1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSet.hpp"
37 #include "gc/shared/c2/barrierSetC2.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "opto/addnode.hpp"
40 #include "opto/block.hpp"
41 #include "opto/c2compiler.hpp"
42 #include "opto/callGenerator.hpp"
43 #include "opto/callnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/cfgnode.hpp"
46 #include "opto/chaitin.hpp"
47 #include "opto/compile.hpp"
48 #include "opto/connode.hpp"
49 #include "opto/convertnode.hpp"
50 #include "opto/divnode.hpp"
51 #include "opto/escape.hpp"
52 #include "opto/idealGraphPrinter.hpp"
53 #include "opto/loopnode.hpp"
54 #include "opto/machnode.hpp"
55 #include "opto/macro.hpp"
56 #include "opto/matcher.hpp"
57 #include "opto/mathexactnode.hpp"
58 #include "opto/memnode.hpp"
59 #include "opto/mulnode.hpp"
60 #include "opto/narrowptrnode.hpp"
61 #include "opto/node.hpp"
62 #include "opto/opcodes.hpp"
63 #include "opto/output.hpp"
64 #include "opto/parse.hpp"
65 #include "opto/phaseX.hpp"
66 #include "opto/rootnode.hpp"
67 #include "opto/runtime.hpp"
68 #include "opto/stringopts.hpp"
69 #include "opto/type.hpp"
70 #include "opto/vectornode.hpp"
71 #include "runtime/arguments.hpp"
72 #include "runtime/sharedRuntime.hpp"
73 #include "runtime/signature.hpp"
74 #include "runtime/stubRoutines.hpp"
75 #include "runtime/timer.hpp"
76 #include "utilities/align.hpp"
77 #include "utilities/copy.hpp"
78 #include "utilities/macros.hpp"
79 #if INCLUDE_G1GC
80 #include "gc/g1/g1ThreadLocalData.hpp"
81 #endif // INCLUDE_G1GC
82 #if INCLUDE_ZGC
83 #include "gc/z/c2/zBarrierSetC2.hpp"
84 #endif
85 #if INCLUDE_SHENANDOAHGC
86 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
87 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
88 #include "gc/shenandoah/brooksPointer.hpp"
89 #endif
90
91
92 // -------------------- Compile::mach_constant_base_node -----------------------
93 // Constant table base node singleton.
94 MachConstantBaseNode* Compile::mach_constant_base_node() {
95 if (_mach_constant_base_node == NULL) {
96 _mach_constant_base_node = new MachConstantBaseNode();
97 _mach_constant_base_node->add_req(C->root());
98 }
99 return _mach_constant_base_node;
100 }
101
102
103 /// Support for intrinsics.
104
105 // Return the index at which m must be inserted (or already exists).
106 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
107 class IntrinsicDescPair {
108 private:
109 ciMethod* _m;
110 bool _is_virtual;
111 public:
112 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
113 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
114 ciMethod* m= elt->method();
115 ciMethod* key_m = key->_m;
116 if (key_m < m) return -1;
117 else if (key_m > m) return 1;
118 else {
119 bool is_virtual = elt->is_virtual();
120 bool key_virtual = key->_is_virtual;
121 if (key_virtual < is_virtual) return -1;
122 else if (key_virtual > is_virtual) return 1;
123 else return 0;
124 }
125 }
126 };
127 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
128 #ifdef ASSERT
129 for (int i = 1; i < _intrinsics->length(); i++) {
130 CallGenerator* cg1 = _intrinsics->at(i-1);
131 CallGenerator* cg2 = _intrinsics->at(i);
132 assert(cg1->method() != cg2->method()
133 ? cg1->method() < cg2->method()
134 : cg1->is_virtual() < cg2->is_virtual(),
135 "compiler intrinsics list must stay sorted");
136 }
137 #endif
138 IntrinsicDescPair pair(m, is_virtual);
139 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
140 }
141
142 void Compile::register_intrinsic(CallGenerator* cg) {
143 if (_intrinsics == NULL) {
144 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
145 }
146 int len = _intrinsics->length();
147 bool found = false;
148 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
149 assert(!found, "registering twice");
150 _intrinsics->insert_before(index, cg);
151 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
152 }
153
154 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
155 assert(m->is_loaded(), "don't try this on unloaded methods");
156 if (_intrinsics != NULL) {
157 bool found = false;
158 int index = intrinsic_insertion_index(m, is_virtual, found);
159 if (found) {
160 return _intrinsics->at(index);
161 }
162 }
163 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
164 if (m->intrinsic_id() != vmIntrinsics::_none &&
165 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
166 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
167 if (cg != NULL) {
168 // Save it for next time:
169 register_intrinsic(cg);
170 return cg;
171 } else {
172 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
173 }
174 }
175 return NULL;
176 }
177
178 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
179 // in library_call.cpp.
180
181
182 #ifndef PRODUCT
183 // statistics gathering...
184
185 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
186 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
187
188 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
189 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
190 int oflags = _intrinsic_hist_flags[id];
191 assert(flags != 0, "what happened?");
192 if (is_virtual) {
193 flags |= _intrinsic_virtual;
194 }
195 bool changed = (flags != oflags);
196 if ((flags & _intrinsic_worked) != 0) {
197 juint count = (_intrinsic_hist_count[id] += 1);
198 if (count == 1) {
199 changed = true; // first time
200 }
201 // increment the overall count also:
202 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
203 }
204 if (changed) {
205 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
206 // Something changed about the intrinsic's virtuality.
207 if ((flags & _intrinsic_virtual) != 0) {
208 // This is the first use of this intrinsic as a virtual call.
209 if (oflags != 0) {
210 // We already saw it as a non-virtual, so note both cases.
211 flags |= _intrinsic_both;
212 }
213 } else if ((oflags & _intrinsic_both) == 0) {
214 // This is the first use of this intrinsic as a non-virtual
215 flags |= _intrinsic_both;
216 }
217 }
218 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
219 }
220 // update the overall flags also:
221 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
222 return changed;
223 }
224
225 static char* format_flags(int flags, char* buf) {
226 buf[0] = 0;
227 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
228 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
229 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
230 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
231 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
232 if (buf[0] == 0) strcat(buf, ",");
233 assert(buf[0] == ',', "must be");
234 return &buf[1];
235 }
236
237 void Compile::print_intrinsic_statistics() {
238 char flagsbuf[100];
239 ttyLocker ttyl;
240 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
241 tty->print_cr("Compiler intrinsic usage:");
242 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
243 if (total == 0) total = 1; // avoid div0 in case of no successes
244 #define PRINT_STAT_LINE(name, c, f) \
245 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
246 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
247 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
248 int flags = _intrinsic_hist_flags[id];
249 juint count = _intrinsic_hist_count[id];
250 if ((flags | count) != 0) {
251 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
252 }
253 }
254 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
255 if (xtty != NULL) xtty->tail("statistics");
256 }
257
258 void Compile::print_statistics() {
259 { ttyLocker ttyl;
260 if (xtty != NULL) xtty->head("statistics type='opto'");
261 Parse::print_statistics();
262 PhaseCCP::print_statistics();
263 PhaseRegAlloc::print_statistics();
264 Scheduling::print_statistics();
265 PhasePeephole::print_statistics();
266 PhaseIdealLoop::print_statistics();
267 if (xtty != NULL) xtty->tail("statistics");
268 }
269 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
270 // put this under its own <statistics> element.
271 print_intrinsic_statistics();
272 }
273 }
274 #endif //PRODUCT
275
276 // Support for bundling info
277 Bundle* Compile::node_bundling(const Node *n) {
278 assert(valid_bundle_info(n), "oob");
279 return &_node_bundling_base[n->_idx];
280 }
281
282 bool Compile::valid_bundle_info(const Node *n) {
283 return (_node_bundling_limit > n->_idx);
284 }
285
286
287 void Compile::gvn_replace_by(Node* n, Node* nn) {
288 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
289 Node* use = n->last_out(i);
290 bool is_in_table = initial_gvn()->hash_delete(use);
291 uint uses_found = 0;
292 for (uint j = 0; j < use->len(); j++) {
293 if (use->in(j) == n) {
294 if (j < use->req())
295 use->set_req(j, nn);
296 else
297 use->set_prec(j, nn);
298 uses_found++;
299 }
300 }
301 if (is_in_table) {
302 // reinsert into table
303 initial_gvn()->hash_find_insert(use);
304 }
305 record_for_igvn(use);
306 i -= uses_found; // we deleted 1 or more copies of this edge
307 }
308 }
309
310
311 static inline bool not_a_node(const Node* n) {
312 if (n == NULL) return true;
313 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
314 if (*(address*)n == badAddress) return true; // kill by Node::destruct
315 return false;
316 }
317
318 // Identify all nodes that are reachable from below, useful.
319 // Use breadth-first pass that records state in a Unique_Node_List,
320 // recursive traversal is slower.
321 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
322 int estimated_worklist_size = live_nodes();
323 useful.map( estimated_worklist_size, NULL ); // preallocate space
324
325 // Initialize worklist
326 if (root() != NULL) { useful.push(root()); }
327 // If 'top' is cached, declare it useful to preserve cached node
328 if( cached_top_node() ) { useful.push(cached_top_node()); }
329
330 // Push all useful nodes onto the list, breadthfirst
331 for( uint next = 0; next < useful.size(); ++next ) {
332 assert( next < unique(), "Unique useful nodes < total nodes");
333 Node *n = useful.at(next);
334 uint max = n->len();
335 for( uint i = 0; i < max; ++i ) {
336 Node *m = n->in(i);
337 if (not_a_node(m)) continue;
338 useful.push(m);
339 }
340 }
341 }
342
343 // Update dead_node_list with any missing dead nodes using useful
344 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
345 void Compile::update_dead_node_list(Unique_Node_List &useful) {
346 uint max_idx = unique();
347 VectorSet& useful_node_set = useful.member_set();
348
349 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
350 // If node with index node_idx is not in useful set,
351 // mark it as dead in dead node list.
352 if (! useful_node_set.test(node_idx) ) {
353 record_dead_node(node_idx);
354 }
355 }
356 }
357
358 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
359 int shift = 0;
360 for (int i = 0; i < inlines->length(); i++) {
361 CallGenerator* cg = inlines->at(i);
362 CallNode* call = cg->call_node();
363 if (shift > 0) {
364 inlines->at_put(i-shift, cg);
365 }
366 if (!useful.member(call)) {
367 shift++;
368 }
369 }
370 inlines->trunc_to(inlines->length()-shift);
371 }
372
373 // Disconnect all useless nodes by disconnecting those at the boundary.
374 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
375 uint next = 0;
376 while (next < useful.size()) {
377 Node *n = useful.at(next++);
378 if (n->is_SafePoint()) {
379 // We're done with a parsing phase. Replaced nodes are not valid
380 // beyond that point.
381 n->as_SafePoint()->delete_replaced_nodes();
382 }
383 // Use raw traversal of out edges since this code removes out edges
384 int max = n->outcnt();
385 for (int j = 0; j < max; ++j) {
386 Node* child = n->raw_out(j);
387 if (! useful.member(child)) {
388 assert(!child->is_top() || child != top(),
389 "If top is cached in Compile object it is in useful list");
390 // Only need to remove this out-edge to the useless node
391 n->raw_del_out(j);
392 --j;
393 --max;
394 }
395 }
396 if (n->outcnt() == 1 && n->has_special_unique_user()) {
397 record_for_igvn(n->unique_out());
398 }
399 #if INCLUDE_SHENANDOAHGC
400 if (n->Opcode() == Op_AddP && CallLeafNode::has_only_shenandoah_wb_pre_uses(n)) {
401 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
402 record_for_igvn(n->fast_out(i));
403 }
404 }
405 #endif
406 }
407 // Remove useless macro and predicate opaq nodes
408 for (int i = C->macro_count()-1; i >= 0; i--) {
409 Node* n = C->macro_node(i);
410 if (!useful.member(n)) {
411 remove_macro_node(n);
412 }
413 }
414 // Remove useless CastII nodes with range check dependency
415 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
416 Node* cast = range_check_cast_node(i);
417 if (!useful.member(cast)) {
418 remove_range_check_cast(cast);
419 }
420 }
421 // Remove useless expensive nodes
422 for (int i = C->expensive_count()-1; i >= 0; i--) {
423 Node* n = C->expensive_node(i);
424 if (!useful.member(n)) {
425 remove_expensive_node(n);
426 }
427 }
428
429 // Remove useless Opaque4 nodes
430 for (int i = opaque4_count() - 1; i >= 0; i--) {
431 Node* opaq = opaque4_node(i);
432 if (!useful.member(opaq)) {
433 remove_opaque4_node(opaq);
434 }
435 }
436 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
437 bs->eliminate_useless_gc_barriers(useful);
438 // clean up the late inline lists
439 remove_useless_late_inlines(&_string_late_inlines, useful);
440 remove_useless_late_inlines(&_boxing_late_inlines, useful);
441 remove_useless_late_inlines(&_late_inlines, useful);
442 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
443 }
444
445 //------------------------------frame_size_in_words-----------------------------
446 // frame_slots in units of words
447 int Compile::frame_size_in_words() const {
448 // shift is 0 in LP32 and 1 in LP64
449 const int shift = (LogBytesPerWord - LogBytesPerInt);
450 int words = _frame_slots >> shift;
451 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
452 return words;
453 }
454
455 // To bang the stack of this compiled method we use the stack size
456 // that the interpreter would need in case of a deoptimization. This
457 // removes the need to bang the stack in the deoptimization blob which
458 // in turn simplifies stack overflow handling.
459 int Compile::bang_size_in_bytes() const {
460 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
461 }
462
463 // ============================================================================
464 //------------------------------CompileWrapper---------------------------------
465 class CompileWrapper : public StackObj {
466 Compile *const _compile;
467 public:
468 CompileWrapper(Compile* compile);
469
470 ~CompileWrapper();
471 };
472
473 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
474 // the Compile* pointer is stored in the current ciEnv:
475 ciEnv* env = compile->env();
476 assert(env == ciEnv::current(), "must already be a ciEnv active");
477 assert(env->compiler_data() == NULL, "compile already active?");
478 env->set_compiler_data(compile);
479 assert(compile == Compile::current(), "sanity");
480
481 compile->set_type_dict(NULL);
482 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
483 compile->clone_map().set_clone_idx(0);
484 compile->set_type_hwm(NULL);
485 compile->set_type_last_size(0);
486 compile->set_last_tf(NULL, NULL);
487 compile->set_indexSet_arena(NULL);
488 compile->set_indexSet_free_block_list(NULL);
489 compile->init_type_arena();
490 Type::Initialize(compile);
491 _compile->set_scratch_buffer_blob(NULL);
492 _compile->begin_method();
493 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
494 }
495 CompileWrapper::~CompileWrapper() {
496 _compile->end_method();
497 if (_compile->scratch_buffer_blob() != NULL)
498 BufferBlob::free(_compile->scratch_buffer_blob());
499 _compile->env()->set_compiler_data(NULL);
500 }
501
502
503 //----------------------------print_compile_messages---------------------------
504 void Compile::print_compile_messages() {
505 #ifndef PRODUCT
506 // Check if recompiling
507 if (_subsume_loads == false && PrintOpto) {
508 // Recompiling without allowing machine instructions to subsume loads
509 tty->print_cr("*********************************************************");
510 tty->print_cr("** Bailout: Recompile without subsuming loads **");
511 tty->print_cr("*********************************************************");
512 }
513 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
514 // Recompiling without escape analysis
515 tty->print_cr("*********************************************************");
516 tty->print_cr("** Bailout: Recompile without escape analysis **");
517 tty->print_cr("*********************************************************");
518 }
519 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
520 // Recompiling without boxing elimination
521 tty->print_cr("*********************************************************");
522 tty->print_cr("** Bailout: Recompile without boxing elimination **");
523 tty->print_cr("*********************************************************");
524 }
525 if (C->directive()->BreakAtCompileOption) {
526 // Open the debugger when compiling this method.
527 tty->print("### Breaking when compiling: ");
528 method()->print_short_name();
529 tty->cr();
530 BREAKPOINT;
531 }
532
533 if( PrintOpto ) {
534 if (is_osr_compilation()) {
535 tty->print("[OSR]%3d", _compile_id);
536 } else {
537 tty->print("%3d", _compile_id);
538 }
539 }
540 #endif
541 }
542
543
544 //-----------------------init_scratch_buffer_blob------------------------------
545 // Construct a temporary BufferBlob and cache it for this compile.
546 void Compile::init_scratch_buffer_blob(int const_size) {
547 // If there is already a scratch buffer blob allocated and the
548 // constant section is big enough, use it. Otherwise free the
549 // current and allocate a new one.
550 BufferBlob* blob = scratch_buffer_blob();
551 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
552 // Use the current blob.
553 } else {
554 if (blob != NULL) {
555 BufferBlob::free(blob);
556 }
557
558 ResourceMark rm;
559 _scratch_const_size = const_size;
560 int locs_size = sizeof(relocInfo) * MAX_locs_size;
561 int slop = 2 * CodeSection::end_slop(); // space between sections
562 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size + slop + locs_size);
563 blob = BufferBlob::create("Compile::scratch_buffer", size);
564 // Record the buffer blob for next time.
565 set_scratch_buffer_blob(blob);
566 // Have we run out of code space?
567 if (scratch_buffer_blob() == NULL) {
568 // Let CompilerBroker disable further compilations.
569 record_failure("Not enough space for scratch buffer in CodeCache");
570 return;
571 }
572 }
573
574 // Initialize the relocation buffers
575 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
576 set_scratch_locs_memory(locs_buf);
577 }
578
579
580 //-----------------------scratch_emit_size-------------------------------------
581 // Helper function that computes size by emitting code
582 uint Compile::scratch_emit_size(const Node* n) {
583 // Start scratch_emit_size section.
584 set_in_scratch_emit_size(true);
585
586 // Emit into a trash buffer and count bytes emitted.
587 // This is a pretty expensive way to compute a size,
588 // but it works well enough if seldom used.
589 // All common fixed-size instructions are given a size
590 // method by the AD file.
591 // Note that the scratch buffer blob and locs memory are
592 // allocated at the beginning of the compile task, and
593 // may be shared by several calls to scratch_emit_size.
594 // The allocation of the scratch buffer blob is particularly
595 // expensive, since it has to grab the code cache lock.
596 BufferBlob* blob = this->scratch_buffer_blob();
597 assert(blob != NULL, "Initialize BufferBlob at start");
598 assert(blob->size() > MAX_inst_size, "sanity");
599 relocInfo* locs_buf = scratch_locs_memory();
600 address blob_begin = blob->content_begin();
601 address blob_end = (address)locs_buf;
602 assert(blob->contains(blob_end), "sanity");
603 CodeBuffer buf(blob_begin, blob_end - blob_begin);
604 buf.initialize_consts_size(_scratch_const_size);
605 buf.initialize_stubs_size(MAX_stubs_size);
606 assert(locs_buf != NULL, "sanity");
607 int lsize = MAX_locs_size / 3;
608 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
609 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
610 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
611 // Mark as scratch buffer.
612 buf.consts()->set_scratch_emit();
613 buf.insts()->set_scratch_emit();
614 buf.stubs()->set_scratch_emit();
615
616 // Do the emission.
