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 C = this;
677 #ifndef PRODUCT
678 if (_printer != NULL) {
679 _printer->set_compile(this);
680 }
681 #endif
682 CompileWrapper cw(this);
683
684 if (CITimeVerbose) {
685 tty->print(" ");
686 target->holder()->name()->print();
687 tty->print(".");
688 target->print_short_name();
689 tty->print(" ");
690 }
691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
692 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
693
694 #ifndef PRODUCT
695 bool print_opto_assembly = directive->PrintOptoAssemblyOption;
696 if (!print_opto_assembly) {
697 bool print_assembly = directive->PrintAssemblyOption;
698 if (print_assembly && !Disassembler::can_decode()) {
699 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
700 print_opto_assembly = true;
701 }
702 }
703 set_print_assembly(print_opto_assembly);
704 set_parsed_irreducible_loop(false);
705
706 if (directive->ReplayInlineOption) {
707 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
708 }
709 #endif
710 set_print_inlining(directive->PrintInliningOption NOT_PRODUCT( || PrintOptoInlining));
711 set_print_intrinsics(directive->PrintIntrinsicsOption);
712 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
713
714 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
715 // Make sure the method being compiled gets its own MDO,
716 // so we can at least track the decompile_count().
717 // Need MDO to record RTM code generation state.
718 method()->ensure_method_data();
719 }
720
721 Init(::AliasLevel);
722
723
724 print_compile_messages();
725
726 _ilt = InlineTree::build_inline_tree_root();
727
728 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
729 assert(num_alias_types() >= AliasIdxRaw, "");
730
731 #define MINIMUM_NODE_HASH 1023
732 // Node list that Iterative GVN will start with
733 Unique_Node_List for_igvn(comp_arena());
734 set_for_igvn(&for_igvn);
735
736 // GVN that will be run immediately on new nodes
737 uint estimated_size = method()->code_size()*4+64;
738 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
739 PhaseGVN gvn(node_arena(), estimated_size);
740 set_initial_gvn(&gvn);
741
742 print_inlining_init();
743 { // Scope for timing the parser
744 TracePhase tp("parse", &timers[_t_parser]);
745
746 // Put top into the hash table ASAP.
747 initial_gvn()->transform_no_reclaim(top());
748
749 // Set up tf(), start(), and find a CallGenerator.
750 CallGenerator* cg = NULL;
751 if (is_osr_compilation()) {
752 const TypeTuple *domain = StartOSRNode::osr_domain();
753 const TypeTuple *range = TypeTuple::make_range(method()->signature());
754 init_tf(TypeFunc::make(domain, range));
755 StartNode* s = new StartOSRNode(root(), domain);
756 initial_gvn()->set_type_bottom(s);
757 init_start(s);
758 cg = CallGenerator::for_osr(method(), entry_bci());
759 } else {
760 // Normal case.
761 init_tf(TypeFunc::make(method()));
762 StartNode* s = new StartNode(root(), tf()->domain());
763 initial_gvn()->set_type_bottom(s);
764 init_start(s);
765 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
766 // With java.lang.ref.reference.get() we must go through the
767 // intrinsic when G1 is enabled - even when get() is the root
768 // method of the compile - so that, if necessary, the value in
769 // the referent field of the reference object gets recorded by
770 // the pre-barrier code.
771 // Specifically, if G1 is enabled, the value in the referent
772 // field is recorded by the G1 SATB pre barrier. This will
773 // result in the referent being marked live and the reference
774 // object removed from the list of discovered references during
775 // reference processing.
776 cg = find_intrinsic(method(), false);
777 }
778 if (cg == NULL) {
779 float past_uses = method()->interpreter_invocation_count();
780 float expected_uses = past_uses;
781 cg = CallGenerator::for_inline(method(), expected_uses);
782 }
783 }
784 if (failing()) return;
785 if (cg == NULL) {
786 record_method_not_compilable_all_tiers("cannot parse method");
787 return;
788 }
789 JVMState* jvms = build_start_state(start(), tf());
790 if ((jvms = cg->generate(jvms)) == NULL) {
791 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
792 record_method_not_compilable("method parse failed");
793 }
794 return;
795 }
796 GraphKit kit(jvms);
797
798 if (!kit.stopped()) {
799 // Accept return values, and transfer control we know not where.
800 // This is done by a special, unique ReturnNode bound to root.
801 return_values(kit.jvms());
802 }
803
804 if (kit.has_exceptions()) {
805 // Any exceptions that escape from this call must be rethrown
806 // to whatever caller is dynamically above us on the stack.
807 // This is done by a special, unique RethrowNode bound to root.
808 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
809 }
810
811 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
812
813 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
814 inline_string_calls(true);
815 }
816
817 if (failing()) return;
818
819 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
820
821 // Remove clutter produced by parsing.
822 if (!failing()) {
823 ResourceMark rm;
824 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
825 }
826 }
827
828 // Note: Large methods are capped off in do_one_bytecode().
829 if (failing()) return;
830
831 // After parsing, node notes are no longer automagic.
832 // They must be propagated by register_new_node_with_optimizer(),
833 // clone(), or the like.
834 set_default_node_notes(NULL);
835
836 for (;;) {
837 int successes = Inline_Warm();
838 if (failing()) return;
839 if (successes == 0) break;
840 }
841
842 // Drain the list.
843 Finish_Warm();
844 #ifndef PRODUCT
845 if (_printer && _printer->should_print(1)) {
846 _printer->print_inlining();
847 }
848 #endif
849
850 if (failing()) return;
851 NOT_PRODUCT( verify_graph_edges(); )
852
853 // Now optimize
854 Optimize();
855 if (failing()) return;
856 NOT_PRODUCT( verify_graph_edges(); )
857
858 #ifndef PRODUCT
859 if (PrintIdeal) {
860 ttyLocker ttyl; // keep the following output all in one block
861 // This output goes directly to the tty, not the compiler log.
862 // To enable tools to match it up with the compilation activity,
863 // be sure to tag this tty output with the compile ID.
864 if (xtty != NULL) {
865 xtty->head("ideal compile_id='%d'%s", compile_id(),
866 is_osr_compilation() ? " compile_kind='osr'" :
867 "");
868 }
869 root()->dump(9999);
870 if (xtty != NULL) {
871 xtty->tail("ideal");
872 }
873 }
874 #endif
875
876 NOT_PRODUCT( verify_barriers(); )
877
878 // Dump compilation data to replay it.
879 if (directive->DumpReplayOption) {
880 env()->dump_replay_data(_compile_id);
881 }
882 if (directive->DumpInlineOption && (ilt() != NULL)) {
883 env()->dump_inline_data(_compile_id);
884 }
885
886 // Now that we know the size of all the monitors we can add a fixed slot
887 // for the original deopt pc.
888
889 _orig_pc_slot = fixed_slots();
890 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
891 set_fixed_slots(next_slot);
892
893 // Compute when to use implicit null checks. Used by matching trap based
894 // nodes and NullCheck optimization.
895 set_allowed_deopt_reasons();
896
897 // Now generate code
898 Code_Gen();
899 if (failing()) return;
900
901 // Check if we want to skip execution of all compiled code.
902 {
903 #ifndef PRODUCT
904 if (OptoNoExecute) {
905 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
906 return;
907 }
908 #endif
909 TracePhase tp("install_code", &timers[_t_registerMethod]);
910
911 if (is_osr_compilation()) {
912 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
913 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
914 } else {
915 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
916 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
917 }
918
919 env()->register_method(_method, _entry_bci,
920 &_code_offsets,
921 _orig_pc_slot_offset_in_bytes,
922 code_buffer(),
923 frame_size_in_words(), _oop_map_set,
924 &_handler_table, &_inc_table,
925 compiler,
926 has_unsafe_access(),
927 SharedRuntime::is_wide_vector(max_vector_size()),
928 _directive,
929 rtm_state()
930 );
931
932 if (log() != NULL) // Print code cache state into compiler log
933 log()->code_cache_state();
934 }
935 }
936
937 //------------------------------Compile----------------------------------------
938 // Compile a runtime stub
939 Compile::Compile( ciEnv* ci_env,
940 TypeFunc_generator generator,
941 address stub_function,
942 const char *stub_name,
943 int is_fancy_jump,
944 bool pass_tls,
945 bool save_arg_registers,
946 bool return_pc,
947 DirectiveSet* directive)
948 : Phase(Compiler),
949 _env(ci_env),
950 _directive(directive),
951 _log(ci_env->log()),
952 _compile_id(0),
953 _save_argument_registers(save_arg_registers),
954 _method(NULL),
955 _stub_name(stub_name),
956 _stub_function(stub_function),
957 _stub_entry_point(NULL),
958 _entry_bci(InvocationEntryBci),
959 _initial_gvn(NULL),
960 _for_igvn(NULL),
961 _warm_calls(NULL),
962 _orig_pc_slot(0),
963 _orig_pc_slot_offset_in_bytes(0),
964 _subsume_loads(true),
965 _do_escape_analysis(false),
966 _eliminate_boxing(false),
967 _failure_reason(NULL),
968 _code_buffer("Compile::Fill_buffer"),
969 _has_method_handle_invokes(false),
970 _mach_constant_base_node(NULL),
971 _node_bundling_limit(0),
972 _node_bundling_base(NULL),
973 _java_calls(0),
974 _inner_loops(0),
975 #ifndef PRODUCT
976 _trace_opto_output(TraceOptoOutput),
977 _in_dump_cnt(0),
978 _printer(NULL),
979 #endif
980 _comp_arena(mtCompiler),
981 _node_arena(mtCompiler),
982 _old_arena(mtCompiler),
983 _Compile_types(mtCompiler),
984 _dead_node_list(comp_arena()),
985 _dead_node_count(0),
986 _congraph(NULL),
987 _replay_inline_data(NULL),
988 _number_of_mh_late_inlines(0),
989 _inlining_progress(false),
990 _inlining_incrementally(false),
991 _print_inlining_list(NULL),
992 _print_inlining_stream(NULL),
993 _print_inlining_idx(0),
994 _print_inlining_output(NULL),
995 _allowed_reasons(0),
996 _interpreter_frame_size(0),
997 _max_node_limit(MaxNodeLimit) {
998 C = this;
999
1000 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1001 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1002
1003 #ifndef PRODUCT
1004 set_print_assembly(PrintFrameConverterAssembly);
1005 set_parsed_irreducible_loop(false);
1006 #endif
1007 set_has_irreducible_loop(false); // no loops
1008
1009 CompileWrapper cw(this);
1010 Init(/*AliasLevel=*/ 0);
1011 init_tf((*generator)());
1012
1013 {
1014 // The following is a dummy for the sake of GraphKit::gen_stub
1015 Unique_Node_List for_igvn(comp_arena());
1016 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1017 PhaseGVN gvn(Thread::current()->resource_area(),255);
1018 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1019 gvn.transform_no_reclaim(top());
1020
1021 GraphKit kit;
1022 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1023 }
1024
1025 NOT_PRODUCT( verify_graph_edges(); )
1026 Code_Gen();
1027 if (failing()) return;
1028
1029
1030 // Entry point will be accessed using compile->stub_entry_point();
1031 if (code_buffer() == NULL) {
1032 Matcher::soft_match_failure();
1033 } else {
1034 if (PrintAssembly && (WizardMode || Verbose))
1035 tty->print_cr("### Stub::%s", stub_name);
1036
1037 if (!failing()) {
1038 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1039
1040 // Make the NMethod
1041 // For now we mark the frame as never safe for profile stackwalking
1042 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1043 code_buffer(),
1044 CodeOffsets::frame_never_safe,
1045 // _code_offsets.value(CodeOffsets::Frame_Complete),
1046 frame_size_in_words(),
1047 _oop_map_set,
1048 save_arg_registers);
1049 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1050
1051 _stub_entry_point = rs->entry_point();
1052 }
1053 }
1054 }
1055
1056 //------------------------------Init-------------------------------------------
1057 // Prepare for a single compilation
1058 void Compile::Init(int aliaslevel) {
1059 _unique = 0;
1060 _regalloc = NULL;
1061
1062 _tf = NULL; // filled in later
1063 _top = NULL; // cached later
1064 _matcher = NULL; // filled in later
1065 _cfg = NULL; // filled in later
1066
1067 set_24_bit_selection_and_mode(Use24BitFP, false);
1068
1069 _node_note_array = NULL;
1070 _default_node_notes = NULL;
1071 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1072
1073 _immutable_memory = NULL; // filled in at first inquiry
1074
1075 // Globally visible Nodes
1076 // First set TOP to NULL to give safe behavior during creation of RootNode
1077 set_cached_top_node(NULL);
1078 set_root(new RootNode());
1079 // Now that you have a Root to point to, create the real TOP
1080 set_cached_top_node( new ConNode(Type::TOP) );
1081 set_recent_alloc(NULL, NULL);
1082
1083 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1084 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1085 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1086 env()->set_dependencies(new Dependencies(env()));
1087
1088 _fixed_slots = 0;
1089 set_has_split_ifs(false);
1090 set_has_loops(has_method() && method()->has_loops()); // first approximation
1091 set_has_stringbuilder(false);
1092 set_has_boxed_value(false);
1093 _trap_can_recompile = false; // no traps emitted yet
1094 _major_progress = true; // start out assuming good things will happen
1095 set_has_unsafe_access(false);
1096 set_max_vector_size(0);
1097 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1098 set_decompile_count(0);
1099
1100 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1101 set_num_loop_opts(LoopOptsCount);
1102 set_do_inlining(Inline);
1103 set_max_inline_size(MaxInlineSize);
1104 set_freq_inline_size(FreqInlineSize);
1105 set_do_scheduling(OptoScheduling);
1106 set_do_count_invocations(false);
1107 set_do_method_data_update(false);
1108
1109 set_do_vector_loop(false);
1110
1111 if (AllowVectorizeOnDemand) {
1112 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1113 set_do_vector_loop(true);
1114 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());})
1115 } else if (has_method() && method()->name() != 0 &&
1116 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1117 set_do_vector_loop(true);
1118 }
1119 }
1120 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1121 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());})
1122
1123 set_age_code(has_method() && method()->profile_aging());
1124 set_rtm_state(NoRTM); // No RTM lock eliding by default
1125 _max_node_limit = _directive->MaxNodeLimitOption;
1126
1127 #if INCLUDE_RTM_OPT
1128 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1129 int rtm_state = method()->method_data()->rtm_state();
1130 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1131 // Don't generate RTM lock eliding code.
1132 set_rtm_state(NoRTM);
1133 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1134 // Generate RTM lock eliding code without abort ratio calculation code.
1135 set_rtm_state(UseRTM);
1136 } else if (UseRTMDeopt) {
1137 // Generate RTM lock eliding code and include abort ratio calculation
1138 // code if UseRTMDeopt is on.
1139 set_rtm_state(ProfileRTM);
1140 }
1141 }
1142 #endif
1143 if (debug_info()->recording_non_safepoints()) {
1144 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1145 (comp_arena(), 8, 0, NULL));
1146 set_default_node_notes(Node_Notes::make(this));
1147 }
1148
1149 // // -- Initialize types before each compile --
1150 // // Update cached type information
1151 // if( _method && _method->constants() )
1152 // Type::update_loaded_types(_method, _method->constants());
1153
1154 // Init alias_type map.
