Tue, 28 Jun 2011 15:50:07 -0700
7047954: VM crashes with assert(is_Mem()) failed
Summary: cast constant array ptrs to bottom
Reviewed-by: never
1 /*
2 * Copyright (c) 1997, 2011, 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 */
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 "utilities/copy.hpp"
65 #ifdef TARGET_ARCH_MODEL_x86_32
66 # include "adfiles/ad_x86_32.hpp"
67 #endif
68 #ifdef TARGET_ARCH_MODEL_x86_64
69 # include "adfiles/ad_x86_64.hpp"
70 #endif
71 #ifdef TARGET_ARCH_MODEL_sparc
72 # include "adfiles/ad_sparc.hpp"
73 #endif
74 #ifdef TARGET_ARCH_MODEL_zero
75 # include "adfiles/ad_zero.hpp"
76 #endif
77 #ifdef TARGET_ARCH_MODEL_arm
78 # include "adfiles/ad_arm.hpp"
79 #endif
80 #ifdef TARGET_ARCH_MODEL_ppc
81 # include "adfiles/ad_ppc.hpp"
82 #endif
85 // -------------------- Compile::mach_constant_base_node -----------------------
86 // Constant table base node singleton.
87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 if (_mach_constant_base_node == NULL) {
89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 _mach_constant_base_node->add_req(C->root());
91 }
92 return _mach_constant_base_node;
93 }
96 /// Support for intrinsics.
98 // Return the index at which m must be inserted (or already exists).
99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 #ifdef ASSERT
102 for (int i = 1; i < _intrinsics->length(); i++) {
103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 CallGenerator* cg2 = _intrinsics->at(i);
105 assert(cg1->method() != cg2->method()
106 ? cg1->method() < cg2->method()
107 : cg1->is_virtual() < cg2->is_virtual(),
108 "compiler intrinsics list must stay sorted");
109 }
110 #endif
111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 int lo = 0, hi = _intrinsics->length()-1;
113 while (lo <= hi) {
114 int mid = (uint)(hi + lo) / 2;
115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 if (m < mid_m) {
117 hi = mid-1;
118 } else if (m > mid_m) {
119 lo = mid+1;
120 } else {
121 // look at minor sort key
122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 if (is_virtual < mid_virt) {
124 hi = mid-1;
125 } else if (is_virtual > mid_virt) {
126 lo = mid+1;
127 } else {
128 return mid; // exact match
129 }
130 }
131 }
132 return lo; // inexact match
133 }
135 void Compile::register_intrinsic(CallGenerator* cg) {
136 if (_intrinsics == NULL) {
137 _intrinsics = new GrowableArray<CallGenerator*>(60);
138 }
139 // This code is stolen from ciObjectFactory::insert.
140 // Really, GrowableArray should have methods for
141 // insert_at, remove_at, and binary_search.
142 int len = _intrinsics->length();
143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 if (index == len) {
145 _intrinsics->append(cg);
146 } else {
147 #ifdef ASSERT
148 CallGenerator* oldcg = _intrinsics->at(index);
149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 #endif
151 _intrinsics->append(_intrinsics->at(len-1));
152 int pos;
153 for (pos = len-2; pos >= index; pos--) {
154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 }
156 _intrinsics->at_put(index, cg);
157 }
158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 }
161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 if (_intrinsics != NULL) {
164 int index = intrinsic_insertion_index(m, is_virtual);
165 if (index < _intrinsics->length()
166 && _intrinsics->at(index)->method() == m
167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 return _intrinsics->at(index);
169 }
170 }
171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 if (cg != NULL) {
176 // Save it for next time:
177 register_intrinsic(cg);
178 return cg;
179 } else {
180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 }
182 }
183 return NULL;
184 }
186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 // in library_call.cpp.
190 #ifndef PRODUCT
191 // statistics gathering...
193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 int oflags = _intrinsic_hist_flags[id];
199 assert(flags != 0, "what happened?");
200 if (is_virtual) {
201 flags |= _intrinsic_virtual;
202 }
203 bool changed = (flags != oflags);
204 if ((flags & _intrinsic_worked) != 0) {
205 juint count = (_intrinsic_hist_count[id] += 1);
206 if (count == 1) {
207 changed = true; // first time
208 }
209 // increment the overall count also:
210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 }
212 if (changed) {
213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 // Something changed about the intrinsic's virtuality.
215 if ((flags & _intrinsic_virtual) != 0) {
216 // This is the first use of this intrinsic as a virtual call.
217 if (oflags != 0) {
218 // We already saw it as a non-virtual, so note both cases.
219 flags |= _intrinsic_both;
220 }
221 } else if ((oflags & _intrinsic_both) == 0) {
222 // This is the first use of this intrinsic as a non-virtual
223 flags |= _intrinsic_both;
224 }
225 }
226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 }
228 // update the overall flags also:
229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 return changed;
231 }
233 static char* format_flags(int flags, char* buf) {
234 buf[0] = 0;
235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 if (buf[0] == 0) strcat(buf, ",");
241 assert(buf[0] == ',', "must be");
242 return &buf[1];
243 }
245 void Compile::print_intrinsic_statistics() {
246 char flagsbuf[100];
247 ttyLocker ttyl;
248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 tty->print_cr("Compiler intrinsic usage:");
250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 #define PRINT_STAT_LINE(name, c, f) \
253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 int flags = _intrinsic_hist_flags[id];
257 juint count = _intrinsic_hist_count[id];
258 if ((flags | count) != 0) {
259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 }
261 }
262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 if (xtty != NULL) xtty->tail("statistics");
264 }
266 void Compile::print_statistics() {
267 { ttyLocker ttyl;
268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 Parse::print_statistics();
270 PhaseCCP::print_statistics();
271 PhaseRegAlloc::print_statistics();
272 Scheduling::print_statistics();
273 PhasePeephole::print_statistics();
274 PhaseIdealLoop::print_statistics();
275 if (xtty != NULL) xtty->tail("statistics");
276 }
277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 // put this under its own <statistics> element.
279 print_intrinsic_statistics();
280 }
281 }
282 #endif //PRODUCT
284 // Support for bundling info
285 Bundle* Compile::node_bundling(const Node *n) {
286 assert(valid_bundle_info(n), "oob");
287 return &_node_bundling_base[n->_idx];
288 }
290 bool Compile::valid_bundle_info(const Node *n) {
291 return (_node_bundling_limit > n->_idx);
292 }
295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 Node* use = n->last_out(i);
298 bool is_in_table = initial_gvn()->hash_delete(use);
299 uint uses_found = 0;
300 for (uint j = 0; j < use->len(); j++) {
301 if (use->in(j) == n) {
302 if (j < use->req())
303 use->set_req(j, nn);
304 else
305 use->set_prec(j, nn);
306 uses_found++;
307 }
308 }
309 if (is_in_table) {
310 // reinsert into table
311 initial_gvn()->hash_find_insert(use);
312 }
313 record_for_igvn(use);
314 i -= uses_found; // we deleted 1 or more copies of this edge
315 }
316 }
321 // Identify all nodes that are reachable from below, useful.
322 // Use breadth-first pass that records state in a Unique_Node_List,
323 // recursive traversal is slower.
324 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
325 int estimated_worklist_size = unique();
326 useful.map( estimated_worklist_size, NULL ); // preallocate space
328 // Initialize worklist
329 if (root() != NULL) { useful.push(root()); }
330 // If 'top' is cached, declare it useful to preserve cached node
331 if( cached_top_node() ) { useful.push(cached_top_node()); }
333 // Push all useful nodes onto the list, breadthfirst
334 for( uint next = 0; next < useful.size(); ++next ) {
335 assert( next < unique(), "Unique useful nodes < total nodes");
336 Node *n = useful.at(next);
337 uint max = n->len();
338 for( uint i = 0; i < max; ++i ) {
339 Node *m = n->in(i);
340 if( m == NULL ) continue;
341 useful.push(m);
342 }
343 }
344 }
346 // Disconnect all useless nodes by disconnecting those at the boundary.
347 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
348 uint next = 0;
349 while( next < useful.size() ) {
350 Node *n = useful.at(next++);
351 // Use raw traversal of out edges since this code removes out edges
352 int max = n->outcnt();
353 for (int j = 0; j < max; ++j ) {
354 Node* child = n->raw_out(j);
355 if( ! useful.member(child) ) {
356 assert( !child->is_top() || child != top(),
357 "If top is cached in Compile object it is in useful list");
358 // Only need to remove this out-edge to the useless node
359 n->raw_del_out(j);
360 --j;
361 --max;
362 }
363 }
364 if (n->outcnt() == 1 && n->has_special_unique_user()) {
365 record_for_igvn( n->unique_out() );
366 }
367 }
368 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
369 }
371 //------------------------------frame_size_in_words-----------------------------
372 // frame_slots in units of words
373 int Compile::frame_size_in_words() const {
374 // shift is 0 in LP32 and 1 in LP64
375 const int shift = (LogBytesPerWord - LogBytesPerInt);
376 int words = _frame_slots >> shift;
377 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
378 return words;
379 }
381 // ============================================================================
382 //------------------------------CompileWrapper---------------------------------
383 class CompileWrapper : public StackObj {
384 Compile *const _compile;
385 public:
386 CompileWrapper(Compile* compile);
388 ~CompileWrapper();
389 };
391 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
392 // the Compile* pointer is stored in the current ciEnv:
393 ciEnv* env = compile->env();
394 assert(env == ciEnv::current(), "must already be a ciEnv active");
395 assert(env->compiler_data() == NULL, "compile already active?");
396 env->set_compiler_data(compile);
397 assert(compile == Compile::current(), "sanity");
399 compile->set_type_dict(NULL);
400 compile->set_type_hwm(NULL);
401 compile->set_type_last_size(0);
402 compile->set_last_tf(NULL, NULL);
403 compile->set_indexSet_arena(NULL);
404 compile->set_indexSet_free_block_list(NULL);
405 compile->init_type_arena();
406 Type::Initialize(compile);
407 _compile->set_scratch_buffer_blob(NULL);
408 _compile->begin_method();
409 }
410 CompileWrapper::~CompileWrapper() {
411 _compile->end_method();
412 if (_compile->scratch_buffer_blob() != NULL)
413 BufferBlob::free(_compile->scratch_buffer_blob());
414 _compile->env()->set_compiler_data(NULL);
415 }
418 //----------------------------print_compile_messages---------------------------
419 void Compile::print_compile_messages() {
420 #ifndef PRODUCT
421 // Check if recompiling
422 if (_subsume_loads == false && PrintOpto) {
423 // Recompiling without allowing machine instructions to subsume loads
424 tty->print_cr("*********************************************************");
425 tty->print_cr("** Bailout: Recompile without subsuming loads **");
426 tty->print_cr("*********************************************************");
427 }
428 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
429 // Recompiling without escape analysis
430 tty->print_cr("*********************************************************");
431 tty->print_cr("** Bailout: Recompile without escape analysis **");
432 tty->print_cr("*********************************************************");
433 }
434 if (env()->break_at_compile()) {
435 // Open the debugger when compiling this method.
436 tty->print("### Breaking when compiling: ");
437 method()->print_short_name();
438 tty->cr();
439 BREAKPOINT;
440 }
442 if( PrintOpto ) {
443 if (is_osr_compilation()) {
444 tty->print("[OSR]%3d", _compile_id);
445 } else {
446 tty->print("%3d", _compile_id);
447 }
448 }
449 #endif
450 }
453 //-----------------------init_scratch_buffer_blob------------------------------
454 // Construct a temporary BufferBlob and cache it for this compile.
455 void Compile::init_scratch_buffer_blob(int const_size) {
456 // If there is already a scratch buffer blob allocated and the
457 // constant section is big enough, use it. Otherwise free the
458 // current and allocate a new one.
459 BufferBlob* blob = scratch_buffer_blob();
460 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
461 // Use the current blob.
462 } else {
463 if (blob != NULL) {
464 BufferBlob::free(blob);
465 }
467 ResourceMark rm;
468 _scratch_const_size = const_size;
469 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
470 blob = BufferBlob::create("Compile::scratch_buffer", size);
471 // Record the buffer blob for next time.
472 set_scratch_buffer_blob(blob);
473 // Have we run out of code space?
474 if (scratch_buffer_blob() == NULL) {
475 // Let CompilerBroker disable further compilations.
476 record_failure("Not enough space for scratch buffer in CodeCache");
477 return;
478 }
479 }
481 // Initialize the relocation buffers
482 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
483 set_scratch_locs_memory(locs_buf);
484 }
487 //-----------------------scratch_emit_size-------------------------------------
488 // Helper function that computes size by emitting code
489 uint Compile::scratch_emit_size(const Node* n) {
490 // Start scratch_emit_size section.
