optimizations produced a valid graph
2178 // at least to this point, even if no loop optimizations were done.
2179 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2180 PhaseIdealLoop::verify(igvn);
2181 }
2182
2183 {
2184 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2185 PhaseMacroExpand mex(igvn);
2186 if (mex.expand_macro_nodes()) {
2187 assert(failing(), "must bail out w/ explicit message");
2188 return;
2189 }
2190 }
2191
2192 } // (End scope of igvn; run destructor if necessary for asserts.)
2193
2194 dump_inlining();
2195 // A method with only infinite loops has no edges entering loops from root
2196 {
2197 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2198 if (final_graph_reshaping()) {
2199 assert(failing(), "must bail out w/ explicit message");
2200 return;
2201 }
2202 }
2203
2204 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2205 }
2206
2207
2208 //------------------------------Code_Gen---------------------------------------
2209 // Given a graph, generate code for it
2210 void Compile::Code_Gen() {
2211 if (failing()) {
2212 return;
2213 }
2214
2215 // Perform instruction selection. You might think we could reclaim Matcher
2216 // memory PDQ, but actually the Matcher is used in generating spill code.
2217 // Internals of the Matcher (including some VectorSets) must remain live
2218 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2219 // set a bit in reclaimed memory.
2220
2221 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2222 // nodes. Mapping is only valid at the root of each matched subtree.
2223 NOT_PRODUCT( verify_graph_edges(); )
2224
2225 Matcher matcher;
2226 _matcher = &matcher;
2227 {
2228 TracePhase t2("matcher", &_t_matcher, true);
2229 matcher.match();
2230 }
2231 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2232 // nodes. Mapping is only valid at the root of each matched subtree.
2233 NOT_PRODUCT( verify_graph_edges(); )
2234
2235 // If you have too many nodes, or if matching has failed, bail out
2236 check_node_count(0, "out of nodes matching instructions");
2237 if (failing()) {
2238 return;
2239 }
2240
2241 // Build a proper-looking CFG
2242 PhaseCFG cfg(node_arena(), root(), matcher);
2243 _cfg = &cfg;
2244 {
2245 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2246 bool success = cfg.do_global_code_motion();
2247 if (!success) {
2248 return;
2249 }
2250
2251 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2252 NOT_PRODUCT( verify_graph_edges(); )
2253 debug_only( cfg.verify(); )
2254 }
2255
2256 PhaseChaitin regalloc(unique(), cfg, matcher);
2257 _regalloc = ®alloc;
2258 {
2259 TracePhase t2("regalloc", &_t_registerAllocation, true);
2260 // Perform register allocation. After Chaitin, use-def chains are
2261 // no longer accurate (at spill code) and so must be ignored.
2262 // Node->LRG->reg mappings are still accurate.
2263 _regalloc->Register_Allocate();
2264
2265 // Bail out if the allocator builds too many nodes
2266 if (failing()) {
2267 return;
2268 }
2269 }
2270
2271 // Prior to register allocation we kept empty basic blocks in case the
2272 // the allocator needed a place to spill. After register allocation we
2273 // are not adding any new instructions. If any basic block is empty, we
2274 // can now safely remove it.
2275 {
2276 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2277 cfg.remove_empty_blocks();
2278 if (do_freq_based_layout()) {
2279 PhaseBlockLayout layout(cfg);
2280 } else {
2281 cfg.set_loop_alignment();
2282 }
2283 cfg.fixup_flow();
2284 }
2285
2286 // Apply peephole optimizations
2287 if( OptoPeephole ) {
2288 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2289 PhasePeephole peep( _regalloc, cfg);
2290 peep.do_transform();
2291 }
2292
2293 // Do late expand if CPU requires this.
2294 if (Matcher::require_postalloc_expand) {
2295 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2296 cfg.postalloc_expand(_regalloc);
2297 }
2298
2299 // Convert Nodes to instruction bits in a buffer
2300 {
2301 // %%%% workspace merge brought two timers together for one job
2302 TracePhase t2a("output", &_t_output, true);
2303 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2304 Output();
2305 }
2306
2307 print_method(PHASE_FINAL_CODE);
2308
2309 // He's dead, Jim.
2310 _cfg = (PhaseCFG*)0xdeadbeef;
2311 _regalloc = (PhaseChaitin*)0xdeadbeef;
2312 }
2313
2314
2315 //------------------------------dump_asm---------------------------------------
2316 // Dump formatted assembly
2317 #ifndef PRODUCT
2318 void Compile::dump_asm(int *pcs, uint pc_limit) {
2319 bool cut_short = false;
2320 tty->print_cr("#");
2321 tty->print("# "); _tf->dump(); tty->cr();
2322 tty->print_cr("#");
2323
2324 // For all blocks
2325 int pc = 0x0; // Program counter
2326 char starts_bundle = ' ';
2327 _regalloc->dump_frame();
2328
2329 Node *n = NULL;
2330 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2331 if (VMThread::should_terminate()) {
2332 cut_short = true;
2333 break;
2334 }
2335 Block* block = _cfg->get_block(i);
2336 if (block->is_connector() && !Verbose) {
2337 continue;
2338 }
2339 n = block->head();
2340 if (pcs && n->_idx < pc_limit) {
2341 tty->print("%3.3x ", pcs[n->_idx]);
2342 } else {
2343 tty->print(" ");
2344 }
2345 block->dump_head(_cfg);
2346 if (block->is_connector()) {
2347 tty->print_cr(" # Empty connector block");
2348 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2349 tty->print_cr(" # Block is sole successor of call");
2350 }
2351
2352 // For all instructions
2353 Node *delay = NULL;
2354 for (uint j = 0; j < block->number_of_nodes(); j++) {
2355 if (VMThread::should_terminate()) {
2356 cut_short = true;
2357 break;
2358 }
2359 n = block->get_node(j);
2360 if (valid_bundle_info(n)) {
2361 Bundle* bundle = node_bundling(n);
2362 if (bundle->used_in_unconditional_delay()) {
2363 delay = n;
2364 continue;
2365 }
2366 if (bundle->starts_bundle()) {
2367 starts_bundle = '+';
2368 }
2369 }
2370
2371 if (WizardMode) {
2372 n->dump();
2373 }
2374
2375 if( !n->is_Region() && // Dont print in the Assembly
2376 !n->is_Phi() && // a few noisely useless nodes
2377 !n->is_Proj() &&
2378 !n->is_MachTemp() &&
2379 !n->is_SafePointScalarObject() &&
2380 !n->is_Catch() && // Would be nice to print exception table targets
2381 !n->is_MergeMem() && // Not very interesting
2382 !n->is_top() && // Debug info table constants
2383 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2384 ) {
2385 if (pcs && n->_idx < pc_limit)
2386 tty->print("%3.3x", pcs[n->_idx]);
2387 else
2388 tty->print(" ");
2389 tty->print(" %c ", starts_bundle);
2390 starts_bundle = ' ';
2391 tty->print("\t");
2392 n->format(_regalloc, tty);
2393 tty->cr();
2394 }
2395
2396 // If we have an instruction with a delay slot, and have seen a delay,
2397 // then back up and print it
2398 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2399 assert(delay != NULL, "no unconditional delay instruction");
2400 if (WizardMode) delay->dump();
2401
2402 if (node_bundling(delay)->starts_bundle())
2403 starts_bundle = '+';
2404 if (pcs && n->_idx < pc_limit)
2405 tty->print("%3.3x", pcs[n->_idx]);
2406 else
2407 tty->print(" ");
2408 tty->print(" %c ", starts_bundle);
2409 starts_bundle = ' ';
2410 tty->print("\t");
2411 delay->format(_regalloc, tty);
2412 tty->print_cr("");
2413 delay = NULL;
2414 }
2415
2416 // Dump the exception table as well
2417 if( n->is_Catch() && (Verbose || WizardMode) ) {
2418 // Print the exception table for this offset
2419 _handler_table.print_subtable_for(pc);
2420 }
2421 }
2422
2423 if (pcs && n->_idx < pc_limit)
2424 tty->print_cr("%3.3x", pcs[n->_idx]);
2425 else
2426 tty->print_cr("");
2427
2428 assert(cut_short || delay == NULL, "no unconditional delay branch");
2429
2430 } // End of per-block dump
2431 tty->print_cr("");
2432
2433 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2434 }
2435 #endif
2436
2437 //------------------------------Final_Reshape_Counts---------------------------
2438 // This class defines counters to help identify when a method
2439 // may/must be executed using hardware with only 24-bit precision.