1155 if (!_do_escape_analysis && aliaslevel == 3)
1156 aliaslevel = 2; // No unique types without escape analysis
1157 _AliasLevel = aliaslevel;
1158 const int grow_ats = 16;
1159 _max_alias_types = grow_ats;
1160 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1161 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1162 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1163 {
1164 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1165 }
1166 // Initialize the first few types.
1167 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1168 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1169 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1170 _num_alias_types = AliasIdxRaw+1;
1171 // Zero out the alias type cache.
1172 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1173 // A NULL adr_type hits in the cache right away. Preload the right answer.
1174 probe_alias_cache(NULL)->_index = AliasIdxTop;
1175
1176 _intrinsics = NULL;
1177 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1178 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1179 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1180 register_library_intrinsics();
1181 }
1182
1183 //---------------------------init_start----------------------------------------
1184 // Install the StartNode on this compile object.
1185 void Compile::init_start(StartNode* s) {
1186 if (failing())
1187 return; // already failing
1188 assert(s == start(), "");
1189 }
1190
1191 /**
1192 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1193 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1194 * the ideal graph.
1195 */
1196 StartNode* Compile::start() const {
1197 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1198 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1199 Node* start = root()->fast_out(i);
1200 if (start->is_Start()) {
1201 return start->as_Start();
1202 }
1203 }
1204 fatal("Did not find Start node!");
1205 return NULL;
1206 }
1207
1208 //-------------------------------immutable_memory-------------------------------------
1209 // Access immutable memory
1210 Node* Compile::immutable_memory() {
1211 if (_immutable_memory != NULL) {
1212 return _immutable_memory;
1213 }
1214 StartNode* s = start();
1215 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1216 Node *p = s->fast_out(i);
1217 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1218 _immutable_memory = p;
1219 return _immutable_memory;
1220 }
1221 }
1222 ShouldNotReachHere();
1223 return NULL;
1224 }
1225
1226 //----------------------set_cached_top_node------------------------------------
1227 // Install the cached top node, and make sure Node::is_top works correctly.
1228 void Compile::set_cached_top_node(Node* tn) {
1229 if (tn != NULL) verify_top(tn);
1230 Node* old_top = _top;
1231 _top = tn;
1232 // Calling Node::setup_is_top allows the nodes the chance to adjust
1233 // their _out arrays.
1234 if (_top != NULL) _top->setup_is_top();
1235 if (old_top != NULL) old_top->setup_is_top();
1236 assert(_top == NULL || top()->is_top(), "");
1237 }
1238
1239 #ifdef ASSERT
1240 uint Compile::count_live_nodes_by_graph_walk() {
1241 Unique_Node_List useful(comp_arena());
1242 // Get useful node list by walking the graph.
1243 identify_useful_nodes(useful);
1244 return useful.size();
1245 }
1246
1247 void Compile::print_missing_nodes() {
1248
1249 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1250 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1251 return;
1252 }
1253
1254 // This is an expensive function. It is executed only when the user
1255 // specifies VerifyIdealNodeCount option or otherwise knows the
1256 // additional work that needs to be done to identify reachable nodes
1257 // by walking the flow graph and find the missing ones using
1258 // _dead_node_list.
1259
1260 Unique_Node_List useful(comp_arena());
1261 // Get useful node list by walking the graph.
1262 identify_useful_nodes(useful);
1263
1264 uint l_nodes = C->live_nodes();
1265 uint l_nodes_by_walk = useful.size();
1266
1267 if (l_nodes != l_nodes_by_walk) {
1268 if (_log != NULL) {
1269 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1270 _log->stamp();
1271 _log->end_head();
1272 }
1273 VectorSet& useful_member_set = useful.member_set();
1274 int last_idx = l_nodes_by_walk;
1275 for (int i = 0; i < last_idx; i++) {
1276 if (useful_member_set.test(i)) {
1277 if (_dead_node_list.test(i)) {
1278 if (_log != NULL) {
1279 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1280 }
1281 if (PrintIdealNodeCount) {
1282 // Print the log message to tty
1283 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1284 useful.at(i)->dump();
1285 }
1286 }
1287 }
1288 else if (! _dead_node_list.test(i)) {
1289 if (_log != NULL) {
1290 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1291 }
1292 if (PrintIdealNodeCount) {
1293 // Print the log message to tty
1294 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1295 }
1296 }
1297 }
1298 if (_log != NULL) {
1299 _log->tail("mismatched_nodes");
1300 }
1301 }
1302 }
1303 void Compile::record_modified_node(Node* n) {
1304 if (_modified_nodes != NULL && !_inlining_incrementally &&
1305 n->outcnt() != 0 && !n->is_Con()) {
1306 _modified_nodes->push(n);
1307 }
1308 }
1309
1310 void Compile::remove_modified_node(Node* n) {
1311 if (_modified_nodes != NULL) {
1312 _modified_nodes->remove(n);
1313 }
1314 }
1315 #endif
1316
1317 #ifndef PRODUCT
1318 void Compile::verify_top(Node* tn) const {
1319 if (tn != NULL) {
1320 assert(tn->is_Con(), "top node must be a constant");
1321 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1322 assert(tn->in(0) != NULL, "must have live top node");
1323 }
1324 }
1325 #endif
1326
1327
1328 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1329
1330 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1331 guarantee(arr != NULL, "");
1332 int num_blocks = arr->length();
1333 if (grow_by < num_blocks) grow_by = num_blocks;
1334 int num_notes = grow_by * _node_notes_block_size;
1335 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1336 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1337 while (num_notes > 0) {
1338 arr->append(notes);
1339 notes += _node_notes_block_size;
1340 num_notes -= _node_notes_block_size;
1341 }
1342 assert(num_notes == 0, "exact multiple, please");
1343 }
1344
1345 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1346 if (source == NULL || dest == NULL) return false;
1347
1348 if (dest->is_Con())
1349 return false; // Do not push debug info onto constants.
1350
1351 #ifdef ASSERT
1352 // Leave a bread crumb trail pointing to the original node:
1353 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1354 dest->set_debug_orig(source);
1355 }
1356 #endif
1357
1358 if (node_note_array() == NULL)
1359 return false; // Not collecting any notes now.
1360
1361 // This is a copy onto a pre-existing node, which may already have notes.
1362 // If both nodes have notes, do not overwrite any pre-existing notes.
1363 Node_Notes* source_notes = node_notes_at(source->_idx);
1364 if (source_notes == NULL || source_notes->is_clear()) return false;
1365 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1366 if (dest_notes == NULL || dest_notes->is_clear()) {
1367 return set_node_notes_at(dest->_idx, source_notes);
1368 }
1369
1370 Node_Notes merged_notes = (*source_notes);
1371 // The order of operations here ensures that dest notes will win...
1372 merged_notes.update_from(dest_notes);
1373 return set_node_notes_at(dest->_idx, &merged_notes);
1374 }
1375
1376
1377 //--------------------------allow_range_check_smearing-------------------------
1378 // Gating condition for coalescing similar range checks.
1379 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1380 // single covering check that is at least as strong as any of them.
1381 // If the optimization succeeds, the simplified (strengthened) range check
1382 // will always succeed. If it fails, we will deopt, and then give up
1383 // on the optimization.
1384 bool Compile::allow_range_check_smearing() const {
1385 // If this method has already thrown a range-check,
1386 // assume it was because we already tried range smearing
1387 // and it failed.
1388 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1389 return !already_trapped;
1390 }
1391
1392
1393 //------------------------------flatten_alias_type-----------------------------
1394 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1395 int offset = tj->offset();
1396 TypePtr::PTR ptr = tj->ptr();
1397
1398 // Known instance (scalarizable allocation) alias only with itself.
1399 bool is_known_inst = tj->isa_oopptr() != NULL &&
1400 tj->is_oopptr()->is_known_instance();
1401
1402 // Process weird unsafe references.
1403 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1404 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1405 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1406 tj = TypeOopPtr::BOTTOM;
1407 ptr = tj->ptr();
1408 offset = tj->offset();
1409 }
1410
1411 // Array pointers need some flattening
1412 const TypeAryPtr *ta = tj->isa_aryptr();
1413 if (ta && ta->is_stable()) {
1414 // Erase stability property for alias analysis.
1415 tj = ta = ta->cast_to_stable(false);
1416 }
1417 if( ta && is_known_inst ) {
1418 if ( offset != Type::OffsetBot &&
1419 offset > arrayOopDesc::length_offset_in_bytes() ) {
1420 offset = Type::OffsetBot; // Flatten constant access into array body only
1421 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1422 }
1423 } else if( ta && _AliasLevel >= 2 ) {
1424 // For arrays indexed by constant indices, we flatten the alias
1425 // space to include all of the array body. Only the header, klass
1426 // and array length can be accessed un-aliased.
1427 if( offset != Type::OffsetBot ) {
1428 if( ta->const_oop() ) { // MethodData* or Method*
1429 offset = Type::OffsetBot; // Flatten constant access into array body
1430 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1431 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1432 // range is OK as-is.
1433 tj = ta = TypeAryPtr::RANGE;
1434 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1435 tj = TypeInstPtr::KLASS; // all klass loads look alike
1436 ta = TypeAryPtr::RANGE; // generic ignored junk
1437 ptr = TypePtr::BotPTR;
1438 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1439 tj = TypeInstPtr::MARK;
1440 ta = TypeAryPtr::RANGE; // generic ignored junk
1441 ptr = TypePtr::BotPTR;
1442 } else { // Random constant offset into array body
1443 offset = Type::OffsetBot; // Flatten constant access into array body
1444 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1445 }
1446 }
1447 // Arrays of fixed size alias with arrays of unknown size.
1448 if (ta->size() != TypeInt::POS) {
1449 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1450 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1451 }
1452 // Arrays of known objects become arrays of unknown objects.
1453 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1454 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1455 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1456 }
1457 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1458 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1459 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1460 }
1461 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1462 // cannot be distinguished by bytecode alone.
1463 if (ta->elem() == TypeInt::BOOL) {
1464 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1465 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1466 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1467 }
1468 // During the 2nd round of IterGVN, NotNull castings are removed.
1469 // Make sure the Bottom and NotNull variants alias the same.
1470 // Also, make sure exact and non-exact variants alias the same.
1471 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1472 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1473 }
1474 }
1475
1476 // Oop pointers need some flattening
1477 const TypeInstPtr *to = tj->isa_instptr();
1478 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1479 ciInstanceKlass *k = to->klass()->as_instance_klass();
1480 if( ptr == TypePtr::Constant ) {
1481 if (to->klass() != ciEnv::current()->Class_klass() ||
1482 offset < k->size_helper() * wordSize) {
1483 // No constant oop pointers (such as Strings); they alias with
1484 // unknown strings.
1485 assert(!is_known_inst, "not scalarizable allocation");
1486 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1487 }
1488 } else if( is_known_inst ) {
1489 tj = to; // Keep NotNull and klass_is_exact for instance type
1490 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1491 // During the 2nd round of IterGVN, NotNull castings are removed.
1492 // Make sure the Bottom and NotNull variants alias the same.
1493 // Also, make sure exact and non-exact variants alias the same.
1494 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1495 }
1496 if (to->speculative() != NULL) {
1497 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1498 }
1499 // Canonicalize the holder of this field
1500 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1501 // First handle header references such as a LoadKlassNode, even if the
1502 // object's klass is unloaded at compile time (4965979).
1503 if (!is_known_inst) { // Do it only for non-instance types
1504 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1505 }
1506 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1507 // Static fields are in the space above the normal instance
1508 // fields in the java.lang.Class instance.
1509 if (to->klass() != ciEnv::current()->Class_klass()) {
1510 to = NULL;
1511 tj = TypeOopPtr::BOTTOM;
1512 offset = tj->offset();
1513 }
1514 } else {
1515 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1516 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1517 if( is_known_inst ) {
1518 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1519 } else {
1520 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1521 }
1522 }
1523 }
1524 }
1525
1526 // Klass pointers to object array klasses need some flattening
1527 const TypeKlassPtr *tk = tj->isa_klassptr();
1528 if( tk ) {
1529 // If we are referencing a field within a Klass, we need
1530 // to assume the worst case of an Object. Both exact and
1531 // inexact types must flatten to the same alias class so
1532 // use NotNull as the PTR.
1533 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1534
1535 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1536 TypeKlassPtr::OBJECT->klass(),
1537 offset);
1538 }
1539
1540 ciKlass* klass = tk->klass();
1541 if( klass->is_obj_array_klass() ) {
1542 ciKlass* k = TypeAryPtr::OOPS->klass();
1543 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1544 k = TypeInstPtr::BOTTOM->klass();
1545 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1546 }
1547
1548 // Check for precise loads from the primary supertype array and force them
1549 // to the supertype cache alias index. Check for generic array loads from
1550 // the primary supertype array and also force them to the supertype cache
1551 // alias index. Since the same load can reach both, we need to merge
1552 // these 2 disparate memories into the same alias class. Since the
1553 // primary supertype array is read-only, there's no chance of confusion
1554 // where we bypass an array load and an array store.
1555 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1556 if (offset == Type::OffsetBot ||
1557 (offset >= primary_supers_offset &&
1558 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1559 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1560 offset = in_bytes(Klass::secondary_super_cache_offset());
1561 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1562 }
1563 }
1564
1565 // Flatten all Raw pointers together.
1566 if (tj->base() == Type::RawPtr)
1567 tj = TypeRawPtr::BOTTOM;
1568
1569 if (tj->base() == Type::AnyPtr)
1570 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1571
1572 // Flatten all to bottom for now
1573 switch( _AliasLevel ) {
1574 case 0:
1575 tj = TypePtr::BOTTOM;
1576 break;
1577 case 1: // Flatten to: oop, static, field or array
1578 switch (tj->base()) {
1579 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1580 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1581 case Type::AryPtr: // do not distinguish arrays at all
1582 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1583 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1584 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1585 default: ShouldNotReachHere();
1586 }
1587 break;
1588 case 2: // No collapsing at level 2; keep all splits
1589 case 3: // No collapsing at level 3; keep all splits
1590 break;
1591 default:
1592 Unimplemented();
1593 }
1594
1595 offset = tj->offset();
1596 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1597
1598 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1599 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1600 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1601 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1602 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1603 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1604 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1605 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1606 assert( tj->ptr() != TypePtr::TopPTR &&
1607 tj->ptr() != TypePtr::AnyNull &&
1608 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1609 // assert( tj->ptr() != TypePtr::Constant ||
1610 // tj->base() == Type::RawPtr ||
1611 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1612
1613 return tj;
1614 }
1615
1616 void Compile::AliasType::Init(int i, const TypePtr* at) {
1617 _index = i;
1618 _adr_type = at;
1619 _field = NULL;
1620 _element = NULL;
1621 _is_rewritable = true; // default
1622 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1623 if (atoop != NULL && atoop->is_known_instance()) {
1624 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1625 _general_index = Compile::current()->get_alias_index(gt);
1626 } else {
1627 _general_index = 0;
1628 }
1629 }
1630
1631 //---------------------------------print_on------------------------------------
1632 #ifndef PRODUCT
1633 void Compile::AliasType::print_on(outputStream* st) {
1634 if (index() < 10)
1635 st->print("@ <%d> ", index());
1636 else st->print("@ <%d>", index());
1637 st->print(is_rewritable() ? " " : " RO");
1638 int offset = adr_type()->offset();
1639 if (offset == Type::OffsetBot)
1640 st->print(" +any");
1641 else st->print(" +%-3d", offset);
1642 st->print(" in ");
1643 adr_type()->dump_on(st);
1644 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1645 if (field() != NULL && tjp) {
1646 if (tjp->klass() != field()->holder() ||
1647 tjp->offset() != field()->offset_in_bytes()) {
1648 st->print(" != ");
1649 field()->print();
1650 st->print(" ***");
1651 }
1652 }
1653 }
1654
1655 void print_alias_types() {
1656 Compile* C = Compile::current();
1657 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1658 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1659 C->alias_type(idx)->print_on(tty);
1660 tty->cr();
1661 }
1662 }
1663 #endif
1664
1665
1666 //----------------------------probe_alias_cache--------------------------------
1667 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1668 intptr_t key = (intptr_t) adr_type;
1669 key ^= key >> logAliasCacheSize;
1670 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1671 }
1672
1673
1674 //-----------------------------grow_alias_types--------------------------------
1675 void Compile::grow_alias_types() {
1676 const int old_ats = _max_alias_types; // how many before?
