1 /*
2 * Copyright (c) 1997, 2017, 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/shenandoah/brooksPointer.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/block.hpp"
40 #include "opto/c2compiler.hpp"
41 #include "opto/callGenerator.hpp"
42 #include "opto/callnode.hpp"
43 #include "opto/castnode.hpp"
44 #include "opto/cfgnode.hpp"
45 #include "opto/chaitin.hpp"
46 #include "opto/compile.hpp"
47 #include "opto/connode.hpp"
48 #include "opto/convertnode.hpp"
49 #include "opto/divnode.hpp"
50 #include "opto/escape.hpp"
51 #include "opto/idealGraphPrinter.hpp"
52 #include "opto/loopnode.hpp"
53 #include "opto/machnode.hpp"
54 #include "opto/macro.hpp"
55 #include "opto/matcher.hpp"
56 #include "opto/mathexactnode.hpp"
57 #include "opto/memnode.hpp"
58 #include "opto/mulnode.hpp"
59 #include "opto/narrowptrnode.hpp"
60 #include "opto/node.hpp"
61 #include "opto/opcodes.hpp"
62 #include "opto/output.hpp"
63 #include "opto/parse.hpp"
64 #include "opto/phaseX.hpp"
65 #include "opto/rootnode.hpp"
66 #include "opto/runtime.hpp"
67 #include "opto/shenandoahSupport.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
79
80 // -------------------- Compile::mach_constant_base_node -----------------------
81 // Constant table base node singleton.
82 MachConstantBaseNode* Compile::mach_constant_base_node() {
83 if (_mach_constant_base_node == NULL) {
84 _mach_constant_base_node = new MachConstantBaseNode();
85 _mach_constant_base_node->add_req(C->root());
86 }
87 return _mach_constant_base_node;
88 }
89
90
91 /// Support for intrinsics.
92
93 // Return the index at which m must be inserted (or already exists).
94 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
95 class IntrinsicDescPair {
96 private:
97 ciMethod* _m;
98 bool _is_virtual;
99 public:
100 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
101 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
102 ciMethod* m= elt->method();
103 ciMethod* key_m = key->_m;
104 if (key_m < m) return -1;
105 else if (key_m > m) return 1;
106 else {
107 bool is_virtual = elt->is_virtual();
108 bool key_virtual = key->_is_virtual;
109 if (key_virtual < is_virtual) return -1;
110 else if (key_virtual > is_virtual) return 1;
111 else return 0;
112 }
113 }
114 };
115 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
116 #ifdef ASSERT
117 for (int i = 1; i < _intrinsics->length(); i++) {
118 CallGenerator* cg1 = _intrinsics->at(i-1);
119 CallGenerator* cg2 = _intrinsics->at(i);
120 assert(cg1->method() != cg2->method()
121 ? cg1->method() < cg2->method()
122 : cg1->is_virtual() < cg2->is_virtual(),
123 "compiler intrinsics list must stay sorted");
124 }
125 #endif
126 IntrinsicDescPair pair(m, is_virtual);
127 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
128 }
129
130 void Compile::register_intrinsic(CallGenerator* cg) {
131 if (_intrinsics == NULL) {
132 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
133 }
134 int len = _intrinsics->length();
135 bool found = false;
136 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
137 assert(!found, "registering twice");
138 _intrinsics->insert_before(index, cg);
139 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
140 }
141
142 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
143 assert(m->is_loaded(), "don't try this on unloaded methods");
144 if (_intrinsics != NULL) {
145 bool found = false;
146 int index = intrinsic_insertion_index(m, is_virtual, found);
147 if (found) {
148 return _intrinsics->at(index);
149 }
150 }
151 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
152 if (m->intrinsic_id() != vmIntrinsics::_none &&
153 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
154 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
155 if (cg != NULL) {
156 // Save it for next time:
157 register_intrinsic(cg);
158 return cg;
159 } else {
160 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
161 }
162 }
163 return NULL;
164 }
165
166 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
167 // in library_call.cpp.
168
169
170 #ifndef PRODUCT
171 // statistics gathering...
172
173 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
174 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
175
176 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
177 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
178 int oflags = _intrinsic_hist_flags[id];
179 assert(flags != 0, "what happened?");
180 if (is_virtual) {
181 flags |= _intrinsic_virtual;
182 }
183 bool changed = (flags != oflags);
184 if ((flags & _intrinsic_worked) != 0) {
185 juint count = (_intrinsic_hist_count[id] += 1);
186 if (count == 1) {
187 changed = true; // first time
188 }
189 // increment the overall count also:
190 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
191 }
192 if (changed) {
193 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
194 // Something changed about the intrinsic's virtuality.
195 if ((flags & _intrinsic_virtual) != 0) {
196 // This is the first use of this intrinsic as a virtual call.
197 if (oflags != 0) {
198 // We already saw it as a non-virtual, so note both cases.
199 flags |= _intrinsic_both;
200 }
201 } else if ((oflags & _intrinsic_both) == 0) {
202 // This is the first use of this intrinsic as a non-virtual
203 flags |= _intrinsic_both;
204 }
205 }
206 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
207 }
208 // update the overall flags also:
209 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
210 return changed;
211 }
212
213 static char* format_flags(int flags, char* buf) {
214 buf[0] = 0;
215 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
216 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
217 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
218 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
219 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
220 if (buf[0] == 0) strcat(buf, ",");
221 assert(buf[0] == ',', "must be");
222 return &buf[1];
223 }
224
225 void Compile::print_intrinsic_statistics() {
226 char flagsbuf[100];
227 ttyLocker ttyl;
228 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
229 tty->print_cr("Compiler intrinsic usage:");
230 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
231 if (total == 0) total = 1; // avoid div0 in case of no successes
232 #define PRINT_STAT_LINE(name, c, f) \
233 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
234 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
235 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
236 int flags = _intrinsic_hist_flags[id];
237 juint count = _intrinsic_hist_count[id];
238 if ((flags | count) != 0) {
239 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
240 }
241 }
242 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
243 if (xtty != NULL) xtty->tail("statistics");
244 }
245
246 void Compile::print_statistics() {
247 { ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='opto'");
249 Parse::print_statistics();
250 PhaseCCP::print_statistics();
251 PhaseRegAlloc::print_statistics();
252 Scheduling::print_statistics();
253 PhasePeephole::print_statistics();
254 PhaseIdealLoop::print_statistics();
255 if (xtty != NULL) xtty->tail("statistics");
256 }
257 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
258 // put this under its own <statistics> element.
259 print_intrinsic_statistics();
260 }
261 }
262 #endif //PRODUCT
263
264 // Support for bundling info
265 Bundle* Compile::node_bundling(const Node *n) {
266 assert(valid_bundle_info(n), "oob");
267 return &_node_bundling_base[n->_idx];
268 }
269
270 bool Compile::valid_bundle_info(const Node *n) {
271 return (_node_bundling_limit > n->_idx);
272 }
273
274
275 void Compile::gvn_replace_by(Node* n, Node* nn) {
276 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
277 Node* use = n->last_out(i);
278 bool is_in_table = initial_gvn()->hash_delete(use);
279 uint uses_found = 0;
280 for (uint j = 0; j < use->len(); j++) {
281 if (use->in(j) == n) {
282 if (j < use->req())
283 use->set_req(j, nn);
284 else
285 use->set_prec(j, nn);
286 uses_found++;
287 }
288 }
289 if (is_in_table) {
290 // reinsert into table
291 initial_gvn()->hash_find_insert(use);
292 }
293 record_for_igvn(use);
294 i -= uses_found; // we deleted 1 or more copies of this edge
295 }
296 }
297
298
299 static inline bool not_a_node(const Node* n) {
300 if (n == NULL) return true;
301 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
302 if (*(address*)n == badAddress) return true; // kill by Node::destruct
303 return false;
304 }
305
306 // Identify all nodes that are reachable from below, useful.
307 // Use breadth-first pass that records state in a Unique_Node_List,
308 // recursive traversal is slower.
309 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
310 int estimated_worklist_size = live_nodes();
311 useful.map( estimated_worklist_size, NULL ); // preallocate space
312
313 // Initialize worklist
314 if (root() != NULL) { useful.push(root()); }
315 // If 'top' is cached, declare it useful to preserve cached node
316 if( cached_top_node() ) { useful.push(cached_top_node()); }
317
318 // Push all useful nodes onto the list, breadthfirst
319 for( uint next = 0; next < useful.size(); ++next ) {
320 assert( next < unique(), "Unique useful nodes < total nodes");
321 Node *n = useful.at(next);
322 uint max = n->len();
323 for( uint i = 0; i < max; ++i ) {
324 Node *m = n->in(i);
325 if (not_a_node(m)) continue;
326 useful.push(m);
327 }
328 }
329 }
330
331 // Update dead_node_list with any missing dead nodes using useful
332 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
333 void Compile::update_dead_node_list(Unique_Node_List &useful) {
334 uint max_idx = unique();
335 VectorSet& useful_node_set = useful.member_set();
336
337 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
338 // If node with index node_idx is not in useful set,
339 // mark it as dead in dead node list.
340 if (! useful_node_set.test(node_idx) ) {
341 record_dead_node(node_idx);
342 }
343 }
344 }
345
346 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
347 int shift = 0;
348 for (int i = 0; i < inlines->length(); i++) {
349 CallGenerator* cg = inlines->at(i);
350 CallNode* call = cg->call_node();
351 if (shift > 0) {
352 inlines->at_put(i-shift, cg);
353 }
354 if (!useful.member(call)) {
355 shift++;
356 }
357 }
358 inlines->trunc_to(inlines->length()-shift);
359 }
360
361 // Disconnect all useless nodes by disconnecting those at the boundary.
362 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
363 uint next = 0;
364 while (next < useful.size()) {
365 Node *n = useful.at(next++);
366 if (n->is_SafePoint()) {
367 // We're done with a parsing phase. Replaced nodes are not valid
368 // beyond that point.
369 n->as_SafePoint()->delete_replaced_nodes();
370 }
371 // Use raw traversal of out edges since this code removes out edges
372 int max = n->outcnt();
373 for (int j = 0; j < max; ++j) {
374 Node* child = n->raw_out(j);
375 if (! useful.member(child)) {
376 assert(!child->is_top() || child != top(),
377 "If top is cached in Compile object it is in useful list");
378 // Only need to remove this out-edge to the useless node
379 n->raw_del_out(j);
380 --j;
381 --max;
382 }
383 }
384 if (n->outcnt() == 1 && n->has_special_unique_user()) {
385 record_for_igvn(n->unique_out());
386 }
387 if (n->Opcode() == Op_AddP && CallLeafNode::has_only_g1_wb_pre_uses(n)) {
388 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
389 record_for_igvn(n->fast_out(i));
390 }
391 }
392 }
393 // Remove useless macro and predicate opaq nodes
394 for (int i = C->macro_count()-1; i >= 0; i--) {
395 Node* n = C->macro_node(i);
396 if (!useful.member(n)) {
397 remove_macro_node(n);
398 }
399 }
400 // Remove useless CastII nodes with range check dependency
401 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
402 Node* cast = range_check_cast_node(i);
403 if (!useful.member(cast)) {
404 remove_range_check_cast(cast);
405 }
406 }
407 // Remove useless expensive node
408 for (int i = C->expensive_count()-1; i >= 0; i--) {
409 Node* n = C->expensive_node(i);
410 if (!useful.member(n)) {
411 remove_expensive_node(n);
412 }
413 }
414 for (int i = C->shenandoah_barriers_count()-1; i >= 0; i--) {
415 ShenandoahWriteBarrierNode* n = C->shenandoah_barrier(i);
416 if (!useful.member(n)) {
417 remove_shenandoah_barrier(n);
418 }
419 }
420 // clean up the late inline lists
421 remove_useless_late_inlines(&_string_late_inlines, useful);
422 remove_useless_late_inlines(&_boxing_late_inlines, useful);
423 remove_useless_late_inlines(&_late_inlines, useful);
424 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
425 }
426
427 //------------------------------frame_size_in_words-----------------------------
428 // frame_slots in units of words
429 int Compile::frame_size_in_words() const {
430 // shift is 0 in LP32 and 1 in LP64
431 const int shift = (LogBytesPerWord - LogBytesPerInt);
432 int words = _frame_slots >> shift;
433 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
434 return words;
435 }
436
437 // To bang the stack of this compiled method we use the stack size
438 // that the interpreter would need in case of a deoptimization. This
439 // removes the need to bang the stack in the deoptimization blob which
440 // in turn simplifies stack overflow handling.
441 int Compile::bang_size_in_bytes() const {
442 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
443 }
444
445 // ============================================================================
446 //------------------------------CompileWrapper---------------------------------
447 class CompileWrapper : public StackObj {
448 Compile *const _compile;
449 public:
450 CompileWrapper(Compile* compile);
451
452 ~CompileWrapper();
453 };
454
455 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
456 // the Compile* pointer is stored in the current ciEnv:
457 ciEnv* env = compile->env();
458 assert(env == ciEnv::current(), "must already be a ciEnv active");
459 assert(env->compiler_data() == NULL, "compile already active?");
460 env->set_compiler_data(compile);
461 assert(compile == Compile::current(), "sanity");
462
463 compile->set_type_dict(NULL);
464 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
465 compile->clone_map().set_clone_idx(0);
466 compile->set_type_hwm(NULL);
467 compile->set_type_last_size(0);
468 compile->set_last_tf(NULL, NULL);
469 compile->set_indexSet_arena(NULL);
470 compile->set_indexSet_free_block_list(NULL);
471 compile->init_type_arena();
472 Type::Initialize(compile);
473 _compile->set_scratch_buffer_blob(NULL);
474 _compile->begin_method();
475 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
476 }
477 CompileWrapper::~CompileWrapper() {
478 _compile->end_method();
479 if (_compile->scratch_buffer_blob() != NULL)
480 BufferBlob::free(_compile->scratch_buffer_blob());
481 _compile->env()->set_compiler_data(NULL);
482 }
483
484
485 //----------------------------print_compile_messages---------------------------
486 void Compile::print_compile_messages() {
487 #ifndef PRODUCT
488 // Check if recompiling
489 if (_subsume_loads == false && PrintOpto) {
490 // Recompiling without allowing machine instructions to subsume loads
491 tty->print_cr("*********************************************************");
492 tty->print_cr("** Bailout: Recompile without subsuming loads **");
493 tty->print_cr("*********************************************************");
494 }
495 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
496 // Recompiling without escape analysis
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without escape analysis **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
502 // Recompiling without boxing elimination
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without boxing elimination **");
505 tty->print_cr("*********************************************************");
506 }
507 if (C->directive()->BreakAtCompileOption) {
508 // Open the debugger when compiling this method.
509 tty->print("### Breaking when compiling: ");
510 method()->print_short_name();
511 tty->cr();
512 BREAKPOINT;
513 }
514
515 if( PrintOpto ) {
516 if (is_osr_compilation()) {
517 tty->print("[OSR]%3d", _compile_id);
518 } else {
519 tty->print("%3d", _compile_id);
520 }
521 }
522 #endif
523 }
524
525
526 //-----------------------init_scratch_buffer_blob------------------------------
527 // Construct a temporary BufferBlob and cache it for this compile.
528 void Compile::init_scratch_buffer_blob(int const_size) {
529 // If there is already a scratch buffer blob allocated and the
530 // constant section is big enough, use it. Otherwise free the
531 // current and allocate a new one.
532 BufferBlob* blob = scratch_buffer_blob();
533 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
534 // Use the current blob.
535 } else {
536 if (blob != NULL) {
537 BufferBlob::free(blob);
538 }
539
540 ResourceMark rm;
541 _scratch_const_size = const_size;
542 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
543 blob = BufferBlob::create("Compile::scratch_buffer", size);
544 // Record the buffer blob for next time.
545 set_scratch_buffer_blob(blob);
546 // Have we run out of code space?
547 if (scratch_buffer_blob() == NULL) {
548 // Let CompilerBroker disable further compilations.
549 record_failure("Not enough space for scratch buffer in CodeCache");
550 return;
551 }
552 }
553
554 // Initialize the relocation buffers
555 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
556 set_scratch_locs_memory(locs_buf);
557 }
558
559
560 //-----------------------scratch_emit_size-------------------------------------
561 // Helper function that computes size by emitting code
562 uint Compile::scratch_emit_size(const Node* n) {
563 // Start scratch_emit_size section.
564 set_in_scratch_emit_size(true);
565
566 // Emit into a trash buffer and count bytes emitted.
567 // This is a pretty expensive way to compute a size,
568 // but it works well enough if seldom used.
569 // All common fixed-size instructions are given a size
570 // method by the AD file.
571 // Note that the scratch buffer blob and locs memory are
572 // allocated at the beginning of the compile task, and
573 // may be shared by several calls to scratch_emit_size.
574 // The allocation of the scratch buffer blob is particularly
575 // expensive, since it has to grab the code cache lock.
576 BufferBlob* blob = this->scratch_buffer_blob();
577 assert(blob != NULL, "Initialize BufferBlob at start");
578 assert(blob->size() > MAX_inst_size, "sanity");
579 relocInfo* locs_buf = scratch_locs_memory();
580 address blob_begin = blob->content_begin();
581 address blob_end = (address)locs_buf;
582 assert(blob->contains(blob_end), "sanity");
583 CodeBuffer buf(blob_begin, blob_end - blob_begin);
584 buf.initialize_consts_size(_scratch_const_size);
585 buf.initialize_stubs_size(MAX_stubs_size);
586 assert(locs_buf != NULL, "sanity");
587 int lsize = MAX_locs_size / 3;
588 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
589 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
590 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
591 // Mark as scratch buffer.
592 buf.consts()->set_scratch_emit();
593 buf.insts()->set_scratch_emit();
594 buf.stubs()->set_scratch_emit();
595
596 // Do the emission.
597
598 Label fakeL; // Fake label for branch instructions.
599 Label* saveL = NULL;
600 uint save_bnum = 0;
601 bool is_branch = n->is_MachBranch();
602 if (is_branch) {
603 MacroAssembler masm(&buf);
604 masm.bind(fakeL);
605 n->as_MachBranch()->save_label(&saveL, &save_bnum);
606 n->as_MachBranch()->label_set(&fakeL, 0);
607 }
608 n->emit(buf, this->regalloc());
609
610 // Emitting into the scratch buffer should not fail
611 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
612
613 if (is_branch) // Restore label.
614 n->as_MachBranch()->label_set(saveL, save_bnum);
615
616 // End scratch_emit_size section.
617 set_in_scratch_emit_size(false);
618
619 return buf.insts_size();
620 }
621
622
623 // ============================================================================
624 //------------------------------Compile standard-------------------------------
625 debug_only( int Compile::_debug_idx = 100000; )
626
627 // Compile a method. entry_bci is -1 for normal compilations and indicates
628 // the continuation bci for on stack replacement.
