Sat, 02 Apr 2011 10:54:15 -0700
7004535: Clone loop predicate during loop unswitch
Summary: Clone loop predicate for clonned loops
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;
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 float past_uses = method()->interpreter_invocation_count();
648 float expected_uses = past_uses;
649 cg = CallGenerator::for_inline(method(), expected_uses);
650 }
651 if (failing()) return;
652 if (cg == NULL) {
653 record_method_not_compilable_all_tiers("cannot parse method");
654 return;
655 }
656 JVMState* jvms = build_start_state(start(), tf());
657 if ((jvms = cg->generate(jvms)) == NULL) {
658 record_method_not_compilable("method parse failed");
659 return;
660 }
661 GraphKit kit(jvms);
663 if (!kit.stopped()) {
664 // Accept return values, and transfer control we know not where.
665 // This is done by a special, unique ReturnNode bound to root.
666 return_values(kit.jvms());
667 }
669 if (kit.has_exceptions()) {
670 // Any exceptions that escape from this call must be rethrown
671 // to whatever caller is dynamically above us on the stack.
672 // This is done by a special, unique RethrowNode bound to root.
673 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
674 }
676 if (!failing() && has_stringbuilder()) {
677 {
678 // remove useless nodes to make the usage analysis simpler
679 ResourceMark rm;
680 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
681 }
683 {
684 ResourceMark rm;
685 print_method("Before StringOpts", 3);
686 PhaseStringOpts pso(initial_gvn(), &for_igvn);
687 print_method("After StringOpts", 3);
688 }
690 // now inline anything that we skipped the first time around
691 while (_late_inlines.length() > 0) {
692 CallGenerator* cg = _late_inlines.pop();
693 cg->do_late_inline();
694 }
695 }
696 assert(_late_inlines.length() == 0, "should have been processed");
698 print_method("Before RemoveUseless", 3);
700 // Remove clutter produced by parsing.
701 if (!failing()) {
702 ResourceMark rm;
703 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
704 }
705 }
707 // Note: Large methods are capped off in do_one_bytecode().
708 if (failing()) return;
710 // After parsing, node notes are no longer automagic.
711 // They must be propagated by register_new_node_with_optimizer(),
712 // clone(), or the like.
713 set_default_node_notes(NULL);
715 for (;;) {
716 int successes = Inline_Warm();
717 if (failing()) return;
718 if (successes == 0) break;
719 }
721 // Drain the list.
722 Finish_Warm();
723 #ifndef PRODUCT
724 if (_printer) {
725 _printer->print_inlining(this);
726 }
727 #endif
729 if (failing()) return;
730 NOT_PRODUCT( verify_graph_edges(); )
732 // Now optimize
733 Optimize();
734 if (failing()) return;
735 NOT_PRODUCT( verify_graph_edges(); )
737 #ifndef PRODUCT
738 if (PrintIdeal) {
739 ttyLocker ttyl; // keep the following output all in one block
740 // This output goes directly to the tty, not the compiler log.
741 // To enable tools to match it up with the compilation activity,
742 // be sure to tag this tty output with the compile ID.
743 if (xtty != NULL) {
744 xtty->head("ideal compile_id='%d'%s", compile_id(),
745 is_osr_compilation() ? " compile_kind='osr'" :
746 "");
747 }
748 root()->dump(9999);
749 if (xtty != NULL) {
750 xtty->tail("ideal");
751 }
752 }
753 #endif
755 // Now that we know the size of all the monitors we can add a fixed slot
756 // for the original deopt pc.
758 _orig_pc_slot = fixed_slots();
759 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
760 set_fixed_slots(next_slot);
762 // Now generate code
763 Code_Gen();
764 if (failing()) return;
766 // Check if we want to skip execution of all compiled code.
