1 /*
2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/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
80
81 // -------------------- Compile::mach_constant_base_node -----------------------
82 // Constant table base node singleton.
mach_constant_base_node()83 MachConstantBaseNode* Compile::mach_constant_base_node() {
84 if (_mach_constant_base_node == NULL) {
85 _mach_constant_base_node = new MachConstantBaseNode();
86 _mach_constant_base_node->add_req(C->root());
87 }
88 return _mach_constant_base_node;
89 }
90
91
92 /// Support for intrinsics.
93
94 // Return the index at which m must be inserted (or already exists).
95 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
96 class IntrinsicDescPair {
97 private:
98 ciMethod* _m;
99 bool _is_virtual;
100 public:
IntrinsicDescPair(ciMethod * m,bool is_virtual)101 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
compare(IntrinsicDescPair * const & key,CallGenerator * const & elt)102 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
103 ciMethod* m= elt->method();
104 ciMethod* key_m = key->_m;
105 if (key_m < m) return -1;
106 else if (key_m > m) return 1;
107 else {
108 bool is_virtual = elt->is_virtual();
109 bool key_virtual = key->_is_virtual;
110 if (key_virtual < is_virtual) return -1;
111 else if (key_virtual > is_virtual) return 1;
112 else return 0;
113 }
114 }
115 };
intrinsic_insertion_index(ciMethod * m,bool is_virtual,bool & found)116 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
117 #ifdef ASSERT
118 for (int i = 1; i < _intrinsics->length(); i++) {
119 CallGenerator* cg1 = _intrinsics->at(i-1);
120 CallGenerator* cg2 = _intrinsics->at(i);
121 assert(cg1->method() != cg2->method()
122 ? cg1->method() < cg2->method()
123 : cg1->is_virtual() < cg2->is_virtual(),
124 "compiler intrinsics list must stay sorted");
125 }
126 #endif
127 IntrinsicDescPair pair(m, is_virtual);
128 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
129 }
130
register_intrinsic(CallGenerator * cg)131 void Compile::register_intrinsic(CallGenerator* cg) {
132 if (_intrinsics == NULL) {
133 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
134 }
135 int len = _intrinsics->length();
136 bool found = false;
137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
138 assert(!found, "registering twice");
139 _intrinsics->insert_before(index, cg);
140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
141 }
142
find_intrinsic(ciMethod * m,bool is_virtual)143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
144 assert(m->is_loaded(), "don't try this on unloaded methods");
145 if (_intrinsics != NULL) {
146 bool found = false;
147 int index = intrinsic_insertion_index(m, is_virtual, found);
148 if (found) {
149 return _intrinsics->at(index);
150 }
151 }
152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
153 if (m->intrinsic_id() != vmIntrinsics::_none &&
154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
156 if (cg != NULL) {
157 // Save it for next time:
158 register_intrinsic(cg);
159 return cg;
160 } else {
161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
162 }
163 }
164 return NULL;
165 }
166
167 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
168 // in library_call.cpp.
169
170
171 #ifndef PRODUCT
172 // statistics gathering...
173
174 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
175 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
176
gather_intrinsic_statistics(vmIntrinsics::ID id,bool is_virtual,int flags)177 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
178 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
179 int oflags = _intrinsic_hist_flags[id];
180 assert(flags != 0, "what happened?");
181 if (is_virtual) {
182 flags |= _intrinsic_virtual;
183 }
184 bool changed = (flags != oflags);
185 if ((flags & _intrinsic_worked) != 0) {
186 juint count = (_intrinsic_hist_count[id] += 1);
187 if (count == 1) {
188 changed = true; // first time
189 }
190 // increment the overall count also:
191 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
192 }
193 if (changed) {
194 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
195 // Something changed about the intrinsic's virtuality.
196 if ((flags & _intrinsic_virtual) != 0) {
197 // This is the first use of this intrinsic as a virtual call.
198 if (oflags != 0) {
199 // We already saw it as a non-virtual, so note both cases.
200 flags |= _intrinsic_both;
201 }
202 } else if ((oflags & _intrinsic_both) == 0) {
203 // This is the first use of this intrinsic as a non-virtual
204 flags |= _intrinsic_both;
205 }
206 }
207 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
208 }
209 // update the overall flags also:
210 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
211 return changed;
212 }
213
format_flags(int flags,char * buf)214 static char* format_flags(int flags, char* buf) {
215 buf[0] = 0;
216 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
217 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
218 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
219 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
220 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
221 if (buf[0] == 0) strcat(buf, ",");
222 assert(buf[0] == ',', "must be");
223 return &buf[1];
224 }
225
print_intrinsic_statistics()226 void Compile::print_intrinsic_statistics() {
227 char flagsbuf[100];
228 ttyLocker ttyl;
229 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
230 tty->print_cr("Compiler intrinsic usage:");
231 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
232 if (total == 0) total = 1; // avoid div0 in case of no successes
233 #define PRINT_STAT_LINE(name, c, f) \
234 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
235 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
236 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
237 int flags = _intrinsic_hist_flags[id];
238 juint count = _intrinsic_hist_count[id];
239 if ((flags | count) != 0) {
240 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
241 }
242 }
243 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
244 if (xtty != NULL) xtty->tail("statistics");
245 }
246
print_statistics()247 void Compile::print_statistics() {
248 { ttyLocker ttyl;
249 if (xtty != NULL) xtty->head("statistics type='opto'");
250 Parse::print_statistics();
251 PhaseCCP::print_statistics();
252 PhaseRegAlloc::print_statistics();
253 Scheduling::print_statistics();
254 PhasePeephole::print_statistics();
255 PhaseIdealLoop::print_statistics();
256 if (xtty != NULL) xtty->tail("statistics");
257 }
258 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
259 // put this under its own <statistics> element.
260 print_intrinsic_statistics();
261 }
262 }
263 #endif //PRODUCT
264
265 // Support for bundling info
node_bundling(const Node * n)266 Bundle* Compile::node_bundling(const Node *n) {
267 assert(valid_bundle_info(n), "oob");
268 return &_node_bundling_base[n->_idx];
269 }
270
valid_bundle_info(const Node * n)271 bool Compile::valid_bundle_info(const Node *n) {
272 return (_node_bundling_limit > n->_idx);
273 }
274
275
gvn_replace_by(Node * n,Node * nn)276 void Compile::gvn_replace_by(Node* n, Node* nn) {
277 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
278 Node* use = n->last_out(i);
279 bool is_in_table = initial_gvn()->hash_delete(use);
280 uint uses_found = 0;
281 for (uint j = 0; j < use->len(); j++) {
282 if (use->in(j) == n) {
283 if (j < use->req())
284 use->set_req(j, nn);
285 else
286 use->set_prec(j, nn);
287 uses_found++;
288 }
289 }
290 if (is_in_table) {
291 // reinsert into table
292 initial_gvn()->hash_find_insert(use);
293 }
294 record_for_igvn(use);
295 i -= uses_found; // we deleted 1 or more copies of this edge
296 }
297 }
298
299
not_a_node(const Node * n)300 static inline bool not_a_node(const Node* n) {
301 if (n == NULL) return true;
302 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
303 if (*(address*)n == badAddress) return true; // kill by Node::destruct
304 return false;
305 }
306
307 // Identify all nodes that are reachable from below, useful.
308 // Use breadth-first pass that records state in a Unique_Node_List,
309 // recursive traversal is slower.
identify_useful_nodes(Unique_Node_List & useful)310 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
311 int estimated_worklist_size = live_nodes();
312 useful.map( estimated_worklist_size, NULL ); // preallocate space
313
314 // Initialize worklist
315 if (root() != NULL) { useful.push(root()); }
316 // If 'top' is cached, declare it useful to preserve cached node
317 if( cached_top_node() ) { useful.push(cached_top_node()); }
318
319 // Push all useful nodes onto the list, breadthfirst
320 for( uint next = 0; next < useful.size(); ++next ) {
321 assert( next < unique(), "Unique useful nodes < total nodes");
322 Node *n = useful.at(next);
323 uint max = n->len();
324 for( uint i = 0; i < max; ++i ) {
325 Node *m = n->in(i);
326 if (not_a_node(m)) continue;
327 useful.push(m);
328 }
329 }
330 }
331
332 // Update dead_node_list with any missing dead nodes using useful
333 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
update_dead_node_list(Unique_Node_List & useful)334 void Compile::update_dead_node_list(Unique_Node_List &useful) {
335 uint max_idx = unique();
336 VectorSet& useful_node_set = useful.member_set();
337
338 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
339 // If node with index node_idx is not in useful set,
340 // mark it as dead in dead node list.
341 if (!useful_node_set.test(node_idx)) {
342 record_dead_node(node_idx);
343 }
344 }
345 }
346
remove_useless_late_inlines(GrowableArray<CallGenerator * > * inlines,Unique_Node_List & useful)347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
348 int shift = 0;
349 for (int i = 0; i < inlines->length(); i++) {
350 CallGenerator* cg = inlines->at(i);
351 CallNode* call = cg->call_node();
352 if (shift > 0) {
353 inlines->at_put(i-shift, cg);
354 }
355 if (!useful.member(call)) {
356 shift++;
357 }
358 }
359 inlines->trunc_to(inlines->length()-shift);
360 }
361
362 // Disconnect all useless nodes by disconnecting those at the boundary.
remove_useless_nodes(Unique_Node_List & useful)363 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
364 uint next = 0;
365 while (next < useful.size()) {
366 Node *n = useful.at(next++);
367 if (n->is_SafePoint()) {
368 // We're done with a parsing phase. Replaced nodes are not valid
369 // beyond that point.
370 n->as_SafePoint()->delete_replaced_nodes();
371 }
372 // Use raw traversal of out edges since this code removes out edges
373 int max = n->outcnt();
374 for (int j = 0; j < max; ++j) {
375 Node* child = n->raw_out(j);
376 if (! useful.member(child)) {
377 assert(!child->is_top() || child != top(),
378 "If top is cached in Compile object it is in useful list");
379 // Only need to remove this out-edge to the useless node
380 n->raw_del_out(j);
381 --j;
382 --max;
383 }
384 }
385 if (n->outcnt() == 1 && n->has_special_unique_user()) {
386 record_for_igvn(n->unique_out());
387 }
388 }
389 // Remove useless macro and predicate opaq nodes
390 for (int i = C->macro_count()-1; i >= 0; i--) {
391 Node* n = C->macro_node(i);
392 if (!useful.member(n)) {
393 remove_macro_node(n);
394 }
395 }
396 // Remove useless CastII nodes with range check dependency
397 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
398 Node* cast = range_check_cast_node(i);
399 if (!useful.member(cast)) {
400 remove_range_check_cast(cast);
401 }
402 }
403 // Remove useless expensive nodes
404 for (int i = C->expensive_count()-1; i >= 0; i--) {
405 Node* n = C->expensive_node(i);
406 if (!useful.member(n)) {
407 remove_expensive_node(n);
408 }
409 }
410 // Remove useless Opaque4 nodes
411 for (int i = opaque4_count() - 1; i >= 0; i--) {
412 Node* opaq = opaque4_node(i);
413 if (!useful.member(opaq)) {
414 remove_opaque4_node(opaq);
415 }
416 }
417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
418 bs->eliminate_useless_gc_barriers(useful, this);
419 // clean up the late inline lists
420 remove_useless_late_inlines(&_string_late_inlines, useful);
421 remove_useless_late_inlines(&_boxing_late_inlines, useful);
422 remove_useless_late_inlines(&_late_inlines, useful);
423 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
424 }
425
426 //------------------------------frame_size_in_words-----------------------------
427 // frame_slots in units of words
frame_size_in_words() const428 int Compile::frame_size_in_words() const {
429 // shift is 0 in LP32 and 1 in LP64
430 const int shift = (LogBytesPerWord - LogBytesPerInt);
431 int words = _frame_slots >> shift;
432 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
433 return words;
434 }
435
436 // To bang the stack of this compiled method we use the stack size
437 // that the interpreter would need in case of a deoptimization. This
438 // removes the need to bang the stack in the deoptimization blob which
439 // in turn simplifies stack overflow handling.
bang_size_in_bytes() const440 int Compile::bang_size_in_bytes() const {
441 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
442 }
443
444 // ============================================================================
445 //------------------------------CompileWrapper---------------------------------
446 class CompileWrapper : public StackObj {
447 Compile *const _compile;
448 public:
449 CompileWrapper(Compile* compile);
450
451 ~CompileWrapper();
452 };
453
CompileWrapper(Compile * compile)454 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
455 // the Compile* pointer is stored in the current ciEnv:
456 ciEnv* env = compile->env();
457 assert(env == ciEnv::current(), "must already be a ciEnv active");
458 assert(env->compiler_data() == NULL, "compile already active?");
459 env->set_compiler_data(compile);
460 assert(compile == Compile::current(), "sanity");
461
462 compile->set_type_dict(NULL);
463 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
464 compile->clone_map().set_clone_idx(0);
465 compile->set_type_last_size(0);
466 compile->set_last_tf(NULL, NULL);
467 compile->set_indexSet_arena(NULL);
468 compile->set_indexSet_free_block_list(NULL);
469 compile->init_type_arena();
470 Type::Initialize(compile);
471 _compile->set_scratch_buffer_blob(NULL);
472 _compile->begin_method();
473 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
474 }
~CompileWrapper()475 CompileWrapper::~CompileWrapper() {
476 _compile->end_method();
477 if (_compile->scratch_buffer_blob() != NULL)
478 BufferBlob::free(_compile->scratch_buffer_blob());
479 _compile->env()->set_compiler_data(NULL);
480 }
481
482
483 //----------------------------print_compile_messages---------------------------
print_compile_messages()484 void Compile::print_compile_messages() {
485 #ifndef PRODUCT
486 // Check if recompiling
487 if (_subsume_loads == false && PrintOpto) {
488 // Recompiling without allowing machine instructions to subsume loads
489 tty->print_cr("*********************************************************");
490 tty->print_cr("** Bailout: Recompile without subsuming loads **");
491 tty->print_cr("*********************************************************");
492 }
493 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
494 // Recompiling without escape analysis
495 tty->print_cr("*********************************************************");
496 tty->print_cr("** Bailout: Recompile without escape analysis **");
497 tty->print_cr("*********************************************************");
498 }
499 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
500 // Recompiling without boxing elimination
501 tty->print_cr("*********************************************************");
502 tty->print_cr("** Bailout: Recompile without boxing elimination **");
503 tty->print_cr("*********************************************************");
504 }
505 if (C->directive()->BreakAtCompileOption) {
506 // Open the debugger when compiling this method.
507 tty->print("### Breaking when compiling: ");
508 method()->print_short_name();
509 tty->cr();
510 BREAKPOINT;
511 }
512
513 if( PrintOpto ) {
514 if (is_osr_compilation()) {
515 tty->print("[OSR]%3d", _compile_id);
516 } else {
517 tty->print("%3d", _compile_id);
518 }
519 }
520 #endif
521 }
522
523
524 //-----------------------init_scratch_buffer_blob------------------------------
525 // Construct a temporary BufferBlob and cache it for this compile.
init_scratch_buffer_blob(int const_size)526 void Compile::init_scratch_buffer_blob(int const_size) {
527 // If there is already a scratch buffer blob allocated and the
528 // constant section is big enough, use it. Otherwise free the
529 // current and allocate a new one.
530 BufferBlob* blob = scratch_buffer_blob();
531 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
532 // Use the current blob.
533 } else {
534 if (blob != NULL) {
535 BufferBlob::free(blob);
536 }
537
538 ResourceMark rm;
539 _scratch_const_size = const_size;
540 int size = C2Compiler::initial_code_buffer_size(const_size);
541 blob = BufferBlob::create("Compile::scratch_buffer", size);
542 // Record the buffer blob for next time.
543 set_scratch_buffer_blob(blob);
544 // Have we run out of code space?
545 if (scratch_buffer_blob() == NULL) {
546 // Let CompilerBroker disable further compilations.
547 record_failure("Not enough space for scratch buffer in CodeCache");
548 return;
549 }
550 }
551
552 // Initialize the relocation buffers
553 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
554 set_scratch_locs_memory(locs_buf);
555 }
556
557
558 //-----------------------scratch_emit_size-------------------------------------
559 // Helper function that computes size by emitting code
scratch_emit_size(const Node * n)560 uint Compile::scratch_emit_size(const Node* n) {
561 // Start scratch_emit_size section.
562 set_in_scratch_emit_size(true);
563
564 // Emit into a trash buffer and count bytes emitted.
565 // This is a pretty expensive way to compute a size,
566 // but it works well enough if seldom used.
567 // All common fixed-size instructions are given a size
568 // method by the AD file.
569 // Note that the scratch buffer blob and locs memory are
570 // allocated at the beginning of the compile task, and
571 // may be shared by several calls to scratch_emit_size.
572 // The allocation of the scratch buffer blob is particularly
573 // expensive, since it has to grab the code cache lock.
574 BufferBlob* blob = this->scratch_buffer_blob();
575 assert(blob != NULL, "Initialize BufferBlob at start");
576 assert(blob->size() > MAX_inst_size, "sanity");
577 relocInfo* locs_buf = scratch_locs_memory();
578 address blob_begin = blob->content_begin();
579 address blob_end = (address)locs_buf;
580 assert(blob->contains(blob_end), "sanity");
581 CodeBuffer buf(blob_begin, blob_end - blob_begin);
582 buf.initialize_consts_size(_scratch_const_size);
583 buf.initialize_stubs_size(MAX_stubs_size);
584 assert(locs_buf != NULL, "sanity");
585 int lsize = MAX_locs_size / 3;
586 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
587 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
588 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
589 // Mark as scratch buffer.
590 buf.consts()->set_scratch_emit();
591 buf.insts()->set_scratch_emit();
592 buf.stubs()->set_scratch_emit();
593
594 // Do the emission.
595
596 Label fakeL; // Fake label for branch instructions.
597 Label* saveL = NULL;
598 uint save_bnum = 0;
599 bool is_branch = n->is_MachBranch();
600 if (is_branch) {
601 MacroAssembler masm(&buf);
602 masm.bind(fakeL);
603 n->as_MachBranch()->save_label(&saveL, &save_bnum);
604 n->as_MachBranch()->label_set(&fakeL, 0);
605 }
606 n->emit(buf, this->regalloc());
607
608 // Emitting into the scratch buffer should not fail
609 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
610
611 if (is_branch) // Restore label.
612 n->as_MachBranch()->label_set(saveL, save_bnum);
613
614 // End scratch_emit_size section.
615 set_in_scratch_emit_size(false);
616
617 return buf.insts_size();
618 }
619
620
621 // ============================================================================
622 //------------------------------Compile standard-------------------------------
debug_only(int Compile::_debug_idx=100000;)623 debug_only( int Compile::_debug_idx = 100000; )
624
625 // Compile a method. entry_bci is -1 for normal compilations and indicates
626 // the continuation bci for on stack replacement.
