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 // To bang the stack of this compiled method we use the stack size
444 // that the interpreter would need in case of a deoptimization. This
445 // removes the need to bang the stack in the deoptimization blob which
446 // in turn simplifies stack overflow handling.
447 int Compile::bang_size_in_bytes() const {
448 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
449 }
450
451 // ============================================================================
452 //------------------------------CompileWrapper---------------------------------
453 class CompileWrapper : public StackObj {
454 Compile *const _compile;
455 public:
456 CompileWrapper(Compile* compile);
457
458 ~CompileWrapper();
459 };
460
461 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
462 // the Compile* pointer is stored in the current ciEnv:
463 ciEnv* env = compile->env();
464 assert(env == ciEnv::current(), "must already be a ciEnv active");
465 assert(env->compiler_data() == NULL, "compile already active?");
466 env->set_compiler_data(compile);
467 assert(compile == Compile::current(), "sanity");
468
469 compile->set_type_dict(NULL);
470 compile->set_type_hwm(NULL);
471 compile->set_type_last_size(0);
472 compile->set_last_tf(NULL, NULL);
473 compile->set_indexSet_arena(NULL);
474 compile->set_indexSet_free_block_list(NULL);
475 compile->init_type_arena();
476 Type::Initialize(compile);
477 _compile->set_scratch_buffer_blob(NULL);
478 _compile->begin_method();
479 }
480 CompileWrapper::~CompileWrapper() {
481 _compile->end_method();
482 if (_compile->scratch_buffer_blob() != NULL)
483 BufferBlob::free(_compile->scratch_buffer_blob());
484 _compile->env()->set_compiler_data(NULL);
485 }
486
487
488 //----------------------------print_compile_messages---------------------------
489 void Compile::print_compile_messages() {
490 #ifndef PRODUCT
491 // Check if recompiling
492 if (_subsume_loads == false && PrintOpto) {
493 // Recompiling without allowing machine instructions to subsume loads
494 tty->print_cr("*********************************************************");
495 tty->print_cr("** Bailout: Recompile without subsuming loads **");
496 tty->print_cr("*********************************************************");
497 }
498 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
499 // Recompiling without escape analysis
500 tty->print_cr("*********************************************************");
501 tty->print_cr("** Bailout: Recompile without escape analysis **");
502 tty->print_cr("*********************************************************");
503 }
504 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
505 // Recompiling without boxing elimination
506 tty->print_cr("*********************************************************");
507 tty->print_cr("** Bailout: Recompile without boxing elimination **");
508 tty->print_cr("*********************************************************");
509 }
510 if (env()->break_at_compile()) {
511 // Open the debugger when compiling this method.
512 tty->print("### Breaking when compiling: ");
513 method()->print_short_name();
514 tty->cr();
515 BREAKPOINT;
516 }
517
518 if( PrintOpto ) {
519 if (is_osr_compilation()) {
520 tty->print("[OSR]%3d", _compile_id);
521 } else {
522 tty->print("%3d", _compile_id);
523 }
524 }
525 #endif
526 }
527
528
529 //-----------------------init_scratch_buffer_blob------------------------------
530 // Construct a temporary BufferBlob and cache it for this compile.
531 void Compile::init_scratch_buffer_blob(int const_size) {
532 // If there is already a scratch buffer blob allocated and the
533 // constant section is big enough, use it. Otherwise free the
534 // current and allocate a new one.
535 BufferBlob* blob = scratch_buffer_blob();
536 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
537 // Use the current blob.
538 } else {
539 if (blob != NULL) {
540 BufferBlob::free(blob);
541 }
542
543 ResourceMark rm;
544 _scratch_const_size = const_size;
545 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
546 blob = BufferBlob::create("Compile::scratch_buffer", size);
547 // Record the buffer blob for next time.
548 set_scratch_buffer_blob(blob);
549 // Have we run out of code space?
550 if (scratch_buffer_blob() == NULL) {
551 // Let CompilerBroker disable further compilations.
552 record_failure("Not enough space for scratch buffer in CodeCache");
553 return;
554 }
555 }
556
557 // Initialize the relocation buffers
558 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
559 set_scratch_locs_memory(locs_buf);
560 }
561
562
563 //-----------------------scratch_emit_size-------------------------------------
564 // Helper function that computes size by emitting code
565 uint Compile::scratch_emit_size(const Node* n) {
566 // Start scratch_emit_size section.
567 set_in_scratch_emit_size(true);
568
569 // Emit into a trash buffer and count bytes emitted.
570 // This is a pretty expensive way to compute a size,
571 // but it works well enough if seldom used.
572 // All common fixed-size instructions are given a size
573 // method by the AD file.
574 // Note that the scratch buffer blob and locs memory are
575 // allocated at the beginning of the compile task, and
576 // may be shared by several calls to scratch_emit_size.
577 // The allocation of the scratch buffer blob is particularly
578 // expensive, since it has to grab the code cache lock.
579 BufferBlob* blob = this->scratch_buffer_blob();
580 assert(blob != NULL, "Initialize BufferBlob at start");
581 assert(blob->size() > MAX_inst_size, "sanity");
582 relocInfo* locs_buf = scratch_locs_memory();
583 address blob_begin = blob->content_begin();
584 address blob_end = (address)locs_buf;
585 assert(blob->content_contains(blob_end), "sanity");
586 CodeBuffer buf(blob_begin, blob_end - blob_begin);
587 buf.initialize_consts_size(_scratch_const_size);
588 buf.initialize_stubs_size(MAX_stubs_size);
589 assert(locs_buf != NULL, "sanity");
590 int lsize = MAX_locs_size / 3;
591 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
592 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
593 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
594
595 // Do the emission.
596
597 Label fakeL; // Fake label for branch instructions.
598 Label* saveL = NULL;
599 uint save_bnum = 0;
600 bool is_branch = n->is_MachBranch();
601 if (is_branch) {
602 MacroAssembler masm(&buf);
603 masm.bind(fakeL);
604 n->as_MachBranch()->save_label(&saveL, &save_bnum);
605 n->as_MachBranch()->label_set(&fakeL, 0);
606 }
607 n->emit(buf, this->regalloc());
608 if (is_branch) // Restore label.
609 n->as_MachBranch()->label_set(saveL, save_bnum);
610
611 // End scratch_emit_size section.
612 set_in_scratch_emit_size(false);
613
614 return buf.insts_size();
615 }
616
617
618 // ============================================================================
619 //------------------------------Compile standard-------------------------------
620 debug_only( int Compile::_debug_idx = 100000; )
621
622 // Compile a method. entry_bci is -1 for normal compilations and indicates
623 // the continuation bci for on stack replacement.
624
625
626 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
627 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
628 : Phase(Compiler),
629 _env(ci_env),
630 _log(ci_env->log()),
631 _compile_id(ci_env->compile_id()),
632 _save_argument_registers(false),
633 _stub_name(NULL),
634 _stub_function(NULL),
635 _stub_entry_point(NULL),
636 _method(target),
637 _entry_bci(osr_bci),
638 _initial_gvn(NULL),
639 _for_igvn(NULL),
640 _warm_calls(NULL),
641 _subsume_loads(subsume_loads),
642 _do_escape_analysis(do_escape_analysis),
643 _eliminate_boxing(eliminate_boxing),
644 _failure_reason(NULL),
645 _code_buffer("Compile::Fill_buffer"),
646 _orig_pc_slot(0),
647 _orig_pc_slot_offset_in_bytes(0),
648 _has_method_handle_invokes(false),
649 _mach_constant_base_node(NULL),
650 _node_bundling_limit(0),
651 _node_bundling_base(NULL),
652 _java_calls(0),
653 _inner_loops(0),
654 _scratch_const_size(-1),
655 _in_scratch_emit_size(false),
656 _dead_node_list(comp_arena()),
657 _dead_node_count(0),
658 #ifndef PRODUCT
659 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
660 _in_dump_cnt(0),
661 _printer(IdealGraphPrinter::printer()),
662 #endif
663 _congraph(NULL),
664 _replay_inline_data(NULL),
665 _late_inlines(comp_arena(), 2, 0, NULL),
666 _string_late_inlines(comp_arena(), 2, 0, NULL),
667 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
668 _late_inlines_pos(0),
669 _number_of_mh_late_inlines(0),
670 _inlining_progress(false),
671 _inlining_incrementally(false),
672 _print_inlining_list(NULL),
673 _print_inlining_stream(NULL),
674 _print_inlining_idx(0),
675 _preserve_jvm_state(0),
676 _interpreter_frame_size(0) {
677 C = this;
678
679 CompileWrapper cw(this);
680 #ifndef PRODUCT
681 if (TimeCompiler2) {
682 tty->print(" ");
683 target->holder()->name()->print();
684 tty->print(".");
685 target->print_short_name();
686 tty->print(" ");
687 }
688 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
689 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
690 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
691 if (!print_opto_assembly) {
692 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
693 if (print_assembly && !Disassembler::can_decode()) {
694 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
695 print_opto_assembly = true;
696 }
697 }
698 set_print_assembly(print_opto_assembly);
699 set_parsed_irreducible_loop(false);
700
701 if (method()->has_option("ReplayInline")) {
702 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
703 }
704 #endif
705 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
706 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
707
708 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
709 // Make sure the method being compiled gets its own MDO,
710 // so we can at least track the decompile_count().
711 // Need MDO to record RTM code generation state.
712 method()->ensure_method_data();
713 }
714
715 Init(::AliasLevel);
716
717
718 print_compile_messages();
719
720 _ilt = InlineTree::build_inline_tree_root();
721
722 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
723 assert(num_alias_types() >= AliasIdxRaw, "");
724
725 #define MINIMUM_NODE_HASH 1023
726 // Node list that Iterative GVN will start with
727 Unique_Node_List for_igvn(comp_arena());
728 set_for_igvn(&for_igvn);
729
730 // GVN that will be run immediately on new nodes
731 uint estimated_size = method()->code_size()*4+64;
732 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
733 PhaseGVN gvn(node_arena(), estimated_size);
734 set_initial_gvn(&gvn);
735
736 print_inlining_init();
737 { // Scope for timing the parser
738 TracePhase t3("parse", &_t_parser, true);
739
740 // Put top into the hash table ASAP.
741 initial_gvn()->transform_no_reclaim(top());
742
743 // Set up tf(), start(), and find a CallGenerator.
744 CallGenerator* cg = NULL;
745 if (is_osr_compilation()) {
746 const TypeTuple *domain = StartOSRNode::osr_domain();
747 const TypeTuple *range = TypeTuple::make_range(method()->signature());
748 init_tf(TypeFunc::make(domain, range));
749 StartNode* s = new (this) StartOSRNode(root(), domain);
750 initial_gvn()->set_type_bottom(s);
751 init_start(s);
752 cg = CallGenerator::for_osr(method(), entry_bci());
753 } else {
754 // Normal case.
755 init_tf(TypeFunc::make(method()));
756 StartNode* s = new (this) StartNode(root(), tf()->domain());
757 initial_gvn()->set_type_bottom(s);
758 init_start(s);
759 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
760 // With java.lang.ref.reference.get() we must go through the
761 // intrinsic when G1 is enabled - even when get() is the root
762 // method of the compile - so that, if necessary, the value in
763 // the referent field of the reference object gets recorded by
764 // the pre-barrier code.
765 // Specifically, if G1 is enabled, the value in the referent
766 // field is recorded by the G1 SATB pre barrier. This will
767 // result in the referent being marked live and the reference
768 // object removed from the list of discovered references during
769 // reference processing.
770 cg = find_intrinsic(method(), false);
771 }
772 if (cg == NULL) {
773 float past_uses = method()->interpreter_invocation_count();
774 float expected_uses = past_uses;
775 cg = CallGenerator::for_inline(method(), expected_uses);
776 }
777 }
778 if (failing()) return;
779 if (cg == NULL) {
780 record_method_not_compilable_all_tiers("cannot parse method");
781 return;
782 }
783 JVMState* jvms = build_start_state(start(), tf());
784 if ((jvms = cg->generate(jvms, NULL)) == NULL) {
785 record_method_not_compilable("method parse failed");
786 return;
787 }
788 GraphKit kit(jvms);
789
790 if (!kit.stopped()) {
791 // Accept return values, and transfer control we know not where.
792 // This is done by a special, unique ReturnNode bound to root.
793 return_values(kit.jvms());
794 }
795
796 if (kit.has_exceptions()) {
797 // Any exceptions that escape from this call must be rethrown
798 // to whatever caller is dynamically above us on the stack.
799 // This is done by a special, unique RethrowNode bound to root.
800 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
801 }
802
803 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
804
805 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
806 inline_string_calls(true);
807 }
808
809 if (failing()) return;
810
811 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
812
813 // Remove clutter produced by parsing.
814 if (!failing()) {
815 ResourceMark rm;
816 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
817 }
818 }
819
820 // Note: Large methods are capped off in do_one_bytecode().
821 if (failing()) return;
822
823 // After parsing, node notes are no longer automagic.
824 // They must be propagated by register_new_node_with_optimizer(),
825 // clone(), or the like.
826 set_default_node_notes(NULL);
827
828 for (;;) {
829 int successes = Inline_Warm();
830 if (failing()) return;
831 if (successes == 0) break;
832 }
833
834 // Drain the list.
835 Finish_Warm();
836 #ifndef PRODUCT
837 if (_printer) {
838 _printer->print_inlining(this);
839 }
840 #endif
841
842 if (failing()) return;
843 NOT_PRODUCT( verify_graph_edges(); )
844
845 // Now optimize
846 Optimize();
847 if (failing()) return;
848 NOT_PRODUCT( verify_graph_edges(); )
849
850 #ifndef PRODUCT
851 if (PrintIdeal) {
852 ttyLocker ttyl; // keep the following output all in one block
853 // This output goes directly to the tty, not the compiler log.
854 // To enable tools to match it up with the compilation activity,
855 // be sure to tag this tty output with the compile ID.
856 if (xtty != NULL) {
857 xtty->head("ideal compile_id='%d'%s", compile_id(),
858 is_osr_compilation() ? " compile_kind='osr'" :
859 "");
860 }
861 root()->dump(9999);
862 if (xtty != NULL) {
863 xtty->tail("ideal");
864 }
865 }
866 #endif
867
868 NOT_PRODUCT( verify_barriers(); )
869
870 // Dump compilation data to replay it.
871 if (method()->has_option("DumpReplay")) {
872 env()->dump_replay_data(_compile_id);
873 }
874 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
875 env()->dump_inline_data(_compile_id);
876 }
877
878 // Now that we know the size of all the monitors we can add a fixed slot
879 // for the original deopt pc.
880
881 _orig_pc_slot = fixed_slots();
882 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
883 set_fixed_slots(next_slot);
884
885 // Compute when to use implicit null checks. Used by matching trap based
886 // nodes and NullCheck optimization.
887 set_allowed_deopt_reasons();
888
889 // Now generate code
890 Code_Gen();
891 if (failing()) return;
892
893 // Check if we want to skip execution of all compiled code.