1677 const int new_ats = old_ats; // how many more?
1678 const int grow_ats = old_ats+new_ats; // how many now?
1679 _max_alias_types = grow_ats;
1680 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1681 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1682 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1683 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1684 }
1685
1686
1687 //--------------------------------find_alias_type------------------------------
1688 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1689 if (_AliasLevel == 0)
1690 return alias_type(AliasIdxBot);
1691
1692 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1693 if (ace->_adr_type == adr_type) {
1694 return alias_type(ace->_index);
1695 }
1696
1697 // Handle special cases.
1698 if (adr_type == NULL) return alias_type(AliasIdxTop);
1699 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1700
1701 // Do it the slow way.
1702 const TypePtr* flat = flatten_alias_type(adr_type);
1703
1704 #ifdef ASSERT
1705 assert(flat == flatten_alias_type(flat), "idempotent");
1706 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1707 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1708 const TypeOopPtr* foop = flat->is_oopptr();
1709 // Scalarizable allocations have exact klass always.
1710 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1711 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1712 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1713 }
1714 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1715 #endif
1716
1717 int idx = AliasIdxTop;
1718 for (int i = 0; i < num_alias_types(); i++) {
1719 if (alias_type(i)->adr_type() == flat) {
1720 idx = i;
1721 break;
1722 }
1723 }
1724
1725 if (idx == AliasIdxTop) {
1726 if (no_create) return NULL;
1727 // Grow the array if necessary.
1728 if (_num_alias_types == _max_alias_types) grow_alias_types();
1729 // Add a new alias type.
1730 idx = _num_alias_types++;
1731 _alias_types[idx]->Init(idx, flat);
1732 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1733 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1734 if (flat->isa_instptr()) {
1735 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1736 && flat->is_instptr()->klass() == env()->Class_klass())
1737 alias_type(idx)->set_rewritable(false);
1738 }
1739 if (flat->isa_aryptr()) {
1740 #ifdef ASSERT
1741 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1742 // (T_BYTE has the weakest alignment and size restrictions...)
1743 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1744 #endif
1745 if (flat->offset() == TypePtr::OffsetBot) {
1746 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1747 }
1748 }
1749 if (flat->isa_klassptr()) {
1750 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1751 alias_type(idx)->set_rewritable(false);
1752 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1753 alias_type(idx)->set_rewritable(false);
1754 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1755 alias_type(idx)->set_rewritable(false);
1756 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1757 alias_type(idx)->set_rewritable(false);
1758 }
1759 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1760 // but the base pointer type is not distinctive enough to identify
1761 // references into JavaThread.)
1762
1763 // Check for final fields.
1764 const TypeInstPtr* tinst = flat->isa_instptr();
1765 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1766 ciField* field;
1767 if (tinst->const_oop() != NULL &&
1768 tinst->klass() == ciEnv::current()->Class_klass() &&
1769 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1770 // static field
1771 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1772 field = k->get_field_by_offset(tinst->offset(), true);
1773 } else {
1774 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1775 field = k->get_field_by_offset(tinst->offset(), false);
1776 }
1777 assert(field == NULL ||
1778 original_field == NULL ||
1779 (field->holder() == original_field->holder() &&
1780 field->offset() == original_field->offset() &&
1781 field->is_static() == original_field->is_static()), "wrong field?");
1782 // Set field() and is_rewritable() attributes.
1783 if (field != NULL) alias_type(idx)->set_field(field);
1784 }
1785 }
1786
1787 // Fill the cache for next time.
1788 ace->_adr_type = adr_type;
1789 ace->_index = idx;
1790 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1791
1792 // Might as well try to fill the cache for the flattened version, too.
1793 AliasCacheEntry* face = probe_alias_cache(flat);
1794 if (face->_adr_type == NULL) {
1795 face->_adr_type = flat;
1796 face->_index = idx;
1797 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1798 }
1799
1800 return alias_type(idx);
1801 }
1802
1803
1804 Compile::AliasType* Compile::alias_type(ciField* field) {
1805 const TypeOopPtr* t;
1806 if (field->is_static())
1807 t = TypeInstPtr::make(field->holder()->java_mirror());
1808 else
1809 t = TypeOopPtr::make_from_klass_raw(field->holder());
1810 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1811 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1812 return atp;
1813 }
1814
1815
1816 //------------------------------have_alias_type--------------------------------
1817 bool Compile::have_alias_type(const TypePtr* adr_type) {
1818 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1819 if (ace->_adr_type == adr_type) {
1820 return true;
1821 }
1822
1823 // Handle special cases.
1824 if (adr_type == NULL) return true;
1825 if (adr_type == TypePtr::BOTTOM) return true;
1826
1827 return find_alias_type(adr_type, true, NULL) != NULL;
1828 }
1829
1830 //-----------------------------must_alias--------------------------------------
1831 // True if all values of the given address type are in the given alias category.
1832 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1833 if (alias_idx == AliasIdxBot) return true; // the universal category
1834 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1835 if (alias_idx == AliasIdxTop) return false; // the empty category
1836 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1837
1838 // the only remaining possible overlap is identity
1839 int adr_idx = get_alias_index(adr_type);
1840 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1841 assert(adr_idx == alias_idx ||
1842 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1843 && adr_type != TypeOopPtr::BOTTOM),
1844 "should not be testing for overlap with an unsafe pointer");
1845 return adr_idx == alias_idx;
1846 }
1847
1848 //------------------------------can_alias--------------------------------------
1849 // True if any values of the given address type are in the given alias category.
1850 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1851 if (alias_idx == AliasIdxTop) return false; // the empty category
1852 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1853 if (alias_idx == AliasIdxBot) return true; // the universal category
1854 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1855
1856 // the only remaining possible overlap is identity
1857 int adr_idx = get_alias_index(adr_type);
1858 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1859 return adr_idx == alias_idx;
1860 }
1861
1862
1863
1864 //---------------------------pop_warm_call-------------------------------------
1865 WarmCallInfo* Compile::pop_warm_call() {
1866 WarmCallInfo* wci = _warm_calls;
1867 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1868 return wci;
1869 }
1870
1871 //----------------------------Inline_Warm--------------------------------------
1872 int Compile::Inline_Warm() {
1873 // If there is room, try to inline some more warm call sites.
1874 // %%% Do a graph index compaction pass when we think we're out of space?
1875 if (!InlineWarmCalls) return 0;
1876
1877 int calls_made_hot = 0;
1878 int room_to_grow = NodeCountInliningCutoff - unique();
1879 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1880 int amount_grown = 0;
1881 WarmCallInfo* call;
1882 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1883 int est_size = (int)call->size();
1884 if (est_size > (room_to_grow - amount_grown)) {
1885 // This one won't fit anyway. Get rid of it.
1886 call->make_cold();
1887 continue;
1888 }
1889 call->make_hot();
1890 calls_made_hot++;
1891 amount_grown += est_size;
1892 amount_to_grow -= est_size;
1893 }
1894
1895 if (calls_made_hot > 0) set_major_progress();
1896 return calls_made_hot;
1897 }
1898
1899
1900 //----------------------------Finish_Warm--------------------------------------
1901 void Compile::Finish_Warm() {
1902 if (!InlineWarmCalls) return;
1903 if (failing()) return;
1904 if (warm_calls() == NULL) return;
1905
1906 // Clean up loose ends, if we are out of space for inlining.
1907 WarmCallInfo* call;
1908 while ((call = pop_warm_call()) != NULL) {
1909 call->make_cold();
1910 }
1911 }
1912
1913 //---------------------cleanup_loop_predicates-----------------------
1914 // Remove the opaque nodes that protect the predicates so that all unused
1915 // checks and uncommon_traps will be eliminated from the ideal graph
1916 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1917 if (predicate_count()==0) return;
1918 for (int i = predicate_count(); i > 0; i--) {
1919 Node * n = predicate_opaque1_node(i-1);
1920 assert(n->Opcode() == Op_Opaque1, "must be");
1921 igvn.replace_node(n, n->in(1));
1922 }
1923 assert(predicate_count()==0, "should be clean!");
1924 }
1925
1926 // StringOpts and late inlining of string methods
1927 void Compile::inline_string_calls(bool parse_time) {
1928 {
1929 // remove useless nodes to make the usage analysis simpler
1930 ResourceMark rm;
1931 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1932 }
1933
1934 {
1935 ResourceMark rm;
1936 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1937 PhaseStringOpts pso(initial_gvn(), for_igvn());
1938 print_method(PHASE_AFTER_STRINGOPTS, 3);
1939 }
1940
1941 // now inline anything that we skipped the first time around
1942 if (!parse_time) {
1943 _late_inlines_pos = _late_inlines.length();
1944 }
1945
1946 while (_string_late_inlines.length() > 0) {
1947 CallGenerator* cg = _string_late_inlines.pop();
1948 cg->do_late_inline();
1949 if (failing()) return;
1950 }
1951 _string_late_inlines.trunc_to(0);
1952 }
1953
1954 // Late inlining of boxing methods
1955 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1956 if (_boxing_late_inlines.length() > 0) {
1957 assert(has_boxed_value(), "inconsistent");
1958
1959 PhaseGVN* gvn = initial_gvn();
1960 set_inlining_incrementally(true);
1961
1962 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1963 for_igvn()->clear();
1964 gvn->replace_with(&igvn);
1965
1966 _late_inlines_pos = _late_inlines.length();
1967
1968 while (_boxing_late_inlines.length() > 0) {
1969 CallGenerator* cg = _boxing_late_inlines.pop();
1970 cg->do_late_inline();
1971 if (failing()) return;
1972 }
1973 _boxing_late_inlines.trunc_to(0);
1974
1975 {
1976 ResourceMark rm;
1977 PhaseRemoveUseless pru(gvn, for_igvn());
1978 }
1979
1980 igvn = PhaseIterGVN(gvn);
1981 igvn.optimize();
1982
1983 set_inlining_progress(false);
1984 set_inlining_incrementally(false);
1985 }
1986 }
1987
1988 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1989 assert(IncrementalInline, "incremental inlining should be on");
1990 PhaseGVN* gvn = initial_gvn();
1991
1992 set_inlining_progress(false);
1993 for_igvn()->clear();
1994 gvn->replace_with(&igvn);
1995
1996 {
1997 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1998 int i = 0;
1999 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2000 CallGenerator* cg = _late_inlines.at(i);
2001 _late_inlines_pos = i+1;
2002 cg->do_late_inline();
2003 if (failing()) return;
2004 }
2005 int j = 0;
2006 for (; i < _late_inlines.length(); i++, j++) {
2007 _late_inlines.at_put(j, _late_inlines.at(i));
2008 }
2009 _late_inlines.trunc_to(j);
2010 }
2011
2012 {
2013 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2014 ResourceMark rm;
2015 PhaseRemoveUseless pru(gvn, for_igvn());
2016 }
2017
2018 {
2019 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2020 igvn = PhaseIterGVN(gvn);
2021 }
2022 }
2023
2024 // Perform incremental inlining until bound on number of live nodes is reached
2025 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2026 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2027
2028 PhaseGVN* gvn = initial_gvn();
2029
2030 set_inlining_incrementally(true);
2031 set_inlining_progress(true);
2032 uint low_live_nodes = 0;
2033
2034 while(inlining_progress() && _late_inlines.length() > 0) {
2035
2036 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2037 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2038 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2039 // PhaseIdealLoop is expensive so we only try it once we are
2040 // out of live nodes and we only try it again if the previous
2041 // helped got the number of nodes down significantly
2042 PhaseIdealLoop ideal_loop( igvn, false, true );
2043 if (failing()) return;
2044 low_live_nodes = live_nodes();
2045 _major_progress = true;
2046 }
2047
2048 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2049 break;
2050 }
2051 }
2052
2053 inline_incrementally_one(igvn);
2054
2055 if (failing()) return;
2056
2057 {
2058 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2059 igvn.optimize();
2060 }
2061
2062 if (failing()) return;
2063 }
2064
2065 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2066
2067 if (_string_late_inlines.length() > 0) {
2068 assert(has_stringbuilder(), "inconsistent");
2069 for_igvn()->clear();
2070 initial_gvn()->replace_with(&igvn);
2071
2072 inline_string_calls(false);
2073
2074 if (failing()) return;
2075
2076 {
2077 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2078 ResourceMark rm;
2079 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2080 }
2081
2082 {
2083 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2084 igvn = PhaseIterGVN(gvn);
2085 igvn.optimize();
2086 }
2087 }
2088
2089 set_inlining_incrementally(false);
2090 }
2091
2092
2093 //------------------------------Optimize---------------------------------------
2094 // Given a graph, optimize it.
2095 void Compile::Optimize() {
2096 TracePhase tp("optimizer", &timers[_t_optimizer]);
2097
2098 #ifndef PRODUCT
2099 if (_directive->BreakAtCompileOption) {
2100 BREAKPOINT;
2101 }
2102
2103 #endif
2104
2105 ResourceMark rm;
2106 int loop_opts_cnt;
2107
2108 print_inlining_reinit();
2109
2110 NOT_PRODUCT( verify_graph_edges(); )
2111
2112 print_method(PHASE_AFTER_PARSING);
2113
2114 {
2115 // Iterative Global Value Numbering, including ideal transforms
2116 // Initialize IterGVN with types and values from parse-time GVN
2117 PhaseIterGVN igvn(initial_gvn());
2118 #ifdef ASSERT
2119 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2120 #endif
2121 {
2122 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2123 igvn.optimize();
2124 }
2125
2126 print_method(PHASE_ITER_GVN1, 2);
2127
2128 if (failing()) return;
2129
2130 inline_incrementally(igvn);
2131
2132 print_method(PHASE_INCREMENTAL_INLINE, 2);
2133
2134 if (failing()) return;
2135
2136 if (eliminate_boxing()) {
2137 // Inline valueOf() methods now.