767 {
768 #ifndef PRODUCT
769 if (OptoNoExecute) {
770 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
771 return;
772 }
773 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
774 #endif
776 if (is_osr_compilation()) {
777 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
778 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
779 } else {
780 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
781 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
782 }
784 env()->register_method(_method, _entry_bci,
785 &_code_offsets,
786 _orig_pc_slot_offset_in_bytes,
787 code_buffer(),
788 frame_size_in_words(), _oop_map_set,
789 &_handler_table, &_inc_table,
790 compiler,
791 env()->comp_level(),
792 true, /*has_debug_info*/
793 has_unsafe_access()
794 );
795 }
796 }
798 //------------------------------Compile----------------------------------------
799 // Compile a runtime stub
800 Compile::Compile( ciEnv* ci_env,
801 TypeFunc_generator generator,
802 address stub_function,
803 const char *stub_name,
804 int is_fancy_jump,
805 bool pass_tls,
806 bool save_arg_registers,
807 bool return_pc )
808 : Phase(Compiler),
809 _env(ci_env),
810 _log(ci_env->log()),
811 _compile_id(-1),
812 _save_argument_registers(save_arg_registers),
813 _method(NULL),
814 _stub_name(stub_name),
815 _stub_function(stub_function),
816 _stub_entry_point(NULL),
817 _entry_bci(InvocationEntryBci),
818 _initial_gvn(NULL),
819 _for_igvn(NULL),
820 _warm_calls(NULL),
821 _orig_pc_slot(0),
822 _orig_pc_slot_offset_in_bytes(0),
823 _subsume_loads(true),
824 _do_escape_analysis(false),
825 _failure_reason(NULL),
826 _code_buffer("Compile::Fill_buffer"),
827 _has_method_handle_invokes(false),
828 _mach_constant_base_node(NULL),
829 _node_bundling_limit(0),
830 _node_bundling_base(NULL),
831 _java_calls(0),
832 _inner_loops(0),
833 #ifndef PRODUCT
834 _trace_opto_output(TraceOptoOutput),
835 _printer(NULL),
836 #endif
837 _congraph(NULL) {
838 C = this;
840 #ifndef PRODUCT
841 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
842 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
843 set_print_assembly(PrintFrameConverterAssembly);
844 set_parsed_irreducible_loop(false);
845 #endif
846 CompileWrapper cw(this);
847 Init(/*AliasLevel=*/ 0);
848 init_tf((*generator)());
850 {
851 // The following is a dummy for the sake of GraphKit::gen_stub
852 Unique_Node_List for_igvn(comp_arena());
853 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
854 PhaseGVN gvn(Thread::current()->resource_area(),255);
855 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
856 gvn.transform_no_reclaim(top());
858 GraphKit kit;
859 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
860 }
862 NOT_PRODUCT( verify_graph_edges(); )
863 Code_Gen();
864 if (failing()) return;
867 // Entry point will be accessed using compile->stub_entry_point();
868 if (code_buffer() == NULL) {
869 Matcher::soft_match_failure();
870 } else {
871 if (PrintAssembly && (WizardMode || Verbose))
872 tty->print_cr("### Stub::%s", stub_name);
874 if (!failing()) {
875 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
877 // Make the NMethod
878 // For now we mark the frame as never safe for profile stackwalking
879 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
880 code_buffer(),
881 CodeOffsets::frame_never_safe,
882 // _code_offsets.value(CodeOffsets::Frame_Complete),
883 frame_size_in_words(),
884 _oop_map_set,
885 save_arg_registers);
886 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
888 _stub_entry_point = rs->entry_point();
889 }
890 }
891 }
893 #ifndef PRODUCT
894 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
895 if(PrintOpto && Verbose) {
896 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
897 }
898 }
899 #endif
901 void Compile::print_codes() {
902 }
904 //------------------------------Init-------------------------------------------
905 // Prepare for a single compilation
906 void Compile::Init(int aliaslevel) {
907 _unique = 0;
908 _regalloc = NULL;
910 _tf = NULL; // filled in later
911 _top = NULL; // cached later
912 _matcher = NULL; // filled in later
913 _cfg = NULL; // filled in later
915 set_24_bit_selection_and_mode(Use24BitFP, false);
917 _node_note_array = NULL;
918 _default_node_notes = NULL;
920 _immutable_memory = NULL; // filled in at first inquiry
922 // Globally visible Nodes
923 // First set TOP to NULL to give safe behavior during creation of RootNode
924 set_cached_top_node(NULL);
925 set_root(new (this, 3) RootNode());
926 // Now that you have a Root to point to, create the real TOP
927 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
928 set_recent_alloc(NULL, NULL);
930 // Create Debug Information Recorder to record scopes, oopmaps, etc.
931 env()->set_oop_recorder(new OopRecorder(comp_arena()));
932 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
933 env()->set_dependencies(new Dependencies(env()));
935 _fixed_slots = 0;
936 set_has_split_ifs(false);
937 set_has_loops(has_method() && method()->has_loops()); // first approximation
938 set_has_stringbuilder(false);
939 _trap_can_recompile = false; // no traps emitted yet
940 _major_progress = true; // start out assuming good things will happen
941 set_has_unsafe_access(false);
942 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
943 set_decompile_count(0);
945 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
946 set_num_loop_opts(LoopOptsCount);
947 set_do_inlining(Inline);
948 set_max_inline_size(MaxInlineSize);
949 set_freq_inline_size(FreqInlineSize);
950 set_do_scheduling(OptoScheduling);
951 set_do_count_invocations(false);
952 set_do_method_data_update(false);
954 if (debug_info()->recording_non_safepoints()) {
955 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
956 (comp_arena(), 8, 0, NULL));
957 set_default_node_notes(Node_Notes::make(this));
958 }
960 // // -- Initialize types before each compile --
961 // // Update cached type information
962 // if( _method && _method->constants() )
963 // Type::update_loaded_types(_method, _method->constants());
965 // Init alias_type map.
966 if (!_do_escape_analysis && aliaslevel == 3)
967 aliaslevel = 2; // No unique types without escape analysis
968 _AliasLevel = aliaslevel;
969 const int grow_ats = 16;
970 _max_alias_types = grow_ats;
971 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
972 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
973 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
974 {
975 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
976 }
977 // Initialize the first few types.