627
628
629 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
630 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
631 : Phase(Compiler),
632 _compile_id(ci_env->compile_id()),
633 _save_argument_registers(false),
634 _subsume_loads(subsume_loads),
635 _do_escape_analysis(do_escape_analysis),
636 _eliminate_boxing(eliminate_boxing),
637 _method(target),
638 _entry_bci(osr_bci),
639 _stub_function(NULL),
640 _stub_name(NULL),
641 _stub_entry_point(NULL),
642 _max_node_limit(MaxNodeLimit),
643 _orig_pc_slot(0),
644 _orig_pc_slot_offset_in_bytes(0),
645 _inlining_progress(false),
646 _inlining_incrementally(false),
647 _do_cleanup(false),
648 _has_reserved_stack_access(target->has_reserved_stack_access()),
649 #ifndef PRODUCT
650 _trace_opto_output(directive->TraceOptoOutputOption),
651 _print_ideal(directive->PrintIdealOption),
652 #endif
653 _has_method_handle_invokes(false),
654 _clinit_barrier_on_entry(false),
655 _comp_arena(mtCompiler),
656 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
657 _env(ci_env),
658 _directive(directive),
659 _log(ci_env->log()),
660 _failure_reason(NULL),
661 _congraph(NULL),
662 #ifndef PRODUCT
663 _printer(IdealGraphPrinter::printer()),
664 #endif
665 _dead_node_list(comp_arena()),
666 _dead_node_count(0),
667 _node_arena(mtCompiler),
668 _old_arena(mtCompiler),
669 _mach_constant_base_node(NULL),
670 _Compile_types(mtCompiler),
671 _initial_gvn(NULL),
672 _for_igvn(NULL),
673 _warm_calls(NULL),
674 _late_inlines(comp_arena(), 2, 0, NULL),
675 _string_late_inlines(comp_arena(), 2, 0, NULL),
676 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
677 _late_inlines_pos(0),
678 _number_of_mh_late_inlines(0),
679 _print_inlining_stream(NULL),
680 _print_inlining_list(NULL),
681 _print_inlining_idx(0),
682 _print_inlining_output(NULL),
683 _replay_inline_data(NULL),
684 _java_calls(0),
685 _inner_loops(0),
686 _interpreter_frame_size(0),
687 _node_bundling_limit(0),
688 _node_bundling_base(NULL),
689 _code_buffer("Compile::Fill_buffer"),
690 _scratch_const_size(-1),
691 _in_scratch_emit_size(false)
692 #ifndef PRODUCT
693 , _in_dump_cnt(0)
694 #endif
695 {
696 C = this;
697 #ifndef PRODUCT
698 if (_printer != NULL) {
699 _printer->set_compile(this);
700 }
701 #endif
702 CompileWrapper cw(this);
703
704 if (CITimeVerbose) {
705 tty->print(" ");
706 target->holder()->name()->print();
707 tty->print(".");
708 target->print_short_name();
709 tty->print(" ");
710 }
711 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
712 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
713
714 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
715 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
716 // We can always print a disassembly, either abstract (hex dump) or
717 // with the help of a suitable hsdis library. Thus, we should not
718 // couple print_assembly and print_opto_assembly controls.
719 // But: always print opto and regular assembly on compile command 'print'.
720 bool print_assembly = directive->PrintAssemblyOption;
721 set_print_assembly(print_opto_assembly || print_assembly);
722 #else
723 set_print_assembly(false); // must initialize.
724 #endif
725
726 #ifndef PRODUCT
727 set_parsed_irreducible_loop(false);
728
729 if (directive->ReplayInlineOption) {
730 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
731 }
732 #endif
733 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
734 set_print_intrinsics(directive->PrintIntrinsicsOption);
735 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
736
737 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
738 // Make sure the method being compiled gets its own MDO,
739 // so we can at least track the decompile_count().
740 // Need MDO to record RTM code generation state.
741 method()->ensure_method_data();
742 }
743
744 Init(::AliasLevel);
745
746
747 print_compile_messages();
748
749 _ilt = InlineTree::build_inline_tree_root();
750
751 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
752 assert(num_alias_types() >= AliasIdxRaw, "");
753
754 #define MINIMUM_NODE_HASH 1023
755 // Node list that Iterative GVN will start with
756 Unique_Node_List for_igvn(comp_arena());
757 set_for_igvn(&for_igvn);
758
759 // GVN that will be run immediately on new nodes
760 uint estimated_size = method()->code_size()*4+64;
761 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
762 PhaseGVN gvn(node_arena(), estimated_size);
763 set_initial_gvn(&gvn);
764
765 print_inlining_init();
766 { // Scope for timing the parser
767 TracePhase tp("parse", &timers[_t_parser]);
768
769 // Put top into the hash table ASAP.
770 initial_gvn()->transform_no_reclaim(top());
771
772 // Set up tf(), start(), and find a CallGenerator.
773 CallGenerator* cg = NULL;
774 if (is_osr_compilation()) {
775 const TypeTuple *domain = StartOSRNode::osr_domain();
776 const TypeTuple *range = TypeTuple::make_range(method()->signature());
777 init_tf(TypeFunc::make(domain, range));
778 StartNode* s = new StartOSRNode(root(), domain);
779 initial_gvn()->set_type_bottom(s);
780 init_start(s);
781 cg = CallGenerator::for_osr(method(), entry_bci());
782 } else {
783 // Normal case.
784 init_tf(TypeFunc::make(method()));
785 StartNode* s = new StartNode(root(), tf()->domain());
786 initial_gvn()->set_type_bottom(s);
787 init_start(s);
788 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
789 // With java.lang.ref.reference.get() we must go through the
790 // intrinsic - even when get() is the root
791 // method of the compile - so that, if necessary, the value in
792 // the referent field of the reference object gets recorded by
793 // the pre-barrier code.
794 cg = find_intrinsic(method(), false);
795 }
796 if (cg == NULL) {
797 float past_uses = method()->interpreter_invocation_count();
798 float expected_uses = past_uses;
799 cg = CallGenerator::for_inline(method(), expected_uses);
800 }
801 }
802 if (failing()) return;
803 if (cg == NULL) {
804 record_method_not_compilable("cannot parse method");
805 return;
806 }
807 JVMState* jvms = build_start_state(start(), tf());
808 if ((jvms = cg->generate(jvms)) == NULL) {
809 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
810 record_method_not_compilable("method parse failed");
811 }
812 return;
813 }
814 GraphKit kit(jvms);
815
816 if (!kit.stopped()) {
817 // Accept return values, and transfer control we know not where.
818 // This is done by a special, unique ReturnNode bound to root.
819 return_values(kit.jvms());
820 }
821
822 if (kit.has_exceptions()) {
823 // Any exceptions that escape from this call must be rethrown
824 // to whatever caller is dynamically above us on the stack.
825 // This is done by a special, unique RethrowNode bound to root.
826 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
827 }
828
829 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
830
831 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
832 inline_string_calls(true);
833 }
834
835 if (failing()) return;
836
837 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
838
839 // Remove clutter produced by parsing.
840 if (!failing()) {
841 ResourceMark rm;
842 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
843 }
844 }
845
846 // Note: Large methods are capped off in do_one_bytecode().
847 if (failing()) return;
848
849 // After parsing, node notes are no longer automagic.
850 // They must be propagated by register_new_node_with_optimizer(),
851 // clone(), or the like.
852 set_default_node_notes(NULL);
853
854 for (;;) {
855 int successes = Inline_Warm();
856 if (failing()) return;
857 if (successes == 0) break;
858 }
859
860 // Drain the list.
861 Finish_Warm();
862 #ifndef PRODUCT
863 if (_printer && _printer->should_print(1)) {
864 _printer->print_inlining();
865 }
866 #endif
867
868 if (failing()) return;
869 NOT_PRODUCT( verify_graph_edges(); )
870
871 // Now optimize
872 Optimize();
873 if (failing()) return;
874 NOT_PRODUCT( verify_graph_edges(); )
875
876 #ifndef PRODUCT
877 if (print_ideal()) {
878 ttyLocker ttyl; // keep the following output all in one block
879 // This output goes directly to the tty, not the compiler log.
880 // To enable tools to match it up with the compilation activity,
881 // be sure to tag this tty output with the compile ID.
882 if (xtty != NULL) {
883 xtty->head("ideal compile_id='%d'%s", compile_id(),
884 is_osr_compilation() ? " compile_kind='osr'" :
885 "");
886 }
887 root()->dump(9999);
888 if (xtty != NULL) {
889 xtty->tail("ideal");
890 }
891 }
892 #endif
893
894 #ifdef ASSERT
895 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
896 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
897 #endif
898
899 // Dump compilation data to replay it.
900 if (directive->DumpReplayOption) {
901 env()->dump_replay_data(_compile_id);
902 }
903 if (directive->DumpInlineOption && (ilt() != NULL)) {
904 env()->dump_inline_data(_compile_id);
905 }
906
907 // Now that we know the size of all the monitors we can add a fixed slot
908 // for the original deopt pc.
909
910 _orig_pc_slot = fixed_slots();
911 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
912 set_fixed_slots(next_slot);
913
914 // Compute when to use implicit null checks. Used by matching trap based
915 // nodes and NullCheck optimization.
916 set_allowed_deopt_reasons();
917
918 // Now generate code
919 Code_Gen();
920 if (failing()) return;
921
922 // Check if we want to skip execution of all compiled code.
923 {
924 #ifndef PRODUCT
925 if (OptoNoExecute) {
926 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
927 return;
928 }
929 #endif
930 TracePhase tp("install_code", &timers[_t_registerMethod]);
931
932 if (is_osr_compilation()) {
933 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
934 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
935 } else {
936 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
937 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
938 }
939
940 env()->register_method(_method, _entry_bci,
941 &_code_offsets,
942 _orig_pc_slot_offset_in_bytes,
943 code_buffer(),
944 frame_size_in_words(), _oop_map_set,
945 &_handler_table, &_inc_table,
946 compiler,
947 has_unsafe_access(),
948 SharedRuntime::is_wide_vector(max_vector_size()),
949 rtm_state()
950 );
951
952 if (log() != NULL) // Print code cache state into compiler log
953 log()->code_cache_state();
954 }
955 }
956
957 //------------------------------Compile----------------------------------------
958 // Compile a runtime stub
Compile(ciEnv * ci_env,TypeFunc_generator generator,address stub_function,const char * stub_name,int is_fancy_jump,bool pass_tls,bool save_arg_registers,bool return_pc,DirectiveSet * directive)959 Compile::Compile( ciEnv* ci_env,
960 TypeFunc_generator generator,
961 address stub_function,
962 const char *stub_name,
963 int is_fancy_jump,
964 bool pass_tls,
965 bool save_arg_registers,
966 bool return_pc,
967 DirectiveSet* directive)
968 : Phase(Compiler),
969 _compile_id(0),
970 _save_argument_registers(save_arg_registers),
971 _subsume_loads(true),
972 _do_escape_analysis(false),
973 _eliminate_boxing(false),
974 _method(NULL),
975 _entry_bci(InvocationEntryBci),
976 _stub_function(stub_function),
977 _stub_name(stub_name),
978 _stub_entry_point(NULL),
979 _max_node_limit(MaxNodeLimit),
980 _orig_pc_slot(0),
981 _orig_pc_slot_offset_in_bytes(0),
982 _inlining_progress(false),
983 _inlining_incrementally(false),
984 _has_reserved_stack_access(false),
985 #ifndef PRODUCT
986 _trace_opto_output(directive->TraceOptoOutputOption),
987 _print_ideal(directive->PrintIdealOption),
988 #endif
989 _has_method_handle_invokes(false),
990 _clinit_barrier_on_entry(false),
991 _comp_arena(mtCompiler),
992 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
993 _env(ci_env),
994 _directive(directive),
995 _log(ci_env->log()),
996 _failure_reason(NULL),
997 _congraph(NULL),
998 #ifndef PRODUCT
999 _printer(NULL),
1000 #endif
1001 _dead_node_list(comp_arena()),
1002 _dead_node_count(0),
1003 _node_arena(mtCompiler),
1004 _old_arena(mtCompiler),
1005 _mach_constant_base_node(NULL),
1006 _Compile_types(mtCompiler),
1007 _initial_gvn(NULL),
1008 _for_igvn(NULL),
1009 _warm_calls(NULL),
1010 _number_of_mh_late_inlines(0),
1011 _print_inlining_stream(NULL),
1012 _print_inlining_list(NULL),
1013 _print_inlining_idx(0),
1014 _print_inlining_output(NULL),
1015 _replay_inline_data(NULL),
1016 _java_calls(0),
1017 _inner_loops(0),
1018 _interpreter_frame_size(0),
1019 _node_bundling_limit(0),
1020 _node_bundling_base(NULL),
1021 _code_buffer("Compile::Fill_buffer"),
1022 #ifndef PRODUCT
1023 _in_dump_cnt(0),
1024 #endif
1025 _allowed_reasons(0) {
1026 C = this;
1027
1028 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1029 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1030
1031 #ifndef PRODUCT
1032 set_print_assembly(PrintFrameConverterAssembly);
1033 set_parsed_irreducible_loop(false);
1034 #else
1035 set_print_assembly(false); // Must initialize.
1036 #endif
1037 set_has_irreducible_loop(false); // no loops
1038
1039 CompileWrapper cw(this);
1040 Init(/*AliasLevel=*/ 0);
1041 init_tf((*generator)());
1042
1043 {
1044 // The following is a dummy for the sake of GraphKit::gen_stub
1045 Unique_Node_List for_igvn(comp_arena());
1046 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1047 PhaseGVN gvn(Thread::current()->resource_area(),255);
1048 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1049 gvn.transform_no_reclaim(top());
1050
1051 GraphKit kit;
1052 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1053 }
1054
1055 NOT_PRODUCT( verify_graph_edges(); )
1056 Code_Gen();
1057 if (failing()) return;
1058
1059
1060 // Entry point will be accessed using compile->stub_entry_point();
1061 if (code_buffer() == NULL) {
1062 Matcher::soft_match_failure();
1063 } else {
1064 if (PrintAssembly && (WizardMode || Verbose))
1065 tty->print_cr("### Stub::%s", stub_name);
1066
1067 if (!failing()) {
1068 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1069
1070 // Make the NMethod
1071 // For now we mark the frame as never safe for profile stackwalking
1072 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1073 code_buffer(),
1074 CodeOffsets::frame_never_safe,
1075 // _code_offsets.value(CodeOffsets::Frame_Complete),
1076 frame_size_in_words(),
1077 _oop_map_set,
1078 save_arg_registers);
1079 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1080
1081 _stub_entry_point = rs->entry_point();
1082 }
1083 }
1084 }
1085
1086 //------------------------------Init-------------------------------------------
1087 // Prepare for a single compilation
Init(int aliaslevel)1088 void Compile::Init(int aliaslevel) {
1089 _unique = 0;
1090 _regalloc = NULL;
1091
1092 _tf = NULL; // filled in later
1093 _top = NULL; // cached later
1094 _matcher = NULL; // filled in later
1095 _cfg = NULL; // filled in later
1096
1097 set_24_bit_selection_and_mode(Use24BitFP, false);
1098
1099 _node_note_array = NULL;
1100 _default_node_notes = NULL;
1101 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1102
1103 _immutable_memory = NULL; // filled in at first inquiry
1104
1105 // Globally visible Nodes
1106 // First set TOP to NULL to give safe behavior during creation of RootNode
1107 set_cached_top_node(NULL);
1108 set_root(new RootNode());
1109 // Now that you have a Root to point to, create the real TOP
1110 set_cached_top_node( new ConNode(Type::TOP) );
1111 set_recent_alloc(NULL, NULL);
1112
1113 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1114 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1115 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1116 env()->set_dependencies(new Dependencies(env()));
1117
1118 _fixed_slots = 0;
1119 set_has_split_ifs(false);
1120 set_has_loops(has_method() && method()->has_loops()); // first approximation
1121 set_has_stringbuilder(false);
1122 set_has_boxed_value(false);
1123 _trap_can_recompile = false; // no traps emitted yet
1124 _major_progress = true; // start out assuming good things will happen
1125 set_has_unsafe_access(false);
1126 set_max_vector_size(0);
1127 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1128 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1129 set_decompile_count(0);
1130
1131 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1132 _loop_opts_cnt = LoopOptsCount;
1133 set_do_inlining(Inline);
1134 set_max_inline_size(MaxInlineSize);
1135 set_freq_inline_size(FreqInlineSize);
1136 set_do_scheduling(OptoScheduling);
1137 set_do_count_invocations(false);
1138 set_do_method_data_update(false);
1139
1140 set_do_vector_loop(false);
1141
1142 if (AllowVectorizeOnDemand) {
1143 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1144 set_do_vector_loop(true);
1145 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());})
1146 } else if (has_method() && method()->name() != 0 &&
1147 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1148 set_do_vector_loop(true);
1149 }
1150 }
1151 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1152 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());})
1153
1154 set_age_code(has_method() && method()->profile_aging());
1155 set_rtm_state(NoRTM); // No RTM lock eliding by default
1156 _max_node_limit = _directive->MaxNodeLimitOption;
1157
1158 #if INCLUDE_RTM_OPT
1159 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1160 int rtm_state = method()->method_data()->rtm_state();
1161 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1162 // Don't generate RTM lock eliding code.
1163 set_rtm_state(NoRTM);
1164 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1165 // Generate RTM lock eliding code without abort ratio calculation code.
1166 set_rtm_state(UseRTM);
1167 } else if (UseRTMDeopt) {
1168 // Generate RTM lock eliding code and include abort ratio calculation
1169 // code if UseRTMDeopt is on.
1170 set_rtm_state(ProfileRTM);
1171 }
1172 }
1173 #endif
1174 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1175 set_clinit_barrier_on_entry(true);
1176 }
1177 if (debug_info()->recording_non_safepoints()) {
1178 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1179 (comp_arena(), 8, 0, NULL));
1180 set_default_node_notes(Node_Notes::make(this));
1181 }
1182
1183 // // -- Initialize types before each compile --
1184 // // Update cached type information
1185 // if( _method && _method->constants() )
1186 // Type::update_loaded_types(_method, _method->constants());
1187
1188 // Init alias_type map.
1189 if (!_do_escape_analysis && aliaslevel == 3)
1190 aliaslevel = 2; // No unique types without escape analysis
1191 _AliasLevel = aliaslevel;
1192 const int grow_ats = 16;
1193 _max_alias_types = grow_ats;
1194 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1195 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1196 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1197 {
1198 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1199 }
1200 // Initialize the first few types.
1201 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1202 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1203 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1204 _num_alias_types = AliasIdxRaw+1;
1205 // Zero out the alias type cache.
1206 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1207 // A NULL adr_type hits in the cache right away. Preload the right answer.
1208 probe_alias_cache(NULL)->_index = AliasIdxTop;
1209
1210 _intrinsics = NULL;
1211 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1212 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1213 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1214 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1215 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1216 register_library_intrinsics();
1217 }
1218
1219 //---------------------------init_start----------------------------------------
1220 // Install the StartNode on this compile object.
init_start(StartNode * s)1221 void Compile::init_start(StartNode* s) {
1222 if (failing())
1223 return; // already failing
1224 assert(s == start(), "");
1225 }
1226
1227 /**
1228 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1229 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1230 * the ideal graph.
1231 */
start() const1232 StartNode* Compile::start() const {
1233 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1234 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1235 Node* start = root()->fast_out(i);
1236 if (start->is_Start()) {
1237 return start->as_Start();
1238 }
1239 }
1240 fatal("Did not find Start node!");
1241 return NULL;
1242 }
1243
1244 //-------------------------------immutable_memory-------------------------------------
1245 // Access immutable memory
immutable_memory()1246 Node* Compile::immutable_memory() {
1247 if (_immutable_memory != NULL) {
1248 return _immutable_memory;
1249 }
1250 StartNode* s = start();
1251 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1252 Node *p = s->fast_out(i);
1253 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1254 _immutable_memory = p;
1255 return _immutable_memory;
1256 }
1257 }
1258 ShouldNotReachHere();
1259 return NULL;
1260 }
1261
1262 //----------------------set_cached_top_node------------------------------------
1263 // Install the cached top node, and make sure Node::is_top works correctly.
set_cached_top_node(Node * tn)1264 void Compile::set_cached_top_node(Node* tn) {
1265 if (tn != NULL) verify_top(tn);
1266 Node* old_top = _top;
1267 _top = tn;
1268 // Calling Node::setup_is_top allows the nodes the chance to adjust
1269 // their _out arrays.