2440 struct Final_Reshape_Counts : public StackObj {
2441 int _call_count; // count non-inlined 'common' calls
2442 int _float_count; // count float ops requiring 24-bit precision
2443 int _double_count; // count double ops requiring more precision
2444 int _java_call_count; // count non-inlined 'java' calls
2445 int _inner_loop_count; // count loops which need alignment
2446 VectorSet _visited; // Visitation flags
2447 Node_List _tests; // Set of IfNodes & PCTableNodes
2448
2449 Final_Reshape_Counts() :
2450 _call_count(0), _float_count(0), _double_count(0),
2451 _java_call_count(0), _inner_loop_count(0),
2452 _visited( Thread::current()->resource_area() ) { }
2453
2454 void inc_call_count () { _call_count ++; }
2455 void inc_float_count () { _float_count ++; }
2456 void inc_double_count() { _double_count++; }
2457 void inc_java_call_count() { _java_call_count++; }
2458 void inc_inner_loop_count() { _inner_loop_count++; }
2459
2460 int get_call_count () const { return _call_count ; }
2461 int get_float_count () const { return _float_count ; }
2462 int get_double_count() const { return _double_count; }
2463 int get_java_call_count() const { return _java_call_count; }
2464 int get_inner_loop_count() const { return _inner_loop_count; }
2465 };
2466
2467 #ifdef ASSERT
2468 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2469 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2470 // Make sure the offset goes inside the instance layout.
2471 return k->contains_field_offset(tp->offset());
2472 // Note that OffsetBot and OffsetTop are very negative.
2473 }
2474 #endif
2475
2476 // Eliminate trivially redundant StoreCMs and accumulate their
2477 // precedence edges.
2478 void Compile::eliminate_redundant_card_marks(Node* n) {
2479 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2480 if (n->in(MemNode::Address)->outcnt() > 1) {
2481 // There are multiple users of the same address so it might be
2482 // possible to eliminate some of the StoreCMs
2483 Node* mem = n->in(MemNode::Memory);
2484 Node* adr = n->in(MemNode::Address);
2485 Node* val = n->in(MemNode::ValueIn);
2486 Node* prev = n;
2487 bool done = false;
2488 // Walk the chain of StoreCMs eliminating ones that match. As
2489 // long as it's a chain of single users then the optimization is
2490 // safe. Eliminating partially redundant StoreCMs would require
2491 // cloning copies down the other paths.
2492 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2493 if (adr == mem->in(MemNode::Address) &&
2494 val == mem->in(MemNode::ValueIn)) {
2495 // redundant StoreCM
2496 if (mem->req() > MemNode::OopStore) {
2497 // Hasn't been processed by this code yet.
2498 n->add_prec(mem->in(MemNode::OopStore));
2499 } else {
2500 // Already converted to precedence edge
2501 for (uint i = mem->req(); i < mem->len(); i++) {
2502 // Accumulate any precedence edges
2503 if (mem->in(i) != NULL) {
2504 n->add_prec(mem->in(i));
2505 }
2506 }
2507 // Everything above this point has been processed.
2508 done = true;
2509 }
2510 // Eliminate the previous StoreCM
2511 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2512 assert(mem->outcnt() == 0, "should be dead");
2513 mem->disconnect_inputs(NULL, this);
2514 } else {
2515 prev = mem;
2516 }
2517 mem = prev->in(MemNode::Memory);
2518 }
2519 }
2520 }
2521
2522 //------------------------------final_graph_reshaping_impl----------------------
2523 // Implement items 1-5 from final_graph_reshaping below.
2524 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2525
2526 if ( n->outcnt() == 0 ) return; // dead node
2527 uint nop = n->Opcode();
2528
2529 // Check for 2-input instruction with "last use" on right input.
2530 // Swap to left input. Implements item (2).
2531 if( n->req() == 3 && // two-input instruction
2532 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2533 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2534 n->in(2)->outcnt() == 1 &&// right use IS a last use
2535 !n->in(2)->is_Con() ) { // right use is not a constant
2536 // Check for commutative opcode
2537 switch( nop ) {
2538 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2539 case Op_MaxI: case Op_MinI:
2540 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2541 case Op_AndL: case Op_XorL: case Op_OrL:
2542 case Op_AndI: case Op_XorI: case Op_OrI: {
2543 // Move "last use" input to left by swapping inputs
2544 n->swap_edges(1, 2);
2545 break;
2546 }
2547 default:
2548 break;
2549 }
2550 }
2551
2552 #ifdef ASSERT
2553 if( n->is_Mem() ) {
2554 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2555 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2556 // oop will be recorded in oop map if load crosses safepoint
2557 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2558 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2559 "raw memory operations should have control edge");
2560 }
2561 #endif
2562 // Count FPU ops and common calls, implements item (3)
2563 switch( nop ) {
2564 // Count all float operations that may use FPU
2565 case Op_AddF:
2566 case Op_SubF:
2567 case Op_MulF:
2568 case Op_DivF:
2569 case Op_NegF:
2570 case Op_ModF:
2571 case Op_ConvI2F:
2572 case Op_ConF:
2573 case Op_CmpF:
2574 case Op_CmpF3:
2575 // case Op_ConvL2F: // longs are split into 32-bit halves
2576 frc.inc_float_count();
2577 break;
2578
2579 case Op_ConvF2D:
2580 case Op_ConvD2F:
2581 frc.inc_float_count();
2582 frc.inc_double_count();
2583 break;
2584
2585 // Count all double operations that may use FPU
2586 case Op_AddD:
2587 case Op_SubD:
2588 case Op_MulD:
2589 case Op_DivD:
2590 case Op_NegD:
2591 case Op_ModD:
2592 case Op_ConvI2D:
2593 case Op_ConvD2I:
2594 // case Op_ConvL2D: // handled by leaf call
2595 // case Op_ConvD2L: // handled by leaf call
2596 case Op_ConD:
2597 case Op_CmpD:
2598 case Op_CmpD3:
2599 frc.inc_double_count();
2600 break;
2601 case Op_Opaque1: // Remove Opaque Nodes before matching
2602 case Op_Opaque2: // Remove Opaque Nodes before matching
2603 case Op_Opaque3:
2604 n->subsume_by(n->in(1), this);
2605 break;
2606 case Op_CallStaticJava:
2607 case Op_CallJava:
2608 case Op_CallDynamicJava:
2609 frc.inc_java_call_count(); // Count java call site;
2610 case Op_CallRuntime:
2611 case Op_CallLeaf:
2612 case Op_CallLeafNoFP: {
2613 assert( n->is_Call(), "" );
2614 CallNode *call = n->as_Call();
2615 // Count call sites where the FP mode bit would have to be flipped.
2616 // Do not count uncommon runtime calls:
2617 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2618 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2619 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2620 frc.inc_call_count(); // Count the call site
2621 } else { // See if uncommon argument is shared
2622 Node *n = call->in(TypeFunc::Parms);
2623 int nop = n->Opcode();
2624 // Clone shared simple arguments to uncommon calls, item (1).