978 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
979 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
980 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
981 _num_alias_types = AliasIdxRaw+1;
982 // Zero out the alias type cache.
983 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
984 // A NULL adr_type hits in the cache right away. Preload the right answer.
985 probe_alias_cache(NULL)->_index = AliasIdxTop;
987 _intrinsics = NULL;
988 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
989 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
990 register_library_intrinsics();
991 }
993 //---------------------------init_start----------------------------------------
994 // Install the StartNode on this compile object.
995 void Compile::init_start(StartNode* s) {
996 if (failing())
997 return; // already failing
998 assert(s == start(), "");
999 }
1001 StartNode* Compile::start() const {
1002 assert(!failing(), "");
1003 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1004 Node* start = root()->fast_out(i);
1005 if( start->is_Start() )
1006 return start->as_Start();
1007 }
1008 ShouldNotReachHere();
1009 return NULL;
1010 }
1012 //-------------------------------immutable_memory-------------------------------------
1013 // Access immutable memory
1014 Node* Compile::immutable_memory() {
1015 if (_immutable_memory != NULL) {
1016 return _immutable_memory;
1017 }
1018 StartNode* s = start();
1019 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1020 Node *p = s->fast_out(i);
1021 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1022 _immutable_memory = p;
1023 return _immutable_memory;
1024 }
1025 }
1026 ShouldNotReachHere();
1027 return NULL;
1028 }
1030 //----------------------set_cached_top_node------------------------------------
1031 // Install the cached top node, and make sure Node::is_top works correctly.
1032 void Compile::set_cached_top_node(Node* tn) {
1033 if (tn != NULL) verify_top(tn);
1034 Node* old_top = _top;
1035 _top = tn;
1036 // Calling Node::setup_is_top allows the nodes the chance to adjust
1037 // their _out arrays.
1038 if (_top != NULL) _top->setup_is_top();
1039 if (old_top != NULL) old_top->setup_is_top();
1040 assert(_top == NULL || top()->is_top(), "");
1041 }
1043 #ifndef PRODUCT
1044 void Compile::verify_top(Node* tn) const {
1045 if (tn != NULL) {
1046 assert(tn->is_Con(), "top node must be a constant");
1047 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1048 assert(tn->in(0) != NULL, "must have live top node");
1049 }
1050 }
1051 #endif
1054 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1056 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1057 guarantee(arr != NULL, "");
1058 int num_blocks = arr->length();
1059 if (grow_by < num_blocks) grow_by = num_blocks;
1060 int num_notes = grow_by * _node_notes_block_size;
1061 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1062 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1063 while (num_notes > 0) {
1064 arr->append(notes);
1065 notes += _node_notes_block_size;
1066 num_notes -= _node_notes_block_size;
1067 }
1068 assert(num_notes == 0, "exact multiple, please");
1069 }
1071 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1072 if (source == NULL || dest == NULL) return false;
1074 if (dest->is_Con())
1075 return false; // Do not push debug info onto constants.
1077 #ifdef ASSERT
1078 // Leave a bread crumb trail pointing to the original node:
1079 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1080 dest->set_debug_orig(source);
1081 }
1082 #endif
1084 if (node_note_array() == NULL)
1085 return false; // Not collecting any notes now.
1087 // This is a copy onto a pre-existing node, which may already have notes.
1088 // If both nodes have notes, do not overwrite any pre-existing notes.
1089 Node_Notes* source_notes = node_notes_at(source->_idx);
1090 if (source_notes == NULL || source_notes->is_clear()) return false;
1091 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1092 if (dest_notes == NULL || dest_notes->is_clear()) {
1093 return set_node_notes_at(dest->_idx, source_notes);
1094 }
1096 Node_Notes merged_notes = (*source_notes);
1097 // The order of operations here ensures that dest notes will win...
1098 merged_notes.update_from(dest_notes);
1099 return set_node_notes_at(dest->_idx, &merged_notes);
1100 }
1103 //--------------------------allow_range_check_smearing-------------------------
1104 // Gating condition for coalescing similar range checks.
1105 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1106 // single covering check that is at least as strong as any of them.
1107 // If the optimization succeeds, the simplified (strengthened) range check
1108 // will always succeed. If it fails, we will deopt, and then give up
1109 // on the optimization.
1110 bool Compile::allow_range_check_smearing() const {
1111 // If this method has already thrown a range-check,
1112 // assume it was because we already tried range smearing
1113 // and it failed.
1114 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1115 return !already_trapped;
1116 }
1119 //------------------------------flatten_alias_type-----------------------------
1120 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1121 int offset = tj->offset();
1122 TypePtr::PTR ptr = tj->ptr();
1124 // Known instance (scalarizable allocation) alias only with itself.
1125 bool is_known_inst = tj->isa_oopptr() != NULL &&
1126 tj->is_oopptr()->is_known_instance();
1128 // Process weird unsafe references.
