1 /*
2 * Copyright (c) 1997, 2019, 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/valuetypenode.hpp"
71 #include "opto/vectornode.hpp"
72 #include "runtime/arguments.hpp"
73 #include "runtime/sharedRuntime.hpp"
74 #include "runtime/signature.hpp"
75 #include "runtime/stubRoutines.hpp"
76 #include "runtime/timer.hpp"
77 #include "utilities/align.hpp"
78 #include "utilities/copy.hpp"
79 #include "utilities/macros.hpp"
80 #if INCLUDE_ZGC
81 #include "gc/z/c2/zBarrierSetC2.hpp"
82 #endif
83
84
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
94
95
96 /// Support for intrinsics.
97
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 class IntrinsicDescPair {
101 private:
102 ciMethod* _m;
103 bool _is_virtual;
104 public:
105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
107 ciMethod* m= elt->method();
108 ciMethod* key_m = key->_m;
109 if (key_m < m) return -1;
110 else if (key_m > m) return 1;
111 else {
112 bool is_virtual = elt->is_virtual();
113 bool key_virtual = key->_is_virtual;
114 if (key_virtual < is_virtual) return -1;
115 else if (key_virtual > is_virtual) return 1;
116 else return 0;
117 }
118 }
119 };
120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
121 #ifdef ASSERT
122 for (int i = 1; i < _intrinsics->length(); i++) {
123 CallGenerator* cg1 = _intrinsics->at(i-1);
124 CallGenerator* cg2 = _intrinsics->at(i);
125 assert(cg1->method() != cg2->method()
126 ? cg1->method() < cg2->method()
127 : cg1->is_virtual() < cg2->is_virtual(),
128 "compiler intrinsics list must stay sorted");
129 }
130 #endif
131 IntrinsicDescPair pair(m, is_virtual);
132 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
133 }
134
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
138 }
139 int len = _intrinsics->length();
140 bool found = false;
141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
142 assert(!found, "registering twice");
143 _intrinsics->insert_before(index, cg);
144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
145 }
146
147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
148 assert(m->is_loaded(), "don't try this on unloaded methods");
149 if (_intrinsics != NULL) {
150 bool found = false;
151 int index = intrinsic_insertion_index(m, is_virtual, found);
152 if (found) {
153 return _intrinsics->at(index);
154 }
155 }
156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
157 if (m->intrinsic_id() != vmIntrinsics::_none &&
158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
160 if (cg != NULL) {
161 // Save it for next time:
162 register_intrinsic(cg);
163 return cg;
164 } else {
165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
166 }
167 }
168 return NULL;
169 }
170
171 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
172 // in library_call.cpp.
173
174
175 #ifndef PRODUCT
176 // statistics gathering...
177
178 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
180
181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
182 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
183 int oflags = _intrinsic_hist_flags[id];
184 assert(flags != 0, "what happened?");
185 if (is_virtual) {
186 flags |= _intrinsic_virtual;
187 }
188 bool changed = (flags != oflags);
189 if ((flags & _intrinsic_worked) != 0) {
190 juint count = (_intrinsic_hist_count[id] += 1);
191 if (count == 1) {
192 changed = true; // first time
193 }
194 // increment the overall count also:
195 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
196 }
197 if (changed) {
198 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
199 // Something changed about the intrinsic's virtuality.
200 if ((flags & _intrinsic_virtual) != 0) {
201 // This is the first use of this intrinsic as a virtual call.
202 if (oflags != 0) {
203 // We already saw it as a non-virtual, so note both cases.
204 flags |= _intrinsic_both;
205 }
206 } else if ((oflags & _intrinsic_both) == 0) {
207 // This is the first use of this intrinsic as a non-virtual
208 flags |= _intrinsic_both;
209 }
210 }
211 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
212 }
213 // update the overall flags also:
214 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
215 return changed;
216 }
217
218 static char* format_flags(int flags, char* buf) {
219 buf[0] = 0;
220 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
221 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
222 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
223 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
224 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
225 if (buf[0] == 0) strcat(buf, ",");
226 assert(buf[0] == ',', "must be");
227 return &buf[1];
228 }
229
230 void Compile::print_intrinsic_statistics() {
231 char flagsbuf[100];
232 ttyLocker ttyl;
233 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
234 tty->print_cr("Compiler intrinsic usage:");
235 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
236 if (total == 0) total = 1; // avoid div0 in case of no successes
237 #define PRINT_STAT_LINE(name, c, f) \
238 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
239 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
240 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
241 int flags = _intrinsic_hist_flags[id];
242 juint count = _intrinsic_hist_count[id];
243 if ((flags | count) != 0) {
244 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
245 }
246 }
247 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
248 if (xtty != NULL) xtty->tail("statistics");
249 }
250
251 void Compile::print_statistics() {
252 { ttyLocker ttyl;
253 if (xtty != NULL) xtty->head("statistics type='opto'");
254 Parse::print_statistics();
255 PhaseCCP::print_statistics();
256 PhaseRegAlloc::print_statistics();
257 Scheduling::print_statistics();
258 PhasePeephole::print_statistics();
259 PhaseIdealLoop::print_statistics();
260 if (xtty != NULL) xtty->tail("statistics");
261 }
262 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
263 // put this under its own <statistics> element.
264 print_intrinsic_statistics();
265 }
266 }
267 #endif //PRODUCT
268
269 // Support for bundling info
270 Bundle* Compile::node_bundling(const Node *n) {
271 assert(valid_bundle_info(n), "oob");
272 return &_node_bundling_base[n->_idx];
273 }
274
275 bool Compile::valid_bundle_info(const Node *n) {
276 return (_node_bundling_limit > n->_idx);
277 }
278
279
280 void Compile::gvn_replace_by(Node* n, Node* nn) {
281 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
282 Node* use = n->last_out(i);
283 bool is_in_table = initial_gvn()->hash_delete(use);
284 uint uses_found = 0;
285 for (uint j = 0; j < use->len(); j++) {
286 if (use->in(j) == n) {
287 if (j < use->req())
288 use->set_req(j, nn);
289 else
290 use->set_prec(j, nn);
291 uses_found++;
292 }
293 }
294 if (is_in_table) {
295 // reinsert into table
296 initial_gvn()->hash_find_insert(use);
297 }
298 record_for_igvn(use);
299 i -= uses_found; // we deleted 1 or more copies of this edge
300 }
301 }
302
303
304 static inline bool not_a_node(const Node* n) {
305 if (n == NULL) return true;
306 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
307 if (*(address*)n == badAddress) return true; // kill by Node::destruct
308 return false;
309 }
310
311 // Identify all nodes that are reachable from below, useful.
312 // Use breadth-first pass that records state in a Unique_Node_List,
313 // recursive traversal is slower.
314 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
315 int estimated_worklist_size = live_nodes();
316 useful.map( estimated_worklist_size, NULL ); // preallocate space
317
318 // Initialize worklist
319 if (root() != NULL) { useful.push(root()); }
320 // If 'top' is cached, declare it useful to preserve cached node
321 if( cached_top_node() ) { useful.push(cached_top_node()); }
322
323 // Push all useful nodes onto the list, breadthfirst
324 for( uint next = 0; next < useful.size(); ++next ) {
325 assert( next < unique(), "Unique useful nodes < total nodes");
326 Node *n = useful.at(next);
327 uint max = n->len();
328 for( uint i = 0; i < max; ++i ) {
329 Node *m = n->in(i);
330 if (not_a_node(m)) continue;
331 useful.push(m);
332 }
333 }
334 }
335
336 // Update dead_node_list with any missing dead nodes using useful
337 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
338 void Compile::update_dead_node_list(Unique_Node_List &useful) {
339 uint max_idx = unique();
340 VectorSet& useful_node_set = useful.member_set();
341
342 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
343 // If node with index node_idx is not in useful set,
344 // mark it as dead in dead node list.
345 if (! useful_node_set.test(node_idx) ) {
346 record_dead_node(node_idx);
347 }
348 }
349 }
350
351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
352 int shift = 0;
353 for (int i = 0; i < inlines->length(); i++) {
354 CallGenerator* cg = inlines->at(i);
355 CallNode* call = cg->call_node();
356 if (shift > 0) {
357 inlines->at_put(i-shift, cg);
358 }
359 if (!useful.member(call)) {
360 shift++;
361 }
362 }
363 inlines->trunc_to(inlines->length()-shift);
364 }
365
366 // Disconnect all useless nodes by disconnecting those at the boundary.
367 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
368 uint next = 0;
369 while (next < useful.size()) {
370 Node *n = useful.at(next++);
371 if (n->is_SafePoint()) {
372 // We're done with a parsing phase. Replaced nodes are not valid
373 // beyond that point.
374 n->as_SafePoint()->delete_replaced_nodes();
375 }
376 // Use raw traversal of out edges since this code removes out edges
377 int max = n->outcnt();
378 for (int j = 0; j < max; ++j) {
379 Node* child = n->raw_out(j);
380 if (! useful.member(child)) {
381 assert(!child->is_top() || child != top(),
382 "If top is cached in Compile object it is in useful list");
383 // Only need to remove this out-edge to the useless node
384 n->raw_del_out(j);
385 --j;
386 --max;
387 }
388 }
389 if (n->outcnt() == 1 && n->has_special_unique_user()) {
390 record_for_igvn(n->unique_out());
391 }
392 }
393 // Remove useless macro and predicate opaq nodes
394 for (int i = C->macro_count()-1; i >= 0; i--) {
395 Node* n = C->macro_node(i);
396 if (!useful.member(n)) {
397 remove_macro_node(n);
398 }
399 }
400 // Remove useless CastII nodes with range check dependency
401 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
402 Node* cast = range_check_cast_node(i);
403 if (!useful.member(cast)) {
404 remove_range_check_cast(cast);
405 }
406 }
407 // Remove useless expensive nodes
408 for (int i = C->expensive_count()-1; i >= 0; i--) {
409 Node* n = C->expensive_node(i);
410 if (!useful.member(n)) {
411 remove_expensive_node(n);
412 }
413 }
414 // Remove useless Opaque4 nodes
415 for (int i = opaque4_count() - 1; i >= 0; i--) {
416 Node* opaq = opaque4_node(i);
417 if (!useful.member(opaq)) {
418 remove_opaque4_node(opaq);
419 }
420 }
421 // Remove useless value type nodes
422 if (_value_type_nodes != NULL) {
423 _value_type_nodes->remove_useless_nodes(useful.member_set());
424 }
425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
426 bs->eliminate_useless_gc_barriers(useful, this);
427 // clean up the late inline lists
428 remove_useless_late_inlines(&_string_late_inlines, useful);
429 remove_useless_late_inlines(&_boxing_late_inlines, useful);
430 remove_useless_late_inlines(&_late_inlines, useful);
431 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
432 }
433
434 //------------------------------frame_size_in_words-----------------------------
435 // frame_slots in units of words
436 int Compile::frame_size_in_words() const {
437 // shift is 0 in LP32 and 1 in LP64
438 const int shift = (LogBytesPerWord - LogBytesPerInt);
439 int words = _frame_slots >> shift;
440 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
441 return words;
442 }
443
444 // To bang the stack of this compiled method we use the stack size
445 // that the interpreter would need in case of a deoptimization. This
446 // removes the need to bang the stack in the deoptimization blob which
447 // in turn simplifies stack overflow handling.
448 int Compile::bang_size_in_bytes() const {
449 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
450 }
451
452 // ============================================================================
453 //------------------------------CompileWrapper---------------------------------
454 class CompileWrapper : public StackObj {
455 Compile *const _compile;
456 public:
457 CompileWrapper(Compile* compile);
458
459 ~CompileWrapper();
460 };
461
462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
463 // the Compile* pointer is stored in the current ciEnv:
464 ciEnv* env = compile->env();
465 assert(env == ciEnv::current(), "must already be a ciEnv active");
466 assert(env->compiler_data() == NULL, "compile already active?");
467 env->set_compiler_data(compile);
468 assert(compile == Compile::current(), "sanity");
469
470 compile->set_type_dict(NULL);
471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
472 compile->clone_map().set_clone_idx(0);
473 compile->set_type_hwm(NULL);
474 compile->set_type_last_size(0);
475 compile->set_last_tf(NULL, NULL);
476 compile->set_indexSet_arena(NULL);
477 compile->set_indexSet_free_block_list(NULL);
478 compile->init_type_arena();
479 Type::Initialize(compile);
480 _compile->set_scratch_buffer_blob(NULL);
481 _compile->begin_method();
482 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
483 }
484 CompileWrapper::~CompileWrapper() {
485 _compile->end_method();
486 if (_compile->scratch_buffer_blob() != NULL)
487 BufferBlob::free(_compile->scratch_buffer_blob());
488 _compile->env()->set_compiler_data(NULL);
489 }
490
491
492 //----------------------------print_compile_messages---------------------------
493 void Compile::print_compile_messages() {
494 #ifndef PRODUCT
495 // Check if recompiling
496 if (_subsume_loads == false && PrintOpto) {
497 // Recompiling without allowing machine instructions to subsume loads
498 tty->print_cr("*********************************************************");
499 tty->print_cr("** Bailout: Recompile without subsuming loads **");
500 tty->print_cr("*********************************************************");
501 }
502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
503 // Recompiling without escape analysis
504 tty->print_cr("*********************************************************");
505 tty->print_cr("** Bailout: Recompile without escape analysis **");
506 tty->print_cr("*********************************************************");
507 }
508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
509 // Recompiling without boxing elimination
510 tty->print_cr("*********************************************************");
511 tty->print_cr("** Bailout: Recompile without boxing elimination **");
512 tty->print_cr("*********************************************************");
513 }
514 if (C->directive()->BreakAtCompileOption) {
515 // Open the debugger when compiling this method.
516 tty->print("### Breaking when compiling: ");
517 method()->print_short_name();
518 tty->cr();
519 BREAKPOINT;
520 }
521
522 if( PrintOpto ) {
523 if (is_osr_compilation()) {
524 tty->print("[OSR]%3d", _compile_id);
525 } else {
526 tty->print("%3d", _compile_id);
527 }
528 }
529 #endif
530 }
531
532
533 //-----------------------init_scratch_buffer_blob------------------------------
534 // Construct a temporary BufferBlob and cache it for this compile.
535 void Compile::init_scratch_buffer_blob(int const_size) {
536 // If there is already a scratch buffer blob allocated and the
537 // constant section is big enough, use it. Otherwise free the
538 // current and allocate a new one.
539 BufferBlob* blob = scratch_buffer_blob();
540 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
541 // Use the current blob.
542 } else {
543 if (blob != NULL) {
544 BufferBlob::free(blob);
545 }
546
547 ResourceMark rm;
548 _scratch_const_size = const_size;
549 int size = C2Compiler::initial_code_buffer_size(const_size);
550 #ifdef ASSERT
551 if (C->has_scalarized_args()) {
552 // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space
553 size += 5120;
554 }
555 #endif
556 blob = BufferBlob::create("Compile::scratch_buffer", size);
557 // Record the buffer blob for next time.
558 set_scratch_buffer_blob(blob);
559 // Have we run out of code space?
560 if (scratch_buffer_blob() == NULL) {
561 // Let CompilerBroker disable further compilations.
562 record_failure("Not enough space for scratch buffer in CodeCache");
563 return;
564 }
565 }
566
567 // Initialize the relocation buffers
568 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
569 set_scratch_locs_memory(locs_buf);
570 }
571
572
573 //-----------------------scratch_emit_size-------------------------------------
574 // Helper function that computes size by emitting code
575 uint Compile::scratch_emit_size(const Node* n) {
576 // Start scratch_emit_size section.
577 set_in_scratch_emit_size(true);
578
579 // Emit into a trash buffer and count bytes emitted.
580 // This is a pretty expensive way to compute a size,
581 // but it works well enough if seldom used.
582 // All common fixed-size instructions are given a size
583 // method by the AD file.
584 // Note that the scratch buffer blob and locs memory are
585 // allocated at the beginning of the compile task, and
586 // may be shared by several calls to scratch_emit_size.
587 // The allocation of the scratch buffer blob is particularly
588 // expensive, since it has to grab the code cache lock.
589 BufferBlob* blob = this->scratch_buffer_blob();
590 assert(blob != NULL, "Initialize BufferBlob at start");
591 assert(blob->size() > MAX_inst_size, "sanity");
592 relocInfo* locs_buf = scratch_locs_memory();
593 address blob_begin = blob->content_begin();
594 address blob_end = (address)locs_buf;
595 assert(blob->contains(blob_end), "sanity");
596 CodeBuffer buf(blob_begin, blob_end - blob_begin);
597 buf.initialize_consts_size(_scratch_const_size);
598 buf.initialize_stubs_size(MAX_stubs_size);
599 assert(locs_buf != NULL, "sanity");
600 int lsize = MAX_locs_size / 3;
601 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
602 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
603 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
604 // Mark as scratch buffer.
605 buf.consts()->set_scratch_emit();
606 buf.insts()->set_scratch_emit();
607 buf.stubs()->set_scratch_emit();
608
609 // Do the emission.
610
611 Label fakeL; // Fake label for branch instructions.
612 Label* saveL = NULL;
613 uint save_bnum = 0;
614 bool is_branch = n->is_MachBranch();
615 if (is_branch) {
616 MacroAssembler masm(&buf);
617 masm.bind(fakeL);
618 n->as_MachBranch()->save_label(&saveL, &save_bnum);
619 n->as_MachBranch()->label_set(&fakeL, 0);
620 } else if (n->is_MachProlog()) {
621 saveL = ((MachPrologNode*)n)->_verified_entry;
622 ((MachPrologNode*)n)->_verified_entry = &fakeL;
623 }
624 n->emit(buf, this->regalloc());
625
626 // Emitting into the scratch buffer should not fail
627 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
628
629 // Restore label.
630 if (is_branch) {
631 n->as_MachBranch()->label_set(saveL, save_bnum);
632 } else if (n->is_MachProlog()) {
633 ((MachPrologNode*)n)->_verified_entry = saveL;
634 }
635
636 // End scratch_emit_size section.
637 set_in_scratch_emit_size(false);
638
639 return buf.insts_size();
640 }
641
642
643 // ============================================================================
644 //------------------------------Compile standard-------------------------------
645 debug_only( int Compile::_debug_idx = 100000; )
646
647 // Compile a method. entry_bci is -1 for normal compilations and indicates
648 // the continuation bci for on stack replacement.
649
650
651 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
652 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
653 : Phase(Compiler),
654 _compile_id(ci_env->compile_id()),
655 _save_argument_registers(false),
656 _subsume_loads(subsume_loads),
657 _do_escape_analysis(do_escape_analysis),
658 _eliminate_boxing(eliminate_boxing),
659 _method(target),
660 _entry_bci(osr_bci),
661 _stub_function(NULL),
662 _stub_name(NULL),
663 _stub_entry_point(NULL),
664 _max_node_limit(MaxNodeLimit),
665 _orig_pc_slot(0),
666 _orig_pc_slot_offset_in_bytes(0),
667 _sp_inc_slot(0),
668 _sp_inc_slot_offset_in_bytes(0),
669 _inlining_progress(false),
670 _inlining_incrementally(false),
671 _do_cleanup(false),
672 _has_reserved_stack_access(target->has_reserved_stack_access()),
673 #ifndef PRODUCT
674 _trace_opto_output(directive->TraceOptoOutputOption),
675 #endif
676 _has_method_handle_invokes(false),
677 _comp_arena(mtCompiler),
678 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
679 _env(ci_env),
680 _directive(directive),
681 _log(ci_env->log()),
682 _failure_reason(NULL),
683 _congraph(NULL),
684 #ifndef PRODUCT
685 _printer(IdealGraphPrinter::printer()),
686 #endif
687 _dead_node_list(comp_arena()),
688 _dead_node_count(0),
689 _node_arena(mtCompiler),
690 _old_arena(mtCompiler),
691 _mach_constant_base_node(NULL),
692 _Compile_types(mtCompiler),
693 _initial_gvn(NULL),
694 _for_igvn(NULL),
695 _warm_calls(NULL),
696 _late_inlines(comp_arena(), 2, 0, NULL),
697 _string_late_inlines(comp_arena(), 2, 0, NULL),
698 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
699 _late_inlines_pos(0),
700 _number_of_mh_late_inlines(0),
701 _print_inlining_stream(NULL),
702 _print_inlining_list(NULL),
703 _print_inlining_idx(0),
704 _print_inlining_output(NULL),
705 _replay_inline_data(NULL),
706 _java_calls(0),
707 _inner_loops(0),
708 _interpreter_frame_size(0),
709 _node_bundling_limit(0),
710 _node_bundling_base(NULL),
711 _code_buffer("Compile::Fill_buffer"),
712 _scratch_const_size(-1),
713 _in_scratch_emit_size(false)
714 #ifndef PRODUCT
715 , _in_dump_cnt(0)
716 #endif
717 {
718 C = this;
719 #ifndef PRODUCT
720 if (_printer != NULL) {
721 _printer->set_compile(this);
722 }
723 #endif
724 CompileWrapper cw(this);
725
726 if (CITimeVerbose) {
727 tty->print(" ");
728 target->holder()->name()->print();
729 tty->print(".");
730 target->print_short_name();
731 tty->print(" ");
732 }
733 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
734 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
735
736 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
737 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
738 // We can always print a disassembly, either abstract (hex dump) or
739 // with the help of a suitable hsdis library. Thus, we should not
740 // couple print_assembly and print_opto_assembly controls.