2625 if( n->outcnt() > 1 &&
2626 !n->is_Proj() &&
2627 nop != Op_CreateEx &&
2628 nop != Op_CheckCastPP &&
2629 nop != Op_DecodeN &&
2630 nop != Op_DecodeNKlass &&
2631 !n->is_Mem() ) {
2632 Node *x = n->clone();
2633 call->set_req( TypeFunc::Parms, x );
2634 }
2635 }
2636 break;
2637 }
2638
2639 case Op_StoreD:
2640 case Op_LoadD:
2641 case Op_LoadD_unaligned:
2642 frc.inc_double_count();
2643 goto handle_mem;
2644 case Op_StoreF:
2645 case Op_LoadF:
2646 frc.inc_float_count();
2647 goto handle_mem;
2648
2649 case Op_StoreCM:
2650 {
2651 // Convert OopStore dependence into precedence edge
2652 Node* prec = n->in(MemNode::OopStore);
2653 n->del_req(MemNode::OopStore);
2654 n->add_prec(prec);
2655 eliminate_redundant_card_marks(n);
2656 }
2657
2658 // fall through
2659
2660 case Op_StoreB:
2661 case Op_StoreC:
2662 case Op_StorePConditional:
2663 case Op_StoreI:
2664 case Op_StoreL:
2665 case Op_StoreIConditional:
2666 case Op_StoreLConditional:
2667 case Op_CompareAndSwapI:
2668 case Op_CompareAndSwapL:
2669 case Op_CompareAndSwapP:
2670 case Op_CompareAndSwapN:
2671 case Op_GetAndAddI:
2672 case Op_GetAndAddL:
2673 case Op_GetAndSetI:
2674 case Op_GetAndSetL:
2675 case Op_GetAndSetP:
2676 case Op_GetAndSetN:
2677 case Op_StoreP:
2678 case Op_StoreN:
2679 case Op_StoreNKlass:
2680 case Op_LoadB:
2681 case Op_LoadUB:
2682 case Op_LoadUS:
2683 case Op_LoadI:
2684 case Op_LoadKlass:
2685 case Op_LoadNKlass:
2686 case Op_LoadL:
2687 case Op_LoadL_unaligned:
2688 case Op_LoadPLocked:
2689 case Op_LoadP:
2690 case Op_LoadN:
2691 case Op_LoadRange:
2692 case Op_LoadS: {
2693 handle_mem:
2694 #ifdef ASSERT
2695 if( VerifyOptoOopOffsets ) {
2696 assert( n->is_Mem(), "" );
2697 MemNode *mem = (MemNode*)n;
2698 // Check to see if address types have grounded out somehow.
2699 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2700 assert( !tp || oop_offset_is_sane(tp), "" );
2701 }
2702 #endif
2703 break;
2704 }
2705
2706 case Op_AddP: { // Assert sane base pointers
2707 Node *addp = n->in(AddPNode::Address);
2708 assert( !addp->is_AddP() ||
2709 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2710 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2711 "Base pointers must match" );
2712 #ifdef _LP64
2713 if ((UseCompressedOops || UseCompressedClassPointers) &&
2714 addp->Opcode() == Op_ConP &&
2715 addp == n->in(AddPNode::Base) &&
2716 n->in(AddPNode::Offset)->is_Con()) {
2717 // Use addressing with narrow klass to load with offset on x86.
2718 // On sparc loading 32-bits constant and decoding it have less
2719 // instructions (4) then load 64-bits constant (7).
2720 // Do this transformation here since IGVN will convert ConN back to ConP.
2721 const Type* t = addp->bottom_type();
2722 if (t->isa_oopptr() || t->isa_klassptr()) {
2723 Node* nn = NULL;
2724
2725 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2726
2727 // Look for existing ConN node of the same exact type.
2728 Node* r = root();
2729 uint cnt = r->outcnt();
2730 for (uint i = 0; i < cnt; i++) {
2731 Node* m = r->raw_out(i);
2732 if (m!= NULL && m->Opcode() == op &&
2733 m->bottom_type()->make_ptr() == t) {
2734 nn = m;
2735 break;
2736 }
2737 }
2738 if (nn != NULL) {
2739 // Decode a narrow oop to match address
2740 // [R12 + narrow_oop_reg<<3 + offset]
2741 if (t->isa_oopptr()) {
2742 nn = new (this) DecodeNNode(nn, t);
2743 } else {
2744 nn = new (this) DecodeNKlassNode(nn, t);
2745 }
2746 n->set_req(AddPNode::Base, nn);
2747 n->set_req(AddPNode::Address, nn);
2748 if (addp->outcnt() == 0) {
2749 addp->disconnect_inputs(NULL, this);
2750 }
2751 }
2752 }
2753 }
2754 #endif
2755 break;
2756 }
2757
2758 #ifdef _LP64
2759 case Op_CastPP:
2760 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2761 Node* in1 = n->in(1);
2762 const Type* t = n->bottom_type();
2763 Node* new_in1 = in1->clone();
2764 new_in1->as_DecodeN()->set_type(t);
2765
2766 if (!Matcher::narrow_oop_use_complex_address()) {
2767 //
2768 // x86, ARM and friends can handle 2 adds in addressing mode
2769 // and Matcher can fold a DecodeN node into address by using
2770 // a narrow oop directly and do implicit NULL check in address:
2771 //
2772 // [R12 + narrow_oop_reg<<3 + offset]
2773 // NullCheck narrow_oop_reg
2774 //
2775 // On other platforms (Sparc) we have to keep new DecodeN node and
2776 // use it to do implicit NULL check in address:
2777 //
2778 // decode_not_null narrow_oop_reg, base_reg
2779 // [base_reg + offset]
2780 // NullCheck base_reg
2781 //
2782 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2783 // to keep the information to which NULL check the new DecodeN node
2784 // corresponds to use it as value in implicit_null_check().
2785 //
2786 new_in1->set_req(0, n->in(0));
2787 }
2788
2789 n->subsume_by(new_in1, this);
2790 if (in1->outcnt() == 0) {
2791 in1->disconnect_inputs(NULL, this);
2792 }
2793 }
2794 break;
2795
2796 case Op_CmpP:
2797 // Do this transformation here to preserve CmpPNode::sub() and
2798 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2799 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2800 Node* in1 = n->in(1);
2801 Node* in2 = n->in(2);
2802 if (!in1->is_DecodeNarrowPtr()) {
2803 in2 = in1;
2804 in1 = n->in(2);
2805 }
2806 assert(in1->is_DecodeNarrowPtr(), "sanity");
2807
2808 Node* new_in2 = NULL;
2809 if (in2->is_DecodeNarrowPtr()) {
2810 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2811 new_in2 = in2->in(1);
2812 } else if (in2->Opcode() == Op_ConP) {
2813 const Type* t = in2->bottom_type();
2814 if (t == TypePtr::NULL_PTR) {
2815 assert(in1->is_DecodeN(), "compare klass to null?");
2816 // Don't convert CmpP null check into CmpN if compressed
2817 // oops implicit null check is not generated.
2818 // This will allow to generate normal oop implicit null check.
2819 if (Matcher::gen_narrow_oop_implicit_null_checks())
2820 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2821 //
2822 // This transformation together with CastPP transformation above
2823 // will generated code for implicit NULL checks for compressed oops.
2824 //
2825 // The original code after Optimize()
2826 //
2827 // LoadN memory, narrow_oop_reg
2828 // decode narrow_oop_reg, base_reg
2829 // CmpP base_reg, NULL
2830 // CastPP base_reg // NotNull
2831 // Load [base_reg + offset], val_reg
2832 //
2833 // after these transformations will be
2834 //
2835 // LoadN memory, narrow_oop_reg
2836 // CmpN narrow_oop_reg, NULL
2837 // decode_not_null narrow_oop_reg, base_reg
2838 // Load [base_reg + offset], val_reg
2839 //
2840 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2841 // since narrow oops can be used in debug info now (see the code in
2842 // final_graph_reshaping_walk()).