1129 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1130 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1131 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1132 tj = TypeOopPtr::BOTTOM;
1133 ptr = tj->ptr();
1134 offset = tj->offset();
1135 }
1137 // Array pointers need some flattening
1138 const TypeAryPtr *ta = tj->isa_aryptr();
1139 if( ta && is_known_inst ) {
1140 if ( offset != Type::OffsetBot &&
1141 offset > arrayOopDesc::length_offset_in_bytes() ) {
1142 offset = Type::OffsetBot; // Flatten constant access into array body only
1143 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1144 }
1145 } else if( ta && _AliasLevel >= 2 ) {
1146 // For arrays indexed by constant indices, we flatten the alias
1147 // space to include all of the array body. Only the header, klass
1148 // and array length can be accessed un-aliased.
1149 if( offset != Type::OffsetBot ) {
1150 if( ta->const_oop() ) { // methodDataOop or methodOop
1151 offset = Type::OffsetBot; // Flatten constant access into array body
1152 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1153 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1154 // range is OK as-is.
1155 tj = ta = TypeAryPtr::RANGE;
1156 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1157 tj = TypeInstPtr::KLASS; // all klass loads look alike
1158 ta = TypeAryPtr::RANGE; // generic ignored junk
1159 ptr = TypePtr::BotPTR;
1160 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1161 tj = TypeInstPtr::MARK;
1162 ta = TypeAryPtr::RANGE; // generic ignored junk
1163 ptr = TypePtr::BotPTR;
1164 } else { // Random constant offset into array body
1165 offset = Type::OffsetBot; // Flatten constant access into array body
1166 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1167 }
1168 }
1169 // Arrays of fixed size alias with arrays of unknown size.
1170 if (ta->size() != TypeInt::POS) {
1171 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1172 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1173 }
1174 // Arrays of known objects become arrays of unknown objects.
1175 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1176 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1177 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1178 }
1179 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1180 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1181 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1182 }
1183 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1184 // cannot be distinguished by bytecode alone.
1185 if (ta->elem() == TypeInt::BOOL) {
1186 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1187 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1188 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1189 }
1190 // During the 2nd round of IterGVN, NotNull castings are removed.
1191 // Make sure the Bottom and NotNull variants alias the same.
1192 // Also, make sure exact and non-exact variants alias the same.
1193 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1194 if (ta->const_oop()) {
1195 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1196 } else {
1197 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1198 }
1199 }
1200 }
1202 // Oop pointers need some flattening
1203 const TypeInstPtr *to = tj->isa_instptr();
1204 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1205 ciInstanceKlass *k = to->klass()->as_instance_klass();
1206 if( ptr == TypePtr::Constant ) {
1207 if (to->klass() != ciEnv::current()->Class_klass() ||
1208 offset < k->size_helper() * wordSize) {
1209 // No constant oop pointers (such as Strings); they alias with
1210 // unknown strings.
1211 assert(!is_known_inst, "not scalarizable allocation");
1212 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1213 }
1214 } else if( is_known_inst ) {
1215 tj = to; // Keep NotNull and klass_is_exact for instance type
1216 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1217 // During the 2nd round of IterGVN, NotNull castings are removed.
1218 // Make sure the Bottom and NotNull variants alias the same.
1219 // Also, make sure exact and non-exact variants alias the same.
1220 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1221 }
1222 // Canonicalize the holder of this field
1223 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1224 // First handle header references such as a LoadKlassNode, even if the
1225 // object's klass is unloaded at compile time (4965979).
1226 if (!is_known_inst) { // Do it only for non-instance types
1227 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1228 }
1229 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1230 // Static fields are in the space above the normal instance
1231 // fields in the java.lang.Class instance.
1232 if (to->klass() != ciEnv::current()->Class_klass()) {
1233 to = NULL;
1234 tj = TypeOopPtr::BOTTOM;
1235 offset = tj->offset();
1236 }
1237 } else {
1238 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1239 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1240 if( is_known_inst ) {
1241 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1242 } else {
1243 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1244 }
1245 }
1246 }
1247 }
1249 // Klass pointers to object array klasses need some flattening
1250 const TypeKlassPtr *tk = tj->isa_klassptr();
1251 if( tk ) {
1252 // If we are referencing a field within a Klass, we need
1253 // to assume the worst case of an Object. Both exact and
1254 // inexact types must flatten to the same alias class.
1255 // Since the flattened result for a klass is defined to be
1256 // precisely java.lang.Object, use a constant ptr.
1257 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1259 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1260 TypeKlassPtr::OBJECT->klass(),
1261 offset);
1262 }
1264 ciKlass* klass = tk->klass();
1265 if( klass->is_obj_array_klass() ) {
1266 ciKlass* k = TypeAryPtr::OOPS->klass();
1267 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1268 k = TypeInstPtr::BOTTOM->klass();
1269 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1270 }
1272 // Check for precise loads from the primary supertype array and force them
1273 // to the supertype cache alias index. Check for generic array loads from
1274 // the primary supertype array and also force them to the supertype cache
1275 // alias index. Since the same load can reach both, we need to merge
1276 // these 2 disparate memories into the same alias class. Since the
1277 // primary supertype array is read-only, there's no chance of confusion
1278 // where we bypass an array load and an array store.
