1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "jfr/jfrEvents.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/block.hpp"
38 #include "opto/c2compiler.hpp"
39 #include "opto/callGenerator.hpp"
40 #include "opto/callnode.hpp"
41 #include "opto/cfgnode.hpp"
42 #include "opto/chaitin.hpp"
43 #include "opto/compile.hpp"
44 #include "opto/connode.hpp"
45 #include "opto/divnode.hpp"
46 #include "opto/escape.hpp"
47 #include "opto/idealGraphPrinter.hpp"
48 #include "opto/loopnode.hpp"
49 #include "opto/machnode.hpp"
50 #include "opto/macro.hpp"
51 #include "opto/matcher.hpp"
52 #include "opto/mathexactnode.hpp"
53 #include "opto/memnode.hpp"
54 #include "opto/mulnode.hpp"
55 #include "opto/node.hpp"
56 #include "opto/opcodes.hpp"
57 #include "opto/output.hpp"
58 #include "opto/parse.hpp"
59 #include "opto/phaseX.hpp"
60 #include "opto/rootnode.hpp"
61 #include "opto/runtime.hpp"
62 #include "opto/stringopts.hpp"
63 #include "opto/type.hpp"
64 #include "opto/vectornode.hpp"
65 #include "runtime/arguments.hpp"
66 #include "runtime/signature.hpp"
67 #include "runtime/stubRoutines.hpp"
68 #include "runtime/timer.hpp"
69 #include "utilities/copy.hpp"
70 #if defined AD_MD_HPP
71 # include AD_MD_HPP
72 #elif defined TARGET_ARCH_MODEL_x86_32
73 # include "adfiles/ad_x86_32.hpp"
74 #elif defined TARGET_ARCH_MODEL_x86_64
75 # include "adfiles/ad_x86_64.hpp"
76 #elif defined TARGET_ARCH_MODEL_aarch64
77 # include "adfiles/ad_aarch64.hpp"
78 #elif defined TARGET_ARCH_MODEL_sparc
79 # include "adfiles/ad_sparc.hpp"
80 #elif defined TARGET_ARCH_MODEL_zero
81 # include "adfiles/ad_zero.hpp"
82 #elif defined TARGET_ARCH_MODEL_ppc_64
83 # include "adfiles/ad_ppc_64.hpp"
84 #endif
85
86 #if INCLUDE_ALL_GCS
87 #include "gc_implementation/shenandoah/shenandoahForwarding.hpp"
88 #include "gc_implementation/shenandoah/c2/shenandoahSupport.hpp"
89 #endif
90
91 // -------------------- Compile::mach_constant_base_node -----------------------
92 // Constant table base node singleton.
93 MachConstantBaseNode* Compile::mach_constant_base_node() {
94 if (_mach_constant_base_node == NULL) {
95 _mach_constant_base_node = new (C) MachConstantBaseNode();
96 _mach_constant_base_node->add_req(C->root());
97 }
98 return _mach_constant_base_node;
99 }
100
101
102 /// Support for intrinsics.
103
104 // Return the index at which m must be inserted (or already exists).
105 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
106 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
107 #ifdef ASSERT
108 for (int i = 1; i < _intrinsics->length(); i++) {
109 CallGenerator* cg1 = _intrinsics->at(i-1);
110 CallGenerator* cg2 = _intrinsics->at(i);
111 assert(cg1->method() != cg2->method()
112 ? cg1->method() < cg2->method()
113 : cg1->is_virtual() < cg2->is_virtual(),
114 "compiler intrinsics list must stay sorted");
115 }
116 #endif
117 // Binary search sorted list, in decreasing intervals [lo, hi].
118 int lo = 0, hi = _intrinsics->length()-1;
119 while (lo <= hi) {
120 int mid = (uint)(hi + lo) / 2;
121 ciMethod* mid_m = _intrinsics->at(mid)->method();
122 if (m < mid_m) {
123 hi = mid-1;
124 } else if (m > mid_m) {
125 lo = mid+1;
126 } else {
127 // look at minor sort key
128 bool mid_virt = _intrinsics->at(mid)->is_virtual();
129 if (is_virtual < mid_virt) {
130 hi = mid-1;
131 } else if (is_virtual > mid_virt) {
132 lo = mid+1;
133 } else {
134 return mid; // exact match
135 }
136 }
137 }
138 return lo; // inexact match
139 }
140
141 void Compile::register_intrinsic(CallGenerator* cg) {
142 if (_intrinsics == NULL) {
143 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
144 }
145 // This code is stolen from ciObjectFactory::insert.
146 // Really, GrowableArray should have methods for
147 // insert_at, remove_at, and binary_search.
148 int len = _intrinsics->length();
149 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
150 if (index == len) {
151 _intrinsics->append(cg);
152 } else {
153 #ifdef ASSERT
154 CallGenerator* oldcg = _intrinsics->at(index);
155 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
156 #endif
157 _intrinsics->append(_intrinsics->at(len-1));
158 int pos;
159 for (pos = len-2; pos >= index; pos--) {
160 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
161 }
162 _intrinsics->at_put(index, cg);
163 }
164 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
165 }
166
167 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
168 assert(m->is_loaded(), "don't try this on unloaded methods");
169 if (_intrinsics != NULL) {
170 int index = intrinsic_insertion_index(m, is_virtual);
171 if (index < _intrinsics->length()
172 && _intrinsics->at(index)->method() == m
173 && _intrinsics->at(index)->is_virtual() == is_virtual) {
174 return _intrinsics->at(index);
175 }
176 }
177 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
178 if (m->intrinsic_id() != vmIntrinsics::_none &&
179 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
180 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
181 if (cg != NULL) {
182 // Save it for next time:
183 register_intrinsic(cg);
184 return cg;
185 } else {
186 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
187 }
188 }
189 return NULL;
190 }
191
192 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
193 // in library_call.cpp.
194
195
196 #ifndef PRODUCT
197 // statistics gathering...
198
199 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
200 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
201
202 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
203 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
204 int oflags = _intrinsic_hist_flags[id];
205 assert(flags != 0, "what happened?");
206 if (is_virtual) {
207 flags |= _intrinsic_virtual;
208 }
209 bool changed = (flags != oflags);
210 if ((flags & _intrinsic_worked) != 0) {
211 juint count = (_intrinsic_hist_count[id] += 1);
212 if (count == 1) {
213 changed = true; // first time
214 }
215 // increment the overall count also:
216 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
217 }
218 if (changed) {
219 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
220 // Something changed about the intrinsic's virtuality.
221 if ((flags & _intrinsic_virtual) != 0) {
222 // This is the first use of this intrinsic as a virtual call.
223 if (oflags != 0) {
224 // We already saw it as a non-virtual, so note both cases.
225 flags |= _intrinsic_both;
226 }
227 } else if ((oflags & _intrinsic_both) == 0) {
228 // This is the first use of this intrinsic as a non-virtual
229 flags |= _intrinsic_both;
230 }
231 }
232 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
233 }
234 // update the overall flags also:
235 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
236 return changed;
237 }
238
239 static char* format_flags(int flags, char* buf) {
240 buf[0] = 0;
241 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
242 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
243 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
244 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
245 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
246 if (buf[0] == 0) strcat(buf, ",");
247 assert(buf[0] == ',', "must be");
248 return &buf[1];
249 }
250
251 void Compile::print_intrinsic_statistics() {
252 char flagsbuf[100];
253 ttyLocker ttyl;
254 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
255 tty->print_cr("Compiler intrinsic usage:");
256 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
257 if (total == 0) total = 1; // avoid div0 in case of no successes
258 #define PRINT_STAT_LINE(name, c, f) \
259 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
260 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
261 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
262 int flags = _intrinsic_hist_flags[id];
263 juint count = _intrinsic_hist_count[id];
264 if ((flags | count) != 0) {
265 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
266 }
267 }
268 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
269 if (xtty != NULL) xtty->tail("statistics");
270 }
271
272 void Compile::print_statistics() {
273 { ttyLocker ttyl;
274 if (xtty != NULL) xtty->head("statistics type='opto'");
275 Parse::print_statistics();
276 PhaseCCP::print_statistics();
277 PhaseRegAlloc::print_statistics();
278 Scheduling::print_statistics();
279 PhasePeephole::print_statistics();
280 PhaseIdealLoop::print_statistics();
281 if (xtty != NULL) xtty->tail("statistics");
282 }
283 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
284 // put this under its own <statistics> element.
285 print_intrinsic_statistics();
286 }
287 }
288 #endif //PRODUCT
289
290 // Support for bundling info
291 Bundle* Compile::node_bundling(const Node *n) {
292 assert(valid_bundle_info(n), "oob");
293 return &_node_bundling_base[n->_idx];
294 }
295
296 bool Compile::valid_bundle_info(const Node *n) {
297 return (_node_bundling_limit > n->_idx);
298 }
299
300
301 void Compile::gvn_replace_by(Node* n, Node* nn) {
302 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
303 Node* use = n->last_out(i);
304 bool is_in_table = initial_gvn()->hash_delete(use);
305 uint uses_found = 0;
306 for (uint j = 0; j < use->len(); j++) {
307 if (use->in(j) == n) {
308 if (j < use->req())
309 use->set_req(j, nn);
310 else
311 use->set_prec(j, nn);
312 uses_found++;
313 }
314 }
315 if (is_in_table) {
316 // reinsert into table
317 initial_gvn()->hash_find_insert(use);
318 }
319 record_for_igvn(use);
320 i -= uses_found; // we deleted 1 or more copies of this edge
321 }
322 }
323
324
325 static inline bool not_a_node(const Node* n) {
326 if (n == NULL) return true;
327 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
328 if (*(address*)n == badAddress) return true; // kill by Node::destruct
329 return false;
330 }
331
332 // Identify all nodes that are reachable from below, useful.
333 // Use breadth-first pass that records state in a Unique_Node_List,
334 // recursive traversal is slower.
335 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
336 int estimated_worklist_size = live_nodes();
337 useful.map( estimated_worklist_size, NULL ); // preallocate space
338
339 // Initialize worklist
340 if (root() != NULL) { useful.push(root()); }
341 // If 'top' is cached, declare it useful to preserve cached node
342 if( cached_top_node() ) { useful.push(cached_top_node()); }
343
344 // Push all useful nodes onto the list, breadthfirst
345 for( uint next = 0; next < useful.size(); ++next ) {
346 assert( next < unique(), "Unique useful nodes < total nodes");
347 Node *n = useful.at(next);
348 uint max = n->len();
349 for( uint i = 0; i < max; ++i ) {
350 Node *m = n->in(i);
351 if (not_a_node(m)) continue;
352 useful.push(m);
353 }
354 }
355 }
356
357 // Update dead_node_list with any missing dead nodes using useful
358 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
359 void Compile::update_dead_node_list(Unique_Node_List &useful) {
360 uint max_idx = unique();
361 VectorSet& useful_node_set = useful.member_set();
362
363 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
364 // If node with index node_idx is not in useful set,
365 // mark it as dead in dead node list.
366 if (! useful_node_set.test(node_idx) ) {
367 record_dead_node(node_idx);
368 }
369 }
370 }
371
372 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
373 int shift = 0;
374 for (int i = 0; i < inlines->length(); i++) {
375 CallGenerator* cg = inlines->at(i);
376 CallNode* call = cg->call_node();
377 if (shift > 0) {
378 inlines->at_put(i-shift, cg);
379 }
380 if (!useful.member(call)) {
381 shift++;
382 }
383 }
384 inlines->trunc_to(inlines->length()-shift);
385 }
386
387 // Disconnect all useless nodes by disconnecting those at the boundary.
388 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
389 uint next = 0;
390 while (next < useful.size()) {
391 Node *n = useful.at(next++);
392 if (n->is_SafePoint()) {
393 // We're done with a parsing phase. Replaced nodes are not valid
394 // beyond that point.
395 n->as_SafePoint()->delete_replaced_nodes();
396 }
397 // Use raw traversal of out edges since this code removes out edges
398 int max = n->outcnt();
399 for (int j = 0; j < max; ++j) {
400 Node* child = n->raw_out(j);
401 if (! useful.member(child)) {
402 assert(!child->is_top() || child != top(),
403 "If top is cached in Compile object it is in useful list");
404 // Only need to remove this out-edge to the useless node
405 n->raw_del_out(j);
406 --j;
407 --max;
408 }
409 }
410 if (n->outcnt() == 1 && n->has_special_unique_user()) {
411 record_for_igvn(n->unique_out());
412 }
413 if (n->Opcode() == Op_AddP && CallLeafNode::has_only_g1_wb_pre_uses(n)) {
414 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
415 record_for_igvn(n->fast_out(i));
416 }
417 }
418 }
419 // Remove useless macro and predicate opaq nodes
420 for (int i = C->macro_count()-1; i >= 0; i--) {
421 Node* n = C->macro_node(i);
422 if (!useful.member(n)) {
423 remove_macro_node(n);
424 }
425 }
426 // Remove useless CastII nodes with range check dependency
427 for (int i = range_check_cast_count() - 1; i >= 0; i--) {
428 Node* cast = range_check_cast_node(i);
429 if (!useful.member(cast)) {
430 remove_range_check_cast(cast);
431 }
432 }
433 // Remove useless expensive node
434 for (int i = C->expensive_count()-1; i >= 0; i--) {
435 Node* n = C->expensive_node(i);
436 if (!useful.member(n)) {
437 remove_expensive_node(n);
438 }
439 }
440 for (int i = C->shenandoah_barriers_count()-1; i >= 0; i--) {
441 ShenandoahLoadReferenceBarrierNode* n = C->shenandoah_barrier(i);
442 if (!useful.member(n)) {
443 remove_shenandoah_barrier(n);
444 }
445 }
446 // clean up the late inline lists
447 remove_useless_late_inlines(&_string_late_inlines, useful);
448 remove_useless_late_inlines(&_boxing_late_inlines, useful);
449 remove_useless_late_inlines(&_late_inlines, useful);
450 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
451 }
452
453 //------------------------------frame_size_in_words-----------------------------
454 // frame_slots in units of words
455 int Compile::frame_size_in_words() const {
456 // shift is 0 in LP32 and 1 in LP64
457 const int shift = (LogBytesPerWord - LogBytesPerInt);
458 int words = _frame_slots >> shift;
459 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
460 return words;
461 }
462
463 // To bang the stack of this compiled method we use the stack size
464 // that the interpreter would need in case of a deoptimization. This
465 // removes the need to bang the stack in the deoptimization blob which
466 // in turn simplifies stack overflow handling.
467 int Compile::bang_size_in_bytes() const {
468 return MAX2(_interpreter_frame_size, frame_size_in_bytes());
469 }
470
471 // ============================================================================
472 //------------------------------CompileWrapper---------------------------------
473 class CompileWrapper : public StackObj {
474 Compile *const _compile;
475 public:
476 CompileWrapper(Compile* compile);
477
478 ~CompileWrapper();
479 };
480
481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
482 // the Compile* pointer is stored in the current ciEnv:
483 ciEnv* env = compile->env();
484 assert(env == ciEnv::current(), "must already be a ciEnv active");
485 assert(env->compiler_data() == NULL, "compile already active?");
486 env->set_compiler_data(compile);
487 assert(compile == Compile::current(), "sanity");
488
489 compile->set_type_dict(NULL);
490 compile->set_type_hwm(NULL);
491 compile->set_type_last_size(0);
492 compile->set_last_tf(NULL, NULL);
493 compile->set_indexSet_arena(NULL);
494 compile->set_indexSet_free_block_list(NULL);
495 compile->init_type_arena();
496 Type::Initialize(compile);
497 _compile->set_scratch_buffer_blob(NULL);
498 _compile->begin_method();
499 }
500 CompileWrapper::~CompileWrapper() {
501 _compile->end_method();
502 if (_compile->scratch_buffer_blob() != NULL)
503 BufferBlob::free(_compile->scratch_buffer_blob());
504 _compile->env()->set_compiler_data(NULL);
505 }
506
507
508 //----------------------------print_compile_messages---------------------------
509 void Compile::print_compile_messages() {
510 #ifndef PRODUCT
511 // Check if recompiling
512 if (_subsume_loads == false && PrintOpto) {
513 // Recompiling without allowing machine instructions to subsume loads
514 tty->print_cr("*********************************************************");
515 tty->print_cr("** Bailout: Recompile without subsuming loads **");
516 tty->print_cr("*********************************************************");
517 }
518 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
519 // Recompiling without escape analysis
520 tty->print_cr("*********************************************************");
521 tty->print_cr("** Bailout: Recompile without escape analysis **");
522 tty->print_cr("*********************************************************");
523 }
524 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
525 // Recompiling without boxing elimination
526 tty->print_cr("*********************************************************");
527 tty->print_cr("** Bailout: Recompile without boxing elimination **");
528 tty->print_cr("*********************************************************");
529 }
530 if (env()->break_at_compile()) {
531 // Open the debugger when compiling this method.
532 tty->print("### Breaking when compiling: ");
533 method()->print_short_name();
534 tty->cr();
535 BREAKPOINT;
536 }
537
538 if( PrintOpto ) {
539 if (is_osr_compilation()) {
540 tty->print("[OSR]%3d", _compile_id);
541 } else {
542 tty->print("%3d", _compile_id);
543 }
544 }
545 #endif
546 }
547
548
549 //-----------------------init_scratch_buffer_blob------------------------------
550 // Construct a temporary BufferBlob and cache it for this compile.