2843 //
2844 // At the end the code will be matched to
2845 // on x86:
2846 //
2847 // Load_narrow_oop memory, narrow_oop_reg
2848 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2849 // NullCheck narrow_oop_reg
2850 //
2851 // and on sparc:
2852 //
2853 // Load_narrow_oop memory, narrow_oop_reg
2854 // decode_not_null narrow_oop_reg, base_reg
2855 // Load [base_reg + offset], val_reg
2856 // NullCheck base_reg
2857 //
2858 } else if (t->isa_oopptr()) {
2859 new_in2 = ConNode::make(this, t->make_narrowoop());
2860 } else if (t->isa_klassptr()) {
2861 new_in2 = ConNode::make(this, t->make_narrowklass());
2862 }
2863 }
2864 if (new_in2 != NULL) {
2865 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2866 n->subsume_by(cmpN, this);
2867 if (in1->outcnt() == 0) {
2868 in1->disconnect_inputs(NULL, this);
2869 }
2870 if (in2->outcnt() == 0) {
2871 in2->disconnect_inputs(NULL, this);
2872 }
2873 }
2874 }
2875 break;
2876
2877 case Op_DecodeN:
2878 case Op_DecodeNKlass:
2879 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2880 // DecodeN could be pinned when it can't be fold into
2881 // an address expression, see the code for Op_CastPP above.
2882 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2883 break;
2884
2885 case Op_EncodeP:
2886 case Op_EncodePKlass: {
2887 Node* in1 = n->in(1);
2888 if (in1->is_DecodeNarrowPtr()) {
2889 n->subsume_by(in1->in(1), this);
2890 } else if (in1->Opcode() == Op_ConP) {
2891 const Type* t = in1->bottom_type();
2892 if (t == TypePtr::NULL_PTR) {
2893 assert(t->isa_oopptr(), "null klass?");
2894 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2895 } else if (t->isa_oopptr()) {
2896 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2897 } else if (t->isa_klassptr()) {
2898 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2899 }
2900 }
2901 if (in1->outcnt() == 0) {
2902 in1->disconnect_inputs(NULL, this);
2903 }
2904 break;
2905 }
2906
2907 case Op_Proj: {
2908 if (OptimizeStringConcat) {
2909 ProjNode* p = n->as_Proj();
2910 if (p->_is_io_use) {
2911 // Separate projections were used for the exception path which
2912 // are normally removed by a late inline. If it wasn't inlined
2913 // then they will hang around and should just be replaced with
2914 // the original one.
2915 Node* proj = NULL;
2916 // Replace with just one
2917 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2918 Node *use = i.get();
2919 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2920 proj = use;
2921 break;
2922 }
2923 }
2924 assert(proj != NULL, "must be found");
2925 p->subsume_by(proj, this);
2926 }
2927 }
2928 break;
2929 }
2930
2931 case Op_Phi:
2932 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2933 // The EncodeP optimization may create Phi with the same edges
2934 // for all paths. It is not handled well by Register Allocator.
2935 Node* unique_in = n->in(1);
2936 assert(unique_in != NULL, "");
2937 uint cnt = n->req();
2938 for (uint i = 2; i < cnt; i++) {
2939 Node* m = n->in(i);
2940 assert(m != NULL, "");
2941 if (unique_in != m)
2942 unique_in = NULL;
2943 }
2944 if (unique_in != NULL) {
2945 n->subsume_by(unique_in, this);
2946 }
2947 }
2948 break;
2949
2950 #endif
2951
2952 case Op_ModI:
2953 if (UseDivMod) {
2954 // Check if a%b and a/b both exist
2955 Node* d = n->find_similar(Op_DivI);
2956 if (d) {
2957 // Replace them with a fused divmod if supported
2958 if (Matcher::has_match_rule(Op_DivModI)) {
2959 DivModINode* divmod = DivModINode::make(this, n);
2960 d->subsume_by(divmod->div_proj(), this);
2961 n->subsume_by(divmod->mod_proj(), this);
2962 } else {
2963 // replace a%b with a-((a/b)*b)
2964 Node* mult = new (this) MulINode(d, d->in(2));
2965 Node* sub = new (this) SubINode(d->in(1), mult);
2966 n->subsume_by(sub, this);
2967 }
2968 }
2969 }
2970 break;
2971
2972 case Op_ModL:
2973 if (UseDivMod) {
2974 // Check if a%b and a/b both exist
2975 Node* d = n->find_similar(Op_DivL);
2976 if (d) {
2977 // Replace them with a fused divmod if supported
2978 if (Matcher::has_match_rule(Op_DivModL)) {
2979 DivModLNode* divmod = DivModLNode::make(this, n);
2980 d->subsume_by(divmod->div_proj(), this);
2981 n->subsume_by(divmod->mod_proj(), this);
2982 } else {
2983 // replace a%b with a-((a/b)*b)
2984 Node* mult = new (this) MulLNode(d, d->in(2));
2985 Node* sub = new (this) SubLNode(d->in(1), mult);
2986 n->subsume_by(sub, this);
2987 }
2988 }
2989 }
2990 break;
2991
2992 case Op_LoadVector:
2993 case Op_StoreVector:
2994 break;
2995
2996 case Op_PackB:
2997 case Op_PackS:
2998 case Op_PackI:
2999 case Op_PackF:
3000 case Op_PackL:
3001 case Op_PackD:
3002 if (n->req()-1 > 2) {
3003 // Replace many operand PackNodes with a binary tree for matching
3004 PackNode* p = (PackNode*) n;
3005 Node* btp = p->binary_tree_pack(this, 1, n->req());
3006 n->subsume_by(btp, this);
3007 }
3008 break;
3009 case Op_Loop:
3010 case Op_CountedLoop:
3011 if (n->as_Loop()->is_inner_loop()) {
3012 frc.inc_inner_loop_count();
3013 }
3014 break;
3015 case Op_LShiftI:
3016 case Op_RShiftI:
3017 case Op_URShiftI:
3018 case Op_LShiftL:
3019 case Op_RShiftL:
3020 case Op_URShiftL:
3021 if (Matcher::need_masked_shift_count) {
3022 // The cpu's shift instructions don't restrict the count to the
3023 // lower 5/6 bits. We need to do the masking ourselves.
3024 Node* in2 = n->in(2);
3025 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3026 const TypeInt* t = in2->find_int_type();
3027 if (t != NULL && t->is_con()) {
3028 juint shift = t->get_con();
3029 if (shift > mask) { // Unsigned cmp
3030 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3031 }
3032 } else {
3033 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3034 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3035 n->set_req(2, shift);
3036 }
3037 }
3038 if (in2->outcnt() == 0) { // Remove dead node
3039 in2->disconnect_inputs(NULL, this);
3040 }
3041 }
3042 break;
3043 case Op_MemBarStoreStore:
3044 case Op_MemBarRelease:
3045 // Break the link with AllocateNode: it is no longer useful and
3046 // confuses register allocation.
3047 if (n->req() > MemBarNode::Precedent) {
3048 n->set_req(MemBarNode::Precedent, top());
3049 }
3050 break;
3051 default:
3052 assert( !n->is_Call(), "" );
3053 assert( !n->is_Mem(), "" );
3054 break;
3055 }
3056
3057 // Collect CFG split points
3058 if (n->is_MultiBranch())
3059 frc._tests.push(n);
3060 }
3061
3062 //------------------------------final_graph_reshaping_walk---------------------
3063 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3064 // requires that the walk visits a node's inputs before visiting the node.
