1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vectornode.hpp"
71 #include "runtime/arguments.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stubRoutines.hpp"
75 #include "runtime/timer.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/copy.hpp"
78 #include "utilities/macros.hpp"
79 #if INCLUDE_G1GC
80 #include "gc/g1/g1ThreadLocalData.hpp"
81 #endif // INCLUDE_G1GC
82 #if INCLUDE_ZGC
83 #include "gc/z/c2/zBarrierSetC2.hpp"
84 #endif
85
86
87 // -------------------- Compile::mach_constant_base_node -----------------------
88 // Constant table base node singleton.
89 MachConstantBaseNode* Compile::mach_constant_base_node() {
90 if (_mach_constant_base_node == NULL) {
91 _mach_constant_base_node = new MachConstantBaseNode();
92 _mach_constant_base_node->add_req(C->root());
93 }
94 return _mach_constant_base_node;
95 }
96
97
98 /// Support for intrinsics.
99
100 // Return the index at which m must be inserted (or already exists).
101 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
102 class IntrinsicDescPair {
103 private:
104 ciMethod* _m;
105 bool _is_virtual;
106 public:
107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
109 ciMethod* m= elt->method();
110 ciMethod* key_m = key->_m;
111 if (key_m < m) return -1;
112 else if (key_m > m) return 1;
113 else {
114 bool is_virtual = elt->is_virtual();
115 bool key_virtual = key->_is_virtual;
116 if (key_virtual < is_virtual) return -1;
117 else if (key_virtual > is_virtual) return 1;
118 else return 0;
119 }
120 }
121 };
122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
123 #ifdef ASSERT
124 for (int i = 1; i < _intrinsics->length(); i++) {
125 CallGenerator* cg1 = _intrinsics->at(i-1);
126 CallGenerator* cg2 = _intrinsics->at(i);
127 assert(cg1->method() != cg2->method()
128 ? cg1->method() < cg2->method()
129 : cg1->is_virtual() < cg2->is_virtual(),
130 "compiler intrinsics list must stay sorted");
131 }
132 #endif
133 IntrinsicDescPair pair(m, is_virtual);
134 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
135 }
136
137 void Compile::register_intrinsic(CallGenerator* cg) {
138 if (_intrinsics == NULL) {
139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
140 }
141 int len = _intrinsics->length();
142 bool found = false;
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
144 assert(!found, "registering twice");
145 _intrinsics->insert_before(index, cg);
146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
147 }
148
149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
150 assert(m->is_loaded(), "don't try this on unloaded methods");
151 if (_intrinsics != NULL) {
152 bool found = false;
153 int index = intrinsic_insertion_index(m, is_virtual, found);
154 if (found) {
155 return _intrinsics->at(index);
156 }
157 }
158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
159 if (m->intrinsic_id() != vmIntrinsics::_none &&
160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
162 if (cg != NULL) {
163 // Save it for next time:
164 register_intrinsic(cg);
165 return cg;
166 } else {
167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
168 }
169 }
170 return NULL;
171 }
172
173 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
174 // in library_call.cpp.
175
176
177 #ifndef PRODUCT
178 // statistics gathering...
179
180 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
182
183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
185 int oflags = _intrinsic_hist_flags[id];
186 assert(flags != 0, "what happened?");
187 if (is_virtual) {
188 flags |= _intrinsic_virtual;
189 }
190 bool changed = (flags != oflags);
191 if ((flags & _intrinsic_worked) != 0) {
192 juint count = (_intrinsic_hist_count[id] += 1);
193 if (count == 1) {
194 changed = true; // first time
195 }
196 // increment the overall count also:
197 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
198 }
199 if (changed) {
200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
201 // Something changed about the intrinsic's virtuality.
202 if ((flags & _intrinsic_virtual) != 0) {
203 // This is the first use of this intrinsic as a virtual call.
204 if (oflags != 0) {
205 // We already saw it as a non-virtual, so note both cases.
206 flags |= _intrinsic_both;
207 }
208 } else if ((oflags & _intrinsic_both) == 0) {
209 // This is the first use of this intrinsic as a non-virtual
210 flags |= _intrinsic_both;
211 }
212 }
213 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
214 }
215 // update the overall flags also:
216 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
217 return changed;
218 }
219
220 static char* format_flags(int flags, char* buf) {
221 buf[0] = 0;
222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
227 if (buf[0] == 0) strcat(buf, ",");
228 assert(buf[0] == ',', "must be");
229 return &buf[1];
230 }
231
232 void Compile::print_intrinsic_statistics() {
233 char flagsbuf[100];
234 ttyLocker ttyl;
235 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
236 tty->print_cr("Compiler intrinsic usage:");
237 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
238 if (total == 0) total = 1; // avoid div0 in case of no successes
239 #define PRINT_STAT_LINE(name, c, f) \
240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
241 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
242 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
243 int flags = _intrinsic_hist_flags[id];
244 juint count = _intrinsic_hist_count[id];
245 if ((flags | count) != 0) {
246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
247 }
248 }
249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
250 if (xtty != NULL) xtty->tail("statistics");
251 }
252
253 void Compile::print_statistics() {
254 { ttyLocker ttyl;
255 if (xtty != NULL) xtty->head("statistics type='opto'");
256 Parse::print_statistics();
257 PhaseCCP::print_statistics();
258 PhaseRegAlloc::print_statistics();
259 Scheduling::print_statistics();
260 PhasePeephole::print_statistics();
261 PhaseIdealLoop::print_statistics();
262 if (xtty != NULL) xtty->tail("statistics");
263 }
264 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
265 // put this under its own <statistics> element.
266 print_intrinsic_statistics();
267 }
268 }
269 #endif //PRODUCT
270
271 // Support for bundling info
272 Bundle* Compile::node_bundling(const Node *n) {
273 assert(valid_bundle_info(n), "oob");
274 return &_node_bundling_base[n->_idx];
275 }
276
277 bool Compile::valid_bundle_info(const Node *n) {
278 return (_node_bundling_limit > n->_idx);
279 }
280
281
282 void Compile::gvn_replace_by(Node* n, Node* nn) {
283 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
284 Node* use = n->last_out(i);
285 bool is_in_table = initial_gvn()->hash_delete(use);
286 uint uses_found = 0;
287 for (uint j = 0; j < use->len(); j++) {
288 if (use->in(j) == n) {
289 if (j < use->req())
290 use->set_req(j, nn);
291 else
292 use->set_prec(j, nn);
293 uses_found++;
294 }
295 }
296 if (is_in_table) {
297 // reinsert into table
298 initial_gvn()->hash_find_insert(use);
299 }
300 record_for_igvn(use);
301 i -= uses_found; // we deleted 1 or more copies of this edge
302 }
303 }
304
305
306 static inline bool not_a_node(const Node* n) {
307 if (n == NULL) return true;
308 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
309 if (*(address*)n == badAddress) return true; // kill by Node::destruct
310 return false;
311 }
312
313 // Identify all nodes that are reachable from below, useful.
314 // Use breadth-first pass that records state in a Unique_Node_List,
315 // recursive traversal is slower.
316 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
317 int estimated_worklist_size = live_nodes();
318 useful.map( estimated_worklist_size, NULL ); // preallocate space
319
320 // Initialize worklist
321 if (root() != NULL) { useful.push(root()); }
322 // If 'top' is cached, declare it useful to preserve cached node
323 if( cached_top_node() ) { useful.push(cached_top_node()); }
324
325 // Push all useful nodes onto the list, breadthfirst
326 for( uint next = 0; next < useful.size(); ++next ) {
327 assert( next < unique(), "Unique useful nodes < total nodes");
328 Node *n = useful.at(next);
329 uint max = n->len();
330 for( uint i = 0; i < max; ++i ) {
331 Node *m = n->in(i);
332 if (not_a_node(m)) continue;
333 useful.push(m);
334 }
335 }
336 }
337
338 // Update dead_node_list with any missing dead nodes using useful
339 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
340 void Compile::update_dead_node_list(Unique_Node_List &useful) {
341 uint max_idx = unique();
342 VectorSet& useful_node_set = useful.member_set();
343
344 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
345 // If node with index node_idx is not in useful set,
346 // mark it as dead in dead node list.
347 if (! useful_node_set.test(node_idx) ) {
348 record_dead_node(node_idx);
349 }
350 }
351 }
352
353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
354 int shift = 0;
355 for (int i = 0; i < inlines->length(); i++) {
356 CallGenerator* cg = inlines->at(i);
357 CallNode* call = cg->call_node();
358 if (shift > 0) {
359 inlines->at_put(i-shift, cg);
360 }
361 if (!useful.member(call)) {
362 shift++;
363 }
364 }
365 inlines->trunc_to(inlines->length()-shift);
366 }
367
368 // Disconnect all useless nodes by disconnecting those at the boundary.
369 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
370 uint next = 0;
371 while (next < useful.size()) {
372 Node *n = useful.at(next++);
373 if (n->is_SafePoint()) {
374 // We're done with a parsing phase. Replaced nodes are not valid
375 // beyond that point.
376 n->as_SafePoint()->delete_replaced_nodes();
377 }
378 // Use raw traversal of out edges since this code removes out edges
379 int max = n->outcnt();
380 for (int j = 0; j < max; ++j) {
381 Node* child = n->raw_out(j);
382 if (! useful.member(child)) {
383 assert(!child->is_top() || child != top(),
384 "If top is cached in Compile object it is in useful list");
385 // Only need to remove this out-edge to the useless node
386 n->raw_del_out(j);
387 --j;
388 --max;
389 }
390 }
391 if (n->outcnt() == 1 && n->has_special_unique_user()) {
392 record_for_igvn(n->unique_out());
393 }
394 }
395 // Remove useless macro and predicate opaq nodes
396 for (int i = C->macro_count()-1; i >= 0; i--) {
397 Node* n = C->macro_node(i);
398 if (!useful.member(n)) {
399 remove_macro_node(n);
400 }
401 }
402 // Remove useless CastII nodes with range check dependency
403 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
404 Node* cast = range_check_cast_node(i);
405 if (!useful.member(cast)) {
406 remove_range_check_cast(cast);
407 }
408 }
409 // Remove useless expensive nodes
410 for (int i = C->expensive_count()-1; i >= 0; i--) {
411 Node* n = C->expensive_node(i);
412 if (!useful.member(n)) {
413 remove_expensive_node(n);
414 }
415 }
416 // Remove useless Opaque4 nodes
417 for (int i = opaque4_count() - 1; i >= 0; i--) {
418 Node* opaq = opaque4_node(i);
419 if (!useful.member(opaq)) {
420 remove_opaque4_node(opaq);
421 }
422 }
423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
424 bs->eliminate_useless_gc_barriers(useful);
425 // clean up the late inline lists
426 remove_useless_late_inlines(&_string_late_inlines, useful);
427 remove_useless_late_inlines(&_boxing_late_inlines, useful);
428 remove_useless_late_inlines(&_late_inlines, useful);
429 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
430 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
431 }
432
433 //------------------------------frame_size_in_words-----------------------------
434 // frame_slots in units of words
435 int Compile::frame_size_in_words() const {
436 // shift is 0 in LP32 and 1 in LP64
437 const int shift = (LogBytesPerWord - LogBytesPerInt);
438 int words = _frame_slots >> shift;
439 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
440 return words;
441 }
442
443 // To bang the stack of this compiled method we use the stack size
444 // that the interpreter would need in case of a deoptimization. This
445 // removes the need to bang the stack in the deoptimization blob which
446 // in turn simplifies stack overflow handling.
447 int Compile::bang_size_in_bytes() const {
448 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
449 }
450
451 // ============================================================================
452 //------------------------------CompileWrapper---------------------------------
453 class CompileWrapper : public StackObj {
454 Compile *const _compile;
455 public:
456 CompileWrapper(Compile* compile);
457
458 ~CompileWrapper();
459 };
460
461 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
462 // the Compile* pointer is stored in the current ciEnv:
463 ciEnv* env = compile->env();
464 assert(env == ciEnv::current(), "must already be a ciEnv active");
465 assert(env->compiler_data() == NULL, "compile already active?");
466 env->set_compiler_data(compile);
467 assert(compile == Compile::current(), "sanity");
468
469 compile->set_type_dict(NULL);
470 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
471 compile->clone_map().set_clone_idx(0);
472 compile->set_type_hwm(NULL);
473 compile->set_type_last_size(0);
474 compile->set_last_tf(NULL, NULL);
475 compile->set_indexSet_arena(NULL);
476 compile->set_indexSet_free_block_list(NULL);
477 compile->init_type_arena();
478 Type::Initialize(compile);
479 _compile->set_scratch_buffer_blob(NULL);
480 _compile->begin_method();
481 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
482 }
483 CompileWrapper::~CompileWrapper() {
484 _compile->end_method();
485 if (_compile->scratch_buffer_blob() != NULL)
486 BufferBlob::free(_compile->scratch_buffer_blob());
487 _compile->env()->set_compiler_data(NULL);
488 }
489
490
491 //----------------------------print_compile_messages---------------------------
492 void Compile::print_compile_messages() {
493 #ifndef PRODUCT
494 // Check if recompiling
495 if (_subsume_loads == false && PrintOpto) {
496 // Recompiling without allowing machine instructions to subsume loads
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without subsuming loads **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
502 // Recompiling without escape analysis
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without escape analysis **");
505 tty->print_cr("*********************************************************");
506 }
507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
508 // Recompiling without boxing elimination
509 tty->print_cr("*********************************************************");
510 tty->print_cr("** Bailout: Recompile without boxing elimination **");
511 tty->print_cr("*********************************************************");
512 }
513 if (C->directive()->BreakAtCompileOption) {
514 // Open the debugger when compiling this method.
515 tty->print("### Breaking when compiling: ");
516 method()->print_short_name();
517 tty->cr();
518 BREAKPOINT;
519 }
520
521 if( PrintOpto ) {
522 if (is_osr_compilation()) {
523 tty->print("[OSR]%3d", _compile_id);
524 } else {
525 tty->print("%3d", _compile_id);
526 }
527 }
528 #endif
529 }
530
531
532 //-----------------------init_scratch_buffer_blob------------------------------
533 // Construct a temporary BufferBlob and cache it for this compile.
534 void Compile::init_scratch_buffer_blob(int const_size) {
535 // If there is already a scratch buffer blob allocated and the
536 // constant section is big enough, use it. Otherwise free the
537 // current and allocate a new one.
538 BufferBlob* blob = scratch_buffer_blob();
539 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
540 // Use the current blob.
541 } else {
542 if (blob != NULL) {
543 BufferBlob::free(blob);
544 }
545
546 ResourceMark rm;
547 _scratch_const_size = const_size;
548 int locs_size = sizeof(relocInfo) * MAX_locs_size;
549 int slop = 2 * CodeSection::end_slop(); // space between sections
550 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size + slop + locs_size);
551 blob = BufferBlob::create("Compile::scratch_buffer", size);
552 // Record the buffer blob for next time.
553 set_scratch_buffer_blob(blob);
554 // Have we run out of code space?
555 if (scratch_buffer_blob() == NULL) {
556 // Let CompilerBroker disable further compilations.
557 record_failure("Not enough space for scratch buffer in CodeCache");
558 return;
559 }
560 }
561
562 // Initialize the relocation buffers
563 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
564 set_scratch_locs_memory(locs_buf);
565 }
566
567
568 //-----------------------scratch_emit_size-------------------------------------
569 // Helper function that computes size by emitting code
570 uint Compile::scratch_emit_size(const Node* n) {
571 // Start scratch_emit_size section.
572 set_in_scratch_emit_size(true);
573
574 // Emit into a trash buffer and count bytes emitted.
575 // This is a pretty expensive way to compute a size,
576 // but it works well enough if seldom used.
577 // All common fixed-size instructions are given a size
578 // method by the AD file.
579 // Note that the scratch buffer blob and locs memory are
580 // allocated at the beginning of the compile task, and
581 // may be shared by several calls to scratch_emit_size.
582 // The allocation of the scratch buffer blob is particularly
583 // expensive, since it has to grab the code cache lock.
584 BufferBlob* blob = this->scratch_buffer_blob();
585 assert(blob != NULL, "Initialize BufferBlob at start");
586 assert(blob->size() > MAX_inst_size, "sanity");
587 relocInfo* locs_buf = scratch_locs_memory();
588 address blob_begin = blob->content_begin();
589 address blob_end = (address)locs_buf;
590 assert(blob->contains(blob_end), "sanity");
591 CodeBuffer buf(blob_begin, blob_end - blob_begin);
592 buf.initialize_consts_size(_scratch_const_size);
593 buf.initialize_stubs_size(MAX_stubs_size);
594 assert(locs_buf != NULL, "sanity");
595 int lsize = MAX_locs_size / 3;
596 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
597 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
598 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
599 // Mark as scratch buffer.
600 buf.consts()->set_scratch_emit();
601 buf.insts()->set_scratch_emit();
602 buf.stubs()->set_scratch_emit();
603
604 // Do the emission.
605
606 Label fakeL; // Fake label for branch instructions.
607 Label* saveL = NULL;
608 uint save_bnum = 0;
609 bool is_branch = n->is_MachBranch();
610 if (is_branch) {
611 MacroAssembler masm(&buf);
612 masm.bind(fakeL);
613 n->as_MachBranch()->save_label(&saveL, &save_bnum);
614 n->as_MachBranch()->label_set(&fakeL, 0);
615 }
616 n->emit(buf, this->regalloc());
617
618 // Emitting into the scratch buffer should not fail
619 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
620
621 if (is_branch) // Restore label.
622 n->as_MachBranch()->label_set(saveL, save_bnum);
623
624 // End scratch_emit_size section.
625 set_in_scratch_emit_size(false);
626
627 return buf.insts_size();
628 }
629
630
631 // ============================================================================
632 //------------------------------Compile standard-------------------------------
633 debug_only( int Compile::_debug_idx = 100000; )
634
635 // Compile a method. entry_bci is -1 for normal compilations and indicates
636 // the continuation bci for on stack replacement.
637
638
639 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
640 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
641 : Phase(Compiler),
642 _compile_id(ci_env->compile_id()),
643 _save_argument_registers(false),
644 _subsume_loads(subsume_loads),
645 _do_escape_analysis(do_escape_analysis),
646 _eliminate_boxing(eliminate_boxing),
647 _method(target),
648 _entry_bci(osr_bci),
649 _stub_function(NULL),
650 _stub_name(NULL),
651 _stub_entry_point(NULL),
652 _max_node_limit(MaxNodeLimit),
653 _orig_pc_slot(0),
654 _orig_pc_slot_offset_in_bytes(0),
655 _inlining_progress(false),
656 _inlining_incrementally(false),
657 _has_reserved_stack_access(target->has_reserved_stack_access()),
658 #ifndef PRODUCT
659 _trace_opto_output(directive->TraceOptoOutputOption),
660 #endif
661 _has_method_handle_invokes(false),
662 _comp_arena(mtCompiler),
663 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
664 _env(ci_env),
665 _directive(directive),
666 _log(ci_env->log()),
667 _failure_reason(NULL),
668 _congraph(NULL),
669 #ifndef PRODUCT
670 _printer(IdealGraphPrinter::printer()),
671 #endif
672 _dead_node_list(comp_arena()),
673 _dead_node_count(0),
674 _node_arena(mtCompiler),
675 _old_arena(mtCompiler),
676 _mach_constant_base_node(NULL),
677 _Compile_types(mtCompiler),
678 _initial_gvn(NULL),
679 _for_igvn(NULL),
680 _warm_calls(NULL),
681 _late_inlines(comp_arena(), 2, 0, NULL),
682 _string_late_inlines(comp_arena(), 2, 0, NULL),
683 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
684 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
685 _late_inlines_pos(0),
686 _number_of_mh_late_inlines(0),
687 _print_inlining_stream(NULL),
688 _print_inlining_list(NULL),
689 _print_inlining_idx(0),
690 _print_inlining_output(NULL),
691 _replay_inline_data(NULL),
692 _java_calls(0),
693 _inner_loops(0),
694 _interpreter_frame_size(0),
695 _node_bundling_limit(0),
696 _node_bundling_base(NULL),
697 _code_buffer("Compile::Fill_buffer"),
698 _scratch_const_size(-1),
699 _in_scratch_emit_size(false)
700 #ifndef PRODUCT
701 , _in_dump_cnt(0)
702 #endif
703 {
704 C = this;
705 #ifndef PRODUCT
706 if (_printer != NULL) {
707 _printer->set_compile(this);
708 }
709 #endif
710 CompileWrapper cw(this);
711
712 if (CITimeVerbose) {
713 tty->print(" ");
714 target->holder()->name()->print();
715 tty->print(".");
716 target->print_short_name();
717 tty->print(" ");
718 }
719 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
720 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
721
722 #ifndef PRODUCT
723 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
724 if (!print_opto_assembly) {
725 bool print_assembly = directive->PrintAssemblyOption;
726 if (print_assembly && !Disassembler::can_decode()) {
727 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
728 print_opto_assembly = true;
729 }
730 }
731 set_print_assembly(print_opto_assembly);
732 set_parsed_irreducible_loop(false);
733
734 if (directive->ReplayInlineOption) {
735 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
736 }
737 #endif
738 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
739 set_print_intrinsics(directive->PrintIntrinsicsOption);
740 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
741
742 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
743 // Make sure the method being compiled gets its own MDO,
744 // so we can at least track the decompile_count().