551 void Compile::init_scratch_buffer_blob(int const_size) {
552 // If there is already a scratch buffer blob allocated and the
553 // constant section is big enough, use it. Otherwise free the
554 // current and allocate a new one.
555 BufferBlob* blob = scratch_buffer_blob();
556 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
557 // Use the current blob.
558 } else {
559 if (blob != NULL) {
560 BufferBlob::free(blob);
561 }
562
563 ResourceMark rm;
564 _scratch_const_size = const_size;
565 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
566 blob = BufferBlob::create("Compile::scratch_buffer", size);
567 // Record the buffer blob for next time.
568 set_scratch_buffer_blob(blob);
569 // Have we run out of code space?
570 if (scratch_buffer_blob() == NULL) {
571 // Let CompilerBroker disable further compilations.
572 record_failure("Not enough space for scratch buffer in CodeCache");
573 return;
574 }
575 }
576
577 // Initialize the relocation buffers
578 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
579 set_scratch_locs_memory(locs_buf);
580 }
581
582
583 //-----------------------scratch_emit_size-------------------------------------
584 // Helper function that computes size by emitting code
585 uint Compile::scratch_emit_size(const Node* n) {
586 // Start scratch_emit_size section.
587 set_in_scratch_emit_size(true);
588
589 // Emit into a trash buffer and count bytes emitted.
590 // This is a pretty expensive way to compute a size,
591 // but it works well enough if seldom used.
592 // All common fixed-size instructions are given a size
593 // method by the AD file.
594 // Note that the scratch buffer blob and locs memory are
595 // allocated at the beginning of the compile task, and
596 // may be shared by several calls to scratch_emit_size.
597 // The allocation of the scratch buffer blob is particularly
598 // expensive, since it has to grab the code cache lock.
599 BufferBlob* blob = this->scratch_buffer_blob();
600 assert(blob != NULL, "Initialize BufferBlob at start");
601 assert(blob->size() > MAX_inst_size, "sanity");
602 relocInfo* locs_buf = scratch_locs_memory();
603 address blob_begin = blob->content_begin();
604 address blob_end = (address)locs_buf;
605 assert(blob->content_contains(blob_end), "sanity");
606 CodeBuffer buf(blob_begin, blob_end - blob_begin);
607 buf.initialize_consts_size(_scratch_const_size);
608 buf.initialize_stubs_size(MAX_stubs_size);
609 assert(locs_buf != NULL, "sanity");
610 int lsize = MAX_locs_size / 3;
611 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
612 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
613 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
614
615 // Do the emission.
616
617 Label fakeL; // Fake label for branch instructions.
618 Label* saveL = NULL;
619 uint save_bnum = 0;
620 bool is_branch = n->is_MachBranch();
621 if (is_branch) {
622 MacroAssembler masm(&buf);
623 masm.bind(fakeL);
624 n->as_MachBranch()->save_label(&saveL, &save_bnum);
625 n->as_MachBranch()->label_set(&fakeL, 0);
626 }
627 n->emit(buf, this->regalloc());
628
629 // Emitting into the scratch buffer should not fail
630 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
631
632 if (is_branch) // Restore label.
633 n->as_MachBranch()->label_set(saveL, save_bnum);
634
635 // End scratch_emit_size section.
636 set_in_scratch_emit_size(false);
637
638 return buf.insts_size();
639 }
640
641
642 // ============================================================================
643 //------------------------------Compile standard-------------------------------
644 debug_only( int Compile::_debug_idx = 100000; )
645
646 // Compile a method. entry_bci is -1 for normal compilations and indicates
647 // the continuation bci for on stack replacement.
648
649
650 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
651 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
652 : Phase(Compiler),
653 _env(ci_env),
654 _log(ci_env->log()),
655 _compile_id(ci_env->compile_id()),
656 _save_argument_registers(false),
657 _stub_name(NULL),
658 _stub_function(NULL),
659 _stub_entry_point(NULL),
660 _method(target),
661 _entry_bci(osr_bci),
662 _initial_gvn(NULL),
663 _for_igvn(NULL),
664 _warm_calls(NULL),
665 _subsume_loads(subsume_loads),
666 _do_escape_analysis(do_escape_analysis),
667 _eliminate_boxing(eliminate_boxing),
668 _failure_reason(NULL),
669 _code_buffer("Compile::Fill_buffer"),
670 _orig_pc_slot(0),
671 _orig_pc_slot_offset_in_bytes(0),
672 _has_method_handle_invokes(false),
673 _mach_constant_base_node(NULL),
674 _node_bundling_limit(0),
675 _node_bundling_base(NULL),
676 _java_calls(0),
677 _inner_loops(0),
678 _scratch_const_size(-1),
679 _in_scratch_emit_size(false),
680 _dead_node_list(comp_arena()),
681 _dead_node_count(0),
682 #ifndef PRODUCT
683 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
684 _in_dump_cnt(0),
685 _printer(IdealGraphPrinter::printer()),
686 #endif
687 _congraph(NULL),
688 _comp_arena(mtCompiler),
689 _node_arena(mtCompiler),
690 _old_arena(mtCompiler),
691 _Compile_types(mtCompiler),
692 _replay_inline_data(NULL),
693 _late_inlines(comp_arena(), 2, 0, NULL),
694 _string_late_inlines(comp_arena(), 2, 0, NULL),
695 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
696 _late_inlines_pos(0),
697 _number_of_mh_late_inlines(0),
698 _inlining_progress(false),
699 _inlining_incrementally(false),
700 _print_inlining_list(NULL),
701 _print_inlining_idx(0),
702 _interpreter_frame_size(0),
703 _max_node_limit(MaxNodeLimit) {
704 C = this;
705
706 CompileWrapper cw(this);
707 #ifndef PRODUCT
708 if (TimeCompiler2) {
709 tty->print(" ");
710 target->holder()->name()->print();
711 tty->print(".");
712 target->print_short_name();
713 tty->print(" ");
714 }
715 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
716 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
717 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
718 if (!print_opto_assembly) {
719 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
720 if (print_assembly && !Disassembler::can_decode()) {
721 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
722 print_opto_assembly = true;
723 }
724 }
725 set_print_assembly(print_opto_assembly);
726 set_parsed_irreducible_loop(false);
727
728 if (method()->has_option("ReplayInline")) {
729 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
730 }
731 #endif
732 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
733 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
734 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
735
736 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
737 // Make sure the method being compiled gets its own MDO,
738 // so we can at least track the decompile_count().
739 // Need MDO to record RTM code generation state.
740 method()->ensure_method_data();
741 }
742
743 Init(::AliasLevel);
744
745
746 print_compile_messages();
747
748 _ilt = InlineTree::build_inline_tree_root();
749
750 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
751 assert(num_alias_types() >= AliasIdxRaw, "");
752
753 #define MINIMUM_NODE_HASH 1023
754 // Node list that Iterative GVN will start with
755 Unique_Node_List for_igvn(comp_arena());
756 set_for_igvn(&for_igvn);
757
758 // GVN that will be run immediately on new nodes
759 uint estimated_size = method()->code_size()*4+64;
760 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
761 PhaseGVN gvn(node_arena(), estimated_size);
762 set_initial_gvn(&gvn);
763
764 if (print_inlining() || print_intrinsics()) {
765 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
766 }
767 { // Scope for timing the parser
768 TracePhase t3("parse", &_t_parser, true);
769
770 // Put top into the hash table ASAP.
771 initial_gvn()->transform_no_reclaim(top());
772
773 // Set up tf(), start(), and find a CallGenerator.
774 CallGenerator* cg = NULL;
775 if (is_osr_compilation()) {
776 const TypeTuple *domain = StartOSRNode::osr_domain();
777 const TypeTuple *range = TypeTuple::make_range(method()->signature());
778 init_tf(TypeFunc::make(domain, range));
779 StartNode* s = new (this) StartOSRNode(root(), domain);
780 initial_gvn()->set_type_bottom(s);
781 init_start(s);
782 cg = CallGenerator::for_osr(method(), entry_bci());
783 } else {
784 // Normal case.
785 init_tf(TypeFunc::make(method()));
786 StartNode* s = new (this) StartNode(root(), tf()->domain());
787 initial_gvn()->set_type_bottom(s);
788 init_start(s);
789 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && (UseG1GC || UseShenandoahGC)) {
790 // With java.lang.ref.reference.get() we must go through the
791 // intrinsic when G1 is enabled - even when get() is the root
792 // method of the compile - so that, if necessary, the value in
793 // the referent field of the reference object gets recorded by
794 // the pre-barrier code.
795 // Specifically, if G1 is enabled, the value in the referent
796 // field is recorded by the G1 SATB pre barrier. This will
797 // result in the referent being marked live and the reference
798 // object removed from the list of discovered references during
799 // reference processing.
800 cg = find_intrinsic(method(), false);
801 }
802 if (cg == NULL) {
803 float past_uses = method()->interpreter_invocation_count();
804 float expected_uses = past_uses;
805 cg = CallGenerator::for_inline(method(), expected_uses);
806 }
807 }
808 if (failing()) return;
809 if (cg == NULL) {
810 record_method_not_compilable_all_tiers("cannot parse method");
811 return;
812 }
813 JVMState* jvms = build_start_state(start(), tf());
814 if ((jvms = cg->generate(jvms)) == NULL) {
815 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
816 record_method_not_compilable("method parse failed");
817 }
818 return;
819 }
820 GraphKit kit(jvms);
821
822 if (!kit.stopped()) {
823 // Accept return values, and transfer control we know not where.
824 // This is done by a special, unique ReturnNode bound to root.
825 return_values(kit.jvms());
826 }
827
828 if (kit.has_exceptions()) {
829 // Any exceptions that escape from this call must be rethrown
830 // to whatever caller is dynamically above us on the stack.
831 // This is done by a special, unique RethrowNode bound to root.
832 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
833 }
834
835 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
836
837 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
838 inline_string_calls(true);
839 }
840
841 if (failing()) return;
842
843 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
844
845 // Remove clutter produced by parsing.
846 if (!failing()) {
847 ResourceMark rm;
848 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
849 }
850 }
851
852 // Note: Large methods are capped off in do_one_bytecode().
853 if (failing()) return;
854
855 // After parsing, node notes are no longer automagic.
856 // They must be propagated by register_new_node_with_optimizer(),
857 // clone(), or the like.
858 set_default_node_notes(NULL);
859
860 for (;;) {
861 int successes = Inline_Warm();
862 if (failing()) return;
863 if (successes == 0) break;
864 }
865
866 // Drain the list.
867 Finish_Warm();
868 #ifndef PRODUCT
869 if (_printer) {
870 _printer->print_inlining(this);
871 }
872 #endif
873
874 if (failing()) return;
875 NOT_PRODUCT( verify_graph_edges(); )
876
877 // Now optimize
878 Optimize();
879 if (failing()) return;
880 NOT_PRODUCT( verify_graph_edges(); )
881
882 #ifndef PRODUCT
883 if (PrintIdeal) {
884 ttyLocker ttyl; // keep the following output all in one block
885 // This output goes directly to the tty, not the compiler log.
886 // To enable tools to match it up with the compilation activity,
887 // be sure to tag this tty output with the compile ID.
888 if (xtty != NULL) {
889 xtty->head("ideal compile_id='%d'%s", compile_id(),
890 is_osr_compilation() ? " compile_kind='osr'" :
891 "");
892 }
893 root()->dump(9999);
894 if (xtty != NULL) {
895 xtty->tail("ideal");
896 }
897 }
898 #endif
899
900 NOT_PRODUCT( verify_barriers(); )
901
902 // Dump compilation data to replay it.
903 if (method()->has_option("DumpReplay")) {
904 env()->dump_replay_data(_compile_id);
905 }
906 if (method()->has_option("DumpInline") && (ilt() != NULL)) {
907 env()->dump_inline_data(_compile_id);
908 }
909
910 // Now that we know the size of all the monitors we can add a fixed slot
911 // for the original deopt pc.
912
913 _orig_pc_slot = fixed_slots();
914 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
915 set_fixed_slots(next_slot);
916
917 // Compute when to use implicit null checks. Used by matching trap based
918 // nodes and NullCheck optimization.
919 set_allowed_deopt_reasons();
920
921 // Now generate code
922 Code_Gen();
923 if (failing()) return;
924
925 // Check if we want to skip execution of all compiled code.
926 {
927 #ifndef PRODUCT
928 if (OptoNoExecute) {
929 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
930 return;
931 }
932 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
933 #endif
934
935 if (is_osr_compilation()) {
936 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
937 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
938 } else {
939 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
940 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
941 }
942
943 env()->register_method(_method, _entry_bci,
944 &_code_offsets,
945 _orig_pc_slot_offset_in_bytes,
946 code_buffer(),
947 frame_size_in_words(), _oop_map_set,
948 &_handler_table, &_inc_table,
949 compiler,
950 env()->comp_level(),
951 has_unsafe_access(),
952 SharedRuntime::is_wide_vector(max_vector_size()),
953 rtm_state()
954 );
955
956 if (log() != NULL) // Print code cache state into compiler log
957 log()->code_cache_state();
958 }
959 }
960
961 //------------------------------Compile----------------------------------------
962 // Compile a runtime stub
963 Compile::Compile( ciEnv* ci_env,
964 TypeFunc_generator generator,
965 address stub_function,
966 const char *stub_name,
967 int is_fancy_jump,
968 bool pass_tls,
969 bool save_arg_registers,
970 bool return_pc )
971 : Phase(Compiler),
972 _env(ci_env),
973 _log(ci_env->log()),
974 _compile_id(0),
975 _save_argument_registers(save_arg_registers),
976 _method(NULL),
977 _stub_name(stub_name),
978 _stub_function(stub_function),
979 _stub_entry_point(NULL),
980 _entry_bci(InvocationEntryBci),
981 _initial_gvn(NULL),
982 _for_igvn(NULL),
983 _warm_calls(NULL),
984 _orig_pc_slot(0),
985 _orig_pc_slot_offset_in_bytes(0),
986 _subsume_loads(true),
987 _do_escape_analysis(false),
988 _eliminate_boxing(false),
989 _failure_reason(NULL),
990 _code_buffer("Compile::Fill_buffer"),
991 _has_method_handle_invokes(false),
992 _mach_constant_base_node(NULL),
993 _node_bundling_limit(0),
994 _node_bundling_base(NULL),
995 _java_calls(0),
996 _inner_loops(0),
997 #ifndef PRODUCT
998 _trace_opto_output(TraceOptoOutput),
999 _in_dump_cnt(0),
1000 _printer(NULL),
1001 #endif
1002 _comp_arena(mtCompiler),
1003 _node_arena(mtCompiler),
1004 _old_arena(mtCompiler),
1005 _Compile_types(mtCompiler),
1006 _dead_node_list(comp_arena()),
1007 _dead_node_count(0),
1008 _congraph(NULL),
1009 _replay_inline_data(NULL),
1010 _number_of_mh_late_inlines(0),
1011 _inlining_progress(false),
1012 _inlining_incrementally(false),
1013 _print_inlining_list(NULL),
1014 _print_inlining_idx(0),
1015 _allowed_reasons(0),
1016 _interpreter_frame_size(0),
1017 _max_node_limit(MaxNodeLimit) {
1018 C = this;
1019
1020 #ifndef PRODUCT
1021 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1022 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1023 set_print_assembly(PrintFrameConverterAssembly);
1024 set_parsed_irreducible_loop(false);
1025 #endif
1026 set_has_irreducible_loop(false); // no loops
1027
1028 CompileWrapper cw(this);
1029 Init(/*AliasLevel=*/ 0);
1030 init_tf((*generator)());
1031
1032 {
1033 // The following is a dummy for the sake of GraphKit::gen_stub
1034 Unique_Node_List for_igvn(comp_arena());
1035 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1036 PhaseGVN gvn(Thread::current()->resource_area(),255);
1037 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1038 gvn.transform_no_reclaim(top());
1039
1040 GraphKit kit;
1041 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1042 }
1043
1044 NOT_PRODUCT( verify_graph_edges(); )
1045 Code_Gen();
1046 if (failing()) return;
1047
1048
1049 // Entry point will be accessed using compile->stub_entry_point();
1050 if (code_buffer() == NULL) {
1051 Matcher::soft_match_failure();
1052 } else {
1053 if (PrintAssembly && (WizardMode || Verbose))
1054 tty->print_cr("### Stub::%s", stub_name);
1055
1056 if (!failing()) {
1057 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1058
1059 // Make the NMethod
1060 // For now we mark the frame as never safe for profile stackwalking
1061 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1062 code_buffer(),
1063 CodeOffsets::frame_never_safe,
1064 // _code_offsets.value(CodeOffsets::Frame_Complete),
1065 frame_size_in_words(),
1066 _oop_map_set,
1067 save_arg_registers);
1068 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1069
1070 _stub_entry_point = rs->entry_point();
1071 }
1072 }
1073 }
1074
1075 //------------------------------Init-------------------------------------------
1076 // Prepare for a single compilation
1077 void Compile::Init(int aliaslevel) {
1078 _unique = 0;
1079 _regalloc = NULL;
1080
1081 _tf = NULL; // filled in later
1082 _top = NULL; // cached later
1083 _matcher = NULL; // filled in later
1084 _cfg = NULL; // filled in later
1085
1086 set_24_bit_selection_and_mode(Use24BitFP, false);
1087
1088 _node_note_array = NULL;
1089 _default_node_notes = NULL;
1090
1091 _immutable_memory = NULL; // filled in at first inquiry
1092
1093 // Globally visible Nodes
1094 // First set TOP to NULL to give safe behavior during creation of RootNode
1095 set_cached_top_node(NULL);
1096 set_root(new (this) RootNode());
1097 // Now that you have a Root to point to, create the real TOP
1098 set_cached_top_node( new (this) ConNode(Type::TOP) );
1099 set_recent_alloc(NULL, NULL);
1100
1101 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1102 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1103 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1104 env()->set_dependencies(new Dependencies(env()));
1105
1106 _fixed_slots = 0;
1107 set_has_split_ifs(false);
1108 set_has_loops(has_method() && method()->has_loops()); // first approximation
1109 set_has_stringbuilder(false);
1110 set_has_boxed_value(false);
1111 _trap_can_recompile = false; // no traps emitted yet
1112 _major_progress = true; // start out assuming good things will happen
1113 set_has_unsafe_access(false);
1114 set_max_vector_size(0);
1115 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1116 set_decompile_count(0);
1117
1118 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1119 set_num_loop_opts(LoopOptsCount);
1120 set_do_inlining(Inline);
1121 set_max_inline_size(MaxInlineSize);
1122 set_freq_inline_size(FreqInlineSize);
1123 set_do_scheduling(OptoScheduling);
1124 set_do_count_invocations(false);
1125 set_do_method_data_update(false);
1126 set_rtm_state(NoRTM); // No RTM lock eliding by default
1127 method_has_option_value("MaxNodeLimit", _max_node_limit);
1128 #if INCLUDE_RTM_OPT
1129 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1130 int rtm_state = method()->method_data()->rtm_state();
1131 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1132 // Don't generate RTM lock eliding code.