894 {
895 #ifndef PRODUCT
896 if (OptoNoExecute) {
897 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
898 return;
899 }
900 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
901 #endif
902
903 if (is_osr_compilation()) {
904 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
905 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
906 } else {
907 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
908 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
909 }
910
911 env()->register_method(_method, _entry_bci,
912 &_code_offsets,
913 _orig_pc_slot_offset_in_bytes,
914 code_buffer(),
915 frame_size_in_words(), _oop_map_set,
916 &_handler_table, &_inc_table,
917 compiler,
918 env()->comp_level(),
919 has_unsafe_access(),
920 SharedRuntime::is_wide_vector(max_vector_size()),
921 rtm_state()
922 );
923
924 if (log() != NULL) // Print code cache state into compiler log
925 log()->code_cache_state();
926 }
927 }
928
929 //------------------------------Compile----------------------------------------
930 // Compile a runtime stub
931 Compile::Compile( ciEnv* ci_env,
932 TypeFunc_generator generator,
933 address stub_function,
934 const char *stub_name,
935 int is_fancy_jump,
936 bool pass_tls,
937 bool save_arg_registers,
938 bool return_pc )
939 : Phase(Compiler),
940 _env(ci_env),
941 _log(ci_env->log()),
942 _compile_id(0),
943 _save_argument_registers(save_arg_registers),
944 _method(NULL),
945 _stub_name(stub_name),
946 _stub_function(stub_function),
947 _stub_entry_point(NULL),
948 _entry_bci(InvocationEntryBci),
949 _initial_gvn(NULL),
950 _for_igvn(NULL),
951 _warm_calls(NULL),
952 _orig_pc_slot(0),
953 _orig_pc_slot_offset_in_bytes(0),
954 _subsume_loads(true),
955 _do_escape_analysis(false),
956 _eliminate_boxing(false),
957 _failure_reason(NULL),
958 _code_buffer("Compile::Fill_buffer"),
959 _has_method_handle_invokes(false),
960 _mach_constant_base_node(NULL),
961 _node_bundling_limit(0),
962 _node_bundling_base(NULL),
963 _java_calls(0),
964 _inner_loops(0),
965 #ifndef PRODUCT
966 _trace_opto_output(TraceOptoOutput),
967 _in_dump_cnt(0),
968 _printer(NULL),
969 #endif
970 _dead_node_list(comp_arena()),
971 _dead_node_count(0),
972 _congraph(NULL),
973 _replay_inline_data(NULL),
974 _number_of_mh_late_inlines(0),
975 _inlining_progress(false),
976 _inlining_incrementally(false),
977 _print_inlining_list(NULL),
978 _print_inlining_stream(NULL),
979 _print_inlining_idx(0),
980 _preserve_jvm_state(0),
981 _allowed_reasons(0),
982 _interpreter_frame_size(0) {
983 C = this;
984
985 #ifndef PRODUCT
986 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
987 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
988 set_print_assembly(PrintFrameConverterAssembly);
989 set_parsed_irreducible_loop(false);
990 #endif
991 CompileWrapper cw(this);
992 Init(/*AliasLevel=*/ 0);
993 init_tf((*generator)());
994
995 {
996 // The following is a dummy for the sake of GraphKit::gen_stub
997 Unique_Node_List for_igvn(comp_arena());
998 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
999 PhaseGVN gvn(Thread::current()->resource_area(),255);
1000 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1001 gvn.transform_no_reclaim(top());
1002
1003 GraphKit kit;
1004 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1005 }
1006
1007 NOT_PRODUCT( verify_graph_edges(); )
1008 Code_Gen();
1009 if (failing()) return;
1010
1011
1012 // Entry point will be accessed using compile->stub_entry_point();
1013 if (code_buffer() == NULL) {
1014 Matcher::soft_match_failure();
1015 } else {
1016 if (PrintAssembly && (WizardMode || Verbose))
1017 tty->print_cr("### Stub::%s", stub_name);
1018
1019 if (!failing()) {
1020 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1021
1022 // Make the NMethod
1023 // For now we mark the frame as never safe for profile stackwalking
1024 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1025 code_buffer(),
1026 CodeOffsets::frame_never_safe,
1027 // _code_offsets.value(CodeOffsets::Frame_Complete),
1028 frame_size_in_words(),
1029 _oop_map_set,
1030 save_arg_registers);
1031 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1032
1033 _stub_entry_point = rs->entry_point();
1034 }
1035 }
1036 }
1037
1038 //------------------------------Init-------------------------------------------
1039 // Prepare for a single compilation
1040 void Compile::Init(int aliaslevel) {
1041 _unique = 0;
1042 _regalloc = NULL;
1043
1044 _tf = NULL; // filled in later
1045 _top = NULL; // cached later
1046 _matcher = NULL; // filled in later
1047 _cfg = NULL; // filled in later
1048
1049 set_24_bit_selection_and_mode(Use24BitFP, false);
1050
1051 _node_note_array = NULL;
1052 _default_node_notes = NULL;
1053
1054 _immutable_memory = NULL; // filled in at first inquiry
1055
1056 // Globally visible Nodes
1057 // First set TOP to NULL to give safe behavior during creation of RootNode
1058 set_cached_top_node(NULL);
1059 set_root(new (this) RootNode());
1060 // Now that you have a Root to point to, create the real TOP
1061 set_cached_top_node( new (this) ConNode(Type::TOP) );
1062 set_recent_alloc(NULL, NULL);
1063
1064 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1065 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1066 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1067 env()->set_dependencies(new Dependencies(env()));
1068
1069 _fixed_slots = 0;
1070 set_has_split_ifs(false);
1071 set_has_loops(has_method() && method()->has_loops()); // first approximation
1072 set_has_stringbuilder(false);
1073 set_has_boxed_value(false);
1074 _trap_can_recompile = false; // no traps emitted yet
1075 _major_progress = true; // start out assuming good things will happen
1076 set_has_unsafe_access(false);
1077 set_max_vector_size(0);
1078 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1079 set_decompile_count(0);
1080
1081 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1082 set_num_loop_opts(LoopOptsCount);
1083 set_do_inlining(Inline);
1084 set_max_inline_size(MaxInlineSize);
1085 set_freq_inline_size(FreqInlineSize);
1086 set_do_scheduling(OptoScheduling);
1087 set_do_count_invocations(false);
1088 set_do_method_data_update(false);
1089 set_rtm_state(NoRTM); // No RTM lock eliding by default
1090 #if INCLUDE_RTM_OPT
1091 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1092 int rtm_state = method()->method_data()->rtm_state();
1093 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1094 // Don't generate RTM lock eliding code.
1095 set_rtm_state(NoRTM);
1096 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1097 // Generate RTM lock eliding code without abort ratio calculation code.
1098 set_rtm_state(UseRTM);
1099 } else if (UseRTMDeopt) {
1100 // Generate RTM lock eliding code and include abort ratio calculation
1101 // code if UseRTMDeopt is on.
1102 set_rtm_state(ProfileRTM);
1103 }
1104 }
1105 #endif
1106 if (debug_info()->recording_non_safepoints()) {
1107 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1108 (comp_arena(), 8, 0, NULL));
1109 set_default_node_notes(Node_Notes::make(this));
1110 }
1111
1112 // // -- Initialize types before each compile --
1113 // // Update cached type information
1114 // if( _method && _method->constants() )
1115 // Type::update_loaded_types(_method, _method->constants());
1116
1117 // Init alias_type map.
1118 if (!_do_escape_analysis && aliaslevel == 3)
1119 aliaslevel = 2; // No unique types without escape analysis
1120 _AliasLevel = aliaslevel;
1121 const int grow_ats = 16;
1122 _max_alias_types = grow_ats;
1123 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1124 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1125 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1126 {
1127 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1128 }
1129 // Initialize the first few types.
1130 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1131 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1132 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1133 _num_alias_types = AliasIdxRaw+1;
1134 // Zero out the alias type cache.
1135 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1136 // A NULL adr_type hits in the cache right away. Preload the right answer.
1137 probe_alias_cache(NULL)->_index = AliasIdxTop;
1138
1139 _intrinsics = NULL;
1140 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1141 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1142 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1143 register_library_intrinsics();
1144 }
1145
1146 //---------------------------init_start----------------------------------------
1147 // Install the StartNode on this compile object.
1148 void Compile::init_start(StartNode* s) {
1149 if (failing())
1150 return; // already failing
1151 assert(s == start(), "");
1152 }
1153
1154 StartNode* Compile::start() const {
1155 assert(!failing(), "");
1156 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1157 Node* start = root()->fast_out(i);
1158 if( start->is_Start() )
1159 return start->as_Start();
1160 }
1161 ShouldNotReachHere();
1162 return NULL;
1163 }
1164
1165 //-------------------------------immutable_memory-------------------------------------
1166 // Access immutable memory
1167 Node* Compile::immutable_memory() {
1168 if (_immutable_memory != NULL) {
1169 return _immutable_memory;
1170 }
1171 StartNode* s = start();
1172 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1173 Node *p = s->fast_out(i);
1174 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1175 _immutable_memory = p;
1176 return _immutable_memory;
1177 }
1178 }
1179 ShouldNotReachHere();
1180 return NULL;
1181 }
1182
1183 //----------------------set_cached_top_node------------------------------------
1184 // Install the cached top node, and make sure Node::is_top works correctly.
1185 void Compile::set_cached_top_node(Node* tn) {
1186 if (tn != NULL) verify_top(tn);
1187 Node* old_top = _top;
1188 _top = tn;
1189 // Calling Node::setup_is_top allows the nodes the chance to adjust
1190 // their _out arrays.
1191 if (_top != NULL) _top->setup_is_top();
1192 if (old_top != NULL) old_top->setup_is_top();
1193 assert(_top == NULL || top()->is_top(), "");
1194 }
1195
1196 #ifdef ASSERT
1197 uint Compile::count_live_nodes_by_graph_walk() {
1198 Unique_Node_List useful(comp_arena());
1199 // Get useful node list by walking the graph.
1200 identify_useful_nodes(useful);
1201 return useful.size();
1202 }
1203
1204 void Compile::print_missing_nodes() {
1205
1206 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1207 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1208 return;
1209 }
1210
1211 // This is an expensive function. It is executed only when the user
1212 // specifies VerifyIdealNodeCount option or otherwise knows the
1213 // additional work that needs to be done to identify reachable nodes
1214 // by walking the flow graph and find the missing ones using
1215 // _dead_node_list.
1216
1217 Unique_Node_List useful(comp_arena());
1218 // Get useful node list by walking the graph.
1219 identify_useful_nodes(useful);
1220
1221 uint l_nodes = C->live_nodes();
1222 uint l_nodes_by_walk = useful.size();
1223
1224 if (l_nodes != l_nodes_by_walk) {
1225 if (_log != NULL) {
1226 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1227 _log->stamp();
1228 _log->end_head();
1229 }
1230 VectorSet& useful_member_set = useful.member_set();
1231 int last_idx = l_nodes_by_walk;
1232 for (int i = 0; i < last_idx; i++) {
1233 if (useful_member_set.test(i)) {
1234 if (_dead_node_list.test(i)) {
1235 if (_log != NULL) {
1236 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1237 }
1238 if (PrintIdealNodeCount) {
1239 // Print the log message to tty
1240 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1241 useful.at(i)->dump();
1242 }
1243 }
1244 }
1245 else if (! _dead_node_list.test(i)) {
1246 if (_log != NULL) {
1247 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1248 }
1249 if (PrintIdealNodeCount) {
1250 // Print the log message to tty
1251 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1252 }
1253 }
1254 }
1255 if (_log != NULL) {
1256 _log->tail("mismatched_nodes");
1257 }
1258 }
1259 }
1260 #endif
1261
1262 #ifndef PRODUCT
1263 void Compile::verify_top(Node* tn) const {
1264 if (tn != NULL) {
1265 assert(tn->is_Con(), "top node must be a constant");
1266 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1267 assert(tn->in(0) != NULL, "must have live top node");
1268 }
1269 }
1270 #endif
1271
1272
1273 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1274
1275 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1276 guarantee(arr != NULL, "");
1277 int num_blocks = arr->length();
1278 if (grow_by < num_blocks) grow_by = num_blocks;
1279 int num_notes = grow_by * _node_notes_block_size;
1280 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1281 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1282 while (num_notes > 0) {
1283 arr->append(notes);
1284 notes += _node_notes_block_size;
1285 num_notes -= _node_notes_block_size;
1286 }
1287 assert(num_notes == 0, "exact multiple, please");
1288 }
1289
1290 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1291 if (source == NULL || dest == NULL) return false;
1292
1293 if (dest->is_Con())
1294 return false; // Do not push debug info onto constants.
1295
1296 #ifdef ASSERT
1297 // Leave a bread crumb trail pointing to the original node:
1298 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1299 dest->set_debug_orig(source);
1300 }
1301 #endif
1302
1303 if (node_note_array() == NULL)
1304 return false; // Not collecting any notes now.
1305
1306 // This is a copy onto a pre-existing node, which may already have notes.
1307 // If both nodes have notes, do not overwrite any pre-existing notes.
1308 Node_Notes* source_notes = node_notes_at(source->_idx);
1309 if (source_notes == NULL || source_notes->is_clear()) return false;
1310 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1311 if (dest_notes == NULL || dest_notes->is_clear()) {
1312 return set_node_notes_at(dest->_idx, source_notes);
1313 }
1314
1315 Node_Notes merged_notes = (*source_notes);
1316 // The order of operations here ensures that dest notes will win...
1317 merged_notes.update_from(dest_notes);
1318 return set_node_notes_at(dest->_idx, &merged_notes);
1319 }
1320
1321
1322 //--------------------------allow_range_check_smearing-------------------------
1323 // Gating condition for coalescing similar range checks.
1324 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1325 // single covering check that is at least as strong as any of them.
1326 // If the optimization succeeds, the simplified (strengthened) range check
1327 // will always succeed. If it fails, we will deopt, and then give up
1328 // on the optimization.
1329 bool Compile::allow_range_check_smearing() const {
1330 // If this method has already thrown a range-check,
1331 // assume it was because we already tried range smearing
1332 // and it failed.
1333 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1334 return !already_trapped;
1335 }
1336
1337
1338 //------------------------------flatten_alias_type-----------------------------
1339 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1340 int offset = tj->offset();
1341 TypePtr::PTR ptr = tj->ptr();
1342
1343 // Known instance (scalarizable allocation) alias only with itself.
1344 bool is_known_inst = tj->isa_oopptr() != NULL &&
1345 tj->is_oopptr()->is_known_instance();
1346
1347 // Process weird unsafe references.
1348 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1349 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1350 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1351 tj = TypeOopPtr::BOTTOM;
1352 ptr = tj->ptr();
1353 offset = tj->offset();
1354 }
1355
1356 // Array pointers need some flattening
1357 const TypeAryPtr *ta = tj->isa_aryptr();
1358 if (ta && ta->is_stable()) {
1359 // Erase stability property for alias analysis.
1360 tj = ta = ta->cast_to_stable(false);
1361 }
1362 if( ta && is_known_inst ) {
1363 if ( offset != Type::OffsetBot &&
1364 offset > arrayOopDesc::length_offset_in_bytes() ) {
1365 offset = Type::OffsetBot; // Flatten constant access into array body only
1366 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1367 }
1368 } else if( ta && _AliasLevel >= 2 ) {
1369 // For arrays indexed by constant indices, we flatten the alias
1370 // space to include all of the array body. Only the header, klass
1371 // and array length can be accessed un-aliased.