629
630
631 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
632 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
633 : Phase(Compiler),
634 _env(ci_env),
635 _directive(directive),
636 _log(ci_env->log()),
637 _compile_id(ci_env->compile_id()),
638 _save_argument_registers(false),
639 _stub_name(NULL),
640 _stub_function(NULL),
641 _stub_entry_point(NULL),
642 _method(target),
643 _entry_bci(osr_bci),
644 _initial_gvn(NULL),
645 _for_igvn(NULL),
646 _warm_calls(NULL),
647 _subsume_loads(subsume_loads),
648 _do_escape_analysis(do_escape_analysis),
649 _eliminate_boxing(eliminate_boxing),
650 _failure_reason(NULL),
651 _code_buffer("Compile::Fill_buffer"),
652 _orig_pc_slot(0),
653 _orig_pc_slot_offset_in_bytes(0),
654 _has_method_handle_invokes(false),
655 _mach_constant_base_node(NULL),
656 _node_bundling_limit(0),
657 _node_bundling_base(NULL),
658 _java_calls(0),
659 _inner_loops(0),
660 _scratch_const_size(-1),
661 _in_scratch_emit_size(false),
662 _dead_node_list(comp_arena()),
663 _dead_node_count(0),
664 #ifndef PRODUCT
665 _trace_opto_output(directive->TraceOptoOutputOption),
666 _in_dump_cnt(0),
667 _printer(IdealGraphPrinter::printer()),
668 #endif
669 _congraph(NULL),
670 _comp_arena(mtCompiler),
671 _node_arena(mtCompiler),
672 _old_arena(mtCompiler),
673 _Compile_types(mtCompiler),
674 _replay_inline_data(NULL),
675 _late_inlines(comp_arena(), 2, 0, NULL),
676 _string_late_inlines(comp_arena(), 2, 0, NULL),
677 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
678 _late_inlines_pos(0),
679 _number_of_mh_late_inlines(0),
680 _inlining_progress(false),
681 _inlining_incrementally(false),
682 _print_inlining_list(NULL),
683 _print_inlining_stream(NULL),
684 _print_inlining_idx(0),
685 _print_inlining_output(NULL),
686 _interpreter_frame_size(0),
687 _max_node_limit(MaxNodeLimit),
688 _has_reserved_stack_access(target->has_reserved_stack_access()) {
689 C = this;
690 #ifndef PRODUCT
691 if (_printer != NULL) {
692 _printer->set_compile(this);
693 }
694 #endif
695 CompileWrapper cw(this);
696
697 if (CITimeVerbose) {
698 tty->print(" ");
699 target->holder()->name()->print();
700 tty->print(".");
701 target->print_short_name();
702 tty->print(" ");
703 }
704 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
705 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
706
707 #ifndef PRODUCT
708 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
709 if (!print_opto_assembly) {
710 bool print_assembly = directive->PrintAssemblyOption;
711 if (print_assembly && !Disassembler::can_decode()) {
712 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
713 print_opto_assembly = true;
714 }
715 }
716 set_print_assembly(print_opto_assembly);
717 set_parsed_irreducible_loop(false);
718
719 if (directive->ReplayInlineOption) {
720 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
721 }
722 #endif
723 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
724 set_print_intrinsics(directive->PrintIntrinsicsOption);
725 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
726
727 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
728 // Make sure the method being compiled gets its own MDO,
729 // so we can at least track the decompile_count().
730 // Need MDO to record RTM code generation state.
731 method()->ensure_method_data();
732 }
733
734 Init(::AliasLevel);
735
736
737 print_compile_messages();
738
739 _ilt = InlineTree::build_inline_tree_root();
740
741 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
742 assert(num_alias_types() >= AliasIdxRaw, "");
743
744 #define MINIMUM_NODE_HASH 1023
745 // Node list that Iterative GVN will start with
746 Unique_Node_List for_igvn(comp_arena());
747 set_for_igvn(&for_igvn);
748
749 // GVN that will be run immediately on new nodes
750 uint estimated_size = method()->code_size()*4+64;
751 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
752 PhaseGVN gvn(node_arena(), estimated_size);
753 set_initial_gvn(&gvn);
754
755 print_inlining_init();
756 { // Scope for timing the parser
757 TracePhase tp("parse", &timers[_t_parser]);
758
759 // Put top into the hash table ASAP.
760 initial_gvn()->transform_no_reclaim(top());
761
762 // Set up tf(), start(), and find a CallGenerator.
763 CallGenerator* cg = NULL;
764 if (is_osr_compilation()) {
765 const TypeTuple *domain = StartOSRNode::osr_domain();
766 const TypeTuple *range = TypeTuple::make_range(method()->signature());
767 init_tf(TypeFunc::make(domain, range));
768 StartNode* s = new StartOSRNode(root(), domain);
769 initial_gvn()->set_type_bottom(s);
770 init_start(s);
771 cg = CallGenerator::for_osr(method(), entry_bci());
772 } else {
773 // Normal case.
774 init_tf(TypeFunc::make(method()));
775 StartNode* s = new StartNode(root(), tf()->domain());
776 initial_gvn()->set_type_bottom(s);
777 init_start(s);
778 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && (UseG1GC || UseShenandoahGC)) {
779 // With java.lang.ref.reference.get() we must go through the
780 // intrinsic when G1 is enabled - even when get() is the root
781 // method of the compile - so that, if necessary, the value in
782 // the referent field of the reference object gets recorded by
783 // the pre-barrier code.
784 // Specifically, if G1 is enabled, the value in the referent
785 // field is recorded by the G1 SATB pre barrier. This will
786 // result in the referent being marked live and the reference
787 // object removed from the list of discovered references during
788 // reference processing.
789 cg = find_intrinsic(method(), false);
790 }
791 if (cg == NULL) {
792 float past_uses = method()->interpreter_invocation_count();
793 float expected_uses = past_uses;
794 cg = CallGenerator::for_inline(method(), expected_uses);
795 }
796 }
797 if (failing()) return;
798 if (cg == NULL) {
799 record_method_not_compilable("cannot parse method");
800 return;
801 }
802 JVMState* jvms = build_start_state(start(), tf());
803 if ((jvms = cg->generate(jvms)) == NULL) {
804 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
805 record_method_not_compilable("method parse failed");
806 }
807 return;
808 }
809 GraphKit kit(jvms);
810
811 if (!kit.stopped()) {
812 // Accept return values, and transfer control we know not where.
813 // This is done by a special, unique ReturnNode bound to root.
814 return_values(kit.jvms());
815 }
816
817 if (kit.has_exceptions()) {
818 // Any exceptions that escape from this call must be rethrown
819 // to whatever caller is dynamically above us on the stack.
820 // This is done by a special, unique RethrowNode bound to root.
821 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
822 }
823
824 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
825
826 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
827 inline_string_calls(true);
828 }
829
830 if (failing()) return;
831
832 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
833
834 // Remove clutter produced by parsing.
835 if (!failing()) {
836 ResourceMark rm;
837 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
838 }
839 }
840
841 // Note: Large methods are capped off in do_one_bytecode().
842 if (failing()) return;
843
844 // After parsing, node notes are no longer automagic.
845 // They must be propagated by register_new_node_with_optimizer(),
846 // clone(), or the like.
847 set_default_node_notes(NULL);
848
849 for (;;) {
850 int successes = Inline_Warm();
851 if (failing()) return;
852 if (successes == 0) break;
853 }
854
855 // Drain the list.
856 Finish_Warm();
857 #ifndef PRODUCT
858 if (_printer && _printer->should_print(1)) {
859 _printer->print_inlining();
860 }
861 #endif
862
863 if (failing()) return;
864 NOT_PRODUCT( verify_graph_edges(); )
865
866 // Now optimize
867 Optimize();
868 if (failing()) return;
869 NOT_PRODUCT( verify_graph_edges(); )
870
871 #ifndef PRODUCT
872 if (PrintIdeal) {
873 ttyLocker ttyl; // keep the following output all in one block
874 // This output goes directly to the tty, not the compiler log.
875 // To enable tools to match it up with the compilation activity,
876 // be sure to tag this tty output with the compile ID.
877 if (xtty != NULL) {
878 xtty->head("ideal compile_id='%d'%s", compile_id(),
879 is_osr_compilation() ? " compile_kind='osr'" :
880 "");
881 }
882 root()->dump(9999);
883 if (xtty != NULL) {
884 xtty->tail("ideal");
885 }
886 }
887 #endif
888
889 NOT_PRODUCT( verify_barriers(); )
890
891 // Dump compilation data to replay it.
892 if (directive->DumpReplayOption) {
893 env()->dump_replay_data(_compile_id);
894 }
895 if (directive->DumpInlineOption && (ilt() != NULL)) {
896 env()->dump_inline_data(_compile_id);
897 }
898
899 // Now that we know the size of all the monitors we can add a fixed slot
900 // for the original deopt pc.
901
902 _orig_pc_slot = fixed_slots();
903 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
904 set_fixed_slots(next_slot);
905
906 // Compute when to use implicit null checks. Used by matching trap based
907 // nodes and NullCheck optimization.
908 set_allowed_deopt_reasons();
909
910 // Now generate code
911 Code_Gen();
912 if (failing()) return;
913
914 // Check if we want to skip execution of all compiled code.
915 {
916 #ifndef PRODUCT
917 if (OptoNoExecute) {
918 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
919 return;
920 }
921 #endif
922 TracePhase tp("install_code", &timers[_t_registerMethod]);
923
924 if (is_osr_compilation()) {
925 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
926 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
927 } else {
928 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
929 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
930 }
931
932 env()->register_method(_method, _entry_bci,
933 &_code_offsets,
934 _orig_pc_slot_offset_in_bytes,
935 code_buffer(),
936 frame_size_in_words(), _oop_map_set,
937 &_handler_table, &_inc_table,
938 compiler,
939 has_unsafe_access(),
940 SharedRuntime::is_wide_vector(max_vector_size()),
941 rtm_state()
942 );
943
944 if (log() != NULL) // Print code cache state into compiler log
945 log()->code_cache_state();
946 }
947 }
948
949 //------------------------------Compile----------------------------------------
950 // Compile a runtime stub
951 Compile::Compile( ciEnv* ci_env,
952 TypeFunc_generator generator,
953 address stub_function,
954 const char *stub_name,
955 int is_fancy_jump,
956 bool pass_tls,
957 bool save_arg_registers,
958 bool return_pc,
959 DirectiveSet* directive)
960 : Phase(Compiler),
961 _env(ci_env),
962 _directive(directive),
963 _log(ci_env->log()),
964 _compile_id(0),
965 _save_argument_registers(save_arg_registers),
966 _method(NULL),
967 _stub_name(stub_name),
968 _stub_function(stub_function),
969 _stub_entry_point(NULL),
970 _entry_bci(InvocationEntryBci),
971 _initial_gvn(NULL),
972 _for_igvn(NULL),
973 _warm_calls(NULL),
974 _orig_pc_slot(0),
975 _orig_pc_slot_offset_in_bytes(0),
976 _subsume_loads(true),
977 _do_escape_analysis(false),
978 _eliminate_boxing(false),
979 _failure_reason(NULL),
980 _code_buffer("Compile::Fill_buffer"),
981 _has_method_handle_invokes(false),
982 _mach_constant_base_node(NULL),
983 _node_bundling_limit(0),
984 _node_bundling_base(NULL),
985 _java_calls(0),
986 _inner_loops(0),
987 #ifndef PRODUCT
988 _trace_opto_output(directive->TraceOptoOutputOption),
989 _in_dump_cnt(0),
990 _printer(NULL),
991 #endif
992 _comp_arena(mtCompiler),
993 _node_arena(mtCompiler),
994 _old_arena(mtCompiler),
995 _Compile_types(mtCompiler),
996 _dead_node_list(comp_arena()),
997 _dead_node_count(0),
998 _congraph(NULL),
999 _replay_inline_data(NULL),
1000 _number_of_mh_late_inlines(0),
1001 _inlining_progress(false),
1002 _inlining_incrementally(false),
1003 _print_inlining_list(NULL),
1004 _print_inlining_stream(NULL),
1005 _print_inlining_idx(0),
1006 _print_inlining_output(NULL),
1007 _allowed_reasons(0),
1008 _interpreter_frame_size(0),
1009 _max_node_limit(MaxNodeLimit),
1010 _has_reserved_stack_access(false) {
1011 C = this;
1012
1013 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1014 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1015
1016 #ifndef PRODUCT
1017 set_print_assembly(PrintFrameConverterAssembly);
1018 set_parsed_irreducible_loop(false);
1019 #endif
1020 set_has_irreducible_loop(false); // no loops
1021
1022 CompileWrapper cw(this);
1023 Init(/*AliasLevel=*/ 0);
1024 init_tf((*generator)());
1025
1026 {
1027 // The following is a dummy for the sake of GraphKit::gen_stub
1028 Unique_Node_List for_igvn(comp_arena());
1029 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1030 PhaseGVN gvn(Thread::current()->resource_area(),255);
1031 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1032 gvn.transform_no_reclaim(top());
1033
1034 GraphKit kit;
1035 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1036 }
1037
1038 NOT_PRODUCT( verify_graph_edges(); )
1039 Code_Gen();
1040 if (failing()) return;
1041
1042
1043 // Entry point will be accessed using compile->stub_entry_point();
1044 if (code_buffer() == NULL) {
1045 Matcher::soft_match_failure();
1046 } else {
1047 if (PrintAssembly && (WizardMode || Verbose))
1048 tty->print_cr("### Stub::%s", stub_name);
1049
1050 if (!failing()) {
1051 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1052
1053 // Make the NMethod
1054 // For now we mark the frame as never safe for profile stackwalking
1055 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1056 code_buffer(),
1057 CodeOffsets::frame_never_safe,
1058 // _code_offsets.value(CodeOffsets::Frame_Complete),
1059 frame_size_in_words(),
1060 _oop_map_set,
1061 save_arg_registers);
1062 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1063
1064 _stub_entry_point = rs->entry_point();
1065 }
1066 }
1067 }
1068
1069 //------------------------------Init-------------------------------------------
1070 // Prepare for a single compilation
1071 void Compile::Init(int aliaslevel) {
1072 _unique = 0;
1073 _regalloc = NULL;
1074
1075 _tf = NULL; // filled in later
1076 _top = NULL; // cached later
1077 _matcher = NULL; // filled in later
1078 _cfg = NULL; // filled in later
1079
1080 set_24_bit_selection_and_mode(Use24BitFP, false);
1081
1082 _node_note_array = NULL;
1083 _default_node_notes = NULL;
1084 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1085
1086 _immutable_memory = NULL; // filled in at first inquiry
1087
1088 // Globally visible Nodes
1089 // First set TOP to NULL to give safe behavior during creation of RootNode
1090 set_cached_top_node(NULL);
1091 set_root(new RootNode());
1092 // Now that you have a Root to point to, create the real TOP
1093 set_cached_top_node( new ConNode(Type::TOP) );
1094 set_recent_alloc(NULL, NULL);
1095
1096 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1097 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1098 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1099 env()->set_dependencies(new Dependencies(env()));
1100
1101 _fixed_slots = 0;
1102 set_has_split_ifs(false);
1103 set_has_loops(has_method() && method()->has_loops()); // first approximation
1104 set_has_stringbuilder(false);
1105 set_has_boxed_value(false);
1106 _trap_can_recompile = false; // no traps emitted yet
1107 _major_progress = true; // start out assuming good things will happen
1108 set_has_unsafe_access(false);
1109 set_max_vector_size(0);
1110 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1111 set_decompile_count(0);
1112
1113 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1114 set_num_loop_opts(LoopOptsCount);
1115 set_do_inlining(Inline);
1116 set_max_inline_size(MaxInlineSize);
1117 set_freq_inline_size(FreqInlineSize);
1118 set_do_scheduling(OptoScheduling);
1119 set_do_count_invocations(false);
1120 set_do_method_data_update(false);
1121
1122 set_do_vector_loop(false);
1123
1124 if (AllowVectorizeOnDemand) {
1125 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1126 set_do_vector_loop(true);
1127 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());})
1128 } else if (has_method() && method()->name() != 0 &&
1129 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1130 set_do_vector_loop(true);
1131 }
1132 }
1133 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1134 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());})
1135
1136 set_age_code(has_method() && method()->profile_aging());
1137 set_rtm_state(NoRTM); // No RTM lock eliding by default
1138 _max_node_limit = _directive->MaxNodeLimitOption;
1139
1140 #if INCLUDE_RTM_OPT
1141 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1142 int rtm_state = method()->method_data()->rtm_state();
1143 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1144 // Don't generate RTM lock eliding code.
1145 set_rtm_state(NoRTM);
1146 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1147 // Generate RTM lock eliding code without abort ratio calculation code.
1148 set_rtm_state(UseRTM);
1149 } else if (UseRTMDeopt) {
1150 // Generate RTM lock eliding code and include abort ratio calculation
1151 // code if UseRTMDeopt is on.
1152 set_rtm_state(ProfileRTM);
1153 }
1154 }
1155 #endif
1156 if (debug_info()->recording_non_safepoints()) {
1157 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1158 (comp_arena(), 8, 0, NULL));
1159 set_default_node_notes(Node_Notes::make(this));
1160 }
1161
1162 // // -- Initialize types before each compile --
1163 // // Update cached type information
1164 // if( _method && _method->constants() )
1165 // Type::update_loaded_types(_method, _method->constants());
1166
1167 // Init alias_type map.
1168 if (!_do_escape_analysis && aliaslevel == 3)
1169 aliaslevel = 2; // No unique types without escape analysis
1170 _AliasLevel = aliaslevel;
1171 const int grow_ats = 16;
1172 _max_alias_types = grow_ats;
1173 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1174 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1175 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1176 {
1177 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1178 }
1179 // Initialize the first few types.
1180 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1181 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1182 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1183 _num_alias_types = AliasIdxRaw+1;
1184 // Zero out the alias type cache.
1185 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1186 // A NULL adr_type hits in the cache right away. Preload the right answer.
1187 probe_alias_cache(NULL)->_index = AliasIdxTop;
1188
1189 _intrinsics = NULL;
1190 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1191 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1192 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1193 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1194 _shenandoah_barriers = new(comp_arena()) GrowableArray<ShenandoahWriteBarrierNode*>(comp_arena(), 8, 0, NULL);
1195 register_library_intrinsics();
1196 }
1197
1198 //---------------------------init_start----------------------------------------
1199 // Install the StartNode on this compile object.
1200 void Compile::init_start(StartNode* s) {
1201 if (failing())
1202 return; // already failing
1203 assert(s == start(), "");
1204 }
1205
1206 /**
1207 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1208 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1209 * the ideal graph.
1210 */
1211 StartNode* Compile::start() const {
1212 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1213 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1214 Node* start = root()->fast_out(i);
1215 if (start->is_Start()) {
1216 return start->as_Start();
1217 }
1218 }
1219 fatal("Did not find Start node!");
1220 return NULL;
1221 }
1222
1223 //-------------------------------immutable_memory-------------------------------------
1224 // Access immutable memory
1225 Node* Compile::immutable_memory() {
1226 if (_immutable_memory != NULL) {
1227 return _immutable_memory;
1228 }
1229 StartNode* s = start();
1230 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1231 Node *p = s->fast_out(i);
1232 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1233 _immutable_memory = p;
1234 return _immutable_memory;
1235 }
1236 }
1237 ShouldNotReachHere();
1238 return NULL;
1239 }
1240
1241 //----------------------set_cached_top_node------------------------------------
1242 // Install the cached top node, and make sure Node::is_top works correctly.
