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_hwm(NULL);
466 compile->set_type_last_size(0);
467 compile->set_last_tf(NULL, NULL);
468 compile->set_indexSet_arena(NULL);
469 compile->set_indexSet_free_block_list(NULL);
470 compile->init_type_arena();
471 Type::Initialize(compile);
472 _compile->set_scratch_buffer_blob(NULL);
473 _compile->begin_method();
474 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
475 }
~CompileWrapper()476 CompileWrapper::~CompileWrapper() {
477 _compile->end_method();
478 if (_compile->scratch_buffer_blob() != NULL)
479 BufferBlob::free(_compile->scratch_buffer_blob());
480 _compile->env()->set_compiler_data(NULL);
481 }
482
483
484 //----------------------------print_compile_messages---------------------------
print_compile_messages()485 void Compile::print_compile_messages() {
486 #ifndef PRODUCT
487 // Check if recompiling
488 if (_subsume_loads == false && PrintOpto) {
489 // Recompiling without allowing machine instructions to subsume loads
490 tty->print_cr("*********************************************************");
491 tty->print_cr("** Bailout: Recompile without subsuming loads **");
492 tty->print_cr("*********************************************************");
493 }
494 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
495 // Recompiling without escape analysis
496 tty->print_cr("*********************************************************");
497 tty->print_cr("** Bailout: Recompile without escape analysis **");
498 tty->print_cr("*********************************************************");
499 }
500 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
501 // Recompiling without boxing elimination
502 tty->print_cr("*********************************************************");
503 tty->print_cr("** Bailout: Recompile without boxing elimination **");
504 tty->print_cr("*********************************************************");
505 }
506 if (C->directive()->BreakAtCompileOption) {
507 // Open the debugger when compiling this method.
508 tty->print("### Breaking when compiling: ");
509 method()->print_short_name();
510 tty->cr();
511 BREAKPOINT;
512 }
513
514 if( PrintOpto ) {
515 if (is_osr_compilation()) {
516 tty->print("[OSR]%3d", _compile_id);
517 } else {
518 tty->print("%3d", _compile_id);
519 }
520 }
521 #endif
522 }
523
524
525 //-----------------------init_scratch_buffer_blob------------------------------
526 // Construct a temporary BufferBlob and cache it for this compile.
init_scratch_buffer_blob(int const_size)527 void Compile::init_scratch_buffer_blob(int const_size) {
528 // If there is already a scratch buffer blob allocated and the
529 // constant section is big enough, use it. Otherwise free the
530 // current and allocate a new one.
531 BufferBlob* blob = scratch_buffer_blob();
532 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
533 // Use the current blob.
534 } else {
535 if (blob != NULL) {
536 BufferBlob::free(blob);
537 }
538
539 ResourceMark rm;
540 _scratch_const_size = const_size;
541 int size = C2Compiler::initial_code_buffer_size(const_size);
542 blob = BufferBlob::create("Compile::scratch_buffer", size);
543 // Record the buffer blob for next time.
544 set_scratch_buffer_blob(blob);
545 // Have we run out of code space?
546 if (scratch_buffer_blob() == NULL) {
547 // Let CompilerBroker disable further compilations.
548 record_failure("Not enough space for scratch buffer in CodeCache");
549 return;
550 }
551 }
552
553 // Initialize the relocation buffers
554 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
555 set_scratch_locs_memory(locs_buf);
556 }
557
558
559 //-----------------------scratch_emit_size-------------------------------------
560 // Helper function that computes size by emitting code
scratch_emit_size(const Node * n)561 uint Compile::scratch_emit_size(const Node* n) {
562 // Start scratch_emit_size section.
563 set_in_scratch_emit_size(true);
564
565 // Emit into a trash buffer and count bytes emitted.
566 // This is a pretty expensive way to compute a size,
567 // but it works well enough if seldom used.
568 // All common fixed-size instructions are given a size
569 // method by the AD file.
570 // Note that the scratch buffer blob and locs memory are
571 // allocated at the beginning of the compile task, and
572 // may be shared by several calls to scratch_emit_size.
573 // The allocation of the scratch buffer blob is particularly
574 // expensive, since it has to grab the code cache lock.
575 BufferBlob* blob = this->scratch_buffer_blob();
576 assert(blob != NULL, "Initialize BufferBlob at start");
577 assert(blob->size() > MAX_inst_size, "sanity");
578 relocInfo* locs_buf = scratch_locs_memory();
579 address blob_begin = blob->content_begin();
580 address blob_end = (address)locs_buf;
581 assert(blob->contains(blob_end), "sanity");
582 CodeBuffer buf(blob_begin, blob_end - blob_begin);
583 buf.initialize_consts_size(_scratch_const_size);
584 buf.initialize_stubs_size(MAX_stubs_size);
585 assert(locs_buf != NULL, "sanity");
586 int lsize = MAX_locs_size / 3;
587 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
588 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
589 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
590 // Mark as scratch buffer.
591 buf.consts()->set_scratch_emit();
592 buf.insts()->set_scratch_emit();
593 buf.stubs()->set_scratch_emit();
594
595 // Do the emission.
596
597 Label fakeL; // Fake label for branch instructions.
598 Label* saveL = NULL;
599 uint save_bnum = 0;
600 bool is_branch = n->is_MachBranch();
601 if (is_branch) {
602 MacroAssembler masm(&buf);
603 masm.bind(fakeL);
604 n->as_MachBranch()->save_label(&saveL, &save_bnum);
605 n->as_MachBranch()->label_set(&fakeL, 0);
606 }
607 n->emit(buf, this->regalloc());
608
609 // Emitting into the scratch buffer should not fail
610 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
611
612 if (is_branch) // Restore label.
613 n->as_MachBranch()->label_set(saveL, save_bnum);
614
615 // End scratch_emit_size section.
616 set_in_scratch_emit_size(false);
617
618 return buf.insts_size();
619 }
620
621
622 // ============================================================================
623 //------------------------------Compile standard-------------------------------
debug_only(int Compile::_debug_idx=100000;)624 debug_only( int Compile::_debug_idx = 100000; )
625
626 // Compile a method. entry_bci is -1 for normal compilations and indicates
627 // the continuation bci for on stack replacement.
628
629
630 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
631 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
632 : Phase(Compiler),
633 _compile_id(ci_env->compile_id()),
634 _save_argument_registers(false),
635 _subsume_loads(subsume_loads),
636 _do_escape_analysis(do_escape_analysis),
637 _eliminate_boxing(eliminate_boxing),
638 _method(target),
639 _entry_bci(osr_bci),
640 _stub_function(NULL),
641 _stub_name(NULL),
642 _stub_entry_point(NULL),
643 _max_node_limit(MaxNodeLimit),
644 _orig_pc_slot(0),
645 _orig_pc_slot_offset_in_bytes(0),
646 _inlining_progress(false),
647 _inlining_incrementally(false),
648 _do_cleanup(false),
649 _has_reserved_stack_access(target->has_reserved_stack_access()),
650 #ifndef PRODUCT
651 _trace_opto_output(directive->TraceOptoOutputOption),
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 (PrintIdeal) {
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 #endif
988 _has_method_handle_invokes(false),
989 _clinit_barrier_on_entry(false),
990 _comp_arena(mtCompiler),
991 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
992 _env(ci_env),
993 _directive(directive),
994 _log(ci_env->log()),
995 _failure_reason(NULL),
996 _congraph(NULL),
997 #ifndef PRODUCT
998 _printer(NULL),
999 #endif
1000 _dead_node_list(comp_arena()),
1001 _dead_node_count(0),
1002 _node_arena(mtCompiler),
1003 _old_arena(mtCompiler),
1004 _mach_constant_base_node(NULL),
1005 _Compile_types(mtCompiler),
1006 _initial_gvn(NULL),
1007 _for_igvn(NULL),
1008 _warm_calls(NULL),
1009 _number_of_mh_late_inlines(0),
1010 _print_inlining_stream(NULL),
1011 _print_inlining_list(NULL),
1012 _print_inlining_idx(0),
1013 _print_inlining_output(NULL),
1014 _replay_inline_data(NULL),
1015 _java_calls(0),
1016 _inner_loops(0),
1017 _interpreter_frame_size(0),
1018 _node_bundling_limit(0),
1019 _node_bundling_base(NULL),
1020 _code_buffer("Compile::Fill_buffer"),
1021 #ifndef PRODUCT
1022 _in_dump_cnt(0),
1023 #endif
1024 _allowed_reasons(0) {
1025 C = this;
1026
1027 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1028 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1029
1030 #ifndef PRODUCT
1031 set_print_assembly(PrintFrameConverterAssembly);
1032 set_parsed_irreducible_loop(false);
1033 #else
1034 set_print_assembly(false); // Must initialize.
1035 #endif
1036 set_has_irreducible_loop(false); // no loops
1037
1038 CompileWrapper cw(this);
1039 Init(/*AliasLevel=*/ 0);
1040 init_tf((*generator)());
1041
1042 {
1043 // The following is a dummy for the sake of GraphKit::gen_stub
1044 Unique_Node_List for_igvn(comp_arena());
1045 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1046 PhaseGVN gvn(Thread::current()->resource_area(),255);
1047 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1048 gvn.transform_no_reclaim(top());
1049
1050 GraphKit kit;
1051 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1052 }
1053
1054 NOT_PRODUCT( verify_graph_edges(); )
1055 Code_Gen();
1056 if (failing()) return;
1057
1058
1059 // Entry point will be accessed using compile->stub_entry_point();
1060 if (code_buffer() == NULL) {
1061 Matcher::soft_match_failure();
1062 } else {
1063 if (PrintAssembly && (WizardMode || Verbose))
1064 tty->print_cr("### Stub::%s", stub_name);
1065
1066 if (!failing()) {
1067 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1068
1069 // Make the NMethod
1070 // For now we mark the frame as never safe for profile stackwalking
1071 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1072 code_buffer(),
1073 CodeOffsets::frame_never_safe,
1074 // _code_offsets.value(CodeOffsets::Frame_Complete),
1075 frame_size_in_words(),
1076 _oop_map_set,
1077 save_arg_registers);
1078 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1079
1080 _stub_entry_point = rs->entry_point();
1081 }
1082 }
1083 }
1084
1085 //------------------------------Init-------------------------------------------
1086 // Prepare for a single compilation
Init(int aliaslevel)1087 void Compile::Init(int aliaslevel) {
1088 _unique = 0;
1089 _regalloc = NULL;
1090
1091 _tf = NULL; // filled in later
1092 _top = NULL; // cached later
1093 _matcher = NULL; // filled in later
1094 _cfg = NULL; // filled in later
1095
1096 set_24_bit_selection_and_mode(Use24BitFP, false);
1097
1098 _node_note_array = NULL;
1099 _default_node_notes = NULL;
1100 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1101
1102 _immutable_memory = NULL; // filled in at first inquiry
1103
1104 // Globally visible Nodes
1105 // First set TOP to NULL to give safe behavior during creation of RootNode
1106 set_cached_top_node(NULL);
1107 set_root(new RootNode());
1108 // Now that you have a Root to point to, create the real TOP
1109 set_cached_top_node( new ConNode(Type::TOP) );
1110 set_recent_alloc(NULL, NULL);
1111
1112 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1113 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1114 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1115 env()->set_dependencies(new Dependencies(env()));
1116
1117 _fixed_slots = 0;
1118 set_has_split_ifs(false);
1119 set_has_loops(has_method() && method()->has_loops()); // first approximation
1120 set_has_stringbuilder(false);
1121 set_has_boxed_value(false);
1122 _trap_can_recompile = false; // no traps emitted yet
1123 _major_progress = true; // start out assuming good things will happen
1124 set_has_unsafe_access(false);
1125 set_max_vector_size(0);
1126 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1127 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1128 set_decompile_count(0);
1129
1130 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1131 _loop_opts_cnt = LoopOptsCount;
1132 set_do_inlining(Inline);
1133 set_max_inline_size(MaxInlineSize);
1134 set_freq_inline_size(FreqInlineSize);
1135 set_do_scheduling(OptoScheduling);
1136 set_do_count_invocations(false);
1137 set_do_method_data_update(false);
1138
1139 set_do_vector_loop(false);
1140
1141 if (AllowVectorizeOnDemand) {
1142 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1143 set_do_vector_loop(true);
1144 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());})
1145 } else if (has_method() && method()->name() != 0 &&
1146 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1147 set_do_vector_loop(true);
1148 }
1149 }
1150 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1151 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());})
1152
1153 set_age_code(has_method() && method()->profile_aging());
1154 set_rtm_state(NoRTM); // No RTM lock eliding by default
1155 _max_node_limit = _directive->MaxNodeLimitOption;
1156
1157 #if INCLUDE_RTM_OPT
1158 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1159 int rtm_state = method()->method_data()->rtm_state();
1160 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1161 // Don't generate RTM lock eliding code.
1162 set_rtm_state(NoRTM);
1163 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1164 // Generate RTM lock eliding code without abort ratio calculation code.
1165 set_rtm_state(UseRTM);
1166 } else if (UseRTMDeopt) {
1167 // Generate RTM lock eliding code and include abort ratio calculation
1168 // code if UseRTMDeopt is on.
1169 set_rtm_state(ProfileRTM);
1170 }
1171 }
1172 #endif
1173 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1174 set_clinit_barrier_on_entry(true);
1175 }
1176 if (debug_info()->recording_non_safepoints()) {
1177 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1178 (comp_arena(), 8, 0, NULL));
1179 set_default_node_notes(Node_Notes::make(this));
1180 }
1181
1182 // // -- Initialize types before each compile --
1183 // // Update cached type information
1184 // if( _method && _method->constants() )
1185 // Type::update_loaded_types(_method, _method->constants());
1186
1187 // Init alias_type map.
1188 if (!_do_escape_analysis && aliaslevel == 3)
1189 aliaslevel = 2; // No unique types without escape analysis
1190 _AliasLevel = aliaslevel;
1191 const int grow_ats = 16;
1192 _max_alias_types = grow_ats;
1193 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1194 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1195 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1196 {
1197 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1198 }
1199 // Initialize the first few types.
1200 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1201 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1202 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1203 _num_alias_types = AliasIdxRaw+1;
1204 // Zero out the alias type cache.
1205 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1206 // A NULL adr_type hits in the cache right away. Preload the right answer.
1207 probe_alias_cache(NULL)->_index = AliasIdxTop;
1208
1209 _intrinsics = NULL;
1210 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1211 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1212 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1213 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1214 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1215 register_library_intrinsics();
1216 }
1217
1218 //---------------------------init_start----------------------------------------
1219 // Install the StartNode on this compile object.
init_start(StartNode * s)1220 void Compile::init_start(StartNode* s) {
1221 if (failing())
1222 return; // already failing
1223 assert(s == start(), "");
1224 }
1225
1226 /**
1227 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1228 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1229 * the ideal graph.
1230 */
start() const1231 StartNode* Compile::start() const {
1232 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1233 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1234 Node* start = root()->fast_out(i);
1235 if (start->is_Start()) {
1236 return start->as_Start();
1237 }
1238 }
1239 fatal("Did not find Start node!");
1240 return NULL;
1241 }
1242
1243 //-------------------------------immutable_memory-------------------------------------
1244 // Access immutable memory
immutable_memory()1245 Node* Compile::immutable_memory() {
1246 if (_immutable_memory != NULL) {
1247 return _immutable_memory;
1248 }
1249 StartNode* s = start();
1250 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1251 Node *p = s->fast_out(i);
1252 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1253 _immutable_memory = p;
1254 return _immutable_memory;
1255 }
1256 }
1257 ShouldNotReachHere();
1258 return NULL;
1259 }
1260
1261 //----------------------set_cached_top_node------------------------------------
1262 // Install the cached top node, and make sure Node::is_top works correctly.
set_cached_top_node(Node * tn)1263 void Compile::set_cached_top_node(Node* tn) {
1264 if (tn != NULL) verify_top(tn);
1265 Node* old_top = _top;
1266 _top = tn;
1267 // Calling Node::setup_is_top allows the nodes the chance to adjust
1268 // their _out arrays.
1269 if (_top != NULL) _top->setup_is_top();
1270 if (old_top != NULL) old_top->setup_is_top();
1271 assert(_top == NULL || top()->is_top(), "");
1272 }
1273
1274 #ifdef ASSERT
count_live_nodes_by_graph_walk()1275 uint Compile::count_live_nodes_by_graph_walk() {
1276 Unique_Node_List useful(comp_arena());
1277 // Get useful node list by walking the graph.
1278 identify_useful_nodes(useful);
1279 return useful.size();
1280 }
1281
print_missing_nodes()1282 void Compile::print_missing_nodes() {
1283
1284 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1285 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1286 return;
1287 }
1288
1289 // This is an expensive function. It is executed only when the user
1290 // specifies VerifyIdealNodeCount option or otherwise knows the
1291 // additional work that needs to be done to identify reachable nodes
1292 // by walking the flow graph and find the missing ones using
1293 // _dead_node_list.