745 // Need MDO to record RTM code generation state.
746 method()->ensure_method_data();
747 }
748
749 Init(::AliasLevel);
750
751
752 print_compile_messages();
753
754 _ilt = InlineTree::build_inline_tree_root();
755
756 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
757 assert(num_alias_types() >= AliasIdxRaw, "");
758
759 #define MINIMUM_NODE_HASH 1023
760 // Node list that Iterative GVN will start with
761 Unique_Node_List for_igvn(comp_arena());
762 set_for_igvn(&for_igvn);
763
764 // GVN that will be run immediately on new nodes
765 uint estimated_size = method()->code_size()*4+64;
766 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
767 PhaseGVN gvn(node_arena(), estimated_size);
768 set_initial_gvn(&gvn);
769
770 print_inlining_init();
771 { // Scope for timing the parser
772 TracePhase tp("parse", &timers[_t_parser]);
773
774 // Put top into the hash table ASAP.
775 initial_gvn()->transform_no_reclaim(top());
776
777 // Set up tf(), start(), and find a CallGenerator.
778 CallGenerator* cg = NULL;
779 if (is_osr_compilation()) {
780 const TypeTuple *domain = StartOSRNode::osr_domain();
781 const TypeTuple *range = TypeTuple::make_range(method()->signature());
782 init_tf(TypeFunc::make(domain, range));
783 StartNode* s = new StartOSRNode(root(), domain);
784 initial_gvn()->set_type_bottom(s);
785 init_start(s);
786 cg = CallGenerator::for_osr(method(), entry_bci());
787 } else {
788 // Normal case.
789 init_tf(TypeFunc::make(method()));
790 StartNode* s = new StartNode(root(), tf()->domain());
791 initial_gvn()->set_type_bottom(s);
792 init_start(s);
793 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
794 // With java.lang.ref.reference.get() we must go through the
795 // intrinsic - even when get() is the root
796 // method of the compile - so that, if necessary, the value in
797 // the referent field of the reference object gets recorded by
798 // the pre-barrier code.
799 cg = find_intrinsic(method(), false);
800 }
801 if (cg == NULL) {
802 float past_uses = method()->interpreter_invocation_count();
803 float expected_uses = past_uses;
804 cg = CallGenerator::for_inline(method(), expected_uses);
805 }
806 }
807 if (failing()) return;
808 if (cg == NULL) {
809 record_method_not_compilable("cannot parse method");
810 return;
811 }
812 JVMState* jvms = build_start_state(start(), tf());
813 if ((jvms = cg->generate(jvms)) == NULL) {
814 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
815 record_method_not_compilable("method parse failed");
816 }
817 return;
818 }
819 GraphKit kit(jvms);
820
821 if (!kit.stopped()) {
822 // Accept return values, and transfer control we know not where.
823 // This is done by a special, unique ReturnNode bound to root.
824 return_values(kit.jvms());
825 }
826
827 if (kit.has_exceptions()) {
828 // Any exceptions that escape from this call must be rethrown
829 // to whatever caller is dynamically above us on the stack.
830 // This is done by a special, unique RethrowNode bound to root.
831 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
832 }
833
834 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
835
836 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
837 inline_string_calls(true);
838 }
839
840 if (failing()) return;
841
842 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
843
844 // Remove clutter produced by parsing.
845 if (!failing()) {
846 ResourceMark rm;
847 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
848 }
849 }
850
851 // Note: Large methods are capped off in do_one_bytecode().
852 if (failing()) return;
853
854 // After parsing, node notes are no longer automagic.
855 // They must be propagated by register_new_node_with_optimizer(),
856 // clone(), or the like.
857 set_default_node_notes(NULL);
858
859 for (;;) {
860 int successes = Inline_Warm();
861 if (failing()) return;
862 if (successes == 0) break;
863 }
864
865 // Drain the list.
866 Finish_Warm();
867 #ifndef PRODUCT
868 if (_printer && _printer->should_print(1)) {
869 _printer->print_inlining();
870 }
871 #endif
872
873 if (failing()) return;
874 NOT_PRODUCT( verify_graph_edges(); )
875
876 // Now optimize
877 Optimize();
878 if (failing()) return;
879 NOT_PRODUCT( verify_graph_edges(); )
880
881 #ifndef PRODUCT
882 if (PrintIdeal) {
883 ttyLocker ttyl; // keep the following output all in one block
884 // This output goes directly to the tty, not the compiler log.
885 // To enable tools to match it up with the compilation activity,
886 // be sure to tag this tty output with the compile ID.
887 if (xtty != NULL) {
888 xtty->head("ideal compile_id='%d'%s", compile_id(),
889 is_osr_compilation() ? " compile_kind='osr'" :
890 "");
891 }
892 root()->dump(9999);
893 if (xtty != NULL) {
894 xtty->tail("ideal");
895 }
896 }
897 #endif
898
899 NOT_PRODUCT( verify_barriers(); )
900
901 // Dump compilation data to replay it.
902 if (directive->DumpReplayOption) {
903 env()->dump_replay_data(_compile_id);
904 }
905 if (directive->DumpInlineOption && (ilt() != NULL)) {
906 env()->dump_inline_data(_compile_id);
907 }
908
909 // Now that we know the size of all the monitors we can add a fixed slot
910 // for the original deopt pc.
911
912 _orig_pc_slot = fixed_slots();
913 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
914 set_fixed_slots(next_slot);
915
916 // Compute when to use implicit null checks. Used by matching trap based
917 // nodes and NullCheck optimization.
918 set_allowed_deopt_reasons();
919
920 // Now generate code
921 Code_Gen();
922 if (failing()) return;
923
924 // Check if we want to skip execution of all compiled code.
925 {
926 #ifndef PRODUCT
927 if (OptoNoExecute) {
928 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
929 return;
930 }
931 #endif
932 TracePhase tp("install_code", &timers[_t_registerMethod]);
933
934 if (is_osr_compilation()) {
935 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
936 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
937 } else {
938 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
939 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
940 }
941
942 env()->register_method(_method, _entry_bci,
943 &_code_offsets,
944 _orig_pc_slot_offset_in_bytes,
945 code_buffer(),
946 frame_size_in_words(), _oop_map_set,
947 &_handler_table, &_inc_table,
948 compiler,
949 has_unsafe_access(),
950 SharedRuntime::is_wide_vector(max_vector_size()),
951 rtm_state()
952 );
953
954 if (log() != NULL) // Print code cache state into compiler log
955 log()->code_cache_state();
956 }
957 }
958
959 //------------------------------Compile----------------------------------------
960 // Compile a runtime stub
961 Compile::Compile( ciEnv* ci_env,
962 TypeFunc_generator generator,
963 address stub_function,
964 const char *stub_name,
965 int is_fancy_jump,
966 bool pass_tls,
967 bool save_arg_registers,
968 bool return_pc,
969 DirectiveSet* directive)
970 : Phase(Compiler),
971 _compile_id(0),
972 _save_argument_registers(save_arg_registers),
973 _subsume_loads(true),
974 _do_escape_analysis(false),
975 _eliminate_boxing(false),
976 _method(NULL),
977 _entry_bci(InvocationEntryBci),
978 _stub_function(stub_function),
979 _stub_name(stub_name),
980 _stub_entry_point(NULL),
981 _max_node_limit(MaxNodeLimit),
982 _orig_pc_slot(0),
983 _orig_pc_slot_offset_in_bytes(0),
984 _inlining_progress(false),
985 _inlining_incrementally(false),
986 _has_reserved_stack_access(false),
987 #ifndef PRODUCT
988 _trace_opto_output(directive->TraceOptoOutputOption),
989 #endif
990 _has_method_handle_invokes(false),
991 _comp_arena(mtCompiler),
992 _env(ci_env),
993 _directive(directive),
994 _log(ci_env->log()),
995 _failure_reason(NULL),
996 _congraph(NULL),
997 #ifndef PRODUCT
998 _printer(NULL),
999 #endif
1000 _dead_node_list(comp_arena()),
1001 _dead_node_count(0),
1002 _node_arena(mtCompiler),
1003 _old_arena(mtCompiler),
1004 _mach_constant_base_node(NULL),
1005 _Compile_types(mtCompiler),
1006 _initial_gvn(NULL),
1007 _for_igvn(NULL),
1008 _warm_calls(NULL),
1009 _number_of_mh_late_inlines(0),
1010 _print_inlining_stream(NULL),
1011 _print_inlining_list(NULL),
1012 _print_inlining_idx(0),
1013 _print_inlining_output(NULL),
1014 _replay_inline_data(NULL),
1015 _java_calls(0),
1016 _inner_loops(0),
1017 _interpreter_frame_size(0),
1018 _node_bundling_limit(0),
1019 _node_bundling_base(NULL),
1020 _code_buffer("Compile::Fill_buffer"),
1021 #ifndef PRODUCT
1022 _in_dump_cnt(0),
1023 #endif
1024 _allowed_reasons(0) {
1025 C = this;
1026
1027 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1028 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1029
1030 #ifndef PRODUCT
1031 set_print_assembly(PrintFrameConverterAssembly);
1032 set_parsed_irreducible_loop(false);
1033 #endif
1034 set_has_irreducible_loop(false); // no loops
1035
1036 CompileWrapper cw(this);
1037 Init(/*AliasLevel=*/ 0);
1038 init_tf((*generator)());
1039
1040 {
1041 // The following is a dummy for the sake of GraphKit::gen_stub
1042 Unique_Node_List for_igvn(comp_arena());
1043 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1044 PhaseGVN gvn(Thread::current()->resource_area(),255);
1045 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1046 gvn.transform_no_reclaim(top());
1047
1048 GraphKit kit;
1049 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1050 }
1051
1052 NOT_PRODUCT( verify_graph_edges(); )
1053 Code_Gen();
1054 if (failing()) return;
1055
1056
1057 // Entry point will be accessed using compile->stub_entry_point();
1058 if (code_buffer() == NULL) {
1059 Matcher::soft_match_failure();
1060 } else {
1061 if (PrintAssembly && (WizardMode || Verbose))
1062 tty->print_cr("### Stub::%s", stub_name);
1063
1064 if (!failing()) {
1065 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1066
1067 // Make the NMethod
1068 // For now we mark the frame as never safe for profile stackwalking
1069 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1070 code_buffer(),
1071 CodeOffsets::frame_never_safe,
1072 // _code_offsets.value(CodeOffsets::Frame_Complete),
1073 frame_size_in_words(),
1074 _oop_map_set,
1075 save_arg_registers);
1076 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1077
1078 _stub_entry_point = rs->entry_point();
1079 }
1080 }
1081 }
1082
1083 //------------------------------Init-------------------------------------------
1084 // Prepare for a single compilation
1085 void Compile::Init(int aliaslevel) {
1086 _unique = 0;
1087 _regalloc = NULL;
1088
1089 _tf = NULL; // filled in later
1090 _top = NULL; // cached later
1091 _matcher = NULL; // filled in later
1092 _cfg = NULL; // filled in later
1093
1094 set_24_bit_selection_and_mode(Use24BitFP, false);
1095
1096 _node_note_array = NULL;
1097 _default_node_notes = NULL;
1098 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1099
1100 _immutable_memory = NULL; // filled in at first inquiry
1101
1102 // Globally visible Nodes
1103 // First set TOP to NULL to give safe behavior during creation of RootNode
1104 set_cached_top_node(NULL);
1105 set_root(new RootNode());
1106 // Now that you have a Root to point to, create the real TOP
1107 set_cached_top_node( new ConNode(Type::TOP) );
1108 set_recent_alloc(NULL, NULL);
1109
1110 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1111 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1112 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1113 env()->set_dependencies(new Dependencies(env()));
1114
1115 _fixed_slots = 0;
1116 set_has_split_ifs(false);
1117 set_has_loops(has_method() && method()->has_loops()); // first approximation
1118 set_has_stringbuilder(false);
1119 set_has_boxed_value(false);
1120 _trap_can_recompile = false; // no traps emitted yet
1121 _major_progress = true; // start out assuming good things will happen
1122 set_has_unsafe_access(false);
1123 set_max_vector_size(0);
1124 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1125 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1126 set_decompile_count(0);
1127
1128 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1129 set_num_loop_opts(LoopOptsCount);
1130 set_do_inlining(Inline);
1131 set_max_inline_size(MaxInlineSize);
1132 set_freq_inline_size(FreqInlineSize);
1133 set_do_scheduling(OptoScheduling);
1134 set_do_count_invocations(false);
1135 set_do_method_data_update(false);
1136
1137 set_do_vector_loop(false);
1138
1139 if (AllowVectorizeOnDemand) {
1140 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1141 set_do_vector_loop(true);
1142 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());})
1143 } else if (has_method() && method()->name() != 0 &&
1144 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1145 set_do_vector_loop(true);
1146 }
1147 }
1148 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1149 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());})
1150
1151 set_age_code(has_method() && method()->profile_aging());
1152 set_rtm_state(NoRTM); // No RTM lock eliding by default
1153 _max_node_limit = _directive->MaxNodeLimitOption;
1154
1155 #if INCLUDE_RTM_OPT
1156 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1157 int rtm_state = method()->method_data()->rtm_state();
1158 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1159 // Don't generate RTM lock eliding code.
1160 set_rtm_state(NoRTM);
1161 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1162 // Generate RTM lock eliding code without abort ratio calculation code.
1163 set_rtm_state(UseRTM);
1164 } else if (UseRTMDeopt) {
1165 // Generate RTM lock eliding code and include abort ratio calculation
1166 // code if UseRTMDeopt is on.
1167 set_rtm_state(ProfileRTM);
1168 }
1169 }
1170 #endif
1171 if (debug_info()->recording_non_safepoints()) {
1172 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1173 (comp_arena(), 8, 0, NULL));
1174 set_default_node_notes(Node_Notes::make(this));
1175 }
1176
1177 // // -- Initialize types before each compile --
1178 // // Update cached type information
1179 // if( _method && _method->constants() )
1180 // Type::update_loaded_types(_method, _method->constants());
1181
1182 // Init alias_type map.
1183 if (!_do_escape_analysis && aliaslevel == 3)
1184 aliaslevel = 2; // No unique types without escape analysis
1185 _AliasLevel = aliaslevel;
1186 const int grow_ats = 16;
1187 _max_alias_types = grow_ats;
1188 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1189 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1190 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1191 {
1192 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1193 }
1194 // Initialize the first few types.
1195 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1196 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1197 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1198 _num_alias_types = AliasIdxRaw+1;
1199 // Zero out the alias type cache.
1200 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1201 // A NULL adr_type hits in the cache right away. Preload the right answer.
1202 probe_alias_cache(NULL)->_index = AliasIdxTop;
1203
1204 _intrinsics = NULL;
1205 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1206 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1207 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1208 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1209 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1210 register_library_intrinsics();
1211 }
1212
1213 //---------------------------init_start----------------------------------------
1214 // Install the StartNode on this compile object.
1215 void Compile::init_start(StartNode* s) {
1216 if (failing())
1217 return; // already failing
1218 assert(s == start(), "");
1219 }
1220
1221 /**
1222 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1223 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1224 * the ideal graph.
1225 */
1226 StartNode* Compile::start() const {
1227 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1228 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1229 Node* start = root()->fast_out(i);
1230 if (start->is_Start()) {
1231 return start->as_Start();
1232 }
1233 }
1234 fatal("Did not find Start node!");
1235 return NULL;
1236 }
1237
1238 //-------------------------------immutable_memory-------------------------------------
1239 // Access immutable memory
1240 Node* Compile::immutable_memory() {
1241 if (_immutable_memory != NULL) {
1242 return _immutable_memory;
1243 }
1244 StartNode* s = start();
1245 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1246 Node *p = s->fast_out(i);
1247 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1248 _immutable_memory = p;
1249 return _immutable_memory;
1250 }
1251 }
1252 ShouldNotReachHere();
1253 return NULL;
1254 }
1255
1256 //----------------------set_cached_top_node------------------------------------
1257 // Install the cached top node, and make sure Node::is_top works correctly.
1258 void Compile::set_cached_top_node(Node* tn) {
1259 if (tn != NULL) verify_top(tn);
1260 Node* old_top = _top;
1261 _top = tn;
1262 // Calling Node::setup_is_top allows the nodes the chance to adjust
1263 // their _out arrays.
1264 if (_top != NULL) _top->setup_is_top();
1265 if (old_top != NULL) old_top->setup_is_top();
1266 assert(_top == NULL || top()->is_top(), "");
1267 }
1268
1269 #ifdef ASSERT
1270 uint Compile::count_live_nodes_by_graph_walk() {
1271 Unique_Node_List useful(comp_arena());
1272 // Get useful node list by walking the graph.
1273 identify_useful_nodes(useful);
1274 return useful.size();
1275 }
1276
1277 void Compile::print_missing_nodes() {
1278
1279 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1280 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1281 return;
1282 }
1283
1284 // This is an expensive function. It is executed only when the user
1285 // specifies VerifyIdealNodeCount option or otherwise knows the
1286 // additional work that needs to be done to identify reachable nodes
1287 // by walking the flow graph and find the missing ones using
1288 // _dead_node_list.
1289
1290 Unique_Node_List useful(comp_arena());
1291 // Get useful node list by walking the graph.
1292 identify_useful_nodes(useful);
1293
1294 uint l_nodes = C->live_nodes();
1295 uint l_nodes_by_walk = useful.size();
1296
1297 if (l_nodes != l_nodes_by_walk) {
1298 if (_log != NULL) {
1299 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1300 _log->stamp();
1301 _log->end_head();
1302 }
1303 VectorSet& useful_member_set = useful.member_set();
1304 int last_idx = l_nodes_by_walk;
1305 for (int i = 0; i < last_idx; i++) {
1306 if (useful_member_set.test(i)) {
1307 if (_dead_node_list.test(i)) {
1308 if (_log != NULL) {
1309 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1310 }
1311 if (PrintIdealNodeCount) {
1312 // Print the log message to tty
1313 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1314 useful.at(i)->dump();
1315 }
1316 }
1317 }
1318 else if (! _dead_node_list.test(i)) {
1319 if (_log != NULL) {
1320 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1321 }
1322 if (PrintIdealNodeCount) {
1323 // Print the log message to tty
1324 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1325 }
1326 }
1327 }
1328 if (_log != NULL) {
1329 _log->tail("mismatched_nodes");
1330 }
1331 }
1332 }
1333 void Compile::record_modified_node(Node* n) {
1334 if (_modified_nodes != NULL && !_inlining_incrementally &&
1335 n->outcnt() != 0 && !n->is_Con()) {
1336 _modified_nodes->push(n);
1337 }
1338 }
1339
1340 void Compile::remove_modified_node(Node* n) {
1341 if (_modified_nodes != NULL) {
1342 _modified_nodes->remove(n);
1343 }
1344 }
1345 #endif
1346
1347 #ifndef PRODUCT
1348 void Compile::verify_top(Node* tn) const {
1349 if (tn != NULL) {
1350 assert(tn->is_Con(), "top node must be a constant");
1351 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1352 assert(tn->in(0) != NULL, "must have live top node");
1353 }
1354 }
1355 #endif
1356
1357
1358 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1359
1360 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1361 guarantee(arr != NULL, "");
1362 int num_blocks = arr->length();
1363 if (grow_by < num_blocks) grow_by = num_blocks;
1364 int num_notes = grow_by * _node_notes_block_size;
1365 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1366 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1367 while (num_notes > 0) {
1368 arr->append(notes);
1369 notes += _node_notes_block_size;
1370 num_notes -= _node_notes_block_size;
1371 }
1372 assert(num_notes == 0, "exact multiple, please");
1373 }
1374
1375 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1376 if (source == NULL || dest == NULL) return false;
1377
1378 if (dest->is_Con())
1379 return false; // Do not push debug info onto constants.
1380
1381 #ifdef ASSERT
1382 // Leave a bread crumb trail pointing to the original node:
1383 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1384 dest->set_debug_orig(source);
1385 }
1386 #endif
1387
1388 if (node_note_array() == NULL)
1389 return false; // Not collecting any notes now.
1390
1391 // This is a copy onto a pre-existing node, which may already have notes.
1392 // If both nodes have notes, do not overwrite any pre-existing notes.
1393 Node_Notes* source_notes = node_notes_at(source->_idx);
1394 if (source_notes == NULL || source_notes->is_clear()) return false;
1395 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1396 if (dest_notes == NULL || dest_notes->is_clear()) {
1397 return set_node_notes_at(dest->_idx, source_notes);
1398 }
1399
1400 Node_Notes merged_notes = (*source_notes);
1401 // The order of operations here ensures that dest notes will win...
