1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vectornode.hpp"
71 #include "runtime/arguments.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stubRoutines.hpp"
75 #include "runtime/timer.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/copy.hpp"
78 #include "utilities/macros.hpp"
79 #if INCLUDE_ZGC
80 #include "gc/z/c2/zBarrierSetC2.hpp"
81 #endif
82
83
84 // -------------------- Compile::mach_constant_base_node -----------------------
85 // Constant table base node singleton.
86 MachConstantBaseNode* Compile::mach_constant_base_node() {
87 if (_mach_constant_base_node == NULL) {
88 _mach_constant_base_node = new MachConstantBaseNode();
89 _mach_constant_base_node->add_req(C->root());
90 }
91 return _mach_constant_base_node;
92 }
93
94
95 /// Support for intrinsics.
96
97 // Return the index at which m must be inserted (or already exists).
98 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
99 class IntrinsicDescPair {
100 private:
101 ciMethod* _m;
102 bool _is_virtual;
103 public:
104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
106 ciMethod* m= elt->method();
107 ciMethod* key_m = key->_m;
108 if (key_m < m) return -1;
109 else if (key_m > m) return 1;
110 else {
111 bool is_virtual = elt->is_virtual();
112 bool key_virtual = key->_is_virtual;
113 if (key_virtual < is_virtual) return -1;
114 else if (key_virtual > is_virtual) return 1;
115 else return 0;
116 }
117 }
118 };
119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
120 #ifdef ASSERT
121 for (int i = 1; i < _intrinsics->length(); i++) {
122 CallGenerator* cg1 = _intrinsics->at(i-1);
123 CallGenerator* cg2 = _intrinsics->at(i);
124 assert(cg1->method() != cg2->method()
125 ? cg1->method() < cg2->method()
126 : cg1->is_virtual() < cg2->is_virtual(),
127 "compiler intrinsics list must stay sorted");
128 }
129 #endif
130 IntrinsicDescPair pair(m, is_virtual);
131 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
132 }
133
134 void Compile::register_intrinsic(CallGenerator* cg) {
135 if (_intrinsics == NULL) {
136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
137 }
138 int len = _intrinsics->length();
139 bool found = false;
140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
141 assert(!found, "registering twice");
142 _intrinsics->insert_before(index, cg);
143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
144 }
145
146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
147 assert(m->is_loaded(), "don't try this on unloaded methods");
148 if (_intrinsics != NULL) {
149 bool found = false;
150 int index = intrinsic_insertion_index(m, is_virtual, found);
151 if (found) {
152 return _intrinsics->at(index);
153 }
154 }
155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
156 if (m->intrinsic_id() != vmIntrinsics::_none &&
157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
159 if (cg != NULL) {
160 // Save it for next time:
161 register_intrinsic(cg);
162 return cg;
163 } else {
164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
165 }
166 }
167 return NULL;
168 }
169
170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
171 // in library_call.cpp.
172
173
174 #ifndef PRODUCT
175 // statistics gathering...
176
177 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
179
180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
182 int oflags = _intrinsic_hist_flags[id];
183 assert(flags != 0, "what happened?");
184 if (is_virtual) {
185 flags |= _intrinsic_virtual;
186 }
187 bool changed = (flags != oflags);
188 if ((flags & _intrinsic_worked) != 0) {
189 juint count = (_intrinsic_hist_count[id] += 1);
190 if (count == 1) {
191 changed = true; // first time
192 }
193 // increment the overall count also:
194 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
195 }
196 if (changed) {
197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
198 // Something changed about the intrinsic's virtuality.
199 if ((flags & _intrinsic_virtual) != 0) {
200 // This is the first use of this intrinsic as a virtual call.
201 if (oflags != 0) {
202 // We already saw it as a non-virtual, so note both cases.
203 flags |= _intrinsic_both;
204 }
205 } else if ((oflags & _intrinsic_both) == 0) {
206 // This is the first use of this intrinsic as a non-virtual
207 flags |= _intrinsic_both;
208 }
209 }
210 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
211 }
212 // update the overall flags also:
213 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
214 return changed;
215 }
216
217 static char* format_flags(int flags, char* buf) {
218 buf[0] = 0;
219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
224 if (buf[0] == 0) strcat(buf, ",");
225 assert(buf[0] == ',', "must be");
226 return &buf[1];
227 }
228
229 void Compile::print_intrinsic_statistics() {
230 char flagsbuf[100];
231 ttyLocker ttyl;
232 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
233 tty->print_cr("Compiler intrinsic usage:");
234 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
235 if (total == 0) total = 1; // avoid div0 in case of no successes
236 #define PRINT_STAT_LINE(name, c, f) \
237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
238 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
239 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
240 int flags = _intrinsic_hist_flags[id];
241 juint count = _intrinsic_hist_count[id];
242 if ((flags | count) != 0) {
243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
244 }
245 }
246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
247 if (xtty != NULL) xtty->tail("statistics");
248 }
249
250 void Compile::print_statistics() {
251 { ttyLocker ttyl;
252 if (xtty != NULL) xtty->head("statistics type='opto'");
253 Parse::print_statistics();
254 PhaseCCP::print_statistics();
255 PhaseRegAlloc::print_statistics();
256 Scheduling::print_statistics();
257 PhasePeephole::print_statistics();
258 PhaseIdealLoop::print_statistics();
259 if (xtty != NULL) xtty->tail("statistics");
260 }
261 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
262 // put this under its own <statistics> element.
263 print_intrinsic_statistics();
264 }
265 }
266 #endif //PRODUCT
267
268 // Support for bundling info
269 Bundle* Compile::node_bundling(const Node *n) {
270 assert(valid_bundle_info(n), "oob");
271 return &_node_bundling_base[n->_idx];
272 }
273
274 bool Compile::valid_bundle_info(const Node *n) {
275 return (_node_bundling_limit > n->_idx);
276 }
277
278
279 void Compile::gvn_replace_by(Node* n, Node* nn) {
280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
281 Node* use = n->last_out(i);
282 bool is_in_table = initial_gvn()->hash_delete(use);
283 uint uses_found = 0;
284 for (uint j = 0; j < use->len(); j++) {
285 if (use->in(j) == n) {
286 if (j < use->req())
287 use->set_req(j, nn);
288 else
289 use->set_prec(j, nn);
290 uses_found++;
291 }
292 }
293 if (is_in_table) {
294 // reinsert into table
295 initial_gvn()->hash_find_insert(use);
296 }
297 record_for_igvn(use);
298 i -= uses_found; // we deleted 1 or more copies of this edge
299 }
300 }
301
302
303 static inline bool not_a_node(const Node* n) {
304 if (n == NULL) return true;
305 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
306 if (*(address*)n == badAddress) return true; // kill by Node::destruct
307 return false;
308 }
309
310 // Identify all nodes that are reachable from below, useful.
311 // Use breadth-first pass that records state in a Unique_Node_List,
312 // recursive traversal is slower.
313 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
314 int estimated_worklist_size = live_nodes();
315 useful.map( estimated_worklist_size, NULL ); // preallocate space
316
317 // Initialize worklist
318 if (root() != NULL) { useful.push(root()); }
319 // If 'top' is cached, declare it useful to preserve cached node
320 if( cached_top_node() ) { useful.push(cached_top_node()); }
321
322 // Push all useful nodes onto the list, breadthfirst
323 for( uint next = 0; next < useful.size(); ++next ) {
324 assert( next < unique(), "Unique useful nodes < total nodes");
325 Node *n = useful.at(next);
326 uint max = n->len();
327 for( uint i = 0; i < max; ++i ) {
328 Node *m = n->in(i);
329 if (not_a_node(m)) continue;
330 useful.push(m);
331 }
332 }
333 }
334
335 // Update dead_node_list with any missing dead nodes using useful
336 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
337 void Compile::update_dead_node_list(Unique_Node_List &useful) {
338 uint max_idx = unique();
339 VectorSet& useful_node_set = useful.member_set();
340
341 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
342 // If node with index node_idx is not in useful set,
343 // mark it as dead in dead node list.
344 if (! useful_node_set.test(node_idx) ) {
345 record_dead_node(node_idx);
346 }
347 }
348 }
349
350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
351 int shift = 0;
352 for (int i = 0; i < inlines->length(); i++) {
353 CallGenerator* cg = inlines->at(i);
354 CallNode* call = cg->call_node();
355 if (shift > 0) {
356 inlines->at_put(i-shift, cg);
357 }
358 if (!useful.member(call)) {
359 shift++;
360 }
361 }
362 inlines->trunc_to(inlines->length()-shift);
363 }
364
365 // Disconnect all useless nodes by disconnecting those at the boundary.
366 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
367 uint next = 0;
368 while (next < useful.size()) {
369 Node *n = useful.at(next++);
370 if (n->is_SafePoint()) {
371 // We're done with a parsing phase. Replaced nodes are not valid
372 // beyond that point.
373 n->as_SafePoint()->delete_replaced_nodes();
374 }
375 // Use raw traversal of out edges since this code removes out edges
376 int max = n->outcnt();
377 for (int j = 0; j < max; ++j) {
378 Node* child = n->raw_out(j);
379 if (! useful.member(child)) {
380 assert(!child->is_top() || child != top(),
381 "If top is cached in Compile object it is in useful list");
382 // Only need to remove this out-edge to the useless node
383 n->raw_del_out(j);
384 --j;
385 --max;
386 }
387 }
388 if (n->outcnt() == 1 && n->has_special_unique_user()) {
389 record_for_igvn(n->unique_out());
390 }
391 }
392 // Remove useless macro and predicate opaq nodes
393 for (int i = C->macro_count()-1; i >= 0; i--) {
394 Node* n = C->macro_node(i);
395 if (!useful.member(n)) {
396 remove_macro_node(n);
397 }
398 }
399 // Remove useless CastII nodes with range check dependency
400 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
401 Node* cast = range_check_cast_node(i);
402 if (!useful.member(cast)) {
403 remove_range_check_cast(cast);
404 }
405 }
406 // Remove useless expensive nodes
407 for (int i = C->expensive_count()-1; i >= 0; i--) {
408 Node* n = C->expensive_node(i);
409 if (!useful.member(n)) {
410 remove_expensive_node(n);
411 }
412 }
413 // Remove useless Opaque4 nodes
414 for (int i = opaque4_count() - 1; i >= 0; i--) {
415 Node* opaq = opaque4_node(i);
416 if (!useful.member(opaq)) {
417 remove_opaque4_node(opaq);
418 }
419 }
420 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
421 bs->eliminate_useless_gc_barriers(useful, this);
422 // clean up the late inline lists
423 remove_useless_late_inlines(&_string_late_inlines, useful);
424 remove_useless_late_inlines(&_boxing_late_inlines, useful);
425 remove_useless_late_inlines(&_late_inlines, useful);
426 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
427 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
428 }
429
430 //------------------------------frame_size_in_words-----------------------------
431 // frame_slots in units of words
432 int Compile::frame_size_in_words() const {
433 // shift is 0 in LP32 and 1 in LP64
434 const int shift = (LogBytesPerWord - LogBytesPerInt);
435 int words = _frame_slots >> shift;
436 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
437 return words;
438 }
439
440 // To bang the stack of this compiled method we use the stack size
441 // that the interpreter would need in case of a deoptimization. This
442 // removes the need to bang the stack in the deoptimization blob which
443 // in turn simplifies stack overflow handling.
444 int Compile::bang_size_in_bytes() const {
445 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
446 }
447
448 // ============================================================================
449 //------------------------------CompileWrapper---------------------------------
450 class CompileWrapper : public StackObj {
451 Compile *const _compile;
452 public:
453 CompileWrapper(Compile* compile);
454
455 ~CompileWrapper();
456 };
457
458 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
459 // the Compile* pointer is stored in the current ciEnv:
460 ciEnv* env = compile->env();
461 assert(env == ciEnv::current(), "must already be a ciEnv active");
462 assert(env->compiler_data() == NULL, "compile already active?");
463 env->set_compiler_data(compile);
464 assert(compile == Compile::current(), "sanity");
465
466 compile->set_type_dict(NULL);
467 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
468 compile->clone_map().set_clone_idx(0);
469 compile->set_type_hwm(NULL);
470 compile->set_type_last_size(0);
471 compile->set_last_tf(NULL, NULL);
472 compile->set_indexSet_arena(NULL);
473 compile->set_indexSet_free_block_list(NULL);
474 compile->init_type_arena();
475 Type::Initialize(compile);
476 _compile->set_scratch_buffer_blob(NULL);
477 _compile->begin_method();
478 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
479 }
480 CompileWrapper::~CompileWrapper() {
481 _compile->end_method();
482 if (_compile->scratch_buffer_blob() != NULL)
483 BufferBlob::free(_compile->scratch_buffer_blob());
484 _compile->env()->set_compiler_data(NULL);
485 }
486
487
488 //----------------------------print_compile_messages---------------------------
489 void Compile::print_compile_messages() {
490 #ifndef PRODUCT
491 // Check if recompiling
492 if (_subsume_loads == false && PrintOpto) {
493 // Recompiling without allowing machine instructions to subsume loads
494 tty->print_cr("*********************************************************");
495 tty->print_cr("** Bailout: Recompile without subsuming loads **");
496 tty->print_cr("*********************************************************");
497 }
498 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
499 // Recompiling without escape analysis
500 tty->print_cr("*********************************************************");
501 tty->print_cr("** Bailout: Recompile without escape analysis **");
502 tty->print_cr("*********************************************************");
503 }
504 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
505 // Recompiling without boxing elimination
506 tty->print_cr("*********************************************************");
507 tty->print_cr("** Bailout: Recompile without boxing elimination **");
508 tty->print_cr("*********************************************************");
509 }
510 if (C->directive()->BreakAtCompileOption) {
511 // Open the debugger when compiling this method.
512 tty->print("### Breaking when compiling: ");
513 method()->print_short_name();
514 tty->cr();
515 BREAKPOINT;
516 }
517
518 if( PrintOpto ) {
519 if (is_osr_compilation()) {
520 tty->print("[OSR]%3d", _compile_id);
521 } else {
522 tty->print("%3d", _compile_id);
523 }
524 }
525 #endif
526 }
527
528
529 //-----------------------init_scratch_buffer_blob------------------------------
530 // Construct a temporary BufferBlob and cache it for this compile.
531 void Compile::init_scratch_buffer_blob(int const_size) {
532 // If there is already a scratch buffer blob allocated and the
533 // constant section is big enough, use it. Otherwise free the
534 // current and allocate a new one.
535 BufferBlob* blob = scratch_buffer_blob();
536 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
537 // Use the current blob.
538 } else {
539 if (blob != NULL) {
540 BufferBlob::free(blob);
541 }
542
543 ResourceMark rm;
544 _scratch_const_size = const_size;
545 int size = C2Compiler::initial_code_buffer_size(const_size);
546 blob = BufferBlob::create("Compile::scratch_buffer", size);
547 // Record the buffer blob for next time.
548 set_scratch_buffer_blob(blob);
549 // Have we run out of code space?
550 if (scratch_buffer_blob() == NULL) {
551 // Let CompilerBroker disable further compilations.
552 record_failure("Not enough space for scratch buffer in CodeCache");
553 return;
554 }
555 }
556
557 // Initialize the relocation buffers
558 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
559 set_scratch_locs_memory(locs_buf);
560 }
561
562
563 //-----------------------scratch_emit_size-------------------------------------
564 // Helper function that computes size by emitting code
565 uint Compile::scratch_emit_size(const Node* n) {
566 // Start scratch_emit_size section.
567 set_in_scratch_emit_size(true);
568
569 // Emit into a trash buffer and count bytes emitted.
570 // This is a pretty expensive way to compute a size,
571 // but it works well enough if seldom used.
572 // All common fixed-size instructions are given a size
573 // method by the AD file.
574 // Note that the scratch buffer blob and locs memory are
575 // allocated at the beginning of the compile task, and
576 // may be shared by several calls to scratch_emit_size.
577 // The allocation of the scratch buffer blob is particularly
578 // expensive, since it has to grab the code cache lock.
579 BufferBlob* blob = this->scratch_buffer_blob();
580 assert(blob != NULL, "Initialize BufferBlob at start");
581 assert(blob->size() > MAX_inst_size, "sanity");
582 relocInfo* locs_buf = scratch_locs_memory();
583 address blob_begin = blob->content_begin();
584 address blob_end = (address)locs_buf;
585 assert(blob->contains(blob_end), "sanity");
586 CodeBuffer buf(blob_begin, blob_end - blob_begin);
587 buf.initialize_consts_size(_scratch_const_size);
588 buf.initialize_stubs_size(MAX_stubs_size);
589 assert(locs_buf != NULL, "sanity");
590 int lsize = MAX_locs_size / 3;
591 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
592 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
593 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
594 // Mark as scratch buffer.
595 buf.consts()->set_scratch_emit();
596 buf.insts()->set_scratch_emit();
597 buf.stubs()->set_scratch_emit();
598
599 // Do the emission.
600
601 Label fakeL; // Fake label for branch instructions.
602 Label* saveL = NULL;
603 uint save_bnum = 0;
604 bool is_branch = n->is_MachBranch();
605 if (is_branch) {
606 MacroAssembler masm(&buf);
607 masm.bind(fakeL);
608 n->as_MachBranch()->save_label(&saveL, &save_bnum);
609 n->as_MachBranch()->label_set(&fakeL, 0);
610 }
611 n->emit(buf, this->regalloc());
612
613 // Emitting into the scratch buffer should not fail
614 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
615
616 if (is_branch) // Restore label.
617 n->as_MachBranch()->label_set(saveL, save_bnum);
618
619 // End scratch_emit_size section.
620 set_in_scratch_emit_size(false);
621
622 return buf.insts_size();
623 }
624
625
626 // ============================================================================
627 //------------------------------Compile standard-------------------------------
628 debug_only( int Compile::_debug_idx = 100000; )
629
630 // Compile a method. entry_bci is -1 for normal compilations and indicates
631 // the continuation bci for on stack replacement.
632
633
634 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
635 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
636 : Phase(Compiler),
637 _compile_id(ci_env->compile_id()),
638 _save_argument_registers(false),
639 _subsume_loads(subsume_loads),
640 _do_escape_analysis(do_escape_analysis),
641 _eliminate_boxing(eliminate_boxing),
642 _method(target),
643 _entry_bci(osr_bci),
644 _stub_function(NULL),
645 _stub_name(NULL),
646 _stub_entry_point(NULL),
647 _max_node_limit(MaxNodeLimit),
648 _orig_pc_slot(0),
649 _orig_pc_slot_offset_in_bytes(0),
650 _inlining_progress(false),
651 _inlining_incrementally(false),
652 _do_cleanup(false),
653 _has_reserved_stack_access(target->has_reserved_stack_access()),
654 #ifndef PRODUCT
655 _trace_opto_output(directive->TraceOptoOutputOption),
656 #endif
657 _has_method_handle_invokes(false),
658 _comp_arena(mtCompiler),
659 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
660 _env(ci_env),
661 _directive(directive),
662 _log(ci_env->log()),
663 _failure_reason(NULL),
664 _congraph(NULL),
665 #ifndef PRODUCT
666 _printer(IdealGraphPrinter::printer()),
667 #endif
668 _dead_node_list(comp_arena()),
669 _dead_node_count(0),
670 _node_arena(mtCompiler),
671 _old_arena(mtCompiler),
672 _mach_constant_base_node(NULL),
673 _Compile_types(mtCompiler),
674 _initial_gvn(NULL),
675 _for_igvn(NULL),
676 _warm_calls(NULL),
677 _late_inlines(comp_arena(), 2, 0, NULL),
678 _string_late_inlines(comp_arena(), 2, 0, NULL),
679 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
680 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
681 _late_inlines_pos(0),
682 _number_of_mh_late_inlines(0),
683 _print_inlining_stream(NULL),
684 _print_inlining_list(NULL),
685 _print_inlining_idx(0),
686 _print_inlining_output(NULL),
687 _replay_inline_data(NULL),
688 _java_calls(0),
689 _inner_loops(0),
690 _interpreter_frame_size(0),
691 _node_bundling_limit(0),
692 _node_bundling_base(NULL),
693 _code_buffer("Compile::Fill_buffer"),
694 _scratch_const_size(-1),
695 _in_scratch_emit_size(false)
696 #ifndef PRODUCT
697 , _in_dump_cnt(0)
698 #endif
699 {
700 C = this;
701 #ifndef PRODUCT
702 if (_printer != NULL) {
703 _printer->set_compile(this);
704 }
705 #endif
706 CompileWrapper cw(this);
707
708 if (CITimeVerbose) {
709 tty->print(" ");
710 target->holder()->name()->print();
711 tty->print(".");
712 target->print_short_name();
713 tty->print(" ");
714 }
715 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
716 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
717
718 #ifndef PRODUCT
719 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
720 if (!print_opto_assembly) {
721 bool print_assembly = directive->PrintAssemblyOption;
722 if (print_assembly && !Disassembler::can_decode()) {
723 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
724 print_opto_assembly = true;
725 }
726 }
727 set_print_assembly(print_opto_assembly);
728 set_parsed_irreducible_loop(false);
729
730 if (directive->ReplayInlineOption) {
731 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
732 }
733 #endif
734 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
735 set_print_intrinsics(directive->PrintIntrinsicsOption);
736 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
737
738 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
739 // Make sure the method being compiled gets its own MDO,
740 // so we can at least track the decompile_count().