1279 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1280 if( offset == Type::OffsetBot ||
1281 off2 < Klass::primary_super_limit()*wordSize ) {
1282 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1283 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1284 }
1285 }
1287 // Flatten all Raw pointers together.
1288 if (tj->base() == Type::RawPtr)
1289 tj = TypeRawPtr::BOTTOM;
1291 if (tj->base() == Type::AnyPtr)
1292 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1294 // Flatten all to bottom for now
1295 switch( _AliasLevel ) {
1296 case 0:
1297 tj = TypePtr::BOTTOM;
1298 break;
1299 case 1: // Flatten to: oop, static, field or array
1300 switch (tj->base()) {
1301 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1302 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1303 case Type::AryPtr: // do not distinguish arrays at all
1304 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1305 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1306 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1307 default: ShouldNotReachHere();
1308 }
1309 break;
1310 case 2: // No collapsing at level 2; keep all splits
1311 case 3: // No collapsing at level 3; keep all splits
1312 break;
1313 default:
1314 Unimplemented();
1315 }
1317 offset = tj->offset();
1318 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1320 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1321 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1322 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1323 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1324 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1325 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1326 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1327 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1328 assert( tj->ptr() != TypePtr::TopPTR &&
1329 tj->ptr() != TypePtr::AnyNull &&
1330 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1331 // assert( tj->ptr() != TypePtr::Constant ||
1332 // tj->base() == Type::RawPtr ||
1333 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1335 return tj;
1336 }
1338 void Compile::AliasType::Init(int i, const TypePtr* at) {
1339 _index = i;
1340 _adr_type = at;
1341 _field = NULL;
1342 _is_rewritable = true; // default
1343 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1344 if (atoop != NULL && atoop->is_known_instance()) {
1345 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1346 _general_index = Compile::current()->get_alias_index(gt);
1347 } else {
1348 _general_index = 0;
1349 }
1350 }
1352 //---------------------------------print_on------------------------------------
1353 #ifndef PRODUCT
1354 void Compile::AliasType::print_on(outputStream* st) {
1355 if (index() < 10)
1356 st->print("@ <%d> ", index());
1357 else st->print("@ <%d>", index());
1358 st->print(is_rewritable() ? " " : " RO");
1359 int offset = adr_type()->offset();
1360 if (offset == Type::OffsetBot)
1361 st->print(" +any");
1362 else st->print(" +%-3d", offset);
1363 st->print(" in ");
1364 adr_type()->dump_on(st);
1365 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1366 if (field() != NULL && tjp) {
1367 if (tjp->klass() != field()->holder() ||
1368 tjp->offset() != field()->offset_in_bytes()) {
1369 st->print(" != ");
1370 field()->print();
1371 st->print(" ***");
1372 }
1373 }
1374 }
1376 void print_alias_types() {
1377 Compile* C = Compile::current();
1378 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1379 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1380 C->alias_type(idx)->print_on(tty);
1381 tty->cr();
1382 }
1383 }
1384 #endif
1387 //----------------------------probe_alias_cache--------------------------------
1388 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1389 intptr_t key = (intptr_t) adr_type;
1390 key ^= key >> logAliasCacheSize;
1391 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1392 }
1395 //-----------------------------grow_alias_types--------------------------------
1396 void Compile::grow_alias_types() {
1397 const int old_ats = _max_alias_types; // how many before?
1398 const int new_ats = old_ats; // how many more?
1399 const int grow_ats = old_ats+new_ats; // how many now?
1400 _max_alias_types = grow_ats;
1401 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1402 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1403 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1404 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1405 }
1408 //--------------------------------find_alias_type------------------------------
1409 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1410 if (_AliasLevel == 0)
1411 return alias_type(AliasIdxBot);
1413 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1414 if (ace->_adr_type == adr_type) {
1415 return alias_type(ace->_index);
1416 }
1418 // Handle special cases.
1419 if (adr_type == NULL) return alias_type(AliasIdxTop);
1420 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1422 // Do it the slow way.
1423 const TypePtr* flat = flatten_alias_type(adr_type);
1425 #ifdef ASSERT
1426 assert(flat == flatten_alias_type(flat), "idempotent");
1427 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1428 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1429 const TypeOopPtr* foop = flat->is_oopptr();
1430 // Scalarizable allocations have exact klass always.
1431 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1432 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1433 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1434 }
1435 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1436 #endif
1438 int idx = AliasIdxTop;
1439 for (int i = 0; i < num_alias_types(); i++) {
1440 if (alias_type(i)->adr_type() == flat) {
1441 idx = i;
1442 break;
1443 }
1444 }
1446 if (idx == AliasIdxTop) {
1447 if (no_create) return NULL;
1448 // Grow the array if necessary.
1449 if (_num_alias_types == _max_alias_types) grow_alias_types();
1450 // Add a new alias type.
