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 = unique();
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->env()->dirset()->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 (env()->dirset()->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(), err_msg_res("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 )
621 : Phase(Compiler),
622 _env(ci_env),
623 _log(ci_env->log()),
624 _compile_id(ci_env->compile_id()),
625 _save_argument_registers(false),
626 _stub_name(NULL),
627 _stub_function(NULL),
628 _stub_entry_point(NULL),
629 _method(target),
630 _entry_bci(osr_bci),
631 _initial_gvn(NULL),
632 _for_igvn(NULL),
633 _warm_calls(NULL),
634 _subsume_loads(subsume_loads),
635 _do_escape_analysis(do_escape_analysis),
636 _eliminate_boxing(eliminate_boxing),
637 _failure_reason(NULL),
638 _code_buffer("Compile::Fill_buffer"),
639 _orig_pc_slot(0),
640 _orig_pc_slot_offset_in_bytes(0),
641 _has_method_handle_invokes(false),
642 _mach_constant_base_node(NULL),
643 _node_bundling_limit(0),
644 _node_bundling_base(NULL),
645 _java_calls(0),
646 _inner_loops(0),
647 _scratch_const_size(-1),
648 _in_scratch_emit_size(false),
649 _dead_node_list(comp_arena()),
650 _dead_node_count(0),
651 #ifndef PRODUCT
652 _trace_opto_output(ci_env->dirset()->TraceOptoOutputOption),
653 _in_dump_cnt(0),
654 _printer(IdealGraphPrinter::printer()),
655 #endif
656 _congraph(NULL),
657 _comp_arena(mtCompiler),
658 _node_arena(mtCompiler),
659 _old_arena(mtCompiler),
660 _Compile_types(mtCompiler),
661 _replay_inline_data(NULL),
662 _late_inlines(comp_arena(), 2, 0, NULL),
663 _string_late_inlines(comp_arena(), 2, 0, NULL),
664 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
665 _late_inlines_pos(0),
666 _number_of_mh_late_inlines(0),
667 _inlining_progress(false),
668 _inlining_incrementally(false),
669 _print_inlining_list(NULL),
670 _print_inlining_stream(NULL),
671 _print_inlining_idx(0),
672 _print_inlining_output(NULL),
673 _interpreter_frame_size(0),
674 _max_node_limit(MaxNodeLimit) {
675 C = this;
676
677 CompileWrapper cw(this);
678
679 if (CITimeVerbose) {
680 tty->print(" ");
681 target->holder()->name()->print();
682 tty->print(".");
683 target->print_short_name();
684 tty->print(" ");
685 }
686 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
687 TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
688
689 #ifndef PRODUCT
690 bool print_opto_assembly = env()->dirset()->PrintOptoAssemblyOption;
691 if (!print_opto_assembly) {
692 bool print_assembly = env()->dirset()->PrintAssemblyOption;
693 if (print_assembly && !Disassembler::can_decode()) {
694 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
695 print_opto_assembly = true;
696 }
697 }
698 set_print_assembly(print_opto_assembly);
699 set_parsed_irreducible_loop(false);
700
701 if (env()->dirset()->ReplayInlineOption) {
702 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
703 }
704 #endif
705 set_print_inlining(env()->dirset()->PrintInliningOption NOT_PRODUCT( || PrintOptoInlining));
706 set_print_intrinsics(env()->dirset()->PrintIntrinsicsOption);
707 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
708
709 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
710 // Make sure the method being compiled gets its own MDO,
711 // so we can at least track the decompile_count().
712 // Need MDO to record RTM code generation state.
713 method()->ensure_method_data();
714 }
715
716 Init(::AliasLevel);
717
718
719 print_compile_messages();
720
721 _ilt = InlineTree::build_inline_tree_root();
722
723 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
724 assert(num_alias_types() >= AliasIdxRaw, "");
725
726 #define MINIMUM_NODE_HASH 1023
727 // Node list that Iterative GVN will start with
728 Unique_Node_List for_igvn(comp_arena());
729 set_for_igvn(&for_igvn);
730
731 // GVN that will be run immediately on new nodes
732 uint estimated_size = method()->code_size()*4+64;
733 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
734 PhaseGVN gvn(node_arena(), estimated_size);
735 set_initial_gvn(&gvn);
736
737 print_inlining_init();
738 { // Scope for timing the parser
739 TracePhase tp("parse", &timers[_t_parser]);
740
741 // Put top into the hash table ASAP.
742 initial_gvn()->transform_no_reclaim(top());
743
744 // Set up tf(), start(), and find a CallGenerator.
745 CallGenerator* cg = NULL;
746 if (is_osr_compilation()) {
747 const TypeTuple *domain = StartOSRNode::osr_domain();
748 const TypeTuple *range = TypeTuple::make_range(method()->signature());
749 init_tf(TypeFunc::make(domain, range));
750 StartNode* s = new StartOSRNode(root(), domain);
751 initial_gvn()->set_type_bottom(s);
752 init_start(s);
753 cg = CallGenerator::for_osr(method(), entry_bci());
754 } else {
755 // Normal case.
756 init_tf(TypeFunc::make(method()));
757 StartNode* s = new StartNode(root(), tf()->domain());
758 initial_gvn()->set_type_bottom(s);
759 init_start(s);
760 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
761 // With java.lang.ref.reference.get() we must go through the
762 // intrinsic when G1 is enabled - even when get() is the root
763 // method of the compile - so that, if necessary, the value in
764 // the referent field of the reference object gets recorded by
765 // the pre-barrier code.
766 // Specifically, if G1 is enabled, the value in the referent
767 // field is recorded by the G1 SATB pre barrier. This will
768 // result in the referent being marked live and the reference
769 // object removed from the list of discovered references during
770 // reference processing.
771 cg = find_intrinsic(method(), false);
772 }
773 if (cg == NULL) {
774 float past_uses = method()->interpreter_invocation_count();
775 float expected_uses = past_uses;
776 cg = CallGenerator::for_inline(method(), expected_uses);
777 }
778 }
779 if (failing()) return;
780 if (cg == NULL) {
781 record_method_not_compilable_all_tiers("cannot parse method");
782 return;
783 }
784 JVMState* jvms = build_start_state(start(), tf());
785 if ((jvms = cg->generate(jvms)) == NULL) {
786 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
787 record_method_not_compilable("method parse failed");
788 }
789 return;
790 }
791 GraphKit kit(jvms);
792
793 if (!kit.stopped()) {
794 // Accept return values, and transfer control we know not where.
795 // This is done by a special, unique ReturnNode bound to root.
796 return_values(kit.jvms());
797 }
798
799 if (kit.has_exceptions()) {
800 // Any exceptions that escape from this call must be rethrown
801 // to whatever caller is dynamically above us on the stack.
802 // This is done by a special, unique RethrowNode bound to root.
803 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
804 }
805
806 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
807
808 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
809 inline_string_calls(true);
810 }
811
812 if (failing()) return;
813
814 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
815
816 // Remove clutter produced by parsing.
817 if (!failing()) {
818 ResourceMark rm;
819 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
820 }
821 }
822
823 // Note: Large methods are capped off in do_one_bytecode().
824 if (failing()) return;
825
826 // After parsing, node notes are no longer automagic.
827 // They must be propagated by register_new_node_with_optimizer(),
828 // clone(), or the like.
829 set_default_node_notes(NULL);
830
831 for (;;) {
832 int successes = Inline_Warm();
833 if (failing()) return;
834 if (successes == 0) break;
835 }
836
837 // Drain the list.
838 Finish_Warm();
839 #ifndef PRODUCT
840 if (_printer && _printer->should_print(1)) {
841 _printer->print_inlining(this);
842 }
843 #endif
844
845 if (failing()) return;
846 NOT_PRODUCT( verify_graph_edges(); )
847
848 // Now optimize
849 Optimize();
850 if (failing()) return;
851 NOT_PRODUCT( verify_graph_edges(); )
852
853 #ifndef PRODUCT
854 if (PrintIdeal) {
855 ttyLocker ttyl; // keep the following output all in one block
856 // This output goes directly to the tty, not the compiler log.
857 // To enable tools to match it up with the compilation activity,
858 // be sure to tag this tty output with the compile ID.
859 if (xtty != NULL) {
860 xtty->head("ideal compile_id='%d'%s", compile_id(),
861 is_osr_compilation() ? " compile_kind='osr'" :
862 "");
863 }
864 root()->dump(9999);
865 if (xtty != NULL) {
866 xtty->tail("ideal");
867 }
868 }
869 #endif
870
871 NOT_PRODUCT( verify_barriers(); )
872
873 // Dump compilation data to replay it.
874 if (env()->dirset()->DumpReplayOption) {
875 env()->dump_replay_data(_compile_id);
876 }
877 if (env()->dirset()->DumpInlineOption && (ilt() != NULL)) {
878 env()->dump_inline_data(_compile_id);
879 }
880
881 // Now that we know the size of all the monitors we can add a fixed slot
882 // for the original deopt pc.
883
884 _orig_pc_slot = fixed_slots();
885 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
886 set_fixed_slots(next_slot);
887
888 // Compute when to use implicit null checks. Used by matching trap based
889 // nodes and NullCheck optimization.
890 set_allowed_deopt_reasons();
891
892 // Now generate code
893 Code_Gen();
894 if (failing()) return;
895
896 // Check if we want to skip execution of all compiled code.
897 {
898 #ifndef PRODUCT
899 if (OptoNoExecute) {
900 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
901 return;
902 }
903 #endif
904 TracePhase tp("install_code", &timers[_t_registerMethod]);
905
906 if (is_osr_compilation()) {
907 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
908 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
909 } else {
910 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
911 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
912 }
913
914 env()->register_method(_method, _entry_bci,
915 &_code_offsets,
916 _orig_pc_slot_offset_in_bytes,
917 code_buffer(),
918 frame_size_in_words(), _oop_map_set,
919 &_handler_table, &_inc_table,
920 compiler,
921 env()->comp_level(),
922 has_unsafe_access(),
923 SharedRuntime::is_wide_vector(max_vector_size()),
924 rtm_state()
925 );
926
927 if (log() != NULL) // Print code cache state into compiler log
928 log()->code_cache_state();
929 }
930 }
931
932 //------------------------------Compile----------------------------------------
933 // Compile a runtime stub
934 Compile::Compile( ciEnv* ci_env,
935 TypeFunc_generator generator,
936 address stub_function,
937 const char *stub_name,
938 int is_fancy_jump,
939 bool pass_tls,
940 bool save_arg_registers,
941 bool return_pc )
942 : Phase(Compiler),
943 _env(ci_env),
944 _log(ci_env->log()),
945 _compile_id(0),
946 _save_argument_registers(save_arg_registers),
947 _method(NULL),
948 _stub_name(stub_name),
949 _stub_function(stub_function),
950 _stub_entry_point(NULL),
951 _entry_bci(InvocationEntryBci),
952 _initial_gvn(NULL),
953 _for_igvn(NULL),
954 _warm_calls(NULL),
955 _orig_pc_slot(0),
956 _orig_pc_slot_offset_in_bytes(0),
957 _subsume_loads(true),
958 _do_escape_analysis(false),
959 _eliminate_boxing(false),
960 _failure_reason(NULL),
961 _code_buffer("Compile::Fill_buffer"),
962 _has_method_handle_invokes(false),
963 _mach_constant_base_node(NULL),
964 _node_bundling_limit(0),
965 _node_bundling_base(NULL),
966 _java_calls(0),
967 _inner_loops(0),
968 #ifndef PRODUCT
969 _trace_opto_output(TraceOptoOutput),
970 _in_dump_cnt(0),
971 _printer(NULL),
972 #endif
973 _comp_arena(mtCompiler),
974 _node_arena(mtCompiler),
975 _old_arena(mtCompiler),
976 _Compile_types(mtCompiler),
977 _dead_node_list(comp_arena()),
978 _dead_node_count(0),
979 _congraph(NULL),
980 _replay_inline_data(NULL),
981 _number_of_mh_late_inlines(0),
982 _inlining_progress(false),
983 _inlining_incrementally(false),
984 _print_inlining_list(NULL),
985 _print_inlining_stream(NULL),
986 _print_inlining_idx(0),
987 _print_inlining_output(NULL),
988 _allowed_reasons(0),
989 _interpreter_frame_size(0),
990 _max_node_limit(MaxNodeLimit) {
991 C = this;
992
993 TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
994 TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
995
996 #ifndef PRODUCT
997 set_print_assembly(PrintFrameConverterAssembly);
998 set_parsed_irreducible_loop(false);
999 #endif
1000 set_has_irreducible_loop(false); // no loops
1001
1002 CompileWrapper cw(this);
1003 Init(/*AliasLevel=*/ 0);
1004 init_tf((*generator)());
1005
1006 {
1007 // The following is a dummy for the sake of GraphKit::gen_stub
1008 Unique_Node_List for_igvn(comp_arena());
1009 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
1010 PhaseGVN gvn(Thread::current()->resource_area(),255);
1011 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
1012 gvn.transform_no_reclaim(top());
1013
1014 GraphKit kit;
1015 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1016 }
1017
1018 NOT_PRODUCT( verify_graph_edges(); )
1019 Code_Gen();
1020 if (failing()) return;
1021
1022
1023 // Entry point will be accessed using compile->stub_entry_point();
1024 if (code_buffer() == NULL) {
1025 Matcher::soft_match_failure();
1026 } else {
1027 if (PrintAssembly && (WizardMode || Verbose))
1028 tty->print_cr("### Stub::%s", stub_name);
1029
1030 if (!failing()) {
1031 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1032
1033 // Make the NMethod
1034 // For now we mark the frame as never safe for profile stackwalking
1035 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1036 code_buffer(),
1037 CodeOffsets::frame_never_safe,
1038 // _code_offsets.value(CodeOffsets::Frame_Complete),
1039 frame_size_in_words(),
1040 _oop_map_set,
1041 save_arg_registers);
1042 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1043
1044 _stub_entry_point = rs->entry_point();
1045 }
1046 }
1047 }
1048
1049 //------------------------------Init-------------------------------------------
1050 // Prepare for a single compilation
1051 void Compile::Init(int aliaslevel) {
1052 _unique = 0;
1053 _regalloc = NULL;
1054
1055 _tf = NULL; // filled in later
1056 _top = NULL; // cached later
1057 _matcher = NULL; // filled in later
1058 _cfg = NULL; // filled in later
1059
1060 set_24_bit_selection_and_mode(Use24BitFP, false);
1061
1062 _node_note_array = NULL;
1063 _default_node_notes = NULL;
1064 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1065
1066 _immutable_memory = NULL; // filled in at first inquiry
1067
1068 // Globally visible Nodes
1069 // First set TOP to NULL to give safe behavior during creation of RootNode
1070 set_cached_top_node(NULL);
1071 set_root(new RootNode());
1072 // Now that you have a Root to point to, create the real TOP
1073 set_cached_top_node( new ConNode(Type::TOP) );
1074 set_recent_alloc(NULL, NULL);
1075
1076 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1077 env()->set_oop_recorder(new OopRecorder(env()->arena()));
1078 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1079 env()->set_dependencies(new Dependencies(env()));
1080
1081 _fixed_slots = 0;
1082 set_has_split_ifs(false);
1083 set_has_loops(has_method() && method()->has_loops()); // first approximation
1084 set_has_stringbuilder(false);
1085 set_has_boxed_value(false);
1086 _trap_can_recompile = false; // no traps emitted yet
1087 _major_progress = true; // start out assuming good things will happen
1088 set_has_unsafe_access(false);
1089 set_max_vector_size(0);
1090 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1091 set_decompile_count(0);
1092
1093 set_do_freq_based_layout(env()->dirset()->BlockLayoutByFrequencyOption);
1094 set_num_loop_opts(LoopOptsCount);
1095 set_do_inlining(Inline);
1096 set_max_inline_size(MaxInlineSize);
1097 set_freq_inline_size(FreqInlineSize);
1098 set_do_scheduling(OptoScheduling);
1099 set_do_count_invocations(false);
1100 set_do_method_data_update(false);
1101
1102 set_do_vector_loop(false);
1103
1104 bool do_vector = false;
1105 if (AllowVectorizeOnDemand) {
1106 if (has_method() && (env()->dirset()->VectorizeOption || env()->dirset()->VectorizeDebugOption)) {
1107 set_do_vector_loop(true);
1108 } else if (has_method() && method()->name() != 0 &&
1109 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1110 set_do_vector_loop(true);
1111 }
1112 #ifndef PRODUCT
1113 if (do_vector_loop() && Verbose) {
1114 tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());
1115 }
1116 #endif
1117 }
1118
1119 set_age_code(has_method() && method()->profile_aging());
1120 set_rtm_state(NoRTM); // No RTM lock eliding by default
1121 _max_node_limit = env()->dirset()->MaxNodeLimitOption;
1122
1123 #if INCLUDE_RTM_OPT
1124 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1125 int rtm_state = method()->method_data()->rtm_state();
1126 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1127 // Don't generate RTM lock eliding code.