1402 merged_notes.update_from(dest_notes);
1403 return set_node_notes_at(dest->_idx, &merged_notes);
1404 }
1405
1406
1407 //--------------------------allow_range_check_smearing-------------------------
1408 // Gating condition for coalescing similar range checks.
1409 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1410 // single covering check that is at least as strong as any of them.
1411 // If the optimization succeeds, the simplified (strengthened) range check
1412 // will always succeed. If it fails, we will deopt, and then give up
1413 // on the optimization.
1414 bool Compile::allow_range_check_smearing() const {
1415 // If this method has already thrown a range-check,
1416 // assume it was because we already tried range smearing
1417 // and it failed.
1418 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1419 return !already_trapped;
1420 }
1421
1422
1423 //------------------------------flatten_alias_type-----------------------------
1424 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1425 int offset = tj->offset();
1426 TypePtr::PTR ptr = tj->ptr();
1427
1428 // Known instance (scalarizable allocation) alias only with itself.
1429 bool is_known_inst = tj->isa_oopptr() != NULL &&
1430 tj->is_oopptr()->is_known_instance();
1431
1432 // Process weird unsafe references.
1433 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1434 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1435 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1436 tj = TypeOopPtr::BOTTOM;
1437 ptr = tj->ptr();
1438 offset = tj->offset();
1439 }
1440
1441 // Array pointers need some flattening
1442 const TypeAryPtr *ta = tj->isa_aryptr();
1443 if (ta && ta->is_stable()) {
1444 // Erase stability property for alias analysis.
1445 tj = ta = ta->cast_to_stable(false);
1446 }
1447 if( ta && is_known_inst ) {
1448 if ( offset != Type::OffsetBot &&
1449 offset > arrayOopDesc::length_offset_in_bytes() ) {
1450 offset = Type::OffsetBot; // Flatten constant access into array body only
1451 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1452 }
1453 } else if( ta && _AliasLevel >= 2 ) {
1454 // For arrays indexed by constant indices, we flatten the alias
1455 // space to include all of the array body. Only the header, klass
1456 // and array length can be accessed un-aliased.
1457 if( offset != Type::OffsetBot ) {
1458 if( ta->const_oop() ) { // MethodData* or Method*
1459 offset = Type::OffsetBot; // Flatten constant access into array body
1460 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1461 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1462 // range is OK as-is.
1463 tj = ta = TypeAryPtr::RANGE;
1464 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1465 tj = TypeInstPtr::KLASS; // all klass loads look alike
1466 ta = TypeAryPtr::RANGE; // generic ignored junk
1467 ptr = TypePtr::BotPTR;
1468 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1469 tj = TypeInstPtr::MARK;
1470 ta = TypeAryPtr::RANGE; // generic ignored junk
1471 ptr = TypePtr::BotPTR;
1472 } else { // Random constant offset into array body
1473 offset = Type::OffsetBot; // Flatten constant access into array body
1474 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1475 }
1476 }
1477 // Arrays of fixed size alias with arrays of unknown size.
1478 if (ta->size() != TypeInt::POS) {
1479 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1480 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1481 }
1482 // Arrays of known objects become arrays of unknown objects.
1483 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1484 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1485 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1486 }
1487 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1488 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1489 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1490 }
1491 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1492 // cannot be distinguished by bytecode alone.
1493 if (ta->elem() == TypeInt::BOOL) {
1494 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1495 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1496 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1497 }
1498 // During the 2nd round of IterGVN, NotNull castings are removed.
1499 // Make sure the Bottom and NotNull variants alias the same.
1500 // Also, make sure exact and non-exact variants alias the same.
1501 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1502 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1503 }
1504 }
1505
1506 // Oop pointers need some flattening
1507 const TypeInstPtr *to = tj->isa_instptr();
1508 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1509 ciInstanceKlass *k = to->klass()->as_instance_klass();
1510 if( ptr == TypePtr::Constant ) {
1511 if (to->klass() != ciEnv::current()->Class_klass() ||
1512 offset < k->size_helper() * wordSize) {
1513 // No constant oop pointers (such as Strings); they alias with
1514 // unknown strings.
1515 assert(!is_known_inst, "not scalarizable allocation");
1516 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1517 }
1518 } else if( is_known_inst ) {
1519 tj = to; // Keep NotNull and klass_is_exact for instance type
1520 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1521 // During the 2nd round of IterGVN, NotNull castings are removed.
1522 // Make sure the Bottom and NotNull variants alias the same.
1523 // Also, make sure exact and non-exact variants alias the same.
1524 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1525 }
1526 if (to->speculative() != NULL) {
1527 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1528 }
1529 // Canonicalize the holder of this field
1530 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1531 // First handle header references such as a LoadKlassNode, even if the
1532 // object's klass is unloaded at compile time (4965979).
1533 if (!is_known_inst) { // Do it only for non-instance types
1534 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1535 }
1536 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1537 // Static fields are in the space above the normal instance
1538 // fields in the java.lang.Class instance.
1539 if (to->klass() != ciEnv::current()->Class_klass()) {
1540 to = NULL;
1541 tj = TypeOopPtr::BOTTOM;
1542 offset = tj->offset();
1543 }
1544 } else {
1545 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1546 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1547 if( is_known_inst ) {
1548 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1549 } else {
1550 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1551 }
1552 }
1553 }
1554 }
1555
1556 // Klass pointers to object array klasses need some flattening
1557 const TypeKlassPtr *tk = tj->isa_klassptr();
1558 if( tk ) {
1559 // If we are referencing a field within a Klass, we need
1560 // to assume the worst case of an Object. Both exact and
1561 // inexact types must flatten to the same alias class so
1562 // use NotNull as the PTR.
1563 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1564
1565 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1566 TypeKlassPtr::OBJECT->klass(),
1567 offset);
1568 }
1569
1570 ciKlass* klass = tk->klass();
1571 if( klass->is_obj_array_klass() ) {
1572 ciKlass* k = TypeAryPtr::OOPS->klass();
1573 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1574 k = TypeInstPtr::BOTTOM->klass();
1575 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1576 }
1577
1578 // Check for precise loads from the primary supertype array and force them
1579 // to the supertype cache alias index. Check for generic array loads from
1580 // the primary supertype array and also force them to the supertype cache
1581 // alias index. Since the same load can reach both, we need to merge
1582 // these 2 disparate memories into the same alias class. Since the
1583 // primary supertype array is read-only, there's no chance of confusion
1584 // where we bypass an array load and an array store.
1585 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1586 if (offset == Type::OffsetBot ||
1587 (offset >= primary_supers_offset &&
1588 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1589 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1590 offset = in_bytes(Klass::secondary_super_cache_offset());
1591 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1592 }
1593 }
1594
1595 // Flatten all Raw pointers together.
1596 if (tj->base() == Type::RawPtr)
1597 tj = TypeRawPtr::BOTTOM;
1598
1599 if (tj->base() == Type::AnyPtr)
1600 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1601
1602 // Flatten all to bottom for now
1603 switch( _AliasLevel ) {
1604 case 0:
1605 tj = TypePtr::BOTTOM;
1606 break;
1607 case 1: // Flatten to: oop, static, field or array
1608 switch (tj->base()) {
1609 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1610 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1611 case Type::AryPtr: // do not distinguish arrays at all
1612 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1613 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1614 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1615 default: ShouldNotReachHere();
1616 }
1617 break;
1618 case 2: // No collapsing at level 2; keep all splits
1619 case 3: // No collapsing at level 3; keep all splits
1620 break;
1621 default:
1622 Unimplemented();
1623 }
1624
1625 offset = tj->offset();
1626 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1627
1628 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1629 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1630 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1631 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1632 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1633 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1634 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1635 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1636 assert( tj->ptr() != TypePtr::TopPTR &&
1637 tj->ptr() != TypePtr::AnyNull &&
1638 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1639 // assert( tj->ptr() != TypePtr::Constant ||
1640 // tj->base() == Type::RawPtr ||
1641 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1642
1643 return tj;
1644 }
1645
1646 void Compile::AliasType::Init(int i, const TypePtr* at) {
1647 _index = i;
1648 _adr_type = at;
1649 _field = NULL;
1650 _element = NULL;
1651 _is_rewritable = true; // default
1652 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1653 if (atoop != NULL && atoop->is_known_instance()) {
1654 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1655 _general_index = Compile::current()->get_alias_index(gt);
1656 } else {
1657 _general_index = 0;
1658 }
1659 }
1660
1661 BasicType Compile::AliasType::basic_type() const {
1662 if (element() != NULL) {
1663 const Type* element = adr_type()->is_aryptr()->elem();
1664 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1665 } if (field() != NULL) {
1666 return field()->layout_type();
1667 } else {
1668 return T_ILLEGAL; // unknown
1669 }
1670 }
1671
1672 //---------------------------------print_on------------------------------------
1673 #ifndef PRODUCT
1674 void Compile::AliasType::print_on(outputStream* st) {
1675 if (index() < 10)
1676 st->print("@ <%d> ", index());
1677 else st->print("@ <%d>", index());
1678 st->print(is_rewritable() ? " " : " RO");
1679 int offset = adr_type()->offset();
1680 if (offset == Type::OffsetBot)
1681 st->print(" +any");
1682 else st->print(" +%-3d", offset);
1683 st->print(" in ");
1684 adr_type()->dump_on(st);
1685 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1686 if (field() != NULL && tjp) {
1687 if (tjp->klass() != field()->holder() ||
1688 tjp->offset() != field()->offset_in_bytes()) {
1689 st->print(" != ");
1690 field()->print();
1691 st->print(" ***");
1692 }
1693 }
1694 }
1695
1696 void print_alias_types() {
1697 Compile* C = Compile::current();
1698 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1699 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1700 C->alias_type(idx)->print_on(tty);
1701 tty->cr();
1702 }
1703 }
1704 #endif
1705
1706
1707 //----------------------------probe_alias_cache--------------------------------
1708 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1709 intptr_t key = (intptr_t) adr_type;
1710 key ^= key >> logAliasCacheSize;
1711 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1712 }
1713
1714
1715 //-----------------------------grow_alias_types--------------------------------
1716 void Compile::grow_alias_types() {
1717 const int old_ats = _max_alias_types; // how many before?
1718 const int new_ats = old_ats; // how many more?
1719 const int grow_ats = old_ats+new_ats; // how many now?
1720 _max_alias_types = grow_ats;
1721 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1722 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1723 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1724 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1725 }
1726
1727
1728 //--------------------------------find_alias_type------------------------------
1729 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1730 if (_AliasLevel == 0)
1731 return alias_type(AliasIdxBot);
1732
1733 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1734 if (ace->_adr_type == adr_type) {
1735 return alias_type(ace->_index);
1736 }
1737
1738 // Handle special cases.
1739 if (adr_type == NULL) return alias_type(AliasIdxTop);
1740 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1741
1742 // Do it the slow way.
1743 const TypePtr* flat = flatten_alias_type(adr_type);
1744
1745 #ifdef ASSERT
1746 {
1747 ResourceMark rm;
1748 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1749 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1750 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1751 Type::str(adr_type));
1752 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1753 const TypeOopPtr* foop = flat->is_oopptr();
1754 // Scalarizable allocations have exact klass always.
1755 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1756 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1757 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1758 Type::str(foop), Type::str(xoop));
1759 }
1760 }
1761 #endif
1762
1763 int idx = AliasIdxTop;
1764 for (int i = 0; i < num_alias_types(); i++) {
1765 if (alias_type(i)->adr_type() == flat) {
1766 idx = i;
1767 break;
1768 }
1769 }
1770
1771 if (idx == AliasIdxTop) {
1772 if (no_create) return NULL;
1773 // Grow the array if necessary.
1774 if (_num_alias_types == _max_alias_types) grow_alias_types();
1775 // Add a new alias type.
1776 idx = _num_alias_types++;
1777 _alias_types[idx]->Init(idx, flat);
1778 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1779 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1780 if (flat->isa_instptr()) {
1781 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1782 && flat->is_instptr()->klass() == env()->Class_klass())
1783 alias_type(idx)->set_rewritable(false);
1784 }
1785 if (flat->isa_aryptr()) {
1786 #ifdef ASSERT
1787 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1788 // (T_BYTE has the weakest alignment and size restrictions...)
1789 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1790 #endif
1791 if (flat->offset() == TypePtr::OffsetBot) {
1792 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1793 }
1794 }
1795 if (flat->isa_klassptr()) {
1796 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1797 alias_type(idx)->set_rewritable(false);
1798 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1799 alias_type(idx)->set_rewritable(false);
1800 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1801 alias_type(idx)->set_rewritable(false);
1802 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1803 alias_type(idx)->set_rewritable(false);
1804 }
1805 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1806 // but the base pointer type is not distinctive enough to identify
1807 // references into JavaThread.)
1808
1809 // Check for final fields.
1810 const TypeInstPtr* tinst = flat->isa_instptr();
1811 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1812 ciField* field;
1813 if (tinst->const_oop() != NULL &&
1814 tinst->klass() == ciEnv::current()->Class_klass() &&
1815 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1816 // static field
1817 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1818 field = k->get_field_by_offset(tinst->offset(), true);
1819 } else {
1820 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1821 field = k->get_field_by_offset(tinst->offset(), false);
1822 }
1823 assert(field == NULL ||
1824 original_field == NULL ||
1825 (field->holder() == original_field->holder() &&
1826 field->offset() == original_field->offset() &&
1827 field->is_static() == original_field->is_static()), "wrong field?");
1828 // Set field() and is_rewritable() attributes.
1829 if (field != NULL) alias_type(idx)->set_field(field);
1830 }
1831 }
1832
1833 // Fill the cache for next time.
1834 ace->_adr_type = adr_type;
1835 ace->_index = idx;
1836 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1837
1838 // Might as well try to fill the cache for the flattened version, too.
1839 AliasCacheEntry* face = probe_alias_cache(flat);
1840 if (face->_adr_type == NULL) {
1841 face->_adr_type = flat;
1842 face->_index = idx;
1843 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1844 }
1845
1846 return alias_type(idx);
1847 }
1848
1849
1850 Compile::AliasType* Compile::alias_type(ciField* field) {
1851 const TypeOopPtr* t;
1852 if (field->is_static())
1853 t = TypeInstPtr::make(field->holder()->java_mirror());
1854 else
1855 t = TypeOopPtr::make_from_klass_raw(field->holder());
1856 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1857 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1858 return atp;
1859 }
1860
1861
1862 //------------------------------have_alias_type--------------------------------
1863 bool Compile::have_alias_type(const TypePtr* adr_type) {
1864 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1865 if (ace->_adr_type == adr_type) {
1866 return true;
1867 }
1868
1869 // Handle special cases.
1870 if (adr_type == NULL) return true;
1871 if (adr_type == TypePtr::BOTTOM) return true;
1872
1873 return find_alias_type(adr_type, true, NULL) != NULL;
1874 }
1875
1876 //-----------------------------must_alias--------------------------------------
1877 // True if all values of the given address type are in the given alias category.
1878 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1879 if (alias_idx == AliasIdxBot) return true; // the universal category
1880 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1881 if (alias_idx == AliasIdxTop) return false; // the empty category
1882 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1883
1884 // the only remaining possible overlap is identity
1885 int adr_idx = get_alias_index(adr_type);
1886 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1887 assert(adr_idx == alias_idx ||
1888 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1889 && adr_type != TypeOopPtr::BOTTOM),
1890 "should not be testing for overlap with an unsafe pointer");
1891 return adr_idx == alias_idx;
1892 }
1893
1894 //------------------------------can_alias--------------------------------------
1895 // True if any values of the given address type are in the given alias category.
1896 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1897 if (alias_idx == AliasIdxTop) return false; // the empty category
1898 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1899 if (alias_idx == AliasIdxBot) return true; // the universal category
1900 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1901
1902 // the only remaining possible overlap is identity
1903 int adr_idx = get_alias_index(adr_type);
1904 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1905 return adr_idx == alias_idx;
1906 }
1907
1908
1909
1910 //---------------------------pop_warm_call-------------------------------------
1911 WarmCallInfo* Compile::pop_warm_call() {
1912 WarmCallInfo* wci = _warm_calls;
1913 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1914 return wci;
1915 }
1916
1917 //----------------------------Inline_Warm--------------------------------------
1918 int Compile::Inline_Warm() {
1919 // If there is room, try to inline some more warm call sites.
1920 // %%% Do a graph index compaction pass when we think we're out of space?
1921 if (!InlineWarmCalls) return 0;
1922
1923 int calls_made_hot = 0;
1924 int room_to_grow = NodeCountInliningCutoff - unique();
1925 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1926 int amount_grown = 0;
1927 WarmCallInfo* call;
1928 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1929 int est_size = (int)call->size();
1930 if (est_size > (room_to_grow - amount_grown)) {
1931 // This one won't fit anyway. Get rid of it.
1932 call->make_cold();
1933 continue;
1934 }
1935 call->make_hot();
1936 calls_made_hot++;
1937 amount_grown += est_size;
1938 amount_to_grow -= est_size;
1939 }
1940
1941 if (calls_made_hot > 0) set_major_progress();
1942 return calls_made_hot;
1943 }
1944
1945
1946 //----------------------------Finish_Warm--------------------------------------
1947 void Compile::Finish_Warm() {
1948 if (!InlineWarmCalls) return;
1949 if (failing()) return;
1950 if (warm_calls() == NULL) return;
1951
1952 // Clean up loose ends, if we are out of space for inlining.
1953 WarmCallInfo* call;
1954 while ((call = pop_warm_call()) != NULL) {
1955 call->make_cold();
1956 }
1957 }
1958
1959 //---------------------cleanup_loop_predicates-----------------------
1960 // Remove the opaque nodes that protect the predicates so that all unused
1961 // checks and uncommon_traps will be eliminated from the ideal graph
1962 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1963 if (predicate_count()==0) return;
1964 for (int i = predicate_count(); i > 0; i--) {
1965 Node * n = predicate_opaque1_node(i-1);
1966 assert(n->Opcode() == Op_Opaque1, "must be");
1967 igvn.replace_node(n, n->in(1));
1968 }
1969 assert(predicate_count()==0, "should be clean!");
1970 }
1971
1972 void Compile::add_range_check_cast(Node* n) {
1973 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1974 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1975 _range_check_casts->append(n);
1976 }
1977
1978 // Remove all range check dependent CastIINodes.
1979 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1980 for (int i = range_check_cast_count(); i > 0; i--) {
1981 Node* cast = range_check_cast_node(i-1);
1982 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1983 igvn.replace_node(cast, cast->in(1));
1984 }
1985 assert(range_check_cast_count() == 0, "should be empty");
1986 }
1987
1988 void Compile::add_opaque4_node(Node* n) {
1989 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1990 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1991 _opaque4_nodes->append(n);
1992 }
1993
1994 // Remove all Opaque4 nodes.