491 set_in_scratch_emit_size(true);
493 // Emit into a trash buffer and count bytes emitted.
494 // This is a pretty expensive way to compute a size,
495 // but it works well enough if seldom used.
496 // All common fixed-size instructions are given a size
497 // method by the AD file.
498 // Note that the scratch buffer blob and locs memory are
499 // allocated at the beginning of the compile task, and
500 // may be shared by several calls to scratch_emit_size.
501 // The allocation of the scratch buffer blob is particularly
502 // expensive, since it has to grab the code cache lock.
503 BufferBlob* blob = this->scratch_buffer_blob();
504 assert(blob != NULL, "Initialize BufferBlob at start");
505 assert(blob->size() > MAX_inst_size, "sanity");
506 relocInfo* locs_buf = scratch_locs_memory();
507 address blob_begin = blob->content_begin();
508 address blob_end = (address)locs_buf;
509 assert(blob->content_contains(blob_end), "sanity");
510 CodeBuffer buf(blob_begin, blob_end - blob_begin);
511 buf.initialize_consts_size(_scratch_const_size);
512 buf.initialize_stubs_size(MAX_stubs_size);
513 assert(locs_buf != NULL, "sanity");
514 int lsize = MAX_locs_size / 3;
515 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
516 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
517 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
519 // Do the emission.
520 n->emit(buf, this->regalloc());
522 // End scratch_emit_size section.
523 set_in_scratch_emit_size(false);
525 return buf.insts_size();
526 }
529 // ============================================================================
530 //------------------------------Compile standard-------------------------------
531 debug_only( int Compile::_debug_idx = 100000; )
533 // Compile a method. entry_bci is -1 for normal compilations and indicates
534 // the continuation bci for on stack replacement.
537 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
538 : Phase(Compiler),
539 _env(ci_env),
540 _log(ci_env->log()),
541 _compile_id(ci_env->compile_id()),
542 _save_argument_registers(false),
543 _stub_name(NULL),
544 _stub_function(NULL),
545 _stub_entry_point(NULL),
546 _method(target),
547 _entry_bci(osr_bci),
548 _initial_gvn(NULL),
549 _for_igvn(NULL),
550 _warm_calls(NULL),
551 _subsume_loads(subsume_loads),
552 _do_escape_analysis(do_escape_analysis),
553 _failure_reason(NULL),
554 _code_buffer("Compile::Fill_buffer"),
555 _orig_pc_slot(0),
556 _orig_pc_slot_offset_in_bytes(0),
557 _has_method_handle_invokes(false),
558 _mach_constant_base_node(NULL),
559 _node_bundling_limit(0),
560 _node_bundling_base(NULL),
561 _java_calls(0),
562 _inner_loops(0),
563 _scratch_const_size(-1),
564 _in_scratch_emit_size(false),
565 #ifndef PRODUCT
566 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
567 _printer(IdealGraphPrinter::printer()),
568 #endif
569 _congraph(NULL) {
570 C = this;
572 CompileWrapper cw(this);
573 #ifndef PRODUCT
574 if (TimeCompiler2) {
575 tty->print(" ");
576 target->holder()->name()->print();
577 tty->print(".");
578 target->print_short_name();
579 tty->print(" ");
580 }
581 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
582 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
583 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
584 if (!print_opto_assembly) {
585 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
586 if (print_assembly && !Disassembler::can_decode()) {
587 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
588 print_opto_assembly = true;
589 }
590 }
591 set_print_assembly(print_opto_assembly);
592 set_parsed_irreducible_loop(false);
593 #endif
595 if (ProfileTraps) {
596 // Make sure the method being compiled gets its own MDO,
597 // so we can at least track the decompile_count().
598 method()->ensure_method_data();
599 }
601 Init(::AliasLevel);
604 print_compile_messages();
606 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
607 _ilt = InlineTree::build_inline_tree_root();
608 else
609 _ilt = NULL;
611 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
612 assert(num_alias_types() >= AliasIdxRaw, "");
614 #define MINIMUM_NODE_HASH 1023
615 // Node list that Iterative GVN will start with
616 Unique_Node_List for_igvn(comp_arena());
617 set_for_igvn(&for_igvn);
619 // GVN that will be run immediately on new nodes
620 uint estimated_size = method()->code_size()*4+64;
621 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
622 PhaseGVN gvn(node_arena(), estimated_size);
623 set_initial_gvn(&gvn);
625 { // Scope for timing the parser
626 TracePhase t3("parse", &_t_parser, true);
628 // Put top into the hash table ASAP.
629 initial_gvn()->transform_no_reclaim(top());
631 // Set up tf(), start(), and find a CallGenerator.
632 CallGenerator* cg = NULL;
633 if (is_osr_compilation()) {
634 const TypeTuple *domain = StartOSRNode::osr_domain();
635 const TypeTuple *range = TypeTuple::make_range(method()->signature());
636 init_tf(TypeFunc::make(domain, range));
637 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
638 initial_gvn()->set_type_bottom(s);
639 init_start(s);
640 cg = CallGenerator::for_osr(method(), entry_bci());
641 } else {
642 // Normal case.
643 init_tf(TypeFunc::make(method()));
644 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
645 initial_gvn()->set_type_bottom(s);
646 init_start(s);
647 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
648 // With java.lang.ref.reference.get() we must go through the
649 // intrinsic when G1 is enabled - even when get() is the root
650 // method of the compile - so that, if necessary, the value in
651 // the referent field of the reference object gets recorded by
652 // the pre-barrier code.
653 // Specifically, if G1 is enabled, the value in the referent
654 // field is recorded by the G1 SATB pre barrier. This will
655 // result in the referent being marked live and the reference
656 // object removed from the list of discovered references during
657 // reference processing.
658 cg = find_intrinsic(method(), false);
659 }
660 if (cg == NULL) {
661 float past_uses = method()->interpreter_invocation_count();
662 float expected_uses = past_uses;
663 cg = CallGenerator::for_inline(method(), expected_uses);
664 }
665 }
666 if (failing()) return;
667 if (cg == NULL) {
668 record_method_not_compilable_all_tiers("cannot parse method");
669 return;
670 }
671 JVMState* jvms = build_start_state(start(), tf());
672 if ((jvms = cg->generate(jvms)) == NULL) {
673 record_method_not_compilable("method parse failed");
674 return;
675 }
676 GraphKit kit(jvms);
678 if (!kit.stopped()) {
679 // Accept return values, and transfer control we know not where.
680 // This is done by a special, unique ReturnNode bound to root.
681 return_values(kit.jvms());
682 }
684 if (kit.has_exceptions()) {
685 // Any exceptions that escape from this call must be rethrown
686 // to whatever caller is dynamically above us on the stack.
687 // This is done by a special, unique RethrowNode bound to root.
688 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
689 }
691 if (!failing() && has_stringbuilder()) {
692 {
693 // remove useless nodes to make the usage analysis simpler
694 ResourceMark rm;
695 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
696 }
698 {
699 ResourceMark rm;
700 print_method("Before StringOpts", 3);
701 PhaseStringOpts pso(initial_gvn(), &for_igvn);
702 print_method("After StringOpts", 3);
703 }
705 // now inline anything that we skipped the first time around
706 while (_late_inlines.length() > 0) {
707 CallGenerator* cg = _late_inlines.pop();
708 cg->do_late_inline();
709 }
710 }
711 assert(_late_inlines.length() == 0, "should have been processed");
713 print_method("Before RemoveUseless", 3);
715 // Remove clutter produced by parsing.
716 if (!failing()) {
717 ResourceMark rm;
718 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
719 }
720 }
722 // Note: Large methods are capped off in do_one_bytecode().
723 if (failing()) return;
725 // After parsing, node notes are no longer automagic.
726 // They must be propagated by register_new_node_with_optimizer(),
727 // clone(), or the like.
728 set_default_node_notes(NULL);
730 for (;;) {
731 int successes = Inline_Warm();
732 if (failing()) return;
733 if (successes == 0) break;
734 }
736 // Drain the list.
737 Finish_Warm();
738 #ifndef PRODUCT
739 if (_printer) {
740 _printer->print_inlining(this);
741 }
742 #endif
744 if (failing()) return;
745 NOT_PRODUCT( verify_graph_edges(); )
747 // Now optimize
748 Optimize();
749 if (failing()) return;
750 NOT_PRODUCT( verify_graph_edges(); )
752 #ifndef PRODUCT
753 if (PrintIdeal) {
754 ttyLocker ttyl; // keep the following output all in one block
755 // This output goes directly to the tty, not the compiler log.
756 // To enable tools to match it up with the compilation activity,
757 // be sure to tag this tty output with the compile ID.
758 if (xtty != NULL) {
759 xtty->head("ideal compile_id='%d'%s", compile_id(),
760 is_osr_compilation() ? " compile_kind='osr'" :
761 "");
762 }
763 root()->dump(9999);
764 if (xtty != NULL) {
765 xtty->tail("ideal");
766 }
767 }
768 #endif
770 // Now that we know the size of all the monitors we can add a fixed slot
771 // for the original deopt pc.
773 _orig_pc_slot = fixed_slots();
774 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
775 set_fixed_slots(next_slot);
777 // Now generate code
778 Code_Gen();
779 if (failing()) return;
781 // Check if we want to skip execution of all compiled code.
782 {
783 #ifndef PRODUCT
784 if (OptoNoExecute) {
785 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
786 return;
787 }
788 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
789 #endif
791 if (is_osr_compilation()) {
792 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
793 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
794 } else {
795 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
796 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
797 }
799 env()->register_method(_method, _entry_bci,
800 &_code_offsets,
801 _orig_pc_slot_offset_in_bytes,
802 code_buffer(),
803 frame_size_in_words(), _oop_map_set,
804 &_handler_table, &_inc_table,
805 compiler,
806 env()->comp_level(),
807 true, /*has_debug_info*/
808 has_unsafe_access()
809 );
810 }
811 }
813 //------------------------------Compile----------------------------------------
814 // Compile a runtime stub
815 Compile::Compile( ciEnv* ci_env,
816 TypeFunc_generator generator,
817 address stub_function,
818 const char *stub_name,
819 int is_fancy_jump,
820 bool pass_tls,
821 bool save_arg_registers,
822 bool return_pc )
823 : Phase(Compiler),
824 _env(ci_env),
825 _log(ci_env->log()),
826 _compile_id(-1),
827 _save_argument_registers(save_arg_registers),
828 _method(NULL),
829 _stub_name(stub_name),
830 _stub_function(stub_function),
831 _stub_entry_point(NULL),
832 _entry_bci(InvocationEntryBci),
833 _initial_gvn(NULL),
834 _for_igvn(NULL),
835 _warm_calls(NULL),
836 _orig_pc_slot(0),
837 _orig_pc_slot_offset_in_bytes(0),
838 _subsume_loads(true),
839 _do_escape_analysis(false),
840 _failure_reason(NULL),
841 _code_buffer("Compile::Fill_buffer"),
842 _has_method_handle_invokes(false),
843 _mach_constant_base_node(NULL),
844 _node_bundling_limit(0),
845 _node_bundling_base(NULL),
846 _java_calls(0),
847 _inner_loops(0),
848 #ifndef PRODUCT
849 _trace_opto_output(TraceOptoOutput),
850 _printer(NULL),
851 #endif
852 _congraph(NULL) {
853 C = this;
855 #ifndef PRODUCT
856 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
857 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
858 set_print_assembly(PrintFrameConverterAssembly);
859 set_parsed_irreducible_loop(false);
860 #endif
861 CompileWrapper cw(this);
862 Init(/*AliasLevel=*/ 0);
863 init_tf((*generator)());
865 {
866 // The following is a dummy for the sake of GraphKit::gen_stub
867 Unique_Node_List for_igvn(comp_arena());
868 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
869 PhaseGVN gvn(Thread::current()->resource_area(),255);
870 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
871 gvn.transform_no_reclaim(top());
873 GraphKit kit;
874 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
875 }
877 NOT_PRODUCT( verify_graph_edges(); )
878 Code_Gen();
879 if (failing()) return;
882 // Entry point will be accessed using compile->stub_entry_point();
883 if (code_buffer() == NULL) {
884 Matcher::soft_match_failure();
885 } else {
886 if (PrintAssembly && (WizardMode || Verbose))
887 tty->print_cr("### Stub::%s", stub_name);
889 if (!failing()) {
890 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
892 // Make the NMethod
893 // For now we mark the frame as never safe for profile stackwalking
894 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
895 code_buffer(),
896 CodeOffsets::frame_never_safe,
897 // _code_offsets.value(CodeOffsets::Frame_Complete),
898 frame_size_in_words(),
899 _oop_map_set,
900 save_arg_registers);
901 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
903 _stub_entry_point = rs->entry_point();
904 }
905 }
906 }
908 #ifndef PRODUCT
909 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
910 if(PrintOpto && Verbose) {
911 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
912 }
913 }
914 #endif
916 void Compile::print_codes() {
917 }
919 //------------------------------Init-------------------------------------------
920 // Prepare for a single compilation
921 void Compile::Init(int aliaslevel) {
922 _unique = 0;
923 _regalloc = NULL;
925 _tf = NULL; // filled in later
926 _top = NULL; // cached later
927 _matcher = NULL; // filled in later
928 _cfg = NULL; // filled in later
930 set_24_bit_selection_and_mode(Use24BitFP, false);
932 _node_note_array = NULL;
933 _default_node_notes = NULL;
935 _immutable_memory = NULL; // filled in at first inquiry
937 // Globally visible Nodes
938 // First set TOP to NULL to give safe behavior during creation of RootNode
939 set_cached_top_node(NULL);
940 set_root(new (this, 3) RootNode());
941 // Now that you have a Root to point to, create the real TOP
942 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
943 set_recent_alloc(NULL, NULL);
945 // Create Debug Information Recorder to record scopes, oopmaps, etc.
