1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciReplay.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/exceptionHandlerTable.hpp"
31 #include "code/nmethod.hpp"
32 #include "compiler/compileBroker.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "compiler/disassembler.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "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/convertnode.hpp"
46 #include "opto/divnode.hpp"
47 #include "opto/escape.hpp"
48 #include "opto/idealGraphPrinter.hpp"
49 #include "opto/loopnode.hpp"
50 #include "opto/machnode.hpp"
51 #include "opto/macro.hpp"
52 #include "opto/matcher.hpp"
53 #include "opto/mathexactnode.hpp"
54 #include "opto/memnode.hpp"
55 #include "opto/mulnode.hpp"
56 #include "opto/narrowptrnode.hpp"
57 #include "opto/node.hpp"
58 #include "opto/opcodes.hpp"
59 #include "opto/output.hpp"
60 #include "opto/parse.hpp"
61 #include "opto/phaseX.hpp"
62 #include "opto/rootnode.hpp"
63 #include "opto/runtime.hpp"
64 #include "opto/stringopts.hpp"
65 #include "opto/type.hpp"
66 #include "opto/vectornode.hpp"
67 #include "runtime/arguments.hpp"
68 #include "runtime/sharedRuntime.hpp"
69 #include "runtime/signature.hpp"
70 #include "runtime/stubRoutines.hpp"
71 #include "runtime/timer.hpp"
72 #include "utilities/copy.hpp"
73
74
75 // -------------------- Compile::mach_constant_base_node -----------------------
76 // Constant table base node singleton.
77 MachConstantBaseNode* Compile::mach_constant_base_node() {
78 if (_mach_constant_base_node == NULL) {
79 _mach_constant_base_node = new MachConstantBaseNode();
80 _mach_constant_base_node->add_req(C->root());
81 }
82 return _mach_constant_base_node;
83 }
84
85
86 /// Support for intrinsics.
87
88 // Return the index at which m must be inserted (or already exists).
89 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
91 #ifdef ASSERT
92 for (int i = 1; i < _intrinsics->length(); i++) {
93 CallGenerator* cg1 = _intrinsics->at(i-1);
94 CallGenerator* cg2 = _intrinsics->at(i);
95 assert(cg1->method() != cg2->method()
96 ? cg1->method() < cg2->method()
97 : cg1->is_virtual() < cg2->is_virtual(),
98 "compiler intrinsics list must stay sorted");
99 }
100 #endif
101 // Binary search sorted list, in decreasing intervals [lo, hi].
102 int lo = 0, hi = _intrinsics->length()-1;
103 while (lo <= hi) {
104 int mid = (uint)(hi + lo) / 2;
105 ciMethod* mid_m = _intrinsics->at(mid)->method();
106 if (m < mid_m) {
107 hi = mid-1;
108 } else if (m > mid_m) {
109 lo = mid+1;
110 } else {
111 // look at minor sort key
112 bool mid_virt = _intrinsics->at(mid)->is_virtual();
113 if (is_virtual < mid_virt) {
114 hi = mid-1;
115 } else if (is_virtual > mid_virt) {
116 lo = mid+1;
117 } else {
118 return mid; // exact match
119 }
120 }
121 }
122 return lo; // inexact match
123 }
124
125 void Compile::register_intrinsic(CallGenerator* cg) {
126 if (_intrinsics == NULL) {
127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
128 }
129 // This code is stolen from ciObjectFactory::insert.
130 // Really, GrowableArray should have methods for
131 // insert_at, remove_at, and binary_search.
132 int len = _intrinsics->length();
133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
134 if (index == len) {
135 _intrinsics->append(cg);
136 } else {
137 #ifdef ASSERT
138 CallGenerator* oldcg = _intrinsics->at(index);
139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
140 #endif
141 _intrinsics->append(_intrinsics->at(len-1));
142 int pos;
143 for (pos = len-2; pos >= index; pos--) {
144 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
145 }
146 _intrinsics->at_put(index, cg);
147 }
148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
149 }
150
151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
152 assert(m->is_loaded(), "don't try this on unloaded methods");
153 if (_intrinsics != NULL) {
154 int index = intrinsic_insertion_index(m, is_virtual);
155 if (index < _intrinsics->length()
156 && _intrinsics->at(index)->method() == m
157 && _intrinsics->at(index)->is_virtual() == is_virtual) {
158 return _intrinsics->at(index);
159 }
160 }
161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
162 if (m->intrinsic_id() != vmIntrinsics::_none &&
163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
165 if (cg != NULL) {
166 // Save it for next time:
167 register_intrinsic(cg);
168 return cg;
169 } else {
170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
171 }
172 }
173 return NULL;
174 }
175
176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
177 // in library_call.cpp.
178
179
180 #ifndef PRODUCT
181 // statistics gathering...
182
183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
185
186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
188 int oflags = _intrinsic_hist_flags[id];
189 assert(flags != 0, "what happened?");
190 if (is_virtual) {
191 flags |= _intrinsic_virtual;
192 }
193 bool changed = (flags != oflags);
194 if ((flags & _intrinsic_worked) != 0) {
195 juint count = (_intrinsic_hist_count[id] += 1);
196 if (count == 1) {
197 changed = true; // first time
198 }
199 // increment the overall count also:
200 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
201 }
202 if (changed) {
203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
204 // Something changed about the intrinsic's virtuality.
205 if ((flags & _intrinsic_virtual) != 0) {
206 // This is the first use of this intrinsic as a virtual call.
207 if (oflags != 0) {
208 // We already saw it as a non-virtual, so note both cases.
209 flags |= _intrinsic_both;
210 }
211 } else if ((oflags & _intrinsic_both) == 0) {
212 // This is the first use of this intrinsic as a non-virtual
213 flags |= _intrinsic_both;
214 }
215 }
216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
217 }
218 // update the overall flags also:
219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
220 return changed;
221 }
222
223 static char* format_flags(int flags, char* buf) {
224 buf[0] = 0;
225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
230 if (buf[0] == 0) strcat(buf, ",");
231 assert(buf[0] == ',', "must be");
232 return &buf[1];
233 }
234
235 void Compile::print_intrinsic_statistics() {
236 char flagsbuf[100];
237 ttyLocker ttyl;
238 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
239 tty->print_cr("Compiler intrinsic usage:");
240 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
241 if (total == 0) total = 1; // avoid div0 in case of no successes
242 #define PRINT_STAT_LINE(name, c, f) \
243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
245 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
246 int flags = _intrinsic_hist_flags[id];
247 juint count = _intrinsic_hist_count[id];
248 if ((flags | count) != 0) {
249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
250 }
251 }
252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
253 if (xtty != NULL) xtty->tail("statistics");
254 }
255
256 void Compile::print_statistics() {
257 { ttyLocker ttyl;
258 if (xtty != NULL) xtty->head("statistics type='opto'");
259 Parse::print_statistics();
260 PhaseCCP::print_statistics();
261 PhaseRegAlloc::print_statistics();
262 Scheduling::print_statistics();
263 PhasePeephole::print_statistics();
264 PhaseIdealLoop::print_statistics();
265 if (xtty != NULL) xtty->tail("statistics");
266 }
267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
268 // put this under its own <statistics> element.
269 print_intrinsic_statistics();
270 }
271 }
272 #endif //PRODUCT
273
274 // Support for bundling info
275 Bundle* Compile::node_bundling(const Node *n) {
276 assert(valid_bundle_info(n), "oob");
277 return &_node_bundling_base[n->_idx];
278 }
279
280 bool Compile::valid_bundle_info(const Node *n) {
281 return (_node_bundling_limit > n->_idx);
282 }
283
284
285 void Compile::gvn_replace_by(Node* n, Node* nn) {
286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
287 Node* use = n->last_out(i);
288 bool is_in_table = initial_gvn()->hash_delete(use);
289 uint uses_found = 0;
290 for (uint j = 0; j < use->len(); j++) {
291 if (use->in(j) == n) {
292 if (j < use->req())
293 use->set_req(j, nn);
294 else
295 use->set_prec(j, nn);
296 uses_found++;
297 }
298 }
299 if (is_in_table) {
300 // reinsert into table
301 initial_gvn()->hash_find_insert(use);
302 }
303 record_for_igvn(use);
304 i -= uses_found; // we deleted 1 or more copies of this edge
305 }
306 }
307
308
309 static inline bool not_a_node(const Node* n) {
310 if (n == NULL) return true;
311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
312 if (*(address*)n == badAddress) return true; // kill by Node::destruct
313 return false;
314 }
315
316 // Identify all nodes that are reachable from below, useful.
317 // Use breadth-first pass that records state in a Unique_Node_List,
318 // recursive traversal is slower.
319 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
320 int estimated_worklist_size = live_nodes();
321 useful.map( estimated_worklist_size, NULL ); // preallocate space
322
323 // Initialize worklist
324 if (root() != NULL) { useful.push(root()); }
325 // If 'top' is cached, declare it useful to preserve cached node
326 if( cached_top_node() ) { useful.push(cached_top_node()); }
327
328 // Push all useful nodes onto the list, breadthfirst
329 for( uint next = 0; next < useful.size(); ++next ) {
330 assert( next < unique(), "Unique useful nodes < total nodes");
331 Node *n = useful.at(next);
332 uint max = n->len();
333 for( uint i = 0; i < max; ++i ) {
334 Node *m = n->in(i);
335 if (not_a_node(m)) continue;
336 useful.push(m);
337 }
338 }
339 }
340
341 // Update dead_node_list with any missing dead nodes using useful
342 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
343 void Compile::update_dead_node_list(Unique_Node_List &useful) {
344 uint max_idx = unique();
345 VectorSet& useful_node_set = useful.member_set();
346
347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
348 // If node with index node_idx is not in useful set,
349 // mark it as dead in dead node list.
350 if (! useful_node_set.test(node_idx) ) {
351 record_dead_node(node_idx);
352 }
353 }
354 }
355
356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
357 int shift = 0;
358 for (int i = 0; i < inlines->length(); i++) {
359 CallGenerator* cg = inlines->at(i);
360 CallNode* call = cg->call_node();
361 if (shift > 0) {
362 inlines->at_put(i-shift, cg);
363 }
364 if (!useful.member(call)) {
365 shift++;
366 }
367 }
368 inlines->trunc_to(inlines->length()-shift);
369 }
370
371 // Disconnect all useless nodes by disconnecting those at the boundary.
372 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
373 uint next = 0;
374 while (next < useful.size()) {
375 Node *n = useful.at(next++);
376 if (n->is_SafePoint()) {
377 // We're done with a parsing phase. Replaced nodes are not valid
378 // beyond that point.
379 n->as_SafePoint()->delete_replaced_nodes();
380 }
381 // Use raw traversal of out edges since this code removes out edges
382 int max = n->outcnt();
383 for (int j = 0; j < max; ++j) {
384 Node* child = n->raw_out(j);
385 if (! useful.member(child)) {
386 assert(!child->is_top() || child != top(),
387 "If top is cached in Compile object it is in useful list");
388 // Only need to remove this out-edge to the useless node
389 n->raw_del_out(j);
390 --j;
391 --max;
392 }
393 }
394 if (n->outcnt() == 1 && n->has_special_unique_user()) {
395 record_for_igvn(n->unique_out());
396 }
397 }
398 // Remove useless macro and predicate opaq nodes
399 for (int i = C->macro_count()-1; i >= 0; i--) {
400 Node* n = C->macro_node(i);
401 if (!useful.member(n)) {
402 remove_macro_node(n);
403 }
404 }
405 // Remove useless expensive node
406 for (int i = C->expensive_count()-1; i >= 0; i--) {
407 Node* n = C->expensive_node(i);
408 if (!useful.member(n)) {
409 remove_expensive_node(n);
410 }
411 }
412 // clean up the late inline lists
413 remove_useless_late_inlines(&_string_late_inlines, useful);
414 remove_useless_late_inlines(&_boxing_late_inlines, useful);
415 remove_useless_late_inlines(&_late_inlines, useful);
416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
417 }
418
419 //------------------------------frame_size_in_words-----------------------------
420 // frame_slots in units of words
421 int Compile::frame_size_in_words() const {
422 // shift is 0 in LP32 and 1 in LP64
423 const int shift = (LogBytesPerWord - LogBytesPerInt);
424 int words = _frame_slots >> shift;
425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
426 return words;
427 }
428
429 // To bang the stack of this compiled method we use the stack size
430 // that the interpreter would need in case of a deoptimization. This
431 // removes the need to bang the stack in the deoptimization blob which
432 // in turn simplifies stack overflow handling.
433 int Compile::bang_size_in_bytes() const {
434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
435 }
436
437 // ============================================================================
438 //------------------------------CompileWrapper---------------------------------
439 class CompileWrapper : public StackObj {
440 Compile *const _compile;
441 public:
442 CompileWrapper(Compile* compile);
443
444 ~CompileWrapper();
445 };
446
447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
448 // the Compile* pointer is stored in the current ciEnv:
449 ciEnv* env = compile->env();
450 assert(env == ciEnv::current(), "must already be a ciEnv active");
451 assert(env->compiler_data() == NULL, "compile already active?");
452 env->set_compiler_data(compile);
453 assert(compile == Compile::current(), "sanity");
454
455 compile->set_type_dict(NULL);
456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
457 compile->clone_map().set_clone_idx(0);
458 compile->set_type_hwm(NULL);
459 compile->set_type_last_size(0);
460 compile->set_last_tf(NULL, NULL);
461 compile->set_indexSet_arena(NULL);
462 compile->set_indexSet_free_block_list(NULL);
463 compile->init_type_arena();
464 Type::Initialize(compile);
465 _compile->set_scratch_buffer_blob(NULL);
466 _compile->begin_method();
467 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
468 }
469 CompileWrapper::~CompileWrapper() {
470 _compile->end_method();
471 if (_compile->scratch_buffer_blob() != NULL)
472 BufferBlob::free(_compile->scratch_buffer_blob());
473 _compile->env()->set_compiler_data(NULL);
474 }
475
476
477 //----------------------------print_compile_messages---------------------------
478 void Compile::print_compile_messages() {
479 #ifndef PRODUCT
480 // Check if recompiling
481 if (_subsume_loads == false && PrintOpto) {
482 // Recompiling without allowing machine instructions to subsume loads
483 tty->print_cr("*********************************************************");
484 tty->print_cr("** Bailout: Recompile without subsuming loads **");
485 tty->print_cr("*********************************************************");
486 }
487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
488 // Recompiling without escape analysis
489 tty->print_cr("*********************************************************");
490 tty->print_cr("** Bailout: Recompile without escape analysis **");
491 tty->print_cr("*********************************************************");
492 }
493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
494 // Recompiling without boxing elimination
495 tty->print_cr("*********************************************************");
496 tty->print_cr("** Bailout: Recompile without boxing elimination **");
497 tty->print_cr("*********************************************************");
498 }
499 if (C->directive()->BreakAtCompileOption) {
500 // Open the debugger when compiling this method.
501 tty->print("### Breaking when compiling: ");
502 method()->print_short_name();
503 tty->cr();
504 BREAKPOINT;
505 }
506
507 if( PrintOpto ) {
508 if (is_osr_compilation()) {
509 tty->print("[OSR]%3d", _compile_id);
510 } else {
511 tty->print("%3d", _compile_id);
512 }
513 }
514 #endif
515 }
516
517
518 //-----------------------init_scratch_buffer_blob------------------------------
519 // Construct a temporary BufferBlob and cache it for this compile.
520 void Compile::init_scratch_buffer_blob(int const_size) {
521 // If there is already a scratch buffer blob allocated and the
522 // constant section is big enough, use it. Otherwise free the
523 // current and allocate a new one.
524 BufferBlob* blob = scratch_buffer_blob();
525 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
526 // Use the current blob.
527 } else {
528 if (blob != NULL) {
529 BufferBlob::free(blob);
530 }
531
532 ResourceMark rm;
533 _scratch_const_size = const_size;
534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
535 blob = BufferBlob::create("Compile::scratch_buffer", size);
536 // Record the buffer blob for next time.
537 set_scratch_buffer_blob(blob);
538 // Have we run out of code space?
539 if (scratch_buffer_blob() == NULL) {
540 // Let CompilerBroker disable further compilations.
541 record_failure("Not enough space for scratch buffer in CodeCache");
542 return;
543 }
544 }
545
546 // Initialize the relocation buffers
547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
548 set_scratch_locs_memory(locs_buf);
549 }
550
551
552 //-----------------------scratch_emit_size-------------------------------------
553 // Helper function that computes size by emitting code
554 uint Compile::scratch_emit_size(const Node* n) {
555 // Start scratch_emit_size section.
556 set_in_scratch_emit_size(true);
557
558 // Emit into a trash buffer and count bytes emitted.
559 // This is a pretty expensive way to compute a size,
560 // but it works well enough if seldom used.
561 // All common fixed-size instructions are given a size
562 // method by the AD file.
563 // Note that the scratch buffer blob and locs memory are
564 // allocated at the beginning of the compile task, and
565 // may be shared by several calls to scratch_emit_size.
566 // The allocation of the scratch buffer blob is particularly
567 // expensive, since it has to grab the code cache lock.
568 BufferBlob* blob = this->scratch_buffer_blob();
569 assert(blob != NULL, "Initialize BufferBlob at start");
570 assert(blob->size() > MAX_inst_size, "sanity");
571 relocInfo* locs_buf = scratch_locs_memory();
572 address blob_begin = blob->content_begin();
573 address blob_end = (address)locs_buf;
574 assert(blob->content_contains(blob_end), "sanity");
575 CodeBuffer buf(blob_begin, blob_end - blob_begin);
576 buf.initialize_consts_size(_scratch_const_size);
577 buf.initialize_stubs_size(MAX_stubs_size);
578 assert(locs_buf != NULL, "sanity");
579 int lsize = MAX_locs_size / 3;
580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
583
584 // Do the emission.