1995 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
1996 for (int i = opaque4_count(); i > 0; i--) {
1997 Node* opaq = opaque4_node(i-1);
1998 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
1999 igvn.replace_node(opaq, opaq->in(2));
2000 }
2001 assert(opaque4_count() == 0, "should be empty");
2002 }
2003
2004 // StringOpts and late inlining of string methods
2005 void Compile::inline_string_calls(bool parse_time) {
2006 {
2007 // remove useless nodes to make the usage analysis simpler
2008 ResourceMark rm;
2009 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2010 }
2011
2012 {
2013 ResourceMark rm;
2014 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2015 PhaseStringOpts pso(initial_gvn(), for_igvn());
2016 print_method(PHASE_AFTER_STRINGOPTS, 3);
2017 }
2018
2019 // now inline anything that we skipped the first time around
2020 if (!parse_time) {
2021 _late_inlines_pos = _late_inlines.length();
2022 }
2023
2024 while (_string_late_inlines.length() > 0) {
2025 CallGenerator* cg = _string_late_inlines.pop();
2026 cg->do_late_inline();
2027 if (failing()) return;
2028 }
2029 _string_late_inlines.trunc_to(0);
2030 }
2031
2032 // Late inlining of boxing methods
2033 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2034 if (_boxing_late_inlines.length() > 0) {
2035 assert(has_boxed_value(), "inconsistent");
2036
2037 PhaseGVN* gvn = initial_gvn();
2038 set_inlining_incrementally(true);
2039
2040 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2041 for_igvn()->clear();
2042 gvn->replace_with(&igvn);
2043
2044 _late_inlines_pos = _late_inlines.length();
2045
2046 while (_boxing_late_inlines.length() > 0) {
2047 CallGenerator* cg = _boxing_late_inlines.pop();
2048 cg->do_late_inline();
2049 if (failing()) return;
2050 }
2051 _boxing_late_inlines.trunc_to(0);
2052
2053 {
2054 ResourceMark rm;
2055 PhaseRemoveUseless pru(gvn, for_igvn());
2056 }
2057
2058 igvn = PhaseIterGVN(gvn);
2059 igvn.optimize();
2060
2061 set_inlining_progress(false);
2062 set_inlining_incrementally(false);
2063 }
2064 }
2065
2066 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2067 assert(IncrementalInline, "incremental inlining should be on");
2068 PhaseGVN* gvn = initial_gvn();
2069
2070 set_inlining_progress(false);
2071 for_igvn()->clear();
2072 gvn->replace_with(&igvn);
2073
2074 {
2075 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2076 int i = 0;
2077 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2078 CallGenerator* cg = _late_inlines.at(i);
2079 _late_inlines_pos = i+1;
2080 cg->do_late_inline();
2081 if (failing()) return;
2082 }
2083 int j = 0;
2084 for (; i < _late_inlines.length(); i++, j++) {
2085 _late_inlines.at_put(j, _late_inlines.at(i));
2086 }
2087 _late_inlines.trunc_to(j);
2088 }
2089
2090 {
2091 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2092 ResourceMark rm;
2093 PhaseRemoveUseless pru(gvn, for_igvn());
2094 }
2095
2096 {
2097 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2098 igvn = PhaseIterGVN(gvn);
2099 }
2100 }
2101
2102 // Perform incremental inlining until bound on number of live nodes is reached
2103 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2104 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2105
2106 PhaseGVN* gvn = initial_gvn();
2107
2108 set_inlining_incrementally(true);
2109 set_inlining_progress(true);
2110 uint low_live_nodes = 0;
2111
2112 while(inlining_progress() && _late_inlines.length() > 0) {
2113
2114 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2115 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2116 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2117 // PhaseIdealLoop is expensive so we only try it once we are
2118 // out of live nodes and we only try it again if the previous
2119 // helped got the number of nodes down significantly
2120 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2121 if (failing()) return;
2122 low_live_nodes = live_nodes();
2123 _major_progress = true;
2124 }
2125
2126 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2127 break;
2128 }
2129 }
2130
2131 inline_incrementally_one(igvn);
2132
2133 if (failing()) return;
2134
2135 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2136
2137 {
2138 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2139 igvn.optimize();
2140 }
2141
2142 if (failing()) return;
2143 }
2144
2145 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2146
2147 if (_string_late_inlines.length() > 0) {
2148 assert(has_stringbuilder(), "inconsistent");
2149 for_igvn()->clear();
2150 initial_gvn()->replace_with(&igvn);
2151
2152 inline_string_calls(false);
2153
2154 if (failing()) return;
2155
2156 {
2157 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2158 ResourceMark rm;
2159 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2160 }
2161
2162 {
2163 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2164 igvn = PhaseIterGVN(gvn);
2165 igvn.optimize();
2166 }
2167 }
2168
2169 set_inlining_incrementally(false);
2170 }
2171
2172
2173 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2174 if(loop_opts_cnt > 0) {
2175 debug_only( int cnt = 0; );
2176 while(major_progress() && (loop_opts_cnt > 0)) {
2177 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2178 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2179 PhaseIdealLoop ideal_loop(igvn, mode);
2180 loop_opts_cnt--;
2181 if (failing()) return false;
2182 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2183 }
2184 }
2185 return true;
2186 }
2187
2188 //------------------------------Optimize---------------------------------------
2189 // Given a graph, optimize it.
2190 void Compile::Optimize() {
2191 TracePhase tp("optimizer", &timers[_t_optimizer]);
2192
2193 #ifndef PRODUCT
2194 if (_directive->BreakAtCompileOption) {
2195 BREAKPOINT;
2196 }
2197
2198 #endif
2199
2200 #ifdef ASSERT
2201 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2202 bs->verify_gc_barriers(true);
2203 #endif
2204
2205 ResourceMark rm;
2206 int loop_opts_cnt;
2207
2208 print_inlining_reinit();
2209
2210 NOT_PRODUCT( verify_graph_edges(); )
2211
2212 print_method(PHASE_AFTER_PARSING);
2213
2214 {
2215 // Iterative Global Value Numbering, including ideal transforms
2216 // Initialize IterGVN with types and values from parse-time GVN
2217 PhaseIterGVN igvn(initial_gvn());
2218 #ifdef ASSERT
2219 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2220 #endif
2221 {
2222 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2223 igvn.optimize();
2224 }
2225
2226 if (failing()) return;
2227
2228 print_method(PHASE_ITER_GVN1, 2);
2229
2230 inline_incrementally(igvn);
2231
2232 print_method(PHASE_INCREMENTAL_INLINE, 2);
2233
2234 if (failing()) return;
2235
2236 if (eliminate_boxing()) {
2237 // Inline valueOf() methods now.
2238 inline_boxing_calls(igvn);
2239
2240 if (AlwaysIncrementalInline) {
2241 inline_incrementally(igvn);
2242 }
2243 if (failing()) return;
2244
2245 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2246 }
2247
2248 // Remove the speculative part of types and clean up the graph from
2249 // the extra CastPP nodes whose only purpose is to carry them. Do
2250 // that early so that optimizations are not disrupted by the extra
2251 // CastPP nodes.
2252 remove_speculative_types(igvn);
2253
2254 // No more new expensive nodes will be added to the list from here
2255 // so keep only the actual candidates for optimizations.
2256 cleanup_expensive_nodes(igvn);
2257
2258 if (has_vbox_nodes()) {
2259 TracePhase tp("expandVectorBoxes", &timers[_t_expandVectorBoxes]);
2260
2261 set_inlining_incrementally(true); // FIXME another way to signal GraphKit it's post-parsing phase?
2262
2263 for_igvn()->clear();
2264 initial_gvn()->replace_with(&igvn);
2265
2266 expand_vunbox_nodes();
2267 scalarize_vbox_nodes();
2268
2269 inline_vector_reboxing_calls();
2270
2271 if (failing()) return;
2272 {
2273 ResourceMark rm;
2274 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2275 }
2276 igvn = PhaseIterGVN(initial_gvn());
2277 igvn.optimize();
2278
2279 print_method(PHASE_ITER_GVN_BEFORE_EA, 2);
2280
2281 if (failing()) return;
2282
2283 ////////////////////////////////////////////////////
2284
2285 for_igvn()->clear();
2286 initial_gvn()->replace_with(&igvn);
2287
2288 expand_vbox_nodes();
2289 eliminate_vbox_alloc_nodes();
2290
2291 if (failing()) return;
2292 {
2293 ResourceMark rm;
2294 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2295 }
2296 igvn = PhaseIterGVN(initial_gvn());
2297 igvn.optimize();
2298 if (failing()) return;
2299
2300 print_method(PHASE_ITER_GVN_BEFORE_EA, 2);
2301
2302 set_inlining_incrementally(false); // FIXME another way to signal GraphKit it's post-parsing phase?
2303 }
2304
2305 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2306 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2307 initial_gvn()->replace_with(&igvn);
2308 for_igvn()->clear();
2309 Unique_Node_List new_worklist(C->comp_arena());
2310 {
2311 ResourceMark rm;
2312 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2313 }
2314 set_for_igvn(&new_worklist);
2315 igvn = PhaseIterGVN(initial_gvn());
2316 igvn.optimize();
2317 }
2318
2319 // FIXME for_igvn() is corrupted from here: new_worklist which is set_for_ignv() was allocated on stack.
2320
2321 // Perform escape analysis
2322 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2323 if (has_loops()) {
2324 // Cleanup graph (remove dead nodes).
2325 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2326 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2327 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2328 if (failing()) return;
2329 }
2330 ConnectionGraph::do_analysis(this, &igvn);
2331
2332 if (failing()) return;
2333
2334 // Optimize out fields loads from scalar replaceable allocations.
2335 igvn.optimize();
2336 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2337
2338 if (failing()) return;
2339
2340 if (congraph() != NULL && macro_count() > 0) {
2341 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2342 PhaseMacroExpand mexp(igvn);
2343 mexp.eliminate_macro_nodes();
2344 igvn.set_delay_transform(false);
2345
2346 igvn.optimize();
2347 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2348
2349 if (failing()) return;
2350 }
2351 }
2352
2353 // Loop transforms on the ideal graph. Range Check Elimination,
2354 // peeling, unrolling, etc.
2355
2356 // Set loop opts counter
2357 loop_opts_cnt = num_loop_opts();
2358 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2359 {
2360 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2361 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2362 loop_opts_cnt--;
2363 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2364 if (failing()) return;
2365 }
2366 // Loop opts pass if partial peeling occurred in previous pass
2367 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2368 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2369 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2370 loop_opts_cnt--;
2371 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2372 if (failing()) return;
2373 }
2374 // Loop opts pass for loop-unrolling before CCP
2375 if(major_progress() && (loop_opts_cnt > 0)) {
2376 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2377 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2378 loop_opts_cnt--;
2379 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2380 }
2381 if (!failing()) {
2382 // Verify that last round of loop opts produced a valid graph
2383 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2384 PhaseIdealLoop::verify(igvn);
2385 }
2386 }
2387 if (failing()) return;
2388
2389 // Conditional Constant Propagation;
2390 PhaseCCP ccp( &igvn );
2391 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2392 {
2393 TracePhase tp("ccp", &timers[_t_ccp]);
2394 ccp.do_transform();
2395 }
2396 print_method(PHASE_CPP1, 2);
2397
2398 assert( true, "Break here to ccp.dump_old2new_map()");
2399
2400 // Iterative Global Value Numbering, including ideal transforms
2401 {
2402 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2403 igvn = ccp;
2404 igvn.optimize();
2405 }
2406
2407 print_method(PHASE_ITER_GVN2, 2);
2408
2409 if (failing()) return;
2410
2411 // Loop transforms on the ideal graph. Range Check Elimination,
2412 // peeling, unrolling, etc.
2413 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2414 return;
2415 }
2416
2417 #if INCLUDE_ZGC
2418 if (UseZGC) {
2419 ZBarrierSetC2::find_dominating_barriers(igvn);
2420 }
2421 #endif
2422
2423 if (failing()) return;
2424
2425 // Ensure that major progress is now clear
2426 C->clear_major_progress();
2427
2428 {
2429 // Verify that all previous optimizations produced a valid graph
2430 // at least to this point, even if no loop optimizations were done.
2431 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2432 PhaseIdealLoop::verify(igvn);
2433 }
2434
2435 if (range_check_cast_count() > 0) {
2436 // No more loop optimizations. Remove all range check dependent CastIINodes.
2437 C->remove_range_check_casts(igvn);
2438 igvn.optimize();
2439 }
2440
2441 #ifdef ASSERT
2442 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2443 bs->verify_gc_barriers(false);
2444 #endif
2445
2446 {
2447 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2448 PhaseMacroExpand mex(igvn);
2449 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2450 if (mex.expand_macro_nodes()) {
2451 assert(failing(), "must bail out w/ explicit message");
2452 return;
2453 }
2454 }
2455
2456 if (opaque4_count() > 0) {
2457 C->remove_opaque4_nodes(igvn);
2458 igvn.optimize();
2459 }
2460
2461 DEBUG_ONLY( _modified_nodes = NULL; )
2462 } // (End scope of igvn; run destructor if necessary for asserts.)
2463
2464 process_print_inlining();
2465 // A method with only infinite loops has no edges entering loops from root
2466 {
2467 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2468 if (final_graph_reshaping()) {
2469 assert(failing(), "must bail out w/ explicit message");
2470 return;
2471 }
2472 }
2473
2474 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2475 }
2476
2477 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
2478 ResourceMark rm;
2479 stringStream ss;
2480 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
2481 if (n != NULL) {
2482 #ifndef PRODUCT
2483 ss.print(": %s %d", n->Name(), n->_idx);
2484 #else
2485 ss.print(": %d %d", n->Opcode(), n->_idx);
2486 #endif // PRODUCT
2487 } else {
2488 ss.print_raw(": NULL");
2489 }
2490 C->print_method(cpt, ss.as_string(), level);
2491 }
2492
2493 void Compile::scalarize_vbox_nodes() {
2494 int macro_idx = C->macro_count() - 1;
2495 while(macro_idx >= 0) {
2496 Node * n = C->macro_node(macro_idx);
2497 assert(n->is_macro(), "only macro nodes expected here");
2498 if (n->Opcode() == Op_VectorBox) {
2499 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(n);
2500 scalarize_vbox_node(vbox);
2501 if (failing()) return;
2502 print_method(PHASE_SCALARIZE_VBOX, vbox);
2503 }
2504 if (C->failing()) return;
2505 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2506 }
2507 }
2508
2509 void Compile::expand_vbox_nodes() {
2510 int macro_idx = C->macro_count() - 1;
2511 while(macro_idx >= 0) {
2512 Node * n = C->macro_node(macro_idx);
2513 assert(n->is_macro(), "only macro nodes expected here");
2514 if (n->Opcode() == Op_VectorBox) {
2515 VectorBoxNode* vbox = static_cast<VectorBoxNode*>(n);
2516 expand_vbox_node(vbox);
2517 if (failing()) return;
2518 print_method(PHASE_EXPAND_VBOX, vbox);
2519 }
2520 if (C->failing()) return;
2521 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2522 }
2523 }
2524
2525 void Compile::expand_vunbox_nodes() {
2526 int macro_idx = C->macro_count() - 1;
2527 while(macro_idx >= 0) {
2528 Node * n = C->macro_node(macro_idx);
2529 assert(n->is_macro(), "only macro nodes expected here");
2530 if (n->Opcode() == Op_VectorUnbox) {
2531 VectorUnboxNode* vec_unbox = static_cast<VectorUnboxNode*>(n);
2532 expand_vunbox_node(vec_unbox);
2533 if (failing()) return;
2534 print_method(PHASE_EXPAND_VUNBOX, vec_unbox);
2535 }
2536 if (C->failing()) return;
2537 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2538 }
2539 }
2540
2541 void Compile::eliminate_vbox_alloc_nodes() {
2542 int macro_idx = C->macro_count() - 1;
2543 while(macro_idx >= 0) {
2544 Node * n = C->macro_node(macro_idx);
2545 assert(n->is_macro(), "only macro nodes expected here");
2546 if (n->Opcode() == Op_VectorBoxAllocate) {
2547 VectorBoxAllocateNode* vbox_alloc = static_cast<VectorBoxAllocateNode*>(n);
2548 eliminate_vbox_alloc_node(vbox_alloc);
2549 if (failing()) return;
2550 print_method(PHASE_ELIMINATE_VBOX_ALLOC, vbox_alloc);
2551 }
2552 if (C->failing()) return;
2553 macro_idx = MIN2(macro_idx - 1, C->macro_count() - 1);
2554 }
2555 }
2556
2557 void Compile::inline_vector_reboxing_calls() {
2558 if (_vector_reboxing_late_inlines.length() > 0) {
2559 PhaseGVN* gvn = initial_gvn();
2560
2561 _late_inlines_pos = _late_inlines.length();
2562 while (_vector_reboxing_late_inlines.length() > 0) {
2563 CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2564 cg->do_late_inline();
2565 if (failing()) return;
2566 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2567 }
2568 _vector_reboxing_late_inlines.trunc_to(0);
2569 }
2570 }
2571
2572 bool Compile::has_vbox_nodes() {
2573 if (_vector_reboxing_late_inlines.length() > 0) {
2574 return true;
2575 }
2576 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2577 Node * n = C->macro_node(macro_idx);
2578 assert(n->is_macro(), "only macro nodes expected here");
2579 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2580 return true;
2581 }
2582 }
2583 return false;
2584 }
2585
2586 static JVMState* clone_jvms(Compile* C, SafePointNode* sfpt) {
2587 JVMState* new_jvms = sfpt->jvms()->clone_shallow(C);
2588 uint size = sfpt->req();
2589 SafePointNode* map = new SafePointNode(size, new_jvms);
2590 for (uint i = 0; i < size; i++) {
2591 map->init_req(i, sfpt->in(i));
2592 }
2593 // Make sure the state is a MergeMem for parsing.
2594 // FIXME not needed?
2595 // if (!map->in(TypeFunc::Memory)->is_MergeMem()) {
2596 // Node* mem = MergeMemNode::make(map->in(TypeFunc::Memory));
2597 // gvn.set_type_bottom(mem);
2598 // map->set_req(TypeFunc::Memory, mem);
2599 // }
2600 new_jvms->set_map(map);
2601 return new_jvms;
2602 }
2603
2604 void Compile::scalarize_vbox_node(VectorBoxNode* vec_box) {
2605 Node* vec_value = vec_box->in(VectorBoxNode::Value);
2606 PhaseGVN& gvn = *initial_gvn();
2607
2608 // Process merged VBAs
2609
2610 Unique_Node_List calls(C->comp_arena());
2611 for (DUIterator_Fast imax, i = vec_box->fast_outs(imax); i < imax; i++) {
2612 Node* use = vec_box->fast_out(i);
2613 if (use->is_CallJava()) {
2614 CallJavaNode* call = use->as_CallJava();
2615 if (call->has_non_debug_use(vec_box) && vec_box->in(VectorBoxNode::Box)->is_Phi()) {
2616 calls.push(call);
2617 }
2618 }
2619 }
2620
2621 while (VectorAPIAggressiveReboxing && calls.size() > 0) {
2622 CallJavaNode* call = calls.pop()->as_CallJava();
2623 // Attach new VBA to the call and use it instead of Phi (VBA ... VBA).
2624
2625 JVMState* jvms = clone_jvms(C, call);
2626 GraphKit kit(jvms);
2627 PhaseGVN& gvn = kit.gvn();
2628
2629 // Adjust JVMS from post-call to pre-call state: put args on stack
2630 uint nargs = call->method()->arg_size();
2631 kit.ensure_stack(kit.sp() + nargs);
2632 for (uint i = TypeFunc::Parms; i < call->tf()->domain()->cnt(); i++) {
2633 kit.push(call->in(i));
2634 }
2635 jvms = kit.sync_jvms();
2636
2637 Node* new_vbox = NULL;
2638 {
2639 PreserveReexecuteState prs(&kit);
2640
2641 kit.jvms()->set_should_reexecute(true);
2642
2643 const TypeInstPtr* vbox_type = vec_box->box_type();
2644 const TypeVect* vect_type = vec_box->vec_type();
2645 Node* vect = vec_box->in(VectorBoxNode::Value);
2646
2647 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type);
2648 kit.set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true);
2649 kit.make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true, /*deoptimize=*/true);
2650 kit.set_i_o(gvn.transform( new ProjNode(alloc, TypeFunc::I_O) ));
2651 kit.set_all_memory(gvn.transform( new ProjNode(alloc, TypeFunc::Memory) ));
2652 Node* ret = gvn.transform(new ProjNode(alloc, TypeFunc::Parms));
2653
2654 new_vbox = gvn.transform(new VectorBoxNode(C, ret, vect, vbox_type, vect_type));
2655
2656 kit.replace_in_map(vec_box, new_vbox);
2657 }
2658
2659 kit.dec_sp(nargs);
2660 jvms = kit.sync_jvms();
2661
2662 bool found = false;
2663 int cnt = _vector_reboxing_late_inlines.length();
2664 for (int i = 0; i < cnt; i++) {
2665 CallGenerator* cg = _vector_reboxing_late_inlines.at(i);
2666 if (cg->call_node() == call) {
2667 ciMethod* m = cg->method();
2668 _vector_reboxing_late_inlines.remove(cg); // remove_at(i);
2669 CallGenerator* new_cg = CallGenerator::for_vector_reboxing_late_inline(
2670 m,
2671 CallGenerator::for_inline(m, m->interpreter_invocation_count()));
2672
2673 JVMState* new_jvms = new_cg->generate(jvms);
2674
2675 Node* new_res = kit.top();
2676 if (m->return_type()->basic_type() != T_VOID) {
2677 new_res = new_cg->call_node()->proj_out(TypeFunc::Parms);
2678 }
2679 kit.replace_call(call, new_res, /*do_replaced_nodes=*/true);
2680
2681 found = true;
2682 break;
2683 }
2684 }
2685
2686 if (!found) {
2687 // FIXME does it cover all cases?