617
618 Label fakeL; // Fake label for branch instructions.
619 Label* saveL = NULL;
620 uint save_bnum = 0;
621 bool is_branch = n->is_MachBranch();
622 if (is_branch) {
623 MacroAssembler masm(&buf);
624 masm.bind(fakeL);
625 n->as_MachBranch()->save_label(&saveL, &save_bnum);
626 n->as_MachBranch()->label_set(&fakeL, 0);
627 }
628 n->emit(buf, this->regalloc());
629
630 // Emitting into the scratch buffer should not fail
631 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
632
633 if (is_branch) // Restore label.
634 n->as_MachBranch()->label_set(saveL, save_bnum);
635
636 // End scratch_emit_size section.
637 set_in_scratch_emit_size(false);
638
639 return buf.insts_size();
640 }
641
642
643 // ============================================================================
644 //------------------------------Compile standard-------------------------------
645 debug_only( int Compile::_debug_idx = 100000; )
646
647 // Compile a method. entry_bci is -1 for normal compilations and indicates
648 // the continuation bci for on stack replacement.
649
650
651 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
652 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
653 : Phase(Compiler),
654 _compile_id(ci_env->compile_id()),
655 _save_argument_registers(false),
656 _subsume_loads(subsume_loads),
657 _do_escape_analysis(do_escape_analysis),
658 _eliminate_boxing(eliminate_boxing),
659 _method(target),
660 _entry_bci(osr_bci),
661 _stub_function(NULL),
662 _stub_name(NULL),
663 _stub_entry_point(NULL),
664 _max_node_limit(MaxNodeLimit),
665 _orig_pc_slot(0),
666 _orig_pc_slot_offset_in_bytes(0),
667 _inlining_progress(false),
668 _inlining_incrementally(false),
669 _has_reserved_stack_access(target->has_reserved_stack_access()),
670 #ifndef PRODUCT
671 _trace_opto_output(directive->TraceOptoOutputOption),
672 #endif
673 _has_method_handle_invokes(false),
674 _comp_arena(mtCompiler),
675 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
676 _env(ci_env),
677 _directive(directive),
678 _log(ci_env->log()),
679 _failure_reason(NULL),
680 _congraph(NULL),
681 #ifndef PRODUCT
682 _printer(IdealGraphPrinter::printer()),
683 #endif
684 _dead_node_list(comp_arena()),
685 _dead_node_count(0),
686 _node_arena(mtCompiler),
687 _old_arena(mtCompiler),
688 _mach_constant_base_node(NULL),
689 _Compile_types(mtCompiler),
690 _initial_gvn(NULL),
691 _for_igvn(NULL),
692 _warm_calls(NULL),
693 _late_inlines(comp_arena(), 2, 0, NULL),
694 _string_late_inlines(comp_arena(), 2, 0, NULL),
695 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
696 _late_inlines_pos(0),
697 _number_of_mh_late_inlines(0),
698 _print_inlining_stream(NULL),
699 _print_inlining_list(NULL),
700 _print_inlining_idx(0),
701 _print_inlining_output(NULL),
702 _replay_inline_data(NULL),
703 _java_calls(0),
704 _inner_loops(0),
705 _interpreter_frame_size(0),
706 _node_bundling_limit(0),
707 _node_bundling_base(NULL),
708 _code_buffer("Compile::Fill_buffer"),
709 _scratch_const_size(-1),
710 _in_scratch_emit_size(false)
711 #ifndef PRODUCT
712 , _in_dump_cnt(0)
713 #endif
714 {
715 C = this;
716 #ifndef PRODUCT
717 if (_printer != NULL) {
718 _printer->set_compile(this);
719 }
720 #endif
721 CompileWrapper cw(this);
722
723 if (CITimeVerbose) {
724 tty->print(" ");
725 target->holder()->name()->print();
726 tty->print(".");
727 target->print_short_name();
728 tty->print(" ");
729 }
730 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
731 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
732
733 #ifndef PRODUCT
734 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
735 if (!print_opto_assembly) {
736 bool print_assembly = directive->PrintAssemblyOption;
737 if (print_assembly && !Disassembler::can_decode()) {
738 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
739 print_opto_assembly = true;
740 }
741 }
742 set_print_assembly(print_opto_assembly);
743 set_parsed_irreducible_loop(false);
744
745 if (directive->ReplayInlineOption) {
746 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
747 }
748 #endif
749 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
750 set_print_intrinsics(directive->PrintIntrinsicsOption);
751 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
752
753 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
754 // Make sure the method being compiled gets its own MDO,
755 // so we can at least track the decompile_count().
756 // Need MDO to record RTM code generation state.
757 method()->ensure_method_data();
758 }
759
760 Init(::AliasLevel);
761
762
763 print_compile_messages();
764
765 _ilt = InlineTree::build_inline_tree_root();
766
767 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
768 assert(num_alias_types() >= AliasIdxRaw, "");
769
770 #define MINIMUM_NODE_HASH 1023
771 // Node list that Iterative GVN will start with
772 Unique_Node_List for_igvn(comp_arena());
773 set_for_igvn(&for_igvn);
774
775 // GVN that will be run immediately on new nodes
776 uint estimated_size = method()->code_size()*4+64;
777 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
778 PhaseGVN gvn(node_arena(), estimated_size);
779 set_initial_gvn(&gvn);
780
781 print_inlining_init();
782 { // Scope for timing the parser
783 TracePhase tp("parse", &timers[_t_parser]);
784
785 // Put top into the hash table ASAP.
786 initial_gvn()->transform_no_reclaim(top());
787
788 // Set up tf(), start(), and find a CallGenerator.
789 CallGenerator* cg = NULL;
790 if (is_osr_compilation()) {
791 const TypeTuple *domain = StartOSRNode::osr_domain();
792 const TypeTuple *range = TypeTuple::make_range(method()->signature());
793 init_tf(TypeFunc::make(domain, range));
794 StartNode* s = new StartOSRNode(root(), domain);
795 initial_gvn()->set_type_bottom(s);
796 init_start(s);
797 cg = CallGenerator::for_osr(method(), entry_bci());
798 } else {
799 // Normal case.
800 init_tf(TypeFunc::make(method()));
801 StartNode* s = new StartNode(root(), tf()->domain());
802 initial_gvn()->set_type_bottom(s);
803 init_start(s);
804 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
805 // With java.lang.ref.reference.get() we must go through the
806 // intrinsic - even when get() is the root
807 // method of the compile - so that, if necessary, the value in
808 // the referent field of the reference object gets recorded by
809 // the pre-barrier code.
810 cg = find_intrinsic(method(), false);
811 }
812 if (cg == NULL) {
813 float past_uses = method()->interpreter_invocation_count();
814 float expected_uses = past_uses;
815 cg = CallGenerator::for_inline(method(), expected_uses);
816 }
817 }
818 if (failing()) return;
819 if (cg == NULL) {
820 record_method_not_compilable("cannot parse method");
821 return;
822 }
823 JVMState* jvms = build_start_state(start(), tf());
824 if ((jvms = cg->generate(jvms)) == NULL) {
825 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
826 record_method_not_compilable("method parse failed");
827 }
828 return;
829 }
830 GraphKit kit(jvms);
831
832 if (!kit.stopped()) {
833 // Accept return values, and transfer control we know not where.
834 // This is done by a special, unique ReturnNode bound to root.
835 return_values(kit.jvms());
836 }
837
838 if (kit.has_exceptions()) {
839 // Any exceptions that escape from this call must be rethrown
840 // to whatever caller is dynamically above us on the stack.
841 // This is done by a special, unique RethrowNode bound to root.
842 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
843 }
844
845 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
846
847 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
848 inline_string_calls(true);
849 }
850
851 if (failing()) return;
852
853 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
854
855 // Remove clutter produced by parsing.
856 if (!failing()) {
857 ResourceMark rm;
858 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
859 }
860 }
861
862 // Note: Large methods are capped off in do_one_bytecode().
863 if (failing()) return;
864
865 // After parsing, node notes are no longer automagic.
866 // They must be propagated by register_new_node_with_optimizer(),
867 // clone(), or the like.
868 set_default_node_notes(NULL);
869
870 for (;;) {
871 int successes = Inline_Warm();
872 if (failing()) return;
873 if (successes == 0) break;
874 }
875
876 // Drain the list.
877 Finish_Warm();
878 #ifndef PRODUCT
879 if (_printer && _printer->should_print(1)) {
880 _printer->print_inlining();
881 }
882 #endif
883
884 if (failing()) return;
885 NOT_PRODUCT( verify_graph_edges(); )
886
887 // Now optimize
888 Optimize();
889 if (failing()) return;
890 NOT_PRODUCT( verify_graph_edges(); )
891
892 #ifndef PRODUCT
893 if (PrintIdeal) {
894 ttyLocker ttyl; // keep the following output all in one block
895 // This output goes directly to the tty, not the compiler log.
896 // To enable tools to match it up with the compilation activity,
897 // be sure to tag this tty output with the compile ID.
898 if (xtty != NULL) {
899 xtty->head("ideal compile_id='%d'%s", compile_id(),
900 is_osr_compilation() ? " compile_kind='osr'" :
901 "");
902 }
903 root()->dump(9999);
904 if (xtty != NULL) {
905 xtty->tail("ideal");
906 }
907 }
908 #endif
909
910 NOT_PRODUCT( verify_barriers(); )
911
912 // Dump compilation data to replay it.
913 if (directive->DumpReplayOption) {
914 env()->dump_replay_data(_compile_id);
915 }
916 if (directive->DumpInlineOption && (ilt() != NULL)) {
917 env()->dump_inline_data(_compile_id);
918 }
919
920 // Now that we know the size of all the monitors we can add a fixed slot
921 // for the original deopt pc.
922
923 _orig_pc_slot = fixed_slots();
924 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
925 set_fixed_slots(next_slot);
926
927 // Compute when to use implicit null checks. Used by matching trap based
928 // nodes and NullCheck optimization.
929 set_allowed_deopt_reasons();
930
931 // Now generate code
932 Code_Gen();
933 if (failing()) return;
934
935 // Check if we want to skip execution of all compiled code.
936 {
937 #ifndef PRODUCT
938 if (OptoNoExecute) {
939 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
940 return;
941 }
942 #endif
943 TracePhase tp("install_code", &timers[_t_registerMethod]);
944
945 if (is_osr_compilation()) {
946 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
947 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
948 } else {
949 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
950 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
951 }
952
953 env()->register_method(_method, _entry_bci,
954 &_code_offsets,
955 _orig_pc_slot_offset_in_bytes,
956 code_buffer(),
957 frame_size_in_words(), _oop_map_set,
958 &_handler_table, &_inc_table,
959 compiler,
960 has_unsafe_access(),
961 SharedRuntime::is_wide_vector(max_vector_size()),
962 rtm_state()
963 );
964
965 if (log() != NULL) // Print code cache state into compiler log
966 log()->code_cache_state();
967 }
968 }
969
970 //------------------------------Compile----------------------------------------
971 // Compile a runtime stub
972 Compile::Compile( ciEnv* ci_env,
973 TypeFunc_generator generator,
974 address stub_function,
975 const char *stub_name,
976 int is_fancy_jump,
977 bool pass_tls,
978 bool save_arg_registers,
979 bool return_pc,
980 DirectiveSet* directive)
981 : Phase(Compiler),
982 _compile_id(0),
983 _save_argument_registers(save_arg_registers),
984 _subsume_loads(true),
985 _do_escape_analysis(false),
986 _eliminate_boxing(false),
987 _method(NULL),
988 _entry_bci(InvocationEntryBci),
989 _stub_function(stub_function),
990 _stub_name(stub_name),
991 _stub_entry_point(NULL),
992 _max_node_limit(MaxNodeLimit),
993 _orig_pc_slot(0),
994 _orig_pc_slot_offset_in_bytes(0),
995 _inlining_progress(false),
996 _inlining_incrementally(false),
997 _has_reserved_stack_access(false),
998 #ifndef PRODUCT
999 _trace_opto_output(directive->TraceOptoOutputOption),
1000 #endif
1001 _has_method_handle_invokes(false),
1002 _comp_arena(mtCompiler),
1003 _env(ci_env),
1004 _directive(directive),
1005 _log(ci_env->log()),
1006 _failure_reason(NULL),
1007 _congraph(NULL),
1008 #ifndef PRODUCT
1009 _printer(NULL),
1010 #endif
1011 _dead_node_list(comp_arena()),
1012 _dead_node_count(0),
1013 _node_arena(mtCompiler),
1014 _old_arena(mtCompiler),
1015 _mach_constant_base_node(NULL),
1016 _Compile_types(mtCompiler),
1017 _initial_gvn(NULL),
1018 _for_igvn(NULL),
1019 _warm_calls(NULL),
1020 _number_of_mh_late_inlines(0),
1021 _print_inlining_stream(NULL),
1022 _print_inlining_list(NULL),
1023 _print_inlining_idx(0),
1024 _print_inlining_output(NULL),
1025 _replay_inline_data(NULL),
1026 _java_calls(0),
1027 _inner_loops(0),
1028 _interpreter_frame_size(0),
1029 _node_bundling_limit(0),
1030 _node_bundling_base(NULL),
1031 _code_buffer("Compile::Fill_buffer"),
1032 #ifndef PRODUCT
1033 _in_dump_cnt(0),
1034 #endif
1035 _allowed_reasons(0) {
1036 C = this;
1037
1038 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1039 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1040
1041 #ifndef PRODUCT
1042 set_print_assembly(PrintFrameConverterAssembly);
1043 set_parsed_irreducible_loop(false);
1044 #endif
1045 set_has_irreducible_loop(false); // no loops
1046
1047 CompileWrapper cw(this);
1048 Init(/*AliasLevel=*/ 0);
1049 init_tf((*generator)());
1050
1051 {
1052 // The following is a dummy for the sake of GraphKit::gen_stub
1053 Unique_Node_List for_igvn(comp_arena());
1054 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1055 PhaseGVN gvn(Thread::current()->resource_area(),255);
1056 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1057 gvn.transform_no_reclaim(top());
1058
1059 GraphKit kit;
1060 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1061 }
1062
1063 NOT_PRODUCT( verify_graph_edges(); )
1064 Code_Gen();
1065 if (failing()) return;
1066
1067
1068 // Entry point will be accessed using compile->stub_entry_point();
1069 if (code_buffer() == NULL) {
1070 Matcher::soft_match_failure();
1071 } else {
1072 if (PrintAssembly && (WizardMode || Verbose))
1073 tty->print_cr("### Stub::%s", stub_name);
1074
1075 if (!failing()) {
1076 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1077
1078 // Make the NMethod
1079 // For now we mark the frame as never safe for profile stackwalking
1080 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1081 code_buffer(),
1082 CodeOffsets::frame_never_safe,
1083 // _code_offsets.value(CodeOffsets::Frame_Complete),
1084 frame_size_in_words(),
1085 _oop_map_set,
1086 save_arg_registers);
1087 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1088
1089 _stub_entry_point = rs->entry_point();
1090 }
1091 }
1092 }
1093
1094 //------------------------------Init-------------------------------------------
1095 // Prepare for a single compilation
1096 void Compile::Init(int aliaslevel) {
1097 _unique = 0;
1098 _regalloc = NULL;
1099
1100 _tf = NULL; // filled in later
1101 _top = NULL; // cached later
1102 _matcher = NULL; // filled in later
1103 _cfg = NULL; // filled in later
1104
1105 set_24_bit_selection_and_mode(Use24BitFP, false);
1106
1107 _node_note_array = NULL;
1108 _default_node_notes = NULL;
1109 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1110
1111 _immutable_memory = NULL; // filled in at first inquiry
1112
1113 // Globally visible Nodes
1114 // First set TOP to NULL to give safe behavior during creation of RootNode
1115 set_cached_top_node(NULL);
1116 set_root(new RootNode());
1117 // Now that you have a Root to point to, create the real TOP
1118 set_cached_top_node( new ConNode(Type::TOP) );
1119 set_recent_alloc(NULL, NULL);
1120
1121 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1122 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1123 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1124 env()->set_dependencies(new Dependencies(env()));
1125
1126 _fixed_slots = 0;
1127 set_has_split_ifs(false);
1128 set_has_loops(has_method() && method()->has_loops()); // first approximation
1129 set_has_stringbuilder(false);
1130 set_has_boxed_value(false);
1131 _trap_can_recompile = false; // no traps emitted yet
1132 _major_progress = true; // start out assuming good things will happen
1133 set_has_unsafe_access(false);
1134 set_max_vector_size(0);
1135 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers
1136 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1137 set_decompile_count(0);
1138
1139 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1140 set_num_loop_opts(LoopOptsCount);
1141 set_do_inlining(Inline);
1142 set_max_inline_size(MaxInlineSize);
1143 set_freq_inline_size(FreqInlineSize);
1144 set_do_scheduling(OptoScheduling);
1145 set_do_count_invocations(false);
1146 set_do_method_data_update(false);
1147
1148 set_do_vector_loop(false);
1149
1150 if (AllowVectorizeOnDemand) {
1151 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1152 set_do_vector_loop(true);
1153 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());})
1154 } else if (has_method() && method()->name() != 0 &&
1155 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1156 set_do_vector_loop(true);
1157 }
1158 }
1159 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1160 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());})
1161
1162 set_age_code(has_method() && method()->profile_aging());
1163 set_rtm_state(NoRTM); // No RTM lock eliding by default
1164 _max_node_limit = _directive->MaxNodeLimitOption;
1165
1166 #if INCLUDE_RTM_OPT
1167 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1168 int rtm_state = method()->method_data()->rtm_state();
1169 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1170 // Don't generate RTM lock eliding code.
1171 set_rtm_state(NoRTM);
1172 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1173 // Generate RTM lock eliding code without abort ratio calculation code.
1174 set_rtm_state(UseRTM);
1175 } else if (UseRTMDeopt) {
1176 // Generate RTM lock eliding code and include abort ratio calculation
1177 // code if UseRTMDeopt is on.
1178 set_rtm_state(ProfileRTM);
1179 }
1180 }
1181 #endif
1182 if (debug_info()->recording_non_safepoints()) {
1183 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1184 (comp_arena(), 8, 0, NULL));
1185 set_default_node_notes(Node_Notes::make(this));
1186 }
1187
1188 // // -- Initialize types before each compile --
1189 // // Update cached type information
1190 // if( _method && _method->constants() )
1191 // Type::update_loaded_types(_method, _method->constants());
1192
1193 // Init alias_type map.
1194 if (!_do_escape_analysis && aliaslevel == 3)
1195 aliaslevel = 2; // No unique types without escape analysis
1196 _AliasLevel = aliaslevel;
1197 const int grow_ats = 16;
1198 _max_alias_types = grow_ats;
1199 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1200 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1201 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1202 {
1203 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1204 }
1205 // Initialize the first few types.
1206 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1207 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1208 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1209 _num_alias_types = AliasIdxRaw+1;
1210 // Zero out the alias type cache.
1211 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1212 // A NULL adr_type hits in the cache right away. Preload the right answer.
1213 probe_alias_cache(NULL)->_index = AliasIdxTop;
1214
1215 _intrinsics = NULL;
1216 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1217 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1218 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1219 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1220 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1221 register_library_intrinsics();
1222 }
1223
1224 //---------------------------init_start----------------------------------------
1225 // Install the StartNode on this compile object.