585
586 Label fakeL; // Fake label for branch instructions.
587 Label* saveL = NULL;
588 uint save_bnum = 0;
589 bool is_branch = n->is_MachBranch();
590 if (is_branch) {
591 MacroAssembler masm(&buf);
592 masm.bind(fakeL);
593 n->as_MachBranch()->save_label(&saveL, &save_bnum);
594 n->as_MachBranch()->label_set(&fakeL, 0);
595 }
596 n->emit(buf, this->regalloc());
597
598 // Emitting into the scratch buffer should not fail
599 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
600
601 if (is_branch) // Restore label.
602 n->as_MachBranch()->label_set(saveL, save_bnum);
603
604 // End scratch_emit_size section.
605 set_in_scratch_emit_size(false);
606
607 return buf.insts_size();
608 }
609
610
611 // ============================================================================
612 //------------------------------Compile standard-------------------------------
613 debug_only( int Compile::_debug_idx = 100000; )
614
615 // Compile a method. entry_bci is -1 for normal compilations and indicates
616 // the continuation bci for on stack replacement.
617
618
619 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
620 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
621 : Phase(Compiler),
622 _env(ci_env),
623 _directive(directive),
624 _log(ci_env->log()),
625 _compile_id(ci_env->compile_id()),
626 _save_argument_registers(false),
627 _stub_name(NULL),
628 _stub_function(NULL),
629 _stub_entry_point(NULL),
630 _method(target),
631 _entry_bci(osr_bci),
632 _initial_gvn(NULL),
633 _for_igvn(NULL),
634 _warm_calls(NULL),
635 _subsume_loads(subsume_loads),
636 _do_escape_analysis(do_escape_analysis),
637 _eliminate_boxing(eliminate_boxing),
638 _failure_reason(NULL),
639 _code_buffer("Compile::Fill_buffer"),
640 _orig_pc_slot(0),
641 _orig_pc_slot_offset_in_bytes(0),
642 _has_method_handle_invokes(false),
643 _mach_constant_base_node(NULL),
644 _node_bundling_limit(0),
645 _node_bundling_base(NULL),
646 _java_calls(0),
647 _inner_loops(0),
648 _scratch_const_size(-1),
649 _in_scratch_emit_size(false),
650 _dead_node_list(comp_arena()),
651 _dead_node_count(0),
652 #ifndef PRODUCT
653 _trace_opto_output(directive->TraceOptoOutputOption),
654 _in_dump_cnt(0),
655 _printer(IdealGraphPrinter::printer()),
656 #endif
657 _congraph(NULL),
658 _comp_arena(mtCompiler),
659 _node_arena(mtCompiler),
660 _old_arena(mtCompiler),
661 _Compile_types(mtCompiler),
662 _replay_inline_data(NULL),
663 _late_inlines(comp_arena(), 2, 0, NULL),
664 _string_late_inlines(comp_arena(), 2, 0, NULL),
665 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
666 _late_inlines_pos(0),
667 _number_of_mh_late_inlines(0),
668 _inlining_progress(false),
669 _inlining_incrementally(false),
670 _print_inlining_list(NULL),
671 _print_inlining_stream(NULL),
672 _print_inlining_idx(0),
673 _print_inlining_output(NULL),
674 _interpreter_frame_size(0),
675 _max_node_limit(MaxNodeLimit),
676 _has_reserved_stack_access(target->has_reserved_stack_access()) {
677 C = this;
678 #ifndef PRODUCT
679 if (_printer != NULL) {
680 _printer->set_compile(this);
681 }
682 #endif
683 CompileWrapper cw(this);
684
685 if (CITimeVerbose) {
686 tty->print(" ");
687 target->holder()->name()->print();
688 tty->print(".");
689 target->print_short_name();
690 tty->print(" ");
691 }
692 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
693 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
694
695 #ifndef PRODUCT
696 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
697 if (!print_opto_assembly) {
698 bool print_assembly = directive->PrintAssemblyOption;
699 if (print_assembly && !Disassembler::can_decode()) {
700 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
701 print_opto_assembly = true;
702 }
703 }
704 set_print_assembly(print_opto_assembly);
705 set_parsed_irreducible_loop(false);
706
707 if (directive->ReplayInlineOption) {
708 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
709 }
710 #endif
711 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
712 set_print_intrinsics(directive->PrintIntrinsicsOption);
713 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
714
715 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
716 // Make sure the method being compiled gets its own MDO,
717 // so we can at least track the decompile_count().
718 // Need MDO to record RTM code generation state.
719 method()->ensure_method_data();
720 }
721
722 Init(::AliasLevel);
723
724
725 print_compile_messages();
726
727 _ilt = InlineTree::build_inline_tree_root();
728
729 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
730 assert(num_alias_types() >= AliasIdxRaw, "");
731
732 #define MINIMUM_NODE_HASH 1023
733 // Node list that Iterative GVN will start with
734 Unique_Node_List for_igvn(comp_arena());
735 set_for_igvn(&for_igvn);
736
737 // GVN that will be run immediately on new nodes
738 uint estimated_size = method()->code_size()*4+64;
739 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
740 PhaseGVN gvn(node_arena(), estimated_size);
741 set_initial_gvn(&gvn);
742
743 print_inlining_init();
744 { // Scope for timing the parser
745 TracePhase tp("parse", &timers[_t_parser]);
746
747 // Put top into the hash table ASAP.
748 initial_gvn()->transform_no_reclaim(top());
749
750 // Set up tf(), start(), and find a CallGenerator.
751 CallGenerator* cg = NULL;
752 if (is_osr_compilation()) {
753 const TypeTuple *domain = StartOSRNode::osr_domain();
754 const TypeTuple *range = TypeTuple::make_range(method()->signature());
755 init_tf(TypeFunc::make(domain, range));
756 StartNode* s = new StartOSRNode(root(), domain);
757 initial_gvn()->set_type_bottom(s);
758 init_start(s);
759 cg = CallGenerator::for_osr(method(), entry_bci());
760 } else {
761 // Normal case.
762 init_tf(TypeFunc::make(method()));
763 StartNode* s = new StartNode(root(), tf()->domain());
764 initial_gvn()->set_type_bottom(s);
765 init_start(s);
766 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
767 // With java.lang.ref.reference.get() we must go through the
768 // intrinsic when G1 is enabled - even when get() is the root
769 // method of the compile - so that, if necessary, the value in
770 // the referent field of the reference object gets recorded by
771 // the pre-barrier code.
772 // Specifically, if G1 is enabled, the value in the referent
773 // field is recorded by the G1 SATB pre barrier. This will
774 // result in the referent being marked live and the reference
775 // object removed from the list of discovered references during
776 // reference processing.
777 cg = find_intrinsic(method(), false);
778 }
779 if (cg == NULL) {
780 float past_uses = method()->interpreter_invocation_count();
781 float expected_uses = past_uses;
782 cg = CallGenerator::for_inline(method(), expected_uses);
783 }
784 }
785 if (failing()) return;
786 if (cg == NULL) {
787 record_method_not_compilable_all_tiers("cannot parse method");
788 return;
789 }
790 JVMState* jvms = build_start_state(start(), tf());
791 if ((jvms = cg->generate(jvms)) == NULL) {
792 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
793 record_method_not_compilable("method parse failed");
794 }
795 return;
796 }
797 GraphKit kit(jvms);
798
799 if (!kit.stopped()) {
800 // Accept return values, and transfer control we know not where.
801 // This is done by a special, unique ReturnNode bound to root.
802 return_values(kit.jvms());
803 }
804
805 if (kit.has_exceptions()) {
806 // Any exceptions that escape from this call must be rethrown
807 // to whatever caller is dynamically above us on the stack.
808 // This is done by a special, unique RethrowNode bound to root.
809 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
810 }
811
812 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
813
814 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
815 inline_string_calls(true);
816 }
817
818 if (failing()) return;
819
820 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
821
822 // Remove clutter produced by parsing.
823 if (!failing()) {
824 ResourceMark rm;
825 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
826 }
827 }
828
829 // Note: Large methods are capped off in do_one_bytecode().
830 if (failing()) return;
831
832 // After parsing, node notes are no longer automagic.
833 // They must be propagated by register_new_node_with_optimizer(),
834 // clone(), or the like.
835 set_default_node_notes(NULL);
836
837 for (;;) {
838 int successes = Inline_Warm();
839 if (failing()) return;
840 if (successes == 0) break;
841 }
842
843 // Drain the list.
844 Finish_Warm();
845 #ifndef PRODUCT
846 if (_printer && _printer->should_print(1)) {
847 _printer->print_inlining();
848 }
849 #endif
850
851 if (failing()) return;
852 NOT_PRODUCT( verify_graph_edges(); )
853
854 // Now optimize
855 Optimize();
856 if (failing()) return;
857 NOT_PRODUCT( verify_graph_edges(); )
858
859 #ifndef PRODUCT
860 if (PrintIdeal) {
861 ttyLocker ttyl; // keep the following output all in one block
862 // This output goes directly to the tty, not the compiler log.
863 // To enable tools to match it up with the compilation activity,
864 // be sure to tag this tty output with the compile ID.
865 if (xtty != NULL) {
866 xtty->head("ideal compile_id='%d'%s", compile_id(),
867 is_osr_compilation() ? " compile_kind='osr'" :
868 "");
869 }
870 root()->dump(9999);
871 if (xtty != NULL) {
872 xtty->tail("ideal");
873 }
874 }
875 #endif
876
877 NOT_PRODUCT( verify_barriers(); )
878
879 // Dump compilation data to replay it.
880 if (directive->DumpReplayOption) {
881 env()->dump_replay_data(_compile_id);
882 }
883 if (directive->DumpInlineOption && (ilt() != NULL)) {
884 env()->dump_inline_data(_compile_id);
885 }
886
887 // Now that we know the size of all the monitors we can add a fixed slot
888 // for the original deopt pc.
889
890 _orig_pc_slot = fixed_slots();
891 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
892 set_fixed_slots(next_slot);
893
894 // Compute when to use implicit null checks. Used by matching trap based
895 // nodes and NullCheck optimization.
896 set_allowed_deopt_reasons();
897
898 // Now generate code
899 Code_Gen();
900 if (failing()) return;
901
902 // Check if we want to skip execution of all compiled code.
903 {
904 #ifndef PRODUCT
905 if (OptoNoExecute) {
906 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
907 return;
908 }
909 #endif
910 TracePhase tp("install_code", &timers[_t_registerMethod]);
911
912 if (is_osr_compilation()) {
913 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
914 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
915 } else {
916 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
917 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
918 }
919
920 env()->register_method(_method, _entry_bci,
921 &_code_offsets,
922 _orig_pc_slot_offset_in_bytes,
923 code_buffer(),
924 frame_size_in_words(), _oop_map_set,
925 &_handler_table, &_inc_table,
926 compiler,
927 has_unsafe_access(),
928 SharedRuntime::is_wide_vector(max_vector_size()),
929 _directive,
930 rtm_state()
931 );
932
933 if (log() != NULL) // Print code cache state into compiler log
934 log()->code_cache_state();
935 }
936 }
937
938 //------------------------------Compile----------------------------------------
939 // Compile a runtime stub
940 Compile::Compile( ciEnv* ci_env,
941 TypeFunc_generator generator,
942 address stub_function,
943 const char *stub_name,
944 int is_fancy_jump,
945 bool pass_tls,
946 bool save_arg_registers,
947 bool return_pc,
948 DirectiveSet* directive)
949 : Phase(Compiler),
950 _env(ci_env),
951 _directive(directive),
952 _log(ci_env->log()),
953 _compile_id(0),
954 _save_argument_registers(save_arg_registers),
955 _method(NULL),
956 _stub_name(stub_name),
957 _stub_function(stub_function),
958 _stub_entry_point(NULL),
959 _entry_bci(InvocationEntryBci),
960 _initial_gvn(NULL),
961 _for_igvn(NULL),
962 _warm_calls(NULL),
963 _orig_pc_slot(0),
964 _orig_pc_slot_offset_in_bytes(0),
965 _subsume_loads(true),
966 _do_escape_analysis(false),
967 _eliminate_boxing(false),
968 _failure_reason(NULL),
969 _code_buffer("Compile::Fill_buffer"),
970 _has_method_handle_invokes(false),
971 _mach_constant_base_node(NULL),
972 _node_bundling_limit(0),
973 _node_bundling_base(NULL),
974 _java_calls(0),
975 _inner_loops(0),
976 #ifndef PRODUCT
977 _trace_opto_output(TraceOptoOutput),
978 _in_dump_cnt(0),
979 _printer(NULL),
980 #endif
981 _comp_arena(mtCompiler),
982 _node_arena(mtCompiler),
983 _old_arena(mtCompiler),
984 _Compile_types(mtCompiler),
985 _dead_node_list(comp_arena()),
986 _dead_node_count(0),
987 _congraph(NULL),
988 _replay_inline_data(NULL),
989 _number_of_mh_late_inlines(0),
990 _inlining_progress(false),
991 _inlining_incrementally(false),
992 _print_inlining_list(NULL),
993 _print_inlining_stream(NULL),
994 _print_inlining_idx(0),
995 _print_inlining_output(NULL),
996 _allowed_reasons(0),
997 _interpreter_frame_size(0),
998 _max_node_limit(MaxNodeLimit) {
999 C = this;
1000
1001 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1002 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1003
1004 #ifndef PRODUCT
1005 set_print_assembly(PrintFrameConverterAssembly);
1006 set_parsed_irreducible_loop(false);
1007 #endif
1008 set_has_irreducible_loop(false); // no loops
1009
1010 CompileWrapper cw(this);
1011 Init(/*AliasLevel=*/ 0);
1012 init_tf((*generator)());
1013
1014 {
1015 // The following is a dummy for the sake of GraphKit::gen_stub
1016 Unique_Node_List for_igvn(comp_arena());
1017 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1018 PhaseGVN gvn(Thread::current()->resource_area(),255);
1019 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1020 gvn.transform_no_reclaim(top());
1021
1022 GraphKit kit;
1023 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1024 }
1025
1026 NOT_PRODUCT( verify_graph_edges(); )
1027 Code_Gen();
1028 if (failing()) return;
1029
1030
1031 // Entry point will be accessed using compile->stub_entry_point();
1032 if (code_buffer() == NULL) {
1033 Matcher::soft_match_failure();
1034 } else {
1035 if (PrintAssembly && (WizardMode || Verbose))
1036 tty->print_cr("### Stub::%s", stub_name);
1037
1038 if (!failing()) {
1039 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1040
1041 // Make the NMethod
1042 // For now we mark the frame as never safe for profile stackwalking
1043 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1044 code_buffer(),
1045 CodeOffsets::frame_never_safe,
1046 // _code_offsets.value(CodeOffsets::Frame_Complete),
1047 frame_size_in_words(),
1048 _oop_map_set,
1049 save_arg_registers);
1050 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1051
1052 _stub_entry_point = rs->entry_point();
1053 }
1054 }
1055 }
1056
1057 //------------------------------Init-------------------------------------------
1058 // Prepare for a single compilation
1059 void Compile::Init(int aliaslevel) {
1060 _unique = 0;
1061 _regalloc = NULL;
1062
1063 _tf = NULL; // filled in later
1064 _top = NULL; // cached later
1065 _matcher = NULL; // filled in later
1066 _cfg = NULL; // filled in later
1067
1068 set_24_bit_selection_and_mode(Use24BitFP, false);
1069
1070 _node_note_array = NULL;
1071 _default_node_notes = NULL;
1072 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1073
1074 _immutable_memory = NULL; // filled in at first inquiry
1075
1076 // Globally visible Nodes
1077 // First set TOP to NULL to give safe behavior during creation of RootNode
1078 set_cached_top_node(NULL);
1079 set_root(new RootNode());
1080 // Now that you have a Root to point to, create the real TOP
1081 set_cached_top_node( new ConNode(Type::TOP) );
1082 set_recent_alloc(NULL, NULL);
1083
1084 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1085 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1086 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1087 env()->set_dependencies(new Dependencies(env()));
1088
1089 _fixed_slots = 0;
1090 set_has_split_ifs(false);
1091 set_has_loops(has_method() && method()->has_loops()); // first approximation
1092 set_has_stringbuilder(false);
1093 set_has_boxed_value(false);
1094 _trap_can_recompile = false; // no traps emitted yet
1095 _major_progress = true; // start out assuming good things will happen
1096 set_has_unsafe_access(false);
1097 set_max_vector_size(0);
1098 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1099 set_decompile_count(0);
1100
1101 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1102 set_num_loop_opts(LoopOptsCount);
1103 set_do_inlining(Inline);
1104 set_max_inline_size(MaxInlineSize);
1105 set_freq_inline_size(FreqInlineSize);
1106 set_do_scheduling(OptoScheduling);
1107 set_do_count_invocations(false);
1108 set_do_method_data_update(false);
1109
1110 set_do_vector_loop(false);
1111
1112 if (AllowVectorizeOnDemand) {
1113 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1114 set_do_vector_loop(true);
1115 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());})
1116 } else if (has_method() && method()->name() != 0 &&
1117 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1118 set_do_vector_loop(true);
1119 }
1120 }
1121 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1122 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());})
1123
1124 set_age_code(has_method() && method()->profile_aging());
1125 set_rtm_state(NoRTM); // No RTM lock eliding by default
1126 _max_node_limit = _directive->MaxNodeLimitOption;
1127
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 register_library_intrinsics();
1182 }
1183
1184 //---------------------------init_start----------------------------------------
1185 // Install the StartNode on this compile object.