2688 CallJavaNode* new_call = call->clone()->as_CallJava();
2689 new_call->set_req(TypeFunc::Control , kit.control());
2690 new_call->set_req(TypeFunc::I_O , kit.i_o());
2691 new_call->set_req(TypeFunc::Memory , kit.reset_memory());
2692 new_call->set_req(TypeFunc::FramePtr , kit.frameptr());
2693
2694 new_call->replace_edge(vec_box, new_vbox);
2695 new_call = gvn.transform(new_call)->as_CallJava();
2696
2697 C->gvn_replace_by(call, new_call);
2698 }
2699 call->disconnect_inputs(NULL, C);
2700 }
2701
2702 // Process debug uses at safepoints
2703 Unique_Node_List safepoints(C->comp_arena());
2704
2705 for (DUIterator_Fast imax, i = vec_box->fast_outs(imax); i < imax; i++) {
2706 Node* use = vec_box->fast_out(i);
2707 if (use->is_SafePoint()) {
2708 SafePointNode* sfpt = use->as_SafePoint();
2709 if (!sfpt->is_Call() || !sfpt->as_Call()->has_non_debug_use(vec_box)) {
2710 safepoints.push(sfpt);
2711 }
2712 }
2713 }
2714
2715 while (safepoints.size() > 0) {
2716 SafePointNode* sfpt = safepoints.pop()->as_SafePoint();
2717
2718 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
2719 Node* sobj = new SafePointScalarObjectNode(vec_box->box_type(),
2720 #ifdef ASSERT
2721 NULL,
2722 #endif // ASSERT
2723 first_ind, /*n_fields=*/1);
2724 sobj->init_req(0, C->root());
2725 sfpt->add_req(vec_value);
2726
2727 sobj = gvn.transform(sobj);
2728
2729 JVMState *jvms = sfpt->jvms();
2730
2731 jvms->set_endoff(sfpt->req());
2732 // Now make a pass over the debug information replacing any references
2733 // to the allocated object with "sobj"
2734 int start = jvms->debug_start();
2735 int end = jvms->debug_end();
2736 sfpt->replace_edges_in_range(vec_box, sobj, start, end);
2737 }
2738 }
2739
2740 void Compile::expand_vbox_node(VectorBoxNode* vec_box) {
2741 if (vec_box->outcnt() > 0) {
2742 Node* vbox = vec_box->in(VectorBoxNode::Box);
2743 Node* vect = vec_box->in(VectorBoxNode::Value);
2744 Node* result = expand_vbox_node_helper(vbox, vect, vec_box->box_type(), vec_box->vec_type());
2745 C->gvn_replace_by(vec_box, result);
2746 }
2747 C->remove_macro_node(vec_box);
2748 }
2749
2750 Node* Compile::expand_vbox_node_helper(Node* vbox,
2751 Node* vect,
2752 const TypeInstPtr* box_type,
2753 const TypeVect* vect_type) {
2754 if (vbox->is_Phi() && vect->is_Phi()) {
2755 assert(vbox->as_Phi()->region() == vect->as_Phi()->region(), "");
2756 Node* new_phi = new PhiNode(vbox->as_Phi()->region(), box_type);
2757 for (uint i = 1; i < vbox->req(); i++) {
2758 Node* new_box = expand_vbox_node_helper(vbox->in(i), vect->in(i), box_type, vect_type);
2759 new_phi->set_req(i, new_box);
2760 }
2761 new_phi = initial_gvn()->transform(new_phi);
2762 return new_phi;
2763 } else if (vbox->is_Proj() && vbox->in(0)->Opcode() == Op_VectorBoxAllocate) {
2764 VectorBoxAllocateNode* vbox_alloc = static_cast<VectorBoxAllocateNode*>(vbox->in(0));
2765 return expand_vbox_alloc_node(vbox_alloc, vect, box_type, vect_type);
2766 } else {
2767 // TODO: ensure that expanded vbox is initialized with the same value (vect).
2768 return vbox; // already expanded
2769 }
2770 }
2771
2772 Node* Compile::expand_vbox_alloc_node(VectorBoxAllocateNode* vbox_alloc,
2773 Node* value,
2774 const TypeInstPtr* box_type,
2775 const TypeVect* vect_type) {
2776 JVMState* jvms = clone_jvms(C, vbox_alloc);
2777 GraphKit kit(jvms);
2778 PhaseGVN& gvn = kit.gvn();
2779
2780 ciInstanceKlass* box_klass = box_type->klass()->as_instance_klass();
2781 BasicType bt = vect_type->element_basic_type();
2782 int num_elem = vect_type->length();
2783
2784 bool is_mask = box_klass->is_vectormask();
2785 //assert(!is_mask || vect_type->element_basic_type() == getMaskBasicType(bt), "consistent vector element type expected");
2786 if (is_mask && bt != T_BOOLEAN) {
2787 value = gvn.transform(new VectorStoreMaskNode(value, bt, num_elem));
2788 // Although type of mask depends on its definition, in terms of storage everything is stored in boolean array.
2789 bt = T_BOOLEAN;
2790 assert(value->as_Vector()->bottom_type()->is_vect()->element_basic_type() == bt,
2791 "must be consistent with mask representation");
2792 }
2793
2794 // Generate the allocate for the Vector object.
2795 const TypeKlassPtr* klass_type = box_type->as_klass_type();
2796 Node* klass_node = kit.makecon(klass_type);
2797 Node* vec_obj = kit.new_instance(klass_node);
2798
2799 // Generate array allocation for the field which holds the values.
2800 const TypeKlassPtr* array_klass = TypeKlassPtr::make(ciTypeArrayKlass::make(bt));
2801 Node* arr = kit.new_array(kit.makecon(array_klass), kit.intcon(num_elem), 1);
2802
2803 // FIXME convert both stores into initializing stores
2804
2805 // Store the vector value into the array.
2806 Node* arr_adr = kit.array_element_address(arr, kit.intcon(0), bt);
2807 const TypePtr* arr_adr_type = arr_adr->bottom_type()->is_ptr();
2808 Node* arr_mem = kit.memory(arr_adr);
2809 Node* vstore = gvn.transform(StoreVectorNode::make(0,
2810 kit.control(),
2811 arr_mem,
2812 arr_adr,
2813 arr_adr_type,
2814 value,
2815 num_elem));
2816 kit.set_memory(vstore, arr_adr_type);
2817
2818
2819 // Store the allocated array into object.
2820 ciField* field = box_klass->get_field_by_name(ciSymbol::make(is_mask ? "bits" : "vec"),
2821 ciSymbol::make(TypeArrayKlass::external_name(bt)),
2822 false);
2823 Node* vec_field = kit.basic_plus_adr(vec_obj, field->offset_in_bytes());
2824 const TypePtr* vec_adr_type = vec_field->bottom_type()->is_ptr();
2825 Node* field_store = gvn.transform(kit.access_store_at(kit.control(),
2826 vec_obj,
2827 vec_field,
2828 vec_adr_type,
2829 arr,
2830 TypeOopPtr::make_from_klass(field->type()->as_klass()),
2831 T_OBJECT,
2832 IN_HEAP));
2833
2834 kit.set_memory(field_store, vec_adr_type);
2835
2836 kit.replace_call(vbox_alloc, vec_obj, true);
2837 C->remove_macro_node(vbox_alloc);
2838 return vec_obj;
2839 }
2840
2841 void Compile::expand_vunbox_node(VectorUnboxNode* vec_unbox) {
2842 if (vec_unbox->outcnt() > 0) {
2843 GraphKit kit;
2844 PhaseGVN& gvn = kit.gvn();
2845
2846 Node* obj = vec_unbox->obj();
2847 const TypeInstPtr* tinst = gvn.type(obj)->isa_instptr();
2848 ciInstanceKlass* from_kls = tinst->klass()->as_instance_klass();
2849 BasicType bt = vec_unbox->vect_type()->element_basic_type();
2850 BasicType masktype = bt;
2851
2852 const char* field_name = "vec";
2853 if (from_kls->is_vectormask()) {
2854 field_name = "bits";
2855 bt = T_BOOLEAN;
2856 } else if (from_kls->is_vectorshuffle()) {
2857 field_name = "reorder";
2858 bt = T_BYTE;
2859 }
2860
2861 ciField* field = from_kls->get_field_by_name(ciSymbol::make(field_name),
2862 ciSymbol::make(TypeArrayKlass::external_name(bt)), false);
2863
2864 int offset = field->offset_in_bytes();
2865 Node* vec_adr = kit.basic_plus_adr(obj, offset);
2866
2867 Node* mem = vec_unbox->mem();
2868 Node* ctrl = vec_unbox->in(0);
2869 Node* vec_field_ld = LoadNode::make(gvn,
2870 ctrl,
2871 mem,
2872 vec_adr,
2873 vec_adr->bottom_type()->is_ptr(),
2874 TypeOopPtr::make_from_klass(field->type()->as_klass()),
2875 T_OBJECT,
2876 MemNode::unordered);
2877 vec_field_ld = gvn.transform(vec_field_ld);
2878
2879 Node* adr = kit.array_element_address(vec_field_ld, gvn.intcon(0), bt);
2880 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
2881 int num_elem = vec_unbox->bottom_type()->is_vect()->length();
2882 Node* vec_val_load = LoadVectorNode::make(0,
2883 ctrl,
2884 mem,
2885 adr,
2886 adr_type,
2887 num_elem,
2888 bt);
2889 vec_val_load = gvn.transform(vec_val_load);
2890
2891 if (from_kls->is_vectormask() && masktype != T_BOOLEAN) {
2892 assert(vec_unbox->bottom_type()->is_vect()->element_basic_type() == masktype, "expect mask type consistency");
2893 vec_val_load = gvn.transform(new VectorLoadMaskNode(vec_val_load, TypeVect::make(masktype, num_elem)));
2894 } else if (from_kls->is_vectorshuffle()) {
2895 assert(vec_unbox->bottom_type()->is_vect()->element_basic_type() == masktype, "expect shuffle type consistency");
2896 vec_val_load = gvn.transform(new VectorLoadShuffleNode(vec_val_load, TypeVect::make(masktype, num_elem)));
2897 }
2898
2899 gvn.hash_delete(vec_unbox);
2900 vec_unbox->disconnect_inputs(NULL, C);
2901 C->gvn_replace_by(vec_unbox, vec_val_load);
2902 }
2903 C->remove_macro_node(vec_unbox);
2904 }
2905
2906 void Compile::eliminate_vbox_alloc_node(VectorBoxAllocateNode* vbox_alloc) {
2907 JVMState* jvms = clone_jvms(C, vbox_alloc);
2908 GraphKit kit(jvms);
2909 // FIXME replace VBA with a safepoint. Otherwise, no safepoints left in tight loops.
2910 kit.replace_call(vbox_alloc, kit.top(), true);
2911 C->remove_macro_node(vbox_alloc);
2912 }
2913
2914 //------------------------------Code_Gen---------------------------------------
2915 // Given a graph, generate code for it
2916 void Compile::Code_Gen() {
2917 if (failing()) {
2918 return;
2919 }
2920
2921 // Perform instruction selection. You might think we could reclaim Matcher
2922 // memory PDQ, but actually the Matcher is used in generating spill code.
2923 // Internals of the Matcher (including some VectorSets) must remain live
2924 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2925 // set a bit in reclaimed memory.
2926
2927 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2928 // nodes. Mapping is only valid at the root of each matched subtree.
2929 NOT_PRODUCT( verify_graph_edges(); )
2930
2931 Matcher matcher;
2932 _matcher = &matcher;
2933 {
2934 TracePhase tp("matcher", &timers[_t_matcher]);
2935 matcher.match();
2936 print_method(PHASE_AFTER_MATCHING, 3);
2937 }
2938 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2939 // nodes. Mapping is only valid at the root of each matched subtree.
2940 NOT_PRODUCT( verify_graph_edges(); )
2941
2942 // If you have too many nodes, or if matching has failed, bail out
2943 check_node_count(0, "out of nodes matching instructions");
2944 if (failing()) {
2945 return;
2946 }
2947
2948 print_method(PHASE_MATCHING, 2);
2949
2950 // Build a proper-looking CFG
2951 PhaseCFG cfg(node_arena(), root(), matcher);
2952 _cfg = &cfg;
2953 {
2954 TracePhase tp("scheduler", &timers[_t_scheduler]);
2955 bool success = cfg.do_global_code_motion();
2956 if (!success) {
2957 return;
2958 }
2959
2960 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2961 NOT_PRODUCT( verify_graph_edges(); )
2962 debug_only( cfg.verify(); )
2963 }
2964
2965 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2966 _regalloc = ®alloc;
2967 {
2968 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2969 // Perform register allocation. After Chaitin, use-def chains are
2970 // no longer accurate (at spill code) and so must be ignored.
2971 // Node->LRG->reg mappings are still accurate.
2972 _regalloc->Register_Allocate();
2973
2974 // Bail out if the allocator builds too many nodes
2975 if (failing()) {
2976 return;
2977 }
2978 }
2979
2980 // Prior to register allocation we kept empty basic blocks in case the
2981 // the allocator needed a place to spill. After register allocation we
2982 // are not adding any new instructions. If any basic block is empty, we
2983 // can now safely remove it.
2984 {
2985 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2986 cfg.remove_empty_blocks();
2987 if (do_freq_based_layout()) {
2988 PhaseBlockLayout layout(cfg);
2989 } else {
2990 cfg.set_loop_alignment();
2991 }
2992 cfg.fixup_flow();
2993 }
2994
2995 // Apply peephole optimizations
2996 if( OptoPeephole ) {
2997 TracePhase tp("peephole", &timers[_t_peephole]);
2998 PhasePeephole peep( _regalloc, cfg);
2999 peep.do_transform();
3000 }
3001
3002 // Do late expand if CPU requires this.
3003 if (Matcher::require_postalloc_expand) {
3004 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
3005 cfg.postalloc_expand(_regalloc);
3006 }
3007
3008 // Convert Nodes to instruction bits in a buffer
3009 {
3010 TraceTime tp("output", &timers[_t_output], CITime);
3011 Output();
3012 }
3013
3014 print_method(PHASE_FINAL_CODE);
3015
3016 // He's dead, Jim.
3017 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
3018 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
3019 }
3020
3021
3022 //------------------------------dump_asm---------------------------------------
3023 // Dump formatted assembly
3024 #ifndef PRODUCT
3025 void Compile::dump_asm(int *pcs, uint pc_limit) {
3026 bool cut_short = false;
3027 tty->print_cr("#");
3028 tty->print("# "); _tf->dump(); tty->cr();
3029 tty->print_cr("#");
3030
3031 // For all blocks
3032 int pc = 0x0; // Program counter
3033 char starts_bundle = ' ';
3034 _regalloc->dump_frame();
3035
3036 Node *n = NULL;
3037 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
3038 if (VMThread::should_terminate()) {
3039 cut_short = true;
3040 break;
3041 }
3042 Block* block = _cfg->get_block(i);
3043 if (block->is_connector() && !Verbose) {
3044 continue;
3045 }
3046 n = block->head();
3047 if (pcs && n->_idx < pc_limit) {
3048 tty->print("%3.3x ", pcs[n->_idx]);
3049 } else {
3050 tty->print(" ");
3051 }
3052 block->dump_head(_cfg);
3053 if (block->is_connector()) {
3054 tty->print_cr(" # Empty connector block");
3055 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
3056 tty->print_cr(" # Block is sole successor of call");
3057 }
3058
3059 // For all instructions
3060 Node *delay = NULL;
3061 for (uint j = 0; j < block->number_of_nodes(); j++) {
3062 if (VMThread::should_terminate()) {
3063 cut_short = true;
3064 break;
3065 }
3066 n = block->get_node(j);
3067 if (valid_bundle_info(n)) {
3068 Bundle* bundle = node_bundling(n);
3069 if (bundle->used_in_unconditional_delay()) {
3070 delay = n;
3071 continue;
3072 }
3073 if (bundle->starts_bundle()) {
3074 starts_bundle = '+';
3075 }
3076 }
3077
3078 if (WizardMode) {
3079 n->dump();
3080 }
3081
3082 if( !n->is_Region() && // Dont print in the Assembly
3083 !n->is_Phi() && // a few noisely useless nodes
3084 !n->is_Proj() &&
3085 !n->is_MachTemp() &&
3086 !n->is_SafePointScalarObject() &&
3087 !n->is_Catch() && // Would be nice to print exception table targets
3088 !n->is_MergeMem() && // Not very interesting
3089 !n->is_top() && // Debug info table constants
3090 !(n->is_Con() && !n->is_Mach())// Debug info table constants
3091 ) {
3092 if (pcs && n->_idx < pc_limit)
3093 tty->print("%3.3x", pcs[n->_idx]);
3094 else
3095 tty->print(" ");
3096 tty->print(" %c ", starts_bundle);
3097 starts_bundle = ' ';
3098 tty->print("\t");
3099 n->format(_regalloc, tty);
3100 tty->cr();
3101 }
3102
3103 // If we have an instruction with a delay slot, and have seen a delay,
3104 // then back up and print it
3105 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
3106 assert(delay != NULL, "no unconditional delay instruction");
3107 if (WizardMode) delay->dump();
3108
3109 if (node_bundling(delay)->starts_bundle())
3110 starts_bundle = '+';
3111 if (pcs && n->_idx < pc_limit)
3112 tty->print("%3.3x", pcs[n->_idx]);
3113 else
3114 tty->print(" ");
3115 tty->print(" %c ", starts_bundle);
3116 starts_bundle = ' ';
3117 tty->print("\t");
3118 delay->format(_regalloc, tty);
3119 tty->cr();
3120 delay = NULL;
3121 }
3122
3123 // Dump the exception table as well
3124 if( n->is_Catch() && (Verbose || WizardMode) ) {
3125 // Print the exception table for this offset
3126 _handler_table.print_subtable_for(pc);
3127 }
3128 }
3129
3130 if (pcs && n->_idx < pc_limit)
3131 tty->print_cr("%3.3x", pcs[n->_idx]);
3132 else
3133 tty->cr();
3134
3135 assert(cut_short || delay == NULL, "no unconditional delay branch");
3136
3137 } // End of per-block dump
3138 tty->cr();
3139
3140 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
3141 }
3142 #endif
3143
3144 //------------------------------Final_Reshape_Counts---------------------------
3145 // This class defines counters to help identify when a method
3146 // may/must be executed using hardware with only 24-bit precision.
3147 struct Final_Reshape_Counts : public StackObj {
3148 int _call_count; // count non-inlined 'common' calls
3149 int _float_count; // count float ops requiring 24-bit precision
3150 int _double_count; // count double ops requiring more precision
3151 int _java_call_count; // count non-inlined 'java' calls
3152 int _inner_loop_count; // count loops which need alignment
3153 VectorSet _visited; // Visitation flags
3154 Node_List _tests; // Set of IfNodes & PCTableNodes
3155
3156 Final_Reshape_Counts() :
3157 _call_count(0), _float_count(0), _double_count(0),
3158 _java_call_count(0), _inner_loop_count(0),
3159 _visited( Thread::current()->resource_area() ) { }
3160
3161 void inc_call_count () { _call_count ++; }
3162 void inc_float_count () { _float_count ++; }
3163 void inc_double_count() { _double_count++; }
3164 void inc_java_call_count() { _java_call_count++; }
3165 void inc_inner_loop_count() { _inner_loop_count++; }
3166
3167 int get_call_count () const { return _call_count ; }
3168 int get_float_count () const { return _float_count ; }
3169 int get_double_count() const { return _double_count; }
3170 int get_java_call_count() const { return _java_call_count; }
3171 int get_inner_loop_count() const { return _inner_loop_count; }
3172 };
3173
3174 #ifdef ASSERT
3175 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
3176 ciInstanceKlass *k = tp->klass()->as_instance_klass();
3177 // Make sure the offset goes inside the instance layout.
3178 return k->contains_field_offset(tp->offset());
3179 // Note that OffsetBot and OffsetTop are very negative.
3180 }
3181 #endif
3182
3183 // Eliminate trivially redundant StoreCMs and accumulate their
3184 // precedence edges.
3185 void Compile::eliminate_redundant_card_marks(Node* n) {
3186 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3187 if (n->in(MemNode::Address)->outcnt() > 1) {
3188 // There are multiple users of the same address so it might be
3189 // possible to eliminate some of the StoreCMs
3190 Node* mem = n->in(MemNode::Memory);
3191 Node* adr = n->in(MemNode::Address);
3192 Node* val = n->in(MemNode::ValueIn);
3193 Node* prev = n;
3194 bool done = false;
3195 // Walk the chain of StoreCMs eliminating ones that match. As
3196 // long as it's a chain of single users then the optimization is
3197 // safe. Eliminating partially redundant StoreCMs would require
3198 // cloning copies down the other paths.
3199 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3200 if (adr == mem->in(MemNode::Address) &&
3201 val == mem->in(MemNode::ValueIn)) {
3202 // redundant StoreCM
3203 if (mem->req() > MemNode::OopStore) {
3204 // Hasn't been processed by this code yet.
3205 n->add_prec(mem->in(MemNode::OopStore));
3206 } else {
3207 // Already converted to precedence edge
3208 for (uint i = mem->req(); i < mem->len(); i++) {
3209 // Accumulate any precedence edges
3210 if (mem->in(i) != NULL) {
3211 n->add_prec(mem->in(i));
3212 }
3213 }
3214 // Everything above this point has been processed.
3215 done = true;
3216 }
3217 // Eliminate the previous StoreCM
3218 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3219 assert(mem->outcnt() == 0, "should be dead");
3220 mem->disconnect_inputs(NULL, this);
3221 } else {
3222 prev = mem;
3223 }
3224 mem = prev->in(MemNode::Memory);
3225 }
3226 }
3227 }
3228
3229 //------------------------------final_graph_reshaping_impl----------------------
3230 // Implement items 1-5 from final_graph_reshaping below.
3231 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
3232
3233 if ( n->outcnt() == 0 ) return; // dead node
3234 uint nop = n->Opcode();
3235
3236 // Check for 2-input instruction with "last use" on right input.
3237 // Swap to left input. Implements item (2).