1128 set_rtm_state(NoRTM);
1129 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1130 // Generate RTM lock eliding code without abort ratio calculation code.
1131 set_rtm_state(UseRTM);
1132 } else if (UseRTMDeopt) {
1133 // Generate RTM lock eliding code and include abort ratio calculation
1134 // code if UseRTMDeopt is on.
1135 set_rtm_state(ProfileRTM);
1136 }
1137 }
1138 #endif
1139 if (debug_info()->recording_non_safepoints()) {
1140 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1141 (comp_arena(), 8, 0, NULL));
1142 set_default_node_notes(Node_Notes::make(this));
1143 }
1144
1145 // // -- Initialize types before each compile --
1146 // // Update cached type information
1147 // if( _method && _method->constants() )
1148 // Type::update_loaded_types(_method, _method->constants());
1149
1150 // Init alias_type map.
1151 if (!_do_escape_analysis && aliaslevel == 3)
1152 aliaslevel = 2; // No unique types without escape analysis
1153 _AliasLevel = aliaslevel;
1154 const int grow_ats = 16;
1155 _max_alias_types = grow_ats;
1156 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1157 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1158 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1159 {
1160 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1161 }
1162 // Initialize the first few types.
1163 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1164 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1165 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1166 _num_alias_types = AliasIdxRaw+1;
1167 // Zero out the alias type cache.
1168 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1169 // A NULL adr_type hits in the cache right away. Preload the right answer.
1170 probe_alias_cache(NULL)->_index = AliasIdxTop;
1171
1172 _intrinsics = NULL;
1173 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1174 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1175 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1176 register_library_intrinsics();
1177 }
1178
1179 //---------------------------init_start----------------------------------------
1180 // Install the StartNode on this compile object.
1181 void Compile::init_start(StartNode* s) {
1182 if (failing())
1183 return; // already failing
1184 assert(s == start(), "");
1185 }
1186
1187 /**
1188 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1189 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1190 * the ideal graph.
1191 */
1192 StartNode* Compile::start() const {
1193 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
1194 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1195 Node* start = root()->fast_out(i);
1196 if (start->is_Start()) {
1197 return start->as_Start();
1198 }
1199 }
1200 fatal("Did not find Start node!");
1201 return NULL;
1202 }
1203
1204 //-------------------------------immutable_memory-------------------------------------
1205 // Access immutable memory
1206 Node* Compile::immutable_memory() {
1207 if (_immutable_memory != NULL) {
1208 return _immutable_memory;
1209 }
1210 StartNode* s = start();
1211 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1212 Node *p = s->fast_out(i);
1213 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1214 _immutable_memory = p;
1215 return _immutable_memory;
1216 }
1217 }
1218 ShouldNotReachHere();
1219 return NULL;
1220 }
1221
1222 //----------------------set_cached_top_node------------------------------------
1223 // Install the cached top node, and make sure Node::is_top works correctly.
1224 void Compile::set_cached_top_node(Node* tn) {
1225 if (tn != NULL) verify_top(tn);
1226 Node* old_top = _top;
1227 _top = tn;
1228 // Calling Node::setup_is_top allows the nodes the chance to adjust
1229 // their _out arrays.
1230 if (_top != NULL) _top->setup_is_top();
1231 if (old_top != NULL) old_top->setup_is_top();
1232 assert(_top == NULL || top()->is_top(), "");
1233 }
1234
1235 #ifdef ASSERT
1236 uint Compile::count_live_nodes_by_graph_walk() {
1237 Unique_Node_List useful(comp_arena());
1238 // Get useful node list by walking the graph.
1239 identify_useful_nodes(useful);
1240 return useful.size();
1241 }
1242
1243 void Compile::print_missing_nodes() {
1244
1245 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1246 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1247 return;
1248 }
1249
1250 // This is an expensive function. It is executed only when the user
1251 // specifies VerifyIdealNodeCount option or otherwise knows the
1252 // additional work that needs to be done to identify reachable nodes
1253 // by walking the flow graph and find the missing ones using
1254 // _dead_node_list.
1255
1256 Unique_Node_List useful(comp_arena());
1257 // Get useful node list by walking the graph.
1258 identify_useful_nodes(useful);
1259
1260 uint l_nodes = C->live_nodes();
1261 uint l_nodes_by_walk = useful.size();
1262
1263 if (l_nodes != l_nodes_by_walk) {
1264 if (_log != NULL) {
1265 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1266 _log->stamp();
1267 _log->end_head();
1268 }
1269 VectorSet& useful_member_set = useful.member_set();
1270 int last_idx = l_nodes_by_walk;
1271 for (int i = 0; i < last_idx; i++) {
1272 if (useful_member_set.test(i)) {
1273 if (_dead_node_list.test(i)) {
1274 if (_log != NULL) {
1275 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1276 }
1277 if (PrintIdealNodeCount) {
1278 // Print the log message to tty
1279 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1280 useful.at(i)->dump();
1281 }
1282 }
1283 }
1284 else if (! _dead_node_list.test(i)) {
1285 if (_log != NULL) {
1286 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1287 }
1288 if (PrintIdealNodeCount) {
1289 // Print the log message to tty
1290 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1291 }
1292 }
1293 }
1294 if (_log != NULL) {
1295 _log->tail("mismatched_nodes");
1296 }
1297 }
1298 }
1299 void Compile::record_modified_node(Node* n) {
1300 if (_modified_nodes != NULL && !_inlining_incrementally &&
1301 n->outcnt() != 0 && !n->is_Con()) {
1302 _modified_nodes->push(n);
1303 }
1304 }
1305
1306 void Compile::remove_modified_node(Node* n) {
1307 if (_modified_nodes != NULL) {
1308 _modified_nodes->remove(n);
1309 }
1310 }
1311 #endif
1312
1313 #ifndef PRODUCT
1314 void Compile::verify_top(Node* tn) const {
1315 if (tn != NULL) {
1316 assert(tn->is_Con(), "top node must be a constant");
1317 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1318 assert(tn->in(0) != NULL, "must have live top node");
1319 }
1320 }
1321 #endif
1322
1323
1324 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1325
1326 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1327 guarantee(arr != NULL, "");
1328 int num_blocks = arr->length();
1329 if (grow_by < num_blocks) grow_by = num_blocks;
1330 int num_notes = grow_by * _node_notes_block_size;
1331 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1332 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1333 while (num_notes > 0) {
1334 arr->append(notes);
1335 notes += _node_notes_block_size;
1336 num_notes -= _node_notes_block_size;
1337 }
1338 assert(num_notes == 0, "exact multiple, please");
1339 }
1340
1341 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1342 if (source == NULL || dest == NULL) return false;
1343
1344 if (dest->is_Con())
1345 return false; // Do not push debug info onto constants.
1346
1347 #ifdef ASSERT
1348 // Leave a bread crumb trail pointing to the original node:
1349 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1350 dest->set_debug_orig(source);
1351 }
1352 #endif
1353
1354 if (node_note_array() == NULL)
1355 return false; // Not collecting any notes now.
1356
1357 // This is a copy onto a pre-existing node, which may already have notes.
1358 // If both nodes have notes, do not overwrite any pre-existing notes.
1359 Node_Notes* source_notes = node_notes_at(source->_idx);
1360 if (source_notes == NULL || source_notes->is_clear()) return false;
1361 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1362 if (dest_notes == NULL || dest_notes->is_clear()) {
1363 return set_node_notes_at(dest->_idx, source_notes);
1364 }
1365
1366 Node_Notes merged_notes = (*source_notes);
1367 // The order of operations here ensures that dest notes will win...
1368 merged_notes.update_from(dest_notes);
1369 return set_node_notes_at(dest->_idx, &merged_notes);
1370 }
1371
1372
1373 //--------------------------allow_range_check_smearing-------------------------
1374 // Gating condition for coalescing similar range checks.
1375 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1376 // single covering check that is at least as strong as any of them.
1377 // If the optimization succeeds, the simplified (strengthened) range check
1378 // will always succeed. If it fails, we will deopt, and then give up
1379 // on the optimization.
1380 bool Compile::allow_range_check_smearing() const {
1381 // If this method has already thrown a range-check,
1382 // assume it was because we already tried range smearing
1383 // and it failed.
1384 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1385 return !already_trapped;
1386 }
1387
1388
1389 //------------------------------flatten_alias_type-----------------------------
1390 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1391 int offset = tj->offset();
1392 TypePtr::PTR ptr = tj->ptr();
1393
1394 // Known instance (scalarizable allocation) alias only with itself.
1395 bool is_known_inst = tj->isa_oopptr() != NULL &&
1396 tj->is_oopptr()->is_known_instance();
1397
1398 // Process weird unsafe references.
1399 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1400 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1401 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1402 tj = TypeOopPtr::BOTTOM;
1403 ptr = tj->ptr();
1404 offset = tj->offset();
1405 }
1406
1407 // Array pointers need some flattening
1408 const TypeAryPtr *ta = tj->isa_aryptr();
1409 if (ta && ta->is_stable()) {
1410 // Erase stability property for alias analysis.
1411 tj = ta = ta->cast_to_stable(false);
1412 }
1413 if( ta && is_known_inst ) {
1414 if ( offset != Type::OffsetBot &&
1415 offset > arrayOopDesc::length_offset_in_bytes() ) {
1416 offset = Type::OffsetBot; // Flatten constant access into array body only
1417 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1418 }
1419 } else if( ta && _AliasLevel >= 2 ) {
1420 // For arrays indexed by constant indices, we flatten the alias
1421 // space to include all of the array body. Only the header, klass
1422 // and array length can be accessed un-aliased.
1423 if( offset != Type::OffsetBot ) {
1424 if( ta->const_oop() ) { // MethodData* or Method*
1425 offset = Type::OffsetBot; // Flatten constant access into array body
1426 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1427 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1428 // range is OK as-is.
1429 tj = ta = TypeAryPtr::RANGE;
1430 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1431 tj = TypeInstPtr::KLASS; // all klass loads look alike
1432 ta = TypeAryPtr::RANGE; // generic ignored junk
1433 ptr = TypePtr::BotPTR;
1434 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1435 tj = TypeInstPtr::MARK;
1436 ta = TypeAryPtr::RANGE; // generic ignored junk
1437 ptr = TypePtr::BotPTR;
1438 } else { // Random constant offset into array body
1439 offset = Type::OffsetBot; // Flatten constant access into array body
1440 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1441 }
1442 }
1443 // Arrays of fixed size alias with arrays of unknown size.
1444 if (ta->size() != TypeInt::POS) {
1445 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1446 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1447 }
1448 // Arrays of known objects become arrays of unknown objects.
1449 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1450 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1451 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1452 }
1453 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1454 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1455 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1456 }
1457 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1458 // cannot be distinguished by bytecode alone.
1459 if (ta->elem() == TypeInt::BOOL) {
1460 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1461 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1462 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1463 }
1464 // During the 2nd round of IterGVN, NotNull castings are removed.
1465 // Make sure the Bottom and NotNull variants alias the same.
1466 // Also, make sure exact and non-exact variants alias the same.
1467 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1468 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1469 }
1470 }
1471
1472 // Oop pointers need some flattening
1473 const TypeInstPtr *to = tj->isa_instptr();
1474 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1475 ciInstanceKlass *k = to->klass()->as_instance_klass();
1476 if( ptr == TypePtr::Constant ) {
1477 if (to->klass() != ciEnv::current()->Class_klass() ||
1478 offset < k->size_helper() * wordSize) {
1479 // No constant oop pointers (such as Strings); they alias with
1480 // unknown strings.
1481 assert(!is_known_inst, "not scalarizable allocation");
1482 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1483 }
1484 } else if( is_known_inst ) {
1485 tj = to; // Keep NotNull and klass_is_exact for instance type
1486 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1487 // During the 2nd round of IterGVN, NotNull castings are removed.
1488 // Make sure the Bottom and NotNull variants alias the same.
1489 // Also, make sure exact and non-exact variants alias the same.
1490 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1491 }
1492 if (to->speculative() != NULL) {
1493 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1494 }
1495 // Canonicalize the holder of this field
1496 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1497 // First handle header references such as a LoadKlassNode, even if the
1498 // object's klass is unloaded at compile time (4965979).
1499 if (!is_known_inst) { // Do it only for non-instance types
1500 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1501 }
1502 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1503 // Static fields are in the space above the normal instance
1504 // fields in the java.lang.Class instance.
1505 if (to->klass() != ciEnv::current()->Class_klass()) {
1506 to = NULL;
1507 tj = TypeOopPtr::BOTTOM;
1508 offset = tj->offset();
1509 }
1510 } else {
1511 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1512 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1513 if( is_known_inst ) {
1514 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1515 } else {
1516 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1517 }
1518 }
1519 }
1520 }
1521
1522 // Klass pointers to object array klasses need some flattening
1523 const TypeKlassPtr *tk = tj->isa_klassptr();
1524 if( tk ) {
1525 // If we are referencing a field within a Klass, we need
1526 // to assume the worst case of an Object. Both exact and
1527 // inexact types must flatten to the same alias class so
1528 // use NotNull as the PTR.
1529 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1530
1531 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1532 TypeKlassPtr::OBJECT->klass(),
1533 offset);
1534 }
1535
1536 ciKlass* klass = tk->klass();
1537 if( klass->is_obj_array_klass() ) {
1538 ciKlass* k = TypeAryPtr::OOPS->klass();
1539 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1540 k = TypeInstPtr::BOTTOM->klass();
1541 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1542 }
1543
1544 // Check for precise loads from the primary supertype array and force them
1545 // to the supertype cache alias index. Check for generic array loads from
1546 // the primary supertype array and also force them to the supertype cache
1547 // alias index. Since the same load can reach both, we need to merge
1548 // these 2 disparate memories into the same alias class. Since the
1549 // primary supertype array is read-only, there's no chance of confusion
1550 // where we bypass an array load and an array store.
1551 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1552 if (offset == Type::OffsetBot ||
1553 (offset >= primary_supers_offset &&
1554 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1555 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1556 offset = in_bytes(Klass::secondary_super_cache_offset());
1557 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1558 }
1559 }
1560
1561 // Flatten all Raw pointers together.