741 // Need MDO to record RTM code generation state.
742 method()->ensure_method_data();
743 }
744
745 Init(::AliasLevel);
746
747
748 print_compile_messages();
749
750 _ilt = InlineTree::build_inline_tree_root();
751
752 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
753 assert(num_alias_types() >= AliasIdxRaw, "");
754
755 #define MINIMUM_NODE_HASH 1023
756 // Node list that Iterative GVN will start with
757 Unique_Node_List for_igvn(comp_arena());
758 set_for_igvn(&for_igvn);
759
760 // GVN that will be run immediately on new nodes
761 uint estimated_size = method()->code_size()*4+64;
762 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
763 PhaseGVN gvn(node_arena(), estimated_size);
764 set_initial_gvn(&gvn);
765
766 print_inlining_init();
767 { // Scope for timing the parser
768 TracePhase tp("parse", &timers[_t_parser]);
769
770 // Put top into the hash table ASAP.
771 initial_gvn()->transform_no_reclaim(top());
772
773 // Set up tf(), start(), and find a CallGenerator.
774 CallGenerator* cg = NULL;
775 if (is_osr_compilation()) {
776 const TypeTuple *domain = StartOSRNode::osr_domain();
777 const TypeTuple *range = TypeTuple::make_range(method()->signature());
778 init_tf(TypeFunc::make(domain, range));
779 StartNode* s = new StartOSRNode(root(), domain);
780 initial_gvn()->set_type_bottom(s);
781 init_start(s);
782 cg = CallGenerator::for_osr(method(), entry_bci());
783 } else {
784 // Normal case.
785 init_tf(TypeFunc::make(method()));
786 StartNode* s = new StartNode(root(), tf()->domain());
787 initial_gvn()->set_type_bottom(s);
788 init_start(s);
789 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
790 // With java.lang.ref.reference.get() we must go through the
791 // intrinsic - even when get() is the root
792 // method of the compile - so that, if necessary, the value in
793 // the referent field of the reference object gets recorded by
794 // the pre-barrier code.
795 cg = find_intrinsic(method(), false);
796 }
797 if (cg == NULL) {
798 float past_uses = method()->interpreter_invocation_count();
799 float expected_uses = past_uses;
800 cg = CallGenerator::for_inline(method(), expected_uses);
801 }
802 }
803 if (failing()) return;
804 if (cg == NULL) {
805 record_method_not_compilable("cannot parse method");
806 return;
807 }
808 JVMState* jvms = build_start_state(start(), tf());
809 if ((jvms = cg->generate(jvms)) == NULL) {
810 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
811 record_method_not_compilable("method parse failed");
812 }
813 return;
814 }
815 GraphKit kit(jvms);
816
817 if (!kit.stopped()) {
818 // Accept return values, and transfer control we know not where.
819 // This is done by a special, unique ReturnNode bound to root.
820 return_values(kit.jvms());
821 }
822
823 if (kit.has_exceptions()) {
824 // Any exceptions that escape from this call must be rethrown
825 // to whatever caller is dynamically above us on the stack.
826 // This is done by a special, unique RethrowNode bound to root.
827 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
828 }
829
830 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
831
832 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
833 inline_string_calls(true);
834 }
835
836 if (failing()) return;
837
838 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
839
840 // Remove clutter produced by parsing.
841 if (!failing()) {
842 ResourceMark rm;
843 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
844 }
845 }
846
847 // Note: Large methods are capped off in do_one_bytecode().
848 if (failing()) return;
849
850 // After parsing, node notes are no longer automagic.
851 // They must be propagated by register_new_node_with_optimizer(),
852 // clone(), or the like.
853 set_default_node_notes(NULL);
854
855 for (;;) {
856 int successes = Inline_Warm();
857 if (failing()) return;
858 if (successes == 0) break;
859 }
860
861 // Drain the list.
862 Finish_Warm();
863 #ifndef PRODUCT
864 if (_printer && _printer->should_print(1)) {
865 _printer->print_inlining();
866 }
867 #endif
868
869 if (failing()) return;
870 NOT_PRODUCT( verify_graph_edges(); )
871
872 // Now optimize
873 Optimize();
874 if (failing()) return;
875 NOT_PRODUCT( verify_graph_edges(); )
876
877 #ifndef PRODUCT
878 if (PrintIdeal) {
879 ttyLocker ttyl; // keep the following output all in one block
880 // This output goes directly to the tty, not the compiler log.
881 // To enable tools to match it up with the compilation activity,
882 // be sure to tag this tty output with the compile ID.
883 if (xtty != NULL) {
884 xtty->head("ideal compile_id='%d'%s", compile_id(),
885 is_osr_compilation() ? " compile_kind='osr'" :
886 "");
887 }
888 root()->dump(9999);
889 if (xtty != NULL) {
890 xtty->tail("ideal");
891 }
892 }
893 #endif
894
895 #ifdef ASSERT
896 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
897 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
898 #endif
899
900 // Dump compilation data to replay it.
901 if (directive->DumpReplayOption) {
902 env()->dump_replay_data(_compile_id);
903 }
904 if (directive->DumpInlineOption && (ilt() != NULL)) {
905 env()->dump_inline_data(_compile_id);
906 }
907
908 // Now that we know the size of all the monitors we can add a fixed slot
909 // for the original deopt pc.
910
911 _orig_pc_slot = fixed_slots();
912 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
913 set_fixed_slots(next_slot);
914
915 // Compute when to use implicit null checks. Used by matching trap based
916 // nodes and NullCheck optimization.
917 set_allowed_deopt_reasons();
918
919 // Now generate code
920 Code_Gen();
921 if (failing()) return;
922
923 // Check if we want to skip execution of all compiled code.
924 {
925 #ifndef PRODUCT
926 if (OptoNoExecute) {
927 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
928 return;
929 }
930 #endif
931 TracePhase tp("install_code", &timers[_t_registerMethod]);
932
933 if (is_osr_compilation()) {
934 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
935 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
936 } else {
937 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
938 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
939 }
940
941 env()->register_method(_method, _entry_bci,
942 &_code_offsets,
943 _orig_pc_slot_offset_in_bytes,
944 code_buffer(),
945 frame_size_in_words(), _oop_map_set,
946 &_handler_table, &_inc_table,
947 compiler,
948 has_unsafe_access(),
949 SharedRuntime::is_wide_vector(max_vector_size()),
950 rtm_state()
951 );
952
953 if (log() != NULL) // Print code cache state into compiler log
954 log()->code_cache_state();
955 }
956 }
957
958 //------------------------------Compile----------------------------------------
959 // Compile a runtime stub
960 Compile::Compile( ciEnv* ci_env,
961 TypeFunc_generator generator,
962 address stub_function,
963 const char *stub_name,
964 int is_fancy_jump,
965 bool pass_tls,
966 bool save_arg_registers,
967 bool return_pc,
968 DirectiveSet* directive)
969 : Phase(Compiler),
970 _compile_id(0),
971 _save_argument_registers(save_arg_registers),
972 _subsume_loads(true),
973 _do_escape_analysis(false),
974 _eliminate_boxing(false),
975 _method(NULL),
976 _entry_bci(InvocationEntryBci),
977 _stub_function(stub_function),
978 _stub_name(stub_name),
979 _stub_entry_point(NULL),
980 _max_node_limit(MaxNodeLimit),
981 _orig_pc_slot(0),
982 _orig_pc_slot_offset_in_bytes(0),
983 _inlining_progress(false),
984 _inlining_incrementally(false),
985 _has_reserved_stack_access(false),
986 #ifndef PRODUCT
987 _trace_opto_output(directive->TraceOptoOutputOption),
988 #endif
989 _has_method_handle_invokes(false),
990 _comp_arena(mtCompiler),
991 _env(ci_env),
992 _directive(directive),
993 _log(ci_env->log()),
994 _failure_reason(NULL),
995 _congraph(NULL),
996 #ifndef PRODUCT
997 _printer(NULL),
998 #endif
999 _dead_node_list(comp_arena()),
1000 _dead_node_count(0),
1001 _node_arena(mtCompiler),
1002 _old_arena(mtCompiler),
1003 _mach_constant_base_node(NULL),
1004 _Compile_types(mtCompiler),
1005 _initial_gvn(NULL),
1006 _for_igvn(NULL),
1007 _warm_calls(NULL),
1008 _number_of_mh_late_inlines(0),
1009 _print_inlining_stream(NULL),
1010 _print_inlining_list(NULL),
1011 _print_inlining_idx(0),
1012 _print_inlining_output(NULL),
1013 _replay_inline_data(NULL),
1014 _java_calls(0),
1015 _inner_loops(0),
1016 _interpreter_frame_size(0),
1017 _node_bundling_limit(0),
1018 _node_bundling_base(NULL),
1019 _code_buffer("Compile::Fill_buffer"),
1020 #ifndef PRODUCT
1021 _in_dump_cnt(0),
1022 #endif
1023 _allowed_reasons(0) {
1024 C = this;
1025
1026 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1027 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1028
1029 #ifndef PRODUCT
1030 set_print_assembly(PrintFrameConverterAssembly);
1031 set_parsed_irreducible_loop(false);
1032 #endif
1033 set_has_irreducible_loop(false); // no loops
1034
1035 CompileWrapper cw(this);
1036 Init(/*AliasLevel=*/ 0);
1037 init_tf((*generator)());
1038
1039 {
1040 // The following is a dummy for the sake of GraphKit::gen_stub
1041 Unique_Node_List for_igvn(comp_arena());
1042 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1043 PhaseGVN gvn(Thread::current()->resource_area(),255);
1044 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1045 gvn.transform_no_reclaim(top());
1046
1047 GraphKit kit;
1048 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1049 }
1050
1051 NOT_PRODUCT( verify_graph_edges(); )
1052 Code_Gen();
1053 if (failing()) return;
1054
1055
1056 // Entry point will be accessed using compile->stub_entry_point();
1057 if (code_buffer() == NULL) {
1058 Matcher::soft_match_failure();
1059 } else {
1060 if (PrintAssembly && (WizardMode || Verbose))
1061 tty->print_cr("### Stub::%s", stub_name);
1062
1063 if (!failing()) {
1064 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1065
1066 // Make the NMethod
1067 // For now we mark the frame as never safe for profile stackwalking
1068 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1069 code_buffer(),
1070 CodeOffsets::frame_never_safe,
1071 // _code_offsets.value(CodeOffsets::Frame_Complete),
1072 frame_size_in_words(),
1073 _oop_map_set,
1074 save_arg_registers);
1075 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1076
1077 _stub_entry_point = rs->entry_point();
1078 }
1079 }
1080 }
1081
1082 //------------------------------Init-------------------------------------------
1083 // Prepare for a single compilation
1084 void Compile::Init(int aliaslevel) {
1085 _unique = 0;
1086 _regalloc = NULL;
1087
1088 _tf = NULL; // filled in later
1089 _top = NULL; // cached later
1090 _matcher = NULL; // filled in later
1091 _cfg = NULL; // filled in later
1092
1093 set_24_bit_selection_and_mode(Use24BitFP, false);
1094
1095 _node_note_array = NULL;
1096 _default_node_notes = NULL;
1097 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1098
1099 _immutable_memory = NULL; // filled in at first inquiry
1100
1101 // Globally visible Nodes
1102 // First set TOP to NULL to give safe behavior during creation of RootNode
1103 set_cached_top_node(NULL);
1104 set_root(new RootNode());
1105 // Now that you have a Root to point to, create the real TOP
1106 set_cached_top_node( new ConNode(Type::TOP) );
1107 set_recent_alloc(NULL, NULL);
1108
1109 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1110 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1111 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1112 env()->set_dependencies(new Dependencies(env()));
1113
1114 _fixed_slots = 0;
1115 set_has_split_ifs(false);
1116 set_has_loops(has_method() && method()->has_loops()); // first approximation
1117 set_has_stringbuilder(false);
1118 set_has_boxed_value(false);
1119 _trap_can_recompile = false; // no traps emitted yet
1120 _major_progress = true; // start out assuming good things will happen
1121 set_has_unsafe_access(false);
1122 set_max_vector_size(0);
1123 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1124 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1125 set_decompile_count(0);
1126
1127 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1128 _loop_opts_cnt = LoopOptsCount;
1129 set_do_inlining(Inline);
1130 set_max_inline_size(MaxInlineSize);
1131 set_freq_inline_size(FreqInlineSize);
1132 set_do_scheduling(OptoScheduling);
1133 set_do_count_invocations(false);
1134 set_do_method_data_update(false);
1135
1136 set_do_vector_loop(false);
1137
1138 if (AllowVectorizeOnDemand) {
1139 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1140 set_do_vector_loop(true);
1141 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1142 } else if (has_method() && method()->name() != 0 &&
1143 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1144 set_do_vector_loop(true);
1145 }
1146 }
1147 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1148 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1149
1150 set_age_code(has_method() && method()->profile_aging());
1151 set_rtm_state(NoRTM); // No RTM lock eliding by default
1152 _max_node_limit = _directive->MaxNodeLimitOption;
1153
1154 #if INCLUDE_RTM_OPT
1155 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1156 int rtm_state = method()->method_data()->rtm_state();
1157 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1158 // Don't generate RTM lock eliding code.
1159 set_rtm_state(NoRTM);
1160 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1161 // Generate RTM lock eliding code without abort ratio calculation code.
1162 set_rtm_state(UseRTM);
1163 } else if (UseRTMDeopt) {
1164 // Generate RTM lock eliding code and include abort ratio calculation
1165 // code if UseRTMDeopt is on.
1166 set_rtm_state(ProfileRTM);
1167 }
1168 }
1169 #endif
1170 if (debug_info()->recording_non_safepoints()) {
1171 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1172 (comp_arena(), 8, 0, NULL));
1173 set_default_node_notes(Node_Notes::make(this));
1174 }
1175
1176 // // -- Initialize types before each compile --
1177 // // Update cached type information
1178 // if( _method && _method->constants() )
1179 // Type::update_loaded_types(_method, _method->constants());
1180
1181 // Init alias_type map.
1182 if (!_do_escape_analysis && aliaslevel == 3)
1183 aliaslevel = 2; // No unique types without escape analysis
1184 _AliasLevel = aliaslevel;
1185 const int grow_ats = 16;
1186 _max_alias_types = grow_ats;
1187 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1188 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1189 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1190 {
1191 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1192 }
1193 // Initialize the first few types.
1194 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1195 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1196 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1197 _num_alias_types = AliasIdxRaw+1;
1198 // Zero out the alias type cache.
1199 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1200 // A NULL adr_type hits in the cache right away. Preload the right answer.
1201 probe_alias_cache(NULL)->_index = AliasIdxTop;
1202
1203 _intrinsics = NULL;
1204 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1205 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1206 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1207 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1208 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1209 register_library_intrinsics();
1210 }
1211
1212 //---------------------------init_start----------------------------------------
1213 // Install the StartNode on this compile object.
1214 void Compile::init_start(StartNode* s) {
1215 if (failing())
1216 return; // already failing
1217 assert(s == start(), "");
1218 }
1219
1220 /**
1221 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1222 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1223 * the ideal graph.
1224 */
1225 StartNode* Compile::start() const {
1226 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1227 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1228 Node* start = root()->fast_out(i);
1229 if (start->is_Start()) {
1230 return start->as_Start();
1231 }
1232 }
1233 fatal("Did not find Start node!");
1234 return NULL;
1235 }
1236
1237 //-------------------------------immutable_memory-------------------------------------
1238 // Access immutable memory
1239 Node* Compile::immutable_memory() {
1240 if (_immutable_memory != NULL) {
1241 return _immutable_memory;
1242 }
1243 StartNode* s = start();
1244 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1245 Node *p = s->fast_out(i);
1246 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1247 _immutable_memory = p;
1248 return _immutable_memory;
1249 }
1250 }
1251 ShouldNotReachHere();
1252 return NULL;
1253 }
1254
1255 //----------------------set_cached_top_node------------------------------------
1256 // Install the cached top node, and make sure Node::is_top works correctly.
1257 void Compile::set_cached_top_node(Node* tn) {
1258 if (tn != NULL) verify_top(tn);
1259 Node* old_top = _top;
1260 _top = tn;
1261 // Calling Node::setup_is_top allows the nodes the chance to adjust
1262 // their _out arrays.
1263 if (_top != NULL) _top->setup_is_top();
1264 if (old_top != NULL) old_top->setup_is_top();
1265 assert(_top == NULL || top()->is_top(), "");
1266 }
1267
1268 #ifdef ASSERT
1269 uint Compile::count_live_nodes_by_graph_walk() {
1270 Unique_Node_List useful(comp_arena());
1271 // Get useful node list by walking the graph.
1272 identify_useful_nodes(useful);
1273 return useful.size();
1274 }
1275
1276 void Compile::print_missing_nodes() {
1277
1278 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1279 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1280 return;
1281 }
1282
1283 // This is an expensive function. It is executed only when the user
1284 // specifies VerifyIdealNodeCount option or otherwise knows the
1285 // additional work that needs to be done to identify reachable nodes
1286 // by walking the flow graph and find the missing ones using
1287 // _dead_node_list.
1288
1289 Unique_Node_List useful(comp_arena());
1290 // Get useful node list by walking the graph.
1291 identify_useful_nodes(useful);
1292
1293 uint l_nodes = C->live_nodes();
1294 uint l_nodes_by_walk = useful.size();
1295
1296 if (l_nodes != l_nodes_by_walk) {
1297 if (_log != NULL) {
1298 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1299 _log->stamp();
1300 _log->end_head();
1301 }
1302 VectorSet& useful_member_set = useful.member_set();
1303 int last_idx = l_nodes_by_walk;
1304 for (int i = 0; i < last_idx; i++) {
1305 if (useful_member_set.test(i)) {
1306 if (_dead_node_list.test(i)) {
1307 if (_log != NULL) {
1308 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1309 }
1310 if (PrintIdealNodeCount) {
1311 // Print the log message to tty
1312 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1313 useful.at(i)->dump();
1314 }
1315 }
1316 }
1317 else if (! _dead_node_list.test(i)) {
1318 if (_log != NULL) {
1319 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1320 }
1321 if (PrintIdealNodeCount) {
1322 // Print the log message to tty
1323 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1324 }
1325 }
1326 }
1327 if (_log != NULL) {
1328 _log->tail("mismatched_nodes");
1329 }
1330 }
1331 }
1332 void Compile::record_modified_node(Node* n) {
1333 if (_modified_nodes != NULL && !_inlining_incrementally &&
1334 n->outcnt() != 0 && !n->is_Con()) {
1335 _modified_nodes->push(n);
1336 }
1337 }
1338
1339 void Compile::remove_modified_node(Node* n) {
1340 if (_modified_nodes != NULL) {
1341 _modified_nodes->remove(n);
1342 }
1343 }
1344 #endif
1345
1346 #ifndef PRODUCT
1347 void Compile::verify_top(Node* tn) const {
1348 if (tn != NULL) {
1349 assert(tn->is_Con(), "top node must be a constant");
1350 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1351 assert(tn->in(0) != NULL, "must have live top node");
1352 }
1353 }
1354 #endif
1355
1356
1357 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1358
1359 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1360 guarantee(arr != NULL, "");
1361 int num_blocks = arr->length();
1362 if (grow_by < num_blocks) grow_by = num_blocks;
1363 int num_notes = grow_by * _node_notes_block_size;
1364 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1365 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1366 while (num_notes > 0) {
1367 arr->append(notes);
1368 notes += _node_notes_block_size;
1369 num_notes -= _node_notes_block_size;
1370 }
1371 assert(num_notes == 0, "exact multiple, please");
1372 }
1373
1374 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1375 if (source == NULL || dest == NULL) return false;
1376
1377 if (dest->is_Con())
1378 return false; // Do not push debug info onto constants.
1379
1380 #ifdef ASSERT
1381 // Leave a bread crumb trail pointing to the original node:
1382 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1383 dest->set_debug_orig(source);
1384 }
1385 #endif
1386
1387 if (node_note_array() == NULL)
1388 return false; // Not collecting any notes now.
1389
1390 // This is a copy onto a pre-existing node, which may already have notes.
1391 // If both nodes have notes, do not overwrite any pre-existing notes.
1392 Node_Notes* source_notes = node_notes_at(source->_idx);
1393 if (source_notes == NULL || source_notes->is_clear()) return false;
1394 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1395 if (dest_notes == NULL || dest_notes->is_clear()) {
1396 return set_node_notes_at(dest->_idx, source_notes);
1397 }
1398
1399 Node_Notes merged_notes = (*source_notes);
1400 // The order of operations here ensures that dest notes will win...
1401 merged_notes.update_from(dest_notes);
1402 return set_node_notes_at(dest->_idx, &merged_notes);
1403 }
1404
1405
1406 //--------------------------allow_range_check_smearing-------------------------
1407 // Gating condition for coalescing similar range checks.
1408 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1409 // single covering check that is at least as strong as any of them.
1410 // If the optimization succeeds, the simplified (strengthened) range check
1411 // will always succeed. If it fails, we will deopt, and then give up
1412 // on the optimization.
1413 bool Compile::allow_range_check_smearing() const {
1414 // If this method has already thrown a range-check,
1415 // assume it was because we already tried range smearing
1416 // and it failed.