946 env()->set_oop_recorder(new OopRecorder(comp_arena()));
947 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
948 env()->set_dependencies(new Dependencies(env()));
950 _fixed_slots = 0;
951 set_has_split_ifs(false);
952 set_has_loops(has_method() && method()->has_loops()); // first approximation
953 set_has_stringbuilder(false);
954 _trap_can_recompile = false; // no traps emitted yet
955 _major_progress = true; // start out assuming good things will happen
956 set_has_unsafe_access(false);
957 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
958 set_decompile_count(0);
960 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
961 set_num_loop_opts(LoopOptsCount);
962 set_do_inlining(Inline);
963 set_max_inline_size(MaxInlineSize);
964 set_freq_inline_size(FreqInlineSize);
965 set_do_scheduling(OptoScheduling);
966 set_do_count_invocations(false);
967 set_do_method_data_update(false);
969 if (debug_info()->recording_non_safepoints()) {
970 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
971 (comp_arena(), 8, 0, NULL));
972 set_default_node_notes(Node_Notes::make(this));
973 }
975 // // -- Initialize types before each compile --
976 // // Update cached type information
977 // if( _method && _method->constants() )
978 // Type::update_loaded_types(_method, _method->constants());
980 // Init alias_type map.
981 if (!_do_escape_analysis && aliaslevel == 3)
982 aliaslevel = 2; // No unique types without escape analysis
983 _AliasLevel = aliaslevel;
984 const int grow_ats = 16;
985 _max_alias_types = grow_ats;
986 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
987 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
988 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
989 {
990 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
991 }
992 // Initialize the first few types.
993 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
994 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
995 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
996 _num_alias_types = AliasIdxRaw+1;
997 // Zero out the alias type cache.
998 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
999 // A NULL adr_type hits in the cache right away. Preload the right answer.
1000 probe_alias_cache(NULL)->_index = AliasIdxTop;
1002 _intrinsics = NULL;
1003 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1004 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
1005 register_library_intrinsics();
1006 }
1008 //---------------------------init_start----------------------------------------
1009 // Install the StartNode on this compile object.
1010 void Compile::init_start(StartNode* s) {
1011 if (failing())
1012 return; // already failing
1013 assert(s == start(), "");
1014 }
1016 StartNode* Compile::start() const {
1017 assert(!failing(), "");
1018 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1019 Node* start = root()->fast_out(i);
1020 if( start->is_Start() )
1021 return start->as_Start();
1022 }
1023 ShouldNotReachHere();
1024 return NULL;
1025 }
1027 //-------------------------------immutable_memory-------------------------------------
1028 // Access immutable memory
1029 Node* Compile::immutable_memory() {
1030 if (_immutable_memory != NULL) {
1031 return _immutable_memory;
1032 }
1033 StartNode* s = start();
1034 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1035 Node *p = s->fast_out(i);
1036 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1037 _immutable_memory = p;
1038 return _immutable_memory;
1039 }
1040 }
1041 ShouldNotReachHere();
1042 return NULL;
1043 }
1045 //----------------------set_cached_top_node------------------------------------
1046 // Install the cached top node, and make sure Node::is_top works correctly.
1047 void Compile::set_cached_top_node(Node* tn) {
1048 if (tn != NULL) verify_top(tn);
1049 Node* old_top = _top;
1050 _top = tn;
1051 // Calling Node::setup_is_top allows the nodes the chance to adjust
1052 // their _out arrays.
1053 if (_top != NULL) _top->setup_is_top();
1054 if (old_top != NULL) old_top->setup_is_top();
1055 assert(_top == NULL || top()->is_top(), "");
1056 }
1058 #ifndef PRODUCT
1059 void Compile::verify_top(Node* tn) const {
1060 if (tn != NULL) {
1061 assert(tn->is_Con(), "top node must be a constant");
1062 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1063 assert(tn->in(0) != NULL, "must have live top node");
1064 }
1065 }
1066 #endif
1069 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1071 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1072 guarantee(arr != NULL, "");
1073 int num_blocks = arr->length();
1074 if (grow_by < num_blocks) grow_by = num_blocks;
1075 int num_notes = grow_by * _node_notes_block_size;
1076 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1077 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1078 while (num_notes > 0) {
1079 arr->append(notes);
1080 notes += _node_notes_block_size;
1081 num_notes -= _node_notes_block_size;
1082 }
1083 assert(num_notes == 0, "exact multiple, please");
1084 }
1086 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1087 if (source == NULL || dest == NULL) return false;
1089 if (dest->is_Con())
1090 return false; // Do not push debug info onto constants.
1092 #ifdef ASSERT
1093 // Leave a bread crumb trail pointing to the original node:
1094 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1095 dest->set_debug_orig(source);
1096 }
1097 #endif
1099 if (node_note_array() == NULL)
1100 return false; // Not collecting any notes now.
1102 // This is a copy onto a pre-existing node, which may already have notes.
1103 // If both nodes have notes, do not overwrite any pre-existing notes.
1104 Node_Notes* source_notes = node_notes_at(source->_idx);
1105 if (source_notes == NULL || source_notes->is_clear()) return false;
1106 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1107 if (dest_notes == NULL || dest_notes->is_clear()) {
1108 return set_node_notes_at(dest->_idx, source_notes);
1109 }
1111 Node_Notes merged_notes = (*source_notes);
1112 // The order of operations here ensures that dest notes will win...
1113 merged_notes.update_from(dest_notes);
1114 return set_node_notes_at(dest->_idx, &merged_notes);
1115 }
1118 //--------------------------allow_range_check_smearing-------------------------
1119 // Gating condition for coalescing similar range checks.
1120 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1121 // single covering check that is at least as strong as any of them.
1122 // If the optimization succeeds, the simplified (strengthened) range check
1123 // will always succeed. If it fails, we will deopt, and then give up
1124 // on the optimization.
1125 bool Compile::allow_range_check_smearing() const {
1126 // If this method has already thrown a range-check,
1127 // assume it was because we already tried range smearing
1128 // and it failed.
1129 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1130 return !already_trapped;
1131 }
1134 //------------------------------flatten_alias_type-----------------------------
1135 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1136 int offset = tj->offset();
1137 TypePtr::PTR ptr = tj->ptr();
1139 // Known instance (scalarizable allocation) alias only with itself.
1140 bool is_known_inst = tj->isa_oopptr() != NULL &&
1141 tj->is_oopptr()->is_known_instance();
1143 // Process weird unsafe references.
1144 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1145 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1146 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1147 tj = TypeOopPtr::BOTTOM;
1148 ptr = tj->ptr();
1149 offset = tj->offset();
1150 }
1152 // Array pointers need some flattening
1153 const TypeAryPtr *ta = tj->isa_aryptr();
1154 if( ta && is_known_inst ) {
1155 if ( offset != Type::OffsetBot &&
1156 offset > arrayOopDesc::length_offset_in_bytes() ) {
1157 offset = Type::OffsetBot; // Flatten constant access into array body only
1158 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1159 }
1160 } else if( ta && _AliasLevel >= 2 ) {
1161 // For arrays indexed by constant indices, we flatten the alias
1162 // space to include all of the array body. Only the header, klass
1163 // and array length can be accessed un-aliased.
1164 if( offset != Type::OffsetBot ) {
1165 if( ta->const_oop() ) { // methodDataOop or methodOop
1166 offset = Type::OffsetBot; // Flatten constant access into array body
1167 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1168 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1169 // range is OK as-is.
1170 tj = ta = TypeAryPtr::RANGE;
1171 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1172 tj = TypeInstPtr::KLASS; // all klass loads look alike
1173 ta = TypeAryPtr::RANGE; // generic ignored junk
1174 ptr = TypePtr::BotPTR;
1175 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1176 tj = TypeInstPtr::MARK;
1177 ta = TypeAryPtr::RANGE; // generic ignored junk
1178 ptr = TypePtr::BotPTR;
1179 } else { // Random constant offset into array body
1180 offset = Type::OffsetBot; // Flatten constant access into array body
1181 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1182 }
1183 }
1184 // Arrays of fixed size alias with arrays of unknown size.
1185 if (ta->size() != TypeInt::POS) {
1186 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1187 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1188 }
1189 // Arrays of known objects become arrays of unknown objects.
1190 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1191 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1192 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1193 }
1194 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1195 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1196 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1197 }
1198 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1199 // cannot be distinguished by bytecode alone.
1200 if (ta->elem() == TypeInt::BOOL) {
1201 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1202 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1203 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1204 }
1205 // During the 2nd round of IterGVN, NotNull castings are removed.
1206 // Make sure the Bottom and NotNull variants alias the same.
1207 // Also, make sure exact and non-exact variants alias the same.
1208 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1209 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1210 }
1211 }
1213 // Oop pointers need some flattening
1214 const TypeInstPtr *to = tj->isa_instptr();
1215 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1216 ciInstanceKlass *k = to->klass()->as_instance_klass();
1217 if( ptr == TypePtr::Constant ) {
1218 if (to->klass() != ciEnv::current()->Class_klass() ||
1219 offset < k->size_helper() * wordSize) {
1220 // No constant oop pointers (such as Strings); they alias with
1221 // unknown strings.
1222 assert(!is_known_inst, "not scalarizable allocation");
1223 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1224 }
1225 } else if( is_known_inst ) {
1226 tj = to; // Keep NotNull and klass_is_exact for instance type
1227 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1228 // During the 2nd round of IterGVN, NotNull castings are removed.
1229 // Make sure the Bottom and NotNull variants alias the same.
1230 // Also, make sure exact and non-exact variants alias the same.
1231 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1232 }
1233 // Canonicalize the holder of this field
1234 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1235 // First handle header references such as a LoadKlassNode, even if the
1236 // object's klass is unloaded at compile time (4965979).
1237 if (!is_known_inst) { // Do it only for non-instance types
1238 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1239 }
1240 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1241 // Static fields are in the space above the normal instance
1242 // fields in the java.lang.Class instance.
1243 if (to->klass() != ciEnv::current()->Class_klass()) {
1244 to = NULL;
1245 tj = TypeOopPtr::BOTTOM;
1246 offset = tj->offset();
1247 }
1248 } else {
1249 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1250 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1251 if( is_known_inst ) {
1252 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1253 } else {
1254 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1255 }
1256 }
1257 }
1258 }
1260 // Klass pointers to object array klasses need some flattening
1261 const TypeKlassPtr *tk = tj->isa_klassptr();
1262 if( tk ) {
1263 // If we are referencing a field within a Klass, we need
1264 // to assume the worst case of an Object. Both exact and
1265 // inexact types must flatten to the same alias class.
1266 // Since the flattened result for a klass is defined to be
1267 // precisely java.lang.Object, use a constant ptr.
1268 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1270 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1271 TypeKlassPtr::OBJECT->klass(),
1272 offset);
1273 }
1275 ciKlass* klass = tk->klass();
1276 if( klass->is_obj_array_klass() ) {
1277 ciKlass* k = TypeAryPtr::OOPS->klass();
1278 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1279 k = TypeInstPtr::BOTTOM->klass();
1280 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1281 }
1283 // Check for precise loads from the primary supertype array and force them
1284 // to the supertype cache alias index. Check for generic array loads from
1285 // the primary supertype array and also force them to the supertype cache
1286 // alias index. Since the same load can reach both, we need to merge
1287 // these 2 disparate memories into the same alias class. Since the
1288 // primary supertype array is read-only, there's no chance of confusion
1289 // where we bypass an array load and an array store.
