1 /*
2 * Copyright (c) 1997, 2014, 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/assembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/exceptionHandlerTable.hpp"
29 #include "code/nmethod.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/block.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/callGenerator.hpp"
36 #include "opto/callnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/chaitin.hpp"
39 #include "opto/compile.hpp"
40 #include "opto/connode.hpp"
41 #include "opto/divnode.hpp"
42 #include "opto/escape.hpp"
43 #include "opto/idealGraphPrinter.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/macro.hpp"
47 #include "opto/matcher.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/node.hpp"
51 #include "opto/opcodes.hpp"
52 #include "opto/output.hpp"
53 #include "opto/parse.hpp"
54 #include "opto/phaseX.hpp"
55 #include "opto/rootnode.hpp"
56 #include "opto/runtime.hpp"
57 #include "opto/stringopts.hpp"
58 #include "opto/type.hpp"
59 #include "opto/vectornode.hpp"
60 #include "runtime/arguments.hpp"
61 #include "runtime/signature.hpp"
62 #include "runtime/stubRoutines.hpp"
63 #include "runtime/timer.hpp"
64 #include "trace/tracing.hpp"
65 #include "utilities/copy.hpp"
66 #ifdef TARGET_ARCH_MODEL_x86_32
67 # include "adfiles/ad_x86_32.hpp"
68 #endif
69 #ifdef TARGET_ARCH_MODEL_x86_64
70 # include "adfiles/ad_x86_64.hpp"
71 #endif
72 #ifdef TARGET_ARCH_MODEL_sparc
73 # include "adfiles/ad_sparc.hpp"
74 #endif
75 #ifdef TARGET_ARCH_MODEL_zero
76 # include "adfiles/ad_zero.hpp"
77 #endif
78 #ifdef TARGET_ARCH_MODEL_arm
79 # include "adfiles/ad_arm.hpp"
80 #endif
81 #ifdef TARGET_ARCH_MODEL_ppc_32
82 # include "adfiles/ad_ppc_32.hpp"
83 #endif
84 #ifdef TARGET_ARCH_MODEL_ppc_64
85 # include "adfiles/ad_ppc_64.hpp"
86 #endif
87 #ifdef TARGET_ARCH_x86
88 # include "compile_x86.hpp"
89 #endif
90 #ifdef TARGET_ARCH_sparc
91 # include "compile_sparc.hpp"
92 #endif
93 #ifdef TARGET_ARCH_zero
94 # include "compile_zero.hpp"
95 #endif
96 #ifdef TARGET_ARCH_ppc
97 # include "compile_ppc.hpp"
98 #endif
99
100
101 // -------------------- Compile::mach_constant_base_node -----------------------
102 // Constant table base node singleton.
103 MachConstantBaseNode* Compile::mach_constant_base_node() {
104 if (_mach_constant_base_node == NULL) {
105 _mach_constant_base_node = new (C) MachConstantBaseNode();
106 _mach_constant_base_node->add_req(C->root());
107 }
108 return _mach_constant_base_node;
109 }
110
111
112 /// Support for intrinsics.
113
114 // Return the index at which m must be inserted (or already exists).
115 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
116 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
117 #ifdef ASSERT
118 for (int i = 1; i < _intrinsics->length(); i++) {
119 CallGenerator* cg1 = _intrinsics->at(i-1);
120 CallGenerator* cg2 = _intrinsics->at(i);
121 assert(cg1->method() != cg2->method()
122 ? cg1->method() < cg2->method()
123 : cg1->is_virtual() < cg2->is_virtual(),
124 "compiler intrinsics list must stay sorted");
125 }
126 #endif
127 // Binary search sorted list, in decreasing intervals [lo, hi].
128 int lo = 0, hi = _intrinsics->length()-1;
129 while (lo <= hi) {
130 int mid = (uint)(hi + lo) / 2;
131 ciMethod* mid_m = _intrinsics->at(mid)->method();
132 if (m < mid_m) {
133 hi = mid-1;
134 } else if (m > mid_m) {
135 lo = mid+1;
136 } else {
137 // look at minor sort key
138 bool mid_virt = _intrinsics->at(mid)->is_virtual();
139 if (is_virtual < mid_virt) {
140 hi = mid-1;
141 } else if (is_virtual > mid_virt) {
142 lo = mid+1;
143 } else {
144 return mid; // exact match
145 }
146 }
147 }
148 return lo; // inexact match
149 }
150
151 void Compile::register_intrinsic(CallGenerator* cg) {
152 if (_intrinsics == NULL) {
153 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
154 }
155 // This code is stolen from ciObjectFactory::insert.
156 // Really, GrowableArray should have methods for
157 // insert_at, remove_at, and binary_search.
158 int len = _intrinsics->length();
159 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
160 if (index == len) {
161 _intrinsics->append(cg);
162 } else {
163 #ifdef ASSERT
164 CallGenerator* oldcg = _intrinsics->at(index);
165 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
166 #endif
167 _intrinsics->append(_intrinsics->at(len-1));
168 int pos;
169 for (pos = len-2; pos >= index; pos--) {
170 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
171 }
172 _intrinsics->at_put(index, cg);
173 }
174 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
175 }
176
177 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
178 assert(m->is_loaded(), "don't try this on unloaded methods");
179 if (_intrinsics != NULL) {
180 int index = intrinsic_insertion_index(m, is_virtual);
181 if (index < _intrinsics->length()
182 && _intrinsics->at(index)->method() == m
183 && _intrinsics->at(index)->is_virtual() == is_virtual) {
184 return _intrinsics->at(index);
185 }
186 }
187 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
188 if (m->intrinsic_id() != vmIntrinsics::_none &&
189 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
190 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
191 if (cg != NULL) {
192 // Save it for next time:
193 register_intrinsic(cg);
194 return cg;
195 } else {
196 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
197 }
198 }
199 return NULL;
200 }
201
202 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
203 // in library_call.cpp.
204
205
206 #ifndef PRODUCT
207 // statistics gathering...
208
209 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
210 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
211
212 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
213 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
214 int oflags = _intrinsic_hist_flags[id];
215 assert(flags != 0, "what happened?");
216 if (is_virtual) {
217 flags |= _intrinsic_virtual;
218 }
219 bool changed = (flags != oflags);
220 if ((flags & _intrinsic_worked) != 0) {
221 juint count = (_intrinsic_hist_count[id] += 1);
222 if (count == 1) {
223 changed = true; // first time
224 }
225 // increment the overall count also:
226 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
227 }
228 if (changed) {
229 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
230 // Something changed about the intrinsic's virtuality.
231 if ((flags & _intrinsic_virtual) != 0) {
232 // This is the first use of this intrinsic as a virtual call.
233 if (oflags != 0) {
234 // We already saw it as a non-virtual, so note both cases.
235 flags |= _intrinsic_both;
236 }
237 } else if ((oflags & _intrinsic_both) == 0) {
238 // This is the first use of this intrinsic as a non-virtual
239 flags |= _intrinsic_both;
240 }
241 }
242 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
243 }
244 // update the overall flags also:
245 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
246 return changed;
247 }
248
249 static char* format_flags(int flags, char* buf) {
250 buf[0] = 0;
251 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
252 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
253 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
254 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
255 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
256 if (buf[0] == 0) strcat(buf, ",");
257 assert(buf[0] == ',', "must be");
258 return &buf[1];
259 }
260
261 void Compile::print_intrinsic_statistics() {
262 char flagsbuf[100];
263 ttyLocker ttyl;
264 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
265 tty->print_cr("Compiler intrinsic usage:");
266 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
267 if (total == 0) total = 1; // avoid div0 in case of no successes
268 #define PRINT_STAT_LINE(name, c, f) \
269 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
270 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
271 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
272 int flags = _intrinsic_hist_flags[id];
273 juint count = _intrinsic_hist_count[id];
274 if ((flags | count) != 0) {
275 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
276 }
277 }
278 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
279 if (xtty != NULL) xtty->tail("statistics");
280 }
281
282 void Compile::print_statistics() {
283 { ttyLocker ttyl;
284 if (xtty != NULL) xtty->head("statistics type='opto'");
285 Parse::print_statistics();
286 PhaseCCP::print_statistics();
287 PhaseRegAlloc::print_statistics();
288 Scheduling::print_statistics();
289 PhasePeephole::print_statistics();
290 PhaseIdealLoop::print_statistics();
291 if (xtty != NULL) xtty->tail("statistics");
292 }
293 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
294 // put this under its own <statistics> element.
295 print_intrinsic_statistics();
296 }
297 }
298 #endif //PRODUCT
299
300 // Support for bundling info
301 Bundle* Compile::node_bundling(const Node *n) {
302 assert(valid_bundle_info(n), "oob");
303 return &_node_bundling_base[n->_idx];
304 }
305
306 bool Compile::valid_bundle_info(const Node *n) {
307 return (_node_bundling_limit > n->_idx);
308 }
309
310
311 void Compile::gvn_replace_by(Node* n, Node* nn) {
312 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
313 Node* use = n->last_out(i);
314 bool is_in_table = initial_gvn()->hash_delete(use);
315 uint uses_found = 0;
316 for (uint j = 0; j < use->len(); j++) {
317 if (use->in(j) == n) {
318 if (j < use->req())
319 use->set_req(j, nn);
320 else
321 use->set_prec(j, nn);
322 uses_found++;
323 }
324 }
325 if (is_in_table) {
326 // reinsert into table
327 initial_gvn()->hash_find_insert(use);
328 }
329 record_for_igvn(use);
330 i -= uses_found; // we deleted 1 or more copies of this edge
331 }
332 }
333
334
335 static inline bool not_a_node(const Node* n) {
336 if (n == NULL) return true;
337 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
338 if (*(address*)n == badAddress) return true; // kill by Node::destruct
339 return false;
340 }
341
342 // Identify all nodes that are reachable from below, useful.
343 // Use breadth-first pass that records state in a Unique_Node_List,
344 // recursive traversal is slower.
345 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
346 int estimated_worklist_size = unique();
347 useful.map( estimated_worklist_size, NULL ); // preallocate space
348
349 // Initialize worklist
350 if (root() != NULL) { useful.push(root()); }
351 // If 'top' is cached, declare it useful to preserve cached node
352 if( cached_top_node() ) { useful.push(cached_top_node()); }
353
354 // Push all useful nodes onto the list, breadthfirst
355 for( uint next = 0; next < useful.size(); ++next ) {
356 assert( next < unique(), "Unique useful nodes < total nodes");
357 Node *n = useful.at(next);
358 uint max = n->len();
359 for( uint i = 0; i < max; ++i ) {
360 Node *m = n->in(i);
361 if (not_a_node(m)) continue;
362 useful.push(m);
363 }
364 }
365 }
366
367 // Update dead_node_list with any missing dead nodes using useful
368 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
369 void Compile::update_dead_node_list(Unique_Node_List &useful) {
370 uint max_idx = unique();
371 VectorSet& useful_node_set = useful.member_set();
372
373 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
374 // If node with index node_idx is not in useful set,
375 // mark it as dead in dead node list.
376 if (! useful_node_set.test(node_idx) ) {
377 record_dead_node(node_idx);
378 }
379 }
380 }
381
382 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
383 int shift = 0;
384 for (int i = 0; i < inlines->length(); i++) {
385 CallGenerator* cg = inlines->at(i);
386 CallNode* call = cg->call_node();
387 if (shift > 0) {
388 inlines->at_put(i-shift, cg);
389 }
390 if (!useful.member(call)) {
391 shift++;
392 }
393 }
394 inlines->trunc_to(inlines->length()-shift);
395 }
396
397 // Disconnect all useless nodes by disconnecting those at the boundary.
398 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
399 uint next = 0;
400 while (next < useful.size()) {
401 Node *n = useful.at(next++);
402 if (n->is_SafePoint()) {
403 // We're done with a parsing phase. Replaced nodes are not valid
404 // beyond that point.
405 n->as_SafePoint()->delete_replaced_nodes();
406 }
407 // Use raw traversal of out edges since this code removes out edges
408 int max = n->outcnt();
409 for (int j = 0; j < max; ++j) {
410 Node* child = n->raw_out(j);
411 if (! useful.member(child)) {
412 assert(!child->is_top() || child != top(),
413 "If top is cached in Compile object it is in useful list");
414 // Only need to remove this out-edge to the useless node
415 n->raw_del_out(j);
416 --j;
417 --max;
418 }
419 }
420 if (n->outcnt() == 1 && n->has_special_unique_user()) {
421 record_for_igvn(n->unique_out());
422 }
423 }
424 // Remove useless macro and predicate opaq nodes
425 for (int i = C->macro_count()-1; i >= 0; i--) {
426 Node* n = C->macro_node(i);
427 if (!useful.member(n)) {
428 remove_macro_node(n);
429 }
430 }
431 // Remove useless expensive node
432 for (int i = C->expensive_count()-1; i >= 0; i--) {
433 Node* n = C->expensive_node(i);
434 if (!useful.member(n)) {
435 remove_expensive_node(n);
436 }
437 }
438 // clean up the late inline lists
439 remove_useless_late_inlines(&_string_late_inlines, useful);
440 remove_useless_late_inlines(&_late_inlines, useful);
441 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
442 }
443
444 //------------------------------frame_size_in_words-----------------------------
445 // frame_slots in units of words
446 int Compile::frame_size_in_words() const {
447 // shift is 0 in LP32 and 1 in LP64
448 const int shift = (LogBytesPerWord - LogBytesPerInt);
449 int words = _frame_slots >> shift;
450 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
451 return words;
452 }
453
454 // ============================================================================
455 //------------------------------CompileWrapper---------------------------------
456 class CompileWrapper : public StackObj {
457 Compile *const _compile;
458 public:
459 CompileWrapper(Compile* compile);
460
461 ~CompileWrapper();
462 };
463
464 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
465 // the Compile* pointer is stored in the current ciEnv:
466 ciEnv* env = compile->env();
467 assert(env == ciEnv::current(), "must already be a ciEnv active");
468 assert(env->compiler_data() == NULL, "compile already active?");
469 env->set_compiler_data(compile);
470 assert(compile == Compile::current(), "sanity");
471
472 compile->set_type_dict(NULL);
473 compile->set_type_hwm(NULL);
474 compile->set_type_last_size(0);
475 compile->set_last_tf(NULL, NULL);
476 compile->set_indexSet_arena(NULL);
477 compile->set_indexSet_free_block_list(NULL);
478 compile->init_type_arena();
479 Type::Initialize(compile);
480 _compile->set_scratch_buffer_blob(NULL);
481 _compile->begin_method();
482 }
483 CompileWrapper::~CompileWrapper() {
484 _compile->end_method();
485 if (_compile->scratch_buffer_blob() != NULL)
486 BufferBlob::free(_compile->scratch_buffer_blob());
487 _compile->env()->set_compiler_data(NULL);
488 }
489
490
491 //----------------------------print_compile_messages---------------------------
492 void Compile::print_compile_messages() {
493 #ifndef PRODUCT
494 // Check if recompiling
495 if (_subsume_loads == false && PrintOpto) {
496 // Recompiling without allowing machine instructions to subsume loads
497 tty->print_cr("*********************************************************");
498 tty->print_cr("** Bailout: Recompile without subsuming loads **");
499 tty->print_cr("*********************************************************");
500 }
501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
502 // Recompiling without escape analysis
503 tty->print_cr("*********************************************************");
504 tty->print_cr("** Bailout: Recompile without escape analysis **");
505 tty->print_cr("*********************************************************");
506 }
507 if (env()->break_at_compile()) {
508 // Open the debugger when compiling this method.