1294
1295 Unique_Node_List useful(comp_arena());
1296 // Get useful node list by walking the graph.
1297 identify_useful_nodes(useful);
1298
1299 uint l_nodes = C->live_nodes();
1300 uint l_nodes_by_walk = useful.size();
1301
1302 if (l_nodes != l_nodes_by_walk) {
1303 if (_log != NULL) {
1304 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1305 _log->stamp();
1306 _log->end_head();
1307 }
1308 VectorSet& useful_member_set = useful.member_set();
1309 int last_idx = l_nodes_by_walk;
1310 for (int i = 0; i < last_idx; i++) {
1311 if (useful_member_set.test(i)) {
1312 if (_dead_node_list.test(i)) {
1313 if (_log != NULL) {
1314 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1315 }
1316 if (PrintIdealNodeCount) {
1317 // Print the log message to tty
1318 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1319 useful.at(i)->dump();
1320 }
1321 }
1322 }
1323 else if (! _dead_node_list.test(i)) {
1324 if (_log != NULL) {
1325 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1326 }
1327 if (PrintIdealNodeCount) {
1328 // Print the log message to tty
1329 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1330 }
1331 }
1332 }
1333 if (_log != NULL) {
1334 _log->tail("mismatched_nodes");
1335 }
1336 }
1337 }
record_modified_node(Node * n)1338 void Compile::record_modified_node(Node* n) {
1339 if (_modified_nodes != NULL && !_inlining_incrementally &&
1340 n->outcnt() != 0 && !n->is_Con()) {
1341 _modified_nodes->push(n);
1342 }
1343 }
1344
remove_modified_node(Node * n)1345 void Compile::remove_modified_node(Node* n) {
1346 if (_modified_nodes != NULL) {
1347 _modified_nodes->remove(n);
1348 }
1349 }
1350 #endif
1351
1352 #ifndef PRODUCT
verify_top(Node * tn) const1353 void Compile::verify_top(Node* tn) const {
1354 if (tn != NULL) {
1355 assert(tn->is_Con(), "top node must be a constant");
1356 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1357 assert(tn->in(0) != NULL, "must have live top node");
1358 }
1359 }
1360 #endif
1361
1362
1363 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1364
grow_node_notes(GrowableArray<Node_Notes * > * arr,int grow_by)1365 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1366 guarantee(arr != NULL, "");
1367 int num_blocks = arr->length();
1368 if (grow_by < num_blocks) grow_by = num_blocks;
1369 int num_notes = grow_by * _node_notes_block_size;
1370 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1371 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1372 while (num_notes > 0) {
1373 arr->append(notes);
1374 notes += _node_notes_block_size;
1375 num_notes -= _node_notes_block_size;
1376 }
1377 assert(num_notes == 0, "exact multiple, please");
1378 }
1379
copy_node_notes_to(Node * dest,Node * source)1380 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1381 if (source == NULL || dest == NULL) return false;
1382
1383 if (dest->is_Con())
1384 return false; // Do not push debug info onto constants.
1385
1386 #ifdef ASSERT
1387 // Leave a bread crumb trail pointing to the original node:
1388 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1389 dest->set_debug_orig(source);
1390 }
1391 #endif
1392
1393 if (node_note_array() == NULL)
1394 return false; // Not collecting any notes now.
1395
1396 // This is a copy onto a pre-existing node, which may already have notes.
1397 // If both nodes have notes, do not overwrite any pre-existing notes.
1398 Node_Notes* source_notes = node_notes_at(source->_idx);
1399 if (source_notes == NULL || source_notes->is_clear()) return false;
1400 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1401 if (dest_notes == NULL || dest_notes->is_clear()) {
1402 return set_node_notes_at(dest->_idx, source_notes);
1403 }
1404
1405 Node_Notes merged_notes = (*source_notes);
1406 // The order of operations here ensures that dest notes will win...
1407 merged_notes.update_from(dest_notes);
1408 return set_node_notes_at(dest->_idx, &merged_notes);
1409 }
1410
1411
1412 //--------------------------allow_range_check_smearing-------------------------
1413 // Gating condition for coalescing similar range checks.
1414 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1415 // single covering check that is at least as strong as any of them.
1416 // If the optimization succeeds, the simplified (strengthened) range check
1417 // will always succeed. If it fails, we will deopt, and then give up
1418 // on the optimization.
allow_range_check_smearing() const1419 bool Compile::allow_range_check_smearing() const {
1420 // If this method has already thrown a range-check,
1421 // assume it was because we already tried range smearing
1422 // and it failed.
1423 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1424 return !already_trapped;
1425 }
1426
1427
1428 //------------------------------flatten_alias_type-----------------------------
flatten_alias_type(const TypePtr * tj) const1429 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1430 int offset = tj->offset();
1431 TypePtr::PTR ptr = tj->ptr();
1432
1433 // Known instance (scalarizable allocation) alias only with itself.
1434 bool is_known_inst = tj->isa_oopptr() != NULL &&
1435 tj->is_oopptr()->is_known_instance();
1436
1437 // Process weird unsafe references.
1438 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1439 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1440 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1441 tj = TypeOopPtr::BOTTOM;
1442 ptr = tj->ptr();
1443 offset = tj->offset();
1444 }
1445
1446 // Array pointers need some flattening
1447 const TypeAryPtr *ta = tj->isa_aryptr();
1448 if (ta && ta->is_stable()) {
1449 // Erase stability property for alias analysis.
1450 tj = ta = ta->cast_to_stable(false);
1451 }
1452 if( ta && is_known_inst ) {
1453 if ( offset != Type::OffsetBot &&
1454 offset > arrayOopDesc::length_offset_in_bytes() ) {
1455 offset = Type::OffsetBot; // Flatten constant access into array body only
1456 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1457 }
1458 } else if( ta && _AliasLevel >= 2 ) {
1459 // For arrays indexed by constant indices, we flatten the alias
1460 // space to include all of the array body. Only the header, klass
1461 // and array length can be accessed un-aliased.
1462 if( offset != Type::OffsetBot ) {
1463 if( ta->const_oop() ) { // MethodData* or Method*
1464 offset = Type::OffsetBot; // Flatten constant access into array body
1465 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1466 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1467 // range is OK as-is.
1468 tj = ta = TypeAryPtr::RANGE;
1469 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1470 tj = TypeInstPtr::KLASS; // all klass loads look alike
1471 ta = TypeAryPtr::RANGE; // generic ignored junk
1472 ptr = TypePtr::BotPTR;
1473 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1474 tj = TypeInstPtr::MARK;
1475 ta = TypeAryPtr::RANGE; // generic ignored junk
1476 ptr = TypePtr::BotPTR;
1477 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1478 ta = tj->isa_aryptr();
1479 } else { // Random constant offset into array body
1480 offset = Type::OffsetBot; // Flatten constant access into array body
1481 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1482 }
1483 }
1484 // Arrays of fixed size alias with arrays of unknown size.
1485 if (ta->size() != TypeInt::POS) {
1486 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1487 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1488 }
1489 // Arrays of known objects become arrays of unknown objects.
1490 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1491 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1492 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1493 }
1494 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1495 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1496 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1497 }
1498 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1499 // cannot be distinguished by bytecode alone.
1500 if (ta->elem() == TypeInt::BOOL) {
1501 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1502 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1503 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1504 }
1505 // During the 2nd round of IterGVN, NotNull castings are removed.
1506 // Make sure the Bottom and NotNull variants alias the same.
1507 // Also, make sure exact and non-exact variants alias the same.
1508 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1509 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1510 }
1511 }
1512
1513 // Oop pointers need some flattening
1514 const TypeInstPtr *to = tj->isa_instptr();
1515 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1516 ciInstanceKlass *k = to->klass()->as_instance_klass();
1517 if( ptr == TypePtr::Constant ) {
1518 if (to->klass() != ciEnv::current()->Class_klass() ||
1519 offset < k->size_helper() * wordSize) {
1520 // No constant oop pointers (such as Strings); they alias with
1521 // unknown strings.
1522 assert(!is_known_inst, "not scalarizable allocation");
1523 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1524 }
1525 } else if( is_known_inst ) {
1526 tj = to; // Keep NotNull and klass_is_exact for instance type
1527 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1528 // During the 2nd round of IterGVN, NotNull castings are removed.
1529 // Make sure the Bottom and NotNull variants alias the same.
1530 // Also, make sure exact and non-exact variants alias the same.
1531 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1532 }
1533 if (to->speculative() != NULL) {
1534 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1535 }
1536 // Canonicalize the holder of this field
1537 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1538 // First handle header references such as a LoadKlassNode, even if the
1539 // object's klass is unloaded at compile time (4965979).
1540 if (!is_known_inst) { // Do it only for non-instance types
1541 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1542 }
1543 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1544 to = tj->is_instptr();
1545 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1546 // Static fields are in the space above the normal instance
1547 // fields in the java.lang.Class instance.
1548 if (to->klass() != ciEnv::current()->Class_klass()) {
1549 to = NULL;
1550 tj = TypeOopPtr::BOTTOM;
1551 offset = tj->offset();
1552 }
1553 } else {
1554 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1555 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1556 if( is_known_inst ) {
1557 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1558 } else {
1559 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1560 }
1561 }
1562 }
1563 }
1564
1565 // Klass pointers to object array klasses need some flattening
1566 const TypeKlassPtr *tk = tj->isa_klassptr();
1567 if( tk ) {
1568 // If we are referencing a field within a Klass, we need
1569 // to assume the worst case of an Object. Both exact and
1570 // inexact types must flatten to the same alias class so
1571 // use NotNull as the PTR.
1572 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1573
1574 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1575 TypeKlassPtr::OBJECT->klass(),
1576 offset);
1577 }
1578
1579 ciKlass* klass = tk->klass();
1580 if( klass->is_obj_array_klass() ) {
1581 ciKlass* k = TypeAryPtr::OOPS->klass();
1582 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1583 k = TypeInstPtr::BOTTOM->klass();
1584 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1585 }
1586
1587 // Check for precise loads from the primary supertype array and force them
1588 // to the supertype cache alias index. Check for generic array loads from
1589 // the primary supertype array and also force them to the supertype cache
1590 // alias index. Since the same load can reach both, we need to merge
1591 // these 2 disparate memories into the same alias class. Since the
1592 // primary supertype array is read-only, there's no chance of confusion
1593 // where we bypass an array load and an array store.
1594 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1595 if (offset == Type::OffsetBot ||
1596 (offset >= primary_supers_offset &&
1597 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1598 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1599 offset = in_bytes(Klass::secondary_super_cache_offset());
1600 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1601 }
1602 }
1603
1604 // Flatten all Raw pointers together.
1605 if (tj->base() == Type::RawPtr)
1606 tj = TypeRawPtr::BOTTOM;
1607
1608 if (tj->base() == Type::AnyPtr)
1609 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1610
1611 // Flatten all to bottom for now
1612 switch( _AliasLevel ) {
1613 case 0:
1614 tj = TypePtr::BOTTOM;
1615 break;
1616 case 1: // Flatten to: oop, static, field or array
1617 switch (tj->base()) {
1618 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1619 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1620 case Type::AryPtr: // do not distinguish arrays at all
1621 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1622 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1623 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1624 default: ShouldNotReachHere();
1625 }
1626 break;
1627 case 2: // No collapsing at level 2; keep all splits
1628 case 3: // No collapsing at level 3; keep all splits
1629 break;
1630 default:
1631 Unimplemented();
1632 }
1633
1634 offset = tj->offset();
1635 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1636
1637 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1638 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1639 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1640 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1641 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1642 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1643 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1644 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1645 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1646 assert( tj->ptr() != TypePtr::TopPTR &&
1647 tj->ptr() != TypePtr::AnyNull &&
1648 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1649 // assert( tj->ptr() != TypePtr::Constant ||
1650 // tj->base() == Type::RawPtr ||
1651 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1652
1653 return tj;
1654 }
1655
Init(int i,const TypePtr * at)1656 void Compile::AliasType::Init(int i, const TypePtr* at) {
1657 _index = i;
1658 _adr_type = at;
1659 _field = NULL;
1660 _element = NULL;
1661 _is_rewritable = true; // default
1662 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1663 if (atoop != NULL && atoop->is_known_instance()) {
1664 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1665 _general_index = Compile::current()->get_alias_index(gt);
1666 } else {
1667 _general_index = 0;
1668 }
1669 }
1670
basic_type() const1671 BasicType Compile::AliasType::basic_type() const {
1672 if (element() != NULL) {
1673 const Type* element = adr_type()->is_aryptr()->elem();
1674 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1675 } if (field() != NULL) {
1676 return field()->layout_type();
1677 } else {
1678 return T_ILLEGAL; // unknown
1679 }
1680 }
1681
1682 //---------------------------------print_on------------------------------------
1683 #ifndef PRODUCT
print_on(outputStream * st)1684 void Compile::AliasType::print_on(outputStream* st) {
1685 if (index() < 10)
1686 st->print("@ <%d> ", index());
1687 else st->print("@ <%d>", index());
1688 st->print(is_rewritable() ? " " : " RO");
1689 int offset = adr_type()->offset();
1690 if (offset == Type::OffsetBot)
1691 st->print(" +any");
1692 else st->print(" +%-3d", offset);
1693 st->print(" in ");
1694 adr_type()->dump_on(st);
1695 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1696 if (field() != NULL && tjp) {
1697 if (tjp->klass() != field()->holder() ||
1698 tjp->offset() != field()->offset_in_bytes()) {
1699 st->print(" != ");
1700 field()->print();
1701 st->print(" ***");
1702 }
1703 }
1704 }
1705
print_alias_types()1706 void print_alias_types() {
1707 Compile* C = Compile::current();
1708 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1709 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1710 C->alias_type(idx)->print_on(tty);
1711 tty->cr();
1712 }
1713 }
1714 #endif
1715
1716
1717 //----------------------------probe_alias_cache--------------------------------
probe_alias_cache(const TypePtr * adr_type)1718 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1719 intptr_t key = (intptr_t) adr_type;
1720 key ^= key >> logAliasCacheSize;
1721 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1722 }
1723
1724
1725 //-----------------------------grow_alias_types--------------------------------
grow_alias_types()1726 void Compile::grow_alias_types() {
1727 const int old_ats = _max_alias_types; // how many before?
1728 const int new_ats = old_ats; // how many more?
1729 const int grow_ats = old_ats+new_ats; // how many now?
1730 _max_alias_types = grow_ats;
1731 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1732 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1733 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1734 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1735 }
1736
1737
1738 //--------------------------------find_alias_type------------------------------
find_alias_type(const TypePtr * adr_type,bool no_create,ciField * original_field)1739 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1740 if (_AliasLevel == 0)
1741 return alias_type(AliasIdxBot);
1742
1743 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1744 if (ace->_adr_type == adr_type) {
1745 return alias_type(ace->_index);
1746 }
1747
1748 // Handle special cases.
1749 if (adr_type == NULL) return alias_type(AliasIdxTop);
1750 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1751
1752 // Do it the slow way.
1753 const TypePtr* flat = flatten_alias_type(adr_type);
1754
1755 #ifdef ASSERT
1756 {
1757 ResourceMark rm;
1758 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1759 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1760 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1761 Type::str(adr_type));
1762 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1763 const TypeOopPtr* foop = flat->is_oopptr();
1764 // Scalarizable allocations have exact klass always.
1765 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1766 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1767 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1768 Type::str(foop), Type::str(xoop));
1769 }
1770 }
1771 #endif
1772
1773 int idx = AliasIdxTop;
1774 for (int i = 0; i < num_alias_types(); i++) {
1775 if (alias_type(i)->adr_type() == flat) {
1776 idx = i;
1777 break;
1778 }
1779 }
1780
1781 if (idx == AliasIdxTop) {
1782 if (no_create) return NULL;
1783 // Grow the array if necessary.
1784 if (_num_alias_types == _max_alias_types) grow_alias_types();
1785 // Add a new alias type.
1786 idx = _num_alias_types++;
1787 _alias_types[idx]->Init(idx, flat);
1788 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1789 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1790 if (flat->isa_instptr()) {
1791 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1792 && flat->is_instptr()->klass() == env()->Class_klass())
1793 alias_type(idx)->set_rewritable(false);
1794 }
1795 if (flat->isa_aryptr()) {
1796 #ifdef ASSERT
1797 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1798 // (T_BYTE has the weakest alignment and size restrictions...)
1799 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1800 #endif
1801 if (flat->offset() == TypePtr::OffsetBot) {
1802 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1803 }
1804 }
1805 if (flat->isa_klassptr()) {
1806 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1807 alias_type(idx)->set_rewritable(false);
1808 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1809 alias_type(idx)->set_rewritable(false);
1810 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1811 alias_type(idx)->set_rewritable(false);
1812 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1813 alias_type(idx)->set_rewritable(false);
1814 }
1815 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1816 // but the base pointer type is not distinctive enough to identify
1817 // references into JavaThread.)