1270 if (_top != NULL) _top->setup_is_top();
1271 if (old_top != NULL) old_top->setup_is_top();
1272 assert(_top == NULL || top()->is_top(), "");
1273 }
1274
1275 #ifdef ASSERT
count_live_nodes_by_graph_walk()1276 uint Compile::count_live_nodes_by_graph_walk() {
1277 Unique_Node_List useful(comp_arena());
1278 // Get useful node list by walking the graph.
1279 identify_useful_nodes(useful);
1280 return useful.size();
1281 }
1282
print_missing_nodes()1283 void Compile::print_missing_nodes() {
1284
1285 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1286 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1287 return;
1288 }
1289
1290 // This is an expensive function. It is executed only when the user
1291 // specifies VerifyIdealNodeCount option or otherwise knows the
1292 // additional work that needs to be done to identify reachable nodes
1293 // by walking the flow graph and find the missing ones using
1294 // _dead_node_list.
1295
1296 Unique_Node_List useful(comp_arena());
1297 // Get useful node list by walking the graph.
1298 identify_useful_nodes(useful);
1299
1300 uint l_nodes = C->live_nodes();
1301 uint l_nodes_by_walk = useful.size();
1302
1303 if (l_nodes != l_nodes_by_walk) {
1304 if (_log != NULL) {
1305 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1306 _log->stamp();
1307 _log->end_head();
1308 }
1309 VectorSet& useful_member_set = useful.member_set();
1310 int last_idx = l_nodes_by_walk;
1311 for (int i = 0; i < last_idx; i++) {
1312 if (useful_member_set.test(i)) {
1313 if (_dead_node_list.test(i)) {
1314 if (_log != NULL) {
1315 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1316 }
1317 if (PrintIdealNodeCount) {
1318 // Print the log message to tty
1319 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1320 useful.at(i)->dump();
1321 }
1322 }
1323 }
1324 else if (! _dead_node_list.test(i)) {
1325 if (_log != NULL) {
1326 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1327 }
1328 if (PrintIdealNodeCount) {
1329 // Print the log message to tty
1330 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1331 }
1332 }
1333 }
1334 if (_log != NULL) {
1335 _log->tail("mismatched_nodes");
1336 }
1337 }
1338 }
record_modified_node(Node * n)1339 void Compile::record_modified_node(Node* n) {
1340 if (_modified_nodes != NULL && !_inlining_incrementally &&
1341 n->outcnt() != 0 && !n->is_Con()) {
1342 _modified_nodes->push(n);
1343 }
1344 }
1345
remove_modified_node(Node * n)1346 void Compile::remove_modified_node(Node* n) {
1347 if (_modified_nodes != NULL) {
1348 _modified_nodes->remove(n);
1349 }
1350 }
1351 #endif
1352
1353 #ifndef PRODUCT
verify_top(Node * tn) const1354 void Compile::verify_top(Node* tn) const {
1355 if (tn != NULL) {
1356 assert(tn->is_Con(), "top node must be a constant");
1357 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1358 assert(tn->in(0) != NULL, "must have live top node");
1359 }
1360 }
1361 #endif
1362
1363
1364 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1365
grow_node_notes(GrowableArray<Node_Notes * > * arr,int grow_by)1366 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1367 guarantee(arr != NULL, "");
1368 int num_blocks = arr->length();
1369 if (grow_by < num_blocks) grow_by = num_blocks;
1370 int num_notes = grow_by * _node_notes_block_size;
1371 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1372 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1373 while (num_notes > 0) {
1374 arr->append(notes);
1375 notes += _node_notes_block_size;
1376 num_notes -= _node_notes_block_size;
1377 }
1378 assert(num_notes == 0, "exact multiple, please");
1379 }
1380
copy_node_notes_to(Node * dest,Node * source)1381 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1382 if (source == NULL || dest == NULL) return false;
1383
1384 if (dest->is_Con())
1385 return false; // Do not push debug info onto constants.
1386
1387 #ifdef ASSERT
1388 // Leave a bread crumb trail pointing to the original node:
1389 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1390 dest->set_debug_orig(source);
1391 }
1392 #endif
1393
1394 if (node_note_array() == NULL)
1395 return false; // Not collecting any notes now.
1396
1397 // This is a copy onto a pre-existing node, which may already have notes.
1398 // If both nodes have notes, do not overwrite any pre-existing notes.
1399 Node_Notes* source_notes = node_notes_at(source->_idx);
1400 if (source_notes == NULL || source_notes->is_clear()) return false;
1401 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1402 if (dest_notes == NULL || dest_notes->is_clear()) {
1403 return set_node_notes_at(dest->_idx, source_notes);
1404 }
1405
1406 Node_Notes merged_notes = (*source_notes);
1407 // The order of operations here ensures that dest notes will win...
1408 merged_notes.update_from(dest_notes);
1409 return set_node_notes_at(dest->_idx, &merged_notes);
1410 }
1411
1412
1413 //--------------------------allow_range_check_smearing-------------------------
1414 // Gating condition for coalescing similar range checks.
1415 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1416 // single covering check that is at least as strong as any of them.
1417 // If the optimization succeeds, the simplified (strengthened) range check
1418 // will always succeed. If it fails, we will deopt, and then give up
1419 // on the optimization.
allow_range_check_smearing() const1420 bool Compile::allow_range_check_smearing() const {
1421 // If this method has already thrown a range-check,
1422 // assume it was because we already tried range smearing
1423 // and it failed.
1424 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1425 return !already_trapped;
1426 }
1427
1428
1429 //------------------------------flatten_alias_type-----------------------------
flatten_alias_type(const TypePtr * tj) const1430 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1431 int offset = tj->offset();
1432 TypePtr::PTR ptr = tj->ptr();
1433
1434 // Known instance (scalarizable allocation) alias only with itself.
1435 bool is_known_inst = tj->isa_oopptr() != NULL &&
1436 tj->is_oopptr()->is_known_instance();
1437
1438 // Process weird unsafe references.
1439 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1440 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1441 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1442 tj = TypeOopPtr::BOTTOM;
1443 ptr = tj->ptr();
1444 offset = tj->offset();
1445 }
1446
1447 // Array pointers need some flattening
1448 const TypeAryPtr *ta = tj->isa_aryptr();
1449 if (ta && ta->is_stable()) {
1450 // Erase stability property for alias analysis.
1451 tj = ta = ta->cast_to_stable(false);
1452 }
1453 if( ta && is_known_inst ) {
1454 if ( offset != Type::OffsetBot &&
1455 offset > arrayOopDesc::length_offset_in_bytes() ) {
1456 offset = Type::OffsetBot; // Flatten constant access into array body only
1457 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1458 }
1459 } else if( ta && _AliasLevel >= 2 ) {
1460 // For arrays indexed by constant indices, we flatten the alias
1461 // space to include all of the array body. Only the header, klass
1462 // and array length can be accessed un-aliased.
1463 if( offset != Type::OffsetBot ) {
1464 if( ta->const_oop() ) { // MethodData* or Method*
1465 offset = Type::OffsetBot; // Flatten constant access into array body
1466 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1467 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1468 // range is OK as-is.
1469 tj = ta = TypeAryPtr::RANGE;
1470 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1471 tj = TypeInstPtr::KLASS; // all klass loads look alike
1472 ta = TypeAryPtr::RANGE; // generic ignored junk
1473 ptr = TypePtr::BotPTR;
1474 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1475 tj = TypeInstPtr::MARK;
1476 ta = TypeAryPtr::RANGE; // generic ignored junk
1477 ptr = TypePtr::BotPTR;
1478 } else { // Random constant offset into array body
1479 offset = Type::OffsetBot; // Flatten constant access into array body
1480 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1481 }
1482 }
1483 // Arrays of fixed size alias with arrays of unknown size.
1484 if (ta->size() != TypeInt::POS) {
1485 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1486 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1487 }
1488 // Arrays of known objects become arrays of unknown objects.
1489 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1490 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1491 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1492 }
1493 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1494 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1495 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1496 }
1497 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1498 // cannot be distinguished by bytecode alone.
1499 if (ta->elem() == TypeInt::BOOL) {
1500 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1501 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1502 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1503 }
1504 // During the 2nd round of IterGVN, NotNull castings are removed.
1505 // Make sure the Bottom and NotNull variants alias the same.
1506 // Also, make sure exact and non-exact variants alias the same.
1507 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1508 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1509 }
1510 }
1511
1512 // Oop pointers need some flattening
1513 const TypeInstPtr *to = tj->isa_instptr();
1514 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1515 ciInstanceKlass *k = to->klass()->as_instance_klass();
1516 if( ptr == TypePtr::Constant ) {
1517 if (to->klass() != ciEnv::current()->Class_klass() ||
1518 offset < k->size_helper() * wordSize) {
1519 // No constant oop pointers (such as Strings); they alias with
1520 // unknown strings.
1521 assert(!is_known_inst, "not scalarizable allocation");
1522 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1523 }
1524 } else if( is_known_inst ) {
1525 tj = to; // Keep NotNull and klass_is_exact for instance type
1526 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1527 // During the 2nd round of IterGVN, NotNull castings are removed.
1528 // Make sure the Bottom and NotNull variants alias the same.
1529 // Also, make sure exact and non-exact variants alias the same.
1530 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1531 }
1532 if (to->speculative() != NULL) {
1533 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1534 }
1535 // Canonicalize the holder of this field
1536 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1537 // First handle header references such as a LoadKlassNode, even if the
1538 // object's klass is unloaded at compile time (4965979).
1539 if (!is_known_inst) { // Do it only for non-instance types
1540 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1541 }
1542 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1543 // Static fields are in the space above the normal instance
1544 // fields in the java.lang.Class instance.
1545 if (to->klass() != ciEnv::current()->Class_klass()) {
1546 to = NULL;
1547 tj = TypeOopPtr::BOTTOM;
1548 offset = tj->offset();
1549 }
1550 } else {
1551 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1552 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1553 if( is_known_inst ) {
1554 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1555 } else {
1556 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1557 }
1558 }
1559 }
1560 }
1561
1562 // Klass pointers to object array klasses need some flattening
1563 const TypeKlassPtr *tk = tj->isa_klassptr();
1564 if( tk ) {
1565 // If we are referencing a field within a Klass, we need
1566 // to assume the worst case of an Object. Both exact and
1567 // inexact types must flatten to the same alias class so
1568 // use NotNull as the PTR.
1569 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1570
1571 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1572 TypeKlassPtr::OBJECT->klass(),
1573 offset);
1574 }
1575
1576 ciKlass* klass = tk->klass();
1577 if( klass->is_obj_array_klass() ) {
1578 ciKlass* k = TypeAryPtr::OOPS->klass();
1579 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1580 k = TypeInstPtr::BOTTOM->klass();
1581 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1582 }
1583
1584 // Check for precise loads from the primary supertype array and force them
1585 // to the supertype cache alias index. Check for generic array loads from
1586 // the primary supertype array and also force them to the supertype cache
1587 // alias index. Since the same load can reach both, we need to merge
1588 // these 2 disparate memories into the same alias class. Since the
1589 // primary supertype array is read-only, there's no chance of confusion
1590 // where we bypass an array load and an array store.
1591 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1592 if (offset == Type::OffsetBot ||
1593 (offset >= primary_supers_offset &&
1594 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1595 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1596 offset = in_bytes(Klass::secondary_super_cache_offset());
1597 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1598 }
1599 }
1600
1601 // Flatten all Raw pointers together.
1602 if (tj->base() == Type::RawPtr)
1603 tj = TypeRawPtr::BOTTOM;
1604
1605 if (tj->base() == Type::AnyPtr)
1606 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1607
1608 // Flatten all to bottom for now
1609 switch( _AliasLevel ) {
1610 case 0:
1611 tj = TypePtr::BOTTOM;
1612 break;
1613 case 1: // Flatten to: oop, static, field or array
1614 switch (tj->base()) {
1615 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1616 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1617 case Type::AryPtr: // do not distinguish arrays at all
1618 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1619 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1620 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1621 default: ShouldNotReachHere();
1622 }
1623 break;
1624 case 2: // No collapsing at level 2; keep all splits
1625 case 3: // No collapsing at level 3; keep all splits
1626 break;
1627 default:
1628 Unimplemented();
1629 }
1630
1631 offset = tj->offset();
1632 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1633
1634 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1635 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1636 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1637 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1638 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1639 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1640 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1641 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1642 assert( tj->ptr() != TypePtr::TopPTR &&
1643 tj->ptr() != TypePtr::AnyNull &&
1644 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1645 // assert( tj->ptr() != TypePtr::Constant ||
1646 // tj->base() == Type::RawPtr ||
1647 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1648
1649 return tj;
1650 }
1651
Init(int i,const TypePtr * at)1652 void Compile::AliasType::Init(int i, const TypePtr* at) {
1653 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1654 _index = i;
1655 _adr_type = at;
1656 _field = NULL;
1657 _element = NULL;
1658 _is_rewritable = true; // default
1659 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1660 if (atoop != NULL && atoop->is_known_instance()) {
1661 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1662 _general_index = Compile::current()->get_alias_index(gt);
1663 } else {
1664 _general_index = 0;
1665 }
1666 }
1667
basic_type() const1668 BasicType Compile::AliasType::basic_type() const {
1669 if (element() != NULL) {
1670 const Type* element = adr_type()->is_aryptr()->elem();
1671 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1672 } if (field() != NULL) {
1673 return field()->layout_type();
1674 } else {
1675 return T_ILLEGAL; // unknown
1676 }
1677 }
1678
1679 //---------------------------------print_on------------------------------------
1680 #ifndef PRODUCT
print_on(outputStream * st)1681 void Compile::AliasType::print_on(outputStream* st) {
1682 if (index() < 10)
1683 st->print("@ <%d> ", index());
1684 else st->print("@ <%d>", index());
1685 st->print(is_rewritable() ? " " : " RO");
1686 int offset = adr_type()->offset();
1687 if (offset == Type::OffsetBot)
1688 st->print(" +any");
1689 else st->print(" +%-3d", offset);
1690 st->print(" in ");
1691 adr_type()->dump_on(st);
1692 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1693 if (field() != NULL && tjp) {
1694 if (tjp->klass() != field()->holder() ||
1695 tjp->offset() != field()->offset_in_bytes()) {
1696 st->print(" != ");
1697 field()->print();
1698 st->print(" ***");
1699 }
1700 }
1701 }
1702
print_alias_types()1703 void print_alias_types() {
1704 Compile* C = Compile::current();
1705 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1706 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1707 C->alias_type(idx)->print_on(tty);
1708 tty->cr();
1709 }
1710 }
1711 #endif
1712
1713
1714 //----------------------------probe_alias_cache--------------------------------
probe_alias_cache(const TypePtr * adr_type)1715 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1716 intptr_t key = (intptr_t) adr_type;
1717 key ^= key >> logAliasCacheSize;
1718 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1719 }
1720
1721
1722 //-----------------------------grow_alias_types--------------------------------
grow_alias_types()1723 void Compile::grow_alias_types() {
1724 const int old_ats = _max_alias_types; // how many before?
1725 const int new_ats = old_ats; // how many more?
1726 const int grow_ats = old_ats+new_ats; // how many now?
1727 _max_alias_types = grow_ats;
1728 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1729 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1730 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1731 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1732 }
1733
1734
1735 //--------------------------------find_alias_type------------------------------
find_alias_type(const TypePtr * adr_type,bool no_create,ciField * original_field)1736 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1737 if (_AliasLevel == 0)
1738 return alias_type(AliasIdxBot);
1739
1740 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1741 if (ace->_adr_type == adr_type) {
1742 return alias_type(ace->_index);
1743 }
1744
1745 // Handle special cases.
1746 if (adr_type == NULL) return alias_type(AliasIdxTop);
1747 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1748
1749 // Do it the slow way.
1750 const TypePtr* flat = flatten_alias_type(adr_type);
1751
1752 #ifdef ASSERT
1753 {
1754 ResourceMark rm;
1755 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1756 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1757 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1758 Type::str(adr_type));
1759 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1760 const TypeOopPtr* foop = flat->is_oopptr();
1761 // Scalarizable allocations have exact klass always.
1762 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1763 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1764 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1765 Type::str(foop), Type::str(xoop));
1766 }
1767 }
1768 #endif
1769
1770 int idx = AliasIdxTop;
1771 for (int i = 0; i < num_alias_types(); i++) {
1772 if (alias_type(i)->adr_type() == flat) {
1773 idx = i;
1774 break;
1775 }
1776 }
1777
1778 if (idx == AliasIdxTop) {
1779 if (no_create) return NULL;
1780 // Grow the array if necessary.
1781 if (_num_alias_types == _max_alias_types) grow_alias_types();
1782 // Add a new alias type.
1783 idx = _num_alias_types++;
1784 _alias_types[idx]->Init(idx, flat);
1785 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1786 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1787 if (flat->isa_instptr()) {
1788 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1789 && flat->is_instptr()->klass() == env()->Class_klass())
1790 alias_type(idx)->set_rewritable(false);
1791 }
1792 if (flat->isa_aryptr()) {
1793 #ifdef ASSERT
1794 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1795 // (T_BYTE has the weakest alignment and size restrictions...)
1796 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1797 #endif
1798 if (flat->offset() == TypePtr::OffsetBot) {
1799 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1800 }
1801 }
1802 if (flat->isa_klassptr()) {
1803 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1804 alias_type(idx)->set_rewritable(false);
1805 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1806 alias_type(idx)->set_rewritable(false);
1807 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1808 alias_type(idx)->set_rewritable(false);
1809 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1810 alias_type(idx)->set_rewritable(false);
1811 }
1812 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1813 // but the base pointer type is not distinctive enough to identify
1814 // references into JavaThread.)
1815
1816 // Check for final fields.
1817 const TypeInstPtr* tinst = flat->isa_instptr();
1818 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1819 ciField* field;
1820 if (tinst->const_oop() != NULL &&
1821 tinst->klass() == ciEnv::current()->Class_klass() &&
1822 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1823 // static field
1824 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1825 field = k->get_field_by_offset(tinst->offset(), true);
1826 } else {
1827 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1828 field = k->get_field_by_offset(tinst->offset(), false);
1829 }
1830 assert(field == NULL ||
1831 original_field == NULL ||
1832 (field->holder() == original_field->holder() &&
1833 field->offset() == original_field->offset() &&
1834 field->is_static() == original_field->is_static()), "wrong field?");
1835 // Set field() and is_rewritable() attributes.
1836 if (field != NULL) alias_type(idx)->set_field(field);
1837 }
1838 }
1839
1840 // Fill the cache for next time.
1841 ace->_adr_type = adr_type;
1842 ace->_index = idx;
1843 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1844
1845 // Might as well try to fill the cache for the flattened version, too.
1846 AliasCacheEntry* face = probe_alias_cache(flat);
1847 if (face->_adr_type == NULL) {
1848 face->_adr_type = flat;
1849 face->_index = idx;
1850 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1851 }
1852
1853 return alias_type(idx);
1854 }
1855
1856
alias_type(ciField * field)1857 Compile::AliasType* Compile::alias_type(ciField* field) {
1858 const TypeOopPtr* t;
1859 if (field->is_static())
1860 t = TypeInstPtr::make(field->holder()->java_mirror());
1861 else
1862 t = TypeOopPtr::make_from_klass_raw(field->holder());
1863 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1864 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1865 return atp;
1866 }
1867
1868
1869 //------------------------------have_alias_type--------------------------------
have_alias_type(const TypePtr * adr_type)1870 bool Compile::have_alias_type(const TypePtr* adr_type) {
1871 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1872 if (ace->_adr_type == adr_type) {
1873 return true;
1874 }
1875
1876 // Handle special cases.