1372 if( offset != Type::OffsetBot ) {
1373 if( ta->const_oop() ) { // MethodData* or Method*
1374 offset = Type::OffsetBot; // Flatten constant access into array body
1375 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1376 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1377 // range is OK as-is.
1378 tj = ta = TypeAryPtr::RANGE;
1379 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1380 tj = TypeInstPtr::KLASS; // all klass loads look alike
1381 ta = TypeAryPtr::RANGE; // generic ignored junk
1382 ptr = TypePtr::BotPTR;
1383 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1384 tj = TypeInstPtr::MARK;
1385 ta = TypeAryPtr::RANGE; // generic ignored junk
1386 ptr = TypePtr::BotPTR;
1387 } else { // Random constant offset into array body
1388 offset = Type::OffsetBot; // Flatten constant access into array body
1389 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1390 }
1391 }
1392 // Arrays of fixed size alias with arrays of unknown size.
1393 if (ta->size() != TypeInt::POS) {
1394 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1395 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1396 }
1397 // Arrays of known objects become arrays of unknown objects.
1398 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1399 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1400 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1401 }
1402 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1403 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1404 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1405 }
1406 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1407 // cannot be distinguished by bytecode alone.
1408 if (ta->elem() == TypeInt::BOOL) {
1409 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1410 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1411 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1412 }
1413 // During the 2nd round of IterGVN, NotNull castings are removed.
1414 // Make sure the Bottom and NotNull variants alias the same.
1415 // Also, make sure exact and non-exact variants alias the same.
1416 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1417 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1418 }
1419 }
1420
1421 // Oop pointers need some flattening
1422 const TypeInstPtr *to = tj->isa_instptr();
1423 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1424 ciInstanceKlass *k = to->klass()->as_instance_klass();
1425 if( ptr == TypePtr::Constant ) {
1426 if (to->klass() != ciEnv::current()->Class_klass() ||
1427 offset < k->size_helper() * wordSize) {
1428 // No constant oop pointers (such as Strings); they alias with
1429 // unknown strings.
1430 assert(!is_known_inst, "not scalarizable allocation");
1431 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1432 }
1433 } else if( is_known_inst ) {
1434 tj = to; // Keep NotNull and klass_is_exact for instance type
1435 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1436 // During the 2nd round of IterGVN, NotNull castings are removed.
1437 // Make sure the Bottom and NotNull variants alias the same.
1438 // Also, make sure exact and non-exact variants alias the same.
1439 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1440 }
1441 if (to->speculative() != NULL) {
1442 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1443 }
1444 // Canonicalize the holder of this field
1445 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1446 // First handle header references such as a LoadKlassNode, even if the
1447 // object's klass is unloaded at compile time (4965979).
1448 if (!is_known_inst) { // Do it only for non-instance types
1449 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1450 }
1451 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1452 // Static fields are in the space above the normal instance
1453 // fields in the java.lang.Class instance.
1454 if (to->klass() != ciEnv::current()->Class_klass()) {
1455 to = NULL;
1456 tj = TypeOopPtr::BOTTOM;
1457 offset = tj->offset();
1458 }
1459 } else {
1460 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1461 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1462 if( is_known_inst ) {
1463 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1464 } else {
1465 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1466 }
1467 }
1468 }
1469 }
1470
1471 // Klass pointers to object array klasses need some flattening
1472 const TypeKlassPtr *tk = tj->isa_klassptr();
1473 if( tk ) {
1474 // If we are referencing a field within a Klass, we need
1475 // to assume the worst case of an Object. Both exact and
1476 // inexact types must flatten to the same alias class so
1477 // use NotNull as the PTR.
1478 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1479
1480 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1481 TypeKlassPtr::OBJECT->klass(),
1482 offset);
1483 }
1484
1485 ciKlass* klass = tk->klass();
1486 if( klass->is_obj_array_klass() ) {
1487 ciKlass* k = TypeAryPtr::OOPS->klass();
1488 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1489 k = TypeInstPtr::BOTTOM->klass();
1490 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1491 }
1492
1493 // Check for precise loads from the primary supertype array and force them
1494 // to the supertype cache alias index. Check for generic array loads from
1495 // the primary supertype array and also force them to the supertype cache
1496 // alias index. Since the same load can reach both, we need to merge
1497 // these 2 disparate memories into the same alias class. Since the
1498 // primary supertype array is read-only, there's no chance of confusion
1499 // where we bypass an array load and an array store.
1500 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1501 if (offset == Type::OffsetBot ||
1502 (offset >= primary_supers_offset &&
1503 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1504 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1505 offset = in_bytes(Klass::secondary_super_cache_offset());
1506 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1507 }
1508 }
1509
1510 // Flatten all Raw pointers together.
1511 if (tj->base() == Type::RawPtr)
1512 tj = TypeRawPtr::BOTTOM;
1513
1514 if (tj->base() == Type::AnyPtr)
1515 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1516
1517 // Flatten all to bottom for now
1518 switch( _AliasLevel ) {
1519 case 0:
1520 tj = TypePtr::BOTTOM;
1521 break;
1522 case 1: // Flatten to: oop, static, field or array
1523 switch (tj->base()) {
1524 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1525 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1526 case Type::AryPtr: // do not distinguish arrays at all
1527 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1528 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1529 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1530 default: ShouldNotReachHere();
1531 }
1532 break;
1533 case 2: // No collapsing at level 2; keep all splits
1534 case 3: // No collapsing at level 3; keep all splits
1535 break;
1536 default:
1537 Unimplemented();
1538 }
1539
1540 offset = tj->offset();
1541 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1542
1543 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1544 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1545 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1546 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1547 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1548 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1549 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1550 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1551 assert( tj->ptr() != TypePtr::TopPTR &&
1552 tj->ptr() != TypePtr::AnyNull &&
1553 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1554 // assert( tj->ptr() != TypePtr::Constant ||
1555 // tj->base() == Type::RawPtr ||
1556 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1557
1558 return tj;
1559 }
1560
1561 void Compile::AliasType::Init(int i, const TypePtr* at) {
1562 _index = i;
1563 _adr_type = at;
1564 _field = NULL;
1565 _element = NULL;
1566 _is_rewritable = true; // default
1567 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1568 if (atoop != NULL && atoop->is_known_instance()) {
1569 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1570 _general_index = Compile::current()->get_alias_index(gt);
1571 } else {
1572 _general_index = 0;
1573 }
1574 }
1575
1576 //---------------------------------print_on------------------------------------
1577 #ifndef PRODUCT
1578 void Compile::AliasType::print_on(outputStream* st) {
1579 if (index() < 10)
1580 st->print("@ <%d> ", index());
1581 else st->print("@ <%d>", index());
1582 st->print(is_rewritable() ? " " : " RO");
1583 int offset = adr_type()->offset();
1584 if (offset == Type::OffsetBot)
1585 st->print(" +any");
1586 else st->print(" +%-3d", offset);
1587 st->print(" in ");
1588 adr_type()->dump_on(st);
1589 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1590 if (field() != NULL && tjp) {
1591 if (tjp->klass() != field()->holder() ||
1592 tjp->offset() != field()->offset_in_bytes()) {
1593 st->print(" != ");
1594 field()->print();
1595 st->print(" ***");
1596 }
1597 }
1598 }
1599
1600 void print_alias_types() {
1601 Compile* C = Compile::current();
1602 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1603 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1604 C->alias_type(idx)->print_on(tty);
1605 tty->cr();
1606 }
1607 }
1608 #endif
1609
1610
1611 //----------------------------probe_alias_cache--------------------------------
1612 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1613 intptr_t key = (intptr_t) adr_type;
1614 key ^= key >> logAliasCacheSize;
1615 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1616 }
1617
1618
1619 //-----------------------------grow_alias_types--------------------------------
1620 void Compile::grow_alias_types() {
1621 const int old_ats = _max_alias_types; // how many before?
1622 const int new_ats = old_ats; // how many more?
1623 const int grow_ats = old_ats+new_ats; // how many now?
1624 _max_alias_types = grow_ats;
1625 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1626 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1627 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1628 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1629 }
1630
1631
1632 //--------------------------------find_alias_type------------------------------
1633 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1634 if (_AliasLevel == 0)
1635 return alias_type(AliasIdxBot);
1636
1637 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1638 if (ace->_adr_type == adr_type) {
1639 return alias_type(ace->_index);
1640 }
1641
1642 // Handle special cases.
1643 if (adr_type == NULL) return alias_type(AliasIdxTop);
1644 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1645
1646 // Do it the slow way.
1647 const TypePtr* flat = flatten_alias_type(adr_type);
1648
1649 #ifdef ASSERT
1650 assert(flat == flatten_alias_type(flat), "idempotent");
1651 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1652 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1653 const TypeOopPtr* foop = flat->is_oopptr();
1654 // Scalarizable allocations have exact klass always.
1655 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1656 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1657 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1658 }
1659 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1660 #endif
1661
1662 int idx = AliasIdxTop;
1663 for (int i = 0; i < num_alias_types(); i++) {
1664 if (alias_type(i)->adr_type() == flat) {
1665 idx = i;
1666 break;
1667 }
1668 }
1669
1670 if (idx == AliasIdxTop) {
1671 if (no_create) return NULL;
1672 // Grow the array if necessary.
1673 if (_num_alias_types == _max_alias_types) grow_alias_types();
1674 // Add a new alias type.
1675 idx = _num_alias_types++;
1676 _alias_types[idx]->Init(idx, flat);
1677 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1678 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1679 if (flat->isa_instptr()) {
1680 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1681 && flat->is_instptr()->klass() == env()->Class_klass())
1682 alias_type(idx)->set_rewritable(false);
1683 }
1684 if (flat->isa_aryptr()) {
1685 #ifdef ASSERT
1686 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1687 // (T_BYTE has the weakest alignment and size restrictions...)
1688 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1689 #endif
1690 if (flat->offset() == TypePtr::OffsetBot) {
1691 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1692 }
1693 }
1694 if (flat->isa_klassptr()) {
1695 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1696 alias_type(idx)->set_rewritable(false);
1697 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1698 alias_type(idx)->set_rewritable(false);
1699 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1700 alias_type(idx)->set_rewritable(false);
1701 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1702 alias_type(idx)->set_rewritable(false);
1703 }
1704 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1705 // but the base pointer type is not distinctive enough to identify
1706 // references into JavaThread.)
1707
1708 // Check for final fields.
1709 const TypeInstPtr* tinst = flat->isa_instptr();
1710 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1711 ciField* field;
1712 if (tinst->const_oop() != NULL &&
1713 tinst->klass() == ciEnv::current()->Class_klass() &&
1714 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1715 // static field
1716 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1717 field = k->get_field_by_offset(tinst->offset(), true);
1718 } else {
1719 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1720 field = k->get_field_by_offset(tinst->offset(), false);
1721 }
1722 assert(field == NULL ||
1723 original_field == NULL ||
1724 (field->holder() == original_field->holder() &&
1725 field->offset() == original_field->offset() &&
1726 field->is_static() == original_field->is_static()), "wrong field?");
1727 // Set field() and is_rewritable() attributes.
1728 if (field != NULL) alias_type(idx)->set_field(field);
1729 }
1730 }
1731
1732 // Fill the cache for next time.
1733 ace->_adr_type = adr_type;
1734 ace->_index = idx;
1735 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1736
1737 // Might as well try to fill the cache for the flattened version, too.
1738 AliasCacheEntry* face = probe_alias_cache(flat);
1739 if (face->_adr_type == NULL) {
1740 face->_adr_type = flat;
1741 face->_index = idx;
1742 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1743 }
1744
1745 return alias_type(idx);
1746 }
1747
1748
1749 Compile::AliasType* Compile::alias_type(ciField* field) {
1750 const TypeOopPtr* t;
1751 if (field->is_static())
1752 t = TypeInstPtr::make(field->holder()->java_mirror());
1753 else
1754 t = TypeOopPtr::make_from_klass_raw(field->holder());
1755 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1756 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1757 return atp;
1758 }
1759
1760
1761 //------------------------------have_alias_type--------------------------------
1762 bool Compile::have_alias_type(const TypePtr* adr_type) {
1763 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1764 if (ace->_adr_type == adr_type) {
1765 return true;
1766 }
1767
1768 // Handle special cases.
1769 if (adr_type == NULL) return true;
1770 if (adr_type == TypePtr::BOTTOM) return true;
1771
1772 return find_alias_type(adr_type, true, NULL) != NULL;
1773 }
1774
1775 //-----------------------------must_alias--------------------------------------
1776 // True if all values of the given address type are in the given alias category.
1777 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1778 if (alias_idx == AliasIdxBot) return true; // the universal category
1779 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1780 if (alias_idx == AliasIdxTop) return false; // the empty category
1781 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1782
1783 // the only remaining possible overlap is identity
1784 int adr_idx = get_alias_index(adr_type);
1785 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1786 assert(adr_idx == alias_idx ||
1787 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1788 && adr_type != TypeOopPtr::BOTTOM),
1789 "should not be testing for overlap with an unsafe pointer");
1790 return adr_idx == alias_idx;
1791 }
1792
1793 //------------------------------can_alias--------------------------------------
1794 // True if any values of the given address type are in the given alias category.
1795 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1796 if (alias_idx == AliasIdxTop) return false; // the empty category
1797 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1798 if (alias_idx == AliasIdxBot) return true; // the universal category
1799 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1800
1801 // the only remaining possible overlap is identity
1802 int adr_idx = get_alias_index(adr_type);
1803 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1804 return adr_idx == alias_idx;
1805 }
1806
1807
1808
1809 //---------------------------pop_warm_call-------------------------------------
1810 WarmCallInfo* Compile::pop_warm_call() {
1811 WarmCallInfo* wci = _warm_calls;
1812 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1813 return wci;
1814 }
1815
1816 //----------------------------Inline_Warm--------------------------------------
1817 int Compile::Inline_Warm() {
1818 // If there is room, try to inline some more warm call sites.
1819 // %%% Do a graph index compaction pass when we think we're out of space?
1820 if (!InlineWarmCalls) return 0;
1821
1822 int calls_made_hot = 0;
1823 int room_to_grow = NodeCountInliningCutoff - unique();
1824 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1825 int amount_grown = 0;
1826 WarmCallInfo* call;
1827 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1828 int est_size = (int)call->size();
1829 if (est_size > (room_to_grow - amount_grown)) {
1830 // This one won't fit anyway. Get rid of it.
1831 call->make_cold();
1832 continue;
1833 }
1834 call->make_hot();
1835 calls_made_hot++;
1836 amount_grown += est_size;
1837 amount_to_grow -= est_size;
1838 }
1839
1840 if (calls_made_hot > 0) set_major_progress();
1841 return calls_made_hot;
1842 }
1843
1844
1845 //----------------------------Finish_Warm--------------------------------------
1846 void Compile::Finish_Warm() {
1847 if (!InlineWarmCalls) return;
1848 if (failing()) return;
1849 if (warm_calls() == NULL) return;
1850
1851 // Clean up loose ends, if we are out of space for inlining.