1243 void Compile::set_cached_top_node(Node* tn) {
1244 if (tn != NULL) verify_top(tn);
1245 Node* old_top = _top;
1246 _top = tn;
1247 // Calling Node::setup_is_top allows the nodes the chance to adjust
1248 // their _out arrays.
1249 if (_top != NULL) _top->setup_is_top();
1250 if (old_top != NULL) old_top->setup_is_top();
1251 assert(_top == NULL || top()->is_top(), "");
1252 }
1253
1254 #ifdef ASSERT
1255 uint Compile::count_live_nodes_by_graph_walk() {
1256 Unique_Node_List useful(comp_arena());
1257 // Get useful node list by walking the graph.
1258 identify_useful_nodes(useful);
1259 return useful.size();
1260 }
1261
1262 void Compile::print_missing_nodes() {
1263
1264 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1265 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1266 return;
1267 }
1268
1269 // This is an expensive function. It is executed only when the user
1270 // specifies VerifyIdealNodeCount option or otherwise knows the
1271 // additional work that needs to be done to identify reachable nodes
1272 // by walking the flow graph and find the missing ones using
1273 // _dead_node_list.
1274
1275 Unique_Node_List useful(comp_arena());
1276 // Get useful node list by walking the graph.
1277 identify_useful_nodes(useful);
1278
1279 uint l_nodes = C->live_nodes();
1280 uint l_nodes_by_walk = useful.size();
1281
1282 if (l_nodes != l_nodes_by_walk) {
1283 if (_log != NULL) {
1284 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1285 _log->stamp();
1286 _log->end_head();
1287 }
1288 VectorSet& useful_member_set = useful.member_set();
1289 int last_idx = l_nodes_by_walk;
1290 for (int i = 0; i < last_idx; i++) {
1291 if (useful_member_set.test(i)) {
1292 if (_dead_node_list.test(i)) {
1293 if (_log != NULL) {
1294 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1295 }
1296 if (PrintIdealNodeCount) {
1297 // Print the log message to tty
1298 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1299 useful.at(i)->dump();
1300 }
1301 }
1302 }
1303 else if (! _dead_node_list.test(i)) {
1304 if (_log != NULL) {
1305 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1306 }
1307 if (PrintIdealNodeCount) {
1308 // Print the log message to tty
1309 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1310 }
1311 }
1312 }
1313 if (_log != NULL) {
1314 _log->tail("mismatched_nodes");
1315 }
1316 }
1317 }
1318 void Compile::record_modified_node(Node* n) {
1319 if (_modified_nodes != NULL && !_inlining_incrementally &&
1320 n->outcnt() != 0 && !n->is_Con()) {
1321 _modified_nodes->push(n);
1322 }
1323 }
1324
1325 void Compile::remove_modified_node(Node* n) {
1326 if (_modified_nodes != NULL) {
1327 _modified_nodes->remove(n);
1328 }
1329 }
1330 #endif
1331
1332 #ifndef PRODUCT
1333 void Compile::verify_top(Node* tn) const {
1334 if (tn != NULL) {
1335 assert(tn->is_Con(), "top node must be a constant");
1336 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1337 assert(tn->in(0) != NULL, "must have live top node");
1338 }
1339 }
1340 #endif
1341
1342
1343 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1344
1345 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1346 guarantee(arr != NULL, "");
1347 int num_blocks = arr->length();
1348 if (grow_by < num_blocks) grow_by = num_blocks;
1349 int num_notes = grow_by * _node_notes_block_size;
1350 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1351 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1352 while (num_notes > 0) {
1353 arr->append(notes);
1354 notes += _node_notes_block_size;
1355 num_notes -= _node_notes_block_size;
1356 }
1357 assert(num_notes == 0, "exact multiple, please");
1358 }
1359
1360 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1361 if (source == NULL || dest == NULL) return false;
1362
1363 if (dest->is_Con())
1364 return false; // Do not push debug info onto constants.
1365
1366 #ifdef ASSERT
1367 // Leave a bread crumb trail pointing to the original node:
1368 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1369 dest->set_debug_orig(source);
1370 }
1371 #endif
1372
1373 if (node_note_array() == NULL)
1374 return false; // Not collecting any notes now.
1375
1376 // This is a copy onto a pre-existing node, which may already have notes.
1377 // If both nodes have notes, do not overwrite any pre-existing notes.
1378 Node_Notes* source_notes = node_notes_at(source->_idx);
1379 if (source_notes == NULL || source_notes->is_clear()) return false;
1380 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1381 if (dest_notes == NULL || dest_notes->is_clear()) {
1382 return set_node_notes_at(dest->_idx, source_notes);
1383 }
1384
1385 Node_Notes merged_notes = (*source_notes);
1386 // The order of operations here ensures that dest notes will win...
1387 merged_notes.update_from(dest_notes);
1388 return set_node_notes_at(dest->_idx, &merged_notes);
1389 }
1390
1391
1392 //--------------------------allow_range_check_smearing-------------------------
1393 // Gating condition for coalescing similar range checks.
1394 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1395 // single covering check that is at least as strong as any of them.
1396 // If the optimization succeeds, the simplified (strengthened) range check
1397 // will always succeed. If it fails, we will deopt, and then give up
1398 // on the optimization.
1399 bool Compile::allow_range_check_smearing() const {
1400 // If this method has already thrown a range-check,
1401 // assume it was because we already tried range smearing
1402 // and it failed.
1403 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1404 return !already_trapped;
1405 }
1406
1407
1408 //------------------------------flatten_alias_type-----------------------------
1409 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1410 int offset = tj->offset();
1411 TypePtr::PTR ptr = tj->ptr();
1412
1413 // Known instance (scalarizable allocation) alias only with itself.
1414 bool is_known_inst = tj->isa_oopptr() != NULL &&
1415 tj->is_oopptr()->is_known_instance();
1416
1417 // Process weird unsafe references.
1418 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1419 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1420 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1421 tj = TypeOopPtr::BOTTOM;
1422 ptr = tj->ptr();
1423 offset = tj->offset();
1424 }
1425
1426 // Array pointers need some flattening
1427 const TypeAryPtr *ta = tj->isa_aryptr();
1428 if (ta && ta->is_stable()) {
1429 // Erase stability property for alias analysis.
1430 tj = ta = ta->cast_to_stable(false);
1431 }
1432 if( ta && is_known_inst ) {
1433 if ( offset != Type::OffsetBot &&
1434 offset > arrayOopDesc::length_offset_in_bytes() ) {
1435 offset = Type::OffsetBot; // Flatten constant access into array body only
1436 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1437 }
1438 } else if( ta && _AliasLevel >= 2 ) {
1439 // For arrays indexed by constant indices, we flatten the alias
1440 // space to include all of the array body. Only the header, klass
1441 // and array length can be accessed un-aliased.
1442 if( offset != Type::OffsetBot ) {
1443 if( ta->const_oop() ) { // MethodData* or Method*
1444 offset = Type::OffsetBot; // Flatten constant access into array body
1445 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1446 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1447 // range is OK as-is.
1448 tj = ta = TypeAryPtr::RANGE;
1449 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1450 tj = TypeInstPtr::KLASS; // all klass loads look alike
1451 ta = TypeAryPtr::RANGE; // generic ignored junk
1452 ptr = TypePtr::BotPTR;
1453 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1454 tj = TypeInstPtr::MARK;
1455 ta = TypeAryPtr::RANGE; // generic ignored junk
1456 ptr = TypePtr::BotPTR;
1457 } else if (offset == BrooksPointer::byte_offset() && UseShenandoahGC) {
1458 // Need to distinguish brooks ptr as is.
1459 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1460 } else { // Random constant offset into array body
1461 offset = Type::OffsetBot; // Flatten constant access into array body
1462 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1463 }
1464 }
1465 // Arrays of fixed size alias with arrays of unknown size.
1466 if (ta->size() != TypeInt::POS) {
1467 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1468 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1469 }
1470 // Arrays of known objects become arrays of unknown objects.
1471 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1472 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1473 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1474 }
1475 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1476 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1477 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1478 }
1479 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1480 // cannot be distinguished by bytecode alone.
1481 if (ta->elem() == TypeInt::BOOL) {
1482 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1483 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1484 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1485 }
1486 // During the 2nd round of IterGVN, NotNull castings are removed.
1487 // Make sure the Bottom and NotNull variants alias the same.
1488 // Also, make sure exact and non-exact variants alias the same.
1489 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1490 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1491 }
1492 }
1493
1494 // Oop pointers need some flattening
1495 const TypeInstPtr *to = tj->isa_instptr();
1496 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1497 ciInstanceKlass *k = to->klass()->as_instance_klass();
1498 if( ptr == TypePtr::Constant ) {
1499 if (to->klass() != ciEnv::current()->Class_klass() ||
1500 offset < k->size_helper() * wordSize) {
1501 // No constant oop pointers (such as Strings); they alias with
1502 // unknown strings.
1503 assert(!is_known_inst, "not scalarizable allocation");
1504 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1505 }
1506 } else if( is_known_inst ) {
1507 tj = to; // Keep NotNull and klass_is_exact for instance type
1508 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1509 // During the 2nd round of IterGVN, NotNull castings are removed.
1510 // Make sure the Bottom and NotNull variants alias the same.
1511 // Also, make sure exact and non-exact variants alias the same.
1512 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1513 }
1514 if (to->speculative() != NULL) {
1515 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1516 }
1517 // Canonicalize the holder of this field
1518 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1519 // First handle header references such as a LoadKlassNode, even if the
1520 // object's klass is unloaded at compile time (4965979).
1521 if (!is_known_inst) { // Do it only for non-instance types
1522 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1523 }
1524 } else if ((offset != BrooksPointer::byte_offset() || !UseShenandoahGC) && (offset < 0 || offset >= k->size_helper() * wordSize)) {
1525 // Static fields are in the space above the normal instance
1526 // fields in the java.lang.Class instance.
1527 if (to->klass() != ciEnv::current()->Class_klass()) {
1528 to = NULL;
1529 tj = TypeOopPtr::BOTTOM;
1530 offset = tj->offset();
1531 }
1532 } else {
1533 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1534 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1535 if( is_known_inst ) {
1536 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1537 } else {
1538 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1539 }
1540 }
1541 }
1542 }
1543
1544 // Klass pointers to object array klasses need some flattening
1545 const TypeKlassPtr *tk = tj->isa_klassptr();
1546 if( tk ) {
1547 // If we are referencing a field within a Klass, we need
1548 // to assume the worst case of an Object. Both exact and
1549 // inexact types must flatten to the same alias class so
1550 // use NotNull as the PTR.
1551 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1552
1553 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1554 TypeKlassPtr::OBJECT->klass(),
1555 offset);
1556 }
1557
1558 ciKlass* klass = tk->klass();
1559 if( klass->is_obj_array_klass() ) {
1560 ciKlass* k = TypeAryPtr::OOPS->klass();
1561 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1562 k = TypeInstPtr::BOTTOM->klass();
1563 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1564 }
1565
1566 // Check for precise loads from the primary supertype array and force them
1567 // to the supertype cache alias index. Check for generic array loads from
1568 // the primary supertype array and also force them to the supertype cache
1569 // alias index. Since the same load can reach both, we need to merge
1570 // these 2 disparate memories into the same alias class. Since the
1571 // primary supertype array is read-only, there's no chance of confusion
1572 // where we bypass an array load and an array store.
1573 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1574 if (offset == Type::OffsetBot ||
1575 (offset >= primary_supers_offset &&
1576 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1577 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1578 offset = in_bytes(Klass::secondary_super_cache_offset());
1579 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1580 }
1581 }
1582
1583 // Flatten all Raw pointers together.
1584 if (tj->base() == Type::RawPtr)
1585 tj = TypeRawPtr::BOTTOM;
1586
1587 if (tj->base() == Type::AnyPtr)
1588 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1589
1590 // Flatten all to bottom for now
1591 switch( _AliasLevel ) {
1592 case 0:
1593 tj = TypePtr::BOTTOM;
1594 break;
1595 case 1: // Flatten to: oop, static, field or array
1596 switch (tj->base()) {
1597 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1598 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1599 case Type::AryPtr: // do not distinguish arrays at all
1600 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1601 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1602 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1603 default: ShouldNotReachHere();
1604 }
1605 break;
1606 case 2: // No collapsing at level 2; keep all splits
1607 case 3: // No collapsing at level 3; keep all splits
1608 break;
1609 default:
1610 Unimplemented();
1611 }
1612
1613 offset = tj->offset();
1614 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1615
1616 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1617 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1618 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1619 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1620 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1621 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1622 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1623 (offset == BrooksPointer::byte_offset() && tj->base() == Type::AryPtr && UseShenandoahGC),
1624 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1625 assert( tj->ptr() != TypePtr::TopPTR &&
1626 tj->ptr() != TypePtr::AnyNull &&
1627 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1628 // assert( tj->ptr() != TypePtr::Constant ||
1629 // tj->base() == Type::RawPtr ||
1630 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1631
1632 return tj;
1633 }
1634
1635 void Compile::AliasType::Init(int i, const TypePtr* at) {
1636 _index = i;
1637 _adr_type = at;
1638 _field = NULL;
1639 _element = NULL;
1640 _is_rewritable = true; // default
1641 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1642 if (atoop != NULL && atoop->is_known_instance()) {
1643 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1644 _general_index = Compile::current()->get_alias_index(gt);
1645 } else {
1646 _general_index = 0;
1647 }
1648 }
1649
1650 BasicType Compile::AliasType::basic_type() const {
1651 if (element() != NULL) {
1652 const Type* element = adr_type()->is_aryptr()->elem();
1653 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1654 } if (field() != NULL) {
1655 return field()->layout_type();
1656 } else {
1657 return T_ILLEGAL; // unknown
1658 }
1659 }
1660
1661 //---------------------------------print_on------------------------------------
1662 #ifndef PRODUCT
1663 void Compile::AliasType::print_on(outputStream* st) {
1664 if (index() < 10)
1665 st->print("@ <%d> ", index());
1666 else st->print("@ <%d>", index());
1667 st->print(is_rewritable() ? " " : " RO");
1668 int offset = adr_type()->offset();
1669 if (offset == Type::OffsetBot)
1670 st->print(" +any");
1671 else st->print(" +%-3d", offset);
1672 st->print(" in ");
1673 adr_type()->dump_on(st);
1674 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1675 if (field() != NULL && tjp) {
1676 if (tjp->klass() != field()->holder() ||
1677 tjp->offset() != field()->offset_in_bytes()) {
1678 st->print(" != ");
1679 field()->print();
1680 st->print(" ***");
1681 }
1682 }
1683 }
1684
1685 void print_alias_types() {
1686 Compile* C = Compile::current();
1687 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1688 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1689 C->alias_type(idx)->print_on(tty);
1690 tty->cr();
1691 }
1692 }
1693 #endif
1694
1695
1696 //----------------------------probe_alias_cache--------------------------------
1697 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1698 intptr_t key = (intptr_t) adr_type;
1699 key ^= key >> logAliasCacheSize;
1700 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1701 }
1702
1703
1704 //-----------------------------grow_alias_types--------------------------------
1705 void Compile::grow_alias_types() {
1706 const int old_ats = _max_alias_types; // how many before?
1707 const int new_ats = old_ats; // how many more?
1708 const int grow_ats = old_ats+new_ats; // how many now?
1709 _max_alias_types = grow_ats;
1710 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1711 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1712 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1713 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1714 }
1715
1716
1717 //--------------------------------find_alias_type------------------------------
1718 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1719 if (_AliasLevel == 0)
1720 return alias_type(AliasIdxBot);
1721
1722 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1723 if (ace->_adr_type == adr_type) {
1724 return alias_type(ace->_index);
1725 }
1726
1727 // Handle special cases.
1728 if (adr_type == NULL) return alias_type(AliasIdxTop);
1729 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1730
1731 // Do it the slow way.
1732 const TypePtr* flat = flatten_alias_type(adr_type);
1733
1734 #ifdef ASSERT
1735 {
1736 ResourceMark rm;
1737 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1738 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1739 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1740 Type::str(adr_type));
1741 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1742 const TypeOopPtr* foop = flat->is_oopptr();
1743 // Scalarizable allocations have exact klass always.
1744 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1745 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1746 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1747 Type::str(foop), Type::str(xoop));
1748 }
1749 }
1750 #endif
1751
1752 int idx = AliasIdxTop;
1753 for (int i = 0; i < num_alias_types(); i++) {
1754 if (alias_type(i)->adr_type() == flat) {
1755 idx = i;
1756 break;
1757 }
1758 }
1759
1760 if (idx == AliasIdxTop) {
1761 if (no_create) return NULL;
1762 // Grow the array if necessary.
1763 if (_num_alias_types == _max_alias_types) grow_alias_types();
1764 // Add a new alias type.
1765 idx = _num_alias_types++;
1766 _alias_types[idx]->Init(idx, flat);
1767 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1768 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1769 if (flat->isa_instptr()) {
1770 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1771 && flat->is_instptr()->klass() == env()->Class_klass())
1772 alias_type(idx)->set_rewritable(false);
1773 }
1774 if (flat->isa_aryptr()) {
1775 #ifdef ASSERT
1776 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1777 // (T_BYTE has the weakest alignment and size restrictions...)
1778 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1779 #endif
1780 if (flat->offset() == TypePtr::OffsetBot) {
1781 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1782 }
1783 }
1784 if (flat->isa_klassptr()) {
1785 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1786 alias_type(idx)->set_rewritable(false);
1787 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1788 alias_type(idx)->set_rewritable(false);
1789 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1790 alias_type(idx)->set_rewritable(false);
1791 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1792 alias_type(idx)->set_rewritable(false);
1793 }
1794 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1795 // but the base pointer type is not distinctive enough to identify
1796 // references into JavaThread.)
1797
1798 // Check for final fields.
1799 const TypeInstPtr* tinst = flat->isa_instptr();
1800 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1801 ciField* field;
1802 if (tinst->const_oop() != NULL &&
1803 tinst->klass() == ciEnv::current()->Class_klass() &&
1804 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1805 // static field
1806 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1807 field = k->get_field_by_offset(tinst->offset(), true);
1808 } else {
1809 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1810 field = k->get_field_by_offset(tinst->offset(), false);
1811 }
1812 assert(field == NULL ||
1813 original_field == NULL ||
1814 (field->holder() == original_field->holder() &&
1815 field->offset() == original_field->offset() &&
1816 field->is_static() == original_field->is_static()), "wrong field?");
1817 // Set field() and is_rewritable() attributes.
1818 if (field != NULL) alias_type(idx)->set_field(field);
1819 }
1820 }
1821
1822 // Fill the cache for next time.