1290 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1291 if( offset == Type::OffsetBot ||
1292 off2 < Klass::primary_super_limit()*wordSize ) {
1293 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1294 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1295 }
1296 }
1298 // Flatten all Raw pointers together.
1299 if (tj->base() == Type::RawPtr)
1300 tj = TypeRawPtr::BOTTOM;
1302 if (tj->base() == Type::AnyPtr)
1303 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1305 // Flatten all to bottom for now
1306 switch( _AliasLevel ) {
1307 case 0:
1308 tj = TypePtr::BOTTOM;
1309 break;
1310 case 1: // Flatten to: oop, static, field or array
1311 switch (tj->base()) {
1312 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1313 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1314 case Type::AryPtr: // do not distinguish arrays at all
1315 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1316 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1317 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1318 default: ShouldNotReachHere();
1319 }
1320 break;
1321 case 2: // No collapsing at level 2; keep all splits
1322 case 3: // No collapsing at level 3; keep all splits
1323 break;
1324 default:
1325 Unimplemented();
1326 }
1328 offset = tj->offset();
1329 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1331 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1332 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1333 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1334 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1335 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1336 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1337 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1338 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1339 assert( tj->ptr() != TypePtr::TopPTR &&
1340 tj->ptr() != TypePtr::AnyNull &&
1341 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1342 // assert( tj->ptr() != TypePtr::Constant ||
1343 // tj->base() == Type::RawPtr ||
1344 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1346 return tj;
1347 }
1349 void Compile::AliasType::Init(int i, const TypePtr* at) {
1350 _index = i;
1351 _adr_type = at;
1352 _field = NULL;
1353 _is_rewritable = true; // default
1354 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1355 if (atoop != NULL && atoop->is_known_instance()) {
1356 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1357 _general_index = Compile::current()->get_alias_index(gt);
1358 } else {
1359 _general_index = 0;
1360 }
1361 }
1363 //---------------------------------print_on------------------------------------
1364 #ifndef PRODUCT
1365 void Compile::AliasType::print_on(outputStream* st) {
1366 if (index() < 10)
1367 st->print("@ <%d> ", index());
1368 else st->print("@ <%d>", index());
1369 st->print(is_rewritable() ? " " : " RO");
1370 int offset = adr_type()->offset();
1371 if (offset == Type::OffsetBot)
1372 st->print(" +any");
1373 else st->print(" +%-3d", offset);
1374 st->print(" in ");
1375 adr_type()->dump_on(st);
1376 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1377 if (field() != NULL && tjp) {
1378 if (tjp->klass() != field()->holder() ||
1379 tjp->offset() != field()->offset_in_bytes()) {
1380 st->print(" != ");
1381 field()->print();
1382 st->print(" ***");
1383 }
1384 }
1385 }
1387 void print_alias_types() {
1388 Compile* C = Compile::current();
1389 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1390 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1391 C->alias_type(idx)->print_on(tty);
1392 tty->cr();
1393 }
1394 }
1395 #endif
1398 //----------------------------probe_alias_cache--------------------------------
1399 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1400 intptr_t key = (intptr_t) adr_type;
1401 key ^= key >> logAliasCacheSize;
1402 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1403 }
1406 //-----------------------------grow_alias_types--------------------------------
1407 void Compile::grow_alias_types() {
1408 const int old_ats = _max_alias_types; // how many before?
1409 const int new_ats = old_ats; // how many more?
1410 const int grow_ats = old_ats+new_ats; // how many now?
1411 _max_alias_types = grow_ats;
1412 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1413 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1414 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1415 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1416 }
1419 //--------------------------------find_alias_type------------------------------
1420 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1421 if (_AliasLevel == 0)
1422 return alias_type(AliasIdxBot);
1424 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1425 if (ace->_adr_type == adr_type) {
1426 return alias_type(ace->_index);
1427 }
1429 // Handle special cases.
1430 if (adr_type == NULL) return alias_type(AliasIdxTop);
1431 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1433 // Do it the slow way.
1434 const TypePtr* flat = flatten_alias_type(adr_type);
1436 #ifdef ASSERT
1437 assert(flat == flatten_alias_type(flat), "idempotent");
1438 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1439 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1440 const TypeOopPtr* foop = flat->is_oopptr();
1441 // Scalarizable allocations have exact klass always.
1442 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1443 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1444 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1445 }
1446 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1447 #endif
1449 int idx = AliasIdxTop;
1450 for (int i = 0; i < num_alias_types(); i++) {
1451 if (alias_type(i)->adr_type() == flat) {
1452 idx = i;
1453 break;
1454 }
1455 }
1457 if (idx == AliasIdxTop) {
1458 if (no_create) return NULL;
1459 // Grow the array if necessary.
1460 if (_num_alias_types == _max_alias_types) grow_alias_types();
1461 // Add a new alias type.
1462 idx = _num_alias_types++;
1463 _alias_types[idx]->Init(idx, flat);
1464 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1465 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1466 if (flat->isa_instptr()) {
1467 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1468 && flat->is_instptr()->klass() == env()->Class_klass())
1469 alias_type(idx)->set_rewritable(false);
1470 }
1471 if (flat->isa_klassptr()) {
1472 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1473 alias_type(idx)->set_rewritable(false);
1474 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1475 alias_type(idx)->set_rewritable(false);
1476 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1477 alias_type(idx)->set_rewritable(false);
1478 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1479 alias_type(idx)->set_rewritable(false);
1480 }
1481 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1482 // but the base pointer type is not distinctive enough to identify
1483 // references into JavaThread.)
1485 // Check for final fields.
1486 const TypeInstPtr* tinst = flat->isa_instptr();
1487 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1488 ciField* field;
1489 if (tinst->const_oop() != NULL &&
1490 tinst->klass() == ciEnv::current()->Class_klass() &&
1491 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1492 // static field
1493 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1494 field = k->get_field_by_offset(tinst->offset(), true);
1495 } else {
1496 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1497 field = k->get_field_by_offset(tinst->offset(), false);
1498 }
1499 assert(field == NULL ||
1500 original_field == NULL ||
1501 (field->holder() == original_field->holder() &&
1502 field->offset() == original_field->offset() &&
1503 field->is_static() == original_field->is_static()), "wrong field?");
1504 // Set field() and is_rewritable() attributes.
1505 if (field != NULL) alias_type(idx)->set_field(field);
1506 }
1507 }
1509 // Fill the cache for next time.
1510 ace->_adr_type = adr_type;
1511 ace->_index = idx;
1512 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1514 // Might as well try to fill the cache for the flattened version, too.
1515 AliasCacheEntry* face = probe_alias_cache(flat);
1516 if (face->_adr_type == NULL) {
1517 face->_adr_type = flat;
1518 face->_index = idx;
1519 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1520 }
1522 return alias_type(idx);
1523 }
1526 Compile::AliasType* Compile::alias_type(ciField* field) {
1527 const TypeOopPtr* t;
1528 if (field->is_static())
1529 t = TypeInstPtr::make(field->holder()->java_mirror());
1530 else
1531 t = TypeOopPtr::make_from_klass_raw(field->holder());
1532 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1533 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1534 return atp;
1535 }
1538 //------------------------------have_alias_type--------------------------------
1539 bool Compile::have_alias_type(const TypePtr* adr_type) {
1540 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1541 if (ace->_adr_type == adr_type) {
1542 return true;
1543 }
1545 // Handle special cases.
1546 if (adr_type == NULL) return true;
1547 if (adr_type == TypePtr::BOTTOM) return true;
1549 return find_alias_type(adr_type, true, NULL) != NULL;
1550 }
1552 //-----------------------------must_alias--------------------------------------
1553 // True if all values of the given address type are in the given alias category.
1554 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1555 if (alias_idx == AliasIdxBot) return true; // the universal category
1556 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1557 if (alias_idx == AliasIdxTop) return false; // the empty category
1558 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1560 // the only remaining possible overlap is identity
1561 int adr_idx = get_alias_index(adr_type);
1562 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1563 assert(adr_idx == alias_idx ||
1564 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1565 && adr_type != TypeOopPtr::BOTTOM),
1566 "should not be testing for overlap with an unsafe pointer");
1567 return adr_idx == alias_idx;
1568 }
1570 //------------------------------can_alias--------------------------------------
1571 // True if any values of the given address type are in the given alias category.
1572 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1573 if (alias_idx == AliasIdxTop) return false; // the empty category
1574 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1575 if (alias_idx == AliasIdxBot) return true; // the universal category
1576 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1578 // the only remaining possible overlap is identity
1579 int adr_idx = get_alias_index(adr_type);
1580 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1581 return adr_idx == alias_idx;
1582 }
1586 //---------------------------pop_warm_call-------------------------------------
1587 WarmCallInfo* Compile::pop_warm_call() {
1588 WarmCallInfo* wci = _warm_calls;
1589 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1590 return wci;
1591 }
1593 //----------------------------Inline_Warm--------------------------------------
1594 int Compile::Inline_Warm() {
1595 // If there is room, try to inline some more warm call sites.
1596 // %%% Do a graph index compaction pass when we think we're out of space?
1597 if (!InlineWarmCalls) return 0;
1599 int calls_made_hot = 0;
1600 int room_to_grow = NodeCountInliningCutoff - unique();
1601 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1602 int amount_grown = 0;
1603 WarmCallInfo* call;
1604 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1605 int est_size = (int)call->size();
1606 if (est_size > (room_to_grow - amount_grown)) {
1607 // This one won't fit anyway. Get rid of it.
1608 call->make_cold();
1609 continue;
1610 }
1611 call->make_hot();
1612 calls_made_hot++;
1613 amount_grown += est_size;
1614 amount_to_grow -= est_size;
1615 }
1617 if (calls_made_hot > 0) set_major_progress();
1618 return calls_made_hot;
1619 }
1622 //----------------------------Finish_Warm--------------------------------------
1623 void Compile::Finish_Warm() {
1624 if (!InlineWarmCalls) return;
1625 if (failing()) return;
1626 if (warm_calls() == NULL) return;
1628 // Clean up loose ends, if we are out of space for inlining.
1629 WarmCallInfo* call;
1630 while ((call = pop_warm_call()) != NULL) {
1631 call->make_cold();
1632 }
1633 }
1635 //---------------------cleanup_loop_predicates-----------------------
1636 // Remove the opaque nodes that protect the predicates so that all unused
1637 // checks and uncommon_traps will be eliminated from the ideal graph
1638 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1639 if (predicate_count()==0) return;
1640 for (int i = predicate_count(); i > 0; i--) {
1641 Node * n = predicate_opaque1_node(i-1);
1642 assert(n->Opcode() == Op_Opaque1, "must be");
1643 igvn.replace_node(n, n->in(1));
1644 }
1645 assert(predicate_count()==0, "should be clean!");
1646 }
1648 //------------------------------Optimize---------------------------------------
1649 // Given a graph, optimize it.
1650 void Compile::Optimize() {
1651 TracePhase t1("optimizer", &_t_optimizer, true);
1653 #ifndef PRODUCT
1654 if (env()->break_at_compile()) {
1655 BREAKPOINT;
1656 }
1658 #endif
1660 ResourceMark rm;
1661 int loop_opts_cnt;
1663 NOT_PRODUCT( verify_graph_edges(); )
1665 print_method("After Parsing");
1667 {
1668 // Iterative Global Value Numbering, including ideal transforms
1669 // Initialize IterGVN with types and values from parse-time GVN
1670 PhaseIterGVN igvn(initial_gvn());
1671 {
1672 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1673 igvn.optimize();
1674 }
1676 print_method("Iter GVN 1", 2);
1678 if (failing()) return;
1680 // Perform escape analysis
1681 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1682 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
1683 ConnectionGraph::do_analysis(this, &igvn);
1685 if (failing()) return;
1687 igvn.optimize();
1688 print_method("Iter GVN 3", 2);
1690 if (failing()) return;
1692 }
1694 // Loop transforms on the ideal graph. Range Check Elimination,
1695 // peeling, unrolling, etc.