1133 set_rtm_state(NoRTM);
1134 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1135 // Generate RTM lock eliding code without abort ratio calculation code.
1136 set_rtm_state(UseRTM);
1137 } else if (UseRTMDeopt) {
1138 // Generate RTM lock eliding code and include abort ratio calculation
1139 // code if UseRTMDeopt is on.
1140 set_rtm_state(ProfileRTM);
1141 }
1142 }
1143 #endif
1144 if (debug_info()->recording_non_safepoints()) {
1145 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1146 (comp_arena(), 8, 0, NULL));
1147 set_default_node_notes(Node_Notes::make(this));
1148 }
1149
1150 // // -- Initialize types before each compile --
1151 // // Update cached type information
1152 // if( _method && _method->constants() )
1153 // Type::update_loaded_types(_method, _method->constants());
1154
1155 // Init alias_type map.
1156 if (!_do_escape_analysis && aliaslevel == 3)
1157 aliaslevel = 2; // No unique types without escape analysis
1158 _AliasLevel = aliaslevel;
1159 const int grow_ats = 16;
1160 _max_alias_types = grow_ats;
1161 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1162 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1163 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1164 {
1165 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1166 }
1167 // Initialize the first few types.
1168 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1169 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1170 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1171 _num_alias_types = AliasIdxRaw+1;
1172 // Zero out the alias type cache.
1173 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1174 // A NULL adr_type hits in the cache right away. Preload the right answer.
1175 probe_alias_cache(NULL)->_index = AliasIdxTop;
1176
1177 _intrinsics = NULL;
1178 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1179 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1180 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1181 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1182 _shenandoah_barriers = new(comp_arena()) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena(), 8, 0, NULL);
1183 register_library_intrinsics();
1184 #ifdef ASSERT
1185 _type_verify_symmetry = true;
1186 #endif
1187 }
1188
1189 //---------------------------init_start----------------------------------------
1190 // Install the StartNode on this compile object.
1191 void Compile::init_start(StartNode* s) {
1192 if (failing())
1193 return; // already failing
1194 assert(s == start(), "");
1195 }
1196
1197 StartNode* Compile::start() const {
1198 assert(!failing(), "");
1199 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1200 Node* start = root()->fast_out(i);
1201 if( start->is_Start() )
1202 return start->as_Start();
1203 }
1204 fatal("Did not find Start node!");
1205 return NULL;
1206 }
1207
1208 //-------------------------------immutable_memory-------------------------------------
1209 // Access immutable memory
1210 Node* Compile::immutable_memory() {
1211 if (_immutable_memory != NULL) {
1212 return _immutable_memory;
1213 }
1214 StartNode* s = start();
1215 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1216 Node *p = s->fast_out(i);
1217 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1218 _immutable_memory = p;
1219 return _immutable_memory;
1220 }
1221 }
1222 ShouldNotReachHere();
1223 return NULL;
1224 }
1225
1226 //----------------------set_cached_top_node------------------------------------
1227 // Install the cached top node, and make sure Node::is_top works correctly.
1228 void Compile::set_cached_top_node(Node* tn) {
1229 if (tn != NULL) verify_top(tn);
1230 Node* old_top = _top;
1231 _top = tn;
1232 // Calling Node::setup_is_top allows the nodes the chance to adjust
1233 // their _out arrays.
1234 if (_top != NULL) _top->setup_is_top();
1235 if (old_top != NULL) old_top->setup_is_top();
1236 assert(_top == NULL || top()->is_top(), "");
1237 }
1238
1239 #ifdef ASSERT
1240 uint Compile::count_live_nodes_by_graph_walk() {
1241 Unique_Node_List useful(comp_arena());
1242 // Get useful node list by walking the graph.
1243 identify_useful_nodes(useful);
1244 return useful.size();
1245 }
1246
1247 void Compile::print_missing_nodes() {
1248
1249 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1250 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1251 return;
1252 }
1253
1254 // This is an expensive function. It is executed only when the user
1255 // specifies VerifyIdealNodeCount option or otherwise knows the
1256 // additional work that needs to be done to identify reachable nodes
1257 // by walking the flow graph and find the missing ones using
1258 // _dead_node_list.
1259
1260 Unique_Node_List useful(comp_arena());
1261 // Get useful node list by walking the graph.
1262 identify_useful_nodes(useful);
1263
1264 uint l_nodes = C->live_nodes();
1265 uint l_nodes_by_walk = useful.size();
1266
1267 if (l_nodes != l_nodes_by_walk) {
1268 if (_log != NULL) {
1269 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1270 _log->stamp();
1271 _log->end_head();
1272 }
1273 VectorSet& useful_member_set = useful.member_set();
1274 int last_idx = l_nodes_by_walk;
1275 for (int i = 0; i < last_idx; i++) {
1276 if (useful_member_set.test(i)) {
1277 if (_dead_node_list.test(i)) {
1278 if (_log != NULL) {
1279 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1280 }
1281 if (PrintIdealNodeCount) {
1282 // Print the log message to tty
1283 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1284 useful.at(i)->dump();
1285 }
1286 }
1287 }
1288 else if (! _dead_node_list.test(i)) {
1289 if (_log != NULL) {
1290 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1291 }
1292 if (PrintIdealNodeCount) {
1293 // Print the log message to tty
1294 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1295 }
1296 }
1297 }
1298 if (_log != NULL) {
1299 _log->tail("mismatched_nodes");
1300 }
1301 }
1302 }
1303 #endif
1304
1305 #ifndef PRODUCT
1306 void Compile::verify_top(Node* tn) const {
1307 if (tn != NULL) {
1308 assert(tn->is_Con(), "top node must be a constant");
1309 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1310 assert(tn->in(0) != NULL, "must have live top node");
1311 }
1312 }
1313 #endif
1314
1315
1316 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1317
1318 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1319 guarantee(arr != NULL, "");
1320 int num_blocks = arr->length();
1321 if (grow_by < num_blocks) grow_by = num_blocks;
1322 int num_notes = grow_by * _node_notes_block_size;
1323 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1324 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1325 while (num_notes > 0) {
1326 arr->append(notes);
1327 notes += _node_notes_block_size;
1328 num_notes -= _node_notes_block_size;
1329 }
1330 assert(num_notes == 0, "exact multiple, please");
1331 }
1332
1333 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1334 if (source == NULL || dest == NULL) return false;
1335
1336 if (dest->is_Con())
1337 return false; // Do not push debug info onto constants.
1338
1339 #ifdef ASSERT
1340 // Leave a bread crumb trail pointing to the original node:
1341 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1342 dest->set_debug_orig(source);
1343 }
1344 #endif
1345
1346 if (node_note_array() == NULL)
1347 return false; // Not collecting any notes now.
1348
1349 // This is a copy onto a pre-existing node, which may already have notes.
1350 // If both nodes have notes, do not overwrite any pre-existing notes.
1351 Node_Notes* source_notes = node_notes_at(source->_idx);
1352 if (source_notes == NULL || source_notes->is_clear()) return false;
1353 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1354 if (dest_notes == NULL || dest_notes->is_clear()) {
1355 return set_node_notes_at(dest->_idx, source_notes);
1356 }
1357
1358 Node_Notes merged_notes = (*source_notes);
1359 // The order of operations here ensures that dest notes will win...
1360 merged_notes.update_from(dest_notes);
1361 return set_node_notes_at(dest->_idx, &merged_notes);
1362 }
1363
1364
1365 //--------------------------allow_range_check_smearing-------------------------
1366 // Gating condition for coalescing similar range checks.
1367 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1368 // single covering check that is at least as strong as any of them.
1369 // If the optimization succeeds, the simplified (strengthened) range check
1370 // will always succeed. If it fails, we will deopt, and then give up
1371 // on the optimization.
1372 bool Compile::allow_range_check_smearing() const {
1373 // If this method has already thrown a range-check,
1374 // assume it was because we already tried range smearing
1375 // and it failed.
1376 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1377 return !already_trapped;
1378 }
1379
1380
1381 //------------------------------flatten_alias_type-----------------------------
1382 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1383 int offset = tj->offset();
1384 TypePtr::PTR ptr = tj->ptr();
1385
1386 // Known instance (scalarizable allocation) alias only with itself.
1387 bool is_known_inst = tj->isa_oopptr() != NULL &&
1388 tj->is_oopptr()->is_known_instance();
1389
1390 // Process weird unsafe references.
1391 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1392 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1393 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1394 tj = TypeOopPtr::BOTTOM;
1395 ptr = tj->ptr();
1396 offset = tj->offset();
1397 }
1398
1399 // Array pointers need some flattening
1400 const TypeAryPtr *ta = tj->isa_aryptr();
1401 if (ta && ta->is_stable()) {
1402 // Erase stability property for alias analysis.
1403 tj = ta = ta->cast_to_stable(false);
1404 }
1405 if( ta && is_known_inst ) {
1406 if ( offset != Type::OffsetBot &&
1407 offset > arrayOopDesc::length_offset_in_bytes() ) {
1408 offset = Type::OffsetBot; // Flatten constant access into array body only
1409 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1410 }
1411 } else if( ta && _AliasLevel >= 2 ) {
1412 // For arrays indexed by constant indices, we flatten the alias
1413 // space to include all of the array body. Only the header, klass
1414 // and array length can be accessed un-aliased.
1415 if( offset != Type::OffsetBot ) {
1416 if( ta->const_oop() ) { // MethodData* or Method*
1417 offset = Type::OffsetBot; // Flatten constant access into array body
1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1419 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1420 // range is OK as-is.
1421 tj = ta = TypeAryPtr::RANGE;
1422 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1423 tj = TypeInstPtr::KLASS; // all klass loads look alike
1424 ta = TypeAryPtr::RANGE; // generic ignored junk
1425 ptr = TypePtr::BotPTR;
1426 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1427 tj = TypeInstPtr::MARK;
1428 ta = TypeAryPtr::RANGE; // generic ignored junk
1429 ptr = TypePtr::BotPTR;
1430 } else { // Random constant offset into array body
1431 offset = Type::OffsetBot; // Flatten constant access into array body
1432 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1433 }
1434 }
1435 // Arrays of fixed size alias with arrays of unknown size.
1436 if (ta->size() != TypeInt::POS) {
1437 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1438 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1439 }
1440 // Arrays of known objects become arrays of unknown objects.
1441 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1442 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1443 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1444 }
1445 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1446 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1447 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1448 }
1449 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1450 // cannot be distinguished by bytecode alone.
1451 if (ta->elem() == TypeInt::BOOL) {
1452 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1453 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1455 }
1456 // During the 2nd round of IterGVN, NotNull castings are removed.
1457 // Make sure the Bottom and NotNull variants alias the same.
1458 // Also, make sure exact and non-exact variants alias the same.
1459 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1460 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1461 }
1462 }
1463
1464 // Oop pointers need some flattening
1465 const TypeInstPtr *to = tj->isa_instptr();
1466 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1467 ciInstanceKlass *k = to->klass()->as_instance_klass();
1468 if( ptr == TypePtr::Constant ) {
1469 if (to->klass() != ciEnv::current()->Class_klass() ||
1470 offset < k->size_helper() * wordSize) {
1471 // No constant oop pointers (such as Strings); they alias with
1472 // unknown strings.
1473 assert(!is_known_inst, "not scalarizable allocation");
1474 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1475 }
1476 } else if( is_known_inst ) {
1477 tj = to; // Keep NotNull and klass_is_exact for instance type
1478 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1479 // During the 2nd round of IterGVN, NotNull castings are removed.
1480 // Make sure the Bottom and NotNull variants alias the same.
1481 // Also, make sure exact and non-exact variants alias the same.
1482 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1483 }
1484 if (to->speculative() != NULL) {
1485 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1486 }
1487 // Canonicalize the holder of this field
1488 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1489 // First handle header references such as a LoadKlassNode, even if the
1490 // object's klass is unloaded at compile time (4965979).
1491 if (!is_known_inst) { // Do it only for non-instance types
1492 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1493 }
1494 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1495 // Static fields are in the space above the normal instance
1496 // fields in the java.lang.Class instance.
1497 if (to->klass() != ciEnv::current()->Class_klass()) {
1498 to = NULL;
1499 tj = TypeOopPtr::BOTTOM;
1500 offset = tj->offset();
1501 }
1502 } else {
1503 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1504 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1505 if( is_known_inst ) {
1506 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1507 } else {
1508 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1509 }
1510 }
1511 }
1512 }
1513
1514 // Klass pointers to object array klasses need some flattening
1515 const TypeKlassPtr *tk = tj->isa_klassptr();
1516 if( tk ) {
1517 // If we are referencing a field within a Klass, we need
1518 // to assume the worst case of an Object. Both exact and
1519 // inexact types must flatten to the same alias class so
1520 // use NotNull as the PTR.
1521 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1522
1523 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1524 TypeKlassPtr::OBJECT->klass(),
1525 offset);
1526 }
1527
1528 ciKlass* klass = tk->klass();
1529 if( klass->is_obj_array_klass() ) {
1530 ciKlass* k = TypeAryPtr::OOPS->klass();
1531 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1532 k = TypeInstPtr::BOTTOM->klass();
1533 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1534 }
1535
1536 // Check for precise loads from the primary supertype array and force them
1537 // to the supertype cache alias index. Check for generic array loads from
1538 // the primary supertype array and also force them to the supertype cache
1539 // alias index. Since the same load can reach both, we need to merge
1540 // these 2 disparate memories into the same alias class. Since the
1541 // primary supertype array is read-only, there's no chance of confusion
1542 // where we bypass an array load and an array store.
1543 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1544 if (offset == Type::OffsetBot ||
1545 (offset >= primary_supers_offset &&
1546 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1547 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1548 offset = in_bytes(Klass::secondary_super_cache_offset());
1549 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1550 }
1551 }
1552
1553 // Flatten all Raw pointers together.
1554 if (tj->base() == Type::RawPtr)
1555 tj = TypeRawPtr::BOTTOM;
1556
1557 if (tj->base() == Type::AnyPtr)
1558 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1559
1560 // Flatten all to bottom for now
1561 switch( _AliasLevel ) {
1562 case 0:
1563 tj = TypePtr::BOTTOM;
1564 break;
1565 case 1: // Flatten to: oop, static, field or array
1566 switch (tj->base()) {
1567 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1568 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1569 case Type::AryPtr: // do not distinguish arrays at all
1570 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1571 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1572 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1573 default: ShouldNotReachHere();
1574 }
1575 break;
1576 case 2: // No collapsing at level 2; keep all splits
1577 case 3: // No collapsing at level 3; keep all splits
1578 break;
1579 default:
1580 Unimplemented();
1581 }
1582
1583 offset = tj->offset();
1584 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1585
1586 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1587 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1588 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1589 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1590 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1591 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1592 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1593 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1594 assert( tj->ptr() != TypePtr::TopPTR &&
1595 tj->ptr() != TypePtr::AnyNull &&
1596 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1597 // assert( tj->ptr() != TypePtr::Constant ||
1598 // tj->base() == Type::RawPtr ||
1599 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1600
1601 return tj;
1602 }
1603
1604 void Compile::AliasType::Init(int i, const TypePtr* at) {
1605 _index = i;
1606 _adr_type = at;
1607 _field = NULL;
1608 _element = NULL;
1609 _is_rewritable = true; // default
1610 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1611 if (atoop != NULL && atoop->is_known_instance()) {
1612 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1613 _general_index = Compile::current()->get_alias_index(gt);
1614 } else {
1615 _general_index = 0;
1616 }
1617 }
1618
1619 BasicType Compile::AliasType::basic_type() const {
1620 if (element() != NULL) {
1621 const Type* element = adr_type()->is_aryptr()->elem();
1622 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1623 } if (field() != NULL) {
1624 return field()->layout_type();
1625 } else {
1626 return T_ILLEGAL; // unknown
1627 }
1628 }
1629
1630 //---------------------------------print_on------------------------------------
1631 #ifndef PRODUCT
1632 void Compile::AliasType::print_on(outputStream* st) {
1633 if (index() < 10)
1634 st->print("@ <%d> ", index());
1635 else st->print("@ <%d>", index());
1636 st->print(is_rewritable() ? " " : " RO");
1637 int offset = adr_type()->offset();
1638 if (offset == Type::OffsetBot)
1639 st->print(" +any");
1640 else st->print(" +%-3d", offset);
1641 st->print(" in ");
1642 adr_type()->dump_on(st);
1643 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1644 if (field() != NULL && tjp) {
1645 if (tjp->klass() != field()->holder() ||
1646 tjp->offset() != field()->offset_in_bytes()) {
1647 st->print(" != ");
1648 field()->print();
1649 st->print(" ***");
1650 }
1651 }
1652 }
1653
1654 void print_alias_types() {
1655 Compile* C = Compile::current();
1656 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1657 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1658 C->alias_type(idx)->print_on(tty);
1659 tty->cr();
1660 }
1661 }
1662 #endif
1663
1664
1665 //----------------------------probe_alias_cache--------------------------------
1666 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1667 intptr_t key = (intptr_t) adr_type;
1668 key ^= key >> logAliasCacheSize;
1669 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1670 }
1671
1672
1673 //-----------------------------grow_alias_types--------------------------------
1674 void Compile::grow_alias_types() {
1675 const int old_ats = _max_alias_types; // how many before?