2138 inline_boxing_calls(igvn);
2139
2140 if (AlwaysIncrementalInline) {
2141 inline_incrementally(igvn);
2142 }
2143
2144 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2145
2146 if (failing()) return;
2147 }
2148
2149 // Remove the speculative part of types and clean up the graph from
2150 // the extra CastPP nodes whose only purpose is to carry them. Do
2151 // that early so that optimizations are not disrupted by the extra
2152 // CastPP nodes.
2153 remove_speculative_types(igvn);
2154
2155 // No more new expensive nodes will be added to the list from here
2156 // so keep only the actual candidates for optimizations.
2157 cleanup_expensive_nodes(igvn);
2158
2159 // Perform escape analysis
2160 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2161 if (has_loops()) {
2162 // Cleanup graph (remove dead nodes).
2163 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2164 PhaseIdealLoop ideal_loop( igvn, false, true );
2165 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2166 if (failing()) return;
2167 }
2168 ConnectionGraph::do_analysis(this, &igvn);
2169
2170 if (failing()) return;
2171
2172 // Optimize out fields loads from scalar replaceable allocations.
2173 igvn.optimize();
2174 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2175
2176 if (failing()) return;
2177
2178 if (congraph() != NULL && macro_count() > 0) {
2179 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2180 PhaseMacroExpand mexp(igvn);
2181 mexp.eliminate_macro_nodes();
2182 igvn.set_delay_transform(false);
2183
2184 igvn.optimize();
2185 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2186
2187 if (failing()) return;
2188 }
2189 }
2190
2191 // Loop transforms on the ideal graph. Range Check Elimination,
2192 // peeling, unrolling, etc.
2193
2194 // Set loop opts counter
2195 loop_opts_cnt = num_loop_opts();
2196 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2197 {
2198 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2199 PhaseIdealLoop ideal_loop( igvn, true );
2200 loop_opts_cnt--;
2201 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2202 if (failing()) return;
2203 }
2204 // Loop opts pass if partial peeling occurred in previous pass
2205 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2206 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2207 PhaseIdealLoop ideal_loop( igvn, false );
2208 loop_opts_cnt--;
2209 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2210 if (failing()) return;
2211 }
2212 // Loop opts pass for loop-unrolling before CCP
2213 if(major_progress() && (loop_opts_cnt > 0)) {
2214 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2215 PhaseIdealLoop ideal_loop( igvn, false );
2216 loop_opts_cnt--;
2217 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2218 }
2219 if (!failing()) {
2220 // Verify that last round of loop opts produced a valid graph
2221 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2222 PhaseIdealLoop::verify(igvn);
2223 }
2224 }
2225 if (failing()) return;
2226
2227 // Conditional Constant Propagation;
2228 PhaseCCP ccp( &igvn );
2229 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2230 {
2231 TracePhase tp("ccp", &timers[_t_ccp]);
2232 ccp.do_transform();
2233 }
2234 print_method(PHASE_CPP1, 2);
2235
2236 assert( true, "Break here to ccp.dump_old2new_map()");
2237
2238 // Iterative Global Value Numbering, including ideal transforms
2239 {
2240 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2241 igvn = ccp;
2242 igvn.optimize();
2243 }
2244
2245 print_method(PHASE_ITER_GVN2, 2);
2246
2247 if (failing()) return;
2248
2249 // Loop transforms on the ideal graph. Range Check Elimination,
2250 // peeling, unrolling, etc.
2251 if(loop_opts_cnt > 0) {
2252 debug_only( int cnt = 0; );
2253 while(major_progress() && (loop_opts_cnt > 0)) {
2254 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2255 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2256 PhaseIdealLoop ideal_loop( igvn, true);
2257 loop_opts_cnt--;
2258 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2259 if (failing()) return;
2260 }
2261 }
2262 // Ensure that major progress is now clear
2263 C->clear_major_progress();
2264
2265 {
2266 // Verify that all previous optimizations produced a valid graph
2267 // at least to this point, even if no loop optimizations were done.
2268 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2269 PhaseIdealLoop::verify(igvn);
2270 }
2271
2272 {
2273 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2274 PhaseMacroExpand mex(igvn);
2275 if (mex.expand_macro_nodes()) {
2276 assert(failing(), "must bail out w/ explicit message");
2277 return;
2278 }
2279 }
2280
2281 DEBUG_ONLY( _modified_nodes = NULL; )
2282 } // (End scope of igvn; run destructor if necessary for asserts.)
2283
2284 process_print_inlining();
2285 // A method with only infinite loops has no edges entering loops from root
2286 {
2287 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2288 if (final_graph_reshaping()) {
2289 assert(failing(), "must bail out w/ explicit message");
2290 return;
2291 }
2292 }
2293
2294 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2295 }
2296
2297
2298 //------------------------------Code_Gen---------------------------------------
2299 // Given a graph, generate code for it
2300 void Compile::Code_Gen() {
2301 if (failing()) {
2302 return;
2303 }
2304
2305 // Perform instruction selection. You might think we could reclaim Matcher
2306 // memory PDQ, but actually the Matcher is used in generating spill code.
2307 // Internals of the Matcher (including some VectorSets) must remain live
2308 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2309 // set a bit in reclaimed memory.
2310
2311 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2312 // nodes. Mapping is only valid at the root of each matched subtree.
2313 NOT_PRODUCT( verify_graph_edges(); )
2314
2315 Matcher matcher;
2316 _matcher = &matcher;
2317 {
2318 TracePhase tp("matcher", &timers[_t_matcher]);
2319 matcher.match();
2320 }
2321 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2322 // nodes. Mapping is only valid at the root of each matched subtree.
2323 NOT_PRODUCT( verify_graph_edges(); )
2324
2325 // If you have too many nodes, or if matching has failed, bail out
2326 check_node_count(0, "out of nodes matching instructions");
2327 if (failing()) {
2328 return;
2329 }
2330
2331 // Build a proper-looking CFG
2332 PhaseCFG cfg(node_arena(), root(), matcher);
2333 _cfg = &cfg;
2334 {
2335 TracePhase tp("scheduler", &timers[_t_scheduler]);
2336 bool success = cfg.do_global_code_motion();
2337 if (!success) {
2338 return;
2339 }
2340
2341 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2342 NOT_PRODUCT( verify_graph_edges(); )
2343 debug_only( cfg.verify(); )
2344 }
2345
2346 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2347 _regalloc = ®alloc;
2348 {
2349 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2350 // Perform register allocation. After Chaitin, use-def chains are
2351 // no longer accurate (at spill code) and so must be ignored.
2352 // Node->LRG->reg mappings are still accurate.
2353 _regalloc->Register_Allocate();
2354
2355 // Bail out if the allocator builds too many nodes
2356 if (failing()) {
2357 return;
2358 }
2359 }
2360
2361 // Prior to register allocation we kept empty basic blocks in case the
2362 // the allocator needed a place to spill. After register allocation we
2363 // are not adding any new instructions. If any basic block is empty, we
2364 // can now safely remove it.
2365 {
2366 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2367 cfg.remove_empty_blocks();
2368 if (do_freq_based_layout()) {
2369 PhaseBlockLayout layout(cfg);
2370 } else {
2371 cfg.set_loop_alignment();
2372 }
2373 cfg.fixup_flow();
2374 }
2375
2376 // Apply peephole optimizations
2377 if( OptoPeephole ) {
2378 TracePhase tp("peephole", &timers[_t_peephole]);
2379 PhasePeephole peep( _regalloc, cfg);
2380 peep.do_transform();
2381 }
2382
2383 // Do late expand if CPU requires this.
2384 if (Matcher::require_postalloc_expand) {
2385 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2386 cfg.postalloc_expand(_regalloc);
2387 }
2388
2389 // Convert Nodes to instruction bits in a buffer
2390 {
2391 TraceTime tp("output", &timers[_t_output], CITime);
2392 Output();
2393 }
2394
2395 print_method(PHASE_FINAL_CODE);
2396
2397 // He's dead, Jim.
2398 _cfg = (PhaseCFG*)0xdeadbeef;
2399 _regalloc = (PhaseChaitin*)0xdeadbeef;
2400 }
2401
2402
2403 //------------------------------dump_asm---------------------------------------
2404 // Dump formatted assembly
2405 #ifndef PRODUCT
2406 void Compile::dump_asm(int *pcs, uint pc_limit) {
2407 bool cut_short = false;
2408 tty->print_cr("#");
2409 tty->print("# "); _tf->dump(); tty->cr();
2410 tty->print_cr("#");
2411
2412 // For all blocks
2413 int pc = 0x0; // Program counter
2414 char starts_bundle = ' ';
2415 _regalloc->dump_frame();
2416
2417 Node *n = NULL;
2418 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2419 if (VMThread::should_terminate()) {
2420 cut_short = true;
2421 break;
2422 }
2423 Block* block = _cfg->get_block(i);
2424 if (block->is_connector() && !Verbose) {
2425 continue;
2426 }
2427 n = block->head();
2428 if (pcs && n->_idx < pc_limit) {
2429 tty->print("%3.3x ", pcs[n->_idx]);
2430 } else {
2431 tty->print(" ");
2432 }
2433 block->dump_head(_cfg);
2434 if (block->is_connector()) {
2435 tty->print_cr(" # Empty connector block");
2436 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2437 tty->print_cr(" # Block is sole successor of call");
2438 }
2439
2440 // For all instructions
2441 Node *delay = NULL;
2442 for (uint j = 0; j < block->number_of_nodes(); j++) {
2443 if (VMThread::should_terminate()) {
2444 cut_short = true;
2445 break;
2446 }
2447 n = block->get_node(j);
2448 if (valid_bundle_info(n)) {
2449 Bundle* bundle = node_bundling(n);
2450 if (bundle->used_in_unconditional_delay()) {
2451 delay = n;
2452 continue;
2453 }
2454 if (bundle->starts_bundle()) {
2455 starts_bundle = '+';
2456 }
2457 }
2458
2459 if (WizardMode) {
2460 n->dump();
2461 }
2462
2463 if( !n->is_Region() && // Dont print in the Assembly
2464 !n->is_Phi() && // a few noisely useless nodes
2465 !n->is_Proj() &&
2466 !n->is_MachTemp() &&
2467 !n->is_SafePointScalarObject() &&
2468 !n->is_Catch() && // Would be nice to print exception table targets
2469 !n->is_MergeMem() && // Not very interesting
2470 !n->is_top() && // Debug info table constants
2471 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2472 ) {
2473 if (pcs && n->_idx < pc_limit)
2474 tty->print("%3.3x", pcs[n->_idx]);
2475 else
2476 tty->print(" ");
2477 tty->print(" %c ", starts_bundle);
2478 starts_bundle = ' ';
2479 tty->print("\t");
2480 n->format(_regalloc, tty);
2481 tty->cr();
2482 }
2483
2484 // If we have an instruction with a delay slot, and have seen a delay,
2485 // then back up and print it
2486 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2487 assert(delay != NULL, "no unconditional delay instruction");
2488 if (WizardMode) delay->dump();
2489
2490 if (node_bundling(delay)->starts_bundle())
2491 starts_bundle = '+';
2492 if (pcs && n->_idx < pc_limit)
2493 tty->print("%3.3x", pcs[n->_idx]);
2494 else
2495 tty->print(" ");
2496 tty->print(" %c ", starts_bundle);
2497 starts_bundle = ' ';
2498 tty->print("\t");
2499 delay->format(_regalloc, tty);
2500 tty->cr();
2501 delay = NULL;
2502 }
2503
2504 // Dump the exception table as well
2505 if( n->is_Catch() && (Verbose || WizardMode) ) {
2506 // Print the exception table for this offset
2507 _handler_table.print_subtable_for(pc);
2508 }
2509 }
2510
2511 if (pcs && n->_idx < pc_limit)
2512 tty->print_cr("%3.3x", pcs[n->_idx]);
2513 else
2514 tty->cr();
2515
2516 assert(cut_short || delay == NULL, "no unconditional delay branch");
2517
2518 } // End of per-block dump
2519 tty->cr();
2520
2521 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2522 }
2523 #endif
2524
2525 //------------------------------Final_Reshape_Counts---------------------------
2526 // This class defines counters to help identify when a method
2527 // may/must be executed using hardware with only 24-bit precision.
2528 struct Final_Reshape_Counts : public StackObj {
2529 int _call_count; // count non-inlined 'common' calls
2530 int _float_count; // count float ops requiring 24-bit precision
2531 int _double_count; // count double ops requiring more precision
2532 int _java_call_count; // count non-inlined 'java' calls
2533 int _inner_loop_count; // count loops which need alignment
2534 VectorSet _visited; // Visitation flags
2535 Node_List _tests; // Set of IfNodes & PCTableNodes
2536
2537 Final_Reshape_Counts() :
2538 _call_count(0), _float_count(0), _double_count(0),
2539 _java_call_count(0), _inner_loop_count(0),
2540 _visited( Thread::current()->resource_area() ) { }
2541
2542 void inc_call_count () { _call_count ++; }
2543 void inc_float_count () { _float_count ++; }
2544 void inc_double_count() { _double_count++; }
2545 void inc_java_call_count() { _java_call_count++; }
2546 void inc_inner_loop_count() { _inner_loop_count++; }
2547
2548 int get_call_count () const { return _call_count ; }
2549 int get_float_count () const { return _float_count ; }
2550 int get_double_count() const { return _double_count; }
2551 int get_java_call_count() const { return _java_call_count; }
2552 int get_inner_loop_count() const { return _inner_loop_count; }
2553 };
2554
2555 #ifdef ASSERT
2556 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2557 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2558 // Make sure the offset goes inside the instance layout.
2559 return k->contains_field_offset(tp->offset());
2560 // Note that OffsetBot and OffsetTop are very negative.
2561 }
2562 #endif
2563
2564 // Eliminate trivially redundant StoreCMs and accumulate their
2565 // precedence edges.
2566 void Compile::eliminate_redundant_card_marks(Node* n) {
2567 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2568 if (n->in(MemNode::Address)->outcnt() > 1) {
2569 // There are multiple users of the same address so it might be
2570 // possible to eliminate some of the StoreCMs
2571 Node* mem = n->in(MemNode::Memory);
2572 Node* adr = n->in(MemNode::Address);
2573 Node* val = n->in(MemNode::ValueIn);
2574 Node* prev = n;
2575 bool done = false;
2576 // Walk the chain of StoreCMs eliminating ones that match. As
2577 // long as it's a chain of single users then the optimization is
2578 // safe. Eliminating partially redundant StoreCMs would require
2579 // cloning copies down the other paths.
2580 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2581 if (adr == mem->in(MemNode::Address) &&
2582 val == mem->in(MemNode::ValueIn)) {
2583 // redundant StoreCM
2584 if (mem->req() > MemNode::OopStore) {
2585 // Hasn't been processed by this code yet.
2586 n->add_prec(mem->in(MemNode::OopStore));
2587 } else {
2588 // Already converted to precedence edge
2589 for (uint i = mem->req(); i < mem->len(); i++) {
2590 // Accumulate any precedence edges
2591 if (mem->in(i) != NULL) {
2592 n->add_prec(mem->in(i));
2593 }
2594 }
2595 // Everything above this point has been processed.