1823 ace->_adr_type = adr_type;
1824 ace->_index = idx;
1825 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1826
1827 // Might as well try to fill the cache for the flattened version, too.
1828 AliasCacheEntry* face = probe_alias_cache(flat);
1829 if (face->_adr_type == NULL) {
1830 face->_adr_type = flat;
1831 face->_index = idx;
1832 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1833 }
1834
1835 return alias_type(idx);
1836 }
1837
1838
1839 Compile::AliasType* Compile::alias_type(ciField* field) {
1840 const TypeOopPtr* t;
1841 if (field->is_static())
1842 t = TypeInstPtr::make(field->holder()->java_mirror());
1843 else
1844 t = TypeOopPtr::make_from_klass_raw(field->holder());
1845 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1846 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1847 return atp;
1848 }
1849
1850
1851 //------------------------------have_alias_type--------------------------------
1852 bool Compile::have_alias_type(const TypePtr* adr_type) {
1853 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1854 if (ace->_adr_type == adr_type) {
1855 return true;
1856 }
1857
1858 // Handle special cases.
1859 if (adr_type == NULL) return true;
1860 if (adr_type == TypePtr::BOTTOM) return true;
1861
1862 return find_alias_type(adr_type, true, NULL) != NULL;
1863 }
1864
1865 //-----------------------------must_alias--------------------------------------
1866 // True if all values of the given address type are in the given alias category.
1867 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1868 if (alias_idx == AliasIdxBot) return true; // the universal category
1869 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1870 if (alias_idx == AliasIdxTop) return false; // the empty category
1871 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1872
1873 // the only remaining possible overlap is identity
1874 int adr_idx = get_alias_index(adr_type);
1875 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1876 assert(adr_idx == alias_idx ||
1877 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1878 && adr_type != TypeOopPtr::BOTTOM),
1879 "should not be testing for overlap with an unsafe pointer");
1880 return adr_idx == alias_idx;
1881 }
1882
1883 //------------------------------can_alias--------------------------------------
1884 // True if any values of the given address type are in the given alias category.
1885 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1886 if (alias_idx == AliasIdxTop) return false; // the empty category
1887 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1888 if (alias_idx == AliasIdxBot) return true; // the universal category
1889 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1890
1891 // the only remaining possible overlap is identity
1892 int adr_idx = get_alias_index(adr_type);
1893 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1894 return adr_idx == alias_idx;
1895 }
1896
1897
1898
1899 //---------------------------pop_warm_call-------------------------------------
1900 WarmCallInfo* Compile::pop_warm_call() {
1901 WarmCallInfo* wci = _warm_calls;
1902 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1903 return wci;
1904 }
1905
1906 //----------------------------Inline_Warm--------------------------------------
1907 int Compile::Inline_Warm() {
1908 // If there is room, try to inline some more warm call sites.
1909 // %%% Do a graph index compaction pass when we think we're out of space?
1910 if (!InlineWarmCalls) return 0;
1911
1912 int calls_made_hot = 0;
1913 int room_to_grow = NodeCountInliningCutoff - unique();
1914 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1915 int amount_grown = 0;
1916 WarmCallInfo* call;
1917 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1918 int est_size = (int)call->size();
1919 if (est_size > (room_to_grow - amount_grown)) {
1920 // This one won't fit anyway. Get rid of it.
1921 call->make_cold();
1922 continue;
1923 }
1924 call->make_hot();
1925 calls_made_hot++;
1926 amount_grown += est_size;
1927 amount_to_grow -= est_size;
1928 }
1929
1930 if (calls_made_hot > 0) set_major_progress();
1931 return calls_made_hot;
1932 }
1933
1934
1935 //----------------------------Finish_Warm--------------------------------------
1936 void Compile::Finish_Warm() {
1937 if (!InlineWarmCalls) return;
1938 if (failing()) return;
1939 if (warm_calls() == NULL) return;
1940
1941 // Clean up loose ends, if we are out of space for inlining.
1942 WarmCallInfo* call;
1943 while ((call = pop_warm_call()) != NULL) {
1944 call->make_cold();
1945 }
1946 }
1947
1948 //---------------------cleanup_loop_predicates-----------------------
1949 // Remove the opaque nodes that protect the predicates so that all unused
1950 // checks and uncommon_traps will be eliminated from the ideal graph
1951 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1952 if (predicate_count()==0) return;
1953 for (int i = predicate_count(); i > 0; i--) {
1954 Node * n = predicate_opaque1_node(i-1);
1955 assert(n->Opcode() == Op_Opaque1, "must be");
1956 igvn.replace_node(n, n->in(1));
1957 }
1958 assert(predicate_count()==0, "should be clean!");
1959 }
1960
1961 void Compile::add_range_check_cast(Node* n) {
1962 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1963 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1964 _range_check_casts->append(n);
1965 }
1966
1967 // Remove all range check dependent CastIINodes.
1968 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1969 for (int i = range_check_cast_count(); i > 0; i--) {
1970 Node* cast = range_check_cast_node(i-1);
1971 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1972 igvn.replace_node(cast, cast->in(1));
1973 }
1974 assert(range_check_cast_count() == 0, "should be empty");
1975 }
1976
1977 // StringOpts and late inlining of string methods
1978 void Compile::inline_string_calls(bool parse_time) {
1979 {
1980 // remove useless nodes to make the usage analysis simpler
1981 ResourceMark rm;
1982 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1983 }
1984
1985 {
1986 ResourceMark rm;
1987 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1988 PhaseStringOpts pso(initial_gvn(), for_igvn());
1989 print_method(PHASE_AFTER_STRINGOPTS, 3);
1990 }
1991
1992 // now inline anything that we skipped the first time around
1993 if (!parse_time) {
1994 _late_inlines_pos = _late_inlines.length();
1995 }
1996
1997 while (_string_late_inlines.length() > 0) {
1998 CallGenerator* cg = _string_late_inlines.pop();
1999 cg->do_late_inline();
2000 if (failing()) return;
2001 }
2002 _string_late_inlines.trunc_to(0);
2003 }
2004
2005 // Late inlining of boxing methods
2006 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2007 if (_boxing_late_inlines.length() > 0) {
2008 assert(has_boxed_value(), "inconsistent");
2009
2010 PhaseGVN* gvn = initial_gvn();
2011 set_inlining_incrementally(true);
2012
2013 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2014 for_igvn()->clear();
2015 gvn->replace_with(&igvn);
2016
2017 _late_inlines_pos = _late_inlines.length();
2018
2019 while (_boxing_late_inlines.length() > 0) {
2020 CallGenerator* cg = _boxing_late_inlines.pop();
2021 cg->do_late_inline();
2022 if (failing()) return;
2023 }
2024 _boxing_late_inlines.trunc_to(0);
2025
2026 {
2027 ResourceMark rm;
2028 PhaseRemoveUseless pru(gvn, for_igvn());
2029 }
2030
2031 igvn = PhaseIterGVN(gvn);
2032 igvn.optimize();
2033
2034 set_inlining_progress(false);
2035 set_inlining_incrementally(false);
2036 }
2037 }
2038
2039 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2040 assert(IncrementalInline, "incremental inlining should be on");
2041 PhaseGVN* gvn = initial_gvn();
2042
2043 set_inlining_progress(false);
2044 for_igvn()->clear();
2045 gvn->replace_with(&igvn);
2046
2047 {
2048 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2049 int i = 0;
2050 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2051 CallGenerator* cg = _late_inlines.at(i);
2052 _late_inlines_pos = i+1;
2053 cg->do_late_inline();
2054 if (failing()) return;
2055 }
2056 int j = 0;
2057 for (; i < _late_inlines.length(); i++, j++) {
2058 _late_inlines.at_put(j, _late_inlines.at(i));
2059 }
2060 _late_inlines.trunc_to(j);
2061 }
2062
2063 {
2064 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2065 ResourceMark rm;
2066 PhaseRemoveUseless pru(gvn, for_igvn());
2067 }
2068
2069 {
2070 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2071 igvn = PhaseIterGVN(gvn);
2072 }
2073 }
2074
2075 // Perform incremental inlining until bound on number of live nodes is reached
2076 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2077 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2078
2079 PhaseGVN* gvn = initial_gvn();
2080
2081 set_inlining_incrementally(true);
2082 set_inlining_progress(true);
2083 uint low_live_nodes = 0;
2084
2085 while(inlining_progress() && _late_inlines.length() > 0) {
2086
2087 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2088 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2089 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2090 // PhaseIdealLoop is expensive so we only try it once we are
2091 // out of live nodes and we only try it again if the previous
2092 // helped got the number of nodes down significantly
2093 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2094 if (failing()) return;
2095 low_live_nodes = live_nodes();
2096 _major_progress = true;
2097 }
2098
2099 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2100 break;
2101 }
2102 }
2103
2104 inline_incrementally_one(igvn);
2105
2106 if (failing()) return;
2107
2108 {
2109 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2110 igvn.optimize();
2111 }
2112
2113 if (failing()) return;
2114 }
2115
2116 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2117
2118 if (_string_late_inlines.length() > 0) {
2119 assert(has_stringbuilder(), "inconsistent");
2120 for_igvn()->clear();
2121 initial_gvn()->replace_with(&igvn);
2122
2123 inline_string_calls(false);
2124
2125 if (failing()) return;
2126
2127 {
2128 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2129 ResourceMark rm;
2130 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2131 }
2132
2133 {
2134 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2135 igvn = PhaseIterGVN(gvn);
2136 igvn.optimize();
2137 }
2138 }
2139
2140 set_inlining_incrementally(false);
2141 }
2142
2143
2144 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2145 if(loop_opts_cnt > 0) {
2146 debug_only( int cnt = 0; );
2147 while(major_progress() && (loop_opts_cnt > 0)) {
2148 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2149 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2150 PhaseIdealLoop ideal_loop(igvn, mode);
2151 loop_opts_cnt--;
2152 if (failing()) return false;
2153 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2154 }
2155 }
2156 return true;
2157 }
2158
2159 //------------------------------Optimize---------------------------------------
2160 // Given a graph, optimize it.
2161 void Compile::Optimize() {
2162 TracePhase tp("optimizer", &timers[_t_optimizer]);
2163
2164 #ifndef PRODUCT
2165 if (_directive->BreakAtCompileOption) {
2166 BREAKPOINT;
2167 }
2168
2169 #endif
2170
2171 ResourceMark rm;
2172 int loop_opts_cnt;
2173
2174 print_inlining_reinit();
2175
2176 NOT_PRODUCT( verify_graph_edges(); )
2177
2178 print_method(PHASE_AFTER_PARSING);
2179
2180 {
2181 // Iterative Global Value Numbering, including ideal transforms
2182 // Initialize IterGVN with types and values from parse-time GVN
2183 PhaseIterGVN igvn(initial_gvn());
2184 #ifdef ASSERT
2185 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2186 #endif
2187 {
2188 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2189 igvn.optimize();
2190 }
2191
2192 print_method(PHASE_ITER_GVN1, 2);
2193
2194 if (failing()) return;
2195
2196 inline_incrementally(igvn);
2197
2198 print_method(PHASE_INCREMENTAL_INLINE, 2);
2199
2200 if (failing()) return;
2201
2202 if (eliminate_boxing()) {
2203 // Inline valueOf() methods now.
2204 inline_boxing_calls(igvn);
2205
2206 if (AlwaysIncrementalInline) {
2207 inline_incrementally(igvn);
2208 }
2209
2210 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2211
2212 if (failing()) return;
2213 }
2214
2215 // Remove the speculative part of types and clean up the graph from
2216 // the extra CastPP nodes whose only purpose is to carry them. Do
2217 // that early so that optimizations are not disrupted by the extra
2218 // CastPP nodes.
2219 remove_speculative_types(igvn);
2220
2221 // No more new expensive nodes will be added to the list from here
2222 // so keep only the actual candidates for optimizations.
2223 cleanup_expensive_nodes(igvn);
2224
2225 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2226 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2227 initial_gvn()->replace_with(&igvn);
2228 for_igvn()->clear();
2229 Unique_Node_List new_worklist(C->comp_arena());
2230 {
2231 ResourceMark rm;
2232 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2233 }
2234 set_for_igvn(&new_worklist);
2235 igvn = PhaseIterGVN(initial_gvn());
2236 igvn.optimize();
2237 }
2238
2239 // Perform escape analysis
2240 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2241 if (has_loops()) {
2242 // Cleanup graph (remove dead nodes).
2243 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2244 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2245 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2246 if (failing()) return;
2247 }
2248 ConnectionGraph::do_analysis(this, &igvn);
2249
2250 if (failing()) return;
2251
2252 // Optimize out fields loads from scalar replaceable allocations.
2253 igvn.optimize();
2254 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2255
2256 if (failing()) return;
2257
2258 if (congraph() != NULL && macro_count() > 0) {
2259 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2260 PhaseMacroExpand mexp(igvn);
2261 mexp.eliminate_macro_nodes();
2262 igvn.set_delay_transform(false);
2263
2264 igvn.optimize();
2265 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2266
2267 if (failing()) return;
2268 }
2269 }
2270
2271 // Loop transforms on the ideal graph. Range Check Elimination,
2272 // peeling, unrolling, etc.
2273
2274 // Set loop opts counter
2275 loop_opts_cnt = num_loop_opts();
2276 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2277 {
2278 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2279 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2280 loop_opts_cnt--;
2281 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2282 if (failing()) return;
2283 }
2284 // Loop opts pass if partial peeling occurred in previous pass
2285 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2286 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2287 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2288 loop_opts_cnt--;
2289 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2290 if (failing()) return;
2291 }
2292 // Loop opts pass for loop-unrolling before CCP
2293 if(major_progress() && (loop_opts_cnt > 0)) {
2294 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2295 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2296 loop_opts_cnt--;
2297 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2298 }
2299 if (!failing()) {
2300 // Verify that last round of loop opts produced a valid graph
2301 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2302 PhaseIdealLoop::verify(igvn);
2303 }
2304 }
2305 if (failing()) return;
2306
2307 // Conditional Constant Propagation;
2308 PhaseCCP ccp( &igvn );
2309 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2310 {
2311 TracePhase tp("ccp", &timers[_t_ccp]);
2312 ccp.do_transform();
2313 }
2314 print_method(PHASE_CPP1, 2);
2315
2316 assert( true, "Break here to ccp.dump_old2new_map()");
2317
2318 // Iterative Global Value Numbering, including ideal transforms
2319 {
2320 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2321 igvn = ccp;
2322 igvn.optimize();
2323 }
2324
2325 print_method(PHASE_ITER_GVN2, 2);
2326
2327 if (failing()) return;
2328
2329 // Loop transforms on the ideal graph. Range Check Elimination,
2330 // peeling, unrolling, etc.
2331 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2332 return;
2333 }
2334 // Ensure that major progress is now clear
2335 C->clear_major_progress();
2336
2337 {
2338 // Verify that all previous optimizations produced a valid graph
2339 // at least to this point, even if no loop optimizations were done.
2340 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2341 PhaseIdealLoop::verify(igvn);
2342 }
2343
2344 if (range_check_cast_count() > 0) {
2345 // No more loop optimizations. Remove all range check dependent CastIINodes.
2346 C->remove_range_check_casts(igvn);
2347 igvn.optimize();
2348 }
2349
2350 #ifdef ASSERT
2351 if (UseShenandoahGC && ShenandoahVerifyOptoBarriers) {
2352 ShenandoahBarrierNode::verify(C->root());
2353 }
2354 #endif
2355
2356 {
2357 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2358 PhaseMacroExpand mex(igvn);
2359 if (mex.expand_macro_nodes()) {
2360 assert(failing(), "must bail out w/ explicit message");
2361 return;
2362 }
2363 }
2364
2365 if (!ShenandoahWriteBarrierNode::expand(this, igvn, loop_opts_cnt)) {
2366 assert(failing(), "must bail out w/ explicit message");
2367 return;
2368 }
2369
2370 DEBUG_ONLY( _modified_nodes = NULL; )
2371 } // (End scope of igvn; run destructor if necessary for asserts.)
2372
2373 process_print_inlining();
2374 // A method with only infinite loops has no edges entering loops from root
2375 {
2376 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2377 if (final_graph_reshaping()) {
2378 assert(failing(), "must bail out w/ explicit message");
2379 return;
2380 }
2381 }
2382
2383 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2384 }
2385
2386
2387 //------------------------------Code_Gen---------------------------------------
2388 // Given a graph, generate code for it
2389 void Compile::Code_Gen() {
2390 if (failing()) {
2391 return;
2392 }
2393
2394 // Perform instruction selection. You might think we could reclaim Matcher
2395 // memory PDQ, but actually the Matcher is used in generating spill code.
2396 // Internals of the Matcher (including some VectorSets) must remain live
2397 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2398 // set a bit in reclaimed memory.
2399
2400 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2401 // nodes. Mapping is only valid at the root of each matched subtree.
2402 NOT_PRODUCT( verify_graph_edges(); )
2403
2404 Matcher matcher;
2405 _matcher = &matcher;
2406 {
2407 TracePhase tp("matcher", &timers[_t_matcher]);
2408 matcher.match();
2409 }
2410 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2411 // nodes. Mapping is only valid at the root of each matched subtree.
2412 NOT_PRODUCT( verify_graph_edges(); )
2413
2414 // If you have too many nodes, or if matching has failed, bail out
2415 check_node_count(0, "out of nodes matching instructions");
2416 if (failing()) {
2417 return;
2418 }
2419
2420 // Build a proper-looking CFG
2421 PhaseCFG cfg(node_arena(), root(), matcher);
2422 _cfg = &cfg;
2423 {
2424 TracePhase tp("scheduler", &timers[_t_scheduler]);
2425 bool success = cfg.do_global_code_motion();
2426 if (!success) {
2427 return;
2428 }
2429
2430 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2431 NOT_PRODUCT( verify_graph_edges(); )
2432 debug_only( cfg.verify(); )
2433 }
2434
2435 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2436 _regalloc = ®alloc;
2437 {
2438 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2439 // Perform register allocation. After Chaitin, use-def chains are
2440 // no longer accurate (at spill code) and so must be ignored.
2441 // Node->LRG->reg mappings are still accurate.
2442 _regalloc->Register_Allocate();
2443
2444 // Bail out if the allocator builds too many nodes
2445 if (failing()) {
2446 return;
2447 }
2448 }
2449
2450 // Prior to register allocation we kept empty basic blocks in case the
2451 // the allocator needed a place to spill. After register allocation we
2452 // are not adding any new instructions. If any basic block is empty, we
2453 // can now safely remove it.
2454 {
2455 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2456 cfg.remove_empty_blocks();
2457 if (do_freq_based_layout()) {
2458 PhaseBlockLayout layout(cfg);
2459 } else {
2460 cfg.set_loop_alignment();
2461 }
2462 cfg.fixup_flow();
2463 }
2464
2465 // Apply peephole optimizations
2466 if( OptoPeephole ) {
2467 TracePhase tp("peephole", &timers[_t_peephole]);
2468 PhasePeephole peep( _regalloc, cfg);
2469 peep.do_transform();
2470 }
2471
2472 // Do late expand if CPU requires this.