1676 const int new_ats = old_ats; // how many more?
1677 const int grow_ats = old_ats+new_ats; // how many now?
1678 _max_alias_types = grow_ats;
1679 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1680 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1681 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1682 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1683 }
1684
1685
1686 //--------------------------------find_alias_type------------------------------
1687 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1688 if (_AliasLevel == 0)
1689 return alias_type(AliasIdxBot);
1690
1691 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1692 if (ace->_adr_type == adr_type) {
1693 return alias_type(ace->_index);
1694 }
1695
1696 // Handle special cases.
1697 if (adr_type == NULL) return alias_type(AliasIdxTop);
1698 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1699
1700 // Do it the slow way.
1701 const TypePtr* flat = flatten_alias_type(adr_type);
1702
1703 #ifdef ASSERT
1704 {
1705 ResourceMark rm;
1706 assert(flat == flatten_alias_type(flat),
1707 err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type),
1708 Type::str(flat), Type::str(flatten_alias_type(flat))));
1709 assert(flat != TypePtr::BOTTOM,
1710 err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type)));
1711 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1712 const TypeOopPtr* foop = flat->is_oopptr();
1713 // Scalarizable allocations have exact klass always.
1714 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1715 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1716 assert(foop == flatten_alias_type(xoop),
1717 err_msg("exactness must not affect alias type: foop = %s; xoop = %s",
1718 Type::str(foop), Type::str(xoop)));
1719 }
1720 }
1721 #endif
1722
1723 int idx = AliasIdxTop;
1724 for (int i = 0; i < num_alias_types(); i++) {
1725 if (alias_type(i)->adr_type() == flat) {
1726 idx = i;
1727 break;
1728 }
1729 }
1730
1731 if (idx == AliasIdxTop) {
1732 if (no_create) return NULL;
1733 // Grow the array if necessary.
1734 if (_num_alias_types == _max_alias_types) grow_alias_types();
1735 // Add a new alias type.
1736 idx = _num_alias_types++;
1737 _alias_types[idx]->Init(idx, flat);
1738 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1739 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1740 if (flat->isa_instptr()) {
1741 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1742 && flat->is_instptr()->klass() == env()->Class_klass())
1743 alias_type(idx)->set_rewritable(false);
1744 }
1745 if (flat->isa_aryptr()) {
1746 #ifdef ASSERT
1747 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1748 // (T_BYTE has the weakest alignment and size restrictions...)
1749 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1750 #endif
1751 if (flat->offset() == TypePtr::OffsetBot) {
1752 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1753 }
1754 }
1755 if (flat->isa_klassptr()) {
1756 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1757 alias_type(idx)->set_rewritable(false);
1758 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1759 alias_type(idx)->set_rewritable(false);
1760 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1761 alias_type(idx)->set_rewritable(false);
1762 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1763 alias_type(idx)->set_rewritable(false);
1764 }
1765 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1766 // but the base pointer type is not distinctive enough to identify
1767 // references into JavaThread.)
1768
1769 // Check for final fields.
1770 const TypeInstPtr* tinst = flat->isa_instptr();
1771 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1772 ciField* field;
1773 if (tinst->const_oop() != NULL &&
1774 tinst->klass() == ciEnv::current()->Class_klass() &&
1775 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1776 // static field
1777 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1778 field = k->get_field_by_offset(tinst->offset(), true);
1779 } else {
1780 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1781 field = k->get_field_by_offset(tinst->offset(), false);
1782 }
1783 assert(field == NULL ||
1784 original_field == NULL ||
1785 (field->holder() == original_field->holder() &&
1786 field->offset() == original_field->offset() &&
1787 field->is_static() == original_field->is_static()), "wrong field?");
1788 // Set field() and is_rewritable() attributes.
1789 if (field != NULL) alias_type(idx)->set_field(field);
1790 }
1791 }
1792
1793 // Fill the cache for next time.
1794 ace->_adr_type = adr_type;
1795 ace->_index = idx;
1796 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1797
1798 // Might as well try to fill the cache for the flattened version, too.
1799 AliasCacheEntry* face = probe_alias_cache(flat);
1800 if (face->_adr_type == NULL) {
1801 face->_adr_type = flat;
1802 face->_index = idx;
1803 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1804 }
1805
1806 return alias_type(idx);
1807 }
1808
1809
1810 Compile::AliasType* Compile::alias_type(ciField* field) {
1811 const TypeOopPtr* t;
1812 if (field->is_static())
1813 t = TypeInstPtr::make(field->holder()->java_mirror());
1814 else
1815 t = TypeOopPtr::make_from_klass_raw(field->holder());
1816 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1817 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1818 return atp;
1819 }
1820
1821
1822 //------------------------------have_alias_type--------------------------------
1823 bool Compile::have_alias_type(const TypePtr* adr_type) {
1824 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1825 if (ace->_adr_type == adr_type) {
1826 return true;
1827 }
1828
1829 // Handle special cases.
1830 if (adr_type == NULL) return true;
1831 if (adr_type == TypePtr::BOTTOM) return true;
1832
1833 return find_alias_type(adr_type, true, NULL) != NULL;
1834 }
1835
1836 //-----------------------------must_alias--------------------------------------
1837 // True if all values of the given address type are in the given alias category.
1838 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1839 if (alias_idx == AliasIdxBot) return true; // the universal category
1840 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1841 if (alias_idx == AliasIdxTop) return false; // the empty category
1842 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1843
1844 // the only remaining possible overlap is identity
1845 int adr_idx = get_alias_index(adr_type);
1846 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1847 assert(adr_idx == alias_idx ||
1848 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1849 && adr_type != TypeOopPtr::BOTTOM),
1850 "should not be testing for overlap with an unsafe pointer");
1851 return adr_idx == alias_idx;
1852 }
1853
1854 //------------------------------can_alias--------------------------------------
1855 // True if any values of the given address type are in the given alias category.
1856 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1857 if (alias_idx == AliasIdxTop) return false; // the empty category
1858 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1859 if (alias_idx == AliasIdxBot) return true; // the universal category
1860 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1861
1862 // the only remaining possible overlap is identity
1863 int adr_idx = get_alias_index(adr_type);
1864 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1865 return adr_idx == alias_idx;
1866 }
1867
1868
1869
1870 //---------------------------pop_warm_call-------------------------------------
1871 WarmCallInfo* Compile::pop_warm_call() {
1872 WarmCallInfo* wci = _warm_calls;
1873 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1874 return wci;
1875 }
1876
1877 //----------------------------Inline_Warm--------------------------------------
1878 int Compile::Inline_Warm() {
1879 // If there is room, try to inline some more warm call sites.
1880 // %%% Do a graph index compaction pass when we think we're out of space?
1881 if (!InlineWarmCalls) return 0;
1882
1883 int calls_made_hot = 0;
1884 int room_to_grow = NodeCountInliningCutoff - unique();
1885 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1886 int amount_grown = 0;
1887 WarmCallInfo* call;
1888 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1889 int est_size = (int)call->size();
1890 if (est_size > (room_to_grow - amount_grown)) {
1891 // This one won't fit anyway. Get rid of it.
1892 call->make_cold();
1893 continue;
1894 }
1895 call->make_hot();
1896 calls_made_hot++;
1897 amount_grown += est_size;
1898 amount_to_grow -= est_size;
1899 }
1900
1901 if (calls_made_hot > 0) set_major_progress();
1902 return calls_made_hot;
1903 }
1904
1905
1906 //----------------------------Finish_Warm--------------------------------------
1907 void Compile::Finish_Warm() {
1908 if (!InlineWarmCalls) return;
1909 if (failing()) return;
1910 if (warm_calls() == NULL) return;
1911
1912 // Clean up loose ends, if we are out of space for inlining.
1913 WarmCallInfo* call;
1914 while ((call = pop_warm_call()) != NULL) {
1915 call->make_cold();
1916 }
1917 }
1918
1919 //---------------------cleanup_loop_predicates-----------------------
1920 // Remove the opaque nodes that protect the predicates so that all unused
1921 // checks and uncommon_traps will be eliminated from the ideal graph
1922 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1923 if (predicate_count()==0) return;
1924 for (int i = predicate_count(); i > 0; i--) {
1925 Node * n = predicate_opaque1_node(i-1);
1926 assert(n->Opcode() == Op_Opaque1, "must be");
1927 igvn.replace_node(n, n->in(1));
1928 }
1929 assert(predicate_count()==0, "should be clean!");
1930 }
1931
1932 void Compile::add_range_check_cast(Node* n) {
1933 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1934 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1935 _range_check_casts->append(n);
1936 }
1937
1938 // Remove all range check dependent CastIINodes.
1939 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1940 for (int i = range_check_cast_count(); i > 0; i--) {
1941 Node* cast = range_check_cast_node(i-1);
1942 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1943 igvn.replace_node(cast, cast->in(1));
1944 }
1945 assert(range_check_cast_count() == 0, "should be empty");
1946 }
1947
1948 // StringOpts and late inlining of string methods
1949 void Compile::inline_string_calls(bool parse_time) {
1950 {
1951 // remove useless nodes to make the usage analysis simpler
1952 ResourceMark rm;
1953 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1954 }
1955
1956 {
1957 ResourceMark rm;
1958 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1959 PhaseStringOpts pso(initial_gvn(), for_igvn());
1960 print_method(PHASE_AFTER_STRINGOPTS, 3);
1961 }
1962
1963 // now inline anything that we skipped the first time around
1964 if (!parse_time) {
1965 _late_inlines_pos = _late_inlines.length();
1966 }
1967
1968 while (_string_late_inlines.length() > 0) {
1969 CallGenerator* cg = _string_late_inlines.pop();
1970 cg->do_late_inline();
1971 if (failing()) return;
1972 }
1973 _string_late_inlines.trunc_to(0);
1974 }
1975
1976 // Late inlining of boxing methods
1977 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1978 if (_boxing_late_inlines.length() > 0) {
1979 assert(has_boxed_value(), "inconsistent");
1980
1981 PhaseGVN* gvn = initial_gvn();
1982 set_inlining_incrementally(true);
1983
1984 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1985 for_igvn()->clear();
1986 gvn->replace_with(&igvn);
1987
1988 _late_inlines_pos = _late_inlines.length();
1989
1990 while (_boxing_late_inlines.length() > 0) {
1991 CallGenerator* cg = _boxing_late_inlines.pop();
1992 cg->do_late_inline();
1993 if (failing()) return;
1994 }
1995 _boxing_late_inlines.trunc_to(0);
1996
1997 {
1998 ResourceMark rm;
1999 PhaseRemoveUseless pru(gvn, for_igvn());
2000 }
2001
2002 igvn = PhaseIterGVN(gvn);
2003 igvn.optimize();
2004
2005 set_inlining_progress(false);
2006 set_inlining_incrementally(false);
2007 }
2008 }
2009
2010 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2011 assert(IncrementalInline, "incremental inlining should be on");
2012 PhaseGVN* gvn = initial_gvn();
2013
2014 set_inlining_progress(false);
2015 for_igvn()->clear();
2016 gvn->replace_with(&igvn);
2017
2018 int i = 0;
2019
2020 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2021 CallGenerator* cg = _late_inlines.at(i);
2022 _late_inlines_pos = i+1;
2023 cg->do_late_inline();
2024 if (failing()) return;
2025 }
2026 int j = 0;
2027 for (; i < _late_inlines.length(); i++, j++) {
2028 _late_inlines.at_put(j, _late_inlines.at(i));
2029 }
2030 _late_inlines.trunc_to(j);
2031
2032 {
2033 ResourceMark rm;
2034 PhaseRemoveUseless pru(gvn, for_igvn());
2035 }
2036
2037 igvn = PhaseIterGVN(gvn);
2038 }
2039
2040 // Perform incremental inlining until bound on number of live nodes is reached
2041 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2042 PhaseGVN* gvn = initial_gvn();
2043
2044 set_inlining_incrementally(true);
2045 set_inlining_progress(true);
2046 uint low_live_nodes = 0;
2047
2048 while(inlining_progress() && _late_inlines.length() > 0) {
2049
2050 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2051 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2052 // PhaseIdealLoop is expensive so we only try it once we are
2053 // out of live nodes and we only try it again if the previous
2054 // helped got the number of nodes down significantly
2055 PhaseIdealLoop ideal_loop( igvn, false, true );
2056 if (failing()) return;
2057 low_live_nodes = live_nodes();
2058 _major_progress = true;
2059 }
2060
2061 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2062 break;
2063 }
2064 }
2065
2066 inline_incrementally_one(igvn);
2067
2068 if (failing()) return;
2069
2070 igvn.optimize();
2071
2072 if (failing()) return;
2073 }
2074
2075 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2076
2077 if (_string_late_inlines.length() > 0) {
2078 assert(has_stringbuilder(), "inconsistent");
2079 for_igvn()->clear();
2080 initial_gvn()->replace_with(&igvn);
2081
2082 inline_string_calls(false);
2083
2084 if (failing()) return;
2085
2086 {
2087 ResourceMark rm;
2088 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2089 }
2090
2091 igvn = PhaseIterGVN(gvn);
2092
2093 igvn.optimize();
2094 }
2095
2096 set_inlining_incrementally(false);
2097 }
2098
2099
2100 // Remove edges from "root" to each SafePoint at a backward branch.
2101 // They were inserted during parsing (see add_safepoint()) to make
2102 // infinite loops without calls or exceptions visible to root, i.e.,
2103 // useful.
2104 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2105 Node *r = root();
2106 if (r != NULL) {
2107 for (uint i = r->req(); i < r->len(); ++i) {
2108 Node *n = r->in(i);
2109 if (n != NULL && n->is_SafePoint()) {
2110 r->rm_prec(i);
2111 if (n->outcnt() == 0) {
2112 igvn.remove_dead_node(n);
2113 }
2114 --i;
2115 }
2116 }
2117 }
2118 }
2119
2120 //------------------------------Optimize---------------------------------------
2121 // Given a graph, optimize it.
2122 void Compile::Optimize() {
2123 TracePhase t1("optimizer", &_t_optimizer, true);
2124
2125 #ifndef PRODUCT
2126 if (env()->break_at_compile()) {
2127 BREAKPOINT;
2128 }
2129
2130 #endif
2131
2132 ResourceMark rm;
2133 int loop_opts_cnt;
2134
2135 NOT_PRODUCT( verify_graph_edges(); )
2136
2137 print_method(PHASE_AFTER_PARSING);
2138
2139 {
2140 // Iterative Global Value Numbering, including ideal transforms
2141 // Initialize IterGVN with types and values from parse-time GVN
2142 PhaseIterGVN igvn(initial_gvn());
2143 {
2144 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2145 igvn.optimize();
2146 }
2147
2148 print_method(PHASE_ITER_GVN1, 2);
2149
2150 if (failing()) return;
2151
2152 {
2153 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2154 inline_incrementally(igvn);
2155 }
2156
2157 print_method(PHASE_INCREMENTAL_INLINE, 2);
2158
2159 if (failing()) return;
2160
2161 if (eliminate_boxing()) {
2162 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2163 // Inline valueOf() methods now.
2164 inline_boxing_calls(igvn);
2165
2166 if (AlwaysIncrementalInline) {
2167 inline_incrementally(igvn);
2168 }
2169
2170 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2171
2172 if (failing()) return;
2173 }
2174
2175 // Now that all inlining is over, cut edge from root to loop
2176 // safepoints
2177 remove_root_to_sfpts_edges(igvn);
2178
2179 // Remove the speculative part of types and clean up the graph from
2180 // the extra CastPP nodes whose only purpose is to carry them. Do
2181 // that early so that optimizations are not disrupted by the extra
2182 // CastPP nodes.
2183 remove_speculative_types(igvn);
2184
2185 // No more new expensive nodes will be added to the list from here
2186 // so keep only the actual candidates for optimizations.
2187 cleanup_expensive_nodes(igvn);
2188
2189 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2190 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2191 initial_gvn()->replace_with(&igvn);
2192 for_igvn()->clear();
2193 Unique_Node_List new_worklist(C->comp_arena());
2194 {
2195 ResourceMark rm;
2196 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2197 }
2198 set_for_igvn(&new_worklist);
2199 igvn = PhaseIterGVN(initial_gvn());
2200 igvn.optimize();
2201 }
2202
2203 // Perform escape analysis
2204 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2205 if (has_loops()) {
2206 // Cleanup graph (remove dead nodes).