1562 if (tj->base() == Type::RawPtr)
1563 tj = TypeRawPtr::BOTTOM;
1564
1565 if (tj->base() == Type::AnyPtr)
1566 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1567
1568 // Flatten all to bottom for now
1569 switch( _AliasLevel ) {
1570 case 0:
1571 tj = TypePtr::BOTTOM;
1572 break;
1573 case 1: // Flatten to: oop, static, field or array
1574 switch (tj->base()) {
1575 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1576 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1577 case Type::AryPtr: // do not distinguish arrays at all
1578 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1579 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1580 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1581 default: ShouldNotReachHere();
1582 }
1583 break;
1584 case 2: // No collapsing at level 2; keep all splits
1585 case 3: // No collapsing at level 3; keep all splits
1586 break;
1587 default:
1588 Unimplemented();
1589 }
1590
1591 offset = tj->offset();
1592 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1593
1594 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1595 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1596 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1597 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1598 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1599 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1600 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1601 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1602 assert( tj->ptr() != TypePtr::TopPTR &&
1603 tj->ptr() != TypePtr::AnyNull &&
1604 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1605 // assert( tj->ptr() != TypePtr::Constant ||
1606 // tj->base() == Type::RawPtr ||
1607 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1608
1609 return tj;
1610 }
1611
1612 void Compile::AliasType::Init(int i, const TypePtr* at) {
1613 _index = i;
1614 _adr_type = at;
1615 _field = NULL;
1616 _element = NULL;
1617 _is_rewritable = true; // default
1618 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1619 if (atoop != NULL && atoop->is_known_instance()) {
1620 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1621 _general_index = Compile::current()->get_alias_index(gt);
1622 } else {
1623 _general_index = 0;
1624 }
1625 }
1626
1627 //---------------------------------print_on------------------------------------
1628 #ifndef PRODUCT
1629 void Compile::AliasType::print_on(outputStream* st) {
1630 if (index() < 10)
1631 st->print("@ <%d> ", index());
1632 else st->print("@ <%d>", index());
1633 st->print(is_rewritable() ? " " : " RO");
1634 int offset = adr_type()->offset();
1635 if (offset == Type::OffsetBot)
1636 st->print(" +any");
1637 else st->print(" +%-3d", offset);
1638 st->print(" in ");
1639 adr_type()->dump_on(st);
1640 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1641 if (field() != NULL && tjp) {
1642 if (tjp->klass() != field()->holder() ||
1643 tjp->offset() != field()->offset_in_bytes()) {
1644 st->print(" != ");
1645 field()->print();
1646 st->print(" ***");
1647 }
1648 }
1649 }
1650
1651 void print_alias_types() {
1652 Compile* C = Compile::current();
1653 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1654 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1655 C->alias_type(idx)->print_on(tty);
1656 tty->cr();
1657 }
1658 }
1659 #endif
1660
1661
1662 //----------------------------probe_alias_cache--------------------------------
1663 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1664 intptr_t key = (intptr_t) adr_type;
1665 key ^= key >> logAliasCacheSize;
1666 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1667 }
1668
1669
1670 //-----------------------------grow_alias_types--------------------------------
1671 void Compile::grow_alias_types() {
1672 const int old_ats = _max_alias_types; // how many before?
1673 const int new_ats = old_ats; // how many more?
1674 const int grow_ats = old_ats+new_ats; // how many now?
1675 _max_alias_types = grow_ats;
1676 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1677 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1678 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1679 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1680 }
1681
1682
1683 //--------------------------------find_alias_type------------------------------
1684 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1685 if (_AliasLevel == 0)
1686 return alias_type(AliasIdxBot);
1687
1688 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1689 if (ace->_adr_type == adr_type) {
1690 return alias_type(ace->_index);
1691 }
1692
1693 // Handle special cases.
1694 if (adr_type == NULL) return alias_type(AliasIdxTop);
1695 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1696
1697 // Do it the slow way.
1698 const TypePtr* flat = flatten_alias_type(adr_type);
1699
1700 #ifdef ASSERT
1701 assert(flat == flatten_alias_type(flat), "idempotent");
1702 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1703 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1704 const TypeOopPtr* foop = flat->is_oopptr();
1705 // Scalarizable allocations have exact klass always.
1706 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1707 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1708 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1709 }
1710 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1711 #endif
1712
1713 int idx = AliasIdxTop;
1714 for (int i = 0; i < num_alias_types(); i++) {
1715 if (alias_type(i)->adr_type() == flat) {
1716 idx = i;
1717 break;
1718 }
1719 }
1720
1721 if (idx == AliasIdxTop) {
1722 if (no_create) return NULL;
1723 // Grow the array if necessary.
1724 if (_num_alias_types == _max_alias_types) grow_alias_types();
1725 // Add a new alias type.
1726 idx = _num_alias_types++;
1727 _alias_types[idx]->Init(idx, flat);
1728 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1729 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1730 if (flat->isa_instptr()) {
1731 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1732 && flat->is_instptr()->klass() == env()->Class_klass())
1733 alias_type(idx)->set_rewritable(false);
1734 }
1735 if (flat->isa_aryptr()) {
1736 #ifdef ASSERT
1737 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1738 // (T_BYTE has the weakest alignment and size restrictions...)
1739 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1740 #endif
1741 if (flat->offset() == TypePtr::OffsetBot) {
1742 alias_type(idx)->set_element(flat->is_aryptr()->elem());
1743 }
1744 }
1745 if (flat->isa_klassptr()) {
1746 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1747 alias_type(idx)->set_rewritable(false);
1748 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1749 alias_type(idx)->set_rewritable(false);
1750 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1751 alias_type(idx)->set_rewritable(false);
1752 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1753 alias_type(idx)->set_rewritable(false);
1754 }
1755 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1756 // but the base pointer type is not distinctive enough to identify
1757 // references into JavaThread.)
1758
1759 // Check for final fields.
1760 const TypeInstPtr* tinst = flat->isa_instptr();
1761 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1762 ciField* field;
1763 if (tinst->const_oop() != NULL &&
1764 tinst->klass() == ciEnv::current()->Class_klass() &&
1765 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1766 // static field
1767 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1768 field = k->get_field_by_offset(tinst->offset(), true);
1769 } else {
1770 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1771 field = k->get_field_by_offset(tinst->offset(), false);
1772 }
1773 assert(field == NULL ||
1774 original_field == NULL ||
1775 (field->holder() == original_field->holder() &&
1776 field->offset() == original_field->offset() &&
1777 field->is_static() == original_field->is_static()), "wrong field?");
1778 // Set field() and is_rewritable() attributes.
1779 if (field != NULL) alias_type(idx)->set_field(field);
1780 }
1781 }
1782
1783 // Fill the cache for next time.
1784 ace->_adr_type = adr_type;
1785 ace->_index = idx;
1786 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1787
1788 // Might as well try to fill the cache for the flattened version, too.
1789 AliasCacheEntry* face = probe_alias_cache(flat);
1790 if (face->_adr_type == NULL) {
1791 face->_adr_type = flat;
1792 face->_index = idx;
1793 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1794 }
1795
1796 return alias_type(idx);
1797 }
1798
1799
1800 Compile::AliasType* Compile::alias_type(ciField* field) {
1801 const TypeOopPtr* t;
1802 if (field->is_static())
1803 t = TypeInstPtr::make(field->holder()->java_mirror());
1804 else
1805 t = TypeOopPtr::make_from_klass_raw(field->holder());
1806 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1807 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1808 return atp;
1809 }
1810
1811
1812 //------------------------------have_alias_type--------------------------------
1813 bool Compile::have_alias_type(const TypePtr* adr_type) {
1814 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1815 if (ace->_adr_type == adr_type) {
1816 return true;
1817 }
1818
1819 // Handle special cases.
1820 if (adr_type == NULL) return true;
1821 if (adr_type == TypePtr::BOTTOM) return true;
1822
1823 return find_alias_type(adr_type, true, NULL) != NULL;
1824 }
1825
1826 //-----------------------------must_alias--------------------------------------
1827 // True if all values of the given address type are in the given alias category.
1828 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1829 if (alias_idx == AliasIdxBot) return true; // the universal category
1830 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1831 if (alias_idx == AliasIdxTop) return false; // the empty category
1832 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1833
1834 // the only remaining possible overlap is identity
1835 int adr_idx = get_alias_index(adr_type);
1836 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1837 assert(adr_idx == alias_idx ||
1838 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1839 && adr_type != TypeOopPtr::BOTTOM),
1840 "should not be testing for overlap with an unsafe pointer");
1841 return adr_idx == alias_idx;
1842 }
1843
1844 //------------------------------can_alias--------------------------------------
1845 // True if any values of the given address type are in the given alias category.
1846 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1847 if (alias_idx == AliasIdxTop) return false; // the empty category
1848 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1849 if (alias_idx == AliasIdxBot) return true; // the universal category
1850 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1851
1852 // the only remaining possible overlap is identity
1853 int adr_idx = get_alias_index(adr_type);
1854 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1855 return adr_idx == alias_idx;
1856 }
1857
1858
1859
1860 //---------------------------pop_warm_call-------------------------------------
1861 WarmCallInfo* Compile::pop_warm_call() {
1862 WarmCallInfo* wci = _warm_calls;
1863 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1864 return wci;
1865 }
1866
1867 //----------------------------Inline_Warm--------------------------------------
1868 int Compile::Inline_Warm() {
1869 // If there is room, try to inline some more warm call sites.
1870 // %%% Do a graph index compaction pass when we think we're out of space?
1871 if (!InlineWarmCalls) return 0;
1872
1873 int calls_made_hot = 0;
1874 int room_to_grow = NodeCountInliningCutoff - unique();
1875 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1876 int amount_grown = 0;
1877 WarmCallInfo* call;
1878 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1879 int est_size = (int)call->size();
1880 if (est_size > (room_to_grow - amount_grown)) {
1881 // This one won't fit anyway. Get rid of it.
1882 call->make_cold();
1883 continue;
1884 }
1885 call->make_hot();
1886 calls_made_hot++;
1887 amount_grown += est_size;
1888 amount_to_grow -= est_size;
1889 }
1890
1891 if (calls_made_hot > 0) set_major_progress();
1892 return calls_made_hot;
1893 }
1894
1895
1896 //----------------------------Finish_Warm--------------------------------------
1897 void Compile::Finish_Warm() {
1898 if (!InlineWarmCalls) return;
1899 if (failing()) return;
1900 if (warm_calls() == NULL) return;
1901
1902 // Clean up loose ends, if we are out of space for inlining.
1903 WarmCallInfo* call;
1904 while ((call = pop_warm_call()) != NULL) {
1905 call->make_cold();
1906 }
1907 }
1908
1909 //---------------------cleanup_loop_predicates-----------------------
1910 // Remove the opaque nodes that protect the predicates so that all unused
1911 // checks and uncommon_traps will be eliminated from the ideal graph
1912 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1913 if (predicate_count()==0) return;
1914 for (int i = predicate_count(); i > 0; i--) {
1915 Node * n = predicate_opaque1_node(i-1);
1916 assert(n->Opcode() == Op_Opaque1, "must be");
1917 igvn.replace_node(n, n->in(1));
1918 }
1919 assert(predicate_count()==0, "should be clean!");
1920 }
1921
1922 // StringOpts and late inlining of string methods
1923 void Compile::inline_string_calls(bool parse_time) {
1924 {
1925 // remove useless nodes to make the usage analysis simpler
1926 ResourceMark rm;
1927 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1928 }
1929
1930 {
1931 ResourceMark rm;
1932 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1933 PhaseStringOpts pso(initial_gvn(), for_igvn());
1934 print_method(PHASE_AFTER_STRINGOPTS, 3);
1935 }
1936
1937 // now inline anything that we skipped the first time around
1938 if (!parse_time) {
1939 _late_inlines_pos = _late_inlines.length();
1940 }
1941
1942 while (_string_late_inlines.length() > 0) {
1943 CallGenerator* cg = _string_late_inlines.pop();
1944 cg->do_late_inline();
1945 if (failing()) return;
1946 }
1947 _string_late_inlines.trunc_to(0);
1948 }
1949
1950 // Late inlining of boxing methods
1951 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1952 if (_boxing_late_inlines.length() > 0) {
1953 assert(has_boxed_value(), "inconsistent");
1954
1955 PhaseGVN* gvn = initial_gvn();
1956 set_inlining_incrementally(true);
1957
1958 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1959 for_igvn()->clear();
1960 gvn->replace_with(&igvn);
1961
1962 _late_inlines_pos = _late_inlines.length();
1963
1964 while (_boxing_late_inlines.length() > 0) {
1965 CallGenerator* cg = _boxing_late_inlines.pop();
1966 cg->do_late_inline();
1967 if (failing()) return;
1968 }
1969 _boxing_late_inlines.trunc_to(0);
1970
1971 {
1972 ResourceMark rm;
1973 PhaseRemoveUseless pru(gvn, for_igvn());
1974 }
1975
1976 igvn = PhaseIterGVN(gvn);
1977 igvn.optimize();
1978
1979 set_inlining_progress(false);
1980 set_inlining_incrementally(false);
1981 }
1982 }
1983
1984 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1985 assert(IncrementalInline, "incremental inlining should be on");
1986 PhaseGVN* gvn = initial_gvn();
1987
1988 set_inlining_progress(false);
1989 for_igvn()->clear();
1990 gvn->replace_with(&igvn);
1991
1992 {
1993 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1994 int i = 0;
1995 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1996 CallGenerator* cg = _late_inlines.at(i);
1997 _late_inlines_pos = i+1;
1998 cg->do_late_inline();
1999 if (failing()) return;
2000 }
2001 int j = 0;
2002 for (; i < _late_inlines.length(); i++, j++) {
2003 _late_inlines.at_put(j, _late_inlines.at(i));
2004 }
2005 _late_inlines.trunc_to(j);
2006 }
2007
2008 {
2009 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2010 ResourceMark rm;
2011 PhaseRemoveUseless pru(gvn, for_igvn());
2012 }
2013
2014 {
2015 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2016 igvn = PhaseIterGVN(gvn);
2017 }
2018 }
2019
2020 // Perform incremental inlining until bound on number of live nodes is reached
2021 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2022 TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2023
2024 PhaseGVN* gvn = initial_gvn();
2025
2026 set_inlining_incrementally(true);
2027 set_inlining_progress(true);
2028 uint low_live_nodes = 0;
2029
2030 while(inlining_progress() && _late_inlines.length() > 0) {
2031
2032 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2033 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2034 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2035 // PhaseIdealLoop is expensive so we only try it once we are
2036 // out of live nodes and we only try it again if the previous
2037 // helped got the number of nodes down significantly
2038 PhaseIdealLoop ideal_loop( igvn, false, true );
2039 if (failing()) return;
2040 low_live_nodes = live_nodes();
2041 _major_progress = true;
2042 }
2043
2044 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2045 break;
2046 }
2047 }
2048
2049 inline_incrementally_one(igvn);
2050
2051 if (failing()) return;
2052
2053 {
2054 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2055 igvn.optimize();
2056 }
2057
2058 if (failing()) return;
2059 }
2060
2061 assert( igvn._worklist.size() == 0, "should be done with igvn" );
2062
2063 if (_string_late_inlines.length() > 0) {
2064 assert(has_stringbuilder(), "inconsistent");
2065 for_igvn()->clear();
2066 initial_gvn()->replace_with(&igvn);
2067
2068 inline_string_calls(false);
2069
2070 if (failing()) return;
2071
2072 {
2073 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2074 ResourceMark rm;
2075 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2076 }
2077
2078 {
2079 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2080 igvn = PhaseIterGVN(gvn);
2081 igvn.optimize();
2082 }
2083 }
2084
2085 set_inlining_incrementally(false);
2086 }
2087
2088
2089 //------------------------------Optimize---------------------------------------
2090 // Given a graph, optimize it.