1877 if (adr_type == NULL) return true;
1878 if (adr_type == TypePtr::BOTTOM) return true;
1879
1880 return find_alias_type(adr_type, true, NULL) != NULL;
1881 }
1882
1883 //-----------------------------must_alias--------------------------------------
1884 // True if all values of the given address type are in the given alias category.
must_alias(const TypePtr * adr_type,int alias_idx)1885 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1886 if (alias_idx == AliasIdxBot) return true; // the universal category
1887 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1888 if (alias_idx == AliasIdxTop) return false; // the empty category
1889 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1890
1891 // the only remaining possible overlap is identity
1892 int adr_idx = get_alias_index(adr_type);
1893 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1894 assert(adr_idx == alias_idx ||
1895 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1896 && adr_type != TypeOopPtr::BOTTOM),
1897 "should not be testing for overlap with an unsafe pointer");
1898 return adr_idx == alias_idx;
1899 }
1900
1901 //------------------------------can_alias--------------------------------------
1902 // True if any values of the given address type are in the given alias category.
can_alias(const TypePtr * adr_type,int alias_idx)1903 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1904 if (alias_idx == AliasIdxTop) return false; // the empty category
1905 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1906 if (alias_idx == AliasIdxBot) return true; // the universal category
1907 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1908
1909 // the only remaining possible overlap is identity
1910 int adr_idx = get_alias_index(adr_type);
1911 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1912 return adr_idx == alias_idx;
1913 }
1914
1915
1916
1917 //---------------------------pop_warm_call-------------------------------------
pop_warm_call()1918 WarmCallInfo* Compile::pop_warm_call() {
1919 WarmCallInfo* wci = _warm_calls;
1920 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1921 return wci;
1922 }
1923
1924 //----------------------------Inline_Warm--------------------------------------
Inline_Warm()1925 int Compile::Inline_Warm() {
1926 // If there is room, try to inline some more warm call sites.
1927 // %%% Do a graph index compaction pass when we think we're out of space?
1928 if (!InlineWarmCalls) return 0;
1929
1930 int calls_made_hot = 0;
1931 int room_to_grow = NodeCountInliningCutoff - unique();
1932 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1933 int amount_grown = 0;
1934 WarmCallInfo* call;
1935 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1936 int est_size = (int)call->size();
1937 if (est_size > (room_to_grow - amount_grown)) {
1938 // This one won't fit anyway. Get rid of it.
1939 call->make_cold();
1940 continue;
1941 }
1942 call->make_hot();
1943 calls_made_hot++;
1944 amount_grown += est_size;
1945 amount_to_grow -= est_size;
1946 }
1947
1948 if (calls_made_hot > 0) set_major_progress();
1949 return calls_made_hot;
1950 }
1951
1952
1953 //----------------------------Finish_Warm--------------------------------------
Finish_Warm()1954 void Compile::Finish_Warm() {
1955 if (!InlineWarmCalls) return;
1956 if (failing()) return;
1957 if (warm_calls() == NULL) return;
1958
1959 // Clean up loose ends, if we are out of space for inlining.
1960 WarmCallInfo* call;
1961 while ((call = pop_warm_call()) != NULL) {
1962 call->make_cold();
1963 }
1964 }
1965
1966 //---------------------cleanup_loop_predicates-----------------------
1967 // Remove the opaque nodes that protect the predicates so that all unused
1968 // checks and uncommon_traps will be eliminated from the ideal graph
cleanup_loop_predicates(PhaseIterGVN & igvn)1969 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1970 if (predicate_count()==0) return;
1971 for (int i = predicate_count(); i > 0; i--) {
1972 Node * n = predicate_opaque1_node(i-1);
1973 assert(n->Opcode() == Op_Opaque1, "must be");
1974 igvn.replace_node(n, n->in(1));
1975 }
1976 assert(predicate_count()==0, "should be clean!");
1977 }
1978
add_range_check_cast(Node * n)1979 void Compile::add_range_check_cast(Node* n) {
1980 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1981 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1982 _range_check_casts->append(n);
1983 }
1984
1985 // Remove all range check dependent CastIINodes.
remove_range_check_casts(PhaseIterGVN & igvn)1986 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1987 for (int i = range_check_cast_count(); i > 0; i--) {
1988 Node* cast = range_check_cast_node(i-1);
1989 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1990 igvn.replace_node(cast, cast->in(1));
1991 }
1992 assert(range_check_cast_count() == 0, "should be empty");
1993 }
1994
add_opaque4_node(Node * n)1995 void Compile::add_opaque4_node(Node* n) {
1996 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1997 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1998 _opaque4_nodes->append(n);
1999 }
2000
2001 // Remove all Opaque4 nodes.
remove_opaque4_nodes(PhaseIterGVN & igvn)2002 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2003 for (int i = opaque4_count(); i > 0; i--) {
2004 Node* opaq = opaque4_node(i-1);
2005 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2006 igvn.replace_node(opaq, opaq->in(2));
2007 }
2008 assert(opaque4_count() == 0, "should be empty");
2009 }
2010
2011 // StringOpts and late inlining of string methods
inline_string_calls(bool parse_time)2012 void Compile::inline_string_calls(bool parse_time) {
2013 {
2014 // remove useless nodes to make the usage analysis simpler
2015 ResourceMark rm;
2016 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2017 }
2018
2019 {
2020 ResourceMark rm;
2021 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2022 PhaseStringOpts pso(initial_gvn(), for_igvn());
2023 print_method(PHASE_AFTER_STRINGOPTS, 3);
2024 }
2025
2026 // now inline anything that we skipped the first time around
2027 if (!parse_time) {
2028 _late_inlines_pos = _late_inlines.length();
2029 }
2030
2031 while (_string_late_inlines.length() > 0) {
2032 CallGenerator* cg = _string_late_inlines.pop();
2033 cg->do_late_inline();
2034 if (failing()) return;
2035 }
2036 _string_late_inlines.trunc_to(0);
2037 }
2038
2039 // Late inlining of boxing methods
inline_boxing_calls(PhaseIterGVN & igvn)2040 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2041 if (_boxing_late_inlines.length() > 0) {
2042 assert(has_boxed_value(), "inconsistent");
2043
2044 PhaseGVN* gvn = initial_gvn();
2045 set_inlining_incrementally(true);
2046
2047 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2048 for_igvn()->clear();
2049 gvn->replace_with(&igvn);
2050
2051 _late_inlines_pos = _late_inlines.length();
2052
2053 while (_boxing_late_inlines.length() > 0) {
2054 CallGenerator* cg = _boxing_late_inlines.pop();
2055 cg->do_late_inline();
2056 if (failing()) return;
2057 }
2058 _boxing_late_inlines.trunc_to(0);
2059
2060 inline_incrementally_cleanup(igvn);
2061
2062 set_inlining_incrementally(false);
2063 }
2064 }
2065
inline_incrementally_one()2066 bool Compile::inline_incrementally_one() {
2067 assert(IncrementalInline, "incremental inlining should be on");
2068
2069 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2070 set_inlining_progress(false);
2071 set_do_cleanup(false);
2072 int i = 0;
2073 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2074 CallGenerator* cg = _late_inlines.at(i);
2075 _late_inlines_pos = i+1;
2076 cg->do_late_inline();
2077 if (failing()) return false;
2078 }
2079 int j = 0;
2080 for (; i < _late_inlines.length(); i++, j++) {
2081 _late_inlines.at_put(j, _late_inlines.at(i));
2082 }
2083 _late_inlines.trunc_to(j);
2084 assert(inlining_progress() || _late_inlines.length() == 0, "");
2085
2086 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2087
2088 set_inlining_progress(false);
2089 set_do_cleanup(false);
2090 return (_late_inlines.length() > 0) && !needs_cleanup;
2091 }
2092
inline_incrementally_cleanup(PhaseIterGVN & igvn)2093 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2094 {
2095 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2096 ResourceMark rm;
2097 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2098 }
2099 {
2100 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2101 igvn = PhaseIterGVN(initial_gvn());
2102 igvn.optimize();
2103 }
2104 }
2105
2106 // Perform incremental inlining until bound on number of live nodes is reached
inline_incrementally(PhaseIterGVN & igvn)2107 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2108 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2109
2110 set_inlining_incrementally(true);
2111 uint low_live_nodes = 0;
2112
2113 while (_late_inlines.length() > 0) {
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::optimize(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; // finish
2128 }
2129 }
2130
2131 for_igvn()->clear();
2132 initial_gvn()->replace_with(&igvn);
2133
2134 while (inline_incrementally_one()) {
2135 assert(!failing(), "inconsistent");
2136 }
2137
2138 if (failing()) return;
2139
2140 inline_incrementally_cleanup(igvn);
2141
2142 if (failing()) return;
2143 }
2144 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2145
2146 if (_string_late_inlines.length() > 0) {
2147 assert(has_stringbuilder(), "inconsistent");
2148 for_igvn()->clear();
2149 initial_gvn()->replace_with(&igvn);
2150
2151 inline_string_calls(false);
2152
2153 if (failing()) return;
2154
2155 inline_incrementally_cleanup(igvn);
2156 }
2157
2158 set_inlining_incrementally(false);
2159 }
2160
2161
optimize_loops(PhaseIterGVN & igvn,LoopOptsMode mode)2162 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2163 if(_loop_opts_cnt > 0) {
2164 debug_only( int cnt = 0; );
2165 while(major_progress() && (_loop_opts_cnt > 0)) {
2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2167 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2168 PhaseIdealLoop::optimize(igvn, mode);
2169 _loop_opts_cnt--;
2170 if (failing()) return false;
2171 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2172 }
2173 }
2174 return true;
2175 }
2176
2177 // Remove edges from "root" to each SafePoint at a backward branch.
2178 // They were inserted during parsing (see add_safepoint()) to make
2179 // infinite loops without calls or exceptions visible to root, i.e.,
2180 // useful.
remove_root_to_sfpts_edges(PhaseIterGVN & igvn)2181 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2182 Node *r = root();
2183 if (r != NULL) {
2184 for (uint i = r->req(); i < r->len(); ++i) {
2185 Node *n = r->in(i);
2186 if (n != NULL && n->is_SafePoint()) {
2187 r->rm_prec(i);
2188 if (n->outcnt() == 0) {
2189 igvn.remove_dead_node(n);
2190 }
2191 --i;
2192 }
2193 }
2194 }
2195 }
2196
2197 //------------------------------Optimize---------------------------------------
2198 // Given a graph, optimize it.
Optimize()2199 void Compile::Optimize() {
2200 TracePhase tp("optimizer", &timers[_t_optimizer]);
2201
2202 #ifndef PRODUCT
2203 if (_directive->BreakAtCompileOption) {
2204 BREAKPOINT;
2205 }
2206
2207 #endif
2208
2209 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2210 #ifdef ASSERT
2211 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2212 #endif
2213
2214 ResourceMark rm;
2215
2216 print_inlining_reinit();
2217
2218 NOT_PRODUCT( verify_graph_edges(); )
2219
2220 print_method(PHASE_AFTER_PARSING);
2221
2222 {
2223 // Iterative Global Value Numbering, including ideal transforms
2224 // Initialize IterGVN with types and values from parse-time GVN
2225 PhaseIterGVN igvn(initial_gvn());
2226 #ifdef ASSERT
2227 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2228 #endif
2229 {
2230 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2231 igvn.optimize();
2232 }
2233
2234 if (failing()) return;
2235
2236 print_method(PHASE_ITER_GVN1, 2);
2237
2238 inline_incrementally(igvn);
2239
2240 print_method(PHASE_INCREMENTAL_INLINE, 2);
2241
2242 if (failing()) return;
2243
2244 if (eliminate_boxing()) {
2245 // Inline valueOf() methods now.
2246 inline_boxing_calls(igvn);
2247
2248 if (AlwaysIncrementalInline) {
2249 inline_incrementally(igvn);
2250 }
2251
2252 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2253
2254 if (failing()) return;
2255 }
2256
2257 // Now that all inlining is over, cut edge from root to loop
2258 // safepoints
2259 remove_root_to_sfpts_edges(igvn);
2260
2261 // Remove the speculative part of types and clean up the graph from
2262 // the extra CastPP nodes whose only purpose is to carry them. Do
2263 // that early so that optimizations are not disrupted by the extra
2264 // CastPP nodes.
2265 remove_speculative_types(igvn);
2266
2267 // No more new expensive nodes will be added to the list from here
2268 // so keep only the actual candidates for optimizations.
2269 cleanup_expensive_nodes(igvn);
2270
2271 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2272 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2273 initial_gvn()->replace_with(&igvn);
2274 for_igvn()->clear();
2275 Unique_Node_List new_worklist(C->comp_arena());
2276 {
2277 ResourceMark rm;
2278 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2279 }
2280 set_for_igvn(&new_worklist);
2281 igvn = PhaseIterGVN(initial_gvn());
2282 igvn.optimize();
2283 }
2284
2285 // Perform escape analysis
2286 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2287 if (has_loops()) {
2288 // Cleanup graph (remove dead nodes).
2289 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2290 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2291 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2292 if (failing()) return;
2293 }
2294 ConnectionGraph::do_analysis(this, &igvn);
2295
2296 if (failing()) return;
2297
2298 // Optimize out fields loads from scalar replaceable allocations.
2299 igvn.optimize();
2300 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2301
2302 if (failing()) return;
2303
2304 if (congraph() != NULL && macro_count() > 0) {
2305 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2306 PhaseMacroExpand mexp(igvn);
2307 mexp.eliminate_macro_nodes();
2308 igvn.set_delay_transform(false);
2309
2310 igvn.optimize();
2311 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2312
2313 if (failing()) return;
2314 }
2315 }
2316
2317 // Loop transforms on the ideal graph. Range Check Elimination,
2318 // peeling, unrolling, etc.
2319
2320 // Set loop opts counter
2321 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2322 {
2323 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2324 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2325 _loop_opts_cnt--;
2326 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2327 if (failing()) return;
2328 }
2329 // Loop opts pass if partial peeling occurred in previous pass
2330 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2331 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2332 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2333 _loop_opts_cnt--;
2334 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2335 if (failing()) return;
2336 }
2337 // Loop opts pass for loop-unrolling before CCP
2338 if(major_progress() && (_loop_opts_cnt > 0)) {
2339 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2340 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2341 _loop_opts_cnt--;
2342 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2343 }
2344 if (!failing()) {
2345 // Verify that last round of loop opts produced a valid graph
2346 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2347 PhaseIdealLoop::verify(igvn);
2348 }
2349 }
2350 if (failing()) return;
2351
2352 // Conditional Constant Propagation;
2353 PhaseCCP ccp( &igvn );
2354 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2355 {
2356 TracePhase tp("ccp", &timers[_t_ccp]);
2357 ccp.do_transform();
2358 }
2359 print_method(PHASE_CPP1, 2);
2360
2361 assert( true, "Break here to ccp.dump_old2new_map()");
2362
2363 // Iterative Global Value Numbering, including ideal transforms
2364 {
2365 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2366 igvn = ccp;
2367 igvn.optimize();
2368 }
2369 print_method(PHASE_ITER_GVN2, 2);
2370
2371 if (failing()) return;
2372
2373 // Loop transforms on the ideal graph. Range Check Elimination,
2374 // peeling, unrolling, etc.
2375 if (!optimize_loops(igvn, LoopOptsDefault)) {
2376 return;
2377 }
2378
2379 if (failing()) return;
2380
2381 // Ensure that major progress is now clear
2382 C->clear_major_progress();
2383
2384 {
2385 // Verify that all previous optimizations produced a valid graph
2386 // at least to this point, even if no loop optimizations were done.
2387 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2388 PhaseIdealLoop::verify(igvn);
2389 }
2390
2391 if (range_check_cast_count() > 0) {
2392 // No more loop optimizations. Remove all range check dependent CastIINodes.
2393 C->remove_range_check_casts(igvn);
2394 igvn.optimize();
2395 }
2396
2397 #ifdef ASSERT
2398 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2399 #endif
2400
2401 {
2402 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2403 PhaseMacroExpand mex(igvn);
2404 if (mex.expand_macro_nodes()) {
2405 assert(failing(), "must bail out w/ explicit message");
2406 return;
2407 }
2408 print_method(PHASE_MACRO_EXPANSION, 2);
2409 }
2410
2411 {
2412 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2413 if (bs->expand_barriers(this, igvn)) {
2414 assert(failing(), "must bail out w/ explicit message");
2415 return;
2416 }
2417 print_method(PHASE_BARRIER_EXPANSION, 2);
2418 }
2419
2420 if (opaque4_count() > 0) {
2421 C->remove_opaque4_nodes(igvn);
2422 igvn.optimize();
2423 }
2424
2425 DEBUG_ONLY( _modified_nodes = NULL; )
2426 } // (End scope of igvn; run destructor if necessary for asserts.)
2427
2428 process_print_inlining();
2429 // A method with only infinite loops has no edges entering loops from root
2430 {
2431 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2432 if (final_graph_reshaping()) {
2433 assert(failing(), "must bail out w/ explicit message");
2434 return;
2435 }
2436 }
2437
2438 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2439 }
2440
2441
2442 //------------------------------Code_Gen---------------------------------------
2443 // Given a graph, generate code for it
Code_Gen()2444 void Compile::Code_Gen() {
2445 if (failing()) {
2446 return;
2447 }
2448
2449 // Perform instruction selection. You might think we could reclaim Matcher
2450 // memory PDQ, but actually the Matcher is used in generating spill code.
2451 // Internals of the Matcher (including some VectorSets) must remain live
2452 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2453 // set a bit in reclaimed memory.
2454
2455 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2456 // nodes. Mapping is only valid at the root of each matched subtree.
2457 NOT_PRODUCT( verify_graph_edges(); )
2458
2459 Matcher matcher;
2460 _matcher = &matcher;
2461 {
2462 TracePhase tp("matcher", &timers[_t_matcher]);
2463 matcher.match();
2464 if (failing()) {
2465 return;
2466 }
2467 }
2468
2469 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2470 // nodes. Mapping is only valid at the root of each matched subtree.
2471 NOT_PRODUCT( verify_graph_edges(); )
2472
2473 // If you have too many nodes, or if matching has failed, bail out
2474 check_node_count(0, "out of nodes matching instructions");
2475 if (failing()) {
2476 return;
2477 }
2478
2479 print_method(PHASE_MATCHING, 2);
2480
2481 // Build a proper-looking CFG
2482 PhaseCFG cfg(node_arena(), root(), matcher);
2483 _cfg = &cfg;
2484 {
2485 TracePhase tp("scheduler", &timers[_t_scheduler]);
2486 bool success = cfg.do_global_code_motion();
2487 if (!success) {
2488 return;
2489 }
2490
2491 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2492 NOT_PRODUCT( verify_graph_edges(); )
2493 debug_only( cfg.verify(); )
2494 }
2495
2496 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2497 _regalloc = ®alloc;
2498 {
2499 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2500 // Perform register allocation. After Chaitin, use-def chains are
2501 // no longer accurate (at spill code) and so must be ignored.
2502 // Node->LRG->reg mappings are still accurate.
2503 _regalloc->Register_Allocate();
2504
2505 // Bail out if the allocator builds too many nodes
2506 if (failing()) {
2507 return;
2508 }
2509 }
2510
2511 // Prior to register allocation we kept empty basic blocks in case the
2512 // the allocator needed a place to spill. After register allocation we
2513 // are not adding any new instructions. If any basic block is empty, we
2514 // can now safely remove it.
2515 {
2516 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2517 cfg.remove_empty_blocks();
2518 if (do_freq_based_layout()) {
2519 PhaseBlockLayout layout(cfg);
2520 } else {
2521 cfg.set_loop_alignment();
2522 }
2523 cfg.fixup_flow();
2524 }
2525
2526 // Apply peephole optimizations
2527 if( OptoPeephole ) {
2528 TracePhase tp("peephole", &timers[_t_peephole]);
2529 PhasePeephole peep( _regalloc, cfg);
2530 peep.do_transform();
2531 }
2532
2533 // Do late expand if CPU requires this.