2207 TracePhase t2("idealLoop", &_t_idealLoop, true);
2208 PhaseIdealLoop ideal_loop( igvn, false, true );
2209 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2210 if (failing()) return;
2211 }
2212 ConnectionGraph::do_analysis(this, &igvn);
2213
2214 if (failing()) return;
2215
2216 // Optimize out fields loads from scalar replaceable allocations.
2217 igvn.optimize();
2218 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2219
2220 if (failing()) return;
2221
2222 if (congraph() != NULL && macro_count() > 0) {
2223 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2224 PhaseMacroExpand mexp(igvn);
2225 mexp.eliminate_macro_nodes();
2226 igvn.set_delay_transform(false);
2227
2228 igvn.optimize();
2229 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2230
2231 if (failing()) return;
2232 }
2233 }
2234
2235 // Loop transforms on the ideal graph. Range Check Elimination,
2236 // peeling, unrolling, etc.
2237
2238 // Set loop opts counter
2239 loop_opts_cnt = num_loop_opts();
2240 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2241 {
2242 TracePhase t2("idealLoop", &_t_idealLoop, true);
2243 PhaseIdealLoop ideal_loop( igvn, true );
2244 loop_opts_cnt--;
2245 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2246 if (failing()) return;
2247 }
2248 // Loop opts pass if partial peeling occurred in previous pass
2249 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2250 TracePhase t3("idealLoop", &_t_idealLoop, true);
2251 PhaseIdealLoop ideal_loop( igvn, false );
2252 loop_opts_cnt--;
2253 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2254 if (failing()) return;
2255 }
2256 // Loop opts pass for loop-unrolling before CCP
2257 if(major_progress() && (loop_opts_cnt > 0)) {
2258 TracePhase t4("idealLoop", &_t_idealLoop, true);
2259 PhaseIdealLoop ideal_loop( igvn, false );
2260 loop_opts_cnt--;
2261 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2262 }
2263 if (!failing()) {
2264 // Verify that last round of loop opts produced a valid graph
2265 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2266 PhaseIdealLoop::verify(igvn);
2267 }
2268 }
2269 if (failing()) return;
2270
2271 // Conditional Constant Propagation;
2272 PhaseCCP ccp( &igvn );
2273 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2274 {
2275 TracePhase t2("ccp", &_t_ccp, true);
2276 ccp.do_transform();
2277 }
2278 print_method(PHASE_CPP1, 2);
2279
2280 assert( true, "Break here to ccp.dump_old2new_map()");
2281
2282 // Iterative Global Value Numbering, including ideal transforms
2283 {
2284 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2285 igvn = ccp;
2286 igvn.optimize();
2287 }
2288
2289 print_method(PHASE_ITER_GVN2, 2);
2290
2291 if (failing()) return;
2292
2293 // Loop transforms on the ideal graph. Range Check Elimination,
2294 // peeling, unrolling, etc.
2295 if(loop_opts_cnt > 0) {
2296 debug_only( int cnt = 0; );
2297 while(major_progress() && (loop_opts_cnt > 0)) {
2298 TracePhase t2("idealLoop", &_t_idealLoop, true);
2299 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2300 PhaseIdealLoop ideal_loop( igvn, true);
2301 loop_opts_cnt--;
2302 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2303 if (failing()) return;
2304 }
2305 }
2306
2307 {
2308 // Verify that all previous optimizations produced a valid graph
2309 // at least to this point, even if no loop optimizations were done.
2310 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2311 PhaseIdealLoop::verify(igvn);
2312 }
2313
2314 if (range_check_cast_count() > 0) {
2315 // No more loop optimizations. Remove all range check dependent CastIINodes.
2316 C->remove_range_check_casts(igvn);
2317 igvn.optimize();
2318 }
2319
2320 #ifdef ASSERT
2321 if (UseShenandoahGC && ShenandoahVerifyOptoBarriers) {
2322 ShenandoahBarrierC2Support::verify(C->root());
2323 }
2324 #endif
2325
2326 {
2327 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2328 PhaseMacroExpand mex(igvn);
2329 if (mex.expand_macro_nodes()) {
2330 assert(failing(), "must bail out w/ explicit message");
2331 return;
2332 }
2333 }
2334
2335 #if INCLUDE_ALL_GCS
2336 if (UseShenandoahGC) {
2337 ShenandoahBarrierC2Support::expand(this, igvn);
2338 }
2339 #endif
2340
2341 } // (End scope of igvn; run destructor if necessary for asserts.)
2342
2343 dump_inlining();
2344 // A method with only infinite loops has no edges entering loops from root
2345 {
2346 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2347 if (final_graph_reshaping()) {
2348 assert(failing(), "must bail out w/ explicit message");
2349 return;
2350 }
2351 }
2352
2353 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2354 }
2355
2356
2357 //------------------------------Code_Gen---------------------------------------
2358 // Given a graph, generate code for it
2359 void Compile::Code_Gen() {
2360 if (failing()) {
2361 return;
2362 }
2363
2364 // Perform instruction selection. You might think we could reclaim Matcher
2365 // memory PDQ, but actually the Matcher is used in generating spill code.
2366 // Internals of the Matcher (including some VectorSets) must remain live
2367 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2368 // set a bit in reclaimed memory.
2369
2370 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2371 // nodes. Mapping is only valid at the root of each matched subtree.
2372 NOT_PRODUCT( verify_graph_edges(); )
2373
2374 Matcher matcher;
2375 _matcher = &matcher;
2376 {
2377 TracePhase t2("matcher", &_t_matcher, true);
2378 matcher.match();
2379 }
2380 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2381 // nodes. Mapping is only valid at the root of each matched subtree.
2382 NOT_PRODUCT( verify_graph_edges(); )
2383
2384 // If you have too many nodes, or if matching has failed, bail out
2385 check_node_count(0, "out of nodes matching instructions");
2386 if (failing()) {
2387 return;
2388 }
2389
2390 // Build a proper-looking CFG
2391 PhaseCFG cfg(node_arena(), root(), matcher);
2392 _cfg = &cfg;
2393 {
2394 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2395 bool success = cfg.do_global_code_motion();
2396 if (!success) {
2397 return;
2398 }
2399
2400 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2401 NOT_PRODUCT( verify_graph_edges(); )
2402 debug_only( cfg.verify(); )
2403 }
2404
2405 PhaseChaitin regalloc(unique(), cfg, matcher);
2406 _regalloc = ®alloc;
2407 {
2408 TracePhase t2("regalloc", &_t_registerAllocation, true);
2409 // Perform register allocation. After Chaitin, use-def chains are
2410 // no longer accurate (at spill code) and so must be ignored.
2411 // Node->LRG->reg mappings are still accurate.
2412 _regalloc->Register_Allocate();
2413
2414 // Bail out if the allocator builds too many nodes
2415 if (failing()) {
2416 return;
2417 }
2418 }
2419
2420 // Prior to register allocation we kept empty basic blocks in case the
2421 // the allocator needed a place to spill. After register allocation we
2422 // are not adding any new instructions. If any basic block is empty, we
2423 // can now safely remove it.
2424 {
2425 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2426 cfg.remove_empty_blocks();
2427 if (do_freq_based_layout()) {
2428 PhaseBlockLayout layout(cfg);
2429 } else {
2430 cfg.set_loop_alignment();
2431 }
2432 cfg.fixup_flow();
2433 }
2434
2435 // Apply peephole optimizations
2436 if( OptoPeephole ) {
2437 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2438 PhasePeephole peep( _regalloc, cfg);
2439 peep.do_transform();
2440 }
2441
2442 // Do late expand if CPU requires this.
2443 if (Matcher::require_postalloc_expand) {
2444 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2445 cfg.postalloc_expand(_regalloc);
2446 }
2447
2448 // Convert Nodes to instruction bits in a buffer
2449 {
2450 // %%%% workspace merge brought two timers together for one job
2451 TracePhase t2a("output", &_t_output, true);
2452 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2453 Output();
2454 }
2455
2456 print_method(PHASE_FINAL_CODE);
2457
2458 // He's dead, Jim.
2459 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef);
2460 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2461 }
2462
2463
2464 //------------------------------dump_asm---------------------------------------
2465 // Dump formatted assembly
2466 #ifndef PRODUCT
2467 void Compile::dump_asm(int *pcs, uint pc_limit) {
2468 bool cut_short = false;
2469 tty->print_cr("#");
2470 tty->print("# "); _tf->dump(); tty->cr();
2471 tty->print_cr("#");
2472
2473 // For all blocks
2474 int pc = 0x0; // Program counter
2475 char starts_bundle = ' ';
2476 _regalloc->dump_frame();
2477
2478 Node *n = NULL;
2479 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2480 if (VMThread::should_terminate()) {
2481 cut_short = true;
2482 break;
2483 }
2484 Block* block = _cfg->get_block(i);
2485 if (block->is_connector() && !Verbose) {
2486 continue;
2487 }
2488 n = block->head();
2489 if (pcs && n->_idx < pc_limit) {
2490 tty->print("%3.3x ", pcs[n->_idx]);
2491 } else {
2492 tty->print(" ");
2493 }
2494 block->dump_head(_cfg);
2495 if (block->is_connector()) {
2496 tty->print_cr(" # Empty connector block");
2497 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2498 tty->print_cr(" # Block is sole successor of call");
2499 }
2500
2501 // For all instructions
2502 Node *delay = NULL;
2503 for (uint j = 0; j < block->number_of_nodes(); j++) {
2504 if (VMThread::should_terminate()) {
2505 cut_short = true;
2506 break;
2507 }
2508 n = block->get_node(j);
2509 if (valid_bundle_info(n)) {
2510 Bundle* bundle = node_bundling(n);
2511 if (bundle->used_in_unconditional_delay()) {
2512 delay = n;
2513 continue;
2514 }
2515 if (bundle->starts_bundle()) {
2516 starts_bundle = '+';
2517 }
2518 }
2519
2520 if (WizardMode) {
2521 n->dump();
2522 }
2523
2524 if( !n->is_Region() && // Dont print in the Assembly
2525 !n->is_Phi() && // a few noisely useless nodes
2526 !n->is_Proj() &&
2527 !n->is_MachTemp() &&
2528 !n->is_SafePointScalarObject() &&
2529 !n->is_Catch() && // Would be nice to print exception table targets
2530 !n->is_MergeMem() && // Not very interesting
2531 !n->is_top() && // Debug info table constants
2532 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2533 ) {
2534 if (pcs && n->_idx < pc_limit)
2535 tty->print("%3.3x", pcs[n->_idx]);
2536 else
2537 tty->print(" ");
2538 tty->print(" %c ", starts_bundle);
2539 starts_bundle = ' ';
2540 tty->print("\t");
2541 n->format(_regalloc, tty);
2542 tty->cr();
2543 }
2544
2545 // If we have an instruction with a delay slot, and have seen a delay,
2546 // then back up and print it
2547 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2548 assert(delay != NULL, "no unconditional delay instruction");
2549 if (WizardMode) delay->dump();
2550
2551 if (node_bundling(delay)->starts_bundle())
2552 starts_bundle = '+';
2553 if (pcs && n->_idx < pc_limit)
2554 tty->print("%3.3x", pcs[n->_idx]);
2555 else
2556 tty->print(" ");
2557 tty->print(" %c ", starts_bundle);
2558 starts_bundle = ' ';
2559 tty->print("\t");
2560 delay->format(_regalloc, tty);
2561 tty->cr();
2562 delay = NULL;
2563 }
2564
2565 // Dump the exception table as well
2566 if( n->is_Catch() && (Verbose || WizardMode) ) {
2567 // Print the exception table for this offset
2568 _handler_table.print_subtable_for(pc);
2569 }
2570 }
2571
2572 if (pcs && n->_idx < pc_limit)
2573 tty->print_cr("%3.3x", pcs[n->_idx]);
2574 else
2575 tty->cr();
2576
2577 assert(cut_short || delay == NULL, "no unconditional delay branch");
2578
2579 } // End of per-block dump
2580 tty->cr();
2581
2582 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2583 }
2584 #endif
2585
2586 //------------------------------Final_Reshape_Counts---------------------------
2587 // This class defines counters to help identify when a method
2588 // may/must be executed using hardware with only 24-bit precision.
2589 struct Final_Reshape_Counts : public StackObj {
2590 int _call_count; // count non-inlined 'common' calls
2591 int _float_count; // count float ops requiring 24-bit precision
2592 int _double_count; // count double ops requiring more precision
2593 int _java_call_count; // count non-inlined 'java' calls
2594 int _inner_loop_count; // count loops which need alignment
2595 VectorSet _visited; // Visitation flags
2596 Node_List _tests; // Set of IfNodes & PCTableNodes
2597
2598 Final_Reshape_Counts() :
2599 _call_count(0), _float_count(0), _double_count(0),
2600 _java_call_count(0), _inner_loop_count(0),
2601 _visited( Thread::current()->resource_area() ) { }
2602
2603 void inc_call_count () { _call_count ++; }
2604 void inc_float_count () { _float_count ++; }
2605 void inc_double_count() { _double_count++; }
2606 void inc_java_call_count() { _java_call_count++; }
2607 void inc_inner_loop_count() { _inner_loop_count++; }
2608
2609 int get_call_count () const { return _call_count ; }
2610 int get_float_count () const { return _float_count ; }
2611 int get_double_count() const { return _double_count; }
2612 int get_java_call_count() const { return _java_call_count; }
2613 int get_inner_loop_count() const { return _inner_loop_count; }
2614 };
2615
2616 #ifdef ASSERT
2617 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2618 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2619 // Make sure the offset goes inside the instance layout.
2620 return k->contains_field_offset(tp->offset());
2621 // Note that OffsetBot and OffsetTop are very negative.
2622 }
2623 #endif
2624
2625 // Eliminate trivially redundant StoreCMs and accumulate their
2626 // precedence edges.
2627 void Compile::eliminate_redundant_card_marks(Node* n) {
2628 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2629 if (n->in(MemNode::Address)->outcnt() > 1) {
2630 // There are multiple users of the same address so it might be
2631 // possible to eliminate some of the StoreCMs
2632 Node* mem = n->in(MemNode::Memory);
2633 Node* adr = n->in(MemNode::Address);
2634 Node* val = n->in(MemNode::ValueIn);
2635 Node* prev = n;
2636 bool done = false;
2637 // Walk the chain of StoreCMs eliminating ones that match. As
2638 // long as it's a chain of single users then the optimization is
2639 // safe. Eliminating partially redundant StoreCMs would require
2640 // cloning copies down the other paths.
2641 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2642 if (adr == mem->in(MemNode::Address) &&
2643 val == mem->in(MemNode::ValueIn)) {
2644 // redundant StoreCM
2645 if (mem->req() > MemNode::OopStore) {
2646 // Hasn't been processed by this code yet.
2647 n->add_prec(mem->in(MemNode::OopStore));
2648 } else {
2649 // Already converted to precedence edge
2650 for (uint i = mem->req(); i < mem->len(); i++) {
2651 // Accumulate any precedence edges
2652 if (mem->in(i) != NULL) {
2653 n->add_prec(mem->in(i));
2654 }
2655 }
2656 // Everything above this point has been processed.
2657 done = true;
2658 }
2659 // Eliminate the previous StoreCM
2660 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2661 assert(mem->outcnt() == 0, "should be dead");
2662 mem->disconnect_inputs(NULL, this);
2663 } else {
2664 prev = mem;
2665 }
2666 mem = prev->in(MemNode::Memory);
2667 }
2668 }
2669 }
2670
2671 //------------------------------final_graph_reshaping_impl----------------------
2672 // Implement items 1-5 from final_graph_reshaping below.
2673 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2674
2675 if ( n->outcnt() == 0 ) return; // dead node
2676 uint nop = n->Opcode();
2677
2678 // Check for 2-input instruction with "last use" on right input.
2679 // Swap to left input. Implements item (2).