2091 void Compile::Optimize() {
2092 TracePhase tp("optimizer", &timers[_t_optimizer]);
2093
2094 #ifndef PRODUCT
2095 if (env()->dirset()->BreakAtCompileOption) {
2096 BREAKPOINT;
2097 }
2098
2099 #endif
2100
2101 ResourceMark rm;
2102 int loop_opts_cnt;
2103
2104 print_inlining_reinit();
2105
2106 NOT_PRODUCT( verify_graph_edges(); )
2107
2108 print_method(PHASE_AFTER_PARSING);
2109
2110 {
2111 // Iterative Global Value Numbering, including ideal transforms
2112 // Initialize IterGVN with types and values from parse-time GVN
2113 PhaseIterGVN igvn(initial_gvn());
2114 #ifdef ASSERT
2115 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2116 #endif
2117 {
2118 TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2119 igvn.optimize();
2120 }
2121
2122 print_method(PHASE_ITER_GVN1, 2);
2123
2124 if (failing()) return;
2125
2126 inline_incrementally(igvn);
2127
2128 print_method(PHASE_INCREMENTAL_INLINE, 2);
2129
2130 if (failing()) return;
2131
2132 if (eliminate_boxing()) {
2133 // Inline valueOf() methods now.
2134 inline_boxing_calls(igvn);
2135
2136 if (AlwaysIncrementalInline) {
2137 inline_incrementally(igvn);
2138 }
2139
2140 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2141
2142 if (failing()) return;
2143 }
2144
2145 // Remove the speculative part of types and clean up the graph from
2146 // the extra CastPP nodes whose only purpose is to carry them. Do
2147 // that early so that optimizations are not disrupted by the extra
2148 // CastPP nodes.
2149 remove_speculative_types(igvn);
2150
2151 // No more new expensive nodes will be added to the list from here
2152 // so keep only the actual candidates for optimizations.
2153 cleanup_expensive_nodes(igvn);
2154
2155 // Perform escape analysis
2156 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2157 if (has_loops()) {
2158 // Cleanup graph (remove dead nodes).
2159 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2160 PhaseIdealLoop ideal_loop( igvn, false, true );
2161 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2162 if (failing()) return;
2163 }
2164 ConnectionGraph::do_analysis(this, &igvn);
2165
2166 if (failing()) return;
2167
2168 // Optimize out fields loads from scalar replaceable allocations.
2169 igvn.optimize();
2170 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2171
2172 if (failing()) return;
2173
2174 if (congraph() != NULL && macro_count() > 0) {
2175 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2176 PhaseMacroExpand mexp(igvn);
2177 mexp.eliminate_macro_nodes();
2178 igvn.set_delay_transform(false);
2179
2180 igvn.optimize();
2181 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2182
2183 if (failing()) return;
2184 }
2185 }
2186
2187 // Loop transforms on the ideal graph. Range Check Elimination,
2188 // peeling, unrolling, etc.
2189
2190 // Set loop opts counter
2191 loop_opts_cnt = num_loop_opts();
2192 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2193 {
2194 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2195 PhaseIdealLoop ideal_loop( igvn, true );
2196 loop_opts_cnt--;
2197 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2198 if (failing()) return;
2199 }
2200 // Loop opts pass if partial peeling occurred in previous pass
2201 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2202 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2203 PhaseIdealLoop ideal_loop( igvn, false );
2204 loop_opts_cnt--;
2205 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2206 if (failing()) return;
2207 }
2208 // Loop opts pass for loop-unrolling before CCP
2209 if(major_progress() && (loop_opts_cnt > 0)) {
2210 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2211 PhaseIdealLoop ideal_loop( igvn, false );
2212 loop_opts_cnt--;
2213 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2214 }
2215 if (!failing()) {
2216 // Verify that last round of loop opts produced a valid graph
2217 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2218 PhaseIdealLoop::verify(igvn);
2219 }
2220 }
2221 if (failing()) return;
2222
2223 // Conditional Constant Propagation;
2224 PhaseCCP ccp( &igvn );
2225 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2226 {
2227 TracePhase tp("ccp", &timers[_t_ccp]);
2228 ccp.do_transform();
2229 }
2230 print_method(PHASE_CPP1, 2);
2231
2232 assert( true, "Break here to ccp.dump_old2new_map()");
2233
2234 // Iterative Global Value Numbering, including ideal transforms
2235 {
2236 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2237 igvn = ccp;
2238 igvn.optimize();
2239 }
2240
2241 print_method(PHASE_ITER_GVN2, 2);
2242
2243 if (failing()) return;
2244
2245 // Loop transforms on the ideal graph. Range Check Elimination,
2246 // peeling, unrolling, etc.
2247 if(loop_opts_cnt > 0) {
2248 debug_only( int cnt = 0; );
2249 while(major_progress() && (loop_opts_cnt > 0)) {
2250 TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2251 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2252 PhaseIdealLoop ideal_loop( igvn, true);
2253 loop_opts_cnt--;
2254 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2255 if (failing()) return;
2256 }
2257 }
2258
2259 {
2260 // Verify that all previous optimizations produced a valid graph
2261 // at least to this point, even if no loop optimizations were done.
2262 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2263 PhaseIdealLoop::verify(igvn);
2264 }
2265
2266 {
2267 TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2268 PhaseMacroExpand mex(igvn);
2269 if (mex.expand_macro_nodes()) {
2270 assert(failing(), "must bail out w/ explicit message");
2271 return;
2272 }
2273 }
2274
2275 DEBUG_ONLY( _modified_nodes = NULL; )
2276 } // (End scope of igvn; run destructor if necessary for asserts.)
2277
2278 process_print_inlining();
2279 // A method with only infinite loops has no edges entering loops from root
2280 {
2281 TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2282 if (final_graph_reshaping()) {
2283 assert(failing(), "must bail out w/ explicit message");
2284 return;
2285 }
2286 }
2287
2288 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2289 }
2290
2291
2292 //------------------------------Code_Gen---------------------------------------
2293 // Given a graph, generate code for it
2294 void Compile::Code_Gen() {
2295 if (failing()) {
2296 return;
2297 }
2298
2299 // Perform instruction selection. You might think we could reclaim Matcher
2300 // memory PDQ, but actually the Matcher is used in generating spill code.
2301 // Internals of the Matcher (including some VectorSets) must remain live
2302 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2303 // set a bit in reclaimed memory.
2304
2305 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2306 // nodes. Mapping is only valid at the root of each matched subtree.
2307 NOT_PRODUCT( verify_graph_edges(); )
2308
2309 Matcher matcher;
2310 _matcher = &matcher;
2311 {
2312 TracePhase tp("matcher", &timers[_t_matcher]);
2313 matcher.match();
2314 }
2315 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2316 // nodes. Mapping is only valid at the root of each matched subtree.
2317 NOT_PRODUCT( verify_graph_edges(); )
2318
2319 // If you have too many nodes, or if matching has failed, bail out
2320 check_node_count(0, "out of nodes matching instructions");
2321 if (failing()) {
2322 return;
2323 }
2324
2325 // Build a proper-looking CFG
2326 PhaseCFG cfg(node_arena(), root(), matcher);
2327 _cfg = &cfg;
2328 {
2329 TracePhase tp("scheduler", &timers[_t_scheduler]);
2330 bool success = cfg.do_global_code_motion();
2331 if (!success) {
2332 return;
2333 }
2334
2335 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2336 NOT_PRODUCT( verify_graph_edges(); )
2337 debug_only( cfg.verify(); )
2338 }
2339
2340 PhaseChaitin regalloc(unique(), cfg, matcher, false);
2341 _regalloc = ®alloc;
2342 {
2343 TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2344 // Perform register allocation. After Chaitin, use-def chains are
2345 // no longer accurate (at spill code) and so must be ignored.
2346 // Node->LRG->reg mappings are still accurate.
2347 _regalloc->Register_Allocate();
2348
2349 // Bail out if the allocator builds too many nodes
2350 if (failing()) {
2351 return;
2352 }
2353 }
2354
2355 // Prior to register allocation we kept empty basic blocks in case the
2356 // the allocator needed a place to spill. After register allocation we
2357 // are not adding any new instructions. If any basic block is empty, we
2358 // can now safely remove it.
2359 {
2360 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2361 cfg.remove_empty_blocks();
2362 if (do_freq_based_layout()) {
2363 PhaseBlockLayout layout(cfg);
2364 } else {
2365 cfg.set_loop_alignment();
2366 }
2367 cfg.fixup_flow();
2368 }
2369
2370 // Apply peephole optimizations
2371 if( OptoPeephole ) {
2372 TracePhase tp("peephole", &timers[_t_peephole]);
2373 PhasePeephole peep( _regalloc, cfg);
2374 peep.do_transform();
2375 }
2376
2377 // Do late expand if CPU requires this.
2378 if (Matcher::require_postalloc_expand) {
2379 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2380 cfg.postalloc_expand(_regalloc);
2381 }
2382
2383 // Convert Nodes to instruction bits in a buffer
2384 {
2385 TraceTime tp("output", &timers[_t_output], CITime);
2386 Output();
2387 }
2388
2389 print_method(PHASE_FINAL_CODE);
2390
2391 // He's dead, Jim.
2392 _cfg = (PhaseCFG*)0xdeadbeef;
2393 _regalloc = (PhaseChaitin*)0xdeadbeef;
2394 }
2395
2396
2397 //------------------------------dump_asm---------------------------------------
2398 // Dump formatted assembly
2399 #ifndef PRODUCT
2400 void Compile::dump_asm(int *pcs, uint pc_limit) {
2401 bool cut_short = false;
2402 tty->print_cr("#");
2403 tty->print("# "); _tf->dump(); tty->cr();
2404 tty->print_cr("#");
2405
2406 // For all blocks
2407 int pc = 0x0; // Program counter
2408 char starts_bundle = ' ';
2409 _regalloc->dump_frame();
2410
2411 Node *n = NULL;
2412 for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2413 if (VMThread::should_terminate()) {
2414 cut_short = true;
2415 break;
2416 }
2417 Block* block = _cfg->get_block(i);
2418 if (block->is_connector() && !Verbose) {
2419 continue;
2420 }
2421 n = block->head();
2422 if (pcs && n->_idx < pc_limit) {
2423 tty->print("%3.3x ", pcs[n->_idx]);
2424 } else {
2425 tty->print(" ");
2426 }
2427 block->dump_head(_cfg);
2428 if (block->is_connector()) {
2429 tty->print_cr(" # Empty connector block");
2430 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2431 tty->print_cr(" # Block is sole successor of call");
2432 }
2433
2434 // For all instructions
2435 Node *delay = NULL;
2436 for (uint j = 0; j < block->number_of_nodes(); j++) {
2437 if (VMThread::should_terminate()) {
2438 cut_short = true;
2439 break;
2440 }
2441 n = block->get_node(j);
2442 if (valid_bundle_info(n)) {
2443 Bundle* bundle = node_bundling(n);
2444 if (bundle->used_in_unconditional_delay()) {
2445 delay = n;
2446 continue;
2447 }
2448 if (bundle->starts_bundle()) {
2449 starts_bundle = '+';
2450 }
2451 }
2452
2453 if (WizardMode) {
2454 n->dump();
2455 }
2456
2457 if( !n->is_Region() && // Dont print in the Assembly
2458 !n->is_Phi() && // a few noisely useless nodes
2459 !n->is_Proj() &&
2460 !n->is_MachTemp() &&
2461 !n->is_SafePointScalarObject() &&
2462 !n->is_Catch() && // Would be nice to print exception table targets
2463 !n->is_MergeMem() && // Not very interesting
2464 !n->is_top() && // Debug info table constants
2465 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2466 ) {
2467 if (pcs && n->_idx < pc_limit)
2468 tty->print("%3.3x", pcs[n->_idx]);
2469 else
2470 tty->print(" ");
2471 tty->print(" %c ", starts_bundle);
2472 starts_bundle = ' ';
2473 tty->print("\t");
2474 n->format(_regalloc, tty);
2475 tty->cr();
2476 }
2477
2478 // If we have an instruction with a delay slot, and have seen a delay,
2479 // then back up and print it
2480 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2481 assert(delay != NULL, "no unconditional delay instruction");
2482 if (WizardMode) delay->dump();
2483
2484 if (node_bundling(delay)->starts_bundle())
2485 starts_bundle = '+';
2486 if (pcs && n->_idx < pc_limit)
2487 tty->print("%3.3x", pcs[n->_idx]);
2488 else
2489 tty->print(" ");
2490 tty->print(" %c ", starts_bundle);
2491 starts_bundle = ' ';
2492 tty->print("\t");
2493 delay->format(_regalloc, tty);
2494 tty->cr();
2495 delay = NULL;
2496 }
2497
2498 // Dump the exception table as well
2499 if( n->is_Catch() && (Verbose || WizardMode) ) {
2500 // Print the exception table for this offset
2501 _handler_table.print_subtable_for(pc);
2502 }
2503 }
2504
2505 if (pcs && n->_idx < pc_limit)
2506 tty->print_cr("%3.3x", pcs[n->_idx]);
2507 else
2508 tty->cr();
2509
2510 assert(cut_short || delay == NULL, "no unconditional delay branch");
2511
2512 } // End of per-block dump
2513 tty->cr();
2514
2515 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2516 }
2517 #endif
2518
2519 //------------------------------Final_Reshape_Counts---------------------------
2520 // This class defines counters to help identify when a method
2521 // may/must be executed using hardware with only 24-bit precision.
2522 struct Final_Reshape_Counts : public StackObj {
2523 int _call_count; // count non-inlined 'common' calls
2524 int _float_count; // count float ops requiring 24-bit precision
2525 int _double_count; // count double ops requiring more precision
2526 int _java_call_count; // count non-inlined 'java' calls
2527 int _inner_loop_count; // count loops which need alignment
2528 VectorSet _visited; // Visitation flags
2529 Node_List _tests; // Set of IfNodes & PCTableNodes
2530
2531 Final_Reshape_Counts() :
2532 _call_count(0), _float_count(0), _double_count(0),
2533 _java_call_count(0), _inner_loop_count(0),
2534 _visited( Thread::current()->resource_area() ) { }
2535
2536 void inc_call_count () { _call_count ++; }
2537 void inc_float_count () { _float_count ++; }
2538 void inc_double_count() { _double_count++; }
2539 void inc_java_call_count() { _java_call_count++; }
2540 void inc_inner_loop_count() { _inner_loop_count++; }
2541
2542 int get_call_count () const { return _call_count ; }
2543 int get_float_count () const { return _float_count ; }
2544 int get_double_count() const { return _double_count; }
2545 int get_java_call_count() const { return _java_call_count; }
2546 int get_inner_loop_count() const { return _inner_loop_count; }
2547 };
2548
2549 #ifdef ASSERT
2550 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2551 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2552 // Make sure the offset goes inside the instance layout.
2553 return k->contains_field_offset(tp->offset());
2554 // Note that OffsetBot and OffsetTop are very negative.
2555 }
2556 #endif
2557
2558 // Eliminate trivially redundant StoreCMs and accumulate their
2559 // precedence edges.
2560 void Compile::eliminate_redundant_card_marks(Node* n) {
2561 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2562 if (n->in(MemNode::Address)->outcnt() > 1) {
2563 // There are multiple users of the same address so it might be
2564 // possible to eliminate some of the StoreCMs
2565 Node* mem = n->in(MemNode::Memory);
2566 Node* adr = n->in(MemNode::Address);
2567 Node* val = n->in(MemNode::ValueIn);
2568 Node* prev = n;
2569 bool done = false;
2570 // Walk the chain of StoreCMs eliminating ones that match. As
2571 // long as it's a chain of single users then the optimization is
2572 // safe. Eliminating partially redundant StoreCMs would require
2573 // cloning copies down the other paths.