2534 if (Matcher::require_postalloc_expand) {
2535 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2536 cfg.postalloc_expand(_regalloc);
2537 }
2538
2539 // Convert Nodes to instruction bits in a buffer
2540 {
2541 TraceTime tp("output", &timers[_t_output], CITime);
2542 Output();
2543 }
2544
2545 print_method(PHASE_FINAL_CODE);
2546
2547 // He's dead, Jim.
2548 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2549 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2550 }
2551
2552
2553 //------------------------------dump_asm---------------------------------------
2554 // Dump formatted assembly
2555 #if defined(SUPPORT_OPTO_ASSEMBLY)
dump_asm_on(outputStream * st,int * pcs,uint pc_limit)2556 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2557
2558 int pc_digits = 3; // #chars required for pc
2559 int sb_chars = 3; // #chars for "start bundle" indicator
2560 int tab_size = 8;
2561 if (pcs != NULL) {
2562 int max_pc = 0;
2563 for (uint i = 0; i < pc_limit; i++) {
2564 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
2565 }
2566 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
2567 }
2568 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
2569
2570 bool cut_short = false;
2571 st->print_cr("#");
2572 st->print("# "); _tf->dump_on(st); st->cr();
2573 st->print_cr("#");
2574
2575 // For all blocks
2576 int pc = 0x0; // Program counter
2577 char starts_bundle = ' ';
2578 _regalloc->dump_frame();
2579
2580 Node *n = NULL;
2581 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2582 if (VMThread::should_terminate()) {
2583 cut_short = true;
2584 break;
2585 }
2586 Block* block = _cfg->get_block(i);
2587 if (block->is_connector() && !Verbose) {
2588 continue;
2589 }
2590 n = block->head();
2591 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2592 pc = pcs[n->_idx];
2593 st->print("%*.*x", pc_digits, pc_digits, pc);
2594 }
2595 st->fill_to(prefix_len);
2596 block->dump_head(_cfg, st);
2597 if (block->is_connector()) {
2598 st->fill_to(prefix_len);
2599 st->print_cr("# Empty connector block");
2600 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2601 st->fill_to(prefix_len);
2602 st->print_cr("# Block is sole successor of call");
2603 }
2604
2605 // For all instructions
2606 Node *delay = NULL;
2607 for (uint j = 0; j < block->number_of_nodes(); j++) {
2608 if (VMThread::should_terminate()) {
2609 cut_short = true;
2610 break;
2611 }
2612 n = block->get_node(j);
2613 if (valid_bundle_info(n)) {
2614 Bundle* bundle = node_bundling(n);
2615 if (bundle->used_in_unconditional_delay()) {
2616 delay = n;
2617 continue;
2618 }
2619 if (bundle->starts_bundle()) {
2620 starts_bundle = '+';
2621 }
2622 }
2623
2624 if (WizardMode) {
2625 n->dump();
2626 }
2627
2628 if( !n->is_Region() && // Dont print in the Assembly
2629 !n->is_Phi() && // a few noisely useless nodes
2630 !n->is_Proj() &&
2631 !n->is_MachTemp() &&
2632 !n->is_SafePointScalarObject() &&
2633 !n->is_Catch() && // Would be nice to print exception table targets
2634 !n->is_MergeMem() && // Not very interesting
2635 !n->is_top() && // Debug info table constants
2636 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2637 ) {
2638 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2639 pc = pcs[n->_idx];
2640 st->print("%*.*x", pc_digits, pc_digits, pc);
2641 } else {
2642 st->fill_to(pc_digits);
2643 }
2644 st->print(" %c ", starts_bundle);
2645 starts_bundle = ' ';
2646 st->fill_to(prefix_len);
2647 n->format(_regalloc, st);
2648 st->cr();
2649 }
2650
2651 // If we have an instruction with a delay slot, and have seen a delay,
2652 // then back up and print it
2653 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2654 // Coverity finding - Explicit null dereferenced.
2655 guarantee(delay != NULL, "no unconditional delay instruction");
2656 if (WizardMode) delay->dump();
2657
2658 if (node_bundling(delay)->starts_bundle())
2659 starts_bundle = '+';
2660 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2661 pc = pcs[n->_idx];
2662 st->print("%*.*x", pc_digits, pc_digits, pc);
2663 } else {
2664 st->fill_to(pc_digits);
2665 }
2666 st->print(" %c ", starts_bundle);
2667 starts_bundle = ' ';
2668 st->fill_to(prefix_len);
2669 delay->format(_regalloc, st);
2670 st->cr();
2671 delay = NULL;
2672 }
2673
2674 // Dump the exception table as well
2675 if( n->is_Catch() && (Verbose || WizardMode) ) {
2676 // Print the exception table for this offset
2677 _handler_table.print_subtable_for(pc);
2678 }
2679 st->bol(); // Make sure we start on a new line
2680 }
2681 st->cr(); // one empty line between blocks
2682 assert(cut_short || delay == NULL, "no unconditional delay branch");
2683 } // End of per-block dump
2684
2685 if (cut_short) st->print_cr("*** disassembly is cut short ***");
2686 }
2687 #endif
2688
2689 //------------------------------Final_Reshape_Counts---------------------------
2690 // This class defines counters to help identify when a method
2691 // may/must be executed using hardware with only 24-bit precision.
2692 struct Final_Reshape_Counts : public StackObj {
2693 int _call_count; // count non-inlined 'common' calls
2694 int _float_count; // count float ops requiring 24-bit precision
2695 int _double_count; // count double ops requiring more precision
2696 int _java_call_count; // count non-inlined 'java' calls
2697 int _inner_loop_count; // count loops which need alignment
2698 VectorSet _visited; // Visitation flags
2699 Node_List _tests; // Set of IfNodes & PCTableNodes
2700
Final_Reshape_CountsFinal_Reshape_Counts2701 Final_Reshape_Counts() :
2702 _call_count(0), _float_count(0), _double_count(0),
2703 _java_call_count(0), _inner_loop_count(0),
2704 _visited( Thread::current()->resource_area() ) { }
2705
inc_call_countFinal_Reshape_Counts2706 void inc_call_count () { _call_count ++; }
inc_float_countFinal_Reshape_Counts2707 void inc_float_count () { _float_count ++; }
inc_double_countFinal_Reshape_Counts2708 void inc_double_count() { _double_count++; }
inc_java_call_countFinal_Reshape_Counts2709 void inc_java_call_count() { _java_call_count++; }
inc_inner_loop_countFinal_Reshape_Counts2710 void inc_inner_loop_count() { _inner_loop_count++; }
2711
get_call_countFinal_Reshape_Counts2712 int get_call_count () const { return _call_count ; }
get_float_countFinal_Reshape_Counts2713 int get_float_count () const { return _float_count ; }
get_double_countFinal_Reshape_Counts2714 int get_double_count() const { return _double_count; }
get_java_call_countFinal_Reshape_Counts2715 int get_java_call_count() const { return _java_call_count; }
get_inner_loop_countFinal_Reshape_Counts2716 int get_inner_loop_count() const { return _inner_loop_count; }
2717 };
2718
2719 #ifdef ASSERT
oop_offset_is_sane(const TypeInstPtr * tp)2720 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2721 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2722 // Make sure the offset goes inside the instance layout.
2723 return k->contains_field_offset(tp->offset());
2724 // Note that OffsetBot and OffsetTop are very negative.
2725 }
2726 #endif
2727
2728 // Eliminate trivially redundant StoreCMs and accumulate their
2729 // precedence edges.
eliminate_redundant_card_marks(Node * n)2730 void Compile::eliminate_redundant_card_marks(Node* n) {
2731 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2732 if (n->in(MemNode::Address)->outcnt() > 1) {
2733 // There are multiple users of the same address so it might be
2734 // possible to eliminate some of the StoreCMs
2735 Node* mem = n->in(MemNode::Memory);
2736 Node* adr = n->in(MemNode::Address);
2737 Node* val = n->in(MemNode::ValueIn);
2738 Node* prev = n;
2739 bool done = false;
2740 // Walk the chain of StoreCMs eliminating ones that match. As
2741 // long as it's a chain of single users then the optimization is
2742 // safe. Eliminating partially redundant StoreCMs would require
2743 // cloning copies down the other paths.
2744 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2745 if (adr == mem->in(MemNode::Address) &&
2746 val == mem->in(MemNode::ValueIn)) {
2747 // redundant StoreCM
2748 if (mem->req() > MemNode::OopStore) {
2749 // Hasn't been processed by this code yet.
2750 n->add_prec(mem->in(MemNode::OopStore));
2751 } else {
2752 // Already converted to precedence edge
2753 for (uint i = mem->req(); i < mem->len(); i++) {
2754 // Accumulate any precedence edges
2755 if (mem->in(i) != NULL) {
2756 n->add_prec(mem->in(i));
2757 }
2758 }
2759 // Everything above this point has been processed.
2760 done = true;
2761 }
2762 // Eliminate the previous StoreCM
2763 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2764 assert(mem->outcnt() == 0, "should be dead");
2765 mem->disconnect_inputs(NULL, this);
2766 } else {
2767 prev = mem;
2768 }
2769 mem = prev->in(MemNode::Memory);
2770 }
2771 }
2772 }
2773
2774 //------------------------------final_graph_reshaping_impl----------------------
2775 // Implement items 1-5 from final_graph_reshaping below.
final_graph_reshaping_impl(Node * n,Final_Reshape_Counts & frc)2776 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2777
2778 if ( n->outcnt() == 0 ) return; // dead node
2779 uint nop = n->Opcode();
2780
2781 // Check for 2-input instruction with "last use" on right input.
2782 // Swap to left input. Implements item (2).
2783 if( n->req() == 3 && // two-input instruction
2784 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2785 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2786 n->in(2)->outcnt() == 1 &&// right use IS a last use
2787 !n->in(2)->is_Con() ) { // right use is not a constant
2788 // Check for commutative opcode
2789 switch( nop ) {
2790 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2791 case Op_MaxI: case Op_MinI:
2792 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2793 case Op_AndL: case Op_XorL: case Op_OrL:
2794 case Op_AndI: case Op_XorI: case Op_OrI: {
2795 // Move "last use" input to left by swapping inputs
2796 n->swap_edges(1, 2);
2797 break;
2798 }
2799 default:
2800 break;
2801 }
2802 }
2803
2804 #ifdef ASSERT
2805 if( n->is_Mem() ) {
2806 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2807 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2808 // oop will be recorded in oop map if load crosses safepoint
2809 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2810 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2811 "raw memory operations should have control edge");
2812 }
2813 if (n->is_MemBar()) {
2814 MemBarNode* mb = n->as_MemBar();
2815 if (mb->trailing_store() || mb->trailing_load_store()) {
2816 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2817 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2818 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2819 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2820 } else if (mb->leading()) {
2821 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2822 }
2823 }
2824 #endif
2825 // Count FPU ops and common calls, implements item (3)
2826 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2827 if (!gc_handled) {
2828 final_graph_reshaping_main_switch(n, frc, nop);
2829 }
2830
2831 // Collect CFG split points
2832 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2833 frc._tests.push(n);
2834 }
2835 }
2836
final_graph_reshaping_main_switch(Node * n,Final_Reshape_Counts & frc,uint nop)2837 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2838 switch( nop ) {
2839 // Count all float operations that may use FPU
2840 case Op_AddF:
2841 case Op_SubF:
2842 case Op_MulF:
2843 case Op_DivF:
2844 case Op_NegF:
2845 case Op_ModF:
2846 case Op_ConvI2F:
2847 case Op_ConF:
2848 case Op_CmpF:
2849 case Op_CmpF3:
2850 // case Op_ConvL2F: // longs are split into 32-bit halves
2851 frc.inc_float_count();
2852 break;
2853
2854 case Op_ConvF2D:
2855 case Op_ConvD2F:
2856 frc.inc_float_count();
2857 frc.inc_double_count();
2858 break;
2859
2860 // Count all double operations that may use FPU
2861 case Op_AddD:
2862 case Op_SubD:
2863 case Op_MulD:
2864 case Op_DivD:
2865 case Op_NegD:
2866 case Op_ModD:
2867 case Op_ConvI2D:
2868 case Op_ConvD2I:
2869 // case Op_ConvL2D: // handled by leaf call
2870 // case Op_ConvD2L: // handled by leaf call
2871 case Op_ConD:
2872 case Op_CmpD:
2873 case Op_CmpD3:
2874 frc.inc_double_count();
2875 break;
2876 case Op_Opaque1: // Remove Opaque Nodes before matching
2877 case Op_Opaque2: // Remove Opaque Nodes before matching
2878 case Op_Opaque3:
2879 n->subsume_by(n->in(1), this);
2880 break;
2881 case Op_CallStaticJava:
2882 case Op_CallJava:
2883 case Op_CallDynamicJava:
2884 frc.inc_java_call_count(); // Count java call site;
2885 case Op_CallRuntime:
2886 case Op_CallLeaf:
2887 case Op_CallLeafNoFP: {
2888 assert (n->is_Call(), "");
2889 CallNode *call = n->as_Call();
2890 // Count call sites where the FP mode bit would have to be flipped.
2891 // Do not count uncommon runtime calls:
2892 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2893 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2894 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2895 frc.inc_call_count(); // Count the call site
2896 } else { // See if uncommon argument is shared
2897 Node *n = call->in(TypeFunc::Parms);
2898 int nop = n->Opcode();
2899 // Clone shared simple arguments to uncommon calls, item (1).
2900 if (n->outcnt() > 1 &&
2901 !n->is_Proj() &&
2902 nop != Op_CreateEx &&
2903 nop != Op_CheckCastPP &&
2904 nop != Op_DecodeN &&
2905 nop != Op_DecodeNKlass &&
2906 !n->is_Mem() &&
2907 !n->is_Phi()) {
2908 Node *x = n->clone();
2909 call->set_req(TypeFunc::Parms, x);
2910 }
2911 }
2912 break;
2913 }
2914
2915 case Op_StoreD:
2916 case Op_LoadD:
2917 case Op_LoadD_unaligned:
2918 frc.inc_double_count();
2919 goto handle_mem;
2920 case Op_StoreF:
2921 case Op_LoadF:
2922 frc.inc_float_count();
2923 goto handle_mem;
2924
2925 case Op_StoreCM:
2926 {
2927 // Convert OopStore dependence into precedence edge
2928 Node* prec = n->in(MemNode::OopStore);
2929 n->del_req(MemNode::OopStore);
2930 n->add_prec(prec);
2931 eliminate_redundant_card_marks(n);
2932 }
2933
2934 // fall through
2935
2936 case Op_StoreB:
2937 case Op_StoreC:
2938 case Op_StorePConditional:
2939 case Op_StoreI:
2940 case Op_StoreL:
2941 case Op_StoreIConditional:
2942 case Op_StoreLConditional:
2943 case Op_CompareAndSwapB:
2944 case Op_CompareAndSwapS:
2945 case Op_CompareAndSwapI:
2946 case Op_CompareAndSwapL:
2947 case Op_CompareAndSwapP:
2948 case Op_CompareAndSwapN:
2949 case Op_WeakCompareAndSwapB:
2950 case Op_WeakCompareAndSwapS:
2951 case Op_WeakCompareAndSwapI:
2952 case Op_WeakCompareAndSwapL:
2953 case Op_WeakCompareAndSwapP:
2954 case Op_WeakCompareAndSwapN:
2955 case Op_CompareAndExchangeB:
2956 case Op_CompareAndExchangeS:
2957 case Op_CompareAndExchangeI:
2958 case Op_CompareAndExchangeL:
2959 case Op_CompareAndExchangeP:
2960 case Op_CompareAndExchangeN:
2961 case Op_GetAndAddS:
2962 case Op_GetAndAddB:
2963 case Op_GetAndAddI:
2964 case Op_GetAndAddL:
2965 case Op_GetAndSetS:
2966 case Op_GetAndSetB:
2967 case Op_GetAndSetI:
2968 case Op_GetAndSetL:
2969 case Op_GetAndSetP:
2970 case Op_GetAndSetN:
2971 case Op_StoreP:
2972 case Op_StoreN:
2973 case Op_StoreNKlass:
2974 case Op_LoadB:
2975 case Op_LoadUB:
2976 case Op_LoadUS:
2977 case Op_LoadI:
2978 case Op_LoadKlass:
2979 case Op_LoadNKlass:
2980 case Op_LoadL:
2981 case Op_LoadL_unaligned:
2982 case Op_LoadPLocked:
2983 case Op_LoadP:
2984 case Op_LoadN:
2985 case Op_LoadRange:
2986 case Op_LoadS: {
2987 handle_mem:
2988 #ifdef ASSERT
2989 if( VerifyOptoOopOffsets ) {
2990 MemNode* mem = n->as_Mem();
2991 // Check to see if address types have grounded out somehow.
2992 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2993 assert( !tp || oop_offset_is_sane(tp), "" );
2994 }
2995 #endif
2996 break;
2997 }
2998
2999 case Op_AddP: { // Assert sane base pointers
3000 Node *addp = n->in(AddPNode::Address);
3001 assert( !addp->is_AddP() ||
3002 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3003 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3004 "Base pointers must match (addp %u)", addp->_idx );
3005 #ifdef _LP64
3006 if ((UseCompressedOops || UseCompressedClassPointers) &&
3007 addp->Opcode() == Op_ConP &&
3008 addp == n->in(AddPNode::Base) &&
3009 n->in(AddPNode::Offset)->is_Con()) {
3010 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3011 // on the platform and on the compressed oops mode.
3012 // Use addressing with narrow klass to load with offset on x86.
3013 // Some platforms can use the constant pool to load ConP.
3014 // Do this transformation here since IGVN will convert ConN back to ConP.
3015 const Type* t = addp->bottom_type();
3016 bool is_oop = t->isa_oopptr() != NULL;
3017 bool is_klass = t->isa_klassptr() != NULL;
3018
3019 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3020 (is_klass && Matcher::const_klass_prefer_decode())) {
3021 Node* nn = NULL;
3022
3023 int op = is_oop ? Op_ConN : Op_ConNKlass;
3024
3025 // Look for existing ConN node of the same exact type.
3026 Node* r = root();
3027 uint cnt = r->outcnt();
3028 for (uint i = 0; i < cnt; i++) {
3029 Node* m = r->raw_out(i);
3030 if (m!= NULL && m->Opcode() == op &&
3031 m->bottom_type()->make_ptr() == t) {
3032 nn = m;
3033 break;
3034 }
3035 }
3036 if (nn != NULL) {
3037 // Decode a narrow oop to match address
3038 // [R12 + narrow_oop_reg<<3 + offset]
3039 if (is_oop) {
3040 nn = new DecodeNNode(nn, t);
3041 } else {
3042 nn = new DecodeNKlassNode(nn, t);
3043 }
3044 // Check for succeeding AddP which uses the same Base.
3045 // Otherwise we will run into the assertion above when visiting that guy.
3046 for (uint i = 0; i < n->outcnt(); ++i) {
3047 Node *out_i = n->raw_out(i);
3048 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3049 out_i->set_req(AddPNode::Base, nn);
3050 #ifdef ASSERT
3051 for (uint j = 0; j < out_i->outcnt(); ++j) {
3052 Node *out_j = out_i->raw_out(j);
3053 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3054 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3055 }
3056 #endif
3057 }
3058 }
3059 n->set_req(AddPNode::Base, nn);
3060 n->set_req(AddPNode::Address, nn);
3061 if (addp->outcnt() == 0) {
3062 addp->disconnect_inputs(NULL, this);
3063 }
3064 }
3065 }
3066 }
3067 #endif
3068 // platform dependent reshaping of the address expression
3069 reshape_address(n->as_AddP());
3070 break;
3071 }
3072
3073 case Op_CastPP: {
3074 // Remove CastPP nodes to gain more freedom during scheduling but
3075 // keep the dependency they encode as control or precedence edges
3076 // (if control is set already) on memory operations. Some CastPP
3077 // nodes don't have a control (don't carry a dependency): skip
3078 // those.