1226 void Compile::init_start(StartNode* s) {
1227 if (failing())
1228 return; // already failing
1229 assert(s == start(), "");
1230 }
1231
1232 /**
1233 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1234 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1235 * the ideal graph.
1236 */
1237 StartNode* Compile::start() const {
1238 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1239 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1240 Node* start = root()->fast_out(i);
1241 if (start->is_Start()) {
1242 return start->as_Start();
1243 }
1244 }
1245 fatal("Did not find Start node!");
1246 return NULL;
1247 }
1248
1249 //-------------------------------immutable_memory-------------------------------------
1250 // Access immutable memory
1251 Node* Compile::immutable_memory() {
1252 if (_immutable_memory != NULL) {
1253 return _immutable_memory;
1254 }
1255 StartNode* s = start();
1256 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1257 Node *p = s->fast_out(i);
1258 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1259 _immutable_memory = p;
1260 return _immutable_memory;
1261 }
1262 }
1263 ShouldNotReachHere();
1264 return NULL;
1265 }
1266
1267 //----------------------set_cached_top_node------------------------------------
1268 // Install the cached top node, and make sure Node::is_top works correctly.
1269 void Compile::set_cached_top_node(Node* tn) {
1270 if (tn != NULL) verify_top(tn);
1271 Node* old_top = _top;
1272 _top = tn;
1273 // Calling Node::setup_is_top allows the nodes the chance to adjust
1274 // their _out arrays.
1275 if (_top != NULL) _top->setup_is_top();
1276 if (old_top != NULL) old_top->setup_is_top();
1277 assert(_top == NULL || top()->is_top(), "");
1278 }
1279
1280 #ifdef ASSERT
1281 uint Compile::count_live_nodes_by_graph_walk() {
1282 Unique_Node_List useful(comp_arena());
1283 // Get useful node list by walking the graph.
1284 identify_useful_nodes(useful);
1285 return useful.size();
1286 }
1287
1288 void Compile::print_missing_nodes() {
1289
1290 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1291 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1292 return;
1293 }
1294
1295 // This is an expensive function. It is executed only when the user
1296 // specifies VerifyIdealNodeCount option or otherwise knows the
1297 // additional work that needs to be done to identify reachable nodes
1298 // by walking the flow graph and find the missing ones using
1299 // _dead_node_list.
1300
1301 Unique_Node_List useful(comp_arena());
1302 // Get useful node list by walking the graph.
1303 identify_useful_nodes(useful);
1304
1305 uint l_nodes = C->live_nodes();
1306 uint l_nodes_by_walk = useful.size();
1307
1308 if (l_nodes != l_nodes_by_walk) {
1309 if (_log != NULL) {
1310 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1311 _log->stamp();
1312 _log->end_head();
1313 }
1314 VectorSet& useful_member_set = useful.member_set();
1315 int last_idx = l_nodes_by_walk;
1316 for (int i = 0; i < last_idx; i++) {
1317 if (useful_member_set.test(i)) {
1318 if (_dead_node_list.test(i)) {
1319 if (_log != NULL) {
1320 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1321 }
1322 if (PrintIdealNodeCount) {
1323 // Print the log message to tty
1324 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1325 useful.at(i)->dump();
1326 }
1327 }
1328 }
1329 else if (! _dead_node_list.test(i)) {
1330 if (_log != NULL) {
1331 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1332 }
1333 if (PrintIdealNodeCount) {
1334 // Print the log message to tty
1335 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1336 }
1337 }
1338 }
1339 if (_log != NULL) {
1340 _log->tail("mismatched_nodes");
1341 }
1342 }
1343 }
1344 void Compile::record_modified_node(Node* n) {
1345 if (_modified_nodes != NULL && !_inlining_incrementally &&
1346 n->outcnt() != 0 && !n->is_Con()) {
1347 _modified_nodes->push(n);
1348 }
1349 }
1350
1351 void Compile::remove_modified_node(Node* n) {
1352 if (_modified_nodes != NULL) {
1353 _modified_nodes->remove(n);
1354 }
1355 }
1356 #endif
1357
1358 #ifndef PRODUCT
1359 void Compile::verify_top(Node* tn) const {
1360 if (tn != NULL) {
1361 assert(tn->is_Con(), "top node must be a constant");
1362 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1363 assert(tn->in(0) != NULL, "must have live top node");
1364 }
1365 }
1366 #endif
1367
1368
1369 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1370
1371 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1372 guarantee(arr != NULL, "");
1373 int num_blocks = arr->length();
1374 if (grow_by < num_blocks) grow_by = num_blocks;
1375 int num_notes = grow_by * _node_notes_block_size;
1376 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1377 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1378 while (num_notes > 0) {
1379 arr->append(notes);
1380 notes += _node_notes_block_size;
1381 num_notes -= _node_notes_block_size;
1382 }
1383 assert(num_notes == 0, "exact multiple, please");
1384 }
1385
1386 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1387 if (source == NULL || dest == NULL) return false;
1388
1389 if (dest->is_Con())
1390 return false; // Do not push debug info onto constants.
1391
1392 #ifdef ASSERT
1393 // Leave a bread crumb trail pointing to the original node:
1394 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1395 dest->set_debug_orig(source);
1396 }
1397 #endif
1398
1399 if (node_note_array() == NULL)
1400 return false; // Not collecting any notes now.
1401
1402 // This is a copy onto a pre-existing node, which may already have notes.
1403 // If both nodes have notes, do not overwrite any pre-existing notes.
1404 Node_Notes* source_notes = node_notes_at(source->_idx);
1405 if (source_notes == NULL || source_notes->is_clear()) return false;
1406 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1407 if (dest_notes == NULL || dest_notes->is_clear()) {
1408 return set_node_notes_at(dest->_idx, source_notes);
1409 }
1410
1411 Node_Notes merged_notes = (*source_notes);
1412 // The order of operations here ensures that dest notes will win...
1413 merged_notes.update_from(dest_notes);
1414 return set_node_notes_at(dest->_idx, &merged_notes);
1415 }
1416
1417
1418 //--------------------------allow_range_check_smearing-------------------------
1419 // Gating condition for coalescing similar range checks.
1420 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1421 // single covering check that is at least as strong as any of them.
1422 // If the optimization succeeds, the simplified (strengthened) range check
1423 // will always succeed. If it fails, we will deopt, and then give up
1424 // on the optimization.
1425 bool Compile::allow_range_check_smearing() const {
1426 // If this method has already thrown a range-check,
1427 // assume it was because we already tried range smearing
1428 // and it failed.
1429 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1430 return !already_trapped;
1431 }
1432
1433
1434 //------------------------------flatten_alias_type-----------------------------
1435 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1436 int offset = tj->offset();
1437 TypePtr::PTR ptr = tj->ptr();
1438
1439 // Known instance (scalarizable allocation) alias only with itself.
1440 bool is_known_inst = tj->isa_oopptr() != NULL &&
1441 tj->is_oopptr()->is_known_instance();
1442
1443 // Process weird unsafe references.
1444 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1445 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1446 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1447 tj = TypeOopPtr::BOTTOM;
1448 ptr = tj->ptr();
1449 offset = tj->offset();
1450 }
1451
1452 // Array pointers need some flattening
1453 const TypeAryPtr *ta = tj->isa_aryptr();
1454 if (ta && ta->is_stable()) {
1455 // Erase stability property for alias analysis.
1456 tj = ta = ta->cast_to_stable(false);
1457 }
1458 if( ta && is_known_inst ) {
1459 if ( offset != Type::OffsetBot &&
1460 offset > arrayOopDesc::length_offset_in_bytes() ) {
1461 offset = Type::OffsetBot; // Flatten constant access into array body only
1462 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1463 }
1464 } else if( ta && _AliasLevel >= 2 ) {
1465 // For arrays indexed by constant indices, we flatten the alias
1466 // space to include all of the array body. Only the header, klass
1467 // and array length can be accessed un-aliased.
1468 if( offset != Type::OffsetBot ) {
1469 if( ta->const_oop() ) { // MethodData* or Method*
1470 offset = Type::OffsetBot; // Flatten constant access into array body
1471 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1472 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1473 // range is OK as-is.
1474 tj = ta = TypeAryPtr::RANGE;
1475 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1476 tj = TypeInstPtr::KLASS; // all klass loads look alike
1477 ta = TypeAryPtr::RANGE; // generic ignored junk
1478 ptr = TypePtr::BotPTR;
1479 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1480 tj = TypeInstPtr::MARK;
1481 ta = TypeAryPtr::RANGE; // generic ignored junk
1482 ptr = TypePtr::BotPTR;
1483 #if INCLUDE_SHENANDOAHGC
1484 } else if (offset == BrooksPointer::byte_offset() && UseShenandoahGC) {
1485 // Need to distinguish brooks ptr as is.
1486 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1487 #endif
1488 } else { // Random constant offset into array body
1489 offset = Type::OffsetBot; // Flatten constant access into array body
1490 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1491 }
1492 }
1493 // Arrays of fixed size alias with arrays of unknown size.
1494 if (ta->size() != TypeInt::POS) {
1495 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1496 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1497 }
1498 // Arrays of known objects become arrays of unknown objects.
1499 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1500 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1501 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1502 }
1503 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1504 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1505 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1506 }
1507 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1508 // cannot be distinguished by bytecode alone.
1509 if (ta->elem() == TypeInt::BOOL) {
1510 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1511 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1512 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1513 }
1514 // During the 2nd round of IterGVN, NotNull castings are removed.
1515 // Make sure the Bottom and NotNull variants alias the same.
1516 // Also, make sure exact and non-exact variants alias the same.
1517 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1518 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1519 }
1520 }
1521
1522 // Oop pointers need some flattening
1523 const TypeInstPtr *to = tj->isa_instptr();
1524 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1525 ciInstanceKlass *k = to->klass()->as_instance_klass();
1526 if( ptr == TypePtr::Constant ) {
1527 if (to->klass() != ciEnv::current()->Class_klass() ||
1528 offset < k->size_helper() * wordSize) {
1529 // No constant oop pointers (such as Strings); they alias with
1530 // unknown strings.
1531 assert(!is_known_inst, "not scalarizable allocation");
1532 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1533 }
1534 } else if( is_known_inst ) {
1535 tj = to; // Keep NotNull and klass_is_exact for instance type
1536 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1537 // During the 2nd round of IterGVN, NotNull castings are removed.
1538 // Make sure the Bottom and NotNull variants alias the same.
1539 // Also, make sure exact and non-exact variants alias the same.
1540 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1541 }
1542 if (to->speculative() != NULL) {
1543 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1544 }
1545 // Canonicalize the holder of this field
1546 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1547 // First handle header references such as a LoadKlassNode, even if the
1548 // object's klass is unloaded at compile time (4965979).
1549 if (!is_known_inst) { // Do it only for non-instance types
1550 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1551 }
1552 } else if (SHENANDOAHGC_ONLY((offset != BrooksPointer::byte_offset() || !UseShenandoahGC) &&) (offset < 0 || offset >= k->size_helper() * wordSize)) {
1553 // Static fields are in the space above the normal instance
1554 // fields in the java.lang.Class instance.
1555 if (to->klass() != ciEnv::current()->Class_klass()) {
1556 to = NULL;
1557 tj = TypeOopPtr::BOTTOM;
1558 offset = tj->offset();
1559 }
1560 } else {
1561 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1562 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1563 if( is_known_inst ) {
1564 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1565 } else {
1566 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1567 }
1568 }
1569 }
1570 }
1571
1572 // Klass pointers to object array klasses need some flattening
1573 const TypeKlassPtr *tk = tj->isa_klassptr();
1574 if( tk ) {
1575 // If we are referencing a field within a Klass, we need
1576 // to assume the worst case of an Object. Both exact and
1577 // inexact types must flatten to the same alias class so
1578 // use NotNull as the PTR.
1579 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1580
1581 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1582 TypeKlassPtr::OBJECT->klass(),
1583 offset);
1584 }
1585
1586 ciKlass* klass = tk->klass();
1587 if( klass->is_obj_array_klass() ) {
1588 ciKlass* k = TypeAryPtr::OOPS->klass();
1589 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1590 k = TypeInstPtr::BOTTOM->klass();
1591 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1592 }
1593
1594 // Check for precise loads from the primary supertype array and force them
1595 // to the supertype cache alias index. Check for generic array loads from
1596 // the primary supertype array and also force them to the supertype cache
1597 // alias index. Since the same load can reach both, we need to merge
1598 // these 2 disparate memories into the same alias class. Since the
1599 // primary supertype array is read-only, there's no chance of confusion
1600 // where we bypass an array load and an array store.
1601 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1602 if (offset == Type::OffsetBot ||
1603 (offset >= primary_supers_offset &&
1604 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1605 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1606 offset = in_bytes(Klass::secondary_super_cache_offset());
1607 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1608 }
1609 }
1610
1611 // Flatten all Raw pointers together.
1612 if (tj->base() == Type::RawPtr)
1613 tj = TypeRawPtr::BOTTOM;
1614
1615 if (tj->base() == Type::AnyPtr)
1616 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1617
1618 // Flatten all to bottom for now
1619 switch( _AliasLevel ) {
1620 case 0:
1621 tj = TypePtr::BOTTOM;
1622 break;
1623 case 1: // Flatten to: oop, static, field or array
1624 switch (tj->base()) {
1625 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1626 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1627 case Type::AryPtr: // do not distinguish arrays at all
1628 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1629 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1630 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1631 default: ShouldNotReachHere();
1632 }
1633 break;
1634 case 2: // No collapsing at level 2; keep all splits
1635 case 3: // No collapsing at level 3; keep all splits
1636 break;
1637 default:
1638 Unimplemented();
1639 }
1640
1641 offset = tj->offset();
1642 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1643
1644 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1645 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1646 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1647 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1648 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1649 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1650 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1651 (UseShenandoahGC SHENANDOAHGC_ONLY(&& offset == BrooksPointer::byte_offset() && tj->base() == Type::AryPtr)),
1652 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1653 assert( tj->ptr() != TypePtr::TopPTR &&
1654 tj->ptr() != TypePtr::AnyNull &&
1655 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1656 // assert( tj->ptr() != TypePtr::Constant ||
1657 // tj->base() == Type::RawPtr ||
1658 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1659
1660 return tj;
1661 }
1662
1663 void Compile::AliasType::Init(int i, const TypePtr* at) {
1664 _index = i;
1665 _adr_type = at;
1666 _field = NULL;
1667 _element = NULL;
1668 _is_rewritable = true; // default
1669 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1670 if (atoop != NULL && atoop->is_known_instance()) {
1671 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1672 _general_index = Compile::current()->get_alias_index(gt);
1673 } else {
1674 _general_index = 0;
1675 }
1676 }
1677
1678 BasicType Compile::AliasType::basic_type() const {
1679 if (element() != NULL) {
1680 const Type* element = adr_type()->is_aryptr()->elem();
1681 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1682 } if (field() != NULL) {
1683 return field()->layout_type();
1684 } else {
1685 return T_ILLEGAL; // unknown
1686 }
1687 }
1688
1689 //---------------------------------print_on------------------------------------
1690 #ifndef PRODUCT
1691 void Compile::AliasType::print_on(outputStream* st) {
1692 if (index() < 10)
1693 st->print("@ <%d> ", index());
1694 else st->print("@ <%d>", index());
1695 st->print(is_rewritable() ? " " : " RO");
1696 int offset = adr_type()->offset();
1697 if (offset == Type::OffsetBot)
1698 st->print(" +any");
1699 else st->print(" +%-3d", offset);
1700 st->print(" in ");
1701 adr_type()->dump_on(st);
1702 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1703 if (field() != NULL && tjp) {
1704 if (tjp->klass() != field()->holder() ||
1705 tjp->offset() != field()->offset_in_bytes()) {
1706 st->print(" != ");
1707 field()->print();
1708 st->print(" ***");
1709 }
1710 }
1711 }
1712
1713 void print_alias_types() {
1714 Compile* C = Compile::current();
1715 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1716 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1717 C->alias_type(idx)->print_on(tty);
1718 tty->cr();
1719 }
1720 }
1721 #endif
1722
1723
1724 //----------------------------probe_alias_cache--------------------------------
1725 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1726 intptr_t key = (intptr_t) adr_type;
1727 key ^= key >> logAliasCacheSize;
1728 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1729 }
1730
1731
1732 //-----------------------------grow_alias_types--------------------------------
1733 void Compile::grow_alias_types() {
1734 const int old_ats = _max_alias_types; // how many before?
1735 const int new_ats = old_ats; // how many more?
1736 const int grow_ats = old_ats+new_ats; // how many now?
1737 _max_alias_types = grow_ats;
1738 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1739 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1740 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1741 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1742 }
1743
1744
1745 //--------------------------------find_alias_type------------------------------
1746 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1747 if (_AliasLevel == 0)
1748 return alias_type(AliasIdxBot);
1749
1750 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1751 if (ace->_adr_type == adr_type) {
1752 return alias_type(ace->_index);
1753 }
1754
1755 // Handle special cases.
1756 if (adr_type == NULL) return alias_type(AliasIdxTop);
1757 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1758
1759 // Do it the slow way.
1760 const TypePtr* flat = flatten_alias_type(adr_type);
1761
1762 #ifdef ASSERT
1763 {
1764 ResourceMark rm;
1765 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1766 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1767 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1768 Type::str(adr_type));
1769 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1770 const TypeOopPtr* foop = flat->is_oopptr();
1771 // Scalarizable allocations have exact klass always.
1772 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1773 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1774 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1775 Type::str(foop), Type::str(xoop));
1776 }
1777 }
1778 #endif
1779
1780 int idx = AliasIdxTop;
1781 for (int i = 0; i < num_alias_types(); i++) {
1782 if (alias_type(i)->adr_type() == flat) {
1783 idx = i;
1784 break;
1785 }
1786 }
1787
1788 if (idx == AliasIdxTop) {
1789 if (no_create) return NULL;
1790 // Grow the array if necessary.
1791 if (_num_alias_types == _max_alias_types) grow_alias_types();
1792 // Add a new alias type.
1793 idx = _num_alias_types++;
1794 _alias_types[idx]->Init(idx, flat);
1795 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1796 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1797 if (flat->isa_instptr()) {
1798 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1799 && flat->is_instptr()->klass() == env()->Class_klass())
1800 alias_type(idx)->set_rewritable(false);
1801 }
1802 if (flat->isa_aryptr()) {
1803 #ifdef ASSERT
1804 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1805 // (T_BYTE has the weakest alignment and size restrictions...)
1806 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1807 #endif
1808 if (flat->offset() == TypePtr::OffsetBot) {
1809 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1810 }
1811 }
1812 if (flat->isa_klassptr()) {
1813 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1814 alias_type(idx)->set_rewritable(false);
1815 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1816 alias_type(idx)->set_rewritable(false);
1817 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1818 alias_type(idx)->set_rewritable(false);
1819 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1820 alias_type(idx)->set_rewritable(false);
1821 }
1822 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1823 // but the base pointer type is not distinctive enough to identify
1824 // references into JavaThread.)
1825
1826 // Check for final fields.