1818
1819 // Check for final fields.
1820 const TypeInstPtr* tinst = flat->isa_instptr();
1821 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1822 ciField* field;
1823 if (tinst->const_oop() != NULL &&
1824 tinst->klass() == ciEnv::current()->Class_klass() &&
1825 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1826 // static field
1827 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1828 field = k->get_field_by_offset(tinst->offset(), true);
1829 } else {
1830 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1831 field = k->get_field_by_offset(tinst->offset(), false);
1832 }
1833 assert(field == NULL ||
1834 original_field == NULL ||
1835 (field->holder() == original_field->holder() &&
1836 field->offset() == original_field->offset() &&
1837 field->is_static() == original_field->is_static()), "wrong field?");
1838 // Set field() and is_rewritable() attributes.
1839 if (field != NULL) alias_type(idx)->set_field(field);
1840 }
1841 }
1842
1843 // Fill the cache for next time.
1844 ace->_adr_type = adr_type;
1845 ace->_index = idx;
1846 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1847
1848 // Might as well try to fill the cache for the flattened version, too.
1849 AliasCacheEntry* face = probe_alias_cache(flat);
1850 if (face->_adr_type == NULL) {
1851 face->_adr_type = flat;
1852 face->_index = idx;
1853 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1854 }
1855
1856 return alias_type(idx);
1857 }
1858
1859
alias_type(ciField * field)1860 Compile::AliasType* Compile::alias_type(ciField* field) {
1861 const TypeOopPtr* t;
1862 if (field->is_static())
1863 t = TypeInstPtr::make(field->holder()->java_mirror());
1864 else
1865 t = TypeOopPtr::make_from_klass_raw(field->holder());
1866 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1867 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1868 return atp;
1869 }
1870
1871
1872 //------------------------------have_alias_type--------------------------------
have_alias_type(const TypePtr * adr_type)1873 bool Compile::have_alias_type(const TypePtr* adr_type) {
1874 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1875 if (ace->_adr_type == adr_type) {
1876 return true;
1877 }
1878
1879 // Handle special cases.
1880 if (adr_type == NULL) return true;
1881 if (adr_type == TypePtr::BOTTOM) return true;
1882
1883 return find_alias_type(adr_type, true, NULL) != NULL;
1884 }
1885
1886 //-----------------------------must_alias--------------------------------------
1887 // True if all values of the given address type are in the given alias category.
must_alias(const TypePtr * adr_type,int alias_idx)1888 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1889 if (alias_idx == AliasIdxBot) return true; // the universal category
1890 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1891 if (alias_idx == AliasIdxTop) return false; // the empty category
1892 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1893
1894 // the only remaining possible overlap is identity
1895 int adr_idx = get_alias_index(adr_type);
1896 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1897 assert(adr_idx == alias_idx ||
1898 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1899 && adr_type != TypeOopPtr::BOTTOM),
1900 "should not be testing for overlap with an unsafe pointer");
1901 return adr_idx == alias_idx;
1902 }
1903
1904 //------------------------------can_alias--------------------------------------
1905 // True if any values of the given address type are in the given alias category.
can_alias(const TypePtr * adr_type,int alias_idx)1906 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1907 if (alias_idx == AliasIdxTop) return false; // the empty category
1908 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1909 if (alias_idx == AliasIdxBot) return true; // the universal category
1910 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1911
1912 // the only remaining possible overlap is identity
1913 int adr_idx = get_alias_index(adr_type);
1914 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1915 return adr_idx == alias_idx;
1916 }
1917
1918
1919
1920 //---------------------------pop_warm_call-------------------------------------
pop_warm_call()1921 WarmCallInfo* Compile::pop_warm_call() {
1922 WarmCallInfo* wci = _warm_calls;
1923 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1924 return wci;
1925 }
1926
1927 //----------------------------Inline_Warm--------------------------------------
Inline_Warm()1928 int Compile::Inline_Warm() {
1929 // If there is room, try to inline some more warm call sites.
1930 // %%% Do a graph index compaction pass when we think we're out of space?
1931 if (!InlineWarmCalls) return 0;
1932
1933 int calls_made_hot = 0;
1934 int room_to_grow = NodeCountInliningCutoff - unique();
1935 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1936 int amount_grown = 0;
1937 WarmCallInfo* call;
1938 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1939 int est_size = (int)call->size();
1940 if (est_size > (room_to_grow - amount_grown)) {
1941 // This one won't fit anyway. Get rid of it.
1942 call->make_cold();
1943 continue;
1944 }
1945 call->make_hot();
1946 calls_made_hot++;
1947 amount_grown += est_size;
1948 amount_to_grow -= est_size;
1949 }
1950
1951 if (calls_made_hot > 0) set_major_progress();
1952 return calls_made_hot;
1953 }
1954
1955
1956 //----------------------------Finish_Warm--------------------------------------
Finish_Warm()1957 void Compile::Finish_Warm() {
1958 if (!InlineWarmCalls) return;
1959 if (failing()) return;
1960 if (warm_calls() == NULL) return;
1961
1962 // Clean up loose ends, if we are out of space for inlining.
1963 WarmCallInfo* call;
1964 while ((call = pop_warm_call()) != NULL) {
1965 call->make_cold();
1966 }
1967 }
1968
1969 //---------------------cleanup_loop_predicates-----------------------
1970 // Remove the opaque nodes that protect the predicates so that all unused
1971 // checks and uncommon_traps will be eliminated from the ideal graph
cleanup_loop_predicates(PhaseIterGVN & igvn)1972 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1973 if (predicate_count()==0) return;
1974 for (int i = predicate_count(); i > 0; i--) {
1975 Node * n = predicate_opaque1_node(i-1);
1976 assert(n->Opcode() == Op_Opaque1, "must be");
1977 igvn.replace_node(n, n->in(1));
1978 }
1979 assert(predicate_count()==0, "should be clean!");
1980 }
1981
add_range_check_cast(Node * n)1982 void Compile::add_range_check_cast(Node* n) {
1983 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1984 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1985 _range_check_casts->append(n);
1986 }
1987
1988 // Remove all range check dependent CastIINodes.
remove_range_check_casts(PhaseIterGVN & igvn)1989 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1990 for (int i = range_check_cast_count(); i > 0; i--) {
1991 Node* cast = range_check_cast_node(i-1);
1992 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1993 igvn.replace_node(cast, cast->in(1));
1994 }
1995 assert(range_check_cast_count() == 0, "should be empty");
1996 }
1997
add_opaque4_node(Node * n)1998 void Compile::add_opaque4_node(Node* n) {
1999 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2000 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2001 _opaque4_nodes->append(n);
2002 }
2003
2004 // Remove all Opaque4 nodes.
remove_opaque4_nodes(PhaseIterGVN & igvn)2005 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2006 for (int i = opaque4_count(); i > 0; i--) {
2007 Node* opaq = opaque4_node(i-1);
2008 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2009 igvn.replace_node(opaq, opaq->in(2));
2010 }
2011 assert(opaque4_count() == 0, "should be empty");
2012 }
2013
2014 // StringOpts and late inlining of string methods
inline_string_calls(bool parse_time)2015 void Compile::inline_string_calls(bool parse_time) {
2016 {
2017 // remove useless nodes to make the usage analysis simpler
2018 ResourceMark rm;
2019 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2020 }
2021
2022 {
2023 ResourceMark rm;
2024 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2025 PhaseStringOpts pso(initial_gvn(), for_igvn());
2026 print_method(PHASE_AFTER_STRINGOPTS, 3);
2027 }
2028
2029 // now inline anything that we skipped the first time around
2030 if (!parse_time) {
2031 _late_inlines_pos = _late_inlines.length();
2032 }
2033
2034 while (_string_late_inlines.length() > 0) {
2035 CallGenerator* cg = _string_late_inlines.pop();
2036 cg->do_late_inline();
2037 if (failing()) return;
2038 }
2039 _string_late_inlines.trunc_to(0);
2040 }
2041
2042 // Late inlining of boxing methods
inline_boxing_calls(PhaseIterGVN & igvn)2043 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2044 if (_boxing_late_inlines.length() > 0) {
2045 assert(has_boxed_value(), "inconsistent");
2046
2047 PhaseGVN* gvn = initial_gvn();
2048 set_inlining_incrementally(true);
2049
2050 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2051 for_igvn()->clear();
2052 gvn->replace_with(&igvn);
2053
2054 _late_inlines_pos = _late_inlines.length();
2055
2056 while (_boxing_late_inlines.length() > 0) {
2057 CallGenerator* cg = _boxing_late_inlines.pop();
2058 cg->do_late_inline();
2059 if (failing()) return;
2060 }
2061 _boxing_late_inlines.trunc_to(0);
2062
2063 inline_incrementally_cleanup(igvn);
2064
2065 set_inlining_incrementally(false);
2066 }
2067 }
2068
inline_incrementally_one()2069 bool Compile::inline_incrementally_one() {
2070 assert(IncrementalInline, "incremental inlining should be on");
2071
2072 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2073 set_inlining_progress(false);
2074 set_do_cleanup(false);
2075 int i = 0;
2076 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2077 CallGenerator* cg = _late_inlines.at(i);
2078 _late_inlines_pos = i+1;
2079 cg->do_late_inline();
2080 if (failing()) return false;
2081 }
2082 int j = 0;
2083 for (; i < _late_inlines.length(); i++, j++) {
2084 _late_inlines.at_put(j, _late_inlines.at(i));
2085 }
2086 _late_inlines.trunc_to(j);
2087 assert(inlining_progress() || _late_inlines.length() == 0, "");
2088
2089 bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2090
2091 set_inlining_progress(false);
2092 set_do_cleanup(false);
2093 return (_late_inlines.length() > 0) && !needs_cleanup;
2094 }
2095
inline_incrementally_cleanup(PhaseIterGVN & igvn)2096 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2097 {
2098 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2099 ResourceMark rm;
2100 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2101 }
2102 {
2103 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2104 igvn = PhaseIterGVN(initial_gvn());
2105 igvn.optimize();
2106 }
2107 }
2108
2109 // Perform incremental inlining until bound on number of live nodes is reached
inline_incrementally(PhaseIterGVN & igvn)2110 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2111 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2112
2113 set_inlining_incrementally(true);
2114 uint low_live_nodes = 0;
2115
2116 while (_late_inlines.length() > 0) {
2117 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2118 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2119 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2120 // PhaseIdealLoop is expensive so we only try it once we are
2121 // out of live nodes and we only try it again if the previous
2122 // helped got the number of nodes down significantly
2123 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2124 if (failing()) return;
2125 low_live_nodes = live_nodes();
2126 _major_progress = true;
2127 }
2128
2129 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2130 break; // finish
2131 }
2132 }
2133
2134 for_igvn()->clear();
2135 initial_gvn()->replace_with(&igvn);
2136
2137 while (inline_incrementally_one()) {
2138 assert(!failing(), "inconsistent");
2139 }
2140
2141 if (failing()) return;
2142
2143 inline_incrementally_cleanup(igvn);
2144
2145 if (failing()) return;
2146 }
2147 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2148
2149 if (_string_late_inlines.length() > 0) {
2150 assert(has_stringbuilder(), "inconsistent");
2151 for_igvn()->clear();
2152 initial_gvn()->replace_with(&igvn);
2153
2154 inline_string_calls(false);
2155
2156 if (failing()) return;
2157
2158 inline_incrementally_cleanup(igvn);
2159 }
2160
2161 set_inlining_incrementally(false);
2162 }
2163
2164
optimize_loops(PhaseIterGVN & igvn,LoopOptsMode mode)2165 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2166 if(_loop_opts_cnt > 0) {
2167 debug_only( int cnt = 0; );
2168 while(major_progress() && (_loop_opts_cnt > 0)) {
2169 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2170 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2171 PhaseIdealLoop::optimize(igvn, mode);
2172 _loop_opts_cnt--;
2173 if (failing()) return false;
2174 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2175 }
2176 }
2177 return true;
2178 }
2179
2180 // Remove edges from "root" to each SafePoint at a backward branch.
2181 // They were inserted during parsing (see add_safepoint()) to make
2182 // infinite loops without calls or exceptions visible to root, i.e.,
2183 // useful.
remove_root_to_sfpts_edges(PhaseIterGVN & igvn)2184 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2185 Node *r = root();
2186 if (r != NULL) {
2187 for (uint i = r->req(); i < r->len(); ++i) {
2188 Node *n = r->in(i);
2189 if (n != NULL && n->is_SafePoint()) {
2190 r->rm_prec(i);
2191 if (n->outcnt() == 0) {
2192 igvn.remove_dead_node(n);
2193 }
2194 --i;
2195 }
2196 }
2197 // Parsing may have added top inputs to the root node (Path
2198 // leading to the Halt node proven dead). Make sure we get a
2199 // chance to clean them up.
2200 igvn._worklist.push(r);
2201 igvn.optimize();
2202 }
2203 }
2204
2205 //------------------------------Optimize---------------------------------------
2206 // Given a graph, optimize it.
Optimize()2207 void Compile::Optimize() {
2208 TracePhase tp("optimizer", &timers[_t_optimizer]);
2209
2210 #ifndef PRODUCT
2211 if (_directive->BreakAtCompileOption) {
2212 BREAKPOINT;
2213 }
2214
2215 #endif
2216
2217 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2218 #ifdef ASSERT
2219 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2220 #endif
2221
2222 ResourceMark rm;
2223
2224 print_inlining_reinit();
2225
2226 NOT_PRODUCT( verify_graph_edges(); )
2227
2228 print_method(PHASE_AFTER_PARSING);
2229
2230 {
2231 // Iterative Global Value Numbering, including ideal transforms
2232 // Initialize IterGVN with types and values from parse-time GVN
2233 PhaseIterGVN igvn(initial_gvn());
2234 #ifdef ASSERT
2235 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2236 #endif
2237 {
2238 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2239 igvn.optimize();
2240 }
2241
2242 if (failing()) return;
2243
2244 print_method(PHASE_ITER_GVN1, 2);
2245
2246 inline_incrementally(igvn);
2247
2248 print_method(PHASE_INCREMENTAL_INLINE, 2);
2249
2250 if (failing()) return;
2251
2252 if (eliminate_boxing()) {
2253 // Inline valueOf() methods now.
2254 inline_boxing_calls(igvn);
2255
2256 if (AlwaysIncrementalInline) {
2257 inline_incrementally(igvn);
2258 }
2259
2260 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2261
2262 if (failing()) return;
2263 }
2264
2265 // Now that all inlining is over, cut edge from root to loop
2266 // safepoints
2267 remove_root_to_sfpts_edges(igvn);
2268
2269 // Remove the speculative part of types and clean up the graph from
2270 // the extra CastPP nodes whose only purpose is to carry them. Do
2271 // that early so that optimizations are not disrupted by the extra
2272 // CastPP nodes.
2273 remove_speculative_types(igvn);
2274
2275 // No more new expensive nodes will be added to the list from here
2276 // so keep only the actual candidates for optimizations.
2277 cleanup_expensive_nodes(igvn);
2278
2279 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2280 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2281 initial_gvn()->replace_with(&igvn);
2282 for_igvn()->clear();
2283 Unique_Node_List new_worklist(C->comp_arena());
2284 {
2285 ResourceMark rm;
2286 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2287 }
2288 set_for_igvn(&new_worklist);
2289 igvn = PhaseIterGVN(initial_gvn());
2290 igvn.optimize();
2291 }
2292
2293 // Perform escape analysis
2294 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2295 if (has_loops()) {
2296 // Cleanup graph (remove dead nodes).
2297 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2298 PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2299 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2300 if (failing()) return;
2301 }
2302 ConnectionGraph::do_analysis(this, &igvn);
2303
2304 if (failing()) return;
2305
2306 // Optimize out fields loads from scalar replaceable allocations.
2307 igvn.optimize();
2308 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2309
2310 if (failing()) return;
2311
2312 if (congraph() != NULL && macro_count() > 0) {
2313 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2314 PhaseMacroExpand mexp(igvn);
2315 mexp.eliminate_macro_nodes();
2316 igvn.set_delay_transform(false);
2317
2318 igvn.optimize();
2319 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2320
2321 if (failing()) return;
2322 }
2323 }
2324
2325 // Loop transforms on the ideal graph. Range Check Elimination,
2326 // peeling, unrolling, etc.