3079 if (n->in(0) != NULL) {
3080 ResourceMark rm;
3081 Unique_Node_List wq;
3082 wq.push(n);
3083 for (uint next = 0; next < wq.size(); ++next) {
3084 Node *m = wq.at(next);
3085 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3086 Node* use = m->fast_out(i);
3087 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3088 use->ensure_control_or_add_prec(n->in(0));
3089 } else {
3090 switch(use->Opcode()) {
3091 case Op_AddP:
3092 case Op_DecodeN:
3093 case Op_DecodeNKlass:
3094 case Op_CheckCastPP:
3095 case Op_CastPP:
3096 wq.push(use);
3097 break;
3098 }
3099 }
3100 }
3101 }
3102 }
3103 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3104 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3105 Node* in1 = n->in(1);
3106 const Type* t = n->bottom_type();
3107 Node* new_in1 = in1->clone();
3108 new_in1->as_DecodeN()->set_type(t);
3109
3110 if (!Matcher::narrow_oop_use_complex_address()) {
3111 //
3112 // x86, ARM and friends can handle 2 adds in addressing mode
3113 // and Matcher can fold a DecodeN node into address by using
3114 // a narrow oop directly and do implicit NULL check in address:
3115 //
3116 // [R12 + narrow_oop_reg<<3 + offset]
3117 // NullCheck narrow_oop_reg
3118 //
3119 // On other platforms (Sparc) we have to keep new DecodeN node and
3120 // use it to do implicit NULL check in address:
3121 //
3122 // decode_not_null narrow_oop_reg, base_reg
3123 // [base_reg + offset]
3124 // NullCheck base_reg
3125 //
3126 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3127 // to keep the information to which NULL check the new DecodeN node
3128 // corresponds to use it as value in implicit_null_check().
3129 //
3130 new_in1->set_req(0, n->in(0));
3131 }
3132
3133 n->subsume_by(new_in1, this);
3134 if (in1->outcnt() == 0) {
3135 in1->disconnect_inputs(NULL, this);
3136 }
3137 } else {
3138 n->subsume_by(n->in(1), this);
3139 if (n->outcnt() == 0) {
3140 n->disconnect_inputs(NULL, this);
3141 }
3142 }
3143 break;
3144 }
3145 #ifdef _LP64
3146 case Op_CmpP:
3147 // Do this transformation here to preserve CmpPNode::sub() and
3148 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3149 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3150 Node* in1 = n->in(1);
3151 Node* in2 = n->in(2);
3152 if (!in1->is_DecodeNarrowPtr()) {
3153 in2 = in1;
3154 in1 = n->in(2);
3155 }
3156 assert(in1->is_DecodeNarrowPtr(), "sanity");
3157
3158 Node* new_in2 = NULL;
3159 if (in2->is_DecodeNarrowPtr()) {
3160 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3161 new_in2 = in2->in(1);
3162 } else if (in2->Opcode() == Op_ConP) {
3163 const Type* t = in2->bottom_type();
3164 if (t == TypePtr::NULL_PTR) {
3165 assert(in1->is_DecodeN(), "compare klass to null?");
3166 // Don't convert CmpP null check into CmpN if compressed
3167 // oops implicit null check is not generated.
3168 // This will allow to generate normal oop implicit null check.
3169 if (Matcher::gen_narrow_oop_implicit_null_checks())
3170 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3171 //
3172 // This transformation together with CastPP transformation above
3173 // will generated code for implicit NULL checks for compressed oops.
3174 //
3175 // The original code after Optimize()
3176 //
3177 // LoadN memory, narrow_oop_reg
3178 // decode narrow_oop_reg, base_reg
3179 // CmpP base_reg, NULL
3180 // CastPP base_reg // NotNull
3181 // Load [base_reg + offset], val_reg
3182 //
3183 // after these transformations will be
3184 //
3185 // LoadN memory, narrow_oop_reg
3186 // CmpN narrow_oop_reg, NULL
3187 // decode_not_null narrow_oop_reg, base_reg
3188 // Load [base_reg + offset], val_reg
3189 //
3190 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3191 // since narrow oops can be used in debug info now (see the code in
3192 // final_graph_reshaping_walk()).
3193 //
3194 // At the end the code will be matched to
3195 // on x86:
3196 //
3197 // Load_narrow_oop memory, narrow_oop_reg
3198 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3199 // NullCheck narrow_oop_reg
3200 //
3201 // and on sparc:
3202 //
3203 // Load_narrow_oop memory, narrow_oop_reg
3204 // decode_not_null narrow_oop_reg, base_reg
3205 // Load [base_reg + offset], val_reg
3206 // NullCheck base_reg
3207 //
3208 } else if (t->isa_oopptr()) {
3209 new_in2 = ConNode::make(t->make_narrowoop());
3210 } else if (t->isa_klassptr()) {
3211 new_in2 = ConNode::make(t->make_narrowklass());
3212 }
3213 }
3214 if (new_in2 != NULL) {
3215 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3216 n->subsume_by(cmpN, this);
3217 if (in1->outcnt() == 0) {
3218 in1->disconnect_inputs(NULL, this);
3219 }
3220 if (in2->outcnt() == 0) {
3221 in2->disconnect_inputs(NULL, this);
3222 }
3223 }
3224 }
3225 break;
3226
3227 case Op_DecodeN:
3228 case Op_DecodeNKlass:
3229 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3230 // DecodeN could be pinned when it can't be fold into
3231 // an address expression, see the code for Op_CastPP above.
3232 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3233 break;
3234
3235 case Op_EncodeP:
3236 case Op_EncodePKlass: {
3237 Node* in1 = n->in(1);
3238 if (in1->is_DecodeNarrowPtr()) {
3239 n->subsume_by(in1->in(1), this);
3240 } else if (in1->Opcode() == Op_ConP) {
3241 const Type* t = in1->bottom_type();
3242 if (t == TypePtr::NULL_PTR) {
3243 assert(t->isa_oopptr(), "null klass?");
3244 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3245 } else if (t->isa_oopptr()) {
3246 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3247 } else if (t->isa_klassptr()) {
3248 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3249 }
3250 }
3251 if (in1->outcnt() == 0) {
3252 in1->disconnect_inputs(NULL, this);
3253 }
3254 break;
3255 }
3256
3257 case Op_Proj: {
3258 if (OptimizeStringConcat) {
3259 ProjNode* p = n->as_Proj();
3260 if (p->_is_io_use) {
3261 // Separate projections were used for the exception path which
3262 // are normally removed by a late inline. If it wasn't inlined
3263 // then they will hang around and should just be replaced with
3264 // the original one.
3265 Node* proj = NULL;
3266 // Replace with just one
3267 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3268 Node *use = i.get();
3269 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3270 proj = use;
3271 break;
3272 }
3273 }
3274 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3275 if (proj != NULL) {
3276 p->subsume_by(proj, this);
3277 }
3278 }
3279 }
3280 break;
3281 }
3282
3283 case Op_Phi:
3284 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3285 // The EncodeP optimization may create Phi with the same edges
3286 // for all paths. It is not handled well by Register Allocator.
3287 Node* unique_in = n->in(1);
3288 assert(unique_in != NULL, "");
3289 uint cnt = n->req();
3290 for (uint i = 2; i < cnt; i++) {
3291 Node* m = n->in(i);
3292 assert(m != NULL, "");
3293 if (unique_in != m)
3294 unique_in = NULL;
3295 }
3296 if (unique_in != NULL) {
3297 n->subsume_by(unique_in, this);
3298 }
3299 }
3300 break;
3301
3302 #endif
3303
3304 #ifdef ASSERT
3305 case Op_CastII:
3306 // Verify that all range check dependent CastII nodes were removed.
3307 if (n->isa_CastII()->has_range_check()) {
3308 n->dump(3);
3309 assert(false, "Range check dependent CastII node was not removed");
3310 }
3311 break;
3312 #endif
3313
3314 case Op_ModI:
3315 if (UseDivMod) {
3316 // Check if a%b and a/b both exist
3317 Node* d = n->find_similar(Op_DivI);
3318 if (d) {
3319 // Replace them with a fused divmod if supported
3320 if (Matcher::has_match_rule(Op_DivModI)) {
3321 DivModINode* divmod = DivModINode::make(n);
3322 d->subsume_by(divmod->div_proj(), this);
3323 n->subsume_by(divmod->mod_proj(), this);
3324 } else {
3325 // replace a%b with a-((a/b)*b)
3326 Node* mult = new MulINode(d, d->in(2));
3327 Node* sub = new SubINode(d->in(1), mult);
3328 n->subsume_by(sub, this);
3329 }
3330 }
3331 }
3332 break;
3333
3334 case Op_ModL:
3335 if (UseDivMod) {
3336 // Check if a%b and a/b both exist
3337 Node* d = n->find_similar(Op_DivL);
3338 if (d) {
3339 // Replace them with a fused divmod if supported
3340 if (Matcher::has_match_rule(Op_DivModL)) {
3341 DivModLNode* divmod = DivModLNode::make(n);
3342 d->subsume_by(divmod->div_proj(), this);
3343 n->subsume_by(divmod->mod_proj(), this);
3344 } else {
3345 // replace a%b with a-((a/b)*b)
3346 Node* mult = new MulLNode(d, d->in(2));
3347 Node* sub = new SubLNode(d->in(1), mult);
3348 n->subsume_by(sub, this);
3349 }
3350 }
3351 }
3352 break;
3353
3354 case Op_LoadVector:
3355 case Op_StoreVector:
3356 break;
3357
3358 case Op_AddReductionVI:
3359 case Op_AddReductionVL:
3360 case Op_AddReductionVF:
3361 case Op_AddReductionVD:
3362 case Op_MulReductionVI:
3363 case Op_MulReductionVL:
3364 case Op_MulReductionVF:
3365 case Op_MulReductionVD:
3366 case Op_MinReductionV:
3367 case Op_MaxReductionV:
3368 break;
3369
3370 case Op_PackB:
3371 case Op_PackS:
3372 case Op_PackI:
3373 case Op_PackF:
3374 case Op_PackL:
3375 case Op_PackD:
3376 if (n->req()-1 > 2) {
3377 // Replace many operand PackNodes with a binary tree for matching
3378 PackNode* p = (PackNode*) n;
3379 Node* btp = p->binary_tree_pack(1, n->req());
3380 n->subsume_by(btp, this);
3381 }
3382 break;
3383 case Op_Loop:
3384 case Op_CountedLoop:
3385 case Op_OuterStripMinedLoop:
3386 if (n->as_Loop()->is_inner_loop()) {
3387 frc.inc_inner_loop_count();
3388 }
3389 n->as_Loop()->verify_strip_mined(0);
3390 break;
3391 case Op_LShiftI:
3392 case Op_RShiftI:
3393 case Op_URShiftI:
3394 case Op_LShiftL:
3395 case Op_RShiftL:
3396 case Op_URShiftL:
3397 if (Matcher::need_masked_shift_count) {
3398 // The cpu's shift instructions don't restrict the count to the
3399 // lower 5/6 bits. We need to do the masking ourselves.
3400 Node* in2 = n->in(2);
3401 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3402 const TypeInt* t = in2->find_int_type();
3403 if (t != NULL && t->is_con()) {
3404 juint shift = t->get_con();
3405 if (shift > mask) { // Unsigned cmp
3406 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3407 }
3408 } else {
3409 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3410 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3411 n->set_req(2, shift);
3412 }
3413 }
3414 if (in2->outcnt() == 0) { // Remove dead node
3415 in2->disconnect_inputs(NULL, this);
3416 }
3417 }
3418 break;
3419 case Op_MemBarStoreStore:
3420 case Op_MemBarRelease:
3421 // Break the link with AllocateNode: it is no longer useful and
3422 // confuses register allocation.
3423 if (n->req() > MemBarNode::Precedent) {
3424 n->set_req(MemBarNode::Precedent, top());
3425 }
3426 break;
3427 case Op_MemBarAcquire: {
3428 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3429 // At parse time, the trailing MemBarAcquire for a volatile load
3430 // is created with an edge to the load. After optimizations,
3431 // that input may be a chain of Phis. If those phis have no
3432 // other use, then the MemBarAcquire keeps them alive and
3433 // register allocation can be confused.
3434 ResourceMark rm;
3435 Unique_Node_List wq;
3436 wq.push(n->in(MemBarNode::Precedent));
3437 n->set_req(MemBarNode::Precedent, top());
3438 while (wq.size() > 0) {
3439 Node* m = wq.pop();
3440 if (m->outcnt() == 0) {
3441 for (uint j = 0; j < m->req(); j++) {
3442 Node* in = m->in(j);
3443 if (in != NULL) {
3444 wq.push(in);
3445 }
3446 }
3447 m->disconnect_inputs(NULL, this);
3448 }
3449 }
3450 }
3451 break;
3452 }
3453 case Op_RangeCheck: {
3454 RangeCheckNode* rc = n->as_RangeCheck();
3455 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3456 n->subsume_by(iff, this);
3457 frc._tests.push(iff);
3458 break;
3459 }
3460 case Op_ConvI2L: {
3461 if (!Matcher::convi2l_type_required) {
3462 // Code generation on some platforms doesn't need accurate
3463 // ConvI2L types. Widening the type can help remove redundant
3464 // address computations.
3465 n->as_Type()->set_type(TypeLong::INT);
3466 ResourceMark rm;
3467 Unique_Node_List wq;
3468 wq.push(n);
3469 for (uint next = 0; next < wq.size(); next++) {
3470 Node *m = wq.at(next);
3471
3472 for(;;) {
3473 // Loop over all nodes with identical inputs edges as m
3474 Node* k = m->find_similar(m->Opcode());
3475 if (k == NULL) {
3476 break;
3477 }
3478 // Push their uses so we get a chance to remove node made
3479 // redundant
3480 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3481 Node* u = k->fast_out(i);
3482 if (u->Opcode() == Op_LShiftL ||
3483 u->Opcode() == Op_AddL ||
3484 u->Opcode() == Op_SubL ||
3485 u->Opcode() == Op_AddP) {
3486 wq.push(u);
3487 }
3488 }
3489 // Replace all nodes with identical edges as m with m
3490 k->subsume_by(m, this);
3491 }
3492 }
3493 }
3494 break;
3495 }
3496 case Op_CmpUL: {
3497 if (!Matcher::has_match_rule(Op_CmpUL)) {
3498 // No support for unsigned long comparisons
3499 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3500 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3501 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3502 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3503 Node* andl = new AndLNode(orl, remove_sign_mask);
3504 Node* cmp = new CmpLNode(andl, n->in(2));
3505 n->subsume_by(cmp, this);
3506 }
3507 break;
3508 }
3509 default:
3510 assert(!n->is_Call(), "");
3511 assert(!n->is_Mem(), "");
3512 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3513 break;
3514 }
3515 }
3516
3517 //------------------------------final_graph_reshaping_walk---------------------
3518 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3519 // requires that the walk visits a node's inputs before visiting the node.
final_graph_reshaping_walk(Node_Stack & nstack,Node * root,Final_Reshape_Counts & frc)3520 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3521 ResourceArea *area = Thread::current()->resource_area();
3522 Unique_Node_List sfpt(area);
3523
3524 frc._visited.set(root->_idx); // first, mark node as visited
3525 uint cnt = root->req();
3526 Node *n = root;
3527 uint i = 0;
3528 while (true) {
3529 if (i < cnt) {
3530 // Place all non-visited non-null inputs onto stack
3531 Node* m = n->in(i);
3532 ++i;
3533 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3534 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3535 // compute worst case interpreter size in case of a deoptimization
3536 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3537
3538 sfpt.push(m);
3539 }
3540 cnt = m->req();
3541 nstack.push(n, i); // put on stack parent and next input's index
3542 n = m;
3543 i = 0;
3544 }
3545 } else {
3546 // Now do post-visit work
3547 final_graph_reshaping_impl( n, frc );
3548 if (nstack.is_empty())
3549 break; // finished
3550 n = nstack.node(); // Get node from stack
3551 cnt = n->req();
3552 i = nstack.index();
3553 nstack.pop(); // Shift to the next node on stack
3554 }
3555 }
3556
3557 // Skip next transformation if compressed oops are not used.
3558 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3559 (!UseCompressedOops && !UseCompressedClassPointers))
3560 return;
3561
3562 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3563 // It could be done for an uncommon traps or any safepoints/calls
3564 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3565 while (sfpt.size() > 0) {
3566 n = sfpt.pop();
3567 JVMState *jvms = n->as_SafePoint()->jvms();
3568 assert(jvms != NULL, "sanity");
3569 int start = jvms->debug_start();
3570 int end = n->req();
3571 bool is_uncommon = (n->is_CallStaticJava() &&
3572 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3573 for (int j = start; j < end; j++) {
3574 Node* in = n->in(j);
3575 if (in->is_DecodeNarrowPtr()) {
3576 bool safe_to_skip = true;
3577 if (!is_uncommon ) {
3578 // Is it safe to skip?
3579 for (uint i = 0; i < in->outcnt(); i++) {
3580 Node* u = in->raw_out(i);
3581 if (!u->is_SafePoint() ||
3582 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3583 safe_to_skip = false;
3584 }
3585 }
3586 }
3587 if (safe_to_skip) {
3588 n->set_req(j, in->in(1));
3589 }
3590 if (in->outcnt() == 0) {
3591 in->disconnect_inputs(NULL, this);
3592 }
3593 }
3594 }
3595 }
3596 }
3597
3598 //------------------------------final_graph_reshaping--------------------------
3599 // Final Graph Reshaping.
3600 //
3601 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3602 // and not commoned up and forced early. Must come after regular
3603 // optimizations to avoid GVN undoing the cloning. Clone constant
3604 // inputs to Loop Phis; these will be split by the allocator anyways.
3605 // Remove Opaque nodes.
3606 // (2) Move last-uses by commutative operations to the left input to encourage
3607 // Intel update-in-place two-address operations and better register usage
3608 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3609 // calls canonicalizing them back.
3610 // (3) Count the number of double-precision FP ops, single-precision FP ops
3611 // and call sites. On Intel, we can get correct rounding either by
3612 // forcing singles to memory (requires extra stores and loads after each
3613 // FP bytecode) or we can set a rounding mode bit (requires setting and
3614 // clearing the mode bit around call sites). The mode bit is only used
3615 // if the relative frequency of single FP ops to calls is low enough.
3616 // This is a key transform for SPEC mpeg_audio.
3617 // (4) Detect infinite loops; blobs of code reachable from above but not
3618 // below. Several of the Code_Gen algorithms fail on such code shapes,
3619 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3620 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3621 // Detection is by looking for IfNodes where only 1 projection is
3622 // reachable from below or CatchNodes missing some targets.
3623 // (5) Assert for insane oop offsets in debug mode.
3624
final_graph_reshaping()3625 bool Compile::final_graph_reshaping() {
3626 // an infinite loop may have been eliminated by the optimizer,
3627 // in which case the graph will be empty.
3628 if (root()->req() == 1) {
3629 record_method_not_compilable("trivial infinite loop");
3630 return true;
3631 }
3632
3633 // Expensive nodes have their control input set to prevent the GVN
3634 // from freely commoning them. There's no GVN beyond this point so
3635 // no need to keep the control input. We want the expensive nodes to
3636 // be freely moved to the least frequent code path by gcm.