3065 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3066 ResourceArea *area = Thread::current()->resource_area();
3067 Unique_Node_List sfpt(area);
3068
3069 frc._visited.set(root->_idx); // first, mark node as visited
3070 uint cnt = root->req();
3071 Node *n = root;
3072 uint i = 0;
3073 while (true) {
3074 if (i < cnt) {
3075 // Place all non-visited non-null inputs onto stack
3076 Node* m = n->in(i);
3077 ++i;
3078 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3079 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
3080 sfpt.push(m);
3081 cnt = m->req();
3082 nstack.push(n, i); // put on stack parent and next input's index
3083 n = m;
3084 i = 0;
3085 }
3086 } else {
3087 // Now do post-visit work
3088 final_graph_reshaping_impl( n, frc );
3089 if (nstack.is_empty())
3090 break; // finished
3091 n = nstack.node(); // Get node from stack
3092 cnt = n->req();
3093 i = nstack.index();
3094 nstack.pop(); // Shift to the next node on stack
3095 }
3096 }
3097
3098 // Skip next transformation if compressed oops are not used.
3099 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3100 (!UseCompressedOops && !UseCompressedClassPointers))
3101 return;
3102
3103 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3104 // It could be done for an uncommon traps or any safepoints/calls
3105 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3106 while (sfpt.size() > 0) {
3107 n = sfpt.pop();
3108 JVMState *jvms = n->as_SafePoint()->jvms();
3109 assert(jvms != NULL, "sanity");
3110 int start = jvms->debug_start();
3111 int end = n->req();
3112 bool is_uncommon = (n->is_CallStaticJava() &&
3113 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3114 for (int j = start; j < end; j++) {
3115 Node* in = n->in(j);
3116 if (in->is_DecodeNarrowPtr()) {
3117 bool safe_to_skip = true;
3118 if (!is_uncommon ) {
3119 // Is it safe to skip?
3120 for (uint i = 0; i < in->outcnt(); i++) {
3121 Node* u = in->raw_out(i);
3122 if (!u->is_SafePoint() ||
3123 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3124 safe_to_skip = false;
3125 }
3126 }
3127 }
3128 if (safe_to_skip) {
3129 n->set_req(j, in->in(1));
3130 }
3131 if (in->outcnt() == 0) {
3132 in->disconnect_inputs(NULL, this);
3133 }
3134 }
3135 }
3136 }
3137 }
3138
3139 //------------------------------final_graph_reshaping--------------------------
3140 // Final Graph Reshaping.
3141 //
3142 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3143 // and not commoned up and forced early. Must come after regular
3144 // optimizations to avoid GVN undoing the cloning. Clone constant
3145 // inputs to Loop Phis; these will be split by the allocator anyways.
3146 // Remove Opaque nodes.
3147 // (2) Move last-uses by commutative operations to the left input to encourage
3148 // Intel update-in-place two-address operations and better register usage
3149 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3150 // calls canonicalizing them back.
3151 // (3) Count the number of double-precision FP ops, single-precision FP ops
3152 // and call sites. On Intel, we can get correct rounding either by
3153 // forcing singles to memory (requires extra stores and loads after each
3154 // FP bytecode) or we can set a rounding mode bit (requires setting and
3155 // clearing the mode bit around call sites). The mode bit is only used
3156 // if the relative frequency of single FP ops to calls is low enough.
3157 // This is a key transform for SPEC mpeg_audio.
3158 // (4) Detect infinite loops; blobs of code reachable from above but not
3159 // below. Several of the Code_Gen algorithms fail on such code shapes,
3160 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3161 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3162 // Detection is by looking for IfNodes where only 1 projection is
3163 // reachable from below or CatchNodes missing some targets.
3164 // (5) Assert for insane oop offsets in debug mode.
3165
3166 bool Compile::final_graph_reshaping() {
3167 // an infinite loop may have been eliminated by the optimizer,
3168 // in which case the graph will be empty.
3169 if (root()->req() == 1) {
3170 record_method_not_compilable("trivial infinite loop");
3171 return true;
3172 }
3173
3174 // Expensive nodes have their control input set to prevent the GVN
3175 // from freely commoning them. There's no GVN beyond this point so
3176 // no need to keep the control input. We want the expensive nodes to
3177 // be freely moved to the least frequent code path by gcm.
3178 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3179 for (int i = 0; i < expensive_count(); i++) {
3180 _expensive_nodes->at(i)->set_req(0, NULL);
3181 }
3182
3183 Final_Reshape_Counts frc;
3184
3185 // Visit everybody reachable!
3186 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3187 Node_Stack nstack(unique() >> 1);
3188 final_graph_reshaping_walk(nstack, root(), frc);
3189
3190 // Check for unreachable (from below) code (i.e., infinite loops).
3191 for( uint i = 0; i < frc._tests.size(); i++ ) {
3192 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3193 // Get number of CFG targets.
3194 // Note that PCTables include exception targets after calls.
3195 uint required_outcnt = n->required_outcnt();
3196 if (n->outcnt() != required_outcnt) {
3197 // Check for a few special cases. Rethrow Nodes never take the
3198 // 'fall-thru' path, so expected kids is 1 less.
3199 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3200 if (n->in(0)->in(0)->is_Call()) {
3201 CallNode *call = n->in(0)->in(0)->as_Call();
3202 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3203 required_outcnt--; // Rethrow always has 1 less kid
3204 } else if (call->req() > TypeFunc::Parms &&
3205 call->is_CallDynamicJava()) {
3206 // Check for null receiver. In such case, the optimizer has
3207 // detected that the virtual call will always result in a null
3208 // pointer exception. The fall-through projection of this CatchNode
3209 // will not be populated.
3210 Node *arg0 = call->in(TypeFunc::Parms);
3211 if (arg0->is_Type() &&
3212 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3213 required_outcnt--;
3214 }
3215 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3216 call->req() > TypeFunc::Parms+1 &&
3217 call->is_CallStaticJava()) {
3218 // Check for negative array length. In such case, the optimizer has
3219 // detected that the allocation attempt will always result in an
3220 // exception. There is no fall-through projection of this CatchNode .
3221 Node *arg1 = call->in(TypeFunc::Parms+1);
3222 if (arg1->is_Type() &&
3223 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3224 required_outcnt--;
3225 }
3226 }
3227 }
3228 }
3229 // Recheck with a better notion of 'required_outcnt'
3230 if (n->outcnt() != required_outcnt) {
3231 record_method_not_compilable("malformed control flow");
3232 return true; // Not all targets reachable!
3233 }
3234 }
3235 // Check that I actually visited all kids. Unreached kids
3236 // must be infinite loops.
3237 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3238 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3239 record_method_not_compilable("infinite loop");
3240 return true; // Found unvisited kid; must be unreach
3241 }
3242 }
3243
3244 // If original bytecodes contained a mixture of floats and doubles
3245 // check if the optimizer has made it homogenous, item (3).
3246 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3247 frc.get_float_count() > 32 &&
3248 frc.get_double_count() == 0 &&
3249 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3250 set_24_bit_selection_and_mode( false, true );
3251 }
3252
3253 set_java_calls(frc.get_java_call_count());
3254 set_inner_loops(frc.get_inner_loop_count());
3255
3256 // No infinite loops, no reason to bail out.
3257 return false;
3258 }
3259
3260 //-----------------------------too_many_traps----------------------------------
3261 // Report if there are too many traps at the current method and bci.
3262 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3263 bool Compile::too_many_traps(ciMethod* method,
3264 int bci,
3265 Deoptimization::DeoptReason reason) {
3266 ciMethodData* md = method->method_data();
3267 if (md->is_empty()) {
3268 // Assume the trap has not occurred, or that it occurred only
3269 // because of a transient condition during start-up in the interpreter.
3270 return false;
3271 }
3272 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3273 if (md->has_trap_at(bci, m, reason) != 0) {
3274 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3275 // Also, if there are multiple reasons, or if there is no per-BCI record,
3276 // assume the worst.
3277 if (log())
3278 log()->elem("observe trap='%s' count='%d'",
3279 Deoptimization::trap_reason_name(reason),
3280 md->trap_count(reason));
3281 return true;
3282 } else {
3283 // Ignore method/bci and see if there have been too many globally.