509 tty->print("### Breaking when compiling: ");
510 method()->print_short_name();
511 tty->cr();
512 BREAKPOINT;
513 }
514
515 if( PrintOpto ) {
516 if (is_osr_compilation()) {
517 tty->print("[OSR]%3d", _compile_id);
518 } else {
519 tty->print("%3d", _compile_id);
520 }
521 }
522 #endif
523 }
524
525
526 //-----------------------init_scratch_buffer_blob------------------------------
527 // Construct a temporary BufferBlob and cache it for this compile.
528 void Compile::init_scratch_buffer_blob(int const_size) {
529 // If there is already a scratch buffer blob allocated and the
530 // constant section is big enough, use it. Otherwise free the
531 // current and allocate a new one.
532 BufferBlob* blob = scratch_buffer_blob();
533 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
534 // Use the current blob.
535 } else {
536 if (blob != NULL) {
537 BufferBlob::free(blob);
538 }
539
540 ResourceMark rm;
541 _scratch_const_size = const_size;
542 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
543 blob = BufferBlob::create("Compile::scratch_buffer", size);
544 // Record the buffer blob for next time.
545 set_scratch_buffer_blob(blob);
546 // Have we run out of code space?
547 if (scratch_buffer_blob() == NULL) {
548 // Let CompilerBroker disable further compilations.
549 record_failure("Not enough space for scratch buffer in CodeCache");
550 return;
551 }
552 }
553
554 // Initialize the relocation buffers
555 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
556 set_scratch_locs_memory(locs_buf);
557 }
558
559
560 //-----------------------scratch_emit_size-------------------------------------
561 // Helper function that computes size by emitting code
562 uint Compile::scratch_emit_size(const Node* n) {
563 // Start scratch_emit_size section.
564 set_in_scratch_emit_size(true);
565
566 // Emit into a trash buffer and count bytes emitted.
567 // This is a pretty expensive way to compute a size,
568 // but it works well enough if seldom used.
569 // All common fixed-size instructions are given a size
570 // method by the AD file.
571 // Note that the scratch buffer blob and locs memory are
572 // allocated at the beginning of the compile task, and
573 // may be shared by several calls to scratch_emit_size.
574 // The allocation of the scratch buffer blob is particularly
575 // expensive, since it has to grab the code cache lock.
576 BufferBlob* blob = this->scratch_buffer_blob();
577 assert(blob != NULL, "Initialize BufferBlob at start");
578 assert(blob->size() > MAX_inst_size, "sanity");
579 relocInfo* locs_buf = scratch_locs_memory();
580 address blob_begin = blob->content_begin();
581 address blob_end = (address)locs_buf;
582 assert(blob->content_contains(blob_end), "sanity");
583 CodeBuffer buf(blob_begin, blob_end - blob_begin);
584 buf.initialize_consts_size(_scratch_const_size);
585 buf.initialize_stubs_size(MAX_stubs_size);
586 assert(locs_buf != NULL, "sanity");
587 int lsize = MAX_locs_size / 3;
588 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
589 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
590 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
591 buf.consts()->set_scratch_emit();
592 buf.insts() ->set_scratch_emit();
593 buf.stubs() ->set_scratch_emit();
594
595 // Do the emission.
596
597 Label fakeL; // Fake label for branch instructions.
598 Label* saveL = NULL;
599 uint save_bnum = 0;
600 bool is_branch = n->is_MachBranch();
601 if (is_branch) {
602 MacroAssembler masm(&buf);
603 masm.bind(fakeL);
604 n->as_MachBranch()->save_label(&saveL, &save_bnum);
605 n->as_MachBranch()->label_set(&fakeL, 0);
606 }
607 n->emit(buf, this->regalloc());
608 if (is_branch) // Restore label.
609 n->as_MachBranch()->label_set(saveL, save_bnum);
610
611 // End scratch_emit_size section.
612 set_in_scratch_emit_size(false);
613
614 return buf.insts_size();
615 }
616
617
618 // ============================================================================
619 //------------------------------Compile standard-------------------------------
620 debug_only( int Compile::_debug_idx = 100000; )
621
622 // Compile a method. entry_bci is -1 for normal compilations and indicates
623 // the continuation bci for on stack replacement.
624
625
626 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
627 : Phase(Compiler),
628 _env(ci_env),
629 _log(ci_env->log()),
630 _compile_id(ci_env->compile_id()),
631 _save_argument_registers(false),
632 _stub_name(NULL),
633 _stub_function(NULL),
634 _stub_entry_point(NULL),
635 _method(target),
636 _entry_bci(osr_bci),
637 _initial_gvn(NULL),
638 _for_igvn(NULL),
639 _warm_calls(NULL),
640 _subsume_loads(subsume_loads),
641 _do_escape_analysis(do_escape_analysis),
642 _failure_reason(NULL),
643 _code_buffer("Compile::Fill_buffer"),
644 _orig_pc_slot(0),
645 _orig_pc_slot_offset_in_bytes(0),
646 _has_method_handle_invokes(false),
647 _mach_constant_base_node(NULL),
648 _node_bundling_limit(0),
649 _node_bundling_base(NULL),
650 _java_calls(0),
651 _inner_loops(0),
652 _scratch_const_size(-1),
653 _in_scratch_emit_size(false),
654 _dead_node_list(comp_arena()),
655 _dead_node_count(0),
656 #ifndef PRODUCT
657 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
658 _in_dump_cnt(0),
659 _printer(IdealGraphPrinter::printer()),
660 #endif
661 _congraph(NULL),
662 _late_inlines(comp_arena(), 2, 0, NULL),
663 _string_late_inlines(comp_arena(), 2, 0, NULL),
664 _late_inlines_pos(0),
665 _number_of_mh_late_inlines(0),
666 _inlining_progress(false),
667 _inlining_incrementally(false),
668 _print_inlining_list(NULL),
669 _print_inlining_idx(0) {
670 C = this;
671
672 CompileWrapper cw(this);
673 #ifndef PRODUCT
674 if (TimeCompiler2) {
675 tty->print(" ");
676 target->holder()->name()->print();
677 tty->print(".");
678 target->print_short_name();
679 tty->print(" ");
680 }
681 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
682 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
683 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
684 if (!print_opto_assembly) {
685 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
686 if (print_assembly && !Disassembler::can_decode()) {
687 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
688 print_opto_assembly = true;
689 }
690 }
691 set_print_assembly(print_opto_assembly);
692 set_parsed_irreducible_loop(false);
693 #endif
694 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
695 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
696 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
697
698 if (ProfileTraps) {
699 // Make sure the method being compiled gets its own MDO,
700 // so we can at least track the decompile_count().
701 method()->ensure_method_data();
702 }
703
704 Init(::AliasLevel);
705
706
707 print_compile_messages();
708
709 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
710 _ilt = InlineTree::build_inline_tree_root();
711 else
712 _ilt = NULL;
713
714 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
715 assert(num_alias_types() >= AliasIdxRaw, "");
716
717 #define MINIMUM_NODE_HASH 1023
718 // Node list that Iterative GVN will start with
719 Unique_Node_List for_igvn(comp_arena());
720 set_for_igvn(&for_igvn);
721
722 // GVN that will be run immediately on new nodes
723 uint estimated_size = method()->code_size()*4+64;
724 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
725 PhaseGVN gvn(node_arena(), estimated_size);
726 set_initial_gvn(&gvn);
727
728 if (print_inlining() || print_intrinsics()) {
729 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
730 }
731 { // Scope for timing the parser
732 TracePhase t3("parse", &_t_parser, true);
733
734 // Put top into the hash table ASAP.
735 initial_gvn()->transform_no_reclaim(top());
736
737 // Set up tf(), start(), and find a CallGenerator.
738 CallGenerator* cg = NULL;
739 if (is_osr_compilation()) {
740 const TypeTuple *domain = StartOSRNode::osr_domain();
741 const TypeTuple *range = TypeTuple::make_range(method()->signature());
742 init_tf(TypeFunc::make(domain, range));
743 StartNode* s = new (this) StartOSRNode(root(), domain);
744 initial_gvn()->set_type_bottom(s);
745 init_start(s);
746 cg = CallGenerator::for_osr(method(), entry_bci());
747 } else {
748 // Normal case.
749 init_tf(TypeFunc::make(method()));
750 StartNode* s = new (this) StartNode(root(), tf()->domain());
751 initial_gvn()->set_type_bottom(s);
752 init_start(s);
753 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
754 // With java.lang.ref.reference.get() we must go through the
755 // intrinsic when G1 is enabled - even when get() is the root
756 // method of the compile - so that, if necessary, the value in
757 // the referent field of the reference object gets recorded by
758 // the pre-barrier code.
759 // Specifically, if G1 is enabled, the value in the referent
760 // field is recorded by the G1 SATB pre barrier. This will
761 // result in the referent being marked live and the reference
762 // object removed from the list of discovered references during
763 // reference processing.
764 cg = find_intrinsic(method(), false);
765 }
766 if (cg == NULL) {
767 float past_uses = method()->interpreter_invocation_count();
768 float expected_uses = past_uses;
769 cg = CallGenerator::for_inline(method(), expected_uses);
770 }
771 }
772 if (failing()) return;
773 if (cg == NULL) {
774 record_method_not_compilable_all_tiers("cannot parse method");
775 return;
776 }
777 JVMState* jvms = build_start_state(start(), tf());
778 if ((jvms = cg->generate(jvms)) == NULL) {
779 record_method_not_compilable("method parse failed");
780 return;
781 }
782 GraphKit kit(jvms);
783
784 if (!kit.stopped()) {
785 // Accept return values, and transfer control we know not where.
786 // This is done by a special, unique ReturnNode bound to root.
787 return_values(kit.jvms());
788 }
789
790 if (kit.has_exceptions()) {
791 // Any exceptions that escape from this call must be rethrown
792 // to whatever caller is dynamically above us on the stack.
793 // This is done by a special, unique RethrowNode bound to root.
794 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
795 }
796
797 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
798
799 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
800 inline_string_calls(true);
801 }
802
803 if (failing()) return;
804
805 print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
806
807 // Remove clutter produced by parsing.
808 if (!failing()) {
809 ResourceMark rm;
810 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
811 }
812 }
813
814 // Note: Large methods are capped off in do_one_bytecode().
815 if (failing()) return;
816
817 // After parsing, node notes are no longer automagic.
818 // They must be propagated by register_new_node_with_optimizer(),
819 // clone(), or the like.
820 set_default_node_notes(NULL);
821
822 for (;;) {
823 int successes = Inline_Warm();
824 if (failing()) return;
825 if (successes == 0) break;
826 }
827
828 // Drain the list.
829 Finish_Warm();
830 #ifndef PRODUCT
831 if (_printer) {
832 _printer->print_inlining(this);
833 }
834 #endif
835
836 if (failing()) return;
837 NOT_PRODUCT( verify_graph_edges(); )
838
839 // Now optimize
840 Optimize();
841 if (failing()) return;
842 NOT_PRODUCT( verify_graph_edges(); )
843
844 #ifndef PRODUCT
845 if (PrintIdeal) {
846 ttyLocker ttyl; // keep the following output all in one block
847 // This output goes directly to the tty, not the compiler log.
848 // To enable tools to match it up with the compilation activity,
849 // be sure to tag this tty output with the compile ID.
850 if (xtty != NULL) {
851 xtty->head("ideal compile_id='%d'%s", compile_id(),
852 is_osr_compilation() ? " compile_kind='osr'" :
853 "");
854 }
855 root()->dump(9999);
856 if (xtty != NULL) {
857 xtty->tail("ideal");
858 }
859 }
860 #endif
861
862 NOT_PRODUCT( verify_barriers(); )
863 // Now that we know the size of all the monitors we can add a fixed slot
864 // for the original deopt pc.
865
866 _orig_pc_slot = fixed_slots();
867 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
868 set_fixed_slots(next_slot);
869
870 // Now generate code
871 Code_Gen();
872 if (failing()) return;
873
874 // Check if we want to skip execution of all compiled code.
875 {
876 #ifndef PRODUCT
877 if (OptoNoExecute) {
878 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
879 return;
880 }
881 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
882 #endif
883
884 if (is_osr_compilation()) {
885 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
886 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
887 } else {
888 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
889 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
890 }
891
892 env()->register_method(_method, _entry_bci,
893 &_code_offsets,
894 _orig_pc_slot_offset_in_bytes,
895 code_buffer(),
896 frame_size_in_words(), _oop_map_set,
897 &_handler_table, &_inc_table,
898 compiler,
899 env()->comp_level(),
900 has_unsafe_access(),
901 SharedRuntime::is_wide_vector(max_vector_size())
902 );
903
904 if (log() != NULL) // Print code cache state into compiler log
905 log()->code_cache_state();
906 }
907 }
908
909 //------------------------------Compile----------------------------------------
910 // Compile a runtime stub
911 Compile::Compile( ciEnv* ci_env,
912 TypeFunc_generator generator,
913 address stub_function,
914 const char *stub_name,
915 int is_fancy_jump,
916 bool pass_tls,
917 bool save_arg_registers,
918 bool return_pc )
919 : Phase(Compiler),
920 _env(ci_env),
921 _log(ci_env->log()),
922 _compile_id(0),
923 _save_argument_registers(save_arg_registers),
924 _method(NULL),
925 _stub_name(stub_name),
926 _stub_function(stub_function),
927 _stub_entry_point(NULL),
928 _entry_bci(InvocationEntryBci),
929 _initial_gvn(NULL),
930 _for_igvn(NULL),
931 _warm_calls(NULL),
932 _orig_pc_slot(0),
933 _orig_pc_slot_offset_in_bytes(0),
934 _subsume_loads(true),
935 _do_escape_analysis(false),
936 _failure_reason(NULL),
937 _code_buffer("Compile::Fill_buffer"),
938 _has_method_handle_invokes(false),
939 _mach_constant_base_node(NULL),
940 _node_bundling_limit(0),
941 _node_bundling_base(NULL),
942 _java_calls(0),
943 _inner_loops(0),
944 #ifndef PRODUCT
945 _trace_opto_output(TraceOptoOutput),
946 _in_dump_cnt(0),
947 _printer(NULL),
948 #endif
949 _dead_node_list(comp_arena()),
950 _dead_node_count(0),
951 _congraph(NULL),
952 _number_of_mh_late_inlines(0),
953 _inlining_progress(false),
954 _inlining_incrementally(false),
955 _print_inlining_list(NULL),
956 _print_inlining_idx(0) {
957 C = this;
958
959 #ifndef PRODUCT
960 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
961 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
962 set_print_assembly(PrintFrameConverterAssembly);
963 set_parsed_irreducible_loop(false);
964 #endif
965 set_has_irreducible_loop(false); // no loops
966
967 CompileWrapper cw(this);
968 Init(/*AliasLevel=*/ 0);
969 init_tf((*generator)());
970
971 {
972 // The following is a dummy for the sake of GraphKit::gen_stub
973 Unique_Node_List for_igvn(comp_arena());
974 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
975 PhaseGVN gvn(Thread::current()->resource_area(),255);
976 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
977 gvn.transform_no_reclaim(top());
978
979 GraphKit kit;
980 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
981 }
982
983 NOT_PRODUCT( verify_graph_edges(); )
984 Code_Gen();
985 if (failing()) return;
986
987
988 // Entry point will be accessed using compile->stub_entry_point();
989 if (code_buffer() == NULL) {
990 Matcher::soft_match_failure();
991 } else {
992 if (PrintAssembly && (WizardMode || Verbose))
993 tty->print_cr("### Stub::%s", stub_name);
994
995 if (!failing()) {
996 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
997
998 // Make the NMethod
999 // For now we mark the frame as never safe for profile stackwalking
1000 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1001 code_buffer(),
1002 CodeOffsets::frame_never_safe,
1003 // _code_offsets.value(CodeOffsets::Frame_Complete),
1004 frame_size_in_words(),
1005 _oop_map_set,
1006 save_arg_registers);
1007 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1008
1009 _stub_entry_point = rs->entry_point();
1010 }
1011 }
1012 }
1013
1014 #ifndef PRODUCT
1015 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
1016 if(PrintOpto && Verbose) {
1017 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
1018 }
1019 }
1020 #endif
1021
1022 void Compile::print_codes() {
1023 }
1024
1025 //------------------------------Init-------------------------------------------
1026 // Prepare for a single compilation
1027 void Compile::Init(int aliaslevel) {
1028 _unique = 0;
1029 _regalloc = NULL;
1030
1031 _tf = NULL; // filled in later
1032 _top = NULL; // cached later
1033 _matcher = NULL; // filled in later
1034 _cfg = NULL; // filled in later
1035
1036 set_24_bit_selection_and_mode(Use24BitFP, false);
1037
1038 _node_note_array = NULL;
1039 _default_node_notes = NULL;
1040
1041 _immutable_memory = NULL; // filled in at first inquiry
1042
1043 // Globally visible Nodes
1044 // First set TOP to NULL to give safe behavior during creation of RootNode
1045 set_cached_top_node(NULL);
1046 set_root(new (this) RootNode());
1047 // Now that you have a Root to point to, create the real TOP
1048 set_cached_top_node( new (this) ConNode(Type::TOP) );
1049 set_recent_alloc(NULL, NULL);
1050
1051 // Create Debug Information Recorder to record scopes, oopmaps, etc.