3238 if( n->req() == 3 && // two-input instruction
3239 n->in(1)->outcnt() > 1 && // left use is NOT a last use
3240 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3241 n->in(2)->outcnt() == 1 &&// right use IS a last use
3242 !n->in(2)->is_Con() ) { // right use is not a constant
3243 // Check for commutative opcode
3244 switch( nop ) {
3245 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
3246 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD:
3247 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD:
3248 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
3249 case Op_AndL: case Op_XorL: case Op_OrL:
3250 case Op_AndI: case Op_XorI: case Op_OrI: {
3251 // Move "last use" input to left by swapping inputs
3252 n->swap_edges(1, 2);
3253 break;
3254 }
3255 default:
3256 break;
3257 }
3258 }
3259
3260 #ifdef ASSERT
3261 if( n->is_Mem() ) {
3262 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3263 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3264 // oop will be recorded in oop map if load crosses safepoint
3265 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3266 LoadNode::is_immutable_value(n->in(MemNode::Address))),
3267 "raw memory operations should have control edge");
3268 }
3269 if (n->is_MemBar()) {
3270 MemBarNode* mb = n->as_MemBar();
3271 if (mb->trailing_store() || mb->trailing_load_store()) {
3272 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3273 Node* mem = mb->in(MemBarNode::Precedent);
3274 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3275 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3276 } else if (mb->leading()) {
3277 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3278 }
3279 }
3280 #endif
3281 // Count FPU ops and common calls, implements item (3)
3282 switch( nop ) {
3283 // Count all float operations that may use FPU
3284 case Op_AddF:
3285 case Op_SubF:
3286 case Op_MulF:
3287 case Op_DivF:
3288 case Op_NegF:
3289 case Op_ModF:
3290 case Op_ConvI2F:
3291 case Op_ConF:
3292 case Op_CmpF:
3293 case Op_CmpF3:
3294 // case Op_ConvL2F: // longs are split into 32-bit halves
3295 frc.inc_float_count();
3296 break;
3297
3298 case Op_ConvF2D:
3299 case Op_ConvD2F:
3300 frc.inc_float_count();
3301 frc.inc_double_count();
3302 break;
3303
3304 // Count all double operations that may use FPU
3305 case Op_AddD:
3306 case Op_SubD:
3307 case Op_MulD:
3308 case Op_DivD:
3309 case Op_NegD:
3310 case Op_ModD:
3311 case Op_ConvI2D:
3312 case Op_ConvD2I:
3313 // case Op_ConvL2D: // handled by leaf call
3314 // case Op_ConvD2L: // handled by leaf call
3315 case Op_ConD:
3316 case Op_CmpD:
3317 case Op_CmpD3:
3318 frc.inc_double_count();
3319 break;
3320 case Op_Opaque1: // Remove Opaque Nodes before matching
3321 case Op_Opaque2: // Remove Opaque Nodes before matching
3322 case Op_Opaque3:
3323 n->subsume_by(n->in(1), this);
3324 break;
3325 case Op_CallStaticJava:
3326 case Op_CallJava:
3327 case Op_CallDynamicJava:
3328 frc.inc_java_call_count(); // Count java call site;
3329 case Op_CallRuntime:
3330 case Op_CallLeaf:
3331 case Op_CallLeafVector:
3332 case Op_CallLeafNoFP: {
3333 assert (n->is_Call(), "");
3334 CallNode *call = n->as_Call();
3335 // Count call sites where the FP mode bit would have to be flipped.
3336 // Do not count uncommon runtime calls:
3337 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3338 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3339 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3340 frc.inc_call_count(); // Count the call site
3341 } else { // See if uncommon argument is shared
3342 Node *n = call->in(TypeFunc::Parms);
3343 int nop = n->Opcode();
3344 // Clone shared simple arguments to uncommon calls, item (1).
3345 if (n->outcnt() > 1 &&
3346 !n->is_Proj() &&
3347 nop != Op_CreateEx &&
3348 nop != Op_CheckCastPP &&
3349 nop != Op_DecodeN &&
3350 nop != Op_DecodeNKlass &&
3351 !n->is_Mem() &&
3352 !n->is_Phi()) {
3353 Node *x = n->clone();
3354 call->set_req(TypeFunc::Parms, x);
3355 }
3356 }
3357 break;
3358 }
3359
3360 case Op_StoreD:
3361 case Op_LoadD:
3362 case Op_LoadD_unaligned:
3363 frc.inc_double_count();
3364 goto handle_mem;
3365 case Op_StoreF:
3366 case Op_LoadF:
3367 frc.inc_float_count();
3368 goto handle_mem;
3369
3370 case Op_StoreCM:
3371 {
3372 // Convert OopStore dependence into precedence edge
3373 Node* prec = n->in(MemNode::OopStore);
3374 n->del_req(MemNode::OopStore);
3375 n->add_prec(prec);
3376 eliminate_redundant_card_marks(n);
3377 }
3378
3379 // fall through
3380
3381 case Op_StoreB:
3382 case Op_StoreC:
3383 case Op_StorePConditional:
3384 case Op_StoreI:
3385 case Op_StoreL:
3386 case Op_StoreIConditional:
3387 case Op_StoreLConditional:
3388 case Op_CompareAndSwapB:
3389 case Op_CompareAndSwapS:
3390 case Op_CompareAndSwapI:
3391 case Op_CompareAndSwapL:
3392 case Op_CompareAndSwapP:
3393 case Op_CompareAndSwapN:
3394 case Op_WeakCompareAndSwapB:
3395 case Op_WeakCompareAndSwapS:
3396 case Op_WeakCompareAndSwapI:
3397 case Op_WeakCompareAndSwapL:
3398 case Op_WeakCompareAndSwapP:
3399 case Op_WeakCompareAndSwapN:
3400 case Op_CompareAndExchangeB:
3401 case Op_CompareAndExchangeS:
3402 case Op_CompareAndExchangeI:
3403 case Op_CompareAndExchangeL:
3404 case Op_CompareAndExchangeP:
3405 case Op_CompareAndExchangeN:
3406 case Op_GetAndAddS:
3407 case Op_GetAndAddB:
3408 case Op_GetAndAddI:
3409 case Op_GetAndAddL:
3410 case Op_GetAndSetS:
3411 case Op_GetAndSetB:
3412 case Op_GetAndSetI:
3413 case Op_GetAndSetL:
3414 case Op_GetAndSetP:
3415 case Op_GetAndSetN:
3416 case Op_StoreP:
3417 case Op_StoreN:
3418 case Op_StoreNKlass:
3419 case Op_LoadB:
3420 case Op_LoadUB:
3421 case Op_LoadUS:
3422 case Op_LoadI:
3423 case Op_LoadKlass:
3424 case Op_LoadNKlass:
3425 case Op_LoadL:
3426 case Op_LoadL_unaligned:
3427 case Op_LoadPLocked:
3428 case Op_LoadP:
3429 #if INCLUDE_ZGC
3430 case Op_LoadBarrierSlowReg:
3431 case Op_LoadBarrierWeakSlowReg:
3432 #endif
3433 case Op_LoadN:
3434 case Op_LoadRange:
3435 case Op_LoadS: {
3436 handle_mem:
3437 #ifdef ASSERT
3438 if( VerifyOptoOopOffsets ) {
3439 assert( n->is_Mem(), "" );
3440 MemNode *mem = (MemNode*)n;
3441 // Check to see if address types have grounded out somehow.
3442 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3443 assert( !tp || oop_offset_is_sane(tp), "" );
3444 }
3445 #endif
3446 break;
3447 }
3448
3449 case Op_AddP: { // Assert sane base pointers
3450 Node *addp = n->in(AddPNode::Address);
3451 assert( !addp->is_AddP() ||
3452 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3453 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3454 "Base pointers must match (addp %u)", addp->_idx );
3455 #ifdef _LP64
3456 if ((UseCompressedOops || UseCompressedClassPointers) &&
3457 addp->Opcode() == Op_ConP &&
3458 addp == n->in(AddPNode::Base) &&
3459 n->in(AddPNode::Offset)->is_Con()) {
3460 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3461 // on the platform and on the compressed oops mode.
3462 // Use addressing with narrow klass to load with offset on x86.
3463 // Some platforms can use the constant pool to load ConP.
3464 // Do this transformation here since IGVN will convert ConN back to ConP.
3465 const Type* t = addp->bottom_type();
3466 bool is_oop = t->isa_oopptr() != NULL;
3467 bool is_klass = t->isa_klassptr() != NULL;
3468
3469 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3470 (is_klass && Matcher::const_klass_prefer_decode())) {
3471 Node* nn = NULL;
3472
3473 int op = is_oop ? Op_ConN : Op_ConNKlass;
3474
3475 // Look for existing ConN node of the same exact type.
3476 Node* r = root();
3477 uint cnt = r->outcnt();
3478 for (uint i = 0; i < cnt; i++) {
3479 Node* m = r->raw_out(i);
3480 if (m!= NULL && m->Opcode() == op &&
3481 m->bottom_type()->make_ptr() == t) {
3482 nn = m;
3483 break;
3484 }
3485 }
3486 if (nn != NULL) {
3487 // Decode a narrow oop to match address
3488 // [R12 + narrow_oop_reg<<3 + offset]
3489 if (is_oop) {
3490 nn = new DecodeNNode(nn, t);
3491 } else {
3492 nn = new DecodeNKlassNode(nn, t);
3493 }
3494 // Check for succeeding AddP which uses the same Base.
3495 // Otherwise we will run into the assertion above when visiting that guy.
3496 for (uint i = 0; i < n->outcnt(); ++i) {
3497 Node *out_i = n->raw_out(i);
3498 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3499 out_i->set_req(AddPNode::Base, nn);
3500 #ifdef ASSERT
3501 for (uint j = 0; j < out_i->outcnt(); ++j) {
3502 Node *out_j = out_i->raw_out(j);
3503 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3504 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3505 }
3506 #endif
3507 }
3508 }
3509 n->set_req(AddPNode::Base, nn);
3510 n->set_req(AddPNode::Address, nn);
3511 if (addp->outcnt() == 0) {
3512 addp->disconnect_inputs(NULL, this);
3513 }
3514 }
3515 }
3516 }
3517 #endif
3518 // platform dependent reshaping of the address expression
3519 reshape_address(n->as_AddP());
3520 break;
3521 }
3522
3523 case Op_CastPP: {
3524 // Remove CastPP nodes to gain more freedom during scheduling but
3525 // keep the dependency they encode as control or precedence edges
3526 // (if control is set already) on memory operations. Some CastPP
3527 // nodes don't have a control (don't carry a dependency): skip
3528 // those.
3529 if (n->in(0) != NULL) {
3530 ResourceMark rm;
3531 Unique_Node_List wq;
3532 wq.push(n);
3533 for (uint next = 0; next < wq.size(); ++next) {
3534 Node *m = wq.at(next);
3535 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3536 Node* use = m->fast_out(i);
3537 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3538 use->ensure_control_or_add_prec(n->in(0));
3539 } else {
3540 switch(use->Opcode()) {
3541 case Op_AddP:
3542 case Op_DecodeN:
3543 case Op_DecodeNKlass:
3544 case Op_CheckCastPP:
3545 case Op_CastPP:
3546 wq.push(use);
3547 break;
3548 }
3549 }
3550 }
3551 }
3552 }
3553 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3554 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3555 Node* in1 = n->in(1);
3556 const Type* t = n->bottom_type();
3557 Node* new_in1 = in1->clone();
3558 new_in1->as_DecodeN()->set_type(t);
3559
3560 if (!Matcher::narrow_oop_use_complex_address()) {
3561 //
3562 // x86, ARM and friends can handle 2 adds in addressing mode
3563 // and Matcher can fold a DecodeN node into address by using
3564 // a narrow oop directly and do implicit NULL check in address:
3565 //
3566 // [R12 + narrow_oop_reg<<3 + offset]
3567 // NullCheck narrow_oop_reg
3568 //
3569 // On other platforms (Sparc) we have to keep new DecodeN node and
3570 // use it to do implicit NULL check in address:
3571 //
3572 // decode_not_null narrow_oop_reg, base_reg
3573 // [base_reg + offset]
3574 // NullCheck base_reg
3575 //
3576 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3577 // to keep the information to which NULL check the new DecodeN node
3578 // corresponds to use it as value in implicit_null_check().
3579 //
3580 new_in1->set_req(0, n->in(0));
3581 }
3582
3583 n->subsume_by(new_in1, this);
3584 if (in1->outcnt() == 0) {
3585 in1->disconnect_inputs(NULL, this);
3586 }
3587 } else {
3588 n->subsume_by(n->in(1), this);
3589 if (n->outcnt() == 0) {
3590 n->disconnect_inputs(NULL, this);
3591 }
3592 }
3593 break;
3594 }
3595 #ifdef _LP64
3596 case Op_CmpP:
3597 // Do this transformation here to preserve CmpPNode::sub() and
3598 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3599 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3600 Node* in1 = n->in(1);
3601 Node* in2 = n->in(2);
3602 if (!in1->is_DecodeNarrowPtr()) {
3603 in2 = in1;
3604 in1 = n->in(2);
3605 }
3606 assert(in1->is_DecodeNarrowPtr(), "sanity");
3607
3608 Node* new_in2 = NULL;
3609 if (in2->is_DecodeNarrowPtr()) {
3610 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3611 new_in2 = in2->in(1);
3612 } else if (in2->Opcode() == Op_ConP) {
3613 const Type* t = in2->bottom_type();
3614 if (t == TypePtr::NULL_PTR) {
3615 assert(in1->is_DecodeN(), "compare klass to null?");
3616 // Don't convert CmpP null check into CmpN if compressed
3617 // oops implicit null check is not generated.
3618 // This will allow to generate normal oop implicit null check.
3619 if (Matcher::gen_narrow_oop_implicit_null_checks())
3620 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3621 //
3622 // This transformation together with CastPP transformation above
3623 // will generated code for implicit NULL checks for compressed oops.
3624 //
3625 // The original code after Optimize()
3626 //
3627 // LoadN memory, narrow_oop_reg
3628 // decode narrow_oop_reg, base_reg
3629 // CmpP base_reg, NULL
3630 // CastPP base_reg // NotNull
3631 // Load [base_reg + offset], val_reg
3632 //
3633 // after these transformations will be
3634 //
3635 // LoadN memory, narrow_oop_reg
3636 // CmpN narrow_oop_reg, NULL
3637 // decode_not_null narrow_oop_reg, base_reg
3638 // Load [base_reg + offset], val_reg
3639 //
3640 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3641 // since narrow oops can be used in debug info now (see the code in
3642 // final_graph_reshaping_walk()).
3643 //
3644 // At the end the code will be matched to
3645 // on x86:
3646 //
3647 // Load_narrow_oop memory, narrow_oop_reg
3648 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3649 // NullCheck narrow_oop_reg
3650 //
3651 // and on sparc:
3652 //
3653 // Load_narrow_oop memory, narrow_oop_reg
3654 // decode_not_null narrow_oop_reg, base_reg
3655 // Load [base_reg + offset], val_reg
3656 // NullCheck base_reg
3657 //
3658 } else if (t->isa_oopptr()) {
3659 new_in2 = ConNode::make(t->make_narrowoop());
3660 } else if (t->isa_klassptr()) {
3661 new_in2 = ConNode::make(t->make_narrowklass());
3662 }
3663 }
3664 if (new_in2 != NULL) {
3665 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3666 n->subsume_by(cmpN, this);
3667 if (in1->outcnt() == 0) {
3668 in1->disconnect_inputs(NULL, this);
3669 }
3670 if (in2->outcnt() == 0) {
3671 in2->disconnect_inputs(NULL, this);
3672 }
3673 }
3674 }
3675 break;
3676
3677 case Op_DecodeN:
3678 case Op_DecodeNKlass:
3679 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3680 // DecodeN could be pinned when it can't be fold into
3681 // an address expression, see the code for Op_CastPP above.
3682 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3683 break;
3684
3685 case Op_EncodeP:
3686 case Op_EncodePKlass: {
3687 Node* in1 = n->in(1);
3688 if (in1->is_DecodeNarrowPtr()) {
3689 n->subsume_by(in1->in(1), this);
3690 } else if (in1->Opcode() == Op_ConP) {
3691 const Type* t = in1->bottom_type();
3692 if (t == TypePtr::NULL_PTR) {
3693 assert(t->isa_oopptr(), "null klass?");
3694 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3695 } else if (t->isa_oopptr()) {
3696 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3697 } else if (t->isa_klassptr()) {
3698 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3699 }
3700 }
3701 if (in1->outcnt() == 0) {
3702 in1->disconnect_inputs(NULL, this);
3703 }
3704 break;
3705 }
3706
3707 case Op_Proj: {
3708 if (OptimizeStringConcat) {
3709 ProjNode* p = n->as_Proj();
3710 if (p->_is_io_use) {
3711 // Separate projections were used for the exception path which
3712 // are normally removed by a late inline. If it wasn't inlined
3713 // then they will hang around and should just be replaced with
3714 // the original one.
3715 Node* proj = NULL;
3716 // Replace with just one
3717 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3718 Node *use = i.get();
3719 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3720 proj = use;
3721 break;
3722 }
3723 }
3724 assert(proj != NULL, "must be found");
3725 p->subsume_by(proj, this);
3726 }
3727 }
3728 break;
3729 }
3730
3731 case Op_Phi:
3732 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3733 // The EncodeP optimization may create Phi with the same edges
3734 // for all paths. It is not handled well by Register Allocator.
3735 Node* unique_in = n->in(1);
3736 assert(unique_in != NULL, "");
3737 uint cnt = n->req();
3738 for (uint i = 2; i < cnt; i++) {
3739 Node* m = n->in(i);
3740 assert(m != NULL, "");
3741 if (unique_in != m)
3742 unique_in = NULL;
3743 }
3744 if (unique_in != NULL) {
3745 n->subsume_by(unique_in, this);
3746 }
3747 }
3748 break;
3749
3750 #endif
3751
3752 #ifdef ASSERT
3753 case Op_CastII:
3754 // Verify that all range check dependent CastII nodes were removed.
3755 if (n->isa_CastII()->has_range_check()) {
3756 n->dump(3);
3757 assert(false, "Range check dependent CastII node was not removed");
3758 }
3759 break;
3760 #endif
3761
3762 case Op_ModI:
3763 if (UseDivMod) {
3764 // Check if a%b and a/b both exist
3765 Node* d = n->find_similar(Op_DivI);
3766 if (d) {
3767 // Replace them with a fused divmod if supported
3768 if (Matcher::has_match_rule(Op_DivModI)) {
3769 DivModINode* divmod = DivModINode::make(n);
3770 d->subsume_by(divmod->div_proj(), this);
3771 n->subsume_by(divmod->mod_proj(), this);
3772 } else {
3773 // replace a%b with a-((a/b)*b)
3774 Node* mult = new MulINode(d, d->in(2));
3775 Node* sub = new SubINode(d->in(1), mult);
3776 n->subsume_by(sub, this);
3777 }
3778 }
3779 }
3780 break;
3781
3782 case Op_ModL:
3783 if (UseDivMod) {
3784 // Check if a%b and a/b both exist
3785 Node* d = n->find_similar(Op_DivL);
3786 if (d) {
3787 // Replace them with a fused divmod if supported
3788 if (Matcher::has_match_rule(Op_DivModL)) {
3789 DivModLNode* divmod = DivModLNode::make(n);
3790 d->subsume_by(divmod->div_proj(), this);
3791 n->subsume_by(divmod->mod_proj(), this);
3792 } else {
3793 // replace a%b with a-((a/b)*b)
3794 Node* mult = new MulLNode(d, d->in(2));
3795 Node* sub = new SubLNode(d->in(1), mult);
3796 n->subsume_by(sub, this);
3797 }
3798 }
3799 }
3800 break;
3801
3802 case Op_LoadVector:
3803 case Op_StoreVector:
3804 break;
3805
3806 case Op_AddReductionVI:
3807 case Op_AddReductionVL:
3808 case Op_AddReductionVF:
3809 case Op_AddReductionVD:
3810 case Op_MulReductionVI:
3811 case Op_MulReductionVL:
3812 case Op_MulReductionVF:
3813 case Op_MulReductionVD:
3814 case Op_MinReductionV:
3815 case Op_MaxReductionV:
3816 case Op_AndReductionV:
3817 case Op_OrReductionV:
3818 case Op_XorReductionV:
3819 case Op_SubReductionV:
3820 case Op_SubReductionVFP:
3821 break;
3822
3823 case Op_PackB:
3824 case Op_PackS:
3825 case Op_PackI:
3826 case Op_PackF:
3827 case Op_PackL:
3828 case Op_PackD:
3829 if (n->req()-1 > 2) {
3830 // Replace many operand PackNodes with a binary tree for matching
3831 PackNode* p = (PackNode*) n;
3832 Node* btp = p->binary_tree_pack(1, n->req());
3833 n->subsume_by(btp, this);
3834 }
3835 break;
3836 case Op_Loop:
3837 case Op_CountedLoop:
3838 case Op_OuterStripMinedLoop:
3839 if (n->as_Loop()->is_inner_loop()) {
3840 frc.inc_inner_loop_count();
3841 }
3842 n->as_Loop()->verify_strip_mined(0);
3843 break;
3844 case Op_LShiftI:
3845 case Op_RShiftI:
3846 case Op_URShiftI:
3847 case Op_LShiftL:
3848 case Op_RShiftL:
3849 case Op_URShiftL:
3850 if (Matcher::need_masked_shift_count) {
3851 // The cpu's shift instructions don't restrict the count to the
3852 // lower 5/6 bits. We need to do the masking ourselves.