2596 done = true;
2597 }
2598 // Eliminate the previous StoreCM
2599 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2600 assert(mem->outcnt() == 0, "should be dead");
2601 mem->disconnect_inputs(NULL, this);
2602 } else {
2603 prev = mem;
2604 }
2605 mem = prev->in(MemNode::Memory);
2606 }
2607 }
2608 }
2609
2610 //------------------------------final_graph_reshaping_impl----------------------
2611 // Implement items 1-5 from final_graph_reshaping below.
2612 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2613
2614 if ( n->outcnt() == 0 ) return; // dead node
2615 uint nop = n->Opcode();
2616
2617 // Check for 2-input instruction with "last use" on right input.
2618 // Swap to left input. Implements item (2).
2619 if( n->req() == 3 && // two-input instruction
2620 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2621 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2622 n->in(2)->outcnt() == 1 &&// right use IS a last use
2623 !n->in(2)->is_Con() ) { // right use is not a constant
2624 // Check for commutative opcode
2625 switch( nop ) {
2626 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2627 case Op_MaxI: case Op_MinI:
2628 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2629 case Op_AndL: case Op_XorL: case Op_OrL:
2630 case Op_AndI: case Op_XorI: case Op_OrI: {
2631 // Move "last use" input to left by swapping inputs
2632 n->swap_edges(1, 2);
2633 break;
2634 }
2635 default:
2636 break;
2637 }
2638 }
2639
2640 #ifdef ASSERT
2641 if( n->is_Mem() ) {
2642 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2643 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2644 // oop will be recorded in oop map if load crosses safepoint
2645 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2646 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2647 "raw memory operations should have control edge");
2648 }
2649 #endif
2650 // Count FPU ops and common calls, implements item (3)
2651 switch( nop ) {
2652 // Count all float operations that may use FPU
2653 case Op_AddF:
2654 case Op_SubF:
2655 case Op_MulF:
2656 case Op_DivF:
2657 case Op_NegF:
2658 case Op_ModF:
2659 case Op_ConvI2F:
2660 case Op_ConF:
2661 case Op_CmpF:
2662 case Op_CmpF3:
2663 // case Op_ConvL2F: // longs are split into 32-bit halves
2664 frc.inc_float_count();
2665 break;
2666
2667 case Op_ConvF2D:
2668 case Op_ConvD2F:
2669 frc.inc_float_count();
2670 frc.inc_double_count();
2671 break;
2672
2673 // Count all double operations that may use FPU
2674 case Op_AddD:
2675 case Op_SubD:
2676 case Op_MulD:
2677 case Op_DivD:
2678 case Op_NegD:
2679 case Op_ModD:
2680 case Op_ConvI2D:
2681 case Op_ConvD2I:
2682 // case Op_ConvL2D: // handled by leaf call
2683 // case Op_ConvD2L: // handled by leaf call
2684 case Op_ConD:
2685 case Op_CmpD:
2686 case Op_CmpD3:
2687 frc.inc_double_count();
2688 break;
2689 case Op_Opaque1: // Remove Opaque Nodes before matching
2690 case Op_Opaque2: // Remove Opaque Nodes before matching
2691 case Op_Opaque3:
2692 n->subsume_by(n->in(1), this);
2693 break;
2694 case Op_CallStaticJava:
2695 case Op_CallJava:
2696 case Op_CallDynamicJava:
2697 frc.inc_java_call_count(); // Count java call site;
2698 case Op_CallRuntime:
2699 case Op_CallLeaf:
2700 case Op_CallLeafNoFP: {
2701 assert( n->is_Call(), "" );
2702 CallNode *call = n->as_Call();
2703 // Count call sites where the FP mode bit would have to be flipped.
2704 // Do not count uncommon runtime calls:
2705 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2706 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2707 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2708 frc.inc_call_count(); // Count the call site
2709 } else { // See if uncommon argument is shared
2710 Node *n = call->in(TypeFunc::Parms);
2711 int nop = n->Opcode();
2712 // Clone shared simple arguments to uncommon calls, item (1).
2713 if( n->outcnt() > 1 &&
2714 !n->is_Proj() &&
2715 nop != Op_CreateEx &&
2716 nop != Op_CheckCastPP &&
2717 nop != Op_DecodeN &&
2718 nop != Op_DecodeNKlass &&
2719 !n->is_Mem() ) {
2720 Node *x = n->clone();
2721 call->set_req( TypeFunc::Parms, x );
2722 }
2723 }
2724 break;
2725 }
2726
2727 case Op_StoreD:
2728 case Op_LoadD:
2729 case Op_LoadD_unaligned:
2730 frc.inc_double_count();
2731 goto handle_mem;
2732 case Op_StoreF:
2733 case Op_LoadF:
2734 frc.inc_float_count();
2735 goto handle_mem;
2736
2737 case Op_StoreCM:
2738 {
2739 // Convert OopStore dependence into precedence edge
2740 Node* prec = n->in(MemNode::OopStore);
2741 n->del_req(MemNode::OopStore);
2742 n->add_prec(prec);
2743 eliminate_redundant_card_marks(n);
2744 }
2745
2746 // fall through
2747
2748 case Op_StoreB:
2749 case Op_StoreC:
2750 case Op_StorePConditional:
2751 case Op_StoreI:
2752 case Op_StoreL:
2753 case Op_StoreIConditional:
2754 case Op_StoreLConditional:
2755 case Op_CompareAndSwapI:
2756 case Op_CompareAndSwapL:
2757 case Op_CompareAndSwapP:
2758 case Op_CompareAndSwapN:
2759 case Op_GetAndAddI:
2760 case Op_GetAndAddL:
2761 case Op_GetAndSetI:
2762 case Op_GetAndSetL:
2763 case Op_GetAndSetP:
2764 case Op_GetAndSetN:
2765 case Op_StoreP:
2766 case Op_StoreN:
2767 case Op_StoreNKlass:
2768 case Op_LoadB:
2769 case Op_LoadUB:
2770 case Op_LoadUS:
2771 case Op_LoadI:
2772 case Op_LoadKlass:
2773 case Op_LoadNKlass:
2774 case Op_LoadL:
2775 case Op_LoadL_unaligned:
2776 case Op_LoadPLocked:
2777 case Op_LoadP:
2778 case Op_LoadN:
2779 case Op_LoadRange:
2780 case Op_LoadS: {
2781 handle_mem:
2782 #ifdef ASSERT
2783 if( VerifyOptoOopOffsets ) {
2784 assert( n->is_Mem(), "" );
2785 MemNode *mem = (MemNode*)n;
2786 // Check to see if address types have grounded out somehow.
2787 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2788 assert( !tp || oop_offset_is_sane(tp), "" );
2789 }
2790 #endif
2791 break;
2792 }
2793
2794 case Op_AddP: { // Assert sane base pointers
2795 Node *addp = n->in(AddPNode::Address);
2796 assert( !addp->is_AddP() ||
2797 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2798 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2799 "Base pointers must match" );
2800 #ifdef _LP64
2801 if ((UseCompressedOops || UseCompressedClassPointers) &&
2802 addp->Opcode() == Op_ConP &&
2803 addp == n->in(AddPNode::Base) &&
2804 n->in(AddPNode::Offset)->is_Con()) {
2805 // Use addressing with narrow klass to load with offset on x86.
2806 // On sparc loading 32-bits constant and decoding it have less
2807 // instructions (4) then load 64-bits constant (7).
2808 // Do this transformation here since IGVN will convert ConN back to ConP.
2809 const Type* t = addp->bottom_type();
2810 if (t->isa_oopptr() || t->isa_klassptr()) {
2811 Node* nn = NULL;
2812
2813 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2814
2815 // Look for existing ConN node of the same exact type.
2816 Node* r = root();
2817 uint cnt = r->outcnt();
2818 for (uint i = 0; i < cnt; i++) {
2819 Node* m = r->raw_out(i);
2820 if (m!= NULL && m->Opcode() == op &&
2821 m->bottom_type()->make_ptr() == t) {
2822 nn = m;
2823 break;
2824 }
2825 }
2826 if (nn != NULL) {
2827 // Decode a narrow oop to match address
2828 // [R12 + narrow_oop_reg<<3 + offset]
2829 if (t->isa_oopptr()) {
2830 nn = new DecodeNNode(nn, t);
2831 } else {
2832 nn = new DecodeNKlassNode(nn, t);
2833 }
2834 n->set_req(AddPNode::Base, nn);
2835 n->set_req(AddPNode::Address, nn);
2836 if (addp->outcnt() == 0) {
2837 addp->disconnect_inputs(NULL, this);
2838 }
2839 }
2840 }
2841 }
2842 #endif
2843 break;
2844 }
2845
2846 case Op_CastPP: {
2847 // Remove CastPP nodes to gain more freedom during scheduling but
2848 // keep the dependency they encode as control or precedence edges
2849 // (if control is set already) on memory operations. Some CastPP
2850 // nodes don't have a control (don't carry a dependency): skip
2851 // those.
2852 if (n->in(0) != NULL) {
2853 ResourceMark rm;
2854 Unique_Node_List wq;
2855 wq.push(n);
2856 for (uint next = 0; next < wq.size(); ++next) {
2857 Node *m = wq.at(next);
2858 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2859 Node* use = m->fast_out(i);
2860 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2861 use->ensure_control_or_add_prec(n->in(0));
2862 } else if (use->in(0) == NULL) {
2863 switch(use->Opcode()) {
2864 case Op_AddP:
2865 case Op_DecodeN:
2866 case Op_DecodeNKlass:
2867 case Op_CheckCastPP:
2868 case Op_CastPP:
2869 wq.push(use);
2870 break;
2871 }
2872 }
2873 }
2874 }
2875 }
2876 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2877 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2878 Node* in1 = n->in(1);
2879 const Type* t = n->bottom_type();
2880 Node* new_in1 = in1->clone();
2881 new_in1->as_DecodeN()->set_type(t);
2882
2883 if (!Matcher::narrow_oop_use_complex_address()) {
2884 //
2885 // x86, ARM and friends can handle 2 adds in addressing mode
2886 // and Matcher can fold a DecodeN node into address by using
2887 // a narrow oop directly and do implicit NULL check in address:
2888 //
2889 // [R12 + narrow_oop_reg<<3 + offset]
2890 // NullCheck narrow_oop_reg
2891 //
2892 // On other platforms (Sparc) we have to keep new DecodeN node and
2893 // use it to do implicit NULL check in address:
2894 //
2895 // decode_not_null narrow_oop_reg, base_reg
2896 // [base_reg + offset]
2897 // NullCheck base_reg
2898 //
2899 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2900 // to keep the information to which NULL check the new DecodeN node
2901 // corresponds to use it as value in implicit_null_check().
2902 //
2903 new_in1->set_req(0, n->in(0));
2904 }
2905
2906 n->subsume_by(new_in1, this);
2907 if (in1->outcnt() == 0) {
2908 in1->disconnect_inputs(NULL, this);
2909 }
2910 } else {
2911 n->subsume_by(n->in(1), this);
2912 if (n->outcnt() == 0) {
2913 n->disconnect_inputs(NULL, this);
2914 }
2915 }
2916 break;
2917 }
2918 #ifdef _LP64
2919 case Op_CmpP:
2920 // Do this transformation here to preserve CmpPNode::sub() and
2921 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2922 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2923 Node* in1 = n->in(1);
2924 Node* in2 = n->in(2);
2925 if (!in1->is_DecodeNarrowPtr()) {
2926 in2 = in1;
2927 in1 = n->in(2);
2928 }
2929 assert(in1->is_DecodeNarrowPtr(), "sanity");
2930
2931 Node* new_in2 = NULL;
2932 if (in2->is_DecodeNarrowPtr()) {
2933 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2934 new_in2 = in2->in(1);
2935 } else if (in2->Opcode() == Op_ConP) {
2936 const Type* t = in2->bottom_type();
2937 if (t == TypePtr::NULL_PTR) {
2938 assert(in1->is_DecodeN(), "compare klass to null?");
2939 // Don't convert CmpP null check into CmpN if compressed
2940 // oops implicit null check is not generated.
2941 // This will allow to generate normal oop implicit null check.
2942 if (Matcher::gen_narrow_oop_implicit_null_checks())
2943 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
2944 //
2945 // This transformation together with CastPP transformation above
2946 // will generated code for implicit NULL checks for compressed oops.
2947 //
2948 // The original code after Optimize()
2949 //
2950 // LoadN memory, narrow_oop_reg
2951 // decode narrow_oop_reg, base_reg
2952 // CmpP base_reg, NULL
2953 // CastPP base_reg // NotNull
2954 // Load [base_reg + offset], val_reg
2955 //
2956 // after these transformations will be
2957 //
2958 // LoadN memory, narrow_oop_reg
2959 // CmpN narrow_oop_reg, NULL
2960 // decode_not_null narrow_oop_reg, base_reg
2961 // Load [base_reg + offset], val_reg
2962 //
2963 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2964 // since narrow oops can be used in debug info now (see the code in
2965 // final_graph_reshaping_walk()).
2966 //
2967 // At the end the code will be matched to
2968 // on x86:
2969 //
2970 // Load_narrow_oop memory, narrow_oop_reg
2971 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2972 // NullCheck narrow_oop_reg
2973 //
2974 // and on sparc:
2975 //
2976 // Load_narrow_oop memory, narrow_oop_reg
2977 // decode_not_null narrow_oop_reg, base_reg
2978 // Load [base_reg + offset], val_reg
2979 // NullCheck base_reg
2980 //
2981 } else if (t->isa_oopptr()) {
2982 new_in2 = ConNode::make(t->make_narrowoop());
2983 } else if (t->isa_klassptr()) {
2984 new_in2 = ConNode::make(t->make_narrowklass());
2985 }
2986 }
2987 if (new_in2 != NULL) {
2988 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
2989 n->subsume_by(cmpN, this);
2990 if (in1->outcnt() == 0) {
2991 in1->disconnect_inputs(NULL, this);
2992 }
2993 if (in2->outcnt() == 0) {
2994 in2->disconnect_inputs(NULL, this);
2995 }
2996 }
2997 }
2998 break;
2999
3000 case Op_DecodeN:
3001 case Op_DecodeNKlass:
3002 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3003 // DecodeN could be pinned when it can't be fold into
3004 // an address expression, see the code for Op_CastPP above.
3005 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3006 break;
3007
3008 case Op_EncodeP:
3009 case Op_EncodePKlass: {
3010 Node* in1 = n->in(1);
3011 if (in1->is_DecodeNarrowPtr()) {
3012 n->subsume_by(in1->in(1), this);
3013 } else if (in1->Opcode() == Op_ConP) {
3014 const Type* t = in1->bottom_type();
3015 if (t == TypePtr::NULL_PTR) {
3016 assert(t->isa_oopptr(), "null klass?");
3017 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3018 } else if (t->isa_oopptr()) {
3019 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3020 } else if (t->isa_klassptr()) {
3021 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3022 }
3023 }
3024 if (in1->outcnt() == 0) {
3025 in1->disconnect_inputs(NULL, this);
3026 }
3027 break;
3028 }
3029
3030 case Op_Proj: {
3031 if (OptimizeStringConcat) {
3032 ProjNode* p = n->as_Proj();
3033 if (p->_is_io_use) {
3034 // Separate projections were used for the exception path which
3035 // are normally removed by a late inline. If it wasn't inlined
3036 // then they will hang around and should just be replaced with
3037 // the original one.