1417 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1418 return !already_trapped;
1419 }
1420
1421
1422 //------------------------------flatten_alias_type-----------------------------
1423 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1424 int offset = tj->offset();
1425 TypePtr::PTR ptr = tj->ptr();
1426
1427 // Known instance (scalarizable allocation) alias only with itself.
1428 bool is_known_inst = tj->isa_oopptr() != NULL &&
1429 tj->is_oopptr()->is_known_instance();
1430
1431 // Process weird unsafe references.
1432 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1433 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1434 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1435 tj = TypeOopPtr::BOTTOM;
1436 ptr = tj->ptr();
1437 offset = tj->offset();
1438 }
1439
1440 // Array pointers need some flattening
1441 const TypeAryPtr *ta = tj->isa_aryptr();
1442 if (ta && ta->is_stable()) {
1443 // Erase stability property for alias analysis.
1444 tj = ta = ta->cast_to_stable(false);
1445 }
1446 if( ta && is_known_inst ) {
1447 if ( offset != Type::OffsetBot &&
1448 offset > arrayOopDesc::length_offset_in_bytes() ) {
1449 offset = Type::OffsetBot; // Flatten constant access into array body only
1450 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1451 }
1452 } else if( ta && _AliasLevel >= 2 ) {
1453 // For arrays indexed by constant indices, we flatten the alias
1454 // space to include all of the array body. Only the header, klass
1455 // and array length can be accessed un-aliased.
1456 if( offset != Type::OffsetBot ) {
1457 if( ta->const_oop() ) { // MethodData* or Method*
1458 offset = Type::OffsetBot; // Flatten constant access into array body
1459 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1460 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1461 // range is OK as-is.
1462 tj = ta = TypeAryPtr::RANGE;
1463 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1464 tj = TypeInstPtr::KLASS; // all klass loads look alike
1465 ta = TypeAryPtr::RANGE; // generic ignored junk
1466 ptr = TypePtr::BotPTR;
1467 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1468 tj = TypeInstPtr::MARK;
1469 ta = TypeAryPtr::RANGE; // generic ignored junk
1470 ptr = TypePtr::BotPTR;
1471 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1472 ta = tj->isa_aryptr();
1473 } else { // Random constant offset into array body
1474 offset = Type::OffsetBot; // Flatten constant access into array body
1475 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1476 }
1477 }
1478 // Arrays of fixed size alias with arrays of unknown size.
1479 if (ta->size() != TypeInt::POS) {
1480 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1481 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1482 }
1483 // Arrays of known objects become arrays of unknown objects.
1484 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1485 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1486 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1487 }
1488 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1489 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1490 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1491 }
1492 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1493 // cannot be distinguished by bytecode alone.
1494 if (ta->elem() == TypeInt::BOOL) {
1495 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1496 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1497 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1498 }
1499 // During the 2nd round of IterGVN, NotNull castings are removed.
1500 // Make sure the Bottom and NotNull variants alias the same.
1501 // Also, make sure exact and non-exact variants alias the same.
1502 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1503 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1504 }
1505 }
1506
1507 // Oop pointers need some flattening
1508 const TypeInstPtr *to = tj->isa_instptr();
1509 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1510 ciInstanceKlass *k = to->klass()->as_instance_klass();
1511 if( ptr == TypePtr::Constant ) {
1512 if (to->klass() != ciEnv::current()->Class_klass() ||
1513 offset < k->size_helper() * wordSize) {
1514 // No constant oop pointers (such as Strings); they alias with
1515 // unknown strings.
1516 assert(!is_known_inst, "not scalarizable allocation");
1517 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1518 }
1519 } else if( is_known_inst ) {
1520 tj = to; // Keep NotNull and klass_is_exact for instance type
1521 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1522 // During the 2nd round of IterGVN, NotNull castings are removed.
1523 // Make sure the Bottom and NotNull variants alias the same.
1524 // Also, make sure exact and non-exact variants alias the same.
1525 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1526 }
1527 if (to->speculative() != NULL) {
1528 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1529 }
1530 // Canonicalize the holder of this field
1531 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1532 // First handle header references such as a LoadKlassNode, even if the
1533 // object's klass is unloaded at compile time (4965979).
1534 if (!is_known_inst) { // Do it only for non-instance types
1535 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1536 }
1537 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1538 to = tj->is_instptr();
1539 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1540 // Static fields are in the space above the normal instance
1541 // fields in the java.lang.Class instance.
1542 if (to->klass() != ciEnv::current()->Class_klass()) {
1543 to = NULL;
1544 tj = TypeOopPtr::BOTTOM;
1545 offset = tj->offset();
1546 }
1547 } else {
1548 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1549 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1550 if( is_known_inst ) {
1551 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1552 } else {
1553 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1554 }
1555 }
1556 }
1557 }
1558
1559 // Klass pointers to object array klasses need some flattening
1560 const TypeKlassPtr *tk = tj->isa_klassptr();
1561 if( tk ) {
1562 // If we are referencing a field within a Klass, we need
1563 // to assume the worst case of an Object. Both exact and
1564 // inexact types must flatten to the same alias class so
1565 // use NotNull as the PTR.
1566 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1567
1568 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1569 TypeKlassPtr::OBJECT->klass(),
1570 offset);
1571 }
1572
1573 ciKlass* klass = tk->klass();
1574 if( klass->is_obj_array_klass() ) {
1575 ciKlass* k = TypeAryPtr::OOPS->klass();
1576 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1577 k = TypeInstPtr::BOTTOM->klass();
1578 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1579 }
1580
1581 // Check for precise loads from the primary supertype array and force them
1582 // to the supertype cache alias index. Check for generic array loads from
1583 // the primary supertype array and also force them to the supertype cache
1584 // alias index. Since the same load can reach both, we need to merge
1585 // these 2 disparate memories into the same alias class. Since the
1586 // primary supertype array is read-only, there's no chance of confusion
1587 // where we bypass an array load and an array store.
1588 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1589 if (offset == Type::OffsetBot ||
1590 (offset >= primary_supers_offset &&
1591 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1592 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1593 offset = in_bytes(Klass::secondary_super_cache_offset());
1594 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1595 }
1596 }
1597
1598 // Flatten all Raw pointers together.
1599 if (tj->base() == Type::RawPtr)
1600 tj = TypeRawPtr::BOTTOM;
1601
1602 if (tj->base() == Type::AnyPtr)
1603 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1604
1605 // Flatten all to bottom for now
1606 switch( _AliasLevel ) {
1607 case 0:
1608 tj = TypePtr::BOTTOM;
1609 break;
1610 case 1: // Flatten to: oop, static, field or array
1611 switch (tj->base()) {
1612 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1613 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1614 case Type::AryPtr: // do not distinguish arrays at all
1615 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1616 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1617 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1618 default: ShouldNotReachHere();
1619 }
1620 break;
1621 case 2: // No collapsing at level 2; keep all splits
1622 case 3: // No collapsing at level 3; keep all splits
1623 break;
1624 default:
1625 Unimplemented();
1626 }
1627
1628 offset = tj->offset();
1629 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1630
1631 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1632 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1633 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1634 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1635 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1636 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1637 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1638 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1639 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1640 assert( tj->ptr() != TypePtr::TopPTR &&
1641 tj->ptr() != TypePtr::AnyNull &&
1642 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1643 // assert( tj->ptr() != TypePtr::Constant ||
1644 // tj->base() == Type::RawPtr ||
1645 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1646
1647 return tj;
1648 }
1649
1650 void Compile::AliasType::Init(int i, const TypePtr* at) {
1651 _index = i;
1652 _adr_type = at;
1653 _field = NULL;
1654 _element = NULL;
1655 _is_rewritable = true; // default
1656 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1657 if (atoop != NULL && atoop->is_known_instance()) {
1658 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1659 _general_index = Compile::current()->get_alias_index(gt);
1660 } else {
1661 _general_index = 0;
1662 }
1663 }
1664
1665 BasicType Compile::AliasType::basic_type() const {
1666 if (element() != NULL) {
1667 const Type* element = adr_type()->is_aryptr()->elem();
1668 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1669 } if (field() != NULL) {
1670 return field()->layout_type();
1671 } else {
1672 return T_ILLEGAL; // unknown
1673 }
1674 }
1675
1676 //---------------------------------print_on------------------------------------
1677 #ifndef PRODUCT
1678 void Compile::AliasType::print_on(outputStream* st) {
1679 if (index() < 10)
1680 st->print("@ <%d> ", index());
1681 else st->print("@ <%d>", index());
1682 st->print(is_rewritable() ? " " : " RO");
1683 int offset = adr_type()->offset();
1684 if (offset == Type::OffsetBot)
1685 st->print(" +any");
1686 else st->print(" +%-3d", offset);
1687 st->print(" in ");
1688 adr_type()->dump_on(st);
1689 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1690 if (field() != NULL && tjp) {
1691 if (tjp->klass() != field()->holder() ||
1692 tjp->offset() != field()->offset_in_bytes()) {
1693 st->print(" != ");
1694 field()->print();
1695 st->print(" ***");
1696 }
1697 }
1698 }
1699
1700 void print_alias_types() {
1701 Compile* C = Compile::current();
1702 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1703 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1704 C->alias_type(idx)->print_on(tty);
1705 tty->cr();
1706 }
1707 }
1708 #endif
1709
1710
1711 //----------------------------probe_alias_cache--------------------------------
1712 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1713 intptr_t key = (intptr_t) adr_type;
1714 key ^= key >> logAliasCacheSize;
1715 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1716 }
1717
1718
1719 //-----------------------------grow_alias_types--------------------------------
1720 void Compile::grow_alias_types() {
1721 const int old_ats = _max_alias_types; // how many before?
1722 const int new_ats = old_ats; // how many more?
1723 const int grow_ats = old_ats+new_ats; // how many now?
1724 _max_alias_types = grow_ats;
1725 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1726 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1727 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1728 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1729 }
1730
1731
1732 //--------------------------------find_alias_type------------------------------
1733 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1734 if (_AliasLevel == 0)
1735 return alias_type(AliasIdxBot);
1736
1737 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1738 if (ace->_adr_type == adr_type) {
1739 return alias_type(ace->_index);
1740 }
1741
1742 // Handle special cases.
1743 if (adr_type == NULL) return alias_type(AliasIdxTop);
1744 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1745
1746 // Do it the slow way.
1747 const TypePtr* flat = flatten_alias_type(adr_type);
1748
1749 #ifdef ASSERT
1750 {
1751 ResourceMark rm;
1752 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1753 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1754 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1755 Type::str(adr_type));
1756 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1757 const TypeOopPtr* foop = flat->is_oopptr();
1758 // Scalarizable allocations have exact klass always.
1759 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1760 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1761 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1762 Type::str(foop), Type::str(xoop));
1763 }
1764 }
1765 #endif
1766
1767 int idx = AliasIdxTop;
1768 for (int i = 0; i < num_alias_types(); i++) {
1769 if (alias_type(i)->adr_type() == flat) {
1770 idx = i;
1771 break;
1772 }
1773 }
1774
1775 if (idx == AliasIdxTop) {
1776 if (no_create) return NULL;
1777 // Grow the array if necessary.
1778 if (_num_alias_types == _max_alias_types) grow_alias_types();
1779 // Add a new alias type.
1780 idx = _num_alias_types++;
1781 _alias_types[idx]->Init(idx, flat);
1782 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1783 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1784 if (flat->isa_instptr()) {
1785 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1786 && flat->is_instptr()->klass() == env()->Class_klass())
1787 alias_type(idx)->set_rewritable(false);
1788 }
1789 if (flat->isa_aryptr()) {
1790 #ifdef ASSERT
1791 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1792 // (T_BYTE has the weakest alignment and size restrictions...)
1793 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1794 #endif
1795 if (flat->offset() == TypePtr::OffsetBot) {
1796 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1797 }
1798 }
1799 if (flat->isa_klassptr()) {
1800 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1801 alias_type(idx)->set_rewritable(false);
1802 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1803 alias_type(idx)->set_rewritable(false);
1804 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1805 alias_type(idx)->set_rewritable(false);
1806 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1807 alias_type(idx)->set_rewritable(false);
1808 }
1809 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1810 // but the base pointer type is not distinctive enough to identify
1811 // references into JavaThread.)
1812
1813 // Check for final fields.
1814 const TypeInstPtr* tinst = flat->isa_instptr();
1815 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1816 ciField* field;
1817 if (tinst->const_oop() != NULL &&
1818 tinst->klass() == ciEnv::current()->Class_klass() &&
1819 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1820 // static field
1821 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1822 field = k->get_field_by_offset(tinst->offset(), true);
1823 } else {
1824 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1825 field = k->get_field_by_offset(tinst->offset(), false);
1826 }
1827 assert(field == NULL ||
1828 original_field == NULL ||
1829 (field->holder() == original_field->holder() &&
1830 field->offset() == original_field->offset() &&
1831 field->is_static() == original_field->is_static()), "wrong field?");
1832 // Set field() and is_rewritable() attributes.
1833 if (field != NULL) alias_type(idx)->set_field(field);
1834 }
1835 }
1836
1837 // Fill the cache for next time.
1838 ace->_adr_type = adr_type;
1839 ace->_index = idx;
1840 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1841
1842 // Might as well try to fill the cache for the flattened version, too.
1843 AliasCacheEntry* face = probe_alias_cache(flat);
1844 if (face->_adr_type == NULL) {
1845 face->_adr_type = flat;
1846 face->_index = idx;
1847 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1848 }
1849
1850 return alias_type(idx);
1851 }
1852
1853
1854 Compile::AliasType* Compile::alias_type(ciField* field) {
1855 const TypeOopPtr* t;
1856 if (field->is_static())
1857 t = TypeInstPtr::make(field->holder()->java_mirror());
1858 else
1859 t = TypeOopPtr::make_from_klass_raw(field->holder());
1860 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1861 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1862 return atp;
1863 }
1864
1865
1866 //------------------------------have_alias_type--------------------------------
1867 bool Compile::have_alias_type(const TypePtr* adr_type) {
1868 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1869 if (ace->_adr_type == adr_type) {
1870 return true;
1871 }
1872
1873 // Handle special cases.
1874 if (adr_type == NULL) return true;
1875 if (adr_type == TypePtr::BOTTOM) return true;
1876
1877 return find_alias_type(adr_type, true, NULL) != NULL;
1878 }
1879
1880 //-----------------------------must_alias--------------------------------------
1881 // True if all values of the given address type are in the given alias category.
1882 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1883 if (alias_idx == AliasIdxBot) return true; // the universal category
1884 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1885 if (alias_idx == AliasIdxTop) return false; // the empty category
1886 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1887
1888 // the only remaining possible overlap is identity
1889 int adr_idx = get_alias_index(adr_type);
1890 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1891 assert(adr_idx == alias_idx ||
1892 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1893 && adr_type != TypeOopPtr::BOTTOM),
1894 "should not be testing for overlap with an unsafe pointer");
1895 return adr_idx == alias_idx;
1896 }
1897
1898 //------------------------------can_alias--------------------------------------
1899 // True if any values of the given address type are in the given alias category.
1900 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1901 if (alias_idx == AliasIdxTop) return false; // the empty category
1902 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1903 if (alias_idx == AliasIdxBot) return true; // the universal category
1904 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1905
1906 // the only remaining possible overlap is identity
1907 int adr_idx = get_alias_index(adr_type);
1908 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1909 return adr_idx == alias_idx;
1910 }
1911
1912
1913
1914 //---------------------------pop_warm_call-------------------------------------
1915 WarmCallInfo* Compile::pop_warm_call() {
1916 WarmCallInfo* wci = _warm_calls;
1917 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1918 return wci;
1919 }
1920
1921 //----------------------------Inline_Warm--------------------------------------
1922 int Compile::Inline_Warm() {
1923 // If there is room, try to inline some more warm call sites.
1924 // %%% Do a graph index compaction pass when we think we're out of space?
1925 if (!InlineWarmCalls) return 0;
1926
1927 int calls_made_hot = 0;
1928 int room_to_grow = NodeCountInliningCutoff - unique();
1929 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1930 int amount_grown = 0;
1931 WarmCallInfo* call;
1932 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1933 int est_size = (int)call->size();
1934 if (est_size > (room_to_grow - amount_grown)) {
1935 // This one won't fit anyway. Get rid of it.
1936 call->make_cold();
1937 continue;
1938 }
1939 call->make_hot();
1940 calls_made_hot++;
1941 amount_grown += est_size;
1942 amount_to_grow -= est_size;
1943 }
1944
1945 if (calls_made_hot > 0) set_major_progress();
1946 return calls_made_hot;
1947 }
1948
1949
1950 //----------------------------Finish_Warm--------------------------------------
1951 void Compile::Finish_Warm() {
1952 if (!InlineWarmCalls) return;
1953 if (failing()) return;
1954 if (warm_calls() == NULL) return;
1955
1956 // Clean up loose ends, if we are out of space for inlining.
1957 WarmCallInfo* call;
1958 while ((call = pop_warm_call()) != NULL) {
1959 call->make_cold();
1960 }
1961 }
1962
1963 //---------------------cleanup_loop_predicates-----------------------
1964 // Remove the opaque nodes that protect the predicates so that all unused
1965 // checks and uncommon_traps will be eliminated from the ideal graph
1966 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1967 if (predicate_count()==0) return;
1968 for (int i = predicate_count(); i > 0; i--) {
1969 Node * n = predicate_opaque1_node(i-1);
1970 assert(n->Opcode() == Op_Opaque1, "must be");
1971 igvn.replace_node(n, n->in(1));
1972 }
1973 assert(predicate_count()==0, "should be clean!");
1974 }
1975
1976 void Compile::add_range_check_cast(Node* n) {
1977 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1978 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1979 _range_check_casts->append(n);
1980 }
1981
1982 // Remove all range check dependent CastIINodes.
1983 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1984 for (int i = range_check_cast_count(); i > 0; i--) {
1985 Node* cast = range_check_cast_node(i-1);
1986 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1987 igvn.replace_node(cast, cast->in(1));
1988 }
1989 assert(range_check_cast_count() == 0, "should be empty");
1990 }
1991
1992 void Compile::add_opaque4_node(Node* n) {
1993 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1994 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1995 _opaque4_nodes->append(n);
1996 }
1997
1998 // Remove all Opaque4 nodes.
1999 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2000 for (int i = opaque4_count(); i > 0; i--) {
2001 Node* opaq = opaque4_node(i-1);
2002 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2003 igvn.replace_node(opaq, opaq->in(2));
2004 }
2005 assert(opaque4_count() == 0, "should be empty");
2006 }
2007
2008 // StringOpts and late inlining of string methods
2009 void Compile::inline_string_calls(bool parse_time) {
2010 {
2011 // remove useless nodes to make the usage analysis simpler
2012 ResourceMark rm;
2013 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2014 }
2015
2016 {
2017 ResourceMark rm;
2018 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2019 PhaseStringOpts pso(initial_gvn(), for_igvn());
2020 print_method(PHASE_AFTER_STRINGOPTS, 3);
2021 }
2022
2023 // now inline anything that we skipped the first time around
2024 if (!parse_time) {
2025 _late_inlines_pos = _late_inlines.length();
2026 }
2027
2028 while (_string_late_inlines.length() > 0) {
2029 CallGenerator* cg = _string_late_inlines.pop();
2030 cg->do_late_inline();
2031 if (failing()) return;
2032 }
2033 _string_late_inlines.trunc_to(0);
2034 }
2035
2036 // Late inlining of boxing methods
2037 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2038 if (_boxing_late_inlines.length() > 0) {
2039 assert(has_boxed_value(), "inconsistent");
2040
2041 PhaseGVN* gvn = initial_gvn();
2042 set_inlining_incrementally(true);
2043
2044 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2045 for_igvn()->clear();
2046 gvn->replace_with(&igvn);
2047
2048 _late_inlines_pos = _late_inlines.length();
2049
2050 while (_boxing_late_inlines.length() > 0) {
2051 CallGenerator* cg = _boxing_late_inlines.pop();
2052 cg->do_late_inline();
2053 if (failing()) return;
2054 }
2055 _boxing_late_inlines.trunc_to(0);
2056
2057 inline_incrementally_cleanup(igvn);
2058
2059 set_inlining_incrementally(false);
2060 }
2061 }
2062
2063 bool Compile::inline_incrementally_one() {
2064 assert(IncrementalInline, "incremental inlining should be on");
2065
2066 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2067 set_inlining_progress(false);
2068 set_do_cleanup(false);
2069 int i = 0;
2070 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2071 CallGenerator* cg = _late_inlines.at(i);
2072 _late_inlines_pos = i+1;
2073 cg->do_late_inline();
2074 if (failing()) return false;
2075 }
2076 int j = 0;
2077 for (; i < _late_inlines.length(); i++, j++) {
2078 _late_inlines.at_put(j, _late_inlines.at(i));
2079 }
2080 _late_inlines.trunc_to(j);
2081 assert(inlining_progress() || _late_inlines.length() == 0, "");
2082
2083 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2084
2085 set_inlining_progress(false);
2086 set_do_cleanup(false);
2087 return (_late_inlines.length() > 0) && !needs_cleanup;
2088 }
2089
2090 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2091 {
2092 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2093 ResourceMark rm;
2094 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2095 }
2096 {
2097 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2098 igvn = PhaseIterGVN(initial_gvn());
2099 igvn.optimize();
2100 }
2101 }
2102
2103 // Perform incremental inlining until bound on number of live nodes is reached
2104 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2105 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2106
2107 set_inlining_incrementally(true);
2108 uint low_live_nodes = 0;
2109
2110 while (_late_inlines.length() > 0) {
2111 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2112 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2113 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2114 // PhaseIdealLoop is expensive so we only try it once we are
2115 // out of live nodes and we only try it again if the previous
2116 // helped got the number of nodes down significantly
2117 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2118 if (failing()) return;
2119 low_live_nodes = live_nodes();
2120 _major_progress = true;
2121 }
2122
2123 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2124 break; // finish
2125 }
2126 }
2127
2128 for_igvn()->clear();
2129 initial_gvn()->replace_with(&igvn);
2130
2131 while (inline_incrementally_one()) {
2132 assert(!failing(), "inconsistent");
2133 }
2134
2135 if (failing()) return;
2136
2137 inline_incrementally_cleanup(igvn);
2138
2139 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2140
2141
2142 if (failing()) return;
2143 }
2144 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2145
2146 if (_string_late_inlines.length() > 0) {
2147 assert(has_stringbuilder(), "inconsistent");
2148 for_igvn()->clear();
2149 initial_gvn()->replace_with(&igvn);
2150
2151 inline_string_calls(false);
2152
2153 if (failing()) return;
2154
2155 inline_incrementally_cleanup(igvn);
2156 }
2157
2158 set_inlining_incrementally(false);
2159 }
2160
2161
2162 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2163 if(_loop_opts_cnt > 0) {
2164 debug_only( int cnt = 0; );
2165 while(major_progress() && (_loop_opts_cnt > 0)) {
2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2167 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2168 PhaseIdealLoop::optimize(igvn, mode);
2169 _loop_opts_cnt--;
2170 if (failing()) return false;
2171 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2172 }
2173 }
2174 return true;
2175 }
2176
2177 // Remove edges from "root" to each SafePoint at a backward branch.