1052 env()->set_oop_recorder(new OopRecorder(comp_arena()));
1053 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1054 env()->set_dependencies(new Dependencies(env()));
1055
1056 _fixed_slots = 0;
1057 set_has_split_ifs(false);
1058 set_has_loops(has_method() && method()->has_loops()); // first approximation
1059 set_has_stringbuilder(false);
1060 _trap_can_recompile = false; // no traps emitted yet
1061 _major_progress = true; // start out assuming good things will happen
1062 set_has_unsafe_access(false);
1063 set_max_vector_size(0);
1064 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1065 set_decompile_count(0);
1066
1067 // Compute when not to trap. Used by matching trap based nodes and
1068 // NullCheck optimization.
1069 _allowed_reasons = 0;
1070 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
1071 assert(rs < BitsPerInt, "recode bit map");
1072 if (!C->too_many_traps((Deoptimization::DeoptReason) rs))
1073 _allowed_reasons |= nth_bit(rs);
1074 }
1075
1076 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1077 set_num_loop_opts(LoopOptsCount);
1078 set_do_inlining(Inline);
1079 set_max_inline_size(MaxInlineSize);
1080 set_freq_inline_size(FreqInlineSize);
1081 set_do_scheduling(OptoScheduling);
1082 set_do_count_invocations(false);
1083 set_do_method_data_update(false);
1084
1085 if (debug_info()->recording_non_safepoints()) {
1086 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1087 (comp_arena(), 8, 0, NULL));
1088 set_default_node_notes(Node_Notes::make(this));
1089 }
1090
1091 // // -- Initialize types before each compile --
1092 // // Update cached type information
1093 // if( _method && _method->constants() )
1094 // Type::update_loaded_types(_method, _method->constants());
1095
1096 // Init alias_type map.
1097 if (!_do_escape_analysis && aliaslevel == 3)
1098 aliaslevel = 2; // No unique types without escape analysis
1099 _AliasLevel = aliaslevel;
1100 const int grow_ats = 16;
1101 _max_alias_types = grow_ats;
1102 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1103 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
1104 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1105 {
1106 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
1107 }
1108 // Initialize the first few types.
1109 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1110 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1111 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1112 _num_alias_types = AliasIdxRaw+1;
1113 // Zero out the alias type cache.
1114 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1115 // A NULL adr_type hits in the cache right away. Preload the right answer.
1116 probe_alias_cache(NULL)->_index = AliasIdxTop;
1117
1118 _intrinsics = NULL;
1119 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1120 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1121 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1122 register_library_intrinsics();
1123 }
1124
1125 //---------------------------init_start----------------------------------------
1126 // Install the StartNode on this compile object.
1127 void Compile::init_start(StartNode* s) {
1128 if (failing())
1129 return; // already failing
1130 assert(s == start(), "");
1131 }
1132
1133 StartNode* Compile::start() const {
1134 assert(!failing(), "");
1135 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1136 Node* start = root()->fast_out(i);
1137 if( start->is_Start() )
1138 return start->as_Start();
1139 }
1140 fatal("Did not find Start node!");
1141 return NULL;
1142 }
1143
1144 //-------------------------------immutable_memory-------------------------------------
1145 // Access immutable memory
1146 Node* Compile::immutable_memory() {
1147 if (_immutable_memory != NULL) {
1148 return _immutable_memory;
1149 }
1150 StartNode* s = start();
1151 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1152 Node *p = s->fast_out(i);
1153 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1154 _immutable_memory = p;
1155 return _immutable_memory;
1156 }
1157 }
1158 ShouldNotReachHere();
1159 return NULL;
1160 }
1161
1162 //----------------------set_cached_top_node------------------------------------
1163 // Install the cached top node, and make sure Node::is_top works correctly.
1164 void Compile::set_cached_top_node(Node* tn) {
1165 if (tn != NULL) verify_top(tn);
1166 Node* old_top = _top;
1167 _top = tn;
1168 // Calling Node::setup_is_top allows the nodes the chance to adjust
1169 // their _out arrays.
1170 if (_top != NULL) _top->setup_is_top();
1171 if (old_top != NULL) old_top->setup_is_top();
1172 assert(_top == NULL || top()->is_top(), "");
1173 }
1174
1175 #ifdef ASSERT
1176 uint Compile::count_live_nodes_by_graph_walk() {
1177 Unique_Node_List useful(comp_arena());
1178 // Get useful node list by walking the graph.
1179 identify_useful_nodes(useful);
1180 return useful.size();
1181 }
1182
1183 void Compile::print_missing_nodes() {
1184
1185 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1186 if ((_log == NULL) && (! PrintIdealNodeCount)) {
1187 return;
1188 }
1189
1190 // This is an expensive function. It is executed only when the user
1191 // specifies VerifyIdealNodeCount option or otherwise knows the
1192 // additional work that needs to be done to identify reachable nodes
1193 // by walking the flow graph and find the missing ones using
1194 // _dead_node_list.
1195
1196 Unique_Node_List useful(comp_arena());
1197 // Get useful node list by walking the graph.
1198 identify_useful_nodes(useful);
1199
1200 uint l_nodes = C->live_nodes();
1201 uint l_nodes_by_walk = useful.size();
1202
1203 if (l_nodes != l_nodes_by_walk) {
1204 if (_log != NULL) {
1205 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1206 _log->stamp();
1207 _log->end_head();
1208 }
1209 VectorSet& useful_member_set = useful.member_set();
1210 int last_idx = l_nodes_by_walk;
1211 for (int i = 0; i < last_idx; i++) {
1212 if (useful_member_set.test(i)) {
1213 if (_dead_node_list.test(i)) {
1214 if (_log != NULL) {
1215 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1216 }
1217 if (PrintIdealNodeCount) {
1218 // Print the log message to tty
1219 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1220 useful.at(i)->dump();
1221 }
1222 }
1223 }
1224 else if (! _dead_node_list.test(i)) {
1225 if (_log != NULL) {
1226 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1227 }
1228 if (PrintIdealNodeCount) {
1229 // Print the log message to tty
1230 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1231 }
1232 }
1233 }
1234 if (_log != NULL) {
1235 _log->tail("mismatched_nodes");
1236 }
1237 }
1238 }
1239 #endif
1240
1241 #ifndef PRODUCT
1242 void Compile::verify_top(Node* tn) const {
1243 if (tn != NULL) {
1244 assert(tn->is_Con(), "top node must be a constant");
1245 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1246 assert(tn->in(0) != NULL, "must have live top node");
1247 }
1248 }
1249 #endif
1250
1251
1252 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1253
1254 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1255 guarantee(arr != NULL, "");
1256 int num_blocks = arr->length();
1257 if (grow_by < num_blocks) grow_by = num_blocks;
1258 int num_notes = grow_by * _node_notes_block_size;
1259 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1260 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1261 while (num_notes > 0) {
1262 arr->append(notes);
1263 notes += _node_notes_block_size;
1264 num_notes -= _node_notes_block_size;
1265 }
1266 assert(num_notes == 0, "exact multiple, please");
1267 }
1268
1269 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1270 if (source == NULL || dest == NULL) return false;
1271
1272 if (dest->is_Con())
1273 return false; // Do not push debug info onto constants.
1274
1275 #ifdef ASSERT
1276 // Leave a bread crumb trail pointing to the original node:
1277 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1278 dest->set_debug_orig(source);
1279 }
1280 #endif
1281
1282 if (node_note_array() == NULL)
1283 return false; // Not collecting any notes now.
1284
1285 // This is a copy onto a pre-existing node, which may already have notes.
1286 // If both nodes have notes, do not overwrite any pre-existing notes.
1287 Node_Notes* source_notes = node_notes_at(source->_idx);
1288 if (source_notes == NULL || source_notes->is_clear()) return false;
1289 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1290 if (dest_notes == NULL || dest_notes->is_clear()) {
1291 return set_node_notes_at(dest->_idx, source_notes);
1292 }
1293
1294 Node_Notes merged_notes = (*source_notes);
1295 // The order of operations here ensures that dest notes will win...
1296 merged_notes.update_from(dest_notes);
1297 return set_node_notes_at(dest->_idx, &merged_notes);
1298 }
1299
1300
1301 //--------------------------allow_range_check_smearing-------------------------
1302 // Gating condition for coalescing similar range checks.
1303 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1304 // single covering check that is at least as strong as any of them.
1305 // If the optimization succeeds, the simplified (strengthened) range check
1306 // will always succeed. If it fails, we will deopt, and then give up
1307 // on the optimization.
1308 bool Compile::allow_range_check_smearing() const {
1309 // If this method has already thrown a range-check,
1310 // assume it was because we already tried range smearing
1311 // and it failed.
1312 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1313 return !already_trapped;
1314 }
1315
1316
1317 //------------------------------flatten_alias_type-----------------------------
1318 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1319 int offset = tj->offset();
1320 TypePtr::PTR ptr = tj->ptr();
1321
1322 // Known instance (scalarizable allocation) alias only with itself.
1323 bool is_known_inst = tj->isa_oopptr() != NULL &&
1324 tj->is_oopptr()->is_known_instance();
1325
1326 // Process weird unsafe references.
1327 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1328 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1329 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1330 tj = TypeOopPtr::BOTTOM;
1331 ptr = tj->ptr();
1332 offset = tj->offset();
1333 }
1334
1335 // Array pointers need some flattening
1336 const TypeAryPtr *ta = tj->isa_aryptr();
1337 if( ta && is_known_inst ) {
1338 if ( offset != Type::OffsetBot &&
1339 offset > arrayOopDesc::length_offset_in_bytes() ) {
1340 offset = Type::OffsetBot; // Flatten constant access into array body only
1341 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1342 }
1343 } else if( ta && _AliasLevel >= 2 ) {
1344 // For arrays indexed by constant indices, we flatten the alias
1345 // space to include all of the array body. Only the header, klass
1346 // and array length can be accessed un-aliased.
1347 if( offset != Type::OffsetBot ) {
1348 if( ta->const_oop() ) { // methodDataOop or methodOop
1349 offset = Type::OffsetBot; // Flatten constant access into array body
1350 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1351 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1352 // range is OK as-is.
1353 tj = ta = TypeAryPtr::RANGE;
1354 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1355 tj = TypeInstPtr::KLASS; // all klass loads look alike
1356 ta = TypeAryPtr::RANGE; // generic ignored junk
1357 ptr = TypePtr::BotPTR;
1358 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1359 tj = TypeInstPtr::MARK;
1360 ta = TypeAryPtr::RANGE; // generic ignored junk
1361 ptr = TypePtr::BotPTR;
1362 } else { // Random constant offset into array body
1363 offset = Type::OffsetBot; // Flatten constant access into array body
1364 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1365 }
1366 }
1367 // Arrays of fixed size alias with arrays of unknown size.
1368 if (ta->size() != TypeInt::POS) {
1369 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1370 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1371 }
1372 // Arrays of known objects become arrays of unknown objects.
1373 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1374 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1375 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1376 }
1377 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1378 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1379 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1380 }
1381 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1382 // cannot be distinguished by bytecode alone.
1383 if (ta->elem() == TypeInt::BOOL) {
1384 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1385 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1386 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1387 }
1388 // During the 2nd round of IterGVN, NotNull castings are removed.
1389 // Make sure the Bottom and NotNull variants alias the same.
1390 // Also, make sure exact and non-exact variants alias the same.
1391 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1392 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1393 }
1394 }
1395
1396 // Oop pointers need some flattening
1397 const TypeInstPtr *to = tj->isa_instptr();
1398 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1399 ciInstanceKlass *k = to->klass()->as_instance_klass();
1400 if( ptr == TypePtr::Constant ) {
1401 if (to->klass() != ciEnv::current()->Class_klass() ||
1402 offset < k->size_helper() * wordSize) {
1403 // No constant oop pointers (such as Strings); they alias with
1404 // unknown strings.
1405 assert(!is_known_inst, "not scalarizable allocation");
1406 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1407 }
1408 } else if( is_known_inst ) {
1409 tj = to; // Keep NotNull and klass_is_exact for instance type
1410 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1411 // During the 2nd round of IterGVN, NotNull castings are removed.
1412 // Make sure the Bottom and NotNull variants alias the same.
1413 // Also, make sure exact and non-exact variants alias the same.
1414 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1415 }
1416 // Canonicalize the holder of this field
1417 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1418 // First handle header references such as a LoadKlassNode, even if the
1419 // object's klass is unloaded at compile time (4965979).
1420 if (!is_known_inst) { // Do it only for non-instance types
1421 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1422 }
1423 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1424 // Static fields are in the space above the normal instance
1425 // fields in the java.lang.Class instance.
1426 if (to->klass() != ciEnv::current()->Class_klass()) {
1427 to = NULL;
1428 tj = TypeOopPtr::BOTTOM;
1429 offset = tj->offset();
1430 }
1431 } else {
1432 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1433 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1434 if( is_known_inst ) {
1435 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1436 } else {
1437 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1438 }
1439 }
1440 }
1441 }
1442
1443 // Klass pointers to object array klasses need some flattening
1444 const TypeKlassPtr *tk = tj->isa_klassptr();
1445 if( tk ) {
1446 // If we are referencing a field within a Klass, we need
1447 // to assume the worst case of an Object. Both exact and
1448 // inexact types must flatten to the same alias class so
1449 // use NotNull as the PTR.