3038 Node* proj = NULL;
3039 // Replace with just one
3040 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3041 Node *use = i.get();
3042 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3043 proj = use;
3044 break;
3045 }
3046 }
3047 assert(proj != NULL, "must be found");
3048 p->subsume_by(proj, this);
3049 }
3050 }
3051 break;
3052 }
3053
3054 case Op_Phi:
3055 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3056 // The EncodeP optimization may create Phi with the same edges
3057 // for all paths. It is not handled well by Register Allocator.
3058 Node* unique_in = n->in(1);
3059 assert(unique_in != NULL, "");
3060 uint cnt = n->req();
3061 for (uint i = 2; i < cnt; i++) {
3062 Node* m = n->in(i);
3063 assert(m != NULL, "");
3064 if (unique_in != m)
3065 unique_in = NULL;
3066 }
3067 if (unique_in != NULL) {
3068 n->subsume_by(unique_in, this);
3069 }
3070 }
3071 break;
3072
3073 #endif
3074
3075 case Op_ModI:
3076 if (UseDivMod) {
3077 // Check if a%b and a/b both exist
3078 Node* d = n->find_similar(Op_DivI);
3079 if (d) {
3080 // Replace them with a fused divmod if supported
3081 if (Matcher::has_match_rule(Op_DivModI)) {
3082 DivModINode* divmod = DivModINode::make(n);
3083 d->subsume_by(divmod->div_proj(), this);
3084 n->subsume_by(divmod->mod_proj(), this);
3085 } else {
3086 // replace a%b with a-((a/b)*b)
3087 Node* mult = new MulINode(d, d->in(2));
3088 Node* sub = new SubINode(d->in(1), mult);
3089 n->subsume_by(sub, this);
3090 }
3091 }
3092 }
3093 break;
3094
3095 case Op_ModL:
3096 if (UseDivMod) {
3097 // Check if a%b and a/b both exist
3098 Node* d = n->find_similar(Op_DivL);
3099 if (d) {
3100 // Replace them with a fused divmod if supported
3101 if (Matcher::has_match_rule(Op_DivModL)) {
3102 DivModLNode* divmod = DivModLNode::make(n);
3103 d->subsume_by(divmod->div_proj(), this);
3104 n->subsume_by(divmod->mod_proj(), this);
3105 } else {
3106 // replace a%b with a-((a/b)*b)
3107 Node* mult = new MulLNode(d, d->in(2));
3108 Node* sub = new SubLNode(d->in(1), mult);
3109 n->subsume_by(sub, this);
3110 }
3111 }
3112 }
3113 break;
3114
3115 case Op_LoadVector:
3116 case Op_StoreVector:
3117 break;
3118
3119 case Op_AddReductionVI:
3120 case Op_AddReductionVL:
3121 case Op_AddReductionVF:
3122 case Op_AddReductionVD:
3123 case Op_MulReductionVI:
3124 case Op_MulReductionVL:
3125 case Op_MulReductionVF:
3126 case Op_MulReductionVD:
3127 break;
3128
3129 case Op_PackB:
3130 case Op_PackS:
3131 case Op_PackI:
3132 case Op_PackF:
3133 case Op_PackL:
3134 case Op_PackD:
3135 if (n->req()-1 > 2) {
3136 // Replace many operand PackNodes with a binary tree for matching
3137 PackNode* p = (PackNode*) n;
3138 Node* btp = p->binary_tree_pack(1, n->req());
3139 n->subsume_by(btp, this);
3140 }
3141 break;
3142 case Op_Loop:
3143 case Op_CountedLoop:
3144 if (n->as_Loop()->is_inner_loop()) {
3145 frc.inc_inner_loop_count();
3146 }
3147 break;
3148 case Op_LShiftI:
3149 case Op_RShiftI:
3150 case Op_URShiftI:
3151 case Op_LShiftL:
3152 case Op_RShiftL:
3153 case Op_URShiftL:
3154 if (Matcher::need_masked_shift_count) {
3155 // The cpu's shift instructions don't restrict the count to the
3156 // lower 5/6 bits. We need to do the masking ourselves.
3157 Node* in2 = n->in(2);
3158 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3159 const TypeInt* t = in2->find_int_type();
3160 if (t != NULL && t->is_con()) {
3161 juint shift = t->get_con();
3162 if (shift > mask) { // Unsigned cmp
3163 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3164 }
3165 } else {
3166 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3167 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3168 n->set_req(2, shift);
3169 }
3170 }
3171 if (in2->outcnt() == 0) { // Remove dead node
3172 in2->disconnect_inputs(NULL, this);
3173 }
3174 }
3175 break;
3176 case Op_MemBarStoreStore:
3177 case Op_MemBarRelease:
3178 // Break the link with AllocateNode: it is no longer useful and
3179 // confuses register allocation.
3180 if (n->req() > MemBarNode::Precedent) {
3181 n->set_req(MemBarNode::Precedent, top());
3182 }
3183 break;
3184 default:
3185 assert( !n->is_Call(), "" );
3186 assert( !n->is_Mem(), "" );
3187 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3188 break;
3189 }
3190
3191 // Collect CFG split points
3192 if (n->is_MultiBranch())
3193 frc._tests.push(n);
3194 }
3195
3196 //------------------------------final_graph_reshaping_walk---------------------
3197 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3198 // requires that the walk visits a node's inputs before visiting the node.
3199 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3200 ResourceArea *area = Thread::current()->resource_area();
3201 Unique_Node_List sfpt(area);
3202
3203 frc._visited.set(root->_idx); // first, mark node as visited
3204 uint cnt = root->req();
3205 Node *n = root;
3206 uint i = 0;
3207 while (true) {
3208 if (i < cnt) {
3209 // Place all non-visited non-null inputs onto stack
3210 Node* m = n->in(i);
3211 ++i;
3212 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3213 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3214 // compute worst case interpreter size in case of a deoptimization
3215 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3216
3217 sfpt.push(m);
3218 }
3219 cnt = m->req();
3220 nstack.push(n, i); // put on stack parent and next input's index
3221 n = m;
3222 i = 0;
3223 }
3224 } else {
3225 // Now do post-visit work
3226 final_graph_reshaping_impl( n, frc );
3227 if (nstack.is_empty())
3228 break; // finished
3229 n = nstack.node(); // Get node from stack
3230 cnt = n->req();
3231 i = nstack.index();
3232 nstack.pop(); // Shift to the next node on stack
3233 }
3234 }
3235
3236 // Skip next transformation if compressed oops are not used.
3237 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3238 (!UseCompressedOops && !UseCompressedClassPointers))
3239 return;
3240
3241 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3242 // It could be done for an uncommon traps or any safepoints/calls
3243 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3244 while (sfpt.size() > 0) {
3245 n = sfpt.pop();
3246 JVMState *jvms = n->as_SafePoint()->jvms();
3247 assert(jvms != NULL, "sanity");
3248 int start = jvms->debug_start();
3249 int end = n->req();
3250 bool is_uncommon = (n->is_CallStaticJava() &&
3251 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3252 for (int j = start; j < end; j++) {
3253 Node* in = n->in(j);
3254 if (in->is_DecodeNarrowPtr()) {
3255 bool safe_to_skip = true;
3256 if (!is_uncommon ) {
3257 // Is it safe to skip?
3258 for (uint i = 0; i < in->outcnt(); i++) {
3259 Node* u = in->raw_out(i);
3260 if (!u->is_SafePoint() ||
3261 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3262 safe_to_skip = false;
3263 }
3264 }
3265 }
3266 if (safe_to_skip) {
3267 n->set_req(j, in->in(1));
3268 }
3269 if (in->outcnt() == 0) {
3270 in->disconnect_inputs(NULL, this);
3271 }
3272 }
3273 }
3274 }
3275 }
3276
3277 //------------------------------final_graph_reshaping--------------------------
3278 // Final Graph Reshaping.
3279 //
3280 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3281 // and not commoned up and forced early. Must come after regular
3282 // optimizations to avoid GVN undoing the cloning. Clone constant
3283 // inputs to Loop Phis; these will be split by the allocator anyways.
3284 // Remove Opaque nodes.
3285 // (2) Move last-uses by commutative operations to the left input to encourage
3286 // Intel update-in-place two-address operations and better register usage
3287 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3288 // calls canonicalizing them back.
3289 // (3) Count the number of double-precision FP ops, single-precision FP ops
3290 // and call sites. On Intel, we can get correct rounding either by
3291 // forcing singles to memory (requires extra stores and loads after each
3292 // FP bytecode) or we can set a rounding mode bit (requires setting and
3293 // clearing the mode bit around call sites). The mode bit is only used
3294 // if the relative frequency of single FP ops to calls is low enough.
3295 // This is a key transform for SPEC mpeg_audio.
3296 // (4) Detect infinite loops; blobs of code reachable from above but not
3297 // below. Several of the Code_Gen algorithms fail on such code shapes,
3298 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3299 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3300 // Detection is by looking for IfNodes where only 1 projection is
3301 // reachable from below or CatchNodes missing some targets.
3302 // (5) Assert for insane oop offsets in debug mode.
3303
3304 bool Compile::final_graph_reshaping() {
3305 // an infinite loop may have been eliminated by the optimizer,
3306 // in which case the graph will be empty.
3307 if (root()->req() == 1) {
3308 record_method_not_compilable("trivial infinite loop");
3309 return true;
3310 }
3311
3312 // Expensive nodes have their control input set to prevent the GVN
3313 // from freely commoning them. There's no GVN beyond this point so
3314 // no need to keep the control input. We want the expensive nodes to
3315 // be freely moved to the least frequent code path by gcm.
3316 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3317 for (int i = 0; i < expensive_count(); i++) {
3318 _expensive_nodes->at(i)->set_req(0, NULL);
3319 }
3320
3321 Final_Reshape_Counts frc;
3322
3323 // Visit everybody reachable!
3324 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3325 Node_Stack nstack(live_nodes() >> 1);
3326 final_graph_reshaping_walk(nstack, root(), frc);
3327
3328 // Check for unreachable (from below) code (i.e., infinite loops).
3329 for( uint i = 0; i < frc._tests.size(); i++ ) {
3330 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3331 // Get number of CFG targets.
3332 // Note that PCTables include exception targets after calls.
3333 uint required_outcnt = n->required_outcnt();
3334 if (n->outcnt() != required_outcnt) {
3335 // Check for a few special cases. Rethrow Nodes never take the
3336 // 'fall-thru' path, so expected kids is 1 less.
3337 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3338 if (n->in(0)->in(0)->is_Call()) {
3339 CallNode *call = n->in(0)->in(0)->as_Call();
3340 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3341 required_outcnt--; // Rethrow always has 1 less kid
3342 } else if (call->req() > TypeFunc::Parms &&
3343 call->is_CallDynamicJava()) {
3344 // Check for null receiver. In such case, the optimizer has
3345 // detected that the virtual call will always result in a null
3346 // pointer exception. The fall-through projection of this CatchNode
3347 // will not be populated.
3348 Node *arg0 = call->in(TypeFunc::Parms);
3349 if (arg0->is_Type() &&
3350 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3351 required_outcnt--;
3352 }
3353 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3354 call->req() > TypeFunc::Parms+1 &&
3355 call->is_CallStaticJava()) {
3356 // Check for negative array length. In such case, the optimizer has
3357 // detected that the allocation attempt will always result in an
3358 // exception. There is no fall-through projection of this CatchNode .
3359 Node *arg1 = call->in(TypeFunc::Parms+1);
3360 if (arg1->is_Type() &&
3361 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3362 required_outcnt--;
3363 }
3364 }
3365 }
3366 }
3367 // Recheck with a better notion of 'required_outcnt'
3368 if (n->outcnt() != required_outcnt) {
3369 record_method_not_compilable("malformed control flow");
3370 return true; // Not all targets reachable!
3371 }
3372 }
3373 // Check that I actually visited all kids. Unreached kids
3374 // must be infinite loops.
3375 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3376 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3377 record_method_not_compilable("infinite loop");
3378 return true; // Found unvisited kid; must be unreach
3379 }
3380 }
3381
3382 // If original bytecodes contained a mixture of floats and doubles
3383 // check if the optimizer has made it homogenous, item (3).
3384 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3385 frc.get_float_count() > 32 &&
3386 frc.get_double_count() == 0 &&
3387 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3388 set_24_bit_selection_and_mode( false, true );
3389 }
3390
3391 set_java_calls(frc.get_java_call_count());
3392 set_inner_loops(frc.get_inner_loop_count());
3393
3394 // No infinite loops, no reason to bail out.
3395 return false;
3396 }
3397
3398 //-----------------------------too_many_traps----------------------------------
3399 // Report if there are too many traps at the current method and bci.
3400 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3401 bool Compile::too_many_traps(ciMethod* method,
3402 int bci,
3403 Deoptimization::DeoptReason reason) {
3404 ciMethodData* md = method->method_data();
3405 if (md->is_empty()) {
3406 // Assume the trap has not occurred, or that it occurred only
3407 // because of a transient condition during start-up in the interpreter.
3408 return false;
3409 }
3410 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3411 if (md->has_trap_at(bci, m, reason) != 0) {
3412 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3413 // Also, if there are multiple reasons, or if there is no per-BCI record,
3414 // assume the worst.
3415 if (log())
3416 log()->elem("observe trap='%s' count='%d'",
3417 Deoptimization::trap_reason_name(reason),
3418 md->trap_count(reason));
3419 return true;
3420 } else {
3421 // Ignore method/bci and see if there have been too many globally.
3422 return too_many_traps(reason, md);
3423 }
3424 }
3425
3426 // Less-accurate variant which does not require a method and bci.
3427 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3428 ciMethodData* logmd) {
3429 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3430 // Too many traps globally.
3431 // Note that we use cumulative trap_count, not just md->trap_count.
3432 if (log()) {
3433 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3434 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3435 Deoptimization::trap_reason_name(reason),
3436 mcount, trap_count(reason));
3437 }
3438 return true;
3439 } else {
3440 // The coast is clear.
3441 return false;
3442 }
3443 }
3444
3445 //--------------------------too_many_recompiles--------------------------------
3446 // Report if there are too many recompiles at the current method and bci.
3447 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3448 // Is not eager to return true, since this will cause the compiler to use
3449 // Action_none for a trap point, to avoid too many recompilations.
3450 bool Compile::too_many_recompiles(ciMethod* method,
3451 int bci,
3452 Deoptimization::DeoptReason reason) {
3453 ciMethodData* md = method->method_data();
3454 if (md->is_empty()) {
3455 // Assume the trap has not occurred, or that it occurred only
3456 // because of a transient condition during start-up in the interpreter.
3457 return false;
3458 }
3459 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3460 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3461 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3462 Deoptimization::DeoptReason per_bc_reason
3463 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3464 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3465 if ((per_bc_reason == Deoptimization::Reason_none
3466 || md->has_trap_at(bci, m, reason) != 0)
3467 // The trap frequency measure we care about is the recompile count:
3468 && md->trap_recompiled_at(bci, m)
3469 && md->overflow_recompile_count() >= bc_cutoff) {
3470 // Do not emit a trap here if it has already caused recompilations.