2178 // They were inserted during parsing (see add_safepoint()) to make
2179 // infinite loops without calls or exceptions visible to root, i.e.,
2180 // useful.
2181 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2182 Node *r = root();
2183 if (r != NULL) {
2184 for (uint i = r->req(); i < r->len(); ++i) {
2185 Node *n = r->in(i);
2186 if (n != NULL && n->is_SafePoint()) {
2187 r->rm_prec(i);
2188 if (n->outcnt() == 0) {
2189 igvn.remove_dead_node(n);
2190 }
2191 --i;
2192 }
2193 }
2194 }
2195 }
2196
2197 //------------------------------Optimize---------------------------------------
2198 // Given a graph, optimize it.
2199 void Compile::Optimize() {
2200 TracePhase tp("optimizer", &timers[_t_optimizer]);
2201
2202 #ifndef PRODUCT
2203 if (_directive->BreakAtCompileOption) {
2204 BREAKPOINT;
2205 }
2206
2207 #endif
2208
2209 #ifdef ASSERT
2210 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2211 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2212 #endif
2213
2214 ResourceMark rm;
2215
2216 print_inlining_reinit();
2217
2218 NOT_PRODUCT( verify_graph_edges(); )
2219
2220 print_method(PHASE_AFTER_PARSING);
2221
2222 {
2223 // Iterative Global Value Numbering, including ideal transforms
2224 // Initialize IterGVN with types and values from parse-time GVN
2225 PhaseIterGVN igvn(initial_gvn());
2226 #ifdef ASSERT
2227 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2228 #endif
2229 {
2230 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2231 igvn.optimize();
2232 }
2233
2234 if (failing()) return;
2235
2236 print_method(PHASE_ITER_GVN1, 2);
2237
2238 inline_incrementally(igvn);
2239
2240 print_method(PHASE_INCREMENTAL_INLINE, 2);
2241
2242 if (failing()) return;
2243
2244 if (eliminate_boxing()) {
2245 // Inline valueOf() methods now.
2246 inline_boxing_calls(igvn);
2247
2248 if (AlwaysIncrementalInline) {
2249 inline_incrementally(igvn);
2250 }
2251 if (failing()) return;
2252
2253 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2254 }
2255
2256 // Now that all inlining is over, cut edge from root to loop
2257 // safepoints
2258 remove_root_to_sfpts_edges(igvn);
2259
2260 // Remove the speculative part of types and clean up the graph from
2261 // the extra CastPP nodes whose only purpose is to carry them. Do
2262 // that early so that optimizations are not disrupted by the extra
2263 // CastPP nodes.
2264 remove_speculative_types(igvn);
2265
2266 // No more new expensive nodes will be added to the list from here
2267 // so keep only the actual candidates for optimizations.
2268 cleanup_expensive_nodes(igvn);
2269
2270 if (has_vbox_nodes()) {
2271 TracePhase tp("expandVectorBoxes", &timers[_t_expandVectorBoxes]);
2272
2273 set_inlining_incrementally(true); // FIXME another way to signal GraphKit it's post-parsing phase?
2274
2275 for_igvn()->clear();
2276 initial_gvn()->replace_with(&igvn);
2277
2278 expand_vunbox_nodes();
2279 scalarize_vbox_nodes();
2280
2281 inline_vector_reboxing_calls();
2282
2283 if (failing()) return;
2284 {
2285 ResourceMark rm;
2286 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2287 }
2288 igvn = PhaseIterGVN(initial_gvn());
2289 igvn.optimize();
2290
2291 print_method(PHASE_ITER_GVN_BEFORE_EA, 2);
2292
2293 if (failing()) return;
2294
2295 ////////////////////////////////////////////////////
2296
2297 for_igvn()->clear();
2298 initial_gvn()->replace_with(&igvn);
2299
2300 expand_vbox_nodes();
2301 eliminate_vbox_alloc_nodes();
2302
2303 if (failing()) return;
2304 {
2305 ResourceMark rm;
2306 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2307 }
2308 igvn = PhaseIterGVN(initial_gvn());
2309 igvn.optimize();
2310 if (failing()) return;
2311
2312 print_method(PHASE_ITER_GVN_BEFORE_EA, 2);
2313
2314 set_inlining_incrementally(false); // FIXME another way to signal GraphKit it's post-parsing phase?
2315 }
2316
2317 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2318 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2319 initial_gvn()->replace_with(&igvn);
2320 for_igvn()->clear();
2321 Unique_Node_List new_worklist(C->comp_arena());
2322 {
2323 ResourceMark rm;
2324 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2325 }
2326 set_for_igvn(&new_worklist);
2327 igvn = PhaseIterGVN(initial_gvn());
2328 igvn.optimize();
2329 }
2330
2331 // FIXME for_igvn() is corrupted from here: new_worklist which is set_for_ignv() was allocated on stack.
2332
2333 // Perform escape analysis
2334 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2335 if (has_loops()) {
2336 // Cleanup graph (remove dead nodes).
2337 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2338 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2339 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2340 if (failing()) return;
2341 }
2342 ConnectionGraph::do_analysis(this, &igvn);
2343
2344 if (failing()) return;
2345
2346 // Optimize out fields loads from scalar replaceable allocations.
2347 igvn.optimize();
2348 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2349
2350 if (failing()) return;
2351
2352 if (congraph() != NULL && macro_count() > 0) {
2353 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2354 PhaseMacroExpand mexp(igvn);
2355 mexp.eliminate_macro_nodes();
2356 igvn.set_delay_transform(false);
2357
2358 igvn.optimize();
2359 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2360
2361 if (failing()) return;
2362 }
2363 }
2364
2365 // Loop transforms on the ideal graph. Range Check Elimination,
2366 // peeling, unrolling, etc.
2367
2368 // Set loop opts counter
2369 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2370 {
2371 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2372 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2373 _loop_opts_cnt--;
2374 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2375 if (failing()) return;
2376 }
2377 // Loop opts pass if partial peeling occurred in previous pass
2378 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2379 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2380 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2381 _loop_opts_cnt--;
2382 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2383 if (failing()) return;
2384 }
2385 // Loop opts pass for loop-unrolling before CCP
2386 if(major_progress() && (_loop_opts_cnt > 0)) {
2387 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2388 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2389 _loop_opts_cnt--;
2390 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2391 }
2392 if (!failing()) {
2393 // Verify that last round of loop opts produced a valid graph
2394 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2395 PhaseIdealLoop::verify(igvn);
2396 }
2397 }
2398 if (failing()) return;
2399
2400 // Conditional Constant Propagation;
2401 PhaseCCP ccp( &igvn );
2402 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2403 {
2404 TracePhase tp("ccp", &timers[_t_ccp]);
2405 ccp.do_transform();
2406 }
2407 print_method(PHASE_CPP1, 2);
2408
2409 assert( true, "Break here to ccp.dump_old2new_map()");
2410
2411 // Iterative Global Value Numbering, including ideal transforms
2412 {
2413 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2414 igvn = ccp;
2415 igvn.optimize();
2416 }
2417
2418 print_method(PHASE_ITER_GVN2, 2);
2419
2420 if (failing()) return;
2421
2422 // Loop transforms on the ideal graph. Range Check Elimination,
2423 // peeling, unrolling, etc.
2424 if (!optimize_loops(igvn, LoopOptsDefault)) {
2425 return;
2426 }
2427
2428 #if INCLUDE_ZGC
2429 if (UseZGC) {
2430 ZBarrierSetC2::find_dominating_barriers(igvn);
2431 }
2432 #endif
2433
2434 if (failing()) return;
2435
2436 // Ensure that major progress is now clear
2437 C->clear_major_progress();
2438
2439 {
2440 // Verify that all previous optimizations produced a valid graph
2441 // at least to this point, even if no loop optimizations were done.
2442 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2443 PhaseIdealLoop::verify(igvn);
2444 }
2445
2446 if (range_check_cast_count() > 0) {
2447 // No more loop optimizations. Remove all range check dependent CastIINodes.
2448 C->remove_range_check_casts(igvn);
2449 igvn.optimize();
2450 }
2451
2452 #ifdef ASSERT
2453 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2454 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand);
2455 #endif
2456
2457 {
2458 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2459 PhaseMacroExpand mex(igvn);
2460 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2461 if (mex.expand_macro_nodes()) {
2462 assert(failing(), "must bail out w/ explicit message");
2463 return;
2464 }
2465 }
2466
2467 {
2468 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2469 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2);
2470 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2471 if (bs->expand_barriers(this, igvn)) {
2472 assert(failing(), "must bail out w/ explicit message");
2473 return;
2474 }
2475 }
2476
2477 if (opaque4_count() > 0) {
2478 C->remove_opaque4_nodes(igvn);
2479 igvn.optimize();
2480 }
2481
2482 DEBUG_ONLY( _modified_nodes = NULL; )
2483 } // (End scope of igvn; run destructor if necessary for asserts.)
2484
2485 process_print_inlining();
2486 // A method with only infinite loops has no edges entering loops from root
2487 {
2488 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2489 if (final_graph_reshaping()) {
2490 assert(failing(), "must bail out w/ explicit message");
2491 return;
2492 }
2493 }
2494
2495 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2496 }
2497
2498 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
2499 ResourceMark rm;
2500 stringStream ss;
2501 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
2502 if (n != NULL) {
2503 #ifndef PRODUCT
2504 ss.print(": %s %d", n->Name(), n->_idx);
2505 #else
2506 ss.print(": %d %d", n->Opcode(), n->_idx);
2507 #endif // PRODUCT
2508 } else {
2509 ss.print_raw(": NULL");
2510 }
2511 C->print_method(cpt, ss.as_string(), level);
2512 }
2513
2514 void Compile::scalarize_vbox_nodes() {
2515 int macro_idx = C->macro_count() - 1;
2516 while(macro_idx >= 0) {
2517 Node * n = C->macro_node(macro_idx);
2518 assert(n->is_macro(), "only macro nodes expected here");
2519 if (n->Opcode() == Op_VectorBox) {
2520 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(n);
2521 scalarize_vbox_node(vbox);
2522 if (failing()) return;
2523 print_method(PHASE_SCALARIZE_VBOX, vbox);
2524 }
2525 if (C->failing()) return;
2526 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2527 }
2528 }
2529
2530 void Compile::expand_vbox_nodes() {
2531 int macro_idx = C->macro_count() - 1;
2532 while(macro_idx >= 0) {
2533 Node * n = C->macro_node(macro_idx);
2534 assert(n->is_macro(), "only macro nodes expected here");
2535 if (n->Opcode() == Op_VectorBox) {
2536 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(n);
2537 expand_vbox_node(vbox);
2538 if (failing()) return;
2539 print_method(PHASE_EXPAND_VBOX, vbox);
2540 }
2541 if (C->failing()) return;
2542 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2543 }
2544 }
2545
2546 void Compile::expand_vunbox_nodes() {
2547 int macro_idx = C->macro_count() - 1;
2548 while(macro_idx >= 0) {
2549 Node * n = C->macro_node(macro_idx);
2550 assert(n->is_macro(), "only macro nodes expected here");
2551 if (n->Opcode() == Op_VectorUnbox) {
2552 VectorUnboxNode* vec_unbox = static_cast<VectorUnboxNode*>(n);
2553 expand_vunbox_node(vec_unbox);
2554 if (failing()) return;
2555 print_method(PHASE_EXPAND_VUNBOX, vec_unbox);
2556 }
2557 if (C->failing()) return;
2558 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2559 }
2560 }
2561
2562 void Compile::eliminate_vbox_alloc_nodes() {
2563 int macro_idx = C->macro_count() - 1;
2564 while(macro_idx >= 0) {
2565 Node * n = C->macro_node(macro_idx);
2566 assert(n->is_macro(), "only macro nodes expected here");
2567 if (n->Opcode() == Op_VectorBoxAllocate) {
2568 VectorBoxAllocateNode* vbox_alloc = static_cast<VectorBoxAllocateNode*>(n);
2569 eliminate_vbox_alloc_node(vbox_alloc);
2570 if (failing()) return;
2571 print_method(PHASE_ELIMINATE_VBOX_ALLOC, vbox_alloc);
2572 }
2573 if (C->failing()) return;
2574 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2575 }
2576 }
2577
2578 void Compile::inline_vector_reboxing_calls() {
2579 if (_vector_reboxing_late_inlines.length() > 0) {
2580 PhaseGVN* gvn = initial_gvn();
2581
2582 _late_inlines_pos = _late_inlines.length();
2583 while (_vector_reboxing_late_inlines.length() > 0) {
2584 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2585 cg->do_late_inline();
2586 if (failing()) return;
2587 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2588 }
2589 _vector_reboxing_late_inlines.trunc_to(0);
2590 }
2591 }
2592
2593 bool Compile::has_vbox_nodes() {
2594 if (_vector_reboxing_late_inlines.length() > 0) {
2595 return true;
2596 }
2597 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2598 Node * n = C->macro_node(macro_idx);
2599 assert(n->is_macro(), "only macro nodes expected here");
2600 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2601 return true;
2602 }
2603 }
2604 return false;
2605 }
2606
2607 static JVMState* clone_jvms(Compile* C, SafePointNode* sfpt) {
2608 JVMState* new_jvms = sfpt->jvms()->clone_shallow(C);
2609 uint size = sfpt->req();
2610 SafePointNode* map = new SafePointNode(size, new_jvms);
2611 for (uint i = 0; i < size; i++) {
2612 map->init_req(i, sfpt->in(i));
2613 }
2614 // Make sure the state is a MergeMem for parsing.
2615 // FIXME not needed?
2616 // if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
2617 // Node* mem = MergeMemNode::make(map->in(TypeFunc::Memory));
2618 // gvn.set_type_bottom(mem);
2619 // map->set_req(TypeFunc::Memory, mem);
2620 // }
2621 new_jvms->set_map(map);
2622 return new_jvms;
2623 }
2624
2625 void Compile::scalarize_vbox_node(VectorBoxNode* vec_box) {
2626 Node* vec_value = vec_box->in(VectorBoxNode::Value);
2627 PhaseGVN& gvn = *initial_gvn();
2628
2629 // Process merged VBAs
2630
2631 Unique_Node_List calls(C->comp_arena());
2632 for (DUIterator_Fast imax, i = vec_box->fast_outs(imax); i < imax; i++) {
2633 Node* use = vec_box->fast_out(i);
2634 if (use->is_CallJava()) {
2635 CallJavaNode* call = use->as_CallJava();
2636 if (call->has_non_debug_use(vec_box) && vec_box->in(VectorBoxNode::Box)->is_Phi()) {
2637 calls.push(call);
2638 }
2639 }
2640 }
2641
2642 while (VectorAPIAggressiveReboxing && calls.size() > 0) {
2643 CallJavaNode* call = calls.pop()->as_CallJava();
2644 // Attach new VBA to the call and use it instead of Phi (VBA ... VBA).
2645
2646 JVMState* jvms = clone_jvms(C, call);
2647 GraphKit kit(jvms);
2648 PhaseGVN& gvn = kit.gvn();
2649
2650 // Adjust JVMS from post-call to pre-call state: put args on stack
2651 uint nargs = call->method()->arg_size();
2652 kit.ensure_stack(kit.sp() + nargs);
2653 for (uint i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
2654 kit.push(call->in(i));
2655 }
2656 jvms = kit.sync_jvms();
2657
2658 Node* new_vbox = NULL;
2659 {
2660 PreserveReexecuteState prs(&kit);
2661
2662 kit.jvms()->set_should_reexecute(true);
2663
2664 const TypeInstPtr* vbox_type = vec_box->box_type();
2665 const TypeVect* vect_type = vec_box->vec_type();
2666 Node* vect = vec_box->in(VectorBoxNode::Value);
2667
2668 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
2669 kit.set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
2670 kit.make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true, /*deoptimize=*/true);
2671 kit.set_i_o(gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
2672 kit.set_all_memory(gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
2673 Node* ret = gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
2674
2675 new_vbox = gvn.transform(new VectorBoxNode(C, ret, vect, vbox_type, vect_type));
2676
2677 kit.replace_in_map(vec_box, new_vbox);
2678 }
2679
2680 kit.dec_sp(nargs);
2681 jvms = kit.sync_jvms();
2682
2683 bool found = false;
2684 int cnt = _vector_reboxing_late_inlines.length();
2685 for (int i = 0; i < cnt; i++) {
2686 CallGenerator* cg = _vector_reboxing_late_inlines.at(i);
2687 if (cg->call_node() == call) {
2688 ciMethod* m = cg->method();
2689 _vector_reboxing_late_inlines.remove(cg); // remove_at(i);
2690 CallGenerator* new_cg = CallGenerator::for_vector_reboxing_late_inline(
2691 m,
2692 CallGenerator::for_inline(m, m->interpreter_invocation_count()));
2693
2694 JVMState* new_jvms = new_cg->generate(jvms);
2695
2696 Node* new_res = kit.top();
2697 if (m->return_type()->basic_type() != T_VOID) {
2698 new_res = new_cg->call_node()->proj_out(TypeFunc::Parms);
2699 }
2700 kit.replace_call(call, new_res, /*do_replaced_nodes=*/true);
2701
2702 found = true;
2703 break;
2704 }
2705 }
2706
2707 if (!found) {
2708 // FIXME does it cover all cases?
2709 CallJavaNode* new_call = call->clone()->as_CallJava();
2710 new_call->set_req(TypeFunc::Control , kit.control());
2711 new_call->set_req(TypeFunc::I_O , kit.i_o());
2712 new_call->set_req(TypeFunc::Memory , kit.reset_memory());
2713 new_call->set_req(TypeFunc::FramePtr , kit.frameptr());
2714
2715 new_call->replace_edge(vec_box, new_vbox);
2716 new_call = gvn.transform(new_call)->as_CallJava();
2717
2718 C->gvn_replace_by(call, new_call);
2719 }
2720 call->disconnect_inputs(NULL, C);
2721 }
2722
2723 // Process debug uses at safepoints
2724 Unique_Node_List safepoints(C->comp_arena());
2725
2726 for (DUIterator_Fast imax, i = vec_box->fast_outs(imax); i < imax; i++) {
2727 Node* use = vec_box->fast_out(i);
2728 if (use->is_SafePoint()) {
2729 SafePointNode* sfpt = use->as_SafePoint();
2730 if (!sfpt->is_Call() || !sfpt->as_Call()->has_non_debug_use(vec_box)) {
2731 safepoints.push(sfpt);
2732 }
2733 }
2734 }
2735
2736 while (safepoints.size() > 0) {
2737 SafePointNode* sfpt = safepoints.pop()->as_SafePoint();
2738
2739 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
2740 Node* sobj = new SafePointScalarObjectNode(vec_box->box_type(),
2741 #ifdef ASSERT
2742 NULL,
2743 #endif // ASSERT
2744 first_ind, /*n_fields=*/1);
2745 sobj->init_req(0, C->root());
2746 sfpt->add_req(vec_value);
2747
2748 sobj = gvn.transform(sobj);
2749
2750 JVMState *jvms = sfpt->jvms();
2751
2752 jvms->set_endoff(sfpt->req());
2753 // Now make a pass over the debug information replacing any references
2754 // to the allocated object with "sobj"
2755 int start = jvms->debug_start();
2756 int end = jvms->debug_end();
2757 sfpt->replace_edges_in_range(vec_box, sobj, start, end);
2758 }
2759 }
2760
2761 void Compile::expand_vbox_node(VectorBoxNode* vec_box) {
2762 if (vec_box->outcnt() > 0) {
2763 Node* vbox = vec_box->in(VectorBoxNode::Box);
2764 Node* vect = vec_box->in(VectorBoxNode::Value);
2765 Node* result = expand_vbox_node_helper(vbox, vect, vec_box->box_type(), vec_box->vec_type());
2766 C->gvn_replace_by(vec_box, result);
2767 }
2768 C->remove_macro_node(vec_box);
2769 }
2770
2771 Node* Compile::expand_vbox_node_helper(Node* vbox,
2772 Node* vect,
2773 const TypeInstPtr* box_type,
2774 const TypeVect* vect_type) {
2775 if (vbox->is_Phi() && vect->is_Phi()) {
2776 assert(vbox->as_Phi()->region() == vect->as_Phi()->region(), "");
2777 Node* new_phi = new PhiNode(vbox->as_Phi()->region(), box_type);
2778 for (uint i = 1; i < vbox->req(); i++) {
2779 Node* new_box = expand_vbox_node_helper(vbox->in(i), vect->in(i), box_type, vect_type);
2780 new_phi->set_req(i, new_box);
2781 }
2782 new_phi = initial_gvn()->transform(new_phi);
2783 return new_phi;
2784 } else if (vbox->is_Proj() && vbox->in(0)->Opcode() == Op_VectorBoxAllocate) {
2785 VectorBoxAllocateNode* vbox_alloc = static_cast<VectorBoxAllocateNode*>(vbox->in(0));
2786 return expand_vbox_alloc_node(vbox_alloc, vect, box_type, vect_type);
2787 } else {
2788 // TODO: ensure that expanded vbox is initialized with the same value (vect).