1451 idx = _num_alias_types++;
1452 _alias_types[idx]->Init(idx, flat);
1453 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1454 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1455 if (flat->isa_instptr()) {
1456 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1457 && flat->is_instptr()->klass() == env()->Class_klass())
1458 alias_type(idx)->set_rewritable(false);
1459 }
1460 if (flat->isa_klassptr()) {
1461 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1462 alias_type(idx)->set_rewritable(false);
1463 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1464 alias_type(idx)->set_rewritable(false);
1465 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1466 alias_type(idx)->set_rewritable(false);
1467 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1468 alias_type(idx)->set_rewritable(false);
1469 }
1470 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1471 // but the base pointer type is not distinctive enough to identify
1472 // references into JavaThread.)
1474 // Check for final fields.
1475 const TypeInstPtr* tinst = flat->isa_instptr();
1476 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1477 ciField* field;
1478 if (tinst->const_oop() != NULL &&
1479 tinst->klass() == ciEnv::current()->Class_klass() &&
1480 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1481 // static field
1482 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1483 field = k->get_field_by_offset(tinst->offset(), true);
1484 } else {
1485 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1486 field = k->get_field_by_offset(tinst->offset(), false);
1487 }
1488 assert(field == NULL ||
1489 original_field == NULL ||
1490 (field->holder() == original_field->holder() &&
1491 field->offset() == original_field->offset() &&
1492 field->is_static() == original_field->is_static()), "wrong field?");
1493 // Set field() and is_rewritable() attributes.
1494 if (field != NULL) alias_type(idx)->set_field(field);
1495 }
1496 }
1498 // Fill the cache for next time.
1499 ace->_adr_type = adr_type;
1500 ace->_index = idx;
1501 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1503 // Might as well try to fill the cache for the flattened version, too.
1504 AliasCacheEntry* face = probe_alias_cache(flat);
1505 if (face->_adr_type == NULL) {
1506 face->_adr_type = flat;
1507 face->_index = idx;
1508 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1509 }
1511 return alias_type(idx);
1512 }
1515 Compile::AliasType* Compile::alias_type(ciField* field) {
1516 const TypeOopPtr* t;
1517 if (field->is_static())
1518 t = TypeInstPtr::make(field->holder()->java_mirror());
1519 else
1520 t = TypeOopPtr::make_from_klass_raw(field->holder());
1521 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1522 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1523 return atp;
1524 }
1527 //------------------------------have_alias_type--------------------------------
1528 bool Compile::have_alias_type(const TypePtr* adr_type) {
1529 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1530 if (ace->_adr_type == adr_type) {
1531 return true;
1532 }
1534 // Handle special cases.
1535 if (adr_type == NULL) return true;
1536 if (adr_type == TypePtr::BOTTOM) return true;
1538 return find_alias_type(adr_type, true, NULL) != NULL;
1539 }
1541 //-----------------------------must_alias--------------------------------------
1542 // True if all values of the given address type are in the given alias category.
1543 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1544 if (alias_idx == AliasIdxBot) return true; // the universal category
1545 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1546 if (alias_idx == AliasIdxTop) return false; // the empty category
1547 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1549 // the only remaining possible overlap is identity
1550 int adr_idx = get_alias_index(adr_type);
1551 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1552 assert(adr_idx == alias_idx ||
1553 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1554 && adr_type != TypeOopPtr::BOTTOM),
1555 "should not be testing for overlap with an unsafe pointer");
1556 return adr_idx == alias_idx;
1557 }
1559 //------------------------------can_alias--------------------------------------
1560 // True if any values of the given address type are in the given alias category.
1561 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1562 if (alias_idx == AliasIdxTop) return false; // the empty category
1563 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1564 if (alias_idx == AliasIdxBot) return true; // the universal category
1565 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1567 // the only remaining possible overlap is identity
1568 int adr_idx = get_alias_index(adr_type);
1569 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1570 return adr_idx == alias_idx;
1571 }
1575 //---------------------------pop_warm_call-------------------------------------
1576 WarmCallInfo* Compile::pop_warm_call() {
1577 WarmCallInfo* wci = _warm_calls;
1578 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1579 return wci;
1580 }
1582 //----------------------------Inline_Warm--------------------------------------
1583 int Compile::Inline_Warm() {
1584 // If there is room, try to inline some more warm call sites.
1585 // %%% Do a graph index compaction pass when we think we're out of space?
1586 if (!InlineWarmCalls) return 0;
1588 int calls_made_hot = 0;
1589 int room_to_grow = NodeCountInliningCutoff - unique();
1590 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1591 int amount_grown = 0;
1592 WarmCallInfo* call;
1593 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1594 int est_size = (int)call->size();
1595 if (est_size > (room_to_grow - amount_grown)) {
1596 // This one won't fit anyway. Get rid of it.