741 // But: always print opto and regular assembly on compile command 'print'.
742 bool print_assembly = directive->PrintAssemblyOption;
743 set_print_assembly(print_opto_assembly || print_assembly);
744 #else
745 set_print_assembly(false); // must initialize.
746 #endif
747
748 #ifndef PRODUCT
749 set_parsed_irreducible_loop(false);
750
751 if (directive->ReplayInlineOption) {
752 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
753 }
754 #endif
755 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
756 set_print_intrinsics(directive->PrintIntrinsicsOption);
757 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
758
759 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
760 // Make sure the method being compiled gets its own MDO,
761 // so we can at least track the decompile_count().
762 // Need MDO to record RTM code generation state.
763 method()->ensure_method_data();
764 }
765
766 Init(::AliasLevel);
767
768
769 print_compile_messages();
770
771 _ilt = InlineTree::build_inline_tree_root();
772
773 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
774 assert(num_alias_types() >= AliasIdxRaw, "");
775
776 #define MINIMUM_NODE_HASH 1023
777 // Node list that Iterative GVN will start with
778 Unique_Node_List for_igvn(comp_arena());
779 set_for_igvn(&for_igvn);
780
781 // GVN that will be run immediately on new nodes
782 uint estimated_size = method()->code_size()*4+64;
783 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
784 PhaseGVN gvn(node_arena(), estimated_size);
785 set_initial_gvn(&gvn);
786
787 print_inlining_init();
788 { // Scope for timing the parser
789 TracePhase tp("parse", &timers[_t_parser]);
790
791 // Put top into the hash table ASAP.
792 initial_gvn()->transform_no_reclaim(top());
793
794 // Set up tf(), start(), and find a CallGenerator.
795 CallGenerator* cg = NULL;
796 if (is_osr_compilation()) {
797 const TypeTuple *domain = StartOSRNode::osr_domain();
798 const TypeTuple *range = TypeTuple::make_range(method()->signature());
799 init_tf(TypeFunc::make(domain, range));
800 StartNode* s = new StartOSRNode(root(), domain);
801 initial_gvn()->set_type_bottom(s);
802 init_start(s);
803 cg = CallGenerator::for_osr(method(), entry_bci());
804 } else {
805 // Normal case.
806 init_tf(TypeFunc::make(method()));
807 StartNode* s = new StartNode(root(), tf()->domain_cc());
808 initial_gvn()->set_type_bottom(s);
809 init_start(s);
810 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
811 // With java.lang.ref.reference.get() we must go through the
812 // intrinsic - even when get() is the root
813 // method of the compile - so that, if necessary, the value in
814 // the referent field of the reference object gets recorded by
815 // the pre-barrier code.
816 cg = find_intrinsic(method(), false);
817 }
818 if (cg == NULL) {
819 float past_uses = method()->interpreter_invocation_count();
820 float expected_uses = past_uses;
821 cg = CallGenerator::for_inline(method(), expected_uses);
822 }
823 }
824 if (failing()) return;
825 if (cg == NULL) {
826 record_method_not_compilable("cannot parse method");
827 return;
828 }
829 JVMState* jvms = build_start_state(start(), tf());
830 if ((jvms = cg->generate(jvms)) == NULL) {
831 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
832 record_method_not_compilable("method parse failed");
833 }
834 return;
835 }
836 GraphKit kit(jvms);
837
838 if (!kit.stopped()) {
839 // Accept return values, and transfer control we know not where.
840 // This is done by a special, unique ReturnNode bound to root.
841 return_values(kit.jvms());
842 }
843
844 if (kit.has_exceptions()) {
845 // Any exceptions that escape from this call must be rethrown
846 // to whatever caller is dynamically above us on the stack.
847 // This is done by a special, unique RethrowNode bound to root.
848 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
849 }
850
851 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
852
853 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
854 inline_string_calls(true);
855 }
856
857 if (failing()) return;
858
859 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
860
861 // Remove clutter produced by parsing.
862 if (!failing()) {
863 ResourceMark rm;
864 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
865 }
866 }
867
868 // Note: Large methods are capped off in do_one_bytecode().
869 if (failing()) return;
870
871 // After parsing, node notes are no longer automagic.
872 // They must be propagated by register_new_node_with_optimizer(),
873 // clone(), or the like.
874 set_default_node_notes(NULL);
875
876 for (;;) {
877 int successes = Inline_Warm();
878 if (failing()) return;
879 if (successes == 0) break;
880 }
881
882 // Drain the list.
883 Finish_Warm();
884 #ifndef PRODUCT
885 if (_printer && _printer->should_print(1)) {
886 _printer->print_inlining();
887 }
888 #endif
889
890 if (failing()) return;
891 NOT_PRODUCT( verify_graph_edges(); )
892
893 // Now optimize
894 Optimize();
895 if (failing()) return;
896 NOT_PRODUCT( verify_graph_edges(); )
897
898 #ifndef PRODUCT
899 if (PrintIdeal) {
900 ttyLocker ttyl; // keep the following output all in one block
901 // This output goes directly to the tty, not the compiler log.
902 // To enable tools to match it up with the compilation activity,
903 // be sure to tag this tty output with the compile ID.
904 if (xtty != NULL) {
905 xtty->head("ideal compile_id='%d'%s", compile_id(),
906 is_osr_compilation() ? " compile_kind='osr'" :
907 "");
908 }
909 root()->dump(9999);
910 if (xtty != NULL) {
911 xtty->tail("ideal");
912 }
913 }
914 #endif
915
916 #ifdef ASSERT
917 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
918 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
919 #endif
920
921 // Dump compilation data to replay it.
922 if (directive->DumpReplayOption) {
923 env()->dump_replay_data(_compile_id);
924 }
925 if (directive->DumpInlineOption && (ilt() != NULL)) {
926 env()->dump_inline_data(_compile_id);
927 }
928
929 // Now that we know the size of all the monitors we can add a fixed slot
930 // for the original deopt pc.
931
932 _orig_pc_slot = fixed_slots();
933 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
934
935 if (needs_stack_repair()) {
936 // One extra slot for the special stack increment value
937 _sp_inc_slot = next_slot;
938 next_slot += 2;
939 }
940
941 set_fixed_slots(next_slot);
942
943 // Compute when to use implicit null checks. Used by matching trap based
944 // nodes and NullCheck optimization.
945 set_allowed_deopt_reasons();
946
947 // Now generate code
948 Code_Gen();
949 if (failing()) return;
950
951 // Check if we want to skip execution of all compiled code.
952 {
953 #ifndef PRODUCT
954 if (OptoNoExecute) {
955 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
956 return;
957 }
958 #endif
959 TracePhase tp("install_code", &timers[_t_registerMethod]);
960
961 if (is_osr_compilation()) {
962 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
963 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
964 } else {
965 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
966 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) {
967 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size);
968 }
969 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry_RO) == -1) {
970 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, _first_block_size);
971 }
972 if (_code_offsets.value(CodeOffsets::Entry) == -1) {
973 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
974 }
975 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
976 }
977
978 env()->register_method(_method, _entry_bci,
979 &_code_offsets,
980 _orig_pc_slot_offset_in_bytes,
981 code_buffer(),
982 frame_size_in_words(), _oop_map_set,
983 &_handler_table, &_inc_table,
984 compiler,
985 has_unsafe_access(),
986 SharedRuntime::is_wide_vector(max_vector_size()),
987 rtm_state()
988 );
989
990 if (log() != NULL) // Print code cache state into compiler log
991 log()->code_cache_state();
992 }
993 }
994
995 //------------------------------Compile----------------------------------------
996 // Compile a runtime stub
997 Compile::Compile( ciEnv* ci_env,
998 TypeFunc_generator generator,
999 address stub_function,
1000 const char *stub_name,
1001 int is_fancy_jump,
1002 bool pass_tls,
1003 bool save_arg_registers,
1004 bool return_pc,
1005 DirectiveSet* directive)
1006 : Phase(Compiler),
1007 _compile_id(0),
1008 _save_argument_registers(save_arg_registers),
1009 _subsume_loads(true),
1010 _do_escape_analysis(false),
1011 _eliminate_boxing(false),
1012 _method(NULL),
1013 _entry_bci(InvocationEntryBci),
1014 _stub_function(stub_function),
1015 _stub_name(stub_name),
1016 _stub_entry_point(NULL),
1017 _max_node_limit(MaxNodeLimit),
1018 _orig_pc_slot(0),
1019 _orig_pc_slot_offset_in_bytes(0),
1020 _sp_inc_slot(0),
1021 _sp_inc_slot_offset_in_bytes(0),
1022 _inlining_progress(false),
1023 _inlining_incrementally(false),
1024 _has_reserved_stack_access(false),
1025 #ifndef PRODUCT
1026 _trace_opto_output(directive->TraceOptoOutputOption),
1027 #endif
1028 _has_method_handle_invokes(false),
1029 _comp_arena(mtCompiler),
1030 _env(ci_env),
1031 _directive(directive),
1032 _log(ci_env->log()),
1033 _failure_reason(NULL),
1034 _congraph(NULL),
1035 #ifndef PRODUCT
1036 _printer(NULL),
1037 #endif
1038 _dead_node_list(comp_arena()),
1039 _dead_node_count(0),
1040 _node_arena(mtCompiler),
1041 _old_arena(mtCompiler),
1042 _mach_constant_base_node(NULL),
1043 _Compile_types(mtCompiler),
1044 _initial_gvn(NULL),
1045 _for_igvn(NULL),
1046 _warm_calls(NULL),
1047 _number_of_mh_late_inlines(0),
1048 _print_inlining_stream(NULL),
1049 _print_inlining_list(NULL),
1050 _print_inlining_idx(0),
1051 _print_inlining_output(NULL),
1052 _replay_inline_data(NULL),
1053 _java_calls(0),
1054 _inner_loops(0),
1055 _interpreter_frame_size(0),
1056 _node_bundling_limit(0),
1057 _node_bundling_base(NULL),
1058 _code_buffer("Compile::Fill_buffer"),
1059 #ifndef PRODUCT
1060 _in_dump_cnt(0),
1061 #endif
1062 _allowed_reasons(0) {
1063 C = this;
1064
1065 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1066 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1067
1068 #ifndef PRODUCT
1069 set_print_assembly(PrintFrameConverterAssembly);
1070 set_parsed_irreducible_loop(false);
1071 #else
1072 set_print_assembly(false); // Must initialize.
1073 #endif
1074 set_has_irreducible_loop(false); // no loops
1075
1076 CompileWrapper cw(this);
1077 Init(/*AliasLevel=*/ 0);
1078 init_tf((*generator)());
1079
1080 {
1081 // The following is a dummy for the sake of GraphKit::gen_stub
1082 Unique_Node_List for_igvn(comp_arena());
1083 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1084 PhaseGVN gvn(Thread::current()->resource_area(),255);
1085 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1086 gvn.transform_no_reclaim(top());
1087
1088 GraphKit kit;
1089 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1090 }
1091
1092 NOT_PRODUCT( verify_graph_edges(); )
1093 Code_Gen();
1094 if (failing()) return;
1095
1096
1097 // Entry point will be accessed using compile->stub_entry_point();
1098 if (code_buffer() == NULL) {
1099 Matcher::soft_match_failure();
1100 } else {
1101 if (PrintAssembly && (WizardMode || Verbose))
1102 tty->print_cr("### Stub::%s", stub_name);
1103
1104 if (!failing()) {
1105 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1106
1107 // Make the NMethod
1108 // For now we mark the frame as never safe for profile stackwalking
1109 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1110 code_buffer(),
1111 CodeOffsets::frame_never_safe,
1112 // _code_offsets.value(CodeOffsets::Frame_Complete),
1113 frame_size_in_words(),
1114 _oop_map_set,
1115 save_arg_registers);
1116 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1117
1118 _stub_entry_point = rs->entry_point();
1119 }
1120 }
1121 }
1122
1123 //------------------------------Init-------------------------------------------
1124 // Prepare for a single compilation
1125 void Compile::Init(int aliaslevel) {
1126 _unique = 0;
1127 _regalloc = NULL;
1128
1129 _tf = NULL; // filled in later
1130 _top = NULL; // cached later
1131 _matcher = NULL; // filled in later
1132 _cfg = NULL; // filled in later
1133
1134 set_24_bit_selection_and_mode(Use24BitFP, false);
1135
1136 _node_note_array = NULL;
1137 _default_node_notes = NULL;
1138 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1139
1140 _immutable_memory = NULL; // filled in at first inquiry
1141
1142 // Globally visible Nodes
1143 // First set TOP to NULL to give safe behavior during creation of RootNode
1144 set_cached_top_node(NULL);
1145 set_root(new RootNode());
1146 // Now that you have a Root to point to, create the real TOP
1147 set_cached_top_node( new ConNode(Type::TOP) );
1148 set_recent_alloc(NULL, NULL);
1149
1150 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1151 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1152 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1153 env()->set_dependencies(new Dependencies(env()));
1154
1155 _fixed_slots = 0;
1156 set_has_split_ifs(false);
1157 set_has_loops(has_method() && method()->has_loops()); // first approximation
1158 set_has_stringbuilder(false);
1159 set_has_boxed_value(false);
1160 _trap_can_recompile = false; // no traps emitted yet
1161 _major_progress = true; // start out assuming good things will happen
1162 set_has_unsafe_access(false);
1163 set_max_vector_size(0);
1164 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1165 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1166 set_decompile_count(0);
1167
1168 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1169 _loop_opts_cnt = LoopOptsCount;
1170 _has_flattened_accesses = false;
1171 _flattened_accesses_share_alias = true;
1172
1173 set_do_inlining(Inline);
1174 set_max_inline_size(MaxInlineSize);
1175 set_freq_inline_size(FreqInlineSize);
1176 set_do_scheduling(OptoScheduling);
1177 set_do_count_invocations(false);
1178 set_do_method_data_update(false);
1179
1180 set_do_vector_loop(false);
1181
1182 if (AllowVectorizeOnDemand) {
1183 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1184 set_do_vector_loop(true);
1185 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());})
1186 } else if (has_method() && method()->name() != 0 &&
1187 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1188 set_do_vector_loop(true);
1189 }
1190 }
1191 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1192 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());})
1193
1194 set_age_code(has_method() && method()->profile_aging());
1195 set_rtm_state(NoRTM); // No RTM lock eliding by default
1196 _max_node_limit = _directive->MaxNodeLimitOption;
1197
1198 #if INCLUDE_RTM_OPT
1199 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1200 int rtm_state = method()->method_data()->rtm_state();
1201 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1202 // Don't generate RTM lock eliding code.
1203 set_rtm_state(NoRTM);
1204 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1205 // Generate RTM lock eliding code without abort ratio calculation code.
1206 set_rtm_state(UseRTM);
1207 } else if (UseRTMDeopt) {
1208 // Generate RTM lock eliding code and include abort ratio calculation
1209 // code if UseRTMDeopt is on.
1210 set_rtm_state(ProfileRTM);
1211 }
1212 }
1213 #endif
1214 if (debug_info()->recording_non_safepoints()) {
1215 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1216 (comp_arena(), 8, 0, NULL));
1217 set_default_node_notes(Node_Notes::make(this));
1218 }
1219
1220 // // -- Initialize types before each compile --
1221 // // Update cached type information
1222 // if( _method && _method->constants() )
1223 // Type::update_loaded_types(_method, _method->constants());
1224
1225 // Init alias_type map.
1226 if (!_do_escape_analysis && aliaslevel == 3)
1227 aliaslevel = 2; // No unique types without escape analysis
1228 _AliasLevel = aliaslevel;
1229 const int grow_ats = 16;
1230 _max_alias_types = grow_ats;
1231 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1232 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1233 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1234 {
1235 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1236 }
1237 // Initialize the first few types.
1238 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1239 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1240 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1241 _num_alias_types = AliasIdxRaw+1;
1242 // Zero out the alias type cache.
1243 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1244 // A NULL adr_type hits in the cache right away. Preload the right answer.
1245 probe_alias_cache(NULL)->_index = AliasIdxTop;
1246
1247 _intrinsics = NULL;
1248 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1249 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1250 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1251 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1252 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1253 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
1254 register_library_intrinsics();
1255 }
1256
1257 //---------------------------init_start----------------------------------------
1258 // Install the StartNode on this compile object.
1259 void Compile::init_start(StartNode* s) {
1260 if (failing())
1261 return; // already failing
1262 assert(s == start(), "");
1263 }
1264
1265 /**
1266 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1267 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1268 * the ideal graph.
1269 */
1270 StartNode* Compile::start() const {
1271 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1272 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1273 Node* start = root()->fast_out(i);
1274 if (start->is_Start()) {
1275 return start->as_Start();
1276 }
1277 }
1278 fatal("Did not find Start node!");
1279 return NULL;
1280 }
1281
1282 //-------------------------------immutable_memory-------------------------------------
1283 // Access immutable memory
1284 Node* Compile::immutable_memory() {
1285 if (_immutable_memory != NULL) {
1286 return _immutable_memory;
1287 }
1288 StartNode* s = start();
1289 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1290 Node *p = s->fast_out(i);
1291 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1292 _immutable_memory = p;
1293 return _immutable_memory;
1294 }
1295 }
1296 ShouldNotReachHere();
1297 return NULL;
1298 }
1299
1300 //----------------------set_cached_top_node------------------------------------
1301 // Install the cached top node, and make sure Node::is_top works correctly.
1302 void Compile::set_cached_top_node(Node* tn) {
1303 if (tn != NULL) verify_top(tn);
1304 Node* old_top = _top;
1305 _top = tn;
1306 // Calling Node::setup_is_top allows the nodes the chance to adjust
1307 // their _out arrays.
1308 if (_top != NULL) _top->setup_is_top();
1309 if (old_top != NULL) old_top->setup_is_top();
1310 assert(_top == NULL || top()->is_top(), "");
1311 }
1312
1313 #ifdef ASSERT
1314 uint Compile::count_live_nodes_by_graph_walk() {
1315 Unique_Node_List useful(comp_arena());
1316 // Get useful node list by walking the graph.
1317 identify_useful_nodes(useful);
1318 return useful.size();
1319 }
1320
1321 void Compile::print_missing_nodes() {
1322
1323 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1324 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1325 return;
1326 }
1327
1328 // This is an expensive function. It is executed only when the user
1329 // specifies VerifyIdealNodeCount option or otherwise knows the
1330 // additional work that needs to be done to identify reachable nodes
1331 // by walking the flow graph and find the missing ones using
1332 // _dead_node_list.
1333
1334 Unique_Node_List useful(comp_arena());
1335 // Get useful node list by walking the graph.
1336 identify_useful_nodes(useful);
1337
1338 uint l_nodes = C->live_nodes();
1339 uint l_nodes_by_walk = useful.size();
1340
1341 if (l_nodes != l_nodes_by_walk) {
1342 if (_log != NULL) {
1343 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1344 _log->stamp();
1345 _log->end_head();
1346 }
1347 VectorSet& useful_member_set = useful.member_set();
1348 int last_idx = l_nodes_by_walk;
1349 for (int i = 0; i < last_idx; i++) {
1350 if (useful_member_set.test(i)) {
1351 if (_dead_node_list.test(i)) {
1352 if (_log != NULL) {
1353 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1354 }
1355 if (PrintIdealNodeCount) {
1356 // Print the log message to tty
1357 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1358 useful.at(i)->dump();
1359 }
1360 }
1361 }
1362 else if (! _dead_node_list.test(i)) {
1363 if (_log != NULL) {
1364 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1365 }
1366 if (PrintIdealNodeCount) {
1367 // Print the log message to tty
1368 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1369 }
1370 }
1371 }
1372 if (_log != NULL) {
1373 _log->tail("mismatched_nodes");
1374 }
1375 }
1376 }
1377 void Compile::record_modified_node(Node* n) {
1378 if (_modified_nodes != NULL && !_inlining_incrementally &&
1379 n->outcnt() != 0 && !n->is_Con()) {
1380 _modified_nodes->push(n);
1381 }
1382 }
1383
1384 void Compile::remove_modified_node(Node* n) {
1385 if (_modified_nodes != NULL) {
1386 _modified_nodes->remove(n);
1387 }
1388 }
1389 #endif
1390
1391 #ifndef PRODUCT
1392 void Compile::verify_top(Node* tn) const {
1393 if (tn != NULL) {
1394 assert(tn->is_Con(), "top node must be a constant");
1395 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1396 assert(tn->in(0) != NULL, "must have live top node");
1397 }
1398 }
1399 #endif
1400
1401
1402 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1403
1404 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1405 guarantee(arr != NULL, "");
1406 int num_blocks = arr->length();
1407 if (grow_by < num_blocks) grow_by = num_blocks;
1408 int num_notes = grow_by * _node_notes_block_size;
1409 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1410 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1411 while (num_notes > 0) {
1412 arr->append(notes);
1413 notes += _node_notes_block_size;
1414 num_notes -= _node_notes_block_size;
1415 }
1416 assert(num_notes == 0, "exact multiple, please");
1417 }
1418
1419 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1420 if (source == NULL || dest == NULL) return false;
1421
1422 if (dest->is_Con())
1423 return false; // Do not push debug info onto constants.