2574 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2575 if (adr == mem->in(MemNode::Address) &&
2576 val == mem->in(MemNode::ValueIn)) {
2577 // redundant StoreCM
2578 if (mem->req() > MemNode::OopStore) {
2579 // Hasn't been processed by this code yet.
2580 n->add_prec(mem->in(MemNode::OopStore));
2581 } else {
2582 // Already converted to precedence edge
2583 for (uint i = mem->req(); i < mem->len(); i++) {
2584 // Accumulate any precedence edges
2585 if (mem->in(i) != NULL) {
2586 n->add_prec(mem->in(i));
2587 }
2588 }
2589 // Everything above this point has been processed.
2590 done = true;
2591 }
2592 // Eliminate the previous StoreCM
2593 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2594 assert(mem->outcnt() == 0, "should be dead");
2595 mem->disconnect_inputs(NULL, this);
2596 } else {
2597 prev = mem;
2598 }
2599 mem = prev->in(MemNode::Memory);
2600 }
2601 }
2602 }
2603
2604 //------------------------------final_graph_reshaping_impl----------------------
2605 // Implement items 1-5 from final_graph_reshaping below.
2606 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2607
2608 if ( n->outcnt() == 0 ) return; // dead node
2609 uint nop = n->Opcode();
2610
2611 // Check for 2-input instruction with "last use" on right input.
2612 // Swap to left input. Implements item (2).
2613 if( n->req() == 3 && // two-input instruction
2614 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2615 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2616 n->in(2)->outcnt() == 1 &&// right use IS a last use
2617 !n->in(2)->is_Con() ) { // right use is not a constant
2618 // Check for commutative opcode
2619 switch( nop ) {
2620 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2621 case Op_MaxI: case Op_MinI:
2622 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2623 case Op_AndL: case Op_XorL: case Op_OrL:
2624 case Op_AndI: case Op_XorI: case Op_OrI: {
2625 // Move "last use" input to left by swapping inputs
2626 n->swap_edges(1, 2);
2627 break;
2628 }
2629 default:
2630 break;
2631 }
2632 }
2633
2634 #ifdef ASSERT
2635 if( n->is_Mem() ) {
2636 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2637 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2638 // oop will be recorded in oop map if load crosses safepoint
2639 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2640 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2641 "raw memory operations should have control edge");
2642 }
2643 #endif
2644 // Count FPU ops and common calls, implements item (3)
2645 switch( nop ) {
2646 // Count all float operations that may use FPU
2647 case Op_AddF:
2648 case Op_SubF:
2649 case Op_MulF:
2650 case Op_DivF:
2651 case Op_NegF:
2652 case Op_ModF:
2653 case Op_ConvI2F:
2654 case Op_ConF:
2655 case Op_CmpF:
2656 case Op_CmpF3:
2657 // case Op_ConvL2F: // longs are split into 32-bit halves
2658 frc.inc_float_count();
2659 break;
2660
2661 case Op_ConvF2D:
2662 case Op_ConvD2F:
2663 frc.inc_float_count();
2664 frc.inc_double_count();
2665 break;
2666
2667 // Count all double operations that may use FPU
2668 case Op_AddD:
2669 case Op_SubD:
2670 case Op_MulD:
2671 case Op_DivD:
2672 case Op_NegD:
2673 case Op_ModD:
2674 case Op_ConvI2D:
2675 case Op_ConvD2I:
2676 // case Op_ConvL2D: // handled by leaf call
2677 // case Op_ConvD2L: // handled by leaf call
2678 case Op_ConD:
2679 case Op_CmpD:
2680 case Op_CmpD3:
2681 frc.inc_double_count();
2682 break;
2683 case Op_Opaque1: // Remove Opaque Nodes before matching
2684 case Op_Opaque2: // Remove Opaque Nodes before matching
2685 case Op_Opaque3:
2686 n->subsume_by(n->in(1), this);
2687 break;
2688 case Op_CallStaticJava:
2689 case Op_CallJava:
2690 case Op_CallDynamicJava:
2691 frc.inc_java_call_count(); // Count java call site;
2692 case Op_CallRuntime:
2693 case Op_CallLeaf:
2694 case Op_CallLeafNoFP: {
2695 assert( n->is_Call(), "" );
2696 CallNode *call = n->as_Call();
2697 // Count call sites where the FP mode bit would have to be flipped.
2698 // Do not count uncommon runtime calls:
2699 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2700 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2701 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2702 frc.inc_call_count(); // Count the call site
2703 } else { // See if uncommon argument is shared
2704 Node *n = call->in(TypeFunc::Parms);
2705 int nop = n->Opcode();
2706 // Clone shared simple arguments to uncommon calls, item (1).
2707 if( n->outcnt() > 1 &&
2708 !n->is_Proj() &&
2709 nop != Op_CreateEx &&
2710 nop != Op_CheckCastPP &&
2711 nop != Op_DecodeN &&
2712 nop != Op_DecodeNKlass &&
2713 !n->is_Mem() ) {
2714 Node *x = n->clone();
2715 call->set_req( TypeFunc::Parms, x );
2716 }
2717 }
2718 break;
2719 }
2720
2721 case Op_StoreD:
2722 case Op_LoadD:
2723 case Op_LoadD_unaligned:
2724 frc.inc_double_count();
2725 goto handle_mem;
2726 case Op_StoreF:
2727 case Op_LoadF:
2728 frc.inc_float_count();
2729 goto handle_mem;
2730
2731 case Op_StoreCM:
2732 {
2733 // Convert OopStore dependence into precedence edge
2734 Node* prec = n->in(MemNode::OopStore);
2735 n->del_req(MemNode::OopStore);
2736 n->add_prec(prec);
2737 eliminate_redundant_card_marks(n);
2738 }
2739
2740 // fall through
2741
2742 case Op_StoreB:
2743 case Op_StoreC:
2744 case Op_StorePConditional:
2745 case Op_StoreI:
2746 case Op_StoreL:
2747 case Op_StoreIConditional:
2748 case Op_StoreLConditional:
2749 case Op_CompareAndSwapI:
2750 case Op_CompareAndSwapL:
2751 case Op_CompareAndSwapP:
2752 case Op_CompareAndSwapN:
2753 case Op_GetAndAddI:
2754 case Op_GetAndAddL:
2755 case Op_GetAndSetI:
2756 case Op_GetAndSetL:
2757 case Op_GetAndSetP:
2758 case Op_GetAndSetN:
2759 case Op_StoreP:
2760 case Op_StoreN:
2761 case Op_StoreNKlass:
2762 case Op_LoadB:
2763 case Op_LoadUB:
2764 case Op_LoadUS:
2765 case Op_LoadI:
2766 case Op_LoadKlass:
2767 case Op_LoadNKlass:
2768 case Op_LoadL:
2769 case Op_LoadL_unaligned:
2770 case Op_LoadPLocked:
2771 case Op_LoadP:
2772 case Op_LoadN:
2773 case Op_LoadRange:
2774 case Op_LoadS: {
2775 handle_mem:
2776 #ifdef ASSERT
2777 if( VerifyOptoOopOffsets ) {
2778 assert( n->is_Mem(), "" );
2779 MemNode *mem = (MemNode*)n;
2780 // Check to see if address types have grounded out somehow.
2781 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2782 assert( !tp || oop_offset_is_sane(tp), "" );
2783 }
2784 #endif
2785 break;
2786 }
2787
2788 case Op_AddP: { // Assert sane base pointers
2789 Node *addp = n->in(AddPNode::Address);
2790 assert( !addp->is_AddP() ||
2791 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2792 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2793 "Base pointers must match" );
2794 #ifdef _LP64
2795 if ((UseCompressedOops || UseCompressedClassPointers) &&
2796 addp->Opcode() == Op_ConP &&
2797 addp == n->in(AddPNode::Base) &&
2798 n->in(AddPNode::Offset)->is_Con()) {
2799 // Use addressing with narrow klass to load with offset on x86.
2800 // On sparc loading 32-bits constant and decoding it have less
2801 // instructions (4) then load 64-bits constant (7).
2802 // Do this transformation here since IGVN will convert ConN back to ConP.
2803 const Type* t = addp->bottom_type();
2804 if (t->isa_oopptr() || t->isa_klassptr()) {
2805 Node* nn = NULL;
2806
2807 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2808
2809 // Look for existing ConN node of the same exact type.
2810 Node* r = root();
2811 uint cnt = r->outcnt();
2812 for (uint i = 0; i < cnt; i++) {
2813 Node* m = r->raw_out(i);
2814 if (m!= NULL && m->Opcode() == op &&
2815 m->bottom_type()->make_ptr() == t) {
2816 nn = m;
2817 break;
2818 }
2819 }
2820 if (nn != NULL) {
2821 // Decode a narrow oop to match address
2822 // [R12 + narrow_oop_reg<<3 + offset]
2823 if (t->isa_oopptr()) {
2824 nn = new DecodeNNode(nn, t);
2825 } else {
2826 nn = new DecodeNKlassNode(nn, t);
2827 }
2828 n->set_req(AddPNode::Base, nn);
2829 n->set_req(AddPNode::Address, nn);
2830 if (addp->outcnt() == 0) {
2831 addp->disconnect_inputs(NULL, this);
2832 }
2833 }
2834 }
2835 }
2836 #endif
2837 break;
2838 }
2839
2840 case Op_CastPP: {
2841 // Remove CastPP nodes to gain more freedom during scheduling but
2842 // keep the dependency they encode as control or precedence edges
2843 // (if control is set already) on memory operations. Some CastPP
2844 // nodes don't have a control (don't carry a dependency): skip
2845 // those.
2846 if (n->in(0) != NULL) {
2847 ResourceMark rm;
2848 Unique_Node_List wq;
2849 wq.push(n);
2850 for (uint next = 0; next < wq.size(); ++next) {
2851 Node *m = wq.at(next);
2852 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2853 Node* use = m->fast_out(i);
2854 if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2855 use->ensure_control_or_add_prec(n->in(0));
2856 } else if (use->in(0) == NULL) {
2857 switch(use->Opcode()) {
2858 case Op_AddP:
2859 case Op_DecodeN:
2860 case Op_DecodeNKlass:
2861 case Op_CheckCastPP:
2862 case Op_CastPP:
2863 wq.push(use);
2864 break;
2865 }
2866 }
2867 }
2868 }
2869 }
2870 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2871 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2872 Node* in1 = n->in(1);
2873 const Type* t = n->bottom_type();
2874 Node* new_in1 = in1->clone();
2875 new_in1->as_DecodeN()->set_type(t);
2876
2877 if (!Matcher::narrow_oop_use_complex_address()) {
2878 //
2879 // x86, ARM and friends can handle 2 adds in addressing mode
2880 // and Matcher can fold a DecodeN node into address by using
2881 // a narrow oop directly and do implicit NULL check in address:
2882 //
2883 // [R12 + narrow_oop_reg<<3 + offset]
2884 // NullCheck narrow_oop_reg
2885 //
2886 // On other platforms (Sparc) we have to keep new DecodeN node and
2887 // use it to do implicit NULL check in address:
2888 //
2889 // decode_not_null narrow_oop_reg, base_reg
2890 // [base_reg + offset]
2891 // NullCheck base_reg
2892 //
2893 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2894 // to keep the information to which NULL check the new DecodeN node
2895 // corresponds to use it as value in implicit_null_check().
2896 //
2897 new_in1->set_req(0, n->in(0));
2898 }
2899
2900 n->subsume_by(new_in1, this);
2901 if (in1->outcnt() == 0) {
2902 in1->disconnect_inputs(NULL, this);
2903 }
2904 } else {
2905 n->subsume_by(n->in(1), this);
2906 if (n->outcnt() == 0) {
2907 n->disconnect_inputs(NULL, this);
2908 }
2909 }
2910 break;
2911 }
2912 #ifdef _LP64
2913 case Op_CmpP:
2914 // Do this transformation here to preserve CmpPNode::sub() and
2915 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2916 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2917 Node* in1 = n->in(1);
2918 Node* in2 = n->in(2);
2919 if (!in1->is_DecodeNarrowPtr()) {
2920 in2 = in1;
2921 in1 = n->in(2);
2922 }
2923 assert(in1->is_DecodeNarrowPtr(), "sanity");
2924
2925 Node* new_in2 = NULL;
2926 if (in2->is_DecodeNarrowPtr()) {
2927 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2928 new_in2 = in2->in(1);
2929 } else if (in2->Opcode() == Op_ConP) {
2930 const Type* t = in2->bottom_type();
2931 if (t == TypePtr::NULL_PTR) {
2932 assert(in1->is_DecodeN(), "compare klass to null?");
2933 // Don't convert CmpP null check into CmpN if compressed
2934 // oops implicit null check is not generated.
2935 // This will allow to generate normal oop implicit null check.
2936 if (Matcher::gen_narrow_oop_implicit_null_checks())
2937 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
2938 //
2939 // This transformation together with CastPP transformation above
2940 // will generated code for implicit NULL checks for compressed oops.
2941 //
2942 // The original code after Optimize()
2943 //
2944 // LoadN memory, narrow_oop_reg
2945 // decode narrow_oop_reg, base_reg
2946 // CmpP base_reg, NULL
2947 // CastPP base_reg // NotNull
2948 // Load [base_reg + offset], val_reg
2949 //
2950 // after these transformations will be
2951 //
2952 // LoadN memory, narrow_oop_reg
2953 // CmpN narrow_oop_reg, NULL
2954 // decode_not_null narrow_oop_reg, base_reg
2955 // Load [base_reg + offset], val_reg
2956 //
2957 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2958 // since narrow oops can be used in debug info now (see the code in
2959 // final_graph_reshaping_walk()).
2960 //
2961 // At the end the code will be matched to
2962 // on x86:
2963 //
2964 // Load_narrow_oop memory, narrow_oop_reg
2965 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2966 // NullCheck narrow_oop_reg
2967 //
2968 // and on sparc:
2969 //
2970 // Load_narrow_oop memory, narrow_oop_reg
2971 // decode_not_null narrow_oop_reg, base_reg
2972 // Load [base_reg + offset], val_reg
2973 // NullCheck base_reg
2974 //
2975 } else if (t->isa_oopptr()) {
2976 new_in2 = ConNode::make(t->make_narrowoop());
2977 } else if (t->isa_klassptr()) {
2978 new_in2 = ConNode::make(t->make_narrowklass());
2979 }
2980 }
2981 if (new_in2 != NULL) {
2982 Node* cmpN = new CmpNNode(in1->in(1), new_in2);
2983 n->subsume_by(cmpN, this);
2984 if (in1->outcnt() == 0) {
2985 in1->disconnect_inputs(NULL, this);
2986 }
2987 if (in2->outcnt() == 0) {
2988 in2->disconnect_inputs(NULL, this);
2989 }
2990 }
2991 }
2992 break;
2993
2994 case Op_DecodeN:
2995 case Op_DecodeNKlass:
2996 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2997 // DecodeN could be pinned when it can't be fold into
2998 // an address expression, see the code for Op_CastPP above.