2473 if (Matcher::require_postalloc_expand) {
2474 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2475 cfg.postalloc_expand(_regalloc);
2476 }
2477
2478 // Convert Nodes to instruction bits in a buffer
2479 {
2480 TraceTime tp("output", &timers[_t_output], CITime);
2481 Output();
2482 }
2483
2484 print_method(PHASE_FINAL_CODE);
2485
2486 // He's dead, Jim.
2487 _cfg = (PhaseCFG*)0xdeadbeef;
2488 _regalloc = (PhaseChaitin*)0xdeadbeef;
2489 }
2490
2491
2492 //------------------------------dump_asm---------------------------------------
2493 // Dump formatted assembly
2494 #ifndef PRODUCT
2495 void Compile::dump_asm(int *pcs, uint pc_limit) {
2496 bool cut_short = false;
2497 tty->print_cr("#");
2498 tty->print("# "); _tf->dump(); tty->cr();
2499 tty->print_cr("#");
2500
2501 // For all blocks
2502 int pc = 0x0; // Program counter
2503 char starts_bundle = ' ';
2504 _regalloc->dump_frame();
2505
2506 Node *n = NULL;
2507 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2508 if (VMThread::should_terminate()) {
2509 cut_short = true;
2510 break;
2511 }
2512 Block* block = _cfg->get_block(i);
2513 if (block->is_connector() && !Verbose) {
2514 continue;
2515 }
2516 n = block->head();
2517 if (pcs && n->_idx < pc_limit) {
2518 tty->print("%3.3x ", pcs[n->_idx]);
2519 } else {
2520 tty->print(" ");
2521 }
2522 block->dump_head(_cfg);
2523 if (block->is_connector()) {
2524 tty->print_cr(" # Empty connector block");
2525 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2526 tty->print_cr(" # Block is sole successor of call");
2527 }
2528
2529 // For all instructions
2530 Node *delay = NULL;
2531 for (uint j = 0; j < block->number_of_nodes(); j++) {
2532 if (VMThread::should_terminate()) {
2533 cut_short = true;
2534 break;
2535 }
2536 n = block->get_node(j);
2537 if (valid_bundle_info(n)) {
2538 Bundle* bundle = node_bundling(n);
2539 if (bundle->used_in_unconditional_delay()) {
2540 delay = n;
2541 continue;
2542 }
2543 if (bundle->starts_bundle()) {
2544 starts_bundle = '+';
2545 }
2546 }
2547
2548 if (WizardMode) {
2549 n->dump();
2550 }
2551
2552 if( !n->is_Region() && // Dont print in the Assembly
2553 !n->is_Phi() && // a few noisely useless nodes
2554 !n->is_Proj() &&
2555 !n->is_MachTemp() &&
2556 !n->is_SafePointScalarObject() &&
2557 !n->is_Catch() && // Would be nice to print exception table targets
2558 !n->is_MergeMem() && // Not very interesting
2559 !n->is_top() && // Debug info table constants
2560 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2561 ) {
2562 if (pcs && n->_idx < pc_limit)
2563 tty->print("%3.3x", pcs[n->_idx]);
2564 else
2565 tty->print(" ");
2566 tty->print(" %c ", starts_bundle);
2567 starts_bundle = ' ';
2568 tty->print("\t");
2569 n->format(_regalloc, tty);
2570 tty->cr();
2571 }
2572
2573 // If we have an instruction with a delay slot, and have seen a delay,
2574 // then back up and print it
2575 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2576 assert(delay != NULL, "no unconditional delay instruction");
2577 if (WizardMode) delay->dump();
2578
2579 if (node_bundling(delay)->starts_bundle())
2580 starts_bundle = '+';
2581 if (pcs && n->_idx < pc_limit)
2582 tty->print("%3.3x", pcs[n->_idx]);
2583 else
2584 tty->print(" ");
2585 tty->print(" %c ", starts_bundle);
2586 starts_bundle = ' ';
2587 tty->print("\t");
2588 delay->format(_regalloc, tty);
2589 tty->cr();
2590 delay = NULL;
2591 }
2592
2593 // Dump the exception table as well
2594 if( n->is_Catch() && (Verbose || WizardMode) ) {
2595 // Print the exception table for this offset
2596 _handler_table.print_subtable_for(pc);
2597 }
2598 }
2599
2600 if (pcs && n->_idx < pc_limit)
2601 tty->print_cr("%3.3x", pcs[n->_idx]);
2602 else
2603 tty->cr();
2604
2605 assert(cut_short || delay == NULL, "no unconditional delay branch");
2606
2607 } // End of per-block dump
2608 tty->cr();
2609
2610 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2611 }
2612 #endif
2613
2614 //------------------------------Final_Reshape_Counts---------------------------
2615 // This class defines counters to help identify when a method
2616 // may/must be executed using hardware with only 24-bit precision.
2617 struct Final_Reshape_Counts : public StackObj {
2618 int _call_count; // count non-inlined 'common' calls
2619 int _float_count; // count float ops requiring 24-bit precision
2620 int _double_count; // count double ops requiring more precision
2621 int _java_call_count; // count non-inlined 'java' calls
2622 int _inner_loop_count; // count loops which need alignment
2623 VectorSet _visited; // Visitation flags
2624 Node_List _tests; // Set of IfNodes & PCTableNodes
2625
2626 Final_Reshape_Counts() :
2627 _call_count(0), _float_count(0), _double_count(0),
2628 _java_call_count(0), _inner_loop_count(0),
2629 _visited( Thread::current()->resource_area() ) { }
2630
2631 void inc_call_count () { _call_count ++; }
2632 void inc_float_count () { _float_count ++; }
2633 void inc_double_count() { _double_count++; }
2634 void inc_java_call_count() { _java_call_count++; }
2635 void inc_inner_loop_count() { _inner_loop_count++; }
2636
2637 int get_call_count () const { return _call_count ; }
2638 int get_float_count () const { return _float_count ; }
2639 int get_double_count() const { return _double_count; }
2640 int get_java_call_count() const { return _java_call_count; }
2641 int get_inner_loop_count() const { return _inner_loop_count; }
2642 };
2643
2644 #ifdef ASSERT
2645 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2646 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2647 // Make sure the offset goes inside the instance layout.
2648 return k->contains_field_offset(tp->offset());
2649 // Note that OffsetBot and OffsetTop are very negative.
2650 }
2651 #endif
2652
2653 // Eliminate trivially redundant StoreCMs and accumulate their
2654 // precedence edges.
2655 void Compile::eliminate_redundant_card_marks(Node* n) {
2656 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2657 if (n->in(MemNode::Address)->outcnt() > 1) {
2658 // There are multiple users of the same address so it might be
2659 // possible to eliminate some of the StoreCMs
2660 Node* mem = n->in(MemNode::Memory);
2661 Node* adr = n->in(MemNode::Address);
2662 Node* val = n->in(MemNode::ValueIn);
2663 Node* prev = n;
2664 bool done = false;
2665 // Walk the chain of StoreCMs eliminating ones that match. As
2666 // long as it's a chain of single users then the optimization is
2667 // safe. Eliminating partially redundant StoreCMs would require
2668 // cloning copies down the other paths.
2669 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2670 if (adr == mem->in(MemNode::Address) &&
2671 val == mem->in(MemNode::ValueIn)) {
2672 // redundant StoreCM
2673 if (mem->req() > MemNode::OopStore) {
2674 // Hasn't been processed by this code yet.
2675 n->add_prec(mem->in(MemNode::OopStore));
2676 } else {
2677 // Already converted to precedence edge
2678 for (uint i = mem->req(); i < mem->len(); i++) {
2679 // Accumulate any precedence edges
2680 if (mem->in(i) != NULL) {
2681 n->add_prec(mem->in(i));
2682 }
2683 }
2684 // Everything above this point has been processed.
2685 done = true;
2686 }
2687 // Eliminate the previous StoreCM
2688 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2689 assert(mem->outcnt() == 0, "should be dead");
2690 mem->disconnect_inputs(NULL, this);
2691 } else {
2692 prev = mem;
2693 }
2694 mem = prev->in(MemNode::Memory);
2695 }
2696 }
2697 }
2698
2699 //------------------------------final_graph_reshaping_impl----------------------
2700 // Implement items 1-5 from final_graph_reshaping below.
2701 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2702
2703 if ( n->outcnt() == 0 ) return; // dead node
2704 uint nop = n->Opcode();
2705
2706 // Check for 2-input instruction with "last use" on right input.
2707 // Swap to left input. Implements item (2).
2708 if( n->req() == 3 && // two-input instruction
2709 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2710 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2711 n->in(2)->outcnt() == 1 &&// right use IS a last use
2712 !n->in(2)->is_Con() ) { // right use is not a constant
2713 // Check for commutative opcode
2714 switch( nop ) {
2715 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2716 case Op_MaxI: case Op_MinI:
2717 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2718 case Op_AndL: case Op_XorL: case Op_OrL:
2719 case Op_AndI: case Op_XorI: case Op_OrI: {
2720 // Move "last use" input to left by swapping inputs
2721 n->swap_edges(1, 2);
2722 break;
2723 }
2724 default:
2725 break;
2726 }
2727 }
2728
2729 #ifdef ASSERT
2730 if( n->is_Mem() ) {
2731 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2732 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2733 // oop will be recorded in oop map if load crosses safepoint
2734 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2735 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2736 "raw memory operations should have control edge");
2737 }
2738 #endif
2739 // Count FPU ops and common calls, implements item (3)
2740 switch( nop ) {
2741 // Count all float operations that may use FPU
2742 case Op_AddF:
2743 case Op_SubF:
2744 case Op_MulF:
2745 case Op_DivF:
2746 case Op_NegF:
2747 case Op_ModF:
2748 case Op_ConvI2F:
2749 case Op_ConF:
2750 case Op_CmpF:
2751 case Op_CmpF3:
2752 // case Op_ConvL2F: // longs are split into 32-bit halves
2753 frc.inc_float_count();
2754 break;
2755
2756 case Op_ConvF2D:
2757 case Op_ConvD2F:
2758 frc.inc_float_count();
2759 frc.inc_double_count();
2760 break;
2761
2762 // Count all double operations that may use FPU
2763 case Op_AddD:
2764 case Op_SubD:
2765 case Op_MulD:
2766 case Op_DivD:
2767 case Op_NegD:
2768 case Op_ModD:
2769 case Op_ConvI2D:
2770 case Op_ConvD2I:
2771 // case Op_ConvL2D: // handled by leaf call
2772 // case Op_ConvD2L: // handled by leaf call
2773 case Op_ConD:
2774 case Op_CmpD:
2775 case Op_CmpD3:
2776 frc.inc_double_count();
2777 break;
2778 case Op_Opaque1: // Remove Opaque Nodes before matching
2779 case Op_Opaque2: // Remove Opaque Nodes before matching
2780 case Op_Opaque3:
2781 n->subsume_by(n->in(1), this);
2782 break;
2783 case Op_CallStaticJava:
2784 case Op_CallJava:
2785 case Op_CallDynamicJava:
2786 frc.inc_java_call_count(); // Count java call site;
2787 case Op_CallRuntime:
2788 case Op_CallLeaf:
2789 case Op_CallLeafNoFP: {
2790 assert( n->is_Call(), "" );
2791 CallNode *call = n->as_Call();
2792 if (UseShenandoahGC && call->is_g1_wb_pre_call()) {
2793 uint cnt = OptoRuntime::g1_wb_pre_Type()->domain()->cnt();
2794 if (call->req() > cnt) {
2795 assert(call->req() == cnt+1, "only one extra input");
2796 Node* addp = call->in(cnt);
2797 assert(!CallLeafNode::has_only_g1_wb_pre_uses(addp), "useless address computation?");
2798 call->del_req(cnt);
2799 }
2800 }
2801 // Count call sites where the FP mode bit would have to be flipped.
2802 // Do not count uncommon runtime calls:
2803 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2804 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2805 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2806 frc.inc_call_count(); // Count the call site
2807 } else { // See if uncommon argument is shared
2808 Node *n = call->in(TypeFunc::Parms);
2809 int nop = n->Opcode();
2810 // Clone shared simple arguments to uncommon calls, item (1).
2811 if( n->outcnt() > 1 &&
2812 !n->is_Proj() &&
2813 nop != Op_CreateEx &&
2814 nop != Op_CheckCastPP &&
2815 nop != Op_DecodeN &&
2816 nop != Op_DecodeNKlass &&
2817 !n->is_Mem() ) {
2818 Node *x = n->clone();
2819 call->set_req( TypeFunc::Parms, x );
2820 }
2821 }
2822 break;
2823 }
2824
2825 case Op_StoreD:
2826 case Op_LoadD:
2827 case Op_LoadD_unaligned:
2828 frc.inc_double_count();
2829 goto handle_mem;
2830 case Op_StoreF:
2831 case Op_LoadF:
2832 frc.inc_float_count();
2833 goto handle_mem;
2834
2835 case Op_StoreCM:
2836 {
2837 // Convert OopStore dependence into precedence edge
2838 Node* prec = n->in(MemNode::OopStore);
2839 n->del_req(MemNode::OopStore);
2840 n->add_prec(prec);
2841 eliminate_redundant_card_marks(n);
2842 }
2843
2844 // fall through
2845
2846 case Op_StoreB:
2847 case Op_StoreC:
2848 case Op_StorePConditional:
2849 case Op_StoreI:
2850 case Op_StoreL:
2851 case Op_StoreIConditional:
2852 case Op_StoreLConditional:
2853 case Op_CompareAndSwapB:
2854 case Op_CompareAndSwapS:
2855 case Op_CompareAndSwapI:
2856 case Op_CompareAndSwapL:
2857 case Op_CompareAndSwapP:
2858 case Op_CompareAndSwapN:
2859 case Op_WeakCompareAndSwapB:
2860 case Op_WeakCompareAndSwapS:
2861 case Op_WeakCompareAndSwapI:
2862 case Op_WeakCompareAndSwapL:
2863 case Op_WeakCompareAndSwapP:
2864 case Op_WeakCompareAndSwapN:
2865 case Op_CompareAndExchangeB:
2866 case Op_CompareAndExchangeS:
2867 case Op_CompareAndExchangeI:
2868 case Op_CompareAndExchangeL:
2869 case Op_CompareAndExchangeP:
2870 case Op_CompareAndExchangeN:
2871 case Op_GetAndAddS:
2872 case Op_GetAndAddB:
2873 case Op_GetAndAddI:
2874 case Op_GetAndAddL:
2875 case Op_GetAndSetS:
2876 case Op_GetAndSetB:
2877 case Op_GetAndSetI:
2878 case Op_GetAndSetL:
2879 case Op_GetAndSetP:
2880 case Op_GetAndSetN:
2881 case Op_StoreP:
2882 case Op_StoreN:
2883 case Op_StoreNKlass:
2884 case Op_LoadB:
2885 case Op_LoadUB:
2886 case Op_LoadUS:
2887 case Op_LoadI:
2888 case Op_LoadKlass:
2889 case Op_LoadNKlass:
2890 case Op_LoadL:
2891 case Op_LoadL_unaligned:
2892 case Op_LoadPLocked:
2893 case Op_LoadP:
2894 case Op_LoadN:
2895 case Op_LoadRange:
2896 case Op_LoadS: {
2897 handle_mem:
2898 #ifdef ASSERT
2899 if( VerifyOptoOopOffsets ) {
2900 assert( n->is_Mem(), "" );
2901 MemNode *mem = (MemNode*)n;
2902 // Check to see if address types have grounded out somehow.
2903 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2904 assert( !tp || oop_offset_is_sane(tp), "" );
2905 }
2906 #endif
2907 break;
2908 }
2909
2910 case Op_AddP: { // Assert sane base pointers
2911 Node *addp = n->in(AddPNode::Address);
2912 assert( !addp->is_AddP() ||
2913 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2914 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2915 "Base pointers must match (addp %u)", addp->_idx );
2916 #ifdef _LP64
2917 if ((UseCompressedOops || UseCompressedClassPointers) &&
2918 addp->Opcode() == Op_ConP &&
2919 addp == n->in(AddPNode::Base) &&
2920 n->in(AddPNode::Offset)->is_Con()) {
2921 // If the transformation of ConP to ConN+DecodeN is beneficial depends
2922 // on the platform and on the compressed oops mode.
2923 // Use addressing with narrow klass to load with offset on x86.
2924 // Some platforms can use the constant pool to load ConP.
2925 // Do this transformation here since IGVN will convert ConN back to ConP.
2926 const Type* t = addp->bottom_type();
2927 bool is_oop = t->isa_oopptr() != NULL;
2928 bool is_klass = t->isa_klassptr() != NULL;
2929
2930 if ((is_oop && Matcher::const_oop_prefer_decode() ) ||
2931 (is_klass && Matcher::const_klass_prefer_decode())) {
2932 Node* nn = NULL;
2933
2934 int op = is_oop ? Op_ConN : Op_ConNKlass;
2935
2936 // Look for existing ConN node of the same exact type.
2937 Node* r = root();
2938 uint cnt = r->outcnt();
2939 for (uint i = 0; i < cnt; i++) {
2940 Node* m = r->raw_out(i);
2941 if (m!= NULL && m->Opcode() == op &&
2942 m->bottom_type()->make_ptr() == t) {
2943 nn = m;
2944 break;
2945 }
2946 }
2947 if (nn != NULL) {
2948 // Decode a narrow oop to match address
2949 // [R12 + narrow_oop_reg<<3 + offset]
2950 if (is_oop) {
2951 nn = new DecodeNNode(nn, t);
2952 } else {
2953 nn = new DecodeNKlassNode(nn, t);
2954 }
2955 // Check for succeeding AddP which uses the same Base.
2956 // Otherwise we will run into the assertion above when visiting that guy.
2957 for (uint i = 0; i < n->outcnt(); ++i) {
2958 Node *out_i = n->raw_out(i);
2959 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
2960 out_i->set_req(AddPNode::Base, nn);
2961 #ifdef ASSERT
2962 for (uint j = 0; j < out_i->outcnt(); ++j) {
2963 Node *out_j = out_i->raw_out(j);
2964 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
2965 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
2966 }
2967 #endif
2968 }
2969 }
2970 n->set_req(AddPNode::Base, nn);
2971 n->set_req(AddPNode::Address, nn);
2972 if (addp->outcnt() == 0) {
2973 addp->disconnect_inputs(NULL, this);
2974 }
2975 }
2976 }
2977 }
2978 #endif
2979 // platform dependent reshaping of the address expression
2980 reshape_address(n->as_AddP());
2981 break;
2982 }
2983
2984 case Op_CastPP: {
2985 // Remove CastPP nodes to gain more freedom during scheduling but
2986 // keep the dependency they encode as control or precedence edges
2987 // (if control is set already) on memory operations. Some CastPP
2988 // nodes don't have a control (don't carry a dependency): skip
2989 // those.