1424
1425 #ifdef ASSERT
1426 // Leave a bread crumb trail pointing to the original node:
1427 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1428 dest->set_debug_orig(source);
1429 }
1430 #endif
1431
1432 if (node_note_array() == NULL)
1433 return false; // Not collecting any notes now.
1434
1435 // This is a copy onto a pre-existing node, which may already have notes.
1436 // If both nodes have notes, do not overwrite any pre-existing notes.
1437 Node_Notes* source_notes = node_notes_at(source->_idx);
1438 if (source_notes == NULL || source_notes->is_clear()) return false;
1439 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1440 if (dest_notes == NULL || dest_notes->is_clear()) {
1441 return set_node_notes_at(dest->_idx, source_notes);
1442 }
1443
1444 Node_Notes merged_notes = (*source_notes);
1445 // The order of operations here ensures that dest notes will win...
1446 merged_notes.update_from(dest_notes);
1447 return set_node_notes_at(dest->_idx, &merged_notes);
1448 }
1449
1450
1451 //--------------------------allow_range_check_smearing-------------------------
1452 // Gating condition for coalescing similar range checks.
1453 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1454 // single covering check that is at least as strong as any of them.
1455 // If the optimization succeeds, the simplified (strengthened) range check
1456 // will always succeed. If it fails, we will deopt, and then give up
1457 // on the optimization.
1458 bool Compile::allow_range_check_smearing() const {
1459 // If this method has already thrown a range-check,
1460 // assume it was because we already tried range smearing
1461 // and it failed.
1462 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1463 return !already_trapped;
1464 }
1465
1466
1467 //------------------------------flatten_alias_type-----------------------------
1468 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1469 int offset = tj->offset();
1470 TypePtr::PTR ptr = tj->ptr();
1471
1472 // Known instance (scalarizable allocation) alias only with itself.
1473 bool is_known_inst = tj->isa_oopptr() != NULL &&
1474 tj->is_oopptr()->is_known_instance();
1475
1476 // Process weird unsafe references.
1477 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1478 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass();
1479 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops");
1480 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1481 tj = TypeOopPtr::BOTTOM;
1482 ptr = tj->ptr();
1483 offset = tj->offset();
1484 }
1485
1486 // Array pointers need some flattening
1487 const TypeAryPtr *ta = tj->isa_aryptr();
1488 if (ta && ta->is_stable()) {
1489 // Erase stability property for alias analysis.
1490 tj = ta = ta->cast_to_stable(false);
1491 }
1492 if( ta && is_known_inst ) {
1493 if ( offset != Type::OffsetBot &&
1494 offset > arrayOopDesc::length_offset_in_bytes() ) {
1495 offset = Type::OffsetBot; // Flatten constant access into array body only
1496 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1497 }
1498 } else if( ta && _AliasLevel >= 2 ) {
1499 // For arrays indexed by constant indices, we flatten the alias
1500 // space to include all of the array body. Only the header, klass
1501 // and array length can be accessed un-aliased.
1502 // For flattened value type array, each field has its own slice so
1503 // we must include the field offset.
1504 if( offset != Type::OffsetBot ) {
1505 if( ta->const_oop() ) { // MethodData* or Method*
1506 offset = Type::OffsetBot; // Flatten constant access into array body
1507 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1508 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1509 // range is OK as-is.
1510 tj = ta = TypeAryPtr::RANGE;
1511 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1512 tj = TypeInstPtr::KLASS; // all klass loads look alike
1513 ta = TypeAryPtr::RANGE; // generic ignored junk
1514 ptr = TypePtr::BotPTR;
1515 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1516 tj = TypeInstPtr::MARK;
1517 ta = TypeAryPtr::RANGE; // generic ignored junk
1518 ptr = TypePtr::BotPTR;
1519 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1520 ta = tj->isa_aryptr();
1521 } else { // Random constant offset into array body
1522 offset = Type::OffsetBot; // Flatten constant access into array body
1523 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1524 }
1525 }
1526 // Arrays of fixed size alias with arrays of unknown size.
1527 if (ta->size() != TypeInt::POS) {
1528 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1529 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1530 }
1531 // Arrays of known objects become arrays of unknown objects.
1532 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1533 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1534 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1535 }
1536 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1537 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1538 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1539 }
1540 // Initially all flattened array accesses share a single slice
1541 if (ta->elem()->isa_valuetype() && ta->elem() != TypeValueType::BOTTOM && _flattened_accesses_share_alias) {
1542 const TypeAry *tary = TypeAry::make(TypeValueType::BOTTOM, ta->size());
1543 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot));
1544 }
1545 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1546 // cannot be distinguished by bytecode alone.
1547 if (ta->elem() == TypeInt::BOOL) {
1548 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1549 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1550 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1551 }
1552 // During the 2nd round of IterGVN, NotNull castings are removed.
1553 // Make sure the Bottom and NotNull variants alias the same.
1554 // Also, make sure exact and non-exact variants alias the same.
1555 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1556 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1557 }
1558 }
1559
1560 // Oop pointers need some flattening
1561 const TypeInstPtr *to = tj->isa_instptr();
1562 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1563 ciInstanceKlass *k = to->klass()->as_instance_klass();
1564 if( ptr == TypePtr::Constant ) {
1565 if (to->klass() != ciEnv::current()->Class_klass() ||
1566 offset < k->size_helper() * wordSize) {
1567 // No constant oop pointers (such as Strings); they alias with
1568 // unknown strings.
1569 assert(!is_known_inst, "not scalarizable allocation");
1570 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1571 }
1572 } else if( is_known_inst ) {
1573 tj = to; // Keep NotNull and klass_is_exact for instance type
1574 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1575 // During the 2nd round of IterGVN, NotNull castings are removed.
1576 // Make sure the Bottom and NotNull variants alias the same.
1577 // Also, make sure exact and non-exact variants alias the same.
1578 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1579 }
1580 if (to->speculative() != NULL) {
1581 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id());
1582 }
1583 // Canonicalize the holder of this field
1584 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1585 // First handle header references such as a LoadKlassNode, even if the
1586 // object's klass is unloaded at compile time (4965979).
1587 if (!is_known_inst) { // Do it only for non-instance types
1588 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1589 }
1590 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1591 to = tj->is_instptr();
1592 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1593 // Static fields are in the space above the normal instance
1594 // fields in the java.lang.Class instance.
1595 if (to->klass() != ciEnv::current()->Class_klass()) {
1596 to = NULL;
1597 tj = TypeOopPtr::BOTTOM;
1598 offset = tj->offset();
1599 }
1600 } else {
1601 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1602 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1603 if( is_known_inst ) {
1604 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id());
1605 } else {
1606 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1607 }
1608 }
1609 }
1610 }
1611
1612 // Klass pointers to object array klasses need some flattening
1613 const TypeKlassPtr *tk = tj->isa_klassptr();
1614 if( tk ) {
1615 // If we are referencing a field within a Klass, we need
1616 // to assume the worst case of an Object. Both exact and
1617 // inexact types must flatten to the same alias class so
1618 // use NotNull as the PTR.
1619 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1620
1621 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1622 TypeKlassPtr::OBJECT->klass(),
1623 Type::Offset(offset));
1624 }
1625
1626 ciKlass* klass = tk->klass();
1627 if (klass != NULL && klass->is_obj_array_klass()) {
1628 ciKlass* k = TypeAryPtr::OOPS->klass();
1629 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1630 k = TypeInstPtr::BOTTOM->klass();
1631 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1632 }
1633
1634 // Check for precise loads from the primary supertype array and force them
1635 // to the supertype cache alias index. Check for generic array loads from
1636 // the primary supertype array and also force them to the supertype cache
1637 // alias index. Since the same load can reach both, we need to merge
1638 // these 2 disparate memories into the same alias class. Since the
1639 // primary supertype array is read-only, there's no chance of confusion
1640 // where we bypass an array load and an array store.
1641 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1642 if (offset == Type::OffsetBot ||
1643 (offset >= primary_supers_offset &&
1644 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1645 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1646 offset = in_bytes(Klass::secondary_super_cache_offset());
1647 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1648 }
1649 }
1650
1651 // Flatten all Raw pointers together.
1652 if (tj->base() == Type::RawPtr)
1653 tj = TypeRawPtr::BOTTOM;
1654
1655 if (tj->base() == Type::AnyPtr)
1656 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1657
1658 // Flatten all to bottom for now
1659 switch( _AliasLevel ) {
1660 case 0:
1661 tj = TypePtr::BOTTOM;
1662 break;
1663 case 1: // Flatten to: oop, static, field or array
1664 switch (tj->base()) {
1665 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1666 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1667 case Type::AryPtr: // do not distinguish arrays at all
1668 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1669 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1670 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1671 default: ShouldNotReachHere();
1672 }
1673 break;
1674 case 2: // No collapsing at level 2; keep all splits
1675 case 3: // No collapsing at level 3; keep all splits
1676 break;
1677 default:
1678 Unimplemented();
1679 }
1680
1681 offset = tj->offset();
1682 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1683
1684 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1685 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1686 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1687 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1688 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1689 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1690 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1691 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1692 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1693 assert( tj->ptr() != TypePtr::TopPTR &&
1694 tj->ptr() != TypePtr::AnyNull &&
1695 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1696 // assert( tj->ptr() != TypePtr::Constant ||
1697 // tj->base() == Type::RawPtr ||
1698 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1699
1700 return tj;
1701 }
1702
1703 void Compile::AliasType::Init(int i, const TypePtr* at) {
1704 _index = i;
1705 _adr_type = at;
1706 _field = NULL;
1707 _element = NULL;
1708 _is_rewritable = true; // default
1709 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1710 if (atoop != NULL && atoop->is_known_instance()) {
1711 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1712 _general_index = Compile::current()->get_alias_index(gt);
1713 } else {
1714 _general_index = 0;
1715 }
1716 }
1717
1718 BasicType Compile::AliasType::basic_type() const {
1719 if (element() != NULL) {
1720 const Type* element = adr_type()->is_aryptr()->elem();
1721 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1722 } if (field() != NULL) {
1723 return field()->layout_type();
1724 } else {
1725 return T_ILLEGAL; // unknown
1726 }
1727 }
1728
1729 //---------------------------------print_on------------------------------------
1730 #ifndef PRODUCT
1731 void Compile::AliasType::print_on(outputStream* st) {
1732 if (index() < 10)
1733 st->print("@ <%d> ", index());
1734 else st->print("@ <%d>", index());
1735 st->print(is_rewritable() ? " " : " RO");
1736 int offset = adr_type()->offset();
1737 if (offset == Type::OffsetBot)
1738 st->print(" +any");
1739 else st->print(" +%-3d", offset);
1740 st->print(" in ");
1741 adr_type()->dump_on(st);
1742 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1743 if (field() != NULL && tjp) {
1744 if (tjp->klass() != field()->holder() ||
1745 tjp->offset() != field()->offset_in_bytes()) {
1746 st->print(" != ");
1747 field()->print();
1748 st->print(" ***");
1749 }
1750 }
1751 }
1752
1753 void print_alias_types() {
1754 Compile* C = Compile::current();
1755 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1756 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1757 C->alias_type(idx)->print_on(tty);
1758 tty->cr();
1759 }
1760 }
1761 #endif
1762
1763
1764 //----------------------------probe_alias_cache--------------------------------
1765 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1766 intptr_t key = (intptr_t) adr_type;
1767 key ^= key >> logAliasCacheSize;
1768 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1769 }
1770
1771
1772 //-----------------------------grow_alias_types--------------------------------
1773 void Compile::grow_alias_types() {
1774 const int old_ats = _max_alias_types; // how many before?
1775 const int new_ats = old_ats; // how many more?
1776 const int grow_ats = old_ats+new_ats; // how many now?
1777 _max_alias_types = grow_ats;
1778 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1779 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1780 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1781 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1782 }
1783
1784
1785 //--------------------------------find_alias_type------------------------------
1786 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) {
1787 if (_AliasLevel == 0)
1788 return alias_type(AliasIdxBot);
1789
1790 AliasCacheEntry* ace = NULL;
1791 if (!uncached) {
1792 ace = probe_alias_cache(adr_type);
1793 if (ace->_adr_type == adr_type) {
1794 return alias_type(ace->_index);
1795 }
1796 }
1797
1798 // Handle special cases.
1799 if (adr_type == NULL) return alias_type(AliasIdxTop);
1800 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1801
1802 // Do it the slow way.
1803 const TypePtr* flat = flatten_alias_type(adr_type);
1804
1805 #ifdef ASSERT
1806 {
1807 ResourceMark rm;
1808 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1809 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1810 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1811 Type::str(adr_type));
1812 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1813 const TypeOopPtr* foop = flat->is_oopptr();
1814 // Scalarizable allocations have exact klass always.
1815 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1816 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1817 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1818 Type::str(foop), Type::str(xoop));
1819 }
1820 }
1821 #endif
1822
1823 int idx = AliasIdxTop;
1824 for (int i = 0; i < num_alias_types(); i++) {
1825 if (alias_type(i)->adr_type() == flat) {
1826 idx = i;
1827 break;
1828 }
1829 }
1830
1831 if (idx == AliasIdxTop) {
1832 if (no_create) return NULL;
1833 // Grow the array if necessary.
1834 if (_num_alias_types == _max_alias_types) grow_alias_types();
1835 // Add a new alias type.
1836 idx = _num_alias_types++;
1837 _alias_types[idx]->Init(idx, flat);
1838 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1839 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1840 if (flat->isa_instptr()) {
1841 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1842 && flat->is_instptr()->klass() == env()->Class_klass())
1843 alias_type(idx)->set_rewritable(false);
1844 }
1845 ciField* field = NULL;
1846 if (flat->isa_aryptr()) {
1847 #ifdef ASSERT
1848 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1849 // (T_BYTE has the weakest alignment and size restrictions...)
1850 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1851 #endif
1852 const Type* elemtype = flat->is_aryptr()->elem();
1853 if (flat->offset() == TypePtr::OffsetBot) {
1854 alias_type(idx)->set_element(elemtype);
1855 }
1856 int field_offset = flat->is_aryptr()->field_offset().get();
1857 if (elemtype->isa_valuetype() &&
1858 elemtype->value_klass() != NULL &&
1859 field_offset != Type::OffsetBot) {
1860 ciValueKlass* vk = elemtype->value_klass();
1861 field_offset += vk->first_field_offset();
1862 field = vk->get_field_by_offset(field_offset, false);
1863 }
1864 }
1865 if (flat->isa_klassptr()) {
1866 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1867 alias_type(idx)->set_rewritable(false);
1868 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1869 alias_type(idx)->set_rewritable(false);
1870 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1871 alias_type(idx)->set_rewritable(false);
1872 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1873 alias_type(idx)->set_rewritable(false);
1874 if (flat->offset() == in_bytes(Klass::layout_helper_offset()))
1875 alias_type(idx)->set_rewritable(false);
1876 }
1877 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1878 // but the base pointer type is not distinctive enough to identify
1879 // references into JavaThread.)
1880
1881 // Check for final fields.
1882 const TypeInstPtr* tinst = flat->isa_instptr();
1883 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1884 if (tinst->const_oop() != NULL &&
1885 tinst->klass() == ciEnv::current()->Class_klass() &&
1886 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1887 // static field
1888 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1889 field = k->get_field_by_offset(tinst->offset(), true);
1890 } else if (tinst->klass()->is_valuetype()) {
1891 // Value type field
1892 ciValueKlass* vk = tinst->value_klass();
1893 field = vk->get_field_by_offset(tinst->offset(), false);
1894 } else {
1895 ciInstanceKlass* k = tinst->klass()->as_instance_klass();
1896 field = k->get_field_by_offset(tinst->offset(), false);
1897 }
1898 }
1899 assert(field == NULL ||
1900 original_field == NULL ||
1901 (field->holder() == original_field->holder() &&
1902 field->offset() == original_field->offset() &&
1903 field->is_static() == original_field->is_static()), "wrong field?");
1904 // Set field() and is_rewritable() attributes.
1905 if (field != NULL) alias_type(idx)->set_field(field);
1906 }
1907
1908 // Fill the cache for next time.
1909 if (!uncached) {
1910 ace->_adr_type = adr_type;
1911 ace->_index = idx;
1912 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1913
1914 // Might as well try to fill the cache for the flattened version, too.
1915 AliasCacheEntry* face = probe_alias_cache(flat);
1916 if (face->_adr_type == NULL) {
1917 face->_adr_type = flat;
1918 face->_index = idx;
1919 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1920 }
1921 }
1922
1923 return alias_type(idx);
1924 }
1925
1926
1927 Compile::AliasType* Compile::alias_type(ciField* field) {
1928 const TypeOopPtr* t;
1929 if (field->is_static())
1930 t = TypeInstPtr::make(field->holder()->java_mirror());
1931 else
1932 t = TypeOopPtr::make_from_klass_raw(field->holder());
1933 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1934 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1935 return atp;
1936 }
1937
1938
1939 //------------------------------have_alias_type--------------------------------
1940 bool Compile::have_alias_type(const TypePtr* adr_type) {
1941 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1942 if (ace->_adr_type == adr_type) {
1943 return true;
1944 }
1945
1946 // Handle special cases.
1947 if (adr_type == NULL) return true;
1948 if (adr_type == TypePtr::BOTTOM) return true;
1949
1950 return find_alias_type(adr_type, true, NULL) != NULL;
1951 }
1952
1953 //-----------------------------must_alias--------------------------------------
1954 // True if all values of the given address type are in the given alias category.
1955 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1956 if (alias_idx == AliasIdxBot) return true; // the universal category
1957 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1958 if (alias_idx == AliasIdxTop) return false; // the empty category
1959 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1960
1961 // the only remaining possible overlap is identity
1962 int adr_idx = get_alias_index(adr_type);
1963 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1964 assert(adr_idx == alias_idx ||
1965 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1966 && adr_type != TypeOopPtr::BOTTOM),
1967 "should not be testing for overlap with an unsafe pointer");
1968 return adr_idx == alias_idx;
1969 }
1970
1971 //------------------------------can_alias--------------------------------------
1972 // True if any values of the given address type are in the given alias category.
1973 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1974 if (alias_idx == AliasIdxTop) return false; // the empty category
1975 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1976 if (alias_idx == AliasIdxBot) return true; // the universal category
1977 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1978
1979 // the only remaining possible overlap is identity
1980 int adr_idx = get_alias_index(adr_type);
1981 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1982 return adr_idx == alias_idx;
1983 }
1984
1985
1986
1987 //---------------------------pop_warm_call-------------------------------------
1988 WarmCallInfo* Compile::pop_warm_call() {
1989 WarmCallInfo* wci = _warm_calls;
1990 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1991 return wci;
1992 }
1993
1994 //----------------------------Inline_Warm--------------------------------------
1995 int Compile::Inline_Warm() {
1996 // If there is room, try to inline some more warm call sites.
1997 // %%% Do a graph index compaction pass when we think we're out of space?
1998 if (!InlineWarmCalls) return 0;
1999
2000 int calls_made_hot = 0;
2001 int room_to_grow = NodeCountInliningCutoff - unique();
2002 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
2003 int amount_grown = 0;
2004 WarmCallInfo* call;
2005 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
2006 int est_size = (int)call->size();
2007 if (est_size > (room_to_grow - amount_grown)) {
2008 // This one won't fit anyway. Get rid of it.
2009 call->make_cold();
2010 continue;
2011 }
2012 call->make_hot();
2013 calls_made_hot++;
2014 amount_grown += est_size;
2015 amount_to_grow -= est_size;
2016 }
2017
2018 if (calls_made_hot > 0) set_major_progress();
2019 return calls_made_hot;
2020 }
2021
2022
2023 //----------------------------Finish_Warm--------------------------------------
2024 void Compile::Finish_Warm() {
2025 if (!InlineWarmCalls) return;
2026 if (failing()) return;
2027 if (warm_calls() == NULL) return;
2028
2029 // Clean up loose ends, if we are out of space for inlining.
2030 WarmCallInfo* call;
2031 while ((call = pop_warm_call()) != NULL) {
2032 call->make_cold();
2033 }
2034 }
2035
2036 //---------------------cleanup_loop_predicates-----------------------
2037 // Remove the opaque nodes that protect the predicates so that all unused
2038 // checks and uncommon_traps will be eliminated from the ideal graph
2039 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
2040 if (predicate_count()==0) return;
2041 for (int i = predicate_count(); i > 0; i--) {
2042 Node * n = predicate_opaque1_node(i-1);
2043 assert(n->Opcode() == Op_Opaque1, "must be");
2044 igvn.replace_node(n, n->in(1));
2045 }
2046 assert(predicate_count()==0, "should be clean!");
2047 }
2048
2049 void Compile::add_range_check_cast(Node* n) {
2050 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2051 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
2052 _range_check_casts->append(n);
2053 }
2054
2055 // Remove all range check dependent CastIINodes.
2056 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
2057 for (int i = range_check_cast_count(); i > 0; i--) {
2058 Node* cast = range_check_cast_node(i-1);
2059 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2060 igvn.replace_node(cast, cast->in(1));
2061 }
2062 assert(range_check_cast_count() == 0, "should be empty");
2063 }
2064
2065 void Compile::add_opaque4_node(Node* n) {
2066 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2067 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2068 _opaque4_nodes->append(n);
2069 }
2070
2071 // Remove all Opaque4 nodes.