1450 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1451
1452 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1453 TypeKlassPtr::OBJECT->klass(),
1454 offset);
1455 }
1456
1457 ciKlass* klass = tk->klass();
1458 if( klass->is_obj_array_klass() ) {
1459 ciKlass* k = TypeAryPtr::OOPS->klass();
1460 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1461 k = TypeInstPtr::BOTTOM->klass();
1462 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1463 }
1464
1465 // Check for precise loads from the primary supertype array and force them
1466 // to the supertype cache alias index. Check for generic array loads from
1467 // the primary supertype array and also force them to the supertype cache
1468 // alias index. Since the same load can reach both, we need to merge
1469 // these 2 disparate memories into the same alias class. Since the
1470 // primary supertype array is read-only, there's no chance of confusion
1471 // where we bypass an array load and an array store.
1472 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1473 if (offset == Type::OffsetBot ||
1474 (offset >= primary_supers_offset &&
1475 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1476 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1477 offset = in_bytes(Klass::secondary_super_cache_offset());
1478 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1479 }
1480 }
1481
1482 // Flatten all Raw pointers together.
1483 if (tj->base() == Type::RawPtr)
1484 tj = TypeRawPtr::BOTTOM;
1485
1486 if (tj->base() == Type::AnyPtr)
1487 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1488
1489 // Flatten all to bottom for now
1490 switch( _AliasLevel ) {
1491 case 0:
1492 tj = TypePtr::BOTTOM;
1493 break;
1494 case 1: // Flatten to: oop, static, field or array
1495 switch (tj->base()) {
1496 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1497 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1498 case Type::AryPtr: // do not distinguish arrays at all
1499 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1500 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1501 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1502 default: ShouldNotReachHere();
1503 }
1504 break;
1505 case 2: // No collapsing at level 2; keep all splits
1506 case 3: // No collapsing at level 3; keep all splits
1507 break;
1508 default:
1509 Unimplemented();
1510 }
1511
1512 offset = tj->offset();
1513 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1514
1515 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1516 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1517 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1518 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1519 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1520 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1521 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1522 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1523 assert( tj->ptr() != TypePtr::TopPTR &&
1524 tj->ptr() != TypePtr::AnyNull &&
1525 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1526 // assert( tj->ptr() != TypePtr::Constant ||
1527 // tj->base() == Type::RawPtr ||
1528 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1529
1530 return tj;
1531 }
1532
1533 void Compile::AliasType::Init(int i, const TypePtr* at) {
1534 _index = i;
1535 _adr_type = at;
1536 _field = NULL;
1537 _is_rewritable = true; // default
1538 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1539 if (atoop != NULL && atoop->is_known_instance()) {
1540 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1541 _general_index = Compile::current()->get_alias_index(gt);
1542 } else {
1543 _general_index = 0;
1544 }
1545 }
1546
1547 //---------------------------------print_on------------------------------------
1548 #ifndef PRODUCT
1549 void Compile::AliasType::print_on(outputStream* st) {
1550 if (index() < 10)
1551 st->print("@ <%d> ", index());
1552 else st->print("@ <%d>", index());
1553 st->print(is_rewritable() ? " " : " RO");
1554 int offset = adr_type()->offset();
1555 if (offset == Type::OffsetBot)
1556 st->print(" +any");
1557 else st->print(" +%-3d", offset);
1558 st->print(" in ");
1559 adr_type()->dump_on(st);
1560 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1561 if (field() != NULL && tjp) {
1562 if (tjp->klass() != field()->holder() ||
1563 tjp->offset() != field()->offset_in_bytes()) {
1564 st->print(" != ");
1565 field()->print();
1566 st->print(" ***");
1567 }
1568 }
1569 }
1570
1571 void print_alias_types() {
1572 Compile* C = Compile::current();
1573 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1574 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1575 C->alias_type(idx)->print_on(tty);
1576 tty->cr();
1577 }
1578 }
1579 #endif
1580
1581
1582 //----------------------------probe_alias_cache--------------------------------
1583 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1584 intptr_t key = (intptr_t) adr_type;
1585 key ^= key >> logAliasCacheSize;
1586 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1587 }
1588
1589
1590 //-----------------------------grow_alias_types--------------------------------
1591 void Compile::grow_alias_types() {
1592 const int old_ats = _max_alias_types; // how many before?
1593 const int new_ats = old_ats; // how many more?
1594 const int grow_ats = old_ats+new_ats; // how many now?
1595 _max_alias_types = grow_ats;
1596 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1597 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1598 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1599 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1600 }
1601
1602
1603 //--------------------------------find_alias_type------------------------------
1604 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1605 if (_AliasLevel == 0)
1606 return alias_type(AliasIdxBot);
1607
1608 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1609 if (ace->_adr_type == adr_type) {
1610 return alias_type(ace->_index);
1611 }
1612
1613 // Handle special cases.
1614 if (adr_type == NULL) return alias_type(AliasIdxTop);
1615 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1616
1617 // Do it the slow way.
1618 const TypePtr* flat = flatten_alias_type(adr_type);
1619
1620 #ifdef ASSERT
1621 assert(flat == flatten_alias_type(flat), "idempotent");
1622 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1623 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1624 const TypeOopPtr* foop = flat->is_oopptr();
1625 // Scalarizable allocations have exact klass always.
1626 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1627 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1628 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1629 }
1630 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1631 #endif
1632
1633 int idx = AliasIdxTop;
1634 for (int i = 0; i < num_alias_types(); i++) {
1635 if (alias_type(i)->adr_type() == flat) {
1636 idx = i;
1637 break;
1638 }
1639 }
1640
1641 if (idx == AliasIdxTop) {
1642 if (no_create) return NULL;
1643 // Grow the array if necessary.
1644 if (_num_alias_types == _max_alias_types) grow_alias_types();
1645 // Add a new alias type.
1646 idx = _num_alias_types++;
1647 _alias_types[idx]->Init(idx, flat);
1648 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1649 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1650 if (flat->isa_instptr()) {
1651 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1652 && flat->is_instptr()->klass() == env()->Class_klass())
1653 alias_type(idx)->set_rewritable(false);
1654 }
1655 if (flat->isa_klassptr()) {
1656 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1657 alias_type(idx)->set_rewritable(false);
1658 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1659 alias_type(idx)->set_rewritable(false);
1660 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1661 alias_type(idx)->set_rewritable(false);
1662 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1663 alias_type(idx)->set_rewritable(false);
1664 }
1665 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1666 // but the base pointer type is not distinctive enough to identify
1667 // references into JavaThread.)
1668
1669 // Check for final fields.
1670 const TypeInstPtr* tinst = flat->isa_instptr();
1671 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1672 ciField* field;
1673 if (tinst->const_oop() != NULL &&
1674 tinst->klass() == ciEnv::current()->Class_klass() &&
1675 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1676 // static field
1677 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1678 field = k->get_field_by_offset(tinst->offset(), true);
1679 } else {
1680 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1681 field = k->get_field_by_offset(tinst->offset(), false);
1682 }
1683 assert(field == NULL ||
1684 original_field == NULL ||
1685 (field->holder() == original_field->holder() &&
1686 field->offset() == original_field->offset() &&
1687 field->is_static() == original_field->is_static()), "wrong field?");
1688 // Set field() and is_rewritable() attributes.
1689 if (field != NULL) alias_type(idx)->set_field(field);
1690 }
1691 }
1692
1693 // Fill the cache for next time.
1694 ace->_adr_type = adr_type;
1695 ace->_index = idx;
1696 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1697
1698 // Might as well try to fill the cache for the flattened version, too.
1699 AliasCacheEntry* face = probe_alias_cache(flat);
1700 if (face->_adr_type == NULL) {
1701 face->_adr_type = flat;
1702 face->_index = idx;
1703 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1704 }
1705
1706 return alias_type(idx);
1707 }
1708
1709
1710 Compile::AliasType* Compile::alias_type(ciField* field) {
1711 const TypeOopPtr* t;
1712 if (field->is_static())
1713 t = TypeInstPtr::make(field->holder()->java_mirror());
1714 else
1715 t = TypeOopPtr::make_from_klass_raw(field->holder());
1716 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1717 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1718 return atp;
1719 }
1720
1721
1722 //------------------------------have_alias_type--------------------------------
1723 bool Compile::have_alias_type(const TypePtr* adr_type) {
1724 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1725 if (ace->_adr_type == adr_type) {
1726 return true;
1727 }
1728
1729 // Handle special cases.
1730 if (adr_type == NULL) return true;
1731 if (adr_type == TypePtr::BOTTOM) return true;
1732
1733 return find_alias_type(adr_type, true, NULL) != NULL;
1734 }
1735
1736 //-----------------------------must_alias--------------------------------------
1737 // True if all values of the given address type are in the given alias category.
1738 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1739 if (alias_idx == AliasIdxBot) return true; // the universal category
1740 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1741 if (alias_idx == AliasIdxTop) return false; // the empty category
1742 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1743
1744 // the only remaining possible overlap is identity
1745 int adr_idx = get_alias_index(adr_type);
1746 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1747 assert(adr_idx == alias_idx ||
1748 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1749 && adr_type != TypeOopPtr::BOTTOM),
1750 "should not be testing for overlap with an unsafe pointer");
1751 return adr_idx == alias_idx;
1752 }
1753
1754 //------------------------------can_alias--------------------------------------
1755 // True if any values of the given address type are in the given alias category.
1756 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1757 if (alias_idx == AliasIdxTop) return false; // the empty category
1758 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1759 if (alias_idx == AliasIdxBot) return true; // the universal category
1760 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1761
1762 // the only remaining possible overlap is identity
1763 int adr_idx = get_alias_index(adr_type);
1764 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1765 return adr_idx == alias_idx;
1766 }
1767
1768
1769
1770 //---------------------------pop_warm_call-------------------------------------
1771 WarmCallInfo* Compile::pop_warm_call() {
1772 WarmCallInfo* wci = _warm_calls;
1773 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1774 return wci;
1775 }
1776
1777 //----------------------------Inline_Warm--------------------------------------
1778 int Compile::Inline_Warm() {
1779 // If there is room, try to inline some more warm call sites.
1780 // %%% Do a graph index compaction pass when we think we're out of space?
1781 if (!InlineWarmCalls) return 0;
1782
1783 int calls_made_hot = 0;
1784 int room_to_grow = NodeCountInliningCutoff - unique();
1785 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1786 int amount_grown = 0;
1787 WarmCallInfo* call;
1788 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1789 int est_size = (int)call->size();
1790 if (est_size > (room_to_grow - amount_grown)) {
1791 // This one won't fit anyway. Get rid of it.
1792 call->make_cold();
1793 continue;
1794 }
1795 call->make_hot();
1796 calls_made_hot++;
1797 amount_grown += est_size;
1798 amount_to_grow -= est_size;
1799 }
1800
1801 if (calls_made_hot > 0) set_major_progress();
1802 return calls_made_hot;
1803 }
1804
1805
1806 //----------------------------Finish_Warm--------------------------------------
1807 void Compile::Finish_Warm() {
1808 if (!InlineWarmCalls) return;
1809 if (failing()) return;
1810 if (warm_calls() == NULL) return;
1811
1812 // Clean up loose ends, if we are out of space for inlining.
1813 WarmCallInfo* call;
1814 while ((call = pop_warm_call()) != NULL) {
1815 call->make_cold();
1816 }
1817 }
1818
1819 //---------------------cleanup_loop_predicates-----------------------
1820 // Remove the opaque nodes that protect the predicates so that all unused
1821 // checks and uncommon_traps will be eliminated from the ideal graph
1822 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1823 if (predicate_count()==0) return;
1824 for (int i = predicate_count(); i > 0; i--) {
1825 Node * n = predicate_opaque1_node(i-1);
1826 assert(n->Opcode() == Op_Opaque1, "must be");
1827 igvn.replace_node(n, n->in(1));
1828 }
1829 assert(predicate_count()==0, "should be clean!");
1830 }
1831
1832 // StringOpts and late inlining of string methods
1833 void Compile::inline_string_calls(bool parse_time) {
1834 {
1835 // remove useless nodes to make the usage analysis simpler
1836 ResourceMark rm;
1837 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1838 }
1839
1840 {
1841 ResourceMark rm;
1842 print_method(PHASE_BEFORE_STRINGOPTS, 3);
1843 PhaseStringOpts pso(initial_gvn(), for_igvn());
1844 print_method(PHASE_AFTER_STRINGOPTS, 3);
1845 }
1846
1847 // now inline anything that we skipped the first time around
1848 if (!parse_time) {
1849 _late_inlines_pos = _late_inlines.length();
1850 }
1851
1852 while (_string_late_inlines.length() > 0) {
1853 CallGenerator* cg = _string_late_inlines.pop();
1854 cg->do_late_inline();
1855 if (failing()) return;
1856 }
1857 _string_late_inlines.trunc_to(0);
1858 }
1859
1860 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1861 assert(IncrementalInline, "incremental inlining should be on");
1862 PhaseGVN* gvn = initial_gvn();
1863
1864 set_inlining_progress(false);
1865 for_igvn()->clear();
1866 gvn->replace_with(&igvn);
1867
1868 int i = 0;
1869
1870 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1871 CallGenerator* cg = _late_inlines.at(i);
1872 _late_inlines_pos = i+1;
1873 cg->do_late_inline();
1874 if (failing()) return;
1875 }
1876 int j = 0;
1877 for (; i < _late_inlines.length(); i++, j++) {
1878 _late_inlines.at_put(j, _late_inlines.at(i));
1879 }
1880 _late_inlines.trunc_to(j);
1881
1882 {
1883 ResourceMark rm;
1884 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn());
1885 }
1886
1887 igvn = PhaseIterGVN(gvn);
1888 }
1889
1890 // Perform incremental inlining until bound on number of live nodes is reached
1891 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1892 PhaseGVN* gvn = initial_gvn();
1893
1894 set_inlining_incrementally(true);
1895 set_inlining_progress(true);
1896 uint low_live_nodes = 0;
1897
1898 while(inlining_progress() && _late_inlines.length() > 0) {
1899
1900 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1901 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1902 // PhaseIdealLoop is expensive so we only try it once we are
1903 // out of live nodes and we only try it again if the previous
1904 // helped got the number of nodes down significantly
1905 PhaseIdealLoop ideal_loop( igvn, false, true );
1906 if (failing()) return;
1907 low_live_nodes = live_nodes();
1908 _major_progress = true;
1909 }
1910
1911 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1912 break;
1913 }
1914 }
1915
1916 inline_incrementally_one(igvn);
1917
1918 if (failing()) return;
1919
1920 igvn.optimize();
1921
1922 if (failing()) return;
1923 }
1924
1925 assert( igvn._worklist.size() == 0, "should be done with igvn" );
1926
1927 if (_string_late_inlines.length() > 0) {
1928 assert(has_stringbuilder(), "inconsistent");
1929 for_igvn()->clear();
1930 initial_gvn()->replace_with(&igvn);
1931
1932 inline_string_calls(false);
1933
1934 if (failing()) return;
1935
1936 {
1937 ResourceMark rm;
1938 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1939 }
1940
1941 igvn = PhaseIterGVN(gvn);
1942
1943 igvn.optimize();
1944 }
1945
1946 set_inlining_incrementally(false);
1947 }
1948
1949
1950 //------------------------------Optimize---------------------------------------
1951 // Given a graph, optimize it.