1827 const TypeInstPtr* tinst = flat->isa_instptr();
1828 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1829 ciField* field;
1830 if (tinst->const_oop() != NULL &&
1831 tinst->klass() == ciEnv::current()->Class_klass() &&
1832 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1833 // static field
1834 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1835 field = k->get_field_by_offset(tinst->offset(), true);
1836 } else {
1837 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1838 field = k->get_field_by_offset(tinst->offset(), false);
1839 }
1840 assert(field == NULL ||
1841 original_field == NULL ||
1842 (field->holder() == original_field->holder() &&
1843 field->offset() == original_field->offset() &&
1844 field->is_static() == original_field->is_static()), "wrong field?");
1845 // Set field() and is_rewritable() attributes.
1846 if (field != NULL) alias_type(idx)->set_field(field);
1847 }
1848 }
1849
1850 // Fill the cache for next time.
1851 ace->_adr_type = adr_type;
1852 ace->_index = idx;
1853 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1854
1855 // Might as well try to fill the cache for the flattened version, too.
1856 AliasCacheEntry* face = probe_alias_cache(flat);
1857 if (face->_adr_type == NULL) {
1858 face->_adr_type = flat;
1859 face->_index = idx;
1860 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1861 }
1862
1863 return alias_type(idx);
1864 }
1865
1866
1867 Compile::AliasType* Compile::alias_type(ciField* field) {
1868 const TypeOopPtr* t;
1869 if (field->is_static())
1870 t = TypeInstPtr::make(field->holder()->java_mirror());
1871 else
1872 t = TypeOopPtr::make_from_klass_raw(field->holder());
1873 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1874 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1875 return atp;
1876 }
1877
1878
1879 //------------------------------have_alias_type--------------------------------
1880 bool Compile::have_alias_type(const TypePtr* adr_type) {
1881 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1882 if (ace->_adr_type == adr_type) {
1883 return true;
1884 }
1885
1886 // Handle special cases.
1887 if (adr_type == NULL) return true;
1888 if (adr_type == TypePtr::BOTTOM) return true;
1889
1890 return find_alias_type(adr_type, true, NULL) != NULL;
1891 }
1892
1893 //-----------------------------must_alias--------------------------------------
1894 // True if all values of the given address type are in the given alias category.
1895 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1896 if (alias_idx == AliasIdxBot) return true; // the universal category
1897 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1898 if (alias_idx == AliasIdxTop) return false; // the empty category
1899 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1900
1901 // the only remaining possible overlap is identity
1902 int adr_idx = get_alias_index(adr_type);
1903 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1904 assert(adr_idx == alias_idx ||
1905 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1906 && adr_type != TypeOopPtr::BOTTOM),
1907 "should not be testing for overlap with an unsafe pointer");
1908 return adr_idx == alias_idx;
1909 }
1910
1911 //------------------------------can_alias--------------------------------------
1912 // True if any values of the given address type are in the given alias category.
1913 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1914 if (alias_idx == AliasIdxTop) return false; // the empty category
1915 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1916 if (alias_idx == AliasIdxBot) return true; // the universal category
1917 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1918
1919 // the only remaining possible overlap is identity
1920 int adr_idx = get_alias_index(adr_type);
1921 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1922 return adr_idx == alias_idx;
1923 }
1924
1925
1926
1927 //---------------------------pop_warm_call-------------------------------------
1928 WarmCallInfo* Compile::pop_warm_call() {
1929 WarmCallInfo* wci = _warm_calls;
1930 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1931 return wci;
1932 }
1933
1934 //----------------------------Inline_Warm--------------------------------------
1935 int Compile::Inline_Warm() {
1936 // If there is room, try to inline some more warm call sites.
1937 // %%% Do a graph index compaction pass when we think we're out of space?
1938 if (!InlineWarmCalls) return 0;
1939
1940 int calls_made_hot = 0;
1941 int room_to_grow = NodeCountInliningCutoff - unique();
1942 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1943 int amount_grown = 0;
1944 WarmCallInfo* call;
1945 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1946 int est_size = (int)call->size();
1947 if (est_size > (room_to_grow - amount_grown)) {
1948 // This one won't fit anyway. Get rid of it.
1949 call->make_cold();
1950 continue;
1951 }
1952 call->make_hot();
1953 calls_made_hot++;
1954 amount_grown += est_size;
1955 amount_to_grow -= est_size;
1956 }
1957
1958 if (calls_made_hot > 0) set_major_progress();
1959 return calls_made_hot;
1960 }
1961
1962
1963 //----------------------------Finish_Warm--------------------------------------
1964 void Compile::Finish_Warm() {
1965 if (!InlineWarmCalls) return;
1966 if (failing()) return;
1967 if (warm_calls() == NULL) return;
1968
1969 // Clean up loose ends, if we are out of space for inlining.
1970 WarmCallInfo* call;
1971 while ((call = pop_warm_call()) != NULL) {
1972 call->make_cold();
1973 }
1974 }
1975
1976 //---------------------cleanup_loop_predicates-----------------------
1977 // Remove the opaque nodes that protect the predicates so that all unused
1978 // checks and uncommon_traps will be eliminated from the ideal graph
1979 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1980 if (predicate_count()==0) return;
1981 for (int i = predicate_count(); i > 0; i--) {
1982 Node * n = predicate_opaque1_node(i-1);
1983 assert(n->Opcode() == Op_Opaque1, "must be");
1984 igvn.replace_node(n, n->in(1));
1985 }
1986 assert(predicate_count()==0, "should be clean!");
1987 }
1988
1989 void Compile::add_range_check_cast(Node* n) {
1990 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1991 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1992 _range_check_casts->append(n);
1993 }
1994
1995 // Remove all range check dependent CastIINodes.
1996 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1997 for (int i = range_check_cast_count(); i > 0; i--) {
1998 Node* cast = range_check_cast_node(i-1);
1999 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2000 igvn.replace_node(cast, cast->in(1));
2001 }
2002 assert(range_check_cast_count() == 0, "should be empty");
2003 }
2004
2005 void Compile::add_opaque4_node(Node* n) {
2006 assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2007 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2008 _opaque4_nodes->append(n);
2009 }
2010
2011 // Remove all Opaque4 nodes.
2012 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2013 for (int i = opaque4_count(); i > 0; i--) {
2014 Node* opaq = opaque4_node(i-1);
2015 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2016 igvn.replace_node(opaq, opaq->in(2));
2017 }
2018 assert(opaque4_count() == 0, "should be empty");
2019 }
2020
2021 // StringOpts and late inlining of string methods
2022 void Compile::inline_string_calls(bool parse_time) {
2023 {
2024 // remove useless nodes to make the usage analysis simpler
2025 ResourceMark rm;
2026 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2027 }
2028
2029 {
2030 ResourceMark rm;
2031 print_method(PHASE_BEFORE_STRINGOPTS, 3);
2032 PhaseStringOpts pso(initial_gvn(), for_igvn());
2033 print_method(PHASE_AFTER_STRINGOPTS, 3);
2034 }
2035
2036 // now inline anything that we skipped the first time around
2037 if (!parse_time) {
2038 _late_inlines_pos = _late_inlines.length();
2039 }
2040
2041 while (_string_late_inlines.length() > 0) {
2042 CallGenerator* cg = _string_late_inlines.pop();
2043 cg->do_late_inline();
2044 if (failing()) return;
2045 }
2046 _string_late_inlines.trunc_to(0);
2047 }
2048
2049 // Late inlining of boxing methods
2050 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2051 if (_boxing_late_inlines.length() > 0) {
2052 assert(has_boxed_value(), "inconsistent");
2053
2054 PhaseGVN* gvn = initial_gvn();
2055 set_inlining_incrementally(true);
2056
2057 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2058 for_igvn()->clear();
2059 gvn->replace_with(&igvn);
2060
2061 _late_inlines_pos = _late_inlines.length();
2062
2063 while (_boxing_late_inlines.length() > 0) {
2064 CallGenerator* cg = _boxing_late_inlines.pop();
2065 cg->do_late_inline();
2066 if (failing()) return;
2067 }
2068 _boxing_late_inlines.trunc_to(0);
2069
2070 {
2071 ResourceMark rm;
2072 PhaseRemoveUseless pru(gvn, for_igvn());
2073 }
2074
2075 igvn = PhaseIterGVN(gvn);
2076 igvn.optimize();
2077
2078 set_inlining_progress(false);
2079 set_inlining_incrementally(false);
2080 }
2081 }
2082
2083 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2084 assert(IncrementalInline, "incremental inlining should be on");
2085 PhaseGVN* gvn = initial_gvn();
2086
2087 set_inlining_progress(false);
2088 for_igvn()->clear();
2089 gvn->replace_with(&igvn);
2090
2091 {
2092 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2093 int i = 0;
2094 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2095 CallGenerator* cg = _late_inlines.at(i);
2096 _late_inlines_pos = i+1;
2097 cg->do_late_inline();
2098 if (failing()) return;
2099 }
2100 int j = 0;
2101 for (; i < _late_inlines.length(); i++, j++) {
2102 _late_inlines.at_put(j, _late_inlines.at(i));
2103 }
2104 _late_inlines.trunc_to(j);
2105 }
2106
2107 {
2108 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2109 ResourceMark rm;
2110 PhaseRemoveUseless pru(gvn, for_igvn());
2111 }
2112
2113 {
2114 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2115 igvn = PhaseIterGVN(gvn);
2116 }
2117 }
2118
2119 // Perform incremental inlining until bound on number of live nodes is reached
2120 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2121 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2122
2123 PhaseGVN* gvn = initial_gvn();
2124
2125 set_inlining_incrementally(true);
2126 set_inlining_progress(true);
2127 uint low_live_nodes = 0;
2128
2129 while(inlining_progress() && _late_inlines.length() > 0) {
2130
2131 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2132 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2133 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2134 // PhaseIdealLoop is expensive so we only try it once we are
2135 // out of live nodes and we only try it again if the previous
2136 // helped got the number of nodes down significantly
2137 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2138 if (failing()) return;
2139 low_live_nodes = live_nodes();
2140 _major_progress = true;
2141 }
2142
2143 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2144 break;
2145 }
2146 }
2147
2148 inline_incrementally_one(igvn);
2149
2150 if (failing()) return;
2151
2152 {
2153 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2154 igvn.optimize();
2155 }
2156
2157 if (failing()) return;
2158 }
2159
2160 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2161
2162 if (_string_late_inlines.length() > 0) {
2163 assert(has_stringbuilder(), "inconsistent");
2164 for_igvn()->clear();
2165 initial_gvn()->replace_with(&igvn);
2166
2167 inline_string_calls(false);
2168
2169 if (failing()) return;
2170
2171 {
2172 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2173 ResourceMark rm;
2174 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2175 }
2176
2177 {
2178 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2179 igvn = PhaseIterGVN(gvn);
2180 igvn.optimize();
2181 }
2182 }
2183
2184 set_inlining_incrementally(false);
2185 }
2186
2187
2188 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2189 if(loop_opts_cnt > 0) {
2190 debug_only( int cnt = 0; );
2191 while(major_progress() && (loop_opts_cnt > 0)) {
2192 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2193 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2194 PhaseIdealLoop ideal_loop(igvn, mode);
2195 loop_opts_cnt--;
2196 if (failing()) return false;
2197 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2198 }
2199 }
2200 return true;
2201 }
2202
2203 //------------------------------Optimize---------------------------------------
2204 // Given a graph, optimize it.
2205 void Compile::Optimize() {
2206 TracePhase tp("optimizer", &timers[_t_optimizer]);
2207
2208 #ifndef PRODUCT
2209 if (_directive->BreakAtCompileOption) {
2210 BREAKPOINT;
2211 }
2212
2213 #endif
2214
2215 #ifdef ASSERT
2216 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2217 bs->verify_gc_barriers(true);
2218 #endif
2219
2220 ResourceMark rm;
2221 int loop_opts_cnt;
2222
2223 print_inlining_reinit();
2224
2225 NOT_PRODUCT( verify_graph_edges(); )
2226
2227 print_method(PHASE_AFTER_PARSING);
2228
2229 {
2230 // Iterative Global Value Numbering, including ideal transforms
2231 // Initialize IterGVN with types and values from parse-time GVN
2232 PhaseIterGVN igvn(initial_gvn());
2233 #ifdef ASSERT
2234 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2235 #endif
2236 {
2237 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2238 igvn.optimize();
2239 }
2240
2241 if (failing()) return;
2242
2243 print_method(PHASE_ITER_GVN1, 2);
2244
2245 inline_incrementally(igvn);
2246
2247 print_method(PHASE_INCREMENTAL_INLINE, 2);
2248
2249 if (failing()) return;
2250
2251 if (eliminate_boxing()) {
2252 // Inline valueOf() methods now.
2253 inline_boxing_calls(igvn);
2254
2255 if (AlwaysIncrementalInline) {
2256 inline_incrementally(igvn);
2257 }
2258
2259 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2260
2261 if (failing()) return;
2262 }
2263
2264 // Remove the speculative part of types and clean up the graph from
2265 // the extra CastPP nodes whose only purpose is to carry them. Do
2266 // that early so that optimizations are not disrupted by the extra
2267 // CastPP nodes.
2268 remove_speculative_types(igvn);
2269
2270 // No more new expensive nodes will be added to the list from here
2271 // so keep only the actual candidates for optimizations.
2272 cleanup_expensive_nodes(igvn);
2273
2274 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2275 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2276 initial_gvn()->replace_with(&igvn);
2277 for_igvn()->clear();
2278 Unique_Node_List new_worklist(C->comp_arena());
2279 {
2280 ResourceMark rm;
2281 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2282 }
2283 set_for_igvn(&new_worklist);
2284 igvn = PhaseIterGVN(initial_gvn());
2285 igvn.optimize();
2286 }
2287
2288 // Perform escape analysis
2289 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2290 if (has_loops()) {
2291 // Cleanup graph (remove dead nodes).
2292 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2293 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2294 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2295 if (failing()) return;
2296 }
2297 ConnectionGraph::do_analysis(this, &igvn);
2298
2299 if (failing()) return;
2300
2301 // Optimize out fields loads from scalar replaceable allocations.
2302 igvn.optimize();
2303 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2304
2305 if (failing()) return;
2306
2307 if (congraph() != NULL && macro_count() > 0) {
2308 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2309 PhaseMacroExpand mexp(igvn);
2310 mexp.eliminate_macro_nodes();
2311 igvn.set_delay_transform(false);
2312
2313 igvn.optimize();
2314 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2315
2316 if (failing()) return;
2317 }
2318 }
2319
2320 // Loop transforms on the ideal graph. Range Check Elimination,
2321 // peeling, unrolling, etc.
2322
2323 // Set loop opts counter
2324 loop_opts_cnt = num_loop_opts();
2325 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2326 {
2327 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2328 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2329 loop_opts_cnt--;
2330 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2331 if (failing()) return;
2332 }
2333 // Loop opts pass if partial peeling occurred in previous pass
2334 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2335 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2336 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2337 loop_opts_cnt--;
2338 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2339 if (failing()) return;
2340 }
2341 // Loop opts pass for loop-unrolling before CCP
2342 if(major_progress() && (loop_opts_cnt > 0)) {
2343 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2344 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2345 loop_opts_cnt--;
2346 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2347 }
2348 if (!failing()) {
2349 // Verify that last round of loop opts produced a valid graph
2350 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2351 PhaseIdealLoop::verify(igvn);
2352 }
2353 }
2354 if (failing()) return;
2355
2356 // Conditional Constant Propagation;
2357 PhaseCCP ccp( &igvn );
2358 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2359 {
2360 TracePhase tp("ccp", &timers[_t_ccp]);
2361 ccp.do_transform();
2362 }
2363 print_method(PHASE_CPP1, 2);
2364
2365 assert( true, "Break here to ccp.dump_old2new_map()");
2366
2367 // Iterative Global Value Numbering, including ideal transforms
2368 {
2369 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2370 igvn = ccp;
2371 igvn.optimize();
2372 }
2373
2374 print_method(PHASE_ITER_GVN2, 2);
2375
2376 if (failing()) return;
2377
2378 // Loop transforms on the ideal graph. Range Check Elimination,
2379 // peeling, unrolling, etc.
2380 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2381 return;
2382 }
2383
2384 #if INCLUDE_ZGC
2385 if (UseZGC) {
2386 ZBarrierSetC2::find_dominating_barriers(igvn);
2387 }
2388 #endif
2389
2390 if (failing()) return;
2391
2392 // Ensure that major progress is now clear
2393 C->clear_major_progress();
2394
2395 {
2396 // Verify that all previous optimizations produced a valid graph
2397 // at least to this point, even if no loop optimizations were done.
2398 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2399 PhaseIdealLoop::verify(igvn);
2400 }
2401
2402 if (range_check_cast_count() > 0) {
2403 // No more loop optimizations. Remove all range check dependent CastIINodes.
2404 C->remove_range_check_casts(igvn);
2405 igvn.optimize();
2406 }
2407
2408 #ifdef ASSERT
2409 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2410 bs->verify_gc_barriers(false);
2411 #endif
2412
2413 {
2414 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2415 PhaseMacroExpand mex(igvn);
2416 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2417 if (mex.expand_macro_nodes()) {
2418 assert(failing(), "must bail out w/ explicit message");
2419 return;
2420 }
2421 }
2422
2423 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2);
2424
2425 #if INCLUDE_SHENANDOAHGC
2426 if (!ShenandoahWriteBarrierNode::expand(this, igvn, loop_opts_cnt)) {
2427 assert(failing(), "must bail out w/ explicit message");
2428 return;
2429 }
2430 #endif
2431
2432 if (opaque4_count() > 0) {
2433 C->remove_opaque4_nodes(igvn);
2434 igvn.optimize();
2435 }
2436
2437 DEBUG_ONLY( _modified_nodes = NULL; )
2438 } // (End scope of igvn; run destructor if necessary for asserts.)
2439
2440 process_print_inlining();
2441 // A method with only infinite loops has no edges entering loops from root
2442 {
2443 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2444 if (final_graph_reshaping()) {
2445 assert(failing(), "must bail out w/ explicit message");
2446 return;
2447 }
2448 }
2449
2450 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2451 }
2452
2453
2454 //------------------------------Code_Gen---------------------------------------
2455 // Given a graph, generate code for it
2456 void Compile::Code_Gen() {
2457 if (failing()) {
2458 return;
2459 }
2460
2461 // Perform instruction selection. You might think we could reclaim Matcher
2462 // memory PDQ, but actually the Matcher is used in generating spill code.
2463 // Internals of the Matcher (including some VectorSets) must remain live
2464 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2465 // set a bit in reclaimed memory.
2466
2467 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2468 // nodes. Mapping is only valid at the root of each matched subtree.
2469 NOT_PRODUCT( verify_graph_edges(); )
2470
2471 Matcher matcher;
2472 _matcher = &matcher;
2473 {
2474 TracePhase tp("matcher", &timers[_t_matcher]);
2475 matcher.match();
2476 }
2477 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2478 // nodes. Mapping is only valid at the root of each matched subtree.
2479 NOT_PRODUCT( verify_graph_edges(); )
2480
2481 // If you have too many nodes, or if matching has failed, bail out
2482 check_node_count(0, "out of nodes matching instructions");
2483 if (failing()) {
2484 return;
2485 }
2486
2487 print_method(PHASE_MATCHING, 2);
2488
2489 // Build a proper-looking CFG
2490 PhaseCFG cfg(node_arena(), root(), matcher);
2491 _cfg = &cfg;
2492 {
2493 TracePhase tp("scheduler", &timers[_t_scheduler]);
2494 bool success = cfg.do_global_code_motion();
2495 if (!success) {
2496 return;
2497 }
2498
2499 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2500 NOT_PRODUCT( verify_graph_edges(); )
2501 debug_only( cfg.verify(); )
2502 }
2503
2504 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2505 _regalloc = ®alloc;
2506 {
2507 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2508 // Perform register allocation. After Chaitin, use-def chains are
2509 // no longer accurate (at spill code) and so must be ignored.