2327
2328 // Set loop opts counter
2329 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2330 {
2331 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2332 PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2333 _loop_opts_cnt--;
2334 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2335 if (failing()) return;
2336 }
2337 // Loop opts pass if partial peeling occurred in previous pass
2338 if(PartialPeelLoop && 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_PHASEIDEALLOOP2, 2);
2343 if (failing()) return;
2344 }
2345 // Loop opts pass for loop-unrolling before CCP
2346 if(major_progress() && (_loop_opts_cnt > 0)) {
2347 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2348 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2349 _loop_opts_cnt--;
2350 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2351 }
2352 if (!failing()) {
2353 // Verify that last round of loop opts produced a valid graph
2354 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2355 PhaseIdealLoop::verify(igvn);
2356 }
2357 }
2358 if (failing()) return;
2359
2360 // Conditional Constant Propagation;
2361 PhaseCCP ccp( &igvn );
2362 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2363 {
2364 TracePhase tp("ccp", &timers[_t_ccp]);
2365 ccp.do_transform();
2366 }
2367 print_method(PHASE_CPP1, 2);
2368
2369 assert( true, "Break here to ccp.dump_old2new_map()");
2370
2371 // Iterative Global Value Numbering, including ideal transforms
2372 {
2373 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2374 igvn = ccp;
2375 igvn.optimize();
2376 }
2377 print_method(PHASE_ITER_GVN2, 2);
2378
2379 if (failing()) return;
2380
2381 // Loop transforms on the ideal graph. Range Check Elimination,
2382 // peeling, unrolling, etc.
2383 if (!optimize_loops(igvn, LoopOptsDefault)) {
2384 return;
2385 }
2386
2387 if (failing()) return;
2388
2389 // Ensure that major progress is now clear
2390 C->clear_major_progress();
2391
2392 {
2393 // Verify that all previous optimizations produced a valid graph
2394 // at least to this point, even if no loop optimizations were done.
2395 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2396 PhaseIdealLoop::verify(igvn);
2397 }
2398
2399 if (range_check_cast_count() > 0) {
2400 // No more loop optimizations. Remove all range check dependent CastIINodes.
2401 C->remove_range_check_casts(igvn);
2402 igvn.optimize();
2403 }
2404
2405 #ifdef ASSERT
2406 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2407 #endif
2408
2409 {
2410 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2411 PhaseMacroExpand mex(igvn);
2412 if (mex.expand_macro_nodes()) {
2413 assert(failing(), "must bail out w/ explicit message");
2414 return;
2415 }
2416 print_method(PHASE_MACRO_EXPANSION, 2);
2417 }
2418
2419 {
2420 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2421 if (bs->expand_barriers(this, igvn)) {
2422 assert(failing(), "must bail out w/ explicit message");
2423 return;
2424 }
2425 print_method(PHASE_BARRIER_EXPANSION, 2);
2426 }
2427
2428 if (opaque4_count() > 0) {
2429 C->remove_opaque4_nodes(igvn);
2430 igvn.optimize();
2431 }
2432
2433 DEBUG_ONLY( _modified_nodes = NULL; )
2434 } // (End scope of igvn; run destructor if necessary for asserts.)
2435
2436 process_print_inlining();
2437 // A method with only infinite loops has no edges entering loops from root
2438 {
2439 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2440 if (final_graph_reshaping()) {
2441 assert(failing(), "must bail out w/ explicit message");
2442 return;
2443 }
2444 }
2445
2446 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2447 }
2448
2449
2450 //------------------------------Code_Gen---------------------------------------
2451 // Given a graph, generate code for it
Code_Gen()2452 void Compile::Code_Gen() {
2453 if (failing()) {
2454 return;
2455 }
2456
2457 // Perform instruction selection. You might think we could reclaim Matcher
2458 // memory PDQ, but actually the Matcher is used in generating spill code.
2459 // Internals of the Matcher (including some VectorSets) must remain live
2460 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2461 // set a bit in reclaimed memory.
2462
2463 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2464 // nodes. Mapping is only valid at the root of each matched subtree.
2465 NOT_PRODUCT( verify_graph_edges(); )
2466
2467 Matcher matcher;
2468 _matcher = &matcher;
2469 {
2470 TracePhase tp("matcher", &timers[_t_matcher]);
2471 matcher.match();
2472 }
2473 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2474 // nodes. Mapping is only valid at the root of each matched subtree.
2475 NOT_PRODUCT( verify_graph_edges(); )
2476
2477 // If you have too many nodes, or if matching has failed, bail out
2478 check_node_count(0, "out of nodes matching instructions");
2479 if (failing()) {
2480 return;
2481 }
2482
2483 print_method(PHASE_MATCHING, 2);
2484
2485 // Build a proper-looking CFG
2486 PhaseCFG cfg(node_arena(), root(), matcher);
2487 _cfg = &cfg;
2488 {
2489 TracePhase tp("scheduler", &timers[_t_scheduler]);
2490 bool success = cfg.do_global_code_motion();
2491 if (!success) {
2492 return;
2493 }
2494
2495 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2496 NOT_PRODUCT( verify_graph_edges(); )
2497 debug_only( cfg.verify(); )
2498 }
2499
2500 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2501 _regalloc = ®alloc;
2502 {
2503 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2504 // Perform register allocation. After Chaitin, use-def chains are
2505 // no longer accurate (at spill code) and so must be ignored.
2506 // Node->LRG->reg mappings are still accurate.
2507 _regalloc->Register_Allocate();
2508
2509 // Bail out if the allocator builds too many nodes
2510 if (failing()) {
2511 return;
2512 }
2513 }
2514
2515 // Prior to register allocation we kept empty basic blocks in case the
2516 // the allocator needed a place to spill. After register allocation we
2517 // are not adding any new instructions. If any basic block is empty, we
2518 // can now safely remove it.
2519 {
2520 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2521 cfg.remove_empty_blocks();
2522 if (do_freq_based_layout()) {
2523 PhaseBlockLayout layout(cfg);
2524 } else {
2525 cfg.set_loop_alignment();
2526 }
2527 cfg.fixup_flow();
2528 }
2529
2530 // Apply peephole optimizations
2531 if( OptoPeephole ) {
2532 TracePhase tp("peephole", &timers[_t_peephole]);
2533 PhasePeephole peep( _regalloc, cfg);
2534 peep.do_transform();
2535 }
2536
2537 // Do late expand if CPU requires this.
2538 if (Matcher::require_postalloc_expand) {
2539 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2540 cfg.postalloc_expand(_regalloc);
2541 }
2542
2543 // Convert Nodes to instruction bits in a buffer
2544 {
2545 TraceTime tp("output", &timers[_t_output], CITime);
2546 Output();
2547 }
2548
2549 print_method(PHASE_FINAL_CODE);
2550
2551 // He's dead, Jim.
2552 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2553 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2554 }
2555
2556
2557 //------------------------------dump_asm---------------------------------------
2558 // Dump formatted assembly
2559 #if defined(SUPPORT_OPTO_ASSEMBLY)
dump_asm_on(outputStream * st,int * pcs,uint pc_limit)2560 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2561
2562 int pc_digits = 3; // #chars required for pc
2563 int sb_chars = 3; // #chars for "start bundle" indicator
2564 int tab_size = 8;
2565 if (pcs != NULL) {
2566 int max_pc = 0;
2567 for (uint i = 0; i < pc_limit; i++) {
2568 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
2569 }
2570 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
2571 }
2572 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
2573
2574 bool cut_short = false;
2575 st->print_cr("#");
2576 st->print("# "); _tf->dump_on(st); st->cr();
2577 st->print_cr("#");
2578
2579 // For all blocks
2580 int pc = 0x0; // Program counter
2581 char starts_bundle = ' ';
2582 _regalloc->dump_frame();
2583
2584 Node *n = NULL;
2585 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2586 if (VMThread::should_terminate()) {
2587 cut_short = true;
2588 break;
2589 }
2590 Block* block = _cfg->get_block(i);
2591 if (block->is_connector() && !Verbose) {
2592 continue;
2593 }
2594 n = block->head();
2595 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2596 pc = pcs[n->_idx];
2597 st->print("%*.*x", pc_digits, pc_digits, pc);
2598 }
2599 st->fill_to(prefix_len);
2600 block->dump_head(_cfg, st);
2601 if (block->is_connector()) {
2602 st->fill_to(prefix_len);
2603 st->print_cr("# Empty connector block");
2604 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2605 st->fill_to(prefix_len);
2606 st->print_cr("# Block is sole successor of call");
2607 }
2608
2609 // For all instructions
2610 Node *delay = NULL;
2611 for (uint j = 0; j < block->number_of_nodes(); j++) {
2612 if (VMThread::should_terminate()) {
2613 cut_short = true;
2614 break;
2615 }
2616 n = block->get_node(j);
2617 if (valid_bundle_info(n)) {
2618 Bundle* bundle = node_bundling(n);
2619 if (bundle->used_in_unconditional_delay()) {
2620 delay = n;
2621 continue;
2622 }
2623 if (bundle->starts_bundle()) {
2624 starts_bundle = '+';
2625 }
2626 }
2627
2628 if (WizardMode) {
2629 n->dump();
2630 }
2631
2632 if( !n->is_Region() && // Dont print in the Assembly
2633 !n->is_Phi() && // a few noisely useless nodes
2634 !n->is_Proj() &&
2635 !n->is_MachTemp() &&
2636 !n->is_SafePointScalarObject() &&
2637 !n->is_Catch() && // Would be nice to print exception table targets
2638 !n->is_MergeMem() && // Not very interesting
2639 !n->is_top() && // Debug info table constants
2640 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2641 ) {
2642 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2643 pc = pcs[n->_idx];
2644 st->print("%*.*x", pc_digits, pc_digits, pc);
2645 } else {
2646 st->fill_to(pc_digits);
2647 }
2648 st->print(" %c ", starts_bundle);
2649 starts_bundle = ' ';
2650 st->fill_to(prefix_len);
2651 n->format(_regalloc, st);
2652 st->cr();
2653 }
2654
2655 // If we have an instruction with a delay slot, and have seen a delay,
2656 // then back up and print it
2657 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2658 // Coverity finding - Explicit null dereferenced.
2659 guarantee(delay != NULL, "no unconditional delay instruction");
2660 if (WizardMode) delay->dump();
2661
2662 if (node_bundling(delay)->starts_bundle())
2663 starts_bundle = '+';
2664 if ((pcs != NULL) && (n->_idx < pc_limit)) {
2665 pc = pcs[n->_idx];
2666 st->print("%*.*x", pc_digits, pc_digits, pc);
2667 } else {
2668 st->fill_to(pc_digits);
2669 }
2670 st->print(" %c ", starts_bundle);
2671 starts_bundle = ' ';
2672 st->fill_to(prefix_len);
2673 delay->format(_regalloc, st);
2674 st->cr();
2675 delay = NULL;
2676 }
2677
2678 // Dump the exception table as well
2679 if( n->is_Catch() && (Verbose || WizardMode) ) {
2680 // Print the exception table for this offset
2681 _handler_table.print_subtable_for(pc);
2682 }
2683 st->bol(); // Make sure we start on a new line
2684 }
2685 st->cr(); // one empty line between blocks
2686 assert(cut_short || delay == NULL, "no unconditional delay branch");
2687 } // End of per-block dump
2688
2689 if (cut_short) st->print_cr("*** disassembly is cut short ***");
2690 }
2691 #endif
2692
2693 //------------------------------Final_Reshape_Counts---------------------------
2694 // This class defines counters to help identify when a method
2695 // may/must be executed using hardware with only 24-bit precision.
2696 struct Final_Reshape_Counts : public StackObj {
2697 int _call_count; // count non-inlined 'common' calls
2698 int _float_count; // count float ops requiring 24-bit precision
2699 int _double_count; // count double ops requiring more precision
2700 int _java_call_count; // count non-inlined 'java' calls
2701 int _inner_loop_count; // count loops which need alignment
2702 VectorSet _visited; // Visitation flags
2703 Node_List _tests; // Set of IfNodes & PCTableNodes
2704
Final_Reshape_CountsFinal_Reshape_Counts2705 Final_Reshape_Counts() :
2706 _call_count(0), _float_count(0), _double_count(0),
2707 _java_call_count(0), _inner_loop_count(0),
2708 _visited( Thread::current()->resource_area() ) { }
2709
inc_call_countFinal_Reshape_Counts2710 void inc_call_count () { _call_count ++; }
inc_float_countFinal_Reshape_Counts2711 void inc_float_count () { _float_count ++; }
inc_double_countFinal_Reshape_Counts2712 void inc_double_count() { _double_count++; }
inc_java_call_countFinal_Reshape_Counts2713 void inc_java_call_count() { _java_call_count++; }
inc_inner_loop_countFinal_Reshape_Counts2714 void inc_inner_loop_count() { _inner_loop_count++; }
2715
get_call_countFinal_Reshape_Counts2716 int get_call_count () const { return _call_count ; }
get_float_countFinal_Reshape_Counts2717 int get_float_count () const { return _float_count ; }
get_double_countFinal_Reshape_Counts2718 int get_double_count() const { return _double_count; }
get_java_call_countFinal_Reshape_Counts2719 int get_java_call_count() const { return _java_call_count; }
get_inner_loop_countFinal_Reshape_Counts2720 int get_inner_loop_count() const { return _inner_loop_count; }
2721 };
2722
2723 #ifdef ASSERT
oop_offset_is_sane(const TypeInstPtr * tp)2724 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2725 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2726 // Make sure the offset goes inside the instance layout.
2727 return k->contains_field_offset(tp->offset());
2728 // Note that OffsetBot and OffsetTop are very negative.
2729 }
2730 #endif
2731
2732 // Eliminate trivially redundant StoreCMs and accumulate their
2733 // precedence edges.
eliminate_redundant_card_marks(Node * n)2734 void Compile::eliminate_redundant_card_marks(Node* n) {
2735 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2736 if (n->in(MemNode::Address)->outcnt() > 1) {
2737 // There are multiple users of the same address so it might be
2738 // possible to eliminate some of the StoreCMs
2739 Node* mem = n->in(MemNode::Memory);
2740 Node* adr = n->in(MemNode::Address);
2741 Node* val = n->in(MemNode::ValueIn);
2742 Node* prev = n;
2743 bool done = false;
2744 // Walk the chain of StoreCMs eliminating ones that match. As
2745 // long as it's a chain of single users then the optimization is
2746 // safe. Eliminating partially redundant StoreCMs would require
2747 // cloning copies down the other paths.
2748 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2749 if (adr == mem->in(MemNode::Address) &&
2750 val == mem->in(MemNode::ValueIn)) {
2751 // redundant StoreCM
2752 if (mem->req() > MemNode::OopStore) {
2753 // Hasn't been processed by this code yet.
2754 n->add_prec(mem->in(MemNode::OopStore));
2755 } else {
2756 // Already converted to precedence edge
2757 for (uint i = mem->req(); i < mem->len(); i++) {
2758 // Accumulate any precedence edges
2759 if (mem->in(i) != NULL) {
2760 n->add_prec(mem->in(i));
2761 }
2762 }
2763 // Everything above this point has been processed.
2764 done = true;
2765 }
2766 // Eliminate the previous StoreCM
2767 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2768 assert(mem->outcnt() == 0, "should be dead");
2769 mem->disconnect_inputs(NULL, this);
2770 } else {
2771 prev = mem;
2772 }
2773 mem = prev->in(MemNode::Memory);
2774 }
2775 }
2776 }
2777
2778 //------------------------------final_graph_reshaping_impl----------------------
2779 // Implement items 1-5 from final_graph_reshaping below.
final_graph_reshaping_impl(Node * n,Final_Reshape_Counts & frc)2780 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2781
2782 if ( n->outcnt() == 0 ) return; // dead node
2783 uint nop = n->Opcode();
2784
2785 // Check for 2-input instruction with "last use" on right input.
2786 // Swap to left input. Implements item (2).