2072 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2073 for (int i = opaque4_count(); i > 0; i--) {
2074 Node* opaq = opaque4_node(i-1);
2075 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2076 igvn.replace_node(opaq, opaq->in(2));
2077 }
2078 assert(opaque4_count() == 0, "should be empty");
2079 }
2080
2081 void Compile::add_value_type(Node* n) {
2082 assert(n->is_ValueTypeBase(), "unexpected node");
2083 if (_value_type_nodes != NULL) {
2084 _value_type_nodes->push(n);
2085 }
2086 }
2087
2088 void Compile::remove_value_type(Node* n) {
2089 assert(n->is_ValueTypeBase(), "unexpected node");
2090 if (_value_type_nodes != NULL) {
2091 _value_type_nodes->remove(n);
2092 }
2093 }
2094
2095 // Does the return value keep otherwise useless value type allocations
2096 // alive?
2097 static bool return_val_keeps_allocations_alive(Node* ret_val) {
2098 ResourceMark rm;
2099 Unique_Node_List wq;
2100 wq.push(ret_val);
2101 bool some_allocations = false;
2102 for (uint i = 0; i < wq.size(); i++) {
2103 Node* n = wq.at(i);
2104 assert(!n->is_ValueTypeBase(), "chain of value type nodes");
2105 if (n->outcnt() > 1) {
2106 // Some other use for the allocation
2107 return false;
2108 } else if (n->is_Phi()) {
2109 for (uint j = 1; j < n->req(); j++) {
2110 wq.push(n->in(j));
2111 }
2112 } else if (n->is_CheckCastPP() &&
2113 n->in(1)->is_Proj() &&
2114 n->in(1)->in(0)->is_Allocate()) {
2115 some_allocations = true;
2116 }
2117 }
2118 return some_allocations;
2119 }
2120
2121 void Compile::process_value_types(PhaseIterGVN &igvn) {
2122 // Make value types scalar in safepoints
2123 while (_value_type_nodes->size() != 0) {
2124 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase();
2125 vt->make_scalar_in_safepoints(&igvn);
2126 if (vt->is_ValueTypePtr()) {
2127 igvn.replace_node(vt, vt->get_oop());
2128 } else if (vt->outcnt() == 0) {
2129 igvn.remove_dead_node(vt);
2130 }
2131 }
2132 _value_type_nodes = NULL;
2133 if (tf()->returns_value_type_as_fields()) {
2134 Node* ret = NULL;
2135 for (uint i = 1; i < root()->req(); i++){
2136 Node* in = root()->in(i);
2137 if (in->Opcode() == Op_Return) {
2138 assert(ret == NULL, "only one return");
2139 ret = in;
2140 }
2141 }
2142 if (ret != NULL) {
2143 Node* ret_val = ret->in(TypeFunc::Parms);
2144 if (igvn.type(ret_val)->isa_oopptr() &&
2145 return_val_keeps_allocations_alive(ret_val)) {
2146 igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn));
2147 assert(ret_val->outcnt() == 0, "should be dead now");
2148 igvn.remove_dead_node(ret_val);
2149 }
2150 }
2151 }
2152 igvn.optimize();
2153 }
2154
2155 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) {
2156 if (!_has_flattened_accesses) {
2157 return;
2158 }
2159 // Initially, all flattened array accesses share the same slice to
2160 // keep dependencies with Object[] array accesses (that could be
2161 // to a flattened array) correct. We're done with parsing so we
2162 // now know all flattened array accesses in this compile
2163 // unit. Let's move flattened array accesses to their own slice,
2164 // one per element field. This should help memory access
2165 // optimizations.
2166 ResourceMark rm;
2167 Unique_Node_List wq;
2168 wq.push(root());
2169
2170 Node_List mergememnodes;
2171 Node_List memnodes;
2172
2173 // Alias index currently shared by all flattened memory accesses
2174 int index = get_alias_index(TypeAryPtr::VALUES);
2175
2176 // Find MergeMem nodes and flattened array accesses
2177 for (uint i = 0; i < wq.size(); i++) {
2178 Node* n = wq.at(i);
2179 if (n->is_Mem()) {
2180 const TypePtr* adr_type = get_adr_type(get_alias_index(n->adr_type()));
2181 if (adr_type == TypeAryPtr::VALUES) {
2182 memnodes.push(n);
2183 }
2184 } else if (n->is_MergeMem()) {
2185 MergeMemNode* mm = n->as_MergeMem();
2186 if (mm->memory_at(index) != mm->base_memory()) {
2187 mergememnodes.push(n);
2188 }
2189 }
2190 for (uint j = 0; j < n->req(); j++) {
2191 Node* m = n->in(j);
2192 if (m != NULL) {
2193 wq.push(m);
2194 }
2195 }
2196 }
2197
2198 if (memnodes.size() > 0) {
2199 _flattened_accesses_share_alias = false;
2200
2201 // We are going to change the slice for the flattened array
2202 // accesses so we need to clear the cache entries that refer to
2203 // them.
2204 for (uint i = 0; i < AliasCacheSize; i++) {
2205 AliasCacheEntry* ace = &_alias_cache[i];
2206 if (ace->_adr_type != NULL &&
2207 ace->_adr_type->isa_aryptr() &&
2208 ace->_adr_type->is_aryptr()->elem()->isa_valuetype()) {
2209 ace->_adr_type = NULL;
2210 ace->_index = 0;
2211 }
2212 }
2213
2214 // Find what aliases we are going to add
2215 int start_alias = num_alias_types()-1;
2216 int stop_alias = 0;
2217
2218 for (uint i = 0; i < memnodes.size(); i++) {
2219 Node* m = memnodes.at(i);
2220 const TypePtr* adr_type = m->adr_type();
2221 #ifdef ASSERT
2222 m->as_Mem()->set_adr_type(adr_type);
2223 #endif
2224 int idx = get_alias_index(adr_type);
2225 start_alias = MIN2(start_alias, idx);
2226 stop_alias = MAX2(stop_alias, idx);
2227 }
2228
2229 assert(stop_alias >= start_alias, "should have expanded aliases");
2230
2231 Node_Stack stack(0);
2232 #ifdef ASSERT
2233 VectorSet seen(Thread::current()->resource_area());
2234 #endif
2235 // Now let's fix the memory graph so each flattened array access
2236 // is moved to the right slice. Start from the MergeMem nodes.
2237 uint last = unique();
2238 for (uint i = 0; i < mergememnodes.size(); i++) {
2239 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem();
2240 Node* n = current->memory_at(index);
2241 MergeMemNode* mm = NULL;
2242 do {
2243 // Follow memory edges through memory accesses, phis and
2244 // narrow membars and push nodes on the stack. Once we hit
2245 // bottom memory, we pop element off the stack one at a
2246 // time, in reverse order, and move them to the right slice
2247 // by changing their memory edges.
2248 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::VALUES) {
2249 assert(!seen.test_set(n->_idx), "");
2250 // Uses (a load for instance) will need to be moved to the
2251 // right slice as well and will get a new memory state
2252 // that we don't know yet. The use could also be the
2253 // backedge of a loop. We put a place holder node between
2254 // the memory node and its uses. We replace that place
2255 // holder with the correct memory state once we know it,
2256 // i.e. when nodes are popped off the stack. Using the
2257 // place holder make the logic work in the presence of
2258 // loops.
2259 if (n->outcnt() > 1) {
2260 Node* place_holder = NULL;
2261 assert(!n->has_out_with(Op_Node), "");
2262 for (DUIterator k = n->outs(); n->has_out(k); k++) {
2263 Node* u = n->out(k);
2264 if (u != current && u->_idx < last) {
2265 bool success = false;
2266 for (uint l = 0; l < u->req(); l++) {
2267 if (!stack.is_empty() && u == stack.node() && l == stack.index()) {
2268 continue;
2269 }
2270 Node* in = u->in(l);
2271 if (in == n) {
2272 if (place_holder == NULL) {
2273 place_holder = new Node(1);
2274 place_holder->init_req(0, n);
2275 }
2276 igvn.replace_input_of(u, l, place_holder);
2277 success = true;
2278 }
2279 }
2280 if (success) {
2281 --k;
2282 }
2283 }
2284 }
2285 }
2286 if (n->is_Phi()) {
2287 stack.push(n, 1);
2288 n = n->in(1);
2289 } else if (n->is_Mem()) {
2290 stack.push(n, n->req());
2291 n = n->in(MemNode::Memory);
2292 } else {
2293 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, "");
2294 stack.push(n, n->req());
2295 n = n->in(0)->in(TypeFunc::Memory);
2296 }
2297 } else {
2298 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), "");
2299 // Build a new MergeMem node to carry the new memory state
2300 // as we build it. IGVN should fold extraneous MergeMem
2301 // nodes.
2302 mm = MergeMemNode::make(n);
2303 igvn.register_new_node_with_optimizer(mm);
2304 while (stack.size() > 0) {
2305 Node* m = stack.node();
2306 uint idx = stack.index();
2307 if (m->is_Mem()) {
2308 // Move memory node to its new slice
2309 const TypePtr* adr_type = m->adr_type();
2310 int alias = get_alias_index(adr_type);
2311 Node* prev = mm->memory_at(alias);
2312 igvn.replace_input_of(m, MemNode::Memory, prev);
2313 mm->set_memory_at(alias, m);
2314 } else if (m->is_Phi()) {
2315 // We need as many new phis as there are new aliases
2316 igvn.replace_input_of(m, idx, mm);
2317 if (idx == m->req()-1) {
2318 Node* r = m->in(0);
2319 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2320 const Type* adr_type = get_adr_type(j);
2321 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2322 continue;
2323 }
2324 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j));
2325 igvn.register_new_node_with_optimizer(phi);
2326 for (uint k = 1; k < m->req(); k++) {
2327 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j));
2328 }
2329 mm->set_memory_at(j, phi);
2330 }
2331 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2332 igvn.register_new_node_with_optimizer(base_phi);
2333 for (uint k = 1; k < m->req(); k++) {
2334 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory());
2335 }
2336 mm->set_base_memory(base_phi);
2337 }
2338 } else {
2339 // This is a MemBarCPUOrder node from
2340 // Parse::array_load()/Parse::array_store(), in the
2341 // branch that handles flattened arrays hidden under
2342 // an Object[] array. We also need one new membar per
2343 // new alias to keep the unknown access that the
2344 // membars protect properly ordered with accesses to
2345 // known flattened array.
2346 assert(m->is_Proj(), "projection expected");
2347 Node* ctrl = m->in(0)->in(TypeFunc::Control);
2348 igvn.replace_input_of(m->in(0), TypeFunc::Control, top());
2349 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2350 const Type* adr_type = get_adr_type(j);
2351 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2352 continue;
2353 }
2354 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL);
2355 igvn.register_new_node_with_optimizer(mb);
2356 Node* mem = mm->memory_at(j);
2357 mb->init_req(TypeFunc::Control, ctrl);
2358 mb->init_req(TypeFunc::Memory, mem);
2359 ctrl = new ProjNode(mb, TypeFunc::Control);
2360 igvn.register_new_node_with_optimizer(ctrl);
2361 mem = new ProjNode(mb, TypeFunc::Memory);
2362 igvn.register_new_node_with_optimizer(mem);
2363 mm->set_memory_at(j, mem);
2364 }
2365 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl);
2366 }
2367 if (idx < m->req()-1) {
2368 idx += 1;
2369 stack.set_index(idx);
2370 n = m->in(idx);
2371 break;
2372 }
2373 // Take care of place holder nodes
2374 if (m->has_out_with(Op_Node)) {
2375 Node* place_holder = m->find_out_with(Op_Node);
2376 if (place_holder != NULL) {
2377 Node* mm_clone = mm->clone();
2378 igvn.register_new_node_with_optimizer(mm_clone);
2379 Node* hook = new Node(1);
2380 hook->init_req(0, mm);
2381 igvn.replace_node(place_holder, mm_clone);
2382 hook->destruct();
2383 }
2384 assert(!m->has_out_with(Op_Node), "place holder should be gone now");
2385 }
2386 stack.pop();
2387 }
2388 }
2389 } while(stack.size() > 0);
2390 // Fix the memory state at the MergeMem we started from
2391 igvn.rehash_node_delayed(current);
2392 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) {
2393 const Type* adr_type = get_adr_type(j);
2394 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) {
2395 continue;
2396 }
2397 current->set_memory_at(j, mm);
2398 }
2399 current->set_memory_at(index, current->base_memory());
2400 }
2401 igvn.optimize();
2402 }
2403 print_method(PHASE_SPLIT_VALUES_ARRAY, 2);
2404 }
2405
2406
2407 // StringOpts and late inlining of string methods
2408 void Compile::inline_string_calls(bool parse_time) {
2409 {
2410 // remove useless nodes to make the usage analysis simpler
2411 ResourceMark rm;
2412 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2413 }
2414
2415 {
2416 ResourceMark rm;
2417 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2418 PhaseStringOpts pso(initial_gvn(), for_igvn());
2419 print_method(PHASE_AFTER_STRINGOPTS, 3);
2420 }
2421
2422 // now inline anything that we skipped the first time around
2423 if (!parse_time) {
2424 _late_inlines_pos = _late_inlines.length();
2425 }
2426
2427 while (_string_late_inlines.length() > 0) {
2428 CallGenerator* cg = _string_late_inlines.pop();
2429 cg->do_late_inline();
2430 if (failing()) return;
2431 }
2432 _string_late_inlines.trunc_to(0);
2433 }
2434
2435 // Late inlining of boxing methods
2436 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2437 if (_boxing_late_inlines.length() > 0) {
2438 assert(has_boxed_value(), "inconsistent");
2439
2440 PhaseGVN* gvn = initial_gvn();
2441 set_inlining_incrementally(true);
2442
2443 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2444 for_igvn()->clear();
2445 gvn->replace_with(&igvn);
2446
2447 _late_inlines_pos = _late_inlines.length();
2448
2449 while (_boxing_late_inlines.length() > 0) {
2450 CallGenerator* cg = _boxing_late_inlines.pop();
2451 cg->do_late_inline();
2452 if (failing()) return;
2453 }
2454 _boxing_late_inlines.trunc_to(0);
2455
2456 inline_incrementally_cleanup(igvn);
2457
2458 set_inlining_incrementally(false);
2459 }
2460 }
2461
2462 bool Compile::inline_incrementally_one() {
2463 assert(IncrementalInline, "incremental inlining should be on");
2464
2465 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2466 set_inlining_progress(false);
2467 set_do_cleanup(false);
2468 int i = 0;
2469 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2470 CallGenerator* cg = _late_inlines.at(i);
2471 _late_inlines_pos = i+1;
2472 cg->do_late_inline();
2473 if (failing()) return false;
2474 }
2475 int j = 0;
2476 for (; i < _late_inlines.length(); i++, j++) {
2477 _late_inlines.at_put(j, _late_inlines.at(i));
2478 }
2479 _late_inlines.trunc_to(j);
2480 assert(inlining_progress() || _late_inlines.length() == 0, "");
2481
2482 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2483
2484 set_inlining_progress(false);
2485 set_do_cleanup(false);
2486 return (_late_inlines.length() > 0) && !needs_cleanup;
2487 }
2488
2489 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2490 {
2491 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2492 ResourceMark rm;
2493 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2494 }
2495 {
2496 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2497 igvn = PhaseIterGVN(initial_gvn());
2498 igvn.optimize();
2499 }
2500 }
2501
2502 // Perform incremental inlining until bound on number of live nodes is reached
2503 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2504 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2505
2506 set_inlining_incrementally(true);
2507 uint low_live_nodes = 0;
2508
2509 while (_late_inlines.length() > 0) {
2510 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2511 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2512 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2513 // PhaseIdealLoop is expensive so we only try it once we are
2514 // out of live nodes and we only try it again if the previous
2515 // helped got the number of nodes down significantly
2516 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2517 if (failing()) return;
2518 low_live_nodes = live_nodes();
2519 _major_progress = true;
2520 }
2521
2522 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2523 break; // finish
2524 }
2525 }
2526
2527 for_igvn()->clear();
2528 initial_gvn()->replace_with(&igvn);
2529
2530 while (inline_incrementally_one()) {
2531 assert(!failing(), "inconsistent");
2532 }
2533
2534 if (failing()) return;
2535
2536 inline_incrementally_cleanup(igvn);
2537
2538 if (failing()) return;
2539 }
2540 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2541
2542 if (_string_late_inlines.length() > 0) {
2543 assert(has_stringbuilder(), "inconsistent");
2544 for_igvn()->clear();
2545 initial_gvn()->replace_with(&igvn);
2546
2547 inline_string_calls(false);
2548
2549 if (failing()) return;
2550
2551 inline_incrementally_cleanup(igvn);
2552 }
2553
2554 set_inlining_incrementally(false);
2555 }
2556
2557
2558 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2559 if(_loop_opts_cnt > 0) {
2560 debug_only( int cnt = 0; );
2561 while(major_progress() && (_loop_opts_cnt > 0)) {
2562 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2563 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2564 PhaseIdealLoop::optimize(igvn, mode);
2565 _loop_opts_cnt--;
2566 if (failing()) return false;
2567 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2568 }
2569 }
2570 return true;
2571 }
2572
2573 // Remove edges from "root" to each SafePoint at a backward branch.
2574 // They were inserted during parsing (see add_safepoint()) to make
2575 // infinite loops without calls or exceptions visible to root, i.e.,
2576 // useful.
2577 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2578 Node *r = root();
2579 if (r != NULL) {
2580 for (uint i = r->req(); i < r->len(); ++i) {
2581 Node *n = r->in(i);
2582 if (n != NULL && n->is_SafePoint()) {
2583 r->rm_prec(i);
2584 if (n->outcnt() == 0) {
2585 igvn.remove_dead_node(n);
2586 }
2587 --i;
2588 }
2589 }
2590 }
2591 }
2592
2593 //------------------------------Optimize---------------------------------------
2594 // Given a graph, optimize it.
2595 void Compile::Optimize() {
2596 TracePhase tp("optimizer", &timers[_t_optimizer]);
2597
2598 #ifndef PRODUCT
2599 if (_directive->BreakAtCompileOption) {
2600 BREAKPOINT;
2601 }
2602
2603 #endif
2604
2605 #ifdef ASSERT
2606 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2607 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2608 #endif
2609
2610 ResourceMark rm;
2611
2612 print_inlining_reinit();
2613
2614 NOT_PRODUCT( verify_graph_edges(); )
2615
2616 print_method(PHASE_AFTER_PARSING);
2617
2618 {
2619 // Iterative Global Value Numbering, including ideal transforms
2620 // Initialize IterGVN with types and values from parse-time GVN
2621 PhaseIterGVN igvn(initial_gvn());
2622 #ifdef ASSERT
2623 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2624 #endif
2625 {
2626 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2627 igvn.optimize();
2628 }
2629
2630 if (failing()) return;
2631
2632 print_method(PHASE_ITER_GVN1, 2);
2633
2634 inline_incrementally(igvn);
2635
2636 print_method(PHASE_INCREMENTAL_INLINE, 2);
2637
2638 if (failing()) return;
2639
2640 if (eliminate_boxing()) {
2641 // Inline valueOf() methods now.
2642 inline_boxing_calls(igvn);
2643
2644 if (AlwaysIncrementalInline) {
2645 inline_incrementally(igvn);
2646 }
2647
2648 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2649
2650 if (failing()) return;
2651 }
2652
2653 // Now that all inlining is over, cut edge from root to loop
2654 // safepoints
2655 remove_root_to_sfpts_edges(igvn);
2656
2657 // Remove the speculative part of types and clean up the graph from
2658 // the extra CastPP nodes whose only purpose is to carry them. Do
2659 // that early so that optimizations are not disrupted by the extra
2660 // CastPP nodes.
2661 remove_speculative_types(igvn);
2662
2663 // No more new expensive nodes will be added to the list from here
2664 // so keep only the actual candidates for optimizations.
2665 cleanup_expensive_nodes(igvn);
2666
2667 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2668 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2669 initial_gvn()->replace_with(&igvn);
2670 for_igvn()->clear();
2671 Unique_Node_List new_worklist(C->comp_arena());
2672 {
2673 ResourceMark rm;
2674 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2675 }
2676 set_for_igvn(&new_worklist);
2677 igvn = PhaseIterGVN(initial_gvn());
2678 igvn.optimize();
2679 }
2680
2681 if (_value_type_nodes->size() > 0) {
2682 // Do this once all inlining is over to avoid getting inconsistent debug info
2683 process_value_types(igvn);
2684 }
2685
2686 adjust_flattened_array_access_aliases(igvn);
2687
2688 // Perform escape analysis
2689 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2690 if (has_loops()) {
2691 // Cleanup graph (remove dead nodes).
2692 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2693 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2694 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2695 if (failing()) return;
2696 }
2697 ConnectionGraph::do_analysis(this, &igvn);
2698
2699 if (failing()) return;
2700
2701 // Optimize out fields loads from scalar replaceable allocations.