1186 void Compile::init_start(StartNode* s) {
1187 if (failing())
1188 return; // already failing
1189 assert(s == start(), "");
1190 }
1191
1192 /**
1193 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1194 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1195 * the ideal graph.
1196 */
1197 StartNode* Compile::start() const {
1198 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
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 }
1205 fatal("Did not find Start node!");
1206 return NULL;
1207 }
1208
1209 //-------------------------------immutable_memory-------------------------------------
1210 // Access immutable memory
1211 Node* Compile::immutable_memory() {
1212 if (_immutable_memory != NULL) {
1213 return _immutable_memory;
1214 }
1215 StartNode* s = start();
1216 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1217 Node *p = s->fast_out(i);
1218 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1219 _immutable_memory = p;
1220 return _immutable_memory;
1221 }
1222 }
1223 ShouldNotReachHere();
1224 return NULL;
1225 }
1226
1227 //----------------------set_cached_top_node------------------------------------
1228 // Install the cached top node, and make sure Node::is_top works correctly.
1229 void Compile::set_cached_top_node(Node* tn) {
1230 if (tn != NULL) verify_top(tn);
1231 Node* old_top = _top;
1232 _top = tn;
1233 // Calling Node::setup_is_top allows the nodes the chance to adjust
1234 // their _out arrays.
1235 if (_top != NULL) _top->setup_is_top();
1236 if (old_top != NULL) old_top->setup_is_top();
1237 assert(_top == NULL || top()->is_top(), "");
1238 }
1239
1240 #ifdef ASSERT
1241 uint Compile::count_live_nodes_by_graph_walk() {
1242 Unique_Node_List useful(comp_arena());
1243 // Get useful node list by walking the graph.
1244 identify_useful_nodes(useful);
1245 return useful.size();
1246 }
1247
1248 void Compile::print_missing_nodes() {
1249
1250 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1251 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1252 return;
1253 }
1254
1255 // This is an expensive function. It is executed only when the user
1256 // specifies VerifyIdealNodeCount option or otherwise knows the
1257 // additional work that needs to be done to identify reachable nodes
1258 // by walking the flow graph and find the missing ones using
1259 // _dead_node_list.
1260
1261 Unique_Node_List useful(comp_arena());
1262 // Get useful node list by walking the graph.
1263 identify_useful_nodes(useful);
1264
1265 uint l_nodes = C->live_nodes();
1266 uint l_nodes_by_walk = useful.size();
1267
1268 if (l_nodes != l_nodes_by_walk) {
1269 if (_log != NULL) {
1270 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1271 _log->stamp();
1272 _log->end_head();
1273 }
1274 VectorSet& useful_member_set = useful.member_set();
1275 int last_idx = l_nodes_by_walk;
1276 for (int i = 0; i < last_idx; i++) {
1277 if (useful_member_set.test(i)) {
1278 if (_dead_node_list.test(i)) {
1279 if (_log != NULL) {
1280 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1281 }
1282 if (PrintIdealNodeCount) {
1283 // Print the log message to tty
1284 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1285 useful.at(i)->dump();
1286 }
1287 }
1288 }
1289 else if (! _dead_node_list.test(i)) {
1290 if (_log != NULL) {
1291 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1292 }
1293 if (PrintIdealNodeCount) {
1294 // Print the log message to tty
1295 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1296 }
1297 }
1298 }
1299 if (_log != NULL) {
1300 _log->tail("mismatched_nodes");
1301 }
1302 }
1303 }
1304 void Compile::record_modified_node(Node* n) {
1305 if (_modified_nodes != NULL && !_inlining_incrementally &&
1306 n->outcnt() != 0 && !n->is_Con()) {
1307 _modified_nodes->push(n);
1308 }
1309 }
1310
1311 void Compile::remove_modified_node(Node* n) {
1312 if (_modified_nodes != NULL) {
1313 _modified_nodes->remove(n);
1314 }
1315 }
1316 #endif
1317
1318 #ifndef PRODUCT
1319 void Compile::verify_top(Node* tn) const {
1320 if (tn != NULL) {
1321 assert(tn->is_Con(), "top node must be a constant");
1322 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1323 assert(tn->in(0) != NULL, "must have live top node");
1324 }
1325 }
1326 #endif
1327
1328
1329 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1330
1331 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1332 guarantee(arr != NULL, "");
1333 int num_blocks = arr->length();
1334 if (grow_by < num_blocks) grow_by = num_blocks;
1335 int num_notes = grow_by * _node_notes_block_size;
1336 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1337 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1338 while (num_notes > 0) {
1339 arr->append(notes);
1340 notes += _node_notes_block_size;
1341 num_notes -= _node_notes_block_size;
1342 }
1343 assert(num_notes == 0, "exact multiple, please");
1344 }
1345
1346 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1347 if (source == NULL || dest == NULL) return false;
1348
1349 if (dest->is_Con())
1350 return false; // Do not push debug info onto constants.
1351
1352 #ifdef ASSERT
1353 // Leave a bread crumb trail pointing to the original node:
1354 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1355 dest->set_debug_orig(source);
1356 }
1357 #endif
1358
1359 if (node_note_array() == NULL)
1360 return false; // Not collecting any notes now.
1361
1362 // This is a copy onto a pre-existing node, which may already have notes.
1363 // If both nodes have notes, do not overwrite any pre-existing notes.
1364 Node_Notes* source_notes = node_notes_at(source->_idx);
1365 if (source_notes == NULL || source_notes->is_clear()) return false;
1366 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1367 if (dest_notes == NULL || dest_notes->is_clear()) {
1368 return set_node_notes_at(dest->_idx, source_notes);
1369 }
1370
1371 Node_Notes merged_notes = (*source_notes);
1372 // The order of operations here ensures that dest notes will win...
1373 merged_notes.update_from(dest_notes);
1374 return set_node_notes_at(dest->_idx, &merged_notes);
1375 }
1376
1377
1378 //--------------------------allow_range_check_smearing-------------------------
1379 // Gating condition for coalescing similar range checks.
1380 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1381 // single covering check that is at least as strong as any of them.
1382 // If the optimization succeeds, the simplified (strengthened) range check
1383 // will always succeed. If it fails, we will deopt, and then give up
1384 // on the optimization.
1385 bool Compile::allow_range_check_smearing() const {
1386 // If this method has already thrown a range-check,
1387 // assume it was because we already tried range smearing
1388 // and it failed.
1389 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1390 return !already_trapped;
1391 }
1392
1393
1394 //------------------------------flatten_alias_type-----------------------------
1395 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1396 int offset = tj->offset();
1397 TypePtr::PTR ptr = tj->ptr();
1398
1399 // Known instance (scalarizable allocation) alias only with itself.
1400 bool is_known_inst = tj->isa_oopptr() != NULL &&
1401 tj->is_oopptr()->is_known_instance();
1402
1403 // Process weird unsafe references.
1404 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1405 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1406 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1407 tj = TypeOopPtr::BOTTOM;
1408 ptr = tj->ptr();
1409 offset = tj->offset();
1410 }
1411
1412 // Array pointers need some flattening
1413 const TypeAryPtr *ta = tj->isa_aryptr();
1414 if (ta && ta->is_stable()) {
1415 // Erase stability property for alias analysis.
1416 tj = ta = ta->cast_to_stable(false);
1417 }
1418 if( ta && is_known_inst ) {
1419 if ( offset != Type::OffsetBot &&
1420 offset > arrayOopDesc::length_offset_in_bytes() ) {
1421 offset = Type::OffsetBot; // Flatten constant access into array body only
1422 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1423 }
1424 } else if( ta && _AliasLevel >= 2 ) {
1425 // For arrays indexed by constant indices, we flatten the alias
1426 // space to include all of the array body. Only the header, klass
1427 // and array length can be accessed un-aliased.
1428 if( offset != Type::OffsetBot ) {
1429 if( ta->const_oop() ) { // MethodData* or Method*
1430 offset = Type::OffsetBot; // Flatten constant access into array body
1431 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1432 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1433 // range is OK as-is.
1434 tj = ta = TypeAryPtr::RANGE;
1435 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1436 tj = TypeInstPtr::KLASS; // all klass loads look alike
1437 ta = TypeAryPtr::RANGE; // generic ignored junk
1438 ptr = TypePtr::BotPTR;
1439 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1440 tj = TypeInstPtr::MARK;
1441 ta = TypeAryPtr::RANGE; // generic ignored junk
1442 ptr = TypePtr::BotPTR;
1443 } else { // Random constant offset into array body
1444 offset = Type::OffsetBot; // Flatten constant access into array body
1445 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1446 }
1447 }
1448 // Arrays of fixed size alias with arrays of unknown size.
1449 if (ta->size() != TypeInt::POS) {
1450 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1452 }
1453 // Arrays of known objects become arrays of unknown objects.
1454 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1455 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1456 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1457 }
1458 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1459 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1460 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1461 }
1462 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1463 // cannot be distinguished by bytecode alone.
1464 if (ta->elem() == TypeInt::BOOL) {
1465 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1466 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1467 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1468 }
1469 // During the 2nd round of IterGVN, NotNull castings are removed.
1470 // Make sure the Bottom and NotNull variants alias the same.
1471 // Also, make sure exact and non-exact variants alias the same.
1472 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1473 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1474 }
1475 }
1476
1477 // Oop pointers need some flattening
1478 const TypeInstPtr *to = tj->isa_instptr();
1479 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1480 ciInstanceKlass *k = to->klass()->as_instance_klass();
1481 if( ptr == TypePtr::Constant ) {
1482 if (to->klass() != ciEnv::current()->Class_klass() ||
1483 offset < k->size_helper() * wordSize) {
1484 // No constant oop pointers (such as Strings); they alias with
1485 // unknown strings.
1486 assert(!is_known_inst, "not scalarizable allocation");
1487 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1488 }
1489 } else if( is_known_inst ) {
1490 tj = to; // Keep NotNull and klass_is_exact for instance type
1491 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1492 // During the 2nd round of IterGVN, NotNull castings are removed.
1493 // Make sure the Bottom and NotNull variants alias the same.
1494 // Also, make sure exact and non-exact variants alias the same.
1495 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1496 }
1497 if (to->speculative() != NULL) {
1498 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1499 }
1500 // Canonicalize the holder of this field
1501 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1502 // First handle header references such as a LoadKlassNode, even if the
1503 // object's klass is unloaded at compile time (4965979).
1504 if (!is_known_inst) { // Do it only for non-instance types
1505 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1506 }
1507 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1508 // Static fields are in the space above the normal instance
1509 // fields in the java.lang.Class instance.
1510 if (to->klass() != ciEnv::current()->Class_klass()) {
1511 to = NULL;
1512 tj = TypeOopPtr::BOTTOM;
1513 offset = tj->offset();
1514 }
1515 } else {
1516 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1517 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1518 if( is_known_inst ) {
1519 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1520 } else {
1521 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1522 }
1523 }
1524 }
1525 }
1526
1527 // Klass pointers to object array klasses need some flattening
1528 const TypeKlassPtr *tk = tj->isa_klassptr();
1529 if( tk ) {
1530 // If we are referencing a field within a Klass, we need
1531 // to assume the worst case of an Object. Both exact and
1532 // inexact types must flatten to the same alias class so
1533 // use NotNull as the PTR.
1534 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1535
1536 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1537 TypeKlassPtr::OBJECT->klass(),
1538 offset);
1539 }
1540
1541 ciKlass* klass = tk->klass();
1542 if( klass->is_obj_array_klass() ) {
1543 ciKlass* k = TypeAryPtr::OOPS->klass();
1544 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1545 k = TypeInstPtr::BOTTOM->klass();
1546 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1547 }
1548
1549 // Check for precise loads from the primary supertype array and force them
1550 // to the supertype cache alias index. Check for generic array loads from
1551 // the primary supertype array and also force them to the supertype cache
1552 // alias index. Since the same load can reach both, we need to merge
1553 // these 2 disparate memories into the same alias class. Since the
1554 // primary supertype array is read-only, there's no chance of confusion
1555 // where we bypass an array load and an array store.
1556 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1557 if (offset == Type::OffsetBot ||
1558 (offset >= primary_supers_offset &&
1559 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1560 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1561 offset = in_bytes(Klass::secondary_super_cache_offset());
1562 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1563 }
1564 }
1565
1566 // Flatten all Raw pointers together.
1567 if (tj->base() == Type::RawPtr)
1568 tj = TypeRawPtr::BOTTOM;
1569
1570 if (tj->base() == Type::AnyPtr)
1571 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1572
1573 // Flatten all to bottom for now
1574 switch( _AliasLevel ) {
1575 case 0:
1576 tj = TypePtr::BOTTOM;
1577 break;
1578 case 1: // Flatten to: oop, static, field or array
1579 switch (tj->base()) {
1580 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1581 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1582 case Type::AryPtr: // do not distinguish arrays at all
1583 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1584 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1585 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1586 default: ShouldNotReachHere();
1587 }
1588 break;
1589 case 2: // No collapsing at level 2; keep all splits
1590 case 3: // No collapsing at level 3; keep all splits
1591 break;
1592 default:
1593 Unimplemented();
1594 }
1595
1596 offset = tj->offset();
1597 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1598
1599 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1600 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1601 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1602 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1603 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1604 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1605 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1606 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1607 assert( tj->ptr() != TypePtr::TopPTR &&
1608 tj->ptr() != TypePtr::AnyNull &&
1609 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1610 // assert( tj->ptr() != TypePtr::Constant ||
1611 // tj->base() == Type::RawPtr ||
1612 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1613
1614 return tj;
1615 }
1616
1617 void Compile::AliasType::Init(int i, const TypePtr* at) {
1618 _index = i;
1619 _adr_type = at;
1620 _field = NULL;
1621 _element = NULL;
1622 _is_rewritable = true; // default
1623 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1624 if (atoop != NULL && atoop->is_known_instance()) {
1625 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1626 _general_index = Compile::current()->get_alias_index(gt);
1627 } else {
1628 _general_index = 0;
1629 }
1630 }
1631
1632 //---------------------------------print_on------------------------------------
1633 #ifndef PRODUCT
1634 void Compile::AliasType::print_on(outputStream* st) {
1635 if (index() < 10)
1636 st->print("@ <%d> ", index());
1637 else st->print("@ <%d>", index());
1638 st->print(is_rewritable() ? " " : " RO");
1639 int offset = adr_type()->offset();
1640 if (offset == Type::OffsetBot)
1641 st->print(" +any");
1642 else st->print(" +%-3d", offset);
1643 st->print(" in ");
1644 adr_type()->dump_on(st);
1645 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1646 if (field() != NULL && tjp) {
1647 if (tjp->klass() != field()->holder() ||
1648 tjp->offset() != field()->offset_in_bytes()) {
1649 st->print(" != ");
1650 field()->print();
1651 st->print(" ***");
1652 }
1653 }
1654 }
1655
1656 void print_alias_types() {
1657 Compile* C = Compile::current();
1658 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1659 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1660 C->alias_type(idx)->print_on(tty);
1661 tty->cr();
1662 }
1663 }
1664 #endif
1665
1666
1667 //----------------------------probe_alias_cache--------------------------------
1668 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1669 intptr_t key = (intptr_t) adr_type;
1670 key ^= key >> logAliasCacheSize;
1671 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1672 }
1673
1674
1675 //-----------------------------grow_alias_types--------------------------------
1676 void Compile::grow_alias_types() {
1677 const int old_ats = _max_alias_types; // how many before?
1678 const int new_ats = old_ats; // how many more?
1679 const int grow_ats = old_ats+new_ats; // how many now?
1680 _max_alias_types = grow_ats;
1681 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1682 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1683 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1684 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1685 }
1686
1687
1688 //--------------------------------find_alias_type------------------------------
1689 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1690 if (_AliasLevel == 0)
1691 return alias_type(AliasIdxBot);
1692
1693 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1694 if (ace->_adr_type == adr_type) {
1695 return alias_type(ace->_index);
1696 }
1697
1698 // Handle special cases.
1699 if (adr_type == NULL) return alias_type(AliasIdxTop);
1700 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1701
1702 // Do it the slow way.
1703 const TypePtr* flat = flatten_alias_type(adr_type);
1704
1705 #ifdef ASSERT
1706 assert(flat == flatten_alias_type(flat), "idempotent");
1707 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1708 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1709 const TypeOopPtr* foop = flat->is_oopptr();
1710 // Scalarizable allocations have exact klass always.
1711 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1712 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1713 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1714 }
1715 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1716 #endif
1717
1718 int idx = AliasIdxTop;
1719 for (int i = 0; i < num_alias_types(); i++) {
1720 if (alias_type(i)->adr_type() == flat) {
1721 idx = i;
1722 break;
1723 }
1724 }
1725
1726 if (idx == AliasIdxTop) {
1727 if (no_create) return NULL;
1728 // Grow the array if necessary.
1729 if (_num_alias_types == _max_alias_types) grow_alias_types();
1730 // Add a new alias type.
1731 idx = _num_alias_types++;
1732 _alias_types[idx]->Init(idx, flat);
1733 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1734 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1735 if (flat->isa_instptr()) {
1736 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1737 && flat->is_instptr()->klass() == env()->Class_klass())
1738 alias_type(idx)->set_rewritable(false);
1739 }
1740 if (flat->isa_aryptr()) {
1741 #ifdef ASSERT
1742 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1743 // (T_BYTE has the weakest alignment and size restrictions...)