2510 // Node->LRG->reg mappings are still accurate.
2511 _regalloc->Register_Allocate();
2512
2513 // Bail out if the allocator builds too many nodes
2514 if (failing()) {
2515 return;
2516 }
2517 }
2518
2519 // Prior to register allocation we kept empty basic blocks in case the
2520 // the allocator needed a place to spill. After register allocation we
2521 // are not adding any new instructions. If any basic block is empty, we
2522 // can now safely remove it.
2523 {
2524 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2525 cfg.remove_empty_blocks();
2526 if (do_freq_based_layout()) {
2527 PhaseBlockLayout layout(cfg);
2528 } else {
2529 cfg.set_loop_alignment();
2530 }
2531 cfg.fixup_flow();
2532 }
2533
2534 // Apply peephole optimizations
2535 if( OptoPeephole ) {
2536 TracePhase tp("peephole", &timers[_t_peephole]);
2537 PhasePeephole peep( _regalloc, cfg);
2538 peep.do_transform();
2539 }
2540
2541 // Do late expand if CPU requires this.
2542 if (Matcher::require_postalloc_expand) {
2543 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2544 cfg.postalloc_expand(_regalloc);
2545 }
2546
2547 // Convert Nodes to instruction bits in a buffer
2548 {
2549 TraceTime tp("output", &timers[_t_output], CITime);
2550 Output();
2551 }
2552
2553 print_method(PHASE_FINAL_CODE);
2554
2555 // He's dead, Jim.
2556 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2557 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2558 }
2559
2560
2561 //------------------------------dump_asm---------------------------------------
2562 // Dump formatted assembly
2563 #ifndef PRODUCT
2564 void Compile::dump_asm(int *pcs, uint pc_limit) {
2565 bool cut_short = false;
2566 tty->print_cr("#");
2567 tty->print("# "); _tf->dump(); tty->cr();
2568 tty->print_cr("#");
2569
2570 // For all blocks
2571 int pc = 0x0; // Program counter
2572 char starts_bundle = ' ';
2573 _regalloc->dump_frame();
2574
2575 Node *n = NULL;
2576 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2577 if (VMThread::should_terminate()) {
2578 cut_short = true;
2579 break;
2580 }
2581 Block* block = _cfg->get_block(i);
2582 if (block->is_connector() && !Verbose) {
2583 continue;
2584 }
2585 n = block->head();
2586 if (pcs && n->_idx < pc_limit) {
2587 tty->print("%3.3x ", pcs[n->_idx]);
2588 } else {
2589 tty->print(" ");
2590 }
2591 block->dump_head(_cfg);
2592 if (block->is_connector()) {
2593 tty->print_cr(" # Empty connector block");
2594 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2595 tty->print_cr(" # Block is sole successor of call");
2596 }
2597
2598 // For all instructions
2599 Node *delay = NULL;
2600 for (uint j = 0; j < block->number_of_nodes(); j++) {
2601 if (VMThread::should_terminate()) {
2602 cut_short = true;
2603 break;
2604 }
2605 n = block->get_node(j);
2606 if (valid_bundle_info(n)) {
2607 Bundle* bundle = node_bundling(n);
2608 if (bundle->used_in_unconditional_delay()) {
2609 delay = n;
2610 continue;
2611 }
2612 if (bundle->starts_bundle()) {
2613 starts_bundle = '+';
2614 }
2615 }
2616
2617 if (WizardMode) {
2618 n->dump();
2619 }
2620
2621 if( !n->is_Region() && // Dont print in the Assembly
2622 !n->is_Phi() && // a few noisely useless nodes
2623 !n->is_Proj() &&
2624 !n->is_MachTemp() &&
2625 !n->is_SafePointScalarObject() &&
2626 !n->is_Catch() && // Would be nice to print exception table targets
2627 !n->is_MergeMem() && // Not very interesting
2628 !n->is_top() && // Debug info table constants
2629 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2630 ) {
2631 if (pcs && n->_idx < pc_limit)
2632 tty->print("%3.3x", pcs[n->_idx]);
2633 else
2634 tty->print(" ");
2635 tty->print(" %c ", starts_bundle);
2636 starts_bundle = ' ';
2637 tty->print("\t");
2638 n->format(_regalloc, tty);
2639 tty->cr();
2640 }
2641
2642 // If we have an instruction with a delay slot, and have seen a delay,
2643 // then back up and print it
2644 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2645 assert(delay != NULL, "no unconditional delay instruction");
2646 if (WizardMode) delay->dump();
2647
2648 if (node_bundling(delay)->starts_bundle())
2649 starts_bundle = '+';
2650 if (pcs && n->_idx < pc_limit)
2651 tty->print("%3.3x", pcs[n->_idx]);
2652 else
2653 tty->print(" ");
2654 tty->print(" %c ", starts_bundle);
2655 starts_bundle = ' ';
2656 tty->print("\t");
2657 delay->format(_regalloc, tty);
2658 tty->cr();
2659 delay = NULL;
2660 }
2661
2662 // Dump the exception table as well
2663 if( n->is_Catch() && (Verbose || WizardMode) ) {
2664 // Print the exception table for this offset
2665 _handler_table.print_subtable_for(pc);
2666 }
2667 }
2668
2669 if (pcs && n->_idx < pc_limit)
2670 tty->print_cr("%3.3x", pcs[n->_idx]);
2671 else
2672 tty->cr();
2673
2674 assert(cut_short || delay == NULL, "no unconditional delay branch");
2675
2676 } // End of per-block dump
2677 tty->cr();
2678
2679 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2680 }
2681 #endif
2682
2683 //------------------------------Final_Reshape_Counts---------------------------
2684 // This class defines counters to help identify when a method
2685 // may/must be executed using hardware with only 24-bit precision.
2686 struct Final_Reshape_Counts : public StackObj {
2687 int _call_count; // count non-inlined 'common' calls
2688 int _float_count; // count float ops requiring 24-bit precision
2689 int _double_count; // count double ops requiring more precision
2690 int _java_call_count; // count non-inlined 'java' calls
2691 int _inner_loop_count; // count loops which need alignment
2692 VectorSet _visited; // Visitation flags
2693 Node_List _tests; // Set of IfNodes & PCTableNodes
2694
2695 Final_Reshape_Counts() :
2696 _call_count(0), _float_count(0), _double_count(0),
2697 _java_call_count(0), _inner_loop_count(0),
2698 _visited( Thread::current()->resource_area() ) { }
2699
2700 void inc_call_count () { _call_count ++; }
2701 void inc_float_count () { _float_count ++; }
2702 void inc_double_count() { _double_count++; }
2703 void inc_java_call_count() { _java_call_count++; }
2704 void inc_inner_loop_count() { _inner_loop_count++; }
2705
2706 int get_call_count () const { return _call_count ; }
2707 int get_float_count () const { return _float_count ; }
2708 int get_double_count() const { return _double_count; }
2709 int get_java_call_count() const { return _java_call_count; }
2710 int get_inner_loop_count() const { return _inner_loop_count; }
2711 };
2712
2713 #ifdef ASSERT
2714 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2715 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2716 // Make sure the offset goes inside the instance layout.
2717 return k->contains_field_offset(tp->offset());
2718 // Note that OffsetBot and OffsetTop are very negative.
2719 }
2720 #endif
2721
2722 // Eliminate trivially redundant StoreCMs and accumulate their
2723 // precedence edges.
2724 void Compile::eliminate_redundant_card_marks(Node* n) {
2725 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2726 if (n->in(MemNode::Address)->outcnt() > 1) {
2727 // There are multiple users of the same address so it might be
2728 // possible to eliminate some of the StoreCMs
2729 Node* mem = n->in(MemNode::Memory);
2730 Node* adr = n->in(MemNode::Address);
2731 Node* val = n->in(MemNode::ValueIn);
2732 Node* prev = n;
2733 bool done = false;
2734 // Walk the chain of StoreCMs eliminating ones that match. As
2735 // long as it's a chain of single users then the optimization is
2736 // safe. Eliminating partially redundant StoreCMs would require
2737 // cloning copies down the other paths.
2738 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2739 if (adr == mem->in(MemNode::Address) &&
2740 val == mem->in(MemNode::ValueIn)) {
2741 // redundant StoreCM
2742 if (mem->req() > MemNode::OopStore) {
2743 // Hasn't been processed by this code yet.
2744 n->add_prec(mem->in(MemNode::OopStore));
2745 } else {
2746 // Already converted to precedence edge
2747 for (uint i = mem->req(); i < mem->len(); i++) {
2748 // Accumulate any precedence edges
2749 if (mem->in(i) != NULL) {
2750 n->add_prec(mem->in(i));
2751 }
2752 }
2753 // Everything above this point has been processed.
2754 done = true;
2755 }
2756 // Eliminate the previous StoreCM
2757 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2758 assert(mem->outcnt() == 0, "should be dead");
2759 mem->disconnect_inputs(NULL, this);
2760 } else {
2761 prev = mem;
2762 }
2763 mem = prev->in(MemNode::Memory);
2764 }
2765 }
2766 }
2767
2768 //------------------------------final_graph_reshaping_impl----------------------
2769 // Implement items 1-5 from final_graph_reshaping below.
2770 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2771
2772 if ( n->outcnt() == 0 ) return; // dead node
2773 uint nop = n->Opcode();
2774
2775 // Check for 2-input instruction with "last use" on right input.
2776 // Swap to left input. Implements item (2).
2777 if( n->req() == 3 && // two-input instruction
2778 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2779 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2780 n->in(2)->outcnt() == 1 &&// right use IS a last use
2781 !n->in(2)->is_Con() ) { // right use is not a constant
2782 // Check for commutative opcode
2783 switch( nop ) {
2784 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2785 case Op_MaxI: case Op_MinI:
2786 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2787 case Op_AndL: case Op_XorL: case Op_OrL:
2788 case Op_AndI: case Op_XorI: case Op_OrI: {
2789 // Move "last use" input to left by swapping inputs
2790 n->swap_edges(1, 2);
2791 break;
2792 }
2793 default:
2794 break;
2795 }
2796 }
2797
2798 #ifdef ASSERT
2799 if( n->is_Mem() ) {
2800 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2801 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2802 // oop will be recorded in oop map if load crosses safepoint
2803 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2804 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2805 "raw memory operations should have control edge");
2806 }
2807 if (n->is_MemBar()) {
2808 MemBarNode* mb = n->as_MemBar();
2809 if (mb->trailing_store() || mb->trailing_load_store()) {
2810 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2811 Node* mem = mb->in(MemBarNode::Precedent);
2812 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2813 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2814 } else if (mb->leading()) {
2815 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2816 }
2817 }
2818 #endif
2819 // Count FPU ops and common calls, implements item (3)
2820 switch( nop ) {
2821 // Count all float operations that may use FPU
2822 case Op_AddF:
2823 case Op_SubF:
2824 case Op_MulF:
2825 case Op_DivF:
2826 case Op_NegF:
2827 case Op_ModF:
2828 case Op_ConvI2F:
2829 case Op_ConF:
2830 case Op_CmpF:
2831 case Op_CmpF3:
2832 // case Op_ConvL2F: // longs are split into 32-bit halves
2833 frc.inc_float_count();
2834 break;
2835
2836 case Op_ConvF2D:
2837 case Op_ConvD2F:
2838 frc.inc_float_count();
2839 frc.inc_double_count();
2840 break;
2841
2842 // Count all double operations that may use FPU
2843 case Op_AddD:
2844 case Op_SubD:
2845 case Op_MulD:
2846 case Op_DivD:
2847 case Op_NegD:
2848 case Op_ModD:
2849 case Op_ConvI2D:
2850 case Op_ConvD2I:
2851 // case Op_ConvL2D: // handled by leaf call
2852 // case Op_ConvD2L: // handled by leaf call
2853 case Op_ConD:
2854 case Op_CmpD:
2855 case Op_CmpD3:
2856 frc.inc_double_count();
2857 break;
2858 case Op_Opaque1: // Remove Opaque Nodes before matching
2859 case Op_Opaque2: // Remove Opaque Nodes before matching
2860 case Op_Opaque3:
2861 n->subsume_by(n->in(1), this);
2862 break;
2863 case Op_CallStaticJava:
2864 case Op_CallJava:
2865 case Op_CallDynamicJava:
2866 frc.inc_java_call_count(); // Count java call site;
2867 case Op_CallRuntime:
2868 case Op_CallLeaf:
2869 case Op_CallLeafNoFP: {
2870 assert (n->is_Call(), "");
2871 CallNode *call = n->as_Call();
2872 #if INCLUDE_SHENANDOAHGC
2873 if (UseShenandoahGC && call->is_shenandoah_wb_pre_call()) {
2874 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
2875 if (call->req() > cnt) {
2876 assert(call->req() == cnt+1, "only one extra input");
2877 Node* addp = call->in(cnt);
2878 assert(!CallLeafNode::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
2879 call->del_req(cnt);
2880 }
2881 }
2882 #endif
2883 // Count call sites where the FP mode bit would have to be flipped.
2884 // Do not count uncommon runtime calls:
2885 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2886 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2887 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2888 frc.inc_call_count(); // Count the call site
2889 } else { // See if uncommon argument is shared
2890 Node *n = call->in(TypeFunc::Parms);
2891 int nop = n->Opcode();
2892 // Clone shared simple arguments to uncommon calls, item (1).
2893 if (n->outcnt() > 1 &&
2894 !n->is_Proj() &&
2895 nop != Op_CreateEx &&
2896 nop != Op_CheckCastPP &&
2897 nop != Op_DecodeN &&
2898 nop != Op_DecodeNKlass &&
2899 !n->is_Mem() &&
2900 !n->is_Phi()) {
2901 Node *x = n->clone();
2902 call->set_req(TypeFunc::Parms, x);
2903 }
2904 }
2905 break;
2906 }
2907
2908 case Op_StoreD:
2909 case Op_LoadD:
2910 case Op_LoadD_unaligned:
2911 frc.inc_double_count();
2912 goto handle_mem;
2913 case Op_StoreF:
2914 case Op_LoadF:
2915 frc.inc_float_count();
2916 goto handle_mem;
2917
2918 case Op_StoreCM:
2919 {
2920 // Convert OopStore dependence into precedence edge
2921 Node* prec = n->in(MemNode::OopStore);
2922 n->del_req(MemNode::OopStore);
2923 n->add_prec(prec);
2924 eliminate_redundant_card_marks(n);
2925 }
2926
2927 // fall through
2928
2929 case Op_StoreB:
2930 case Op_StoreC:
2931 case Op_StorePConditional:
2932 case Op_StoreI:
2933 case Op_StoreL:
2934 case Op_StoreIConditional:
2935 case Op_StoreLConditional:
2936 case Op_CompareAndSwapB:
2937 case Op_CompareAndSwapS:
2938 case Op_CompareAndSwapI:
2939 case Op_CompareAndSwapL:
2940 case Op_CompareAndSwapP:
2941 case Op_CompareAndSwapN:
2942 case Op_WeakCompareAndSwapB:
2943 case Op_WeakCompareAndSwapS:
2944 case Op_WeakCompareAndSwapI:
2945 case Op_WeakCompareAndSwapL:
2946 case Op_WeakCompareAndSwapP:
2947 case Op_WeakCompareAndSwapN:
2948 case Op_CompareAndExchangeB:
2949 case Op_CompareAndExchangeS:
2950 case Op_CompareAndExchangeI:
2951 case Op_CompareAndExchangeL:
2952 case Op_CompareAndExchangeP:
2953 case Op_CompareAndExchangeN:
2954 case Op_GetAndAddS:
2955 case Op_GetAndAddB:
2956 case Op_GetAndAddI:
2957 case Op_GetAndAddL:
2958 case Op_GetAndSetS:
2959 case Op_GetAndSetB:
2960 case Op_GetAndSetI:
2961 case Op_GetAndSetL:
2962 case Op_GetAndSetP:
2963 case Op_GetAndSetN:
2964 case Op_StoreP:
2965 case Op_StoreN:
2966 case Op_StoreNKlass:
2967 case Op_LoadB:
2968 case Op_LoadUB:
2969 case Op_LoadUS:
2970 case Op_LoadI:
2971 case Op_LoadKlass:
2972 case Op_LoadNKlass:
2973 case Op_LoadL:
2974 case Op_LoadL_unaligned:
2975 case Op_LoadPLocked:
2976 case Op_LoadP:
2977 #if INCLUDE_ZGC
2978 case Op_LoadBarrierSlowReg:
2979 case Op_LoadBarrierWeakSlowReg:
2980 #endif
2981 case Op_LoadN:
2982 case Op_LoadRange:
2983 case Op_LoadS: {
2984 handle_mem:
2985 #ifdef ASSERT
2986 if( VerifyOptoOopOffsets ) {
2987 assert( n->is_Mem(), "" );
2988 MemNode *mem = (MemNode*)n;
2989 // Check to see if address types have grounded out somehow.
2990 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2991 assert( !tp || oop_offset_is_sane(tp), "" );
2992 }
2993 #endif
2994 break;
2995 }
2996
2997 case Op_AddP: { // Assert sane base pointers
2998 Node *addp = n->in(AddPNode::Address);
2999 assert( !addp->is_AddP() ||
3000 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3001 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3002 "Base pointers must match (addp %u)", addp->_idx );
3003 #ifdef _LP64
3004 if ((UseCompressedOops || UseCompressedClassPointers) &&
3005 addp->Opcode() == Op_ConP &&
3006 addp == n->in(AddPNode::Base) &&
3007 n->in(AddPNode::Offset)->is_Con()) {
3008 // If the transformation of ConP to ConN+DecodeN is beneficial depends
3009 // on the platform and on the compressed oops mode.
3010 // Use addressing with narrow klass to load with offset on x86.
3011 // Some platforms can use the constant pool to load ConP.
3012 // Do this transformation here since IGVN will convert ConN back to ConP.
3013 const Type* t = addp->bottom_type();
3014 bool is_oop = t->isa_oopptr() != NULL;
3015 bool is_klass = t->isa_klassptr() != NULL;
3016
3017 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
3018 (is_klass && Matcher::const_klass_prefer_decode())) {
3019 Node* nn = NULL;
3020
3021 int op = is_oop ? Op_ConN : Op_ConNKlass;
3022
3023 // Look for existing ConN node of the same exact type.
3024 Node* r = root();
3025 uint cnt = r->outcnt();
3026 for (uint i = 0; i < cnt; i++) {
3027 Node* m = r->raw_out(i);
3028 if (m!= NULL && m->Opcode() == op &&
3029 m->bottom_type()->make_ptr() == t) {
3030 nn = m;
3031 break;
3032 }
3033 }
3034 if (nn != NULL) {
3035 // Decode a narrow oop to match address
3036 // [R12 + narrow_oop_reg<<3 + offset]
3037 if (is_oop) {
3038 nn = new DecodeNNode(nn, t);
3039 } else {
3040 nn = new DecodeNKlassNode(nn, t);
3041 }
3042 // Check for succeeding AddP which uses the same Base.