2990 if (n->in(0) != NULL) {
2991 ResourceMark rm;
2992 Unique_Node_List wq;
2993 wq.push(n);
2994 for (uint next = 0; next < wq.size(); ++next) {
2995 Node *m = wq.at(next);
2996 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2997 Node* use = m->fast_out(i);
2998 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) {
2999 use->ensure_control_or_add_prec(n->in(0));
3000 } else {
3001 switch(use->Opcode()) {
3002 case Op_AddP:
3003 case Op_DecodeN:
3004 case Op_DecodeNKlass:
3005 case Op_CheckCastPP:
3006 case Op_CastPP:
3007 wq.push(use);
3008 break;
3009 }
3010 }
3011 }
3012 }
3013 }
3014 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3015 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3016 Node* in1 = n->in(1);
3017 const Type* t = n->bottom_type();
3018 Node* new_in1 = in1->clone();
3019 new_in1->as_DecodeN()->set_type(t);
3020
3021 if (!Matcher::narrow_oop_use_complex_address()) {
3022 //
3023 // x86, ARM and friends can handle 2 adds in addressing mode
3024 // and Matcher can fold a DecodeN node into address by using
3025 // a narrow oop directly and do implicit NULL check in address:
3026 //
3027 // [R12 + narrow_oop_reg<<3 + offset]
3028 // NullCheck narrow_oop_reg
3029 //
3030 // On other platforms (Sparc) we have to keep new DecodeN node and
3031 // use it to do implicit NULL check in address:
3032 //
3033 // decode_not_null narrow_oop_reg, base_reg
3034 // [base_reg + offset]
3035 // NullCheck base_reg
3036 //
3037 // Pin the new DecodeN node to non-null path on these platform (Sparc)
3038 // to keep the information to which NULL check the new DecodeN node
3039 // corresponds to use it as value in implicit_null_check().
3040 //
3041 new_in1->set_req(0, n->in(0));
3042 }
3043
3044 n->subsume_by(new_in1, this);
3045 if (in1->outcnt() == 0) {
3046 in1->disconnect_inputs(NULL, this);
3047 }
3048 } else {
3049 n->subsume_by(n->in(1), this);
3050 if (n->outcnt() == 0) {
3051 n->disconnect_inputs(NULL, this);
3052 }
3053 }
3054 break;
3055 }
3056 #ifdef _LP64
3057 case Op_CmpP:
3058 // Do this transformation here to preserve CmpPNode::sub() and
3059 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3060 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3061 Node* in1 = n->in(1);
3062 Node* in2 = n->in(2);
3063 if (!in1->is_DecodeNarrowPtr()) {
3064 in2 = in1;
3065 in1 = n->in(2);
3066 }
3067 assert(in1->is_DecodeNarrowPtr(), "sanity");
3068
3069 Node* new_in2 = NULL;
3070 if (in2->is_DecodeNarrowPtr()) {
3071 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3072 new_in2 = in2->in(1);
3073 } else if (in2->Opcode() == Op_ConP) {
3074 const Type* t = in2->bottom_type();
3075 if (t == TypePtr::NULL_PTR) {
3076 assert(in1->is_DecodeN(), "compare klass to null?");
3077 // Don't convert CmpP null check into CmpN if compressed
3078 // oops implicit null check is not generated.
3079 // This will allow to generate normal oop implicit null check.
3080 if (Matcher::gen_narrow_oop_implicit_null_checks())
3081 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3082 //
3083 // This transformation together with CastPP transformation above
3084 // will generated code for implicit NULL checks for compressed oops.
3085 //
3086 // The original code after Optimize()
3087 //
3088 // LoadN memory, narrow_oop_reg
3089 // decode narrow_oop_reg, base_reg
3090 // CmpP base_reg, NULL
3091 // CastPP base_reg // NotNull
3092 // Load [base_reg + offset], val_reg
3093 //
3094 // after these transformations will be
3095 //
3096 // LoadN memory, narrow_oop_reg
3097 // CmpN narrow_oop_reg, NULL
3098 // decode_not_null narrow_oop_reg, base_reg
3099 // Load [base_reg + offset], val_reg
3100 //
3101 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3102 // since narrow oops can be used in debug info now (see the code in
3103 // final_graph_reshaping_walk()).
3104 //
3105 // At the end the code will be matched to
3106 // on x86:
3107 //
3108 // Load_narrow_oop memory, narrow_oop_reg
3109 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3110 // NullCheck narrow_oop_reg
3111 //
3112 // and on sparc:
3113 //
3114 // Load_narrow_oop memory, narrow_oop_reg
3115 // decode_not_null narrow_oop_reg, base_reg
3116 // Load [base_reg + offset], val_reg
3117 // NullCheck base_reg
3118 //
3119 } else if (t->isa_oopptr()) {
3120 new_in2 = ConNode::make(t->make_narrowoop());
3121 } else if (t->isa_klassptr()) {
3122 new_in2 = ConNode::make(t->make_narrowklass());
3123 }
3124 }
3125 if (new_in2 != NULL) {
3126 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3127 n->subsume_by(cmpN, this);
3128 if (in1->outcnt() == 0) {
3129 in1->disconnect_inputs(NULL, this);
3130 }
3131 if (in2->outcnt() == 0) {
3132 in2->disconnect_inputs(NULL, this);
3133 }
3134 }
3135 }
3136 break;
3137
3138 case Op_DecodeN:
3139 case Op_DecodeNKlass:
3140 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3141 // DecodeN could be pinned when it can't be fold into
3142 // an address expression, see the code for Op_CastPP above.
3143 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3144 break;
3145
3146 case Op_EncodeP:
3147 case Op_EncodePKlass: {
3148 Node* in1 = n->in(1);
3149 if (in1->is_DecodeNarrowPtr()) {
3150 n->subsume_by(in1->in(1), this);
3151 } else if (in1->Opcode() == Op_ConP) {
3152 const Type* t = in1->bottom_type();
3153 if (t == TypePtr::NULL_PTR) {
3154 assert(t->isa_oopptr(), "null klass?");
3155 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3156 } else if (t->isa_oopptr()) {
3157 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3158 } else if (t->isa_klassptr()) {
3159 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3160 }
3161 }
3162 if (in1->outcnt() == 0) {
3163 in1->disconnect_inputs(NULL, this);
3164 }
3165 break;
3166 }
3167
3168 case Op_Proj: {
3169 if (OptimizeStringConcat) {
3170 ProjNode* p = n->as_Proj();
3171 if (p->_is_io_use) {
3172 // Separate projections were used for the exception path which
3173 // are normally removed by a late inline. If it wasn't inlined
3174 // then they will hang around and should just be replaced with
3175 // the original one.
3176 Node* proj = NULL;
3177 // Replace with just one
3178 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3179 Node *use = i.get();
3180 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3181 proj = use;
3182 break;
3183 }
3184 }
3185 assert(proj != NULL, "must be found");
3186 p->subsume_by(proj, this);
3187 }
3188 }
3189 break;
3190 }
3191
3192 case Op_Phi:
3193 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3194 // The EncodeP optimization may create Phi with the same edges
3195 // for all paths. It is not handled well by Register Allocator.
3196 Node* unique_in = n->in(1);
3197 assert(unique_in != NULL, "");
3198 uint cnt = n->req();
3199 for (uint i = 2; i < cnt; i++) {
3200 Node* m = n->in(i);
3201 assert(m != NULL, "");
3202 if (unique_in != m)
3203 unique_in = NULL;
3204 }
3205 if (unique_in != NULL) {
3206 n->subsume_by(unique_in, this);
3207 }
3208 }
3209 break;
3210
3211 #endif
3212
3213 #ifdef ASSERT
3214 case Op_CastII:
3215 // Verify that all range check dependent CastII nodes were removed.
3216 if (n->isa_CastII()->has_range_check()) {
3217 n->dump(3);
3218 assert(false, "Range check dependent CastII node was not removed");
3219 }
3220 break;
3221 #endif
3222
3223 case Op_ModI:
3224 if (UseDivMod) {
3225 // Check if a%b and a/b both exist
3226 Node* d = n->find_similar(Op_DivI);
3227 if (d) {
3228 // Replace them with a fused divmod if supported
3229 if (Matcher::has_match_rule(Op_DivModI)) {
3230 DivModINode* divmod = DivModINode::make(n);
3231 d->subsume_by(divmod->div_proj(), this);
3232 n->subsume_by(divmod->mod_proj(), this);
3233 } else {
3234 // replace a%b with a-((a/b)*b)
3235 Node* mult = new MulINode(d, d->in(2));
3236 Node* sub = new SubINode(d->in(1), mult);
3237 n->subsume_by(sub, this);
3238 }
3239 }
3240 }
3241 break;
3242
3243 case Op_ModL:
3244 if (UseDivMod) {
3245 // Check if a%b and a/b both exist
3246 Node* d = n->find_similar(Op_DivL);
3247 if (d) {
3248 // Replace them with a fused divmod if supported
3249 if (Matcher::has_match_rule(Op_DivModL)) {
3250 DivModLNode* divmod = DivModLNode::make(n);
3251 d->subsume_by(divmod->div_proj(), this);
3252 n->subsume_by(divmod->mod_proj(), this);
3253 } else {
3254 // replace a%b with a-((a/b)*b)
3255 Node* mult = new MulLNode(d, d->in(2));
3256 Node* sub = new SubLNode(d->in(1), mult);
3257 n->subsume_by(sub, this);
3258 }
3259 }
3260 }
3261 break;
3262
3263 case Op_LoadVector:
3264 case Op_StoreVector:
3265 break;
3266
3267 case Op_AddReductionVI:
3268 case Op_AddReductionVL:
3269 case Op_AddReductionVF:
3270 case Op_AddReductionVD:
3271 case Op_MulReductionVI:
3272 case Op_MulReductionVL:
3273 case Op_MulReductionVF:
3274 case Op_MulReductionVD:
3275 break;
3276
3277 case Op_PackB:
3278 case Op_PackS:
3279 case Op_PackI:
3280 case Op_PackF:
3281 case Op_PackL:
3282 case Op_PackD:
3283 if (n->req()-1 > 2) {
3284 // Replace many operand PackNodes with a binary tree for matching
3285 PackNode* p = (PackNode*) n;
3286 Node* btp = p->binary_tree_pack(1, n->req());
3287 n->subsume_by(btp, this);
3288 }
3289 break;
3290 case Op_Loop:
3291 case Op_CountedLoop:
3292 if (n->as_Loop()->is_inner_loop()) {
3293 frc.inc_inner_loop_count();
3294 }
3295 n->as_Loop()->verify_strip_mined(0);
3296 break;
3297 case Op_LShiftI:
3298 case Op_RShiftI:
3299 case Op_URShiftI:
3300 case Op_LShiftL:
3301 case Op_RShiftL:
3302 case Op_URShiftL:
3303 if (Matcher::need_masked_shift_count) {
3304 // The cpu's shift instructions don't restrict the count to the
3305 // lower 5/6 bits. We need to do the masking ourselves.
3306 Node* in2 = n->in(2);
3307 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3308 const TypeInt* t = in2->find_int_type();
3309 if (t != NULL && t->is_con()) {
3310 juint shift = t->get_con();
3311 if (shift > mask) { // Unsigned cmp
3312 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3313 }
3314 } else {
3315 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3316 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3317 n->set_req(2, shift);
3318 }
3319 }
3320 if (in2->outcnt() == 0) { // Remove dead node
3321 in2->disconnect_inputs(NULL, this);
3322 }
3323 }
3324 break;
3325 case Op_MemBarStoreStore:
3326 case Op_MemBarRelease:
3327 // Break the link with AllocateNode: it is no longer useful and
3328 // confuses register allocation.
3329 if (n->req() > MemBarNode::Precedent) {
3330 n->set_req(MemBarNode::Precedent, top());
3331 }
3332 break;
3333 case Op_ShenandoahReadBarrier:
3334 break;
3335 case Op_ShenandoahWriteBarrier:
3336 assert(!ShenandoahWriteBarrierToIR, "should have been expanded already");
3337 break;
3338 case Op_RangeCheck: {
3339 RangeCheckNode* rc = n->as_RangeCheck();
3340 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3341 n->subsume_by(iff, this);
3342 frc._tests.push(iff);
3343 break;
3344 }
3345 case Op_ConvI2L: {
3346 if (!Matcher::convi2l_type_required) {
3347 // Code generation on some platforms doesn't need accurate
3348 // ConvI2L types. Widening the type can help remove redundant
3349 // address computations.
3350 n->as_Type()->set_type(TypeLong::INT);
3351 ResourceMark rm;
3352 Node_List wq;
3353 wq.push(n);
3354 for (uint next = 0; next < wq.size(); next++) {
3355 Node *m = wq.at(next);
3356
3357 for(;;) {
3358 // Loop over all nodes with identical inputs edges as m
3359 Node* k = m->find_similar(m->Opcode());
3360 if (k == NULL) {
3361 break;
3362 }
3363 // Push their uses so we get a chance to remove node made
3364 // redundant
3365 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3366 Node* u = k->fast_out(i);
3367 assert(!wq.contains(u), "shouldn't process one node several times");
3368 if (u->Opcode() == Op_LShiftL ||
3369 u->Opcode() == Op_AddL ||
3370 u->Opcode() == Op_SubL ||
3371 u->Opcode() == Op_AddP) {
3372 wq.push(u);
3373 }
3374 }
3375 // Replace all nodes with identical edges as m with m
3376 k->subsume_by(m, this);
3377 }
3378 }
3379 }
3380 break;
3381 }
3382 case Op_If: {
3383 #ifdef ASSERT
3384 if (ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(n->as_If())) {
3385 Node* c = n->in(0);
3386 int count = 0;
3387 for (;;) {
3388 assert(c->is_Proj(), "proj expected");
3389 MemBarNode* mb = c->in(0)->as_MemBar();
3390 c = c->in(0)->in(0);
3391 count++;
3392 assert(mb->outcnt() == 2, "lost a projection?");
3393
3394 if (mb->adr_type() == TypeRawPtr::BOTTOM) {
3395 break;
3396 }
3397 }
3398 assert(count >= 2, "at least 2 membars");
3399 }
3400 #endif
3401 break;
3402 }
3403 default:
3404 assert( !n->is_Call(), "" );
3405 assert( !n->is_Mem(), "" );
3406 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3407 break;
3408 }
3409
3410 // Collect CFG split points
3411 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3412 frc._tests.push(n);
3413 }
3414 }
3415
3416 //------------------------------final_graph_reshaping_walk---------------------
3417 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3418 // requires that the walk visits a node's inputs before visiting the node.
3419 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3420 ResourceArea *area = Thread::current()->resource_area();
3421 Unique_Node_List sfpt(area);
3422
3423 frc._visited.set(root->_idx); // first, mark node as visited
3424 uint cnt = root->req();
3425 Node *n = root;
3426 uint i = 0;
3427 while (true) {
3428 if (i < cnt) {
3429 // Place all non-visited non-null inputs onto stack
3430 Node* m = n->in(i);
3431 ++i;
3432 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3433 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3434 // compute worst case interpreter size in case of a deoptimization
3435 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3436
3437 sfpt.push(m);
3438 }
3439 cnt = m->req();
3440 nstack.push(n, i); // put on stack parent and next input's index
3441 n = m;
3442 i = 0;
3443 }
3444 } else {
3445 // Now do post-visit work
3446 final_graph_reshaping_impl( n, frc );
3447 if (nstack.is_empty())
3448 break; // finished
3449 n = nstack.node(); // Get node from stack
3450 cnt = n->req();
3451 i = nstack.index();
3452 nstack.pop(); // Shift to the next node on stack
3453 }
3454 }
3455
3456 // Skip next transformation if compressed oops are not used.
3457 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3458 (!UseCompressedOops && !UseCompressedClassPointers))
3459 return;
3460
3461 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3462 // It could be done for an uncommon traps or any safepoints/calls
3463 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3464 while (sfpt.size() > 0) {
3465 n = sfpt.pop();
3466 JVMState *jvms = n->as_SafePoint()->jvms();
3467 assert(jvms != NULL, "sanity");
3468 int start = jvms->debug_start();
3469 int end = n->req();
3470 bool is_uncommon = (n->is_CallStaticJava() &&
3471 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3472 for (int j = start; j < end; j++) {
3473 Node* in = n->in(j);
3474 if (in->is_DecodeNarrowPtr()) {
3475 bool safe_to_skip = true;
3476 if (!is_uncommon ) {
3477 // Is it safe to skip?
3478 for (uint i = 0; i < in->outcnt(); i++) {
3479 Node* u = in->raw_out(i);
3480 if (!u->is_SafePoint() ||
3481 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3482 safe_to_skip = false;
3483 }
3484 }
3485 }
3486 if (safe_to_skip) {
3487 n->set_req(j, in->in(1));
3488 }
3489 if (in->outcnt() == 0) {
3490 in->disconnect_inputs(NULL, this);
3491 }
3492 }
3493 }
3494 }
3495 }
3496
3497 //------------------------------final_graph_reshaping--------------------------
3498 // Final Graph Reshaping.
3499 //
3500 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3501 // and not commoned up and forced early. Must come after regular
3502 // optimizations to avoid GVN undoing the cloning. Clone constant
3503 // inputs to Loop Phis; these will be split by the allocator anyways.
3504 // Remove Opaque nodes.
3505 // (2) Move last-uses by commutative operations to the left input to encourage
3506 // Intel update-in-place two-address operations and better register usage
3507 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3508 // calls canonicalizing them back.
3509 // (3) Count the number of double-precision FP ops, single-precision FP ops
3510 // and call sites. On Intel, we can get correct rounding either by
3511 // forcing singles to memory (requires extra stores and loads after each
3512 // FP bytecode) or we can set a rounding mode bit (requires setting and
3513 // clearing the mode bit around call sites). The mode bit is only used
3514 // if the relative frequency of single FP ops to calls is low enough.
3515 // This is a key transform for SPEC mpeg_audio.
3516 // (4) Detect infinite loops; blobs of code reachable from above but not
3517 // below. Several of the Code_Gen algorithms fail on such code shapes,
3518 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3519 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3520 // Detection is by looking for IfNodes where only 1 projection is
3521 // reachable from below or CatchNodes missing some targets.
3522 // (5) Assert for insane oop offsets in debug mode.
3523
3524 bool Compile::final_graph_reshaping() {
3525 // an infinite loop may have been eliminated by the optimizer,
3526 // in which case the graph will be empty.
3527 if (root()->req() == 1) {
3528 record_method_not_compilable("trivial infinite loop");
3529 return true;
3530 }
3531
3532 // Expensive nodes have their control input set to prevent the GVN
3533 // from freely commoning them. There's no GVN beyond this point so
3534 // no need to keep the control input. We want the expensive nodes to
3535 // be freely moved to the least frequent code path by gcm.
3536 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3537 for (int i = 0; i < expensive_count(); i++) {
3538 _expensive_nodes->at(i)->set_req(0, NULL);
3539 }
3540
3541 Final_Reshape_Counts frc;
3542
3543 // Visit everybody reachable!