2789 return vbox; // already expanded
2790 }
2791 }
2792
2793 static const char * get_field_name(ciInstanceKlass * box_class) {
2794 if(box_class->is_vectormask())
2795 return "bits";
2796 else if (box_class->is_vectorshuffle())
2797 return "reorder";
2798 else
2799 return "vec";
2800 }
2801
2802 Node* Compile::expand_vbox_alloc_node(VectorBoxAllocateNode* vbox_alloc,
2803 Node* value,
2804 const TypeInstPtr* box_type,
2805 const TypeVect* vect_type) {
2806 JVMState* jvms = clone_jvms(C, vbox_alloc);
2807 GraphKit kit(jvms);
2808 PhaseGVN& gvn = kit.gvn();
2809
2810 ciInstanceKlass* box_klass = box_type->klass()->as_instance_klass();
2811 BasicType bt = vect_type->element_basic_type();
2812 int num_elem = vect_type->length();
2813
2814 bool is_mask = box_klass->is_vectormask();
2815 //assert(!is_mask || vect_type->element_basic_type() == getMaskBasicType(bt), "consistent vector element type expected");
2816 if (is_mask && bt != T_BOOLEAN) {
2817 value = gvn.transform(new VectorStoreMaskNode(value, bt, num_elem));
2818 // Although type of mask depends on its definition, in terms of storage everything is stored in boolean array.
2819 bt = T_BOOLEAN;
2820 assert(value->as_Vector()->bottom_type()->is_vect()->element_basic_type() == bt,
2821 "must be consistent with mask representation");
2822 }
2823
2824 // Generate the allocate for the Vector object.
2825 const TypeKlassPtr* klass_type = box_type->as_klass_type();
2826 Node* klass_node = kit.makecon(klass_type);
2827 Node* vec_obj = kit.new_instance(klass_node);
2828
2829 // Generate array allocation for the field which holds the values.
2830 const TypeKlassPtr* array_klass = TypeKlassPtr::make(ciTypeArrayKlass::make(bt));
2831 Node* arr = kit.new_array(kit.makecon(array_klass), kit.intcon(num_elem), 1);
2832
2833 // FIXME convert both stores into initializing stores
2834
2835 // Store the vector value into the array.
2836 Node* arr_adr = kit.array_element_address(arr, kit.intcon(0), bt);
2837 const TypePtr* arr_adr_type = arr_adr->bottom_type()->is_ptr();
2838 Node* arr_mem = kit.memory(arr_adr);
2839 Node* vstore = gvn.transform(StoreVectorNode::make(0,
2840 kit.control(),
2841 arr_mem,
2842 arr_adr,
2843 arr_adr_type,
2844 value,
2845 num_elem));
2846 kit.set_memory(vstore, arr_adr_type);
2847
2848
2849 // Store the allocated array into object.
2850 ciField* field = box_klass->get_field_by_name(ciSymbol::make(get_field_name(box_klass)),
2851 ciSymbol::make(TypeArrayKlass::external_name(bt)),
2852 false);
2853 Node* vec_field = kit.basic_plus_adr(vec_obj, field->offset_in_bytes());
2854 const TypePtr* vec_adr_type = vec_field->bottom_type()->is_ptr();
2855 Node* field_store = gvn.transform(kit.access_store_at(vec_obj,
2856 vec_field,
2857 vec_adr_type,
2858 arr,
2859 TypeOopPtr::make_from_klass(field->type()->as_klass()),
2860 T_OBJECT,
2861 IN_HEAP));
2862
2863 kit.set_memory(field_store, vec_adr_type);
2864
2865 kit.replace_call(vbox_alloc, vec_obj, true);
2866 C->remove_macro_node(vbox_alloc);
2867 return vec_obj;
2868 }
2869
2870 void Compile::expand_vunbox_node(VectorUnboxNode* vec_unbox) {
2871 if (vec_unbox->outcnt() > 0) {
2872 GraphKit kit;
2873 PhaseGVN& gvn = kit.gvn();
2874
2875 Node* obj = vec_unbox->obj();
2876 const TypeInstPtr* tinst = gvn.type(obj)->isa_instptr();
2877 ciInstanceKlass* from_kls = tinst->klass()->as_instance_klass();
2878 BasicType bt = vec_unbox->vect_type()->element_basic_type();
2879 BasicType masktype = bt;
2880
2881 const char* field_name = "vec";
2882 if (from_kls->is_vectormask()) {
2883 field_name = "bits";
2884 bt = T_BOOLEAN;
2885 } else if (from_kls->is_vectorshuffle()) {
2886 field_name = "reorder";
2887 bt = T_BYTE;
2888 }
2889
2890 ciField* field = from_kls->get_field_by_name(ciSymbol::make(field_name),
2891 ciSymbol::make(TypeArrayKlass::external_name(bt)), false);
2892
2893 int offset = field->offset_in_bytes();
2894 Node* vec_adr = kit.basic_plus_adr(obj, offset);
2895
2896 Node* mem = vec_unbox->mem();
2897 Node* ctrl = vec_unbox->in(0);
2898 Node* vec_field_ld = LoadNode::make(gvn,
2899 ctrl,
2900 mem,
2901 vec_adr,
2902 vec_adr->bottom_type()->is_ptr(),
2903 TypeOopPtr::make_from_klass(field->type()->as_klass()),
2904 T_OBJECT,
2905 MemNode::unordered);
2906 vec_field_ld = gvn.transform(vec_field_ld);
2907
2908 Node* adr = kit.array_element_address(vec_field_ld, gvn.intcon(0), bt);
2909 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
2910 int num_elem = vec_unbox->bottom_type()->is_vect()->length();
2911 Node* vec_val_load = LoadVectorNode::make(0,
2912 ctrl,
2913 mem,
2914 adr,
2915 adr_type,
2916 num_elem,
2917 bt);
2918 vec_val_load = gvn.transform(vec_val_load);
2919
2920 if (from_kls->is_vectormask() && masktype != T_BOOLEAN) {
2921 assert(vec_unbox->bottom_type()->is_vect()->element_basic_type() == masktype, "expect mask type consistency");
2922 vec_val_load = gvn.transform(new VectorLoadMaskNode(vec_val_load, TypeVect::make(masktype, num_elem)));
2923 } else if (from_kls->is_vectorshuffle()) {
2924 assert(vec_unbox->bottom_type()->is_vect()->element_basic_type() == masktype, "expect shuffle type consistency");
2925 vec_val_load = gvn.transform(new VectorLoadShuffleNode(vec_val_load, TypeVect::make(masktype, num_elem)));
2926 }
2927
2928 gvn.hash_delete(vec_unbox);
2929 vec_unbox->disconnect_inputs(NULL, C);
2930 C->gvn_replace_by(vec_unbox, vec_val_load);
2931 }
2932 C->remove_macro_node(vec_unbox);
2933 }
2934
2935 void Compile::eliminate_vbox_alloc_node(VectorBoxAllocateNode* vbox_alloc) {
2936 JVMState* jvms = clone_jvms(C, vbox_alloc);
2937 GraphKit kit(jvms);
2938 // FIXME replace VBA with a safepoint. Otherwise, no safepoints left in tight loops.
2939 kit.replace_call(vbox_alloc, kit.top(), true);
2940 C->remove_macro_node(vbox_alloc);
2941 }
2942
2943 //------------------------------Code_Gen---------------------------------------
2944 // Given a graph, generate code for it
2945 void Compile::Code_Gen() {
2946 if (failing()) {
2947 return;
2948 }
2949
2950 // Perform instruction selection. You might think we could reclaim Matcher
2951 // memory PDQ, but actually the Matcher is used in generating spill code.
2952 // Internals of the Matcher (including some VectorSets) must remain live
2953 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2954 // set a bit in reclaimed memory.
2955
2956 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2957 // nodes. Mapping is only valid at the root of each matched subtree.
2958 NOT_PRODUCT( verify_graph_edges(); )
2959
2960 Matcher matcher;
2961 _matcher = &matcher;
2962 {
2963 TracePhase tp("matcher", &timers[_t_matcher]);
2964 matcher.match();
2965 print_method(PHASE_AFTER_MATCHING, 3);
2966 }
2967 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2968 // nodes. Mapping is only valid at the root of each matched subtree.
2969 NOT_PRODUCT( verify_graph_edges(); )
2970
2971 // If you have too many nodes, or if matching has failed, bail out
2972 check_node_count(0, "out of nodes matching instructions");
2973 if (failing()) {
2974 return;
2975 }
2976
2977 print_method(PHASE_MATCHING, 2);
2978
2979 // Build a proper-looking CFG
2980 PhaseCFG cfg(node_arena(), root(), matcher);
2981 _cfg = &cfg;
2982 {
2983 TracePhase tp("scheduler", &timers[_t_scheduler]);
2984 bool success = cfg.do_global_code_motion();
2985 if (!success) {
2986 return;
2987 }
2988
2989 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2990 NOT_PRODUCT( verify_graph_edges(); )
2991 debug_only( cfg.verify(); )
2992 }
2993
2994 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2995 _regalloc = ®alloc;
2996 {
2997 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2998 // Perform register allocation. After Chaitin, use-def chains are
2999 // no longer accurate (at spill code) and so must be ignored.
3000 // Node->LRG->reg mappings are still accurate.
3001 _regalloc->Register_Allocate();
3002
3003 // Bail out if the allocator builds too many nodes
3004 if (failing()) {
3005 return;
3006 }
3007 }
3008
3009 // Prior to register allocation we kept empty basic blocks in case the
3010 // the allocator needed a place to spill. After register allocation we
3011 // are not adding any new instructions. If any basic block is empty, we
3012 // can now safely remove it.
3013 {
3014 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
3015 cfg.remove_empty_blocks();
3016 if (do_freq_based_layout()) {
3017 PhaseBlockLayout layout(cfg);
3018 } else {
3019 cfg.set_loop_alignment();
3020 }
3021 cfg.fixup_flow();
3022 }
3023
3024 // Apply peephole optimizations
3025 if( OptoPeephole ) {
3026 TracePhase tp("peephole", &timers[_t_peephole]);
3027 PhasePeephole peep( _regalloc, cfg);
3028 peep.do_transform();
3029 }
3030
3031 // Do late expand if CPU requires this.
3032 if (Matcher::require_postalloc_expand) {
3033 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3034 cfg.postalloc_expand(_regalloc);
3035 }
3036
3037 // Convert Nodes to instruction bits in a buffer
3038 {
3039 TraceTime tp("output", &timers[_t_output], CITime);
3040 Output();
3041 }
3042
3043 print_method(PHASE_FINAL_CODE);
3044
3045 // He's dead, Jim.
3046 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3047 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3048 }
3049
3050
3051 //------------------------------dump_asm---------------------------------------
3052 // Dump formatted assembly
3053 #ifndef PRODUCT
3054 void Compile::dump_asm(int *pcs, uint pc_limit) {
3055 bool cut_short = false;
3056 tty->print_cr("#");
3057 tty->print("# "); _tf->dump(); tty->cr();
3058 tty->print_cr("#");
3059
3060 // For all blocks
3061 int pc = 0x0; // Program counter
3062 char starts_bundle = ' ';
3063 _regalloc->dump_frame();
3064
3065 Node *n = NULL;
3066 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
3067 if (VMThread::should_terminate()) {
3068 cut_short = true;
3069 break;
3070 }
3071 Block* block = _cfg->get_block(i);
3072 if (block->is_connector() && !Verbose) {
3073 continue;
3074 }
3075 n = block->head();
3076 if (pcs && n->_idx < pc_limit) {
3077 tty->print("%3.3x ", pcs[n->_idx]);
3078 } else {
3079 tty->print(" ");
3080 }
3081 block->dump_head(_cfg);
3082 if (block->is_connector()) {
3083 tty->print_cr(" # Empty connector block");
3084 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3085 tty->print_cr(" # Block is sole successor of call");
3086 }
3087
3088 // For all instructions
3089 Node *delay = NULL;
3090 for (uint j = 0; j < block->number_of_nodes(); j++) {
3091 if (VMThread::should_terminate()) {
3092 cut_short = true;
3093 break;
3094 }
3095 n = block->get_node(j);
3096 if (valid_bundle_info(n)) {
3097 Bundle* bundle = node_bundling(n);
3098 if (bundle->used_in_unconditional_delay()) {
3099 delay = n;
3100 continue;
3101 }
3102 if (bundle->starts_bundle()) {
3103 starts_bundle = '+';
3104 }
3105 }
3106
3107 if (WizardMode) {
3108 n->dump();
3109 }
3110
3111 if( !n->is_Region() && // Dont print in the Assembly
3112 !n->is_Phi() && // a few noisely useless nodes
3113 !n->is_Proj() &&
3114 !n->is_MachTemp() &&
3115 !n->is_SafePointScalarObject() &&
3116 !n->is_Catch() && // Would be nice to print exception table targets
3117 !n->is_MergeMem() && // Not very interesting
3118 !n->is_top() && // Debug info table constants
3119 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3120 ) {
3121 if (pcs && n->_idx < pc_limit)
3122 tty->print("%3.3x", pcs[n->_idx]);
3123 else
3124 tty->print(" ");
3125 tty->print(" %c ", starts_bundle);
3126 starts_bundle = ' ';
3127 tty->print("\t");
3128 n->format(_regalloc, tty);
3129 tty->cr();
3130 }
3131
3132 // If we have an instruction with a delay slot, and have seen a delay,
3133 // then back up and print it
3134 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3135 assert(delay != NULL, "no unconditional delay instruction");
3136 if (WizardMode) delay->dump();
3137
3138 if (node_bundling(delay)->starts_bundle())
3139 starts_bundle = '+';
3140 if (pcs && n->_idx < pc_limit)
3141 tty->print("%3.3x", pcs[n->_idx]);
3142 else
3143 tty->print(" ");
3144 tty->print(" %c ", starts_bundle);
3145 starts_bundle = ' ';
3146 tty->print("\t");
3147 delay->format(_regalloc, tty);
3148 tty->cr();
3149 delay = NULL;
3150 }
3151
3152 // Dump the exception table as well
3153 if( n->is_Catch() && (Verbose || WizardMode) ) {
3154 // Print the exception table for this offset
3155 _handler_table.print_subtable_for(pc);
3156 }
3157 }
3158
3159 if (pcs && n->_idx < pc_limit)
3160 tty->print_cr("%3.3x", pcs[n->_idx]);
3161 else
3162 tty->cr();
3163
3164 assert(cut_short || delay == NULL, "no unconditional delay branch");
3165
3166 } // End of per-block dump
3167 tty->cr();
3168
3169 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
3170 }
3171 #endif
3172
3173 //------------------------------Final_Reshape_Counts---------------------------
3174 // This class defines counters to help identify when a method
3175 // may/must be executed using hardware with only 24-bit precision.
3176 struct Final_Reshape_Counts : public StackObj {
3177 int _call_count; // count non-inlined 'common' calls
3178 int _float_count; // count float ops requiring 24-bit precision
3179 int _double_count; // count double ops requiring more precision
3180 int _java_call_count; // count non-inlined 'java' calls
3181 int _inner_loop_count; // count loops which need alignment
3182 VectorSet _visited; // Visitation flags
3183 Node_List _tests; // Set of IfNodes & PCTableNodes
3184
3185 Final_Reshape_Counts() :
3186 _call_count(0), _float_count(0), _double_count(0),
3187 _java_call_count(0), _inner_loop_count(0),
3188 _visited( Thread::current()->resource_area() ) { }
3189
3190 void inc_call_count () { _call_count ++; }
3191 void inc_float_count () { _float_count ++; }
3192 void inc_double_count() { _double_count++; }
3193 void inc_java_call_count() { _java_call_count++; }
3194 void inc_inner_loop_count() { _inner_loop_count++; }
3195
3196 int get_call_count () const { return _call_count ; }
3197 int get_float_count () const { return _float_count ; }
3198 int get_double_count() const { return _double_count; }
3199 int get_java_call_count() const { return _java_call_count; }
3200 int get_inner_loop_count() const { return _inner_loop_count; }
3201 };
3202
3203 #ifdef ASSERT
3204 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
3205 ciInstanceKlass *k = tp->klass()->as_instance_klass();
3206 // Make sure the offset goes inside the instance layout.
3207 return k->contains_field_offset(tp->offset());
3208 // Note that OffsetBot and OffsetTop are very negative.
3209 }
3210 #endif
3211
3212 // Eliminate trivially redundant StoreCMs and accumulate their
3213 // precedence edges.
3214 void Compile::eliminate_redundant_card_marks(Node* n) {
3215 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3216 if (n->in(MemNode::Address)->outcnt() > 1) {
3217 // There are multiple users of the same address so it might be
3218 // possible to eliminate some of the StoreCMs
3219 Node* mem = n->in(MemNode::Memory);
3220 Node* adr = n->in(MemNode::Address);
3221 Node* val = n->in(MemNode::ValueIn);
3222 Node* prev = n;
3223 bool done = false;
3224 // Walk the chain of StoreCMs eliminating ones that match. As
3225 // long as it's a chain of single users then the optimization is
3226 // safe. Eliminating partially redundant StoreCMs would require
3227 // cloning copies down the other paths.
3228 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3229 if (adr == mem->in(MemNode::Address) &&
3230 val == mem->in(MemNode::ValueIn)) {
3231 // redundant StoreCM
3232 if (mem->req() > MemNode::OopStore) {
3233 // Hasn't been processed by this code yet.
3234 n->add_prec(mem->in(MemNode::OopStore));
3235 } else {
3236 // Already converted to precedence edge
3237 for (uint i = mem->req(); i < mem->len(); i++) {
3238 // Accumulate any precedence edges
3239 if (mem->in(i) != NULL) {
3240 n->add_prec(mem->in(i));
3241 }
3242 }
3243 // Everything above this point has been processed.
3244 done = true;
3245 }
3246 // Eliminate the previous StoreCM
3247 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3248 assert(mem->outcnt() == 0, "should be dead");
3249 mem->disconnect_inputs(NULL, this);
3250 } else {
3251 prev = mem;
3252 }
3253 mem = prev->in(MemNode::Memory);
3254 }
3255 }
3256 }
3257
3258 //------------------------------final_graph_reshaping_impl----------------------
3259 // Implement items 1-5 from final_graph_reshaping below.
3260 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3261
3262 if ( n->outcnt() == 0 ) return; // dead node
3263 uint nop = n->Opcode();
3264
3265 // Check for 2-input instruction with "last use" on right input.
3266 // Swap to left input. Implements item (2).