1744 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1745 #endif
1746 if (flat->offset() == TypePtr::OffsetBot) {
1747 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1748 }
1749 }
1750 if (flat->isa_klassptr()) {
1751 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1752 alias_type(idx)->set_rewritable(false);
1753 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1754 alias_type(idx)->set_rewritable(false);
1755 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1756 alias_type(idx)->set_rewritable(false);
1757 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1758 alias_type(idx)->set_rewritable(false);
1759 }
1760 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1761 // but the base pointer type is not distinctive enough to identify
1762 // references into JavaThread.)
1763
1764 // Check for final fields.
1765 const TypeInstPtr* tinst = flat->isa_instptr();
1766 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1767 ciField* field;
1768 if (tinst->const_oop() != NULL &&
1769 tinst->klass() == ciEnv::current()->Class_klass() &&
1770 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1771 // static field
1772 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1773 field = k->get_field_by_offset(tinst->offset(), true);
1774 } else {
1775 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1776 field = k->get_field_by_offset(tinst->offset(), false);
1777 }
1778 assert(field == NULL ||
1779 original_field == NULL ||
1780 (field->holder() == original_field->holder() &&
1781 field->offset() == original_field->offset() &&
1782 field->is_static() == original_field->is_static()), "wrong field?");
1783 // Set field() and is_rewritable() attributes.
1784 if (field != NULL) alias_type(idx)->set_field(field);
1785 }
1786 }
1787
1788 // Fill the cache for next time.
1789 ace->_adr_type = adr_type;
1790 ace->_index = idx;
1791 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1792
1793 // Might as well try to fill the cache for the flattened version, too.
1794 AliasCacheEntry* face = probe_alias_cache(flat);
1795 if (face->_adr_type == NULL) {
1796 face->_adr_type = flat;
1797 face->_index = idx;
1798 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1799 }
1800
1801 return alias_type(idx);
1802 }
1803
1804
1805 Compile::AliasType* Compile::alias_type(ciField* field) {
1806 const TypeOopPtr* t;
1807 if (field->is_static())
1808 t = TypeInstPtr::make(field->holder()->java_mirror());
1809 else
1810 t = TypeOopPtr::make_from_klass_raw(field->holder());
1811 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1812 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1813 return atp;
1814 }
1815
1816
1817 //------------------------------have_alias_type--------------------------------
1818 bool Compile::have_alias_type(const TypePtr* adr_type) {
1819 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1820 if (ace->_adr_type == adr_type) {
1821 return true;
1822 }
1823
1824 // Handle special cases.
1825 if (adr_type == NULL) return true;
1826 if (adr_type == TypePtr::BOTTOM) return true;
1827
1828 return find_alias_type(adr_type, true, NULL) != NULL;
1829 }
1830
1831 //-----------------------------must_alias--------------------------------------
1832 // True if all values of the given address type are in the given alias category.
1833 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1834 if (alias_idx == AliasIdxBot) return true; // the universal category
1835 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1836 if (alias_idx == AliasIdxTop) return false; // the empty category
1837 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1838
1839 // the only remaining possible overlap is identity
1840 int adr_idx = get_alias_index(adr_type);
1841 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1842 assert(adr_idx == alias_idx ||
1843 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1844 && adr_type != TypeOopPtr::BOTTOM),
1845 "should not be testing for overlap with an unsafe pointer");
1846 return adr_idx == alias_idx;
1847 }
1848
1849 //------------------------------can_alias--------------------------------------
1850 // True if any values of the given address type are in the given alias category.
1851 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1852 if (alias_idx == AliasIdxTop) return false; // the empty category
1853 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1854 if (alias_idx == AliasIdxBot) return true; // the universal category
1855 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1856
1857 // the only remaining possible overlap is identity
1858 int adr_idx = get_alias_index(adr_type);
1859 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1860 return adr_idx == alias_idx;
1861 }
1862
1863
1864
1865 //---------------------------pop_warm_call-------------------------------------
1866 WarmCallInfo* Compile::pop_warm_call() {
1867 WarmCallInfo* wci = _warm_calls;
1868 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1869 return wci;
1870 }
1871
1872 //----------------------------Inline_Warm--------------------------------------
1873 int Compile::Inline_Warm() {
1874 // If there is room, try to inline some more warm call sites.
1875 // %%% Do a graph index compaction pass when we think we're out of space?
1876 if (!InlineWarmCalls) return 0;
1877
1878 int calls_made_hot = 0;
1879 int room_to_grow = NodeCountInliningCutoff - unique();
1880 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1881 int amount_grown = 0;
1882 WarmCallInfo* call;
1883 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1884 int est_size = (int)call->size();
1885 if (est_size > (room_to_grow - amount_grown)) {
1886 // This one won't fit anyway. Get rid of it.
1887 call->make_cold();
1888 continue;
1889 }
1890 call->make_hot();
1891 calls_made_hot++;
1892 amount_grown += est_size;
1893 amount_to_grow -= est_size;
1894 }
1895
1896 if (calls_made_hot > 0) set_major_progress();
1897 return calls_made_hot;
1898 }
1899
1900
1901 //----------------------------Finish_Warm--------------------------------------
1902 void Compile::Finish_Warm() {
1903 if (!InlineWarmCalls) return;
1904 if (failing()) return;
1905 if (warm_calls() == NULL) return;
1906
1907 // Clean up loose ends, if we are out of space for inlining.
1908 WarmCallInfo* call;
1909 while ((call = pop_warm_call()) != NULL) {
1910 call->make_cold();
1911 }
1912 }
1913
1914 //---------------------cleanup_loop_predicates-----------------------
1915 // Remove the opaque nodes that protect the predicates so that all unused
1916 // checks and uncommon_traps will be eliminated from the ideal graph
1917 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1918 if (predicate_count()==0) return;
1919 for (int i = predicate_count(); i > 0; i--) {
1920 Node * n = predicate_opaque1_node(i-1);
1921 assert(n->Opcode() == Op_Opaque1, "must be");
1922 igvn.replace_node(n, n->in(1));
1923 }
1924 assert(predicate_count()==0, "should be clean!");
1925 }
1926
1927 // StringOpts and late inlining of string methods
1928 void Compile::inline_string_calls(bool parse_time) {
1929 {
1930 // remove useless nodes to make the usage analysis simpler
1931 ResourceMark rm;
1932 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1933 }
1934
1935 {
1936 ResourceMark rm;
1937 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1938 PhaseStringOpts pso(initial_gvn(), for_igvn());
1939 print_method(PHASE_AFTER_STRINGOPTS, 3);
1940 }
1941
1942 // now inline anything that we skipped the first time around
1943 if (!parse_time) {
1944 _late_inlines_pos = _late_inlines.length();
1945 }
1946
1947 while (_string_late_inlines.length() > 0) {
1948 CallGenerator* cg = _string_late_inlines.pop();
1949 cg->do_late_inline();
1950 if (failing()) return;
1951 }
1952 _string_late_inlines.trunc_to(0);
1953 }
1954
1955 // Late inlining of boxing methods
1956 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1957 if (_boxing_late_inlines.length() > 0) {
1958 assert(has_boxed_value(), "inconsistent");
1959
1960 PhaseGVN* gvn = initial_gvn();
1961 set_inlining_incrementally(true);
1962
1963 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1964 for_igvn()->clear();
1965 gvn->replace_with(&igvn);
1966
1967 _late_inlines_pos = _late_inlines.length();
1968
1969 while (_boxing_late_inlines.length() > 0) {
1970 CallGenerator* cg = _boxing_late_inlines.pop();
1971 cg->do_late_inline();
1972 if (failing()) return;
1973 }
1974 _boxing_late_inlines.trunc_to(0);
1975
1976 {
1977 ResourceMark rm;
1978 PhaseRemoveUseless pru(gvn, for_igvn());
1979 }
1980
1981 igvn = PhaseIterGVN(gvn);
1982 igvn.optimize();
1983
1984 set_inlining_progress(false);
1985 set_inlining_incrementally(false);
1986 }
1987 }
1988
1989 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1990 assert(IncrementalInline, "incremental inlining should be on");
1991 PhaseGVN* gvn = initial_gvn();
1992
1993 set_inlining_progress(false);
1994 for_igvn()->clear();
1995 gvn->replace_with(&igvn);
1996
1997 {
1998 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1999 int i = 0;
2000 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2001 CallGenerator* cg = _late_inlines.at(i);
2002 _late_inlines_pos = i+1;
2003 cg->do_late_inline();
2004 if (failing()) return;
2005 }
2006 int j = 0;
2007 for (; i < _late_inlines.length(); i++, j++) {
2008 _late_inlines.at_put(j, _late_inlines.at(i));
2009 }
2010 _late_inlines.trunc_to(j);
2011 }
2012
2013 {
2014 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2015 ResourceMark rm;
2016 PhaseRemoveUseless pru(gvn, for_igvn());
2017 }
2018
2019 {
2020 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2021 igvn = PhaseIterGVN(gvn);
2022 }
2023 }
2024
2025 // Perform incremental inlining until bound on number of live nodes is reached
2026 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2027 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2028
2029 PhaseGVN* gvn = initial_gvn();
2030
2031 set_inlining_incrementally(true);
2032 set_inlining_progress(true);
2033 uint low_live_nodes = 0;
2034
2035 while(inlining_progress() && _late_inlines.length() > 0) {
2036
2037 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2038 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2039 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2040 // PhaseIdealLoop is expensive so we only try it once we are
2041 // out of live nodes and we only try it again if the previous
2042 // helped got the number of nodes down significantly
2043 PhaseIdealLoop ideal_loop( igvn, false, true );
2044 if (failing()) return;
2045 low_live_nodes = live_nodes();
2046 _major_progress = true;
2047 }
2048
2049 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2050 break;
2051 }
2052 }
2053
2054 inline_incrementally_one(igvn);
2055
2056 if (failing()) return;
2057
2058 {
2059 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2060 igvn.optimize();
2061 }
2062
2063 if (failing()) return;
2064 }
2065
2066 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2067
2068 if (_string_late_inlines.length() > 0) {
2069 assert(has_stringbuilder(), "inconsistent");
2070 for_igvn()->clear();
2071 initial_gvn()->replace_with(&igvn);
2072
2073 inline_string_calls(false);
2074
2075 if (failing()) return;
2076
2077 {
2078 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2079 ResourceMark rm;
2080 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2081 }
2082
2083 {
2084 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2085 igvn = PhaseIterGVN(gvn);
2086 igvn.optimize();
2087 }
2088 }
2089
2090 set_inlining_incrementally(false);
2091 }
2092
2093
2094 //------------------------------Optimize---------------------------------------
2095 // Given a graph, optimize it.
2096 void Compile::Optimize() {
2097 TracePhase tp("optimizer", &timers[_t_optimizer]);
2098
2099 #ifndef PRODUCT
2100 if (_directive->BreakAtCompileOption) {
2101 BREAKPOINT;
2102 }
2103
2104 #endif
2105
2106 ResourceMark rm;
2107 int loop_opts_cnt;
2108
2109 print_inlining_reinit();
2110
2111 NOT_PRODUCT( verify_graph_edges(); )
2112
2113 print_method(PHASE_AFTER_PARSING);
2114
2115 {
2116 // Iterative Global Value Numbering, including ideal transforms
2117 // Initialize IterGVN with types and values from parse-time GVN
2118 PhaseIterGVN igvn(initial_gvn());
2119 #ifdef ASSERT
2120 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2121 #endif
2122 {
2123 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2124 igvn.optimize();
2125 }
2126
2127 print_method(PHASE_ITER_GVN1, 2);
2128
2129 if (failing()) return;
2130
2131 inline_incrementally(igvn);
2132
2133 print_method(PHASE_INCREMENTAL_INLINE, 2);
2134
2135 if (failing()) return;
2136
2137 if (eliminate_boxing()) {
2138 // Inline valueOf() methods now.
2139 inline_boxing_calls(igvn);
2140
2141 if (AlwaysIncrementalInline) {
2142 inline_incrementally(igvn);
2143 }
2144
2145 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2146
2147 if (failing()) return;
2148 }
2149
2150 // Remove the speculative part of types and clean up the graph from
2151 // the extra CastPP nodes whose only purpose is to carry them. Do
2152 // that early so that optimizations are not disrupted by the extra
2153 // CastPP nodes.
2154 remove_speculative_types(igvn);
2155
2156 // No more new expensive nodes will be added to the list from here
2157 // so keep only the actual candidates for optimizations.
2158 cleanup_expensive_nodes(igvn);
2159
2160 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2161 Compile::TracePhase tp("", &timers[_t_renumberLive]);
2162 initial_gvn()->replace_with(&igvn);
2163 for_igvn()->clear();
2164 Unique_Node_List new_worklist(C->comp_arena());
2165 {
2166 ResourceMark rm;
2167 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2168 }
2169 set_for_igvn(&new_worklist);
2170 igvn = PhaseIterGVN(initial_gvn());
2171 igvn.optimize();
2172 }
2173
2174 // Perform escape analysis
2175 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2176 if (has_loops()) {
2177 // Cleanup graph (remove dead nodes).
2178 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2179 PhaseIdealLoop ideal_loop( igvn, false, true );
2180 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2181 if (failing()) return;
2182 }
2183 ConnectionGraph::do_analysis(this, &igvn);
2184
2185 if (failing()) return;
2186
2187 // Optimize out fields loads from scalar replaceable allocations.
2188 igvn.optimize();
2189 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2190
2191 if (failing()) return;
2192
2193 if (congraph() != NULL && macro_count() > 0) {
2194 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2195 PhaseMacroExpand mexp(igvn);
2196 mexp.eliminate_macro_nodes();
2197 igvn.set_delay_transform(false);
2198
2199 igvn.optimize();
2200 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2201
2202 if (failing()) return;
2203 }
2204 }
2205
2206 // Loop transforms on the ideal graph. Range Check Elimination,
2207 // peeling, unrolling, etc.
2208
2209 // Set loop opts counter
2210 loop_opts_cnt = num_loop_opts();
2211 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2212 {
2213 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2214 PhaseIdealLoop ideal_loop( igvn, true );
2215 loop_opts_cnt--;
2216 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2217 if (failing()) return;
2218 }
2219 // Loop opts pass if partial peeling occurred in previous pass
2220 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2221 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2222 PhaseIdealLoop ideal_loop( igvn, false );
2223 loop_opts_cnt--;
2224 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2225 if (failing()) return;
2226 }
2227 // Loop opts pass for loop-unrolling before CCP
2228 if(major_progress() && (loop_opts_cnt > 0)) {
2229 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2230 PhaseIdealLoop ideal_loop( igvn, false );
2231 loop_opts_cnt--;
2232 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2233 }
2234 if (!failing()) {
2235 // Verify that last round of loop opts produced a valid graph
2236 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2237 PhaseIdealLoop::verify(igvn);
2238 }
2239 }
2240 if (failing()) return;
2241
2242 // Conditional Constant Propagation;
2243 PhaseCCP ccp( &igvn );
2244 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2245 {
2246 TracePhase tp("ccp", &timers[_t_ccp]);
2247 ccp.do_transform();
2248 }
2249 print_method(PHASE_CPP1, 2);
2250
2251 assert( true, "Break here to ccp.dump_old2new_map()");
2252
2253 // Iterative Global Value Numbering, including ideal transforms
2254 {
2255 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2256 igvn = ccp;
2257 igvn.optimize();
2258 }
2259
2260 print_method(PHASE_ITER_GVN2, 2);
2261
2262 if (failing()) return;
2263
2264 // Loop transforms on the ideal graph. Range Check Elimination,
2265 // peeling, unrolling, etc.
2266 if(loop_opts_cnt > 0) {
2267 debug_only( int cnt = 0; );
2268 while(major_progress() && (loop_opts_cnt > 0)) {
2269 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2270 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2271 PhaseIdealLoop ideal_loop( igvn, true);
2272 loop_opts_cnt--;
2273 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2274 if (failing()) return;
2275 }
2276 }
2277 // Ensure that major progress is now clear
2278 C->clear_major_progress();
2279
2280 {
2281 // Verify that all previous optimizations produced a valid graph
2282 // at least to this point, even if no loop optimizations were done.
2283 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2284 PhaseIdealLoop::verify(igvn);
2285 }
2286
2287 {
2288 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2289 PhaseMacroExpand mex(igvn);
2290 if (mex.expand_macro_nodes()) {
2291 assert(failing(), "must bail out w/ explicit message");
2292 return;
2293 }
2294 }
2295
2296 DEBUG_ONLY( _modified_nodes = NULL; )
2297 } // (End scope of igvn; run destructor if necessary for asserts.)
2298
2299 process_print_inlining();
2300 // A method with only infinite loops has no edges entering loops from root
2301 {
2302 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2303 if (final_graph_reshaping()) {
2304 assert(failing(), "must bail out w/ explicit message");
2305 return;
2306 }
2307 }
2308
2309 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2310 }
2311
2312
2313 //------------------------------Code_Gen---------------------------------------
2314 // Given a graph, generate code for it
2315 void Compile::Code_Gen() {
2316 if (failing()) {
2317 return;
2318 }
2319
2320 // Perform instruction selection. You might think we could reclaim Matcher
2321 // memory PDQ, but actually the Matcher is used in generating spill code.
2322 // Internals of the Matcher (including some VectorSets) must remain live
2323 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2324 // set a bit in reclaimed memory.
2325
2326 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2327 // nodes. Mapping is only valid at the root of each matched subtree.
2328 NOT_PRODUCT( verify_graph_edges(); )
2329
2330 Matcher matcher;
2331 _matcher = &matcher;
2332 {
2333 TracePhase tp("matcher", &timers[_t_matcher]);
2334 matcher.match();
2335 }
2336 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2337 // nodes. Mapping is only valid at the root of each matched subtree.