1852 WarmCallInfo* call;
1853 while ((call = pop_warm_call()) != NULL) {
1854 call->make_cold();
1855 }
1856 }
1857
1858 //---------------------cleanup_loop_predicates-----------------------
1859 // Remove the opaque nodes that protect the predicates so that all unused
1860 // checks and uncommon_traps will be eliminated from the ideal graph
1861 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1862 if (predicate_count()==0) return;
1863 for (int i = predicate_count(); i > 0; i--) {
1864 Node * n = predicate_opaque1_node(i-1);
1865 assert(n->Opcode() == Op_Opaque1, "must be");
1866 igvn.replace_node(n, n->in(1));
1867 }
1868 assert(predicate_count()==0, "should be clean!");
1869 }
1870
1871 // StringOpts and late inlining of string methods
1872 void Compile::inline_string_calls(bool parse_time) {
1873 {
1874 // remove useless nodes to make the usage analysis simpler
1875 ResourceMark rm;
1876 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1877 }
1878
1879 {
1880 ResourceMark rm;
1881 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1882 PhaseStringOpts pso(initial_gvn(), for_igvn());
1883 print_method(PHASE_AFTER_STRINGOPTS, 3);
1884 }
1885
1886 // now inline anything that we skipped the first time around
1887 if (!parse_time) {
1888 _late_inlines_pos = _late_inlines.length();
1889 }
1890
1891 while (_string_late_inlines.length() > 0) {
1892 CallGenerator* cg = _string_late_inlines.pop();
1893 cg->do_late_inline();
1894 if (failing()) return;
1895 }
1896 _string_late_inlines.trunc_to(0);
1897 }
1898
1899 // Late inlining of boxing methods
1900 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1901 if (_boxing_late_inlines.length() > 0) {
1902 assert(has_boxed_value(), "inconsistent");
1903
1904 PhaseGVN* gvn = initial_gvn();
1905 set_inlining_incrementally(true);
1906
1907 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1908 for_igvn()->clear();
1909 gvn->replace_with(&igvn);
1910
1911 while (_boxing_late_inlines.length() > 0) {
1912 CallGenerator* cg = _boxing_late_inlines.pop();
1913 cg->do_late_inline();
1914 if (failing()) return;
1915 }
1916 _boxing_late_inlines.trunc_to(0);
1917
1918 {
1919 ResourceMark rm;
1920 PhaseRemoveUseless pru(gvn, for_igvn());
1921 }
1922
1923 igvn = PhaseIterGVN(gvn);
1924 igvn.optimize();
1925
1926 set_inlining_progress(false);
1927 set_inlining_incrementally(false);
1928 }
1929 }
1930
1931 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1932 assert(IncrementalInline, "incremental inlining should be on");
1933 PhaseGVN* gvn = initial_gvn();
1934
1935 set_inlining_progress(false);
1936 for_igvn()->clear();
1937 gvn->replace_with(&igvn);
1938
1939 int i = 0;
1940
1941 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1942 CallGenerator* cg = _late_inlines.at(i);
1943 _late_inlines_pos = i+1;
1944 cg->do_late_inline();
1945 if (failing()) return;
1946 }
1947 int j = 0;
1948 for (; i < _late_inlines.length(); i++, j++) {
1949 _late_inlines.at_put(j, _late_inlines.at(i));
1950 }
1951 _late_inlines.trunc_to(j);
1952
1953 {
1954 ResourceMark rm;
1955 PhaseRemoveUseless pru(gvn, for_igvn());
1956 }
1957
1958 igvn = PhaseIterGVN(gvn);
1959 }
1960
1961 // Perform incremental inlining until bound on number of live nodes is reached
1962 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1963 PhaseGVN* gvn = initial_gvn();
1964
1965 set_inlining_incrementally(true);
1966 set_inlining_progress(true);
1967 uint low_live_nodes = 0;
1968
1969 while(inlining_progress() && _late_inlines.length() > 0) {
1970
1971 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1972 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1973 // PhaseIdealLoop is expensive so we only try it once we are
1974 // out of loop and we only try it again if the previous helped
1975 // got the number of nodes down significantly
1976 PhaseIdealLoop ideal_loop( igvn, false, true );
1977 if (failing()) return;
1978 low_live_nodes = live_nodes();
1979 _major_progress = true;
1980 }
1981
1982 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1983 break;
1984 }
1985 }
1986
1987 inline_incrementally_one(igvn);
1988
1989 if (failing()) return;
1990
1991 igvn.optimize();
1992
1993 if (failing()) return;
1994 }
1995
1996 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1997
1998 if (_string_late_inlines.length() > 0) {
1999 assert(has_stringbuilder(), "inconsistent");
2000 for_igvn()->clear();
2001 initial_gvn()->replace_with(&igvn);
2002
2003 inline_string_calls(false);
2004
2005 if (failing()) return;
2006
2007 {
2008 ResourceMark rm;
2009 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2010 }
2011
2012 igvn = PhaseIterGVN(gvn);
2013
2014 igvn.optimize();
2015 }
2016
2017 set_inlining_incrementally(false);
2018 }
2019
2020
2021 //------------------------------Optimize---------------------------------------
2022 // Given a graph, optimize it.
2023 void Compile::Optimize() {
2024 TracePhase t1("optimizer", &_t_optimizer, true);
2025
2026 #ifndef PRODUCT
2027 if (env()->break_at_compile()) {
2028 BREAKPOINT;
2029 }
2030
2031 #endif
2032
2033 ResourceMark rm;
2034 int loop_opts_cnt;
2035
2036 print_inlining_reinit();
2037
2038 NOT_PRODUCT( verify_graph_edges(); )
2039
2040 print_method(PHASE_AFTER_PARSING);
2041
2042 {
2043 // Iterative Global Value Numbering, including ideal transforms
2044 // Initialize IterGVN with types and values from parse-time GVN
2045 PhaseIterGVN igvn(initial_gvn());
2046 {
2047 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2048 igvn.optimize();
2049 }
2050
2051 print_method(PHASE_ITER_GVN1, 2);
2052
2053 if (failing()) return;
2054
2055 {
2056 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2057 inline_incrementally(igvn);
2058 }
2059
2060 print_method(PHASE_INCREMENTAL_INLINE, 2);
2061
2062 if (failing()) return;
2063
2064 if (eliminate_boxing()) {
2065 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2066 // Inline valueOf() methods now.
2067 inline_boxing_calls(igvn);
2068
2069 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2070
2071 if (failing()) return;
2072 }
2073
2074 // Remove the speculative part of types and clean up the graph from
2075 // the extra CastPP nodes whose only purpose is to carry them. Do
2076 // that early so that optimizations are not disrupted by the extra
2077 // CastPP nodes.
2078 remove_speculative_types(igvn);
2079
2080 // No more new expensive nodes will be added to the list from here
2081 // so keep only the actual candidates for optimizations.
2082 cleanup_expensive_nodes(igvn);
2083
2084 // Perform escape analysis
2085 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2086 if (has_loops()) {
2087 // Cleanup graph (remove dead nodes).
2088 TracePhase t2("idealLoop", &_t_idealLoop, true);
2089 PhaseIdealLoop ideal_loop( igvn, false, true );
2090 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2091 if (failing()) return;
2092 }
2093 ConnectionGraph::do_analysis(this, &igvn);
2094
2095 if (failing()) return;
2096
2097 // Optimize out fields loads from scalar replaceable allocations.
2098 igvn.optimize();
2099 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2100
2101 if (failing()) return;
2102
2103 if (congraph() != NULL && macro_count() > 0) {
2104 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2105 PhaseMacroExpand mexp(igvn);
2106 mexp.eliminate_macro_nodes();
2107 igvn.set_delay_transform(false);
2108
2109 igvn.optimize();
2110 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2111
2112 if (failing()) return;
2113 }
2114 }
2115
2116 // Loop transforms on the ideal graph. Range Check Elimination,
2117 // peeling, unrolling, etc.
2118
2119 // Set loop opts counter
2120 loop_opts_cnt = num_loop_opts();
2121 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2122 {
2123 TracePhase t2("idealLoop", &_t_idealLoop, true);
2124 PhaseIdealLoop ideal_loop( igvn, true );
2125 loop_opts_cnt--;
2126 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2127 if (failing()) return;
2128 }
2129 // Loop opts pass if partial peeling occurred in previous pass
2130 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2131 TracePhase t3("idealLoop", &_t_idealLoop, true);
2132 PhaseIdealLoop ideal_loop( igvn, false );
2133 loop_opts_cnt--;
2134 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2135 if (failing()) return;
2136 }
2137 // Loop opts pass for loop-unrolling before CCP
2138 if(major_progress() && (loop_opts_cnt > 0)) {
2139 TracePhase t4("idealLoop", &_t_idealLoop, true);
2140 PhaseIdealLoop ideal_loop( igvn, false );
2141 loop_opts_cnt--;
2142 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2143 }
2144 if (!failing()) {
2145 // Verify that last round of loop opts produced a valid graph
2146 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2147 PhaseIdealLoop::verify(igvn);
2148 }
2149 }
2150 if (failing()) return;
2151
2152 // Conditional Constant Propagation;
2153 PhaseCCP ccp( &igvn );
2154 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2155 {
2156 TracePhase t2("ccp", &_t_ccp, true);
2157 ccp.do_transform();
2158 }
2159 print_method(PHASE_CPP1, 2);
2160
2161 assert( true, "Break here to ccp.dump_old2new_map()");
2162
2163 // Iterative Global Value Numbering, including ideal transforms
2164 {
2165 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2166 igvn = ccp;
2167 igvn.optimize();
2168 }
2169
2170 print_method(PHASE_ITER_GVN2, 2);
2171
2172 if (failing()) return;
2173
2174 // Loop transforms on the ideal graph. Range Check Elimination,
2175 // peeling, unrolling, etc.
2176 if(loop_opts_cnt > 0) {
2177 debug_only( int cnt = 0; );
2178 while(major_progress() && (loop_opts_cnt > 0)) {
2179 TracePhase t2("idealLoop", &_t_idealLoop, true);
2180 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2181 PhaseIdealLoop ideal_loop( igvn, true);
2182 loop_opts_cnt--;
2183 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2184 if (failing()) return;
2185 }
2186 }
2187
2188 {
2189 // Verify that all previous optimizations produced a valid graph
2190 // at least to this point, even if no loop optimizations were done.
2191 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2192 PhaseIdealLoop::verify(igvn);
2193 }
2194
2195 {
2196 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2197 PhaseMacroExpand mex(igvn);
2198 if (mex.expand_macro_nodes()) {
2199 assert(failing(), "must bail out w/ explicit message");
2200 return;
2201 }
2202 }
2203
2204 } // (End scope of igvn; run destructor if necessary for asserts.)
2205
2206 dump_inlining();
2207 // A method with only infinite loops has no edges entering loops from root
2208 {
2209 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2210 if (final_graph_reshaping()) {
2211 assert(failing(), "must bail out w/ explicit message");
2212 return;
2213 }
2214 }
2215
2216 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2217 }
2218
2219
2220 //------------------------------Code_Gen---------------------------------------
2221 // Given a graph, generate code for it
2222 void Compile::Code_Gen() {
2223 if (failing()) {
2224 return;
2225 }
2226
2227 // Perform instruction selection. You might think we could reclaim Matcher
2228 // memory PDQ, but actually the Matcher is used in generating spill code.
2229 // Internals of the Matcher (including some VectorSets) must remain live
2230 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2231 // set a bit in reclaimed memory.
2232
2233 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2234 // nodes. Mapping is only valid at the root of each matched subtree.
2235 NOT_PRODUCT( verify_graph_edges(); )
2236
2237 Matcher matcher;
2238 _matcher = &matcher;
2239 {
2240 TracePhase t2("matcher", &_t_matcher, true);
2241 matcher.match();
2242 }
2243 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2244 // nodes. Mapping is only valid at the root of each matched subtree.
2245 NOT_PRODUCT( verify_graph_edges(); )
2246
2247 // If you have too many nodes, or if matching has failed, bail out
2248 check_node_count(0, "out of nodes matching instructions");
2249 if (failing()) {
2250 return;
2251 }
2252
2253 // Build a proper-looking CFG
2254 PhaseCFG cfg(node_arena(), root(), matcher);
2255 _cfg = &cfg;
2256 {
2257 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2258 bool success = cfg.do_global_code_motion();
2259 if (!success) {
2260 return;
2261 }
2262
2263 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2264 NOT_PRODUCT( verify_graph_edges(); )
2265 debug_only( cfg.verify(); )
2266 }
2267
2268 PhaseChaitin regalloc(unique(), cfg, matcher);
2269 _regalloc = ®alloc;
2270 {
2271 TracePhase t2("regalloc", &_t_registerAllocation, true);
2272 // Perform register allocation. After Chaitin, use-def chains are
2273 // no longer accurate (at spill code) and so must be ignored.
2274 // Node->LRG->reg mappings are still accurate.
2275 _regalloc->Register_Allocate();
2276
2277 // Bail out if the allocator builds too many nodes
2278 if (failing()) {
2279 return;
2280 }
2281 }
2282
2283 // Prior to register allocation we kept empty basic blocks in case the
2284 // the allocator needed a place to spill. After register allocation we
2285 // are not adding any new instructions. If any basic block is empty, we
2286 // can now safely remove it.
2287 {
2288 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2289 cfg.remove_empty_blocks();
2290 if (do_freq_based_layout()) {
2291 PhaseBlockLayout layout(cfg);
2292 } else {
2293 cfg.set_loop_alignment();
2294 }
2295 cfg.fixup_flow();
2296 }
2297
2298 // Apply peephole optimizations
2299 if( OptoPeephole ) {
2300 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2301 PhasePeephole peep( _regalloc, cfg);
2302 peep.do_transform();
2303 }
2304
2305 // Do late expand if CPU requires this.
2306 if (Matcher::require_postalloc_expand) {
2307 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2308 cfg.postalloc_expand(_regalloc);
2309 }
2310
2311 // Convert Nodes to instruction bits in a buffer
2312 {
2313 // %%%% workspace merge brought two timers together for one job
2314 TracePhase t2a("output", &_t_output, true);
2315 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2316 Output();
2317 }
2318
2319 print_method(PHASE_FINAL_CODE);
2320
2321 // He's dead, Jim.