1597 call->make_cold();
1598 continue;
1599 }
1600 call->make_hot();
1601 calls_made_hot++;
1602 amount_grown += est_size;
1603 amount_to_grow -= est_size;
1604 }
1606 if (calls_made_hot > 0) set_major_progress();
1607 return calls_made_hot;
1608 }
1611 //----------------------------Finish_Warm--------------------------------------
1612 void Compile::Finish_Warm() {
1613 if (!InlineWarmCalls) return;
1614 if (failing()) return;
1615 if (warm_calls() == NULL) return;
1617 // Clean up loose ends, if we are out of space for inlining.
1618 WarmCallInfo* call;
1619 while ((call = pop_warm_call()) != NULL) {
1620 call->make_cold();
1621 }
1622 }
1624 //---------------------cleanup_loop_predicates-----------------------
1625 // Remove the opaque nodes that protect the predicates so that all unused
1626 // checks and uncommon_traps will be eliminated from the ideal graph
1627 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1628 if (predicate_count()==0) return;
1629 for (int i = predicate_count(); i > 0; i--) {
1630 Node * n = predicate_opaque1_node(i-1);
1631 assert(n->Opcode() == Op_Opaque1, "must be");
1632 igvn.replace_node(n, n->in(1));
1633 }
1634 assert(predicate_count()==0, "should be clean!");
1635 }
1637 //------------------------------Optimize---------------------------------------
1638 // Given a graph, optimize it.
1639 void Compile::Optimize() {
1640 TracePhase t1("optimizer", &_t_optimizer, true);
1642 #ifndef PRODUCT
1643 if (env()->break_at_compile()) {
1644 BREAKPOINT;
1645 }
1647 #endif
1649 ResourceMark rm;
1650 int loop_opts_cnt;
1652 NOT_PRODUCT( verify_graph_edges(); )
1654 print_method("After Parsing");
1656 {
1657 // Iterative Global Value Numbering, including ideal transforms
1658 // Initialize IterGVN with types and values from parse-time GVN
1659 PhaseIterGVN igvn(initial_gvn());
1660 {
1661 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1662 igvn.optimize();
1663 }
1665 print_method("Iter GVN 1", 2);
1667 if (failing()) return;
1669 // Perform escape analysis
1670 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1671 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
1672 ConnectionGraph::do_analysis(this, &igvn);
1674 if (failing()) return;
1676 igvn.optimize();
1677 print_method("Iter GVN 3", 2);
1679 if (failing()) return;
1681 }
1683 // Loop transforms on the ideal graph. Range Check Elimination,
1684 // peeling, unrolling, etc.
1686 // Set loop opts counter
1687 loop_opts_cnt = num_loop_opts();
1688 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1689 {
1690 TracePhase t2("idealLoop", &_t_idealLoop, true);
1691 PhaseIdealLoop ideal_loop( igvn, true );
1692 loop_opts_cnt--;
1693 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1694 if (failing()) return;
1695 }
1696 // Loop opts pass if partial peeling occurred in previous pass
1697 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1698 TracePhase t3("idealLoop", &_t_idealLoop, true);
1699 PhaseIdealLoop ideal_loop( igvn, false );
1700 loop_opts_cnt--;
1701 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1702 if (failing()) return;
1703 }
1704 // Loop opts pass for loop-unrolling before CCP
1705 if(major_progress() && (loop_opts_cnt > 0)) {
1706 TracePhase t4("idealLoop", &_t_idealLoop, true);
1707 PhaseIdealLoop ideal_loop( igvn, false );
1708 loop_opts_cnt--;
1709 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1710 }
1711 if (!failing()) {
1712 // Verify that last round of loop opts produced a valid graph
1713 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1714 PhaseIdealLoop::verify(igvn);
1715 }
1716 }
1717 if (failing()) return;
1719 // Conditional Constant Propagation;
1720 PhaseCCP ccp( &igvn );
1721 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1722 {
1723 TracePhase t2("ccp", &_t_ccp, true);
1724 ccp.do_transform();
1725 }
1726 print_method("PhaseCPP 1", 2);
1728 assert( true, "Break here to ccp.dump_old2new_map()");
1730 // Iterative Global Value Numbering, including ideal transforms
1731 {
1732 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1733 igvn = ccp;
1734 igvn.optimize();
1735 }
1737 print_method("Iter GVN 2", 2);
1739 if (failing()) return;
1741 // Loop transforms on the ideal graph. Range Check Elimination,
1742 // peeling, unrolling, etc.
1743 if(loop_opts_cnt > 0) {
1744 debug_only( int cnt = 0; );
1745 while(major_progress() && (loop_opts_cnt > 0)) {
1746 TracePhase t2("idealLoop", &_t_idealLoop, true);
1747 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1748 PhaseIdealLoop ideal_loop( igvn, true);
1749 loop_opts_cnt--;
1750 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1751 if (failing()) return;
1752 }
1753 }
1755 {
1756 // Verify that all previous optimizations produced a valid graph
1757 // at least to this point, even if no loop optimizations were done.
1758 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1759 PhaseIdealLoop::verify(igvn);
1760 }
1762 {
1763 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1764 PhaseMacroExpand mex(igvn);
1765 if (mex.expand_macro_nodes()) {
1766 assert(failing(), "must bail out w/ explicit message");
1767 return;
1768 }
1769 }
1771 } // (End scope of igvn; run destructor if necessary for asserts.)