2680 if( n->req() == 3 && // two-input instruction
2681 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2682 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2683 n->in(2)->outcnt() == 1 &&// right use IS a last use
2684 !n->in(2)->is_Con() ) { // right use is not a constant
2685 // Check for commutative opcode
2686 switch( nop ) {
2687 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2688 case Op_MaxI: case Op_MinI:
2689 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2690 case Op_AndL: case Op_XorL: case Op_OrL:
2691 case Op_AndI: case Op_XorI: case Op_OrI: {
2692 // Move "last use" input to left by swapping inputs
2693 n->swap_edges(1, 2);
2694 break;
2695 }
2696 default:
2697 break;
2698 }
2699 }
2700
2701 #ifdef ASSERT
2702 if( n->is_Mem() ) {
2703 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2704 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2705 // oop will be recorded in oop map if load crosses safepoint
2706 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2707 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2708 "raw memory operations should have control edge");
2709 }
2710 if (n->is_MemBar()) {
2711 MemBarNode* mb = n->as_MemBar();
2712 if (mb->trailing_store() || mb->trailing_load_store()) {
2713 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2714 Node* mem = mb->in(MemBarNode::Precedent);
2715 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2716 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2717 } else if (mb->leading()) {
2718 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2719 }
2720 }
2721 #endif
2722 // Count FPU ops and common calls, implements item (3)
2723 switch( nop ) {
2724 // Count all float operations that may use FPU
2725 case Op_AddF:
2726 case Op_SubF:
2727 case Op_MulF:
2728 case Op_DivF:
2729 case Op_NegF:
2730 case Op_ModF:
2731 case Op_ConvI2F:
2732 case Op_ConF:
2733 case Op_CmpF:
2734 case Op_CmpF3:
2735 // case Op_ConvL2F: // longs are split into 32-bit halves
2736 frc.inc_float_count();
2737 break;
2738
2739 case Op_ConvF2D:
2740 case Op_ConvD2F:
2741 frc.inc_float_count();
2742 frc.inc_double_count();
2743 break;
2744
2745 // Count all double operations that may use FPU
2746 case Op_AddD:
2747 case Op_SubD:
2748 case Op_MulD:
2749 case Op_DivD:
2750 case Op_NegD:
2751 case Op_ModD:
2752 case Op_ConvI2D:
2753 case Op_ConvD2I:
2754 // case Op_ConvL2D: // handled by leaf call
2755 // case Op_ConvD2L: // handled by leaf call
2756 case Op_ConD:
2757 case Op_CmpD:
2758 case Op_CmpD3:
2759 frc.inc_double_count();
2760 break;
2761 case Op_Opaque1: // Remove Opaque Nodes before matching
2762 case Op_Opaque2: // Remove Opaque Nodes before matching
2763 case Op_Opaque3:
2764 n->subsume_by(n->in(1), this);
2765 break;
2766 case Op_CallStaticJava:
2767 case Op_CallJava:
2768 case Op_CallDynamicJava:
2769 frc.inc_java_call_count(); // Count java call site;
2770 case Op_CallRuntime:
2771 case Op_CallLeaf:
2772 case Op_CallLeafNoFP: {
2773 assert( n->is_Call(), "" );
2774 CallNode *call = n->as_Call();
2775 if (UseShenandoahGC && call->is_g1_wb_pre_call()) {
2776 uint cnt = OptoRuntime::g1_wb_pre_Type()->domain()->cnt();
2777 if (call->req() > cnt) {
2778 assert(call->req() == cnt+1, "only one extra input");
2779 Node* addp = call->in(cnt);
2780 assert(!CallLeafNode::has_only_g1_wb_pre_uses(addp), "useless address computation?");
2781 call->del_req(cnt);
2782 }
2783 }
2784 // Count call sites where the FP mode bit would have to be flipped.
2785 // Do not count uncommon runtime calls:
2786 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2787 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2788 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2789 frc.inc_call_count(); // Count the call site
2790 } else { // See if uncommon argument is shared
2791 Node *n = call->in(TypeFunc::Parms);
2792 int nop = n->Opcode();
2793 // Clone shared simple arguments to uncommon calls, item (1).
2794 if( n->outcnt() > 1 &&
2795 !n->is_Proj() &&
2796 nop != Op_CreateEx &&
2797 nop != Op_CheckCastPP &&
2798 nop != Op_DecodeN &&
2799 nop != Op_DecodeNKlass &&
2800 !n->is_Mem() ) {
2801 Node *x = n->clone();
2802 call->set_req( TypeFunc::Parms, x );
2803 }
2804 }
2805 break;
2806 }
2807
2808 case Op_StoreD:
2809 case Op_LoadD:
2810 case Op_LoadD_unaligned:
2811 frc.inc_double_count();
2812 goto handle_mem;
2813 case Op_StoreF:
2814 case Op_LoadF:
2815 frc.inc_float_count();
2816 goto handle_mem;
2817
2818 case Op_StoreCM:
2819 {
2820 // Convert OopStore dependence into precedence edge
2821 Node* prec = n->in(MemNode::OopStore);
2822 n->del_req(MemNode::OopStore);
2823 n->add_prec(prec);
2824 eliminate_redundant_card_marks(n);
2825 }
2826
2827 // fall through
2828
2829 case Op_StoreB:
2830 case Op_StoreC:
2831 case Op_StorePConditional:
2832 case Op_StoreI:
2833 case Op_StoreL:
2834 case Op_StoreIConditional:
2835 case Op_StoreLConditional:
2836 case Op_CompareAndSwapI:
2837 case Op_CompareAndSwapL:
2838 case Op_CompareAndSwapP:
2839 case Op_CompareAndSwapN:
2840 case Op_GetAndAddI:
2841 case Op_GetAndAddL:
2842 case Op_GetAndSetI:
2843 case Op_GetAndSetL:
2844 case Op_GetAndSetP:
2845 case Op_GetAndSetN:
2846 case Op_StoreP:
2847 case Op_StoreN:
2848 case Op_StoreNKlass:
2849 case Op_LoadB:
2850 case Op_LoadUB:
2851 case Op_LoadUS:
2852 case Op_LoadI:
2853 case Op_LoadKlass:
2854 case Op_LoadNKlass:
2855 case Op_LoadL:
2856 case Op_LoadL_unaligned:
2857 case Op_LoadPLocked:
2858 case Op_LoadP:
2859 case Op_LoadN:
2860 case Op_LoadRange:
2861 case Op_LoadS: {
2862 handle_mem:
2863 #ifdef ASSERT
2864 if( VerifyOptoOopOffsets ) {
2865 assert( n->is_Mem(), "" );
2866 MemNode *mem = (MemNode*)n;
2867 // Check to see if address types have grounded out somehow.
2868 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2869 assert( !tp || oop_offset_is_sane(tp), "" );
2870 }
2871 #endif
2872 break;
2873 }
2874
2875 case Op_AddP: { // Assert sane base pointers
2876 Node *addp = n->in(AddPNode::Address);
2877 assert( !addp->is_AddP() ||
2878 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2879 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2880 "Base pointers must match" );
2881 #ifdef _LP64
2882 if ((UseCompressedOops || UseCompressedClassPointers) &&
2883 addp->Opcode() == Op_ConP &&
2884 addp == n->in(AddPNode::Base) &&
2885 n->in(AddPNode::Offset)->is_Con()) {
2886 // Use addressing with narrow klass to load with offset on x86.
2887 // On sparc loading 32-bits constant and decoding it have less
2888 // instructions (4) then load 64-bits constant (7).
2889 // Do this transformation here since IGVN will convert ConN back to ConP.
2890 const Type* t = addp->bottom_type();
2891 if (t->isa_oopptr() || t->isa_klassptr()) {
2892 Node* nn = NULL;
2893
2894 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2895
2896 // Look for existing ConN node of the same exact type.
2897 Node* r = root();
2898 uint cnt = r->outcnt();
2899 for (uint i = 0; i < cnt; i++) {
2900 Node* m = r->raw_out(i);
2901 if (m!= NULL && m->Opcode() == op &&
2902 m->bottom_type()->make_ptr() == t) {
2903 nn = m;
2904 break;
2905 }
2906 }
2907 if (nn != NULL) {
2908 // Decode a narrow oop to match address
2909 // [R12 + narrow_oop_reg<<3 + offset]
2910 if (t->isa_oopptr()) {
2911 nn = new (this) DecodeNNode(nn, t);
2912 } else {
2913 nn = new (this) DecodeNKlassNode(nn, t);
2914 }
2915 n->set_req(AddPNode::Base, nn);
2916 n->set_req(AddPNode::Address, nn);
2917 if (addp->outcnt() == 0) {
2918 addp->disconnect_inputs(NULL, this);
2919 }
2920 }
2921 }
2922 }
2923 #endif
2924 break;
2925 }
2926
2927 case Op_CastPP: {
2928 // Remove CastPP nodes to gain more freedom during scheduling but
2929 // keep the dependency they encode as control or precedence edges
2930 // (if control is set already) on memory operations. Some CastPP
2931 // nodes don't have a control (don't carry a dependency): skip
2932 // those.
2933 if (n->in(0) != NULL) {
2934 ResourceMark rm;
2935 Unique_Node_List wq;
2936 wq.push(n);
2937 for (uint next = 0; next < wq.size(); ++next) {
2938 Node *m = wq.at(next);
2939 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2940 Node* use = m->fast_out(i);
2941 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
2942 use->ensure_control_or_add_prec(n->in(0));
2943 } else if (use->in(0) == NULL) {
2944 switch(use->Opcode()) {
2945 case Op_AddP:
2946 case Op_DecodeN:
2947 case Op_DecodeNKlass:
2948 case Op_CheckCastPP:
2949 case Op_CastPP:
2950 wq.push(use);
2951 break;
2952 }
2953 }
2954 }
2955 }
2956 }
2957 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2958 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2959 Node* in1 = n->in(1);
2960 const Type* t = n->bottom_type();
2961 Node* new_in1 = in1->clone();
2962 new_in1->as_DecodeN()->set_type(t);
2963
2964 if (!Matcher::narrow_oop_use_complex_address()) {
2965 //
2966 // x86, ARM and friends can handle 2 adds in addressing mode
2967 // and Matcher can fold a DecodeN node into address by using
2968 // a narrow oop directly and do implicit NULL check in address:
2969 //
2970 // [R12 + narrow_oop_reg<<3 + offset]
2971 // NullCheck narrow_oop_reg
2972 //
2973 // On other platforms (Sparc) we have to keep new DecodeN node and
2974 // use it to do implicit NULL check in address:
2975 //
2976 // decode_not_null narrow_oop_reg, base_reg
2977 // [base_reg + offset]
2978 // NullCheck base_reg
2979 //
2980 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2981 // to keep the information to which NULL check the new DecodeN node
2982 // corresponds to use it as value in implicit_null_check().
2983 //
2984 new_in1->set_req(0, n->in(0));
2985 }
2986
2987 n->subsume_by(new_in1, this);
2988 if (in1->outcnt() == 0) {
2989 in1->disconnect_inputs(NULL, this);
2990 }
2991 } else {
2992 n->subsume_by(n->in(1), this);
2993 if (n->outcnt() == 0) {
2994 n->disconnect_inputs(NULL, this);
2995 }
2996 }
2997 break;
2998 }
2999 #ifdef _LP64
3000 case Op_CmpP:
3001 // Do this transformation here to preserve CmpPNode::sub() and
3002 // other TypePtr related Ideal optimizations (for example, ptr nullness).
3003 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3004 Node* in1 = n->in(1);
3005 Node* in2 = n->in(2);
3006 if (!in1->is_DecodeNarrowPtr()) {
3007 in2 = in1;
3008 in1 = n->in(2);
3009 }
3010 assert(in1->is_DecodeNarrowPtr(), "sanity");
3011
3012 Node* new_in2 = NULL;
3013 if (in2->is_DecodeNarrowPtr()) {
3014 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3015 new_in2 = in2->in(1);
3016 } else if (in2->Opcode() == Op_ConP) {
3017 const Type* t = in2->bottom_type();
3018 if (t == TypePtr::NULL_PTR) {
3019 assert(in1->is_DecodeN(), "compare klass to null?");
3020 // Don't convert CmpP null check into CmpN if compressed
3021 // oops implicit null check is not generated.
3022 // This will allow to generate normal oop implicit null check.
3023 if (Matcher::gen_narrow_oop_implicit_null_checks())
3024 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
3025 //
3026 // This transformation together with CastPP transformation above
3027 // will generated code for implicit NULL checks for compressed oops.
3028 //
3029 // The original code after Optimize()
3030 //
3031 // LoadN memory, narrow_oop_reg
3032 // decode narrow_oop_reg, base_reg
3033 // CmpP base_reg, NULL
3034 // CastPP base_reg // NotNull
3035 // Load [base_reg + offset], val_reg
3036 //
3037 // after these transformations will be
3038 //
3039 // LoadN memory, narrow_oop_reg
3040 // CmpN narrow_oop_reg, NULL
3041 // decode_not_null narrow_oop_reg, base_reg
3042 // Load [base_reg + offset], val_reg
3043 //
3044 // and the uncommon path (== NULL) will use narrow_oop_reg directly
3045 // since narrow oops can be used in debug info now (see the code in
3046 // final_graph_reshaping_walk()).
3047 //
3048 // At the end the code will be matched to
3049 // on x86:
3050 //
3051 // Load_narrow_oop memory, narrow_oop_reg
3052 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3053 // NullCheck narrow_oop_reg
3054 //
3055 // and on sparc:
3056 //
3057 // Load_narrow_oop memory, narrow_oop_reg
3058 // decode_not_null narrow_oop_reg, base_reg
3059 // Load [base_reg + offset], val_reg
3060 // NullCheck base_reg
3061 //
3062 } else if (t->isa_oopptr()) {
3063 new_in2 = ConNode::make(this, t->make_narrowoop());
3064 } else if (t->isa_klassptr()) {
3065 new_in2 = ConNode::make(this, t->make_narrowklass());
3066 }
3067 }
3068 if (new_in2 != NULL) {
3069 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
3070 n->subsume_by(cmpN, this);
3071 if (in1->outcnt() == 0) {
3072 in1->disconnect_inputs(NULL, this);
3073 }
3074 if (in2->outcnt() == 0) {
3075 in2->disconnect_inputs(NULL, this);
3076 }
3077 }
3078 }
3079 break;
3080
3081 case Op_DecodeN:
3082 case Op_DecodeNKlass:
3083 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3084 // DecodeN could be pinned when it can't be fold into
3085 // an address expression, see the code for Op_CastPP above.
3086 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3087 break;
3088
3089 case Op_EncodeP:
3090 case Op_EncodePKlass: {
3091 Node* in1 = n->in(1);
3092 if (in1->is_DecodeNarrowPtr()) {
3093 n->subsume_by(in1->in(1), this);
3094 } else if (in1->Opcode() == Op_ConP) {
3095 const Type* t = in1->bottom_type();
3096 if (t == TypePtr::NULL_PTR) {
3097 assert(t->isa_oopptr(), "null klass?");
3098 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
3099 } else if (t->isa_oopptr()) {
3100 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
3101 } else if (t->isa_klassptr()) {
3102 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
3103 }
3104 }
3105 if (in1->outcnt() == 0) {
3106 in1->disconnect_inputs(NULL, this);
3107 }
3108 break;
3109 }
3110
3111 case Op_Proj: {
3112 if (OptimizeStringConcat) {
3113 ProjNode* p = n->as_Proj();
3114 if (p->_is_io_use) {
3115 // Separate projections were used for the exception path which
3116 // are normally removed by a late inline. If it wasn't inlined
3117 // then they will hang around and should just be replaced with
3118 // the original one.
3119 Node* proj = NULL;
3120 // Replace with just one
3121 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3122 Node *use = i.get();
3123 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3124 proj = use;
3125 break;
3126 }
3127 }
3128 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3129 if (proj != NULL) {
3130 p->subsume_by(proj, this);
3131 }
3132 }
3133 }
3134 break;
3135 }
3136
3137 case Op_Phi:
3138 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3139 // The EncodeP optimization may create Phi with the same edges
3140 // for all paths. It is not handled well by Register Allocator.
3141 Node* unique_in = n->in(1);
3142 assert(unique_in != NULL, "");
3143 uint cnt = n->req();
3144 for (uint i = 2; i < cnt; i++) {
3145 Node* m = n->in(i);
3146 assert(m != NULL, "");
3147 if (unique_in != m)
3148 unique_in = NULL;
3149 }
3150 if (unique_in != NULL) {
3151 n->subsume_by(unique_in, this);
3152 }
3153 }
3154 break;
3155
3156 #endif
3157
3158 #ifdef ASSERT
3159 case Op_CastII:
3160 // Verify that all range check dependent CastII nodes were removed.
3161 if (n->isa_CastII()->has_range_check()) {
3162 n->dump(3);
3163 assert(false, "Range check dependent CastII node was not removed");
3164 }
3165 break;
3166 #endif
3167
3168 case Op_ModI:
3169 if (UseDivMod) {
3170 // Check if a%b and a/b both exist
3171 Node* d = n->find_similar(Op_DivI);
3172 if (d) {
3173 // Replace them with a fused divmod if supported
3174 if (Matcher::has_match_rule(Op_DivModI)) {
3175 DivModINode* divmod = DivModINode::make(this, n);
3176 d->subsume_by(divmod->div_proj(), this);
3177 n->subsume_by(divmod->mod_proj(), this);
3178 } else {
3179 // replace a%b with a-((a/b)*b)
3180 Node* mult = new (this) MulINode(d, d->in(2));
3181 Node* sub = new (this) SubINode(d->in(1), mult);
3182 n->subsume_by(sub, this);
3183 }
3184 }
3185 }
3186 break;
3187
3188 case Op_ModL:
3189 if (UseDivMod) {
3190 // Check if a%b and a/b both exist
3191 Node* d = n->find_similar(Op_DivL);
3192 if (d) {
3193 // Replace them with a fused divmod if supported
3194 if (Matcher::has_match_rule(Op_DivModL)) {
3195 DivModLNode* divmod = DivModLNode::make(this, n);
3196 d->subsume_by(divmod->div_proj(), this);
3197 n->subsume_by(divmod->mod_proj(), this);
3198 } else {
3199 // replace a%b with a-((a/b)*b)
3200 Node* mult = new (this) MulLNode(d, d->in(2));
3201 Node* sub = new (this) SubLNode(d->in(1), mult);
3202 n->subsume_by(sub, this);
3203 }
3204 }
3205 }
3206 break;
3207
3208 case Op_LoadVector:
3209 case Op_StoreVector:
3210 break;
3211
3212 case Op_PackB:
3213 case Op_PackS:
3214 case Op_PackI:
3215 case Op_PackF:
3216 case Op_PackL:
3217 case Op_PackD:
3218 if (n->req()-1 > 2) {
3219 // Replace many operand PackNodes with a binary tree for matching
3220 PackNode* p = (PackNode*) n;
3221 Node* btp = p->binary_tree_pack(this, 1, n->req());
3222 n->subsume_by(btp, this);
3223 }
3224 break;
3225 case Op_Loop:
3226 case Op_CountedLoop:
3227 if (n->as_Loop()->is_inner_loop()) {
3228 frc.inc_inner_loop_count();
3229 }
3230 break;
3231 case Op_LShiftI:
3232 case Op_RShiftI:
3233 case Op_URShiftI:
3234 case Op_LShiftL:
3235 case Op_RShiftL:
3236 case Op_URShiftL:
3237 if (Matcher::need_masked_shift_count) {
3238 // The cpu's shift instructions don't restrict the count to the
3239 // lower 5/6 bits. We need to do the masking ourselves.