2322 _cfg = (PhaseCFG*)0xdeadbeef;
2323 _regalloc = (PhaseChaitin*)0xdeadbeef;
2324 }
2325
2326
2327 //------------------------------dump_asm---------------------------------------
2328 // Dump formatted assembly
2329 #ifndef PRODUCT
2330 void Compile::dump_asm(int *pcs, uint pc_limit) {
2331 bool cut_short = false;
2332 tty->print_cr("#");
2333 tty->print("# "); _tf->dump(); tty->cr();
2334 tty->print_cr("#");
2335
2336 // For all blocks
2337 int pc = 0x0; // Program counter
2338 char starts_bundle = ' ';
2339 _regalloc->dump_frame();
2340
2341 Node *n = NULL;
2342 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2343 if (VMThread::should_terminate()) {
2344 cut_short = true;
2345 break;
2346 }
2347 Block* block = _cfg->get_block(i);
2348 if (block->is_connector() && !Verbose) {
2349 continue;
2350 }
2351 n = block->head();
2352 if (pcs && n->_idx < pc_limit) {
2353 tty->print("%3.3x ", pcs[n->_idx]);
2354 } else {
2355 tty->print(" ");
2356 }
2357 block->dump_head(_cfg);
2358 if (block->is_connector()) {
2359 tty->print_cr(" # Empty connector block");
2360 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2361 tty->print_cr(" # Block is sole successor of call");
2362 }
2363
2364 // For all instructions
2365 Node *delay = NULL;
2366 for (uint j = 0; j < block->number_of_nodes(); j++) {
2367 if (VMThread::should_terminate()) {
2368 cut_short = true;
2369 break;
2370 }
2371 n = block->get_node(j);
2372 if (valid_bundle_info(n)) {
2373 Bundle* bundle = node_bundling(n);
2374 if (bundle->used_in_unconditional_delay()) {
2375 delay = n;
2376 continue;
2377 }
2378 if (bundle->starts_bundle()) {
2379 starts_bundle = '+';
2380 }
2381 }
2382
2383 if (WizardMode) {
2384 n->dump();
2385 }
2386
2387 if( !n->is_Region() && // Dont print in the Assembly
2388 !n->is_Phi() && // a few noisely useless nodes
2389 !n->is_Proj() &&
2390 !n->is_MachTemp() &&
2391 !n->is_SafePointScalarObject() &&
2392 !n->is_Catch() && // Would be nice to print exception table targets
2393 !n->is_MergeMem() && // Not very interesting
2394 !n->is_top() && // Debug info table constants
2395 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2396 ) {
2397 if (pcs && n->_idx < pc_limit)
2398 tty->print("%3.3x", pcs[n->_idx]);
2399 else
2400 tty->print(" ");
2401 tty->print(" %c ", starts_bundle);
2402 starts_bundle = ' ';
2403 tty->print("\t");
2404 n->format(_regalloc, tty);
2405 tty->cr();
2406 }
2407
2408 // If we have an instruction with a delay slot, and have seen a delay,
2409 // then back up and print it
2410 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2411 assert(delay != NULL, "no unconditional delay instruction");
2412 if (WizardMode) delay->dump();
2413
2414 if (node_bundling(delay)->starts_bundle())
2415 starts_bundle = '+';
2416 if (pcs && n->_idx < pc_limit)
2417 tty->print("%3.3x", pcs[n->_idx]);
2418 else
2419 tty->print(" ");
2420 tty->print(" %c ", starts_bundle);
2421 starts_bundle = ' ';
2422 tty->print("\t");
2423 delay->format(_regalloc, tty);
2424 tty->print_cr("");
2425 delay = NULL;
2426 }
2427
2428 // Dump the exception table as well
2429 if( n->is_Catch() && (Verbose || WizardMode) ) {
2430 // Print the exception table for this offset
2431 _handler_table.print_subtable_for(pc);
2432 }
2433 }
2434
2435 if (pcs && n->_idx < pc_limit)
2436 tty->print_cr("%3.3x", pcs[n->_idx]);
2437 else
2438 tty->print_cr("");
2439
2440 assert(cut_short || delay == NULL, "no unconditional delay branch");
2441
2442 } // End of per-block dump
2443 tty->print_cr("");
2444
2445 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2446 }
2447 #endif
2448
2449 //------------------------------Final_Reshape_Counts---------------------------
2450 // This class defines counters to help identify when a method
2451 // may/must be executed using hardware with only 24-bit precision.
2452 struct Final_Reshape_Counts : public StackObj {
2453 int _call_count; // count non-inlined 'common' calls
2454 int _float_count; // count float ops requiring 24-bit precision
2455 int _double_count; // count double ops requiring more precision
2456 int _java_call_count; // count non-inlined 'java' calls
2457 int _inner_loop_count; // count loops which need alignment
2458 VectorSet _visited; // Visitation flags
2459 Node_List _tests; // Set of IfNodes & PCTableNodes
2460
2461 Final_Reshape_Counts() :
2462 _call_count(0), _float_count(0), _double_count(0),
2463 _java_call_count(0), _inner_loop_count(0),
2464 _visited( Thread::current()->resource_area() ) { }
2465
2466 void inc_call_count () { _call_count ++; }
2467 void inc_float_count () { _float_count ++; }
2468 void inc_double_count() { _double_count++; }
2469 void inc_java_call_count() { _java_call_count++; }
2470 void inc_inner_loop_count() { _inner_loop_count++; }
2471
2472 int get_call_count () const { return _call_count ; }
2473 int get_float_count () const { return _float_count ; }
2474 int get_double_count() const { return _double_count; }
2475 int get_java_call_count() const { return _java_call_count; }
2476 int get_inner_loop_count() const { return _inner_loop_count; }
2477 };
2478
2479 #ifdef ASSERT
2480 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2481 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2482 // Make sure the offset goes inside the instance layout.
2483 return k->contains_field_offset(tp->offset());
2484 // Note that OffsetBot and OffsetTop are very negative.
2485 }
2486 #endif
2487
2488 // Eliminate trivially redundant StoreCMs and accumulate their
2489 // precedence edges.
2490 void Compile::eliminate_redundant_card_marks(Node* n) {
2491 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2492 if (n->in(MemNode::Address)->outcnt() > 1) {
2493 // There are multiple users of the same address so it might be
2494 // possible to eliminate some of the StoreCMs
2495 Node* mem = n->in(MemNode::Memory);
2496 Node* adr = n->in(MemNode::Address);
2497 Node* val = n->in(MemNode::ValueIn);
2498 Node* prev = n;
2499 bool done = false;
2500 // Walk the chain of StoreCMs eliminating ones that match. As
2501 // long as it's a chain of single users then the optimization is
2502 // safe. Eliminating partially redundant StoreCMs would require
2503 // cloning copies down the other paths.
2504 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2505 if (adr == mem->in(MemNode::Address) &&
2506 val == mem->in(MemNode::ValueIn)) {
2507 // redundant StoreCM
2508 if (mem->req() > MemNode::OopStore) {
2509 // Hasn't been processed by this code yet.
2510 n->add_prec(mem->in(MemNode::OopStore));
2511 } else {
2512 // Already converted to precedence edge
2513 for (uint i = mem->req(); i < mem->len(); i++) {
2514 // Accumulate any precedence edges
2515 if (mem->in(i) != NULL) {
2516 n->add_prec(mem->in(i));
2517 }
2518 }
2519 // Everything above this point has been processed.
2520 done = true;
2521 }
2522 // Eliminate the previous StoreCM
2523 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2524 assert(mem->outcnt() == 0, "should be dead");
2525 mem->disconnect_inputs(NULL, this);
2526 } else {
2527 prev = mem;
2528 }
2529 mem = prev->in(MemNode::Memory);
2530 }
2531 }
2532 }
2533
2534 //------------------------------final_graph_reshaping_impl----------------------
2535 // Implement items 1-5 from final_graph_reshaping below.
2536 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2537
2538 if ( n->outcnt() == 0 ) return; // dead node
2539 uint nop = n->Opcode();
2540
2541 // Check for 2-input instruction with "last use" on right input.
2542 // Swap to left input. Implements item (2).
2543 if( n->req() == 3 && // two-input instruction
2544 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2545 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2546 n->in(2)->outcnt() == 1 &&// right use IS a last use
2547 !n->in(2)->is_Con() ) { // right use is not a constant
2548 // Check for commutative opcode
2549 switch( nop ) {
2550 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2551 case Op_MaxI: case Op_MinI:
2552 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2553 case Op_AndL: case Op_XorL: case Op_OrL:
2554 case Op_AndI: case Op_XorI: case Op_OrI: {
2555 // Move "last use" input to left by swapping inputs
2556 n->swap_edges(1, 2);
2557 break;
2558 }
2559 default:
2560 break;
2561 }
2562 }
2563
2564 #ifdef ASSERT
2565 if( n->is_Mem() ) {
2566 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2567 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2568 // oop will be recorded in oop map if load crosses safepoint
2569 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2570 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2571 "raw memory operations should have control edge");
2572 }
2573 #endif
2574 // Count FPU ops and common calls, implements item (3)
2575 switch( nop ) {
2576 // Count all float operations that may use FPU
2577 case Op_AddF:
2578 case Op_SubF:
2579 case Op_MulF:
2580 case Op_DivF:
2581 case Op_NegF:
2582 case Op_ModF:
2583 case Op_ConvI2F:
2584 case Op_ConF:
2585 case Op_CmpF:
2586 case Op_CmpF3:
2587 // case Op_ConvL2F: // longs are split into 32-bit halves
2588 frc.inc_float_count();
2589 break;
2590
2591 case Op_ConvF2D:
2592 case Op_ConvD2F:
2593 frc.inc_float_count();
2594 frc.inc_double_count();
2595 break;
2596
2597 // Count all double operations that may use FPU
2598 case Op_AddD:
2599 case Op_SubD:
2600 case Op_MulD:
2601 case Op_DivD:
2602 case Op_NegD:
2603 case Op_ModD:
2604 case Op_ConvI2D:
2605 case Op_ConvD2I:
2606 // case Op_ConvL2D: // handled by leaf call
2607 // case Op_ConvD2L: // handled by leaf call
2608 case Op_ConD:
2609 case Op_CmpD:
2610 case Op_CmpD3:
2611 frc.inc_double_count();
2612 break;
2613 case Op_Opaque1: // Remove Opaque Nodes before matching
2614 case Op_Opaque2: // Remove Opaque Nodes before matching
2615 case Op_Opaque3:
2616 n->subsume_by(n->in(1), this);
2617 break;
2618 case Op_CallStaticJava:
2619 case Op_CallJava:
2620 case Op_CallDynamicJava:
2621 frc.inc_java_call_count(); // Count java call site;
2622 case Op_CallRuntime:
2623 case Op_CallLeaf:
2624 case Op_CallLeafNoFP: {
2625 assert( n->is_Call(), "" );
2626 CallNode *call = n->as_Call();
2627 // Count call sites where the FP mode bit would have to be flipped.
2628 // Do not count uncommon runtime calls:
2629 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2630 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2631 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2632 frc.inc_call_count(); // Count the call site
2633 } else { // See if uncommon argument is shared
2634 Node *n = call->in(TypeFunc::Parms);
2635 int nop = n->Opcode();
2636 // Clone shared simple arguments to uncommon calls, item (1).
2637 if( n->outcnt() > 1 &&
2638 !n->is_Proj() &&
2639 nop != Op_CreateEx &&
2640 nop != Op_CheckCastPP &&
2641 nop != Op_DecodeN &&
2642 nop != Op_DecodeNKlass &&
2643 !n->is_Mem() ) {
2644 Node *x = n->clone();
2645 call->set_req( TypeFunc::Parms, x );
2646 }
2647 }
2648 break;
2649 }
2650
2651 case Op_StoreD:
2652 case Op_LoadD:
2653 case Op_LoadD_unaligned:
2654 frc.inc_double_count();
2655 goto handle_mem;
2656 case Op_StoreF:
2657 case Op_LoadF:
2658 frc.inc_float_count();
2659 goto handle_mem;
2660
2661 case Op_StoreCM:
2662 {
2663 // Convert OopStore dependence into precedence edge
2664 Node* prec = n->in(MemNode::OopStore);
2665 n->del_req(MemNode::OopStore);
2666 n->add_prec(prec);
2667 eliminate_redundant_card_marks(n);
2668 }
2669
2670 // fall through
2671
2672 case Op_StoreB:
2673 case Op_StoreC:
2674 case Op_StorePConditional:
2675 case Op_StoreI:
2676 case Op_StoreL:
2677 case Op_StoreIConditional:
2678 case Op_StoreLConditional:
2679 case Op_CompareAndSwapI:
2680 case Op_CompareAndSwapL:
2681 case Op_CompareAndSwapP:
2682 case Op_CompareAndSwapN:
2683 case Op_GetAndAddI:
2684 case Op_GetAndAddL:
2685 case Op_GetAndSetI:
2686 case Op_GetAndSetL:
2687 case Op_GetAndSetP:
2688 case Op_GetAndSetN:
2689 case Op_StoreP:
2690 case Op_StoreN:
2691 case Op_StoreNKlass:
2692 case Op_LoadB:
2693 case Op_LoadUB:
2694 case Op_LoadUS:
2695 case Op_LoadI:
2696 case Op_LoadKlass:
2697 case Op_LoadNKlass:
2698 case Op_LoadL:
2699 case Op_LoadL_unaligned:
2700 case Op_LoadPLocked:
2701 case Op_LoadP:
2702 case Op_LoadN:
2703 case Op_LoadRange:
2704 case Op_LoadS: {
2705 handle_mem:
2706 #ifdef ASSERT
2707 if( VerifyOptoOopOffsets ) {
2708 assert( n->is_Mem(), "" );
2709 MemNode *mem = (MemNode*)n;
2710 // Check to see if address types have grounded out somehow.
2711 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2712 assert( !tp || oop_offset_is_sane(tp), "" );
2713 }
2714 #endif
2715 break;
2716 }
2717
2718 case Op_AddP: { // Assert sane base pointers
2719 Node *addp = n->in(AddPNode::Address);
2720 assert( !addp->is_AddP() ||
2721 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2722 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2723 "Base pointers must match" );
2724 #ifdef _LP64
2725 if ((UseCompressedOops || UseCompressedClassPointers) &&
2726 addp->Opcode() == Op_ConP &&
2727 addp == n->in(AddPNode::Base) &&
2728 n->in(AddPNode::Offset)->is_Con()) {
2729 // Use addressing with narrow klass to load with offset on x86.
2730 // On sparc loading 32-bits constant and decoding it have less
2731 // instructions (4) then load 64-bits constant (7).
2732 // Do this transformation here since IGVN will convert ConN back to ConP.
2733 const Type* t = addp->bottom_type();
2734 if (t->isa_oopptr() || t->isa_klassptr()) {
2735 Node* nn = NULL;
2736
2737 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2738
2739 // Look for existing ConN node of the same exact type.
2740 Node* r = root();
2741 uint cnt = r->outcnt();
2742 for (uint i = 0; i < cnt; i++) {
2743 Node* m = r->raw_out(i);
2744 if (m!= NULL && m->Opcode() == op &&
2745 m->bottom_type()->make_ptr() == t) {
2746 nn = m;
2747 break;
2748 }
2749 }
2750 if (nn != NULL) {
2751 // Decode a narrow oop to match address
2752 // [R12 + narrow_oop_reg<<3 + offset]
2753 if (t->isa_oopptr()) {
2754 nn = new (this) DecodeNNode(nn, t);
2755 } else {
2756 nn = new (this) DecodeNKlassNode(nn, t);
2757 }
2758 n->set_req(AddPNode::Base, nn);
2759 n->set_req(AddPNode::Address, nn);
2760 if (addp->outcnt() == 0) {
2761 addp->disconnect_inputs(NULL, this);
2762 }
2763 }
2764 }
2765 }
2766 #endif
2767 break;
2768 }
2769
2770 #ifdef _LP64
2771 case Op_CastPP:
2772 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2773 Node* in1 = n->in(1);
2774 const Type* t = n->bottom_type();
2775 Node* new_in1 = in1->clone();
2776 new_in1->as_DecodeN()->set_type(t);
2777
2778 if (!Matcher::narrow_oop_use_complex_address()) {
2779 //
2780 // x86, ARM and friends can handle 2 adds in addressing mode
2781 // and Matcher can fold a DecodeN node into address by using
2782 // a narrow oop directly and do implicit NULL check in address:
2783 //
2784 // [R12 + narrow_oop_reg<<3 + offset]
2785 // NullCheck narrow_oop_reg
2786 //
2787 // On other platforms (Sparc) we have to keep new DecodeN node and
2788 // use it to do implicit NULL check in address:
2789 //
2790 // decode_not_null narrow_oop_reg, base_reg
2791 // [base_reg + offset]
2792 // NullCheck base_reg
2793 //
2794 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2795 // to keep the information to which NULL check the new DecodeN node
2796 // corresponds to use it as value in implicit_null_check().