2338 NOT_PRODUCT( verify_graph_edges(); )
2339
2340 // If you have too many nodes, or if matching has failed, bail out
2341 check_node_count(0, "out of nodes matching instructions");
2342 if (failing()) {
2343 return;
2344 }
2345
2346 // Build a proper-looking CFG
2347 PhaseCFG cfg(node_arena(), root(), matcher);
2348 _cfg = &cfg;
2349 {
2350 TracePhase tp("scheduler", &timers[_t_scheduler]);
2351 bool success = cfg.do_global_code_motion();
2352 if (!success) {
2353 return;
2354 }
2355
2356 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2357 NOT_PRODUCT( verify_graph_edges(); )
2358 debug_only( cfg.verify(); )
2359 }
2360
2361 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2362 _regalloc = ®alloc;
2363 {
2364 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2365 // Perform register allocation. After Chaitin, use-def chains are
2366 // no longer accurate (at spill code) and so must be ignored.
2367 // Node->LRG->reg mappings are still accurate.
2368 _regalloc->Register_Allocate();
2369
2370 // Bail out if the allocator builds too many nodes
2371 if (failing()) {
2372 return;
2373 }
2374 }
2375
2376 // Prior to register allocation we kept empty basic blocks in case the
2377 // the allocator needed a place to spill. After register allocation we
2378 // are not adding any new instructions. If any basic block is empty, we
2379 // can now safely remove it.
2380 {
2381 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2382 cfg.remove_empty_blocks();
2383 if (do_freq_based_layout()) {
2384 PhaseBlockLayout layout(cfg);
2385 } else {
2386 cfg.set_loop_alignment();
2387 }
2388 cfg.fixup_flow();
2389 }
2390
2391 // Apply peephole optimizations
2392 if( OptoPeephole ) {
2393 TracePhase tp("peephole", &timers[_t_peephole]);
2394 PhasePeephole peep( _regalloc, cfg);
2395 peep.do_transform();
2396 }
2397
2398 // Do late expand if CPU requires this.
2399 if (Matcher::require_postalloc_expand) {
2400 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2401 cfg.postalloc_expand(_regalloc);
2402 }
2403
2404 // Convert Nodes to instruction bits in a buffer
2405 {
2406 TraceTime tp("output", &timers[_t_output], CITime);
2407 Output();
2408 }
2409
2410 print_method(PHASE_FINAL_CODE);
2411
2412 // He's dead, Jim.
2413 _cfg = (PhaseCFG*)0xdeadbeef;
2414 _regalloc = (PhaseChaitin*)0xdeadbeef;
2415 }
2416
2417
2418 //------------------------------dump_asm---------------------------------------
2419 // Dump formatted assembly
2420 #ifndef PRODUCT
2421 void Compile::dump_asm(int *pcs, uint pc_limit) {
2422 bool cut_short = false;
2423 tty->print_cr("#");
2424 tty->print("# "); _tf->dump(); tty->cr();
2425 tty->print_cr("#");
2426
2427 // For all blocks
2428 int pc = 0x0; // Program counter
2429 char starts_bundle = ' ';
2430 _regalloc->dump_frame();
2431
2432 Node *n = NULL;
2433 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2434 if (VMThread::should_terminate()) {
2435 cut_short = true;
2436 break;
2437 }
2438 Block* block = _cfg->get_block(i);
2439 if (block->is_connector() && !Verbose) {
2440 continue;
2441 }
2442 n = block->head();
2443 if (pcs && n->_idx < pc_limit) {
2444 tty->print("%3.3x ", pcs[n->_idx]);
2445 } else {
2446 tty->print(" ");
2447 }
2448 block->dump_head(_cfg);
2449 if (block->is_connector()) {
2450 tty->print_cr(" # Empty connector block");
2451 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2452 tty->print_cr(" # Block is sole successor of call");
2453 }
2454
2455 // For all instructions
2456 Node *delay = NULL;
2457 for (uint j = 0; j < block->number_of_nodes(); j++) {
2458 if (VMThread::should_terminate()) {
2459 cut_short = true;
2460 break;
2461 }
2462 n = block->get_node(j);
2463 if (valid_bundle_info(n)) {
2464 Bundle* bundle = node_bundling(n);
2465 if (bundle->used_in_unconditional_delay()) {
2466 delay = n;
2467 continue;
2468 }
2469 if (bundle->starts_bundle()) {
2470 starts_bundle = '+';
2471 }
2472 }
2473
2474 if (WizardMode) {
2475 n->dump();
2476 }
2477
2478 if( !n->is_Region() && // Dont print in the Assembly
2479 !n->is_Phi() && // a few noisely useless nodes
2480 !n->is_Proj() &&
2481 !n->is_MachTemp() &&
2482 !n->is_SafePointScalarObject() &&
2483 !n->is_Catch() && // Would be nice to print exception table targets
2484 !n->is_MergeMem() && // Not very interesting
2485 !n->is_top() && // Debug info table constants
2486 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2487 ) {
2488 if (pcs && n->_idx < pc_limit)
2489 tty->print("%3.3x", pcs[n->_idx]);
2490 else
2491 tty->print(" ");
2492 tty->print(" %c ", starts_bundle);
2493 starts_bundle = ' ';
2494 tty->print("\t");
2495 n->format(_regalloc, tty);
2496 tty->cr();
2497 }
2498
2499 // If we have an instruction with a delay slot, and have seen a delay,
2500 // then back up and print it
2501 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2502 assert(delay != NULL, "no unconditional delay instruction");
2503 if (WizardMode) delay->dump();
2504
2505 if (node_bundling(delay)->starts_bundle())
2506 starts_bundle = '+';
2507 if (pcs && n->_idx < pc_limit)
2508 tty->print("%3.3x", pcs[n->_idx]);
2509 else
2510 tty->print(" ");
2511 tty->print(" %c ", starts_bundle);
2512 starts_bundle = ' ';
2513 tty->print("\t");
2514 delay->format(_regalloc, tty);
2515 tty->cr();
2516 delay = NULL;
2517 }
2518
2519 // Dump the exception table as well
2520 if( n->is_Catch() && (Verbose || WizardMode) ) {
2521 // Print the exception table for this offset
2522 _handler_table.print_subtable_for(pc);
2523 }
2524 }
2525
2526 if (pcs && n->_idx < pc_limit)
2527 tty->print_cr("%3.3x", pcs[n->_idx]);
2528 else
2529 tty->cr();
2530
2531 assert(cut_short || delay == NULL, "no unconditional delay branch");
2532
2533 } // End of per-block dump
2534 tty->cr();
2535
2536 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2537 }
2538 #endif
2539
2540 //------------------------------Final_Reshape_Counts---------------------------
2541 // This class defines counters to help identify when a method
2542 // may/must be executed using hardware with only 24-bit precision.
2543 struct Final_Reshape_Counts : public StackObj {
2544 int _call_count; // count non-inlined 'common' calls
2545 int _float_count; // count float ops requiring 24-bit precision
2546 int _double_count; // count double ops requiring more precision
2547 int _java_call_count; // count non-inlined 'java' calls
2548 int _inner_loop_count; // count loops which need alignment
2549 VectorSet _visited; // Visitation flags
2550 Node_List _tests; // Set of IfNodes & PCTableNodes
2551
2552 Final_Reshape_Counts() :
2553 _call_count(0), _float_count(0), _double_count(0),
2554 _java_call_count(0), _inner_loop_count(0),
2555 _visited( Thread::current()->resource_area() ) { }
2556
2557 void inc_call_count () { _call_count ++; }
2558 void inc_float_count () { _float_count ++; }
2559 void inc_double_count() { _double_count++; }
2560 void inc_java_call_count() { _java_call_count++; }
2561 void inc_inner_loop_count() { _inner_loop_count++; }
2562
2563 int get_call_count () const { return _call_count ; }
2564 int get_float_count () const { return _float_count ; }
2565 int get_double_count() const { return _double_count; }
2566 int get_java_call_count() const { return _java_call_count; }
2567 int get_inner_loop_count() const { return _inner_loop_count; }
2568 };
2569
2570 #ifdef ASSERT
2571 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2572 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2573 // Make sure the offset goes inside the instance layout.
2574 return k->contains_field_offset(tp->offset());
2575 // Note that OffsetBot and OffsetTop are very negative.
2576 }
2577 #endif
2578
2579 // Eliminate trivially redundant StoreCMs and accumulate their
2580 // precedence edges.
2581 void Compile::eliminate_redundant_card_marks(Node* n) {
2582 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2583 if (n->in(MemNode::Address)->outcnt() > 1) {
2584 // There are multiple users of the same address so it might be
2585 // possible to eliminate some of the StoreCMs
2586 Node* mem = n->in(MemNode::Memory);
2587 Node* adr = n->in(MemNode::Address);
2588 Node* val = n->in(MemNode::ValueIn);
2589 Node* prev = n;
2590 bool done = false;
2591 // Walk the chain of StoreCMs eliminating ones that match. As
2592 // long as it's a chain of single users then the optimization is
2593 // safe. Eliminating partially redundant StoreCMs would require
2594 // cloning copies down the other paths.
2595 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2596 if (adr == mem->in(MemNode::Address) &&
2597 val == mem->in(MemNode::ValueIn)) {
2598 // redundant StoreCM
2599 if (mem->req() > MemNode::OopStore) {
2600 // Hasn't been processed by this code yet.
2601 n->add_prec(mem->in(MemNode::OopStore));
2602 } else {
2603 // Already converted to precedence edge
2604 for (uint i = mem->req(); i < mem->len(); i++) {
2605 // Accumulate any precedence edges
2606 if (mem->in(i) != NULL) {
2607 n->add_prec(mem->in(i));
2608 }
2609 }
2610 // Everything above this point has been processed.
2611 done = true;
2612 }
2613 // Eliminate the previous StoreCM
2614 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2615 assert(mem->outcnt() == 0, "should be dead");
2616 mem->disconnect_inputs(NULL, this);
2617 } else {
2618 prev = mem;
2619 }
2620 mem = prev->in(MemNode::Memory);
2621 }
2622 }
2623 }
2624
2625 //------------------------------final_graph_reshaping_impl----------------------
2626 // Implement items 1-5 from final_graph_reshaping below.
2627 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2628
2629 if ( n->outcnt() == 0 ) return; // dead node
2630 uint nop = n->Opcode();
2631
2632 // Check for 2-input instruction with "last use" on right input.
2633 // Swap to left input. Implements item (2).
2634 if( n->req() == 3 && // two-input instruction
2635 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2636 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2637 n->in(2)->outcnt() == 1 &&// right use IS a last use
2638 !n->in(2)->is_Con() ) { // right use is not a constant
2639 // Check for commutative opcode
2640 switch( nop ) {
2641 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2642 case Op_MaxI: case Op_MinI:
2643 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2644 case Op_AndL: case Op_XorL: case Op_OrL:
2645 case Op_AndI: case Op_XorI: case Op_OrI: {
2646 // Move "last use" input to left by swapping inputs
2647 n->swap_edges(1, 2);
2648 break;
2649 }
2650 default:
2651 break;
2652 }
2653 }
2654
2655 #ifdef ASSERT
2656 if( n->is_Mem() ) {
2657 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2658 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2659 // oop will be recorded in oop map if load crosses safepoint
2660 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2661 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2662 "raw memory operations should have control edge");
2663 }
2664 #endif
2665 // Count FPU ops and common calls, implements item (3)
2666 switch( nop ) {
2667 // Count all float operations that may use FPU
2668 case Op_AddF:
2669 case Op_SubF:
2670 case Op_MulF:
2671 case Op_DivF:
2672 case Op_NegF:
2673 case Op_ModF:
2674 case Op_ConvI2F:
2675 case Op_ConF:
2676 case Op_CmpF:
2677 case Op_CmpF3:
2678 // case Op_ConvL2F: // longs are split into 32-bit halves
2679 frc.inc_float_count();
2680 break;
2681
2682 case Op_ConvF2D:
2683 case Op_ConvD2F:
2684 frc.inc_float_count();
2685 frc.inc_double_count();
2686 break;
2687
2688 // Count all double operations that may use FPU
2689 case Op_AddD:
2690 case Op_SubD:
2691 case Op_MulD:
2692 case Op_DivD:
2693 case Op_NegD:
2694 case Op_ModD:
2695 case Op_ConvI2D:
2696 case Op_ConvD2I:
2697 // case Op_ConvL2D: // handled by leaf call
2698 // case Op_ConvD2L: // handled by leaf call
2699 case Op_ConD:
2700 case Op_CmpD:
2701 case Op_CmpD3:
2702 frc.inc_double_count();
2703 break;
2704 case Op_Opaque1: // Remove Opaque Nodes before matching
2705 case Op_Opaque2: // Remove Opaque Nodes before matching
2706 case Op_Opaque3:
2707 n->subsume_by(n->in(1), this);
2708 break;
2709 case Op_CallStaticJava:
2710 case Op_CallJava:
2711 case Op_CallDynamicJava:
2712 frc.inc_java_call_count(); // Count java call site;
2713 case Op_CallRuntime:
2714 case Op_CallLeaf:
2715 case Op_CallLeafNoFP: {
2716 assert( n->is_Call(), "" );
2717 CallNode *call = n->as_Call();
2718 // Count call sites where the FP mode bit would have to be flipped.
2719 // Do not count uncommon runtime calls:
2720 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2721 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2722 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2723 frc.inc_call_count(); // Count the call site
2724 } else { // See if uncommon argument is shared
2725 Node *n = call->in(TypeFunc::Parms);
2726 int nop = n->Opcode();
2727 // Clone shared simple arguments to uncommon calls, item (1).
2728 if( n->outcnt() > 1 &&
2729 !n->is_Proj() &&
2730 nop != Op_CreateEx &&
2731 nop != Op_CheckCastPP &&
2732 nop != Op_DecodeN &&
2733 nop != Op_DecodeNKlass &&
2734 !n->is_Mem() ) {
2735 Node *x = n->clone();
2736 call->set_req( TypeFunc::Parms, x );
2737 }
2738 }
2739 break;
2740 }
2741
2742 case Op_StoreD:
2743 case Op_LoadD:
2744 case Op_LoadD_unaligned:
2745 frc.inc_double_count();
2746 goto handle_mem;
2747 case Op_StoreF:
2748 case Op_LoadF:
2749 frc.inc_float_count();
2750 goto handle_mem;
2751
2752 case Op_StoreCM:
2753 {
2754 // Convert OopStore dependence into precedence edge
2755 Node* prec = n->in(MemNode::OopStore);
2756 n->del_req(MemNode::OopStore);
2757 n->add_prec(prec);
2758 eliminate_redundant_card_marks(n);
2759 }
2760
2761 // fall through
2762
2763 case Op_StoreB:
2764 case Op_StoreC:
2765 case Op_StorePConditional:
2766 case Op_StoreI:
2767 case Op_StoreL:
2768 case Op_StoreIConditional:
2769 case Op_StoreLConditional:
2770 case Op_CompareAndSwapI:
2771 case Op_CompareAndSwapL:
2772 case Op_CompareAndSwapP:
2773 case Op_CompareAndSwapN:
2774 case Op_GetAndAddI:
2775 case Op_GetAndAddL:
2776 case Op_GetAndSetI:
2777 case Op_GetAndSetL:
2778 case Op_GetAndSetP:
2779 case Op_GetAndSetN:
2780 case Op_StoreP:
2781 case Op_StoreN:
2782 case Op_StoreNKlass:
2783 case Op_LoadB:
2784 case Op_LoadUB:
2785 case Op_LoadUS:
2786 case Op_LoadI:
2787 case Op_LoadKlass:
2788 case Op_LoadNKlass:
2789 case Op_LoadL:
2790 case Op_LoadL_unaligned:
2791 case Op_LoadPLocked:
2792 case Op_LoadP:
2793 case Op_LoadN:
2794 case Op_LoadRange:
2795 case Op_LoadS: {
2796 handle_mem:
2797 #ifdef ASSERT
2798 if( VerifyOptoOopOffsets ) {
2799 assert( n->is_Mem(), "" );
2800 MemNode *mem = (MemNode*)n;
2801 // Check to see if address types have grounded out somehow.
2802 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2803 assert( !tp || oop_offset_is_sane(tp), "" );
2804 }
2805 #endif
2806 break;
2807 }
2808
2809 case Op_AddP: { // Assert sane base pointers
2810 Node *addp = n->in(AddPNode::Address);
2811 assert( !addp->is_AddP() ||
2812 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2813 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2814 "Base pointers must match" );
2815 #ifdef _LP64
2816 if ((UseCompressedOops || UseCompressedClassPointers) &&
2817 addp->Opcode() == Op_ConP &&
2818 addp == n->in(AddPNode::Base) &&
2819 n->in(AddPNode::Offset)->is_Con()) {
2820 // Use addressing with narrow klass to load with offset on x86.
2821 // On sparc loading 32-bits constant and decoding it have less
2822 // instructions (4) then load 64-bits constant (7).
2823 // Do this transformation here since IGVN will convert ConN back to ConP.
2824 const Type* t = addp->bottom_type();
2825 if (t->isa_oopptr() || t->isa_klassptr()) {
2826 Node* nn = NULL;
2827
2828 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2829
2830 // Look for existing ConN node of the same exact type.