3267 if( n->req() == 3 && // two-input instruction
3268 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3269 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3270 n->in(2)->outcnt() == 1 &&// right use IS a last use
3271 !n->in(2)->is_Con() ) { // right use is not a constant
3272 // Check for commutative opcode
3273 switch( nop ) {
3274 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3275 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3276 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3277 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3278 case Op_AndL: case Op_XorL: case Op_OrL:
3279 case Op_AndI: case Op_XorI: case Op_OrI: {
3280 // Move "last use" input to left by swapping inputs
3281 n->swap_edges(1, 2);
3282 break;
3283 }
3284 default:
3285 break;
3286 }
3287 }
3288
3289 #ifdef ASSERT
3290 if( n->is_Mem() ) {
3291 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3292 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3293 // oop will be recorded in oop map if load crosses safepoint
3294 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3295 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3296 "raw memory operations should have control edge");
3297 }
3298 if (n->is_MemBar()) {
3299 MemBarNode* mb = n->as_MemBar();
3300 if (mb->trailing_store() || mb->trailing_load_store()) {
3301 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3302 Node* mem = mb->in(MemBarNode::Precedent);
3303 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3304 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3305 } else if (mb->leading()) {
3306 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3307 }
3308 }
3309 #endif
3310 // Count FPU ops and common calls, implements item (3)
3311 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3312 if (!gc_handled) {
3313 final_graph_reshaping_main_switch(n, frc, nop);
3314 }
3315
3316 // Collect CFG split points
3317 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3318 frc._tests.push(n);
3319 }
3320 }
3321
3322 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3323 switch( nop ) {
3324 // Count all float operations that may use FPU
3325 case Op_AddF:
3326 case Op_SubF:
3327 case Op_MulF:
3328 case Op_DivF:
3329 case Op_NegF:
3330 case Op_ModF:
3331 case Op_ConvI2F:
3332 case Op_ConF:
3333 case Op_CmpF:
3334 case Op_CmpF3:
3335 // case Op_ConvL2F: // longs are split into 32-bit halves
3336 frc.inc_float_count();
3337 break;
3338
3339 case Op_ConvF2D:
3340 case Op_ConvD2F:
3341 frc.inc_float_count();
3342 frc.inc_double_count();
3343 break;
3344
3345 // Count all double operations that may use FPU
3346 case Op_AddD:
3347 case Op_SubD:
3348 case Op_MulD:
3349 case Op_DivD:
3350 case Op_NegD:
3351 case Op_ModD:
3352 case Op_ConvI2D:
3353 case Op_ConvD2I:
3354 // case Op_ConvL2D: // handled by leaf call
3355 // case Op_ConvD2L: // handled by leaf call
3356 case Op_ConD:
3357 case Op_CmpD:
3358 case Op_CmpD3:
3359 frc.inc_double_count();
3360 break;
3361 case Op_Opaque1: // Remove Opaque Nodes before matching
3362 case Op_Opaque2: // Remove Opaque Nodes before matching
3363 case Op_Opaque3:
3364 n->subsume_by(n->in(1), this);
3365 break;
3366 case Op_CallStaticJava:
3367 case Op_CallJava:
3368 case Op_CallDynamicJava:
3369 frc.inc_java_call_count(); // Count java call site;
3370 case Op_CallRuntime:
3371 case Op_CallLeaf:
3372 case Op_CallLeafVector:
3373 case Op_CallLeafNoFP: {
3374 assert (n->is_Call(), "");
3375 CallNode *call = n->as_Call();
3376 // Count call sites where the FP mode bit would have to be flipped.
3377 // Do not count uncommon runtime calls:
3378 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3379 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3380 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3381 frc.inc_call_count(); // Count the call site
3382 } else { // See if uncommon argument is shared
3383 Node *n = call->in(TypeFunc::Parms);
3384 int nop = n->Opcode();
3385 // Clone shared simple arguments to uncommon calls, item (1).
3386 if (n->outcnt() > 1 &&
3387 !n->is_Proj() &&
3388 nop != Op_CreateEx &&
3389 nop != Op_CheckCastPP &&
3390 nop != Op_DecodeN &&
3391 nop != Op_DecodeNKlass &&
3392 !n->is_Mem() &&
3393 !n->is_Phi()) {
3394 Node *x = n->clone();
3395 call->set_req(TypeFunc::Parms, x);
3396 }
3397 }
3398 break;
3399 }
3400
3401 case Op_StoreD:
3402 case Op_LoadD:
3403 case Op_LoadD_unaligned:
3404 frc.inc_double_count();
3405 goto handle_mem;
3406 case Op_StoreF:
3407 case Op_LoadF:
3408 frc.inc_float_count();
3409 goto handle_mem;
3410
3411 case Op_StoreCM:
3412 {
3413 // Convert OopStore dependence into precedence edge
3414 Node* prec = n->in(MemNode::OopStore);
3415 n->del_req(MemNode::OopStore);
3416 n->add_prec(prec);
3417 eliminate_redundant_card_marks(n);
3418 }
3419
3420 // fall through
3421
3422 case Op_StoreB:
3423 case Op_StoreC:
3424 case Op_StorePConditional:
3425 case Op_StoreI:
3426 case Op_StoreL:
3427 case Op_StoreIConditional:
3428 case Op_StoreLConditional:
3429 case Op_CompareAndSwapB:
3430 case Op_CompareAndSwapS:
3431 case Op_CompareAndSwapI:
3432 case Op_CompareAndSwapL:
3433 case Op_CompareAndSwapP:
3434 case Op_CompareAndSwapN:
3435 case Op_WeakCompareAndSwapB:
3436 case Op_WeakCompareAndSwapS:
3437 case Op_WeakCompareAndSwapI:
3438 case Op_WeakCompareAndSwapL:
3439 case Op_WeakCompareAndSwapP:
3440 case Op_WeakCompareAndSwapN:
3441 case Op_CompareAndExchangeB:
3442 case Op_CompareAndExchangeS:
3443 case Op_CompareAndExchangeI:
3444 case Op_CompareAndExchangeL:
3445 case Op_CompareAndExchangeP:
3446 case Op_CompareAndExchangeN:
3447 case Op_GetAndAddS:
3448 case Op_GetAndAddB:
3449 case Op_GetAndAddI:
3450 case Op_GetAndAddL:
3451 case Op_GetAndSetS:
3452 case Op_GetAndSetB:
3453 case Op_GetAndSetI:
3454 case Op_GetAndSetL:
3455 case Op_GetAndSetP:
3456 case Op_GetAndSetN:
3457 case Op_StoreP:
3458 case Op_StoreN:
3459 case Op_StoreNKlass:
3460 case Op_LoadB:
3461 case Op_LoadUB:
3462 case Op_LoadUS:
3463 case Op_LoadI:
3464 case Op_LoadKlass:
3465 case Op_LoadNKlass:
3466 case Op_LoadL:
3467 case Op_LoadL_unaligned:
3468 case Op_LoadPLocked:
3469 case Op_LoadP:
3470 case Op_LoadN:
3471 case Op_LoadRange:
3472 case Op_LoadS: {
3473 handle_mem:
3474 #ifdef ASSERT
3475 if( VerifyOptoOopOffsets ) {
3476 MemNode* mem = n->as_Mem();
3477 // Check to see if address types have grounded out somehow.
3478 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3479 assert( !tp || oop_offset_is_sane(tp), "" );
3480 }
3481 #endif
3482 break;
3483 }
3484
3485 case Op_AddP: { // Assert sane base pointers
3486 Node *addp = n->in(AddPNode::Address);
3487 assert( !addp->is_AddP() ||
3488 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3489 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3490 "Base pointers must match (addp %u)", addp->_idx );
3491 #ifdef _LP64
3492 if ((UseCompressedOops || UseCompressedClassPointers) &&
3493 addp->Opcode() == Op_ConP &&
3494 addp == n->in(AddPNode::Base) &&
3495 n->in(AddPNode::Offset)->is_Con()) {
3496 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3497 // on the platform and on the compressed oops mode.
3498 // Use addressing with narrow klass to load with offset on x86.
3499 // Some platforms can use the constant pool to load ConP.
3500 // Do this transformation here since IGVN will convert ConN back to ConP.
3501 const Type* t = addp->bottom_type();
3502 bool is_oop = t->isa_oopptr() != NULL;
3503 bool is_klass = t->isa_klassptr() != NULL;
3504
3505 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3506 (is_klass && Matcher::const_klass_prefer_decode())) {
3507 Node* nn = NULL;
3508
3509 int op = is_oop ? Op_ConN : Op_ConNKlass;
3510
3511 // Look for existing ConN node of the same exact type.
3512 Node* r = root();
3513 uint cnt = r->outcnt();
3514 for (uint i = 0; i < cnt; i++) {
3515 Node* m = r->raw_out(i);
3516 if (m!= NULL && m->Opcode() == op &&
3517 m->bottom_type()->make_ptr() == t) {
3518 nn = m;
3519 break;
3520 }
3521 }
3522 if (nn != NULL) {
3523 // Decode a narrow oop to match address
3524 // [R12 + narrow_oop_reg<<3 + offset]
3525 if (is_oop) {
3526 nn = new DecodeNNode(nn, t);
3527 } else {
3528 nn = new DecodeNKlassNode(nn, t);
3529 }
3530 // Check for succeeding AddP which uses the same Base.
3531 // Otherwise we will run into the assertion above when visiting that guy.
3532 for (uint i = 0; i < n->outcnt(); ++i) {
3533 Node *out_i = n->raw_out(i);
3534 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3535 out_i->set_req(AddPNode::Base, nn);
3536 #ifdef ASSERT
3537 for (uint j = 0; j < out_i->outcnt(); ++j) {
3538 Node *out_j = out_i->raw_out(j);
3539 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3540 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3541 }
3542 #endif
3543 }
3544 }
3545 n->set_req(AddPNode::Base, nn);
3546 n->set_req(AddPNode::Address, nn);
3547 if (addp->outcnt() == 0) {
3548 addp->disconnect_inputs(NULL, this);
3549 }
3550 }
3551 }
3552 }
3553 #endif
3554 // platform dependent reshaping of the address expression
3555 reshape_address(n->as_AddP());
3556 break;
3557 }
3558
3559 case Op_CastPP: {
3560 // Remove CastPP nodes to gain more freedom during scheduling but
3561 // keep the dependency they encode as control or precedence edges
3562 // (if control is set already) on memory operations. Some CastPP
3563 // nodes don't have a control (don't carry a dependency): skip
3564 // those.
3565 if (n->in(0) != NULL) {
3566 ResourceMark rm;
3567 Unique_Node_List wq;
3568 wq.push(n);
3569 for (uint next = 0; next < wq.size(); ++next) {
3570 Node *m = wq.at(next);
3571 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3572 Node* use = m->fast_out(i);
3573 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) {
3574 use->ensure_control_or_add_prec(n->in(0));
3575 } else {
3576 switch(use->Opcode()) {
3577 case Op_AddP:
3578 case Op_DecodeN:
3579 case Op_DecodeNKlass:
3580 case Op_CheckCastPP:
3581 case Op_CastPP:
3582 wq.push(use);
3583 break;
3584 }
3585 }
3586 }
3587 }
3588 }
3589 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3590 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3591 Node* in1 = n->in(1);
3592 const Type* t = n->bottom_type();
3593 Node* new_in1 = in1->clone();
3594 new_in1->as_DecodeN()->set_type(t);
3595
3596 if (!Matcher::narrow_oop_use_complex_address()) {
3597 //
3598 // x86, ARM and friends can handle 2 adds in addressing mode
3599 // and Matcher can fold a DecodeN node into address by using
3600 // a narrow oop directly and do implicit NULL check in address:
3601 //
3602 // [R12 + narrow_oop_reg<<3 + offset]
3603 // NullCheck narrow_oop_reg
3604 //
3605 // On other platforms (Sparc) we have to keep new DecodeN node and
3606 // use it to do implicit NULL check in address:
3607 //
3608 // decode_not_null narrow_oop_reg, base_reg
3609 // [base_reg + offset]
3610 // NullCheck base_reg
3611 //
3612 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3613 // to keep the information to which NULL check the new DecodeN node
3614 // corresponds to use it as value in implicit_null_check().
3615 //
3616 new_in1->set_req(0, n->in(0));
3617 }
3618
3619 n->subsume_by(new_in1, this);
3620 if (in1->outcnt() == 0) {
3621 in1->disconnect_inputs(NULL, this);
3622 }
3623 } else {
3624 n->subsume_by(n->in(1), this);
3625 if (n->outcnt() == 0) {
3626 n->disconnect_inputs(NULL, this);
3627 }
3628 }
3629 break;
3630 }
3631 #ifdef _LP64
3632 case Op_CmpP:
3633 // Do this transformation here to preserve CmpPNode::sub() and
3634 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3635 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3636 Node* in1 = n->in(1);
3637 Node* in2 = n->in(2);
3638 if (!in1->is_DecodeNarrowPtr()) {
3639 in2 = in1;
3640 in1 = n->in(2);
3641 }
3642 assert(in1->is_DecodeNarrowPtr(), "sanity");
3643
3644 Node* new_in2 = NULL;
3645 if (in2->is_DecodeNarrowPtr()) {
3646 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3647 new_in2 = in2->in(1);
3648 } else if (in2->Opcode() == Op_ConP) {
3649 const Type* t = in2->bottom_type();
3650 if (t == TypePtr::NULL_PTR) {
3651 assert(in1->is_DecodeN(), "compare klass to null?");
3652 // Don't convert CmpP null check into CmpN if compressed
3653 // oops implicit null check is not generated.
3654 // This will allow to generate normal oop implicit null check.
3655 if (Matcher::gen_narrow_oop_implicit_null_checks())
3656 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3657 //
3658 // This transformation together with CastPP transformation above
3659 // will generated code for implicit NULL checks for compressed oops.
3660 //
3661 // The original code after Optimize()
3662 //
3663 // LoadN memory, narrow_oop_reg
3664 // decode narrow_oop_reg, base_reg
3665 // CmpP base_reg, NULL
3666 // CastPP base_reg // NotNull
3667 // Load [base_reg + offset], val_reg
3668 //
3669 // after these transformations will be
3670 //
3671 // LoadN memory, narrow_oop_reg
3672 // CmpN narrow_oop_reg, NULL
3673 // decode_not_null narrow_oop_reg, base_reg
3674 // Load [base_reg + offset], val_reg
3675 //
3676 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3677 // since narrow oops can be used in debug info now (see the code in
3678 // final_graph_reshaping_walk()).
3679 //
3680 // At the end the code will be matched to
3681 // on x86:
3682 //
3683 // Load_narrow_oop memory, narrow_oop_reg
3684 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3685 // NullCheck narrow_oop_reg
3686 //
3687 // and on sparc:
3688 //
3689 // Load_narrow_oop memory, narrow_oop_reg
3690 // decode_not_null narrow_oop_reg, base_reg
3691 // Load [base_reg + offset], val_reg
3692 // NullCheck base_reg
3693 //
3694 } else if (t->isa_oopptr()) {
3695 new_in2 = ConNode::make(t->make_narrowoop());
3696 } else if (t->isa_klassptr()) {
3697 new_in2 = ConNode::make(t->make_narrowklass());
3698 }
3699 }
3700 if (new_in2 != NULL) {
3701 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3702 n->subsume_by(cmpN, this);
3703 if (in1->outcnt() == 0) {
3704 in1->disconnect_inputs(NULL, this);
3705 }
3706 if (in2->outcnt() == 0) {
3707 in2->disconnect_inputs(NULL, this);
3708 }
3709 }
3710 }
3711 break;
3712
3713 case Op_DecodeN:
3714 case Op_DecodeNKlass:
3715 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3716 // DecodeN could be pinned when it can't be fold into
3717 // an address expression, see the code for Op_CastPP above.
3718 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3719 break;
3720
3721 case Op_EncodeP:
3722 case Op_EncodePKlass: {
3723 Node* in1 = n->in(1);
3724 if (in1->is_DecodeNarrowPtr()) {
3725 n->subsume_by(in1->in(1), this);
3726 } else if (in1->Opcode() == Op_ConP) {
3727 const Type* t = in1->bottom_type();
3728 if (t == TypePtr::NULL_PTR) {
3729 assert(t->isa_oopptr(), "null klass?");
3730 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3731 } else if (t->isa_oopptr()) {
3732 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3733 } else if (t->isa_klassptr()) {
3734 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3735 }
3736 }
3737 if (in1->outcnt() == 0) {
3738 in1->disconnect_inputs(NULL, this);
3739 }
3740 break;
3741 }
3742
3743 case Op_Proj: {
3744 if (OptimizeStringConcat) {
3745 ProjNode* p = n->as_Proj();
3746 if (p->_is_io_use) {
3747 // Separate projections were used for the exception path which
3748 // are normally removed by a late inline. If it wasn't inlined
3749 // then they will hang around and should just be replaced with
3750 // the original one.
3751 Node* proj = NULL;
3752 // Replace with just one
3753 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3754 Node *use = i.get();
3755 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3756 proj = use;
3757 break;
3758 }
3759 }
3760 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3761 if (proj != NULL) {
3762 p->subsume_by(proj, this);
3763 }
3764 }
3765 }
3766 break;
3767 }
3768
3769 case Op_Phi:
3770 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3771 // The EncodeP optimization may create Phi with the same edges
3772 // for all paths. It is not handled well by Register Allocator.
3773 Node* unique_in = n->in(1);
3774 assert(unique_in != NULL, "");
3775 uint cnt = n->req();
3776 for (uint i = 2; i < cnt; i++) {
3777 Node* m = n->in(i);
3778 assert(m != NULL, "");
3779 if (unique_in != m)
3780 unique_in = NULL;
3781 }
3782 if (unique_in != NULL) {
3783 n->subsume_by(unique_in, this);
3784 }
3785 }
3786 break;
3787
3788 #endif
3789
3790 #ifdef ASSERT
3791 case Op_CastII:
3792 // Verify that all range check dependent CastII nodes were removed.
3793 if (n->isa_CastII()->has_range_check()) {
3794 n->dump(3);
3795 assert(false, "Range check dependent CastII node was not removed");
3796 }
3797 break;
3798 #endif
3799
3800 case Op_ModI:
3801 if (UseDivMod) {
3802 // Check if a%b and a/b both exist
3803 Node* d = n->find_similar(Op_DivI);
3804 if (d) {
3805 // Replace them with a fused divmod if supported
3806 if (Matcher::has_match_rule(Op_DivModI)) {
3807 DivModINode* divmod = DivModINode::make(n);
3808 d->subsume_by(divmod->div_proj(), this);
3809 n->subsume_by(divmod->mod_proj(), this);
3810 } else {
3811 // replace a%b with a-((a/b)*b)
3812 Node* mult = new MulINode(d, d->in(2));
3813 Node* sub = new SubINode(d->in(1), mult);
3814 n->subsume_by(sub, this);
3815 }
3816 }
3817 }
3818 break;
3819
3820 case Op_ModL:
3821 if (UseDivMod) {
3822 // Check if a%b and a/b both exist
3823 Node* d = n->find_similar(Op_DivL);
3824 if (d) {
3825 // Replace them with a fused divmod if supported
3826 if (Matcher::has_match_rule(Op_DivModL)) {
3827 DivModLNode* divmod = DivModLNode::make(n);
3828 d->subsume_by(divmod->div_proj(), this);
3829 n->subsume_by(divmod->mod_proj(), this);
3830 } else {
3831 // replace a%b with a-((a/b)*b)
3832 Node* mult = new MulLNode(d, d->in(2));
3833 Node* sub = new SubLNode(d->in(1), mult);
3834 n->subsume_by(sub, this);
3835 }
3836 }
3837 }
3838 break;
3839
3840 case Op_LoadVector:
3841 case Op_StoreVector:
3842 case Op_LoadVectorGather:
3843 case Op_StoreVectorScatter:
3844 break;
3845
3846 case Op_AddReductionVI:
3847 case Op_AddReductionVL:
3848 case Op_AddReductionVF:
3849 case Op_AddReductionVD:
3850 case Op_MulReductionVI:
3851 case Op_MulReductionVL:
3852 case Op_MulReductionVF:
3853 case Op_MulReductionVD:
3854 case Op_MinReductionV:
3855 case Op_MaxReductionV:
3856 case Op_AndReductionV:
3857 case Op_OrReductionV:
3858 case Op_XorReductionV:
3859 case Op_SubReductionV:
3860 case Op_SubReductionVFP:
3861 break;
3862
3863 case Op_PackB:
3864 case Op_PackS:
3865 case Op_PackI:
3866 case Op_PackF:
3867 case Op_PackL:
3868 case Op_PackD:
3869 if (n->req()-1 > 2) {
3870 // Replace many operand PackNodes with a binary tree for matching
3871 PackNode* p = (PackNode*) n;
3872 Node* btp = p->binary_tree_pack(1, n->req());
3873 n->subsume_by(btp, this);
3874 }
3875 break;
3876 case Op_Loop:
3877 case Op_CountedLoop:
3878 case Op_OuterStripMinedLoop:
3879 if (n->as_Loop()->is_inner_loop()) {
3880 frc.inc_inner_loop_count();
3881 }
3882 n->as_Loop()->verify_strip_mined(0);
3883 break;
3884 case Op_LShiftI:
3885 case Op_RShiftI:
3886 case Op_URShiftI:
3887 case Op_LShiftL:
3888 case Op_RShiftL:
3889 case Op_URShiftL:
3890 if (Matcher::need_masked_shift_count) {
3891 // The cpu's shift instructions don't restrict the count to the
3892 // lower 5/6 bits. We need to do the masking ourselves.
3893 Node* in2 = n->in(2);
3894 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3895 const TypeInt* t = in2->find_int_type();
3896 if (t != NULL && t->is_con()) {
3897 juint shift = t->get_con();
3898 if (shift > mask) { // Unsigned cmp
3899 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3900 }
3901 } else {
3902 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3903 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3904 n->set_req(2, shift);
3905 }
3906 }
3907 if (in2->outcnt() == 0) { // Remove dead node
3908 in2->disconnect_inputs(NULL, this);
3909 }
3910 }
3911 break;
3912 case Op_MemBarStoreStore:
3913 case Op_MemBarRelease:
3914 // Break the link with AllocateNode: it is no longer useful and
3915 // confuses register allocation.