3240 Node* in2 = n->in(2);
3241 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3242 const TypeInt* t = in2->find_int_type();
3243 if (t != NULL && t->is_con()) {
3244 juint shift = t->get_con();
3245 if (shift > mask) { // Unsigned cmp
3246 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3247 }
3248 } else {
3249 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3250 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3251 n->set_req(2, shift);
3252 }
3253 }
3254 if (in2->outcnt() == 0) { // Remove dead node
3255 in2->disconnect_inputs(NULL, this);
3256 }
3257 }
3258 break;
3259 case Op_MemBarStoreStore:
3260 case Op_MemBarRelease:
3261 // Break the link with AllocateNode: it is no longer useful and
3262 // confuses register allocation.
3263 if (n->req() > MemBarNode::Precedent) {
3264 n->set_req(MemBarNode::Precedent, top());
3265 }
3266 break;
3267 case Op_ShenandoahLoadReferenceBarrier:
3268 assert(false, "should have been expanded already");
3269 break;
3270 default:
3271 assert( !n->is_Call(), "" );
3272 assert( !n->is_Mem(), "" );
3273 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3274 break;
3275 }
3276
3277 // Collect CFG split points
3278 if (n->is_MultiBranch())
3279 frc._tests.push(n);
3280 }
3281
3282 //------------------------------final_graph_reshaping_walk---------------------
3283 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3284 // requires that the walk visits a node's inputs before visiting the node.
3285 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3286 ResourceArea *area = Thread::current()->resource_area();
3287 Unique_Node_List sfpt(area);
3288
3289 frc._visited.set(root->_idx); // first, mark node as visited
3290 uint cnt = root->req();
3291 Node *n = root;
3292 uint i = 0;
3293 while (true) {
3294 if (i < cnt) {
3295 // Place all non-visited non-null inputs onto stack
3296 Node* m = n->in(i);
3297 ++i;
3298 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3299 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3300 // compute worst case interpreter size in case of a deoptimization
3301 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3302
3303 sfpt.push(m);
3304 }
3305 cnt = m->req();
3306 nstack.push(n, i); // put on stack parent and next input's index
3307 n = m;
3308 i = 0;
3309 }
3310 } else {
3311 // Now do post-visit work
3312 final_graph_reshaping_impl( n, frc );
3313 if (nstack.is_empty())
3314 break; // finished
3315 n = nstack.node(); // Get node from stack
3316 cnt = n->req();
3317 i = nstack.index();
3318 nstack.pop(); // Shift to the next node on stack
3319 }
3320 }
3321
3322 // Skip next transformation if compressed oops are not used.
3323 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3324 (!UseCompressedOops && !UseCompressedClassPointers))
3325 return;
3326
3327 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3328 // It could be done for an uncommon traps or any safepoints/calls
3329 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3330 while (sfpt.size() > 0) {
3331 n = sfpt.pop();
3332 JVMState *jvms = n->as_SafePoint()->jvms();
3333 assert(jvms != NULL, "sanity");
3334 int start = jvms->debug_start();
3335 int end = n->req();
3336 bool is_uncommon = (n->is_CallStaticJava() &&
3337 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3338 for (int j = start; j < end; j++) {
3339 Node* in = n->in(j);
3340 if (in->is_DecodeNarrowPtr()) {
3341 bool safe_to_skip = true;
3342 if (!is_uncommon ) {
3343 // Is it safe to skip?
3344 for (uint i = 0; i < in->outcnt(); i++) {
3345 Node* u = in->raw_out(i);
3346 if (!u->is_SafePoint() ||
3347 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3348 safe_to_skip = false;
3349 }
3350 }
3351 }
3352 if (safe_to_skip) {
3353 n->set_req(j, in->in(1));
3354 }
3355 if (in->outcnt() == 0) {
3356 in->disconnect_inputs(NULL, this);
3357 }
3358 }
3359 }
3360 }
3361 }
3362
3363 //------------------------------final_graph_reshaping--------------------------
3364 // Final Graph Reshaping.
3365 //
3366 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3367 // and not commoned up and forced early. Must come after regular
3368 // optimizations to avoid GVN undoing the cloning. Clone constant
3369 // inputs to Loop Phis; these will be split by the allocator anyways.
3370 // Remove Opaque nodes.
3371 // (2) Move last-uses by commutative operations to the left input to encourage
3372 // Intel update-in-place two-address operations and better register usage
3373 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3374 // calls canonicalizing them back.
3375 // (3) Count the number of double-precision FP ops, single-precision FP ops
3376 // and call sites. On Intel, we can get correct rounding either by
3377 // forcing singles to memory (requires extra stores and loads after each
3378 // FP bytecode) or we can set a rounding mode bit (requires setting and
3379 // clearing the mode bit around call sites). The mode bit is only used
3380 // if the relative frequency of single FP ops to calls is low enough.
3381 // This is a key transform for SPEC mpeg_audio.
3382 // (4) Detect infinite loops; blobs of code reachable from above but not
3383 // below. Several of the Code_Gen algorithms fail on such code shapes,
3384 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3385 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3386 // Detection is by looking for IfNodes where only 1 projection is
3387 // reachable from below or CatchNodes missing some targets.
3388 // (5) Assert for insane oop offsets in debug mode.
3389
3390 bool Compile::final_graph_reshaping() {
3391 // an infinite loop may have been eliminated by the optimizer,
3392 // in which case the graph will be empty.
3393 if (root()->req() == 1) {
3394 record_method_not_compilable("trivial infinite loop");
3395 return true;
3396 }
3397
3398 // Expensive nodes have their control input set to prevent the GVN
3399 // from freely commoning them. There's no GVN beyond this point so
3400 // no need to keep the control input. We want the expensive nodes to
3401 // be freely moved to the least frequent code path by gcm.
3402 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3403 for (int i = 0; i < expensive_count(); i++) {
3404 _expensive_nodes->at(i)->set_req(0, NULL);
3405 }
3406
3407 Final_Reshape_Counts frc;
3408
3409 // Visit everybody reachable!
3410 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3411 Node_Stack nstack(live_nodes() >> 1);
3412 final_graph_reshaping_walk(nstack, root(), frc);
3413
3414 // Check for unreachable (from below) code (i.e., infinite loops).
3415 for( uint i = 0; i < frc._tests.size(); i++ ) {
3416 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3417 // Get number of CFG targets.
3418 // Note that PCTables include exception targets after calls.
3419 uint required_outcnt = n->required_outcnt();
3420 if (n->outcnt() != required_outcnt) {
3421 // Check for a few special cases. Rethrow Nodes never take the
3422 // 'fall-thru' path, so expected kids is 1 less.
3423 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3424 if (n->in(0)->in(0)->is_Call()) {
3425 CallNode *call = n->in(0)->in(0)->as_Call();
3426 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3427 required_outcnt--; // Rethrow always has 1 less kid
3428 } else if (call->req() > TypeFunc::Parms &&
3429 call->is_CallDynamicJava()) {
3430 // Check for null receiver. In such case, the optimizer has
3431 // detected that the virtual call will always result in a null
3432 // pointer exception. The fall-through projection of this CatchNode
3433 // will not be populated.
3434 Node *arg0 = call->in(TypeFunc::Parms);
3435 if (arg0->is_Type() &&
3436 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3437 required_outcnt--;
3438 }
3439 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3440 call->req() > TypeFunc::Parms+1 &&
3441 call->is_CallStaticJava()) {
3442 // Check for negative array length. In such case, the optimizer has
3443 // detected that the allocation attempt will always result in an
3444 // exception. There is no fall-through projection of this CatchNode .
3445 Node *arg1 = call->in(TypeFunc::Parms+1);
3446 if (arg1->is_Type() &&
3447 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3448 required_outcnt--;
3449 }
3450 }
3451 }
3452 }
3453 // Recheck with a better notion of 'required_outcnt'
3454 if (n->outcnt() != required_outcnt) {
3455 record_method_not_compilable("malformed control flow");
3456 return true; // Not all targets reachable!
3457 }
3458 }
3459 // Check that I actually visited all kids. Unreached kids
3460 // must be infinite loops.
3461 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3462 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3463 record_method_not_compilable("infinite loop");
3464 return true; // Found unvisited kid; must be unreach
3465 }
3466 }
3467
3468 // If original bytecodes contained a mixture of floats and doubles
3469 // check if the optimizer has made it homogenous, item (3).
3470 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3471 frc.get_float_count() > 32 &&
3472 frc.get_double_count() == 0 &&
3473 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3474 set_24_bit_selection_and_mode( false, true );
3475 }
3476
3477 set_java_calls(frc.get_java_call_count());
3478 set_inner_loops(frc.get_inner_loop_count());
3479
3480 // No infinite loops, no reason to bail out.
3481 return false;
3482 }
3483
3484 //-----------------------------too_many_traps----------------------------------
3485 // Report if there are too many traps at the current method and bci.
3486 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3487 bool Compile::too_many_traps(ciMethod* method,
3488 int bci,
3489 Deoptimization::DeoptReason reason) {
3490 ciMethodData* md = method->method_data();
3491 if (md->is_empty()) {
3492 // Assume the trap has not occurred, or that it occurred only
3493 // because of a transient condition during start-up in the interpreter.
3494 return false;
3495 }
3496 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3497 if (md->has_trap_at(bci, m, reason) != 0) {
3498 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3499 // Also, if there are multiple reasons, or if there is no per-BCI record,
3500 // assume the worst.
3501 if (log())
3502 log()->elem("observe trap='%s' count='%d'",
3503 Deoptimization::trap_reason_name(reason),
3504 md->trap_count(reason));
3505 return true;
3506 } else {
3507 // Ignore method/bci and see if there have been too many globally.
3508 return too_many_traps(reason, md);
3509 }
3510 }
3511
3512 // Less-accurate variant which does not require a method and bci.
3513 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3514 ciMethodData* logmd) {
3515 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3516 // Too many traps globally.
3517 // Note that we use cumulative trap_count, not just md->trap_count.
3518 if (log()) {
3519 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3520 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3521 Deoptimization::trap_reason_name(reason),
3522 mcount, trap_count(reason));
3523 }
3524 return true;
3525 } else {
3526 // The coast is clear.
3527 return false;
3528 }
3529 }
3530
3531 //--------------------------too_many_recompiles--------------------------------
3532 // Report if there are too many recompiles at the current method and bci.
3533 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3534 // Is not eager to return true, since this will cause the compiler to use
3535 // Action_none for a trap point, to avoid too many recompilations.
3536 bool Compile::too_many_recompiles(ciMethod* method,
3537 int bci,
3538 Deoptimization::DeoptReason reason) {
3539 ciMethodData* md = method->method_data();
3540 if (md->is_empty()) {
3541 // Assume the trap has not occurred, or that it occurred only
3542 // because of a transient condition during start-up in the interpreter.
3543 return false;
3544 }
3545 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3546 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3547 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3548 Deoptimization::DeoptReason per_bc_reason
3549 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3550 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3551 if ((per_bc_reason == Deoptimization::Reason_none
3552 || md->has_trap_at(bci, m, reason) != 0)
3553 // The trap frequency measure we care about is the recompile count:
3554 && md->trap_recompiled_at(bci, m)
3555 && md->overflow_recompile_count() >= bc_cutoff) {
3556 // Do not emit a trap here if it has already caused recompilations.
3557 // Also, if there are multiple reasons, or if there is no per-BCI record,
3558 // assume the worst.
3559 if (log())
3560 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3561 Deoptimization::trap_reason_name(reason),
3562 md->trap_count(reason),
3563 md->overflow_recompile_count());
3564 return true;
3565 } else if (trap_count(reason) != 0
3566 && decompile_count() >= m_cutoff) {
3567 // Too many recompiles globally, and we have seen this sort of trap.
3568 // Use cumulative decompile_count, not just md->decompile_count.
3569 if (log())
3570 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3571 Deoptimization::trap_reason_name(reason),
3572 md->trap_count(reason), trap_count(reason),
3573 md->decompile_count(), decompile_count());
3574 return true;
3575 } else {
3576 // The coast is clear.
3577 return false;
3578 }
3579 }
3580
3581 // Compute when not to trap. Used by matching trap based nodes and
3582 // NullCheck optimization.
3583 void Compile::set_allowed_deopt_reasons() {
3584 _allowed_reasons = 0;
3585 if (is_method_compilation()) {
3586 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3587 assert(rs < BitsPerInt, "recode bit map");
3588 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3589 _allowed_reasons |= nth_bit(rs);
3590 }
3591 }
3592 }
3593 }
3594
3595 #ifndef PRODUCT
3596 //------------------------------verify_graph_edges---------------------------
3597 // Walk the Graph and verify that there is a one-to-one correspondence
3598 // between Use-Def edges and Def-Use edges in the graph.
3599 void Compile::verify_graph_edges(bool no_dead_code) {
3600 if (VerifyGraphEdges) {
3601 ResourceArea *area = Thread::current()->resource_area();
3602 Unique_Node_List visited(area);
3603 // Call recursive graph walk to check edges
3604 _root->verify_edges(visited);
3605 if (no_dead_code) {
3606 // Now make sure that no visited node is used by an unvisited node.
3607 bool dead_nodes = false;
3608 Unique_Node_List checked(area);
3609 while (visited.size() > 0) {
3610 Node* n = visited.pop();
3611 checked.push(n);
3612 for (uint i = 0; i < n->outcnt(); i++) {
3613 Node* use = n->raw_out(i);
3614 if (checked.member(use)) continue; // already checked
3615 if (visited.member(use)) continue; // already in the graph
3616 if (use->is_Con()) continue; // a dead ConNode is OK
3617 // At this point, we have found a dead node which is DU-reachable.
3618 if (!dead_nodes) {
3619 tty->print_cr("*** Dead nodes reachable via DU edges:");
3620 dead_nodes = true;
3621 }
3622 use->dump(2);
3623 tty->print_cr("---");
3624 checked.push(use); // No repeats; pretend it is now checked.
3625 }
3626 }
3627 assert(!dead_nodes, "using nodes must be reachable from root");
3628 }
3629 }
3630 }
3631
3632 // Verify GC barriers consistency
3633 // Currently supported:
3634 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3635 void Compile::verify_barriers() {
3636 if (UseG1GC || UseShenandoahGC) {
3637 // Verify G1 pre-barriers
3638 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3639
3640 ResourceArea *area = Thread::current()->resource_area();
3641 Unique_Node_List visited(area);
3642 Node_List worklist(area);
3643 // We're going to walk control flow backwards starting from the Root
3644 worklist.push(_root);
3645 while (worklist.size() > 0) {
3646 Node* x = worklist.pop();
3647 if (x == NULL || x == top()) continue;
3648 if (visited.member(x)) {
3649 continue;
3650 } else {
3651 visited.push(x);
3652 }
3653
3654 if (x->is_Region()) {
3655 for (uint i = 1; i < x->req(); i++) {
3656 worklist.push(x->in(i));
3657 }
3658 } else {
3659 worklist.push(x->in(0));
3660 // We are looking for the pattern:
3661 // /->ThreadLocal
3662 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3663 // \->ConI(0)
3664 // We want to verify that the If and the LoadB have the same control
3665 // See GraphKit::g1_write_barrier_pre()
3666 if (x->is_If()) {
3667 IfNode *iff = x->as_If();
3668 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3669 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3670 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3671 && cmp->in(1)->is_Load()) {
3672 LoadNode* load = cmp->in(1)->as_Load();
3673 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3674 && load->in(2)->in(3)->is_Con()
3675 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3676
3677 Node* if_ctrl = iff->in(0);
3678 Node* load_ctrl = load->in(0);
3679
3680 if (if_ctrl != load_ctrl) {
3681 // Skip possible CProj->NeverBranch in infinite loops
3682 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3683 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3684 if_ctrl = if_ctrl->in(0)->in(0);
3685 }
3686 }
3687 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3688 }
3689 }
3690 }
3691 }
3692 }
3693 }
3694 }
3695 }
3696
3697 #endif
3698
3699 // The Compile object keeps track of failure reasons separately from the ciEnv.
3700 // This is required because there is not quite a 1-1 relation between the
3701 // ciEnv and its compilation task and the Compile object. Note that one
3702 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3703 // to backtrack and retry without subsuming loads. Other than this backtracking
3704 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3705 // by the logic in C2Compiler.
3706 void Compile::record_failure(const char* reason) {
3707 if (log() != NULL) {
3708 log()->elem("failure reason='%s' phase='compile'", reason);
3709 }
3710 if (_failure_reason == NULL) {
3711 // Record the first failure reason.
