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