3637 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3638 for (int i = 0; i < expensive_count(); i++) {
3639 _expensive_nodes->at(i)->set_req(0, NULL);
3640 }
3641
3642 Final_Reshape_Counts frc;
3643
3644 // Visit everybody reachable!
3645 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3646 Node_Stack nstack(live_nodes() >> 1);
3647 final_graph_reshaping_walk(nstack, root(), frc);
3648
3649 // Check for unreachable (from below) code (i.e., infinite loops).
3650 for( uint i = 0; i < frc._tests.size(); i++ ) {
3651 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3652 // Get number of CFG targets.
3653 // Note that PCTables include exception targets after calls.
3654 uint required_outcnt = n->required_outcnt();
3655 if (n->outcnt() != required_outcnt) {
3656 // Check for a few special cases. Rethrow Nodes never take the
3657 // 'fall-thru' path, so expected kids is 1 less.
3658 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3659 if (n->in(0)->in(0)->is_Call()) {
3660 CallNode *call = n->in(0)->in(0)->as_Call();
3661 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3662 required_outcnt--; // Rethrow always has 1 less kid
3663 } else if (call->req() > TypeFunc::Parms &&
3664 call->is_CallDynamicJava()) {
3665 // Check for null receiver. In such case, the optimizer has
3666 // detected that the virtual call will always result in a null
3667 // pointer exception. The fall-through projection of this CatchNode
3668 // will not be populated.
3669 Node *arg0 = call->in(TypeFunc::Parms);
3670 if (arg0->is_Type() &&
3671 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3672 required_outcnt--;
3673 }
3674 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3675 call->req() > TypeFunc::Parms+1 &&
3676 call->is_CallStaticJava()) {
3677 // Check for negative array length. In such case, the optimizer has
3678 // detected that the allocation attempt will always result in an
3679 // exception. There is no fall-through projection of this CatchNode .
3680 Node *arg1 = call->in(TypeFunc::Parms+1);
3681 if (arg1->is_Type() &&
3682 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3683 required_outcnt--;
3684 }
3685 }
3686 }
3687 }
3688 // Recheck with a better notion of 'required_outcnt'
3689 if (n->outcnt() != required_outcnt) {
3690 record_method_not_compilable("malformed control flow");
3691 return true; // Not all targets reachable!
3692 }
3693 }
3694 // Check that I actually visited all kids. Unreached kids
3695 // must be infinite loops.
3696 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3697 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3698 record_method_not_compilable("infinite loop");
3699 return true; // Found unvisited kid; must be unreach
3700 }
3701
3702 // Here so verification code in final_graph_reshaping_walk()
3703 // always see an OuterStripMinedLoopEnd
3704 if (n->is_OuterStripMinedLoopEnd()) {
3705 IfNode* init_iff = n->as_If();
3706 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3707 n->subsume_by(iff, this);
3708 }
3709 }
3710
3711 // If original bytecodes contained a mixture of floats and doubles
3712 // check if the optimizer has made it homogenous, item (3).
3713 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3714 frc.get_float_count() > 32 &&
3715 frc.get_double_count() == 0 &&
3716 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3717 set_24_bit_selection_and_mode( false, true );
3718 }
3719
3720 set_java_calls(frc.get_java_call_count());
3721 set_inner_loops(frc.get_inner_loop_count());
3722
3723 // No infinite loops, no reason to bail out.
3724 return false;
3725 }
3726
3727 //-----------------------------too_many_traps----------------------------------
3728 // Report if there are too many traps at the current method and bci.
3729 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
too_many_traps(ciMethod * method,int bci,Deoptimization::DeoptReason reason)3730 bool Compile::too_many_traps(ciMethod* method,
3731 int bci,
3732 Deoptimization::DeoptReason reason) {
3733 ciMethodData* md = method->method_data();
3734 if (md->is_empty()) {
3735 // Assume the trap has not occurred, or that it occurred only
3736 // because of a transient condition during start-up in the interpreter.
3737 return false;
3738 }
3739 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3740 if (md->has_trap_at(bci, m, reason) != 0) {
3741 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3742 // Also, if there are multiple reasons, or if there is no per-BCI record,
3743 // assume the worst.
3744 if (log())
3745 log()->elem("observe trap='%s' count='%d'",
3746 Deoptimization::trap_reason_name(reason),
3747 md->trap_count(reason));
3748 return true;
3749 } else {
3750 // Ignore method/bci and see if there have been too many globally.
3751 return too_many_traps(reason, md);
3752 }
3753 }
3754
3755 // Less-accurate variant which does not require a method and bci.
too_many_traps(Deoptimization::DeoptReason reason,ciMethodData * logmd)3756 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3757 ciMethodData* logmd) {
3758 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3759 // Too many traps globally.
3760 // Note that we use cumulative trap_count, not just md->trap_count.
3761 if (log()) {
3762 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3763 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3764 Deoptimization::trap_reason_name(reason),
3765 mcount, trap_count(reason));
3766 }
3767 return true;
3768 } else {
3769 // The coast is clear.
3770 return false;
3771 }
3772 }
3773
3774 //--------------------------too_many_recompiles--------------------------------
3775 // Report if there are too many recompiles at the current method and bci.
3776 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3777 // Is not eager to return true, since this will cause the compiler to use
3778 // Action_none for a trap point, to avoid too many recompilations.
too_many_recompiles(ciMethod * method,int bci,Deoptimization::DeoptReason reason)3779 bool Compile::too_many_recompiles(ciMethod* method,
3780 int bci,
3781 Deoptimization::DeoptReason reason) {
3782 ciMethodData* md = method->method_data();
3783 if (md->is_empty()) {
3784 // Assume the trap has not occurred, or that it occurred only
3785 // because of a transient condition during start-up in the interpreter.
3786 return false;
3787 }
3788 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3789 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3790 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3791 Deoptimization::DeoptReason per_bc_reason
3792 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3793 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3794 if ((per_bc_reason == Deoptimization::Reason_none
3795 || md->has_trap_at(bci, m, reason) != 0)
3796 // The trap frequency measure we care about is the recompile count:
3797 && md->trap_recompiled_at(bci, m)
3798 && md->overflow_recompile_count() >= bc_cutoff) {
3799 // Do not emit a trap here if it has already caused recompilations.
3800 // Also, if there are multiple reasons, or if there is no per-BCI record,
3801 // assume the worst.
3802 if (log())
3803 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3804 Deoptimization::trap_reason_name(reason),
3805 md->trap_count(reason),
3806 md->overflow_recompile_count());
3807 return true;
3808 } else if (trap_count(reason) != 0
3809 && decompile_count() >= m_cutoff) {
3810 // Too many recompiles globally, and we have seen this sort of trap.
3811 // Use cumulative decompile_count, not just md->decompile_count.
3812 if (log())
3813 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3814 Deoptimization::trap_reason_name(reason),
3815 md->trap_count(reason), trap_count(reason),
3816 md->decompile_count(), decompile_count());
3817 return true;
3818 } else {
3819 // The coast is clear.
3820 return false;
3821 }
3822 }
3823
3824 // Compute when not to trap. Used by matching trap based nodes and
3825 // NullCheck optimization.
set_allowed_deopt_reasons()3826 void Compile::set_allowed_deopt_reasons() {
3827 _allowed_reasons = 0;
3828 if (is_method_compilation()) {
3829 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3830 assert(rs < BitsPerInt, "recode bit map");
3831 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3832 _allowed_reasons |= nth_bit(rs);
3833 }
3834 }
3835 }
3836 }
3837
needs_clinit_barrier(ciMethod * method,ciMethod * accessing_method)3838 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3839 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3840 }
3841
needs_clinit_barrier(ciField * field,ciMethod * accessing_method)3842 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3843 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3844 }
3845
needs_clinit_barrier(ciInstanceKlass * holder,ciMethod * accessing_method)3846 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3847 if (holder->is_initialized()) {
3848 return false;
3849 }
3850 if (holder->is_being_initialized()) {
3851 if (accessing_method->holder() == holder) {
3852 // Access inside a class. The barrier can be elided when access happens in <clinit>,
3853 // <init>, or a static method. In all those cases, there was an initialization
3854 // barrier on the holder klass passed.
3855 if (accessing_method->is_static_initializer() ||
3856 accessing_method->is_object_initializer() ||
3857 accessing_method->is_static()) {
3858 return false;
3859 }
3860 } else if (accessing_method->holder()->is_subclass_of(holder)) {
3861 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3862 // In case of <init> or a static method, the barrier is on the subclass is not enough:
3863 // child class can become fully initialized while its parent class is still being initialized.
3864 if (accessing_method->is_static_initializer()) {
3865 return false;
3866 }
3867 }
3868 ciMethod* root = method(); // the root method of compilation
3869 if (root != accessing_method) {
3870 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3871 }
3872 }
3873 return true;
3874 }
3875
3876 #ifndef PRODUCT
3877 //------------------------------verify_graph_edges---------------------------
3878 // Walk the Graph and verify that there is a one-to-one correspondence
3879 // between Use-Def edges and Def-Use edges in the graph.
verify_graph_edges(bool no_dead_code)3880 void Compile::verify_graph_edges(bool no_dead_code) {
3881 if (VerifyGraphEdges) {
3882 ResourceArea *area = Thread::current()->resource_area();
3883 Unique_Node_List visited(area);
3884 // Call recursive graph walk to check edges
3885 _root->verify_edges(visited);
3886 if (no_dead_code) {
3887 // Now make sure that no visited node is used by an unvisited node.
3888 bool dead_nodes = false;
3889 Unique_Node_List checked(area);
3890 while (visited.size() > 0) {
3891 Node* n = visited.pop();
3892 checked.push(n);
3893 for (uint i = 0; i < n->outcnt(); i++) {
3894 Node* use = n->raw_out(i);
3895 if (checked.member(use)) continue; // already checked
3896 if (visited.member(use)) continue; // already in the graph
3897 if (use->is_Con()) continue; // a dead ConNode is OK
3898 // At this point, we have found a dead node which is DU-reachable.
3899 if (!dead_nodes) {
3900 tty->print_cr("*** Dead nodes reachable via DU edges:");
3901 dead_nodes = true;
3902 }
3903 use->dump(2);
3904 tty->print_cr("---");
3905 checked.push(use); // No repeats; pretend it is now checked.
3906 }
3907 }
3908 assert(!dead_nodes, "using nodes must be reachable from root");
3909 }
3910 }
3911 }
3912 #endif
3913
3914 // The Compile object keeps track of failure reasons separately from the ciEnv.
3915 // This is required because there is not quite a 1-1 relation between the
3916 // ciEnv and its compilation task and the Compile object. Note that one
3917 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3918 // to backtrack and retry without subsuming loads. Other than this backtracking
3919 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3920 // by the logic in C2Compiler.
record_failure(const char * reason)3921 void Compile::record_failure(const char* reason) {
3922 if (log() != NULL) {
3923 log()->elem("failure reason='%s' phase='compile'", reason);
3924 }
3925 if (_failure_reason == NULL) {
3926 // Record the first failure reason.
3927 _failure_reason = reason;
3928 }
3929
3930 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3931 C->print_method(PHASE_FAILURE);
3932 }
3933 _root = NULL; // flush the graph, too
3934 }
3935
TracePhase(const char * name,elapsedTimer * accumulator)3936 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3937 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3938 _phase_name(name), _dolog(CITimeVerbose)
3939 {
3940 if (_dolog) {
3941 C = Compile::current();
3942 _log = C->log();
3943 } else {
3944 C = NULL;
3945 _log = NULL;
3946 }
3947 if (_log != NULL) {
3948 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3949 _log->stamp();
3950 _log->end_head();
3951 }
3952 }
3953
~TracePhase()3954 Compile::TracePhase::~TracePhase() {
3955
3956 C = Compile::current();
3957 if (_dolog) {
3958 _log = C->log();
3959 } else {
3960 _log = NULL;
3961 }
3962
3963 #ifdef ASSERT
3964 if (PrintIdealNodeCount) {
3965 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3966 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3967 }
3968
3969 if (VerifyIdealNodeCount) {
3970 Compile::current()->print_missing_nodes();
3971 }
3972 #endif
3973
3974 if (_log != NULL) {
3975 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3976 }
3977 }
3978
3979 //=============================================================================
3980 // Two Constant's are equal when the type and the value are equal.
operator ==(const Constant & other)3981 bool Compile::Constant::operator==(const Constant& other) {
3982 if (type() != other.type() ) return false;
3983 if (can_be_reused() != other.can_be_reused()) return false;
3984 // For floating point values we compare the bit pattern.
3985 switch (type()) {
3986 case T_INT:
3987 case T_FLOAT: return (_v._value.i == other._v._value.i);
3988 case T_LONG:
3989 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3990 case T_OBJECT:
3991 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3992 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3993 case T_METADATA: return (_v._metadata == other._v._metadata);
3994 default: ShouldNotReachHere(); return false;
3995 }
3996 }
3997
type_to_size_in_bytes(BasicType t)3998 static int type_to_size_in_bytes(BasicType t) {
3999 switch (t) {
4000 case T_INT: return sizeof(jint );
4001 case T_LONG: return sizeof(jlong );
4002 case T_FLOAT: return sizeof(jfloat );
4003 case T_DOUBLE: return sizeof(jdouble);
4004 case T_METADATA: return sizeof(Metadata*);
4005 // We use T_VOID as marker for jump-table entries (labels) which
4006 // need an internal word relocation.
4007 case T_VOID:
4008 case T_ADDRESS:
4009 case T_OBJECT: return sizeof(jobject);
4010 default:
4011 ShouldNotReachHere();
4012 return -1;
4013 }
4014 }
4015
qsort_comparator(Constant * a,Constant * b)4016 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4017 // sort descending
4018 if (a->freq() > b->freq()) return -1;
4019 if (a->freq() < b->freq()) return 1;
4020 return 0;
4021 }
4022
calculate_offsets_and_size()4023 void Compile::ConstantTable::calculate_offsets_and_size() {
4024 // First, sort the array by frequencies.
4025 _constants.sort(qsort_comparator);
4026
4027 #ifdef ASSERT
4028 // Make sure all jump-table entries were sorted to the end of the
4029 // array (they have a negative frequency).
4030 bool found_void = false;
4031 for (int i = 0; i < _constants.length(); i++) {
4032 Constant con = _constants.at(i);
4033 if (con.type() == T_VOID)
4034 found_void = true; // jump-tables
4035 else
4036 assert(!found_void, "wrong sorting");
4037 }
4038 #endif
4039
4040 int offset = 0;
4041 for (int i = 0; i < _constants.length(); i++) {
4042 Constant* con = _constants.adr_at(i);
4043
4044 // Align offset for type.
4045 int typesize = type_to_size_in_bytes(con->type());
4046 offset = align_up(offset, typesize);
4047 con->set_offset(offset); // set constant's offset
4048
4049 if (con->type() == T_VOID) {
4050 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4051 offset = offset + typesize * n->outcnt(); // expand jump-table
4052 } else {
4053 offset = offset + typesize;
4054 }
4055 }
4056
4057 // Align size up to the next section start (which is insts; see
4058 // CodeBuffer::align_at_start).
4059 assert(_size == -1, "already set?");
4060 _size = align_up(offset, (int)CodeEntryAlignment);
4061 }
4062
emit(CodeBuffer & cb)4063 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4064 MacroAssembler _masm(&cb);
4065 for (int i = 0; i < _constants.length(); i++) {
4066 Constant con = _constants.at(i);
4067 address constant_addr = NULL;
4068 switch (con.type()) {
4069 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4070 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4071 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4072 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4073 case T_OBJECT: {
4074 jobject obj = con.get_jobject();
4075 int oop_index = _masm.oop_recorder()->find_index(obj);
4076 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4077 break;
4078 }
4079 case T_ADDRESS: {
4080 address addr = (address) con.get_jobject();
4081 constant_addr = _masm.address_constant(addr);
4082 break;
4083 }
4084 // We use T_VOID as marker for jump-table entries (labels) which
4085 // need an internal word relocation.
4086 case T_VOID: {
4087 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4088 // Fill the jump-table with a dummy word. The real value is
4089 // filled in later in fill_jump_table.
4090 address dummy = (address) n;
4091 constant_addr = _masm.address_constant(dummy);
4092 // Expand jump-table
4093 for (uint i = 1; i < n->outcnt(); i++) {
4094 address temp_addr = _masm.address_constant(dummy + i);
4095 assert(temp_addr, "consts section too small");
4096 }
4097 break;
4098 }
4099 case T_METADATA: {
4100 Metadata* obj = con.get_metadata();
4101 int metadata_index = _masm.oop_recorder()->find_index(obj);
4102 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4103 break;
4104 }
4105 default: ShouldNotReachHere();
4106 }
4107 assert(constant_addr, "consts section too small");
4108 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4109 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4110 }
4111 }
4112
find_offset(Constant & con) const4113 int Compile::ConstantTable::find_offset(Constant& con) const {
4114 int idx = _constants.find(con);
4115 guarantee(idx != -1, "constant must be in constant table");
4116 int offset = _constants.at(idx).offset();
4117 guarantee(offset != -1, "constant table not emitted yet?");
4118 return offset;
4119 }
4120
add(Constant & con)4121 void Compile::ConstantTable::add(Constant& con) {
4122 if (con.can_be_reused()) {
4123 int idx = _constants.find(con);
4124 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4125 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4126 return;
4127 }
4128 }
4129 (void) _constants.append(con);
4130 }
4131
add(MachConstantNode * n,BasicType type,jvalue value)4132 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4133 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4134 Constant con(type, value, b->_freq);
4135 add(con);
4136 return con;
4137 }
4138
add(Metadata * metadata)4139 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4140 Constant con(metadata);
4141 add(con);
4142 return con;
4143 }
4144
add(MachConstantNode * n,MachOper * oper)4145 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4146 jvalue value;
4147 BasicType type = oper->type()->basic_type();
4148 switch (type) {
4149 case T_LONG: value.j = oper->constantL(); break;
4150 case T_FLOAT: value.f = oper->constantF(); break;
4151 case T_DOUBLE: value.d = oper->constantD(); break;
4152 case T_OBJECT:
4153 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4154 case T_METADATA: return add((Metadata*)oper->constant()); break;
4155 default: guarantee(false, "unhandled type: %s", type2name(type));
4156 }
4157 return add(n, type, value);
4158 }
4159
add_jump_table(MachConstantNode * n)4160 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4161 jvalue value;
4162 // We can use the node pointer here to identify the right jump-table
4163 // as this method is called from Compile::Fill_buffer right before
4164 // the MachNodes are emitted and the jump-table is filled (means the
4165 // MachNode pointers do not change anymore).
4166 value.l = (jobject) n;
4167 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4168 add(con);
4169 return con;
4170 }
4171
fill_jump_table(CodeBuffer & cb,MachConstantNode * n,GrowableArray<Label * > labels) const4172 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4173 // If called from Compile::scratch_emit_size do nothing.
4174 if (Compile::current()->in_scratch_emit_size()) return;
4175
4176 assert(labels.is_nonempty(), "must be");
4177 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4178
4179 // Since MachConstantNode::constant_offset() also contains
4180 // table_base_offset() we need to subtract the table_base_offset()
4181 // to get the plain offset into the constant table.
4182 int offset = n->constant_offset() - table_base_offset();
4183
4184 MacroAssembler _masm(&cb);
4185 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4186
4187 for (uint i = 0; i < n->outcnt(); i++) {
4188 address* constant_addr = &jump_table_base[i];
4189 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));
4190 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4191 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4192 }
4193 }
4194
4195 //----------------------------static_subtype_check-----------------------------
4196 // Shortcut important common cases when superklass is exact:
4197 // (0) superklass is java.lang.Object (can occur in reflective code)
4198 // (1) subklass is already limited to a subtype of superklass => always ok
4199 // (2) subklass does not overlap with superklass => always fail
4200 // (3) superklass has NO subtypes and we can check with a simple compare.
static_subtype_check(ciKlass * superk,ciKlass * subk)4201 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4202 if (StressReflectiveCode) {
4203 return SSC_full_test; // Let caller generate the general case.