3284 return too_many_traps(reason, md);
3285 }
3286 }
3287
3288 // Less-accurate variant which does not require a method and bci.
3289 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3290 ciMethodData* logmd) {
3291 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3292 // Too many traps globally.
3293 // Note that we use cumulative trap_count, not just md->trap_count.
3294 if (log()) {
3295 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3296 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3297 Deoptimization::trap_reason_name(reason),
3298 mcount, trap_count(reason));
3299 }
3300 return true;
3301 } else {
3302 // The coast is clear.
3303 return false;
3304 }
3305 }
3306
3307 //--------------------------too_many_recompiles--------------------------------
3308 // Report if there are too many recompiles at the current method and bci.
3309 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3310 // Is not eager to return true, since this will cause the compiler to use
3311 // Action_none for a trap point, to avoid too many recompilations.
3312 bool Compile::too_many_recompiles(ciMethod* method,
3313 int bci,
3314 Deoptimization::DeoptReason reason) {
3315 ciMethodData* md = method->method_data();
3316 if (md->is_empty()) {
3317 // Assume the trap has not occurred, or that it occurred only
3318 // because of a transient condition during start-up in the interpreter.
3319 return false;
3320 }
3321 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3322 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3323 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3324 Deoptimization::DeoptReason per_bc_reason
3325 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3326 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3327 if ((per_bc_reason == Deoptimization::Reason_none
3328 || md->has_trap_at(bci, m, reason) != 0)
3329 // The trap frequency measure we care about is the recompile count:
3330 && md->trap_recompiled_at(bci, m)
3331 && md->overflow_recompile_count() >= bc_cutoff) {
3332 // Do not emit a trap here if it has already caused recompilations.
3333 // Also, if there are multiple reasons, or if there is no per-BCI record,
3334 // assume the worst.
3335 if (log())
3336 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3337 Deoptimization::trap_reason_name(reason),
3338 md->trap_count(reason),
3339 md->overflow_recompile_count());
3340 return true;
3341 } else if (trap_count(reason) != 0
3342 && decompile_count() >= m_cutoff) {
3343 // Too many recompiles globally, and we have seen this sort of trap.
3344 // Use cumulative decompile_count, not just md->decompile_count.
3345 if (log())
3346 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3347 Deoptimization::trap_reason_name(reason),
3348 md->trap_count(reason), trap_count(reason),
3349 md->decompile_count(), decompile_count());
3350 return true;
3351 } else {
3352 // The coast is clear.
3353 return false;
3354 }
3355 }
3356
3357 // Compute when not to trap. Used by matching trap based nodes and
3358 // NullCheck optimization.
3359 void Compile::set_allowed_deopt_reasons() {
3360 _allowed_reasons = 0;
3361 if (is_method_compilation()) {
3362 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3363 assert(rs < BitsPerInt, "recode bit map");
3364 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3365 _allowed_reasons |= nth_bit(rs);
3366 }
3367 }
3368 }
3369 }
3370
3371 #ifndef PRODUCT
3372 //------------------------------verify_graph_edges---------------------------
3373 // Walk the Graph and verify that there is a one-to-one correspondence
3374 // between Use-Def edges and Def-Use edges in the graph.
3375 void Compile::verify_graph_edges(bool no_dead_code) {
3376 if (VerifyGraphEdges) {
3377 ResourceArea *area = Thread::current()->resource_area();
3378 Unique_Node_List visited(area);
3379 // Call recursive graph walk to check edges
3380 _root->verify_edges(visited);
3381 if (no_dead_code) {
3382 // Now make sure that no visited node is used by an unvisited node.
3383 bool dead_nodes = 0;
3384 Unique_Node_List checked(area);
3385 while (visited.size() > 0) {
3386 Node* n = visited.pop();
3387 checked.push(n);
3388 for (uint i = 0; i < n->outcnt(); i++) {
3389 Node* use = n->raw_out(i);
3390 if (checked.member(use)) continue; // already checked
3391 if (visited.member(use)) continue; // already in the graph
3392 if (use->is_Con()) continue; // a dead ConNode is OK
3393 // At this point, we have found a dead node which is DU-reachable.
3394 if (dead_nodes++ == 0)
3395 tty->print_cr("*** Dead nodes reachable via DU edges:");
3396 use->dump(2);
3397 tty->print_cr("---");
3398 checked.push(use); // No repeats; pretend it is now checked.
3399 }
3400 }
3401 assert(dead_nodes == 0, "using nodes must be reachable from root");
3402 }
3403 }
3404 }
3405
3406 // Verify GC barriers consistency
3407 // Currently supported:
3408 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3409 void Compile::verify_barriers() {
3410 if (UseG1GC) {
3411 // Verify G1 pre-barriers
3412 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3413
3414 ResourceArea *area = Thread::current()->resource_area();
3415 Unique_Node_List visited(area);
3416 Node_List worklist(area);
3417 // We're going to walk control flow backwards starting from the Root
3418 worklist.push(_root);
3419 while (worklist.size() > 0) {
3420 Node* x = worklist.pop();
3421 if (x == NULL || x == top()) continue;
3422 if (visited.member(x)) {
3423 continue;
3424 } else {
3425 visited.push(x);
3426 }
3427
3428 if (x->is_Region()) {
3429 for (uint i = 1; i < x->req(); i++) {
3430 worklist.push(x->in(i));
3431 }
3432 } else {
3433 worklist.push(x->in(0));
3434 // We are looking for the pattern:
3435 // /->ThreadLocal
3436 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3437 // \->ConI(0)
3438 // We want to verify that the If and the LoadB have the same control
3439 // See GraphKit::g1_write_barrier_pre()
3440 if (x->is_If()) {
3441 IfNode *iff = x->as_If();
3442 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3443 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3444 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3445 && cmp->in(1)->is_Load()) {
3446 LoadNode* load = cmp->in(1)->as_Load();
3447 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3448 && load->in(2)->in(3)->is_Con()
3449 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3450
3451 Node* if_ctrl = iff->in(0);
3452 Node* load_ctrl = load->in(0);
3453
3454 if (if_ctrl != load_ctrl) {
3455 // Skip possible CProj->NeverBranch in infinite loops
3456 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3457 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3458 if_ctrl = if_ctrl->in(0)->in(0);
3459 }
3460 }
3461 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3462 }
3463 }
3464 }
3465 }
3466 }
3467 }
3468 }
3469 }
3470
3471 #endif
3472
3473 // The Compile object keeps track of failure reasons separately from the ciEnv.
3474 // This is required because there is not quite a 1-1 relation between the
3475 // ciEnv and its compilation task and the Compile object. Note that one
3476 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3477 // to backtrack and retry without subsuming loads. Other than this backtracking
3478 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3479 // by the logic in C2Compiler.
3480 void Compile::record_failure(const char* reason) {
3481 if (log() != NULL) {
3482 log()->elem("failure reason='%s' phase='compile'", reason);
3483 }
3484 if (_failure_reason == NULL) {
3485 // Record the first failure reason.
3486 _failure_reason = reason;
3487 }
3488
3489 EventCompilerFailure event;
3490 if (event.should_commit()) {
3491 event.set_compileID(Compile::compile_id());
3492 event.set_failure(reason);
3493 event.commit();
3494 }
3495
3496 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3497 C->print_method(PHASE_FAILURE);
3498 }
3499 _root = NULL; // flush the graph, too
3500 }
3501
3502 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3503 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3504 _phase_name(name), _dolog(dolog)
3505 {
3506 if (dolog) {
3507 C = Compile::current();
3508 _log = C->log();
3509 } else {
3510 C = NULL;
3511 _log = NULL;
3512 }
3513 if (_log != NULL) {
3514 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3515 _log->stamp();
3516 _log->end_head();
3517 }
3518 }
3519
3520 Compile::TracePhase::~TracePhase() {
3521
3522 C = Compile::current();
3523 if (_dolog) {
3524 _log = C->log();
3525 } else {
3526 _log = NULL;
3527 }
3528
3529 #ifdef ASSERT
3530 if (PrintIdealNodeCount) {
3531 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3532 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3533 }
3534
3535 if (VerifyIdealNodeCount) {
3536 Compile::current()->print_missing_nodes();
3537 }
3538 #endif
3539
3540 if (_log != NULL) {
3541 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3542 }
3543 }
3544
3545 //=============================================================================
3546 // Two Constant's are equal when the type and the value are equal.