3043 // Otherwise we will run into the assertion above when visiting that guy.
3044 for (uint i = 0; i < n->outcnt(); ++i) {
3045 Node *out_i = n->raw_out(i);
3046 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3047 out_i->set_req(AddPNode::Base, nn);
3048 #ifdef ASSERT
3049 for (uint j = 0; j < out_i->outcnt(); ++j) {
3050 Node *out_j = out_i->raw_out(j);
3051 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3052 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3053 }
3054 #endif
3055 }
3056 }
3057 n->set_req(AddPNode::Base, nn);
3058 n->set_req(AddPNode::Address, nn);
3059 if (addp->outcnt() == 0) {
3060 addp->disconnect_inputs(NULL, this);
3061 }
3062 }
3063 }
3064 }
3065 #endif
3066 // platform dependent reshaping of the address expression
3067 reshape_address(n->as_AddP());
3068 break;
3069 }
3070
3071 case Op_CastPP: {
3072 // Remove CastPP nodes to gain more freedom during scheduling but
3073 // keep the dependency they encode as control or precedence edges
3074 // (if control is set already) on memory operations. Some CastPP
3075 // nodes don't have a control (don't carry a dependency): skip
3076 // those.
3077 if (n->in(0) != NULL) {
3078 ResourceMark rm;
3079 Unique_Node_List wq;
3080 wq.push(n);
3081 for (uint next = 0; next < wq.size(); ++next) {
3082 Node *m = wq.at(next);
3083 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3084 Node* use = m->fast_out(i);
3085 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) {
3086 use->ensure_control_or_add_prec(n->in(0));
3087 } else {
3088 switch(use->Opcode()) {
3089 case Op_AddP:
3090 case Op_DecodeN:
3091 case Op_DecodeNKlass:
3092 case Op_CheckCastPP:
3093 case Op_CastPP:
3094 wq.push(use);
3095 break;
3096 }
3097 }
3098 }
3099 }
3100 }
3101 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3102 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3103 Node* in1 = n->in(1);
3104 const Type* t = n->bottom_type();
3105 Node* new_in1 = in1->clone();
3106 new_in1->as_DecodeN()->set_type(t);
3107
3108 if (!Matcher::narrow_oop_use_complex_address()) {
3109 //
3110 // x86, ARM and friends can handle 2 adds in addressing mode
3111 // and Matcher can fold a DecodeN node into address by using
3112 // a narrow oop directly and do implicit NULL check in address:
3113 //
3114 // [R12 + narrow_oop_reg<<3 + offset]
3115 // NullCheck narrow_oop_reg
3116 //
3117 // On other platforms (Sparc) we have to keep new DecodeN node and
3118 // use it to do implicit NULL check in address:
3119 //
3120 // decode_not_null narrow_oop_reg, base_reg
3121 // [base_reg + offset]
3122 // NullCheck base_reg
3123 //
3124 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3125 // to keep the information to which NULL check the new DecodeN node
3126 // corresponds to use it as value in implicit_null_check().
3127 //
3128 new_in1->set_req(0, n->in(0));
3129 }
3130
3131 n->subsume_by(new_in1, this);
3132 if (in1->outcnt() == 0) {
3133 in1->disconnect_inputs(NULL, this);
3134 }
3135 } else {
3136 n->subsume_by(n->in(1), this);
3137 if (n->outcnt() == 0) {
3138 n->disconnect_inputs(NULL, this);
3139 }
3140 }
3141 break;
3142 }
3143 #ifdef _LP64
3144 case Op_CmpP:
3145 // Do this transformation here to preserve CmpPNode::sub() and
3146 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3147 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3148 Node* in1 = n->in(1);
3149 Node* in2 = n->in(2);
3150 if (!in1->is_DecodeNarrowPtr()) {
3151 in2 = in1;
3152 in1 = n->in(2);
3153 }
3154 assert(in1->is_DecodeNarrowPtr(), "sanity");
3155
3156 Node* new_in2 = NULL;
3157 if (in2->is_DecodeNarrowPtr()) {
3158 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3159 new_in2 = in2->in(1);
3160 } else if (in2->Opcode() == Op_ConP) {
3161 const Type* t = in2->bottom_type();
3162 if (t == TypePtr::NULL_PTR) {
3163 assert(in1->is_DecodeN(), "compare klass to null?");
3164 // Don't convert CmpP null check into CmpN if compressed
3165 // oops implicit null check is not generated.
3166 // This will allow to generate normal oop implicit null check.
3167 if (Matcher::gen_narrow_oop_implicit_null_checks())
3168 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3169 //
3170 // This transformation together with CastPP transformation above
3171 // will generated code for implicit NULL checks for compressed oops.
3172 //
3173 // The original code after Optimize()
3174 //
3175 // LoadN memory, narrow_oop_reg
3176 // decode narrow_oop_reg, base_reg
3177 // CmpP base_reg, NULL
3178 // CastPP base_reg // NotNull
3179 // Load [base_reg + offset], val_reg
3180 //
3181 // after these transformations will be
3182 //
3183 // LoadN memory, narrow_oop_reg
3184 // CmpN narrow_oop_reg, NULL
3185 // decode_not_null narrow_oop_reg, base_reg
3186 // Load [base_reg + offset], val_reg
3187 //
3188 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3189 // since narrow oops can be used in debug info now (see the code in
3190 // final_graph_reshaping_walk()).
3191 //
3192 // At the end the code will be matched to
3193 // on x86:
3194 //
3195 // Load_narrow_oop memory, narrow_oop_reg
3196 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3197 // NullCheck narrow_oop_reg
3198 //
3199 // and on sparc:
3200 //
3201 // Load_narrow_oop memory, narrow_oop_reg
3202 // decode_not_null narrow_oop_reg, base_reg
3203 // Load [base_reg + offset], val_reg
3204 // NullCheck base_reg
3205 //
3206 } else if (t->isa_oopptr()) {
3207 new_in2 = ConNode::make(t->make_narrowoop());
3208 } else if (t->isa_klassptr()) {
3209 new_in2 = ConNode::make(t->make_narrowklass());
3210 }
3211 }
3212 if (new_in2 != NULL) {
3213 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3214 n->subsume_by(cmpN, this);
3215 if (in1->outcnt() == 0) {
3216 in1->disconnect_inputs(NULL, this);
3217 }
3218 if (in2->outcnt() == 0) {
3219 in2->disconnect_inputs(NULL, this);
3220 }
3221 }
3222 }
3223 break;
3224
3225 case Op_DecodeN:
3226 case Op_DecodeNKlass:
3227 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3228 // DecodeN could be pinned when it can't be fold into
3229 // an address expression, see the code for Op_CastPP above.
3230 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3231 break;
3232
3233 case Op_EncodeP:
3234 case Op_EncodePKlass: {
3235 Node* in1 = n->in(1);
3236 if (in1->is_DecodeNarrowPtr()) {
3237 n->subsume_by(in1->in(1), this);
3238 } else if (in1->Opcode() == Op_ConP) {
3239 const Type* t = in1->bottom_type();
3240 if (t == TypePtr::NULL_PTR) {
3241 assert(t->isa_oopptr(), "null klass?");
3242 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3243 } else if (t->isa_oopptr()) {
3244 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3245 } else if (t->isa_klassptr()) {
3246 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3247 }
3248 }
3249 if (in1->outcnt() == 0) {
3250 in1->disconnect_inputs(NULL, this);
3251 }
3252 break;
3253 }
3254
3255 case Op_Proj: {
3256 if (OptimizeStringConcat) {
3257 ProjNode* p = n->as_Proj();
3258 if (p->_is_io_use) {
3259 // Separate projections were used for the exception path which
3260 // are normally removed by a late inline. If it wasn't inlined
3261 // then they will hang around and should just be replaced with
3262 // the original one.
3263 Node* proj = NULL;
3264 // Replace with just one
3265 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3266 Node *use = i.get();
3267 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3268 proj = use;
3269 break;
3270 }
3271 }
3272 assert(proj != NULL, "must be found");
3273 p->subsume_by(proj, this);
3274 }
3275 }
3276 break;
3277 }
3278
3279 case Op_Phi:
3280 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3281 // The EncodeP optimization may create Phi with the same edges
3282 // for all paths. It is not handled well by Register Allocator.
3283 Node* unique_in = n->in(1);
3284 assert(unique_in != NULL, "");
3285 uint cnt = n->req();
3286 for (uint i = 2; i < cnt; i++) {
3287 Node* m = n->in(i);
3288 assert(m != NULL, "");
3289 if (unique_in != m)
3290 unique_in = NULL;
3291 }
3292 if (unique_in != NULL) {
3293 n->subsume_by(unique_in, this);
3294 }
3295 }
3296 break;
3297
3298 #endif
3299
3300 #ifdef ASSERT
3301 case Op_CastII:
3302 // Verify that all range check dependent CastII nodes were removed.
3303 if (n->isa_CastII()->has_range_check()) {
3304 n->dump(3);
3305 assert(false, "Range check dependent CastII node was not removed");
3306 }
3307 break;
3308 #endif
3309
3310 case Op_ModI:
3311 if (UseDivMod) {
3312 // Check if a%b and a/b both exist
3313 Node* d = n->find_similar(Op_DivI);
3314 if (d) {
3315 // Replace them with a fused divmod if supported
3316 if (Matcher::has_match_rule(Op_DivModI)) {
3317 DivModINode* divmod = DivModINode::make(n);
3318 d->subsume_by(divmod->div_proj(), this);
3319 n->subsume_by(divmod->mod_proj(), this);
3320 } else {
3321 // replace a%b with a-((a/b)*b)
3322 Node* mult = new MulINode(d, d->in(2));
3323 Node* sub = new SubINode(d->in(1), mult);
3324 n->subsume_by(sub, this);
3325 }
3326 }
3327 }
3328 break;
3329
3330 case Op_ModL:
3331 if (UseDivMod) {
3332 // Check if a%b and a/b both exist
3333 Node* d = n->find_similar(Op_DivL);
3334 if (d) {
3335 // Replace them with a fused divmod if supported
3336 if (Matcher::has_match_rule(Op_DivModL)) {
3337 DivModLNode* divmod = DivModLNode::make(n);
3338 d->subsume_by(divmod->div_proj(), this);
3339 n->subsume_by(divmod->mod_proj(), this);
3340 } else {
3341 // replace a%b with a-((a/b)*b)
3342 Node* mult = new MulLNode(d, d->in(2));
3343 Node* sub = new SubLNode(d->in(1), mult);
3344 n->subsume_by(sub, this);
3345 }
3346 }
3347 }
3348 break;
3349
3350 case Op_LoadVector:
3351 case Op_StoreVector:
3352 break;
3353
3354 case Op_AddReductionVI:
3355 case Op_AddReductionVL:
3356 case Op_AddReductionVF:
3357 case Op_AddReductionVD:
3358 case Op_MulReductionVI:
3359 case Op_MulReductionVL:
3360 case Op_MulReductionVF:
3361 case Op_MulReductionVD:
3362 break;
3363
3364 case Op_PackB:
3365 case Op_PackS:
3366 case Op_PackI:
3367 case Op_PackF:
3368 case Op_PackL:
3369 case Op_PackD:
3370 if (n->req()-1 > 2) {
3371 // Replace many operand PackNodes with a binary tree for matching
3372 PackNode* p = (PackNode*) n;
3373 Node* btp = p->binary_tree_pack(1, n->req());
3374 n->subsume_by(btp, this);
3375 }
3376 break;
3377 case Op_Loop:
3378 case Op_CountedLoop:
3379 case Op_OuterStripMinedLoop:
3380 if (n->as_Loop()->is_inner_loop()) {
3381 frc.inc_inner_loop_count();
3382 }
3383 n->as_Loop()->verify_strip_mined(0);
3384 break;
3385 case Op_LShiftI:
3386 case Op_RShiftI:
3387 case Op_URShiftI:
3388 case Op_LShiftL:
3389 case Op_RShiftL:
3390 case Op_URShiftL:
3391 if (Matcher::need_masked_shift_count) {
3392 // The cpu's shift instructions don't restrict the count to the
3393 // lower 5/6 bits. We need to do the masking ourselves.
3394 Node* in2 = n->in(2);
3395 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3396 const TypeInt* t = in2->find_int_type();
3397 if (t != NULL && t->is_con()) {
3398 juint shift = t->get_con();
3399 if (shift > mask) { // Unsigned cmp
3400 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3401 }
3402 } else {
3403 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3404 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3405 n->set_req(2, shift);
3406 }
3407 }
3408 if (in2->outcnt() == 0) { // Remove dead node
3409 in2->disconnect_inputs(NULL, this);
3410 }
3411 }
3412 break;
3413 case Op_MemBarStoreStore:
3414 case Op_MemBarRelease:
3415 // Break the link with AllocateNode: it is no longer useful and
3416 // confuses register allocation.
3417 if (n->req() > MemBarNode::Precedent) {
3418 n->set_req(MemBarNode::Precedent, top());
3419 }
3420 break;
3421 #if INCLUDE_SHENANDOAHGC
3422 case Op_ShenandoahReadBarrier:
3423 break;
3424 case Op_ShenandoahWriteBarrier:
3425 assert(false, "should have been expanded already");
3426 break;
3427 #endif
3428 case Op_RangeCheck: {
3429 RangeCheckNode* rc = n->as_RangeCheck();
3430 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3431 n->subsume_by(iff, this);
3432 frc._tests.push(iff);
3433 break;
3434 }
3435 case Op_ConvI2L: {
3436 if (!Matcher::convi2l_type_required) {
3437 // Code generation on some platforms doesn't need accurate
3438 // ConvI2L types. Widening the type can help remove redundant
3439 // address computations.
3440 n->as_Type()->set_type(TypeLong::INT);
3441 ResourceMark rm;
3442 Node_List wq;
3443 wq.push(n);
3444 for (uint next = 0; next < wq.size(); next++) {
3445 Node *m = wq.at(next);
3446
3447 for(;;) {
3448 // Loop over all nodes with identical inputs edges as m
3449 Node* k = m->find_similar(m->Opcode());
3450 if (k == NULL) {
3451 break;
3452 }
3453 // Push their uses so we get a chance to remove node made
3454 // redundant
3455 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3456 Node* u = k->fast_out(i);
3457 assert(!wq.contains(u), "shouldn't process one node several times");
3458 if (u->Opcode() == Op_LShiftL ||
3459 u->Opcode() == Op_AddL ||
3460 u->Opcode() == Op_SubL ||
3461 u->Opcode() == Op_AddP) {
3462 wq.push(u);
3463 }
3464 }
3465 // Replace all nodes with identical edges as m with m
3466 k->subsume_by(m, this);
3467 }
3468 }
3469 }
3470 break;
3471 }
3472 case Op_CmpUL: {
3473 if (!Matcher::has_match_rule(Op_CmpUL)) {
3474 // We don't support unsigned long comparisons. Set 'max_idx_expr'
3475 // to max_julong if < 0 to make the signed comparison fail.
3476 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3477 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3478 Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3479 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3480 Node* andl = new AndLNode(orl, remove_sign_mask);
3481 Node* cmp = new CmpLNode(andl, n->in(2));
3482 n->subsume_by(cmp, this);
3483 }
3484 break;
3485 }
3486 default:
3487 assert( !n->is_Call(), "" );
3488 assert( !n->is_Mem(), "" );
3489 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3490 break;
3491 }
3492
3493 // Collect CFG split points
3494 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3495 frc._tests.push(n);
3496 }
3497 }
3498
3499 //------------------------------final_graph_reshaping_walk---------------------
3500 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3501 // requires that the walk visits a node's inputs before visiting the node.
3502 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3503 ResourceArea *area = Thread::current()->resource_area();
3504 Unique_Node_List sfpt(area);
3505
3506 frc._visited.set(root->_idx); // first, mark node as visited
3507 uint cnt = root->req();
3508 Node *n = root;
3509 uint i = 0;
3510 while (true) {
3511 if (i < cnt) {
3512 // Place all non-visited non-null inputs onto stack
3513 Node* m = n->in(i);
3514 ++i;
3515 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3516 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3517 // compute worst case interpreter size in case of a deoptimization
3518 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3519
3520 sfpt.push(m);
3521 }
3522 cnt = m->req();
3523 nstack.push(n, i); // put on stack parent and next input's index
3524 n = m;
3525 i = 0;
3526 }
3527 } else {
3528 // Now do post-visit work
3529 final_graph_reshaping_impl( n, frc );
3530 if (nstack.is_empty())
3531 break; // finished
3532 n = nstack.node(); // Get node from stack
3533 cnt = n->req();
3534 i = nstack.index();
3535 nstack.pop(); // Shift to the next node on stack
3536 }
3537 }
3538
3539 // Skip next transformation if compressed oops are not used.
3540 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3541 (!UseCompressedOops && !UseCompressedClassPointers))
3542 return;
3543
3544 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3545 // It could be done for an uncommon traps or any safepoints/calls
3546 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3547 while (sfpt.size() > 0) {
3548 n = sfpt.pop();
3549 JVMState *jvms = n->as_SafePoint()->jvms();
3550 assert(jvms != NULL, "sanity");
3551 int start = jvms->debug_start();
3552 int end = n->req();
3553 bool is_uncommon = (n->is_CallStaticJava() &&
3554 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3555 for (int j = start; j < end; j++) {
3556 Node* in = n->in(j);
3557 if (in->is_DecodeNarrowPtr()) {
3558 bool safe_to_skip = true;
3559 if (!is_uncommon ) {
3560 // Is it safe to skip?
3561 for (uint i = 0; i < in->outcnt(); i++) {
3562 Node* u = in->raw_out(i);
3563 if (!u->is_SafePoint() ||
3564 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3565 safe_to_skip = false;
3566 }
3567 }
3568 }
3569 if (safe_to_skip) {
3570 n->set_req(j, in->in(1));
3571 }
3572 if (in->outcnt() == 0) {
3573 in->disconnect_inputs(NULL, this);
3574 }
3575 }
3576 }
3577 }
3578 }
3579
3580 //------------------------------final_graph_reshaping--------------------------
3581 // Final Graph Reshaping.
3582 //
3583 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3584 // and not commoned up and forced early. Must come after regular
3585 // optimizations to avoid GVN undoing the cloning. Clone constant
3586 // inputs to Loop Phis; these will be split by the allocator anyways.
3587 // Remove Opaque nodes.
3588 // (2) Move last-uses by commutative operations to the left input to encourage
3589 // Intel update-in-place two-address operations and better register usage
3590 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3591 // calls canonicalizing them back.
3592 // (3) Count the number of double-precision FP ops, single-precision FP ops
3593 // and call sites. On Intel, we can get correct rounding either by
3594 // forcing singles to memory (requires extra stores and loads after each
3595 // FP bytecode) or we can set a rounding mode bit (requires setting and
3596 // clearing the mode bit around call sites). The mode bit is only used
3597 // if the relative frequency of single FP ops to calls is low enough.
3598 // This is a key transform for SPEC mpeg_audio.
3599 // (4) Detect infinite loops; blobs of code reachable from above but not
3600 // below. Several of the Code_Gen algorithms fail on such code shapes,
3601 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3602 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3603 // Detection is by looking for IfNodes where only 1 projection is
3604 // reachable from below or CatchNodes missing some targets.