2831 Node* r = root();
2832 uint cnt = r->outcnt();
2833 for (uint i = 0; i < cnt; i++) {
2834 Node* m = r->raw_out(i);
2835 if (m!= NULL && m->Opcode() == op &&
2836 m->bottom_type()->make_ptr() == t) {
2837 nn = m;
2838 break;
2839 }
2840 }
2841 if (nn != NULL) {
2842 // Decode a narrow oop to match address
2843 // [R12 + narrow_oop_reg<<3 + offset]
2844 if (t->isa_oopptr()) {
2845 nn = new DecodeNNode(nn, t);
2846 } else {
2847 nn = new DecodeNKlassNode(nn, t);
2848 }
2849 n->set_req(AddPNode::Base, nn);
2850 n->set_req(AddPNode::Address, nn);
2851 if (addp->outcnt() == 0) {
2852 addp->disconnect_inputs(NULL, this);
2853 }
2854 }
2855 }
2856 }
2857 #endif
2858 break;
2859 }
2860
2861 case Op_CastPP: {
2862 // Remove CastPP nodes to gain more freedom during scheduling but
2863 // keep the dependency they encode as control or precedence edges
2864 // (if control is set already) on memory operations. Some CastPP
2865 // nodes don't have a control (don't carry a dependency): skip
2866 // those.
2867 if (n->in(0) != NULL) {
2868 ResourceMark rm;
2869 Unique_Node_List wq;
2870 wq.push(n);
2871 for (uint next = 0; next < wq.size(); ++next) {
2872 Node *m = wq.at(next);
2873 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2874 Node* use = m->fast_out(i);
2875 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2876 use->ensure_control_or_add_prec(n->in(0));
2877 } else {
2878 switch(use->Opcode()) {
2879 case Op_AddP:
2880 case Op_DecodeN:
2881 case Op_DecodeNKlass:
2882 case Op_CheckCastPP:
2883 case Op_CastPP:
2884 wq.push(use);
2885 break;
2886 }
2887 }
2888 }
2889 }
2890 }
2891 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2892 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2893 Node* in1 = n->in(1);
2894 const Type* t = n->bottom_type();
2895 Node* new_in1 = in1->clone();
2896 new_in1->as_DecodeN()->set_type(t);
2897
2898 if (!Matcher::narrow_oop_use_complex_address()) {
2899 //
2900 // x86, ARM and friends can handle 2 adds in addressing mode
2901 // and Matcher can fold a DecodeN node into address by using
2902 // a narrow oop directly and do implicit NULL check in address:
2903 //
2904 // [R12 + narrow_oop_reg<<3 + offset]
2905 // NullCheck narrow_oop_reg
2906 //
2907 // On other platforms (Sparc) we have to keep new DecodeN node and
2908 // use it to do implicit NULL check in address:
2909 //
2910 // decode_not_null narrow_oop_reg, base_reg
2911 // [base_reg + offset]
2912 // NullCheck base_reg
2913 //
2914 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2915 // to keep the information to which NULL check the new DecodeN node
2916 // corresponds to use it as value in implicit_null_check().
2917 //
2918 new_in1->set_req(0, n->in(0));
2919 }
2920
2921 n->subsume_by(new_in1, this);
2922 if (in1->outcnt() == 0) {
2923 in1->disconnect_inputs(NULL, this);
2924 }
2925 } else {
2926 n->subsume_by(n->in(1), this);
2927 if (n->outcnt() == 0) {
2928 n->disconnect_inputs(NULL, this);
2929 }
2930 }
2931 break;
2932 }
2933 #ifdef _LP64
2934 case Op_CmpP:
2935 // Do this transformation here to preserve CmpPNode::sub() and
2936 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2937 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2938 Node* in1 = n->in(1);
2939 Node* in2 = n->in(2);
2940 if (!in1->is_DecodeNarrowPtr()) {
2941 in2 = in1;
2942 in1 = n->in(2);
2943 }
2944 assert(in1->is_DecodeNarrowPtr(), "sanity");
2945
2946 Node* new_in2 = NULL;
2947 if (in2->is_DecodeNarrowPtr()) {
2948 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2949 new_in2 = in2->in(1);
2950 } else if (in2->Opcode() == Op_ConP) {
2951 const Type* t = in2->bottom_type();
2952 if (t == TypePtr::NULL_PTR) {
2953 assert(in1->is_DecodeN(), "compare klass to null?");
2954 // Don't convert CmpP null check into CmpN if compressed
2955 // oops implicit null check is not generated.
2956 // This will allow to generate normal oop implicit null check.
2957 if (Matcher::gen_narrow_oop_implicit_null_checks())
2958 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
2959 //
2960 // This transformation together with CastPP transformation above
2961 // will generated code for implicit NULL checks for compressed oops.
2962 //
2963 // The original code after Optimize()
2964 //
2965 // LoadN memory, narrow_oop_reg
2966 // decode narrow_oop_reg, base_reg
2967 // CmpP base_reg, NULL
2968 // CastPP base_reg // NotNull
2969 // Load [base_reg + offset], val_reg
2970 //
2971 // after these transformations will be
2972 //
2973 // LoadN memory, narrow_oop_reg
2974 // CmpN narrow_oop_reg, NULL
2975 // decode_not_null narrow_oop_reg, base_reg
2976 // Load [base_reg + offset], val_reg
2977 //
2978 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2979 // since narrow oops can be used in debug info now (see the code in
2980 // final_graph_reshaping_walk()).
2981 //
2982 // At the end the code will be matched to
2983 // on x86:
2984 //
2985 // Load_narrow_oop memory, narrow_oop_reg
2986 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2987 // NullCheck narrow_oop_reg
2988 //
2989 // and on sparc:
2990 //
2991 // Load_narrow_oop memory, narrow_oop_reg
2992 // decode_not_null narrow_oop_reg, base_reg
2993 // Load [base_reg + offset], val_reg
2994 // NullCheck base_reg
2995 //
2996 } else if (t->isa_oopptr()) {
2997 new_in2 = ConNode::make(t->make_narrowoop());
2998 } else if (t->isa_klassptr()) {
2999 new_in2 = ConNode::make(t->make_narrowklass());
3000 }
3001 }
3002 if (new_in2 != NULL) {
3003 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3004 n->subsume_by(cmpN, this);
3005 if (in1->outcnt() == 0) {
3006 in1->disconnect_inputs(NULL, this);
3007 }
3008 if (in2->outcnt() == 0) {
3009 in2->disconnect_inputs(NULL, this);
3010 }
3011 }
3012 }
3013 break;
3014
3015 case Op_DecodeN:
3016 case Op_DecodeNKlass:
3017 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3018 // DecodeN could be pinned when it can't be fold into
3019 // an address expression, see the code for Op_CastPP above.
3020 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3021 break;
3022
3023 case Op_EncodeP:
3024 case Op_EncodePKlass: {
3025 Node* in1 = n->in(1);
3026 if (in1->is_DecodeNarrowPtr()) {
3027 n->subsume_by(in1->in(1), this);
3028 } else if (in1->Opcode() == Op_ConP) {
3029 const Type* t = in1->bottom_type();
3030 if (t == TypePtr::NULL_PTR) {
3031 assert(t->isa_oopptr(), "null klass?");
3032 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3033 } else if (t->isa_oopptr()) {
3034 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3035 } else if (t->isa_klassptr()) {
3036 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3037 }
3038 }
3039 if (in1->outcnt() == 0) {
3040 in1->disconnect_inputs(NULL, this);
3041 }
3042 break;
3043 }
3044
3045 case Op_Proj: {
3046 if (OptimizeStringConcat) {
3047 ProjNode* p = n->as_Proj();
3048 if (p->_is_io_use) {
3049 // Separate projections were used for the exception path which
3050 // are normally removed by a late inline. If it wasn't inlined
3051 // then they will hang around and should just be replaced with
3052 // the original one.
3053 Node* proj = NULL;
3054 // Replace with just one
3055 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3056 Node *use = i.get();
3057 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3058 proj = use;
3059 break;
3060 }
3061 }
3062 assert(proj != NULL, "must be found");
3063 p->subsume_by(proj, this);
3064 }
3065 }
3066 break;
3067 }
3068
3069 case Op_Phi:
3070 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3071 // The EncodeP optimization may create Phi with the same edges
3072 // for all paths. It is not handled well by Register Allocator.
3073 Node* unique_in = n->in(1);
3074 assert(unique_in != NULL, "");
3075 uint cnt = n->req();
3076 for (uint i = 2; i < cnt; i++) {
3077 Node* m = n->in(i);
3078 assert(m != NULL, "");
3079 if (unique_in != m)
3080 unique_in = NULL;
3081 }
3082 if (unique_in != NULL) {
3083 n->subsume_by(unique_in, this);
3084 }
3085 }
3086 break;
3087
3088 #endif
3089
3090 case Op_ModI:
3091 if (UseDivMod) {
3092 // Check if a%b and a/b both exist
3093 Node* d = n->find_similar(Op_DivI);
3094 if (d) {
3095 // Replace them with a fused divmod if supported
3096 if (Matcher::has_match_rule(Op_DivModI)) {
3097 DivModINode* divmod = DivModINode::make(n);
3098 d->subsume_by(divmod->div_proj(), this);
3099 n->subsume_by(divmod->mod_proj(), this);
3100 } else {
3101 // replace a%b with a-((a/b)*b)
3102 Node* mult = new MulINode(d, d->in(2));
3103 Node* sub = new SubINode(d->in(1), mult);
3104 n->subsume_by(sub, this);
3105 }
3106 }
3107 }
3108 break;
3109
3110 case Op_ModL:
3111 if (UseDivMod) {
3112 // Check if a%b and a/b both exist
3113 Node* d = n->find_similar(Op_DivL);
3114 if (d) {
3115 // Replace them with a fused divmod if supported
3116 if (Matcher::has_match_rule(Op_DivModL)) {
3117 DivModLNode* divmod = DivModLNode::make(n);
3118 d->subsume_by(divmod->div_proj(), this);
3119 n->subsume_by(divmod->mod_proj(), this);
3120 } else {
3121 // replace a%b with a-((a/b)*b)
3122 Node* mult = new MulLNode(d, d->in(2));
3123 Node* sub = new SubLNode(d->in(1), mult);
3124 n->subsume_by(sub, this);
3125 }
3126 }
3127 }
3128 break;
3129
3130 case Op_LoadVector:
3131 case Op_StoreVector:
3132 break;
3133
3134 case Op_AddReductionVI:
3135 case Op_AddReductionVL:
3136 case Op_AddReductionVF:
3137 case Op_AddReductionVD:
3138 case Op_MulReductionVI:
3139 case Op_MulReductionVL:
3140 case Op_MulReductionVF:
3141 case Op_MulReductionVD:
3142 break;
3143
3144 case Op_PackB:
3145 case Op_PackS:
3146 case Op_PackI:
3147 case Op_PackF:
3148 case Op_PackL:
3149 case Op_PackD:
3150 if (n->req()-1 > 2) {
3151 // Replace many operand PackNodes with a binary tree for matching
3152 PackNode* p = (PackNode*) n;
3153 Node* btp = p->binary_tree_pack(1, n->req());
3154 n->subsume_by(btp, this);
3155 }
3156 break;
3157 case Op_Loop:
3158 case Op_CountedLoop:
3159 if (n->as_Loop()->is_inner_loop()) {
3160 frc.inc_inner_loop_count();
3161 }
3162 break;
3163 case Op_LShiftI:
3164 case Op_RShiftI:
3165 case Op_URShiftI:
3166 case Op_LShiftL:
3167 case Op_RShiftL:
3168 case Op_URShiftL:
3169 if (Matcher::need_masked_shift_count) {
3170 // The cpu's shift instructions don't restrict the count to the
3171 // lower 5/6 bits. We need to do the masking ourselves.
3172 Node* in2 = n->in(2);
3173 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3174 const TypeInt* t = in2->find_int_type();
3175 if (t != NULL && t->is_con()) {
3176 juint shift = t->get_con();
3177 if (shift > mask) { // Unsigned cmp
3178 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3179 }
3180 } else {
3181 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3182 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3183 n->set_req(2, shift);
3184 }
3185 }
3186 if (in2->outcnt() == 0) { // Remove dead node
3187 in2->disconnect_inputs(NULL, this);
3188 }
3189 }
3190 break;
3191 case Op_MemBarStoreStore:
3192 case Op_MemBarRelease:
3193 // Break the link with AllocateNode: it is no longer useful and
3194 // confuses register allocation.
3195 if (n->req() > MemBarNode::Precedent) {
3196 n->set_req(MemBarNode::Precedent, top());
3197 }
3198 break;
3199 case Op_RangeCheck: {
3200 RangeCheckNode* rc = n->as_RangeCheck();
3201 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3202 n->subsume_by(iff, this);
3203 frc._tests.push(iff);
3204 break;
3205 }
3206 default:
3207 assert( !n->is_Call(), "" );
3208 assert( !n->is_Mem(), "" );
3209 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3210 break;
3211 }
3212
3213 // Collect CFG split points
3214 if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3215 frc._tests.push(n);
3216 }
3217 }
3218
3219 //------------------------------final_graph_reshaping_walk---------------------
3220 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3221 // requires that the walk visits a node's inputs before visiting the node.
3222 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3223 ResourceArea *area = Thread::current()->resource_area();
3224 Unique_Node_List sfpt(area);
3225
3226 frc._visited.set(root->_idx); // first, mark node as visited
3227 uint cnt = root->req();
3228 Node *n = root;
3229 uint i = 0;
3230 while (true) {
3231 if (i < cnt) {
3232 // Place all non-visited non-null inputs onto stack
3233 Node* m = n->in(i);
3234 ++i;
3235 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3236 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3237 // compute worst case interpreter size in case of a deoptimization
3238 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3239
3240 sfpt.push(m);
3241 }
3242 cnt = m->req();
3243 nstack.push(n, i); // put on stack parent and next input's index
3244 n = m;
3245 i = 0;
3246 }
3247 } else {
3248 // Now do post-visit work
3249 final_graph_reshaping_impl( n, frc );
3250 if (nstack.is_empty())
3251 break; // finished
3252 n = nstack.node(); // Get node from stack
3253 cnt = n->req();
3254 i = nstack.index();
3255 nstack.pop(); // Shift to the next node on stack
3256 }
3257 }
3258
3259 // Skip next transformation if compressed oops are not used.
3260 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3261 (!UseCompressedOops && !UseCompressedClassPointers))
3262 return;
3263
3264 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3265 // It could be done for an uncommon traps or any safepoints/calls
3266 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3267 while (sfpt.size() > 0) {
3268 n = sfpt.pop();
3269 JVMState *jvms = n->as_SafePoint()->jvms();
3270 assert(jvms != NULL, "sanity");
3271 int start = jvms->debug_start();
3272 int end = n->req();
3273 bool is_uncommon = (n->is_CallStaticJava() &&
3274 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3275 for (int j = start; j < end; j++) {
3276 Node* in = n->in(j);
3277 if (in->is_DecodeNarrowPtr()) {
3278 bool safe_to_skip = true;
3279 if (!is_uncommon ) {
3280 // Is it safe to skip?
3281 for (uint i = 0; i < in->outcnt(); i++) {
3282 Node* u = in->raw_out(i);
3283 if (!u->is_SafePoint() ||
3284 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3285 safe_to_skip = false;
3286 }
3287 }
3288 }
3289 if (safe_to_skip) {
3290 n->set_req(j, in->in(1));
3291 }
3292 if (in->outcnt() == 0) {
3293 in->disconnect_inputs(NULL, this);
3294 }
3295 }
3296 }
3297 }
3298 }
3299
3300 //------------------------------final_graph_reshaping--------------------------
3301 // Final Graph Reshaping.
3302 //
3303 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3304 // and not commoned up and forced early. Must come after regular
3305 // optimizations to avoid GVN undoing the cloning. Clone constant
3306 // inputs to Loop Phis; these will be split by the allocator anyways.
3307 // Remove Opaque nodes.
3308 // (2) Move last-uses by commutative operations to the left input to encourage
3309 // Intel update-in-place two-address operations and better register usage
3310 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3311 // calls canonicalizing them back.
3312 // (3) Count the number of double-precision FP ops, single-precision FP ops
3313 // and call sites. On Intel, we can get correct rounding either by
3314 // forcing singles to memory (requires extra stores and loads after each
3315 // FP bytecode) or we can set a rounding mode bit (requires setting and
3316 // clearing the mode bit around call sites). The mode bit is only used
3317 // if the relative frequency of single FP ops to calls is low enough.
3318 // This is a key transform for SPEC mpeg_audio.
3319 // (4) Detect infinite loops; blobs of code reachable from above but not
3320 // below. Several of the Code_Gen algorithms fail on such code shapes,
3321 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3322 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3323 // Detection is by looking for IfNodes where only 1 projection is
3324 // reachable from below or CatchNodes missing some targets.
3325 // (5) Assert for insane oop offsets in debug mode.
3326
3327 bool Compile::final_graph_reshaping() {
3328 // an infinite loop may have been eliminated by the optimizer,
3329 // in which case the graph will be empty.
3330 if (root()->req() == 1) {
3331 record_method_not_compilable("trivial infinite loop");
3332 return true;
3333 }
3334
3335 // Expensive nodes have their control input set to prevent the GVN
3336 // from freely commoning them. There's no GVN beyond this point so
3337 // no need to keep the control input. We want the expensive nodes to
3338 // be freely moved to the least frequent code path by gcm.
3339 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3340 for (int i = 0; i < expensive_count(); i++) {
3341 _expensive_nodes->at(i)->set_req(0, NULL);
3342 }
3343
3344 Final_Reshape_Counts frc;
3345
3346 // Visit everybody reachable!
3347 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3348 Node_Stack nstack(live_nodes() >> 1);
3349 final_graph_reshaping_walk(nstack, root(), frc);
3350
3351 // Check for unreachable (from below) code (i.e., infinite loops).
3352 for( uint i = 0; i < frc._tests.size(); i++ ) {
3353 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3354 // Get number of CFG targets.
3355 // Note that PCTables include exception targets after calls.