2702 igvn.optimize();
2703 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2704
2705 if (failing()) return;
2706
2707 if (congraph() != NULL && macro_count() > 0) {
2708 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2709 PhaseMacroExpand mexp(igvn);
2710 mexp.eliminate_macro_nodes();
2711 igvn.set_delay_transform(false);
2712
2713 igvn.optimize();
2714 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2715
2716 if (failing()) return;
2717 }
2718 }
2719
2720 // Loop transforms on the ideal graph. Range Check Elimination,
2721 // peeling, unrolling, etc.
2722
2723 // Set loop opts counter
2724 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2725 {
2726 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2727 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2728 _loop_opts_cnt--;
2729 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2730 if (failing()) return;
2731 }
2732 // Loop opts pass if partial peeling occurred in previous pass
2733 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2734 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2735 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2736 _loop_opts_cnt--;
2737 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2738 if (failing()) return;
2739 }
2740 // Loop opts pass for loop-unrolling before CCP
2741 if(major_progress() && (_loop_opts_cnt > 0)) {
2742 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2743 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2744 _loop_opts_cnt--;
2745 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2746 }
2747 if (!failing()) {
2748 // Verify that last round of loop opts produced a valid graph
2749 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2750 PhaseIdealLoop::verify(igvn);
2751 }
2752 }
2753 if (failing()) return;
2754
2755 // Conditional Constant Propagation;
2756 PhaseCCP ccp( &igvn );
2757 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2758 {
2759 TracePhase tp("ccp", &timers[_t_ccp]);
2760 ccp.do_transform();
2761 }
2762 print_method(PHASE_CPP1, 2);
2763
2764 assert( true, "Break here to ccp.dump_old2new_map()");
2765
2766 // Iterative Global Value Numbering, including ideal transforms
2767 {
2768 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2769 igvn = ccp;
2770 igvn.optimize();
2771 }
2772
2773 print_method(PHASE_ITER_GVN2, 2);
2774
2775 if (failing()) return;
2776
2777 // Loop transforms on the ideal graph. Range Check Elimination,
2778 // peeling, unrolling, etc.
2779 if (!optimize_loops(igvn, LoopOptsDefault)) {
2780 return;
2781 }
2782
2783 #if INCLUDE_ZGC
2784 if (UseZGC) {
2785 ZBarrierSetC2::find_dominating_barriers(igvn);
2786 }
2787 #endif
2788
2789 if (failing()) return;
2790
2791 // Ensure that major progress is now clear
2792 C->clear_major_progress();
2793
2794 {
2795 // Verify that all previous optimizations produced a valid graph
2796 // at least to this point, even if no loop optimizations were done.
2797 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2798 PhaseIdealLoop::verify(igvn);
2799 }
2800
2801 if (range_check_cast_count() > 0) {
2802 // No more loop optimizations. Remove all range check dependent CastIINodes.
2803 C->remove_range_check_casts(igvn);
2804 igvn.optimize();
2805 }
2806
2807 #ifdef ASSERT
2808 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2809 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand);
2810 #endif
2811
2812 {
2813 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2814 PhaseMacroExpand mex(igvn);
2815 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2816 if (mex.expand_macro_nodes()) {
2817 assert(failing(), "must bail out w/ explicit message");
2818 return;
2819 }
2820 }
2821
2822 {
2823 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2824 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2);
2825 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2826 if (bs->expand_barriers(this, igvn)) {
2827 assert(failing(), "must bail out w/ explicit message");
2828 return;
2829 }
2830 }
2831
2832 if (opaque4_count() > 0) {
2833 C->remove_opaque4_nodes(igvn);
2834 igvn.optimize();
2835 }
2836
2837 DEBUG_ONLY( _modified_nodes = NULL; )
2838 } // (End scope of igvn; run destructor if necessary for asserts.)
2839
2840 process_print_inlining();
2841 // A method with only infinite loops has no edges entering loops from root
2842 {
2843 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2844 if (final_graph_reshaping()) {
2845 assert(failing(), "must bail out w/ explicit message");
2846 return;
2847 }
2848 }
2849
2850 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2851 }
2852
2853 //------------------------------Code_Gen---------------------------------------
2854 // Given a graph, generate code for it
2855 void Compile::Code_Gen() {
2856 if (failing()) {
2857 return;
2858 }
2859
2860 // Perform instruction selection. You might think we could reclaim Matcher
2861 // memory PDQ, but actually the Matcher is used in generating spill code.
2862 // Internals of the Matcher (including some VectorSets) must remain live
2863 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2864 // set a bit in reclaimed memory.
2865
2866 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2867 // nodes. Mapping is only valid at the root of each matched subtree.
2868 NOT_PRODUCT( verify_graph_edges(); )
2869
2870 Matcher matcher;
2871 _matcher = &matcher;
2872 {
2873 TracePhase tp("matcher", &timers[_t_matcher]);
2874 matcher.match();
2875 }
2876 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2877 // nodes. Mapping is only valid at the root of each matched subtree.
2878 NOT_PRODUCT( verify_graph_edges(); )
2879
2880 // If you have too many nodes, or if matching has failed, bail out
2881 check_node_count(0, "out of nodes matching instructions");
2882 if (failing()) {
2883 return;
2884 }
2885
2886 print_method(PHASE_MATCHING, 2);
2887
2888 // Build a proper-looking CFG
2889 PhaseCFG cfg(node_arena(), root(), matcher);
2890 _cfg = &cfg;
2891 {
2892 TracePhase tp("scheduler", &timers[_t_scheduler]);
2893 bool success = cfg.do_global_code_motion();
2894 if (!success) {
2895 return;
2896 }
2897
2898 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2899 NOT_PRODUCT( verify_graph_edges(); )
2900 debug_only( cfg.verify(); )
2901 }
2902
2903 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2904 _regalloc = ®alloc;
2905 {
2906 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2907 // Perform register allocation. After Chaitin, use-def chains are
2908 // no longer accurate (at spill code) and so must be ignored.
2909 // Node->LRG->reg mappings are still accurate.
2910 _regalloc->Register_Allocate();
2911
2912 // Bail out if the allocator builds too many nodes
2913 if (failing()) {
2914 return;
2915 }
2916 }
2917
2918 // Prior to register allocation we kept empty basic blocks in case the
2919 // the allocator needed a place to spill. After register allocation we
2920 // are not adding any new instructions. If any basic block is empty, we
2921 // can now safely remove it.
2922 {
2923 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2924 cfg.remove_empty_blocks();
2925 if (do_freq_based_layout()) {
2926 PhaseBlockLayout layout(cfg);
2927 } else {
2928 cfg.set_loop_alignment();
2929 }
2930 cfg.fixup_flow();
2931 }
2932
2933 // Apply peephole optimizations
2934 if( OptoPeephole ) {
2935 TracePhase tp("peephole", &timers[_t_peephole]);
2936 PhasePeephole peep( _regalloc, cfg);
2937 peep.do_transform();
2938 }
2939
2940 // Do late expand if CPU requires this.
2941 if (Matcher::require_postalloc_expand) {
2942 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2943 cfg.postalloc_expand(_regalloc);
2944 }
2945
2946 // Convert Nodes to instruction bits in a buffer
2947 {
2948 TraceTime tp("output", &timers[_t_output], CITime);
2949 Output();
2950 }
2951
2952 print_method(PHASE_FINAL_CODE);
2953
2954 // He's dead, Jim.
2955 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2956 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2957 }
2958
2959
2960 //------------------------------dump_asm---------------------------------------
2961 // Dump formatted assembly
2962 #if defined(SUPPORT_OPTO_ASSEMBLY)
2963 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2964
2965 int pc_digits = 3; // #chars required for pc
2966 int sb_chars = 3; // #chars for "start bundle" indicator
2967 int tab_size = 8;
2968 if (pcs != NULL) {
2969 int max_pc = 0;
2970 for (uint i = 0; i < pc_limit; i++) {
2971 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
2972 }
2973 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
2974 }
2975 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
2976
2977 bool cut_short = false;
2978 st->print_cr("#");
2979 st->print("# "); _tf->dump_on(st); st->cr();
2980 st->print_cr("#");
2981
2982 // For all blocks
2983 int pc = 0x0; // Program counter
2984 char starts_bundle = ' ';
2985 _regalloc->dump_frame();
2986
2987 Node *n = NULL;
2988 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2989 if (VMThread::should_terminate()) {
2990 cut_short = true;
2991 break;
2992 }
2993 Block* block = _cfg->get_block(i);
2994 if (block->is_connector() && !Verbose) {
2995 continue;
2996 }
2997 n = block->head();
2998 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2999 pc = pcs[n->_idx];
3000 st->print("%*.*x", pc_digits, pc_digits, pc);
3001 }
3002 st->fill_to(prefix_len);
3003 block->dump_head(_cfg, st);
3004 if (block->is_connector()) {
3005 st->fill_to(prefix_len);
3006 st->print_cr("# Empty connector block");
3007 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3008 st->fill_to(prefix_len);
3009 st->print_cr("# Block is sole successor of call");
3010 }
3011
3012 // For all instructions
3013 Node *delay = NULL;
3014 for (uint j = 0; j < block->number_of_nodes(); j++) {
3015 if (VMThread::should_terminate()) {
3016 cut_short = true;
3017 break;
3018 }
3019 n = block->get_node(j);
3020 if (valid_bundle_info(n)) {
3021 Bundle* bundle = node_bundling(n);
3022 if (bundle->used_in_unconditional_delay()) {
3023 delay = n;
3024 continue;
3025 }
3026 if (bundle->starts_bundle()) {
3027 starts_bundle = '+';
3028 }
3029 }
3030
3031 if (WizardMode) {
3032 n->dump();
3033 }
3034
3035 if( !n->is_Region() && // Dont print in the Assembly
3036 !n->is_Phi() && // a few noisely useless nodes
3037 !n->is_Proj() &&
3038 !n->is_MachTemp() &&
3039 !n->is_SafePointScalarObject() &&
3040 !n->is_Catch() && // Would be nice to print exception table targets
3041 !n->is_MergeMem() && // Not very interesting
3042 !n->is_top() && // Debug info table constants
3043 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3044 ) {
3045 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3046 pc = pcs[n->_idx];
3047 st->print("%*.*x", pc_digits, pc_digits, pc);
3048 } else {
3049 st->fill_to(pc_digits);
3050 }
3051 st->print(" %c ", starts_bundle);
3052 starts_bundle = ' ';
3053 st->fill_to(prefix_len);
3054 n->format(_regalloc, st);
3055 st->cr();
3056 }
3057
3058 // If we have an instruction with a delay slot, and have seen a delay,
3059 // then back up and print it
3060 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3061 // Coverity finding - Explicit null dereferenced.
3062 guarantee(delay != NULL, "no unconditional delay instruction");
3063 if (WizardMode) delay->dump();
3064
3065 if (node_bundling(delay)->starts_bundle())
3066 starts_bundle = '+';
3067 if ((pcs != NULL) && (n->_idx < pc_limit)) {
3068 pc = pcs[n->_idx];
3069 st->print("%*.*x", pc_digits, pc_digits, pc);
3070 } else {
3071 st->fill_to(pc_digits);
3072 }
3073 st->print(" %c ", starts_bundle);
3074 starts_bundle = ' ';
3075 st->fill_to(prefix_len);
3076 delay->format(_regalloc, st);
3077 st->cr();
3078 delay = NULL;
3079 }
3080
3081 // Dump the exception table as well
3082 if( n->is_Catch() && (Verbose || WizardMode) ) {
3083 // Print the exception table for this offset
3084 _handler_table.print_subtable_for(pc);
3085 }
3086 st->bol(); // Make sure we start on a new line
3087 }
3088 st->cr(); // one empty line between blocks
3089 assert(cut_short || delay == NULL, "no unconditional delay branch");
3090 } // End of per-block dump
3091
3092 if (cut_short) st->print_cr("*** disassembly is cut short ***");
3093 }
3094 #endif
3095
3096 //------------------------------Final_Reshape_Counts---------------------------
3097 // This class defines counters to help identify when a method
3098 // may/must be executed using hardware with only 24-bit precision.
3099 struct Final_Reshape_Counts : public StackObj {
3100 int _call_count; // count non-inlined 'common' calls
3101 int _float_count; // count float ops requiring 24-bit precision
3102 int _double_count; // count double ops requiring more precision
3103 int _java_call_count; // count non-inlined 'java' calls
3104 int _inner_loop_count; // count loops which need alignment
3105 VectorSet _visited; // Visitation flags
3106 Node_List _tests; // Set of IfNodes & PCTableNodes
3107
3108 Final_Reshape_Counts() :
3109 _call_count(0), _float_count(0), _double_count(0),
3110 _java_call_count(0), _inner_loop_count(0),
3111 _visited( Thread::current()->resource_area() ) { }
3112
3113 void inc_call_count () { _call_count ++; }
3114 void inc_float_count () { _float_count ++; }
3115 void inc_double_count() { _double_count++; }
3116 void inc_java_call_count() { _java_call_count++; }
3117 void inc_inner_loop_count() { _inner_loop_count++; }
3118
3119 int get_call_count () const { return _call_count ; }
3120 int get_float_count () const { return _float_count ; }
3121 int get_double_count() const { return _double_count; }
3122 int get_java_call_count() const { return _java_call_count; }
3123 int get_inner_loop_count() const { return _inner_loop_count; }
3124 };
3125
3126 #ifdef ASSERT
3127 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
3128 ciInstanceKlass *k = tp->klass()->as_instance_klass();
3129 // Make sure the offset goes inside the instance layout.
3130 return k->contains_field_offset(tp->offset());
3131 // Note that OffsetBot and OffsetTop are very negative.
3132 }
3133 #endif
3134
3135 // Eliminate trivially redundant StoreCMs and accumulate their
3136 // precedence edges.
3137 void Compile::eliminate_redundant_card_marks(Node* n) {
3138 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3139 if (n->in(MemNode::Address)->outcnt() > 1) {
3140 // There are multiple users of the same address so it might be
3141 // possible to eliminate some of the StoreCMs
3142 Node* mem = n->in(MemNode::Memory);
3143 Node* adr = n->in(MemNode::Address);
3144 Node* val = n->in(MemNode::ValueIn);
3145 Node* prev = n;
3146 bool done = false;
3147 // Walk the chain of StoreCMs eliminating ones that match. As
3148 // long as it's a chain of single users then the optimization is
3149 // safe. Eliminating partially redundant StoreCMs would require
3150 // cloning copies down the other paths.
3151 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3152 if (adr == mem->in(MemNode::Address) &&
3153 val == mem->in(MemNode::ValueIn)) {
3154 // redundant StoreCM
3155 if (mem->req() > MemNode::OopStore) {
3156 // Hasn't been processed by this code yet.
3157 n->add_prec(mem->in(MemNode::OopStore));
3158 } else {
3159 // Already converted to precedence edge
3160 for (uint i = mem->req(); i < mem->len(); i++) {
3161 // Accumulate any precedence edges
3162 if (mem->in(i) != NULL) {
3163 n->add_prec(mem->in(i));
3164 }
3165 }
3166 // Everything above this point has been processed.
3167 done = true;
3168 }
3169 // Eliminate the previous StoreCM
3170 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3171 assert(mem->outcnt() == 0, "should be dead");
3172 mem->disconnect_inputs(NULL, this);
3173 } else {
3174 prev = mem;
3175 }
3176 mem = prev->in(MemNode::Memory);
3177 }
3178 }
3179 }
3180
3181
3182 //------------------------------final_graph_reshaping_impl----------------------
3183 // Implement items 1-5 from final_graph_reshaping below.
3184 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3185
3186 if ( n->outcnt() == 0 ) return; // dead node
3187 uint nop = n->Opcode();
3188
3189 // Check for 2-input instruction with "last use" on right input.
3190 // Swap to left input. Implements item (2).
3191 if( n->req() == 3 && // two-input instruction
3192 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3193 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3194 n->in(2)->outcnt() == 1 &&// right use IS a last use
3195 !n->in(2)->is_Con() ) { // right use is not a constant
3196 // Check for commutative opcode
3197 switch( nop ) {
3198 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3199 case Op_MaxI: case Op_MinI:
3200 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3201 case Op_AndL: case Op_XorL: case Op_OrL:
3202 case Op_AndI: case Op_XorI: case Op_OrI: {
3203 // Move "last use" input to left by swapping inputs
3204 n->swap_edges(1, 2);
3205 break;
3206 }
3207 default:
3208 break;
3209 }
3210 }
3211
3212 #ifdef ASSERT
3213 if( n->is_Mem() ) {
3214 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3215 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3216 // oop will be recorded in oop map if load crosses safepoint
3217 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3218 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3219 "raw memory operations should have control edge");
3220 }
3221 if (n->is_MemBar()) {
3222 MemBarNode* mb = n->as_MemBar();
3223 if (mb->trailing_store() || mb->trailing_load_store()) {
3224 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3225 Node* mem = mb->in(MemBarNode::Precedent);
3226 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3227 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3228 } else if (mb->leading()) {
3229 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3230 }
3231 }
3232 #endif
3233 // Count FPU ops and common calls, implements item (3)
3234 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
3235 if (!gc_handled) {
3236 final_graph_reshaping_main_switch(n, frc, nop);
3237 }
3238
3239 // Collect CFG split points
3240 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3241 frc._tests.push(n);
3242 }
3243 }
3244
3245 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
3246 switch( nop ) {
3247 // Count all float operations that may use FPU
3248 case Op_AddF:
3249 case Op_SubF:
3250 case Op_MulF:
3251 case Op_DivF:
3252 case Op_NegF:
3253 case Op_ModF:
3254 case Op_ConvI2F:
3255 case Op_ConF:
3256 case Op_CmpF:
3257 case Op_CmpF3:
3258 // case Op_ConvL2F: // longs are split into 32-bit halves
3259 frc.inc_float_count();
3260 break;
3261
3262 case Op_ConvF2D:
3263 case Op_ConvD2F:
3264 frc.inc_float_count();
3265 frc.inc_double_count();
3266 break;
3267
3268 // Count all double operations that may use FPU
3269 case Op_AddD:
3270 case Op_SubD:
3271 case Op_MulD:
3272 case Op_DivD:
3273 case Op_NegD:
3274 case Op_ModD:
3275 case Op_ConvI2D:
3276 case Op_ConvD2I:
3277 // case Op_ConvL2D: // handled by leaf call
3278 // case Op_ConvD2L: // handled by leaf call
3279 case Op_ConD:
3280 case Op_CmpD:
3281 case Op_CmpD3:
3282 frc.inc_double_count();
3283 break;
3284 case Op_Opaque1: // Remove Opaque Nodes before matching
3285 case Op_Opaque2: // Remove Opaque Nodes before matching
3286 case Op_Opaque3:
3287 n->subsume_by(n->in(1), this);
3288 break;
3289 case Op_CallStaticJava:
3290 case Op_CallJava:
3291 case Op_CallDynamicJava:
3292 frc.inc_java_call_count(); // Count java call site;
3293 case Op_CallRuntime:
3294 case Op_CallLeaf:
3295 case Op_CallLeafNoFP: {
3296 assert (n->is_Call(), "");
3297 CallNode *call = n->as_Call();
3298 // Count call sites where the FP mode bit would have to be flipped.
3299 // Do not count uncommon runtime calls:
3300 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3301 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3302 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3303 frc.inc_call_count(); // Count the call site
3304 } else { // See if uncommon argument is shared
3305 Node *n = call->in(TypeFunc::Parms);
3306 int nop = n->Opcode();
3307 // Clone shared simple arguments to uncommon calls, item (1).
3308 if (n->outcnt() > 1 &&
3309 !n->is_Proj() &&
3310 nop != Op_CreateEx &&
3311 nop != Op_CheckCastPP &&
3312 nop != Op_DecodeN &&
3313 nop != Op_DecodeNKlass &&
3314 !n->is_Mem() &&
3315 !n->is_Phi()) {
3316 Node *x = n->clone();
3317 call->set_req(TypeFunc::Parms, x);
3318 }
3319 }
3320 break;
3321 }
3322
3323 case Op_StoreD:
3324 case Op_LoadD:
3325 case Op_LoadD_unaligned:
3326 frc.inc_double_count();
3327 goto handle_mem;
3328 case Op_StoreF:
3329 case Op_LoadF:
3330 frc.inc_float_count();
3331 goto handle_mem;
3332
3333 case Op_StoreCM:
3334 {
3335 // Convert OopStore dependence into precedence edge
3336 Node* prec = n->in(MemNode::OopStore);
3337 n->del_req(MemNode::OopStore);
3338 n->add_prec(prec);
3339 eliminate_redundant_card_marks(n);
3340 }
3341
3342 // fall through
3343
3344 case Op_StoreB:
3345 case Op_StoreC:
3346 case Op_StorePConditional:
3347 case Op_StoreI:
3348 case Op_StoreL:
3349 case Op_StoreIConditional:
3350 case Op_StoreLConditional:
3351 case Op_CompareAndSwapB:
3352 case Op_CompareAndSwapS:
3353 case Op_CompareAndSwapI:
3354 case Op_CompareAndSwapL:
3355 case Op_CompareAndSwapP:
3356 case Op_CompareAndSwapN:
3357 case Op_WeakCompareAndSwapB:
3358 case Op_WeakCompareAndSwapS:
3359 case Op_WeakCompareAndSwapI:
3360 case Op_WeakCompareAndSwapL:
3361 case Op_WeakCompareAndSwapP:
3362 case Op_WeakCompareAndSwapN:
3363 case Op_CompareAndExchangeB:
3364 case Op_CompareAndExchangeS:
3365 case Op_CompareAndExchangeI:
3366 case Op_CompareAndExchangeL:
3367 case Op_CompareAndExchangeP:
3368 case Op_CompareAndExchangeN:
3369 case Op_GetAndAddS:
3370 case Op_GetAndAddB:
3371 case Op_GetAndAddI:
3372 case Op_GetAndAddL:
3373 case Op_GetAndSetS:
3374 case Op_GetAndSetB:
3375 case Op_GetAndSetI:
3376 case Op_GetAndSetL:
3377 case Op_GetAndSetP:
3378 case Op_GetAndSetN:
3379 case Op_StoreP:
3380 case Op_StoreN:
3381 case Op_StoreNKlass:
3382 case Op_LoadB:
3383 case Op_LoadUB:
3384 case Op_LoadUS:
3385 case Op_LoadI:
3386 case Op_LoadKlass:
3387 case Op_LoadNKlass:
3388 case Op_LoadL:
3389 case Op_LoadL_unaligned:
3390 case Op_LoadPLocked:
3391 case Op_LoadP:
3392 case Op_LoadN:
3393 case Op_LoadRange:
3394 case Op_LoadS: {
3395 handle_mem:
3396 #ifdef ASSERT
3397 if( VerifyOptoOopOffsets ) {
3398 MemNode* mem = n->as_Mem();
3399 // Check to see if address types have grounded out somehow.