1952 void Compile::Optimize() {
1953 TracePhase t1("optimizer", &_t_optimizer, true);
1954
1955 #ifndef PRODUCT
1956 if (env()->break_at_compile()) {
1957 BREAKPOINT;
1958 }
1959
1960 #endif
1961
1962 ResourceMark rm;
1963 int loop_opts_cnt;
1964
1965 NOT_PRODUCT( verify_graph_edges(); )
1966
1967 print_method(PHASE_AFTER_PARSING);
1968
1969 {
1970 // Iterative Global Value Numbering, including ideal transforms
1971 // Initialize IterGVN with types and values from parse-time GVN
1972 PhaseIterGVN igvn(initial_gvn());
1973 {
1974 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1975 igvn.optimize();
1976 }
1977
1978 print_method(PHASE_ITER_GVN1, 2);
1979
1980 if (failing()) return;
1981
1982 inline_incrementally(igvn);
1983
1984 print_method(PHASE_INCREMENTAL_INLINE, 2);
1985
1986 if (failing()) return;
1987
1988 // No more new expensive nodes will be added to the list from here
1989 // so keep only the actual candidates for optimizations.
1990 cleanup_expensive_nodes(igvn);
1991
1992 // Perform escape analysis
1993 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1994 if (has_loops()) {
1995 // Cleanup graph (remove dead nodes).
1996 TracePhase t2("idealLoop", &_t_idealLoop, true);
1997 PhaseIdealLoop ideal_loop( igvn, false, true );
1998 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
1999 if (failing()) return;
2000 }
2001 ConnectionGraph::do_analysis(this, &igvn);
2002
2003 if (failing()) return;
2004
2005 // Optimize out fields loads from scalar replaceable allocations.
2006 igvn.optimize();
2007 print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2008
2009 if (failing()) return;
2010
2011 if (congraph() != NULL && macro_count() > 0) {
2012 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2013 PhaseMacroExpand mexp(igvn);
2014 mexp.eliminate_macro_nodes();
2015 igvn.set_delay_transform(false);
2016
2017 igvn.optimize();
2018 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2019
2020 if (failing()) return;
2021 }
2022 }
2023
2024 // Loop transforms on the ideal graph. Range Check Elimination,
2025 // peeling, unrolling, etc.
2026
2027 // Set loop opts counter
2028 loop_opts_cnt = num_loop_opts();
2029 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2030 {
2031 TracePhase t2("idealLoop", &_t_idealLoop, true);
2032 PhaseIdealLoop ideal_loop( igvn, true );
2033 loop_opts_cnt--;
2034 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2035 if (failing()) return;
2036 }
2037 // Loop opts pass if partial peeling occurred in previous pass
2038 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2039 TracePhase t3("idealLoop", &_t_idealLoop, true);
2040 PhaseIdealLoop ideal_loop( igvn, false );
2041 loop_opts_cnt--;
2042 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2043 if (failing()) return;
2044 }
2045 // Loop opts pass for loop-unrolling before CCP
2046 if(major_progress() && (loop_opts_cnt > 0)) {
2047 TracePhase t4("idealLoop", &_t_idealLoop, true);
2048 PhaseIdealLoop ideal_loop( igvn, false );
2049 loop_opts_cnt--;
2050 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2051 }
2052 if (!failing()) {
2053 // Verify that last round of loop opts produced a valid graph
2054 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2055 PhaseIdealLoop::verify(igvn);
2056 }
2057 }
2058 if (failing()) return;
2059
2060 // Conditional Constant Propagation;
2061 PhaseCCP ccp( &igvn );
2062 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2063 {
2064 TracePhase t2("ccp", &_t_ccp, true);
2065 ccp.do_transform();
2066 }
2067 print_method(PHASE_CPP1, 2);
2068
2069 assert( true, "Break here to ccp.dump_old2new_map()");
2070
2071 // Iterative Global Value Numbering, including ideal transforms
2072 {
2073 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2074 igvn = ccp;
2075 igvn.optimize();
2076 }
2077
2078 print_method(PHASE_ITER_GVN2, 2);
2079
2080 if (failing()) return;
2081
2082 // Loop transforms on the ideal graph. Range Check Elimination,
2083 // peeling, unrolling, etc.
2084 if(loop_opts_cnt > 0) {
2085 debug_only( int cnt = 0; );
2086 while(major_progress() && (loop_opts_cnt > 0)) {
2087 TracePhase t2("idealLoop", &_t_idealLoop, true);
2088 assert( cnt++ < 40, "infinite cycle in loop optimization" );
2089 PhaseIdealLoop ideal_loop( igvn, true);
2090 loop_opts_cnt--;
2091 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2092 if (failing()) return;
2093 }
2094 }
2095
2096 {
2097 // Verify that all previous optimizations produced a valid graph
2098 // at least to this point, even if no loop optimizations were done.
2099 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2100 PhaseIdealLoop::verify(igvn);
2101 }
2102
2103 {
2104 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2105 PhaseMacroExpand mex(igvn);
2106 if (mex.expand_macro_nodes()) {
2107 assert(failing(), "must bail out w/ explicit message");
2108 return;
2109 }
2110 }
2111
2112 } // (End scope of igvn; run destructor if necessary for asserts.)
2113
2114 dump_inlining();
2115 // A method with only infinite loops has no edges entering loops from root
2116 {
2117 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2118 if (final_graph_reshaping()) {
2119 assert(failing(), "must bail out w/ explicit message");
2120 return;
2121 }
2122 }
2123
2124 print_method(PHASE_OPTIMIZE_FINISHED, 2);
2125 }
2126
2127
2128 //------------------------------Code_Gen---------------------------------------
2129 // Given a graph, generate code for it
2130 void Compile::Code_Gen() {
2131 if (failing()) return;
2132
2133 // Perform instruction selection. You might think we could reclaim Matcher
2134 // memory PDQ, but actually the Matcher is used in generating spill code.
2135 // Internals of the Matcher (including some VectorSets) must remain live
2136 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2137 // set a bit in reclaimed memory.
2138
2139 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2140 // nodes. Mapping is only valid at the root of each matched subtree.
2141 NOT_PRODUCT( verify_graph_edges(); )
2142
2143 Node_List proj_list;
2144 Matcher m(proj_list);
2145 _matcher = &m;
2146 {
2147 TracePhase t2("matcher", &_t_matcher, true);
2148 m.match();
2149 }
2150 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2151 // nodes. Mapping is only valid at the root of each matched subtree.
2152 NOT_PRODUCT( verify_graph_edges(); )
2153
2154 // If you have too many nodes, or if matching has failed, bail out
2155 check_node_count(0, "out of nodes matching instructions");
2156 if (failing()) return;
2157
2158 // Platform dependent post matching hook (used on ppc).
2159 PdCompile::pd_post_matching_hook(this);
2160
2161 // Build a proper-looking CFG
2162 PhaseCFG cfg(node_arena(), root(), m);
2163 _cfg = &cfg;
2164 {
2165 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2166 cfg.Dominators();
2167 if (failing()) return;
2168
2169 NOT_PRODUCT( verify_graph_edges(); )
2170
2171 cfg.Estimate_Block_Frequency();
2172 cfg.GlobalCodeMotion(m,unique(),proj_list);
2173 if (failing()) return;
2174
2175 print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2176
2177 NOT_PRODUCT( verify_graph_edges(); )
2178
2179 debug_only( cfg.verify(); )
2180 }
2181 NOT_PRODUCT( verify_graph_edges(); )
2182
2183 PhaseChaitin regalloc(unique(),cfg,m);
2184 _regalloc = ®alloc;
2185 {
2186 TracePhase t2("regalloc", &_t_registerAllocation, true);
2187 // Perform any platform dependent preallocation actions. This is used,
2188 // for example, to avoid taking an implicit null pointer exception
2189 // using the frame pointer on win95.
2190 _regalloc->pd_preallocate_hook();
2191
2192 // Perform register allocation. After Chaitin, use-def chains are
2193 // no longer accurate (at spill code) and so must be ignored.
2194 // Node->LRG->reg mappings are still accurate.
2195 _regalloc->Register_Allocate();
2196
2197 // Bail out if the allocator builds too many nodes
2198 if (failing()) return;
2199 }
2200
2201 // Prior to register allocation we kept empty basic blocks in case the
2202 // the allocator needed a place to spill. After register allocation we
2203 // are not adding any new instructions. If any basic block is empty, we
2204 // can now safely remove it.
2205 {
2206 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2207 cfg.remove_empty();
2208 if (do_freq_based_layout()) {
2209 PhaseBlockLayout layout(cfg);
2210 } else {
2211 cfg.set_loop_alignment();
2212 }
2213 cfg.fixup_flow();
2214 }
2215
2216 // Perform any platform dependent postallocation verifications.
2217 debug_only( _regalloc->pd_postallocate_verify_hook(); )
2218
2219 // Apply peephole optimizations
2220 if( OptoPeephole ) {
2221 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2222 PhasePeephole peep( _regalloc, cfg);
2223 peep.do_transform();
2224 }
2225
2226 // Do late expand if CPU requires this.
2227 if (Matcher::require_late_expand) {
2228 cfg.LateExpand(_regalloc);
2229 }
2230
2231 // Convert Nodes to instruction bits in a buffer
2232 {
2233 // %%%% workspace merge brought two timers together for one job
2234 TracePhase t2a("output", &_t_output, true);
2235 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2236 Output();
2237 }
2238
2239 print_method(PHASE_FINAL_CODE);
2240
2241 // He's dead, Jim.
2242 _cfg = (PhaseCFG*)0xdeadbeef;
2243 _regalloc = (PhaseChaitin*)0xdeadbeef;
2244 }
2245
2246
2247 //------------------------------dump_asm---------------------------------------
2248 // Dump formatted assembly
2249 #ifndef PRODUCT
2250 void Compile::dump_asm(int *pcs, uint pc_limit) {
2251 bool cut_short = false;
2252 tty->print_cr("#");
2253 tty->print("# "); _tf->dump(); tty->cr();
2254 tty->print_cr("#");
2255
2256 // For all blocks
2257 int pc = 0x0; // Program counter
2258 char starts_bundle = ' ';
2259 _regalloc->dump_frame();
2260
2261 Node *n = NULL;
2262 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
2263 if (VMThread::should_terminate()) { cut_short = true; break; }
2264 Block *b = _cfg->_blocks[i];
2265 if (b->is_connector() && !Verbose) continue;
2266 n = b->_nodes[0];
2267 if (pcs && n->_idx < pc_limit)
2268 tty->print("%3.3x ", pcs[n->_idx]);
2269 else
2270 tty->print(" ");
2271 b->dump_head( &_cfg->_bbs );
2272 if (b->is_connector()) {
2273 tty->print_cr(" # Empty connector block");
2274 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2275 tty->print_cr(" # Block is sole successor of call");
2276 }
2277
2278 // For all instructions
2279 Node *delay = NULL;
2280 for( uint j = 0; j<b->_nodes.size(); j++ ) {
2281 if (VMThread::should_terminate()) { cut_short = true; break; }
2282 n = b->_nodes[j];
2283 if (valid_bundle_info(n)) {
2284 Bundle *bundle = node_bundling(n);
2285 if (bundle->used_in_unconditional_delay()) {
2286 delay = n;
2287 continue;
2288 }
2289 if (bundle->starts_bundle())
2290 starts_bundle = '+';
2291 }
2292
2293 if (WizardMode) n->dump();
2294
2295 if( !n->is_Region() && // Dont print in the Assembly
2296 !n->is_Phi() && // a few noisely useless nodes
2297 !n->is_Proj() &&
2298 !n->is_MachTemp() &&
2299 !n->is_SafePointScalarObject() &&
2300 !n->is_Catch() && // Would be nice to print exception table targets
2301 !n->is_MergeMem() && // Not very interesting
2302 !n->is_top() && // Debug info table constants
2303 !(n->is_Con() && !n->is_Mach())// Debug info table constants
2304 ) {
2305 if (pcs && n->_idx < pc_limit)
2306 tty->print("%3.3x", pcs[n->_idx]);
2307 else
2308 tty->print(" ");
2309 tty->print(" %c ", starts_bundle);
2310 starts_bundle = ' ';
2311 tty->print("\t");
2312 n->format(_regalloc, tty);
2313 tty->cr();
2314 }
2315
2316 // If we have an instruction with a delay slot, and have seen a delay,
2317 // then back up and print it
2318 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2319 assert(delay != NULL, "no unconditional delay instruction");
2320 if (WizardMode) delay->dump();
2321
2322 if (node_bundling(delay)->starts_bundle())
2323 starts_bundle = '+';
2324 if (pcs && n->_idx < pc_limit)
2325 tty->print("%3.3x", pcs[n->_idx]);
2326 else
2327 tty->print(" ");
2328 tty->print(" %c ", starts_bundle);
2329 starts_bundle = ' ';
2330 tty->print("\t");
2331 delay->format(_regalloc, tty);
2332 tty->print_cr("");
2333 delay = NULL;
2334 }
2335
2336 // Dump the exception table as well
2337 if( n->is_Catch() && (Verbose || WizardMode) ) {
2338 // Print the exception table for this offset
2339 _handler_table.print_subtable_for(pc);
2340 }
2341 }
2342
2343 if (pcs && n->_idx < pc_limit)
2344 tty->print_cr("%3.3x", pcs[n->_idx]);
2345 else
2346 tty->print_cr("");
2347
2348 assert(cut_short || delay == NULL, "no unconditional delay branch");
2349
2350 } // End of per-block dump
2351 tty->print_cr("");
2352
2353 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2354 }
2355 #endif
2356
2357 //------------------------------Final_Reshape_Counts---------------------------
2358 // This class defines counters to help identify when a method
2359 // may/must be executed using hardware with only 24-bit precision.
2360 struct Final_Reshape_Counts : public StackObj {
2361 int _call_count; // count non-inlined 'common' calls
2362 int _float_count; // count float ops requiring 24-bit precision
2363 int _double_count; // count double ops requiring more precision
2364 int _java_call_count; // count non-inlined 'java' calls
2365 int _inner_loop_count; // count loops which need alignment
2366 VectorSet _visited; // Visitation flags
2367 Node_List _tests; // Set of IfNodes & PCTableNodes
2368
2369 Final_Reshape_Counts() :
2370 _call_count(0), _float_count(0), _double_count(0),
2371 _java_call_count(0), _inner_loop_count(0),
2372 _visited( Thread::current()->resource_area() ) { }
2373
2374 void inc_call_count () { _call_count ++; }
2375 void inc_float_count () { _float_count ++; }
2376 void inc_double_count() { _double_count++; }
2377 void inc_java_call_count() { _java_call_count++; }
2378 void inc_inner_loop_count() { _inner_loop_count++; }
2379
2380 int get_call_count () const { return _call_count ; }
2381 int get_float_count () const { return _float_count ; }
2382 int get_double_count() const { return _double_count; }
2383 int get_java_call_count() const { return _java_call_count; }
2384 int get_inner_loop_count() const { return _inner_loop_count; }
2385 };
2386
2387 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2388 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2389 // Make sure the offset goes inside the instance layout.
2390 return k->contains_field_offset(tp->offset());
2391 // Note that OffsetBot and OffsetTop are very negative.
2392 }
2393
2394 // Eliminate trivially redundant StoreCMs and accumulate their
2395 // precedence edges.
2396 void Compile::eliminate_redundant_card_marks(Node* n) {
2397 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2398 if (n->in(MemNode::Address)->outcnt() > 1) {
2399 // There are multiple users of the same address so it might be
2400 // possible to eliminate some of the StoreCMs
2401 Node* mem = n->in(MemNode::Memory);
2402 Node* adr = n->in(MemNode::Address);
2403 Node* val = n->in(MemNode::ValueIn);
2404 Node* prev = n;
2405 bool done = false;
2406 // Walk the chain of StoreCMs eliminating ones that match. As
2407 // long as it's a chain of single users then the optimization is
2408 // safe. Eliminating partially redundant StoreCMs would require
2409 // cloning copies down the other paths.