3605 // (5) Assert for insane oop offsets in debug mode.
3606
3607 bool Compile::final_graph_reshaping() {
3608 // an infinite loop may have been eliminated by the optimizer,
3609 // in which case the graph will be empty.
3610 if (root()->req() == 1) {
3611 record_method_not_compilable("trivial infinite loop");
3612 return true;
3613 }
3614
3615 // Expensive nodes have their control input set to prevent the GVN
3616 // from freely commoning them. There's no GVN beyond this point so
3617 // no need to keep the control input. We want the expensive nodes to
3618 // be freely moved to the least frequent code path by gcm.
3619 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3620 for (int i = 0; i < expensive_count(); i++) {
3621 _expensive_nodes->at(i)->set_req(0, NULL);
3622 }
3623
3624 Final_Reshape_Counts frc;
3625
3626 // Visit everybody reachable!
3627 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3628 Node_Stack nstack(live_nodes() >> 1);
3629 final_graph_reshaping_walk(nstack, root(), frc);
3630
3631 // Check for unreachable (from below) code (i.e., infinite loops).
3632 for( uint i = 0; i < frc._tests.size(); i++ ) {
3633 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3634 // Get number of CFG targets.
3635 // Note that PCTables include exception targets after calls.
3636 uint required_outcnt = n->required_outcnt();
3637 if (n->outcnt() != required_outcnt) {
3638 // Check for a few special cases. Rethrow Nodes never take the
3639 // 'fall-thru' path, so expected kids is 1 less.
3640 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3641 if (n->in(0)->in(0)->is_Call()) {
3642 CallNode *call = n->in(0)->in(0)->as_Call();
3643 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3644 required_outcnt--; // Rethrow always has 1 less kid
3645 } else if (call->req() > TypeFunc::Parms &&
3646 call->is_CallDynamicJava()) {
3647 // Check for null receiver. In such case, the optimizer has
3648 // detected that the virtual call will always result in a null
3649 // pointer exception. The fall-through projection of this CatchNode
3650 // will not be populated.
3651 Node *arg0 = call->in(TypeFunc::Parms);
3652 if (arg0->is_Type() &&
3653 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3654 required_outcnt--;
3655 }
3656 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3657 call->req() > TypeFunc::Parms+1 &&
3658 call->is_CallStaticJava()) {
3659 // Check for negative array length. In such case, the optimizer has
3660 // detected that the allocation attempt will always result in an
3661 // exception. There is no fall-through projection of this CatchNode .
3662 Node *arg1 = call->in(TypeFunc::Parms+1);
3663 if (arg1->is_Type() &&
3664 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3665 required_outcnt--;
3666 }
3667 }
3668 }
3669 }
3670 // Recheck with a better notion of 'required_outcnt'
3671 if (n->outcnt() != required_outcnt) {
3672 record_method_not_compilable("malformed control flow");
3673 return true; // Not all targets reachable!
3674 }
3675 }
3676 // Check that I actually visited all kids. Unreached kids
3677 // must be infinite loops.
3678 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3679 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3680 record_method_not_compilable("infinite loop");
3681 return true; // Found unvisited kid; must be unreach
3682 }
3683
3684 // Here so verification code in final_graph_reshaping_walk()
3685 // always see an OuterStripMinedLoopEnd
3686 if (n->is_OuterStripMinedLoopEnd()) {
3687 IfNode* init_iff = n->as_If();
3688 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3689 n->subsume_by(iff, this);
3690 }
3691 }
3692
3693 // If original bytecodes contained a mixture of floats and doubles
3694 // check if the optimizer has made it homogenous, item (3).
3695 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3696 frc.get_float_count() > 32 &&
3697 frc.get_double_count() == 0 &&
3698 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3699 set_24_bit_selection_and_mode( false, true );
3700 }
3701
3702 set_java_calls(frc.get_java_call_count());
3703 set_inner_loops(frc.get_inner_loop_count());
3704
3705 // No infinite loops, no reason to bail out.
3706 return false;
3707 }
3708
3709 //-----------------------------too_many_traps----------------------------------
3710 // Report if there are too many traps at the current method and bci.
3711 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3712 bool Compile::too_many_traps(ciMethod* method,
3713 int bci,
3714 Deoptimization::DeoptReason reason) {
3715 ciMethodData* md = method->method_data();
3716 if (md->is_empty()) {
3717 // Assume the trap has not occurred, or that it occurred only
3718 // because of a transient condition during start-up in the interpreter.
3719 return false;
3720 }
3721 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3722 if (md->has_trap_at(bci, m, reason) != 0) {
3723 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3724 // Also, if there are multiple reasons, or if there is no per-BCI record,
3725 // assume the worst.
3726 if (log())
3727 log()->elem("observe trap='%s' count='%d'",
3728 Deoptimization::trap_reason_name(reason),
3729 md->trap_count(reason));
3730 return true;
3731 } else {
3732 // Ignore method/bci and see if there have been too many globally.
3733 return too_many_traps(reason, md);
3734 }
3735 }
3736
3737 // Less-accurate variant which does not require a method and bci.
3738 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3739 ciMethodData* logmd) {
3740 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3741 // Too many traps globally.
3742 // Note that we use cumulative trap_count, not just md->trap_count.
3743 if (log()) {
3744 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3745 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3746 Deoptimization::trap_reason_name(reason),
3747 mcount, trap_count(reason));
3748 }
3749 return true;
3750 } else {
3751 // The coast is clear.
3752 return false;
3753 }
3754 }
3755
3756 //--------------------------too_many_recompiles--------------------------------
3757 // Report if there are too many recompiles at the current method and bci.
3758 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3759 // Is not eager to return true, since this will cause the compiler to use
3760 // Action_none for a trap point, to avoid too many recompilations.
3761 bool Compile::too_many_recompiles(ciMethod* method,
3762 int bci,
3763 Deoptimization::DeoptReason reason) {
3764 ciMethodData* md = method->method_data();
3765 if (md->is_empty()) {
3766 // Assume the trap has not occurred, or that it occurred only
3767 // because of a transient condition during start-up in the interpreter.
3768 return false;
3769 }
3770 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3771 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3772 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3773 Deoptimization::DeoptReason per_bc_reason
3774 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3775 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3776 if ((per_bc_reason == Deoptimization::Reason_none
3777 || md->has_trap_at(bci, m, reason) != 0)
3778 // The trap frequency measure we care about is the recompile count:
3779 && md->trap_recompiled_at(bci, m)
3780 && md->overflow_recompile_count() >= bc_cutoff) {
3781 // Do not emit a trap here if it has already caused recompilations.
3782 // Also, if there are multiple reasons, or if there is no per-BCI record,
3783 // assume the worst.
3784 if (log())
3785 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3786 Deoptimization::trap_reason_name(reason),
3787 md->trap_count(reason),
3788 md->overflow_recompile_count());
3789 return true;
3790 } else if (trap_count(reason) != 0
3791 && decompile_count() >= m_cutoff) {
3792 // Too many recompiles globally, and we have seen this sort of trap.
3793 // Use cumulative decompile_count, not just md->decompile_count.
3794 if (log())
3795 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3796 Deoptimization::trap_reason_name(reason),
3797 md->trap_count(reason), trap_count(reason),
3798 md->decompile_count(), decompile_count());
3799 return true;
3800 } else {
3801 // The coast is clear.
3802 return false;
3803 }
3804 }
3805
3806 // Compute when not to trap. Used by matching trap based nodes and
3807 // NullCheck optimization.
3808 void Compile::set_allowed_deopt_reasons() {
3809 _allowed_reasons = 0;
3810 if (is_method_compilation()) {
3811 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3812 assert(rs < BitsPerInt, "recode bit map");
3813 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3814 _allowed_reasons |= nth_bit(rs);
3815 }
3816 }
3817 }
3818 }
3819
3820 #ifndef PRODUCT
3821 //------------------------------verify_graph_edges---------------------------
3822 // Walk the Graph and verify that there is a one-to-one correspondence
3823 // between Use-Def edges and Def-Use edges in the graph.
3824 void Compile::verify_graph_edges(bool no_dead_code) {
3825 if (VerifyGraphEdges) {
3826 ResourceArea *area = Thread::current()->resource_area();
3827 Unique_Node_List visited(area);
3828 // Call recursive graph walk to check edges
3829 _root->verify_edges(visited);
3830 if (no_dead_code) {
3831 // Now make sure that no visited node is used by an unvisited node.
3832 bool dead_nodes = false;
3833 Unique_Node_List checked(area);
3834 while (visited.size() > 0) {
3835 Node* n = visited.pop();
3836 checked.push(n);
3837 for (uint i = 0; i < n->outcnt(); i++) {
3838 Node* use = n->raw_out(i);
3839 if (checked.member(use)) continue; // already checked
3840 if (visited.member(use)) continue; // already in the graph
3841 if (use->is_Con()) continue; // a dead ConNode is OK
3842 // At this point, we have found a dead node which is DU-reachable.
3843 if (!dead_nodes) {
3844 tty->print_cr("*** Dead nodes reachable via DU edges:");
3845 dead_nodes = true;
3846 }
3847 use->dump(2);
3848 tty->print_cr("---");
3849 checked.push(use); // No repeats; pretend it is now checked.
3850 }
3851 }
3852 assert(!dead_nodes, "using nodes must be reachable from root");
3853 }
3854 }
3855 }
3856
3857 // Verify GC barriers consistency
3858 // Currently supported:
3859 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3860 void Compile::verify_barriers() {
3861 #if INCLUDE_G1GC || INCLUDE_SHENANDOAHGC
3862 if (UseG1GC || UseShenandoahGC) {
3863 // Verify G1 pre-barriers
3864
3865 #if INCLUDE_G1GC && INCLUDE_SHENANDOAHGC
3866 const int marking_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_active_offset()
3867 : ShenandoahThreadLocalData::satb_mark_queue_active_offset());
3868 #elif INCLUDE_G1GC
3869 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
3870 #else
3871 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
3872 #endif
3873
3874 ResourceArea *area = Thread::current()->resource_area();
3875 Unique_Node_List visited(area);
3876 Node_List worklist(area);
3877 // We're going to walk control flow backwards starting from the Root
3878 worklist.push(_root);
3879 while (worklist.size() > 0) {
3880 Node* x = worklist.pop();
3881 if (x == NULL || x == top()) continue;
3882 if (visited.member(x)) {
3883 continue;
3884 } else {
3885 visited.push(x);
3886 }
3887
3888 if (x->is_Region()) {
3889 for (uint i = 1; i < x->req(); i++) {
3890 worklist.push(x->in(i));
3891 }
3892 } else {
3893 worklist.push(x->in(0));
3894 // We are looking for the pattern:
3895 // /->ThreadLocal
3896 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3897 // \->ConI(0)
3898 // We want to verify that the If and the LoadB have the same control
3899 // See GraphKit::g1_write_barrier_pre()
3900 if (x->is_If()) {
3901 IfNode *iff = x->as_If();
3902 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3903 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3904 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3905 && cmp->in(1)->is_Load()) {
3906 LoadNode* load = cmp->in(1)->as_Load();
3907 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3908 && load->in(2)->in(3)->is_Con()
3909 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3910
3911 Node* if_ctrl = iff->in(0);
3912 Node* load_ctrl = load->in(0);
3913
3914 if (if_ctrl != load_ctrl) {
3915 // Skip possible CProj->NeverBranch in infinite loops
3916 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3917 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3918 if_ctrl = if_ctrl->in(0)->in(0);
3919 }
3920 }
3921 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3922 }
3923 }
3924 }
3925 }
3926 }
3927 }
3928 }
3929 #endif
3930 }
3931
3932 #endif
3933
3934 // The Compile object keeps track of failure reasons separately from the ciEnv.
3935 // This is required because there is not quite a 1-1 relation between the
3936 // ciEnv and its compilation task and the Compile object. Note that one
3937 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3938 // to backtrack and retry without subsuming loads. Other than this backtracking
3939 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3940 // by the logic in C2Compiler.
3941 void Compile::record_failure(const char* reason) {
3942 if (log() != NULL) {
3943 log()->elem("failure reason='%s' phase='compile'", reason);
3944 }
3945 if (_failure_reason == NULL) {
3946 // Record the first failure reason.
3947 _failure_reason = reason;
3948 }
3949
3950 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3951 C->print_method(PHASE_FAILURE);
3952 }
3953 _root = NULL; // flush the graph, too
3954 }
3955
3956 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3957 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3958 _phase_name(name), _dolog(CITimeVerbose)
3959 {
3960 if (_dolog) {
3961 C = Compile::current();
3962 _log = C->log();
3963 } else {
3964 C = NULL;
3965 _log = NULL;
3966 }
3967 if (_log != NULL) {
3968 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3969 _log->stamp();
3970 _log->end_head();
3971 }
3972 }
3973
3974 Compile::TracePhase::~TracePhase() {
3975
3976 C = Compile::current();
3977 if (_dolog) {
3978 _log = C->log();
3979 } else {
3980 _log = NULL;
3981 }
3982
3983 #ifdef ASSERT
3984 if (PrintIdealNodeCount) {
3985 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3986 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3987 }
3988
3989 if (VerifyIdealNodeCount) {
3990 Compile::current()->print_missing_nodes();
3991 }
3992 #endif
3993
3994 if (_log != NULL) {
3995 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3996 }
3997 }
3998
3999 //=============================================================================
4000 // Two Constant's are equal when the type and the value are equal.
4001 bool Compile::Constant::operator==(const Constant& other) {
4002 if (type() != other.type() ) return false;
4003 if (can_be_reused() != other.can_be_reused()) return false;
4004 // For floating point values we compare the bit pattern.
4005 switch (type()) {
4006 case T_INT:
4007 case T_FLOAT: return (_v._value.i == other._v._value.i);
4008 case T_LONG:
4009 case T_DOUBLE: return (_v._value.j == other._v._value.j);
4010 case T_OBJECT:
4011 case T_ADDRESS: return (_v._value.l == other._v._value.l);
4012 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
4013 case T_METADATA: return (_v._metadata == other._v._metadata);
4014 default: ShouldNotReachHere(); return false;
4015 }
4016 }
4017
4018 static int type_to_size_in_bytes(BasicType t) {
4019 switch (t) {
4020 case T_INT: return sizeof(jint );
4021 case T_LONG: return sizeof(jlong );
4022 case T_FLOAT: return sizeof(jfloat );
4023 case T_DOUBLE: return sizeof(jdouble);
4024 case T_METADATA: return sizeof(Metadata*);
4025 // We use T_VOID as marker for jump-table entries (labels) which
4026 // need an internal word relocation.
4027 case T_VOID:
4028 case T_ADDRESS:
4029 case T_OBJECT: return sizeof(jobject);
4030 default:
4031 ShouldNotReachHere();
4032 return -1;
4033 }
4034 }
4035
4036 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4037 // sort descending
4038 if (a->freq() > b->freq()) return -1;
4039 if (a->freq() < b->freq()) return 1;
4040 return 0;
4041 }
4042
4043 void Compile::ConstantTable::calculate_offsets_and_size() {
4044 // First, sort the array by frequencies.
4045 _constants.sort(qsort_comparator);
4046
4047 #ifdef ASSERT
4048 // Make sure all jump-table entries were sorted to the end of the
4049 // array (they have a negative frequency).
4050 bool found_void = false;
4051 for (int i = 0; i < _constants.length(); i++) {
4052 Constant con = _constants.at(i);
4053 if (con.type() == T_VOID)
4054 found_void = true; // jump-tables
4055 else
4056 assert(!found_void, "wrong sorting");
4057 }
4058 #endif
4059
4060 int offset = 0;
4061 for (int i = 0; i < _constants.length(); i++) {
4062 Constant* con = _constants.adr_at(i);
4063
4064 // Align offset for type.
4065 int typesize = type_to_size_in_bytes(con->type());
4066 offset = align_up(offset, typesize);
4067 con->set_offset(offset); // set constant's offset
4068
4069 if (con->type() == T_VOID) {
4070 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4071 offset = offset + typesize * n->outcnt(); // expand jump-table
4072 } else {
4073 offset = offset + typesize;
4074 }
4075 }
4076
4077 // Align size up to the next section start (which is insts; see
4078 // CodeBuffer::align_at_start).
4079 assert(_size == -1, "already set?");
4080 _size = align_up(offset, (int)CodeEntryAlignment);
4081 }
4082
4083 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4084 MacroAssembler _masm(&cb);
4085 for (int i = 0; i < _constants.length(); i++) {
4086 Constant con = _constants.at(i);
4087 address constant_addr = NULL;
4088 switch (con.type()) {
4089 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
4090 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
4091 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4092 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4093 case T_OBJECT: {
4094 jobject obj = con.get_jobject();
4095 int oop_index = _masm.oop_recorder()->find_index(obj);
4096 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4097 break;
4098 }
4099 case T_ADDRESS: {
4100 address addr = (address) con.get_jobject();
4101 constant_addr = _masm.address_constant(addr);
4102 break;
4103 }
4104 // We use T_VOID as marker for jump-table entries (labels) which
4105 // need an internal word relocation.
4106 case T_VOID: {
4107 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4108 // Fill the jump-table with a dummy word. The real value is
4109 // filled in later in fill_jump_table.
4110 address dummy = (address) n;
4111 constant_addr = _masm.address_constant(dummy);
4112 // Expand jump-table
4113 for (uint i = 1; i < n->outcnt(); i++) {
4114 address temp_addr = _masm.address_constant(dummy + i);
4115 assert(temp_addr, "consts section too small");
4116 }
4117 break;
4118 }
4119 case T_METADATA: {
4120 Metadata* obj = con.get_metadata();
4121 int metadata_index = _masm.oop_recorder()->find_index(obj);
4122 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4123 break;
4124 }
4125 default: ShouldNotReachHere();
4126 }
4127 assert(constant_addr, "consts section too small");
4128 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4129 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4130 }
4131 }
4132
4133 int Compile::ConstantTable::find_offset(Constant& con) const {
4134 int idx = _constants.find(con);
4135 guarantee(idx != -1, "constant must be in constant table");
4136 int offset = _constants.at(idx).offset();
4137 guarantee(offset != -1, "constant table not emitted yet?");
4138 return offset;
4139 }
4140
4141 void Compile::ConstantTable::add(Constant& con) {
4142 if (con.can_be_reused()) {
4143 int idx = _constants.find(con);
4144 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4145 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4146 return;
4147 }
4148 }
4149 (void) _constants.append(con);
4150 }
4151
4152 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4153 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4154 Constant con(type, value, b->_freq);
4155 add(con);
4156 return con;
4157 }
4158
4159 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4160 Constant con(metadata);
4161 add(con);
4162 return con;
4163 }
4164
4165 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4166 jvalue value;
4167 BasicType type = oper->type()->basic_type();
4168 switch (type) {
4169 case T_LONG: value.j = oper->constantL(); break;
4170 case T_FLOAT: value.f = oper->constantF(); break;
4171 case T_DOUBLE: value.d = oper->constantD(); break;
4172 case T_OBJECT:
4173 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4174 case T_METADATA: return add((Metadata*)oper->constant()); break;
4175 default: guarantee(false, "unhandled type: %s", type2name(type));
4176 }
4177 return add(n, type, value);
4178 }
4179
4180 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4181 jvalue value;
4182 // We can use the node pointer here to identify the right jump-table
4183 // as this method is called from Compile::Fill_buffer right before
4184 // the MachNodes are emitted and the jump-table is filled (means the
4185 // MachNode pointers do not change anymore).