3853 Node* in2 = n->in(2);
3854 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3855 const TypeInt* t = in2->find_int_type();
3856 if (t != NULL && t->is_con()) {
3857 juint shift = t->get_con();
3858 if (shift > mask) { // Unsigned cmp
3859 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3860 }
3861 } else {
3862 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3863 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3864 n->set_req(2, shift);
3865 }
3866 }
3867 if (in2->outcnt() == 0) { // Remove dead node
3868 in2->disconnect_inputs(NULL, this);
3869 }
3870 }
3871 break;
3872 case Op_MemBarStoreStore:
3873 case Op_MemBarRelease:
3874 // Break the link with AllocateNode: it is no longer useful and
3875 // confuses register allocation.
3876 if (n->req() > MemBarNode::Precedent) {
3877 n->set_req(MemBarNode::Precedent, top());
3878 }
3879 break;
3880 case Op_RangeCheck: {
3881 RangeCheckNode* rc = n->as_RangeCheck();
3882 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3883 n->subsume_by(iff, this);
3884 frc._tests.push(iff);
3885 break;
3886 }
3887 case Op_ConvI2L: {
3888 if (!Matcher::convi2l_type_required) {
3889 // Code generation on some platforms doesn't need accurate
3890 // ConvI2L types. Widening the type can help remove redundant
3891 // address computations.
3892 n->as_Type()->set_type(TypeLong::INT);
3893 ResourceMark rm;
3894 Node_List wq;
3895 wq.push(n);
3896 for (uint next = 0; next < wq.size(); next++) {
3897 Node *m = wq.at(next);
3898
3899 for(;;) {
3900 // Loop over all nodes with identical inputs edges as m
3901 Node* k = m->find_similar(m->Opcode());
3902 if (k == NULL) {
3903 break;
3904 }
3905 // Push their uses so we get a chance to remove node made
3906 // redundant
3907 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3908 Node* u = k->fast_out(i);
3909 assert(!wq.contains(u), "shouldn't process one node several times");
3910 if (u->Opcode() == Op_LShiftL ||
3911 u->Opcode() == Op_AddL ||
3912 u->Opcode() == Op_SubL ||
3913 u->Opcode() == Op_AddP) {
3914 wq.push(u);
3915 }
3916 }
3917 // Replace all nodes with identical edges as m with m
3918 k->subsume_by(m, this);
3919 }
3920 }
3921 }
3922 break;
3923 }
3924 case Op_CmpUL: {
3925 if (!Matcher::has_match_rule(Op_CmpUL)) {
3926 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3927 // to max_julong if < 0 to make the signed comparison fail.
3928 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3929 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3930 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3931 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3932 Node* andl = new AndLNode(orl, remove_sign_mask);
3933 Node* cmp = new CmpLNode(andl, n->in(2));
3934 n->subsume_by(cmp, this);
3935 }
3936 break;
3937 }
3938 default:
3939 assert( !n->is_Call(), "" );
3940 assert( !n->is_Mem(), "" );
3941 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3942 break;
3943 }
3944
3945 // Collect CFG split points
3946 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3947 frc._tests.push(n);
3948 }
3949 }
3950
3951 //------------------------------final_graph_reshaping_walk---------------------
3952 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3953 // requires that the walk visits a node's inputs before visiting the node.
3954 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3955 ResourceArea *area = Thread::current()->resource_area();
3956 Unique_Node_List sfpt(area);
3957
3958 frc._visited.set(root->_idx); // first, mark node as visited
3959 uint cnt = root->req();
3960 Node *n = root;
3961 uint i = 0;
3962 while (true) {
3963 if (i < cnt) {
3964 // Place all non-visited non-null inputs onto stack
3965 Node* m = n->in(i);
3966 ++i;
3967 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3968 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3969 // compute worst case interpreter size in case of a deoptimization
3970 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3971
3972 sfpt.push(m);
3973 }
3974 cnt = m->req();
3975 nstack.push(n, i); // put on stack parent and next input's index
3976 n = m;
3977 i = 0;
3978 }
3979 } else {
3980 // Now do post-visit work
3981 final_graph_reshaping_impl( n, frc );
3982 if (nstack.is_empty())
3983 break; // finished
3984 n = nstack.node(); // Get node from stack
3985 cnt = n->req();
3986 i = nstack.index();
3987 nstack.pop(); // Shift to the next node on stack
3988 }
3989 }
3990
3991 // Skip next transformation if compressed oops are not used.
3992 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3993 (!UseCompressedOops && !UseCompressedClassPointers))
3994 return;
3995
3996 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3997 // It could be done for an uncommon traps or any safepoints/calls
3998 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3999 while (sfpt.size() > 0) {
4000 n = sfpt.pop();
4001 JVMState *jvms = n->as_SafePoint()->jvms();
4002 assert(jvms != NULL, "sanity");
4003 int start = jvms->debug_start();
4004 int end = n->req();
4005 bool is_uncommon = (n->is_CallStaticJava() &&
4006 n->as_CallStaticJava()->uncommon_trap_request() != 0);
4007 for (int j = start; j < end; j++) {
4008 Node* in = n->in(j);
4009 if (in->is_DecodeNarrowPtr()) {
4010 bool safe_to_skip = true;
4011 if (!is_uncommon ) {
4012 // Is it safe to skip?
4013 for (uint i = 0; i < in->outcnt(); i++) {
4014 Node* u = in->raw_out(i);
4015 if (!u->is_SafePoint() ||
4016 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
4017 safe_to_skip = false;
4018 }
4019 }
4020 }
4021 if (safe_to_skip) {
4022 n->set_req(j, in->in(1));
4023 }
4024 if (in->outcnt() == 0) {
4025 in->disconnect_inputs(NULL, this);
4026 }
4027 }
4028 }
4029 }
4030 }
4031
4032 //------------------------------final_graph_reshaping--------------------------
4033 // Final Graph Reshaping.
4034 //
4035 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
4036 // and not commoned up and forced early. Must come after regular
4037 // optimizations to avoid GVN undoing the cloning. Clone constant
4038 // inputs to Loop Phis; these will be split by the allocator anyways.
4039 // Remove Opaque nodes.
4040 // (2) Move last-uses by commutative operations to the left input to encourage
4041 // Intel update-in-place two-address operations and better register usage
4042 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
4043 // calls canonicalizing them back.
4044 // (3) Count the number of double-precision FP ops, single-precision FP ops
4045 // and call sites. On Intel, we can get correct rounding either by
4046 // forcing singles to memory (requires extra stores and loads after each
4047 // FP bytecode) or we can set a rounding mode bit (requires setting and
4048 // clearing the mode bit around call sites). The mode bit is only used
4049 // if the relative frequency of single FP ops to calls is low enough.
4050 // This is a key transform for SPEC mpeg_audio.
4051 // (4) Detect infinite loops; blobs of code reachable from above but not
4052 // below. Several of the Code_Gen algorithms fail on such code shapes,
4053 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
4054 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
4055 // Detection is by looking for IfNodes where only 1 projection is
4056 // reachable from below or CatchNodes missing some targets.
4057 // (5) Assert for insane oop offsets in debug mode.
4058
4059 bool Compile::final_graph_reshaping() {
4060 // an infinite loop may have been eliminated by the optimizer,
4061 // in which case the graph will be empty.
4062 if (root()->req() == 1) {
4063 record_method_not_compilable("trivial infinite loop");
4064 return true;
4065 }
4066
4067 // Expensive nodes have their control input set to prevent the GVN
4068 // from freely commoning them. There's no GVN beyond this point so
4069 // no need to keep the control input. We want the expensive nodes to
4070 // be freely moved to the least frequent code path by gcm.
4071 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
4072 for (int i = 0; i < expensive_count(); i++) {
4073 _expensive_nodes->at(i)->set_req(0, NULL);
4074 }
4075
4076 Final_Reshape_Counts frc;
4077
4078 // Visit everybody reachable!
4079 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
4080 Node_Stack nstack(live_nodes() >> 1);
4081 final_graph_reshaping_walk(nstack, root(), frc);
4082
4083 // Check for unreachable (from below) code (i.e., infinite loops).
4084 for( uint i = 0; i < frc._tests.size(); i++ ) {
4085 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
4086 // Get number of CFG targets.
4087 // Note that PCTables include exception targets after calls.
4088 uint required_outcnt = n->required_outcnt();
4089 if (n->outcnt() != required_outcnt) {
4090 // Check for a few special cases. Rethrow Nodes never take the
4091 // 'fall-thru' path, so expected kids is 1 less.
4092 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
4093 if (n->in(0)->in(0)->is_Call()) {
4094 CallNode *call = n->in(0)->in(0)->as_Call();
4095 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
4096 required_outcnt--; // Rethrow always has 1 less kid
4097 } else if (call->req() > TypeFunc::Parms &&
4098 call->is_CallDynamicJava()) {
4099 // Check for null receiver. In such case, the optimizer has
4100 // detected that the virtual call will always result in a null
4101 // pointer exception. The fall-through projection of this CatchNode
4102 // will not be populated.
4103 Node *arg0 = call->in(TypeFunc::Parms);
4104 if (arg0->is_Type() &&
4105 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
4106 required_outcnt--;
4107 }
4108 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
4109 call->req() > TypeFunc::Parms+1 &&
4110 call->is_CallStaticJava()) {
4111 // Check for negative array length. In such case, the optimizer has
4112 // detected that the allocation attempt will always result in an
4113 // exception. There is no fall-through projection of this CatchNode .
4114 Node *arg1 = call->in(TypeFunc::Parms+1);
4115 if (arg1->is_Type() &&
4116 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
4117 required_outcnt--;
4118 }
4119 }
4120 }
4121 }
4122 // Recheck with a better notion of 'required_outcnt'
4123 if (n->outcnt() != required_outcnt) {
4124 record_method_not_compilable("malformed control flow");
4125 return true; // Not all targets reachable!
4126 }
4127 }
4128 // Check that I actually visited all kids. Unreached kids
4129 // must be infinite loops.
4130 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4131 if (!frc._visited.test(n->fast_out(j)->_idx)) {
4132 record_method_not_compilable("infinite loop");
4133 return true; // Found unvisited kid; must be unreach
4134 }
4135
4136 // Here so verification code in final_graph_reshaping_walk()
4137 // always see an OuterStripMinedLoopEnd
4138 if (n->is_OuterStripMinedLoopEnd()) {
4139 IfNode* init_iff = n->as_If();
4140 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4141 n->subsume_by(iff, this);
4142 }
4143 }
4144
4145 // If original bytecodes contained a mixture of floats and doubles
4146 // check if the optimizer has made it homogenous, item (3).
4147 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
4148 frc.get_float_count() > 32 &&
4149 frc.get_double_count() == 0 &&
4150 (10 * frc.get_call_count() < frc.get_float_count()) ) {
4151 set_24_bit_selection_and_mode( false, true );
4152 }
4153
4154 set_java_calls(frc.get_java_call_count());
4155 set_inner_loops(frc.get_inner_loop_count());
4156
4157 // No infinite loops, no reason to bail out.
4158 return false;
4159 }
4160
4161 //-----------------------------too_many_traps----------------------------------
4162 // Report if there are too many traps at the current method and bci.
4163 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4164 bool Compile::too_many_traps(ciMethod* method,
4165 int bci,
4166 Deoptimization::DeoptReason reason) {
4167 ciMethodData* md = method->method_data();
4168 if (md->is_empty()) {
4169 // Assume the trap has not occurred, or that it occurred only
4170 // because of a transient condition during start-up in the interpreter.
4171 return false;
4172 }
4173 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4174 if (md->has_trap_at(bci, m, reason) != 0) {
4175 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4176 // Also, if there are multiple reasons, or if there is no per-BCI record,
4177 // assume the worst.
4178 if (log())
4179 log()->elem("observe trap='%s' count='%d'",
4180 Deoptimization::trap_reason_name(reason),
4181 md->trap_count(reason));
4182 return true;
4183 } else {
4184 // Ignore method/bci and see if there have been too many globally.
4185 return too_many_traps(reason, md);
4186 }
4187 }
4188
4189 // Less-accurate variant which does not require a method and bci.
4190 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4191 ciMethodData* logmd) {
4192 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4193 // Too many traps globally.
4194 // Note that we use cumulative trap_count, not just md->trap_count.
4195 if (log()) {
4196 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4197 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4198 Deoptimization::trap_reason_name(reason),
4199 mcount, trap_count(reason));
4200 }
4201 return true;
4202 } else {
4203 // The coast is clear.
4204 return false;
4205 }
4206 }
4207
4208 //--------------------------too_many_recompiles--------------------------------
4209 // Report if there are too many recompiles at the current method and bci.
4210 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4211 // Is not eager to return true, since this will cause the compiler to use
4212 // Action_none for a trap point, to avoid too many recompilations.
4213 bool Compile::too_many_recompiles(ciMethod* method,
4214 int bci,
4215 Deoptimization::DeoptReason reason) {
4216 ciMethodData* md = method->method_data();
4217 if (md->is_empty()) {
4218 // Assume the trap has not occurred, or that it occurred only
4219 // because of a transient condition during start-up in the interpreter.
4220 return false;
4221 }
4222 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4223 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4224 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
4225 Deoptimization::DeoptReason per_bc_reason
4226 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4227 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4228 if ((per_bc_reason == Deoptimization::Reason_none
4229 || md->has_trap_at(bci, m, reason) != 0)
4230 // The trap frequency measure we care about is the recompile count:
4231 && md->trap_recompiled_at(bci, m)
4232 && md->overflow_recompile_count() >= bc_cutoff) {
4233 // Do not emit a trap here if it has already caused recompilations.
4234 // Also, if there are multiple reasons, or if there is no per-BCI record,
4235 // assume the worst.
4236 if (log())
4237 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4238 Deoptimization::trap_reason_name(reason),
4239 md->trap_count(reason),
4240 md->overflow_recompile_count());
4241 return true;
4242 } else if (trap_count(reason) != 0
4243 && decompile_count() >= m_cutoff) {
4244 // Too many recompiles globally, and we have seen this sort of trap.
4245 // Use cumulative decompile_count, not just md->decompile_count.
4246 if (log())
4247 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4248 Deoptimization::trap_reason_name(reason),
4249 md->trap_count(reason), trap_count(reason),
4250 md->decompile_count(), decompile_count());
4251 return true;
4252 } else {
4253 // The coast is clear.
4254 return false;
4255 }
4256 }
4257
4258 // Compute when not to trap. Used by matching trap based nodes and
4259 // NullCheck optimization.
4260 void Compile::set_allowed_deopt_reasons() {
4261 _allowed_reasons = 0;
4262 if (is_method_compilation()) {
4263 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4264 assert(rs < BitsPerInt, "recode bit map");
4265 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4266 _allowed_reasons |= nth_bit(rs);
4267 }
4268 }
4269 }
4270 }
4271
4272 #ifndef PRODUCT
4273 //------------------------------verify_graph_edges---------------------------
4274 // Walk the Graph and verify that there is a one-to-one correspondence
4275 // between Use-Def edges and Def-Use edges in the graph.
4276 void Compile::verify_graph_edges(bool no_dead_code) {
4277 if (VerifyGraphEdges) {
4278 ResourceArea *area = Thread::current()->resource_area();
4279 Unique_Node_List visited(area);
4280 // Call recursive graph walk to check edges
4281 _root->verify_edges(visited);
4282 if (no_dead_code) {
4283 // Now make sure that no visited node is used by an unvisited node.
4284 bool dead_nodes = false;
4285 Unique_Node_List checked(area);
4286 while (visited.size() > 0) {
4287 Node* n = visited.pop();
4288 checked.push(n);
4289 for (uint i = 0; i < n->outcnt(); i++) {
4290 Node* use = n->raw_out(i);
4291 if (checked.member(use)) continue; // already checked
4292 if (visited.member(use)) continue; // already in the graph
4293 if (use->is_Con()) continue; // a dead ConNode is OK
4294 // At this point, we have found a dead node which is DU-reachable.
4295 if (!dead_nodes) {
4296 tty->print_cr("*** Dead nodes reachable via DU edges:");
4297 dead_nodes = true;
4298 }
4299 use->dump(2);
4300 tty->print_cr("---");
4301 checked.push(use); // No repeats; pretend it is now checked.
4302 }
4303 }
4304 assert(!dead_nodes, "using nodes must be reachable from root");
4305 }
4306 }
4307 }
4308
4309 // Verify GC barriers consistency
4310 // Currently supported:
4311 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
4312 void Compile::verify_barriers() {
4313 #if INCLUDE_G1GC
4314 if (UseG1GC) {
4315 // Verify G1 pre-barriers
4316 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
4317
4318 ResourceArea *area = Thread::current()->resource_area();
4319 Unique_Node_List visited(area);
4320 Node_List worklist(area);
4321 // We're going to walk control flow backwards starting from the Root
4322 worklist.push(_root);
4323 while (worklist.size() > 0) {
4324 Node* x = worklist.pop();
4325 if (x == NULL || x == top()) continue;
4326 if (visited.member(x)) {
4327 continue;
4328 } else {
4329 visited.push(x);
4330 }
4331
4332 if (x->is_Region()) {
4333 for (uint i = 1; i < x->req(); i++) {
4334 worklist.push(x->in(i));
4335 }
4336 } else {
4337 worklist.push(x->in(0));
4338 // We are looking for the pattern:
4339 // /->ThreadLocal
4340 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
4341 // \->ConI(0)
4342 // We want to verify that the If and the LoadB have the same control
4343 // See GraphKit::g1_write_barrier_pre()
4344 if (x->is_If()) {
4345 IfNode *iff = x->as_If();
4346 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
4347 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
4348 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
4349 && cmp->in(1)->is_Load()) {
4350 LoadNode* load = cmp->in(1)->as_Load();
4351 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
4352 && load->in(2)->in(3)->is_Con()
4353 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
4354
4355 Node* if_ctrl = iff->in(0);
4356 Node* load_ctrl = load->in(0);
4357
4358 if (if_ctrl != load_ctrl) {
4359 // Skip possible CProj->NeverBranch in infinite loops
4360 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
4361 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
4362 if_ctrl = if_ctrl->in(0)->in(0);
4363 }
4364 }
4365 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
4366 }
4367 }
4368 }
4369 }
4370 }
4371 }
4372 }
4373 #endif
4374 }
4375
4376 #endif
4377
4378 // The Compile object keeps track of failure reasons separately from the ciEnv.
4379 // This is required because there is not quite a 1-1 relation between the
4380 // ciEnv and its compilation task and the Compile object. Note that one
4381 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4382 // to backtrack and retry without subsuming loads. Other than this backtracking
4383 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4384 // by the logic in C2Compiler.
4385 void Compile::record_failure(const char* reason) {
4386 if (log() != NULL) {
4387 log()->elem("failure reason='%s' phase='compile'", reason);
4388 }
4389 if (_failure_reason == NULL) {
4390 // Record the first failure reason.
4391 _failure_reason = reason;
4392 }
4393
4394 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4395 C->print_method(PHASE_FAILURE);
4396 }
4397 _root = NULL; // flush the graph, too
4398 }
4399
4400 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4401 : TraceTime(name, accumulator, CITime, CITimeVerbose),
4402 _phase_name(name), _dolog(CITimeVerbose)
4403 {
4404 if (_dolog) {
4405 C = Compile::current();
4406 _log = C->log();
4407 } else {
4408 C = NULL;
4409 _log = NULL;
4410 }
4411 if (_log != NULL) {
4412 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4413 _log->stamp();
4414 _log->end_head();
4415 }
4416 }
4417
4418 Compile::TracePhase::~TracePhase() {
4419
4420 C = Compile::current();
4421 if (_dolog) {
4422 _log = C->log();
4423 } else {
4424 _log = NULL;
4425 }
4426
4427 #ifdef ASSERT
4428 if (PrintIdealNodeCount) {
4429 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4430 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4431 }
4432
4433 if (VerifyIdealNodeCount) {
4434 Compile::current()->print_missing_nodes();
4435 }
4436 #endif
4437
4438 if (_log != NULL) {
4439 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4440 }
4441 }
4442
4443 //=============================================================================
4444 // Two Constant's are equal when the type and the value are equal.
4445 bool Compile::Constant::operator==(const Constant& other) {
4446 if (type() != other.type() ) return false;
4447 if (can_be_reused() != other.can_be_reused()) return false;
4448 // For floating point values we compare the bit pattern.