2410 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2411 if (adr == mem->in(MemNode::Address) &&
2412 val == mem->in(MemNode::ValueIn)) {
2413 // redundant StoreCM
2414 if (mem->req() > MemNode::OopStore) {
2415 // Hasn't been processed by this code yet.
2416 n->add_prec(mem->in(MemNode::OopStore));
2417 } else {
2418 // Already converted to precedence edge
2419 for (uint i = mem->req(); i < mem->len(); i++) {
2420 // Accumulate any precedence edges
2421 if (mem->in(i) != NULL) {
2422 n->add_prec(mem->in(i));
2423 }
2424 }
2425 // Everything above this point has been processed.
2426 done = true;
2427 }
2428 // Eliminate the previous StoreCM
2429 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2430 assert(mem->outcnt() == 0, "should be dead");
2431 mem->disconnect_inputs(NULL, this);
2432 } else {
2433 prev = mem;
2434 }
2435 mem = prev->in(MemNode::Memory);
2436 }
2437 }
2438 }
2439
2440 //------------------------------final_graph_reshaping_impl----------------------
2441 // Implement items 1-5 from final_graph_reshaping below.
2442 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2443
2444 if ( n->outcnt() == 0 ) return; // dead node
2445 uint nop = n->Opcode();
2446
2447 // Check for 2-input instruction with "last use" on right input.
2448 // Swap to left input. Implements item (2).
2449 if( n->req() == 3 && // two-input instruction
2450 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2451 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2452 n->in(2)->outcnt() == 1 &&// right use IS a last use
2453 !n->in(2)->is_Con() ) { // right use is not a constant
2454 // Check for commutative opcode
2455 switch( nop ) {
2456 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2457 case Op_MaxI: case Op_MinI:
2458 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2459 case Op_AndL: case Op_XorL: case Op_OrL:
2460 case Op_AndI: case Op_XorI: case Op_OrI: {
2461 // Move "last use" input to left by swapping inputs
2462 n->swap_edges(1, 2);
2463 break;
2464 }
2465 default:
2466 break;
2467 }
2468 }
2469
2470 #ifdef ASSERT
2471 if( n->is_Mem() ) {
2472 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2473 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2474 // oop will be recorded in oop map if load crosses safepoint
2475 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2476 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2477 "raw memory operations should have control edge");
2478 }
2479 #endif
2480 // Count FPU ops and common calls, implements item (3)
2481 switch( nop ) {
2482 // Count all float operations that may use FPU
2483 case Op_AddF:
2484 case Op_SubF:
2485 case Op_MulF:
2486 case Op_DivF:
2487 case Op_NegF:
2488 case Op_ModF:
2489 case Op_ConvI2F:
2490 case Op_ConF:
2491 case Op_CmpF:
2492 case Op_CmpF3:
2493 // case Op_ConvL2F: // longs are split into 32-bit halves
2494 frc.inc_float_count();
2495 break;
2496
2497 case Op_ConvF2D:
2498 case Op_ConvD2F:
2499 frc.inc_float_count();
2500 frc.inc_double_count();
2501 break;
2502
2503 // Count all double operations that may use FPU
2504 case Op_AddD:
2505 case Op_SubD:
2506 case Op_MulD:
2507 case Op_DivD:
2508 case Op_NegD:
2509 case Op_ModD:
2510 case Op_ConvI2D:
2511 case Op_ConvD2I:
2512 // case Op_ConvL2D: // handled by leaf call
2513 // case Op_ConvD2L: // handled by leaf call
2514 case Op_ConD:
2515 case Op_CmpD:
2516 case Op_CmpD3:
2517 frc.inc_double_count();
2518 break;
2519 case Op_Opaque1: // Remove Opaque Nodes before matching
2520 case Op_Opaque2: // Remove Opaque Nodes before matching
2521 n->subsume_by(n->in(1), this);
2522 break;
2523 case Op_CallStaticJava:
2524 case Op_CallJava:
2525 case Op_CallDynamicJava:
2526 frc.inc_java_call_count(); // Count java call site;
2527 case Op_CallRuntime:
2528 case Op_CallLeaf:
2529 case Op_CallLeafNoFP: {
2530 assert( n->is_Call(), "" );
2531 CallNode *call = n->as_Call();
2532 // Count call sites where the FP mode bit would have to be flipped.
2533 // Do not count uncommon runtime calls:
2534 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2535 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2536 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2537 frc.inc_call_count(); // Count the call site
2538 } else { // See if uncommon argument is shared
2539 Node *n = call->in(TypeFunc::Parms);
2540 int nop = n->Opcode();
2541 // Clone shared simple arguments to uncommon calls, item (1).
2542 if( n->outcnt() > 1 &&
2543 !n->is_Proj() &&
2544 nop != Op_CreateEx &&
2545 nop != Op_CheckCastPP &&
2546 nop != Op_DecodeN &&
2547 !n->is_Mem() ) {
2548 Node *x = n->clone();
2549 call->set_req( TypeFunc::Parms, x );
2550 }
2551 }
2552 break;
2553 }
2554
2555 case Op_StoreD:
2556 case Op_LoadD:
2557 case Op_LoadD_unaligned:
2558 frc.inc_double_count();
2559 goto handle_mem;
2560 case Op_StoreF:
2561 case Op_LoadF:
2562 frc.inc_float_count();
2563 goto handle_mem;
2564
2565 case Op_StoreCM:
2566 {
2567 // Convert OopStore dependence into precedence edge
2568 Node* prec = n->in(MemNode::OopStore);
2569 n->del_req(MemNode::OopStore);
2570 n->add_prec(prec);
2571 eliminate_redundant_card_marks(n);
2572 }
2573
2574 // fall through
2575
2576 case Op_StoreB:
2577 case Op_StoreC:
2578 case Op_StorePConditional:
2579 case Op_StoreI:
2580 case Op_StoreL:
2581 case Op_StoreIConditional:
2582 case Op_StoreLConditional:
2583 case Op_CompareAndSwapI:
2584 case Op_CompareAndSwapL:
2585 case Op_CompareAndSwapP:
2586 case Op_CompareAndSwapN:
2587 case Op_GetAndAddI:
2588 case Op_GetAndAddL:
2589 case Op_GetAndSetI:
2590 case Op_GetAndSetL:
2591 case Op_GetAndSetP:
2592 case Op_GetAndSetN:
2593 case Op_StoreP:
2594 case Op_StoreN:
2595 case Op_LoadB:
2596 case Op_LoadUB:
2597 case Op_LoadUS:
2598 case Op_LoadI:
2599 case Op_LoadKlass:
2600 case Op_LoadNKlass:
2601 case Op_LoadL:
2602 case Op_LoadL_unaligned:
2603 case Op_LoadPLocked:
2604 case Op_LoadP:
2605 case Op_LoadN:
2606 case Op_LoadRange:
2607 case Op_LoadS: {
2608 handle_mem:
2609 #ifdef ASSERT
2610 if( VerifyOptoOopOffsets ) {
2611 assert( n->is_Mem(), "" );
2612 MemNode *mem = (MemNode*)n;
2613 // Check to see if address types have grounded out somehow.
2614 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2615 assert( !tp || oop_offset_is_sane(tp), "" );
2616 }
2617 #endif
2618 break;
2619 }
2620
2621 case Op_AddP: { // Assert sane base pointers
2622 Node *addp = n->in(AddPNode::Address);
2623 assert( !addp->is_AddP() ||
2624 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2625 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2626 "Base pointers must match" );
2627 #ifdef _LP64
2628 if (UseCompressedOops &&
2629 addp->Opcode() == Op_ConP &&
2630 addp == n->in(AddPNode::Base) &&
2631 n->in(AddPNode::Offset)->is_Con()) {
2632 // Use addressing with narrow klass to load with offset on x86.
2633 // On sparc loading 32-bits constant and decoding it have less
2634 // instructions (4) then load 64-bits constant (7).
2635 // Do this transformation here since IGVN will convert ConN back to ConP.
2636 const Type* t = addp->bottom_type();
2637 if (t->isa_oopptr()) {
2638 Node* nn = NULL;
2639
2640 // Look for existing ConN node of the same exact type.
2641 Node* r = root();
2642 uint cnt = r->outcnt();
2643 for (uint i = 0; i < cnt; i++) {
2644 Node* m = r->raw_out(i);
2645 if (m!= NULL && m->Opcode() == Op_ConN &&
2646 m->bottom_type()->make_ptr() == t) {
2647 nn = m;
2648 break;
2649 }
2650 }
2651 if (nn != NULL) {
2652 // Decode a narrow oop to match address
2653 // [R12 + narrow_oop_reg<<3 + offset]
2654 nn = new (this) DecodeNNode(nn, t);
2655 n->set_req(AddPNode::Base, nn);
2656 n->set_req(AddPNode::Address, nn);
2657 if (addp->outcnt() == 0) {
2658 addp->disconnect_inputs(NULL, this);
2659 }
2660 }
2661 }
2662 }
2663 #endif
2664 break;
2665 }
2666
2667 #ifdef _LP64
2668 case Op_CastPP:
2669 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2670 Node* in1 = n->in(1);
2671 const Type* t = n->bottom_type();
2672 Node* new_in1 = in1->clone();
2673 new_in1->as_DecodeN()->set_type(t);
2674
2675 if (!Matcher::narrow_oop_use_complex_address()) {
2676 //
2677 // x86, ARM and friends can handle 2 adds in addressing mode
2678 // and Matcher can fold a DecodeN node into address by using
2679 // a narrow oop directly and do implicit NULL check in address:
2680 //
2681 // [R12 + narrow_oop_reg<<3 + offset]
2682 // NullCheck narrow_oop_reg
2683 //
2684 // On other platforms (Sparc) we have to keep new DecodeN node and
2685 // use it to do implicit NULL check in address:
2686 //
2687 // decode_not_null narrow_oop_reg, base_reg
2688 // [base_reg + offset]
2689 // NullCheck base_reg
2690 //
2691 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2692 // to keep the information to which NULL check the new DecodeN node
2693 // corresponds to use it as value in implicit_null_check().
2694 //
2695 new_in1->set_req(0, n->in(0));
2696 }
2697
2698 n->subsume_by(new_in1, this);
2699 if (in1->outcnt() == 0) {
2700 in1->disconnect_inputs(NULL, this);
2701 }
2702 }
2703 break;
2704
2705 case Op_CmpP:
2706 // Do this transformation here to preserve CmpPNode::sub() and
2707 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2708 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2709 Node* in1 = n->in(1);
2710 Node* in2 = n->in(2);
2711 if (!in1->is_DecodeN()) {
2712 in2 = in1;
2713 in1 = n->in(2);
2714 }
2715 assert(in1->is_DecodeN(), "sanity");
2716
2717 Node* new_in2 = NULL;
2718 if (in2->is_DecodeN()) {
2719 new_in2 = in2->in(1);
2720 } else if (in2->Opcode() == Op_ConP) {
2721 const Type* t = in2->bottom_type();
2722 if (t == TypePtr::NULL_PTR) {
2723 // Don't convert CmpP null check into CmpN if compressed
2724 // oops implicit null check is not generated.
2725 // This will allow to generate normal oop implicit null check.
2726 if (Matcher::gen_narrow_oop_implicit_null_checks())
2727 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2728 //
2729 // This transformation together with CastPP transformation above
2730 // will generated code for implicit NULL checks for compressed oops.
2731 //
2732 // The original code after Optimize()
2733 //
2734 // LoadN memory, narrow_oop_reg
2735 // decode narrow_oop_reg, base_reg
2736 // CmpP base_reg, NULL
2737 // CastPP base_reg // NotNull
2738 // Load [base_reg + offset], val_reg
2739 //
2740 // after these transformations will be
2741 //
2742 // LoadN memory, narrow_oop_reg
2743 // CmpN narrow_oop_reg, NULL
2744 // decode_not_null narrow_oop_reg, base_reg
2745 // Load [base_reg + offset], val_reg
2746 //
2747 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2748 // since narrow oops can be used in debug info now (see the code in
2749 // final_graph_reshaping_walk()).
2750 //
2751 // At the end the code will be matched to
2752 // on x86:
2753 //
2754 // Load_narrow_oop memory, narrow_oop_reg
2755 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2756 // NullCheck narrow_oop_reg
2757 //
2758 // and on sparc:
2759 //
2760 // Load_narrow_oop memory, narrow_oop_reg
2761 // decode_not_null narrow_oop_reg, base_reg
2762 // Load [base_reg + offset], val_reg
2763 // NullCheck base_reg
2764 //
2765 } else if (t->isa_oopptr()) {
2766 new_in2 = ConNode::make(this, t->make_narrowoop());
2767 }
2768 }
2769 if (new_in2 != NULL) {
2770 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2771 n->subsume_by(cmpN, this);
2772 if (in1->outcnt() == 0) {
2773 in1->disconnect_inputs(NULL, this);
2774 }
2775 if (in2->outcnt() == 0) {
2776 in2->disconnect_inputs(NULL, this);
2777 }
2778 }
2779 }
2780 break;
2781
2782 case Op_DecodeN:
2783 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2784 // DecodeN could be pinned when it can't be fold into
2785 // an address expression, see the code for Op_CastPP above.
2786 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2787 break;
2788
2789 case Op_EncodeP: {
2790 Node* in1 = n->in(1);
2791 if (in1->is_DecodeN()) {
2792 n->subsume_by(in1->in(1), this);
2793 } else if (in1->Opcode() == Op_ConP) {
2794 const Type* t = in1->bottom_type();
2795 if (t == TypePtr::NULL_PTR) {
2796 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2797 } else if (t->isa_oopptr()) {
2798 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2799 }
2800 }
2801 if (in1->outcnt() == 0) {
2802 in1->disconnect_inputs(NULL, this);
2803 }
2804 break;
2805 }
2806
2807 case Op_Proj: {
2808 if (OptimizeStringConcat) {
2809 ProjNode* p = n->as_Proj();
2810 if (p->_is_io_use) {
2811 // Separate projections were used for the exception path which
2812 // are normally removed by a late inline. If it wasn't inlined
2813 // then they will hang around and should just be replaced with
2814 // the original one.
2815 Node* proj = NULL;
2816 // Replace with just one
2817 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2818 Node *use = i.get();
2819 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2820 proj = use;
2821 break;
2822 }
2823 }
2824 assert(proj != NULL, "must be found");
2825 p->subsume_by(proj, this);
2826 }
2827 }
2828 break;
2829 }
2830
2831 case Op_Phi:
2832 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2833 // The EncodeP optimization may create Phi with the same edges
2834 // for all paths. It is not handled well by Register Allocator.