2787 if( n->req() == 3 && // two-input instruction
2788 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2789 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2790 n->in(2)->outcnt() == 1 &&// right use IS a last use
2791 !n->in(2)->is_Con() ) { // right use is not a constant
2792 // Check for commutative opcode
2793 switch( nop ) {
2794 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2795 case Op_MaxI: case Op_MinI:
2796 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2797 case Op_AndL: case Op_XorL: case Op_OrL:
2798 case Op_AndI: case Op_XorI: case Op_OrI: {
2799 // Move "last use" input to left by swapping inputs
2800 n->swap_edges(1, 2);
2801 break;
2802 }
2803 default:
2804 break;
2805 }
2806 }
2807
2808 #ifdef ASSERT
2809 if( n->is_Mem() ) {
2810 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2811 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2812 // oop will be recorded in oop map if load crosses safepoint
2813 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2814 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2815 "raw memory operations should have control edge");
2816 }
2817 if (n->is_MemBar()) {
2818 MemBarNode* mb = n->as_MemBar();
2819 if (mb->trailing_store() || mb->trailing_load_store()) {
2820 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2821 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2822 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2823 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2824 } else if (mb->leading()) {
2825 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2826 }
2827 }
2828 #endif
2829 // Count FPU ops and common calls, implements item (3)
2830 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2831 if (!gc_handled) {
2832 final_graph_reshaping_main_switch(n, frc, nop);
2833 }
2834
2835 // Collect CFG split points
2836 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2837 frc._tests.push(n);
2838 }
2839 }
2840
final_graph_reshaping_main_switch(Node * n,Final_Reshape_Counts & frc,uint nop)2841 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2842 switch( nop ) {
2843 // Count all float operations that may use FPU
2844 case Op_AddF:
2845 case Op_SubF:
2846 case Op_MulF:
2847 case Op_DivF:
2848 case Op_NegF:
2849 case Op_ModF:
2850 case Op_ConvI2F:
2851 case Op_ConF:
2852 case Op_CmpF:
2853 case Op_CmpF3:
2854 // case Op_ConvL2F: // longs are split into 32-bit halves
2855 frc.inc_float_count();
2856 break;
2857
2858 case Op_ConvF2D:
2859 case Op_ConvD2F:
2860 frc.inc_float_count();
2861 frc.inc_double_count();
2862 break;
2863
2864 // Count all double operations that may use FPU
2865 case Op_AddD:
2866 case Op_SubD:
2867 case Op_MulD:
2868 case Op_DivD:
2869 case Op_NegD:
2870 case Op_ModD:
2871 case Op_ConvI2D:
2872 case Op_ConvD2I:
2873 // case Op_ConvL2D: // handled by leaf call
2874 // case Op_ConvD2L: // handled by leaf call
2875 case Op_ConD:
2876 case Op_CmpD:
2877 case Op_CmpD3:
2878 frc.inc_double_count();
2879 break;
2880 case Op_Opaque1: // Remove Opaque Nodes before matching
2881 case Op_Opaque2: // Remove Opaque Nodes before matching
2882 case Op_Opaque3:
2883 n->subsume_by(n->in(1), this);
2884 break;
2885 case Op_CallStaticJava:
2886 case Op_CallJava:
2887 case Op_CallDynamicJava:
2888 frc.inc_java_call_count(); // Count java call site;
2889 case Op_CallRuntime:
2890 case Op_CallLeaf:
2891 case Op_CallLeafNoFP: {
2892 assert (n->is_Call(), "");
2893 CallNode *call = n->as_Call();
2894 // Count call sites where the FP mode bit would have to be flipped.
2895 // Do not count uncommon runtime calls:
2896 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2897 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2898 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2899 frc.inc_call_count(); // Count the call site
2900 } else { // See if uncommon argument is shared
2901 Node *n = call->in(TypeFunc::Parms);
2902 int nop = n->Opcode();
2903 // Clone shared simple arguments to uncommon calls, item (1).
2904 if (n->outcnt() > 1 &&
2905 !n->is_Proj() &&
2906 nop != Op_CreateEx &&
2907 nop != Op_CheckCastPP &&
2908 nop != Op_DecodeN &&
2909 nop != Op_DecodeNKlass &&
2910 !n->is_Mem() &&
2911 !n->is_Phi()) {
2912 Node *x = n->clone();
2913 call->set_req(TypeFunc::Parms, x);
2914 }
2915 }
2916 break;
2917 }
2918
2919 case Op_StoreD:
2920 case Op_LoadD:
2921 case Op_LoadD_unaligned:
2922 frc.inc_double_count();
2923 goto handle_mem;
2924 case Op_StoreF:
2925 case Op_LoadF:
2926 frc.inc_float_count();
2927 goto handle_mem;
2928
2929 case Op_StoreCM:
2930 {
2931 // Convert OopStore dependence into precedence edge
2932 Node* prec = n->in(MemNode::OopStore);
2933 n->del_req(MemNode::OopStore);
2934 n->add_prec(prec);
2935 eliminate_redundant_card_marks(n);
2936 }
2937
2938 // fall through
2939
2940 case Op_StoreB:
2941 case Op_StoreC:
2942 case Op_StorePConditional:
2943 case Op_StoreI:
2944 case Op_StoreL:
2945 case Op_StoreIConditional:
2946 case Op_StoreLConditional:
2947 case Op_CompareAndSwapB:
2948 case Op_CompareAndSwapS:
2949 case Op_CompareAndSwapI:
2950 case Op_CompareAndSwapL:
2951 case Op_CompareAndSwapP:
2952 case Op_CompareAndSwapN:
2953 case Op_WeakCompareAndSwapB:
2954 case Op_WeakCompareAndSwapS:
2955 case Op_WeakCompareAndSwapI:
2956 case Op_WeakCompareAndSwapL:
2957 case Op_WeakCompareAndSwapP:
2958 case Op_WeakCompareAndSwapN:
2959 case Op_CompareAndExchangeB:
2960 case Op_CompareAndExchangeS:
2961 case Op_CompareAndExchangeI:
2962 case Op_CompareAndExchangeL:
2963 case Op_CompareAndExchangeP:
2964 case Op_CompareAndExchangeN:
2965 case Op_GetAndAddS:
2966 case Op_GetAndAddB:
2967 case Op_GetAndAddI:
2968 case Op_GetAndAddL:
2969 case Op_GetAndSetS:
2970 case Op_GetAndSetB:
2971 case Op_GetAndSetI:
2972 case Op_GetAndSetL:
2973 case Op_GetAndSetP:
2974 case Op_GetAndSetN:
2975 case Op_StoreP:
2976 case Op_StoreN:
2977 case Op_StoreNKlass:
2978 case Op_LoadB:
2979 case Op_LoadUB:
2980 case Op_LoadUS:
2981 case Op_LoadI:
2982 case Op_LoadKlass:
2983 case Op_LoadNKlass:
2984 case Op_LoadL:
2985 case Op_LoadL_unaligned:
2986 case Op_LoadPLocked:
2987 case Op_LoadP:
2988 case Op_LoadN:
2989 case Op_LoadRange:
2990 case Op_LoadS: {
2991 handle_mem:
2992 #ifdef ASSERT
2993 if( VerifyOptoOopOffsets ) {
2994 MemNode* mem = n->as_Mem();
2995 // Check to see if address types have grounded out somehow.
2996 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2997 assert( !tp || oop_offset_is_sane(tp), "" );
2998 }
2999 #endif
3000 break;
3001 }
3002
3003 case Op_AddP: { // Assert sane base pointers
3004 Node *addp = n->in(AddPNode::Address);
3005 assert( !addp->is_AddP() ||
3006 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3007 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3008 "Base pointers must match (addp %u)", addp->_idx );
3009 #ifdef _LP64
3010 if ((UseCompressedOops || UseCompressedClassPointers) &&
3011 addp->Opcode() == Op_ConP &&
3012 addp == n->in(AddPNode::Base) &&
3013 n->in(AddPNode::Offset)->is_Con()) {
3014 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3015 // on the platform and on the compressed oops mode.
3016 // Use addressing with narrow klass to load with offset on x86.
3017 // Some platforms can use the constant pool to load ConP.
3018 // Do this transformation here since IGVN will convert ConN back to ConP.
3019 const Type* t = addp->bottom_type();
3020 bool is_oop = t->isa_oopptr() != NULL;
3021 bool is_klass = t->isa_klassptr() != NULL;
3022
3023 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3024 (is_klass && Matcher::const_klass_prefer_decode())) {
3025 Node* nn = NULL;
3026
3027 int op = is_oop ? Op_ConN : Op_ConNKlass;
3028
3029 // Look for existing ConN node of the same exact type.
3030 Node* r = root();
3031 uint cnt = r->outcnt();
3032 for (uint i = 0; i < cnt; i++) {
3033 Node* m = r->raw_out(i);
3034 if (m!= NULL && m->Opcode() == op &&
3035 m->bottom_type()->make_ptr() == t) {
3036 nn = m;
3037 break;
3038 }
3039 }
3040 if (nn != NULL) {
3041 // Decode a narrow oop to match address
3042 // [R12 + narrow_oop_reg<<3 + offset]
3043 if (is_oop) {
3044 nn = new DecodeNNode(nn, t);
3045 } else {
3046 nn = new DecodeNKlassNode(nn, t);
3047 }
3048 // Check for succeeding AddP which uses the same Base.
3049 // Otherwise we will run into the assertion above when visiting that guy.
3050 for (uint i = 0; i < n->outcnt(); ++i) {
3051 Node *out_i = n->raw_out(i);
3052 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3053 out_i->set_req(AddPNode::Base, nn);
3054 #ifdef ASSERT
3055 for (uint j = 0; j < out_i->outcnt(); ++j) {
3056 Node *out_j = out_i->raw_out(j);
3057 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3058 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3059 }
3060 #endif
3061 }
3062 }
3063 n->set_req(AddPNode::Base, nn);
3064 n->set_req(AddPNode::Address, nn);
3065 if (addp->outcnt() == 0) {
3066 addp->disconnect_inputs(NULL, this);
3067 }
3068 }
3069 }
3070 }
3071 #endif
3072 // platform dependent reshaping of the address expression
3073 reshape_address(n->as_AddP());
3074 break;
3075 }
3076
3077 case Op_CastPP: {
3078 // Remove CastPP nodes to gain more freedom during scheduling but
3079 // keep the dependency they encode as control or precedence edges
3080 // (if control is set already) on memory operations. Some CastPP
3081 // nodes don't have a control (don't carry a dependency): skip
3082 // those.
3083 if (n->in(0) != NULL) {
3084 ResourceMark rm;
3085 Unique_Node_List wq;
3086 wq.push(n);
3087 for (uint next = 0; next < wq.size(); ++next) {
3088 Node *m = wq.at(next);
3089 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3090 Node* use = m->fast_out(i);
3091 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3092 use->ensure_control_or_add_prec(n->in(0));
3093 } else {
3094 switch(use->Opcode()) {
3095 case Op_AddP:
3096 case Op_DecodeN:
3097 case Op_DecodeNKlass:
3098 case Op_CheckCastPP:
3099 case Op_CastPP:
3100 wq.push(use);
3101 break;
3102 }
3103 }
3104 }
3105 }
3106 }
3107 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3108 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3109 Node* in1 = n->in(1);
3110 const Type* t = n->bottom_type();
3111 Node* new_in1 = in1->clone();
3112 new_in1->as_DecodeN()->set_type(t);
3113
3114 if (!Matcher::narrow_oop_use_complex_address()) {
3115 //
3116 // x86, ARM and friends can handle 2 adds in addressing mode
3117 // and Matcher can fold a DecodeN node into address by using
3118 // a narrow oop directly and do implicit NULL check in address:
3119 //
3120 // [R12 + narrow_oop_reg<<3 + offset]
3121 // NullCheck narrow_oop_reg
3122 //
3123 // On other platforms (Sparc) we have to keep new DecodeN node and
3124 // use it to do implicit NULL check in address:
3125 //
3126 // decode_not_null narrow_oop_reg, base_reg
3127 // [base_reg + offset]
3128 // NullCheck base_reg
3129 //
3130 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3131 // to keep the information to which NULL check the new DecodeN node
3132 // corresponds to use it as value in implicit_null_check().
3133 //
3134 new_in1->set_req(0, n->in(0));
3135 }
3136
3137 n->subsume_by(new_in1, this);
3138 if (in1->outcnt() == 0) {
3139 in1->disconnect_inputs(NULL, this);
3140 }
3141 } else {
3142 n->subsume_by(n->in(1), this);
3143 if (n->outcnt() == 0) {
3144 n->disconnect_inputs(NULL, this);
3145 }
3146 }
3147 break;
3148 }
3149 #ifdef _LP64
3150 case Op_CmpP:
3151 // Do this transformation here to preserve CmpPNode::sub() and
3152 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3153 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3154 Node* in1 = n->in(1);
3155 Node* in2 = n->in(2);
3156 if (!in1->is_DecodeNarrowPtr()) {
3157 in2 = in1;
3158 in1 = n->in(2);
3159 }
3160 assert(in1->is_DecodeNarrowPtr(), "sanity");
3161
3162 Node* new_in2 = NULL;
3163 if (in2->is_DecodeNarrowPtr()) {
3164 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3165 new_in2 = in2->in(1);
3166 } else if (in2->Opcode() == Op_ConP) {
3167 const Type* t = in2->bottom_type();
3168 if (t == TypePtr::NULL_PTR) {
3169 assert(in1->is_DecodeN(), "compare klass to null?");
3170 // Don't convert CmpP null check into CmpN if compressed
3171 // oops implicit null check is not generated.
3172 // This will allow to generate normal oop implicit null check.
3173 if (Matcher::gen_narrow_oop_implicit_null_checks())
3174 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3175 //
3176 // This transformation together with CastPP transformation above
3177 // will generated code for implicit NULL checks for compressed oops.
3178 //
3179 // The original code after Optimize()
3180 //
3181 // LoadN memory, narrow_oop_reg
3182 // decode narrow_oop_reg, base_reg
3183 // CmpP base_reg, NULL
3184 // CastPP base_reg // NotNull
3185 // Load [base_reg + offset], val_reg
3186 //
3187 // after these transformations will be
3188 //
3189 // LoadN memory, narrow_oop_reg
3190 // CmpN narrow_oop_reg, NULL
3191 // decode_not_null narrow_oop_reg, base_reg
3192 // Load [base_reg + offset], val_reg
3193 //
3194 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3195 // since narrow oops can be used in debug info now (see the code in
3196 // final_graph_reshaping_walk()).
3197 //
3198 // At the end the code will be matched to
3199 // on x86:
3200 //
3201 // Load_narrow_oop memory, narrow_oop_reg
3202 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3203 // NullCheck narrow_oop_reg
3204 //
3205 // and on sparc:
3206 //
3207 // Load_narrow_oop memory, narrow_oop_reg
3208 // decode_not_null narrow_oop_reg, base_reg
3209 // Load [base_reg + offset], val_reg
3210 // NullCheck base_reg
3211 //
3212 } else if (t->isa_oopptr()) {
3213 new_in2 = ConNode::make(t->make_narrowoop());
3214 } else if (t->isa_klassptr()) {
3215 new_in2 = ConNode::make(t->make_narrowklass());
3216 }
3217 }
3218 if (new_in2 != NULL) {
3219 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3220 n->subsume_by(cmpN, this);
3221 if (in1->outcnt() == 0) {
3222 in1->disconnect_inputs(NULL, this);
3223 }
3224 if (in2->outcnt() == 0) {
3225 in2->disconnect_inputs(NULL, this);
3226 }
3227 }
3228 }
3229 break;
3230
3231 case Op_DecodeN:
3232 case Op_DecodeNKlass:
3233 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3234 // DecodeN could be pinned when it can't be fold into
3235 // an address expression, see the code for Op_CastPP above.
3236 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3237 break;
3238
3239 case Op_EncodeP:
3240 case Op_EncodePKlass: {
3241 Node* in1 = n->in(1);
3242 if (in1->is_DecodeNarrowPtr()) {
3243 n->subsume_by(in1->in(1), this);
3244 } else if (in1->Opcode() == Op_ConP) {
3245 const Type* t = in1->bottom_type();
3246 if (t == TypePtr::NULL_PTR) {
3247 assert(t->isa_oopptr(), "null klass?");
3248 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3249 } else if (t->isa_oopptr()) {
3250 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3251 } else if (t->isa_klassptr()) {
3252 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3253 }
3254 }
3255 if (in1->outcnt() == 0) {
3256 in1->disconnect_inputs(NULL, this);
3257 }
3258 break;
3259 }
3260
3261 case Op_Proj: {
3262 if (OptimizeStringConcat) {
3263 ProjNode* p = n->as_Proj();
3264 if (p->_is_io_use) {
3265 // Separate projections were used for the exception path which
3266 // are normally removed by a late inline. If it wasn't inlined
3267 // then they will hang around and should just be replaced with
3268 // the original one.
3269 Node* proj = NULL;
3270 // Replace with just one
3271 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3272 Node *use = i.get();
3273 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3274 proj = use;
3275 break;
3276 }
3277 }
3278 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3279 if (proj != NULL) {
3280 p->subsume_by(proj, this);
3281 }
3282 }
3283 }
3284 break;
3285 }
3286
3287 case Op_Phi:
3288 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3289 // The EncodeP optimization may create Phi with the same edges
3290 // for all paths. It is not handled well by Register Allocator.
3291 Node* unique_in = n->in(1);
3292 assert(unique_in != NULL, "");
3293 uint cnt = n->req();
3294 for (uint i = 2; i < cnt; i++) {
3295 Node* m = n->in(i);
3296 assert(m != NULL, "");
3297 if (unique_in != m)
3298 unique_in = NULL;
3299 }
3300 if (unique_in != NULL) {
3301 n->subsume_by(unique_in, this);
3302 }
3303 }
3304 break;
3305
3306 #endif
3307
3308 #ifdef ASSERT
3309 case Op_CastII:
3310 // Verify that all range check dependent CastII nodes were removed.