4449 switch (type()) {
4450 case T_INT:
4451 case T_FLOAT: return (_v._value.i == other._v._value.i);
4452 case T_LONG:
4453 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4454 case T_OBJECT:
4455 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4456 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4457 case T_METADATA: return (_v._metadata == other._v._metadata);
4458 default: ShouldNotReachHere(); return false;
4459 }
4460 }
4461
4462 static int type_to_size_in_bytes(BasicType t) {
4463 switch (t) {
4464 case T_INT: return sizeof(jint );
4465 case T_LONG: return sizeof(jlong );
4466 case T_FLOAT: return sizeof(jfloat );
4467 case T_DOUBLE: return sizeof(jdouble);
4468 case T_METADATA: return sizeof(Metadata*);
4469 // We use T_VOID as marker for jump-table entries (labels) which
4470 // need an internal word relocation.
4471 case T_VOID:
4472 case T_ADDRESS:
4473 case T_OBJECT: return sizeof(jobject);
4474 default:
4475 ShouldNotReachHere();
4476 return -1;
4477 }
4478 }
4479
4480 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4481 // sort descending
4482 if (a->freq() > b->freq()) return -1;
4483 if (a->freq() < b->freq()) return 1;
4484 return 0;
4485 }
4486
4487 void Compile::ConstantTable::calculate_offsets_and_size() {
4488 // First, sort the array by frequencies.
4489 _constants.sort(qsort_comparator);
4490
4491 #ifdef ASSERT
4492 // Make sure all jump-table entries were sorted to the end of the
4493 // array (they have a negative frequency).
4494 bool found_void = false;
4495 for (int i = 0; i < _constants.length(); i++) {
4496 Constant con = _constants.at(i);
4497 if (con.type() == T_VOID)
4498 found_void = true; // jump-tables
4499 else
4500 assert(!found_void, "wrong sorting");
4501 }
4502 #endif
4503
4504 int offset = 0;
4505 for (int i = 0; i < _constants.length(); i++) {
4506 Constant* con = _constants.adr_at(i);
4507
4508 // Align offset for type.
4509 int typesize = type_to_size_in_bytes(con->type());
4510 offset = align_up(offset, typesize);
4511 con->set_offset(offset); // set constant's offset
4512
4513 if (con->type() == T_VOID) {
4514 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4515 offset = offset + typesize * n->outcnt(); // expand jump-table
4516 } else {
4517 offset = offset + typesize;
4518 }
4519 }
4520
4521 // Align size up to the next section start (which is insts; see
4522 // CodeBuffer::align_at_start).
4523 assert(_size == -1, "already set?");
4524 _size = align_up(offset, (int)CodeEntryAlignment);
4525 }
4526
4527 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4528 MacroAssembler _masm(&cb);
4529 for (int i = 0; i < _constants.length(); i++) {
4530 Constant con = _constants.at(i);
4531 address constant_addr = NULL;
4532 switch (con.type()) {
4533 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4534 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4535 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4536 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4537 case T_OBJECT: {
4538 jobject obj = con.get_jobject();
4539 int oop_index = _masm.oop_recorder()->find_index(obj);
4540 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4541 break;
4542 }
4543 case T_ADDRESS: {
4544 address addr = (address) con.get_jobject();
4545 constant_addr = _masm.address_constant(addr);
4546 break;
4547 }
4548 // We use T_VOID as marker for jump-table entries (labels) which
4549 // need an internal word relocation.
4550 case T_VOID: {
4551 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4552 // Fill the jump-table with a dummy word. The real value is
4553 // filled in later in fill_jump_table.
4554 address dummy = (address) n;
4555 constant_addr = _masm.address_constant(dummy);
4556 // Expand jump-table
4557 for (uint i = 1; i < n->outcnt(); i++) {
4558 address temp_addr = _masm.address_constant(dummy + i);
4559 assert(temp_addr, "consts section too small");
4560 }
4561 break;
4562 }
4563 case T_METADATA: {
4564 Metadata* obj = con.get_metadata();
4565 int metadata_index = _masm.oop_recorder()->find_index(obj);
4566 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4567 break;
4568 }
4569 default: ShouldNotReachHere();
4570 }
4571 assert(constant_addr, "consts section too small");
4572 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4573 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4574 }
4575 }
4576
4577 int Compile::ConstantTable::find_offset(Constant& con) const {
4578 int idx = _constants.find(con);
4579 guarantee(idx != -1, "constant must be in constant table");
4580 int offset = _constants.at(idx).offset();
4581 guarantee(offset != -1, "constant table not emitted yet?");
4582 return offset;
4583 }
4584
4585 void Compile::ConstantTable::add(Constant& con) {
4586 if (con.can_be_reused()) {
4587 int idx = _constants.find(con);
4588 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4589 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4590 return;
4591 }
4592 }
4593 (void) _constants.append(con);
4594 }
4595
4596 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4597 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4598 Constant con(type, value, b->_freq);
4599 add(con);
4600 return con;
4601 }
4602
4603 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4604 Constant con(metadata);
4605 add(con);
4606 return con;
4607 }
4608
4609 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4610 jvalue value;
4611 BasicType type = oper->type()->basic_type();
4612 switch (type) {
4613 case T_LONG: value.j = oper->constantL(); break;
4614 case T_FLOAT: value.f = oper->constantF(); break;
4615 case T_DOUBLE: value.d = oper->constantD(); break;
4616 case T_OBJECT:
4617 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4618 case T_METADATA: return add((Metadata*)oper->constant()); break;
4619 default: guarantee(false, "unhandled type: %s", type2name(type));
4620 }
4621 return add(n, type, value);
4622 }
4623
4624 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4625 jvalue value;
4626 // We can use the node pointer here to identify the right jump-table
4627 // as this method is called from Compile::Fill_buffer right before
4628 // the MachNodes are emitted and the jump-table is filled (means the
4629 // MachNode pointers do not change anymore).
4630 value.l = (jobject) n;
4631 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4632 add(con);
4633 return con;
4634 }
4635
4636 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4637 // If called from Compile::scratch_emit_size do nothing.
4638 if (Compile::current()->in_scratch_emit_size()) return;
4639
4640 assert(labels.is_nonempty(), "must be");
4641 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4642
4643 // Since MachConstantNode::constant_offset() also contains
4644 // table_base_offset() we need to subtract the table_base_offset()
4645 // to get the plain offset into the constant table.
4646 int offset = n->constant_offset() - table_base_offset();
4647
4648 MacroAssembler _masm(&cb);
4649 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4650
4651 for (uint i = 0; i < n->outcnt(); i++) {
4652 address* constant_addr = &jump_table_base[i];
4653 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));
4654 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4655 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4656 }
4657 }
4658
4659 //----------------------------static_subtype_check-----------------------------
4660 // Shortcut important common cases when superklass is exact:
4661 // (0) superklass is java.lang.Object (can occur in reflective code)
4662 // (1) subklass is already limited to a subtype of superklass => always ok
4663 // (2) subklass does not overlap with superklass => always fail
4664 // (3) superklass has NO subtypes and we can check with a simple compare.
4665 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4666 if (StressReflectiveCode) {
4667 return SSC_full_test; // Let caller generate the general case.
4668 }
4669
4670 if (superk == env()->Object_klass()) {
4671 return SSC_always_true; // (0) this test cannot fail
4672 }
4673
4674 ciType* superelem = superk;
4675 if (superelem->is_array_klass())
4676 superelem = superelem->as_array_klass()->base_element_type();
4677
4678 if (!subk->is_interface()) { // cannot trust static interface types yet
4679 if (subk->is_subtype_of(superk)) {
4680 return SSC_always_true; // (1) false path dead; no dynamic test needed
4681 }
4682 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4683 !superk->is_subtype_of(subk)) {
4684 return SSC_always_false;
4685 }
4686 }
4687
4688 // If casting to an instance klass, it must have no subtypes
4689 if (superk->is_interface()) {
4690 // Cannot trust interfaces yet.
4691 // %%% S.B. superk->nof_implementors() == 1
4692 } else if (superelem->is_instance_klass()) {
4693 ciInstanceKlass* ik = superelem->as_instance_klass();
4694 if (!ik->has_subklass() && !ik->is_interface()) {
4695 if (!ik->is_final()) {
4696 // Add a dependency if there is a chance of a later subclass.
4697 dependencies()->assert_leaf_type(ik);
4698 }
4699 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4700 }
4701 } else {
4702 // A primitive array type has no subtypes.
4703 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4704 }
4705
4706 return SSC_full_test;
4707 }
4708
4709 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4710 #ifdef _LP64
4711 // The scaled index operand to AddP must be a clean 64-bit value.
4712 // Java allows a 32-bit int to be incremented to a negative
4713 // value, which appears in a 64-bit register as a large
4714 // positive number. Using that large positive number as an
4715 // operand in pointer arithmetic has bad consequences.
4716 // On the other hand, 32-bit overflow is rare, and the possibility
4717 // can often be excluded, if we annotate the ConvI2L node with
4718 // a type assertion that its value is known to be a small positive
4719 // number. (The prior range check has ensured this.)
4720 // This assertion is used by ConvI2LNode::Ideal.
4721 int index_max = max_jint - 1; // array size is max_jint, index is one less
4722 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4723 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4724 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4725 #endif
4726 return idx;
4727 }
4728
4729 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4730 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4731 if (ctrl != NULL) {
4732 // Express control dependency by a CastII node with a narrow type.
4733 value = new CastIINode(value, itype, false, true /* range check dependency */);
4734 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4735 // node from floating above the range check during loop optimizations. Otherwise, the
4736 // ConvI2L node may be eliminated independently of the range check, causing the data path
4737 // to become TOP while the control path is still there (although it's unreachable).
4738 value->set_req(0, ctrl);
4739 // Save CastII node to remove it after loop optimizations.
4740 phase->C->add_range_check_cast(value);
4741 value = phase->transform(value);
4742 }
4743 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4744 return phase->transform(new ConvI2LNode(value, ltype));
4745 }
4746
4747 // The message about the current inlining is accumulated in
4748 // _print_inlining_stream and transfered into the _print_inlining_list
4749 // once we know whether inlining succeeds or not. For regular
4750 // inlining, messages are appended to the buffer pointed by
4751 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4752 // a new buffer is added after _print_inlining_idx in the list. This
4753 // way we can update the inlining message for late inlining call site
4754 // when the inlining is attempted again.
4755 void Compile::print_inlining_init() {
4756 if (print_inlining() || print_intrinsics()) {
4757 _print_inlining_stream = new stringStream();
4758 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4759 }
4760 }
4761
4762 void Compile::print_inlining_reinit() {
4763 if (print_inlining() || print_intrinsics()) {
4764 // Re allocate buffer when we change ResourceMark
4765 _print_inlining_stream = new stringStream();
4766 }
4767 }
4768
4769 void Compile::print_inlining_reset() {
4770 _print_inlining_stream->reset();
4771 }
4772
4773 void Compile::print_inlining_commit() {
4774 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4775 // Transfer the message from _print_inlining_stream to the current
4776 // _print_inlining_list buffer and clear _print_inlining_stream.
4777 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4778 print_inlining_reset();
4779 }
4780
4781 void Compile::print_inlining_push() {
4782 // Add new buffer to the _print_inlining_list at current position
4783 _print_inlining_idx++;
4784 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4785 }
4786
4787 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4788 return _print_inlining_list->at(_print_inlining_idx);
4789 }
4790
4791 void Compile::print_inlining_update(CallGenerator* cg) {
4792 if (print_inlining() || print_intrinsics()) {
4793 if (!cg->is_late_inline()) {
4794 if (print_inlining_current().cg() != NULL) {
4795 print_inlining_push();
4796 }
4797 print_inlining_commit();
4798 } else {
4799 if (print_inlining_current().cg() != cg &&
4800 (print_inlining_current().cg() != NULL ||
4801 print_inlining_current().ss()->size() != 0)) {
4802 print_inlining_push();
4803 }
4804 print_inlining_commit();
4805 print_inlining_current().set_cg(cg);
4806 }
4807 }
4808 }
4809
4810 void Compile::print_inlining_move_to(CallGenerator* cg) {
4811 // We resume inlining at a late inlining call site. Locate the
4812 // corresponding inlining buffer so that we can update it.
4813 if (print_inlining()) {
4814 for (int i = 0; i < _print_inlining_list->length(); i++) {
4815 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4816 _print_inlining_idx = i;
4817 return;
4818 }
4819 }
4820 ShouldNotReachHere();
4821 }
4822 }
4823
4824 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4825 if (print_inlining()) {
4826 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4827 assert(print_inlining_current().cg() == cg, "wrong entry");
4828 // replace message with new message
4829 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4830 print_inlining_commit();
4831 print_inlining_current().set_cg(cg);
4832 }
4833 }
4834
4835 void Compile::print_inlining_assert_ready() {
4836 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4837 }
4838
4839 void Compile::process_print_inlining() {
4840 bool do_print_inlining = print_inlining() || print_intrinsics();
4841 if (do_print_inlining || log() != NULL) {
4842 // Print inlining message for candidates that we couldn't inline
4843 // for lack of space
4844 for (int i = 0; i < _late_inlines.length(); i++) {
4845 CallGenerator* cg = _late_inlines.at(i);
4846 if (!cg->is_mh_late_inline()) {
4847 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4848 if (do_print_inlining) {
4849 cg->print_inlining_late(msg);
4850 }
4851 log_late_inline_failure(cg, msg);
4852 }
4853 }
4854 }
4855 if (do_print_inlining) {
4856 ResourceMark rm;
4857 stringStream ss;
4858 for (int i = 0; i < _print_inlining_list->length(); i++) {
4859 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4860 }
4861 size_t end = ss.size();
4862 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4863 strncpy(_print_inlining_output, ss.base(), end+1);
4864 _print_inlining_output[end] = 0;
4865 }
4866 }
4867
4868 void Compile::dump_print_inlining() {
4869 if (_print_inlining_output != NULL) {
4870 tty->print_raw(_print_inlining_output);
4871 }
4872 }
4873
4874 void Compile::log_late_inline(CallGenerator* cg) {
4875 if (log() != NULL) {
4876 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4877 cg->unique_id());
4878 JVMState* p = cg->call_node()->jvms();
4879 while (p != NULL) {
4880 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4881 p = p->caller();
4882 }
4883 log()->tail("late_inline");
4884 }
4885 }
4886
4887 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4888 log_late_inline(cg);
4889 if (log() != NULL) {
4890 log()->inline_fail(msg);
4891 }
4892 }
4893
4894 void Compile::log_inline_id(CallGenerator* cg) {
4895 if (log() != NULL) {
4896 // The LogCompilation tool needs a unique way to identify late
4897 // inline call sites. This id must be unique for this call site in
4898 // this compilation. Try to have it unique across compilations as
4899 // well because it can be convenient when grepping through the log
4900 // file.
4901 // Distinguish OSR compilations from others in case CICountOSR is
4902 // on.
4903 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4904 cg->set_unique_id(id);
4905 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4906 }
4907 }
4908
4909 void Compile::log_inline_failure(const char* msg) {
4910 if (C->log() != NULL) {
4911 C->log()->inline_fail(msg);
4912 }
4913 }
4914
4915
4916 // Dump inlining replay data to the stream.
4917 // Don't change thread state and acquire any locks.
4918 void Compile::dump_inline_data(outputStream* out) {
4919 InlineTree* inl_tree = ilt();
4920 if (inl_tree != NULL) {
4921 out->print(" inline %d", inl_tree->count());
4922 inl_tree->dump_replay_data(out);
4923 }
4924 }
4925
4926 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4927 if (n1->Opcode() < n2->Opcode()) return -1;
4928 else if (n1->Opcode() > n2->Opcode()) return 1;
4929
4930 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4931 for (uint i = 1; i < n1->req(); i++) {
4932 if (n1->in(i) < n2->in(i)) return -1;
4933 else if (n1->in(i) > n2->in(i)) return 1;
4934 }
4935
4936 return 0;
4937 }
4938
4939 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4940 Node* n1 = *n1p;
4941 Node* n2 = *n2p;
4942
4943 return cmp_expensive_nodes(n1, n2);
4944 }
4945
4946 void Compile::sort_expensive_nodes() {
4947 if (!expensive_nodes_sorted()) {
4948 _expensive_nodes->sort(cmp_expensive_nodes);
4949 }
4950 }
4951
4952 bool Compile::expensive_nodes_sorted() const {
4953 for (int i = 1; i < _expensive_nodes->length(); i++) {
4954 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4955 return false;
4956 }
4957 }
4958 return true;
4959 }
4960
4961 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4962 if (_expensive_nodes->length() == 0) {
4963 return false;
4964 }
4965
4966 assert(OptimizeExpensiveOps, "optimization off?");
4967
4968 // Take this opportunity to remove dead nodes from the list
4969 int j = 0;
4970 for (int i = 0; i < _expensive_nodes->length(); i++) {
4971 Node* n = _expensive_nodes->at(i);
4972 if (!n->is_unreachable(igvn)) {
4973 assert(n->is_expensive(), "should be expensive");
4974 _expensive_nodes->at_put(j, n);
4975 j++;
4976 }
4977 }
4978 _expensive_nodes->trunc_to(j);
4979
4980 // Then sort the list so that similar nodes are next to each other
4981 // and check for at least two nodes of identical kind with same data
4982 // inputs.
4983 sort_expensive_nodes();
4984
4985 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4986 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4987 return true;
4988 }
4989 }
4990
4991 return false;
4992 }
4993
4994 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4995 if (_expensive_nodes->length() == 0) {
4996 return;
4997 }
4998
4999 assert(OptimizeExpensiveOps, "optimization off?");
5000
5001 // Sort to bring similar nodes next to each other and clear the
5002 // control input of nodes for which there's only a single copy.
5003 sort_expensive_nodes();
5004
5005 int j = 0;
5006 int identical = 0;
5007 int i = 0;
5008 bool modified = false;
5009 for (; i < _expensive_nodes->length()-1; i++) {
5010 assert(j <= i, "can't write beyond current index");
5011 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
5012 identical++;
5013 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5014 continue;
5015 }
5016 if (identical > 0) {
5017 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5018 identical = 0;
5019 } else {
5020 Node* n = _expensive_nodes->at(i);
5021 igvn.replace_input_of(n, 0, NULL);
5022 igvn.hash_insert(n);
5023 modified = true;
5024 }
5025 }
5026 if (identical > 0) {
5027 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
5028 } else if (_expensive_nodes->length() >= 1) {
5029 Node* n = _expensive_nodes->at(i);
5030 igvn.replace_input_of(n, 0, NULL);
5031 igvn.hash_insert(n);
5032 modified = true;
5033 }
5034 _expensive_nodes->trunc_to(j);
5035 if (modified) {
5036 igvn.optimize();
5037 }
5038 }
5039
5040 void Compile::add_expensive_node(Node * n) {
5041 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
5042 assert(n->is_expensive(), "expensive nodes with non-null control here only");
5043 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
5044 if (OptimizeExpensiveOps) {
5045 _expensive_nodes->append(n);
5046 } else {
5047 // Clear control input and let IGVN optimize expensive nodes if
5048 // OptimizeExpensiveOps is off.
5049 n->set_req(0, NULL);
5050 }
5051 }
5052
5053 /**
5054 * Remove the speculative part of types and clean up the graph
5055 */
5056 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
5057 if (UseTypeSpeculation) {
5058 Unique_Node_List worklist;
5059 worklist.push(root());
5060 int modified = 0;
5061 // Go over all type nodes that carry a speculative type, drop the
5062 // speculative part of the type and enqueue the node for an igvn
5063 // which may optimize it out.
5064 for (uint next = 0; next < worklist.size(); ++next) {
5065 Node *n = worklist.at(next);
5066 if (n->is_Type()) {
5067 TypeNode* tn = n->as_Type();
5068 const Type* t = tn->type();
5069 const Type* t_no_spec = t->remove_speculative();
5070 if (t_no_spec != t) {
5071 bool in_hash = igvn.hash_delete(n);
5072 assert(in_hash, "node should be in igvn hash table");
5073 tn->set_type(t_no_spec);
5074 igvn.hash_insert(n);
5075 igvn._worklist.push(n); // give it a chance to go away
5076 modified++;
5077 }
5078 }
5079 uint max = n->len();
5080 for( uint i = 0; i < max; ++i ) {
5081 Node *m = n->in(i);
5082 if (not_a_node(m)) continue;
5083 worklist.push(m);
5084 }
5085 }
5086 // Drop the speculative part of all types in the igvn's type table
5087 igvn.remove_speculative_types();
5088 if (modified > 0) {
5089 igvn.optimize();
5090 }
5091 #ifdef ASSERT
5092 // Verify that after the IGVN is over no speculative type has resurfaced
5093 worklist.clear();
5094 worklist.push(root());
5095 for (uint next = 0; next < worklist.size(); ++next) {
5096 Node *n = worklist.at(next);
5097 const Type* t = igvn.type_or_null(n);
5098 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
5099 if (n->is_Type()) {
5100 t = n->as_Type()->type();
5101 assert(t == t->remove_speculative(), "no more speculative types");
5102 }
5103 uint max = n->len();
5104 for( uint i = 0; i < max; ++i ) {
5105 Node *m = n->in(i);
5106 if (not_a_node(m)) continue;
5107 worklist.push(m);
5108 }
5109 }
5110 igvn.check_no_speculative_types();
5111 #endif
5112 }
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 }