3400 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3401 assert( !tp || oop_offset_is_sane(tp), "" );
3402 }
3403 #endif
3404 if (nop == Op_LoadKlass || nop == Op_LoadNKlass) {
3405 const TypeKlassPtr* tk = n->bottom_type()->make_ptr()->is_klassptr();
3406 assert(!tk->klass_is_exact(), "should have been folded");
3407 if (tk->klass()->is_obj_array_klass() || tk->klass()->is_java_lang_Object()) {
3408 bool maybe_value_array = tk->klass()->is_java_lang_Object();
3409 if (!maybe_value_array) {
3410 ciArrayKlass* ak = tk->klass()->as_array_klass();
3411 ciKlass* elem = ak->element_klass();
3412 maybe_value_array = elem->is_java_lang_Object() || elem->is_interface() || elem->is_valuetype();
3413 }
3414 if (maybe_value_array) {
3415 // Array load klass needs to filter out property bits (but not
3416 // GetNullFreePropertyNode which needs to extract the null free
3417 // bits)
3418 uint last = unique();
3419 Node* pointer = NULL;
3420 if (nop == Op_LoadKlass) {
3421 Node* cast = new CastP2XNode(NULL, n);
3422 Node* masked = new LShiftXNode(cast, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3423 masked = new RShiftXNode(masked, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits)));
3424 pointer = new CastX2PNode(masked);
3425 pointer = new CheckCastPPNode(NULL, pointer, n->bottom_type());
3426 } else {
3427 Node* cast = new CastN2INode(n);
3428 Node* masked = new AndINode(cast, new ConINode(TypeInt::make(oopDesc::compressed_klass_mask())));
3429 pointer = new CastI2NNode(masked, n->bottom_type());
3430 }
3431 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3432 Node* u = n->fast_out(i);
3433 if (u->_idx < last && u->Opcode() != Op_GetNullFreeProperty) {
3434 int nb = u->replace_edge(n, pointer);
3435 --i, imax -= nb;
3436 }
3437 }
3438 }
3439 }
3440 }
3441 break;
3442 }
3443
3444 case Op_AddP: { // Assert sane base pointers
3445 Node *addp = n->in(AddPNode::Address);
3446 assert( !addp->is_AddP() ||
3447 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3448 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3449 "Base pointers must match (addp %u)", addp->_idx );
3450 #ifdef _LP64
3451 if ((UseCompressedOops || UseCompressedClassPointers) &&
3452 addp->Opcode() == Op_ConP &&
3453 addp == n->in(AddPNode::Base) &&
3454 n->in(AddPNode::Offset)->is_Con()) {
3455 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3456 // on the platform and on the compressed oops mode.
3457 // Use addressing with narrow klass to load with offset on x86.
3458 // Some platforms can use the constant pool to load ConP.
3459 // Do this transformation here since IGVN will convert ConN back to ConP.
3460 const Type* t = addp->bottom_type();
3461 bool is_oop = t->isa_oopptr() != NULL;
3462 bool is_klass = t->isa_klassptr() != NULL;
3463
3464 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3465 (is_klass && Matcher::const_klass_prefer_decode())) {
3466 Node* nn = NULL;
3467
3468 int op = is_oop ? Op_ConN : Op_ConNKlass;
3469
3470 // Look for existing ConN node of the same exact type.
3471 Node* r = root();
3472 uint cnt = r->outcnt();
3473 for (uint i = 0; i < cnt; i++) {
3474 Node* m = r->raw_out(i);
3475 if (m!= NULL && m->Opcode() == op &&
3476 m->bottom_type()->make_ptr() == t) {
3477 nn = m;
3478 break;
3479 }
3480 }
3481 if (nn != NULL) {
3482 // Decode a narrow oop to match address
3483 // [R12 + narrow_oop_reg<<3 + offset]
3484 if (is_oop) {
3485 nn = new DecodeNNode(nn, t);
3486 } else {
3487 nn = new DecodeNKlassNode(nn, t);
3488 }
3489 // Check for succeeding AddP which uses the same Base.
3490 // Otherwise we will run into the assertion above when visiting that guy.
3491 for (uint i = 0; i < n->outcnt(); ++i) {
3492 Node *out_i = n->raw_out(i);
3493 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3494 out_i->set_req(AddPNode::Base, nn);
3495 #ifdef ASSERT
3496 for (uint j = 0; j < out_i->outcnt(); ++j) {
3497 Node *out_j = out_i->raw_out(j);
3498 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3499 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3500 }
3501 #endif
3502 }
3503 }
3504 n->set_req(AddPNode::Base, nn);
3505 n->set_req(AddPNode::Address, nn);
3506 if (addp->outcnt() == 0) {
3507 addp->disconnect_inputs(NULL, this);
3508 }
3509 }
3510 }
3511 }
3512 #endif
3513 // platform dependent reshaping of the address expression
3514 reshape_address(n->as_AddP());
3515 break;
3516 }
3517
3518 case Op_CastPP: {
3519 // Remove CastPP nodes to gain more freedom during scheduling but
3520 // keep the dependency they encode as control or precedence edges
3521 // (if control is set already) on memory operations. Some CastPP
3522 // nodes don't have a control (don't carry a dependency): skip
3523 // those.
3524 if (n->in(0) != NULL) {
3525 ResourceMark rm;
3526 Unique_Node_List wq;
3527 wq.push(n);
3528 for (uint next = 0; next < wq.size(); ++next) {
3529 Node *m = wq.at(next);
3530 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3531 Node* use = m->fast_out(i);
3532 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3533 use->ensure_control_or_add_prec(n->in(0));
3534 } else {
3535 switch(use->Opcode()) {
3536 case Op_AddP:
3537 case Op_DecodeN:
3538 case Op_DecodeNKlass:
3539 case Op_CheckCastPP:
3540 case Op_CastPP:
3541 wq.push(use);
3542 break;
3543 }
3544 }
3545 }
3546 }
3547 }
3548 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3549 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3550 Node* in1 = n->in(1);
3551 const Type* t = n->bottom_type();
3552 Node* new_in1 = in1->clone();
3553 new_in1->as_DecodeN()->set_type(t);
3554
3555 if (!Matcher::narrow_oop_use_complex_address()) {
3556 //
3557 // x86, ARM and friends can handle 2 adds in addressing mode
3558 // and Matcher can fold a DecodeN node into address by using
3559 // a narrow oop directly and do implicit NULL check in address:
3560 //
3561 // [R12 + narrow_oop_reg<<3 + offset]
3562 // NullCheck narrow_oop_reg
3563 //
3564 // On other platforms (Sparc) we have to keep new DecodeN node and
3565 // use it to do implicit NULL check in address:
3566 //
3567 // decode_not_null narrow_oop_reg, base_reg
3568 // [base_reg + offset]
3569 // NullCheck base_reg
3570 //
3571 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3572 // to keep the information to which NULL check the new DecodeN node
3573 // corresponds to use it as value in implicit_null_check().
3574 //
3575 new_in1->set_req(0, n->in(0));
3576 }
3577
3578 n->subsume_by(new_in1, this);
3579 if (in1->outcnt() == 0) {
3580 in1->disconnect_inputs(NULL, this);
3581 }
3582 } else {
3583 n->subsume_by(n->in(1), this);
3584 if (n->outcnt() == 0) {
3585 n->disconnect_inputs(NULL, this);
3586 }
3587 }
3588 break;
3589 }
3590 #ifdef _LP64
3591 case Op_CmpP:
3592 // Do this transformation here to preserve CmpPNode::sub() and
3593 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3594 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3595 Node* in1 = n->in(1);
3596 Node* in2 = n->in(2);
3597 if (!in1->is_DecodeNarrowPtr()) {
3598 in2 = in1;
3599 in1 = n->in(2);
3600 }
3601 assert(in1->is_DecodeNarrowPtr(), "sanity");
3602
3603 Node* new_in2 = NULL;
3604 if (in2->is_DecodeNarrowPtr()) {
3605 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3606 new_in2 = in2->in(1);
3607 } else if (in2->Opcode() == Op_ConP) {
3608 const Type* t = in2->bottom_type();
3609 if (t == TypePtr::NULL_PTR) {
3610 assert(in1->is_DecodeN(), "compare klass to null?");
3611 // Don't convert CmpP null check into CmpN if compressed
3612 // oops implicit null check is not generated.
3613 // This will allow to generate normal oop implicit null check.
3614 if (Matcher::gen_narrow_oop_implicit_null_checks())
3615 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3616 //
3617 // This transformation together with CastPP transformation above
3618 // will generated code for implicit NULL checks for compressed oops.
3619 //
3620 // The original code after Optimize()
3621 //
3622 // LoadN memory, narrow_oop_reg
3623 // decode narrow_oop_reg, base_reg
3624 // CmpP base_reg, NULL
3625 // CastPP base_reg // NotNull
3626 // Load [base_reg + offset], val_reg
3627 //
3628 // after these transformations will be
3629 //
3630 // LoadN memory, narrow_oop_reg
3631 // CmpN narrow_oop_reg, NULL
3632 // decode_not_null narrow_oop_reg, base_reg
3633 // Load [base_reg + offset], val_reg
3634 //
3635 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3636 // since narrow oops can be used in debug info now (see the code in
3637 // final_graph_reshaping_walk()).
3638 //
3639 // At the end the code will be matched to
3640 // on x86:
3641 //
3642 // Load_narrow_oop memory, narrow_oop_reg
3643 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3644 // NullCheck narrow_oop_reg
3645 //
3646 // and on sparc:
3647 //
3648 // Load_narrow_oop memory, narrow_oop_reg
3649 // decode_not_null narrow_oop_reg, base_reg
3650 // Load [base_reg + offset], val_reg
3651 // NullCheck base_reg
3652 //
3653 } else if (t->isa_oopptr()) {
3654 new_in2 = ConNode::make(t->make_narrowoop());
3655 } else if (t->isa_klassptr()) {
3656 new_in2 = ConNode::make(t->make_narrowklass());
3657 }
3658 }
3659 if (new_in2 != NULL) {
3660 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3661 n->subsume_by(cmpN, this);
3662 if (in1->outcnt() == 0) {
3663 in1->disconnect_inputs(NULL, this);
3664 }
3665 if (in2->outcnt() == 0) {
3666 in2->disconnect_inputs(NULL, this);
3667 }
3668 }
3669 }
3670 break;
3671
3672 case Op_DecodeN:
3673 case Op_DecodeNKlass:
3674 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3675 // DecodeN could be pinned when it can't be fold into
3676 // an address expression, see the code for Op_CastPP above.
3677 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3678 break;
3679
3680 case Op_EncodeP:
3681 case Op_EncodePKlass: {
3682 Node* in1 = n->in(1);
3683 if (in1->is_DecodeNarrowPtr()) {
3684 n->subsume_by(in1->in(1), this);
3685 } else if (in1->Opcode() == Op_ConP) {
3686 const Type* t = in1->bottom_type();
3687 if (t == TypePtr::NULL_PTR) {
3688 assert(t->isa_oopptr(), "null klass?");
3689 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3690 } else if (t->isa_oopptr()) {
3691 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3692 } else if (t->isa_klassptr()) {
3693 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3694 }
3695 }
3696 if (in1->outcnt() == 0) {
3697 in1->disconnect_inputs(NULL, this);
3698 }
3699 break;
3700 }
3701
3702 case Op_Proj: {
3703 if (OptimizeStringConcat) {
3704 ProjNode* p = n->as_Proj();
3705 if (p->_is_io_use) {
3706 // Separate projections were used for the exception path which
3707 // are normally removed by a late inline. If it wasn't inlined
3708 // then they will hang around and should just be replaced with
3709 // the original one.
3710 Node* proj = NULL;
3711 // Replace with just one
3712 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3713 Node *use = i.get();
3714 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3715 proj = use;
3716 break;
3717 }
3718 }
3719 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3720 if (proj != NULL) {
3721 p->subsume_by(proj, this);
3722 }
3723 }
3724 }
3725 break;
3726 }
3727
3728 case Op_Phi:
3729 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3730 // The EncodeP optimization may create Phi with the same edges
3731 // for all paths. It is not handled well by Register Allocator.
3732 Node* unique_in = n->in(1);
3733 assert(unique_in != NULL, "");
3734 uint cnt = n->req();
3735 for (uint i = 2; i < cnt; i++) {
3736 Node* m = n->in(i);
3737 assert(m != NULL, "");
3738 if (unique_in != m)
3739 unique_in = NULL;
3740 }
3741 if (unique_in != NULL) {
3742 n->subsume_by(unique_in, this);
3743 }
3744 }
3745 break;
3746
3747 #endif
3748
3749 #ifdef ASSERT
3750 case Op_CastII:
3751 // Verify that all range check dependent CastII nodes were removed.
3752 if (n->isa_CastII()->has_range_check()) {
3753 n->dump(3);
3754 assert(false, "Range check dependent CastII node was not removed");
3755 }
3756 break;
3757 #endif
3758
3759 case Op_ModI:
3760 if (UseDivMod) {
3761 // Check if a%b and a/b both exist
3762 Node* d = n->find_similar(Op_DivI);
3763 if (d) {
3764 // Replace them with a fused divmod if supported
3765 if (Matcher::has_match_rule(Op_DivModI)) {
3766 DivModINode* divmod = DivModINode::make(n);
3767 d->subsume_by(divmod->div_proj(), this);
3768 n->subsume_by(divmod->mod_proj(), this);
3769 } else {
3770 // replace a%b with a-((a/b)*b)
3771 Node* mult = new MulINode(d, d->in(2));
3772 Node* sub = new SubINode(d->in(1), mult);
3773 n->subsume_by(sub, this);
3774 }
3775 }
3776 }
3777 break;
3778
3779 case Op_ModL:
3780 if (UseDivMod) {
3781 // Check if a%b and a/b both exist
3782 Node* d = n->find_similar(Op_DivL);
3783 if (d) {
3784 // Replace them with a fused divmod if supported
3785 if (Matcher::has_match_rule(Op_DivModL)) {
3786 DivModLNode* divmod = DivModLNode::make(n);
3787 d->subsume_by(divmod->div_proj(), this);
3788 n->subsume_by(divmod->mod_proj(), this);
3789 } else {
3790 // replace a%b with a-((a/b)*b)
3791 Node* mult = new MulLNode(d, d->in(2));
3792 Node* sub = new SubLNode(d->in(1), mult);
3793 n->subsume_by(sub, this);
3794 }
3795 }
3796 }
3797 break;
3798
3799 case Op_LoadVector:
3800 case Op_StoreVector:
3801 break;
3802
3803 case Op_AddReductionVI:
3804 case Op_AddReductionVL:
3805 case Op_AddReductionVF:
3806 case Op_AddReductionVD:
3807 case Op_MulReductionVI:
3808 case Op_MulReductionVL:
3809 case Op_MulReductionVF:
3810 case Op_MulReductionVD:
3811 case Op_MinReductionV:
3812 case Op_MaxReductionV:
3813 break;
3814
3815 case Op_PackB:
3816 case Op_PackS:
3817 case Op_PackI:
3818 case Op_PackF:
3819 case Op_PackL:
3820 case Op_PackD:
3821 if (n->req()-1 > 2) {
3822 // Replace many operand PackNodes with a binary tree for matching
3823 PackNode* p = (PackNode*) n;
3824 Node* btp = p->binary_tree_pack(1, n->req());
3825 n->subsume_by(btp, this);
3826 }
3827 break;
3828 case Op_Loop:
3829 case Op_CountedLoop:
3830 case Op_OuterStripMinedLoop:
3831 if (n->as_Loop()->is_inner_loop()) {
3832 frc.inc_inner_loop_count();
3833 }
3834 n->as_Loop()->verify_strip_mined(0);
3835 break;
3836 case Op_LShiftI:
3837 case Op_RShiftI:
3838 case Op_URShiftI:
3839 case Op_LShiftL:
3840 case Op_RShiftL:
3841 case Op_URShiftL:
3842 if (Matcher::need_masked_shift_count) {
3843 // The cpu's shift instructions don't restrict the count to the
3844 // lower 5/6 bits. We need to do the masking ourselves.
3845 Node* in2 = n->in(2);
3846 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3847 const TypeInt* t = in2->find_int_type();
3848 if (t != NULL && t->is_con()) {
3849 juint shift = t->get_con();
3850 if (shift > mask) { // Unsigned cmp
3851 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3852 }
3853 } else {
3854 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3855 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3856 n->set_req(2, shift);
3857 }
3858 }
3859 if (in2->outcnt() == 0) { // Remove dead node
3860 in2->disconnect_inputs(NULL, this);
3861 }
3862 }
3863 break;
3864 case Op_MemBarStoreStore:
3865 case Op_MemBarRelease:
3866 // Break the link with AllocateNode: it is no longer useful and
3867 // confuses register allocation.
3868 if (n->req() > MemBarNode::Precedent) {
3869 n->set_req(MemBarNode::Precedent, top());
3870 }
3871 break;
3872 case Op_MemBarAcquire: {
3873 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3874 // At parse time, the trailing MemBarAcquire for a volatile load
3875 // is created with an edge to the load. After optimizations,
3876 // that input may be a chain of Phis. If those phis have no
3877 // other use, then the MemBarAcquire keeps them alive and
3878 // register allocation can be confused.
3879 ResourceMark rm;
3880 Unique_Node_List wq;
3881 wq.push(n->in(MemBarNode::Precedent));
3882 n->set_req(MemBarNode::Precedent, top());
3883 while (wq.size() > 0) {
3884 Node* m = wq.pop();
3885 if (m->outcnt() == 0) {
3886 for (uint j = 0; j < m->req(); j++) {
3887 Node* in = m->in(j);
3888 if (in != NULL) {
3889 wq.push(in);
3890 }
3891 }
3892 m->disconnect_inputs(NULL, this);
3893 }
3894 }
3895 }
3896 break;
3897 }
3898 case Op_RangeCheck: {
3899 RangeCheckNode* rc = n->as_RangeCheck();
3900 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3901 n->subsume_by(iff, this);
3902 frc._tests.push(iff);
3903 break;
3904 }
3905 case Op_ConvI2L: {
3906 if (!Matcher::convi2l_type_required) {
3907 // Code generation on some platforms doesn't need accurate
3908 // ConvI2L types. Widening the type can help remove redundant
3909 // address computations.