2835 Node* unique_in = n->in(1);
2836 assert(unique_in != NULL, "");
2837 uint cnt = n->req();
2838 for (uint i = 2; i < cnt; i++) {
2839 Node* m = n->in(i);
2840 assert(m != NULL, "");
2841 if (unique_in != m)
2842 unique_in = NULL;
2843 }
2844 if (unique_in != NULL) {
2845 n->subsume_by(unique_in, this);
2846 }
2847 }
2848 break;
2849
2850 #endif
2851
2852 case Op_ModI:
2853 if (UseDivMod) {
2854 // Check if a%b and a/b both exist
2855 Node* d = n->find_similar(Op_DivI);
2856 if (d) {
2857 // Replace them with a fused divmod if supported
2858 if (Matcher::has_match_rule(Op_DivModI)) {
2859 DivModINode* divmod = DivModINode::make(this, n);
2860 d->subsume_by(divmod->div_proj(), this);
2861 n->subsume_by(divmod->mod_proj(), this);
2862 } else {
2863 // replace a%b with a-((a/b)*b)
2864 Node* mult = new (this) MulINode(d, d->in(2));
2865 Node* sub = new (this) SubINode(d->in(1), mult);
2866 n->subsume_by(sub, this);
2867 }
2868 }
2869 }
2870 break;
2871
2872 case Op_ModL:
2873 if (UseDivMod) {
2874 // Check if a%b and a/b both exist
2875 Node* d = n->find_similar(Op_DivL);
2876 if (d) {
2877 // Replace them with a fused divmod if supported
2878 if (Matcher::has_match_rule(Op_DivModL)) {
2879 DivModLNode* divmod = DivModLNode::make(this, n);
2880 d->subsume_by(divmod->div_proj(), this);
2881 n->subsume_by(divmod->mod_proj(), this);
2882 } else {
2883 // replace a%b with a-((a/b)*b)
2884 Node* mult = new (this) MulLNode(d, d->in(2));
2885 Node* sub = new (this) SubLNode(d->in(1), mult);
2886 n->subsume_by(sub, this);
2887 }
2888 }
2889 }
2890 break;
2891
2892 case Op_LoadVector:
2893 case Op_StoreVector:
2894 break;
2895
2896 case Op_PackB:
2897 case Op_PackS:
2898 case Op_PackI:
2899 case Op_PackF:
2900 case Op_PackL:
2901 case Op_PackD:
2902 if (n->req()-1 > 2) {
2903 // Replace many operand PackNodes with a binary tree for matching
2904 PackNode* p = (PackNode*) n;
2905 Node* btp = p->binary_tree_pack(this, 1, n->req());
2906 n->subsume_by(btp, this);
2907 }
2908 break;
2909 case Op_Loop:
2910 case Op_CountedLoop:
2911 if (n->as_Loop()->is_inner_loop()) {
2912 frc.inc_inner_loop_count();
2913 }
2914 break;
2915 case Op_LShiftI:
2916 case Op_RShiftI:
2917 case Op_URShiftI:
2918 case Op_LShiftL:
2919 case Op_RShiftL:
2920 case Op_URShiftL:
2921 if (Matcher::need_masked_shift_count) {
2922 // The cpu's shift instructions don't restrict the count to the
2923 // lower 5/6 bits. We need to do the masking ourselves.
2924 Node* in2 = n->in(2);
2925 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2926 const TypeInt* t = in2->find_int_type();
2927 if (t != NULL && t->is_con()) {
2928 juint shift = t->get_con();
2929 if (shift > mask) { // Unsigned cmp
2930 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
2931 }
2932 } else {
2933 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2934 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
2935 n->set_req(2, shift);
2936 }
2937 }
2938 if (in2->outcnt() == 0) { // Remove dead node
2939 in2->disconnect_inputs(NULL, this);
2940 }
2941 }
2942 break;
2943 case Op_MemBarStoreStore:
2944 // Break the link with AllocateNode: it is no longer useful and
2945 // confuses register allocation.
2946 if (n->req() > MemBarNode::Precedent) {
2947 n->set_req(MemBarNode::Precedent, top());
2948 }
2949 break;
2950 default:
2951 assert( !n->is_Call(), "" );
2952 assert( !n->is_Mem(), "" );
2953 break;
2954 }
2955
2956 // Collect CFG split points
2957 if (n->is_MultiBranch())
2958 frc._tests.push(n);
2959 }
2960
2961 //------------------------------final_graph_reshaping_walk---------------------
2962 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2963 // requires that the walk visits a node's inputs before visiting the node.
2964 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2965 ResourceArea *area = Thread::current()->resource_area();
2966 Unique_Node_List sfpt(area);
2967
2968 frc._visited.set(root->_idx); // first, mark node as visited
2969 uint cnt = root->req();
2970 Node *n = root;
2971 uint i = 0;
2972 while (true) {
2973 if (i < cnt) {
2974 // Place all non-visited non-null inputs onto stack
2975 Node* m = n->in(i);
2976 ++i;
2977 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2978 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2979 sfpt.push(m);
2980 cnt = m->req();
2981 nstack.push(n, i); // put on stack parent and next input's index
2982 n = m;
2983 i = 0;
2984 }
2985 } else {
2986 // Now do post-visit work
2987 final_graph_reshaping_impl( n, frc );
2988 if (nstack.is_empty())
2989 break; // finished
2990 n = nstack.node(); // Get node from stack
2991 cnt = n->req();
2992 i = nstack.index();
2993 nstack.pop(); // Shift to the next node on stack
2994 }
2995 }
2996
2997 // Skip next transformation if compressed oops are not used.
2998 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2999 return;
3000
3001 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
3002 // It could be done for an uncommon traps or any safepoints/calls
3003 // if the DecodeN node is referenced only in a debug info.
3004 while (sfpt.size() > 0) {
3005 n = sfpt.pop();
3006 JVMState *jvms = n->as_SafePoint()->jvms();
3007 assert(jvms != NULL, "sanity");
3008 int start = jvms->debug_start();
3009 int end = n->req();
3010 bool is_uncommon = (n->is_CallStaticJava() &&
3011 n->as_CallStaticJava()->uncommon_trap_request() != 0);
3012 for (int j = start; j < end; j++) {
3013 Node* in = n->in(j);
3014 if (in->is_DecodeN()) {
3015 bool safe_to_skip = true;
3016 if (!is_uncommon ) {
3017 // Is it safe to skip?
3018 for (uint i = 0; i < in->outcnt(); i++) {
3019 Node* u = in->raw_out(i);
3020 if (!u->is_SafePoint() ||
3021 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3022 safe_to_skip = false;
3023 }
3024 }
3025 }
3026 if (safe_to_skip) {
3027 n->set_req(j, in->in(1));
3028 }
3029 if (in->outcnt() == 0) {
3030 in->disconnect_inputs(NULL, this);
3031 }
3032 }
3033 }
3034 }
3035 }
3036
3037 //------------------------------final_graph_reshaping--------------------------
3038 // Final Graph Reshaping.
3039 //
3040 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3041 // and not commoned up and forced early. Must come after regular
3042 // optimizations to avoid GVN undoing the cloning. Clone constant
3043 // inputs to Loop Phis; these will be split by the allocator anyways.
3044 // Remove Opaque nodes.
3045 // (2) Move last-uses by commutative operations to the left input to encourage
3046 // Intel update-in-place two-address operations and better register usage
3047 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
3048 // calls canonicalizing them back.
3049 // (3) Count the number of double-precision FP ops, single-precision FP ops
3050 // and call sites. On Intel, we can get correct rounding either by
3051 // forcing singles to memory (requires extra stores and loads after each
3052 // FP bytecode) or we can set a rounding mode bit (requires setting and
3053 // clearing the mode bit around call sites). The mode bit is only used
3054 // if the relative frequency of single FP ops to calls is low enough.
3055 // This is a key transform for SPEC mpeg_audio.
3056 // (4) Detect infinite loops; blobs of code reachable from above but not
3057 // below. Several of the Code_Gen algorithms fail on such code shapes,
3058 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
3059 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
3060 // Detection is by looking for IfNodes where only 1 projection is
3061 // reachable from below or CatchNodes missing some targets.
3062 // (5) Assert for insane oop offsets in debug mode.
3063
3064 bool Compile::final_graph_reshaping() {
3065 // an infinite loop may have been eliminated by the optimizer,
3066 // in which case the graph will be empty.
3067 if (root()->req() == 1) {
3068 record_method_not_compilable("trivial infinite loop");
3069 return true;
3070 }
3071
3072 // Expensive nodes have their control input set to prevent the GVN
3073 // from freely commoning them. There's no GVN beyond this point so
3074 // no need to keep the control input. We want the expensive nodes to
3075 // be freely moved to the least frequent code path by gcm.
3076 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3077 for (int i = 0; i < expensive_count(); i++) {
3078 _expensive_nodes->at(i)->set_req(0, NULL);
3079 }
3080
3081 Final_Reshape_Counts frc;
3082
3083 // Visit everybody reachable!
3084 // Allocate stack of size C->unique()/2 to avoid frequent realloc
3085 Node_Stack nstack(unique() >> 1);
3086 final_graph_reshaping_walk(nstack, root(), frc);
3087
3088 // Check for unreachable (from below) code (i.e., infinite loops).
3089 for( uint i = 0; i < frc._tests.size(); i++ ) {
3090 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3091 // Get number of CFG targets.
3092 // Note that PCTables include exception targets after calls.
3093 uint required_outcnt = n->required_outcnt();
3094 if (n->outcnt() != required_outcnt) {
3095 // Check for a few special cases. Rethrow Nodes never take the
3096 // 'fall-thru' path, so expected kids is 1 less.
3097 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3098 if (n->in(0)->in(0)->is_Call()) {
3099 CallNode *call = n->in(0)->in(0)->as_Call();
3100 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3101 required_outcnt--; // Rethrow always has 1 less kid
3102 } else if (call->req() > TypeFunc::Parms &&
3103 call->is_CallDynamicJava()) {
3104 // Check for null receiver. In such case, the optimizer has
3105 // detected that the virtual call will always result in a null
3106 // pointer exception. The fall-through projection of this CatchNode
3107 // will not be populated.
3108 Node *arg0 = call->in(TypeFunc::Parms);
3109 if (arg0->is_Type() &&
3110 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3111 required_outcnt--;
3112 }
3113 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3114 call->req() > TypeFunc::Parms+1 &&
3115 call->is_CallStaticJava()) {
3116 // Check for negative array length. In such case, the optimizer has
3117 // detected that the allocation attempt will always result in an
3118 // exception. There is no fall-through projection of this CatchNode .
3119 Node *arg1 = call->in(TypeFunc::Parms+1);
3120 if (arg1->is_Type() &&
3121 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3122 required_outcnt--;
3123 }
3124 }
3125 }
3126 }
3127 // Recheck with a better notion of 'required_outcnt'
3128 if (n->outcnt() != required_outcnt) {
3129 record_method_not_compilable("malformed control flow");
3130 return true; // Not all targets reachable!
3131 }
3132 }
3133 // Check that I actually visited all kids. Unreached kids
3134 // must be infinite loops.
3135 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3136 if (!frc._visited.test(n->fast_out(j)->_idx)) {
3137 record_method_not_compilable("infinite loop");
3138 return true; // Found unvisited kid; must be unreach
3139 }
3140 }
3141
3142 // If original bytecodes contained a mixture of floats and doubles
3143 // check if the optimizer has made it homogenous, item (3).
3144 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3145 frc.get_float_count() > 32 &&
3146 frc.get_double_count() == 0 &&
3147 (10 * frc.get_call_count() < frc.get_float_count()) ) {
3148 set_24_bit_selection_and_mode( false, true );
3149 }
3150
3151 set_java_calls(frc.get_java_call_count());
3152 set_inner_loops(frc.get_inner_loop_count());
3153
3154 // No infinite loops, no reason to bail out.
3155 return false;
3156 }
3157
3158 //-----------------------------too_many_traps----------------------------------
3159 // Report if there are too many traps at the current method and bci.
3160 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3161 bool Compile::too_many_traps(ciMethod* method,
3162 int bci,
3163 Deoptimization::DeoptReason reason) {
3164 ciMethodData* md = method->method_data();
3165 if (md->is_empty()) {
3166 // Assume the trap has not occurred, or that it occurred only
3167 // because of a transient condition during start-up in the interpreter.
3168 return false;
3169 }
3170 if (md->has_trap_at(bci, reason) != 0) {
3171 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3172 // Also, if there are multiple reasons, or if there is no per-BCI record,
3173 // assume the worst.
3174 if (log())
3175 log()->elem("observe trap='%s' count='%d'",
3176 Deoptimization::trap_reason_name(reason),
3177 md->trap_count(reason));
3178 return true;
3179 } else {
3180 // Ignore method/bci and see if there have been too many globally.
3181 return too_many_traps(reason, md);
3182 }
3183 }
3184
3185 // Less-accurate variant which does not require a method and bci.
3186 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3187 ciMethodData* logmd) {
3188 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
3189 // Too many traps globally.
3190 // Note that we use cumulative trap_count, not just md->trap_count.
3191 if (log()) {
3192 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3193 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3194 Deoptimization::trap_reason_name(reason),
3195 mcount, trap_count(reason));
3196 }
3197 return true;
3198 } else {
3199 // The coast is clear.
3200 return false;
3201 }
3202 }
3203
3204 //--------------------------too_many_recompiles--------------------------------
3205 // Report if there are too many recompiles at the current method and bci.
3206 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3207 // Is not eager to return true, since this will cause the compiler to use
3208 // Action_none for a trap point, to avoid too many recompilations.
3209 bool Compile::too_many_recompiles(ciMethod* method,
3210 int bci,
3211 Deoptimization::DeoptReason reason) {
3212 ciMethodData* md = method->method_data();
3213 if (md->is_empty()) {
3214 // Assume the trap has not occurred, or that it occurred only
3215 // because of a transient condition during start-up in the interpreter.
3216 return false;
3217 }
3218 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3219 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3220 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
3221 Deoptimization::DeoptReason per_bc_reason
3222 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3223 if ((per_bc_reason == Deoptimization::Reason_none
3224 || md->has_trap_at(bci, reason) != 0)
3225 // The trap frequency measure we care about is the recompile count:
3226 && md->trap_recompiled_at(bci)
3227 && md->overflow_recompile_count() >= bc_cutoff) {
3228 // Do not emit a trap here if it has already caused recompilations.
3229 // Also, if there are multiple reasons, or if there is no per-BCI record,
3230 // assume the worst.
3231 if (log())
3232 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3233 Deoptimization::trap_reason_name(reason),
3234 md->trap_count(reason),
3235 md->overflow_recompile_count());
3236 return true;
3237 } else if (trap_count(reason) != 0
3238 && decompile_count() >= m_cutoff) {
3239 // Too many recompiles globally, and we have seen this sort of trap.
3240 // Use cumulative decompile_count, not just md->decompile_count.
3241 if (log())
3242 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3243 Deoptimization::trap_reason_name(reason),
3244 md->trap_count(reason), trap_count(reason),
3245 md->decompile_count(), decompile_count());
3246 return true;
3247 } else {
3248 // The coast is clear.
3249 return false;
3250 }
3251 }
3252
3253
3254 #ifndef PRODUCT
3255 //------------------------------verify_graph_edges---------------------------
3256 // Walk the Graph and verify that there is a one-to-one correspondence
3257 // between Use-Def edges and Def-Use edges in the graph.
3258 void Compile::verify_graph_edges(bool no_dead_code) {
3259 if (VerifyGraphEdges) {
3260 ResourceArea *area = Thread::current()->resource_area();
3261 Unique_Node_List visited(area);
3262 // Call recursive graph walk to check edges
3263 _root->verify_edges(visited);
3264 if (no_dead_code) {
3265 // Now make sure that no visited node is used by an unvisited node.
3266 bool dead_nodes = 0;
3267 Unique_Node_List checked(area);
3268 while (visited.size() > 0) {
3269 Node* n = visited.pop();
3270 checked.push(n);
3271 for (uint i = 0; i < n->outcnt(); i++) {
3272 Node* use = n->raw_out(i);
3273 if (checked.member(use)) continue; // already checked
3274 if (visited.member(use)) continue; // already in the graph
3275 if (use->is_Con()) continue; // a dead ConNode is OK
3276 // At this point, we have found a dead node which is DU-reachable.