3471 // Also, if there are multiple reasons, or if there is no per-BCI record,
3472 // assume the worst.
3473 if (log())
3474 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3475 Deoptimization::trap_reason_name(reason),
3476 md->trap_count(reason),
3477 md->overflow_recompile_count());
3478 return true;
3479 } else if (trap_count(reason) != 0
3480 && decompile_count() >= m_cutoff) {
3481 // Too many recompiles globally, and we have seen this sort of trap.
3482 // Use cumulative decompile_count, not just md->decompile_count.
3483 if (log())
3484 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3485 Deoptimization::trap_reason_name(reason),
3486 md->trap_count(reason), trap_count(reason),
3487 md->decompile_count(), decompile_count());
3488 return true;
3489 } else {
3490 // The coast is clear.
3491 return false;
3492 }
3493 }
3494
3495 // Compute when not to trap. Used by matching trap based nodes and
3496 // NullCheck optimization.
3497 void Compile::set_allowed_deopt_reasons() {
3498 _allowed_reasons = 0;
3499 if (is_method_compilation()) {
3500 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3501 assert(rs < BitsPerInt, "recode bit map");
3502 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3503 _allowed_reasons |= nth_bit(rs);
3504 }
3505 }
3506 }
3507 }
3508
3509 #ifndef PRODUCT
3510 //------------------------------verify_graph_edges---------------------------
3511 // Walk the Graph and verify that there is a one-to-one correspondence
3512 // between Use-Def edges and Def-Use edges in the graph.
3513 void Compile::verify_graph_edges(bool no_dead_code) {
3514 if (VerifyGraphEdges) {
3515 ResourceArea *area = Thread::current()->resource_area();
3516 Unique_Node_List visited(area);
3517 // Call recursive graph walk to check edges
3518 _root->verify_edges(visited);
3519 if (no_dead_code) {
3520 // Now make sure that no visited node is used by an unvisited node.
3521 bool dead_nodes = false;
3522 Unique_Node_List checked(area);
3523 while (visited.size() > 0) {
3524 Node* n = visited.pop();
3525 checked.push(n);
3526 for (uint i = 0; i < n->outcnt(); i++) {
3527 Node* use = n->raw_out(i);
3528 if (checked.member(use)) continue; // already checked
3529 if (visited.member(use)) continue; // already in the graph
3530 if (use->is_Con()) continue; // a dead ConNode is OK
3531 // At this point, we have found a dead node which is DU-reachable.
3532 if (!dead_nodes) {
3533 tty->print_cr("*** Dead nodes reachable via DU edges:");
3534 dead_nodes = true;
3535 }
3536 use->dump(2);
3537 tty->print_cr("---");
3538 checked.push(use); // No repeats; pretend it is now checked.
3539 }
3540 }
3541 assert(!dead_nodes, "using nodes must be reachable from root");
3542 }
3543 }
3544 }
3545
3546 // Verify GC barriers consistency
3547 // Currently supported:
3548 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3549 void Compile::verify_barriers() {
3550 if (UseG1GC) {
3551 // Verify G1 pre-barriers
3552 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3553
3554 ResourceArea *area = Thread::current()->resource_area();
3555 Unique_Node_List visited(area);
3556 Node_List worklist(area);
3557 // We're going to walk control flow backwards starting from the Root
3558 worklist.push(_root);
3559 while (worklist.size() > 0) {
3560 Node* x = worklist.pop();
3561 if (x == NULL || x == top()) continue;
3562 if (visited.member(x)) {
3563 continue;
3564 } else {
3565 visited.push(x);
3566 }
3567
3568 if (x->is_Region()) {
3569 for (uint i = 1; i < x->req(); i++) {
3570 worklist.push(x->in(i));
3571 }
3572 } else {
3573 worklist.push(x->in(0));
3574 // We are looking for the pattern:
3575 // /->ThreadLocal
3576 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3577 // \->ConI(0)
3578 // We want to verify that the If and the LoadB have the same control
3579 // See GraphKit::g1_write_barrier_pre()
3580 if (x->is_If()) {
3581 IfNode *iff = x->as_If();
3582 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3583 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3584 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3585 && cmp->in(1)->is_Load()) {
3586 LoadNode* load = cmp->in(1)->as_Load();
3587 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3588 && load->in(2)->in(3)->is_Con()
3589 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3590
3591 Node* if_ctrl = iff->in(0);
3592 Node* load_ctrl = load->in(0);
3593
3594 if (if_ctrl != load_ctrl) {
3595 // Skip possible CProj->NeverBranch in infinite loops
3596 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3597 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3598 if_ctrl = if_ctrl->in(0)->in(0);
3599 }
3600 }
3601 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3602 }
3603 }
3604 }
3605 }
3606 }
3607 }
3608 }
3609 }
3610
3611 #endif
3612
3613 // The Compile object keeps track of failure reasons separately from the ciEnv.
3614 // This is required because there is not quite a 1-1 relation between the
3615 // ciEnv and its compilation task and the Compile object. Note that one
3616 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3617 // to backtrack and retry without subsuming loads. Other than this backtracking
3618 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3619 // by the logic in C2Compiler.
3620 void Compile::record_failure(const char* reason) {
3621 if (log() != NULL) {
3622 log()->elem("failure reason='%s' phase='compile'", reason);
3623 }
3624 if (_failure_reason == NULL) {
3625 // Record the first failure reason.
3626 _failure_reason = reason;
3627 }
3628
3629 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3630 C->print_method(PHASE_FAILURE);
3631 }
3632 _root = NULL; // flush the graph, too
3633 }
3634
3635 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3636 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3637 _phase_name(name), _dolog(CITimeVerbose)
3638 {
3639 if (_dolog) {
3640 C = Compile::current();
3641 _log = C->log();
3642 } else {
3643 C = NULL;
3644 _log = NULL;
3645 }
3646 if (_log != NULL) {
3647 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3648 _log->stamp();
3649 _log->end_head();
3650 }
3651 }
3652
3653 Compile::TracePhase::~TracePhase() {
3654
3655 C = Compile::current();
3656 if (_dolog) {
3657 _log = C->log();
3658 } else {
3659 _log = NULL;
3660 }
3661
3662 #ifdef ASSERT
3663 if (PrintIdealNodeCount) {
3664 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3665 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3666 }
3667
3668 if (VerifyIdealNodeCount) {
3669 Compile::current()->print_missing_nodes();
3670 }
3671 #endif
3672
3673 if (_log != NULL) {
3674 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3675 }
3676 }
3677
3678 //=============================================================================
3679 // Two Constant's are equal when the type and the value are equal.
3680 bool Compile::Constant::operator==(const Constant& other) {
3681 if (type() != other.type() ) return false;
3682 if (can_be_reused() != other.can_be_reused()) return false;
3683 // For floating point values we compare the bit pattern.
3684 switch (type()) {
3685 case T_FLOAT: return (_v._value.i == other._v._value.i);
3686 case T_LONG:
3687 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3688 case T_OBJECT:
3689 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3690 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3691 case T_METADATA: return (_v._metadata == other._v._metadata);
3692 default: ShouldNotReachHere();
3693 }
3694 return false;
3695 }
3696
3697 static int type_to_size_in_bytes(BasicType t) {
3698 switch (t) {
3699 case T_LONG: return sizeof(jlong );
3700 case T_FLOAT: return sizeof(jfloat );
3701 case T_DOUBLE: return sizeof(jdouble);
3702 case T_METADATA: return sizeof(Metadata*);
3703 // We use T_VOID as marker for jump-table entries (labels) which
3704 // need an internal word relocation.
3705 case T_VOID:
3706 case T_ADDRESS:
3707 case T_OBJECT: return sizeof(jobject);
3708 }
3709
3710 ShouldNotReachHere();
3711 return -1;
3712 }
3713
3714 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3715 // sort descending
3716 if (a->freq() > b->freq()) return -1;
3717 if (a->freq() < b->freq()) return 1;
3718 return 0;
3719 }
3720
3721 void Compile::ConstantTable::calculate_offsets_and_size() {
3722 // First, sort the array by frequencies.
3723 _constants.sort(qsort_comparator);
3724
3725 #ifdef ASSERT
3726 // Make sure all jump-table entries were sorted to the end of the
3727 // array (they have a negative frequency).
3728 bool found_void = false;
3729 for (int i = 0; i < _constants.length(); i++) {
3730 Constant con = _constants.at(i);
3731 if (con.type() == T_VOID)
3732 found_void = true; // jump-tables
3733 else
3734 assert(!found_void, "wrong sorting");
3735 }
3736 #endif
3737
3738 int offset = 0;
3739 for (int i = 0; i < _constants.length(); i++) {
3740 Constant* con = _constants.adr_at(i);
3741
3742 // Align offset for type.
3743 int typesize = type_to_size_in_bytes(con->type());
3744 offset = align_size_up(offset, typesize);
3745 con->set_offset(offset); // set constant's offset
3746
3747 if (con->type() == T_VOID) {
3748 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3749 offset = offset + typesize * n->outcnt(); // expand jump-table
3750 } else {
3751 offset = offset + typesize;
3752 }
3753 }
3754
3755 // Align size up to the next section start (which is insts; see
3756 // CodeBuffer::align_at_start).
3757 assert(_size == -1, "already set?");
3758 _size = align_size_up(offset, CodeEntryAlignment);
3759 }
3760
3761 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3762 MacroAssembler _masm(&cb);
3763 for (int i = 0; i < _constants.length(); i++) {
3764 Constant con = _constants.at(i);
3765 address constant_addr;
3766 switch (con.type()) {
3767 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3768 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3769 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3770 case T_OBJECT: {
3771 jobject obj = con.get_jobject();
3772 int oop_index = _masm.oop_recorder()->find_index(obj);
3773 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3774 break;
3775 }
3776 case T_ADDRESS: {
3777 address addr = (address) con.get_jobject();
3778 constant_addr = _masm.address_constant(addr);
3779 break;
3780 }
3781 // We use T_VOID as marker for jump-table entries (labels) which
3782 // need an internal word relocation.
3783 case T_VOID: {
3784 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3785 // Fill the jump-table with a dummy word. The real value is
3786 // filled in later in fill_jump_table.
3787 address dummy = (address) n;
3788 constant_addr = _masm.address_constant(dummy);
3789 // Expand jump-table
3790 for (uint i = 1; i < n->outcnt(); i++) {
3791 address temp_addr = _masm.address_constant(dummy + i);
3792 assert(temp_addr, "consts section too small");
3793 }
3794 break;
3795 }
3796 case T_METADATA: {
3797 Metadata* obj = con.get_metadata();
3798 int metadata_index = _masm.oop_recorder()->find_index(obj);
3799 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3800 break;
3801 }
3802 default: ShouldNotReachHere();
3803 }
3804 assert(constant_addr, "consts section too small");
3805 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3806 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
3807 }
3808 }
3809
3810 int Compile::ConstantTable::find_offset(Constant& con) const {
3811 int idx = _constants.find(con);
3812 assert(idx != -1, "constant must be in constant table");
3813 int offset = _constants.at(idx).offset();
3814 assert(offset != -1, "constant table not emitted yet?");
3815 return offset;
3816 }
3817
3818 void Compile::ConstantTable::add(Constant& con) {
3819 if (con.can_be_reused()) {
3820 int idx = _constants.find(con);
3821 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3822 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3823 return;
3824 }
3825 }
3826 (void) _constants.append(con);
3827 }
3828
3829 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3830 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3831 Constant con(type, value, b->_freq);
3832 add(con);
3833 return con;
3834 }
3835
3836 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3837 Constant con(metadata);
3838 add(con);
3839 return con;
3840 }
3841
3842 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3843 jvalue value;
3844 BasicType type = oper->type()->basic_type();
3845 switch (type) {
3846 case T_LONG: value.j = oper->constantL(); break;
3847 case T_FLOAT: value.f = oper->constantF(); break;
3848 case T_DOUBLE: value.d = oper->constantD(); break;
3849 case T_OBJECT:
3850 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3851 case T_METADATA: return add((Metadata*)oper->constant()); break;
3852 default: guarantee(false, "unhandled type: %s", type2name(type));
3853 }
3854 return add(n, type, value);
3855 }
3856
3857 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3858 jvalue value;
3859 // We can use the node pointer here to identify the right jump-table
3860 // as this method is called from Compile::Fill_buffer right before
3861 // the MachNodes are emitted and the jump-table is filled (means the
3862 // MachNode pointers do not change anymore).
3863 value.l = (jobject) n;
3864 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3865 add(con);
3866 return con;
3867 }
3868
3869 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3870 // If called from Compile::scratch_emit_size do nothing.
3871 if (Compile::current()->in_scratch_emit_size()) return;
3872
3873 assert(labels.is_nonempty(), "must be");
3874 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
3875
3876 // Since MachConstantNode::constant_offset() also contains
3877 // table_base_offset() we need to subtract the table_base_offset()
3878 // to get the plain offset into the constant table.
3879 int offset = n->constant_offset() - table_base_offset();
3880
3881 MacroAssembler _masm(&cb);
3882 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3883
3884 for (uint i = 0; i < n->outcnt(); i++) {
3885 address* constant_addr = &jump_table_base[i];
3886 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));
3887 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3888 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3889 }
3890 }
3891
3892 //----------------------------static_subtype_check-----------------------------
3893 // Shortcut important common cases when superklass is exact:
3894 // (0) superklass is java.lang.Object (can occur in reflective code)
3895 // (1) subklass is already limited to a subtype of superklass => always ok
3896 // (2) subklass does not overlap with superklass => always fail
3897 // (3) superklass has NO subtypes and we can check with a simple compare.
3898 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3899 if (StressReflectiveCode) {
3900 return SSC_full_test; // Let caller generate the general case.
3901 }
3902
3903 if (superk == env()->Object_klass()) {
3904 return SSC_always_true; // (0) this test cannot fail
3905 }
3906
3907 ciType* superelem = superk;
3908 if (superelem->is_array_klass())
3909 superelem = superelem->as_array_klass()->base_element_type();
3910
3911 if (!subk->is_interface()) { // cannot trust static interface types yet
3912 if (subk->is_subtype_of(superk)) {
3913 return SSC_always_true; // (1) false path dead; no dynamic test needed
3914 }
3915 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
3916 !superk->is_subtype_of(subk)) {
3917 return SSC_always_false;
3918 }
3919 }
3920
3921 // If casting to an instance klass, it must have no subtypes
3922 if (superk->is_interface()) {
3923 // Cannot trust interfaces yet.
3924 // %%% S.B. superk->nof_implementors() == 1
3925 } else if (superelem->is_instance_klass()) {
3926 ciInstanceKlass* ik = superelem->as_instance_klass();
3927 if (!ik->has_subklass() && !ik->is_interface()) {
3928 if (!ik->is_final()) {
3929 // Add a dependency if there is a chance of a later subclass.
3930 dependencies()->assert_leaf_type(ik);
3931 }
3932 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3933 }
3934 } else {
3935 // A primitive array type has no subtypes.
3936 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3937 }
3938
3939 return SSC_full_test;
3940 }
3941
3942 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) {
3943 #ifdef _LP64
3944 // The scaled index operand to AddP must be a clean 64-bit value.