3547 bool Compile::Constant::operator==(const Constant& other) {
3548 if (type() != other.type() ) return false;
3549 if (can_be_reused() != other.can_be_reused()) return false;
3550 // For floating point values we compare the bit pattern.
3551 switch (type()) {
3552 case T_FLOAT: return (_v._value.i == other._v._value.i);
3553 case T_LONG:
3554 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3555 case T_OBJECT:
3556 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3557 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3558 case T_METADATA: return (_v._metadata == other._v._metadata);
3559 default: ShouldNotReachHere();
3560 }
3561 return false;
3562 }
3563
3564 static int type_to_size_in_bytes(BasicType t) {
3565 switch (t) {
3566 case T_LONG: return sizeof(jlong );
3567 case T_FLOAT: return sizeof(jfloat );
3568 case T_DOUBLE: return sizeof(jdouble);
3569 case T_METADATA: return sizeof(Metadata*);
3570 // We use T_VOID as marker for jump-table entries (labels) which
3571 // need an internal word relocation.
3572 case T_VOID:
3573 case T_ADDRESS:
3574 case T_OBJECT: return sizeof(jobject);
3575 }
3576
3577 ShouldNotReachHere();
3578 return -1;
3579 }
3580
3581 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3582 // sort descending
3583 if (a->freq() > b->freq()) return -1;
3584 if (a->freq() < b->freq()) return 1;
3585 return 0;
3586 }
3587
3588 void Compile::ConstantTable::calculate_offsets_and_size() {
3589 // First, sort the array by frequencies.
3590 _constants.sort(qsort_comparator);
3591
3592 #ifdef ASSERT
3593 // Make sure all jump-table entries were sorted to the end of the
3594 // array (they have a negative frequency).
3595 bool found_void = false;
3596 for (int i = 0; i < _constants.length(); i++) {
3597 Constant con = _constants.at(i);
3598 if (con.type() == T_VOID)
3599 found_void = true; // jump-tables
3600 else
3601 assert(!found_void, "wrong sorting");
3602 }
3603 #endif
3604
3605 int offset = 0;
3606 for (int i = 0; i < _constants.length(); i++) {
3607 Constant* con = _constants.adr_at(i);
3608
3609 // Align offset for type.
3610 int typesize = type_to_size_in_bytes(con->type());
3611 offset = align_size_up(offset, typesize);
3612 con->set_offset(offset); // set constant's offset
3613
3614 if (con->type() == T_VOID) {
3615 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3616 offset = offset + typesize * n->outcnt(); // expand jump-table
3617 } else {
3618 offset = offset + typesize;
3619 }
3620 }
3621
3622 // Align size up to the next section start (which is insts; see
3623 // CodeBuffer::align_at_start).
3624 assert(_size == -1, "already set?");
3625 _size = align_size_up(offset, CodeEntryAlignment);
3626 }
3627
3628 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3629 MacroAssembler _masm(&cb);
3630 for (int i = 0; i < _constants.length(); i++) {
3631 Constant con = _constants.at(i);
3632 address constant_addr;
3633 switch (con.type()) {
3634 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3635 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3636 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3637 case T_OBJECT: {
3638 jobject obj = con.get_jobject();
3639 int oop_index = _masm.oop_recorder()->find_index(obj);
3640 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3641 break;
3642 }
3643 case T_ADDRESS: {
3644 address addr = (address) con.get_jobject();
3645 constant_addr = _masm.address_constant(addr);
3646 break;
3647 }
3648 // We use T_VOID as marker for jump-table entries (labels) which
3649 // need an internal word relocation.
3650 case T_VOID: {
3651 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3652 // Fill the jump-table with a dummy word. The real value is
3653 // filled in later in fill_jump_table.
3654 address dummy = (address) n;
3655 constant_addr = _masm.address_constant(dummy);
3656 // Expand jump-table
3657 for (uint i = 1; i < n->outcnt(); i++) {
3658 address temp_addr = _masm.address_constant(dummy + i);
3659 assert(temp_addr, "consts section too small");
3660 }
3661 break;
3662 }
3663 case T_METADATA: {
3664 Metadata* obj = con.get_metadata();
3665 int metadata_index = _masm.oop_recorder()->find_index(obj);
3666 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3667 break;
3668 }
3669 default: ShouldNotReachHere();
3670 }
3671 assert(constant_addr, "consts section too small");
3672 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3673 }
3674 }
3675
3676 int Compile::ConstantTable::find_offset(Constant& con) const {
3677 int idx = _constants.find(con);
3678 assert(idx != -1, "constant must be in constant table");
3679 int offset = _constants.at(idx).offset();
3680 assert(offset != -1, "constant table not emitted yet?");
3681 return offset;
3682 }
3683
3684 void Compile::ConstantTable::add(Constant& con) {
3685 if (con.can_be_reused()) {
3686 int idx = _constants.find(con);
3687 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3688 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3689 return;
3690 }
3691 }
3692 (void) _constants.append(con);
3693 }
3694
3695 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3696 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3697 Constant con(type, value, b->_freq);
3698 add(con);
3699 return con;
3700 }
3701
3702 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3703 Constant con(metadata);
3704 add(con);
3705 return con;
3706 }
3707
3708 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3709 jvalue value;
3710 BasicType type = oper->type()->basic_type();
3711 switch (type) {
3712 case T_LONG: value.j = oper->constantL(); break;
3713 case T_FLOAT: value.f = oper->constantF(); break;
3714 case T_DOUBLE: value.d = oper->constantD(); break;
3715 case T_OBJECT:
3716 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3717 case T_METADATA: return add((Metadata*)oper->constant()); break;
3718 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3719 }
3720 return add(n, type, value);
3721 }
3722
3723 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3724 jvalue value;
3725 // We can use the node pointer here to identify the right jump-table
3726 // as this method is called from Compile::Fill_buffer right before
3727 // the MachNodes are emitted and the jump-table is filled (means the
3728 // MachNode pointers do not change anymore).
3729 value.l = (jobject) n;
3730 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3731 add(con);
3732 return con;
3733 }
3734
3735 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3736 // If called from Compile::scratch_emit_size do nothing.
3737 if (Compile::current()->in_scratch_emit_size()) return;
3738
3739 assert(labels.is_nonempty(), "must be");
3740 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3741
3742 // Since MachConstantNode::constant_offset() also contains
3743 // table_base_offset() we need to subtract the table_base_offset()
3744 // to get the plain offset into the constant table.
3745 int offset = n->constant_offset() - table_base_offset();
3746
3747 MacroAssembler _masm(&cb);
3748 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3749
3750 for (uint i = 0; i < n->outcnt(); i++) {
3751 address* constant_addr = &jump_table_base[i];
3752 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)));
3753 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3754 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3755 }
3756 }
3757
3758 void Compile::dump_inlining() {
3759 if (print_inlining() || print_intrinsics()) {
3760 // Print inlining message for candidates that we couldn't inline
3761 // for lack of space or non constant receiver
3762 for (int i = 0; i < _late_inlines.length(); i++) {
3763 CallGenerator* cg = _late_inlines.at(i);
3764 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3765 }
3766 Unique_Node_List useful;
3767 useful.push(root());
3768 for (uint next = 0; next < useful.size(); ++next) {
3769 Node* n = useful.at(next);
3770 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3771 CallNode* call = n->as_Call();
3772 CallGenerator* cg = call->generator();
3773 cg->print_inlining_late("receiver not constant");
3774 }
3775 uint max = n->len();
3776 for ( uint i = 0; i < max; ++i ) {
3777 Node *m = n->in(i);
3778 if ( m == NULL ) continue;
3779 useful.push(m);
3780 }
3781 }
3782 for (int i = 0; i < _print_inlining_list->length(); i++) {
3783 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3784 }
3785 }
3786 }
3787
3788 // Dump inlining replay data to the stream.