3910 n->as_Type()->set_type(TypeLong::INT);
3911 ResourceMark rm;
3912 Node_List wq;
3913 wq.push(n);
3914 for (uint next = 0; next < wq.size(); next++) {
3915 Node *m = wq.at(next);
3916
3917 for(;;) {
3918 // Loop over all nodes with identical inputs edges as m
3919 Node* k = m->find_similar(m->Opcode());
3920 if (k == NULL) {
3921 break;
3922 }
3923 // Push their uses so we get a chance to remove node made
3924 // redundant
3925 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3926 Node* u = k->fast_out(i);
3927 assert(!wq.contains(u), "shouldn't process one node several times");
3928 if (u->Opcode() == Op_LShiftL ||
3929 u->Opcode() == Op_AddL ||
3930 u->Opcode() == Op_SubL ||
3931 u->Opcode() == Op_AddP) {
3932 wq.push(u);
3933 }
3934 }
3935 // Replace all nodes with identical edges as m with m
3936 k->subsume_by(m, this);
3937 }
3938 }
3939 }
3940 break;
3941 }
3942 case Op_CmpUL: {
3943 if (!Matcher::has_match_rule(Op_CmpUL)) {
3944 // No support for unsigned long comparisons
3945 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3946 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3947 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3948 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3949 Node* andl = new AndLNode(orl, remove_sign_mask);
3950 Node* cmp = new CmpLNode(andl, n->in(2));
3951 n->subsume_by(cmp, this);
3952 }
3953 break;
3954 }
3955 #ifdef ASSERT
3956 case Op_ValueTypePtr:
3957 case Op_ValueType: {
3958 n->dump(-1);
3959 assert(false, "value type node was not removed");
3960 break;
3961 }
3962 #endif
3963 case Op_GetNullFreeProperty: {
3964 // Extract the null free bits
3965 uint last = unique();
3966 Node* null_free = NULL;
3967 if (n->in(1)->Opcode() == Op_LoadKlass) {
3968 Node* cast = new CastP2XNode(NULL, n->in(1));
3969 null_free = new AndLNode(cast, new ConLNode(TypeLong::make(((jlong)1)<<(oopDesc::wide_storage_props_shift + ArrayStorageProperties::null_free_bit))));
3970 } else {
3971 assert(n->in(1)->Opcode() == Op_LoadNKlass, "not a compressed klass?");
3972 Node* cast = new CastN2INode(n->in(1));
3973 null_free = new AndINode(cast, new ConINode(TypeInt::make(1<<(oopDesc::narrow_storage_props_shift + ArrayStorageProperties::null_free_bit))));
3974 }
3975 n->replace_by(null_free);
3976 break;
3977 }
3978 default:
3979 assert(!n->is_Call(), "");
3980 assert(!n->is_Mem(), "");
3981 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3982 break;
3983 }
3984 }
3985
3986 //------------------------------final_graph_reshaping_walk---------------------
3987 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3988 // requires that the walk visits a node's inputs before visiting the node.
3989 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3990 ResourceArea *area = Thread::current()->resource_area();
3991 Unique_Node_List sfpt(area);
3992
3993 frc._visited.set(root->_idx); // first, mark node as visited
3994 uint cnt = root->req();
3995 Node *n = root;
3996 uint i = 0;
3997 while (true) {
3998 if (i < cnt) {
3999 // Place all non-visited non-null inputs onto stack
4000 Node* m = n->in(i);
4001 ++i;
4002 if (m != NULL && !frc._visited.test_set(m->_idx)) {
4003 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
4004 // compute worst case interpreter size in case of a deoptimization
4005 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
4006
4007 sfpt.push(m);
4008 }
4009 cnt = m->req();
4010 nstack.push(n, i); // put on stack parent and next input's index
4011 n = m;
4012 i = 0;
4013 }
4014 } else {
4015 // Now do post-visit work
4016 final_graph_reshaping_impl( n, frc );
4017 if (nstack.is_empty())
4018 break; // finished
4019 n = nstack.node(); // Get node from stack
4020 cnt = n->req();
4021 i = nstack.index();
4022 nstack.pop(); // Shift to the next node on stack
4023 }
4024 }
4025
4026 // Skip next transformation if compressed oops are not used.
4027 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
4028 (!UseCompressedOops && !UseCompressedClassPointers))
4029 return;
4030
4031 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
4032 // It could be done for an uncommon traps or any safepoints/calls
4033 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
4034 while (sfpt.size() > 0) {
4035 n = sfpt.pop();
4036 JVMState *jvms = n->as_SafePoint()->jvms();
4037 assert(jvms != NULL, "sanity");
4038 int start = jvms->debug_start();
4039 int end = n->req();
4040 bool is_uncommon = (n->is_CallStaticJava() &&
4041 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4042 for (int j = start; j < end; j++) {
4043 Node* in = n->in(j);
4044 if (in->is_DecodeNarrowPtr()) {
4045 bool safe_to_skip = true;
4046 if (!is_uncommon ) {
4047 // Is it safe to skip?
4048 for (uint i = 0; i < in->outcnt(); i++) {
4049 Node* u = in->raw_out(i);
4050 if (!u->is_SafePoint() ||
4051 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4052 safe_to_skip = false;
4053 }
4054 }
4055 }
4056 if (safe_to_skip) {
4057 n->set_req(j, in->in(1));
4058 }
4059 if (in->outcnt() == 0) {
4060 in->disconnect_inputs(NULL, this);
4061 }
4062 }
4063 }
4064 }
4065 }
4066
4067 //------------------------------final_graph_reshaping--------------------------
4068 // Final Graph Reshaping.
4069 //
4070 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4071 // and not commoned up and forced early. Must come after regular
4072 // optimizations to avoid GVN undoing the cloning. Clone constant
4073 // inputs to Loop Phis; these will be split by the allocator anyways.
4074 // Remove Opaque nodes.
4075 // (2) Move last-uses by commutative operations to the left input to encourage
4076 // Intel update-in-place two-address operations and better register usage
4077 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4078 // calls canonicalizing them back.
4079 // (3) Count the number of double-precision FP ops, single-precision FP ops
4080 // and call sites. On Intel, we can get correct rounding either by
4081 // forcing singles to memory (requires extra stores and loads after each
4082 // FP bytecode) or we can set a rounding mode bit (requires setting and
4083 // clearing the mode bit around call sites). The mode bit is only used
4084 // if the relative frequency of single FP ops to calls is low enough.
4085 // This is a key transform for SPEC mpeg_audio.
4086 // (4) Detect infinite loops; blobs of code reachable from above but not
4087 // below. Several of the Code_Gen algorithms fail on such code shapes,
4088 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4089 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4090 // Detection is by looking for IfNodes where only 1 projection is
4091 // reachable from below or CatchNodes missing some targets.
4092 // (5) Assert for insane oop offsets in debug mode.
4093
4094 bool Compile::final_graph_reshaping() {
4095 // an infinite loop may have been eliminated by the optimizer,
4096 // in which case the graph will be empty.
4097 if (root()->req() == 1) {
4098 record_method_not_compilable("trivial infinite loop");
4099 return true;
4100 }
4101
4102 // Expensive nodes have their control input set to prevent the GVN
4103 // from freely commoning them. There's no GVN beyond this point so
4104 // no need to keep the control input. We want the expensive nodes to
4105 // be freely moved to the least frequent code path by gcm.
4106 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4107 for (int i = 0; i < expensive_count(); i++) {
4108 _expensive_nodes->at(i)->set_req(0, NULL);
4109 }
4110
4111 Final_Reshape_Counts frc;
4112
4113 // Visit everybody reachable!
4114 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4115 Node_Stack nstack(live_nodes() >> 1);
4116 final_graph_reshaping_walk(nstack, root(), frc);
4117
4118 // Check for unreachable (from below) code (i.e., infinite loops).
4119 for( uint i = 0; i < frc._tests.size(); i++ ) {
4120 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4121 // Get number of CFG targets.
4122 // Note that PCTables include exception targets after calls.
4123 uint required_outcnt = n->required_outcnt();
4124 if (n->outcnt() != required_outcnt) {
4125 // Check for a few special cases. Rethrow Nodes never take the
4126 // 'fall-thru' path, so expected kids is 1 less.
4127 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4128 if (n->in(0)->in(0)->is_Call()) {
4129 CallNode *call = n->in(0)->in(0)->as_Call();
4130 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4131 required_outcnt--; // Rethrow always has 1 less kid
4132 } else if (call->req() > TypeFunc::Parms &&
4133 call->is_CallDynamicJava()) {
4134 // Check for null receiver. In such case, the optimizer has
4135 // detected that the virtual call will always result in a null
4136 // pointer exception. The fall-through projection of this CatchNode
4137 // will not be populated.
4138 Node *arg0 = call->in(TypeFunc::Parms);
4139 if (arg0->is_Type() &&
4140 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4141 required_outcnt--;
4142 }
4143 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4144 call->req() > TypeFunc::Parms+1 &&
4145 call->is_CallStaticJava()) {
4146 // Check for negative array length. In such case, the optimizer has
4147 // detected that the allocation attempt will always result in an
4148 // exception. There is no fall-through projection of this CatchNode .
4149 Node *arg1 = call->in(TypeFunc::Parms+1);
4150 if (arg1->is_Type() &&
4151 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4152 required_outcnt--;
4153 }
4154 }
4155 }
4156 }
4157 // Recheck with a better notion of 'required_outcnt'
4158 if (n->outcnt() != required_outcnt) {
4159 record_method_not_compilable("malformed control flow");
4160 return true; // Not all targets reachable!
4161 }
4162 }
4163 // Check that I actually visited all kids. Unreached kids
4164 // must be infinite loops.
4165 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4166 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4167 record_method_not_compilable("infinite loop");
4168 return true; // Found unvisited kid; must be unreach
4169 }
4170
4171 // Here so verification code in final_graph_reshaping_walk()
4172 // always see an OuterStripMinedLoopEnd
4173 if (n->is_OuterStripMinedLoopEnd()) {
4174 IfNode* init_iff = n->as_If();
4175 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4176 n->subsume_by(iff, this);
4177 }
4178 }
4179
4180 // If original bytecodes contained a mixture of floats and doubles
4181 // check if the optimizer has made it homogenous, item (3).
4182 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
4183 frc.get_float_count() > 32 &&
4184 frc.get_double_count() == 0 &&
4185 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4186 set_24_bit_selection_and_mode( false, true );
4187 }
4188
4189 set_java_calls(frc.get_java_call_count());
4190 set_inner_loops(frc.get_inner_loop_count());
4191
4192 // No infinite loops, no reason to bail out.
4193 return false;
4194 }
4195
4196 //-----------------------------too_many_traps----------------------------------
4197 // Report if there are too many traps at the current method and bci.
4198 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4199 bool Compile::too_many_traps(ciMethod* method,
4200 int bci,
4201 Deoptimization::DeoptReason reason) {
4202 ciMethodData* md = method->method_data();
4203 if (md->is_empty()) {
4204 // Assume the trap has not occurred, or that it occurred only
4205 // because of a transient condition during start-up in the interpreter.
4206 return false;
4207 }
4208 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4209 if (md->has_trap_at(bci, m, reason) != 0) {
4210 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4211 // Also, if there are multiple reasons, or if there is no per-BCI record,
4212 // assume the worst.
4213 if (log())
4214 log()->elem("observe trap='%s' count='%d'",
4215 Deoptimization::trap_reason_name(reason),
4216 md->trap_count(reason));
4217 return true;
4218 } else {
4219 // Ignore method/bci and see if there have been too many globally.
4220 return too_many_traps(reason, md);
4221 }
4222 }
4223
4224 // Less-accurate variant which does not require a method and bci.
4225 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4226 ciMethodData* logmd) {
4227 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4228 // Too many traps globally.
4229 // Note that we use cumulative trap_count, not just md->trap_count.
4230 if (log()) {
4231 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4232 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4233 Deoptimization::trap_reason_name(reason),
4234 mcount, trap_count(reason));
4235 }
4236 return true;
4237 } else {
4238 // The coast is clear.
4239 return false;
4240 }
4241 }
4242
4243 //--------------------------too_many_recompiles--------------------------------
4244 // Report if there are too many recompiles at the current method and bci.
4245 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4246 // Is not eager to return true, since this will cause the compiler to use
4247 // Action_none for a trap point, to avoid too many recompilations.
4248 bool Compile::too_many_recompiles(ciMethod* method,
4249 int bci,
4250 Deoptimization::DeoptReason reason) {
4251 ciMethodData* md = method->method_data();
4252 if (md->is_empty()) {
4253 // Assume the trap has not occurred, or that it occurred only
4254 // because of a transient condition during start-up in the interpreter.
4255 return false;
4256 }
4257 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4258 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4259 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4260 Deoptimization::DeoptReason per_bc_reason
4261 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4262 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4263 if ((per_bc_reason == Deoptimization::Reason_none
4264 || md->has_trap_at(bci, m, reason) != 0)
4265 // The trap frequency measure we care about is the recompile count:
4266 && md->trap_recompiled_at(bci, m)
4267 && md->overflow_recompile_count() >= bc_cutoff) {
4268 // Do not emit a trap here if it has already caused recompilations.
4269 // Also, if there are multiple reasons, or if there is no per-BCI record,
4270 // assume the worst.
4271 if (log())
4272 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4273 Deoptimization::trap_reason_name(reason),
4274 md->trap_count(reason),
4275 md->overflow_recompile_count());
4276 return true;
4277 } else if (trap_count(reason) != 0
4278 && decompile_count() >= m_cutoff) {
4279 // Too many recompiles globally, and we have seen this sort of trap.
4280 // Use cumulative decompile_count, not just md->decompile_count.
4281 if (log())
4282 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4283 Deoptimization::trap_reason_name(reason),
4284 md->trap_count(reason), trap_count(reason),
4285 md->decompile_count(), decompile_count());
4286 return true;
4287 } else {
4288 // The coast is clear.
4289 return false;
4290 }
4291 }
4292
4293 // Compute when not to trap. Used by matching trap based nodes and
4294 // NullCheck optimization.
4295 void Compile::set_allowed_deopt_reasons() {
4296 _allowed_reasons = 0;
4297 if (is_method_compilation()) {
4298 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4299 assert(rs < BitsPerInt, "recode bit map");
4300 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4301 _allowed_reasons |= nth_bit(rs);
4302 }
4303 }
4304 }
4305 }
4306
4307 bool Compile::is_compiling_clinit_for(ciKlass* k) {
4308 ciMethod* root = method(); // the root method of compilation
4309 return root->is_class_initializer() && root->holder() == k; // access in the context of clinit
4310 }
4311
4312 #ifndef PRODUCT
4313 //------------------------------verify_graph_edges---------------------------
4314 // Walk the Graph and verify that there is a one-to-one correspondence
4315 // between Use-Def edges and Def-Use edges in the graph.
4316 void Compile::verify_graph_edges(bool no_dead_code) {
4317 if (VerifyGraphEdges) {
4318 ResourceArea *area = Thread::current()->resource_area();
4319 Unique_Node_List visited(area);
4320 // Call recursive graph walk to check edges
4321 _root->verify_edges(visited);
4322 if (no_dead_code) {
4323 // Now make sure that no visited node is used by an unvisited node.
4324 bool dead_nodes = false;
4325 Unique_Node_List checked(area);
4326 while (visited.size() > 0) {
4327 Node* n = visited.pop();
4328 checked.push(n);
4329 for (uint i = 0; i < n->outcnt(); i++) {
4330 Node* use = n->raw_out(i);
4331 if (checked.member(use)) continue; // already checked
4332 if (visited.member(use)) continue; // already in the graph
4333 if (use->is_Con()) continue; // a dead ConNode is OK
4334 // At this point, we have found a dead node which is DU-reachable.
4335 if (!dead_nodes) {
4336 tty->print_cr("*** Dead nodes reachable via DU edges:");
4337 dead_nodes = true;
4338 }
4339 use->dump(2);
4340 tty->print_cr("---");
4341 checked.push(use); // No repeats; pretend it is now checked.
4342 }
4343 }
4344 assert(!dead_nodes, "using nodes must be reachable from root");
4345 }
4346 }
4347 }
4348 #endif
4349
4350 // The Compile object keeps track of failure reasons separately from the ciEnv.
4351 // This is required because there is not quite a 1-1 relation between the
4352 // ciEnv and its compilation task and the Compile object. Note that one
4353 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4354 // to backtrack and retry without subsuming loads. Other than this backtracking
4355 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4356 // by the logic in C2Compiler.
4357 void Compile::record_failure(const char* reason) {
4358 if (log() != NULL) {
4359 log()->elem("failure reason='%s' phase='compile'", reason);
4360 }
4361 if (_failure_reason == NULL) {
4362 // Record the first failure reason.
4363 _failure_reason = reason;
4364 }
4365
4366 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4367 C->print_method(PHASE_FAILURE);
4368 }
4369 _root = NULL; // flush the graph, too
4370 }
4371
4372 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4373 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4374 _phase_name(name), _dolog(CITimeVerbose)
4375 {
4376 if (_dolog) {
4377 C = Compile::current();
4378 _log = C->log();
4379 } else {
4380 C = NULL;
4381 _log = NULL;
4382 }
4383 if (_log != NULL) {
4384 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4385 _log->stamp();
4386 _log->end_head();
4387 }
4388 }
4389
4390 Compile::TracePhase::~TracePhase() {
4391
4392 C = Compile::current();
4393 if (_dolog) {
4394 _log = C->log();
4395 } else {
4396 _log = NULL;
4397 }
4398
4399 #ifdef ASSERT
4400 if (PrintIdealNodeCount) {
4401 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4402 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4403 }
4404
4405 if (VerifyIdealNodeCount) {
4406 Compile::current()->print_missing_nodes();
4407 }
4408 #endif
4409
4410 if (_log != NULL) {
4411 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4412 }
4413 }
4414
4415 //=============================================================================
4416 // Two Constant's are equal when the type and the value are equal.
4417 bool Compile::Constant::operator==(const Constant& other) {
4418 if (type() != other.type() ) return false;
4419 if (can_be_reused() != other.can_be_reused()) return false;
4420 // For floating point values we compare the bit pattern.
4421 switch (type()) {
4422 case T_INT:
4423 case T_FLOAT: return (_v._value.i == other._v._value.i);
4424 case T_LONG:
4425 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4426 case T_OBJECT:
4427 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4428 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4429 case T_METADATA: return (_v._metadata == other._v._metadata);
4430 default: ShouldNotReachHere(); return false;
4431 }
4432 }
4433
4434 static int type_to_size_in_bytes(BasicType t) {
4435 switch (t) {
4436 case T_INT: return sizeof(jint );
4437 case T_LONG: return sizeof(jlong );
4438 case T_FLOAT: return sizeof(jfloat );
4439 case T_DOUBLE: return sizeof(jdouble);
4440 case T_METADATA: return sizeof(Metadata*);
4441 // We use T_VOID as marker for jump-table entries (labels) which
4442 // need an internal word relocation.
4443 case T_VOID:
4444 case T_ADDRESS:
4445 case T_OBJECT: return sizeof(jobject);
4446 default:
4447 ShouldNotReachHere();
4448 return -1;
4449 }
4450 }
4451
4452 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4453 // sort descending
4454 if (a->freq() > b->freq()) return -1;
4455 if (a->freq() < b->freq()) return 1;
4456 return 0;
4457 }
4458
4459 void Compile::ConstantTable::calculate_offsets_and_size() {
4460 // First, sort the array by frequencies.
4461 _constants.sort(qsort_comparator);
4462
4463 #ifdef ASSERT
4464 // Make sure all jump-table entries were sorted to the end of the
4465 // array (they have a negative frequency).
4466 bool found_void = false;
4467 for (int i = 0; i < _constants.length(); i++) {
4468 Constant con = _constants.at(i);
4469 if (con.type() == T_VOID)
4470 found_void = true; // jump-tables
4471 else
4472 assert(!found_void, "wrong sorting");
4473 }
4474 #endif
4475
4476 int offset = 0;
4477 for (int i = 0; i < _constants.length(); i++) {
4478 Constant* con = _constants.adr_at(i);
4479
4480 // Align offset for type.
4481 int typesize = type_to_size_in_bytes(con->type());
4482 offset = align_up(offset, typesize);
4483 con->set_offset(offset); // set constant's offset
4484
4485 if (con->type() == T_VOID) {
4486 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4487 offset = offset + typesize * n->outcnt(); // expand jump-table
4488 } else {
4489 offset = offset + typesize;
4490 }
4491 }
4492
4493 // Align size up to the next section start (which is insts; see
4494 // CodeBuffer::align_at_start).
4495 assert(_size == -1, "already set?");
4496 _size = align_up(offset, (int)CodeEntryAlignment);
4497 }
4498
4499 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4500 MacroAssembler _masm(&cb);
4501 for (int i = 0; i < _constants.length(); i++) {
4502 Constant con = _constants.at(i);
4503 address constant_addr = NULL;
4504 switch (con.type()) {
4505 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4506 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4507 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4508 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4509 case T_OBJECT: {
4510 jobject obj = con.get_jobject();
4511 int oop_index = _masm.oop_recorder()->find_index(obj);
4512 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4513 break;
4514 }
4515 case T_ADDRESS: {
4516 address addr = (address) con.get_jobject();
4517 constant_addr = _masm.address_constant(addr);
4518 break;
4519 }
4520 // We use T_VOID as marker for jump-table entries (labels) which
4521 // need an internal word relocation.
4522 case T_VOID: {
4523 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4524 // Fill the jump-table with a dummy word. The real value is
4525 // filled in later in fill_jump_table.