3945 // Java allows a 32-bit int to be incremented to a negative
3946 // value, which appears in a 64-bit register as a large
3947 // positive number. Using that large positive number as an
3948 // operand in pointer arithmetic has bad consequences.
3949 // On the other hand, 32-bit overflow is rare, and the possibility
3950 // can often be excluded, if we annotate the ConvI2L node with
3951 // a type assertion that its value is known to be a small positive
3952 // number. (The prior range check has ensured this.)
3953 // This assertion is used by ConvI2LNode::Ideal.
3954 int index_max = max_jint - 1; // array size is max_jint, index is one less
3955 if (sizetype != NULL) index_max = sizetype->_hi - 1;
3956 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
3957 idx = phase->transform(new ConvI2LNode(idx, lidxtype));
3958 #endif
3959 return idx;
3960 }
3961
3962 // The message about the current inlining is accumulated in
3963 // _print_inlining_stream and transfered into the _print_inlining_list
3964 // once we know whether inlining succeeds or not. For regular
3965 // inlining, messages are appended to the buffer pointed by
3966 // _print_inlining_idx in the _print_inlining_list. For late inlining,
3967 // a new buffer is added after _print_inlining_idx in the list. This
3968 // way we can update the inlining message for late inlining call site
3969 // when the inlining is attempted again.
3970 void Compile::print_inlining_init() {
3971 if (print_inlining() || print_intrinsics()) {
3972 _print_inlining_stream = new stringStream();
3973 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
3974 }
3975 }
3976
3977 void Compile::print_inlining_reinit() {
3978 if (print_inlining() || print_intrinsics()) {
3979 // Re allocate buffer when we change ResourceMark
3980 _print_inlining_stream = new stringStream();
3981 }
3982 }
3983
3984 void Compile::print_inlining_reset() {
3985 _print_inlining_stream->reset();
3986 }
3987
3988 void Compile::print_inlining_commit() {
3989 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
3990 // Transfer the message from _print_inlining_stream to the current
3991 // _print_inlining_list buffer and clear _print_inlining_stream.
3992 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
3993 print_inlining_reset();
3994 }
3995
3996 void Compile::print_inlining_push() {
3997 // Add new buffer to the _print_inlining_list at current position
3998 _print_inlining_idx++;
3999 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4000 }
4001
4002 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4003 return _print_inlining_list->at(_print_inlining_idx);
4004 }
4005
4006 void Compile::print_inlining_update(CallGenerator* cg) {
4007 if (print_inlining() || print_intrinsics()) {
4008 if (!cg->is_late_inline()) {
4009 if (print_inlining_current().cg() != NULL) {
4010 print_inlining_push();
4011 }
4012 print_inlining_commit();
4013 } else {
4014 if (print_inlining_current().cg() != cg &&
4015 (print_inlining_current().cg() != NULL ||
4016 print_inlining_current().ss()->size() != 0)) {
4017 print_inlining_push();
4018 }
4019 print_inlining_commit();
4020 print_inlining_current().set_cg(cg);
4021 }
4022 }
4023 }
4024
4025 void Compile::print_inlining_move_to(CallGenerator* cg) {
4026 // We resume inlining at a late inlining call site. Locate the
4027 // corresponding inlining buffer so that we can update it.
4028 if (print_inlining()) {
4029 for (int i = 0; i < _print_inlining_list->length(); i++) {
4030 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4031 _print_inlining_idx = i;
4032 return;
4033 }
4034 }
4035 ShouldNotReachHere();
4036 }
4037 }
4038
4039 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4040 if (print_inlining()) {
4041 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4042 assert(print_inlining_current().cg() == cg, "wrong entry");
4043 // replace message with new message
4044 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4045 print_inlining_commit();
4046 print_inlining_current().set_cg(cg);
4047 }
4048 }
4049
4050 void Compile::print_inlining_assert_ready() {
4051 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4052 }
4053
4054 void Compile::process_print_inlining() {
4055 bool do_print_inlining = print_inlining() || print_intrinsics();
4056 if (do_print_inlining || log() != NULL) {
4057 // Print inlining message for candidates that we couldn't inline
4058 // for lack of space
4059 for (int i = 0; i < _late_inlines.length(); i++) {
4060 CallGenerator* cg = _late_inlines.at(i);
4061 if (!cg->is_mh_late_inline()) {
4062 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4063 if (do_print_inlining) {
4064 cg->print_inlining_late(msg);
4065 }
4066 log_late_inline_failure(cg, msg);
4067 }
4068 }
4069 }
4070 if (do_print_inlining) {
4071 ResourceMark rm;
4072 stringStream ss;
4073 for (int i = 0; i < _print_inlining_list->length(); i++) {
4074 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4075 }
4076 size_t end = ss.size();
4077 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4078 strncpy(_print_inlining_output, ss.base(), end+1);
4079 _print_inlining_output[end] = 0;
4080 }
4081 }
4082
4083 void Compile::dump_print_inlining() {
4084 if (_print_inlining_output != NULL) {
4085 tty->print_raw(_print_inlining_output);
4086 }
4087 }
4088
4089 void Compile::log_late_inline(CallGenerator* cg) {
4090 if (log() != NULL) {
4091 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4092 cg->unique_id());
4093 JVMState* p = cg->call_node()->jvms();
4094 while (p != NULL) {
4095 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4096 p = p->caller();
4097 }
4098 log()->tail("late_inline");
4099 }
4100 }
4101
4102 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4103 log_late_inline(cg);
4104 if (log() != NULL) {
4105 log()->inline_fail(msg);
4106 }
4107 }
4108
4109 void Compile::log_inline_id(CallGenerator* cg) {
4110 if (log() != NULL) {
4111 // The LogCompilation tool needs a unique way to identify late
4112 // inline call sites. This id must be unique for this call site in
4113 // this compilation. Try to have it unique across compilations as
4114 // well because it can be convenient when grepping through the log
4115 // file.
4116 // Distinguish OSR compilations from others in case CICountOSR is
4117 // on.
4118 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4119 cg->set_unique_id(id);
4120 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4121 }
4122 }
4123
4124 void Compile::log_inline_failure(const char* msg) {
4125 if (C->log() != NULL) {
4126 C->log()->inline_fail(msg);
4127 }
4128 }
4129
4130
4131 // Dump inlining replay data to the stream.
4132 // Don't change thread state and acquire any locks.
4133 void Compile::dump_inline_data(outputStream* out) {
4134 InlineTree* inl_tree = ilt();
4135 if (inl_tree != NULL) {
4136 out->print(" inline %d", inl_tree->count());
4137 inl_tree->dump_replay_data(out);
4138 }
4139 }
4140
4141 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4142 if (n1->Opcode() < n2->Opcode()) return -1;
4143 else if (n1->Opcode() > n2->Opcode()) return 1;
4144
4145 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4146 for (uint i = 1; i < n1->req(); i++) {
4147 if (n1->in(i) < n2->in(i)) return -1;
4148 else if (n1->in(i) > n2->in(i)) return 1;
4149 }
4150
4151 return 0;
4152 }
4153
4154 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4155 Node* n1 = *n1p;
4156 Node* n2 = *n2p;
4157
4158 return cmp_expensive_nodes(n1, n2);
4159 }
4160
4161 void Compile::sort_expensive_nodes() {
4162 if (!expensive_nodes_sorted()) {
4163 _expensive_nodes->sort(cmp_expensive_nodes);
4164 }
4165 }
4166
4167 bool Compile::expensive_nodes_sorted() const {
4168 for (int i = 1; i < _expensive_nodes->length(); i++) {
4169 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4170 return false;
4171 }
4172 }
4173 return true;
4174 }
4175
4176 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4177 if (_expensive_nodes->length() == 0) {
4178 return false;
4179 }
4180
4181 assert(OptimizeExpensiveOps, "optimization off?");
4182
4183 // Take this opportunity to remove dead nodes from the list
4184 int j = 0;
4185 for (int i = 0; i < _expensive_nodes->length(); i++) {
4186 Node* n = _expensive_nodes->at(i);
4187 if (!n->is_unreachable(igvn)) {
4188 assert(n->is_expensive(), "should be expensive");
4189 _expensive_nodes->at_put(j, n);
4190 j++;
4191 }
4192 }
4193 _expensive_nodes->trunc_to(j);
4194
4195 // Then sort the list so that similar nodes are next to each other
4196 // and check for at least two nodes of identical kind with same data
4197 // inputs.
4198 sort_expensive_nodes();
4199
4200 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4201 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4202 return true;
4203 }
4204 }
4205
4206 return false;
4207 }
4208
4209 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4210 if (_expensive_nodes->length() == 0) {
4211 return;
4212 }
4213
4214 assert(OptimizeExpensiveOps, "optimization off?");
4215
4216 // Sort to bring similar nodes next to each other and clear the
4217 // control input of nodes for which there's only a single copy.
4218 sort_expensive_nodes();
4219
4220 int j = 0;
4221 int identical = 0;
4222 int i = 0;
4223 bool modified = false;
4224 for (; i < _expensive_nodes->length()-1; i++) {
4225 assert(j <= i, "can't write beyond current index");
4226 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4227 identical++;
4228 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4229 continue;
4230 }
4231 if (identical > 0) {
4232 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4233 identical = 0;
4234 } else {
4235 Node* n = _expensive_nodes->at(i);
4236 igvn.replace_input_of(n, 0, NULL);
4237 igvn.hash_insert(n);
4238 modified = true;
4239 }
4240 }
4241 if (identical > 0) {
4242 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4243 } else if (_expensive_nodes->length() >= 1) {
4244 Node* n = _expensive_nodes->at(i);
4245 igvn.replace_input_of(n, 0, NULL);
4246 igvn.hash_insert(n);
4247 modified = true;
4248 }
4249 _expensive_nodes->trunc_to(j);
4250 if (modified) {
4251 igvn.optimize();
4252 }
4253 }
4254
4255 void Compile::add_expensive_node(Node * n) {
4256 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4257 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4258 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4259 if (OptimizeExpensiveOps) {
4260 _expensive_nodes->append(n);
4261 } else {
4262 // Clear control input and let IGVN optimize expensive nodes if
4263 // OptimizeExpensiveOps is off.
4264 n->set_req(0, NULL);
4265 }
4266 }
4267
4268 /**
4269 * Remove the speculative part of types and clean up the graph
4270 */
4271 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4272 if (UseTypeSpeculation) {
4273 Unique_Node_List worklist;
4274 worklist.push(root());
4275 int modified = 0;
4276 // Go over all type nodes that carry a speculative type, drop the
4277 // speculative part of the type and enqueue the node for an igvn
4278 // which may optimize it out.
4279 for (uint next = 0; next < worklist.size(); ++next) {
4280 Node *n = worklist.at(next);
4281 if (n->is_Type()) {
4282 TypeNode* tn = n->as_Type();
4283 const Type* t = tn->type();
4284 const Type* t_no_spec = t->remove_speculative();
4285 if (t_no_spec != t) {
4286 bool in_hash = igvn.hash_delete(n);
4287 assert(in_hash, "node should be in igvn hash table");
4288 tn->set_type(t_no_spec);
4289 igvn.hash_insert(n);
4290 igvn._worklist.push(n); // give it a chance to go away
4291 modified++;
4292 }
4293 }
4294 uint max = n->len();
4295 for( uint i = 0; i < max; ++i ) {
4296 Node *m = n->in(i);
4297 if (not_a_node(m)) continue;
4298 worklist.push(m);
4299 }
4300 }
4301 // Drop the speculative part of all types in the igvn's type table
4302 igvn.remove_speculative_types();
4303 if (modified > 0) {
4304 igvn.optimize();
4305 }
4306 #ifdef ASSERT
4307 // Verify that after the IGVN is over no speculative type has resurfaced
4308 worklist.clear();
4309 worklist.push(root());
4310 for (uint next = 0; next < worklist.size(); ++next) {
4311 Node *n = worklist.at(next);
4312 const Type* t = igvn.type_or_null(n);
4313 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4314 if (n->is_Type()) {
4315 t = n->as_Type()->type();
4316 assert(t == t->remove_speculative(), "no more speculative types");
4317 }
4318 uint max = n->len();
4319 for( uint i = 0; i < max; ++i ) {
4320 Node *m = n->in(i);
4321 if (not_a_node(m)) continue;
4322 worklist.push(m);
4323 }
4324 }
4325 igvn.check_no_speculative_types();
4326 #endif
4327 }
4328 }
4329
4330 // Auxiliary method to support randomized stressing/fuzzing.
4331 //
4332 // This method can be called the arbitrary number of times, with current count
4333 // as the argument. The logic allows selecting a single candidate from the
4334 // running list of candidates as follows:
4335 // int count = 0;
4336 // Cand* selected = null;
4337 // while(cand = cand->next()) {
4338 // if (randomized_select(++count)) {
4339 // selected = cand;
4340 // }
4341 // }
4342 //
4343 // Including count equalizes the chances any candidate is "selected".
4344 // This is useful when we don't have the complete list of candidates to choose
4345 // from uniformly. In this case, we need to adjust the randomicity of the
4346 // selection, or else we will end up biasing the selection towards the latter
4347 // candidates.
4348 //
4349 // Quick back-envelope calculation shows that for the list of n candidates
4350 // the equal probability for the candidate to persist as "best" can be
4351 // achieved by replacing it with "next" k-th candidate with the probability
4352 // of 1/k. It can be easily shown that by the end of the run, the
4353 // probability for any candidate is converged to 1/n, thus giving the
4354 // uniform distribution among all the candidates.
4355 //
4356 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4357 #define RANDOMIZED_DOMAIN_POW 29
4358 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4359 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4360 bool Compile::randomized_select(int count) {
4361 assert(count > 0, "only positive");
4362 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4363 }
4364
4365 CloneMap& Compile::clone_map() { return _clone_map; }
4366 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4367
4368 void NodeCloneInfo::dump() const {
4369 tty->print(" {%d:%d} ", idx(), gen());
4370 }
4371
4372 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4373 uint64_t val = value(old->_idx);
4374 NodeCloneInfo cio(val);
4375 assert(val != 0, "old node should be in the map");
4376 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4377 insert(nnn->_idx, cin.get());
4378 #ifndef PRODUCT
4379 if (is_debug()) {
4380 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4381 }
4382 #endif
4383 }
4384
4385 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4386 NodeCloneInfo cio(value(old->_idx));
4387 if (cio.get() == 0) {
4388 cio.set(old->_idx, 0);
4389 insert(old->_idx, cio.get());
4390 #ifndef PRODUCT
4391 if (is_debug()) {
4392 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4393 }
4394 #endif
4395 }
4396 clone(old, nnn, gen);
4397 }
4398
4399 int CloneMap::max_gen() const {
4400 int g = 0;
4401 DictI di(_dict);
4402 for(; di.test(); ++di) {
4403 int t = gen(di._key);
4404 if (g < t) {
4405 g = t;
4406 #ifndef PRODUCT
4407 if (is_debug()) {
4408 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4409 }
4410 #endif
4411 }
4412 }
4413 return g;
4414 }
4415
4416 void CloneMap::dump(node_idx_t key) const {
4417 uint64_t val = value(key);
4418 if (val != 0) {
4419 NodeCloneInfo ni(val);
4420 ni.dump();
4421 }
4422 }