4186 value.l = (jobject) n;
4187 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4188 add(con);
4189 return con;
4190 }
4191
4192 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4193 // If called from Compile::scratch_emit_size do nothing.
4194 if (Compile::current()->in_scratch_emit_size()) return;
4195
4196 assert(labels.is_nonempty(), "must be");
4197 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4198
4199 // Since MachConstantNode::constant_offset() also contains
4200 // table_base_offset() we need to subtract the table_base_offset()
4201 // to get the plain offset into the constant table.
4202 int offset = n->constant_offset() - table_base_offset();
4203
4204 MacroAssembler _masm(&cb);
4205 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4206
4207 for (uint i = 0; i < n->outcnt(); i++) {
4208 address* constant_addr = &jump_table_base[i];
4209 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));
4210 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4211 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4212 }
4213 }
4214
4215 //----------------------------static_subtype_check-----------------------------
4216 // Shortcut important common cases when superklass is exact:
4217 // (0) superklass is java.lang.Object (can occur in reflective code)
4218 // (1) subklass is already limited to a subtype of superklass => always ok
4219 // (2) subklass does not overlap with superklass => always fail
4220 // (3) superklass has NO subtypes and we can check with a simple compare.
4221 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4222 if (StressReflectiveCode) {
4223 return SSC_full_test; // Let caller generate the general case.
4224 }
4225
4226 if (superk == env()->Object_klass()) {
4227 return SSC_always_true; // (0) this test cannot fail
4228 }
4229
4230 ciType* superelem = superk;
4231 if (superelem->is_array_klass())
4232 superelem = superelem->as_array_klass()->base_element_type();
4233
4234 if (!subk->is_interface()) { // cannot trust static interface types yet
4235 if (subk->is_subtype_of(superk)) {
4236 return SSC_always_true; // (1) false path dead; no dynamic test needed
4237 }
4238 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4239 !superk->is_subtype_of(subk)) {
4240 return SSC_always_false;
4241 }
4242 }
4243
4244 // If casting to an instance klass, it must have no subtypes
4245 if (superk->is_interface()) {
4246 // Cannot trust interfaces yet.
4247 // %%% S.B. superk->nof_implementors() == 1
4248 } else if (superelem->is_instance_klass()) {
4249 ciInstanceKlass* ik = superelem->as_instance_klass();
4250 if (!ik->has_subklass() && !ik->is_interface()) {
4251 if (!ik->is_final()) {
4252 // Add a dependency if there is a chance of a later subclass.
4253 dependencies()->assert_leaf_type(ik);
4254 }
4255 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4256 }
4257 } else {
4258 // A primitive array type has no subtypes.
4259 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4260 }
4261
4262 return SSC_full_test;
4263 }
4264
4265 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4266 #ifdef _LP64
4267 // The scaled index operand to AddP must be a clean 64-bit value.
4268 // Java allows a 32-bit int to be incremented to a negative
4269 // value, which appears in a 64-bit register as a large
4270 // positive number. Using that large positive number as an
4271 // operand in pointer arithmetic has bad consequences.
4272 // On the other hand, 32-bit overflow is rare, and the possibility
4273 // can often be excluded, if we annotate the ConvI2L node with
4274 // a type assertion that its value is known to be a small positive
4275 // number. (The prior range check has ensured this.)
4276 // This assertion is used by ConvI2LNode::Ideal.
4277 int index_max = max_jint - 1; // array size is max_jint, index is one less
4278 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4279 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4280 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4281 #endif
4282 return idx;
4283 }
4284
4285 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4286 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4287 if (ctrl != NULL) {
4288 // Express control dependency by a CastII node with a narrow type.
4289 value = new CastIINode(value, itype, false, true /* range check dependency */);
4290 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4291 // node from floating above the range check during loop optimizations. Otherwise, the
4292 // ConvI2L node may be eliminated independently of the range check, causing the data path
4293 // to become TOP while the control path is still there (although it's unreachable).
4294 value->set_req(0, ctrl);
4295 // Save CastII node to remove it after loop optimizations.
4296 phase->C->add_range_check_cast(value);
4297 value = phase->transform(value);
4298 }
4299 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4300 return phase->transform(new ConvI2LNode(value, ltype));
4301 }
4302
4303 // The message about the current inlining is accumulated in
4304 // _print_inlining_stream and transfered into the _print_inlining_list
4305 // once we know whether inlining succeeds or not. For regular
4306 // inlining, messages are appended to the buffer pointed by
4307 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4308 // a new buffer is added after _print_inlining_idx in the list. This
4309 // way we can update the inlining message for late inlining call site
4310 // when the inlining is attempted again.
4311 void Compile::print_inlining_init() {
4312 if (print_inlining() || print_intrinsics()) {
4313 _print_inlining_stream = new stringStream();
4314 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4315 }
4316 }
4317
4318 void Compile::print_inlining_reinit() {
4319 if (print_inlining() || print_intrinsics()) {
4320 // Re allocate buffer when we change ResourceMark
4321 _print_inlining_stream = new stringStream();
4322 }
4323 }
4324
4325 void Compile::print_inlining_reset() {
4326 _print_inlining_stream->reset();
4327 }
4328
4329 void Compile::print_inlining_commit() {
4330 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4331 // Transfer the message from _print_inlining_stream to the current
4332 // _print_inlining_list buffer and clear _print_inlining_stream.
4333 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4334 print_inlining_reset();
4335 }
4336
4337 void Compile::print_inlining_push() {
4338 // Add new buffer to the _print_inlining_list at current position
4339 _print_inlining_idx++;
4340 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4341 }
4342
4343 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4344 return _print_inlining_list->at(_print_inlining_idx);
4345 }
4346
4347 void Compile::print_inlining_update(CallGenerator* cg) {
4348 if (print_inlining() || print_intrinsics()) {
4349 if (!cg->is_late_inline()) {
4350 if (print_inlining_current().cg() != NULL) {
4351 print_inlining_push();
4352 }
4353 print_inlining_commit();
4354 } else {
4355 if (print_inlining_current().cg() != cg &&
4356 (print_inlining_current().cg() != NULL ||
4357 print_inlining_current().ss()->size() != 0)) {
4358 print_inlining_push();
4359 }
4360 print_inlining_commit();
4361 print_inlining_current().set_cg(cg);
4362 }
4363 }
4364 }
4365
4366 void Compile::print_inlining_move_to(CallGenerator* cg) {
4367 // We resume inlining at a late inlining call site. Locate the
4368 // corresponding inlining buffer so that we can update it.
4369 if (print_inlining()) {
4370 for (int i = 0; i < _print_inlining_list->length(); i++) {
4371 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4372 _print_inlining_idx = i;
4373 return;
4374 }
4375 }
4376 ShouldNotReachHere();
4377 }
4378 }
4379
4380 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4381 if (print_inlining()) {
4382 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4383 assert(print_inlining_current().cg() == cg, "wrong entry");
4384 // replace message with new message
4385 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4386 print_inlining_commit();
4387 print_inlining_current().set_cg(cg);
4388 }
4389 }
4390
4391 void Compile::print_inlining_assert_ready() {
4392 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4393 }
4394
4395 void Compile::process_print_inlining() {
4396 bool do_print_inlining = print_inlining() || print_intrinsics();
4397 if (do_print_inlining || log() != NULL) {
4398 // Print inlining message for candidates that we couldn't inline
4399 // for lack of space
4400 for (int i = 0; i < _late_inlines.length(); i++) {
4401 CallGenerator* cg = _late_inlines.at(i);
4402 if (!cg->is_mh_late_inline()) {
4403 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4404 if (do_print_inlining) {
4405 cg->print_inlining_late(msg);
4406 }
4407 log_late_inline_failure(cg, msg);
4408 }
4409 }
4410 }
4411 if (do_print_inlining) {
4412 ResourceMark rm;
4413 stringStream ss;
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 }
4417 size_t end = ss.size();
4418 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4419 strncpy(_print_inlining_output, ss.base(), end+1);
4420 _print_inlining_output[end] = 0;
4421 }
4422 }
4423
4424 void Compile::dump_print_inlining() {
4425 if (_print_inlining_output != NULL) {
4426 tty->print_raw(_print_inlining_output);
4427 }
4428 }
4429
4430 void Compile::log_late_inline(CallGenerator* cg) {
4431 if (log() != NULL) {
4432 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4433 cg->unique_id());
4434 JVMState* p = cg->call_node()->jvms();
4435 while (p != NULL) {
4436 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4437 p = p->caller();
4438 }
4439 log()->tail("late_inline");
4440 }
4441 }
4442
4443 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4444 log_late_inline(cg);
4445 if (log() != NULL) {
4446 log()->inline_fail(msg);
4447 }
4448 }
4449
4450 void Compile::log_inline_id(CallGenerator* cg) {
4451 if (log() != NULL) {
4452 // The LogCompilation tool needs a unique way to identify late
4453 // inline call sites. This id must be unique for this call site in
4454 // this compilation. Try to have it unique across compilations as
4455 // well because it can be convenient when grepping through the log
4456 // file.
4457 // Distinguish OSR compilations from others in case CICountOSR is
4458 // on.
4459 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4460 cg->set_unique_id(id);
4461 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4462 }
4463 }
4464
4465 void Compile::log_inline_failure(const char* msg) {
4466 if (C->log() != NULL) {
4467 C->log()->inline_fail(msg);
4468 }
4469 }
4470
4471
4472 // Dump inlining replay data to the stream.
4473 // Don't change thread state and acquire any locks.
4474 void Compile::dump_inline_data(outputStream* out) {
4475 InlineTree* inl_tree = ilt();
4476 if (inl_tree != NULL) {
4477 out->print(" inline %d", inl_tree->count());
4478 inl_tree->dump_replay_data(out);
4479 }
4480 }
4481
4482 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4483 if (n1->Opcode() < n2->Opcode()) return -1;
4484 else if (n1->Opcode() > n2->Opcode()) return 1;
4485
4486 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4487 for (uint i = 1; i < n1->req(); i++) {
4488 if (n1->in(i) < n2->in(i)) return -1;
4489 else if (n1->in(i) > n2->in(i)) return 1;
4490 }
4491
4492 return 0;
4493 }
4494
4495 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4496 Node* n1 = *n1p;
4497 Node* n2 = *n2p;
4498
4499 return cmp_expensive_nodes(n1, n2);
4500 }
4501
4502 void Compile::sort_expensive_nodes() {
4503 if (!expensive_nodes_sorted()) {
4504 _expensive_nodes->sort(cmp_expensive_nodes);
4505 }
4506 }
4507
4508 bool Compile::expensive_nodes_sorted() const {
4509 for (int i = 1; i < _expensive_nodes->length(); i++) {
4510 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4511 return false;
4512 }
4513 }
4514 return true;
4515 }
4516
4517 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4518 if (_expensive_nodes->length() == 0) {
4519 return false;
4520 }
4521
4522 assert(OptimizeExpensiveOps, "optimization off?");
4523
4524 // Take this opportunity to remove dead nodes from the list
4525 int j = 0;
4526 for (int i = 0; i < _expensive_nodes->length(); i++) {
4527 Node* n = _expensive_nodes->at(i);
4528 if (!n->is_unreachable(igvn)) {
4529 assert(n->is_expensive(), "should be expensive");
4530 _expensive_nodes->at_put(j, n);
4531 j++;
4532 }
4533 }
4534 _expensive_nodes->trunc_to(j);
4535
4536 // Then sort the list so that similar nodes are next to each other
4537 // and check for at least two nodes of identical kind with same data
4538 // inputs.
4539 sort_expensive_nodes();
4540
4541 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4542 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4543 return true;
4544 }
4545 }
4546
4547 return false;
4548 }
4549
4550 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4551 if (_expensive_nodes->length() == 0) {
4552 return;
4553 }
4554
4555 assert(OptimizeExpensiveOps, "optimization off?");
4556
4557 // Sort to bring similar nodes next to each other and clear the
4558 // control input of nodes for which there's only a single copy.
4559 sort_expensive_nodes();
4560
4561 int j = 0;
4562 int identical = 0;
4563 int i = 0;
4564 bool modified = false;
4565 for (; i < _expensive_nodes->length()-1; i++) {
4566 assert(j <= i, "can't write beyond current index");
4567 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4568 identical++;
4569 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4570 continue;
4571 }
4572 if (identical > 0) {
4573 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4574 identical = 0;
4575 } else {
4576 Node* n = _expensive_nodes->at(i);
4577 igvn.replace_input_of(n, 0, NULL);
4578 igvn.hash_insert(n);
4579 modified = true;
4580 }
4581 }
4582 if (identical > 0) {
4583 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4584 } else if (_expensive_nodes->length() >= 1) {
4585 Node* n = _expensive_nodes->at(i);
4586 igvn.replace_input_of(n, 0, NULL);
4587 igvn.hash_insert(n);
4588 modified = true;
4589 }
4590 _expensive_nodes->trunc_to(j);
4591 if (modified) {
4592 igvn.optimize();
4593 }
4594 }
4595
4596 void Compile::add_expensive_node(Node * n) {
4597 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4598 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4599 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4600 if (OptimizeExpensiveOps) {
4601 _expensive_nodes->append(n);
4602 } else {
4603 // Clear control input and let IGVN optimize expensive nodes if
4604 // OptimizeExpensiveOps is off.
4605 n->set_req(0, NULL);
4606 }
4607 }
4608
4609 /**
4610 * Remove the speculative part of types and clean up the graph
4611 */
4612 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4613 if (UseTypeSpeculation) {
4614 Unique_Node_List worklist;
4615 worklist.push(root());
4616 int modified = 0;
4617 // Go over all type nodes that carry a speculative type, drop the
4618 // speculative part of the type and enqueue the node for an igvn
4619 // which may optimize it out.
4620 for (uint next = 0; next < worklist.size(); ++next) {
4621 Node *n = worklist.at(next);
4622 if (n->is_Type()) {
4623 TypeNode* tn = n->as_Type();
4624 const Type* t = tn->type();
4625 const Type* t_no_spec = t->remove_speculative();
4626 if (t_no_spec != t) {
4627 bool in_hash = igvn.hash_delete(n);
4628 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4629 tn->set_type(t_no_spec);
4630 igvn.hash_insert(n);
4631 igvn._worklist.push(n); // give it a chance to go away
4632 modified++;
4633 }
4634 }
4635 uint max = n->len();
4636 for( uint i = 0; i < max; ++i ) {
4637 Node *m = n->in(i);
4638 if (not_a_node(m)) continue;
4639 worklist.push(m);
4640 }
4641 }
4642 // Drop the speculative part of all types in the igvn's type table
4643 igvn.remove_speculative_types();
4644 if (modified > 0) {
4645 igvn.optimize();
4646 }
4647 #ifdef ASSERT
4648 // Verify that after the IGVN is over no speculative type has resurfaced
4649 worklist.clear();
4650 worklist.push(root());
4651 for (uint next = 0; next < worklist.size(); ++next) {
4652 Node *n = worklist.at(next);
4653 const Type* t = igvn.type_or_null(n);
4654 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4655 if (n->is_Type()) {
4656 t = n->as_Type()->type();
4657 assert(t == t->remove_speculative(), "no more speculative types");
4658 }
4659 uint max = n->len();
4660 for( uint i = 0; i < max; ++i ) {
4661 Node *m = n->in(i);
4662 if (not_a_node(m)) continue;
4663 worklist.push(m);
4664 }
4665 }
4666 igvn.check_no_speculative_types();
4667 #endif
4668 }
4669 }
4670
4671 // Auxiliary method to support randomized stressing/fuzzing.
4672 //
4673 // This method can be called the arbitrary number of times, with current count
4674 // as the argument. The logic allows selecting a single candidate from the
4675 // running list of candidates as follows:
4676 // int count = 0;
4677 // Cand* selected = null;
4678 // while(cand = cand->next()) {
4679 // if (randomized_select(++count)) {
4680 // selected = cand;
4681 // }
4682 // }
4683 //
4684 // Including count equalizes the chances any candidate is "selected".
4685 // This is useful when we don't have the complete list of candidates to choose
4686 // from uniformly. In this case, we need to adjust the randomicity of the
4687 // selection, or else we will end up biasing the selection towards the latter
4688 // candidates.
4689 //
4690 // Quick back-envelope calculation shows that for the list of n candidates
4691 // the equal probability for the candidate to persist as "best" can be
4692 // achieved by replacing it with "next" k-th candidate with the probability
4693 // of 1/k. It can be easily shown that by the end of the run, the
4694 // probability for any candidate is converged to 1/n, thus giving the
4695 // uniform distribution among all the candidates.
4696 //
4697 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4698 #define RANDOMIZED_DOMAIN_POW 29
4699 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4700 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4701 bool Compile::randomized_select(int count) {
4702 assert(count > 0, "only positive");
4703 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4704 }
4705
4706 CloneMap& Compile::clone_map() { return _clone_map; }
4707 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4708
4709 void NodeCloneInfo::dump() const {
4710 tty->print(" {%d:%d} ", idx(), gen());
4711 }
4712
4713 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4714 uint64_t val = value(old->_idx);
4715 NodeCloneInfo cio(val);
4716 assert(val != 0, "old node should be in the map");
4717 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4718 insert(nnn->_idx, cin.get());
4719 #ifndef PRODUCT
4720 if (is_debug()) {
4721 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4722 }
4723 #endif
4724 }
4725
4726 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4727 NodeCloneInfo cio(value(old->_idx));
4728 if (cio.get() == 0) {
4729 cio.set(old->_idx, 0);
4730 insert(old->_idx, cio.get());
4731 #ifndef PRODUCT
4732 if (is_debug()) {
4733 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4734 }
4735 #endif
4736 }
4737 clone(old, nnn, gen);
4738 }
4739
4740 int CloneMap::max_gen() const {
4741 int g = 0;
4742 DictI di(_dict);
4743 for(; di.test(); ++di) {
4744 int t = gen(di._key);
4745 if (g < t) {
4746 g = t;
4747 #ifndef PRODUCT
4748 if (is_debug()) {
4749 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4750 }
4751 #endif
4752 }
4753 }
4754 return g;
4755 }
4756
4757 void CloneMap::dump(node_idx_t key) const {
4758 uint64_t val = value(key);
4759 if (val != 0) {
4760 NodeCloneInfo ni(val);
4761 ni.dump();
4762 }
4763 }
4764
4765 #if INCLUDE_SHENANDOAHGC
4766 void Compile::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) {
4767 assert(UseShenandoahGC && call->is_shenandoah_wb_pre_call(), "");
4768 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
4769 c = c->unique_ctrl_out();
4770 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4771 c = c->unique_ctrl_out();
4772 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4773 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4774 assert(iff->is_If(), "expect test");
4775 if (!iff->is_shenandoah_marking_if(igvn)) {
4776 c = c->unique_ctrl_out();
4777 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4778 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4779 assert(iff->is_shenandoah_marking_if(igvn), "expect marking test");
4780 }
4781 Node* cmpx = iff->in(1)->in(1);
4782 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
4783 igvn->rehash_node_delayed(call);
4784 call->del_req(call->req()-1);
4785 }
4786 #endif