1697 // Set loop opts counter
1698 loop_opts_cnt = num_loop_opts();
1699 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1700 {
1701 TracePhase t2("idealLoop", &_t_idealLoop, true);
1702 PhaseIdealLoop ideal_loop( igvn, true );
1703 loop_opts_cnt--;
1704 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1705 if (failing()) return;
1706 }
1707 // Loop opts pass if partial peeling occurred in previous pass
1708 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1709 TracePhase t3("idealLoop", &_t_idealLoop, true);
1710 PhaseIdealLoop ideal_loop( igvn, false );
1711 loop_opts_cnt--;
1712 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1713 if (failing()) return;
1714 }
1715 // Loop opts pass for loop-unrolling before CCP
1716 if(major_progress() && (loop_opts_cnt > 0)) {
1717 TracePhase t4("idealLoop", &_t_idealLoop, true);
1718 PhaseIdealLoop ideal_loop( igvn, false );
1719 loop_opts_cnt--;
1720 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1721 }
1722 if (!failing()) {
1723 // Verify that last round of loop opts produced a valid graph
1724 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1725 PhaseIdealLoop::verify(igvn);
1726 }
1727 }
1728 if (failing()) return;
1730 // Conditional Constant Propagation;
1731 PhaseCCP ccp( &igvn );
1732 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1733 {
1734 TracePhase t2("ccp", &_t_ccp, true);
1735 ccp.do_transform();
1736 }
1737 print_method("PhaseCPP 1", 2);
1739 assert( true, "Break here to ccp.dump_old2new_map()");
1741 // Iterative Global Value Numbering, including ideal transforms
1742 {
1743 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1744 igvn = ccp;
1745 igvn.optimize();
1746 }
1748 print_method("Iter GVN 2", 2);
1750 if (failing()) return;
1752 // Loop transforms on the ideal graph. Range Check Elimination,
1753 // peeling, unrolling, etc.
1754 if(loop_opts_cnt > 0) {
1755 debug_only( int cnt = 0; );
1756 while(major_progress() && (loop_opts_cnt > 0)) {
1757 TracePhase t2("idealLoop", &_t_idealLoop, true);
1758 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1759 PhaseIdealLoop ideal_loop( igvn, true);
1760 loop_opts_cnt--;
1761 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1762 if (failing()) return;
1763 }
1764 }
1766 {
1767 // Verify that all previous optimizations produced a valid graph
1768 // at least to this point, even if no loop optimizations were done.
1769 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1770 PhaseIdealLoop::verify(igvn);
1771 }
1773 {
1774 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1775 PhaseMacroExpand mex(igvn);
1776 if (mex.expand_macro_nodes()) {
1777 assert(failing(), "must bail out w/ explicit message");
1778 return;
1779 }
1780 }
1782 } // (End scope of igvn; run destructor if necessary for asserts.)
1784 // A method with only infinite loops has no edges entering loops from root
1785 {
1786 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1787 if (final_graph_reshaping()) {
1788 assert(failing(), "must bail out w/ explicit message");
1789 return;
1790 }
1791 }
1793 print_method("Optimize finished", 2);
1794 }
1797 //------------------------------Code_Gen---------------------------------------
1798 // Given a graph, generate code for it
1799 void Compile::Code_Gen() {
1800 if (failing()) return;
1802 // Perform instruction selection. You might think we could reclaim Matcher
1803 // memory PDQ, but actually the Matcher is used in generating spill code.
1804 // Internals of the Matcher (including some VectorSets) must remain live
1805 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1806 // set a bit in reclaimed memory.
1808 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1809 // nodes. Mapping is only valid at the root of each matched subtree.
1810 NOT_PRODUCT( verify_graph_edges(); )
1812 Node_List proj_list;
1813 Matcher m(proj_list);
1814 _matcher = &m;
1815 {
1816 TracePhase t2("matcher", &_t_matcher, true);
1817 m.match();
1818 }
1819 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1820 // nodes. Mapping is only valid at the root of each matched subtree.
1821 NOT_PRODUCT( verify_graph_edges(); )
1823 // If you have too many nodes, or if matching has failed, bail out
1824 check_node_count(0, "out of nodes matching instructions");
1825 if (failing()) return;
1827 // Build a proper-looking CFG
1828 PhaseCFG cfg(node_arena(), root(), m);
1829 _cfg = &cfg;
1830 {
1831 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1832 cfg.Dominators();
1833 if (failing()) return;
1835 NOT_PRODUCT( verify_graph_edges(); )
1837 cfg.Estimate_Block_Frequency();
1838 cfg.GlobalCodeMotion(m,unique(),proj_list);
1840 print_method("Global code motion", 2);
1842 if (failing()) return;
1843 NOT_PRODUCT( verify_graph_edges(); )
1845 debug_only( cfg.verify(); )
1846 }
1847 NOT_PRODUCT( verify_graph_edges(); )
1849 PhaseChaitin regalloc(unique(),cfg,m);
1850 _regalloc = ®alloc;
1851 {
1852 TracePhase t2("regalloc", &_t_registerAllocation, true);
1853 // Perform any platform dependent preallocation actions. This is used,
1854 // for example, to avoid taking an implicit null pointer exception
1855 // using the frame pointer on win95.
1856 _regalloc->pd_preallocate_hook();
1858 // Perform register allocation. After Chaitin, use-def chains are
1859 // no longer accurate (at spill code) and so must be ignored.
1860 // Node->LRG->reg mappings are still accurate.
1861 _regalloc->Register_Allocate();
1863 // Bail out if the allocator builds too many nodes
1864 if (failing()) return;
1865 }
1867 // Prior to register allocation we kept empty basic blocks in case the
1868 // the allocator needed a place to spill. After register allocation we
1869 // are not adding any new instructions. If any basic block is empty, we
1870 // can now safely remove it.
1871 {
1872 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1873 cfg.remove_empty();
1874 if (do_freq_based_layout()) {
1875 PhaseBlockLayout layout(cfg);
1876 } else {
1877 cfg.set_loop_alignment();
1878 }
1879 cfg.fixup_flow();
1880 }
1882 // Perform any platform dependent postallocation verifications.
1883 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1885 // Apply peephole optimizations
1886 if( OptoPeephole ) {
1887 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1888 PhasePeephole peep( _regalloc, cfg);
1889 peep.do_transform();
1890 }
1892 // Convert Nodes to instruction bits in a buffer
1893 {
1894 // %%%% workspace merge brought two timers together for one job
1895 TracePhase t2a("output", &_t_output, true);
1896 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1897 Output();
1898 }
1900 print_method("Final Code");
1902 // He's dead, Jim.
1903 _cfg = (PhaseCFG*)0xdeadbeef;
1904 _regalloc = (PhaseChaitin*)0xdeadbeef;
1905 }
1908 //------------------------------dump_asm---------------------------------------
1909 // Dump formatted assembly
1910 #ifndef PRODUCT
1911 void Compile::dump_asm(int *pcs, uint pc_limit) {
1912 bool cut_short = false;
1913 tty->print_cr("#");
1914 tty->print("# "); _tf->dump(); tty->cr();
1915 tty->print_cr("#");
1917 // For all blocks
1918 int pc = 0x0; // Program counter
1919 char starts_bundle = ' ';
1920 _regalloc->dump_frame();
1922 Node *n = NULL;
1923 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1924 if (VMThread::should_terminate()) { cut_short = true; break; }
1925 Block *b = _cfg->_blocks[i];
1926 if (b->is_connector() && !Verbose) continue;
1927 n = b->_nodes[0];
1928 if (pcs && n->_idx < pc_limit)
1929 tty->print("%3.3x ", pcs[n->_idx]);
1930 else
1931 tty->print(" ");
1932 b->dump_head( &_cfg->_bbs );
1933 if (b->is_connector()) {
1934 tty->print_cr(" # Empty connector block");
1935 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1936 tty->print_cr(" # Block is sole successor of call");
1937 }
1939 // For all instructions
1940 Node *delay = NULL;
1941 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1942 if (VMThread::should_terminate()) { cut_short = true; break; }
1943 n = b->_nodes[j];
1944 if (valid_bundle_info(n)) {
1945 Bundle *bundle = node_bundling(n);
1946 if (bundle->used_in_unconditional_delay()) {
1947 delay = n;
1948 continue;
1949 }
1950 if (bundle->starts_bundle())
1951 starts_bundle = '+';
1952 }
1954 if (WizardMode) n->dump();
1956 if( !n->is_Region() && // Dont print in the Assembly
1957 !n->is_Phi() && // a few noisely useless nodes
1958 !n->is_Proj() &&
1959 !n->is_MachTemp() &&
1960 !n->is_SafePointScalarObject() &&
1961 !n->is_Catch() && // Would be nice to print exception table targets
1962 !n->is_MergeMem() && // Not very interesting
1963 !n->is_top() && // Debug info table constants
1964 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1965 ) {
1966 if (pcs && n->_idx < pc_limit)
1967 tty->print("%3.3x", pcs[n->_idx]);
1968 else
1969 tty->print(" ");
1970 tty->print(" %c ", starts_bundle);
1971 starts_bundle = ' ';
1972 tty->print("\t");
1973 n->format(_regalloc, tty);
1974 tty->cr();
1975 }
1977 // If we have an instruction with a delay slot, and have seen a delay,
1978 // then back up and print it
1979 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1980 assert(delay != NULL, "no unconditional delay instruction");
1981 if (WizardMode) delay->dump();
1983 if (node_bundling(delay)->starts_bundle())
1984 starts_bundle = '+';
1985 if (pcs && n->_idx < pc_limit)
1986 tty->print("%3.3x", pcs[n->_idx]);
1987 else
1988 tty->print(" ");
1989 tty->print(" %c ", starts_bundle);
1990 starts_bundle = ' ';
1991 tty->print("\t");
1992 delay->format(_regalloc, tty);
1993 tty->print_cr("");
1994 delay = NULL;
1995 }
1997 // Dump the exception table as well
1998 if( n->is_Catch() && (Verbose || WizardMode) ) {
1999 // Print the exception table for this offset
2000 _handler_table.print_subtable_for(pc);
2001 }
2002 }
2004 if (pcs && n->_idx < pc_limit)
2005 tty->print_cr("%3.3x", pcs[n->_idx]);
2006 else
2007 tty->print_cr("");
2009 assert(cut_short || delay == NULL, "no unconditional delay branch");
2011 } // End of per-block dump
2012 tty->print_cr("");
2014 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2015 }
2016 #endif
2018 //------------------------------Final_Reshape_Counts---------------------------
2019 // This class defines counters to help identify when a method
2020 // may/must be executed using hardware with only 24-bit precision.
2021 struct Final_Reshape_Counts : public StackObj {
2022 int _call_count; // count non-inlined 'common' calls
2023 int _float_count; // count float ops requiring 24-bit precision
2024 int _double_count; // count double ops requiring more precision
2025 int _java_call_count; // count non-inlined 'java' calls
2026 int _inner_loop_count; // count loops which need alignment
2027 VectorSet _visited; // Visitation flags
2028 Node_List _tests; // Set of IfNodes & PCTableNodes
2030 Final_Reshape_Counts() :
2031 _call_count(0), _float_count(0), _double_count(0),
2032 _java_call_count(0), _inner_loop_count(0),
2033 _visited( Thread::current()->resource_area() ) { }
2035 void inc_call_count () { _call_count ++; }
2036 void inc_float_count () { _float_count ++; }
2037 void inc_double_count() { _double_count++; }
2038 void inc_java_call_count() { _java_call_count++; }
2039 void inc_inner_loop_count() { _inner_loop_count++; }
2041 int get_call_count () const { return _call_count ; }
2042 int get_float_count () const { return _float_count ; }
2043 int get_double_count() const { return _double_count; }
2044 int get_java_call_count() const { return _java_call_count; }
2045 int get_inner_loop_count() const { return _inner_loop_count; }
2046 };
2048 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2049 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2050 // Make sure the offset goes inside the instance layout.
2051 return k->contains_field_offset(tp->offset());
2052 // Note that OffsetBot and OffsetTop are very negative.
2053 }
2055 // Eliminate trivially redundant StoreCMs and accumulate their
2056 // precedence edges.
2057 static void eliminate_redundant_card_marks(Node* n) {
2058 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2059 if (n->in(MemNode::Address)->outcnt() > 1) {
2060 // There are multiple users of the same address so it might be
2061 // possible to eliminate some of the StoreCMs
2062 Node* mem = n->in(MemNode::Memory);
2063 Node* adr = n->in(MemNode::Address);
2064 Node* val = n->in(MemNode::ValueIn);
2065 Node* prev = n;
2066 bool done = false;
2067 // Walk the chain of StoreCMs eliminating ones that match. As
2068 // long as it's a chain of single users then the optimization is
2069 // safe. Eliminating partially redundant StoreCMs would require
2070 // cloning copies down the other paths.
2071 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2072 if (adr == mem->in(MemNode::Address) &&
2073 val == mem->in(MemNode::ValueIn)) {
2074 // redundant StoreCM
2075 if (mem->req() > MemNode::OopStore) {
2076 // Hasn't been processed by this code yet.
2077 n->add_prec(mem->in(MemNode::OopStore));
2078 } else {
2079 // Already converted to precedence edge
2080 for (uint i = mem->req(); i < mem->len(); i++) {
2081 // Accumulate any precedence edges
2082 if (mem->in(i) != NULL) {
2083 n->add_prec(mem->in(i));
2084 }
2085 }
2086 // Everything above this point has been processed.