4526 address dummy = (address) n;
4527 constant_addr = _masm.address_constant(dummy);
4528 // Expand jump-table
4529 for (uint i = 1; i < n->outcnt(); i++) {
4530 address temp_addr = _masm.address_constant(dummy + i);
4531 assert(temp_addr, "consts section too small");
4532 }
4533 break;
4534 }
4535 case T_METADATA: {
4536 Metadata* obj = con.get_metadata();
4537 int metadata_index = _masm.oop_recorder()->find_index(obj);
4538 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4539 break;
4540 }
4541 default: ShouldNotReachHere();
4542 }
4543 assert(constant_addr, "consts section too small");
4544 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4545 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4546 }
4547 }
4548
4549 int Compile::ConstantTable::find_offset(Constant& con) const {
4550 int idx = _constants.find(con);
4551 guarantee(idx != -1, "constant must be in constant table");
4552 int offset = _constants.at(idx).offset();
4553 guarantee(offset != -1, "constant table not emitted yet?");
4554 return offset;
4555 }
4556
4557 void Compile::ConstantTable::add(Constant& con) {
4558 if (con.can_be_reused()) {
4559 int idx = _constants.find(con);
4560 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4561 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4562 return;
4563 }
4564 }
4565 (void) _constants.append(con);
4566 }
4567
4568 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4569 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4570 Constant con(type, value, b->_freq);
4571 add(con);
4572 return con;
4573 }
4574
4575 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4576 Constant con(metadata);
4577 add(con);
4578 return con;
4579 }
4580
4581 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4582 jvalue value;
4583 BasicType type = oper->type()->basic_type();
4584 switch (type) {
4585 case T_LONG: value.j = oper->constantL(); break;
4586 case T_FLOAT: value.f = oper->constantF(); break;
4587 case T_DOUBLE: value.d = oper->constantD(); break;
4588 case T_OBJECT:
4589 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4590 case T_METADATA: return add((Metadata*)oper->constant()); break;
4591 default: guarantee(false, "unhandled type: %s", type2name(type));
4592 }
4593 return add(n, type, value);
4594 }
4595
4596 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4597 jvalue value;
4598 // We can use the node pointer here to identify the right jump-table
4599 // as this method is called from Compile::Fill_buffer right before
4600 // the MachNodes are emitted and the jump-table is filled (means the
4601 // MachNode pointers do not change anymore).
4602 value.l = (jobject) n;
4603 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4604 add(con);
4605 return con;
4606 }
4607
4608 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4609 // If called from Compile::scratch_emit_size do nothing.
4610 if (Compile::current()->in_scratch_emit_size()) return;
4611
4612 assert(labels.is_nonempty(), "must be");
4613 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4614
4615 // Since MachConstantNode::constant_offset() also contains
4616 // table_base_offset() we need to subtract the table_base_offset()
4617 // to get the plain offset into the constant table.
4618 int offset = n->constant_offset() - table_base_offset();
4619
4620 MacroAssembler _masm(&cb);
4621 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4622
4623 for (uint i = 0; i < n->outcnt(); i++) {
4624 address* constant_addr = &jump_table_base[i];
4625 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));
4626 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4627 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4628 }
4629 }
4630
4631 //----------------------------static_subtype_check-----------------------------
4632 // Shortcut important common cases when superklass is exact:
4633 // (0) superklass is java.lang.Object (can occur in reflective code)
4634 // (1) subklass is already limited to a subtype of superklass => always ok
4635 // (2) subklass does not overlap with superklass => always fail
4636 // (3) superklass has NO subtypes and we can check with a simple compare.
4637 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4638 if (StressReflectiveCode || superk == NULL || subk == NULL) {
4639 return SSC_full_test; // Let caller generate the general case.
4640 }
4641
4642 if (superk == env()->Object_klass()) {
4643 return SSC_always_true; // (0) this test cannot fail
4644 }
4645
4646 ciType* superelem = superk;
4647 if (superelem->is_array_klass()) {
4648 ciArrayKlass* ak = superelem->as_array_klass();
4649 superelem = superelem->as_array_klass()->base_element_type();
4650 }
4651
4652 if (!subk->is_interface()) { // cannot trust static interface types yet
4653 if (subk->is_subtype_of(superk)) {
4654 return SSC_always_true; // (1) false path dead; no dynamic test needed
4655 }
4656 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4657 !superk->is_subtype_of(subk)) {
4658 return SSC_always_false;
4659 }
4660 }
4661
4662 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays.
4663 // [V is a subtype of [V? but the klass for [V is not equal to the klass for [V?. Perform a full test.
4664 if (superk->is_obj_array_klass() && !superk->as_array_klass()->storage_properties().is_null_free() && superk->as_array_klass()->element_klass()->is_valuetype()) {
4665 return SSC_full_test;
4666 }
4667 // If casting to an instance klass, it must have no subtypes
4668 if (superk->is_interface()) {
4669 // Cannot trust interfaces yet.
4670 // %%% S.B. superk->nof_implementors() == 1
4671 } else if (superelem->is_instance_klass()) {
4672 ciInstanceKlass* ik = superelem->as_instance_klass();
4673 if (!ik->has_subklass() && !ik->is_interface()) {
4674 if (!ik->is_final()) {
4675 // Add a dependency if there is a chance of a later subclass.
4676 dependencies()->assert_leaf_type(ik);
4677 }
4678 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4679 }
4680 } else {
4681 // A primitive array type has no subtypes.
4682 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4683 }
4684
4685 return SSC_full_test;
4686 }
4687
4688 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4689 #ifdef _LP64
4690 // The scaled index operand to AddP must be a clean 64-bit value.
4691 // Java allows a 32-bit int to be incremented to a negative
4692 // value, which appears in a 64-bit register as a large
4693 // positive number. Using that large positive number as an
4694 // operand in pointer arithmetic has bad consequences.
4695 // On the other hand, 32-bit overflow is rare, and the possibility
4696 // can often be excluded, if we annotate the ConvI2L node with
4697 // a type assertion that its value is known to be a small positive
4698 // number. (The prior range check has ensured this.)
4699 // This assertion is used by ConvI2LNode::Ideal.
4700 int index_max = max_jint - 1; // array size is max_jint, index is one less
4701 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4702 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4703 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4704 #endif
4705 return idx;
4706 }
4707
4708 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4709 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4710 if (ctrl != NULL) {
4711 // Express control dependency by a CastII node with a narrow type.
4712 value = new CastIINode(value, itype, false, true /* range check dependency */);
4713 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4714 // node from floating above the range check during loop optimizations. Otherwise, the
4715 // ConvI2L node may be eliminated independently of the range check, causing the data path
4716 // to become TOP while the control path is still there (although it's unreachable).
4717 value->set_req(0, ctrl);
4718 // Save CastII node to remove it after loop optimizations.
4719 phase->C->add_range_check_cast(value);
4720 value = phase->transform(value);
4721 }
4722 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4723 return phase->transform(new ConvI2LNode(value, ltype));
4724 }
4725
4726 // The message about the current inlining is accumulated in
4727 // _print_inlining_stream and transfered into the _print_inlining_list
4728 // once we know whether inlining succeeds or not. For regular
4729 // inlining, messages are appended to the buffer pointed by
4730 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4731 // a new buffer is added after _print_inlining_idx in the list. This
4732 // way we can update the inlining message for late inlining call site
4733 // when the inlining is attempted again.
4734 void Compile::print_inlining_init() {
4735 if (print_inlining() || print_intrinsics()) {
4736 _print_inlining_stream = new stringStream();
4737 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4738 }
4739 }
4740
4741 void Compile::print_inlining_reinit() {
4742 if (print_inlining() || print_intrinsics()) {
4743 // Re allocate buffer when we change ResourceMark
4744 _print_inlining_stream = new stringStream();
4745 }
4746 }
4747
4748 void Compile::print_inlining_reset() {
4749 _print_inlining_stream->reset();
4750 }
4751
4752 void Compile::print_inlining_commit() {
4753 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4754 // Transfer the message from _print_inlining_stream to the current
4755 // _print_inlining_list buffer and clear _print_inlining_stream.
4756 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4757 print_inlining_reset();
4758 }
4759
4760 void Compile::print_inlining_push() {
4761 // Add new buffer to the _print_inlining_list at current position
4762 _print_inlining_idx++;
4763 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4764 }
4765
4766 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4767 return _print_inlining_list->at(_print_inlining_idx);
4768 }
4769
4770 void Compile::print_inlining_update(CallGenerator* cg) {
4771 if (print_inlining() || print_intrinsics()) {
4772 if (!cg->is_late_inline()) {
4773 if (print_inlining_current().cg() != NULL) {
4774 print_inlining_push();
4775 }
4776 print_inlining_commit();
4777 } else {
4778 if (print_inlining_current().cg() != cg &&
4779 (print_inlining_current().cg() != NULL ||
4780 print_inlining_current().ss()->size() != 0)) {
4781 print_inlining_push();
4782 }
4783 print_inlining_commit();
4784 print_inlining_current().set_cg(cg);
4785 }
4786 }
4787 }
4788
4789 void Compile::print_inlining_move_to(CallGenerator* cg) {
4790 // We resume inlining at a late inlining call site. Locate the
4791 // corresponding inlining buffer so that we can update it.
4792 if (print_inlining()) {
4793 for (int i = 0; i < _print_inlining_list->length(); i++) {
4794 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4795 _print_inlining_idx = i;
4796 return;
4797 }
4798 }
4799 ShouldNotReachHere();
4800 }
4801 }
4802
4803 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4804 if (print_inlining()) {
4805 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4806 assert(print_inlining_current().cg() == cg, "wrong entry");
4807 // replace message with new message
4808 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4809 print_inlining_commit();
4810 print_inlining_current().set_cg(cg);
4811 }
4812 }
4813
4814 void Compile::print_inlining_assert_ready() {
4815 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4816 }
4817
4818 void Compile::process_print_inlining() {
4819 bool do_print_inlining = print_inlining() || print_intrinsics();
4820 if (do_print_inlining || log() != NULL) {
4821 // Print inlining message for candidates that we couldn't inline
4822 // for lack of space
4823 for (int i = 0; i < _late_inlines.length(); i++) {
4824 CallGenerator* cg = _late_inlines.at(i);
4825 if (!cg->is_mh_late_inline()) {
4826 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4827 if (do_print_inlining) {
4828 cg->print_inlining_late(msg);
4829 }
4830 log_late_inline_failure(cg, msg);
4831 }
4832 }
4833 }
4834 if (do_print_inlining) {
4835 ResourceMark rm;
4836 stringStream ss;
4837 for (int i = 0; i < _print_inlining_list->length(); i++) {
4838 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4839 }
4840 size_t end = ss.size();
4841 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4842 strncpy(_print_inlining_output, ss.base(), end+1);
4843 _print_inlining_output[end] = 0;
4844 }
4845 }
4846
4847 void Compile::dump_print_inlining() {
4848 if (_print_inlining_output != NULL) {
4849 tty->print_raw(_print_inlining_output);
4850 }
4851 }
4852
4853 void Compile::log_late_inline(CallGenerator* cg) {
4854 if (log() != NULL) {
4855 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4856 cg->unique_id());
4857 JVMState* p = cg->call_node()->jvms();
4858 while (p != NULL) {
4859 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4860 p = p->caller();
4861 }
4862 log()->tail("late_inline");
4863 }
4864 }
4865
4866 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4867 log_late_inline(cg);
4868 if (log() != NULL) {
4869 log()->inline_fail(msg);
4870 }
4871 }
4872
4873 void Compile::log_inline_id(CallGenerator* cg) {
4874 if (log() != NULL) {
4875 // The LogCompilation tool needs a unique way to identify late
4876 // inline call sites. This id must be unique for this call site in
4877 // this compilation. Try to have it unique across compilations as
4878 // well because it can be convenient when grepping through the log
4879 // file.
4880 // Distinguish OSR compilations from others in case CICountOSR is
4881 // on.
4882 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4883 cg->set_unique_id(id);
4884 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4885 }
4886 }
4887
4888 void Compile::log_inline_failure(const char* msg) {
4889 if (C->log() != NULL) {
4890 C->log()->inline_fail(msg);
4891 }
4892 }
4893
4894
4895 // Dump inlining replay data to the stream.
4896 // Don't change thread state and acquire any locks.
4897 void Compile::dump_inline_data(outputStream* out) {
4898 InlineTree* inl_tree = ilt();
4899 if (inl_tree != NULL) {
4900 out->print(" inline %d", inl_tree->count());
4901 inl_tree->dump_replay_data(out);
4902 }
4903 }
4904
4905 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4906 if (n1->Opcode() < n2->Opcode()) return -1;
4907 else if (n1->Opcode() > n2->Opcode()) return 1;
4908
4909 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4910 for (uint i = 1; i < n1->req(); i++) {
4911 if (n1->in(i) < n2->in(i)) return -1;
4912 else if (n1->in(i) > n2->in(i)) return 1;
4913 }
4914
4915 return 0;
4916 }
4917
4918 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4919 Node* n1 = *n1p;
4920 Node* n2 = *n2p;
4921
4922 return cmp_expensive_nodes(n1, n2);
4923 }
4924
4925 void Compile::sort_expensive_nodes() {
4926 if (!expensive_nodes_sorted()) {
4927 _expensive_nodes->sort(cmp_expensive_nodes);
4928 }
4929 }
4930
4931 bool Compile::expensive_nodes_sorted() const {
4932 for (int i = 1; i < _expensive_nodes->length(); i++) {
4933 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4934 return false;
4935 }
4936 }
4937 return true;
4938 }
4939
4940 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4941 if (_expensive_nodes->length() == 0) {
4942 return false;
4943 }
4944
4945 assert(OptimizeExpensiveOps, "optimization off?");
4946
4947 // Take this opportunity to remove dead nodes from the list
4948 int j = 0;
4949 for (int i = 0; i < _expensive_nodes->length(); i++) {
4950 Node* n = _expensive_nodes->at(i);
4951 if (!n->is_unreachable(igvn)) {
4952 assert(n->is_expensive(), "should be expensive");
4953 _expensive_nodes->at_put(j, n);
4954 j++;
4955 }
4956 }
4957 _expensive_nodes->trunc_to(j);
4958
4959 // Then sort the list so that similar nodes are next to each other
4960 // and check for at least two nodes of identical kind with same data
4961 // inputs.
4962 sort_expensive_nodes();
4963
4964 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4965 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4966 return true;
4967 }
4968 }
4969
4970 return false;
4971 }
4972
4973 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4974 if (_expensive_nodes->length() == 0) {
4975 return;
4976 }
4977
4978 assert(OptimizeExpensiveOps, "optimization off?");
4979
4980 // Sort to bring similar nodes next to each other and clear the
4981 // control input of nodes for which there's only a single copy.
4982 sort_expensive_nodes();
4983
4984 int j = 0;
4985 int identical = 0;
4986 int i = 0;
4987 bool modified = false;
4988 for (; i < _expensive_nodes->length()-1; i++) {
4989 assert(j <= i, "can't write beyond current index");
4990 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4991 identical++;
4992 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4993 continue;
4994 }
4995 if (identical > 0) {
4996 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4997 identical = 0;
4998 } else {
4999 Node* n = _expensive_nodes->at(i);
5000 igvn.replace_input_of(n, 0, NULL);
5001 igvn.hash_insert(n);
5002 modified = true;
5003 }
5004 }
5005 if (identical > 0) {
5006 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5007 } else if (_expensive_nodes->length() >= 1) {
5008 Node* n = _expensive_nodes->at(i);
5009 igvn.replace_input_of(n, 0, NULL);
5010 igvn.hash_insert(n);
5011 modified = true;
5012 }
5013 _expensive_nodes->trunc_to(j);
5014 if (modified) {
5015 igvn.optimize();
5016 }
5017 }
5018
5019 void Compile::add_expensive_node(Node * n) {
5020 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
5021 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5022 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5023 if (OptimizeExpensiveOps) {
5024 _expensive_nodes->append(n);
5025 } else {
5026 // Clear control input and let IGVN optimize expensive nodes if
5027 // OptimizeExpensiveOps is off.
5028 n->set_req(0, NULL);
5029 }
5030 }
5031
5032 /**
5033 * Remove the speculative part of types and clean up the graph
5034 */
5035 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5036 if (UseTypeSpeculation) {
5037 Unique_Node_List worklist;
5038 worklist.push(root());
5039 int modified = 0;
5040 // Go over all type nodes that carry a speculative type, drop the
5041 // speculative part of the type and enqueue the node for an igvn
5042 // which may optimize it out.
5043 for (uint next = 0; next < worklist.size(); ++next) {
5044 Node *n = worklist.at(next);
5045 if (n->is_Type()) {
5046 TypeNode* tn = n->as_Type();
5047 const Type* t = tn->type();
5048 const Type* t_no_spec = t->remove_speculative();
5049 if (t_no_spec != t) {
5050 bool in_hash = igvn.hash_delete(n);
5051 assert(in_hash, "node should be in igvn hash table");
5052 tn->set_type(t_no_spec);
5053 igvn.hash_insert(n);
5054 igvn._worklist.push(n); // give it a chance to go away
5055 modified++;
5056 }
5057 }
5058 uint max = n->len();
5059 for( uint i = 0; i < max; ++i ) {
5060 Node *m = n->in(i);
5061 if (not_a_node(m)) continue;
5062 worklist.push(m);
5063 }
5064 }
5065 // Drop the speculative part of all types in the igvn's type table
5066 igvn.remove_speculative_types();
5067 if (modified > 0) {
5068 igvn.optimize();
5069 }
5070 #ifdef ASSERT
5071 // Verify that after the IGVN is over no speculative type has resurfaced
5072 worklist.clear();
5073 worklist.push(root());
5074 for (uint next = 0; next < worklist.size(); ++next) {
5075 Node *n = worklist.at(next);
5076 const Type* t = igvn.type_or_null(n);
5077 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5078 if (n->is_Type()) {
5079 t = n->as_Type()->type();
5080 assert(t == t->remove_speculative(), "no more speculative types");
5081 }
5082 uint max = n->len();
5083 for( uint i = 0; i < max; ++i ) {
5084 Node *m = n->in(i);
5085 if (not_a_node(m)) continue;
5086 worklist.push(m);
5087 }
5088 }
5089 igvn.check_no_speculative_types();
5090 #endif
5091 }
5092 }
5093
5094 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
5095 const TypeInstPtr* ta = phase->type(a)->isa_instptr();
5096 const TypeInstPtr* tb = phase->type(b)->isa_instptr();
5097 if (!EnableValhalla || ta == NULL || tb == NULL ||
5098 ta->is_zero_type() || tb->is_zero_type() ||
5099 !ta->can_be_value_type() || !tb->can_be_value_type()) {
5100 // Use old acmp if one operand is null or not a value type
5101 return new CmpPNode(a, b);
5102 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) {
5103 // We know that one operand is a value type. Therefore,
5104 // new acmp will only return true if both operands are NULL.
5105 // Check if both operands are null by or'ing the oops.
5106 a = phase->transform(new CastP2XNode(NULL, a));
5107 b = phase->transform(new CastP2XNode(NULL, b));
5108 a = phase->transform(new OrXNode(a, b));
5109 return new CmpXNode(a, phase->MakeConX(0));
5110 }
5111 // Use new acmp
5112 return NULL;
5113 }
5114
5115 // Auxiliary method to support randomized stressing/fuzzing.
5116 //
5117 // This method can be called the arbitrary number of times, with current count
5118 // as the argument. The logic allows selecting a single candidate from the
5119 // running list of candidates as follows:
5120 // int count = 0;
5121 // Cand* selected = null;
5122 // while(cand = cand->next()) {
5123 // if (randomized_select(++count)) {
5124 // selected = cand;
5125 // }
5126 // }
5127 //
5128 // Including count equalizes the chances any candidate is "selected".
5129 // This is useful when we don't have the complete list of candidates to choose
5130 // from uniformly. In this case, we need to adjust the randomicity of the
5131 // selection, or else we will end up biasing the selection towards the latter
5132 // candidates.
5133 //
5134 // Quick back-envelope calculation shows that for the list of n candidates
5135 // the equal probability for the candidate to persist as "best" can be
5136 // achieved by replacing it with "next" k-th candidate with the probability
5137 // of 1/k. It can be easily shown that by the end of the run, the
5138 // probability for any candidate is converged to 1/n, thus giving the
5139 // uniform distribution among all the candidates.
5140 //
5141 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
5142 #define RANDOMIZED_DOMAIN_POW 29
5143 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
5144 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
5145 bool Compile::randomized_select(int count) {
5146 assert(count > 0, "only positive");
5147 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
5148 }
5149
5150 CloneMap& Compile::clone_map() { return _clone_map; }
5151 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
5152
5153 void NodeCloneInfo::dump() const {
5154 tty->print(" {%d:%d} ", idx(), gen());
5155 }
5156
5157 void CloneMap::clone(Node* old, Node* nnn, int gen) {
5158 uint64_t val = value(old->_idx);
5159 NodeCloneInfo cio(val);
5160 assert(val != 0, "old node should be in the map");
5161 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
5162 insert(nnn->_idx, cin.get());
5163 #ifndef PRODUCT
5164 if (is_debug()) {
5165 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
5166 }
5167 #endif
5168 }
5169
5170 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
5171 NodeCloneInfo cio(value(old->_idx));
5172 if (cio.get() == 0) {
5173 cio.set(old->_idx, 0);
5174 insert(old->_idx, cio.get());
5175 #ifndef PRODUCT
5176 if (is_debug()) {
5177 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
5178 }
5179 #endif
5180 }
5181 clone(old, nnn, gen);
5182 }
5183
5184 int CloneMap::max_gen() const {
5185 int g = 0;
5186 DictI di(_dict);
5187 for(; di.test(); ++di) {
5188 int t = gen(di._key);
5189 if (g < t) {
5190 g = t;
5191 #ifndef PRODUCT
5192 if (is_debug()) {
5193 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5194 }
5195 #endif
5196 }
5197 }
5198 return g;
5199 }
5200
5201 void CloneMap::dump(node_idx_t key) const {
5202 uint64_t val = value(key);
5203 if (val != 0) {
5204 NodeCloneInfo ni(val);
5205 ni.dump();
5206 }
5207 }