3916 if (n->req() > MemBarNode::Precedent) {
3917 n->set_req(MemBarNode::Precedent, top());
3918 }
3919 break;
3920 case Op_MemBarAcquire: {
3921 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3922 // At parse time, the trailing MemBarAcquire for a volatile load
3923 // is created with an edge to the load. After optimizations,
3924 // that input may be a chain of Phis. If those phis have no
3925 // other use, then the MemBarAcquire keeps them alive and
3926 // register allocation can be confused.
3927 ResourceMark rm;
3928 Unique_Node_List wq;
3929 wq.push(n->in(MemBarNode::Precedent));
3930 n->set_req(MemBarNode::Precedent, top());
3931 while (wq.size() > 0) {
3932 Node* m = wq.pop();
3933 if (m->outcnt() == 0) {
3934 for (uint j = 0; j < m->req(); j++) {
3935 Node* in = m->in(j);
3936 if (in != NULL) {
3937 wq.push(in);
3938 }
3939 }
3940 m->disconnect_inputs(NULL, this);
3941 }
3942 }
3943 }
3944 break;
3945 }
3946 case Op_RangeCheck: {
3947 RangeCheckNode* rc = n->as_RangeCheck();
3948 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3949 n->subsume_by(iff, this);
3950 frc._tests.push(iff);
3951 break;
3952 }
3953 case Op_ConvI2L: {
3954 if (!Matcher::convi2l_type_required) {
3955 // Code generation on some platforms doesn't need accurate
3956 // ConvI2L types. Widening the type can help remove redundant
3957 // address computations.
3958 n->as_Type()->set_type(TypeLong::INT);
3959 ResourceMark rm;
3960 Node_List wq;
3961 wq.push(n);
3962 for (uint next = 0; next < wq.size(); next++) {
3963 Node *m = wq.at(next);
3964
3965 for(;;) {
3966 // Loop over all nodes with identical inputs edges as m
3967 Node* k = m->find_similar(m->Opcode());
3968 if (k == NULL) {
3969 break;
3970 }
3971 // Push their uses so we get a chance to remove node made
3972 // redundant
3973 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3974 Node* u = k->fast_out(i);
3975 assert(!wq.contains(u), "shouldn't process one node several times");
3976 if (u->Opcode() == Op_LShiftL ||
3977 u->Opcode() == Op_AddL ||
3978 u->Opcode() == Op_SubL ||
3979 u->Opcode() == Op_AddP) {
3980 wq.push(u);
3981 }
3982 }
3983 // Replace all nodes with identical edges as m with m
3984 k->subsume_by(m, this);
3985 }
3986 }
3987 }
3988 break;
3989 }
3990 case Op_CmpUL: {
3991 if (!Matcher::has_match_rule(Op_CmpUL)) {
3992 // No support for unsigned long comparisons
3993 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3994 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3995 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3996 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3997 Node* andl = new AndLNode(orl, remove_sign_mask);
3998 Node* cmp = new CmpLNode(andl, n->in(2));
3999 n->subsume_by(cmp, this);
4000 }
4001 break;
4002 }
4003 default:
4004 assert(!n->is_Call(), "");
4005 assert(!n->is_Mem(), "");
4006 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
4007 break;
4008 }
4009 }
4010
4011 //------------------------------final_graph_reshaping_walk---------------------
4012 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
4013 // requires that the walk visits a node's inputs before visiting the node.
4014 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
4015 ResourceArea *area = Thread::current()->resource_area();
4016 Unique_Node_List sfpt(area);
4017
4018 frc._visited.set(root->_idx); // first, mark node as visited
4019 uint cnt = root->req();
4020 Node *n = root;
4021 uint i = 0;
4022 while (true) {
4023 if (i < cnt) {
4024 // Place all non-visited non-null inputs onto stack
4025 Node* m = n->in(i);
4026 ++i;
4027 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4028 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4029 // compute worst case interpreter size in case of a deoptimization
4030 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4031
4032 sfpt.push(m);
4033 }
4034 cnt = m->req();
4035 nstack.push(n, i); // put on stack parent and next input's index
4036 n = m;
4037 i = 0;
4038 }
4039 } else {
4040 // Now do post-visit work
4041 final_graph_reshaping_impl( n, frc );
4042 if (nstack.is_empty())
4043 break; // finished
4044 n = nstack.node(); // Get node from stack
4045 cnt = n->req();
4046 i = nstack.index();
4047 nstack.pop(); // Shift to the next node on stack
4048 }
4049 }
4050
4051 // Skip next transformation if compressed oops are not used.
4052 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4053 (!UseCompressedOops && !UseCompressedClassPointers))
4054 return;
4055
4056 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4057 // It could be done for an uncommon traps or any safepoints/calls
4058 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4059 while (sfpt.size() > 0) {
4060 n = sfpt.pop();
4061 JVMState *jvms = n->as_SafePoint()->jvms();
4062 assert(jvms != NULL, "sanity");
4063 int start = jvms->debug_start();
4064 int end = n->req();
4065 bool is_uncommon = (n->is_CallStaticJava() &&
4066 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4067 for (int j = start; j < end; j++) {
4068 Node* in = n->in(j);
4069 if (in->is_DecodeNarrowPtr()) {
4070 bool safe_to_skip = true;
4071 if (!is_uncommon ) {
4072 // Is it safe to skip?
4073 for (uint i = 0; i < in->outcnt(); i++) {
4074 Node* u = in->raw_out(i);
4075 if (!u->is_SafePoint() ||
4076 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4077 safe_to_skip = false;
4078 }
4079 }
4080 }
4081 if (safe_to_skip) {
4082 n->set_req(j, in->in(1));
4083 }
4084 if (in->outcnt() == 0) {
4085 in->disconnect_inputs(NULL, this);
4086 }
4087 }
4088 }
4089 }
4090 }
4091
4092 //------------------------------final_graph_reshaping--------------------------
4093 // Final Graph Reshaping.
4094 //
4095 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4096 // and not commoned up and forced early. Must come after regular
4097 // optimizations to avoid GVN undoing the cloning. Clone constant
4098 // inputs to Loop Phis; these will be split by the allocator anyways.
4099 // Remove Opaque nodes.
4100 // (2) Move last-uses by commutative operations to the left input to encourage
4101 // Intel update-in-place two-address operations and better register usage
4102 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4103 // calls canonicalizing them back.
4104 // (3) Count the number of double-precision FP ops, single-precision FP ops
4105 // and call sites. On Intel, we can get correct rounding either by
4106 // forcing singles to memory (requires extra stores and loads after each
4107 // FP bytecode) or we can set a rounding mode bit (requires setting and
4108 // clearing the mode bit around call sites). The mode bit is only used
4109 // if the relative frequency of single FP ops to calls is low enough.
4110 // This is a key transform for SPEC mpeg_audio.
4111 // (4) Detect infinite loops; blobs of code reachable from above but not
4112 // below. Several of the Code_Gen algorithms fail on such code shapes,
4113 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4114 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4115 // Detection is by looking for IfNodes where only 1 projection is
4116 // reachable from below or CatchNodes missing some targets.
4117 // (5) Assert for insane oop offsets in debug mode.
4118
4119 bool Compile::final_graph_reshaping() {
4120 // an infinite loop may have been eliminated by the optimizer,
4121 // in which case the graph will be empty.
4122 if (root()->req() == 1) {
4123 record_method_not_compilable("trivial infinite loop");
4124 return true;
4125 }
4126
4127 // Expensive nodes have their control input set to prevent the GVN
4128 // from freely commoning them. There's no GVN beyond this point so
4129 // no need to keep the control input. We want the expensive nodes to
4130 // be freely moved to the least frequent code path by gcm.
4131 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4132 for (int i = 0; i < expensive_count(); i++) {
4133 _expensive_nodes->at(i)->set_req(0, NULL);
4134 }
4135
4136 Final_Reshape_Counts frc;
4137
4138 // Visit everybody reachable!
4139 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4140 Node_Stack nstack(live_nodes() >> 1);
4141 final_graph_reshaping_walk(nstack, root(), frc);
4142
4143 // Check for unreachable (from below) code (i.e., infinite loops).
4144 for( uint i = 0; i < frc._tests.size(); i++ ) {
4145 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4146 // Get number of CFG targets.
4147 // Note that PCTables include exception targets after calls.
4148 uint required_outcnt = n->required_outcnt();
4149 if (n->outcnt() != required_outcnt) {
4150 // Check for a few special cases. Rethrow Nodes never take the
4151 // 'fall-thru' path, so expected kids is 1 less.
4152 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4153 if (n->in(0)->in(0)->is_Call()) {
4154 CallNode *call = n->in(0)->in(0)->as_Call();
4155 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4156 required_outcnt--; // Rethrow always has 1 less kid
4157 } else if (call->req() > TypeFunc::Parms &&
4158 call->is_CallDynamicJava()) {
4159 // Check for null receiver. In such case, the optimizer has
4160 // detected that the virtual call will always result in a null
4161 // pointer exception. The fall-through projection of this CatchNode
4162 // will not be populated.
4163 Node *arg0 = call->in(TypeFunc::Parms);
4164 if (arg0->is_Type() &&
4165 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4166 required_outcnt--;
4167 }
4168 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4169 call->req() > TypeFunc::Parms+1 &&
4170 call->is_CallStaticJava()) {
4171 // Check for negative array length. In such case, the optimizer has
4172 // detected that the allocation attempt will always result in an
4173 // exception. There is no fall-through projection of this CatchNode .
4174 Node *arg1 = call->in(TypeFunc::Parms+1);
4175 if (arg1->is_Type() &&
4176 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4177 required_outcnt--;
4178 }
4179 }
4180 }
4181 }
4182 // Recheck with a better notion of 'required_outcnt'
4183 if (n->outcnt() != required_outcnt) {
4184 record_method_not_compilable("malformed control flow");
4185 return true; // Not all targets reachable!
4186 }
4187 }
4188 // Check that I actually visited all kids. Unreached kids
4189 // must be infinite loops.
4190 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4191 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4192 record_method_not_compilable("infinite loop");
4193 return true; // Found unvisited kid; must be unreach
4194 }
4195
4196 // Here so verification code in final_graph_reshaping_walk()
4197 // always see an OuterStripMinedLoopEnd
4198 if (n->is_OuterStripMinedLoopEnd()) {
4199 IfNode* init_iff = n->as_If();
4200 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4201 n->subsume_by(iff, this);
4202 }
4203 }
4204
4205 // If original bytecodes contained a mixture of floats and doubles
4206 // check if the optimizer has made it homogenous, item (3).
4207 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
4208 frc.get_float_count() > 32 &&
4209 frc.get_double_count() == 0 &&
4210 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4211 set_24_bit_selection_and_mode( false, true );
4212 }
4213
4214 set_java_calls(frc.get_java_call_count());
4215 set_inner_loops(frc.get_inner_loop_count());
4216
4217 // No infinite loops, no reason to bail out.
4218 return false;
4219 }
4220
4221 //-----------------------------too_many_traps----------------------------------
4222 // Report if there are too many traps at the current method and bci.
4223 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4224 bool Compile::too_many_traps(ciMethod* method,
4225 int bci,
4226 Deoptimization::DeoptReason reason) {
4227 ciMethodData* md = method->method_data();
4228 if (md->is_empty()) {
4229 // Assume the trap has not occurred, or that it occurred only
4230 // because of a transient condition during start-up in the interpreter.
4231 return false;
4232 }
4233 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4234 if (md->has_trap_at(bci, m, reason) != 0) {
4235 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4236 // Also, if there are multiple reasons, or if there is no per-BCI record,
4237 // assume the worst.
4238 if (log())
4239 log()->elem("observe trap='%s' count='%d'",
4240 Deoptimization::trap_reason_name(reason),
4241 md->trap_count(reason));
4242 return true;
4243 } else {
4244 // Ignore method/bci and see if there have been too many globally.
4245 return too_many_traps(reason, md);
4246 }
4247 }
4248
4249 // Less-accurate variant which does not require a method and bci.
4250 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4251 ciMethodData* logmd) {
4252 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4253 // Too many traps globally.
4254 // Note that we use cumulative trap_count, not just md->trap_count.
4255 if (log()) {
4256 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4257 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4258 Deoptimization::trap_reason_name(reason),
4259 mcount, trap_count(reason));
4260 }
4261 return true;
4262 } else {
4263 // The coast is clear.
4264 return false;
4265 }
4266 }
4267
4268 //--------------------------too_many_recompiles--------------------------------
4269 // Report if there are too many recompiles at the current method and bci.
4270 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4271 // Is not eager to return true, since this will cause the compiler to use
4272 // Action_none for a trap point, to avoid too many recompilations.
4273 bool Compile::too_many_recompiles(ciMethod* method,
4274 int bci,
4275 Deoptimization::DeoptReason reason) {
4276 ciMethodData* md = method->method_data();
4277 if (md->is_empty()) {
4278 // Assume the trap has not occurred, or that it occurred only
4279 // because of a transient condition during start-up in the interpreter.
4280 return false;
4281 }
4282 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4283 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4284 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4285 Deoptimization::DeoptReason per_bc_reason
4286 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4287 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4288 if ((per_bc_reason == Deoptimization::Reason_none
4289 || md->has_trap_at(bci, m, reason) != 0)
4290 // The trap frequency measure we care about is the recompile count:
4291 && md->trap_recompiled_at(bci, m)
4292 && md->overflow_recompile_count() >= bc_cutoff) {
4293 // Do not emit a trap here if it has already caused recompilations.
4294 // Also, if there are multiple reasons, or if there is no per-BCI record,
4295 // assume the worst.
4296 if (log())
4297 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4298 Deoptimization::trap_reason_name(reason),
4299 md->trap_count(reason),
4300 md->overflow_recompile_count());
4301 return true;
4302 } else if (trap_count(reason) != 0
4303 && decompile_count() >= m_cutoff) {
4304 // Too many recompiles globally, and we have seen this sort of trap.
4305 // Use cumulative decompile_count, not just md->decompile_count.
4306 if (log())
4307 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4308 Deoptimization::trap_reason_name(reason),
4309 md->trap_count(reason), trap_count(reason),
4310 md->decompile_count(), decompile_count());
4311 return true;
4312 } else {
4313 // The coast is clear.
4314 return false;
4315 }
4316 }
4317
4318 // Compute when not to trap. Used by matching trap based nodes and
4319 // NullCheck optimization.
4320 void Compile::set_allowed_deopt_reasons() {
4321 _allowed_reasons = 0;
4322 if (is_method_compilation()) {
4323 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4324 assert(rs < BitsPerInt, "recode bit map");
4325 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4326 _allowed_reasons |= nth_bit(rs);
4327 }
4328 }
4329 }
4330 }
4331
4332 bool Compile::is_compiling_clinit_for(ciKlass* k) {
4333 ciMethod* root = method(); // the root method of compilation
4334 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit
4335 }
4336
4337 #ifndef PRODUCT
4338 //------------------------------verify_graph_edges---------------------------
4339 // Walk the Graph and verify that there is a one-to-one correspondence
4340 // between Use-Def edges and Def-Use edges in the graph.
4341 void Compile::verify_graph_edges(bool no_dead_code) {
4342 if (VerifyGraphEdges) {
4343 ResourceArea *area = Thread::current()->resource_area();
4344 Unique_Node_List visited(area);
4345 // Call recursive graph walk to check edges
4346 _root->verify_edges(visited);
4347 if (no_dead_code) {
4348 // Now make sure that no visited node is used by an unvisited node.
4349 bool dead_nodes = false;
4350 Unique_Node_List checked(area);
4351 while (visited.size() > 0) {
4352 Node* n = visited.pop();
4353 checked.push(n);
4354 for (uint i = 0; i < n->outcnt(); i++) {
4355 Node* use = n->raw_out(i);
4356 if (checked.member(use)) continue; // already checked
4357 if (visited.member(use)) continue; // already in the graph
4358 if (use->is_Con()) continue; // a dead ConNode is OK
4359 // At this point, we have found a dead node which is DU-reachable.
4360 if (!dead_nodes) {
4361 tty->print_cr("*** Dead nodes reachable via DU edges:");
4362 dead_nodes = true;
4363 }
4364 use->dump(2);
4365 tty->print_cr("---");
4366 checked.push(use); // No repeats; pretend it is now checked.
4367 }
4368 }
4369 assert(!dead_nodes, "using nodes must be reachable from root");
4370 }
4371 }
4372 }
4373 #endif
4374
4375 // The Compile object keeps track of failure reasons separately from the ciEnv.
4376 // This is required because there is not quite a 1-1 relation between the
4377 // ciEnv and its compilation task and the Compile object. Note that one
4378 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4379 // to backtrack and retry without subsuming loads. Other than this backtracking
4380 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4381 // by the logic in C2Compiler.
4382 void Compile::record_failure(const char* reason) {
4383 if (log() != NULL) {
4384 log()->elem("failure reason='%s' phase='compile'", reason);
4385 }
4386 if (_failure_reason == NULL) {
4387 // Record the first failure reason.
4388 _failure_reason = reason;
4389 }
4390
4391 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4392 C->print_method(PHASE_FAILURE);
4393 }
4394 _root = NULL; // flush the graph, too
4395 }
4396
4397 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4398 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4399 _phase_name(name), _dolog(CITimeVerbose)
4400 {
4401 if (_dolog) {
4402 C = Compile::current();
4403 _log = C->log();
4404 } else {
4405 C = NULL;
4406 _log = NULL;
4407 }
4408 if (_log != NULL) {
4409 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4410 _log->stamp();
4411 _log->end_head();
4412 }
4413 }
4414
4415 Compile::TracePhase::~TracePhase() {
4416
4417 C = Compile::current();
4418 if (_dolog) {
4419 _log = C->log();
4420 } else {
4421 _log = NULL;
4422 }
4423
4424 #ifdef ASSERT
4425 if (PrintIdealNodeCount) {
4426 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4427 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4428 }
4429
4430 if (VerifyIdealNodeCount) {
4431 Compile::current()->print_missing_nodes();
4432 }
4433 #endif
4434
4435 if (_log != NULL) {
4436 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4437 }
4438 }
4439
4440 //=============================================================================
4441 // Two Constant's are equal when the type and the value are equal.
4442 bool Compile::Constant::operator==(const Constant& other) {
4443 if (type() != other.type() ) return false;
4444 if (can_be_reused() != other.can_be_reused()) return false;
4445 // For floating point values we compare the bit pattern.
4446 switch (type()) {
4447 case T_INT:
4448 case T_FLOAT: return (_v._value.i == other._v._value.i);
4449 case T_LONG:
4450 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4451 case T_OBJECT:
4452 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4453 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4454 case T_METADATA: return (_v._metadata == other._v._metadata);
4455 default: ShouldNotReachHere(); return false;
4456 }
4457 }
4458
4459 static int type_to_size_in_bytes(BasicType t) {
4460 switch (t) {
4461 case T_INT: return sizeof(jint );
4462 case T_LONG: return sizeof(jlong );
4463 case T_FLOAT: return sizeof(jfloat );
4464 case T_DOUBLE: return sizeof(jdouble);
4465 case T_METADATA: return sizeof(Metadata*);
4466 // We use T_VOID as marker for jump-table entries (labels) which
4467 // need an internal word relocation.
4468 case T_VOID:
4469 case T_ADDRESS:
4470 case T_OBJECT: return sizeof(jobject);
4471 default:
4472 ShouldNotReachHere();
4473 return -1;
4474 }
4475 }
4476
4477 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4478 // sort descending
4479 if (a->freq() > b->freq()) return -1;
4480 if (a->freq() < b->freq()) return 1;
4481 return 0;
4482 }
4483
4484 void Compile::ConstantTable::calculate_offsets_and_size() {
4485 // First, sort the array by frequencies.
4486 _constants.sort(qsort_comparator);
4487
4488 #ifdef ASSERT
4489 // Make sure all jump-table entries were sorted to the end of the
4490 // array (they have a negative frequency).
4491 bool found_void = false;
4492 for (int i = 0; i < _constants.length(); i++) {
4493 Constant con = _constants.at(i);
4494 if (con.type() == T_VOID)
4495 found_void = true; // jump-tables
4496 else
4497 assert(!found_void, "wrong sorting");
4498 }
4499 #endif
4500
4501 int offset = 0;
4502 for (int i = 0; i < _constants.length(); i++) {
4503 Constant* con = _constants.adr_at(i);
4504
4505 // Align offset for type.
4506 int typesize = type_to_size_in_bytes(con->type());
4507 offset = align_up(offset, typesize);
4508 con->set_offset(offset); // set constant's offset
4509
4510 if (con->type() == T_VOID) {
4511 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4512 offset = offset + typesize * n->outcnt(); // expand jump-table
4513 } else {
4514 offset = offset + typesize;
4515 }
4516 }
4517
4518 // Align size up to the next section start (which is insts; see
4519 // CodeBuffer::align_at_start).