3311 if (n->isa_CastII()->has_range_check()) {
3312 n->dump(3);
3313 assert(false, "Range check dependent CastII node was not removed");
3314 }
3315 break;
3316 #endif
3317
3318 case Op_ModI:
3319 if (UseDivMod) {
3320 // Check if a%b and a/b both exist
3321 Node* d = n->find_similar(Op_DivI);
3322 if (d) {
3323 // Replace them with a fused divmod if supported
3324 if (Matcher::has_match_rule(Op_DivModI)) {
3325 DivModINode* divmod = DivModINode::make(n);
3326 d->subsume_by(divmod->div_proj(), this);
3327 n->subsume_by(divmod->mod_proj(), this);
3328 } else {
3329 // replace a%b with a-((a/b)*b)
3330 Node* mult = new MulINode(d, d->in(2));
3331 Node* sub = new SubINode(d->in(1), mult);
3332 n->subsume_by(sub, this);
3333 }
3334 }
3335 }
3336 break;
3337
3338 case Op_ModL:
3339 if (UseDivMod) {
3340 // Check if a%b and a/b both exist
3341 Node* d = n->find_similar(Op_DivL);
3342 if (d) {
3343 // Replace them with a fused divmod if supported
3344 if (Matcher::has_match_rule(Op_DivModL)) {
3345 DivModLNode* divmod = DivModLNode::make(n);
3346 d->subsume_by(divmod->div_proj(), this);
3347 n->subsume_by(divmod->mod_proj(), this);
3348 } else {
3349 // replace a%b with a-((a/b)*b)
3350 Node* mult = new MulLNode(d, d->in(2));
3351 Node* sub = new SubLNode(d->in(1), mult);
3352 n->subsume_by(sub, this);
3353 }
3354 }
3355 }
3356 break;
3357
3358 case Op_LoadVector:
3359 case Op_StoreVector:
3360 break;
3361
3362 case Op_AddReductionVI:
3363 case Op_AddReductionVL:
3364 case Op_AddReductionVF:
3365 case Op_AddReductionVD:
3366 case Op_MulReductionVI:
3367 case Op_MulReductionVL:
3368 case Op_MulReductionVF:
3369 case Op_MulReductionVD:
3370 case Op_MinReductionV:
3371 case Op_MaxReductionV:
3372 break;
3373
3374 case Op_PackB:
3375 case Op_PackS:
3376 case Op_PackI:
3377 case Op_PackF:
3378 case Op_PackL:
3379 case Op_PackD:
3380 if (n->req()-1 > 2) {
3381 // Replace many operand PackNodes with a binary tree for matching
3382 PackNode* p = (PackNode*) n;
3383 Node* btp = p->binary_tree_pack(1, n->req());
3384 n->subsume_by(btp, this);
3385 }
3386 break;
3387 case Op_Loop:
3388 case Op_CountedLoop:
3389 case Op_OuterStripMinedLoop:
3390 if (n->as_Loop()->is_inner_loop()) {
3391 frc.inc_inner_loop_count();
3392 }
3393 n->as_Loop()->verify_strip_mined(0);
3394 break;
3395 case Op_LShiftI:
3396 case Op_RShiftI:
3397 case Op_URShiftI:
3398 case Op_LShiftL:
3399 case Op_RShiftL:
3400 case Op_URShiftL:
3401 if (Matcher::need_masked_shift_count) {
3402 // The cpu's shift instructions don't restrict the count to the
3403 // lower 5/6 bits. We need to do the masking ourselves.
3404 Node* in2 = n->in(2);
3405 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3406 const TypeInt* t = in2->find_int_type();
3407 if (t != NULL && t->is_con()) {
3408 juint shift = t->get_con();
3409 if (shift > mask) { // Unsigned cmp
3410 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3411 }
3412 } else {
3413 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3414 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3415 n->set_req(2, shift);
3416 }
3417 }
3418 if (in2->outcnt() == 0) { // Remove dead node
3419 in2->disconnect_inputs(NULL, this);
3420 }
3421 }
3422 break;
3423 case Op_MemBarStoreStore:
3424 case Op_MemBarRelease:
3425 // Break the link with AllocateNode: it is no longer useful and
3426 // confuses register allocation.
3427 if (n->req() > MemBarNode::Precedent) {
3428 n->set_req(MemBarNode::Precedent, top());
3429 }
3430 break;
3431 case Op_MemBarAcquire: {
3432 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3433 // At parse time, the trailing MemBarAcquire for a volatile load
3434 // is created with an edge to the load. After optimizations,
3435 // that input may be a chain of Phis. If those phis have no
3436 // other use, then the MemBarAcquire keeps them alive and
3437 // register allocation can be confused.
3438 ResourceMark rm;
3439 Unique_Node_List wq;
3440 wq.push(n->in(MemBarNode::Precedent));
3441 n->set_req(MemBarNode::Precedent, top());
3442 while (wq.size() > 0) {
3443 Node* m = wq.pop();
3444 if (m->outcnt() == 0) {
3445 for (uint j = 0; j < m->req(); j++) {
3446 Node* in = m->in(j);
3447 if (in != NULL) {
3448 wq.push(in);
3449 }
3450 }
3451 m->disconnect_inputs(NULL, this);
3452 }
3453 }
3454 }
3455 break;
3456 }
3457 case Op_RangeCheck: {
3458 RangeCheckNode* rc = n->as_RangeCheck();
3459 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3460 n->subsume_by(iff, this);
3461 frc._tests.push(iff);
3462 break;
3463 }
3464 case Op_ConvI2L: {
3465 if (!Matcher::convi2l_type_required) {
3466 // Code generation on some platforms doesn't need accurate
3467 // ConvI2L types. Widening the type can help remove redundant
3468 // address computations.
3469 n->as_Type()->set_type(TypeLong::INT);
3470 ResourceMark rm;
3471 Unique_Node_List wq;
3472 wq.push(n);
3473 for (uint next = 0; next < wq.size(); next++) {
3474 Node *m = wq.at(next);
3475
3476 for(;;) {
3477 // Loop over all nodes with identical inputs edges as m
3478 Node* k = m->find_similar(m->Opcode());
3479 if (k == NULL) {
3480 break;
3481 }
3482 // Push their uses so we get a chance to remove node made
3483 // redundant
3484 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3485 Node* u = k->fast_out(i);
3486 if (u->Opcode() == Op_LShiftL ||
3487 u->Opcode() == Op_AddL ||
3488 u->Opcode() == Op_SubL ||
3489 u->Opcode() == Op_AddP) {
3490 wq.push(u);
3491 }
3492 }
3493 // Replace all nodes with identical edges as m with m
3494 k->subsume_by(m, this);
3495 }
3496 }
3497 }
3498 break;
3499 }
3500 case Op_CmpUL: {
3501 if (!Matcher::has_match_rule(Op_CmpUL)) {
3502 // No support for unsigned long comparisons
3503 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3504 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3505 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3506 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3507 Node* andl = new AndLNode(orl, remove_sign_mask);
3508 Node* cmp = new CmpLNode(andl, n->in(2));
3509 n->subsume_by(cmp, this);
3510 }
3511 break;
3512 }
3513 default:
3514 assert(!n->is_Call(), "");
3515 assert(!n->is_Mem(), "");
3516 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3517 break;
3518 }
3519 }
3520
3521 //------------------------------final_graph_reshaping_walk---------------------
3522 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3523 // 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)3524 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3525 ResourceArea *area = Thread::current()->resource_area();
3526 Unique_Node_List sfpt(area);
3527
3528 frc._visited.set(root->_idx); // first, mark node as visited
3529 uint cnt = root->req();
3530 Node *n = root;
3531 uint i = 0;
3532 while (true) {
3533 if (i < cnt) {
3534 // Place all non-visited non-null inputs onto stack
3535 Node* m = n->in(i);
3536 ++i;
3537 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3538 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3539 // compute worst case interpreter size in case of a deoptimization
3540 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3541
3542 sfpt.push(m);
3543 }
3544 cnt = m->req();
3545 nstack.push(n, i); // put on stack parent and next input's index
3546 n = m;
3547 i = 0;
3548 }
3549 } else {
3550 // Now do post-visit work
3551 final_graph_reshaping_impl( n, frc );
3552 if (nstack.is_empty())
3553 break; // finished
3554 n = nstack.node(); // Get node from stack
3555 cnt = n->req();
3556 i = nstack.index();
3557 nstack.pop(); // Shift to the next node on stack
3558 }
3559 }
3560
3561 // Skip next transformation if compressed oops are not used.
3562 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3563 (!UseCompressedOops && !UseCompressedClassPointers))
3564 return;
3565
3566 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3567 // It could be done for an uncommon traps or any safepoints/calls
3568 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3569 while (sfpt.size() > 0) {
3570 n = sfpt.pop();
3571 JVMState *jvms = n->as_SafePoint()->jvms();
3572 assert(jvms != NULL, "sanity");
3573 int start = jvms->debug_start();
3574 int end = n->req();
3575 bool is_uncommon = (n->is_CallStaticJava() &&
3576 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3577 for (int j = start; j < end; j++) {
3578 Node* in = n->in(j);
3579 if (in->is_DecodeNarrowPtr()) {
3580 bool safe_to_skip = true;
3581 if (!is_uncommon ) {
3582 // Is it safe to skip?
3583 for (uint i = 0; i < in->outcnt(); i++) {
3584 Node* u = in->raw_out(i);
3585 if (!u->is_SafePoint() ||
3586 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3587 safe_to_skip = false;
3588 }
3589 }
3590 }
3591 if (safe_to_skip) {
3592 n->set_req(j, in->in(1));
3593 }
3594 if (in->outcnt() == 0) {
3595 in->disconnect_inputs(NULL, this);
3596 }
3597 }
3598 }
3599 }
3600 }
3601
3602 //------------------------------final_graph_reshaping--------------------------
3603 // Final Graph Reshaping.
3604 //
3605 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3606 // and not commoned up and forced early. Must come after regular
3607 // optimizations to avoid GVN undoing the cloning. Clone constant
3608 // inputs to Loop Phis; these will be split by the allocator anyways.
3609 // Remove Opaque nodes.
3610 // (2) Move last-uses by commutative operations to the left input to encourage
3611 // Intel update-in-place two-address operations and better register usage
3612 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3613 // calls canonicalizing them back.
3614 // (3) Count the number of double-precision FP ops, single-precision FP ops
3615 // and call sites. On Intel, we can get correct rounding either by
3616 // forcing singles to memory (requires extra stores and loads after each
3617 // FP bytecode) or we can set a rounding mode bit (requires setting and
3618 // clearing the mode bit around call sites). The mode bit is only used
3619 // if the relative frequency of single FP ops to calls is low enough.
3620 // This is a key transform for SPEC mpeg_audio.
3621 // (4) Detect infinite loops; blobs of code reachable from above but not
3622 // below. Several of the Code_Gen algorithms fail on such code shapes,
3623 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3624 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3625 // Detection is by looking for IfNodes where only 1 projection is
3626 // reachable from below or CatchNodes missing some targets.
3627 // (5) Assert for insane oop offsets in debug mode.
3628
final_graph_reshaping()3629 bool Compile::final_graph_reshaping() {
3630 // an infinite loop may have been eliminated by the optimizer,
3631 // in which case the graph will be empty.
3632 if (root()->req() == 1) {
3633 record_method_not_compilable("trivial infinite loop");
3634 return true;
3635 }
3636
3637 // Expensive nodes have their control input set to prevent the GVN
3638 // from freely commoning them. There's no GVN beyond this point so
3639 // no need to keep the control input. We want the expensive nodes to
3640 // be freely moved to the least frequent code path by gcm.
3641 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3642 for (int i = 0; i < expensive_count(); i++) {
3643 _expensive_nodes->at(i)->set_req(0, NULL);
3644 }
3645
3646 Final_Reshape_Counts frc;
3647
3648 // Visit everybody reachable!
3649 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3650 Node_Stack nstack(live_nodes() >> 1);
3651 final_graph_reshaping_walk(nstack, root(), frc);
3652
3653 // Check for unreachable (from below) code (i.e., infinite loops).
3654 for( uint i = 0; i < frc._tests.size(); i++ ) {
3655 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3656 // Get number of CFG targets.
3657 // Note that PCTables include exception targets after calls.
3658 uint required_outcnt = n->required_outcnt();
3659 if (n->outcnt() != required_outcnt) {
3660 // Check for a few special cases. Rethrow Nodes never take the
3661 // 'fall-thru' path, so expected kids is 1 less.
3662 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3663 if (n->in(0)->in(0)->is_Call()) {
3664 CallNode *call = n->in(0)->in(0)->as_Call();
3665 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3666 required_outcnt--; // Rethrow always has 1 less kid
3667 } else if (call->req() > TypeFunc::Parms &&
3668 call->is_CallDynamicJava()) {
3669 // Check for null receiver. In such case, the optimizer has
3670 // detected that the virtual call will always result in a null
3671 // pointer exception. The fall-through projection of this CatchNode
3672 // will not be populated.
3673 Node *arg0 = call->in(TypeFunc::Parms);
3674 if (arg0->is_Type() &&
3675 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3676 required_outcnt--;
3677 }
3678 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3679 call->req() > TypeFunc::Parms+1 &&
3680 call->is_CallStaticJava()) {
3681 // Check for negative array length. In such case, the optimizer has
3682 // detected that the allocation attempt will always result in an
3683 // exception. There is no fall-through projection of this CatchNode .
3684 Node *arg1 = call->in(TypeFunc::Parms+1);
3685 if (arg1->is_Type() &&
3686 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3687 required_outcnt--;
3688 }
3689 }
3690 }
3691 }
3692 // Recheck with a better notion of 'required_outcnt'
3693 if (n->outcnt() != required_outcnt) {
3694 record_method_not_compilable("malformed control flow");
3695 return true; // Not all targets reachable!
3696 }
3697 }
3698 // Check that I actually visited all kids. Unreached kids
3699 // must be infinite loops.
3700 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3701 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3702 record_method_not_compilable("infinite loop");
3703 return true; // Found unvisited kid; must be unreach
3704 }
3705
3706 // Here so verification code in final_graph_reshaping_walk()
3707 // always see an OuterStripMinedLoopEnd
3708 if (n->is_OuterStripMinedLoopEnd()) {
3709 IfNode* init_iff = n->as_If();
3710 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3711 n->subsume_by(iff, this);
3712 }
3713 }
3714
3715 // If original bytecodes contained a mixture of floats and doubles
3716 // check if the optimizer has made it homogenous, item (3).
3717 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3718 frc.get_float_count() > 32 &&
3719 frc.get_double_count() == 0 &&
3720 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3721 set_24_bit_selection_and_mode( false, true );
3722 }
3723
3724 set_java_calls(frc.get_java_call_count());
3725 set_inner_loops(frc.get_inner_loop_count());
3726
3727 // No infinite loops, no reason to bail out.
3728 return false;
3729 }
3730
3731 //-----------------------------too_many_traps----------------------------------
3732 // Report if there are too many traps at the current method and bci.
3733 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
too_many_traps(ciMethod * method,int bci,Deoptimization::DeoptReason reason)3734 bool Compile::too_many_traps(ciMethod* method,
3735 int bci,
3736 Deoptimization::DeoptReason reason) {
3737 ciMethodData* md = method->method_data();
3738 if (md->is_empty()) {
3739 // Assume the trap has not occurred, or that it occurred only
3740 // because of a transient condition during start-up in the interpreter.
3741 return false;
3742 }
3743 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3744 if (md->has_trap_at(bci, m, reason) != 0) {
3745 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3746 // Also, if there are multiple reasons, or if there is no per-BCI record,
3747 // assume the worst.
3748 if (log())
3749 log()->elem("observe trap='%s' count='%d'",
3750 Deoptimization::trap_reason_name(reason),
3751 md->trap_count(reason));
3752 return true;
3753 } else {
3754 // Ignore method/bci and see if there have been too many globally.
3755 return too_many_traps(reason, md);
3756 }
3757 }
3758
3759 // Less-accurate variant which does not require a method and bci.
too_many_traps(Deoptimization::DeoptReason reason,ciMethodData * logmd)3760 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3761 ciMethodData* logmd) {
3762 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3763 // Too many traps globally.
3764 // Note that we use cumulative trap_count, not just md->trap_count.
3765 if (log()) {
3766 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3767 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3768 Deoptimization::trap_reason_name(reason),
3769 mcount, trap_count(reason));
3770 }
3771 return true;
3772 } else {
3773 // The coast is clear.
3774 return false;
3775 }
3776 }
3777
3778 //--------------------------too_many_recompiles--------------------------------
3779 // Report if there are too many recompiles at the current method and bci.
3780 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3781 // Is not eager to return true, since this will cause the compiler to use
3782 // Action_none for a trap point, to avoid too many recompilations.
too_many_recompiles(ciMethod * method,int bci,Deoptimization::DeoptReason reason)3783 bool Compile::too_many_recompiles(ciMethod* method,
3784 int bci,
3785 Deoptimization::DeoptReason reason) {
3786 ciMethodData* md = method->method_data();
3787 if (md->is_empty()) {
3788 // Assume the trap has not occurred, or that it occurred only
3789 // because of a transient condition during start-up in the interpreter.