2797 //
2798 new_in1->set_req(0, n->in(0));
2799 }
2800
2801 n->subsume_by(new_in1, this);
2802 if (in1->outcnt() == 0) {
2803 in1->disconnect_inputs(NULL, this);
2804 }
2805 }
2806 break;
2807
2808 case Op_CmpP:
2809 // Do this transformation here to preserve CmpPNode::sub() and
2810 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2811 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2812 Node* in1 = n->in(1);
2813 Node* in2 = n->in(2);
2814 if (!in1->is_DecodeNarrowPtr()) {
2815 in2 = in1;
2816 in1 = n->in(2);
2817 }
2818 assert(in1->is_DecodeNarrowPtr(), "sanity");
2819
2820 Node* new_in2 = NULL;
2821 if (in2->is_DecodeNarrowPtr()) {
2822 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2823 new_in2 = in2->in(1);
2824 } else if (in2->Opcode() == Op_ConP) {
2825 const Type* t = in2->bottom_type();
2826 if (t == TypePtr::NULL_PTR) {
2827 assert(in1->is_DecodeN(), "compare klass to null?");
2828 // Don't convert CmpP null check into CmpN if compressed
2829 // oops implicit null check is not generated.
2830 // This will allow to generate normal oop implicit null check.
2831 if (Matcher::gen_narrow_oop_implicit_null_checks())
2832 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2833 //
2834 // This transformation together with CastPP transformation above
2835 // will generated code for implicit NULL checks for compressed oops.
2836 //
2837 // The original code after Optimize()
2838 //
2839 // LoadN memory, narrow_oop_reg
2840 // decode narrow_oop_reg, base_reg
2841 // CmpP base_reg, NULL
2842 // CastPP base_reg // NotNull
2843 // Load [base_reg + offset], val_reg
2844 //
2845 // after these transformations will be
2846 //
2847 // LoadN memory, narrow_oop_reg
2848 // CmpN narrow_oop_reg, NULL
2849 // decode_not_null narrow_oop_reg, base_reg
2850 // Load [base_reg + offset], val_reg
2851 //
2852 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2853 // since narrow oops can be used in debug info now (see the code in
2854 // final_graph_reshaping_walk()).
2855 //
2856 // At the end the code will be matched to
2857 // on x86:
2858 //
2859 // Load_narrow_oop memory, narrow_oop_reg
2860 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2861 // NullCheck narrow_oop_reg
2862 //
2863 // and on sparc:
2864 //
2865 // Load_narrow_oop memory, narrow_oop_reg
2866 // decode_not_null narrow_oop_reg, base_reg
2867 // Load [base_reg + offset], val_reg
2868 // NullCheck base_reg
2869 //
2870 } else if (t->isa_oopptr()) {
2871 new_in2 = ConNode::make(this, t->make_narrowoop());
2872 } else if (t->isa_klassptr()) {
2873 new_in2 = ConNode::make(this, t->make_narrowklass());
2874 }
2875 }
2876 if (new_in2 != NULL) {
2877 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2878 n->subsume_by(cmpN, this);
2879 if (in1->outcnt() == 0) {
2880 in1->disconnect_inputs(NULL, this);
2881 }
2882 if (in2->outcnt() == 0) {
2883 in2->disconnect_inputs(NULL, this);
2884 }
2885 }
2886 }
2887 break;
2888
2889 case Op_DecodeN:
2890 case Op_DecodeNKlass:
2891 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2892 // DecodeN could be pinned when it can't be fold into
2893 // an address expression, see the code for Op_CastPP above.
2894 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2895 break;
2896
2897 case Op_EncodeP:
2898 case Op_EncodePKlass: {
2899 Node* in1 = n->in(1);
2900 if (in1->is_DecodeNarrowPtr()) {
2901 n->subsume_by(in1->in(1), this);
2902 } else if (in1->Opcode() == Op_ConP) {
2903 const Type* t = in1->bottom_type();
2904 if (t == TypePtr::NULL_PTR) {
2905 assert(t->isa_oopptr(), "null klass?");
2906 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2907 } else if (t->isa_oopptr()) {
2908 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2909 } else if (t->isa_klassptr()) {
2910 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2911 }
2912 }
2913 if (in1->outcnt() == 0) {
2914 in1->disconnect_inputs(NULL, this);
2915 }
2916 break;
2917 }
2918
2919 case Op_Proj: {
2920 if (OptimizeStringConcat) {
2921 ProjNode* p = n->as_Proj();
2922 if (p->_is_io_use) {
2923 // Separate projections were used for the exception path which
2924 // are normally removed by a late inline. If it wasn't inlined
2925 // then they will hang around and should just be replaced with
2926 // the original one.
2927 Node* proj = NULL;
2928 // Replace with just one
2929 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2930 Node *use = i.get();
2931 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2932 proj = use;
2933 break;
2934 }
2935 }
2936 assert(proj != NULL, "must be found");
2937 p->subsume_by(proj, this);
2938 }
2939 }
2940 break;
2941 }
2942
2943 case Op_Phi:
2944 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2945 // The EncodeP optimization may create Phi with the same edges
2946 // for all paths. It is not handled well by Register Allocator.
2947 Node* unique_in = n->in(1);
2948 assert(unique_in != NULL, "");
2949 uint cnt = n->req();
2950 for (uint i = 2; i < cnt; i++) {
2951 Node* m = n->in(i);
2952 assert(m != NULL, "");
2953 if (unique_in != m)
2954 unique_in = NULL;
2955 }
2956 if (unique_in != NULL) {
2957 n->subsume_by(unique_in, this);
2958 }
2959 }
2960 break;
2961
2962 #endif
2963
2964 case Op_ModI:
2965 if (UseDivMod) {
2966 // Check if a%b and a/b both exist
2967 Node* d = n->find_similar(Op_DivI);
2968 if (d) {
2969 // Replace them with a fused divmod if supported
2970 if (Matcher::has_match_rule(Op_DivModI)) {
2971 DivModINode* divmod = DivModINode::make(this, n);
2972 d->subsume_by(divmod->div_proj(), this);
2973 n->subsume_by(divmod->mod_proj(), this);
2974 } else {
2975 // replace a%b with a-((a/b)*b)
2976 Node* mult = new (this) MulINode(d, d->in(2));
2977 Node* sub = new (this) SubINode(d->in(1), mult);
2978 n->subsume_by(sub, this);
2979 }
2980 }
2981 }
2982 break;
2983
2984 case Op_ModL:
2985 if (UseDivMod) {
2986 // Check if a%b and a/b both exist
2987 Node* d = n->find_similar(Op_DivL);
2988 if (d) {
2989 // Replace them with a fused divmod if supported
2990 if (Matcher::has_match_rule(Op_DivModL)) {
2991 DivModLNode* divmod = DivModLNode::make(this, n);
2992 d->subsume_by(divmod->div_proj(), this);
2993 n->subsume_by(divmod->mod_proj(), this);
2994 } else {
2995 // replace a%b with a-((a/b)*b)
2996 Node* mult = new (this) MulLNode(d, d->in(2));
2997 Node* sub = new (this) SubLNode(d->in(1), mult);
2998 n->subsume_by(sub, this);
2999 }
3000 }
3001 }
3002 break;
3003
3004 case Op_LoadVector:
3005 case Op_StoreVector:
3006 break;
3007
3008 case Op_PackB:
3009 case Op_PackS:
3010 case Op_PackI:
3011 case Op_PackF:
3012 case Op_PackL:
3013 case Op_PackD:
3014 if (n->req()-1 > 2) {
3015 // Replace many operand PackNodes with a binary tree for matching
3016 PackNode* p = (PackNode*) n;
3017 Node* btp = p->binary_tree_pack(this, 1, n->req());
3018 n->subsume_by(btp, this);
3019 }
3020 break;
3021 case Op_Loop:
3022 case Op_CountedLoop:
3023 if (n->as_Loop()->is_inner_loop()) {
3024 frc.inc_inner_loop_count();
3025 }
3026 break;
3027 case Op_LShiftI:
3028 case Op_RShiftI:
3029 case Op_URShiftI:
3030 case Op_LShiftL:
3031 case Op_RShiftL:
3032 case Op_URShiftL:
3033 if (Matcher::need_masked_shift_count) {
3034 // The cpu's shift instructions don't restrict the count to the
3035 // lower 5/6 bits. We need to do the masking ourselves.
3036 Node* in2 = n->in(2);
3037 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3038 const TypeInt* t = in2->find_int_type();
3039 if (t != NULL && t->is_con()) {
3040 juint shift = t->get_con();
3041 if (shift > mask) { // Unsigned cmp
3042 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3043 }
3044 } else {
3045 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3046 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3047 n->set_req(2, shift);
3048 }
3049 }
3050 if (in2->outcnt() == 0) { // Remove dead node
3051 in2->disconnect_inputs(NULL, this);
3052 }
3053 }
3054 break;
3055 case Op_MemBarStoreStore:
3056 case Op_MemBarRelease:
3057 // Break the link with AllocateNode: it is no longer useful and
3058 // confuses register allocation.
3059 if (n->req() > MemBarNode::Precedent) {
3060 n->set_req(MemBarNode::Precedent, top());
3061 }
3062 break;
3063 default:
3064 assert( !n->is_Call(), "" );
3065 assert( !n->is_Mem(), "" );
3066 break;
3067 }
3068
3069 // Collect CFG split points
3070 if (n->is_MultiBranch())
3071 frc._tests.push(n);
3072 }
3073
3074 //------------------------------final_graph_reshaping_walk---------------------
3075 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3076 // requires that the walk visits a node's inputs before visiting the node.
3077 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3078 ResourceArea *area = Thread::current()->resource_area();
3079 Unique_Node_List sfpt(area);
3080
3081 frc._visited.set(root->_idx); // first, mark node as visited
3082 uint cnt = root->req();
3083 Node *n = root;
3084 uint i = 0;
3085 while (true) {
3086 if (i < cnt) {
3087 // Place all non-visited non-null inputs onto stack
3088 Node* m = n->in(i);
3089 ++i;
3090 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3091 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3092 // compute worst case interpreter size in case of a deoptimization
3093 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3094
3095 sfpt.push(m);
3096 }
3097 cnt = m->req();
3098 nstack.push(n, i); // put on stack parent and next input's index
3099 n = m;
3100 i = 0;
3101 }
3102 } else {
3103 // Now do post-visit work
3104 final_graph_reshaping_impl( n, frc );
3105 if (nstack.is_empty())
3106 break; // finished
3107 n = nstack.node(); // Get node from stack
3108 cnt = n->req();
3109 i = nstack.index();
3110 nstack.pop(); // Shift to the next node on stack
3111 }
3112 }
3113
3114 // Skip next transformation if compressed oops are not used.
3115 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3116 (!UseCompressedOops && !UseCompressedClassPointers))
3117 return;
3118
3119 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3120 // It could be done for an uncommon traps or any safepoints/calls
3121 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3122 while (sfpt.size() > 0) {
3123 n = sfpt.pop();
3124 JVMState *jvms = n->as_SafePoint()->jvms();
3125 assert(jvms != NULL, "sanity");
3126 int start = jvms->debug_start();
3127 int end = n->req();
3128 bool is_uncommon = (n->is_CallStaticJava() &&
3129 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3130 for (int j = start; j < end; j++) {
3131 Node* in = n->in(j);
3132 if (in->is_DecodeNarrowPtr()) {
3133 bool safe_to_skip = true;
3134 if (!is_uncommon ) {
3135 // Is it safe to skip?
3136 for (uint i = 0; i < in->outcnt(); i++) {
3137 Node* u = in->raw_out(i);
3138 if (!u->is_SafePoint() ||
3139 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3140 safe_to_skip = false;
3141 }
3142 }
3143 }
3144 if (safe_to_skip) {
3145 n->set_req(j, in->in(1));
3146 }
3147 if (in->outcnt() == 0) {
3148 in->disconnect_inputs(NULL, this);
3149 }
3150 }
3151 }
3152 }
3153 }
3154
3155 //------------------------------final_graph_reshaping--------------------------
3156 // Final Graph Reshaping.
3157 //
3158 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3159 // and not commoned up and forced early. Must come after regular
3160 // optimizations to avoid GVN undoing the cloning. Clone constant
3161 // inputs to Loop Phis; these will be split by the allocator anyways.
3162 // Remove Opaque nodes.
3163 // (2) Move last-uses by commutative operations to the left input to encourage
3164 // Intel update-in-place two-address operations and better register usage
3165 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3166 // calls canonicalizing them back.
3167 // (3) Count the number of double-precision FP ops, single-precision FP ops
3168 // and call sites. On Intel, we can get correct rounding either by
3169 // forcing singles to memory (requires extra stores and loads after each
3170 // FP bytecode) or we can set a rounding mode bit (requires setting and
3171 // clearing the mode bit around call sites). The mode bit is only used
3172 // if the relative frequency of single FP ops to calls is low enough.
3173 // This is a key transform for SPEC mpeg_audio.
3174 // (4) Detect infinite loops; blobs of code reachable from above but not
3175 // below. Several of the Code_Gen algorithms fail on such code shapes,
3176 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3177 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3178 // Detection is by looking for IfNodes where only 1 projection is
3179 // reachable from below or CatchNodes missing some targets.
3180 // (5) Assert for insane oop offsets in debug mode.
3181
3182 bool Compile::final_graph_reshaping() {
3183 // an infinite loop may have been eliminated by the optimizer,
3184 // in which case the graph will be empty.
3185 if (root()->req() == 1) {
3186 record_method_not_compilable("trivial infinite loop");
3187 return true;
3188 }
3189
3190 // Expensive nodes have their control input set to prevent the GVN
3191 // from freely commoning them. There's no GVN beyond this point so
3192 // no need to keep the control input. We want the expensive nodes to
3193 // be freely moved to the least frequent code path by gcm.
3194 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3195 for (int i = 0; i < expensive_count(); i++) {
3196 _expensive_nodes->at(i)->set_req(0, NULL);
3197 }
3198
3199 Final_Reshape_Counts frc;
3200
3201 // Visit everybody reachable!
3202 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3203 Node_Stack nstack(unique() >> 1);
3204 final_graph_reshaping_walk(nstack, root(), frc);
3205
3206 // Check for unreachable (from below) code (i.e., infinite loops).
3207 for( uint i = 0; i < frc._tests.size(); i++ ) {
3208 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3209 // Get number of CFG targets.
3210 // Note that PCTables include exception targets after calls.
3211 uint required_outcnt = n->required_outcnt();
3212 if (n->outcnt() != required_outcnt) {
3213 // Check for a few special cases. Rethrow Nodes never take the
3214 // 'fall-thru' path, so expected kids is 1 less.