4204 }
4205
4206 if (superk == env()->Object_klass()) {
4207 return SSC_always_true; // (0) this test cannot fail
4208 }
4209
4210 ciType* superelem = superk;
4211 if (superelem->is_array_klass())
4212 superelem = superelem->as_array_klass()->base_element_type();
4213
4214 if (!subk->is_interface()) { // cannot trust static interface types yet
4215 if (subk->is_subtype_of(superk)) {
4216 return SSC_always_true; // (1) false path dead; no dynamic test needed
4217 }
4218 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4219 !superk->is_subtype_of(subk)) {
4220 return SSC_always_false;
4221 }
4222 }
4223
4224 // If casting to an instance klass, it must have no subtypes
4225 if (superk->is_interface()) {
4226 // Cannot trust interfaces yet.
4227 // %%% S.B. superk->nof_implementors() == 1
4228 } else if (superelem->is_instance_klass()) {
4229 ciInstanceKlass* ik = superelem->as_instance_klass();
4230 if (!ik->has_subklass() && !ik->is_interface()) {
4231 if (!ik->is_final()) {
4232 // Add a dependency if there is a chance of a later subclass.
4233 dependencies()->assert_leaf_type(ik);
4234 }
4235 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4236 }
4237 } else {
4238 // A primitive array type has no subtypes.
4239 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4240 }
4241
4242 return SSC_full_test;
4243 }
4244
conv_I2X_index(PhaseGVN * phase,Node * idx,const TypeInt * sizetype,Node * ctrl)4245 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4246 #ifdef _LP64
4247 // The scaled index operand to AddP must be a clean 64-bit value.
4248 // Java allows a 32-bit int to be incremented to a negative
4249 // value, which appears in a 64-bit register as a large
4250 // positive number. Using that large positive number as an
4251 // operand in pointer arithmetic has bad consequences.
4252 // On the other hand, 32-bit overflow is rare, and the possibility
4253 // can often be excluded, if we annotate the ConvI2L node with
4254 // a type assertion that its value is known to be a small positive
4255 // number. (The prior range check has ensured this.)
4256 // This assertion is used by ConvI2LNode::Ideal.
4257 int index_max = max_jint - 1; // array size is max_jint, index is one less
4258 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4259 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4260 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4261 #endif
4262 return idx;
4263 }
4264
4265 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
constrained_convI2L(PhaseGVN * phase,Node * value,const TypeInt * itype,Node * ctrl)4266 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4267 if (ctrl != NULL) {
4268 // Express control dependency by a CastII node with a narrow type.
4269 value = new CastIINode(value, itype, false, true /* range check dependency */);
4270 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4271 // node from floating above the range check during loop optimizations. Otherwise, the
4272 // ConvI2L node may be eliminated independently of the range check, causing the data path
4273 // to become TOP while the control path is still there (although it's unreachable).
4274 value->set_req(0, ctrl);
4275 // Save CastII node to remove it after loop optimizations.
4276 phase->C->add_range_check_cast(value);
4277 value = phase->transform(value);
4278 }
4279 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4280 return phase->transform(new ConvI2LNode(value, ltype));
4281 }
4282
print_inlining_stream_free()4283 void Compile::print_inlining_stream_free() {
4284 if (_print_inlining_stream != NULL) {
4285 _print_inlining_stream->~stringStream();
4286 _print_inlining_stream = NULL;
4287 }
4288 }
4289
4290 // The message about the current inlining is accumulated in
4291 // _print_inlining_stream and transfered into the _print_inlining_list
4292 // once we know whether inlining succeeds or not. For regular
4293 // inlining, messages are appended to the buffer pointed by
4294 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4295 // a new buffer is added after _print_inlining_idx in the list. This
4296 // way we can update the inlining message for late inlining call site
4297 // when the inlining is attempted again.
print_inlining_init()4298 void Compile::print_inlining_init() {
4299 if (print_inlining() || print_intrinsics()) {
4300 // print_inlining_init is actually called several times.
4301 print_inlining_stream_free();
4302 _print_inlining_stream = new stringStream();
4303 // Watch out: The memory initialized by the constructor call PrintInliningBuffer()
4304 // will be copied into the only initial element. The default destructor of
4305 // PrintInliningBuffer will be called when leaving the scope here. If it
4306 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss
4307 // would be destructed, too!
4308 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4309 }
4310 }
4311
print_inlining_reinit()4312 void Compile::print_inlining_reinit() {
4313 if (print_inlining() || print_intrinsics()) {
4314 print_inlining_stream_free();
4315 // Re allocate buffer when we change ResourceMark
4316 _print_inlining_stream = new stringStream();
4317 }
4318 }
4319
print_inlining_reset()4320 void Compile::print_inlining_reset() {
4321 _print_inlining_stream->reset();
4322 }
4323
print_inlining_commit()4324 void Compile::print_inlining_commit() {
4325 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4326 // Transfer the message from _print_inlining_stream to the current
4327 // _print_inlining_list buffer and clear _print_inlining_stream.
4328 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4329 print_inlining_reset();
4330 }
4331
print_inlining_push()4332 void Compile::print_inlining_push() {
4333 // Add new buffer to the _print_inlining_list at current position
4334 _print_inlining_idx++;
4335 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4336 }
4337
print_inlining_current()4338 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4339 return _print_inlining_list->at(_print_inlining_idx);
4340 }
4341
print_inlining_update(CallGenerator * cg)4342 void Compile::print_inlining_update(CallGenerator* cg) {
4343 if (print_inlining() || print_intrinsics()) {
4344 if (!cg->is_late_inline()) {
4345 if (print_inlining_current().cg() != NULL) {
4346 print_inlining_push();
4347 }
4348 print_inlining_commit();
4349 } else {
4350 if (print_inlining_current().cg() != cg &&
4351 (print_inlining_current().cg() != NULL ||
4352 print_inlining_current().ss()->size() != 0)) {
4353 print_inlining_push();
4354 }
4355 print_inlining_commit();
4356 print_inlining_current().set_cg(cg);
4357 }
4358 }
4359 }
4360
print_inlining_move_to(CallGenerator * cg)4361 void Compile::print_inlining_move_to(CallGenerator* cg) {
4362 // We resume inlining at a late inlining call site. Locate the
4363 // corresponding inlining buffer so that we can update it.
4364 if (print_inlining()) {
4365 for (int i = 0; i < _print_inlining_list->length(); i++) {
4366 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4367 _print_inlining_idx = i;
4368 return;
4369 }
4370 }
4371 ShouldNotReachHere();
4372 }
4373 }
4374
print_inlining_update_delayed(CallGenerator * cg)4375 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4376 if (print_inlining()) {
4377 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4378 assert(print_inlining_current().cg() == cg, "wrong entry");
4379 // replace message with new message
4380 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4381 print_inlining_commit();
4382 print_inlining_current().set_cg(cg);
4383 }
4384 }
4385
print_inlining_assert_ready()4386 void Compile::print_inlining_assert_ready() {
4387 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4388 }
4389
process_print_inlining()4390 void Compile::process_print_inlining() {
4391 bool do_print_inlining = print_inlining() || print_intrinsics();
4392 if (do_print_inlining || log() != NULL) {
4393 // Print inlining message for candidates that we couldn't inline
4394 // for lack of space
4395 for (int i = 0; i < _late_inlines.length(); i++) {
4396 CallGenerator* cg = _late_inlines.at(i);
4397 if (!cg->is_mh_late_inline()) {
4398 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4399 if (do_print_inlining) {
4400 cg->print_inlining_late(msg);
4401 }
4402 log_late_inline_failure(cg, msg);
4403 }
4404 }
4405 }
4406 if (do_print_inlining) {
4407 ResourceMark rm;
4408 stringStream ss;
4409 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4410 for (int i = 0; i < _print_inlining_list->length(); i++) {
4411 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4412 _print_inlining_list->at(i).freeStream();
4413 }
4414 // Reset _print_inlining_list, it only contains destructed objects.
4415 // It is on the arena, so it will be freed when the arena is reset.
4416 _print_inlining_list = NULL;
4417 // _print_inlining_stream won't be used anymore, either.
4418 print_inlining_stream_free();
4419 size_t end = ss.size();
4420 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4421 strncpy(_print_inlining_output, ss.base(), end+1);
4422 _print_inlining_output[end] = 0;
4423 }
4424 }
4425
dump_print_inlining()4426 void Compile::dump_print_inlining() {
4427 if (_print_inlining_output != NULL) {
4428 tty->print_raw(_print_inlining_output);
4429 }
4430 }
4431
log_late_inline(CallGenerator * cg)4432 void Compile::log_late_inline(CallGenerator* cg) {
4433 if (log() != NULL) {
4434 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4435 cg->unique_id());
4436 JVMState* p = cg->call_node()->jvms();
4437 while (p != NULL) {
4438 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4439 p = p->caller();
4440 }
4441 log()->tail("late_inline");
4442 }
4443 }
4444
log_late_inline_failure(CallGenerator * cg,const char * msg)4445 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4446 log_late_inline(cg);
4447 if (log() != NULL) {
4448 log()->inline_fail(msg);
4449 }
4450 }
4451
log_inline_id(CallGenerator * cg)4452 void Compile::log_inline_id(CallGenerator* cg) {
4453 if (log() != NULL) {
4454 // The LogCompilation tool needs a unique way to identify late
4455 // inline call sites. This id must be unique for this call site in
4456 // this compilation. Try to have it unique across compilations as
4457 // well because it can be convenient when grepping through the log
4458 // file.
4459 // Distinguish OSR compilations from others in case CICountOSR is
4460 // on.
4461 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4462 cg->set_unique_id(id);
4463 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4464 }
4465 }
4466
log_inline_failure(const char * msg)4467 void Compile::log_inline_failure(const char* msg) {
4468 if (C->log() != NULL) {
4469 C->log()->inline_fail(msg);
4470 }
4471 }
4472
4473
4474 // Dump inlining replay data to the stream.
4475 // Don't change thread state and acquire any locks.
dump_inline_data(outputStream * out)4476 void Compile::dump_inline_data(outputStream* out) {
4477 InlineTree* inl_tree = ilt();
4478 if (inl_tree != NULL) {
4479 out->print(" inline %d", inl_tree->count());
4480 inl_tree->dump_replay_data(out);
4481 }
4482 }
4483
cmp_expensive_nodes(Node * n1,Node * n2)4484 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4485 if (n1->Opcode() < n2->Opcode()) return -1;
4486 else if (n1->Opcode() > n2->Opcode()) return 1;
4487
4488 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4489 for (uint i = 1; i < n1->req(); i++) {
4490 if (n1->in(i) < n2->in(i)) return -1;
4491 else if (n1->in(i) > n2->in(i)) return 1;
4492 }
4493
4494 return 0;
4495 }
4496
cmp_expensive_nodes(Node ** n1p,Node ** n2p)4497 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4498 Node* n1 = *n1p;
4499 Node* n2 = *n2p;
4500
4501 return cmp_expensive_nodes(n1, n2);
4502 }
4503
sort_expensive_nodes()4504 void Compile::sort_expensive_nodes() {
4505 if (!expensive_nodes_sorted()) {
4506 _expensive_nodes->sort(cmp_expensive_nodes);
4507 }
4508 }
4509
expensive_nodes_sorted() const4510 bool Compile::expensive_nodes_sorted() const {
4511 for (int i = 1; i < _expensive_nodes->length(); i++) {
4512 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4513 return false;
4514 }
4515 }
4516 return true;
4517 }
4518
should_optimize_expensive_nodes(PhaseIterGVN & igvn)4519 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4520 if (_expensive_nodes->length() == 0) {
4521 return false;
4522 }
4523
4524 assert(OptimizeExpensiveOps, "optimization off?");
4525
4526 // Take this opportunity to remove dead nodes from the list
4527 int j = 0;
4528 for (int i = 0; i < _expensive_nodes->length(); i++) {
4529 Node* n = _expensive_nodes->at(i);
4530 if (!n->is_unreachable(igvn)) {
4531 assert(n->is_expensive(), "should be expensive");
4532 _expensive_nodes->at_put(j, n);
4533 j++;
4534 }
4535 }
4536 _expensive_nodes->trunc_to(j);
4537
4538 // Then sort the list so that similar nodes are next to each other
4539 // and check for at least two nodes of identical kind with same data
4540 // inputs.
4541 sort_expensive_nodes();
4542
4543 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4544 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4545 return true;
4546 }
4547 }
4548
4549 return false;
4550 }
4551
cleanup_expensive_nodes(PhaseIterGVN & igvn)4552 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4553 if (_expensive_nodes->length() == 0) {
4554 return;
4555 }
4556
4557 assert(OptimizeExpensiveOps, "optimization off?");
4558
4559 // Sort to bring similar nodes next to each other and clear the
4560 // control input of nodes for which there's only a single copy.
4561 sort_expensive_nodes();
4562
4563 int j = 0;
4564 int identical = 0;
4565 int i = 0;
4566 bool modified = false;
4567 for (; i < _expensive_nodes->length()-1; i++) {
4568 assert(j <= i, "can't write beyond current index");
4569 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4570 identical++;
4571 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4572 continue;
4573 }
4574 if (identical > 0) {
4575 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4576 identical = 0;
4577 } else {
4578 Node* n = _expensive_nodes->at(i);
4579 igvn.replace_input_of(n, 0, NULL);
4580 igvn.hash_insert(n);
4581 modified = true;
4582 }
4583 }
4584 if (identical > 0) {
4585 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4586 } else if (_expensive_nodes->length() >= 1) {
4587 Node* n = _expensive_nodes->at(i);
4588 igvn.replace_input_of(n, 0, NULL);
4589 igvn.hash_insert(n);
4590 modified = true;
4591 }
4592 _expensive_nodes->trunc_to(j);
4593 if (modified) {
4594 igvn.optimize();
4595 }
4596 }
4597
add_expensive_node(Node * n)4598 void Compile::add_expensive_node(Node * n) {
4599 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4600 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4601 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4602 if (OptimizeExpensiveOps) {
4603 _expensive_nodes->append(n);
4604 } else {
4605 // Clear control input and let IGVN optimize expensive nodes if
4606 // OptimizeExpensiveOps is off.
4607 n->set_req(0, NULL);
4608 }
4609 }
4610
4611 /**
4612 * Remove the speculative part of types and clean up the graph
4613 */
remove_speculative_types(PhaseIterGVN & igvn)4614 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4615 if (UseTypeSpeculation) {
4616 Unique_Node_List worklist;
4617 worklist.push(root());
4618 int modified = 0;
4619 // Go over all type nodes that carry a speculative type, drop the
4620 // speculative part of the type and enqueue the node for an igvn
4621 // which may optimize it out.
4622 for (uint next = 0; next < worklist.size(); ++next) {
4623 Node *n = worklist.at(next);
4624 if (n->is_Type()) {
4625 TypeNode* tn = n->as_Type();
4626 const Type* t = tn->type();
4627 const Type* t_no_spec = t->remove_speculative();
4628 if (t_no_spec != t) {
4629 bool in_hash = igvn.hash_delete(n);
4630 assert(in_hash, "node should be in igvn hash table");
4631 tn->set_type(t_no_spec);
4632 igvn.hash_insert(n);
4633 igvn._worklist.push(n); // give it a chance to go away
4634 modified++;
4635 }
4636 }
4637 uint max = n->len();
4638 for( uint i = 0; i < max; ++i ) {
4639 Node *m = n->in(i);
4640 if (not_a_node(m)) continue;
4641 worklist.push(m);
4642 }
4643 }
4644 // Drop the speculative part of all types in the igvn's type table
4645 igvn.remove_speculative_types();
4646 if (modified > 0) {
4647 igvn.optimize();
4648 }
4649 #ifdef ASSERT
4650 // Verify that after the IGVN is over no speculative type has resurfaced
4651 worklist.clear();
4652 worklist.push(root());
4653 for (uint next = 0; next < worklist.size(); ++next) {
4654 Node *n = worklist.at(next);
4655 const Type* t = igvn.type_or_null(n);
4656 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4657 if (n->is_Type()) {
4658 t = n->as_Type()->type();
4659 assert(t == t->remove_speculative(), "no more speculative types");
4660 }
4661 uint max = n->len();
4662 for( uint i = 0; i < max; ++i ) {
4663 Node *m = n->in(i);
4664 if (not_a_node(m)) continue;
4665 worklist.push(m);
4666 }
4667 }
4668 igvn.check_no_speculative_types();
4669 #endif
4670 }
4671 }
4672
4673 // Auxiliary method to support randomized stressing/fuzzing.
4674 //
4675 // This method can be called the arbitrary number of times, with current count
4676 // as the argument. The logic allows selecting a single candidate from the
4677 // running list of candidates as follows:
4678 // int count = 0;
4679 // Cand* selected = null;
4680 // while(cand = cand->next()) {
4681 // if (randomized_select(++count)) {
4682 // selected = cand;
4683 // }
4684 // }
4685 //
4686 // Including count equalizes the chances any candidate is "selected".
4687 // This is useful when we don't have the complete list of candidates to choose
4688 // from uniformly. In this case, we need to adjust the randomicity of the
4689 // selection, or else we will end up biasing the selection towards the latter
4690 // candidates.
4691 //
4692 // Quick back-envelope calculation shows that for the list of n candidates
4693 // the equal probability for the candidate to persist as "best" can be
4694 // achieved by replacing it with "next" k-th candidate with the probability
4695 // of 1/k. It can be easily shown that by the end of the run, the
4696 // probability for any candidate is converged to 1/n, thus giving the
4697 // uniform distribution among all the candidates.
4698 //
4699 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4700 #define RANDOMIZED_DOMAIN_POW 29
4701 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4702 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
randomized_select(int count)4703 bool Compile::randomized_select(int count) {
4704 assert(count > 0, "only positive");
4705 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4706 }
4707
clone_map()4708 CloneMap& Compile::clone_map() { return _clone_map; }
set_clone_map(Dict * d)4709 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4710
dump() const4711 void NodeCloneInfo::dump() const {
4712 tty->print(" {%d:%d} ", idx(), gen());
4713 }
4714
clone(Node * old,Node * nnn,int gen)4715 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4716 uint64_t val = value(old->_idx);
4717 NodeCloneInfo cio(val);
4718 assert(val != 0, "old node should be in the map");
4719 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4720 insert(nnn->_idx, cin.get());
4721 #ifndef PRODUCT
4722 if (is_debug()) {
4723 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4724 }
4725 #endif
4726 }
4727
verify_insert_and_clone(Node * old,Node * nnn,int gen)4728 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4729 NodeCloneInfo cio(value(old->_idx));
4730 if (cio.get() == 0) {
4731 cio.set(old->_idx, 0);
4732 insert(old->_idx, cio.get());
4733 #ifndef PRODUCT
4734 if (is_debug()) {
4735 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4736 }
4737 #endif
4738 }
4739 clone(old, nnn, gen);
4740 }
4741
max_gen() const4742 int CloneMap::max_gen() const {
4743 int g = 0;
4744 DictI di(_dict);
4745 for(; di.test(); ++di) {
4746 int t = gen(di._key);
4747 if (g < t) {
4748 g = t;
4749 #ifndef PRODUCT
4750 if (is_debug()) {
4751 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4752 }
4753 #endif
4754 }
4755 }
4756 return g;
4757 }
4758
dump(node_idx_t key) const4759 void CloneMap::dump(node_idx_t key) const {
4760 uint64_t val = value(key);
4761 if (val != 0) {
4762 NodeCloneInfo ni(val);
4763 ni.dump();
4764 }
4765 }
4766