3356 uint required_outcnt = n->required_outcnt();
3357 if (n->outcnt() != required_outcnt) {
3358 // Check for a few special cases. Rethrow Nodes never take the
3359 // 'fall-thru' path, so expected kids is 1 less.
3360 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3361 if (n->in(0)->in(0)->is_Call()) {
3362 CallNode *call = n->in(0)->in(0)->as_Call();
3363 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3364 required_outcnt--; // Rethrow always has 1 less kid
3365 } else if (call->req() > TypeFunc::Parms &&
3366 call->is_CallDynamicJava()) {
3367 // Check for null receiver. In such case, the optimizer has
3368 // detected that the virtual call will always result in a null
3369 // pointer exception. The fall-through projection of this CatchNode
3370 // will not be populated.
3371 Node *arg0 = call->in(TypeFunc::Parms);
3372 if (arg0->is_Type() &&
3373 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3374 required_outcnt--;
3375 }
3376 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3377 call->req() > TypeFunc::Parms+1 &&
3378 call->is_CallStaticJava()) {
3379 // Check for negative array length. In such case, the optimizer has
3380 // detected that the allocation attempt will always result in an
3381 // exception. There is no fall-through projection of this CatchNode .
3382 Node *arg1 = call->in(TypeFunc::Parms+1);
3383 if (arg1->is_Type() &&
3384 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3385 required_outcnt--;
3386 }
3387 }
3388 }
3389 }
3390 // Recheck with a better notion of 'required_outcnt'
3391 if (n->outcnt() != required_outcnt) {
3392 record_method_not_compilable("malformed control flow");
3393 return true; // Not all targets reachable!
3394 }
3395 }
3396 // Check that I actually visited all kids. Unreached kids
3397 // must be infinite loops.
3398 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3399 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3400 record_method_not_compilable("infinite loop");
3401 return true; // Found unvisited kid; must be unreach
3402 }
3403 }
3404
3405 // If original bytecodes contained a mixture of floats and doubles
3406 // check if the optimizer has made it homogenous, item (3).
3407 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3408 frc.get_float_count() > 32 &&
3409 frc.get_double_count() == 0 &&
3410 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3411 set_24_bit_selection_and_mode( false, true );
3412 }
3413
3414 set_java_calls(frc.get_java_call_count());
3415 set_inner_loops(frc.get_inner_loop_count());
3416
3417 // No infinite loops, no reason to bail out.
3418 return false;
3419 }
3420
3421 //-----------------------------too_many_traps----------------------------------
3422 // Report if there are too many traps at the current method and bci.
3423 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3424 bool Compile::too_many_traps(ciMethod* method,
3425 int bci,
3426 Deoptimization::DeoptReason reason) {
3427 ciMethodData* md = method->method_data();
3428 if (md->is_empty()) {
3429 // Assume the trap has not occurred, or that it occurred only
3430 // because of a transient condition during start-up in the interpreter.
3431 return false;
3432 }
3433 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3434 if (md->has_trap_at(bci, m, reason) != 0) {
3435 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3436 // Also, if there are multiple reasons, or if there is no per-BCI record,
3437 // assume the worst.
3438 if (log())
3439 log()->elem("observe trap='%s' count='%d'",
3440 Deoptimization::trap_reason_name(reason),
3441 md->trap_count(reason));
3442 return true;
3443 } else {
3444 // Ignore method/bci and see if there have been too many globally.
3445 return too_many_traps(reason, md);
3446 }
3447 }
3448
3449 // Less-accurate variant which does not require a method and bci.
3450 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3451 ciMethodData* logmd) {
3452 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3453 // Too many traps globally.
3454 // Note that we use cumulative trap_count, not just md->trap_count.
3455 if (log()) {
3456 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3457 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3458 Deoptimization::trap_reason_name(reason),
3459 mcount, trap_count(reason));
3460 }
3461 return true;
3462 } else {
3463 // The coast is clear.
3464 return false;
3465 }
3466 }
3467
3468 //--------------------------too_many_recompiles--------------------------------
3469 // Report if there are too many recompiles at the current method and bci.
3470 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3471 // Is not eager to return true, since this will cause the compiler to use
3472 // Action_none for a trap point, to avoid too many recompilations.
3473 bool Compile::too_many_recompiles(ciMethod* method,
3474 int bci,
3475 Deoptimization::DeoptReason reason) {
3476 ciMethodData* md = method->method_data();
3477 if (md->is_empty()) {
3478 // Assume the trap has not occurred, or that it occurred only
3479 // because of a transient condition during start-up in the interpreter.
3480 return false;
3481 }
3482 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3483 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3484 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3485 Deoptimization::DeoptReason per_bc_reason
3486 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3487 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3488 if ((per_bc_reason == Deoptimization::Reason_none
3489 || md->has_trap_at(bci, m, reason) != 0)
3490 // The trap frequency measure we care about is the recompile count:
3491 && md->trap_recompiled_at(bci, m)
3492 && md->overflow_recompile_count() >= bc_cutoff) {
3493 // Do not emit a trap here if it has already caused recompilations.
3494 // Also, if there are multiple reasons, or if there is no per-BCI record,
3495 // assume the worst.
3496 if (log())
3497 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3498 Deoptimization::trap_reason_name(reason),
3499 md->trap_count(reason),
3500 md->overflow_recompile_count());
3501 return true;
3502 } else if (trap_count(reason) != 0
3503 && decompile_count() >= m_cutoff) {
3504 // Too many recompiles globally, and we have seen this sort of trap.
3505 // Use cumulative decompile_count, not just md->decompile_count.
3506 if (log())
3507 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3508 Deoptimization::trap_reason_name(reason),
3509 md->trap_count(reason), trap_count(reason),
3510 md->decompile_count(), decompile_count());
3511 return true;
3512 } else {
3513 // The coast is clear.
3514 return false;
3515 }
3516 }
3517
3518 // Compute when not to trap. Used by matching trap based nodes and
3519 // NullCheck optimization.
3520 void Compile::set_allowed_deopt_reasons() {
3521 _allowed_reasons = 0;
3522 if (is_method_compilation()) {
3523 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3524 assert(rs < BitsPerInt, "recode bit map");
3525 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3526 _allowed_reasons |= nth_bit(rs);
3527 }
3528 }
3529 }
3530 }
3531
3532 #ifndef PRODUCT
3533 //------------------------------verify_graph_edges---------------------------
3534 // Walk the Graph and verify that there is a one-to-one correspondence
3535 // between Use-Def edges and Def-Use edges in the graph.
3536 void Compile::verify_graph_edges(bool no_dead_code) {
3537 if (VerifyGraphEdges) {
3538 ResourceArea *area = Thread::current()->resource_area();
3539 Unique_Node_List visited(area);
3540 // Call recursive graph walk to check edges
3541 _root->verify_edges(visited);
3542 if (no_dead_code) {
3543 // Now make sure that no visited node is used by an unvisited node.
3544 bool dead_nodes = false;
3545 Unique_Node_List checked(area);
3546 while (visited.size() > 0) {
3547 Node* n = visited.pop();
3548 checked.push(n);
3549 for (uint i = 0; i < n->outcnt(); i++) {
3550 Node* use = n->raw_out(i);
3551 if (checked.member(use)) continue; // already checked
3552 if (visited.member(use)) continue; // already in the graph
3553 if (use->is_Con()) continue; // a dead ConNode is OK
3554 // At this point, we have found a dead node which is DU-reachable.
3555 if (!dead_nodes) {
3556 tty->print_cr("*** Dead nodes reachable via DU edges:");
3557 dead_nodes = true;
3558 }
3559 use->dump(2);
3560 tty->print_cr("---");
3561 checked.push(use); // No repeats; pretend it is now checked.
3562 }
3563 }
3564 assert(!dead_nodes, "using nodes must be reachable from root");
3565 }
3566 }
3567 }
3568
3569 // Verify GC barriers consistency
3570 // Currently supported:
3571 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3572 void Compile::verify_barriers() {
3573 if (UseG1GC) {
3574 // Verify G1 pre-barriers
3575 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3576
3577 ResourceArea *area = Thread::current()->resource_area();
3578 Unique_Node_List visited(area);
3579 Node_List worklist(area);
3580 // We're going to walk control flow backwards starting from the Root
3581 worklist.push(_root);
3582 while (worklist.size() > 0) {
3583 Node* x = worklist.pop();
3584 if (x == NULL || x == top()) continue;
3585 if (visited.member(x)) {
3586 continue;
3587 } else {
3588 visited.push(x);
3589 }
3590
3591 if (x->is_Region()) {
3592 for (uint i = 1; i < x->req(); i++) {
3593 worklist.push(x->in(i));
3594 }
3595 } else {
3596 worklist.push(x->in(0));
3597 // We are looking for the pattern:
3598 // /->ThreadLocal
3599 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3600 // \->ConI(0)
3601 // We want to verify that the If and the LoadB have the same control
3602 // See GraphKit::g1_write_barrier_pre()
3603 if (x->is_If()) {
3604 IfNode *iff = x->as_If();
3605 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3606 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3607 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3608 && cmp->in(1)->is_Load()) {
3609 LoadNode* load = cmp->in(1)->as_Load();
3610 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3611 && load->in(2)->in(3)->is_Con()
3612 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3613
3614 Node* if_ctrl = iff->in(0);
3615 Node* load_ctrl = load->in(0);
3616
3617 if (if_ctrl != load_ctrl) {
3618 // Skip possible CProj->NeverBranch in infinite loops
3619 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3620 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3621 if_ctrl = if_ctrl->in(0)->in(0);
3622 }
3623 }
3624 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3625 }
3626 }
3627 }
3628 }
3629 }
3630 }
3631 }
3632 }
3633
3634 #endif
3635
3636 // The Compile object keeps track of failure reasons separately from the ciEnv.
3637 // This is required because there is not quite a 1-1 relation between the
3638 // ciEnv and its compilation task and the Compile object. Note that one
3639 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3640 // to backtrack and retry without subsuming loads. Other than this backtracking
3641 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3642 // by the logic in C2Compiler.
3643 void Compile::record_failure(const char* reason) {
3644 if (log() != NULL) {
3645 log()->elem("failure reason='%s' phase='compile'", reason);
3646 }
3647 if (_failure_reason == NULL) {
3648 // Record the first failure reason.
3649 _failure_reason = reason;
3650 }
3651
3652 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3653 C->print_method(PHASE_FAILURE);
3654 }
3655 _root = NULL; // flush the graph, too
3656 }
3657
3658 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3659 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3660 _phase_name(name), _dolog(CITimeVerbose)
3661 {
3662 if (_dolog) {
3663 C = Compile::current();
3664 _log = C->log();
3665 } else {
3666 C = NULL;
3667 _log = NULL;
3668 }
3669 if (_log != NULL) {
3670 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3671 _log->stamp();
3672 _log->end_head();
3673 }
3674 }
3675
3676 Compile::TracePhase::~TracePhase() {
3677
3678 C = Compile::current();
3679 if (_dolog) {
3680 _log = C->log();
3681 } else {
3682 _log = NULL;
3683 }
3684
3685 #ifdef ASSERT
3686 if (PrintIdealNodeCount) {
3687 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3688 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3689 }
3690
3691 if (VerifyIdealNodeCount) {
3692 Compile::current()->print_missing_nodes();
3693 }
3694 #endif
3695
3696 if (_log != NULL) {
3697 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3698 }
3699 }
3700
3701 //=============================================================================
3702 // Two Constant's are equal when the type and the value are equal.
3703 bool Compile::Constant::operator==(const Constant& other) {
3704 if (type() != other.type() ) return false;
3705 if (can_be_reused() != other.can_be_reused()) return false;
3706 // For floating point values we compare the bit pattern.
3707 switch (type()) {
3708 case T_FLOAT: return (_v._value.i == other._v._value.i);
3709 case T_LONG:
3710 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3711 case T_OBJECT:
3712 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3713 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3714 case T_METADATA: return (_v._metadata == other._v._metadata);
3715 default: ShouldNotReachHere();
3716 }
3717 return false;
3718 }
3719
3720 static int type_to_size_in_bytes(BasicType t) {
3721 switch (t) {
3722 case T_LONG: return sizeof(jlong );
3723 case T_FLOAT: return sizeof(jfloat );
3724 case T_DOUBLE: return sizeof(jdouble);
3725 case T_METADATA: return sizeof(Metadata*);
3726 // We use T_VOID as marker for jump-table entries (labels) which
3727 // need an internal word relocation.
3728 case T_VOID:
3729 case T_ADDRESS:
3730 case T_OBJECT: return sizeof(jobject);
3731 }
3732
3733 ShouldNotReachHere();
3734 return -1;
3735 }
3736
3737 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3738 // sort descending
3739 if (a->freq() > b->freq()) return -1;
3740 if (a->freq() < b->freq()) return 1;
3741 return 0;
3742 }
3743
3744 void Compile::ConstantTable::calculate_offsets_and_size() {
3745 // First, sort the array by frequencies.
3746 _constants.sort(qsort_comparator);
3747
3748 #ifdef ASSERT
3749 // Make sure all jump-table entries were sorted to the end of the
3750 // array (they have a negative frequency).
3751 bool found_void = false;
3752 for (int i = 0; i < _constants.length(); i++) {
3753 Constant con = _constants.at(i);
3754 if (con.type() == T_VOID)
3755 found_void = true; // jump-tables
3756 else
3757 assert(!found_void, "wrong sorting");
3758 }
3759 #endif
3760
3761 int offset = 0;
3762 for (int i = 0; i < _constants.length(); i++) {
3763 Constant* con = _constants.adr_at(i);
3764
3765 // Align offset for type.
3766 int typesize = type_to_size_in_bytes(con->type());
3767 offset = align_size_up(offset, typesize);
3768 con->set_offset(offset); // set constant's offset
3769
3770 if (con->type() == T_VOID) {
3771 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3772 offset = offset + typesize * n->outcnt(); // expand jump-table
3773 } else {
3774 offset = offset + typesize;
3775 }
3776 }
3777
3778 // Align size up to the next section start (which is insts; see
3779 // CodeBuffer::align_at_start).
3780 assert(_size == -1, "already set?");
3781 _size = align_size_up(offset, CodeEntryAlignment);
3782 }
3783
3784 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3785 MacroAssembler _masm(&cb);
3786 for (int i = 0; i < _constants.length(); i++) {
3787 Constant con = _constants.at(i);
3788 address constant_addr = NULL;
3789 switch (con.type()) {
3790 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3791 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3792 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3793 case T_OBJECT: {
3794 jobject obj = con.get_jobject();
3795 int oop_index = _masm.oop_recorder()->find_index(obj);
3796 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3797 break;
3798 }
3799 case T_ADDRESS: {
3800 address addr = (address) con.get_jobject();
3801 constant_addr = _masm.address_constant(addr);
3802 break;
3803 }
3804 // We use T_VOID as marker for jump-table entries (labels) which
3805 // need an internal word relocation.
3806 case T_VOID: {
3807 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3808 // Fill the jump-table with a dummy word. The real value is
3809 // filled in later in fill_jump_table.
3810 address dummy = (address) n;
3811 constant_addr = _masm.address_constant(dummy);
3812 // Expand jump-table
3813 for (uint i = 1; i < n->outcnt(); i++) {
3814 address temp_addr = _masm.address_constant(dummy + i);
3815 assert(temp_addr, "consts section too small");
3816 }
3817 break;
3818 }
3819 case T_METADATA: {
3820 Metadata* obj = con.get_metadata();
3821 int metadata_index = _masm.oop_recorder()->find_index(obj);
3822 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3823 break;
3824 }
3825 default: ShouldNotReachHere();
3826 }
3827 assert(constant_addr, "consts section too small");
3828 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3829 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
3830 }
3831 }
3832
3833 int Compile::ConstantTable::find_offset(Constant& con) const {
3834 int idx = _constants.find(con);
3835 assert(idx != -1, "constant must be in constant table");
3836 int offset = _constants.at(idx).offset();
3837 assert(offset != -1, "constant table not emitted yet?");
3838 return offset;
3839 }
3840
3841 void Compile::ConstantTable::add(Constant& con) {
3842 if (con.can_be_reused()) {
3843 int idx = _constants.find(con);
3844 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3845 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3846 return;
3847 }
3848 }
3849 (void) _constants.append(con);
3850 }
3851
3852 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3853 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3854 Constant con(type, value, b->_freq);
3855 add(con);
3856 return con;
3857 }
3858
3859 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3860 Constant con(metadata);
3861 add(con);
3862 return con;
3863 }
3864
3865 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3866 jvalue value;
3867 BasicType type = oper->type()->basic_type();
3868 switch (type) {
3869 case T_LONG: value.j = oper->constantL(); break;
3870 case T_FLOAT: value.f = oper->constantF(); break;
3871 case T_DOUBLE: value.d = oper->constantD(); break;
3872 case T_OBJECT:
3873 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3874 case T_METADATA: return add((Metadata*)oper->constant()); break;
3875 default: guarantee(false, "unhandled type: %s", type2name(type));
3876 }
3877 return add(n, type, value);
3878 }
3879
3880 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3881 jvalue value;
3882 // We can use the node pointer here to identify the right jump-table
3883 // as this method is called from Compile::Fill_buffer right before
3884 // the MachNodes are emitted and the jump-table is filled (means the
3885 // MachNode pointers do not change anymore).
3886 value.l = (jobject) n;
3887 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3888 add(con);
3889 return con;
3890 }
3891
3892 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3893 // If called from Compile::scratch_emit_size do nothing.
3894 if (Compile::current()->in_scratch_emit_size()) return;
3895
3896 assert(labels.is_nonempty(), "must be");
3897 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
3898
3899 // Since MachConstantNode::constant_offset() also contains
3900 // table_base_offset() we need to subtract the table_base_offset()
3901 // to get the plain offset into the constant table.