3215 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3216 if (n->in(0)->in(0)->is_Call()) {
3217 CallNode *call = n->in(0)->in(0)->as_Call();
3218 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3219 required_outcnt--; // Rethrow always has 1 less kid
3220 } else if (call->req() > TypeFunc::Parms &&
3221 call->is_CallDynamicJava()) {
3222 // Check for null receiver. In such case, the optimizer has
3223 // detected that the virtual call will always result in a null
3224 // pointer exception. The fall-through projection of this CatchNode
3225 // will not be populated.
3226 Node *arg0 = call->in(TypeFunc::Parms);
3227 if (arg0->is_Type() &&
3228 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3229 required_outcnt--;
3230 }
3231 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3232 call->req() > TypeFunc::Parms+1 &&
3233 call->is_CallStaticJava()) {
3234 // Check for negative array length. In such case, the optimizer has
3235 // detected that the allocation attempt will always result in an
3236 // exception. There is no fall-through projection of this CatchNode .
3237 Node *arg1 = call->in(TypeFunc::Parms+1);
3238 if (arg1->is_Type() &&
3239 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3240 required_outcnt--;
3241 }
3242 }
3243 }
3244 }
3245 // Recheck with a better notion of 'required_outcnt'
3246 if (n->outcnt() != required_outcnt) {
3247 record_method_not_compilable("malformed control flow");
3248 return true; // Not all targets reachable!
3249 }
3250 }
3251 // Check that I actually visited all kids. Unreached kids
3252 // must be infinite loops.
3253 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3254 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3255 record_method_not_compilable("infinite loop");
3256 return true; // Found unvisited kid; must be unreach
3257 }
3258 }
3259
3260 // If original bytecodes contained a mixture of floats and doubles
3261 // check if the optimizer has made it homogenous, item (3).
3262 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3263 frc.get_float_count() > 32 &&
3264 frc.get_double_count() == 0 &&
3265 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3266 set_24_bit_selection_and_mode( false, true );
3267 }
3268
3269 set_java_calls(frc.get_java_call_count());
3270 set_inner_loops(frc.get_inner_loop_count());
3271
3272 // No infinite loops, no reason to bail out.
3273 return false;
3274 }
3275
3276 //-----------------------------too_many_traps----------------------------------
3277 // Report if there are too many traps at the current method and bci.
3278 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3279 bool Compile::too_many_traps(ciMethod* method,
3280 int bci,
3281 Deoptimization::DeoptReason reason) {
3282 ciMethodData* md = method->method_data();
3283 if (md->is_empty()) {
3284 // Assume the trap has not occurred, or that it occurred only
3285 // because of a transient condition during start-up in the interpreter.
3286 return false;
3287 }
3288 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3289 if (md->has_trap_at(bci, m, reason) != 0) {
3290 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3291 // Also, if there are multiple reasons, or if there is no per-BCI record,
3292 // assume the worst.
3293 if (log())
3294 log()->elem("observe trap='%s' count='%d'",
3295 Deoptimization::trap_reason_name(reason),
3296 md->trap_count(reason));
3297 return true;
3298 } else {
3299 // Ignore method/bci and see if there have been too many globally.
3300 return too_many_traps(reason, md);
3301 }
3302 }
3303
3304 // Less-accurate variant which does not require a method and bci.
3305 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3306 ciMethodData* logmd) {
3307 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3308 // Too many traps globally.
3309 // Note that we use cumulative trap_count, not just md->trap_count.
3310 if (log()) {
3311 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3312 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3313 Deoptimization::trap_reason_name(reason),
3314 mcount, trap_count(reason));
3315 }
3316 return true;
3317 } else {
3318 // The coast is clear.
3319 return false;
3320 }
3321 }
3322
3323 //--------------------------too_many_recompiles--------------------------------
3324 // Report if there are too many recompiles at the current method and bci.
3325 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3326 // Is not eager to return true, since this will cause the compiler to use
3327 // Action_none for a trap point, to avoid too many recompilations.
3328 bool Compile::too_many_recompiles(ciMethod* method,
3329 int bci,
3330 Deoptimization::DeoptReason reason) {
3331 ciMethodData* md = method->method_data();
3332 if (md->is_empty()) {
3333 // Assume the trap has not occurred, or that it occurred only
3334 // because of a transient condition during start-up in the interpreter.
3335 return false;
3336 }
3337 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3338 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3339 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3340 Deoptimization::DeoptReason per_bc_reason
3341 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3342 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3343 if ((per_bc_reason == Deoptimization::Reason_none
3344 || md->has_trap_at(bci, m, reason) != 0)
3345 // The trap frequency measure we care about is the recompile count:
3346 && md->trap_recompiled_at(bci, m)
3347 && md->overflow_recompile_count() >= bc_cutoff) {
3348 // Do not emit a trap here if it has already caused recompilations.
3349 // Also, if there are multiple reasons, or if there is no per-BCI record,
3350 // assume the worst.
3351 if (log())
3352 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3353 Deoptimization::trap_reason_name(reason),
3354 md->trap_count(reason),
3355 md->overflow_recompile_count());
3356 return true;
3357 } else if (trap_count(reason) != 0
3358 && decompile_count() >= m_cutoff) {
3359 // Too many recompiles globally, and we have seen this sort of trap.
3360 // Use cumulative decompile_count, not just md->decompile_count.
3361 if (log())
3362 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3363 Deoptimization::trap_reason_name(reason),
3364 md->trap_count(reason), trap_count(reason),
3365 md->decompile_count(), decompile_count());
3366 return true;
3367 } else {
3368 // The coast is clear.
3369 return false;
3370 }
3371 }
3372
3373 // Compute when not to trap. Used by matching trap based nodes and
3374 // NullCheck optimization.
3375 void Compile::set_allowed_deopt_reasons() {
3376 _allowed_reasons = 0;
3377 if (is_method_compilation()) {
3378 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3379 assert(rs < BitsPerInt, "recode bit map");
3380 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3381 _allowed_reasons |= nth_bit(rs);
3382 }
3383 }
3384 }
3385 }
3386
3387 #ifndef PRODUCT
3388 //------------------------------verify_graph_edges---------------------------
3389 // Walk the Graph and verify that there is a one-to-one correspondence
3390 // between Use-Def edges and Def-Use edges in the graph.
3391 void Compile::verify_graph_edges(bool no_dead_code) {
3392 if (VerifyGraphEdges) {
3393 ResourceArea *area = Thread::current()->resource_area();
3394 Unique_Node_List visited(area);
3395 // Call recursive graph walk to check edges
3396 _root->verify_edges(visited);
3397 if (no_dead_code) {
3398 // Now make sure that no visited node is used by an unvisited node.
3399 bool dead_nodes = 0;
3400 Unique_Node_List checked(area);
3401 while (visited.size() > 0) {
3402 Node* n = visited.pop();
3403 checked.push(n);
3404 for (uint i = 0; i < n->outcnt(); i++) {
3405 Node* use = n->raw_out(i);
3406 if (checked.member(use)) continue; // already checked
3407 if (visited.member(use)) continue; // already in the graph
3408 if (use->is_Con()) continue; // a dead ConNode is OK
3409 // At this point, we have found a dead node which is DU-reachable.
3410 if (dead_nodes++ == 0)
3411 tty->print_cr("*** Dead nodes reachable via DU edges:");
3412 use->dump(2);
3413 tty->print_cr("---");
3414 checked.push(use); // No repeats; pretend it is now checked.
3415 }
3416 }
3417 assert(dead_nodes == 0, "using nodes must be reachable from root");
3418 }
3419 }
3420 }
3421
3422 // Verify GC barriers consistency
3423 // Currently supported:
3424 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3425 void Compile::verify_barriers() {
3426 if (UseG1GC) {
3427 // Verify G1 pre-barriers
3428 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3429
3430 ResourceArea *area = Thread::current()->resource_area();
3431 Unique_Node_List visited(area);
3432 Node_List worklist(area);
3433 // We're going to walk control flow backwards starting from the Root
3434 worklist.push(_root);
3435 while (worklist.size() > 0) {
3436 Node* x = worklist.pop();
3437 if (x == NULL || x == top()) continue;
3438 if (visited.member(x)) {
3439 continue;
3440 } else {
3441 visited.push(x);
3442 }
3443
3444 if (x->is_Region()) {
3445 for (uint i = 1; i < x->req(); i++) {
3446 worklist.push(x->in(i));
3447 }
3448 } else {
3449 worklist.push(x->in(0));
3450 // We are looking for the pattern:
3451 // /->ThreadLocal
3452 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3453 // \->ConI(0)
3454 // We want to verify that the If and the LoadB have the same control
3455 // See GraphKit::g1_write_barrier_pre()
3456 if (x->is_If()) {
3457 IfNode *iff = x->as_If();
3458 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3459 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3460 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3461 && cmp->in(1)->is_Load()) {
3462 LoadNode* load = cmp->in(1)->as_Load();
3463 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3464 && load->in(2)->in(3)->is_Con()
3465 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3466
3467 Node* if_ctrl = iff->in(0);
3468 Node* load_ctrl = load->in(0);
3469
3470 if (if_ctrl != load_ctrl) {
3471 // Skip possible CProj->NeverBranch in infinite loops
3472 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3473 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3474 if_ctrl = if_ctrl->in(0)->in(0);
3475 }
3476 }
3477 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3478 }
3479 }
3480 }
3481 }
3482 }
3483 }
3484 }
3485 }
3486
3487 #endif
3488
3489 // The Compile object keeps track of failure reasons separately from the ciEnv.
3490 // This is required because there is not quite a 1-1 relation between the
3491 // ciEnv and its compilation task and the Compile object. Note that one
3492 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3493 // to backtrack and retry without subsuming loads. Other than this backtracking
3494 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3495 // by the logic in C2Compiler.
3496 void Compile::record_failure(const char* reason) {
3497 if (log() != NULL) {
3498 log()->elem("failure reason='%s' phase='compile'", reason);
3499 }
3500 if (_failure_reason == NULL) {
3501 // Record the first failure reason.
3502 _failure_reason = reason;
3503 }
3504
3505 EventCompilerFailure event;
3506 if (event.should_commit()) {
3507 event.set_compileID(Compile::compile_id());
3508 event.set_failure(reason);
3509 event.commit();
3510 }
3511
3512 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3513 C->print_method(PHASE_FAILURE);
3514 }
3515 _root = NULL; // flush the graph, too
3516 }
3517
3518 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3519 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3520 _phase_name(name), _dolog(dolog)
3521 {
3522 if (dolog) {
3523 C = Compile::current();
3524 _log = C->log();
3525 } else {
3526 C = NULL;
3527 _log = NULL;
3528 }
3529 if (_log != NULL) {
3530 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3531 _log->stamp();
3532 _log->end_head();
3533 }
3534 }
3535
3536 Compile::TracePhase::~TracePhase() {
3537
3538 C = Compile::current();
3539 if (_dolog) {
3540 _log = C->log();
3541 } else {
3542 _log = NULL;
3543 }
3544
3545 #ifdef ASSERT
3546 if (PrintIdealNodeCount) {
3547 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3548 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3549 }
3550
3551 if (VerifyIdealNodeCount) {
3552 Compile::current()->print_missing_nodes();
3553 }
3554 #endif
3555
3556 if (_log != NULL) {
3557 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3558 }
3559 }
3560
3561 //=============================================================================
3562 // Two Constant's are equal when the type and the value are equal.
3563 bool Compile::Constant::operator==(const Constant& other) {
3564 if (type() != other.type() ) return false;
3565 if (can_be_reused() != other.can_be_reused()) return false;
3566 // For floating point values we compare the bit pattern.
3567 switch (type()) {
3568 case T_FLOAT: return (_v._value.i == other._v._value.i);
3569 case T_LONG:
3570 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3571 case T_OBJECT:
3572 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3573 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3574 case T_METADATA: return (_v._metadata == other._v._metadata);
3575 default: ShouldNotReachHere();
3576 }
3577 return false;
3578 }
3579
3580 static int type_to_size_in_bytes(BasicType t) {
3581 switch (t) {
3582 case T_LONG: return sizeof(jlong );
3583 case T_FLOAT: return sizeof(jfloat );
3584 case T_DOUBLE: return sizeof(jdouble);
3585 case T_METADATA: return sizeof(Metadata*);
3586 // We use T_VOID as marker for jump-table entries (labels) which
3587 // need an internal word relocation.
3588 case T_VOID:
3589 case T_ADDRESS:
3590 case T_OBJECT: return sizeof(jobject);
3591 }
3592
3593 ShouldNotReachHere();
3594 return -1;
3595 }
3596
3597 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3598 // sort descending
3599 if (a->freq() > b->freq()) return -1;
3600 if (a->freq() < b->freq()) return 1;
3601 return 0;
3602 }
3603
3604 void Compile::ConstantTable::calculate_offsets_and_size() {
3605 // First, sort the array by frequencies.
3606 _constants.sort(qsort_comparator);
3607
3608 #ifdef ASSERT
3609 // Make sure all jump-table entries were sorted to the end of the
3610 // array (they have a negative frequency).
3611 bool found_void = false;
3612 for (int i = 0; i < _constants.length(); i++) {
3613 Constant con = _constants.at(i);
3614 if (con.type() == T_VOID)
3615 found_void = true; // jump-tables
3616 else
3617 assert(!found_void, "wrong sorting");
3618 }
3619 #endif
3620
3621 int offset = 0;
3622 for (int i = 0; i < _constants.length(); i++) {
3623 Constant* con = _constants.adr_at(i);
3624
3625 // Align offset for type.
3626 int typesize = type_to_size_in_bytes(con->type());
3627 offset = align_size_up(offset, typesize);
3628 con->set_offset(offset); // set constant's offset
3629
3630 if (con->type() == T_VOID) {
3631 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3632 offset = offset + typesize * n->outcnt(); // expand jump-table
3633 } else {
3634 offset = offset + typesize;
3635 }
3636 }
3637
3638 // Align size up to the next section start (which is insts; see
3639 // CodeBuffer::align_at_start).
3640 assert(_size == -1, "already set?");
3641 _size = align_size_up(offset, CodeEntryAlignment);
3642 }
3643
3644 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3645 MacroAssembler _masm(&cb);
3646 for (int i = 0; i < _constants.length(); i++) {
3647 Constant con = _constants.at(i);
3648 address constant_addr;
3649 switch (con.type()) {
3650 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3651 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3652 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3653 case T_OBJECT: {
3654 jobject obj = con.get_jobject();
3655 int oop_index = _masm.oop_recorder()->find_index(obj);
3656 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3657 break;
3658 }
3659 case T_ADDRESS: {
3660 address addr = (address) con.get_jobject();
3661 constant_addr = _masm.address_constant(addr);
3662 break;
3663 }
3664 // We use T_VOID as marker for jump-table entries (labels) which
3665 // need an internal word relocation.
3666 case T_VOID: {
3667 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3668 // Fill the jump-table with a dummy word. The real value is
3669 // filled in later in fill_jump_table.