2087 done = true;
2088 }
2089 // Eliminate the previous StoreCM
2090 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2091 assert(mem->outcnt() == 0, "should be dead");
2092 mem->disconnect_inputs(NULL);
2093 } else {
2094 prev = mem;
2095 }
2096 mem = prev->in(MemNode::Memory);
2097 }
2098 }
2099 }
2101 //------------------------------final_graph_reshaping_impl----------------------
2102 // Implement items 1-5 from final_graph_reshaping below.
2103 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
2105 if ( n->outcnt() == 0 ) return; // dead node
2106 uint nop = n->Opcode();
2108 // Check for 2-input instruction with "last use" on right input.
2109 // Swap to left input. Implements item (2).
2110 if( n->req() == 3 && // two-input instruction
2111 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2112 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2113 n->in(2)->outcnt() == 1 &&// right use IS a last use
2114 !n->in(2)->is_Con() ) { // right use is not a constant
2115 // Check for commutative opcode
2116 switch( nop ) {
2117 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2118 case Op_MaxI: case Op_MinI:
2119 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2120 case Op_AndL: case Op_XorL: case Op_OrL:
2121 case Op_AndI: case Op_XorI: case Op_OrI: {
2122 // Move "last use" input to left by swapping inputs
2123 n->swap_edges(1, 2);
2124 break;
2125 }
2126 default:
2127 break;
2128 }
2129 }
2131 #ifdef ASSERT
2132 if( n->is_Mem() ) {
2133 Compile* C = Compile::current();
2134 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
2135 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2136 // oop will be recorded in oop map if load crosses safepoint
2137 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2138 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2139 "raw memory operations should have control edge");
2140 }
2141 #endif
2142 // Count FPU ops and common calls, implements item (3)
2143 switch( nop ) {
2144 // Count all float operations that may use FPU
2145 case Op_AddF:
2146 case Op_SubF:
2147 case Op_MulF:
2148 case Op_DivF:
2149 case Op_NegF:
2150 case Op_ModF:
2151 case Op_ConvI2F:
2152 case Op_ConF:
2153 case Op_CmpF:
2154 case Op_CmpF3:
2155 // case Op_ConvL2F: // longs are split into 32-bit halves
2156 frc.inc_float_count();
2157 break;
2159 case Op_ConvF2D:
2160 case Op_ConvD2F:
2161 frc.inc_float_count();
2162 frc.inc_double_count();
2163 break;
2165 // Count all double operations that may use FPU
2166 case Op_AddD:
2167 case Op_SubD:
2168 case Op_MulD:
2169 case Op_DivD:
2170 case Op_NegD:
2171 case Op_ModD:
2172 case Op_ConvI2D:
2173 case Op_ConvD2I:
2174 // case Op_ConvL2D: // handled by leaf call
2175 // case Op_ConvD2L: // handled by leaf call
2176 case Op_ConD:
2177 case Op_CmpD:
2178 case Op_CmpD3:
2179 frc.inc_double_count();
2180 break;
2181 case Op_Opaque1: // Remove Opaque Nodes before matching
2182 case Op_Opaque2: // Remove Opaque Nodes before matching
2183 n->subsume_by(n->in(1));
2184 break;
2185 case Op_CallStaticJava:
2186 case Op_CallJava:
2187 case Op_CallDynamicJava:
2188 frc.inc_java_call_count(); // Count java call site;
2189 case Op_CallRuntime:
2190 case Op_CallLeaf:
2191 case Op_CallLeafNoFP: {
2192 assert( n->is_Call(), "" );
2193 CallNode *call = n->as_Call();
2194 // Count call sites where the FP mode bit would have to be flipped.
2195 // Do not count uncommon runtime calls:
2196 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2197 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2198 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2199 frc.inc_call_count(); // Count the call site
2200 } else { // See if uncommon argument is shared
2201 Node *n = call->in(TypeFunc::Parms);
2202 int nop = n->Opcode();
2203 // Clone shared simple arguments to uncommon calls, item (1).
2204 if( n->outcnt() > 1 &&
2205 !n->is_Proj() &&
2206 nop != Op_CreateEx &&
2207 nop != Op_CheckCastPP &&
2208 nop != Op_DecodeN &&
2209 !n->is_Mem() ) {
2210 Node *x = n->clone();
2211 call->set_req( TypeFunc::Parms, x );
2212 }
2213 }
2214 break;
2215 }
2217 case Op_StoreD:
2218 case Op_LoadD:
2219 case Op_LoadD_unaligned:
2220 frc.inc_double_count();
2221 goto handle_mem;
2222 case Op_StoreF:
2223 case Op_LoadF:
2224 frc.inc_float_count();
2225 goto handle_mem;
2227 case Op_StoreCM:
2228 {
2229 // Convert OopStore dependence into precedence edge
2230 Node* prec = n->in(MemNode::OopStore);
2231 n->del_req(MemNode::OopStore);
2232 n->add_prec(prec);
2233 eliminate_redundant_card_marks(n);
2234 }
2236 // fall through
2238 case Op_StoreB:
2239 case Op_StoreC:
2240 case Op_StorePConditional:
2241 case Op_StoreI:
2242 case Op_StoreL:
2243 case Op_StoreIConditional:
2244 case Op_StoreLConditional:
2245 case Op_CompareAndSwapI:
2246 case Op_CompareAndSwapL:
2247 case Op_CompareAndSwapP:
2248 case Op_CompareAndSwapN:
2249 case Op_StoreP:
2250 case Op_StoreN:
2251 case Op_LoadB:
2252 case Op_LoadUB:
2253 case Op_LoadUS:
2254 case Op_LoadI:
2255 case Op_LoadUI2L:
2256 case Op_LoadKlass:
2257 case Op_LoadNKlass:
2258 case Op_LoadL:
2259 case Op_LoadL_unaligned:
2260 case Op_LoadPLocked:
2261 case Op_LoadLLocked:
2262 case Op_LoadP:
2263 case Op_LoadN:
2264 case Op_LoadRange:
2265 case Op_LoadS: {
2266 handle_mem:
2267 #ifdef ASSERT
2268 if( VerifyOptoOopOffsets ) {
2269 assert( n->is_Mem(), "" );
2270 MemNode *mem = (MemNode*)n;
2271 // Check to see if address types have grounded out somehow.
2272 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2273 assert( !tp || oop_offset_is_sane(tp), "" );
2274 }
2275 #endif
2276 break;
2277 }
2279 case Op_AddP: { // Assert sane base pointers
2280 Node *addp = n->in(AddPNode::Address);
2281 assert( !addp->is_AddP() ||
2282 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2283 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2284 "Base pointers must match" );
2285 #ifdef _LP64
2286 if (UseCompressedOops &&
2287 addp->Opcode() == Op_ConP &&
2288 addp == n->in(AddPNode::Base) &&
2289 n->in(AddPNode::Offset)->is_Con()) {
2290 // Use addressing with narrow klass to load with offset on x86.
2291 // On sparc loading 32-bits constant and decoding it have less
2292 // instructions (4) then load 64-bits constant (7).
2293 // Do this transformation here since IGVN will convert ConN back to ConP.
2294 const Type* t = addp->bottom_type();
2295 if (t->isa_oopptr()) {
2296 Node* nn = NULL;
2298 // Look for existing ConN node of the same exact type.
2299 Compile* C = Compile::current();
2300 Node* r = C->root();
2301 uint cnt = r->outcnt();
2302 for (uint i = 0; i < cnt; i++) {
2303 Node* m = r->raw_out(i);
2304 if (m!= NULL && m->Opcode() == Op_ConN &&
2305 m->bottom_type()->make_ptr() == t) {
2306 nn = m;
2307 break;
2308 }
2309 }
2310 if (nn != NULL) {
2311 // Decode a narrow oop to match address
2312 // [R12 + narrow_oop_reg<<3 + offset]
2313 nn = new (C, 2) DecodeNNode(nn, t);
2314 n->set_req(AddPNode::Base, nn);
2315 n->set_req(AddPNode::Address, nn);
2316 if (addp->outcnt() == 0) {
2317 addp->disconnect_inputs(NULL);
2318 }
2319 }
2320 }
2321 }
2322 #endif
2323 break;
2324 }
2326 #ifdef _LP64
2327 case Op_CastPP:
2328 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2329 Compile* C = Compile::current();
2330 Node* in1 = n->in(1);
2331 const Type* t = n->bottom_type();
2332 Node* new_in1 = in1->clone();
2333 new_in1->as_DecodeN()->set_type(t);
2335 if (!Matcher::narrow_oop_use_complex_address()) {
2336 //
2337 // x86, ARM and friends can handle 2 adds in addressing mode
2338 // and Matcher can fold a DecodeN node into address by using
2339 // a narrow oop directly and do implicit NULL check in address:
2340 //
2341 // [R12 + narrow_oop_reg<<3 + offset]
2342 // NullCheck narrow_oop_reg
2343 //
2344 // On other platforms (Sparc) we have to keep new DecodeN node and
2345 // use it to do implicit NULL check in address:
2346 //
2347 // decode_not_null narrow_oop_reg, base_reg
2348 // [base_reg + offset]
2349 // NullCheck base_reg
2350 //
2351 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2352 // to keep the information to which NULL check the new DecodeN node
2353 // corresponds to use it as value in implicit_null_check().
2354 //
2355 new_in1->set_req(0, n->in(0));
2356 }
2358 n->subsume_by(new_in1);
2359 if (in1->outcnt() == 0) {
2360 in1->disconnect_inputs(NULL);
2361 }
2362 }
2363 break;
2365 case Op_CmpP:
2366 // Do this transformation here to preserve CmpPNode::sub() and
2367 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2368 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2369 Node* in1 = n->in(1);
2370 Node* in2 = n->in(2);
2371 if (!in1->is_DecodeN()) {
2372 in2 = in1;
2373 in1 = n->in(2);
2374 }
2375 assert(in1->is_DecodeN(), "sanity");
2377 Compile* C = Compile::current();
2378 Node* new_in2 = NULL;
2379 if (in2->is_DecodeN()) {
2380 new_in2 = in2->in(1);
2381 } else if (in2->Opcode() == Op_ConP) {
2382 const Type* t = in2->bottom_type();
2383 if (t == TypePtr::NULL_PTR) {
2384 // Don't convert CmpP null check into CmpN if compressed
2385 // oops implicit null check is not generated.
2386 // This will allow to generate normal oop implicit null check.
2387 if (Matcher::gen_narrow_oop_implicit_null_checks())
2388 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2389 //
2390 // This transformation together with CastPP transformation above
2391 // will generated code for implicit NULL checks for compressed oops.
2392 //
2393 // The original code after Optimize()
2394 //
2395 // LoadN memory, narrow_oop_reg
2396 // decode narrow_oop_reg, base_reg
2397 // CmpP base_reg, NULL
2398 // CastPP base_reg // NotNull
2399 // Load [base_reg + offset], val_reg
2400 //
2401 // after these transformations will be
2402 //
2403 // LoadN memory, narrow_oop_reg
2404 // CmpN narrow_oop_reg, NULL
2405 // decode_not_null narrow_oop_reg, base_reg
2406 // Load [base_reg + offset], val_reg
2407 //
2408 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2409 // since narrow oops can be used in debug info now (see the code in
2410 // final_graph_reshaping_walk()).
2411 //
2412 // At the end the code will be matched to
2413 // on x86:
2414 //
2415 // Load_narrow_oop memory, narrow_oop_reg
2416 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2417 // NullCheck narrow_oop_reg
2418 //
2419 // and on sparc:
2420 //
2421 // Load_narrow_oop memory, narrow_oop_reg
2422 // decode_not_null narrow_oop_reg, base_reg
2423 // Load [base_reg + offset], val_reg
2424 // NullCheck base_reg
2425 //
2426 } else if (t->isa_oopptr()) {
2427 new_in2 = ConNode::make(C, t->make_narrowoop());
2428 }
2429 }
2430 if (new_in2 != NULL) {
2431 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2432 n->subsume_by( cmpN );
2433 if (in1->outcnt() == 0) {
2434 in1->disconnect_inputs(NULL);
2435 }
2436 if (in2->outcnt() == 0) {
2437 in2->disconnect_inputs(NULL);
2438 }
2439 }
2440 }
2441 break;
2443 case Op_DecodeN:
2444 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2445 // DecodeN could be pinned when it can't be fold into
2446 // an address expression, see the code for Op_CastPP above.