3544 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3545 Node_Stack nstack(live_nodes() >> 1);
3546 final_graph_reshaping_walk(nstack, root(), frc);
3547
3548 // Check for unreachable (from below) code (i.e., infinite loops).
3549 for( uint i = 0; i < frc._tests.size(); i++ ) {
3550 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3551 // Get number of CFG targets.
3552 // Note that PCTables include exception targets after calls.
3553 uint required_outcnt = n->required_outcnt();
3554 if (n->outcnt() != required_outcnt) {
3555 // Check for a few special cases. Rethrow Nodes never take the
3556 // 'fall-thru' path, so expected kids is 1 less.
3557 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3558 if (n->in(0)->in(0)->is_Call()) {
3559 CallNode *call = n->in(0)->in(0)->as_Call();
3560 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3561 required_outcnt--; // Rethrow always has 1 less kid
3562 } else if (call->req() > TypeFunc::Parms &&
3563 call->is_CallDynamicJava()) {
3564 // Check for null receiver. In such case, the optimizer has
3565 // detected that the virtual call will always result in a null
3566 // pointer exception. The fall-through projection of this CatchNode
3567 // will not be populated.
3568 Node *arg0 = call->in(TypeFunc::Parms);
3569 if (arg0->is_Type() &&
3570 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3571 required_outcnt--;
3572 }
3573 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3574 call->req() > TypeFunc::Parms+1 &&
3575 call->is_CallStaticJava()) {
3576 // Check for negative array length. In such case, the optimizer has
3577 // detected that the allocation attempt will always result in an
3578 // exception. There is no fall-through projection of this CatchNode .
3579 Node *arg1 = call->in(TypeFunc::Parms+1);
3580 if (arg1->is_Type() &&
3581 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3582 required_outcnt--;
3583 }
3584 }
3585 }
3586 }
3587 // Recheck with a better notion of 'required_outcnt'
3588 if (n->outcnt() != required_outcnt) {
3589 record_method_not_compilable("malformed control flow");
3590 return true; // Not all targets reachable!
3591 }
3592 }
3593 // Check that I actually visited all kids. Unreached kids
3594 // must be infinite loops.
3595 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3596 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3597 record_method_not_compilable("infinite loop");
3598 return true; // Found unvisited kid; must be unreach
3599 }
3600 }
3601
3602 // If original bytecodes contained a mixture of floats and doubles
3603 // check if the optimizer has made it homogenous, item (3).
3604 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3605 frc.get_float_count() > 32 &&
3606 frc.get_double_count() == 0 &&
3607 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3608 set_24_bit_selection_and_mode( false, true );
3609 }
3610
3611 set_java_calls(frc.get_java_call_count());
3612 set_inner_loops(frc.get_inner_loop_count());
3613
3614 // No infinite loops, no reason to bail out.
3615 return false;
3616 }
3617
3618 //-----------------------------too_many_traps----------------------------------
3619 // Report if there are too many traps at the current method and bci.
3620 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3621 bool Compile::too_many_traps(ciMethod* method,
3622 int bci,
3623 Deoptimization::DeoptReason reason) {
3624 ciMethodData* md = method->method_data();
3625 if (md->is_empty()) {
3626 // Assume the trap has not occurred, or that it occurred only
3627 // because of a transient condition during start-up in the interpreter.
3628 return false;
3629 }
3630 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3631 if (md->has_trap_at(bci, m, reason) != 0) {
3632 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3633 // Also, if there are multiple reasons, or if there is no per-BCI record,
3634 // assume the worst.
3635 if (log())
3636 log()->elem("observe trap='%s' count='%d'",
3637 Deoptimization::trap_reason_name(reason),
3638 md->trap_count(reason));
3639 return true;
3640 } else {
3641 // Ignore method/bci and see if there have been too many globally.
3642 return too_many_traps(reason, md);
3643 }
3644 }
3645
3646 // Less-accurate variant which does not require a method and bci.
3647 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3648 ciMethodData* logmd) {
3649 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3650 // Too many traps globally.
3651 // Note that we use cumulative trap_count, not just md->trap_count.
3652 if (log()) {
3653 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3654 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3655 Deoptimization::trap_reason_name(reason),
3656 mcount, trap_count(reason));
3657 }
3658 return true;
3659 } else {
3660 // The coast is clear.
3661 return false;
3662 }
3663 }
3664
3665 //--------------------------too_many_recompiles--------------------------------
3666 // Report if there are too many recompiles at the current method and bci.
3667 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3668 // Is not eager to return true, since this will cause the compiler to use
3669 // Action_none for a trap point, to avoid too many recompilations.
3670 bool Compile::too_many_recompiles(ciMethod* method,
3671 int bci,
3672 Deoptimization::DeoptReason reason) {
3673 ciMethodData* md = method->method_data();
3674 if (md->is_empty()) {
3675 // Assume the trap has not occurred, or that it occurred only
3676 // because of a transient condition during start-up in the interpreter.
3677 return false;
3678 }
3679 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3680 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3681 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3682 Deoptimization::DeoptReason per_bc_reason
3683 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3684 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3685 if ((per_bc_reason == Deoptimization::Reason_none
3686 || md->has_trap_at(bci, m, reason) != 0)
3687 // The trap frequency measure we care about is the recompile count:
3688 && md->trap_recompiled_at(bci, m)
3689 && md->overflow_recompile_count() >= bc_cutoff) {
3690 // Do not emit a trap here if it has already caused recompilations.
3691 // Also, if there are multiple reasons, or if there is no per-BCI record,
3692 // assume the worst.
3693 if (log())
3694 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3695 Deoptimization::trap_reason_name(reason),
3696 md->trap_count(reason),
3697 md->overflow_recompile_count());
3698 return true;
3699 } else if (trap_count(reason) != 0
3700 && decompile_count() >= m_cutoff) {
3701 // Too many recompiles globally, and we have seen this sort of trap.
3702 // Use cumulative decompile_count, not just md->decompile_count.
3703 if (log())
3704 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3705 Deoptimization::trap_reason_name(reason),
3706 md->trap_count(reason), trap_count(reason),
3707 md->decompile_count(), decompile_count());
3708 return true;
3709 } else {
3710 // The coast is clear.
3711 return false;
3712 }
3713 }
3714
3715 // Compute when not to trap. Used by matching trap based nodes and
3716 // NullCheck optimization.
3717 void Compile::set_allowed_deopt_reasons() {
3718 _allowed_reasons = 0;
3719 if (is_method_compilation()) {
3720 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3721 assert(rs < BitsPerInt, "recode bit map");
3722 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3723 _allowed_reasons |= nth_bit(rs);
3724 }
3725 }
3726 }
3727 }
3728
3729 #ifndef PRODUCT
3730 //------------------------------verify_graph_edges---------------------------
3731 // Walk the Graph and verify that there is a one-to-one correspondence
3732 // between Use-Def edges and Def-Use edges in the graph.
3733 void Compile::verify_graph_edges(bool no_dead_code) {
3734 if (VerifyGraphEdges) {
3735 ResourceArea *area = Thread::current()->resource_area();
3736 Unique_Node_List visited(area);
3737 // Call recursive graph walk to check edges
3738 _root->verify_edges(visited);
3739 if (no_dead_code) {
3740 // Now make sure that no visited node is used by an unvisited node.
3741 bool dead_nodes = false;
3742 Unique_Node_List checked(area);
3743 while (visited.size() > 0) {
3744 Node* n = visited.pop();
3745 checked.push(n);
3746 for (uint i = 0; i < n->outcnt(); i++) {
3747 Node* use = n->raw_out(i);
3748 if (checked.member(use)) continue; // already checked
3749 if (visited.member(use)) continue; // already in the graph
3750 if (use->is_Con()) continue; // a dead ConNode is OK
3751 // At this point, we have found a dead node which is DU-reachable.
3752 if (!dead_nodes) {
3753 tty->print_cr("*** Dead nodes reachable via DU edges:");
3754 dead_nodes = true;
3755 }
3756 use->dump(2);
3757 tty->print_cr("---");
3758 checked.push(use); // No repeats; pretend it is now checked.
3759 }
3760 }
3761 assert(!dead_nodes, "using nodes must be reachable from root");
3762 }
3763 }
3764 }
3765
3766 // Verify GC barriers consistency
3767 // Currently supported:
3768 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3769 void Compile::verify_barriers() {
3770 if (UseG1GC || UseShenandoahGC) {
3771 // Verify G1 pre-barriers
3772 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3773
3774 ResourceArea *area = Thread::current()->resource_area();
3775 Unique_Node_List visited(area);
3776 Node_List worklist(area);
3777 // We're going to walk control flow backwards starting from the Root
3778 worklist.push(_root);
3779 while (worklist.size() > 0) {
3780 Node* x = worklist.pop();
3781 if (x == NULL || x == top()) continue;
3782 if (visited.member(x)) {
3783 continue;
3784 } else {
3785 visited.push(x);
3786 }
3787
3788 if (x->is_Region()) {
3789 for (uint i = 1; i < x->req(); i++) {
3790 worklist.push(x->in(i));
3791 }
3792 } else {
3793 worklist.push(x->in(0));
3794 // We are looking for the pattern:
3795 // /->ThreadLocal
3796 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3797 // \->ConI(0)
3798 // We want to verify that the If and the LoadB have the same control
3799 // See GraphKit::g1_write_barrier_pre()
3800 if (x->is_If()) {
3801 IfNode *iff = x->as_If();
3802 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3803 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3804 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3805 && cmp->in(1)->is_Load()) {
3806 LoadNode* load = cmp->in(1)->as_Load();
3807 if (load->is_g1_marking_load()) {
3808
3809 Node* if_ctrl = iff->in(0);
3810 Node* load_ctrl = load->in(0);
3811
3812 if (if_ctrl != load_ctrl) {
3813 // Skip possible CProj->NeverBranch in infinite loops
3814 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3815 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3816 if_ctrl = if_ctrl->in(0)->in(0);
3817 }
3818 }
3819 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3820 }
3821 }
3822 }
3823 }
3824 }
3825 }
3826 }
3827 }
3828
3829 #endif
3830
3831 // The Compile object keeps track of failure reasons separately from the ciEnv.
3832 // This is required because there is not quite a 1-1 relation between the
3833 // ciEnv and its compilation task and the Compile object. Note that one
3834 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3835 // to backtrack and retry without subsuming loads. Other than this backtracking
3836 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3837 // by the logic in C2Compiler.
3838 void Compile::record_failure(const char* reason) {
3839 if (log() != NULL) {
3840 log()->elem("failure reason='%s' phase='compile'", reason);
3841 }
3842 if (_failure_reason == NULL) {
3843 // Record the first failure reason.
3844 _failure_reason = reason;
3845 }
3846
3847 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3848 C->print_method(PHASE_FAILURE);
3849 }
3850 _root = NULL; // flush the graph, too
3851 }
3852
3853 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3854 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3855 _phase_name(name), _dolog(CITimeVerbose)
3856 {
3857 if (_dolog) {
3858 C = Compile::current();
3859 _log = C->log();
3860 } else {
3861 C = NULL;
3862 _log = NULL;
3863 }
3864 if (_log != NULL) {
3865 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3866 _log->stamp();
3867 _log->end_head();
3868 }
3869 }
3870
3871 Compile::TracePhase::~TracePhase() {
3872
3873 C = Compile::current();
3874 if (_dolog) {
3875 _log = C->log();
3876 } else {
3877 _log = NULL;
3878 }
3879
3880 #ifdef ASSERT
3881 if (PrintIdealNodeCount) {
3882 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3883 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3884 }
3885
3886 if (VerifyIdealNodeCount) {
3887 Compile::current()->print_missing_nodes();
3888 }
3889 #endif
3890
3891 if (_log != NULL) {
3892 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3893 }
3894 }
3895
3896 //=============================================================================
3897 // Two Constant's are equal when the type and the value are equal.
3898 bool Compile::Constant::operator==(const Constant& other) {
3899 if (type() != other.type() ) return false;
3900 if (can_be_reused() != other.can_be_reused()) return false;
3901 // For floating point values we compare the bit pattern.
3902 switch (type()) {
3903 case T_INT:
3904 case T_FLOAT: return (_v._value.i == other._v._value.i);
3905 case T_LONG:
3906 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3907 case T_OBJECT:
3908 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3909 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3910 case T_METADATA: return (_v._metadata == other._v._metadata);
3911 default: ShouldNotReachHere(); return false;
3912 }
3913 }
3914
3915 static int type_to_size_in_bytes(BasicType t) {
3916 switch (t) {
3917 case T_INT: return sizeof(jint );
3918 case T_LONG: return sizeof(jlong );
3919 case T_FLOAT: return sizeof(jfloat );
3920 case T_DOUBLE: return sizeof(jdouble);
3921 case T_METADATA: return sizeof(Metadata*);
3922 // We use T_VOID as marker for jump-table entries (labels) which
3923 // need an internal word relocation.
3924 case T_VOID:
3925 case T_ADDRESS:
3926 case T_OBJECT: return sizeof(jobject);
3927 default:
3928 ShouldNotReachHere();
3929 return -1;
3930 }
3931 }
3932
3933 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3934 // sort descending
3935 if (a->freq() > b->freq()) return -1;
3936 if (a->freq() < b->freq()) return 1;
3937 return 0;
3938 }
3939
3940 void Compile::ConstantTable::calculate_offsets_and_size() {
3941 // First, sort the array by frequencies.
3942 _constants.sort(qsort_comparator);
3943
3944 #ifdef ASSERT
3945 // Make sure all jump-table entries were sorted to the end of the
3946 // array (they have a negative frequency).
3947 bool found_void = false;
3948 for (int i = 0; i < _constants.length(); i++) {
3949 Constant con = _constants.at(i);
3950 if (con.type() == T_VOID)
3951 found_void = true; // jump-tables
3952 else
3953 assert(!found_void, "wrong sorting");
3954 }
3955 #endif
3956
3957 int offset = 0;
3958 for (int i = 0; i < _constants.length(); i++) {
3959 Constant* con = _constants.adr_at(i);
3960
3961 // Align offset for type.
3962 int typesize = type_to_size_in_bytes(con->type());
3963 offset = align_up(offset, typesize);
3964 con->set_offset(offset); // set constant's offset
3965
3966 if (con->type() == T_VOID) {
3967 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3968 offset = offset + typesize * n->outcnt(); // expand jump-table
3969 } else {
3970 offset = offset + typesize;
3971 }
3972 }
3973
3974 // Align size up to the next section start (which is insts; see
3975 // CodeBuffer::align_at_start).
3976 assert(_size == -1, "already set?");
3977 _size = align_up(offset, (int)CodeEntryAlignment);
3978 }
3979
3980 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3981 MacroAssembler _masm(&cb);
3982 for (int i = 0; i < _constants.length(); i++) {
3983 Constant con = _constants.at(i);
3984 address constant_addr = NULL;
3985 switch (con.type()) {
3986 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break;
3987 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3988 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3989 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3990 case T_OBJECT: {
3991 jobject obj = con.get_jobject();
3992 int oop_index = _masm.oop_recorder()->find_index(obj);
3993 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3994 break;
3995 }
3996 case T_ADDRESS: {
3997 address addr = (address) con.get_jobject();
3998 constant_addr = _masm.address_constant(addr);
3999 break;
4000 }
4001 // We use T_VOID as marker for jump-table entries (labels) which
4002 // need an internal word relocation.
4003 case T_VOID: {
4004 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4005 // Fill the jump-table with a dummy word. The real value is
4006 // filled in later in fill_jump_table.
4007 address dummy = (address) n;
4008 constant_addr = _masm.address_constant(dummy);
4009 // Expand jump-table
4010 for (uint i = 1; i < n->outcnt(); i++) {
4011 address temp_addr = _masm.address_constant(dummy + i);
4012 assert(temp_addr, "consts section too small");
4013 }
4014 break;
4015 }
4016 case T_METADATA: {
4017 Metadata* obj = con.get_metadata();
4018 int metadata_index = _masm.oop_recorder()->find_index(obj);
4019 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4020 break;
4021 }
4022 default: ShouldNotReachHere();
4023 }
4024 assert(constant_addr, "consts section too small");
4025 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4026 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4027 }
4028 }
4029
4030 int Compile::ConstantTable::find_offset(Constant& con) const {
4031 int idx = _constants.find(con);
4032 assert(idx != -1, "constant must be in constant table");
4033 int offset = _constants.at(idx).offset();
4034 assert(offset != -1, "constant table not emitted yet?");
4035 return offset;
4036 }
4037
4038 void Compile::ConstantTable::add(Constant& con) {
4039 if (con.can_be_reused()) {
4040 int idx = _constants.find(con);
4041 if (idx != -1 && _constants.at(idx).can_be_reused()) {
4042 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
4043 return;
4044 }
4045 }
4046 (void) _constants.append(con);
4047 }
4048
4049 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4050 Block* b = Compile::current()->cfg()->get_block_for_node(n);
4051 Constant con(type, value, b->_freq);
4052 add(con);
4053 return con;
4054 }
4055
4056 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4057 Constant con(metadata);
4058 add(con);
4059 return con;
4060 }
4061
4062 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4063 jvalue value;
4064 BasicType type = oper->type()->basic_type();
4065 switch (type) {
4066 case T_LONG: value.j = oper->constantL(); break;
4067 case T_FLOAT: value.f = oper->constantF(); break;
4068 case T_DOUBLE: value.d = oper->constantD(); break;
4069 case T_OBJECT:
4070 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4071 case T_METADATA: return add((Metadata*)oper->constant()); break;
4072 default: guarantee(false, "unhandled type: %s", type2name(type));
4073 }
4074 return add(n, type, value);
4075 }
4076
4077 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4078 jvalue value;
4079 // We can use the node pointer here to identify the right jump-table
4080 // as this method is called from Compile::Fill_buffer right before
4081 // the MachNodes are emitted and the jump-table is filled (means the
4082 // MachNode pointers do not change anymore).
4083 value.l = (jobject) n;
4084 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
4085 add(con);
4086 return con;
4087 }
4088
4089 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4090 // If called from Compile::scratch_emit_size do nothing.
4091 if (Compile::current()->in_scratch_emit_size()) return;
4092
4093 assert(labels.is_nonempty(), "must be");
4094 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4095
4096 // Since MachConstantNode::constant_offset() also contains
4097 // table_base_offset() we need to subtract the table_base_offset()
4098 // to get the plain offset into the constant table.
4099 int offset = n->constant_offset() - table_base_offset();
4100
4101 MacroAssembler _masm(&cb);
4102 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4103
4104 for (uint i = 0; i < n->outcnt(); i++) {
4105 address* constant_addr = &jump_table_base[i];
4106 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));
4107 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4108 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4109 }
4110 }
4111
4112 //----------------------------static_subtype_check-----------------------------
4113 // Shortcut important common cases when superklass is exact:
4114 // (0) superklass is java.lang.Object (can occur in reflective code)
4115 // (1) subklass is already limited to a subtype of superklass => always ok
4116 // (2) subklass does not overlap with superklass => always fail
4117 // (3) superklass has NO subtypes and we can check with a simple compare.