2999 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3000 break;
3001
3002 case Op_EncodeP:
3003 case Op_EncodePKlass: {
3004 Node* in1 = n->in(1);
3005 if (in1->is_DecodeNarrowPtr()) {
3006 n->subsume_by(in1->in(1), this);
3007 } else if (in1->Opcode() == Op_ConP) {
3008 const Type* t = in1->bottom_type();
3009 if (t == TypePtr::NULL_PTR) {
3010 assert(t->isa_oopptr(), "null klass?");
3011 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3012 } else if (t->isa_oopptr()) {
3013 n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3014 } else if (t->isa_klassptr()) {
3015 n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3016 }
3017 }
3018 if (in1->outcnt() == 0) {
3019 in1->disconnect_inputs(NULL, this);
3020 }
3021 break;
3022 }
3023
3024 case Op_Proj: {
3025 if (OptimizeStringConcat) {
3026 ProjNode* p = n->as_Proj();
3027 if (p->_is_io_use) {
3028 // Separate projections were used for the exception path which
3029 // are normally removed by a late inline. If it wasn't inlined
3030 // then they will hang around and should just be replaced with
3031 // the original one.
3032 Node* proj = NULL;
3033 // Replace with just one
3034 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3035 Node *use = i.get();
3036 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3037 proj = use;
3038 break;
3039 }
3040 }
3041 assert(proj != NULL, "must be found");
3042 p->subsume_by(proj, this);
3043 }
3044 }
3045 break;
3046 }
3047
3048 case Op_Phi:
3049 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3050 // The EncodeP optimization may create Phi with the same edges
3051 // for all paths. It is not handled well by Register Allocator.
3052 Node* unique_in = n->in(1);
3053 assert(unique_in != NULL, "");
3054 uint cnt = n->req();
3055 for (uint i = 2; i < cnt; i++) {
3056 Node* m = n->in(i);
3057 assert(m != NULL, "");
3058 if (unique_in != m)
3059 unique_in = NULL;
3060 }
3061 if (unique_in != NULL) {
3062 n->subsume_by(unique_in, this);
3063 }
3064 }
3065 break;
3066
3067 #endif
3068
3069 case Op_ModI:
3070 if (UseDivMod) {
3071 // Check if a%b and a/b both exist
3072 Node* d = n->find_similar(Op_DivI);
3073 if (d) {
3074 // Replace them with a fused divmod if supported
3075 if (Matcher::has_match_rule(Op_DivModI)) {
3076 DivModINode* divmod = DivModINode::make(n);
3077 d->subsume_by(divmod->div_proj(), this);
3078 n->subsume_by(divmod->mod_proj(), this);
3079 } else {
3080 // replace a%b with a-((a/b)*b)
3081 Node* mult = new MulINode(d, d->in(2));
3082 Node* sub = new SubINode(d->in(1), mult);
3083 n->subsume_by(sub, this);
3084 }
3085 }
3086 }
3087 break;
3088
3089 case Op_ModL:
3090 if (UseDivMod) {
3091 // Check if a%b and a/b both exist
3092 Node* d = n->find_similar(Op_DivL);
3093 if (d) {
3094 // Replace them with a fused divmod if supported
3095 if (Matcher::has_match_rule(Op_DivModL)) {
3096 DivModLNode* divmod = DivModLNode::make(n);
3097 d->subsume_by(divmod->div_proj(), this);
3098 n->subsume_by(divmod->mod_proj(), this);
3099 } else {
3100 // replace a%b with a-((a/b)*b)
3101 Node* mult = new MulLNode(d, d->in(2));
3102 Node* sub = new SubLNode(d->in(1), mult);
3103 n->subsume_by(sub, this);
3104 }
3105 }
3106 }
3107 break;
3108
3109 case Op_LoadVector:
3110 case Op_StoreVector:
3111 break;
3112
3113 case Op_AddReductionVI:
3114 case Op_AddReductionVL:
3115 case Op_AddReductionVF:
3116 case Op_AddReductionVD:
3117 case Op_MulReductionVI:
3118 case Op_MulReductionVL:
3119 case Op_MulReductionVF:
3120 case Op_MulReductionVD:
3121 break;
3122
3123 case Op_PackB:
3124 case Op_PackS:
3125 case Op_PackI:
3126 case Op_PackF:
3127 case Op_PackL:
3128 case Op_PackD:
3129 if (n->req()-1 > 2) {
3130 // Replace many operand PackNodes with a binary tree for matching
3131 PackNode* p = (PackNode*) n;
3132 Node* btp = p->binary_tree_pack(1, n->req());
3133 n->subsume_by(btp, this);
3134 }
3135 break;
3136 case Op_Loop:
3137 case Op_CountedLoop:
3138 if (n->as_Loop()->is_inner_loop()) {
3139 frc.inc_inner_loop_count();
3140 }
3141 break;
3142 case Op_LShiftI:
3143 case Op_RShiftI:
3144 case Op_URShiftI:
3145 case Op_LShiftL:
3146 case Op_RShiftL:
3147 case Op_URShiftL:
3148 if (Matcher::need_masked_shift_count) {
3149 // The cpu's shift instructions don't restrict the count to the
3150 // lower 5/6 bits. We need to do the masking ourselves.
3151 Node* in2 = n->in(2);
3152 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3153 const TypeInt* t = in2->find_int_type();
3154 if (t != NULL && t->is_con()) {
3155 juint shift = t->get_con();
3156 if (shift > mask) { // Unsigned cmp
3157 n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3158 }
3159 } else {
3160 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3161 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3162 n->set_req(2, shift);
3163 }
3164 }
3165 if (in2->outcnt() == 0) { // Remove dead node
3166 in2->disconnect_inputs(NULL, this);
3167 }
3168 }
3169 break;
3170 case Op_MemBarStoreStore:
3171 case Op_MemBarRelease:
3172 // Break the link with AllocateNode: it is no longer useful and
3173 // confuses register allocation.
3174 if (n->req() > MemBarNode::Precedent) {
3175 n->set_req(MemBarNode::Precedent, top());
3176 }
3177 break;
3178 default:
3179 assert( !n->is_Call(), "" );
3180 assert( !n->is_Mem(), "" );
3181 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3182 break;
3183 }
3184
3185 // Collect CFG split points
3186 if (n->is_MultiBranch())
3187 frc._tests.push(n);
3188 }
3189
3190 //------------------------------final_graph_reshaping_walk---------------------
3191 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3192 // requires that the walk visits a node's inputs before visiting the node.
3193 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3194 ResourceArea *area = Thread::current()->resource_area();
3195 Unique_Node_List sfpt(area);
3196
3197 frc._visited.set(root->_idx); // first, mark node as visited
3198 uint cnt = root->req();
3199 Node *n = root;
3200 uint i = 0;
3201 while (true) {
3202 if (i < cnt) {
3203 // Place all non-visited non-null inputs onto stack
3204 Node* m = n->in(i);
3205 ++i;
3206 if (m != NULL && !frc._visited.test_set(m->_idx)) {
3207 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3208 // compute worst case interpreter size in case of a deoptimization
3209 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3210
3211 sfpt.push(m);
3212 }
3213 cnt = m->req();
3214 nstack.push(n, i); // put on stack parent and next input's index
3215 n = m;
3216 i = 0;
3217 }
3218 } else {
3219 // Now do post-visit work
3220 final_graph_reshaping_impl( n, frc );
3221 if (nstack.is_empty())
3222 break; // finished
3223 n = nstack.node(); // Get node from stack
3224 cnt = n->req();
3225 i = nstack.index();
3226 nstack.pop(); // Shift to the next node on stack
3227 }
3228 }
3229
3230 // Skip next transformation if compressed oops are not used.
3231 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3232 (!UseCompressedOops && !UseCompressedClassPointers))
3233 return;
3234
3235 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3236 // It could be done for an uncommon traps or any safepoints/calls
3237 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3238 while (sfpt.size() > 0) {
3239 n = sfpt.pop();
3240 JVMState *jvms = n->as_SafePoint()->jvms();
3241 assert(jvms != NULL, "sanity");
3242 int start = jvms->debug_start();
3243 int end = n->req();
3244 bool is_uncommon = (n->is_CallStaticJava() &&
3245 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3246 for (int j = start; j < end; j++) {
3247 Node* in = n->in(j);
3248 if (in->is_DecodeNarrowPtr()) {
3249 bool safe_to_skip = true;
3250 if (!is_uncommon ) {
3251 // Is it safe to skip?
3252 for (uint i = 0; i < in->outcnt(); i++) {
3253 Node* u = in->raw_out(i);
3254 if (!u->is_SafePoint() ||
3255 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3256 safe_to_skip = false;
3257 }
3258 }
3259 }
3260 if (safe_to_skip) {
3261 n->set_req(j, in->in(1));
3262 }
3263 if (in->outcnt() == 0) {
3264 in->disconnect_inputs(NULL, this);
3265 }
3266 }
3267 }
3268 }
3269 }
3270
3271 //------------------------------final_graph_reshaping--------------------------
3272 // Final Graph Reshaping.
3273 //
3274 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3275 // and not commoned up and forced early. Must come after regular
3276 // optimizations to avoid GVN undoing the cloning. Clone constant
3277 // inputs to Loop Phis; these will be split by the allocator anyways.
3278 // Remove Opaque nodes.
3279 // (2) Move last-uses by commutative operations to the left input to encourage
3280 // Intel update-in-place two-address operations and better register usage
3281 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3282 // calls canonicalizing them back.
3283 // (3) Count the number of double-precision FP ops, single-precision FP ops
3284 // and call sites. On Intel, we can get correct rounding either by
3285 // forcing singles to memory (requires extra stores and loads after each
3286 // FP bytecode) or we can set a rounding mode bit (requires setting and
3287 // clearing the mode bit around call sites). The mode bit is only used
3288 // if the relative frequency of single FP ops to calls is low enough.
3289 // This is a key transform for SPEC mpeg_audio.
3290 // (4) Detect infinite loops; blobs of code reachable from above but not
3291 // below. Several of the Code_Gen algorithms fail on such code shapes,
3292 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3293 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3294 // Detection is by looking for IfNodes where only 1 projection is
3295 // reachable from below or CatchNodes missing some targets.
3296 // (5) Assert for insane oop offsets in debug mode.
3297
3298 bool Compile::final_graph_reshaping() {
3299 // an infinite loop may have been eliminated by the optimizer,
3300 // in which case the graph will be empty.
3301 if (root()->req() == 1) {
3302 record_method_not_compilable("trivial infinite loop");
3303 return true;
3304 }
3305
3306 // Expensive nodes have their control input set to prevent the GVN
3307 // from freely commoning them. There's no GVN beyond this point so
3308 // no need to keep the control input. We want the expensive nodes to
3309 // be freely moved to the least frequent code path by gcm.
3310 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3311 for (int i = 0; i < expensive_count(); i++) {
3312 _expensive_nodes->at(i)->set_req(0, NULL);
3313 }
3314
3315 Final_Reshape_Counts frc;
3316
3317 // Visit everybody reachable!
3318 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3319 Node_Stack nstack(live_nodes() >> 1);
3320 final_graph_reshaping_walk(nstack, root(), frc);
3321
3322 // Check for unreachable (from below) code (i.e., infinite loops).
3323 for( uint i = 0; i < frc._tests.size(); i++ ) {
3324 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3325 // Get number of CFG targets.
3326 // Note that PCTables include exception targets after calls.
3327 uint required_outcnt = n->required_outcnt();
3328 if (n->outcnt() != required_outcnt) {
3329 // Check for a few special cases. Rethrow Nodes never take the
3330 // 'fall-thru' path, so expected kids is 1 less.
3331 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3332 if (n->in(0)->in(0)->is_Call()) {
3333 CallNode *call = n->in(0)->in(0)->as_Call();
3334 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3335 required_outcnt--; // Rethrow always has 1 less kid
3336 } else if (call->req() > TypeFunc::Parms &&
3337 call->is_CallDynamicJava()) {
3338 // Check for null receiver. In such case, the optimizer has
3339 // detected that the virtual call will always result in a null
3340 // pointer exception. The fall-through projection of this CatchNode
3341 // will not be populated.
3342 Node *arg0 = call->in(TypeFunc::Parms);
3343 if (arg0->is_Type() &&
3344 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3345 required_outcnt--;
3346 }
3347 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3348 call->req() > TypeFunc::Parms+1 &&
3349 call->is_CallStaticJava()) {
3350 // Check for negative array length. In such case, the optimizer has
3351 // detected that the allocation attempt will always result in an
3352 // exception. There is no fall-through projection of this CatchNode .
3353 Node *arg1 = call->in(TypeFunc::Parms+1);
3354 if (arg1->is_Type() &&
3355 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3356 required_outcnt--;
3357 }
3358 }
3359 }
3360 }
3361 // Recheck with a better notion of 'required_outcnt'
3362 if (n->outcnt() != required_outcnt) {
3363 record_method_not_compilable("malformed control flow");
3364 return true; // Not all targets reachable!
3365 }
3366 }
3367 // Check that I actually visited all kids. Unreached kids
3368 // must be infinite loops.
3369 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3370 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3371 record_method_not_compilable("infinite loop");
3372 return true; // Found unvisited kid; must be unreach
3373 }
3374 }
3375
3376 // If original bytecodes contained a mixture of floats and doubles
3377 // check if the optimizer has made it homogenous, item (3).
3378 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3379 frc.get_float_count() > 32 &&
3380 frc.get_double_count() == 0 &&
3381 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3382 set_24_bit_selection_and_mode( false, true );
3383 }
3384
3385 set_java_calls(frc.get_java_call_count());
3386 set_inner_loops(frc.get_inner_loop_count());
3387
3388 // No infinite loops, no reason to bail out.
3389 return false;
3390 }
3391
3392 //-----------------------------too_many_traps----------------------------------
3393 // Report if there are too many traps at the current method and bci.
3394 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3395 bool Compile::too_many_traps(ciMethod* method,
3396 int bci,
3397 Deoptimization::DeoptReason reason) {
3398 ciMethodData* md = method->method_data();
3399 if (md->is_empty()) {
3400 // Assume the trap has not occurred, or that it occurred only
3401 // because of a transient condition during start-up in the interpreter.
3402 return false;
3403 }
3404 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3405 if (md->has_trap_at(bci, m, reason) != 0) {
3406 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3407 // Also, if there are multiple reasons, or if there is no per-BCI record,
3408 // assume the worst.
3409 if (log())
3410 log()->elem("observe trap='%s' count='%d'",
3411 Deoptimization::trap_reason_name(reason),
3412 md->trap_count(reason));
3413 return true;
3414 } else {
3415 // Ignore method/bci and see if there have been too many globally.
3416 return too_many_traps(reason, md);
3417 }
3418 }
3419
3420 // Less-accurate variant which does not require a method and bci.
3421 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3422 ciMethodData* logmd) {
3423 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3424 // Too many traps globally.
3425 // Note that we use cumulative trap_count, not just md->trap_count.
3426 if (log()) {
3427 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3428 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3429 Deoptimization::trap_reason_name(reason),
3430 mcount, trap_count(reason));
3431 }
3432 return true;
3433 } else {
3434 // The coast is clear.
3435 return false;
3436 }
3437 }
3438
3439 //--------------------------too_many_recompiles--------------------------------
3440 // Report if there are too many recompiles at the current method and bci.
3441 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3442 // Is not eager to return true, since this will cause the compiler to use
3443 // Action_none for a trap point, to avoid too many recompilations.
3444 bool Compile::too_many_recompiles(ciMethod* method,
3445 int bci,
3446 Deoptimization::DeoptReason reason) {
3447 ciMethodData* md = method->method_data();
3448 if (md->is_empty()) {
3449 // Assume the trap has not occurred, or that it occurred only
3450 // because of a transient condition during start-up in the interpreter.