3790 return false;
3791 }
3792 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3793 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3794 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3795 Deoptimization::DeoptReason per_bc_reason
3796 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3797 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3798 if ((per_bc_reason == Deoptimization::Reason_none
3799 || md->has_trap_at(bci, m, reason) != 0)
3800 // The trap frequency measure we care about is the recompile count:
3801 && md->trap_recompiled_at(bci, m)
3802 && md->overflow_recompile_count() >= bc_cutoff) {
3803 // Do not emit a trap here if it has already caused recompilations.
3804 // Also, if there are multiple reasons, or if there is no per-BCI record,
3805 // assume the worst.
3806 if (log())
3807 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3808 Deoptimization::trap_reason_name(reason),
3809 md->trap_count(reason),
3810 md->overflow_recompile_count());
3811 return true;
3812 } else if (trap_count(reason) != 0
3813 && decompile_count() >= m_cutoff) {
3814 // Too many recompiles globally, and we have seen this sort of trap.
3815 // Use cumulative decompile_count, not just md->decompile_count.
3816 if (log())
3817 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3818 Deoptimization::trap_reason_name(reason),
3819 md->trap_count(reason), trap_count(reason),
3820 md->decompile_count(), decompile_count());
3821 return true;
3822 } else {
3823 // The coast is clear.
3824 return false;
3825 }
3826 }
3827
3828 // Compute when not to trap. Used by matching trap based nodes and
3829 // NullCheck optimization.
set_allowed_deopt_reasons()3830 void Compile::set_allowed_deopt_reasons() {
3831 _allowed_reasons = 0;
3832 if (is_method_compilation()) {
3833 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3834 assert(rs < BitsPerInt, "recode bit map");
3835 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3836 _allowed_reasons |= nth_bit(rs);
3837 }
3838 }
3839 }
3840 }
3841
needs_clinit_barrier(ciMethod * method,ciMethod * accessing_method)3842 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3843 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3844 }
3845
needs_clinit_barrier(ciField * field,ciMethod * accessing_method)3846 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3847 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3848 }
3849
needs_clinit_barrier(ciInstanceKlass * holder,ciMethod * accessing_method)3850 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3851 if (holder->is_initialized()) {
3852 return false;
3853 }
3854 if (holder->is_being_initialized()) {
3855 if (accessing_method->holder() == holder) {
3856 // Access inside a class. The barrier can be elided when access happens in <clinit>,
3857 // <init>, or a static method. In all those cases, there was an initialization
3858 // barrier on the holder klass passed.
3859 if (accessing_method->is_static_initializer() ||
3860 accessing_method->is_object_initializer() ||
3861 accessing_method->is_static()) {
3862 return false;
3863 }
3864 } else if (accessing_method->holder()->is_subclass_of(holder)) {
3865 // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3866 // In case of <init> or a static method, the barrier is on the subclass is not enough:
3867 // child class can become fully initialized while its parent class is still being initialized.
3868 if (accessing_method->is_static_initializer()) {
3869 return false;
3870 }
3871 }
3872 ciMethod* root = method(); // the root method of compilation
3873 if (root != accessing_method) {
3874 return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3875 }
3876 }
3877 return true;
3878 }
3879
3880 #ifndef PRODUCT
3881 //------------------------------verify_graph_edges---------------------------
3882 // Walk the Graph and verify that there is a one-to-one correspondence
3883 // between Use-Def edges and Def-Use edges in the graph.
verify_graph_edges(bool no_dead_code)3884 void Compile::verify_graph_edges(bool no_dead_code) {
3885 if (VerifyGraphEdges) {
3886 ResourceArea *area = Thread::current()->resource_area();
3887 Unique_Node_List visited(area);
3888 // Call recursive graph walk to check edges
3889 _root->verify_edges(visited);
3890 if (no_dead_code) {
3891 // Now make sure that no visited node is used by an unvisited node.
3892 bool dead_nodes = false;
3893 Unique_Node_List checked(area);
3894 while (visited.size() > 0) {
3895 Node* n = visited.pop();
3896 checked.push(n);
3897 for (uint i = 0; i < n->outcnt(); i++) {
3898 Node* use = n->raw_out(i);
3899 if (checked.member(use)) continue; // already checked
3900 if (visited.member(use)) continue; // already in the graph
3901 if (use->is_Con()) continue; // a dead ConNode is OK
3902 // At this point, we have found a dead node which is DU-reachable.
3903 if (!dead_nodes) {
3904 tty->print_cr("*** Dead nodes reachable via DU edges:");
3905 dead_nodes = true;
3906 }
3907 use->dump(2);
3908 tty->print_cr("---");
3909 checked.push(use); // No repeats; pretend it is now checked.
3910 }
3911 }
3912 assert(!dead_nodes, "using nodes must be reachable from root");
3913 }
3914 }
3915 }
3916 #endif
3917
3918 // The Compile object keeps track of failure reasons separately from the ciEnv.
3919 // This is required because there is not quite a 1-1 relation between the
3920 // ciEnv and its compilation task and the Compile object. Note that one
3921 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3922 // to backtrack and retry without subsuming loads. Other than this backtracking
3923 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3924 // by the logic in C2Compiler.
record_failure(const char * reason)3925 void Compile::record_failure(const char* reason) {
3926 if (log() != NULL) {
3927 log()->elem("failure reason='%s' phase='compile'", reason);
3928 }
3929 if (_failure_reason == NULL) {
3930 // Record the first failure reason.
3931 _failure_reason = reason;
3932 }
3933
3934 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3935 C->print_method(PHASE_FAILURE);
3936 }
3937 _root = NULL; // flush the graph, too
3938 }
3939
TracePhase(const char * name,elapsedTimer * accumulator)3940 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3941 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3942 _phase_name(name), _dolog(CITimeVerbose)
3943 {
3944 if (_dolog) {
3945 C = Compile::current();
3946 _log = C->log();
3947 } else {
3948 C = NULL;
3949 _log = NULL;
3950 }
3951 if (_log != NULL) {
3952 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3953 _log->stamp();
3954 _log->end_head();
3955 }
3956 }
3957
~TracePhase()3958 Compile::TracePhase::~TracePhase() {
3959
3960 C = Compile::current();
3961 if (_dolog) {
3962 _log = C->log();
3963 } else {
3964 _log = NULL;
3965 }
3966
3967 #ifdef ASSERT
3968 if (PrintIdealNodeCount) {
3969 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3970 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3971 }
3972
3973 if (VerifyIdealNodeCount) {
3974 Compile::current()->print_missing_nodes();
3975 }
3976 #endif
3977
3978 if (_log != NULL) {
3979 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3980 }
3981 }
3982
3983 //=============================================================================
3984 // Two Constant's are equal when the type and the value are equal.
operator ==(const Constant & other)3985 bool Compile::Constant::operator==(const Constant& other) {
3986 if (type() != other.type() ) return false;
3987 if (can_be_reused() != other.can_be_reused()) return false;
3988 // For floating point values we compare the bit pattern.
3989 switch (type()) {
3990 case T_INT:
3991 case T_FLOAT: return (_v._value.i == other._v._value.i);
3992 case T_LONG:
3993 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3994 case T_OBJECT:
3995 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3996 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3997 case T_METADATA: return (_v._metadata == other._v._metadata);
3998 default: ShouldNotReachHere(); return false;
3999 }
4000 }
4001
type_to_size_in_bytes(BasicType t)4002 static int type_to_size_in_bytes(BasicType t) {
4003 switch (t) {
4004 case T_INT: return sizeof(jint );
4005 case T_LONG: return sizeof(jlong );
4006 case T_FLOAT: return sizeof(jfloat );
4007 case T_DOUBLE: return sizeof(jdouble);
4008 case T_METADATA: return sizeof(Metadata*);
4009 // We use T_VOID as marker for jump-table entries (labels) which
4010 // need an internal word relocation.
4011 case T_VOID:
4012 case T_ADDRESS:
4013 case T_OBJECT: return sizeof(jobject);
4014 default:
4015 ShouldNotReachHere();
4016 return -1;
4017 }
4018 }
4019
qsort_comparator(Constant * a,Constant * b)4020 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4021 // sort descending
4022 if (a->freq() > b->freq()) return -1;
4023 if (a->freq() < b->freq()) return 1;
4024 return 0;
4025 }
4026
calculate_offsets_and_size()4027 void Compile::ConstantTable::calculate_offsets_and_size() {
4028 // First, sort the array by frequencies.
4029 _constants.sort(qsort_comparator);
4030
4031 #ifdef ASSERT
4032 // Make sure all jump-table entries were sorted to the end of the
4033 // array (they have a negative frequency).
4034 bool found_void = false;
4035 for (int i = 0; i < _constants.length(); i++) {
4036 Constant con = _constants.at(i);
4037 if (con.type() == T_VOID)
4038 found_void = true; // jump-tables
4039 else
4040 assert(!found_void, "wrong sorting");
4041 }
4042 #endif
4043
4044 int offset = 0;
4045 for (int i = 0; i < _constants.length(); i++) {
4046 Constant* con = _constants.adr_at(i);
4047
4048 // Align offset for type.
4049 int typesize = type_to_size_in_bytes(con->type());
4050 offset = align_up(offset, typesize);
4051 con->set_offset(offset); // set constant's offset
4052
4053 if (con->type() == T_VOID) {
4054 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4055 offset = offset + typesize * n->outcnt(); // expand jump-table
4056 } else {
4057 offset = offset + typesize;
4058 }
4059 }
4060
4061 // Align size up to the next section start (which is insts; see
4062 // CodeBuffer::align_at_start).
4063 assert(_size == -1, "already set?");
4064 _size = align_up(offset, (int)CodeEntryAlignment);
4065 }
4066
emit(CodeBuffer & cb)4067 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4068 MacroAssembler _masm(&cb);
4069 for (int i = 0; i < _constants.length(); i++) {
4070 Constant con = _constants.at(i);
4071 address constant_addr = NULL;
4072 switch (con.type()) {
4073 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4074 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4075 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4076 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4077 case T_OBJECT: {
4078 jobject obj = con.get_jobject();
4079 int oop_index = _masm.oop_recorder()->find_index(obj);
4080 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4081 break;
4082 }
4083 case T_ADDRESS: {
4084 address addr = (address) con.get_jobject();
4085 constant_addr = _masm.address_constant(addr);
4086 break;
4087 }
4088 // We use T_VOID as marker for jump-table entries (labels) which
4089 // need an internal word relocation.
4090 case T_VOID: {
4091 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4092 // Fill the jump-table with a dummy word. The real value is
4093 // filled in later in fill_jump_table.
4094 address dummy = (address) n;
4095 constant_addr = _masm.address_constant(dummy);
4096 // Expand jump-table
4097 for (uint i = 1; i < n->outcnt(); i++) {
4098 address temp_addr = _masm.address_constant(dummy + i);
4099 assert(temp_addr, "consts section too small");
4100 }
4101 break;
4102 }
4103 case T_METADATA: {
4104 Metadata* obj = con.get_metadata();
4105 int metadata_index = _masm.oop_recorder()->find_index(obj);
4106 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4107 break;
4108 }
4109 default: ShouldNotReachHere();
4110 }
4111 assert(constant_addr, "consts section too small");
4112 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4113 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4114 }
4115 }
4116
find_offset(Constant & con) const4117 int Compile::ConstantTable::find_offset(Constant& con) const {
4118 int idx = _constants.find(con);
4119 guarantee(idx != -1, "constant must be in constant table");
4120 int offset = _constants.at(idx).offset();
4121 guarantee(offset != -1, "constant table not emitted yet?");
4122 return offset;
4123 }
4124
add(Constant & con)4125 void Compile::ConstantTable::add(Constant& con) {
4126 if (con.can_be_reused()) {
4127 int idx = _constants.find(con);
4128 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4129 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4130 return;
4131 }
4132 }
4133 (void) _constants.append(con);
4134 }
4135
add(MachConstantNode * n,BasicType type,jvalue value)4136 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4137 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4138 Constant con(type, value, b->_freq);
4139 add(con);
4140 return con;
4141 }
4142
add(Metadata * metadata)4143 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4144 Constant con(metadata);
4145 add(con);
4146 return con;
4147 }
4148
add(MachConstantNode * n,MachOper * oper)4149 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4150 jvalue value;
4151 BasicType type = oper->type()->basic_type();
4152 switch (type) {
4153 case T_LONG: value.j = oper->constantL(); break;
4154 case T_FLOAT: value.f = oper->constantF(); break;
4155 case T_DOUBLE: value.d = oper->constantD(); break;
4156 case T_OBJECT:
4157 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4158 case T_METADATA: return add((Metadata*)oper->constant()); break;
4159 default: guarantee(false, "unhandled type: %s", type2name(type));
4160 }
4161 return add(n, type, value);
4162 }
4163
add_jump_table(MachConstantNode * n)4164 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4165 jvalue value;
4166 // We can use the node pointer here to identify the right jump-table
4167 // as this method is called from Compile::Fill_buffer right before
4168 // the MachNodes are emitted and the jump-table is filled (means the
4169 // MachNode pointers do not change anymore).
4170 value.l = (jobject) n;
4171 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4172 add(con);
4173 return con;
4174 }
4175
fill_jump_table(CodeBuffer & cb,MachConstantNode * n,GrowableArray<Label * > labels) const4176 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4177 // If called from Compile::scratch_emit_size do nothing.
4178 if (Compile::current()->in_scratch_emit_size()) return;
4179
4180 assert(labels.is_nonempty(), "must be");
4181 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4182
4183 // Since MachConstantNode::constant_offset() also contains
4184 // table_base_offset() we need to subtract the table_base_offset()
4185 // to get the plain offset into the constant table.
4186 int offset = n->constant_offset() - table_base_offset();
4187
4188 MacroAssembler _masm(&cb);
4189 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4190
4191 for (uint i = 0; i < n->outcnt(); i++) {
4192 address* constant_addr = &jump_table_base[i];
4193 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));
4194 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4195 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4196 }
4197 }
4198
4199 //----------------------------static_subtype_check-----------------------------
4200 // Shortcut important common cases when superklass is exact:
4201 // (0) superklass is java.lang.Object (can occur in reflective code)
4202 // (1) subklass is already limited to a subtype of superklass => always ok
4203 // (2) subklass does not overlap with superklass => always fail
4204 // (3) superklass has NO subtypes and we can check with a simple compare.
static_subtype_check(ciKlass * superk,ciKlass * subk)4205 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4206 if (StressReflectiveCode) {
4207 return SSC_full_test; // Let caller generate the general case.
4208 }
4209
4210 if (superk == env()->Object_klass()) {
4211 return SSC_always_true; // (0) this test cannot fail
4212 }
4213
4214 ciType* superelem = superk;
4215 if (superelem->is_array_klass())
4216 superelem = superelem->as_array_klass()->base_element_type();
4217
4218 if (!subk->is_interface()) { // cannot trust static interface types yet
4219 if (subk->is_subtype_of(superk)) {
4220 return SSC_always_true; // (1) false path dead; no dynamic test needed
4221 }
4222 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4223 !superk->is_subtype_of(subk)) {
4224 return SSC_always_false;
4225 }
4226 }
4227
4228 // If casting to an instance klass, it must have no subtypes
4229 if (superk->is_interface()) {
4230 // Cannot trust interfaces yet.
4231 // %%% S.B. superk->nof_implementors() == 1
4232 } else if (superelem->is_instance_klass()) {
4233 ciInstanceKlass* ik = superelem->as_instance_klass();
4234 if (!ik->has_subklass() && !ik->is_interface()) {
4235 if (!ik->is_final()) {
4236 // Add a dependency if there is a chance of a later subclass.
4237 dependencies()->assert_leaf_type(ik);
4238 }
4239 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4240 }
4241 } else {
4242 // A primitive array type has no subtypes.
4243 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4244 }
4245
4246 return SSC_full_test;
4247 }
4248
conv_I2X_index(PhaseGVN * phase,Node * idx,const TypeInt * sizetype,Node * ctrl)4249 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4250 #ifdef _LP64
4251 // The scaled index operand to AddP must be a clean 64-bit value.
4252 // Java allows a 32-bit int to be incremented to a negative
4253 // value, which appears in a 64-bit register as a large
4254 // positive number. Using that large positive number as an
4255 // operand in pointer arithmetic has bad consequences.
4256 // On the other hand, 32-bit overflow is rare, and the possibility
4257 // can often be excluded, if we annotate the ConvI2L node with
4258 // a type assertion that its value is known to be a small positive
4259 // number. (The prior range check has ensured this.)
4260 // This assertion is used by ConvI2LNode::Ideal.
4261 int index_max = max_jint - 1; // array size is max_jint, index is one less
4262 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4263 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4264 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4265 #endif
4266 return idx;
4267 }
4268
4269 // 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)4270 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4271 if (ctrl != NULL) {
4272 // Express control dependency by a CastII node with a narrow type.
4273 value = new CastIINode(value, itype, false, true /* range check dependency */);
4274 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4275 // node from floating above the range check during loop optimizations. Otherwise, the
4276 // ConvI2L node may be eliminated independently of the range check, causing the data path
4277 // to become TOP while the control path is still there (although it's unreachable).