4520 assert(_size == -1, "already set?");
4521 _size = align_up(offset, (int)CodeEntryAlignment);
4522 }
4523
4524 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4525 MacroAssembler _masm(&cb);
4526 for (int i = 0; i < _constants.length(); i++) {
4527 Constant con = _constants.at(i);
4528 address constant_addr = NULL;
4529 switch (con.type()) {
4530 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4531 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4532 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4533 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4534 case T_OBJECT: {
4535 jobject obj = con.get_jobject();
4536 int oop_index = _masm.oop_recorder()->find_index(obj);
4537 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4538 break;
4539 }
4540 case T_ADDRESS: {
4541 address addr = (address) con.get_jobject();
4542 constant_addr = _masm.address_constant(addr);
4543 break;
4544 }
4545 // We use T_VOID as marker for jump-table entries (labels) which
4546 // need an internal word relocation.
4547 case T_VOID: {
4548 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4549 // Fill the jump-table with a dummy word. The real value is
4550 // filled in later in fill_jump_table.
4551 address dummy = (address) n;
4552 constant_addr = _masm.address_constant(dummy);
4553 // Expand jump-table
4554 for (uint i = 1; i < n->outcnt(); i++) {
4555 address temp_addr = _masm.address_constant(dummy + i);
4556 assert(temp_addr, "consts section too small");
4557 }
4558 break;
4559 }
4560 case T_METADATA: {
4561 Metadata* obj = con.get_metadata();
4562 int metadata_index = _masm.oop_recorder()->find_index(obj);
4563 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4564 break;
4565 }
4566 default: ShouldNotReachHere();
4567 }
4568 assert(constant_addr, "consts section too small");
4569 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4570 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4571 }
4572 }
4573
4574 int Compile::ConstantTable::find_offset(Constant& con) const {
4575 int idx = _constants.find(con);
4576 guarantee(idx != -1, "constant must be in constant table");
4577 int offset = _constants.at(idx).offset();
4578 guarantee(offset != -1, "constant table not emitted yet?");
4579 return offset;
4580 }
4581
4582 void Compile::ConstantTable::add(Constant& con) {
4583 if (con.can_be_reused()) {
4584 int idx = _constants.find(con);
4585 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4586 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4587 return;
4588 }
4589 }
4590 (void) _constants.append(con);
4591 }
4592
4593 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4594 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4595 Constant con(type, value, b->_freq);
4596 add(con);
4597 return con;
4598 }
4599
4600 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4601 Constant con(metadata);
4602 add(con);
4603 return con;
4604 }
4605
4606 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4607 jvalue value;
4608 BasicType type = oper->type()->basic_type();
4609 switch (type) {
4610 case T_LONG: value.j = oper->constantL(); break;
4611 case T_FLOAT: value.f = oper->constantF(); break;
4612 case T_DOUBLE: value.d = oper->constantD(); break;
4613 case T_OBJECT:
4614 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4615 case T_METADATA: return add((Metadata*)oper->constant()); break;
4616 default: guarantee(false, "unhandled type: %s", type2name(type));
4617 }
4618 return add(n, type, value);
4619 }
4620
4621 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4622 jvalue value;
4623 // We can use the node pointer here to identify the right jump-table
4624 // as this method is called from Compile::Fill_buffer right before
4625 // the MachNodes are emitted and the jump-table is filled (means the
4626 // MachNode pointers do not change anymore).
4627 value.l = (jobject) n;
4628 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4629 add(con);
4630 return con;
4631 }
4632
4633 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4634 // If called from Compile::scratch_emit_size do nothing.
4635 if (Compile::current()->in_scratch_emit_size()) return;
4636
4637 assert(labels.is_nonempty(), "must be");
4638 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4639
4640 // Since MachConstantNode::constant_offset() also contains
4641 // table_base_offset() we need to subtract the table_base_offset()
4642 // to get the plain offset into the constant table.
4643 int offset = n->constant_offset() - table_base_offset();
4644
4645 MacroAssembler _masm(&cb);
4646 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4647
4648 for (uint i = 0; i < n->outcnt(); i++) {
4649 address* constant_addr = &jump_table_base[i];
4650 assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i));
4651 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4652 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4653 }
4654 }
4655
4656 //----------------------------static_subtype_check-----------------------------
4657 // Shortcut important common cases when superklass is exact:
4658 // (0) superklass is java.lang.Object (can occur in reflective code)
4659 // (1) subklass is already limited to a subtype of superklass => always ok
4660 // (2) subklass does not overlap with superklass => always fail
4661 // (3) superklass has NO subtypes and we can check with a simple compare.
4662 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4663 if (StressReflectiveCode) {
4664 return SSC_full_test; // Let caller generate the general case.
4665 }
4666
4667 if (superk == env()->Object_klass()) {
4668 return SSC_always_true; // (0) this test cannot fail
4669 }
4670
4671 ciType* superelem = superk;
4672 if (superelem->is_array_klass())
4673 superelem = superelem->as_array_klass()->base_element_type();
4674
4675 if (!subk->is_interface()) { // cannot trust static interface types yet
4676 if (subk->is_subtype_of(superk)) {
4677 return SSC_always_true; // (1) false path dead; no dynamic test needed
4678 }
4679 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4680 !superk->is_subtype_of(subk)) {
4681 return SSC_always_false;
4682 }
4683 }
4684
4685 // If casting to an instance klass, it must have no subtypes
4686 if (superk->is_interface()) {
4687 // Cannot trust interfaces yet.
4688 // %%% S.B. superk->nof_implementors() == 1
4689 } else if (superelem->is_instance_klass()) {
4690 ciInstanceKlass* ik = superelem->as_instance_klass();
4691 if (!ik->has_subklass() && !ik->is_interface()) {
4692 if (!ik->is_final()) {
4693 // Add a dependency if there is a chance of a later subclass.
4694 dependencies()->assert_leaf_type(ik);
4695 }
4696 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4697 }
4698 } else {
4699 // A primitive array type has no subtypes.
4700 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4701 }
4702
4703 return SSC_full_test;
4704 }
4705
4706 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4707 #ifdef _LP64
4708 // The scaled index operand to AddP must be a clean 64-bit value.
4709 // Java allows a 32-bit int to be incremented to a negative
4710 // value, which appears in a 64-bit register as a large
4711 // positive number. Using that large positive number as an
4712 // operand in pointer arithmetic has bad consequences.
4713 // On the other hand, 32-bit overflow is rare, and the possibility
4714 // can often be excluded, if we annotate the ConvI2L node with
4715 // a type assertion that its value is known to be a small positive
4716 // number. (The prior range check has ensured this.)
4717 // This assertion is used by ConvI2LNode::Ideal.
4718 int index_max = max_jint - 1; // array size is max_jint, index is one less
4719 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4720 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4721 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4722 #endif
4723 return idx;
4724 }
4725
4726 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4727 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4728 if (ctrl != NULL) {
4729 // Express control dependency by a CastII node with a narrow type.
4730 value = new CastIINode(value, itype, false, true /* range check dependency */);
4731 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4732 // node from floating above the range check during loop optimizations. Otherwise, the
4733 // ConvI2L node may be eliminated independently of the range check, causing the data path
4734 // to become TOP while the control path is still there (although it's unreachable).
4735 value->set_req(0, ctrl);
4736 // Save CastII node to remove it after loop optimizations.
4737 phase->C->add_range_check_cast(value);
4738 value = phase->transform(value);
4739 }
4740 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4741 return phase->transform(new ConvI2LNode(value, ltype));
4742 }
4743
4744 // The message about the current inlining is accumulated in
4745 // _print_inlining_stream and transfered into the _print_inlining_list
4746 // once we know whether inlining succeeds or not. For regular
4747 // inlining, messages are appended to the buffer pointed by
4748 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4749 // a new buffer is added after _print_inlining_idx in the list. This
4750 // way we can update the inlining message for late inlining call site
4751 // when the inlining is attempted again.
4752 void Compile::print_inlining_init() {
4753 if (print_inlining() || print_intrinsics()) {
4754 _print_inlining_stream = new stringStream();
4755 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4756 }
4757 }
4758
4759 void Compile::print_inlining_reinit() {
4760 if (print_inlining() || print_intrinsics()) {
4761 // Re allocate buffer when we change ResourceMark
4762 _print_inlining_stream = new stringStream();
4763 }
4764 }
4765
4766 void Compile::print_inlining_reset() {
4767 _print_inlining_stream->reset();
4768 }
4769
4770 void Compile::print_inlining_commit() {
4771 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4772 // Transfer the message from _print_inlining_stream to the current
4773 // _print_inlining_list buffer and clear _print_inlining_stream.
4774 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4775 print_inlining_reset();
4776 }
4777
4778 void Compile::print_inlining_push() {
4779 // Add new buffer to the _print_inlining_list at current position
4780 _print_inlining_idx++;
4781 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4782 }
4783
4784 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4785 return _print_inlining_list->at(_print_inlining_idx);
4786 }
4787
4788 void Compile::print_inlining_update(CallGenerator* cg) {
4789 if (print_inlining() || print_intrinsics()) {
4790 if (!cg->is_late_inline()) {
4791 if (print_inlining_current().cg() != NULL) {
4792 print_inlining_push();
4793 }
4794 print_inlining_commit();
4795 } else {
4796 if (print_inlining_current().cg() != cg &&
4797 (print_inlining_current().cg() != NULL ||
4798 print_inlining_current().ss()->size() != 0)) {
4799 print_inlining_push();
4800 }
4801 print_inlining_commit();
4802 print_inlining_current().set_cg(cg);
4803 }
4804 }
4805 }
4806
4807 void Compile::print_inlining_move_to(CallGenerator* cg) {
4808 // We resume inlining at a late inlining call site. Locate the
4809 // corresponding inlining buffer so that we can update it.
4810 if (print_inlining()) {
4811 for (int i = 0; i < _print_inlining_list->length(); i++) {
4812 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4813 _print_inlining_idx = i;
4814 return;
4815 }
4816 }
4817 ShouldNotReachHere();
4818 }
4819 }
4820
4821 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4822 if (print_inlining()) {
4823 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4824 assert(print_inlining_current().cg() == cg, "wrong entry");
4825 // replace message with new message
4826 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4827 print_inlining_commit();
4828 print_inlining_current().set_cg(cg);
4829 }
4830 }
4831
4832 void Compile::print_inlining_assert_ready() {
4833 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4834 }
4835
4836 void Compile::process_print_inlining() {
4837 bool do_print_inlining = print_inlining() || print_intrinsics();
4838 if (do_print_inlining || log() != NULL) {
4839 // Print inlining message for candidates that we couldn't inline
4840 // for lack of space
4841 for (int i = 0; i < _late_inlines.length(); i++) {
4842 CallGenerator* cg = _late_inlines.at(i);
4843 if (!cg->is_mh_late_inline()) {
4844 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4845 if (do_print_inlining) {
4846 cg->print_inlining_late(msg);
4847 }
4848 log_late_inline_failure(cg, msg);
4849 }
4850 }
4851 }
4852 if (do_print_inlining) {
4853 ResourceMark rm;
4854 stringStream ss;
4855 for (int i = 0; i < _print_inlining_list->length(); i++) {
4856 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4857 }
4858 size_t end = ss.size();
4859 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4860 strncpy(_print_inlining_output, ss.base(), end+1);
4861 _print_inlining_output[end] = 0;
4862 }
4863 }
4864
4865 void Compile::dump_print_inlining() {
4866 if (_print_inlining_output != NULL) {
4867 tty->print_raw(_print_inlining_output);
4868 }
4869 }
4870
4871 void Compile::log_late_inline(CallGenerator* cg) {
4872 if (log() != NULL) {
4873 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4874 cg->unique_id());
4875 JVMState* p = cg->call_node()->jvms();
4876 while (p != NULL) {
4877 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4878 p = p->caller();
4879 }
4880 log()->tail("late_inline");
4881 }
4882 }
4883
4884 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4885 log_late_inline(cg);
4886 if (log() != NULL) {
4887 log()->inline_fail(msg);
4888 }
4889 }
4890
4891 void Compile::log_inline_id(CallGenerator* cg) {
4892 if (log() != NULL) {
4893 // The LogCompilation tool needs a unique way to identify late
4894 // inline call sites. This id must be unique for this call site in
4895 // this compilation. Try to have it unique across compilations as
4896 // well because it can be convenient when grepping through the log
4897 // file.
4898 // Distinguish OSR compilations from others in case CICountOSR is
4899 // on.
4900 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4901 cg->set_unique_id(id);
4902 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4903 }
4904 }
4905
4906 void Compile::log_inline_failure(const char* msg) {
4907 if (C->log() != NULL) {
4908 C->log()->inline_fail(msg);
4909 }
4910 }
4911
4912
4913 // Dump inlining replay data to the stream.
4914 // Don't change thread state and acquire any locks.
4915 void Compile::dump_inline_data(outputStream* out) {
4916 InlineTree* inl_tree = ilt();
4917 if (inl_tree != NULL) {
4918 out->print(" inline %d", inl_tree->count());
4919 inl_tree->dump_replay_data(out);
4920 }
4921 }
4922
4923 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4924 if (n1->Opcode() < n2->Opcode()) return -1;
4925 else if (n1->Opcode() > n2->Opcode()) return 1;
4926
4927 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4928 for (uint i = 1; i < n1->req(); i++) {
4929 if (n1->in(i) < n2->in(i)) return -1;
4930 else if (n1->in(i) > n2->in(i)) return 1;
4931 }
4932
4933 return 0;
4934 }
4935
4936 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4937 Node* n1 = *n1p;
4938 Node* n2 = *n2p;
4939
4940 return cmp_expensive_nodes(n1, n2);
4941 }
4942
4943 void Compile::sort_expensive_nodes() {
4944 if (!expensive_nodes_sorted()) {
4945 _expensive_nodes->sort(cmp_expensive_nodes);
4946 }
4947 }
4948
4949 bool Compile::expensive_nodes_sorted() const {
4950 for (int i = 1; i < _expensive_nodes->length(); i++) {
4951 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4952 return false;
4953 }
4954 }
4955 return true;
4956 }
4957
4958 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4959 if (_expensive_nodes->length() == 0) {
4960 return false;
4961 }
4962
4963 assert(OptimizeExpensiveOps, "optimization off?");
4964
4965 // Take this opportunity to remove dead nodes from the list
4966 int j = 0;
4967 for (int i = 0; i < _expensive_nodes->length(); i++) {
4968 Node* n = _expensive_nodes->at(i);
4969 if (!n->is_unreachable(igvn)) {
4970 assert(n->is_expensive(), "should be expensive");
4971 _expensive_nodes->at_put(j, n);
4972 j++;
4973 }
4974 }
4975 _expensive_nodes->trunc_to(j);
4976
4977 // Then sort the list so that similar nodes are next to each other
4978 // and check for at least two nodes of identical kind with same data
4979 // inputs.
4980 sort_expensive_nodes();
4981
4982 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4983 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4984 return true;
4985 }
4986 }
4987
4988 return false;
4989 }
4990
4991 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4992 if (_expensive_nodes->length() == 0) {
4993 return;
4994 }
4995
4996 assert(OptimizeExpensiveOps, "optimization off?");
4997
4998 // Sort to bring similar nodes next to each other and clear the
4999 // control input of nodes for which there's only a single copy.
5000 sort_expensive_nodes();
5001
5002 int j = 0;
5003 int identical = 0;
5004 int i = 0;
5005 bool modified = false;
5006 for (; i < _expensive_nodes->length()-1; i++) {
5007 assert(j <= i, "can't write beyond current index");
5008 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
5009 identical++;
5010 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5011 continue;
5012 }
5013 if (identical > 0) {
5014 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5015 identical = 0;
5016 } else {
5017 Node* n = _expensive_nodes->at(i);
5018 igvn.replace_input_of(n, 0, NULL);
5019 igvn.hash_insert(n);
5020 modified = true;
5021 }
5022 }
5023 if (identical > 0) {
5024 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5025 } else if (_expensive_nodes->length() >= 1) {
5026 Node* n = _expensive_nodes->at(i);
5027 igvn.replace_input_of(n, 0, NULL);
5028 igvn.hash_insert(n);
5029 modified = true;
5030 }
5031 _expensive_nodes->trunc_to(j);
5032 if (modified) {
5033 igvn.optimize();
5034 }
5035 }
5036
5037 void Compile::add_expensive_node(Node * n) {
5038 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
5039 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5040 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5041 if (OptimizeExpensiveOps) {
5042 _expensive_nodes->append(n);
5043 } else {
5044 // Clear control input and let IGVN optimize expensive nodes if
5045 // OptimizeExpensiveOps is off.
5046 n->set_req(0, NULL);
5047 }
5048 }
5049
5050 /**
5051 * Remove the speculative part of types and clean up the graph
5052 */
5053 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5054 if (UseTypeSpeculation) {
5055 Unique_Node_List worklist;
5056 worklist.push(root());
5057 int modified = 0;
5058 // Go over all type nodes that carry a speculative type, drop the
5059 // speculative part of the type and enqueue the node for an igvn
5060 // which may optimize it out.
5061 for (uint next = 0; next < worklist.size(); ++next) {
5062 Node *n = worklist.at(next);
5063 if (n->is_Type()) {
5064 TypeNode* tn = n->as_Type();
5065 const Type* t = tn->type();
5066 const Type* t_no_spec = t->remove_speculative();
5067 if (t_no_spec != t) {
5068 bool in_hash = igvn.hash_delete(n);
5069 assert(in_hash, "node should be in igvn hash table");
5070 tn->set_type(t_no_spec);
5071 igvn.hash_insert(n);
5072 igvn._worklist.push(n); // give it a chance to go away
5073 modified++;
5074 }
5075 }
5076 uint max = n->len();
5077 for( uint i = 0; i < max; ++i ) {
5078 Node *m = n->in(i);
5079 if (not_a_node(m)) continue;
5080 worklist.push(m);
5081 }
5082 }
5083 // Drop the speculative part of all types in the igvn's type table
5084 igvn.remove_speculative_types();
5085 if (modified > 0) {
5086 igvn.optimize();
5087 }
5088 #ifdef ASSERT
5089 // Verify that after the IGVN is over no speculative type has resurfaced
5090 worklist.clear();
5091 worklist.push(root());
5092 for (uint next = 0; next < worklist.size(); ++next) {
5093 Node *n = worklist.at(next);
5094 const Type* t = igvn.type_or_null(n);
5095 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5096 if (n->is_Type()) {
5097 t = n->as_Type()->type();
5098 assert(t == t->remove_speculative(), "no more speculative types");
5099 }
5100 uint max = n->len();
5101 for( uint i = 0; i < max; ++i ) {
5102 Node *m = n->in(i);
5103 if (not_a_node(m)) continue;
5104 worklist.push(m);
5105 }
5106 }
5107 igvn.check_no_speculative_types();
5108 #endif
5109 }
5110 }
5111
5112 // Auxiliary method to support randomized stressing/fuzzing.
5113 //
5114 // This method can be called the arbitrary number of times, with current count
5115 // as the argument. The logic allows selecting a single candidate from the
5116 // running list of candidates as follows:
5117 // int count = 0;
5118 // Cand* selected = null;
5119 // while(cand = cand->next()) {
5120 // if (randomized_select(++count)) {
5121 // selected = cand;
5122 // }
5123 // }
5124 //
5125 // Including count equalizes the chances any candidate is "selected".
5126 // This is useful when we don't have the complete list of candidates to choose
5127 // from uniformly. In this case, we need to adjust the randomicity of the
5128 // selection, or else we will end up biasing the selection towards the latter
5129 // candidates.
5130 //
5131 // Quick back-envelope calculation shows that for the list of n candidates
5132 // the equal probability for the candidate to persist as "best" can be
5133 // achieved by replacing it with "next" k-th candidate with the probability
5134 // of 1/k. It can be easily shown that by the end of the run, the
5135 // probability for any candidate is converged to 1/n, thus giving the
5136 // uniform distribution among all the candidates.
5137 //
5138 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5139 #define RANDOMIZED_DOMAIN_POW 29
5140 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5141 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5142 bool Compile::randomized_select(int count) {
5143 assert(count > 0, "only positive");
5144 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5145 }
5146
5147 CloneMap& Compile::clone_map() { return _clone_map; }
5148 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5149
5150 void NodeCloneInfo::dump() const {
5151 tty->print(" {%d:%d} ", idx(), gen());
5152 }
5153
5154 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5155 uint64_t val = value(old->_idx);
5156 NodeCloneInfo cio(val);
5157 assert(val != 0, "old node should be in the map");
5158 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5159 insert(nnn->_idx, cin.get());
5160 #ifndef PRODUCT
5161 if (is_debug()) {
5162 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5163 }
5164 #endif
5165 }
5166
5167 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5168 NodeCloneInfo cio(value(old->_idx));
5169 if (cio.get() == 0) {
5170 cio.set(old->_idx, 0);
5171 insert(old->_idx, cio.get());
5172 #ifndef PRODUCT
5173 if (is_debug()) {
5174 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5175 }
5176 #endif
5177 }
5178 clone(old, nnn, gen);
5179 }
5180
5181 int CloneMap::max_gen() const {
5182 int g = 0;
5183 DictI di(_dict);
5184 for(; di.test(); ++di) {
5185 int t = gen(di._key);
5186 if (g < t) {
5187 g = t;
5188 #ifndef PRODUCT
5189 if (is_debug()) {
5190 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5191 }
5192 #endif
5193 }
5194 }
5195 return g;
5196 }
5197
5198 void CloneMap::dump(node_idx_t key) const {
5199 uint64_t val = value(key);
5200 if (val != 0) {
5201 NodeCloneInfo ni(val);
5202 ni.dump();
5203 }
5204 }