2447 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2448 break;
2450 case Op_EncodeP: {
2451 Node* in1 = n->in(1);
2452 if (in1->is_DecodeN()) {
2453 n->subsume_by(in1->in(1));
2454 } else if (in1->Opcode() == Op_ConP) {
2455 Compile* C = Compile::current();
2456 const Type* t = in1->bottom_type();
2457 if (t == TypePtr::NULL_PTR) {
2458 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2459 } else if (t->isa_oopptr()) {
2460 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2461 }
2462 }
2463 if (in1->outcnt() == 0) {
2464 in1->disconnect_inputs(NULL);
2465 }
2466 break;
2467 }
2469 case Op_Proj: {
2470 if (OptimizeStringConcat) {
2471 ProjNode* p = n->as_Proj();
2472 if (p->_is_io_use) {
2473 // Separate projections were used for the exception path which
2474 // are normally removed by a late inline. If it wasn't inlined
2475 // then they will hang around and should just be replaced with
2476 // the original one.
2477 Node* proj = NULL;
2478 // Replace with just one
2479 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2480 Node *use = i.get();
2481 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2482 proj = use;
2483 break;
2484 }
2485 }
2486 assert(p != NULL, "must be found");
2487 p->subsume_by(proj);
2488 }
2489 }
2490 break;
2491 }
2493 case Op_Phi:
2494 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2495 // The EncodeP optimization may create Phi with the same edges
2496 // for all paths. It is not handled well by Register Allocator.
2497 Node* unique_in = n->in(1);
2498 assert(unique_in != NULL, "");
2499 uint cnt = n->req();
2500 for (uint i = 2; i < cnt; i++) {
2501 Node* m = n->in(i);
2502 assert(m != NULL, "");
2503 if (unique_in != m)
2504 unique_in = NULL;
2505 }
2506 if (unique_in != NULL) {
2507 n->subsume_by(unique_in);
2508 }
2509 }
2510 break;
2512 #endif
2514 case Op_ModI:
2515 if (UseDivMod) {
2516 // Check if a%b and a/b both exist
2517 Node* d = n->find_similar(Op_DivI);
2518 if (d) {
2519 // Replace them with a fused divmod if supported
2520 Compile* C = Compile::current();
2521 if (Matcher::has_match_rule(Op_DivModI)) {
2522 DivModINode* divmod = DivModINode::make(C, n);
2523 d->subsume_by(divmod->div_proj());
2524 n->subsume_by(divmod->mod_proj());
2525 } else {
2526 // replace a%b with a-((a/b)*b)
2527 Node* mult = new (C, 3) MulINode(d, d->in(2));
2528 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2529 n->subsume_by( sub );
2530 }
2531 }
2532 }
2533 break;
2535 case Op_ModL:
2536 if (UseDivMod) {
2537 // Check if a%b and a/b both exist
2538 Node* d = n->find_similar(Op_DivL);
2539 if (d) {
2540 // Replace them with a fused divmod if supported
2541 Compile* C = Compile::current();
2542 if (Matcher::has_match_rule(Op_DivModL)) {
2543 DivModLNode* divmod = DivModLNode::make(C, n);
2544 d->subsume_by(divmod->div_proj());
2545 n->subsume_by(divmod->mod_proj());
2546 } else {
2547 // replace a%b with a-((a/b)*b)
2548 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2549 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2550 n->subsume_by( sub );
2551 }
2552 }
2553 }
2554 break;
2556 case Op_Load16B:
2557 case Op_Load8B:
2558 case Op_Load4B:
2559 case Op_Load8S:
2560 case Op_Load4S:
2561 case Op_Load2S:
2562 case Op_Load8C:
2563 case Op_Load4C:
2564 case Op_Load2C:
2565 case Op_Load4I:
2566 case Op_Load2I:
2567 case Op_Load2L:
2568 case Op_Load4F:
2569 case Op_Load2F:
2570 case Op_Load2D:
2571 case Op_Store16B:
2572 case Op_Store8B:
2573 case Op_Store4B:
2574 case Op_Store8C:
2575 case Op_Store4C:
2576 case Op_Store2C:
2577 case Op_Store4I:
2578 case Op_Store2I:
2579 case Op_Store2L:
2580 case Op_Store4F:
2581 case Op_Store2F:
2582 case Op_Store2D:
2583 break;
2585 case Op_PackB:
2586 case Op_PackS:
2587 case Op_PackC:
2588 case Op_PackI:
2589 case Op_PackF:
2590 case Op_PackL:
2591 case Op_PackD:
2592 if (n->req()-1 > 2) {
2593 // Replace many operand PackNodes with a binary tree for matching
2594 PackNode* p = (PackNode*) n;
2595 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2596 n->subsume_by(btp);
2597 }
2598 break;
2599 case Op_Loop:
2600 case Op_CountedLoop:
2601 if (n->as_Loop()->is_inner_loop()) {
2602 frc.inc_inner_loop_count();
2603 }
2604 break;
2605 case Op_LShiftI:
2606 case Op_RShiftI:
2607 case Op_URShiftI:
2608 case Op_LShiftL:
2609 case Op_RShiftL:
2610 case Op_URShiftL:
2611 if (Matcher::need_masked_shift_count) {
2612 // The cpu's shift instructions don't restrict the count to the
2613 // lower 5/6 bits. We need to do the masking ourselves.
2614 Node* in2 = n->in(2);
2615 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2616 const TypeInt* t = in2->find_int_type();
2617 if (t != NULL && t->is_con()) {
2618 juint shift = t->get_con();
2619 if (shift > mask) { // Unsigned cmp
2620 Compile* C = Compile::current();
2621 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask)));
2622 }
2623 } else {
2624 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2625 Compile* C = Compile::current();
2626 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask)));
2627 n->set_req(2, shift);
2628 }
2629 }
2630 if (in2->outcnt() == 0) { // Remove dead node
2631 in2->disconnect_inputs(NULL);
2632 }
2633 }
2634 break;
2635 default:
2636 assert( !n->is_Call(), "" );
2637 assert( !n->is_Mem(), "" );
2638 break;
2639 }
2641 // Collect CFG split points
2642 if (n->is_MultiBranch())
2643 frc._tests.push(n);
2644 }
2646 //------------------------------final_graph_reshaping_walk---------------------
2647 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2648 // requires that the walk visits a node's inputs before visiting the node.
2649 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2650 ResourceArea *area = Thread::current()->resource_area();
2651 Unique_Node_List sfpt(area);
2653 frc._visited.set(root->_idx); // first, mark node as visited
2654 uint cnt = root->req();
2655 Node *n = root;
2656 uint i = 0;
2657 while (true) {
2658 if (i < cnt) {
2659 // Place all non-visited non-null inputs onto stack
2660 Node* m = n->in(i);
2661 ++i;
2662 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2663 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2664 sfpt.push(m);
2665 cnt = m->req();
2666 nstack.push(n, i); // put on stack parent and next input's index
2667 n = m;
2668 i = 0;
2669 }
2670 } else {
2671 // Now do post-visit work
2672 final_graph_reshaping_impl( n, frc );
2673 if (nstack.is_empty())
2674 break; // finished
2675 n = nstack.node(); // Get node from stack
2676 cnt = n->req();
2677 i = nstack.index();
2678 nstack.pop(); // Shift to the next node on stack
2679 }
2680 }
2682 // Skip next transformation if compressed oops are not used.
2683 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2684 return;
2686 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2687 // It could be done for an uncommon traps or any safepoints/calls
2688 // if the DecodeN node is referenced only in a debug info.
2689 while (sfpt.size() > 0) {
2690 n = sfpt.pop();
2691 JVMState *jvms = n->as_SafePoint()->jvms();
2692 assert(jvms != NULL, "sanity");
2693 int start = jvms->debug_start();
2694 int end = n->req();
2695 bool is_uncommon = (n->is_CallStaticJava() &&
2696 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2697 for (int j = start; j < end; j++) {
2698 Node* in = n->in(j);
2699 if (in->is_DecodeN()) {
2700 bool safe_to_skip = true;
2701 if (!is_uncommon ) {
2702 // Is it safe to skip?
2703 for (uint i = 0; i < in->outcnt(); i++) {
2704 Node* u = in->raw_out(i);
2705 if (!u->is_SafePoint() ||
2706 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2707 safe_to_skip = false;
2708 }
2709 }
2710 }
2711 if (safe_to_skip) {
2712 n->set_req(j, in->in(1));
2713 }
2714 if (in->outcnt() == 0) {
2715 in->disconnect_inputs(NULL);
2716 }
2717 }
2718 }
2719 }
2720 }
2722 //------------------------------final_graph_reshaping--------------------------
2723 // Final Graph Reshaping.
2724 //
2725 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2726 // and not commoned up and forced early. Must come after regular
2727 // optimizations to avoid GVN undoing the cloning. Clone constant
2728 // inputs to Loop Phis; these will be split by the allocator anyways.
2729 // Remove Opaque nodes.
2730 // (2) Move last-uses by commutative operations to the left input to encourage
2731 // Intel update-in-place two-address operations and better register usage
2732 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2733 // calls canonicalizing them back.
2734 // (3) Count the number of double-precision FP ops, single-precision FP ops
2735 // and call sites. On Intel, we can get correct rounding either by
2736 // forcing singles to memory (requires extra stores and loads after each
2737 // FP bytecode) or we can set a rounding mode bit (requires setting and
2738 // clearing the mode bit around call sites). The mode bit is only used
2739 // if the relative frequency of single FP ops to calls is low enough.
2740 // This is a key transform for SPEC mpeg_audio.
2741 // (4) Detect infinite loops; blobs of code reachable from above but not
2742 // below. Several of the Code_Gen algorithms fail on such code shapes,
2743 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2744 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2745 // Detection is by looking for IfNodes where only 1 projection is
2746 // reachable from below or CatchNodes missing some targets.
2747 // (5) Assert for insane oop offsets in debug mode.
2749 bool Compile::final_graph_reshaping() {
2750 // an infinite loop may have been eliminated by the optimizer,
2751 // in which case the graph will be empty.
2752 if (root()->req() == 1) {
2753 record_method_not_compilable("trivial infinite loop");
2754 return true;
2755 }
2757 Final_Reshape_Counts frc;
2759 // Visit everybody reachable!
2760 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2761 Node_Stack nstack(unique() >> 1);
2762 final_graph_reshaping_walk(nstack, root(), frc);
2764 // Check for unreachable (from below) code (i.e., infinite loops).
2765 for( uint i = 0; i < frc._tests.size(); i++ ) {
2766 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2767 // Get number of CFG targets.
2768 // Note that PCTables include exception targets after calls.
2769 uint required_outcnt = n->required_outcnt();
2770 if (n->outcnt() != required_outcnt) {
2771 // Check for a few special cases. Rethrow Nodes never take the
2772 // 'fall-thru' path, so expected kids is 1 less.
2773 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2774 if (n->in(0)->in(0)->is_Call()) {
2775 CallNode *call = n->in(0)->in(0)->as_Call();
2776 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2777 required_outcnt--; // Rethrow always has 1 less kid
2778 } else if (call->req() > TypeFunc::Parms &&
2779 call->is_CallDynamicJava()) {
2780 // Check for null receiver. In such case, the optimizer has
2781 // detected that the virtual call will always result in a null
2782 // pointer exception. The fall-through projection of this CatchNode
2783 // will not be populated.
2784 Node *arg0 = call->in(TypeFunc::Parms);
2785 if (arg0->is_Type() &&
2786 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2787 required_outcnt--;
2788 }
2789 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2790 call->req() > TypeFunc::Parms+1 &&
2791 call->is_CallStaticJava()) {
2792 // Check for negative array length. In such case, the optimizer has
2793 // detected that the allocation attempt will always result in an
2794 // exception. There is no fall-through projection of this CatchNode .
2795 Node *arg1 = call->in(TypeFunc::Parms+1);
2796 if (arg1->is_Type() &&
2797 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2798 required_outcnt--;
2799 }
2800 }
2801 }
2802 }
2803 // Recheck with a better notion of 'required_outcnt'
2804 if (n->outcnt() != required_outcnt) {
2805 record_method_not_compilable("malformed control flow");
2806 return true; // Not all targets reachable!
2807 }
2808 }
2809 // Check that I actually visited all kids. Unreached kids
2810 // must be infinite loops.
2811 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2812 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2813 record_method_not_compilable("infinite loop");
2814 return true; // Found unvisited kid; must be unreach
2815 }
2816 }
2818 // If original bytecodes contained a mixture of floats and doubles
2819 // check if the optimizer has made it homogenous, item (3).
2820 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2821 frc.get_float_count() > 32 &&
2822 frc.get_double_count() == 0 &&
2823 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2824 set_24_bit_selection_and_mode( false, true );
2825 }
2827 set_java_calls(frc.get_java_call_count());
2828 set_inner_loops(frc.get_inner_loop_count());
2830 // No infinite loops, no reason to bail out.