3670 address dummy = (address) n;
3671 constant_addr = _masm.address_constant(dummy);
3672 // Expand jump-table
3673 for (uint i = 1; i < n->outcnt(); i++) {
3674 address temp_addr = _masm.address_constant(dummy + i);
3675 assert(temp_addr, "consts section too small");
3676 }
3677 break;
3678 }
3679 case T_METADATA: {
3680 Metadata* obj = con.get_metadata();
3681 int metadata_index = _masm.oop_recorder()->find_index(obj);
3682 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3683 break;
3684 }
3685 default: ShouldNotReachHere();
3686 }
3687 assert(constant_addr, "consts section too small");
3688 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3689 }
3690 }
3691
3692 int Compile::ConstantTable::find_offset(Constant& con) const {
3693 int idx = _constants.find(con);
3694 assert(idx != -1, "constant must be in constant table");
3695 int offset = _constants.at(idx).offset();
3696 assert(offset != -1, "constant table not emitted yet?");
3697 return offset;
3698 }
3699
3700 void Compile::ConstantTable::add(Constant& con) {
3701 if (con.can_be_reused()) {
3702 int idx = _constants.find(con);
3703 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3704 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3705 return;
3706 }
3707 }
3708 (void) _constants.append(con);
3709 }
3710
3711 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3712 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3713 Constant con(type, value, b->_freq);
3714 add(con);
3715 return con;
3716 }
3717
3718 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3719 Constant con(metadata);
3720 add(con);
3721 return con;
3722 }
3723
3724 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3725 jvalue value;
3726 BasicType type = oper->type()->basic_type();
3727 switch (type) {
3728 case T_LONG: value.j = oper->constantL(); break;
3729 case T_FLOAT: value.f = oper->constantF(); break;
3730 case T_DOUBLE: value.d = oper->constantD(); break;
3731 case T_OBJECT:
3732 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3733 case T_METADATA: return add((Metadata*)oper->constant()); break;
3734 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3735 }
3736 return add(n, type, value);
3737 }
3738
3739 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3740 jvalue value;
3741 // We can use the node pointer here to identify the right jump-table
3742 // as this method is called from Compile::Fill_buffer right before
3743 // the MachNodes are emitted and the jump-table is filled (means the
3744 // MachNode pointers do not change anymore).
3745 value.l = (jobject) n;
3746 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3747 add(con);
3748 return con;
3749 }
3750
3751 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3752 // If called from Compile::scratch_emit_size do nothing.
3753 if (Compile::current()->in_scratch_emit_size()) return;
3754
3755 assert(labels.is_nonempty(), "must be");
3756 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3757
3758 // Since MachConstantNode::constant_offset() also contains
3759 // table_base_offset() we need to subtract the table_base_offset()
3760 // to get the plain offset into the constant table.
3761 int offset = n->constant_offset() - table_base_offset();
3762
3763 MacroAssembler _masm(&cb);
3764 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3765
3766 for (uint i = 0; i < n->outcnt(); i++) {
3767 address* constant_addr = &jump_table_base[i];
3768 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)));
3769 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3770 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3771 }
3772 }
3773
3774 // The message about the current inlining is accumulated in
3775 // _print_inlining_stream and transfered into the _print_inlining_list
3776 // once we know whether inlining succeeds or not. For regular
3777 // inlining, messages are appended to the buffer pointed by
3778 // _print_inlining_idx in the _print_inlining_list. For late inlining,
3779 // a new buffer is added after _print_inlining_idx in the list. This
3780 // way we can update the inlining message for late inlining call site
3781 // when the inlining is attempted again.
3782 void Compile::print_inlining_init() {
3783 if (print_inlining() || print_intrinsics()) {
3784 _print_inlining_stream = new stringStream();
3785 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
3786 }
3787 }
3788
3789 void Compile::print_inlining_reinit() {
3790 if (print_inlining() || print_intrinsics()) {
3791 // Re allocate buffer when we change ResourceMark
3792 _print_inlining_stream = new stringStream();
3793 }
3794 }
3795
3796 void Compile::print_inlining_reset() {
3797 _print_inlining_stream->reset();
3798 }
3799
3800 void Compile::print_inlining_commit() {
3801 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
3802 // Transfer the message from _print_inlining_stream to the current
3803 // _print_inlining_list buffer and clear _print_inlining_stream.
3804 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
3805 print_inlining_reset();
3806 }
3807
3808 void Compile::print_inlining_push() {
3809 // Add new buffer to the _print_inlining_list at current position
3810 _print_inlining_idx++;
3811 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
3812 }
3813
3814 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
3815 return _print_inlining_list->at(_print_inlining_idx);
3816 }
3817
3818 void Compile::print_inlining_update(CallGenerator* cg) {
3819 if (print_inlining() || print_intrinsics()) {
3820 if (!cg->is_late_inline()) {
3821 if (print_inlining_current().cg() != NULL) {
3822 print_inlining_push();
3823 }
3824 print_inlining_commit();
3825 } else {
3826 if (print_inlining_current().cg() != cg &&
3827 (print_inlining_current().cg() != NULL ||
3828 print_inlining_current().ss()->size() != 0)) {
3829 print_inlining_push();
3830 }
3831 print_inlining_commit();
3832 print_inlining_current().set_cg(cg);
3833 }
3834 }
3835 }
3836
3837 void Compile::print_inlining_move_to(CallGenerator* cg) {
3838 // We resume inlining at a late inlining call site. Locate the
3839 // corresponding inlining buffer so that we can update it.
3840 if (print_inlining()) {
3841 for (int i = 0; i < _print_inlining_list->length(); i++) {
3842 if (_print_inlining_list->adr_at(i)->cg() == cg) {
3843 _print_inlining_idx = i;
3844 return;
3845 }
3846 }
3847 ShouldNotReachHere();
3848 }
3849 }
3850
3851 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
3852 if (print_inlining()) {
3853 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
3854 assert(print_inlining_current().cg() == cg, "wrong entry");
3855 // replace message with new message
3856 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
3857 print_inlining_commit();
3858 print_inlining_current().set_cg(cg);
3859 }
3860 }
3861
3862 void Compile::print_inlining_assert_ready() {
3863 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
3864 }
3865
3866 void Compile::dump_inlining() {
3867 bool do_print_inlining = print_inlining() || print_intrinsics();
3868 if (do_print_inlining || log() != NULL) {
3869 // Print inlining message for candidates that we couldn't inline
3870 // for lack of space
3871 for (int i = 0; i < _late_inlines.length(); i++) {
3872 CallGenerator* cg = _late_inlines.at(i);
3873 if (!cg->is_mh_late_inline()) {
3874 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
3875 if (do_print_inlining) {
3876 cg->print_inlining_late(msg);
3877 }
3878 log_late_inline_failure(cg, msg);
3879 }
3880 }
3881 }
3882 if (do_print_inlining) {
3883 for (int i = 0; i < _print_inlining_list->length(); i++) {
3884 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3885 }
3886 }
3887 }
3888
3889 void Compile::log_late_inline(CallGenerator* cg) {
3890 if (log() != NULL) {
3891 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
3892 cg->unique_id());
3893 JVMState* p = cg->call_node()->jvms();
3894 while (p != NULL) {
3895 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
3896 p = p->caller();
3897 }
3898 log()->tail("late_inline");
3899 }
3900 }
3901
3902 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
3903 log_late_inline(cg);
3904 if (log() != NULL) {
3905 log()->inline_fail(msg);
3906 }
3907 }
3908
3909 void Compile::log_inline_id(CallGenerator* cg) {
3910 if (log() != NULL) {
3911 // The LogCompilation tool needs a unique way to identify late
3912 // inline call sites. This id must be unique for this call site in
3913 // this compilation. Try to have it unique across compilations as
3914 // well because it can be convenient when grepping through the log
3915 // file.
3916 // Distinguish OSR compilations from others in case CICountOSR is
3917 // on.
3918 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
3919 cg->set_unique_id(id);
3920 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
3921 }
3922 }
3923
3924 void Compile::log_inline_failure(const char* msg) {
3925 if (C->log() != NULL) {
3926 C->log()->inline_fail(msg);
3927 }
3928 }
3929
3930
3931 // Dump inlining replay data to the stream.
3932 // Don't change thread state and acquire any locks.
3933 void Compile::dump_inline_data(outputStream* out) {
3934 InlineTree* inl_tree = ilt();
3935 if (inl_tree != NULL) {
3936 out->print(" inline %d", inl_tree->count());
3937 inl_tree->dump_replay_data(out);
3938 }
3939 }
3940
3941 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3942 if (n1->Opcode() < n2->Opcode()) return -1;
3943 else if (n1->Opcode() > n2->Opcode()) return 1;
3944
3945 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()));
3946 for (uint i = 1; i < n1->req(); i++) {
3947 if (n1->in(i) < n2->in(i)) return -1;
3948 else if (n1->in(i) > n2->in(i)) return 1;
3949 }
3950
3951 return 0;
3952 }
3953
3954 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3955 Node* n1 = *n1p;
3956 Node* n2 = *n2p;
3957
3958 return cmp_expensive_nodes(n1, n2);
3959 }
3960
3961 void Compile::sort_expensive_nodes() {
3962 if (!expensive_nodes_sorted()) {
3963 _expensive_nodes->sort(cmp_expensive_nodes);
3964 }
3965 }
3966
3967 bool Compile::expensive_nodes_sorted() const {
3968 for (int i = 1; i < _expensive_nodes->length(); i++) {
3969 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3970 return false;
3971 }
3972 }
3973 return true;
3974 }
3975
3976 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3977 if (_expensive_nodes->length() == 0) {
3978 return false;
3979 }
3980
3981 assert(OptimizeExpensiveOps, "optimization off?");
3982
3983 // Take this opportunity to remove dead nodes from the list
3984 int j = 0;
3985 for (int i = 0; i < _expensive_nodes->length(); i++) {
3986 Node* n = _expensive_nodes->at(i);
3987 if (!n->is_unreachable(igvn)) {
3988 assert(n->is_expensive(), "should be expensive");
3989 _expensive_nodes->at_put(j, n);
3990 j++;
3991 }
3992 }
3993 _expensive_nodes->trunc_to(j);
3994
3995 // Then sort the list so that similar nodes are next to each other
3996 // and check for at least two nodes of identical kind with same data
3997 // inputs.
3998 sort_expensive_nodes();
3999
4000 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4001 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4002 return true;
4003 }
4004 }
4005
4006 return false;
4007 }
4008
4009 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4010 if (_expensive_nodes->length() == 0) {
4011 return;
4012 }
4013
4014 assert(OptimizeExpensiveOps, "optimization off?");
4015
4016 // Sort to bring similar nodes next to each other and clear the
4017 // control input of nodes for which there's only a single copy.
4018 sort_expensive_nodes();
4019
4020 int j = 0;
4021 int identical = 0;
4022 int i = 0;
4023 for (; i < _expensive_nodes->length()-1; i++) {
4024 assert(j <= i, "can't write beyond current index");
4025 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4026 identical++;
4027 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4028 continue;
4029 }
4030 if (identical > 0) {
4031 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4032 identical = 0;
4033 } else {
4034 Node* n = _expensive_nodes->at(i);
4035 igvn.hash_delete(n);
4036 n->set_req(0, NULL);
4037 igvn.hash_insert(n);
4038 }
4039 }
4040 if (identical > 0) {
4041 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4042 } else if (_expensive_nodes->length() >= 1) {
4043 Node* n = _expensive_nodes->at(i);
4044 igvn.hash_delete(n);
4045 n->set_req(0, NULL);
4046 igvn.hash_insert(n);
4047 }
4048 _expensive_nodes->trunc_to(j);
4049 }
4050
4051 void Compile::add_expensive_node(Node * n) {
4052 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4053 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4054 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4055 if (OptimizeExpensiveOps) {
4056 _expensive_nodes->append(n);
4057 } else {
4058 // Clear control input and let IGVN optimize expensive nodes if
4059 // OptimizeExpensiveOps is off.
4060 n->set_req(0, NULL);
4061 }
4062 }
4063
4064 /**
4065 * Remove the speculative part of types and clean up the graph
4066 */
4067 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4068 if (UseTypeSpeculation) {
4069 Unique_Node_List worklist;
4070 worklist.push(root());
4071 int modified = 0;
4072 // Go over all type nodes that carry a speculative type, drop the
4073 // speculative part of the type and enqueue the node for an igvn
4074 // which may optimize it out.
4075 for (uint next = 0; next < worklist.size(); ++next) {
4076 Node *n = worklist.at(next);
4077 if (n->is_Type()) {
4078 TypeNode* tn = n->as_Type();
4079 const Type* t = tn->type();
4080 const Type* t_no_spec = t->remove_speculative();
4081 if (t_no_spec != t) {
4082 bool in_hash = igvn.hash_delete(n);
4083 assert(in_hash, "node should be in igvn hash table");
4084 tn->set_type(t_no_spec);
4085 igvn.hash_insert(n);
4086 igvn._worklist.push(n); // give it a chance to go away
4087 modified++;
4088 }
4089 }
4090 uint max = n->len();
4091 for( uint i = 0; i < max; ++i ) {
4092 Node *m = n->in(i);
4093 if (not_a_node(m)) continue;
4094 worklist.push(m);
4095 }
4096 }
4097 // Drop the speculative part of all types in the igvn's type table
4098 igvn.remove_speculative_types();
4099 if (modified > 0) {
4100 igvn.optimize();
4101 }
4102 #ifdef ASSERT
4103 // Verify that after the IGVN is over no speculative type has resurfaced
4104 worklist.clear();
4105 worklist.push(root());
4106 for (uint next = 0; next < worklist.size(); ++next) {
4107 Node *n = worklist.at(next);
4108 const Type* t = igvn.type_or_null(n);
4109 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4110 if (n->is_Type()) {
4111 t = n->as_Type()->type();
4112 assert(t == t->remove_speculative(), "no more speculative types");
4113 }
4114 uint max = n->len();
4115 for( uint i = 0; i < max; ++i ) {
4116 Node *m = n->in(i);
4117 if (not_a_node(m)) continue;
4118 worklist.push(m);
4119 }
4120 }
4121 igvn.check_no_speculative_types();
4122 #endif
4123 }
4124 }
4125
4126 // Auxiliary method to support randomized stressing/fuzzing.
4127 //
4128 // This method can be called the arbitrary number of times, with current count
4129 // as the argument. The logic allows selecting a single candidate from the
4130 // running list of candidates as follows:
4131 // int count = 0;
4132 // Cand* selected = null;
4133 // while(cand = cand->next()) {
4134 // if (randomized_select(++count)) {
4135 // selected = cand;
4136 // }
4137 // }
4138 //
4139 // Including count equalizes the chances any candidate is "selected".
4140 // This is useful when we don't have the complete list of candidates to choose
4141 // from uniformly. In this case, we need to adjust the randomicity of the
4142 // selection, or else we will end up biasing the selection towards the latter
4143 // candidates.
4144 //
4145 // Quick back-envelope calculation shows that for the list of n candidates
4146 // the equal probability for the candidate to persist as "best" can be
4147 // achieved by replacing it with "next" k-th candidate with the probability
4148 // of 1/k. It can be easily shown that by the end of the run, the
4149 // probability for any candidate is converged to 1/n, thus giving the
4150 // uniform distribution among all the candidates.
4151 //
4152 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4153 #define RANDOMIZED_DOMAIN_POW 29
4154 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4155 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4156 bool Compile::randomized_select(int count) {
4157 assert(count > 0, "only positive");
4158 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4159 }