4278 value->set_req(0, ctrl);
4279 // Save CastII node to remove it after loop optimizations.
4280 phase->C->add_range_check_cast(value);
4281 value = phase->transform(value);
4282 }
4283 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4284 return phase->transform(new ConvI2LNode(value, ltype));
4285 }
4286
print_inlining_stream_free()4287 void Compile::print_inlining_stream_free() {
4288 if (_print_inlining_stream != NULL) {
4289 _print_inlining_stream->~stringStream();
4290 _print_inlining_stream = NULL;
4291 }
4292 }
4293
4294 // The message about the current inlining is accumulated in
4295 // _print_inlining_stream and transfered into the _print_inlining_list
4296 // once we know whether inlining succeeds or not. For regular
4297 // inlining, messages are appended to the buffer pointed by
4298 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4299 // a new buffer is added after _print_inlining_idx in the list. This
4300 // way we can update the inlining message for late inlining call site
4301 // when the inlining is attempted again.
print_inlining_init()4302 void Compile::print_inlining_init() {
4303 if (print_inlining() || print_intrinsics()) {
4304 // print_inlining_init is actually called several times.
4305 print_inlining_stream_free();
4306 _print_inlining_stream = new stringStream();
4307 // Watch out: The memory initialized by the constructor call PrintInliningBuffer()
4308 // will be copied into the only initial element. The default destructor of
4309 // PrintInliningBuffer will be called when leaving the scope here. If it
4310 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss
4311 // would be destructed, too!
4312 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4313 }
4314 }
4315
print_inlining_reinit()4316 void Compile::print_inlining_reinit() {
4317 if (print_inlining() || print_intrinsics()) {
4318 print_inlining_stream_free();
4319 // Re allocate buffer when we change ResourceMark
4320 _print_inlining_stream = new stringStream();
4321 }
4322 }
4323
print_inlining_reset()4324 void Compile::print_inlining_reset() {
4325 _print_inlining_stream->reset();
4326 }
4327
print_inlining_commit()4328 void Compile::print_inlining_commit() {
4329 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4330 // Transfer the message from _print_inlining_stream to the current
4331 // _print_inlining_list buffer and clear _print_inlining_stream.
4332 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4333 print_inlining_reset();
4334 }
4335
print_inlining_push()4336 void Compile::print_inlining_push() {
4337 // Add new buffer to the _print_inlining_list at current position
4338 _print_inlining_idx++;
4339 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4340 }
4341
print_inlining_current()4342 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4343 return _print_inlining_list->at(_print_inlining_idx);
4344 }
4345
print_inlining_update(CallGenerator * cg)4346 void Compile::print_inlining_update(CallGenerator* cg) {
4347 if (print_inlining() || print_intrinsics()) {
4348 if (!cg->is_late_inline()) {
4349 if (print_inlining_current().cg() != NULL) {
4350 print_inlining_push();
4351 }
4352 print_inlining_commit();
4353 } else {
4354 if (print_inlining_current().cg() != cg &&
4355 (print_inlining_current().cg() != NULL ||
4356 print_inlining_current().ss()->size() != 0)) {
4357 print_inlining_push();
4358 }
4359 print_inlining_commit();
4360 print_inlining_current().set_cg(cg);
4361 }
4362 }
4363 }
4364
print_inlining_move_to(CallGenerator * cg)4365 void Compile::print_inlining_move_to(CallGenerator* cg) {
4366 // We resume inlining at a late inlining call site. Locate the
4367 // corresponding inlining buffer so that we can update it.
4368 if (print_inlining()) {
4369 for (int i = 0; i < _print_inlining_list->length(); i++) {
4370 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4371 _print_inlining_idx = i;
4372 return;
4373 }
4374 }
4375 ShouldNotReachHere();
4376 }
4377 }
4378
print_inlining_update_delayed(CallGenerator * cg)4379 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4380 if (print_inlining()) {
4381 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4382 assert(print_inlining_current().cg() == cg, "wrong entry");
4383 // replace message with new message
4384 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4385 print_inlining_commit();
4386 print_inlining_current().set_cg(cg);
4387 }
4388 }
4389
print_inlining_assert_ready()4390 void Compile::print_inlining_assert_ready() {
4391 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4392 }
4393
process_print_inlining()4394 void Compile::process_print_inlining() {
4395 bool do_print_inlining = print_inlining() || print_intrinsics();
4396 if (do_print_inlining || log() != NULL) {
4397 // Print inlining message for candidates that we couldn't inline
4398 // for lack of space
4399 for (int i = 0; i < _late_inlines.length(); i++) {
4400 CallGenerator* cg = _late_inlines.at(i);
4401 if (!cg->is_mh_late_inline()) {
4402 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4403 if (do_print_inlining) {
4404 cg->print_inlining_late(msg);
4405 }
4406 log_late_inline_failure(cg, msg);
4407 }
4408 }
4409 }
4410 if (do_print_inlining) {
4411 ResourceMark rm;
4412 stringStream ss;
4413 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4414 for (int i = 0; i < _print_inlining_list->length(); i++) {
4415 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4416 _print_inlining_list->at(i).freeStream();
4417 }
4418 // Reset _print_inlining_list, it only contains destructed objects.
4419 // It is on the arena, so it will be freed when the arena is reset.
4420 _print_inlining_list = NULL;
4421 // _print_inlining_stream won't be used anymore, either.
4422 print_inlining_stream_free();
4423 size_t end = ss.size();
4424 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4425 strncpy(_print_inlining_output, ss.base(), end+1);
4426 _print_inlining_output[end] = 0;
4427 }
4428 }
4429
dump_print_inlining()4430 void Compile::dump_print_inlining() {
4431 if (_print_inlining_output != NULL) {
4432 tty->print_raw(_print_inlining_output);
4433 }
4434 }
4435
log_late_inline(CallGenerator * cg)4436 void Compile::log_late_inline(CallGenerator* cg) {
4437 if (log() != NULL) {
4438 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4439 cg->unique_id());
4440 JVMState* p = cg->call_node()->jvms();
4441 while (p != NULL) {
4442 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4443 p = p->caller();
4444 }
4445 log()->tail("late_inline");
4446 }
4447 }
4448
log_late_inline_failure(CallGenerator * cg,const char * msg)4449 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4450 log_late_inline(cg);
4451 if (log() != NULL) {
4452 log()->inline_fail(msg);
4453 }
4454 }
4455
log_inline_id(CallGenerator * cg)4456 void Compile::log_inline_id(CallGenerator* cg) {
4457 if (log() != NULL) {
4458 // The LogCompilation tool needs a unique way to identify late
4459 // inline call sites. This id must be unique for this call site in
4460 // this compilation. Try to have it unique across compilations as
4461 // well because it can be convenient when grepping through the log
4462 // file.
4463 // Distinguish OSR compilations from others in case CICountOSR is
4464 // on.
4465 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4466 cg->set_unique_id(id);
4467 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4468 }
4469 }
4470
log_inline_failure(const char * msg)4471 void Compile::log_inline_failure(const char* msg) {
4472 if (C->log() != NULL) {
4473 C->log()->inline_fail(msg);
4474 }
4475 }
4476
4477
4478 // Dump inlining replay data to the stream.
4479 // Don't change thread state and acquire any locks.
dump_inline_data(outputStream * out)4480 void Compile::dump_inline_data(outputStream* out) {
4481 InlineTree* inl_tree = ilt();
4482 if (inl_tree != NULL) {
4483 out->print(" inline %d", inl_tree->count());
4484 inl_tree->dump_replay_data(out);
4485 }
4486 }
4487
cmp_expensive_nodes(Node * n1,Node * n2)4488 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4489 if (n1->Opcode() < n2->Opcode()) return -1;
4490 else if (n1->Opcode() > n2->Opcode()) return 1;
4491
4492 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4493 for (uint i = 1; i < n1->req(); i++) {
4494 if (n1->in(i) < n2->in(i)) return -1;
4495 else if (n1->in(i) > n2->in(i)) return 1;
4496 }
4497
4498 return 0;
4499 }
4500
cmp_expensive_nodes(Node ** n1p,Node ** n2p)4501 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4502 Node* n1 = *n1p;
4503 Node* n2 = *n2p;
4504
4505 return cmp_expensive_nodes(n1, n2);
4506 }
4507
sort_expensive_nodes()4508 void Compile::sort_expensive_nodes() {
4509 if (!expensive_nodes_sorted()) {
4510 _expensive_nodes->sort(cmp_expensive_nodes);
4511 }
4512 }
4513
expensive_nodes_sorted() const4514 bool Compile::expensive_nodes_sorted() const {
4515 for (int i = 1; i < _expensive_nodes->length(); i++) {
4516 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4517 return false;
4518 }
4519 }
4520 return true;
4521 }
4522
should_optimize_expensive_nodes(PhaseIterGVN & igvn)4523 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4524 if (_expensive_nodes->length() == 0) {
4525 return false;
4526 }
4527
4528 assert(OptimizeExpensiveOps, "optimization off?");
4529
4530 // Take this opportunity to remove dead nodes from the list
4531 int j = 0;
4532 for (int i = 0; i < _expensive_nodes->length(); i++) {
4533 Node* n = _expensive_nodes->at(i);
4534 if (!n->is_unreachable(igvn)) {
4535 assert(n->is_expensive(), "should be expensive");
4536 _expensive_nodes->at_put(j, n);
4537 j++;
4538 }
4539 }
4540 _expensive_nodes->trunc_to(j);
4541
4542 // Then sort the list so that similar nodes are next to each other
4543 // and check for at least two nodes of identical kind with same data
4544 // inputs.
4545 sort_expensive_nodes();
4546
4547 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4548 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4549 return true;
4550 }
4551 }
4552
4553 return false;
4554 }
4555
cleanup_expensive_nodes(PhaseIterGVN & igvn)4556 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4557 if (_expensive_nodes->length() == 0) {
4558 return;
4559 }
4560
4561 assert(OptimizeExpensiveOps, "optimization off?");
4562
4563 // Sort to bring similar nodes next to each other and clear the
4564 // control input of nodes for which there's only a single copy.
4565 sort_expensive_nodes();
4566
4567 int j = 0;
4568 int identical = 0;
4569 int i = 0;
4570 bool modified = false;
4571 for (; i < _expensive_nodes->length()-1; i++) {
4572 assert(j <= i, "can't write beyond current index");
4573 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4574 identical++;
4575 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4576 continue;
4577 }
4578 if (identical > 0) {
4579 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4580 identical = 0;
4581 } else {
4582 Node* n = _expensive_nodes->at(i);
4583 igvn.replace_input_of(n, 0, NULL);
4584 igvn.hash_insert(n);
4585 modified = true;
4586 }
4587 }
4588 if (identical > 0) {
4589 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4590 } else if (_expensive_nodes->length() >= 1) {
4591 Node* n = _expensive_nodes->at(i);
4592 igvn.replace_input_of(n, 0, NULL);
4593 igvn.hash_insert(n);
4594 modified = true;
4595 }
4596 _expensive_nodes->trunc_to(j);
4597 if (modified) {
4598 igvn.optimize();
4599 }
4600 }
4601
add_expensive_node(Node * n)4602 void Compile::add_expensive_node(Node * n) {
4603 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4604 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4605 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4606 if (OptimizeExpensiveOps) {
4607 _expensive_nodes->append(n);
4608 } else {
4609 // Clear control input and let IGVN optimize expensive nodes if
4610 // OptimizeExpensiveOps is off.
4611 n->set_req(0, NULL);
4612 }
4613 }
4614
4615 /**
4616 * Remove the speculative part of types and clean up the graph
4617 */
remove_speculative_types(PhaseIterGVN & igvn)4618 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4619 if (UseTypeSpeculation) {
4620 Unique_Node_List worklist;
4621 worklist.push(root());
4622 int modified = 0;
4623 // Go over all type nodes that carry a speculative type, drop the
4624 // speculative part of the type and enqueue the node for an igvn
4625 // which may optimize it out.
4626 for (uint next = 0; next < worklist.size(); ++next) {
4627 Node *n = worklist.at(next);
4628 if (n->is_Type()) {
4629 TypeNode* tn = n->as_Type();
4630 const Type* t = tn->type();
4631 const Type* t_no_spec = t->remove_speculative();
4632 if (t_no_spec != t) {
4633 bool in_hash = igvn.hash_delete(n);
4634 assert(in_hash, "node should be in igvn hash table");
4635 tn->set_type(t_no_spec);
4636 igvn.hash_insert(n);
4637 igvn._worklist.push(n); // give it a chance to go away
4638 modified++;
4639 }
4640 }
4641 uint max = n->len();
4642 for( uint i = 0; i < max; ++i ) {
4643 Node *m = n->in(i);
4644 if (not_a_node(m)) continue;
4645 worklist.push(m);
4646 }
4647 }
4648 // Drop the speculative part of all types in the igvn's type table
4649 igvn.remove_speculative_types();
4650 if (modified > 0) {
4651 igvn.optimize();
4652 }
4653 #ifdef ASSERT
4654 // Verify that after the IGVN is over no speculative type has resurfaced
4655 worklist.clear();
4656 worklist.push(root());
4657 for (uint next = 0; next < worklist.size(); ++next) {
4658 Node *n = worklist.at(next);
4659 const Type* t = igvn.type_or_null(n);
4660 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4661 if (n->is_Type()) {
4662 t = n->as_Type()->type();
4663 assert(t == t->remove_speculative(), "no more speculative types");
4664 }
4665 uint max = n->len();
4666 for( uint i = 0; i < max; ++i ) {
4667 Node *m = n->in(i);
4668 if (not_a_node(m)) continue;
4669 worklist.push(m);
4670 }
4671 }
4672 igvn.check_no_speculative_types();
4673 #endif
4674 }
4675 }
4676
4677 // Auxiliary method to support randomized stressing/fuzzing.
4678 //
4679 // This method can be called the arbitrary number of times, with current count
4680 // as the argument. The logic allows selecting a single candidate from the
4681 // running list of candidates as follows:
4682 // int count = 0;
4683 // Cand* selected = null;
4684 // while(cand = cand->next()) {
4685 // if (randomized_select(++count)) {
4686 // selected = cand;
4687 // }
4688 // }
4689 //
4690 // Including count equalizes the chances any candidate is "selected".
4691 // This is useful when we don't have the complete list of candidates to choose
4692 // from uniformly. In this case, we need to adjust the randomicity of the
4693 // selection, or else we will end up biasing the selection towards the latter
4694 // candidates.
4695 //
4696 // Quick back-envelope calculation shows that for the list of n candidates
4697 // the equal probability for the candidate to persist as "best" can be
4698 // achieved by replacing it with "next" k-th candidate with the probability
4699 // of 1/k. It can be easily shown that by the end of the run, the
4700 // probability for any candidate is converged to 1/n, thus giving the
4701 // uniform distribution among all the candidates.
4702 //
4703 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4704 #define RANDOMIZED_DOMAIN_POW 29
4705 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4706 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
randomized_select(int count)4707 bool Compile::randomized_select(int count) {
4708 assert(count > 0, "only positive");
4709 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4710 }
4711
clone_map()4712 CloneMap& Compile::clone_map() { return _clone_map; }
set_clone_map(Dict * d)4713 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4714
dump() const4715 void NodeCloneInfo::dump() const {
4716 tty->print(" {%d:%d} ", idx(), gen());
4717 }
4718
clone(Node * old,Node * nnn,int gen)4719 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4720 uint64_t val = value(old->_idx);
4721 NodeCloneInfo cio(val);
4722 assert(val != 0, "old node should be in the map");
4723 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4724 insert(nnn->_idx, cin.get());
4725 #ifndef PRODUCT
4726 if (is_debug()) {
4727 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4728 }
4729 #endif
4730 }
4731
verify_insert_and_clone(Node * old,Node * nnn,int gen)4732 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4733 NodeCloneInfo cio(value(old->_idx));
4734 if (cio.get() == 0) {
4735 cio.set(old->_idx, 0);
4736 insert(old->_idx, cio.get());
4737 #ifndef PRODUCT
4738 if (is_debug()) {
4739 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4740 }
4741 #endif
4742 }
4743 clone(old, nnn, gen);
4744 }
4745
max_gen() const4746 int CloneMap::max_gen() const {
4747 int g = 0;
4748 DictI di(_dict);
4749 for(; di.test(); ++di) {
4750 int t = gen(di._key);
4751 if (g < t) {
4752 g = t;
4753 #ifndef PRODUCT
4754 if (is_debug()) {
4755 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4756 }
4757 #endif
4758 }
4759 }
4760 return g;
4761 }
4762
dump(node_idx_t key) const4763 void CloneMap::dump(node_idx_t key) const {
4764 uint64_t val = value(key);
4765 if (val != 0) {
4766 NodeCloneInfo ni(val);
4767 ni.dump();
4768 }
4769 }
4770