3789 // Don't change thread state and acquire any locks.
3790 void Compile::dump_inline_data(outputStream* out) {
3791 InlineTree* inl_tree = ilt();
3792 if (inl_tree != NULL) {
3793 out->print(" inline %d", inl_tree->count());
3794 inl_tree->dump_replay_data(out);
3795 }
3796 }
3797
3798 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3799 if (n1->Opcode() < n2->Opcode()) return -1;
3800 else if (n1->Opcode() > n2->Opcode()) return 1;
3801
3802 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()));
3803 for (uint i = 1; i < n1->req(); i++) {
3804 if (n1->in(i) < n2->in(i)) return -1;
3805 else if (n1->in(i) > n2->in(i)) return 1;
3806 }
3807
3808 return 0;
3809 }
3810
3811 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3812 Node* n1 = *n1p;
3813 Node* n2 = *n2p;
3814
3815 return cmp_expensive_nodes(n1, n2);
3816 }
3817
3818 void Compile::sort_expensive_nodes() {
3819 if (!expensive_nodes_sorted()) {
3820 _expensive_nodes->sort(cmp_expensive_nodes);
3821 }
3822 }
3823
3824 bool Compile::expensive_nodes_sorted() const {
3825 for (int i = 1; i < _expensive_nodes->length(); i++) {
3826 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3827 return false;
3828 }
3829 }
3830 return true;
3831 }
3832
3833 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3834 if (_expensive_nodes->length() == 0) {
3835 return false;
3836 }
3837
3838 assert(OptimizeExpensiveOps, "optimization off?");
3839
3840 // Take this opportunity to remove dead nodes from the list
3841 int j = 0;
3842 for (int i = 0; i < _expensive_nodes->length(); i++) {
3843 Node* n = _expensive_nodes->at(i);
3844 if (!n->is_unreachable(igvn)) {
3845 assert(n->is_expensive(), "should be expensive");
3846 _expensive_nodes->at_put(j, n);
3847 j++;
3848 }
3849 }
3850 _expensive_nodes->trunc_to(j);
3851
3852 // Then sort the list so that similar nodes are next to each other
3853 // and check for at least two nodes of identical kind with same data
3854 // inputs.
3855 sort_expensive_nodes();
3856
3857 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3858 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3859 return true;
3860 }
3861 }
3862
3863 return false;
3864 }
3865
3866 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3867 if (_expensive_nodes->length() == 0) {
3868 return;
3869 }
3870
3871 assert(OptimizeExpensiveOps, "optimization off?");
3872
3873 // Sort to bring similar nodes next to each other and clear the
3874 // control input of nodes for which there's only a single copy.
3875 sort_expensive_nodes();
3876
3877 int j = 0;
3878 int identical = 0;
3879 int i = 0;
3880 for (; i < _expensive_nodes->length()-1; i++) {
3881 assert(j <= i, "can't write beyond current index");
3882 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3883 identical++;
3884 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3885 continue;
3886 }
3887 if (identical > 0) {
3888 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3889 identical = 0;
3890 } else {
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 }
3897 if (identical > 0) {
3898 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3899 } else if (_expensive_nodes->length() >= 1) {
3900 Node* n = _expensive_nodes->at(i);
3901 igvn.hash_delete(n);
3902 n->set_req(0, NULL);
3903 igvn.hash_insert(n);
3904 }
3905 _expensive_nodes->trunc_to(j);
3906 }
3907
3908 void Compile::add_expensive_node(Node * n) {
3909 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3910 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3911 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3912 if (OptimizeExpensiveOps) {
3913 _expensive_nodes->append(n);
3914 } else {
3915 // Clear control input and let IGVN optimize expensive nodes if
3916 // OptimizeExpensiveOps is off.
3917 n->set_req(0, NULL);
3918 }
3919 }
3920
3921 /**
3922 * Remove the speculative part of types and clean up the graph
3923 */
3924 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
3925 if (UseTypeSpeculation) {
3926 Unique_Node_List worklist;
3927 worklist.push(root());
3928 int modified = 0;
3929 // Go over all type nodes that carry a speculative type, drop the
3930 // speculative part of the type and enqueue the node for an igvn
3931 // which may optimize it out.
3932 for (uint next = 0; next < worklist.size(); ++next) {
3933 Node *n = worklist.at(next);
3934 if (n->is_Type()) {
3935 TypeNode* tn = n->as_Type();
3936 const Type* t = tn->type();
3937 const Type* t_no_spec = t->remove_speculative();
3938 if (t_no_spec != t) {
3939 bool in_hash = igvn.hash_delete(n);
3940 assert(in_hash, "node should be in igvn hash table");
3941 tn->set_type(t_no_spec);
3942 igvn.hash_insert(n);
3943 igvn._worklist.push(n); // give it a chance to go away
3944 modified++;
3945 }
3946 }
3947 uint max = n->len();
3948 for( uint i = 0; i < max; ++i ) {
3949 Node *m = n->in(i);
3950 if (not_a_node(m)) continue;
3951 worklist.push(m);
3952 }
3953 }
3954 // Drop the speculative part of all types in the igvn's type table
3955 igvn.remove_speculative_types();
3956 if (modified > 0) {
3957 igvn.optimize();
3958 }
3959 #ifdef ASSERT
3960 // Verify that after the IGVN is over no speculative type has resurfaced
3961 worklist.clear();
3962 worklist.push(root());
3963 for (uint next = 0; next < worklist.size(); ++next) {
3964 Node *n = worklist.at(next);
3965 const Type* t = igvn.type(n);
3966 assert(t == t->remove_speculative(), "no more speculative types");
3967 if (n->is_Type()) {
3968 t = n->as_Type()->type();
3969 assert(t == t->remove_speculative(), "no more speculative types");
3970 }
3971 uint max = n->len();
3972 for( uint i = 0; i < max; ++i ) {
3973 Node *m = n->in(i);
3974 if (not_a_node(m)) continue;
3975 worklist.push(m);
3976 }
3977 }
3978 igvn.check_no_speculative_types();
3979 #endif
3980 }
3981 }
3982
3983 // Auxiliary method to support randomized stressing/fuzzing.
3984 //
3985 // This method can be called the arbitrary number of times, with current count
3986 // as the argument. The logic allows selecting a single candidate from the
3987 // running list of candidates as follows:
3988 // int count = 0;
3989 // Cand* selected = null;
3990 // while(cand = cand->next()) {
3991 // if (randomized_select(++count)) {
3992 // selected = cand;
3993 // }
3994 // }
3995 //
3996 // Including count equalizes the chances any candidate is "selected".
3997 // This is useful when we don't have the complete list of candidates to choose
3998 // from uniformly. In this case, we need to adjust the randomicity of the
3999 // selection, or else we will end up biasing the selection towards the latter
4000 // candidates.
4001 //
4002 // Quick back-envelope calculation shows that for the list of n candidates
4003 // the equal probability for the candidate to persist as "best" can be
4004 // achieved by replacing it with "next" k-th candidate with the probability
4005 // of 1/k. It can be easily shown that by the end of the run, the
4006 // probability for any candidate is converged to 1/n, thus giving the
4007 // uniform distribution among all the candidates.
4008 //
4009 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4010 #define RANDOMIZED_DOMAIN_POW 29
4011 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4012 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4013 bool Compile::randomized_select(int count) {
4014 assert(count > 0, "only positive");
4015 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4016 }