3902 int offset = n->constant_offset() - table_base_offset();
3903
3904 MacroAssembler _masm(&cb);
3905 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3906
3907 for (uint i = 0; i < n->outcnt(); i++) {
3908 address* constant_addr = &jump_table_base[i];
3909 assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i));
3910 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3911 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3912 }
3913 }
3914
3915 //----------------------------static_subtype_check-----------------------------
3916 // Shortcut important common cases when superklass is exact:
3917 // (0) superklass is java.lang.Object (can occur in reflective code)
3918 // (1) subklass is already limited to a subtype of superklass => always ok
3919 // (2) subklass does not overlap with superklass => always fail
3920 // (3) superklass has NO subtypes and we can check with a simple compare.
3921 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3922 if (StressReflectiveCode) {
3923 return SSC_full_test; // Let caller generate the general case.
3924 }
3925
3926 if (superk == env()->Object_klass()) {
3927 return SSC_always_true; // (0) this test cannot fail
3928 }
3929
3930 ciType* superelem = superk;
3931 if (superelem->is_array_klass())
3932 superelem = superelem->as_array_klass()->base_element_type();
3933
3934 if (!subk->is_interface()) { // cannot trust static interface types yet
3935 if (subk->is_subtype_of(superk)) {
3936 return SSC_always_true; // (1) false path dead; no dynamic test needed
3937 }
3938 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
3939 !superk->is_subtype_of(subk)) {
3940 return SSC_always_false;
3941 }
3942 }
3943
3944 // If casting to an instance klass, it must have no subtypes
3945 if (superk->is_interface()) {
3946 // Cannot trust interfaces yet.
3947 // %%% S.B. superk->nof_implementors() == 1
3948 } else if (superelem->is_instance_klass()) {
3949 ciInstanceKlass* ik = superelem->as_instance_klass();
3950 if (!ik->has_subklass() && !ik->is_interface()) {
3951 if (!ik->is_final()) {
3952 // Add a dependency if there is a chance of a later subclass.
3953 dependencies()->assert_leaf_type(ik);
3954 }
3955 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3956 }
3957 } else {
3958 // A primitive array type has no subtypes.
3959 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3960 }
3961
3962 return SSC_full_test;
3963 }
3964
3965 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) {
3966 #ifdef _LP64
3967 // The scaled index operand to AddP must be a clean 64-bit value.
3968 // Java allows a 32-bit int to be incremented to a negative
3969 // value, which appears in a 64-bit register as a large
3970 // positive number. Using that large positive number as an
3971 // operand in pointer arithmetic has bad consequences.
3972 // On the other hand, 32-bit overflow is rare, and the possibility
3973 // can often be excluded, if we annotate the ConvI2L node with
3974 // a type assertion that its value is known to be a small positive
3975 // number. (The prior range check has ensured this.)
3976 // This assertion is used by ConvI2LNode::Ideal.
3977 int index_max = max_jint - 1; // array size is max_jint, index is one less
3978 if (sizetype != NULL) index_max = sizetype->_hi - 1;
3979 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
3980 idx = phase->transform(new ConvI2LNode(idx, lidxtype));
3981 #endif
3982 return idx;
3983 }
3984
3985 // The message about the current inlining is accumulated in
3986 // _print_inlining_stream and transfered into the _print_inlining_list
3987 // once we know whether inlining succeeds or not. For regular
3988 // inlining, messages are appended to the buffer pointed by
3989 // _print_inlining_idx in the _print_inlining_list. For late inlining,
3990 // a new buffer is added after _print_inlining_idx in the list. This
3991 // way we can update the inlining message for late inlining call site
3992 // when the inlining is attempted again.
3993 void Compile::print_inlining_init() {
3994 if (print_inlining() || print_intrinsics()) {
3995 _print_inlining_stream = new stringStream();
3996 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
3997 }
3998 }
3999
4000 void Compile::print_inlining_reinit() {
4001 if (print_inlining() || print_intrinsics()) {
4002 // Re allocate buffer when we change ResourceMark
4003 _print_inlining_stream = new stringStream();
4004 }
4005 }
4006
4007 void Compile::print_inlining_reset() {
4008 _print_inlining_stream->reset();
4009 }
4010
4011 void Compile::print_inlining_commit() {
4012 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4013 // Transfer the message from _print_inlining_stream to the current
4014 // _print_inlining_list buffer and clear _print_inlining_stream.
4015 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4016 print_inlining_reset();
4017 }
4018
4019 void Compile::print_inlining_push() {
4020 // Add new buffer to the _print_inlining_list at current position
4021 _print_inlining_idx++;
4022 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4023 }
4024
4025 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4026 return _print_inlining_list->at(_print_inlining_idx);
4027 }
4028
4029 void Compile::print_inlining_update(CallGenerator* cg) {
4030 if (print_inlining() || print_intrinsics()) {
4031 if (!cg->is_late_inline()) {
4032 if (print_inlining_current().cg() != NULL) {
4033 print_inlining_push();
4034 }
4035 print_inlining_commit();
4036 } else {
4037 if (print_inlining_current().cg() != cg &&
4038 (print_inlining_current().cg() != NULL ||
4039 print_inlining_current().ss()->size() != 0)) {
4040 print_inlining_push();
4041 }
4042 print_inlining_commit();
4043 print_inlining_current().set_cg(cg);
4044 }
4045 }
4046 }
4047
4048 void Compile::print_inlining_move_to(CallGenerator* cg) {
4049 // We resume inlining at a late inlining call site. Locate the
4050 // corresponding inlining buffer so that we can update it.
4051 if (print_inlining()) {
4052 for (int i = 0; i < _print_inlining_list->length(); i++) {
4053 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4054 _print_inlining_idx = i;
4055 return;
4056 }
4057 }
4058 ShouldNotReachHere();
4059 }
4060 }
4061
4062 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4063 if (print_inlining()) {
4064 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4065 assert(print_inlining_current().cg() == cg, "wrong entry");
4066 // replace message with new message
4067 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4068 print_inlining_commit();
4069 print_inlining_current().set_cg(cg);
4070 }
4071 }
4072
4073 void Compile::print_inlining_assert_ready() {
4074 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4075 }
4076
4077 void Compile::process_print_inlining() {
4078 bool do_print_inlining = print_inlining() || print_intrinsics();
4079 if (do_print_inlining || log() != NULL) {
4080 // Print inlining message for candidates that we couldn't inline
4081 // for lack of space
4082 for (int i = 0; i < _late_inlines.length(); i++) {
4083 CallGenerator* cg = _late_inlines.at(i);
4084 if (!cg->is_mh_late_inline()) {
4085 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4086 if (do_print_inlining) {
4087 cg->print_inlining_late(msg);
4088 }
4089 log_late_inline_failure(cg, msg);
4090 }
4091 }
4092 }
4093 if (do_print_inlining) {
4094 ResourceMark rm;
4095 stringStream ss;
4096 for (int i = 0; i < _print_inlining_list->length(); i++) {
4097 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4098 }
4099 size_t end = ss.size();
4100 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4101 strncpy(_print_inlining_output, ss.base(), end+1);
4102 _print_inlining_output[end] = 0;
4103 }
4104 }
4105
4106 void Compile::dump_print_inlining() {
4107 if (_print_inlining_output != NULL) {
4108 tty->print_raw(_print_inlining_output);
4109 }
4110 }
4111
4112 void Compile::log_late_inline(CallGenerator* cg) {
4113 if (log() != NULL) {
4114 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4115 cg->unique_id());
4116 JVMState* p = cg->call_node()->jvms();
4117 while (p != NULL) {
4118 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4119 p = p->caller();
4120 }
4121 log()->tail("late_inline");
4122 }
4123 }
4124
4125 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4126 log_late_inline(cg);
4127 if (log() != NULL) {
4128 log()->inline_fail(msg);
4129 }
4130 }
4131
4132 void Compile::log_inline_id(CallGenerator* cg) {
4133 if (log() != NULL) {
4134 // The LogCompilation tool needs a unique way to identify late
4135 // inline call sites. This id must be unique for this call site in
4136 // this compilation. Try to have it unique across compilations as
4137 // well because it can be convenient when grepping through the log
4138 // file.
4139 // Distinguish OSR compilations from others in case CICountOSR is
4140 // on.
4141 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4142 cg->set_unique_id(id);
4143 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4144 }
4145 }
4146
4147 void Compile::log_inline_failure(const char* msg) {
4148 if (C->log() != NULL) {
4149 C->log()->inline_fail(msg);
4150 }
4151 }
4152
4153
4154 // Dump inlining replay data to the stream.
4155 // Don't change thread state and acquire any locks.
4156 void Compile::dump_inline_data(outputStream* out) {
4157 InlineTree* inl_tree = ilt();
4158 if (inl_tree != NULL) {
4159 out->print(" inline %d", inl_tree->count());
4160 inl_tree->dump_replay_data(out);
4161 }
4162 }
4163
4164 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4165 if (n1->Opcode() < n2->Opcode()) return -1;
4166 else if (n1->Opcode() > n2->Opcode()) return 1;
4167
4168 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4169 for (uint i = 1; i < n1->req(); i++) {
4170 if (n1->in(i) < n2->in(i)) return -1;
4171 else if (n1->in(i) > n2->in(i)) return 1;
4172 }
4173
4174 return 0;
4175 }
4176
4177 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4178 Node* n1 = *n1p;
4179 Node* n2 = *n2p;
4180
4181 return cmp_expensive_nodes(n1, n2);
4182 }
4183
4184 void Compile::sort_expensive_nodes() {
4185 if (!expensive_nodes_sorted()) {
4186 _expensive_nodes->sort(cmp_expensive_nodes);
4187 }
4188 }
4189
4190 bool Compile::expensive_nodes_sorted() const {
4191 for (int i = 1; i < _expensive_nodes->length(); i++) {
4192 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4193 return false;
4194 }
4195 }
4196 return true;
4197 }
4198
4199 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4200 if (_expensive_nodes->length() == 0) {
4201 return false;
4202 }
4203
4204 assert(OptimizeExpensiveOps, "optimization off?");
4205
4206 // Take this opportunity to remove dead nodes from the list
4207 int j = 0;
4208 for (int i = 0; i < _expensive_nodes->length(); i++) {
4209 Node* n = _expensive_nodes->at(i);
4210 if (!n->is_unreachable(igvn)) {
4211 assert(n->is_expensive(), "should be expensive");
4212 _expensive_nodes->at_put(j, n);
4213 j++;
4214 }
4215 }
4216 _expensive_nodes->trunc_to(j);
4217
4218 // Then sort the list so that similar nodes are next to each other
4219 // and check for at least two nodes of identical kind with same data
4220 // inputs.
4221 sort_expensive_nodes();
4222
4223 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4224 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4225 return true;
4226 }
4227 }
4228
4229 return false;
4230 }
4231
4232 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4233 if (_expensive_nodes->length() == 0) {
4234 return;
4235 }
4236
4237 assert(OptimizeExpensiveOps, "optimization off?");
4238
4239 // Sort to bring similar nodes next to each other and clear the
4240 // control input of nodes for which there's only a single copy.
4241 sort_expensive_nodes();
4242
4243 int j = 0;
4244 int identical = 0;
4245 int i = 0;
4246 bool modified = false;
4247 for (; i < _expensive_nodes->length()-1; i++) {
4248 assert(j <= i, "can't write beyond current index");
4249 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4250 identical++;
4251 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4252 continue;
4253 }
4254 if (identical > 0) {
4255 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4256 identical = 0;
4257 } else {
4258 Node* n = _expensive_nodes->at(i);
4259 igvn.replace_input_of(n, 0, NULL);
4260 igvn.hash_insert(n);
4261 modified = true;
4262 }
4263 }
4264 if (identical > 0) {
4265 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4266 } else if (_expensive_nodes->length() >= 1) {
4267 Node* n = _expensive_nodes->at(i);
4268 igvn.replace_input_of(n, 0, NULL);
4269 igvn.hash_insert(n);
4270 modified = true;
4271 }
4272 _expensive_nodes->trunc_to(j);
4273 if (modified) {
4274 igvn.optimize();
4275 }
4276 }
4277
4278 void Compile::add_expensive_node(Node * n) {
4279 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4280 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4281 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4282 if (OptimizeExpensiveOps) {
4283 _expensive_nodes->append(n);
4284 } else {
4285 // Clear control input and let IGVN optimize expensive nodes if
4286 // OptimizeExpensiveOps is off.
4287 n->set_req(0, NULL);
4288 }
4289 }
4290
4291 /**
4292 * Remove the speculative part of types and clean up the graph
4293 */
4294 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4295 if (UseTypeSpeculation) {
4296 Unique_Node_List worklist;
4297 worklist.push(root());
4298 int modified = 0;
4299 // Go over all type nodes that carry a speculative type, drop the
4300 // speculative part of the type and enqueue the node for an igvn
4301 // which may optimize it out.
4302 for (uint next = 0; next < worklist.size(); ++next) {
4303 Node *n = worklist.at(next);
4304 if (n->is_Type()) {
4305 TypeNode* tn = n->as_Type();
4306 const Type* t = tn->type();
4307 const Type* t_no_spec = t->remove_speculative();
4308 if (t_no_spec != t) {
4309 bool in_hash = igvn.hash_delete(n);
4310 assert(in_hash, "node should be in igvn hash table");
4311 tn->set_type(t_no_spec);
4312 igvn.hash_insert(n);
4313 igvn._worklist.push(n); // give it a chance to go away
4314 modified++;
4315 }
4316 }
4317 uint max = n->len();
4318 for( uint i = 0; i < max; ++i ) {
4319 Node *m = n->in(i);
4320 if (not_a_node(m)) continue;
4321 worklist.push(m);
4322 }
4323 }
4324 // Drop the speculative part of all types in the igvn's type table
4325 igvn.remove_speculative_types();
4326 if (modified > 0) {
4327 igvn.optimize();
4328 }
4329 #ifdef ASSERT
4330 // Verify that after the IGVN is over no speculative type has resurfaced
4331 worklist.clear();
4332 worklist.push(root());
4333 for (uint next = 0; next < worklist.size(); ++next) {
4334 Node *n = worklist.at(next);
4335 const Type* t = igvn.type_or_null(n);
4336 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4337 if (n->is_Type()) {
4338 t = n->as_Type()->type();
4339 assert(t == t->remove_speculative(), "no more speculative types");
4340 }
4341 uint max = n->len();
4342 for( uint i = 0; i < max; ++i ) {
4343 Node *m = n->in(i);
4344 if (not_a_node(m)) continue;
4345 worklist.push(m);
4346 }
4347 }
4348 igvn.check_no_speculative_types();
4349 #endif
4350 }
4351 }
4352
4353 // Auxiliary method to support randomized stressing/fuzzing.
4354 //
4355 // This method can be called the arbitrary number of times, with current count
4356 // as the argument. The logic allows selecting a single candidate from the
4357 // running list of candidates as follows:
4358 // int count = 0;
4359 // Cand* selected = null;
4360 // while(cand = cand->next()) {
4361 // if (randomized_select(++count)) {
4362 // selected = cand;
4363 // }
4364 // }
4365 //
4366 // Including count equalizes the chances any candidate is "selected".
4367 // This is useful when we don't have the complete list of candidates to choose
4368 // from uniformly. In this case, we need to adjust the randomicity of the
4369 // selection, or else we will end up biasing the selection towards the latter
4370 // candidates.
4371 //
4372 // Quick back-envelope calculation shows that for the list of n candidates
4373 // the equal probability for the candidate to persist as "best" can be
4374 // achieved by replacing it with "next" k-th candidate with the probability
4375 // of 1/k. It can be easily shown that by the end of the run, the
4376 // probability for any candidate is converged to 1/n, thus giving the
4377 // uniform distribution among all the candidates.
4378 //
4379 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4380 #define RANDOMIZED_DOMAIN_POW 29
4381 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4382 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4383 bool Compile::randomized_select(int count) {
4384 assert(count > 0, "only positive");
4385 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4386 }
4387
4388 CloneMap& Compile::clone_map() { return _clone_map; }
4389 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4390
4391 void NodeCloneInfo::dump() const {
4392 tty->print(" {%d:%d} ", idx(), gen());
4393 }
4394
4395 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4396 uint64_t val = value(old->_idx);
4397 NodeCloneInfo cio(val);
4398 assert(val != 0, "old node should be in the map");
4399 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4400 insert(nnn->_idx, cin.get());
4401 #ifndef PRODUCT
4402 if (is_debug()) {
4403 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4404 }
4405 #endif
4406 }
4407
4408 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4409 NodeCloneInfo cio(value(old->_idx));
4410 if (cio.get() == 0) {
4411 cio.set(old->_idx, 0);
4412 insert(old->_idx, cio.get());
4413 #ifndef PRODUCT
4414 if (is_debug()) {
4415 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4416 }
4417 #endif
4418 }
4419 clone(old, nnn, gen);
4420 }
4421
4422 int CloneMap::max_gen() const {
4423 int g = 0;
4424 DictI di(_dict);
4425 for(; di.test(); ++di) {
4426 int t = gen(di._key);
4427 if (g < t) {
4428 g = t;
4429 #ifndef PRODUCT
4430 if (is_debug()) {
4431 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4432 }
4433 #endif
4434 }
4435 }
4436 return g;
4437 }
4438
4439 void CloneMap::dump(node_idx_t key) const {
4440 uint64_t val = value(key);
4441 if (val != 0) {
4442 NodeCloneInfo ni(val);
4443 ni.dump();
4444 }
4445 }