3712 _failure_reason = reason;
3713 }
3714
3715 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3716 C->print_method(PHASE_FAILURE);
3717 }
3718 _root = NULL; // flush the graph, too
3719 }
3720
3721 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3722 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3723 _phase_name(name), _dolog(dolog)
3724 {
3725 if (dolog) {
3726 C = Compile::current();
3727 _log = C->log();
3728 } else {
3729 C = NULL;
3730 _log = NULL;
3731 }
3732 if (_log != NULL) {
3733 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3734 _log->stamp();
3735 _log->end_head();
3736 }
3737 }
3738
3739 Compile::TracePhase::~TracePhase() {
3740
3741 C = Compile::current();
3742 if (_dolog) {
3743 _log = C->log();
3744 } else {
3745 _log = NULL;
3746 }
3747
3748 #ifdef ASSERT
3749 if (PrintIdealNodeCount) {
3750 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3751 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3752 }
3753
3754 if (VerifyIdealNodeCount) {
3755 Compile::current()->print_missing_nodes();
3756 }
3757 #endif
3758
3759 if (_log != NULL) {
3760 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3761 }
3762 }
3763
3764 //=============================================================================
3765 // Two Constant's are equal when the type and the value are equal.
3766 bool Compile::Constant::operator==(const Constant& other) {
3767 if (type() != other.type() ) return false;
3768 if (can_be_reused() != other.can_be_reused()) return false;
3769 // For floating point values we compare the bit pattern.
3770 switch (type()) {
3771 case T_FLOAT: return (_v._value.i == other._v._value.i);
3772 case T_LONG:
3773 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3774 case T_OBJECT:
3775 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3776 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3777 case T_METADATA: return (_v._metadata == other._v._metadata);
3778 default: ShouldNotReachHere();
3779 }
3780 return false;
3781 }
3782
3783 static int type_to_size_in_bytes(BasicType t) {
3784 switch (t) {
3785 case T_LONG: return sizeof(jlong );
3786 case T_FLOAT: return sizeof(jfloat );
3787 case T_DOUBLE: return sizeof(jdouble);
3788 case T_METADATA: return sizeof(Metadata*);
3789 // We use T_VOID as marker for jump-table entries (labels) which
3790 // need an internal word relocation.
3791 case T_VOID:
3792 case T_ADDRESS:
3793 case T_OBJECT: return sizeof(jobject);
3794 }
3795
3796 ShouldNotReachHere();
3797 return -1;
3798 }
3799
3800 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3801 // sort descending
3802 if (a->freq() > b->freq()) return -1;
3803 if (a->freq() < b->freq()) return 1;
3804 return 0;
3805 }
3806
3807 void Compile::ConstantTable::calculate_offsets_and_size() {
3808 // First, sort the array by frequencies.
3809 _constants.sort(qsort_comparator);
3810
3811 #ifdef ASSERT
3812 // Make sure all jump-table entries were sorted to the end of the
3813 // array (they have a negative frequency).
3814 bool found_void = false;
3815 for (int i = 0; i < _constants.length(); i++) {
3816 Constant con = _constants.at(i);
3817 if (con.type() == T_VOID)
3818 found_void = true; // jump-tables
3819 else
3820 assert(!found_void, "wrong sorting");
3821 }
3822 #endif
3823
3824 int offset = 0;
3825 for (int i = 0; i < _constants.length(); i++) {
3826 Constant* con = _constants.adr_at(i);
3827
3828 // Align offset for type.
3829 int typesize = type_to_size_in_bytes(con->type());
3830 offset = align_size_up(offset, typesize);
3831 con->set_offset(offset); // set constant's offset
3832
3833 if (con->type() == T_VOID) {
3834 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3835 offset = offset + typesize * n->outcnt(); // expand jump-table
3836 } else {
3837 offset = offset + typesize;
3838 }
3839 }
3840
3841 // Align size up to the next section start (which is insts; see
3842 // CodeBuffer::align_at_start).
3843 assert(_size == -1, "already set?");
3844 _size = align_size_up(offset, CodeEntryAlignment);
3845 }
3846
3847 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3848 MacroAssembler _masm(&cb);
3849 for (int i = 0; i < _constants.length(); i++) {
3850 Constant con = _constants.at(i);
3851 address constant_addr = NULL;
3852 switch (con.type()) {
3853 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3854 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3855 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3856 case T_OBJECT: {
3857 jobject obj = con.get_jobject();
3858 int oop_index = _masm.oop_recorder()->find_index(obj);
3859 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3860 break;
3861 }
3862 case T_ADDRESS: {
3863 address addr = (address) con.get_jobject();
3864 constant_addr = _masm.address_constant(addr);
3865 break;
3866 }
3867 // We use T_VOID as marker for jump-table entries (labels) which
3868 // need an internal word relocation.
3869 case T_VOID: {
3870 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3871 // Fill the jump-table with a dummy word. The real value is
3872 // filled in later in fill_jump_table.
3873 address dummy = (address) n;
3874 constant_addr = _masm.address_constant(dummy);
3875 // Expand jump-table
3876 for (uint i = 1; i < n->outcnt(); i++) {
3877 address temp_addr = _masm.address_constant(dummy + i);
3878 assert(temp_addr, "consts section too small");
3879 }
3880 break;
3881 }
3882 case T_METADATA: {
3883 Metadata* obj = con.get_metadata();
3884 int metadata_index = _masm.oop_recorder()->find_index(obj);
3885 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3886 break;
3887 }
3888 default: ShouldNotReachHere();
3889 }
3890 assert(constant_addr, "consts section too small");
3891 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3892 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3893 }
3894 }
3895
3896 int Compile::ConstantTable::find_offset(Constant& con) const {
3897 int idx = _constants.find(con);
3898 assert(idx != -1, "constant must be in constant table");
3899 int offset = _constants.at(idx).offset();
3900 assert(offset != -1, "constant table not emitted yet?");
3901 return offset;
3902 }
3903
3904 void Compile::ConstantTable::add(Constant& con) {
3905 if (con.can_be_reused()) {
3906 int idx = _constants.find(con);
3907 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3908 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3909 return;
3910 }
3911 }
3912 (void) _constants.append(con);
3913 }
3914
3915 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3916 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3917 Constant con(type, value, b->_freq);
3918 add(con);
3919 return con;
3920 }
3921
3922 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3923 Constant con(metadata);
3924 add(con);
3925 return con;
3926 }
3927
3928 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3929 jvalue value;
3930 BasicType type = oper->type()->basic_type();
3931 switch (type) {
3932 case T_LONG: value.j = oper->constantL(); break;
3933 case T_FLOAT: value.f = oper->constantF(); break;
3934 case T_DOUBLE: value.d = oper->constantD(); break;
3935 case T_OBJECT:
3936 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3937 case T_METADATA: return add((Metadata*)oper->constant()); break;
3938 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3939 }
3940 return add(n, type, value);
3941 }
3942
3943 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3944 jvalue value;
3945 // We can use the node pointer here to identify the right jump-table
3946 // as this method is called from Compile::Fill_buffer right before
3947 // the MachNodes are emitted and the jump-table is filled (means the
3948 // MachNode pointers do not change anymore).
3949 value.l = (jobject) n;
3950 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3951 add(con);
3952 return con;
3953 }
3954
3955 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3956 // If called from Compile::scratch_emit_size do nothing.
3957 if (Compile::current()->in_scratch_emit_size()) return;
3958
3959 assert(labels.is_nonempty(), "must be");
3960 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3961
3962 // Since MachConstantNode::constant_offset() also contains
3963 // table_base_offset() we need to subtract the table_base_offset()
3964 // to get the plain offset into the constant table.
3965 int offset = n->constant_offset() - table_base_offset();
3966
3967 MacroAssembler _masm(&cb);
3968 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3969
3970 for (uint i = 0; i < n->outcnt(); i++) {
3971 address* constant_addr = &jump_table_base[i];
3972 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3973 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3974 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3975 }
3976 }
3977
3978 void Compile::dump_inlining() {
3979 if (print_inlining() || print_intrinsics()) {
3980 // Print inlining message for candidates that we couldn't inline
3981 // for lack of space or non constant receiver
3982 for (int i = 0; i < _late_inlines.length(); i++) {
3983 CallGenerator* cg = _late_inlines.at(i);
3984 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3985 }
3986 Unique_Node_List useful;
3987 useful.push(root());
3988 for (uint next = 0; next < useful.size(); ++next) {
3989 Node* n = useful.at(next);
3990 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3991 CallNode* call = n->as_Call();
3992 CallGenerator* cg = call->generator();
3993 cg->print_inlining_late("receiver not constant");
3994 }
3995 uint max = n->len();
3996 for ( uint i = 0; i < max; ++i ) {
3997 Node *m = n->in(i);
3998 if ( m == NULL ) continue;
3999 useful.push(m);
4000 }
4001 }
4002 for (int i = 0; i < _print_inlining_list->length(); i++) {
4003 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4004 }
4005 }
4006 }
4007
4008 // Dump inlining replay data to the stream.
4009 // Don't change thread state and acquire any locks.
4010 void Compile::dump_inline_data(outputStream* out) {
4011 InlineTree* inl_tree = ilt();
4012 if (inl_tree != NULL) {
4013 out->print(" inline %d", inl_tree->count());
4014 inl_tree->dump_replay_data(out);
4015 }
4016 }
4017
4018 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4019 if (n1->Opcode() < n2->Opcode()) return -1;
4020 else if (n1->Opcode() > n2->Opcode()) return 1;
4021
4022 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
4023 for (uint i = 1; i < n1->req(); i++) {
4024 if (n1->in(i) < n2->in(i)) return -1;
4025 else if (n1->in(i) > n2->in(i)) return 1;
4026 }
4027
4028 return 0;
4029 }
4030
4031 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4032 Node* n1 = *n1p;
4033 Node* n2 = *n2p;
4034
4035 return cmp_expensive_nodes(n1, n2);
4036 }
4037
4038 void Compile::sort_expensive_nodes() {
4039 if (!expensive_nodes_sorted()) {
4040 _expensive_nodes->sort(cmp_expensive_nodes);
4041 }
4042 }
4043
4044 bool Compile::expensive_nodes_sorted() const {
4045 for (int i = 1; i < _expensive_nodes->length(); i++) {
4046 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4047 return false;
4048 }
4049 }
4050 return true;
4051 }
4052
4053 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4054 if (_expensive_nodes->length() == 0) {
4055 return false;
4056 }
4057
4058 assert(OptimizeExpensiveOps, "optimization off?");
4059
4060 // Take this opportunity to remove dead nodes from the list
4061 int j = 0;
4062 for (int i = 0; i < _expensive_nodes->length(); i++) {
4063 Node* n = _expensive_nodes->at(i);
4064 if (!n->is_unreachable(igvn)) {
4065 assert(n->is_expensive(), "should be expensive");
4066 _expensive_nodes->at_put(j, n);
4067 j++;
4068 }
4069 }
4070 _expensive_nodes->trunc_to(j);
4071
4072 // Then sort the list so that similar nodes are next to each other
4073 // and check for at least two nodes of identical kind with same data
4074 // inputs.
4075 sort_expensive_nodes();
4076
4077 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4078 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4079 return true;
4080 }
4081 }
4082
4083 return false;
4084 }
4085
4086 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4087 if (_expensive_nodes->length() == 0) {
4088 return;
4089 }
4090
4091 assert(OptimizeExpensiveOps, "optimization off?");
4092
4093 // Sort to bring similar nodes next to each other and clear the
4094 // control input of nodes for which there's only a single copy.
4095 sort_expensive_nodes();
4096
4097 int j = 0;
4098 int identical = 0;
4099 int i = 0;
4100 for (; i < _expensive_nodes->length()-1; i++) {
4101 assert(j <= i, "can't write beyond current index");
4102 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4103 identical++;
4104 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4105 continue;
4106 }
4107 if (identical > 0) {
4108 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4109 identical = 0;
4110 } else {
4111 Node* n = _expensive_nodes->at(i);
4112 igvn.hash_delete(n);
4113 n->set_req(0, NULL);
4114 igvn.hash_insert(n);
4115 }
4116 }
4117 if (identical > 0) {
4118 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4119 } else if (_expensive_nodes->length() >= 1) {
4120 Node* n = _expensive_nodes->at(i);
4121 igvn.hash_delete(n);
4122 n->set_req(0, NULL);
4123 igvn.hash_insert(n);
4124 }
4125 _expensive_nodes->trunc_to(j);
4126 }
4127
4128 void Compile::add_expensive_node(Node * n) {
4129 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4130 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4131 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4132 if (OptimizeExpensiveOps) {
4133 _expensive_nodes->append(n);
4134 } else {
4135 // Clear control input and let IGVN optimize expensive nodes if
4136 // OptimizeExpensiveOps is off.
4137 n->set_req(0, NULL);
4138 }
4139 }
4140
4141 /**
4142 * Remove the speculative part of types and clean up the graph
4143 */
4144 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4145 if (UseTypeSpeculation) {
4146 Unique_Node_List worklist;
4147 worklist.push(root());
4148 int modified = 0;
4149 // Go over all type nodes that carry a speculative type, drop the
4150 // speculative part of the type and enqueue the node for an igvn
4151 // which may optimize it out.
4152 for (uint next = 0; next < worklist.size(); ++next) {
4153 Node *n = worklist.at(next);
4154 if (n->is_Type()) {
4155 TypeNode* tn = n->as_Type();
4156 const Type* t = tn->type();
4157 const Type* t_no_spec = t->remove_speculative();
4158 if (t_no_spec != t) {
4159 bool in_hash = igvn.hash_delete(n);
4160 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4161 tn->set_type(t_no_spec);
4162 igvn.hash_insert(n);
4163 igvn._worklist.push(n); // give it a chance to go away
4164 modified++;
4165 }
4166 }
4167 uint max = n->len();
4168 for( uint i = 0; i < max; ++i ) {
4169 Node *m = n->in(i);
4170 if (not_a_node(m)) continue;
4171 worklist.push(m);
4172 }
4173 }
4174 // Drop the speculative part of all types in the igvn's type table
4175 igvn.remove_speculative_types();
4176 if (modified > 0) {
4177 igvn.optimize();
4178 }
4179 #ifdef ASSERT
4180 // Verify that after the IGVN is over no speculative type has resurfaced
4181 worklist.clear();
4182 worklist.push(root());
4183 for (uint next = 0; next < worklist.size(); ++next) {
4184 Node *n = worklist.at(next);
4185 const Type* t = igvn.type_or_null(n);
4186 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4187 if (n->is_Type()) {
4188 t = n->as_Type()->type();
4189 assert(t == t->remove_speculative(), "no more speculative types");
4190 }
4191 uint max = n->len();
4192 for( uint i = 0; i < max; ++i ) {
4193 Node *m = n->in(i);
4194 if (not_a_node(m)) continue;
4195 worklist.push(m);
4196 }
4197 }
4198 igvn.check_no_speculative_types();
4199 #endif
4200 }
4201 }
4202
4203 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4204 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4205 if (ctrl != NULL) {
4206 // Express control dependency by a CastII node with a narrow type.
4207 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4208 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4209 // node from floating above the range check during loop optimizations. Otherwise, the
4210 // ConvI2L node may be eliminated independently of the range check, causing the data path
4211 // to become TOP while the control path is still there (although it's unreachable).
4212 value->set_req(0, ctrl);
4213 // Save CastII node to remove it after loop optimizations.
4214 phase->C->add_range_check_cast(value);
4215 value = phase->transform(value);
4216 }
4217 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4218 return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4219 }
4220
4221 // Auxiliary method to support randomized stressing/fuzzing.
4222 //
4223 // This method can be called the arbitrary number of times, with current count
4224 // as the argument. The logic allows selecting a single candidate from the
4225 // running list of candidates as follows:
4226 // int count = 0;
4227 // Cand* selected = null;
4228 // while(cand = cand->next()) {
4229 // if (randomized_select(++count)) {
4230 // selected = cand;
4231 // }
4232 // }
4233 //
4234 // Including count equalizes the chances any candidate is "selected".
4235 // This is useful when we don't have the complete list of candidates to choose
4236 // from uniformly. In this case, we need to adjust the randomicity of the
4237 // selection, or else we will end up biasing the selection towards the latter
4238 // candidates.
4239 //
4240 // Quick back-envelope calculation shows that for the list of n candidates
4241 // the equal probability for the candidate to persist as "best" can be
4242 // achieved by replacing it with "next" k-th candidate with the probability
4243 // of 1/k. It can be easily shown that by the end of the run, the
4244 // probability for any candidate is converged to 1/n, thus giving the
4245 // uniform distribution among all the candidates.
4246 //
4247 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4248 #define RANDOMIZED_DOMAIN_POW 29
4249 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4250 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4251 bool Compile::randomized_select(int count) {
4252 assert(count > 0, "only positive");
4253 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4254 }
4255
4256 void Compile::shenandoah_eliminate_g1_wb_pre(Node* call, PhaseIterGVN* igvn) {
4257 assert(UseShenandoahGC && call->is_g1_wb_pre_call(), "");
4258 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
4259 c = c->unique_ctrl_out();
4260 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4261 c = c->unique_ctrl_out();
4262 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4263 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4264 assert(iff->is_If(), "expect test");
4265 if (!iff->is_shenandoah_marking_if(igvn)) {
4266 c = c->unique_ctrl_out();
4267 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4268 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4269 assert(iff->is_shenandoah_marking_if(igvn), "expect marking test");
4270 }
4271 Node* cmpx = iff->in(1)->in(1);
4272 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
4273 igvn->rehash_node_delayed(call);
4274 call->del_req(call->req()-1);
4275 }