3451 return false;
3452 }
3453 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3454 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3455 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3456 Deoptimization::DeoptReason per_bc_reason
3457 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3458 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3459 if ((per_bc_reason == Deoptimization::Reason_none
3460 || md->has_trap_at(bci, m, reason) != 0)
3461 // The trap frequency measure we care about is the recompile count:
3462 && md->trap_recompiled_at(bci, m)
3463 && md->overflow_recompile_count() >= bc_cutoff) {
3464 // Do not emit a trap here if it has already caused recompilations.
3465 // Also, if there are multiple reasons, or if there is no per-BCI record,
3466 // assume the worst.
3467 if (log())
3468 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3469 Deoptimization::trap_reason_name(reason),
3470 md->trap_count(reason),
3471 md->overflow_recompile_count());
3472 return true;
3473 } else if (trap_count(reason) != 0
3474 && decompile_count() >= m_cutoff) {
3475 // Too many recompiles globally, and we have seen this sort of trap.
3476 // Use cumulative decompile_count, not just md->decompile_count.
3477 if (log())
3478 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3479 Deoptimization::trap_reason_name(reason),
3480 md->trap_count(reason), trap_count(reason),
3481 md->decompile_count(), decompile_count());
3482 return true;
3483 } else {
3484 // The coast is clear.
3485 return false;
3486 }
3487 }
3488
3489 // Compute when not to trap. Used by matching trap based nodes and
3490 // NullCheck optimization.
3491 void Compile::set_allowed_deopt_reasons() {
3492 _allowed_reasons = 0;
3493 if (is_method_compilation()) {
3494 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3495 assert(rs < BitsPerInt, "recode bit map");
3496 if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3497 _allowed_reasons |= nth_bit(rs);
3498 }
3499 }
3500 }
3501 }
3502
3503 #ifndef PRODUCT
3504 //------------------------------verify_graph_edges---------------------------
3505 // Walk the Graph and verify that there is a one-to-one correspondence
3506 // between Use-Def edges and Def-Use edges in the graph.
3507 void Compile::verify_graph_edges(bool no_dead_code) {
3508 if (VerifyGraphEdges) {
3509 ResourceArea *area = Thread::current()->resource_area();
3510 Unique_Node_List visited(area);
3511 // Call recursive graph walk to check edges
3512 _root->verify_edges(visited);
3513 if (no_dead_code) {
3514 // Now make sure that no visited node is used by an unvisited node.
3515 bool dead_nodes = false;
3516 Unique_Node_List checked(area);
3517 while (visited.size() > 0) {
3518 Node* n = visited.pop();
3519 checked.push(n);
3520 for (uint i = 0; i < n->outcnt(); i++) {
3521 Node* use = n->raw_out(i);
3522 if (checked.member(use)) continue; // already checked
3523 if (visited.member(use)) continue; // already in the graph
3524 if (use->is_Con()) continue; // a dead ConNode is OK
3525 // At this point, we have found a dead node which is DU-reachable.
3526 if (!dead_nodes) {
3527 tty->print_cr("*** Dead nodes reachable via DU edges:");
3528 dead_nodes = true;
3529 }
3530 use->dump(2);
3531 tty->print_cr("---");
3532 checked.push(use); // No repeats; pretend it is now checked.
3533 }
3534 }
3535 assert(!dead_nodes, "using nodes must be reachable from root");
3536 }
3537 }
3538 }
3539
3540 // Verify GC barriers consistency
3541 // Currently supported:
3542 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3543 void Compile::verify_barriers() {
3544 if (UseG1GC) {
3545 // Verify G1 pre-barriers
3546 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3547
3548 ResourceArea *area = Thread::current()->resource_area();
3549 Unique_Node_List visited(area);
3550 Node_List worklist(area);
3551 // We're going to walk control flow backwards starting from the Root
3552 worklist.push(_root);
3553 while (worklist.size() > 0) {
3554 Node* x = worklist.pop();
3555 if (x == NULL || x == top()) continue;
3556 if (visited.member(x)) {
3557 continue;
3558 } else {
3559 visited.push(x);
3560 }
3561
3562 if (x->is_Region()) {
3563 for (uint i = 1; i < x->req(); i++) {
3564 worklist.push(x->in(i));
3565 }
3566 } else {
3567 worklist.push(x->in(0));
3568 // We are looking for the pattern:
3569 // /->ThreadLocal
3570 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3571 // \->ConI(0)
3572 // We want to verify that the If and the LoadB have the same control
3573 // See GraphKit::g1_write_barrier_pre()
3574 if (x->is_If()) {
3575 IfNode *iff = x->as_If();
3576 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3577 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3578 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3579 && cmp->in(1)->is_Load()) {
3580 LoadNode* load = cmp->in(1)->as_Load();
3581 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3582 && load->in(2)->in(3)->is_Con()
3583 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3584
3585 Node* if_ctrl = iff->in(0);
3586 Node* load_ctrl = load->in(0);
3587
3588 if (if_ctrl != load_ctrl) {
3589 // Skip possible CProj->NeverBranch in infinite loops
3590 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3591 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3592 if_ctrl = if_ctrl->in(0)->in(0);
3593 }
3594 }
3595 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3596 }
3597 }
3598 }
3599 }
3600 }
3601 }
3602 }
3603 }
3604
3605 #endif
3606
3607 // The Compile object keeps track of failure reasons separately from the ciEnv.
3608 // This is required because there is not quite a 1-1 relation between the
3609 // ciEnv and its compilation task and the Compile object. Note that one
3610 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3611 // to backtrack and retry without subsuming loads. Other than this backtracking
3612 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3613 // by the logic in C2Compiler.
3614 void Compile::record_failure(const char* reason) {
3615 if (log() != NULL) {
3616 log()->elem("failure reason='%s' phase='compile'", reason);
3617 }
3618 if (_failure_reason == NULL) {
3619 // Record the first failure reason.
3620 _failure_reason = reason;
3621 }
3622
3623 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3624 C->print_method(PHASE_FAILURE);
3625 }
3626 _root = NULL; // flush the graph, too
3627 }
3628
3629 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3630 : TraceTime(name, accumulator, CITime, CITimeVerbose),
3631 _phase_name(name), _dolog(CITimeVerbose)
3632 {
3633 if (_dolog) {
3634 C = Compile::current();
3635 _log = C->log();
3636 } else {
3637 C = NULL;
3638 _log = NULL;
3639 }
3640 if (_log != NULL) {
3641 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3642 _log->stamp();
3643 _log->end_head();
3644 }
3645 }
3646
3647 Compile::TracePhase::~TracePhase() {
3648
3649 C = Compile::current();
3650 if (_dolog) {
3651 _log = C->log();
3652 } else {
3653 _log = NULL;
3654 }
3655
3656 #ifdef ASSERT
3657 if (PrintIdealNodeCount) {
3658 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3659 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3660 }
3661
3662 if (VerifyIdealNodeCount) {
3663 Compile::current()->print_missing_nodes();
3664 }
3665 #endif
3666
3667 if (_log != NULL) {
3668 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3669 }
3670 }
3671
3672 //=============================================================================
3673 // Two Constant's are equal when the type and the value are equal.
3674 bool Compile::Constant::operator==(const Constant& other) {
3675 if (type() != other.type() ) return false;
3676 if (can_be_reused() != other.can_be_reused()) return false;
3677 // For floating point values we compare the bit pattern.
3678 switch (type()) {
3679 case T_FLOAT: return (_v._value.i == other._v._value.i);
3680 case T_LONG:
3681 case T_DOUBLE: return (_v._value.j == other._v._value.j);
3682 case T_OBJECT:
3683 case T_ADDRESS: return (_v._value.l == other._v._value.l);
3684 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
3685 case T_METADATA: return (_v._metadata == other._v._metadata);
3686 default: ShouldNotReachHere();
3687 }
3688 return false;
3689 }
3690
3691 static int type_to_size_in_bytes(BasicType t) {
3692 switch (t) {
3693 case T_LONG: return sizeof(jlong );
3694 case T_FLOAT: return sizeof(jfloat );
3695 case T_DOUBLE: return sizeof(jdouble);
3696 case T_METADATA: return sizeof(Metadata*);
3697 // We use T_VOID as marker for jump-table entries (labels) which
3698 // need an internal word relocation.
3699 case T_VOID:
3700 case T_ADDRESS:
3701 case T_OBJECT: return sizeof(jobject);
3702 }
3703
3704 ShouldNotReachHere();
3705 return -1;
3706 }
3707
3708 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3709 // sort descending
3710 if (a->freq() > b->freq()) return -1;
3711 if (a->freq() < b->freq()) return 1;
3712 return 0;
3713 }
3714
3715 void Compile::ConstantTable::calculate_offsets_and_size() {
3716 // First, sort the array by frequencies.
3717 _constants.sort(qsort_comparator);
3718
3719 #ifdef ASSERT
3720 // Make sure all jump-table entries were sorted to the end of the
3721 // array (they have a negative frequency).
3722 bool found_void = false;
3723 for (int i = 0; i < _constants.length(); i++) {
3724 Constant con = _constants.at(i);
3725 if (con.type() == T_VOID)
3726 found_void = true; // jump-tables
3727 else
3728 assert(!found_void, "wrong sorting");
3729 }
3730 #endif
3731
3732 int offset = 0;
3733 for (int i = 0; i < _constants.length(); i++) {
3734 Constant* con = _constants.adr_at(i);
3735
3736 // Align offset for type.
3737 int typesize = type_to_size_in_bytes(con->type());
3738 offset = align_size_up(offset, typesize);
3739 con->set_offset(offset); // set constant's offset
3740
3741 if (con->type() == T_VOID) {
3742 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3743 offset = offset + typesize * n->outcnt(); // expand jump-table
3744 } else {
3745 offset = offset + typesize;
3746 }
3747 }
3748
3749 // Align size up to the next section start (which is insts; see
3750 // CodeBuffer::align_at_start).
3751 assert(_size == -1, "already set?");
3752 _size = align_size_up(offset, CodeEntryAlignment);
3753 }
3754
3755 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3756 MacroAssembler _masm(&cb);
3757 for (int i = 0; i < _constants.length(); i++) {
3758 Constant con = _constants.at(i);
3759 address constant_addr;
3760 switch (con.type()) {
3761 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3762 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3763 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3764 case T_OBJECT: {
3765 jobject obj = con.get_jobject();
3766 int oop_index = _masm.oop_recorder()->find_index(obj);
3767 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3768 break;
3769 }
3770 case T_ADDRESS: {
3771 address addr = (address) con.get_jobject();
3772 constant_addr = _masm.address_constant(addr);
3773 break;
3774 }
3775 // We use T_VOID as marker for jump-table entries (labels) which
3776 // need an internal word relocation.
3777 case T_VOID: {
3778 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3779 // Fill the jump-table with a dummy word. The real value is
3780 // filled in later in fill_jump_table.
3781 address dummy = (address) n;
3782 constant_addr = _masm.address_constant(dummy);
3783 // Expand jump-table
3784 for (uint i = 1; i < n->outcnt(); i++) {
3785 address temp_addr = _masm.address_constant(dummy + i);
3786 assert(temp_addr, "consts section too small");
3787 }
3788 break;
3789 }
3790 case T_METADATA: {
3791 Metadata* obj = con.get_metadata();
3792 int metadata_index = _masm.oop_recorder()->find_index(obj);
3793 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3794 break;
3795 }
3796 default: ShouldNotReachHere();
3797 }
3798 assert(constant_addr, "consts section too small");
3799 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3800 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3801 }
3802 }
3803
3804 int Compile::ConstantTable::find_offset(Constant& con) const {
3805 int idx = _constants.find(con);
3806 assert(idx != -1, "constant must be in constant table");
3807 int offset = _constants.at(idx).offset();
3808 assert(offset != -1, "constant table not emitted yet?");
3809 return offset;
3810 }
3811
3812 void Compile::ConstantTable::add(Constant& con) {
3813 if (con.can_be_reused()) {
3814 int idx = _constants.find(con);
3815 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3816 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3817 return;
3818 }
3819 }
3820 (void) _constants.append(con);
3821 }
3822
3823 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3824 Block* b = Compile::current()->cfg()->get_block_for_node(n);
3825 Constant con(type, value, b->_freq);
3826 add(con);
3827 return con;
3828 }
3829
3830 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3831 Constant con(metadata);
3832 add(con);
3833 return con;
3834 }
3835
3836 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3837 jvalue value;
3838 BasicType type = oper->type()->basic_type();
3839 switch (type) {
3840 case T_LONG: value.j = oper->constantL(); break;
3841 case T_FLOAT: value.f = oper->constantF(); break;
3842 case T_DOUBLE: value.d = oper->constantD(); break;
3843 case T_OBJECT:
3844 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3845 case T_METADATA: return add((Metadata*)oper->constant()); break;
3846 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3847 }
3848 return add(n, type, value);
3849 }
3850
3851 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3852 jvalue value;
3853 // We can use the node pointer here to identify the right jump-table
3854 // as this method is called from Compile::Fill_buffer right before
3855 // the MachNodes are emitted and the jump-table is filled (means the
3856 // MachNode pointers do not change anymore).
3857 value.l = (jobject) n;
3858 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3859 add(con);
3860 return con;
3861 }
3862
3863 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3864 // If called from Compile::scratch_emit_size do nothing.
3865 if (Compile::current()->in_scratch_emit_size()) return;
3866
3867 assert(labels.is_nonempty(), "must be");
3868 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3869
3870 // Since MachConstantNode::constant_offset() also contains
3871 // table_base_offset() we need to subtract the table_base_offset()
3872 // to get the plain offset into the constant table.
3873 int offset = n->constant_offset() - table_base_offset();
3874
3875 MacroAssembler _masm(&cb);
3876 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3877
3878 for (uint i = 0; i < n->outcnt(); i++) {
3879 address* constant_addr = &jump_table_base[i];
3880 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3881 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3882 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3883 }
3884 }
3885
3886 //----------------------------static_subtype_check-----------------------------
3887 // Shortcut important common cases when superklass is exact:
3888 // (0) superklass is java.lang.Object (can occur in reflective code)
3889 // (1) subklass is already limited to a subtype of superklass => always ok
3890 // (2) subklass does not overlap with superklass => always fail
3891 // (3) superklass has NO subtypes and we can check with a simple compare.
3892 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3893 if (StressReflectiveCode) {
3894 return SSC_full_test; // Let caller generate the general case.
3895 }
3896
3897 if (superk == env()->Object_klass()) {
3898 return SSC_always_true; // (0) this test cannot fail
3899 }
3900
3901 ciType* superelem = superk;
3902 if (superelem->is_array_klass())
3903 superelem = superelem->as_array_klass()->base_element_type();
3904
3905 if (!subk->is_interface()) { // cannot trust static interface types yet
3906 if (subk->is_subtype_of(superk)) {
3907 return SSC_always_true; // (1) false path dead; no dynamic test needed
3908 }
3909 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
3910 !superk->is_subtype_of(subk)) {
3911 return SSC_always_false;
3912 }
3913 }
3914
3915 // If casting to an instance klass, it must have no subtypes
3916 if (superk->is_interface()) {
3917 // Cannot trust interfaces yet.
3918 // %%% S.B. superk->nof_implementors() == 1
3919 } else if (superelem->is_instance_klass()) {
3920 ciInstanceKlass* ik = superelem->as_instance_klass();
3921 if (!ik->has_subklass() && !ik->is_interface()) {
3922 if (!ik->is_final()) {
3923 // Add a dependency if there is a chance of a later subclass.
3924 dependencies()->assert_leaf_type(ik);
3925 }
3926 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3927 }
3928 } else {
3929 // A primitive array type has no subtypes.