4118 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4119 if (StressReflectiveCode) {
4120 return SSC_full_test; // Let caller generate the general case.
4121 }
4122
4123 if (superk == env()->Object_klass()) {
4124 return SSC_always_true; // (0) this test cannot fail
4125 }
4126
4127 ciType* superelem = superk;
4128 if (superelem->is_array_klass())
4129 superelem = superelem->as_array_klass()->base_element_type();
4130
4131 if (!subk->is_interface()) { // cannot trust static interface types yet
4132 if (subk->is_subtype_of(superk)) {
4133 return SSC_always_true; // (1) false path dead; no dynamic test needed
4134 }
4135 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4136 !superk->is_subtype_of(subk)) {
4137 return SSC_always_false;
4138 }
4139 }
4140
4141 // If casting to an instance klass, it must have no subtypes
4142 if (superk->is_interface()) {
4143 // Cannot trust interfaces yet.
4144 // %%% S.B. superk->nof_implementors() == 1
4145 } else if (superelem->is_instance_klass()) {
4146 ciInstanceKlass* ik = superelem->as_instance_klass();
4147 if (!ik->has_subklass() && !ik->is_interface()) {
4148 if (!ik->is_final()) {
4149 // Add a dependency if there is a chance of a later subclass.
4150 dependencies()->assert_leaf_type(ik);
4151 }
4152 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4153 }
4154 } else {
4155 // A primitive array type has no subtypes.
4156 return SSC_easy_test; // (3) caller can do a simple ptr comparison
4157 }
4158
4159 return SSC_full_test;
4160 }
4161
4162 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4163 #ifdef _LP64
4164 // The scaled index operand to AddP must be a clean 64-bit value.
4165 // Java allows a 32-bit int to be incremented to a negative
4166 // value, which appears in a 64-bit register as a large
4167 // positive number. Using that large positive number as an
4168 // operand in pointer arithmetic has bad consequences.
4169 // On the other hand, 32-bit overflow is rare, and the possibility
4170 // can often be excluded, if we annotate the ConvI2L node with
4171 // a type assertion that its value is known to be a small positive
4172 // number. (The prior range check has ensured this.)
4173 // This assertion is used by ConvI2LNode::Ideal.
4174 int index_max = max_jint - 1; // array size is max_jint, index is one less
4175 if (sizetype != NULL) index_max = sizetype->_hi - 1;
4176 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4177 idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4178 #endif
4179 return idx;
4180 }
4181
4182 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4183 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4184 if (ctrl != NULL) {
4185 // Express control dependency by a CastII node with a narrow type.
4186 value = new CastIINode(value, itype, false, true /* range check dependency */);
4187 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4188 // node from floating above the range check during loop optimizations. Otherwise, the
4189 // ConvI2L node may be eliminated independently of the range check, causing the data path
4190 // to become TOP while the control path is still there (although it's unreachable).
4191 value->set_req(0, ctrl);
4192 // Save CastII node to remove it after loop optimizations.
4193 phase->C->add_range_check_cast(value);
4194 value = phase->transform(value);
4195 }
4196 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4197 return phase->transform(new ConvI2LNode(value, ltype));
4198 }
4199
4200 // The message about the current inlining is accumulated in
4201 // _print_inlining_stream and transfered into the _print_inlining_list
4202 // once we know whether inlining succeeds or not. For regular
4203 // inlining, messages are appended to the buffer pointed by
4204 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4205 // a new buffer is added after _print_inlining_idx in the list. This
4206 // way we can update the inlining message for late inlining call site
4207 // when the inlining is attempted again.
4208 void Compile::print_inlining_init() {
4209 if (print_inlining() || print_intrinsics()) {
4210 _print_inlining_stream = new stringStream();
4211 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4212 }
4213 }
4214
4215 void Compile::print_inlining_reinit() {
4216 if (print_inlining() || print_intrinsics()) {
4217 // Re allocate buffer when we change ResourceMark
4218 _print_inlining_stream = new stringStream();
4219 }
4220 }
4221
4222 void Compile::print_inlining_reset() {
4223 _print_inlining_stream->reset();
4224 }
4225
4226 void Compile::print_inlining_commit() {
4227 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4228 // Transfer the message from _print_inlining_stream to the current
4229 // _print_inlining_list buffer and clear _print_inlining_stream.
4230 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4231 print_inlining_reset();
4232 }
4233
4234 void Compile::print_inlining_push() {
4235 // Add new buffer to the _print_inlining_list at current position
4236 _print_inlining_idx++;
4237 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4238 }
4239
4240 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4241 return _print_inlining_list->at(_print_inlining_idx);
4242 }
4243
4244 void Compile::print_inlining_update(CallGenerator* cg) {
4245 if (print_inlining() || print_intrinsics()) {
4246 if (!cg->is_late_inline()) {
4247 if (print_inlining_current().cg() != NULL) {
4248 print_inlining_push();
4249 }
4250 print_inlining_commit();
4251 } else {
4252 if (print_inlining_current().cg() != cg &&
4253 (print_inlining_current().cg() != NULL ||
4254 print_inlining_current().ss()->size() != 0)) {
4255 print_inlining_push();
4256 }
4257 print_inlining_commit();
4258 print_inlining_current().set_cg(cg);
4259 }
4260 }
4261 }
4262
4263 void Compile::print_inlining_move_to(CallGenerator* cg) {
4264 // We resume inlining at a late inlining call site. Locate the
4265 // corresponding inlining buffer so that we can update it.
4266 if (print_inlining()) {
4267 for (int i = 0; i < _print_inlining_list->length(); i++) {
4268 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4269 _print_inlining_idx = i;
4270 return;
4271 }
4272 }
4273 ShouldNotReachHere();
4274 }
4275 }
4276
4277 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4278 if (print_inlining()) {
4279 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4280 assert(print_inlining_current().cg() == cg, "wrong entry");
4281 // replace message with new message
4282 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4283 print_inlining_commit();
4284 print_inlining_current().set_cg(cg);
4285 }
4286 }
4287
4288 void Compile::print_inlining_assert_ready() {
4289 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4290 }
4291
4292 void Compile::process_print_inlining() {
4293 bool do_print_inlining = print_inlining() || print_intrinsics();
4294 if (do_print_inlining || log() != NULL) {
4295 // Print inlining message for candidates that we couldn't inline
4296 // for lack of space
4297 for (int i = 0; i < _late_inlines.length(); i++) {
4298 CallGenerator* cg = _late_inlines.at(i);
4299 if (!cg->is_mh_late_inline()) {
4300 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4301 if (do_print_inlining) {
4302 cg->print_inlining_late(msg);
4303 }
4304 log_late_inline_failure(cg, msg);
4305 }
4306 }
4307 }
4308 if (do_print_inlining) {
4309 ResourceMark rm;
4310 stringStream ss;
4311 for (int i = 0; i < _print_inlining_list->length(); i++) {
4312 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4313 }
4314 size_t end = ss.size();
4315 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4316 strncpy(_print_inlining_output, ss.base(), end+1);
4317 _print_inlining_output[end] = 0;
4318 }
4319 }
4320
4321 void Compile::dump_print_inlining() {
4322 if (_print_inlining_output != NULL) {
4323 tty->print_raw(_print_inlining_output);
4324 }
4325 }
4326
4327 void Compile::log_late_inline(CallGenerator* cg) {
4328 if (log() != NULL) {
4329 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4330 cg->unique_id());
4331 JVMState* p = cg->call_node()->jvms();
4332 while (p != NULL) {
4333 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4334 p = p->caller();
4335 }
4336 log()->tail("late_inline");
4337 }
4338 }
4339
4340 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4341 log_late_inline(cg);
4342 if (log() != NULL) {
4343 log()->inline_fail(msg);
4344 }
4345 }
4346
4347 void Compile::log_inline_id(CallGenerator* cg) {
4348 if (log() != NULL) {
4349 // The LogCompilation tool needs a unique way to identify late
4350 // inline call sites. This id must be unique for this call site in
4351 // this compilation. Try to have it unique across compilations as
4352 // well because it can be convenient when grepping through the log
4353 // file.
4354 // Distinguish OSR compilations from others in case CICountOSR is
4355 // on.
4356 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4357 cg->set_unique_id(id);
4358 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4359 }
4360 }
4361
4362 void Compile::log_inline_failure(const char* msg) {
4363 if (C->log() != NULL) {
4364 C->log()->inline_fail(msg);
4365 }
4366 }
4367
4368
4369 // Dump inlining replay data to the stream.
4370 // Don't change thread state and acquire any locks.
4371 void Compile::dump_inline_data(outputStream* out) {
4372 InlineTree* inl_tree = ilt();
4373 if (inl_tree != NULL) {
4374 out->print(" inline %d", inl_tree->count());
4375 inl_tree->dump_replay_data(out);
4376 }
4377 }
4378
4379 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4380 if (n1->Opcode() < n2->Opcode()) return -1;
4381 else if (n1->Opcode() > n2->Opcode()) return 1;
4382
4383 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4384 for (uint i = 1; i < n1->req(); i++) {
4385 if (n1->in(i) < n2->in(i)) return -1;
4386 else if (n1->in(i) > n2->in(i)) return 1;
4387 }
4388
4389 return 0;
4390 }
4391
4392 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4393 Node* n1 = *n1p;
4394 Node* n2 = *n2p;
4395
4396 return cmp_expensive_nodes(n1, n2);
4397 }
4398
4399 void Compile::sort_expensive_nodes() {
4400 if (!expensive_nodes_sorted()) {
4401 _expensive_nodes->sort(cmp_expensive_nodes);
4402 }
4403 }
4404
4405 bool Compile::expensive_nodes_sorted() const {
4406 for (int i = 1; i < _expensive_nodes->length(); i++) {
4407 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4408 return false;
4409 }
4410 }
4411 return true;
4412 }
4413
4414 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4415 if (_expensive_nodes->length() == 0) {
4416 return false;
4417 }
4418
4419 assert(OptimizeExpensiveOps, "optimization off?");
4420
4421 // Take this opportunity to remove dead nodes from the list
4422 int j = 0;
4423 for (int i = 0; i < _expensive_nodes->length(); i++) {
4424 Node* n = _expensive_nodes->at(i);
4425 if (!n->is_unreachable(igvn)) {
4426 assert(n->is_expensive(), "should be expensive");
4427 _expensive_nodes->at_put(j, n);
4428 j++;
4429 }
4430 }
4431 _expensive_nodes->trunc_to(j);
4432
4433 // Then sort the list so that similar nodes are next to each other
4434 // and check for at least two nodes of identical kind with same data
4435 // inputs.
4436 sort_expensive_nodes();
4437
4438 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4439 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4440 return true;
4441 }
4442 }
4443
4444 return false;
4445 }
4446
4447 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4448 if (_expensive_nodes->length() == 0) {
4449 return;
4450 }
4451
4452 assert(OptimizeExpensiveOps, "optimization off?");
4453
4454 // Sort to bring similar nodes next to each other and clear the
4455 // control input of nodes for which there's only a single copy.
4456 sort_expensive_nodes();
4457
4458 int j = 0;
4459 int identical = 0;
4460 int i = 0;
4461 bool modified = false;
4462 for (; i < _expensive_nodes->length()-1; i++) {
4463 assert(j <= i, "can't write beyond current index");
4464 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4465 identical++;
4466 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4467 continue;
4468 }
4469 if (identical > 0) {
4470 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4471 identical = 0;
4472 } else {
4473 Node* n = _expensive_nodes->at(i);
4474 igvn.replace_input_of(n, 0, NULL);
4475 igvn.hash_insert(n);
4476 modified = true;
4477 }
4478 }
4479 if (identical > 0) {
4480 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4481 } else if (_expensive_nodes->length() >= 1) {
4482 Node* n = _expensive_nodes->at(i);
4483 igvn.replace_input_of(n, 0, NULL);
4484 igvn.hash_insert(n);
4485 modified = true;
4486 }
4487 _expensive_nodes->trunc_to(j);
4488 if (modified) {
4489 igvn.optimize();
4490 }
4491 }
4492
4493 void Compile::add_expensive_node(Node * n) {
4494 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4495 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4496 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4497 if (OptimizeExpensiveOps) {
4498 _expensive_nodes->append(n);
4499 } else {
4500 // Clear control input and let IGVN optimize expensive nodes if
4501 // OptimizeExpensiveOps is off.
4502 n->set_req(0, NULL);
4503 }
4504 }
4505
4506 /**
4507 * Remove the speculative part of types and clean up the graph
4508 */
4509 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4510 if (UseTypeSpeculation) {
4511 Unique_Node_List worklist;
4512 worklist.push(root());
4513 int modified = 0;
4514 // Go over all type nodes that carry a speculative type, drop the
4515 // speculative part of the type and enqueue the node for an igvn
4516 // which may optimize it out.
4517 for (uint next = 0; next < worklist.size(); ++next) {
4518 Node *n = worklist.at(next);
4519 if (n->is_Type()) {
4520 TypeNode* tn = n->as_Type();
4521 const Type* t = tn->type();
4522 const Type* t_no_spec = t->remove_speculative();
4523 if (t_no_spec != t) {
4524 bool in_hash = igvn.hash_delete(n);
4525 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4526 tn->set_type(t_no_spec);
4527 igvn.hash_insert(n);
4528 igvn._worklist.push(n); // give it a chance to go away
4529 modified++;
4530 }
4531 }
4532 uint max = n->len();
4533 for( uint i = 0; i < max; ++i ) {
4534 Node *m = n->in(i);
4535 if (not_a_node(m)) continue;
4536 worklist.push(m);
4537 }
4538 }
4539 // Drop the speculative part of all types in the igvn's type table
4540 igvn.remove_speculative_types();
4541 if (modified > 0) {
4542 igvn.optimize();
4543 }
4544 #ifdef ASSERT
4545 // Verify that after the IGVN is over no speculative type has resurfaced
4546 worklist.clear();
4547 worklist.push(root());
4548 for (uint next = 0; next < worklist.size(); ++next) {
4549 Node *n = worklist.at(next);
4550 const Type* t = igvn.type_or_null(n);
4551 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4552 if (n->is_Type()) {
4553 t = n->as_Type()->type();
4554 assert(t == t->remove_speculative(), "no more speculative types");
4555 }
4556 uint max = n->len();
4557 for( uint i = 0; i < max; ++i ) {
4558 Node *m = n->in(i);
4559 if (not_a_node(m)) continue;
4560 worklist.push(m);
4561 }
4562 }
4563 igvn.check_no_speculative_types();
4564 #endif
4565 }
4566 }
4567
4568 // Auxiliary method to support randomized stressing/fuzzing.
4569 //
4570 // This method can be called the arbitrary number of times, with current count
4571 // as the argument. The logic allows selecting a single candidate from the
4572 // running list of candidates as follows:
4573 // int count = 0;
4574 // Cand* selected = null;
4575 // while(cand = cand->next()) {
4576 // if (randomized_select(++count)) {
4577 // selected = cand;
4578 // }
4579 // }
4580 //
4581 // Including count equalizes the chances any candidate is "selected".
4582 // This is useful when we don't have the complete list of candidates to choose
4583 // from uniformly. In this case, we need to adjust the randomicity of the
4584 // selection, or else we will end up biasing the selection towards the latter
4585 // candidates.
4586 //
4587 // Quick back-envelope calculation shows that for the list of n candidates
4588 // the equal probability for the candidate to persist as "best" can be
4589 // achieved by replacing it with "next" k-th candidate with the probability
4590 // of 1/k. It can be easily shown that by the end of the run, the
4591 // probability for any candidate is converged to 1/n, thus giving the
4592 // uniform distribution among all the candidates.
4593 //
4594 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4595 #define RANDOMIZED_DOMAIN_POW 29
4596 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4597 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4598 bool Compile::randomized_select(int count) {
4599 assert(count > 0, "only positive");
4600 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4601 }
4602
4603 CloneMap& Compile::clone_map() { return _clone_map; }
4604 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4605
4606 void NodeCloneInfo::dump() const {
4607 tty->print(" {%d:%d} ", idx(), gen());
4608 }
4609
4610 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4611 uint64_t val = value(old->_idx);
4612 NodeCloneInfo cio(val);
4613 assert(val != 0, "old node should be in the map");
4614 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4615 insert(nnn->_idx, cin.get());
4616 #ifndef PRODUCT
4617 if (is_debug()) {
4618 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4619 }
4620 #endif
4621 }
4622
4623 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4624 NodeCloneInfo cio(value(old->_idx));
4625 if (cio.get() == 0) {
4626 cio.set(old->_idx, 0);
4627 insert(old->_idx, cio.get());
4628 #ifndef PRODUCT
4629 if (is_debug()) {
4630 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4631 }
4632 #endif
4633 }
4634 clone(old, nnn, gen);
4635 }
4636
4637 int CloneMap::max_gen() const {
4638 int g = 0;
4639 DictI di(_dict);
4640 for(; di.test(); ++di) {
4641 int t = gen(di._key);
4642 if (g < t) {
4643 g = t;
4644 #ifndef PRODUCT
4645 if (is_debug()) {
4646 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4647 }
4648 #endif
4649 }
4650 }
4651 return g;
4652 }
4653
4654 void CloneMap::dump(node_idx_t key) const {
4655 uint64_t val = value(key);
4656 if (val != 0) {
4657 NodeCloneInfo ni(val);
4658 ni.dump();
4659 }
4660 }
4661
4662 void Compile::shenandoah_eliminate_matrix_update(Node* p2x, PhaseIterGVN* igvn) {
4663 assert(UseShenandoahGC && p2x->Opcode() == Op_CastP2X, "");
4664 ResourceMark rm;
4665 Unique_Node_List wq;
4666
4667 wq.push(p2x);
4668 for (uint next = 0; next < wq.size(); next++) {
4669 Node *n = wq.at(next);
4670 if (n->is_Store()) {
4671 // do nothing
4672 } else if (n->is_Load()) {
4673 igvn->replace_node(n, igvn->intcon(1));
4674 } else {
4675 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4676 Node* u = n->fast_out(i);
4677 wq.push(u);
4678 }
4679 }
4680 }
4681 igvn->replace_node(p2x, C->top());
4682 }
4683
4684 void Compile::shenandoah_eliminate_g1_wb_pre(Node* call, PhaseIterGVN* igvn) {
4685 assert(UseShenandoahGC && call->is_g1_wb_pre_call(), "");
4686 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
4687 c = c->unique_ctrl_out();
4688 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4689 c = c->unique_ctrl_out();
4690 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4691 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4692 assert(iff->is_If(), "expect test");
4693 if (!iff->is_shenandoah_marking_if(igvn)) {
4694 c = c->unique_ctrl_out();
4695 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4696 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4697 assert(iff->is_shenandoah_marking_if(igvn), "expect marking test");
4698 }
4699 Node* cmpx = iff->in(1)->in(1);
4700 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
4701 igvn->rehash_node_delayed(call);
4702 call->del_req(call->req()-1);
4703 }