3277 if (dead_nodes++ == 0)
3278 tty->print_cr("*** Dead nodes reachable via DU edges:");
3279 use->dump(2);
3280 tty->print_cr("---");
3281 checked.push(use); // No repeats; pretend it is now checked.
3282 }
3283 }
3284 assert(dead_nodes == 0, "using nodes must be reachable from root");
3285 }
3286 }
3287 }
3288
3289 // Verify GC barriers consistency
3290 // Currently supported:
3291 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3292 void Compile::verify_barriers() {
3293 if (UseG1GC) {
3294 // Verify G1 pre-barriers
3295 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3296
3297 ResourceArea *area = Thread::current()->resource_area();
3298 Unique_Node_List visited(area);
3299 Node_List worklist(area);
3300 // We're going to walk control flow backwards starting from the Root
3301 worklist.push(_root);
3302 while (worklist.size() > 0) {
3303 Node* x = worklist.pop();
3304 if (x == NULL || x == top()) continue;
3305 if (visited.member(x)) {
3306 continue;
3307 } else {
3308 visited.push(x);
3309 }
3310
3311 if (x->is_Region()) {
3312 for (uint i = 1; i < x->req(); i++) {
3313 worklist.push(x->in(i));
3314 }
3315 } else {
3316 worklist.push(x->in(0));
3317 // We are looking for the pattern:
3318 // /->ThreadLocal
3319 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3320 // \->ConI(0)
3321 // We want to verify that the If and the LoadB have the same control
3322 // See GraphKit::g1_write_barrier_pre()
3323 if (x->is_If()) {
3324 IfNode *iff = x->as_If();
3325 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3326 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3327 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3328 && cmp->in(1)->is_Load()) {
3329 LoadNode* load = cmp->in(1)->as_Load();
3330 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3331 && load->in(2)->in(3)->is_Con()
3332 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3333
3334 Node* if_ctrl = iff->in(0);
3335 Node* load_ctrl = load->in(0);
3336
3337 if (if_ctrl != load_ctrl) {
3338 // Skip possible CProj->NeverBranch in infinite loops
3339 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3340 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3341 if_ctrl = if_ctrl->in(0)->in(0);
3342 }
3343 }
3344 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3345 }
3346 }
3347 }
3348 }
3349 }
3350 }
3351 }
3352 }
3353
3354 #endif
3355
3356 // The Compile object keeps track of failure reasons separately from the ciEnv.
3357 // This is required because there is not quite a 1-1 relation between the
3358 // ciEnv and its compilation task and the Compile object. Note that one
3359 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3360 // to backtrack and retry without subsuming loads. Other than this backtracking
3361 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3362 // by the logic in C2Compiler.
3363 void Compile::record_failure(const char* reason) {
3364 if (log() != NULL) {
3365 log()->elem("failure reason='%s' phase='compile'", reason);
3366 }
3367 if (_failure_reason == NULL) {
3368 // Record the first failure reason.
3369 _failure_reason = reason;
3370 }
3371
3372 EventCompilerFailure event;
3373 if (event.should_commit()) {
3374 event.set_compileID(Compile::compile_id());
3375 event.set_failure(reason);
3376 event.commit();
3377 }
3378
3379 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3380 C->print_method(PHASE_FAILURE);
3381 }
3382 _root = NULL; // flush the graph, too
3383 }
3384
3385 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3386 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3387 _phase_name(name), _dolog(dolog)
3388 {
3389 if (dolog) {
3390 C = Compile::current();
3391 _log = C->log();
3392 } else {
3393 C = NULL;
3394 _log = NULL;
3395 }
3396 if (_log != NULL) {
3397 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3398 _log->stamp();
3399 _log->end_head();
3400 }
3401 }
3402
3403 Compile::TracePhase::~TracePhase() {
3404
3405 C = Compile::current();
3406 if (_dolog) {
3407 _log = C->log();
3408 } else {
3409 _log = NULL;
3410 }
3411
3412 #ifdef ASSERT
3413 if (PrintIdealNodeCount) {
3414 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3415 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3416 }
3417
3418 if (VerifyIdealNodeCount) {
3419 Compile::current()->print_missing_nodes();
3420 }
3421 #endif
3422
3423 if (_log != NULL) {
3424 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3425 }
3426 }
3427
3428 //=============================================================================
3429 // Two Constant's are equal when the type and the value are equal.
3430 bool Compile::Constant::operator==(const Constant& other) {
3431 if (type() != other.type() ) return false;
3432 if (can_be_reused() != other.can_be_reused()) return false;
3433 // For floating point values we compare the bit pattern.
3434 switch (type()) {
3435 case T_FLOAT: return (_value.i == other._value.i);
3436 case T_LONG:
3437 case T_DOUBLE: return (_value.j == other._value.j);
3438 case T_OBJECT:
3439 case T_ADDRESS: return (_value.l == other._value.l);
3440 case T_VOID: return (_value.l == other._value.l); // jump-table entries
3441 default: ShouldNotReachHere();
3442 }
3443 return false;
3444 }
3445
3446 static int type_to_size_in_bytes(BasicType t) {
3447 switch (t) {
3448 case T_LONG: return sizeof(jlong );
3449 case T_FLOAT: return sizeof(jfloat );
3450 case T_DOUBLE: return sizeof(jdouble);
3451 // We use T_VOID as marker for jump-table entries (labels) which
3452 // need an internal word relocation.
3453 case T_VOID:
3454 case T_ADDRESS:
3455 case T_OBJECT: return sizeof(jobject);
3456 }
3457
3458 ShouldNotReachHere();
3459 return -1;
3460 }
3461
3462 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3463 // sort descending
3464 if (a->freq() > b->freq()) return -1;
3465 if (a->freq() < b->freq()) return 1;
3466 return 0;
3467 }
3468
3469 void Compile::ConstantTable::calculate_offsets_and_size() {
3470 // First, sort the array by frequencies.
3471 _constants.sort(qsort_comparator);
3472
3473 #ifdef ASSERT
3474 // Make sure all jump-table entries were sorted to the end of the
3475 // array (they have a negative frequency).
3476 bool found_void = false;
3477 for (int i = 0; i < _constants.length(); i++) {
3478 Constant con = _constants.at(i);
3479 if (con.type() == T_VOID)
3480 found_void = true; // jump-tables
3481 else
3482 assert(!found_void, "wrong sorting");
3483 }
3484 #endif
3485
3486 int offset = 0;
3487 for (int i = 0; i < _constants.length(); i++) {
3488 Constant* con = _constants.adr_at(i);
3489
3490 // Align offset for type.
3491 int typesize = type_to_size_in_bytes(con->type());
3492 offset = align_size_up(offset, typesize);
3493 con->set_offset(offset); // set constant's offset
3494
3495 if (con->type() == T_VOID) {
3496 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3497 offset = offset + typesize * n->outcnt(); // expand jump-table
3498 } else {
3499 offset = offset + typesize;
3500 }
3501 }
3502
3503 // Align size up to the next section start (which is insts; see
3504 // CodeBuffer::align_at_start).
3505 assert(_size == -1, "already set?");
3506 _size = align_size_up(offset, CodeEntryAlignment);
3507 }
3508
3509 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3510 MacroAssembler _masm(&cb);
3511 for (int i = 0; i < _constants.length(); i++) {
3512 Constant con = _constants.at(i);
3513 address constant_addr;
3514 switch (con.type()) {
3515 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3516 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3517 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3518 case T_OBJECT: {
3519 jobject obj = con.get_jobject();
3520 int oop_index = _masm.oop_recorder()->find_index(obj);
3521 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3522 break;
3523 }
3524 case T_ADDRESS: {
3525 address addr = (address) con.get_jobject();
3526 constant_addr = _masm.address_constant(addr);
3527 break;
3528 }
3529 // We use T_VOID as marker for jump-table entries (labels) which
3530 // need an internal word relocation.
3531 case T_VOID: {
3532 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3533 // Fill the jump-table with a dummy word. The real value is
3534 // filled in later in fill_jump_table.
3535 address dummy = (address) n;
3536 constant_addr = _masm.address_constant(dummy);
3537 // Expand jump-table
3538 for (uint i = 1; i < n->outcnt(); i++) {
3539 address temp_addr = _masm.address_constant(dummy + i);
3540 assert(temp_addr, "consts section too small");
3541 }
3542 break;
3543 }
3544 default: ShouldNotReachHere();
3545 }
3546 assert(constant_addr, "consts section too small");
3547 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3548 }
3549 }
3550
3551 int Compile::ConstantTable::find_offset(Constant& con) const {
3552 int idx = _constants.find(con);
3553 assert(idx != -1, "constant must be in constant table");
3554 int offset = _constants.at(idx).offset();
3555 assert(offset != -1, "constant table not emitted yet?");
3556 return offset;
3557 }
3558
3559 void Compile::ConstantTable::add(Constant& con) {
3560 if (con.can_be_reused()) {
3561 int idx = _constants.find(con);
3562 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3563 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
3564 return;
3565 }
3566 }
3567 (void) _constants.append(con);
3568 }
3569
3570 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3571 Block* b = Compile::current()->cfg()->_bbs[n->_idx];
3572 Constant con(type, value, b->_freq);
3573 add(con);
3574 return con;
3575 }
3576
3577 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3578 jvalue value;
3579 BasicType type = oper->type()->basic_type();
3580 switch (type) {
3581 case T_LONG: value.j = oper->constantL(); break;
3582 case T_FLOAT: value.f = oper->constantF(); break;
3583 case T_DOUBLE: value.d = oper->constantD(); break;
3584 case T_OBJECT:
3585 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3586 default: ShouldNotReachHere();
3587 }
3588 return add(n, type, value);
3589 }
3590
3591 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3592 jvalue value;
3593 // We can use the node pointer here to identify the right jump-table
3594 // as this method is called from Compile::Fill_buffer right before
3595 // the MachNodes are emitted and the jump-table is filled (means the
3596 // MachNode pointers do not change anymore).
3597 value.l = (jobject) n;
3598 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
3599 add(con);
3600 return con;
3601 }
3602
3603 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3604 // If called from Compile::scratch_emit_size do nothing.
3605 if (Compile::current()->in_scratch_emit_size()) return;
3606
3607 assert(labels.is_nonempty(), "must be");
3608 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3609
3610 // Since MachConstantNode::constant_offset() also contains
3611 // table_base_offset() we need to subtract the table_base_offset()
3612 // to get the plain offset into the constant table.
3613 int offset = n->constant_offset() - table_base_offset();
3614
3615 MacroAssembler _masm(&cb);
3616 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3617
3618 for (uint i = 0; i < n->outcnt(); i++) {
3619 address* constant_addr = &jump_table_base[i];
3620 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
3621 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3622 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3623 }
3624 }
3625
3626 void Compile::dump_inlining() {
3627 if (print_inlining() || print_intrinsics()) {
3628 // Print inlining message for candidates that we couldn't inline
3629 // for lack of space or non constant receiver
3630 for (int i = 0; i < _late_inlines.length(); i++) {
3631 CallGenerator* cg = _late_inlines.at(i);
3632 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3633 }
3634 Unique_Node_List useful;
3635 useful.push(root());
3636 for (uint next = 0; next < useful.size(); ++next) {
3637 Node* n = useful.at(next);
3638 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3639 CallNode* call = n->as_Call();
3640 CallGenerator* cg = call->generator();
3641 cg->print_inlining_late("receiver not constant");
3642 }
3643 uint max = n->len();
3644 for ( uint i = 0; i < max; ++i ) {
3645 Node *m = n->in(i);
3646 if ( m == NULL ) continue;
3647 useful.push(m);
3648 }
3649 }
3650 for (int i = 0; i < _print_inlining_list->length(); i++) {
3651 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string());
3652 }
3653 }
3654 }
3655
3656 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3657 if (n1->Opcode() < n2->Opcode()) return -1;
3658 else if (n1->Opcode() > n2->Opcode()) return 1;
3659
3660 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()));
3661 for (uint i = 1; i < n1->req(); i++) {
3662 if (n1->in(i) < n2->in(i)) return -1;
3663 else if (n1->in(i) > n2->in(i)) return 1;
3664 }
3665
3666 return 0;
3667 }
3668
3669 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3670 Node* n1 = *n1p;
3671 Node* n2 = *n2p;
3672
3673 return cmp_expensive_nodes(n1, n2);
3674 }
3675
3676 void Compile::sort_expensive_nodes() {
3677 if (!expensive_nodes_sorted()) {
3678 _expensive_nodes->sort(cmp_expensive_nodes);
3679 }
3680 }
3681
3682 bool Compile::expensive_nodes_sorted() const {
3683 for (int i = 1; i < _expensive_nodes->length(); i++) {
3684 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3685 return false;
3686 }
3687 }
3688 return true;
3689 }
3690
3691 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3692 if (_expensive_nodes->length() == 0) {
3693 return false;
3694 }
3695
3696 assert(OptimizeExpensiveOps, "optimization off?");
3697
3698 // Take this opportunity to remove dead nodes from the list
3699 int j = 0;
3700 for (int i = 0; i < _expensive_nodes->length(); i++) {
3701 Node* n = _expensive_nodes->at(i);
3702 if (!n->is_unreachable(igvn)) {
3703 assert(n->is_expensive(), "should be expensive");
3704 _expensive_nodes->at_put(j, n);
3705 j++;
3706 }
3707 }
3708 _expensive_nodes->trunc_to(j);
3709
3710 // Then sort the list so that similar nodes are next to each other
3711 // and check for at least two nodes of identical kind with same data
3712 // inputs.
3713 sort_expensive_nodes();
3714
3715 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3716 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3717 return true;
3718 }
3719 }
3720
3721 return false;
3722 }
3723
3724 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3725 if (_expensive_nodes->length() == 0) {
3726 return;
3727 }
3728
3729 assert(OptimizeExpensiveOps, "optimization off?");
3730
3731 // Sort to bring similar nodes next to each other and clear the
3732 // control input of nodes for which there's only a single copy.
3733 sort_expensive_nodes();
3734
3735 int j = 0;
3736 int identical = 0;
3737 int i = 0;
3738 for (; i < _expensive_nodes->length()-1; i++) {
3739 assert(j <= i, "can't write beyond current index");
3740 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3741 identical++;
3742 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3743 continue;
3744 }
3745 if (identical > 0) {
3746 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3747 identical = 0;
3748 } else {
3749 Node* n = _expensive_nodes->at(i);
3750 igvn.hash_delete(n);
3751 n->set_req(0, NULL);
3752 igvn.hash_insert(n);
3753 }
3754 }
3755 if (identical > 0) {
3756 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3757 } else if (_expensive_nodes->length() >= 1) {
3758 Node* n = _expensive_nodes->at(i);
3759 igvn.hash_delete(n);
3760 n->set_req(0, NULL);
3761 igvn.hash_insert(n);
3762 }
3763 _expensive_nodes->trunc_to(j);
3764 }
3765
3766 void Compile::add_expensive_node(Node * n) {
3767 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3768 assert(n->is_expensive(), "expensive nodes with non-null control here only");
3769 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
3770 if (OptimizeExpensiveOps) {
3771 _expensive_nodes->append(n);
3772 } else {
3773 // Clear control input and let IGVN optimize expensive nodes if
3774 // OptimizeExpensiveOps is off.
3775 n->set_req(0, NULL);
3776 }
3777 }