3930 return SSC_easy_test; // (3) caller can do a simple ptr comparison
3931 }
3932
3933 return SSC_full_test;
3934 }
3935
3936 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) {
3937 #ifdef _LP64
3938 // The scaled index operand to AddP must be a clean 64-bit value.
3939 // Java allows a 32-bit int to be incremented to a negative
3940 // value, which appears in a 64-bit register as a large
3941 // positive number. Using that large positive number as an
3942 // operand in pointer arithmetic has bad consequences.
3943 // On the other hand, 32-bit overflow is rare, and the possibility
3944 // can often be excluded, if we annotate the ConvI2L node with
3945 // a type assertion that its value is known to be a small positive
3946 // number. (The prior range check has ensured this.)
3947 // This assertion is used by ConvI2LNode::Ideal.
3948 int index_max = max_jint - 1; // array size is max_jint, index is one less
3949 if (sizetype != NULL) index_max = sizetype->_hi - 1;
3950 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
3951 idx = phase->transform(new ConvI2LNode(idx, lidxtype));
3952 #endif
3953 return idx;
3954 }
3955
3956 // The message about the current inlining is accumulated in
3957 // _print_inlining_stream and transfered into the _print_inlining_list
3958 // once we know whether inlining succeeds or not. For regular
3959 // inlining, messages are appended to the buffer pointed by
3960 // _print_inlining_idx in the _print_inlining_list. For late inlining,
3961 // a new buffer is added after _print_inlining_idx in the list. This
3962 // way we can update the inlining message for late inlining call site
3963 // when the inlining is attempted again.
3964 void Compile::print_inlining_init() {
3965 if (print_inlining() || print_intrinsics()) {
3966 _print_inlining_stream = new stringStream();
3967 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
3968 }
3969 }
3970
3971 void Compile::print_inlining_reinit() {
3972 if (print_inlining() || print_intrinsics()) {
3973 // Re allocate buffer when we change ResourceMark
3974 _print_inlining_stream = new stringStream();
3975 }
3976 }
3977
3978 void Compile::print_inlining_reset() {
3979 _print_inlining_stream->reset();
3980 }
3981
3982 void Compile::print_inlining_commit() {
3983 assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
3984 // Transfer the message from _print_inlining_stream to the current
3985 // _print_inlining_list buffer and clear _print_inlining_stream.
3986 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
3987 print_inlining_reset();
3988 }
3989
3990 void Compile::print_inlining_push() {
3991 // Add new buffer to the _print_inlining_list at current position
3992 _print_inlining_idx++;
3993 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
3994 }
3995
3996 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
3997 return _print_inlining_list->at(_print_inlining_idx);
3998 }
3999
4000 void Compile::print_inlining_update(CallGenerator* cg) {
4001 if (print_inlining() || print_intrinsics()) {
4002 if (!cg->is_late_inline()) {
4003 if (print_inlining_current().cg() != NULL) {
4004 print_inlining_push();
4005 }
4006 print_inlining_commit();
4007 } else {
4008 if (print_inlining_current().cg() != cg &&
4009 (print_inlining_current().cg() != NULL ||
4010 print_inlining_current().ss()->size() != 0)) {
4011 print_inlining_push();
4012 }
4013 print_inlining_commit();
4014 print_inlining_current().set_cg(cg);
4015 }
4016 }
4017 }
4018
4019 void Compile::print_inlining_move_to(CallGenerator* cg) {
4020 // We resume inlining at a late inlining call site. Locate the
4021 // corresponding inlining buffer so that we can update it.
4022 if (print_inlining()) {
4023 for (int i = 0; i < _print_inlining_list->length(); i++) {
4024 if (_print_inlining_list->adr_at(i)->cg() == cg) {
4025 _print_inlining_idx = i;
4026 return;
4027 }
4028 }
4029 ShouldNotReachHere();
4030 }
4031 }
4032
4033 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4034 if (print_inlining()) {
4035 assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4036 assert(print_inlining_current().cg() == cg, "wrong entry");
4037 // replace message with new message
4038 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4039 print_inlining_commit();
4040 print_inlining_current().set_cg(cg);
4041 }
4042 }
4043
4044 void Compile::print_inlining_assert_ready() {
4045 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4046 }
4047
4048 void Compile::process_print_inlining() {
4049 bool do_print_inlining = print_inlining() || print_intrinsics();
4050 if (do_print_inlining || log() != NULL) {
4051 // Print inlining message for candidates that we couldn't inline
4052 // for lack of space
4053 for (int i = 0; i < _late_inlines.length(); i++) {
4054 CallGenerator* cg = _late_inlines.at(i);
4055 if (!cg->is_mh_late_inline()) {
4056 const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4057 if (do_print_inlining) {
4058 cg->print_inlining_late(msg);
4059 }
4060 log_late_inline_failure(cg, msg);
4061 }
4062 }
4063 }
4064 if (do_print_inlining) {
4065 ResourceMark rm;
4066 stringStream ss;
4067 for (int i = 0; i < _print_inlining_list->length(); i++) {
4068 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4069 }
4070 size_t end = ss.size();
4071 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4072 strncpy(_print_inlining_output, ss.base(), end+1);
4073 _print_inlining_output[end] = 0;
4074 }
4075 }
4076
4077 void Compile::dump_print_inlining() {
4078 if (_print_inlining_output != NULL) {
4079 tty->print_raw(_print_inlining_output);
4080 }
4081 }
4082
4083 void Compile::log_late_inline(CallGenerator* cg) {
4084 if (log() != NULL) {
4085 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4086 cg->unique_id());
4087 JVMState* p = cg->call_node()->jvms();
4088 while (p != NULL) {
4089 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4090 p = p->caller();
4091 }
4092 log()->tail("late_inline");
4093 }
4094 }
4095
4096 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4097 log_late_inline(cg);
4098 if (log() != NULL) {
4099 log()->inline_fail(msg);
4100 }
4101 }
4102
4103 void Compile::log_inline_id(CallGenerator* cg) {
4104 if (log() != NULL) {
4105 // The LogCompilation tool needs a unique way to identify late
4106 // inline call sites. This id must be unique for this call site in
4107 // this compilation. Try to have it unique across compilations as
4108 // well because it can be convenient when grepping through the log
4109 // file.
4110 // Distinguish OSR compilations from others in case CICountOSR is
4111 // on.
4112 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4113 cg->set_unique_id(id);
4114 log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4115 }
4116 }
4117
4118 void Compile::log_inline_failure(const char* msg) {
4119 if (C->log() != NULL) {
4120 C->log()->inline_fail(msg);
4121 }
4122 }
4123
4124
4125 // Dump inlining replay data to the stream.
4126 // Don't change thread state and acquire any locks.
4127 void Compile::dump_inline_data(outputStream* out) {
4128 InlineTree* inl_tree = ilt();
4129 if (inl_tree != NULL) {
4130 out->print(" inline %d", inl_tree->count());
4131 inl_tree->dump_replay_data(out);
4132 }
4133 }
4134
4135 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4136 if (n1->Opcode() < n2->Opcode()) return -1;
4137 else if (n1->Opcode() > n2->Opcode()) return 1;
4138
4139 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
4140 for (uint i = 1; i < n1->req(); i++) {
4141 if (n1->in(i) < n2->in(i)) return -1;
4142 else if (n1->in(i) > n2->in(i)) return 1;
4143 }
4144
4145 return 0;
4146 }
4147
4148 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4149 Node* n1 = *n1p;
4150 Node* n2 = *n2p;
4151
4152 return cmp_expensive_nodes(n1, n2);
4153 }
4154
4155 void Compile::sort_expensive_nodes() {
4156 if (!expensive_nodes_sorted()) {
4157 _expensive_nodes->sort(cmp_expensive_nodes);
4158 }
4159 }
4160
4161 bool Compile::expensive_nodes_sorted() const {
4162 for (int i = 1; i < _expensive_nodes->length(); i++) {
4163 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4164 return false;
4165 }
4166 }
4167 return true;
4168 }
4169
4170 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4171 if (_expensive_nodes->length() == 0) {
4172 return false;
4173 }
4174
4175 assert(OptimizeExpensiveOps, "optimization off?");
4176
4177 // Take this opportunity to remove dead nodes from the list
4178 int j = 0;
4179 for (int i = 0; i < _expensive_nodes->length(); i++) {
4180 Node* n = _expensive_nodes->at(i);
4181 if (!n->is_unreachable(igvn)) {
4182 assert(n->is_expensive(), "should be expensive");
4183 _expensive_nodes->at_put(j, n);
4184 j++;
4185 }
4186 }
4187 _expensive_nodes->trunc_to(j);
4188
4189 // Then sort the list so that similar nodes are next to each other
4190 // and check for at least two nodes of identical kind with same data
4191 // inputs.
4192 sort_expensive_nodes();
4193
4194 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4195 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4196 return true;
4197 }
4198 }
4199
4200 return false;
4201 }
4202
4203 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4204 if (_expensive_nodes->length() == 0) {
4205 return;
4206 }
4207
4208 assert(OptimizeExpensiveOps, "optimization off?");
4209
4210 // Sort to bring similar nodes next to each other and clear the
4211 // control input of nodes for which there's only a single copy.
4212 sort_expensive_nodes();
4213
4214 int j = 0;
4215 int identical = 0;
4216 int i = 0;
4217 bool modified = false;
4218 for (; i < _expensive_nodes->length()-1; i++) {
4219 assert(j <= i, "can't write beyond current index");
4220 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4221 identical++;
4222 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4223 continue;
4224 }
4225 if (identical > 0) {
4226 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4227 identical = 0;
4228 } else {
4229 Node* n = _expensive_nodes->at(i);
4230 igvn.replace_input_of(n, 0, NULL);
4231 igvn.hash_insert(n);
4232 modified = true;
4233 }
4234 }
4235 if (identical > 0) {
4236 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4237 } else if (_expensive_nodes->length() >= 1) {
4238 Node* n = _expensive_nodes->at(i);
4239 igvn.replace_input_of(n, 0, NULL);
4240 igvn.hash_insert(n);
4241 modified = true;
4242 }
4243 _expensive_nodes->trunc_to(j);
4244 if (modified) {
4245 igvn.optimize();
4246 }
4247 }
4248
4249 void Compile::add_expensive_node(Node * n) {
4250 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4251 assert(n->is_expensive(), "expensive nodes with non-null control here only");
4252 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4253 if (OptimizeExpensiveOps) {
4254 _expensive_nodes->append(n);
4255 } else {
4256 // Clear control input and let IGVN optimize expensive nodes if
4257 // OptimizeExpensiveOps is off.
4258 n->set_req(0, NULL);
4259 }
4260 }
4261
4262 /**
4263 * Remove the speculative part of types and clean up the graph
4264 */
4265 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4266 if (UseTypeSpeculation) {
4267 Unique_Node_List worklist;
4268 worklist.push(root());
4269 int modified = 0;
4270 // Go over all type nodes that carry a speculative type, drop the
4271 // speculative part of the type and enqueue the node for an igvn
4272 // which may optimize it out.
4273 for (uint next = 0; next < worklist.size(); ++next) {
4274 Node *n = worklist.at(next);
4275 if (n->is_Type()) {
4276 TypeNode* tn = n->as_Type();
4277 const Type* t = tn->type();
4278 const Type* t_no_spec = t->remove_speculative();
4279 if (t_no_spec != t) {
4280 bool in_hash = igvn.hash_delete(n);
4281 assert(in_hash, "node should be in igvn hash table");
4282 tn->set_type(t_no_spec);
4283 igvn.hash_insert(n);
4284 igvn._worklist.push(n); // give it a chance to go away
4285 modified++;
4286 }
4287 }
4288 uint max = n->len();
4289 for( uint i = 0; i < max; ++i ) {
4290 Node *m = n->in(i);
4291 if (not_a_node(m)) continue;
4292 worklist.push(m);
4293 }
4294 }
4295 // Drop the speculative part of all types in the igvn's type table
4296 igvn.remove_speculative_types();
4297 if (modified > 0) {
4298 igvn.optimize();
4299 }
4300 #ifdef ASSERT
4301 // Verify that after the IGVN is over no speculative type has resurfaced
4302 worklist.clear();
4303 worklist.push(root());
4304 for (uint next = 0; next < worklist.size(); ++next) {
4305 Node *n = worklist.at(next);
4306 const Type* t = igvn.type_or_null(n);
4307 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4308 if (n->is_Type()) {
4309 t = n->as_Type()->type();
4310 assert(t == t->remove_speculative(), "no more speculative types");
4311 }
4312 uint max = n->len();
4313 for( uint i = 0; i < max; ++i ) {
4314 Node *m = n->in(i);
4315 if (not_a_node(m)) continue;
4316 worklist.push(m);
4317 }
4318 }
4319 igvn.check_no_speculative_types();
4320 #endif
4321 }
4322 }
4323
4324 // Auxiliary method to support randomized stressing/fuzzing.
4325 //
4326 // This method can be called the arbitrary number of times, with current count
4327 // as the argument. The logic allows selecting a single candidate from the
4328 // running list of candidates as follows:
4329 // int count = 0;
4330 // Cand* selected = null;
4331 // while(cand = cand->next()) {
4332 // if (randomized_select(++count)) {
4333 // selected = cand;
4334 // }
4335 // }
4336 //
4337 // Including count equalizes the chances any candidate is "selected".
4338 // This is useful when we don't have the complete list of candidates to choose
4339 // from uniformly. In this case, we need to adjust the randomicity of the
4340 // selection, or else we will end up biasing the selection towards the latter
4341 // candidates.
4342 //
4343 // Quick back-envelope calculation shows that for the list of n candidates
4344 // the equal probability for the candidate to persist as "best" can be
4345 // achieved by replacing it with "next" k-th candidate with the probability
4346 // of 1/k. It can be easily shown that by the end of the run, the
4347 // probability for any candidate is converged to 1/n, thus giving the
4348 // uniform distribution among all the candidates.
4349 //
4350 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4351 #define RANDOMIZED_DOMAIN_POW 29
4352 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4353 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4354 bool Compile::randomized_select(int count) {
4355 assert(count > 0, "only positive");
4356 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4357 }
4358
4359 CloneMap& Compile::clone_map() { return _clone_map; }
4360 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; }
4361
4362 void NodeCloneInfo::dump() const {
4363 tty->print(" {%d:%d} ", idx(), gen());
4364 }
4365
4366 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4367 uint64_t val = value(old->_idx);
4368 NodeCloneInfo cio(val);
4369 assert(val != 0, "old node should be in the map");
4370 NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4371 insert(nnn->_idx, cin.get());
4372 #ifndef PRODUCT
4373 if (is_debug()) {
4374 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4375 }
4376 #endif
4377 }
4378
4379 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4380 NodeCloneInfo cio(value(old->_idx));
4381 if (cio.get() == 0) {
4382 cio.set(old->_idx, 0);
4383 insert(old->_idx, cio.get());
4384 #ifndef PRODUCT
4385 if (is_debug()) {
4386 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4387 }
4388 #endif
4389 }
4390 clone(old, nnn, gen);
4391 }
4392
4393 int CloneMap::max_gen() const {
4394 int g = 0;
4395 DictI di(_dict);
4396 for(; di.test(); ++di) {
4397 int t = gen(di._key);
4398 if (g < t) {
4399 g = t;
4400 #ifndef PRODUCT
4401 if (is_debug()) {
4402 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4403 }
4404 #endif
4405 }
4406 }
4407 return g;
4408 }
4409
4410 void CloneMap::dump(node_idx_t key) const {
4411 uint64_t val = value(key);
4412 if (val != 0) {
4413 NodeCloneInfo ni(val);
4414 ni.dump();
4415 }
4416 }