2831 return false;
2832 }
2834 //-----------------------------too_many_traps----------------------------------
2835 // Report if there are too many traps at the current method and bci.
2836 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2837 bool Compile::too_many_traps(ciMethod* method,
2838 int bci,
2839 Deoptimization::DeoptReason reason) {
2840 ciMethodData* md = method->method_data();
2841 if (md->is_empty()) {
2842 // Assume the trap has not occurred, or that it occurred only
2843 // because of a transient condition during start-up in the interpreter.
2844 return false;
2845 }
2846 if (md->has_trap_at(bci, reason) != 0) {
2847 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2848 // Also, if there are multiple reasons, or if there is no per-BCI record,
2849 // assume the worst.
2850 if (log())
2851 log()->elem("observe trap='%s' count='%d'",
2852 Deoptimization::trap_reason_name(reason),
2853 md->trap_count(reason));
2854 return true;
2855 } else {
2856 // Ignore method/bci and see if there have been too many globally.
2857 return too_many_traps(reason, md);
2858 }
2859 }
2861 // Less-accurate variant which does not require a method and bci.
2862 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2863 ciMethodData* logmd) {
2864 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2865 // Too many traps globally.
2866 // Note that we use cumulative trap_count, not just md->trap_count.
2867 if (log()) {
2868 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2869 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2870 Deoptimization::trap_reason_name(reason),
2871 mcount, trap_count(reason));
2872 }
2873 return true;
2874 } else {
2875 // The coast is clear.
2876 return false;
2877 }
2878 }
2880 //--------------------------too_many_recompiles--------------------------------
2881 // Report if there are too many recompiles at the current method and bci.
2882 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2883 // Is not eager to return true, since this will cause the compiler to use
2884 // Action_none for a trap point, to avoid too many recompilations.
2885 bool Compile::too_many_recompiles(ciMethod* method,
2886 int bci,
2887 Deoptimization::DeoptReason reason) {
2888 ciMethodData* md = method->method_data();
2889 if (md->is_empty()) {
2890 // Assume the trap has not occurred, or that it occurred only
2891 // because of a transient condition during start-up in the interpreter.
2892 return false;
2893 }
2894 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2895 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2896 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2897 Deoptimization::DeoptReason per_bc_reason
2898 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2899 if ((per_bc_reason == Deoptimization::Reason_none
2900 || md->has_trap_at(bci, reason) != 0)
2901 // The trap frequency measure we care about is the recompile count:
2902 && md->trap_recompiled_at(bci)
2903 && md->overflow_recompile_count() >= bc_cutoff) {
2904 // Do not emit a trap here if it has already caused recompilations.
2905 // Also, if there are multiple reasons, or if there is no per-BCI record,
2906 // assume the worst.
2907 if (log())
2908 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2909 Deoptimization::trap_reason_name(reason),
2910 md->trap_count(reason),
2911 md->overflow_recompile_count());
2912 return true;
2913 } else if (trap_count(reason) != 0
2914 && decompile_count() >= m_cutoff) {
2915 // Too many recompiles globally, and we have seen this sort of trap.
2916 // Use cumulative decompile_count, not just md->decompile_count.
2917 if (log())
2918 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2919 Deoptimization::trap_reason_name(reason),
2920 md->trap_count(reason), trap_count(reason),
2921 md->decompile_count(), decompile_count());
2922 return true;
2923 } else {
2924 // The coast is clear.
2925 return false;
2926 }
2927 }
2930 #ifndef PRODUCT
2931 //------------------------------verify_graph_edges---------------------------
2932 // Walk the Graph and verify that there is a one-to-one correspondence
2933 // between Use-Def edges and Def-Use edges in the graph.
2934 void Compile::verify_graph_edges(bool no_dead_code) {
2935 if (VerifyGraphEdges) {
2936 ResourceArea *area = Thread::current()->resource_area();
2937 Unique_Node_List visited(area);
2938 // Call recursive graph walk to check edges
2939 _root->verify_edges(visited);
2940 if (no_dead_code) {
2941 // Now make sure that no visited node is used by an unvisited node.
2942 bool dead_nodes = 0;
2943 Unique_Node_List checked(area);
2944 while (visited.size() > 0) {
2945 Node* n = visited.pop();
2946 checked.push(n);
2947 for (uint i = 0; i < n->outcnt(); i++) {
2948 Node* use = n->raw_out(i);
2949 if (checked.member(use)) continue; // already checked
2950 if (visited.member(use)) continue; // already in the graph
2951 if (use->is_Con()) continue; // a dead ConNode is OK
2952 // At this point, we have found a dead node which is DU-reachable.
2953 if (dead_nodes++ == 0)
2954 tty->print_cr("*** Dead nodes reachable via DU edges:");
2955 use->dump(2);
2956 tty->print_cr("---");
2957 checked.push(use); // No repeats; pretend it is now checked.
2958 }
2959 }
2960 assert(dead_nodes == 0, "using nodes must be reachable from root");
2961 }
2962 }
2963 }
2964 #endif
2966 // The Compile object keeps track of failure reasons separately from the ciEnv.
2967 // This is required because there is not quite a 1-1 relation between the
2968 // ciEnv and its compilation task and the Compile object. Note that one
2969 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2970 // to backtrack and retry without subsuming loads. Other than this backtracking
2971 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2972 // by the logic in C2Compiler.
2973 void Compile::record_failure(const char* reason) {
2974 if (log() != NULL) {
2975 log()->elem("failure reason='%s' phase='compile'", reason);
2976 }
2977 if (_failure_reason == NULL) {
2978 // Record the first failure reason.
2979 _failure_reason = reason;
2980 }
2981 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2982 C->print_method(_failure_reason);
2983 }
2984 _root = NULL; // flush the graph, too
2985 }
2987 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2988 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2989 {
2990 if (dolog) {
2991 C = Compile::current();
2992 _log = C->log();
2993 } else {
2994 C = NULL;
2995 _log = NULL;
2996 }
2997 if (_log != NULL) {
2998 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2999 _log->stamp();
3000 _log->end_head();
3001 }
3002 }
3004 Compile::TracePhase::~TracePhase() {
3005 if (_log != NULL) {
3006 _log->done("phase nodes='%d'", C->unique());
3007 }
3008 }
3010 //=============================================================================
3011 // Two Constant's are equal when the type and the value are equal.
3012 bool Compile::Constant::operator==(const Constant& other) {
3013 if (type() != other.type() ) return false;
3014 if (can_be_reused() != other.can_be_reused()) return false;
3015 // For floating point values we compare the bit pattern.
3016 switch (type()) {
3017 case T_FLOAT: return (_value.i == other._value.i);
3018 case T_LONG:
3019 case T_DOUBLE: return (_value.j == other._value.j);
3020 case T_OBJECT:
3021 case T_ADDRESS: return (_value.l == other._value.l);
3022 case T_VOID: return (_value.l == other._value.l); // jump-table entries
3023 default: ShouldNotReachHere();
3024 }
3025 return false;
3026 }
3028 // Emit constants grouped in the following order:
3029 static BasicType type_order[] = {
3030 T_FLOAT, // 32-bit
3031 T_OBJECT, // 32 or 64-bit
3032 T_ADDRESS, // 32 or 64-bit
3033 T_DOUBLE, // 64-bit
3034 T_LONG, // 64-bit
3035 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
3036 T_ILLEGAL
3037 };
3039 static int type_to_size_in_bytes(BasicType t) {
3040 switch (t) {
3041 case T_LONG: return sizeof(jlong );
3042 case T_FLOAT: return sizeof(jfloat );
3043 case T_DOUBLE: return sizeof(jdouble);
3044 // We use T_VOID as marker for jump-table entries (labels) which
3045 // need an interal word relocation.
3046 case T_VOID:
3047 case T_ADDRESS:
3048 case T_OBJECT: return sizeof(jobject);
3049 }
3051 ShouldNotReachHere();
3052 return -1;
3053 }
3055 void Compile::ConstantTable::calculate_offsets_and_size() {
3056 int size = 0;
3057 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3058 BasicType type = type_order[t];
3060 for (int i = 0; i < _constants.length(); i++) {
3061 Constant con = _constants.at(i);
3062 if (con.type() != type) continue; // Skip other types.
3064 // Align size for type.
3065 int typesize = type_to_size_in_bytes(con.type());
3066 size = align_size_up(size, typesize);
3068 // Set offset.
3069 con.set_offset(size);
3070 _constants.at_put(i, con);
3072 // Add type size.
3073 size = size + typesize;
3074 }
3075 }
3077 // Align size up to the next section start (which is insts; see
3078 // CodeBuffer::align_at_start).
3079 assert(_size == -1, "already set?");
3080 _size = align_size_up(size, CodeEntryAlignment);
3082 if (Matcher::constant_table_absolute_addressing) {
3083 set_table_base_offset(0); // No table base offset required
3084 } else {
3085 if (UseRDPCForConstantTableBase) {
3086 // table base offset is set in MachConstantBaseNode::emit
3087 } else {
3088 // When RDPC is not used, the table base is set into the middle of
3089 // the constant table.
3090 int half_size = _size / 2;
3091 assert(half_size * 2 == _size, "sanity");
3092 set_table_base_offset(-half_size);
3093 }
3094 }
3095 }
3097 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3098 MacroAssembler _masm(&cb);
3099 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3100 BasicType type = type_order[t];
3102 for (int i = 0; i < _constants.length(); i++) {
3103 Constant con = _constants.at(i);
3104 if (con.type() != type) continue; // Skip other types.
3106 address constant_addr;
3107 switch (con.type()) {
3108 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3109 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3110 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3111 case T_OBJECT: {
3112 jobject obj = con.get_jobject();
3113 int oop_index = _masm.oop_recorder()->find_index(obj);
3114 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3115 break;
3116 }
3117 case T_ADDRESS: {
3118 address addr = (address) con.get_jobject();
3119 constant_addr = _masm.address_constant(addr);
3120 break;
3121 }
3122 // We use T_VOID as marker for jump-table entries (labels) which
3123 // need an interal word relocation.
3124 case T_VOID: {
3125 // Write a dummy word. The real value is filled in later
3126 // in fill_jump_table_in_constant_table.
3127 address addr = (address) con.get_jobject();
3128 constant_addr = _masm.address_constant(addr);
3129 break;
3130 }
3131 default: ShouldNotReachHere();
3132 }
3133 assert(constant_addr != NULL, "consts section too small");
3134 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3135 }
3136 }
3137 }
3139 int Compile::ConstantTable::find_offset(Constant& con) const {
3140 int idx = _constants.find(con);
3141 assert(idx != -1, "constant must be in constant table");
3142 int offset = _constants.at(idx).offset();
3143 assert(offset != -1, "constant table not emitted yet?");
3144 return offset;
3145 }
3147 void Compile::ConstantTable::add(Constant& con) {
3148 if (con.can_be_reused()) {
3149 int idx = _constants.find(con);
3150 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3151 return;
3152 }
3153 }
3154 (void) _constants.append(con);
3155 }
3157 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
3158 Constant con(type, value);
3159 add(con);
3160 return con;
3161 }
3163 Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
3164 jvalue value;
3165 BasicType type = oper->type()->basic_type();
3166 switch (type) {
3167 case T_LONG: value.j = oper->constantL(); break;
3168 case T_FLOAT: value.f = oper->constantF(); break;
3169 case T_DOUBLE: value.d = oper->constantD(); break;
3170 case T_OBJECT:
3171 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3172 default: ShouldNotReachHere();
3173 }
3174 return add(type, value);
3175 }
3177 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
3178 jvalue value;
3179 // We can use the node pointer here to identify the right jump-table
3180 // as this method is called from Compile::Fill_buffer right before
3181 // the MachNodes are emitted and the jump-table is filled (means the
3182 // MachNode pointers do not change anymore).
3183 value.l = (jobject) n;
3184 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused.
3185 for (uint i = 0; i < n->outcnt(); i++) {
3186 add(con);
3187 }
3188 return con;
3189 }
3191 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3192 // If called from Compile::scratch_emit_size do nothing.
3193 if (Compile::current()->in_scratch_emit_size()) return;
3195 assert(labels.is_nonempty(), "must be");
3196 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
3198 // Since MachConstantNode::constant_offset() also contains
3199 // table_base_offset() we need to subtract the table_base_offset()
3200 // to get the plain offset into the constant table.
3201 int offset = n->constant_offset() - table_base_offset();
3203 MacroAssembler _masm(&cb);
3204 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3206 for (int i = 0; i < labels.length(); i++) {
3207 address* constant_addr = &jump_table_base[i];
3208 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
3209 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3210 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3211 }
3212 }