1773 // A method with only infinite loops has no edges entering loops from root
1774 {
1775 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1776 if (final_graph_reshaping()) {
1777 assert(failing(), "must bail out w/ explicit message");
1778 return;
1779 }
1780 }
1782 print_method("Optimize finished", 2);
1783 }
1786 //------------------------------Code_Gen---------------------------------------
1787 // Given a graph, generate code for it
1788 void Compile::Code_Gen() {
1789 if (failing()) return;
1791 // Perform instruction selection. You might think we could reclaim Matcher
1792 // memory PDQ, but actually the Matcher is used in generating spill code.
1793 // Internals of the Matcher (including some VectorSets) must remain live
1794 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1795 // set a bit in reclaimed memory.
1797 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1798 // nodes. Mapping is only valid at the root of each matched subtree.
1799 NOT_PRODUCT( verify_graph_edges(); )
1801 Node_List proj_list;
1802 Matcher m(proj_list);
1803 _matcher = &m;
1804 {
1805 TracePhase t2("matcher", &_t_matcher, true);
1806 m.match();
1807 }
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 // If you have too many nodes, or if matching has failed, bail out
1813 check_node_count(0, "out of nodes matching instructions");
1814 if (failing()) return;
1816 // Build a proper-looking CFG
1817 PhaseCFG cfg(node_arena(), root(), m);
1818 _cfg = &cfg;
1819 {
1820 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1821 cfg.Dominators();
1822 if (failing()) return;
1824 NOT_PRODUCT( verify_graph_edges(); )
1826 cfg.Estimate_Block_Frequency();
1827 cfg.GlobalCodeMotion(m,unique(),proj_list);
1829 print_method("Global code motion", 2);
1831 if (failing()) return;
1832 NOT_PRODUCT( verify_graph_edges(); )
1834 debug_only( cfg.verify(); )
1835 }
1836 NOT_PRODUCT( verify_graph_edges(); )
1838 PhaseChaitin regalloc(unique(),cfg,m);
1839 _regalloc = ®alloc;
1840 {
1841 TracePhase t2("regalloc", &_t_registerAllocation, true);
1842 // Perform any platform dependent preallocation actions. This is used,
1843 // for example, to avoid taking an implicit null pointer exception
1844 // using the frame pointer on win95.
1845 _regalloc->pd_preallocate_hook();
1847 // Perform register allocation. After Chaitin, use-def chains are
1848 // no longer accurate (at spill code) and so must be ignored.
1849 // Node->LRG->reg mappings are still accurate.
1850 _regalloc->Register_Allocate();
1852 // Bail out if the allocator builds too many nodes
1853 if (failing()) return;
1854 }
1856 // Prior to register allocation we kept empty basic blocks in case the
1857 // the allocator needed a place to spill. After register allocation we
1858 // are not adding any new instructions. If any basic block is empty, we
1859 // can now safely remove it.
1860 {
1861 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1862 cfg.remove_empty();
1863 if (do_freq_based_layout()) {
1864 PhaseBlockLayout layout(cfg);
1865 } else {
1866 cfg.set_loop_alignment();
1867 }
1868 cfg.fixup_flow();
1869 }
1871 // Perform any platform dependent postallocation verifications.
1872 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1874 // Apply peephole optimizations
1875 if( OptoPeephole ) {
1876 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1877 PhasePeephole peep( _regalloc, cfg);
1878 peep.do_transform();
1879 }
1881 // Convert Nodes to instruction bits in a buffer
1882 {
1883 // %%%% workspace merge brought two timers together for one job
1884 TracePhase t2a("output", &_t_output, true);
1885 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1886 Output();
1887 }
1889 print_method("Final Code");
1891 // He's dead, Jim.
1892 _cfg = (PhaseCFG*)0xdeadbeef;
1893 _regalloc = (PhaseChaitin*)0xdeadbeef;
1894 }
1897 //------------------------------dump_asm---------------------------------------
1898 // Dump formatted assembly
1899 #ifndef PRODUCT
1900 void Compile::dump_asm(int *pcs, uint pc_limit) {
1901 bool cut_short = false;
1902 tty->print_cr("#");
1903 tty->print("# "); _tf->dump(); tty->cr();
1904 tty->print_cr("#");
1906 // For all blocks
1907 int pc = 0x0; // Program counter
1908 char starts_bundle = ' ';
1909 _regalloc->dump_frame();
1911 Node *n = NULL;
1912 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1913 if (VMThread::should_terminate()) { cut_short = true; break; }
1914 Block *b = _cfg->_blocks[i];
1915 if (b->is_connector() && !Verbose) continue;
1916 n = b->_nodes[0];
1917 if (pcs && n->_idx < pc_limit)
1918 tty->print("%3.3x ", pcs[n->_idx]);
1919 else
1920 tty->print(" ");
1921 b->dump_head( &_cfg->_bbs );
1922 if (b->is_connector()) {
1923 tty->print_cr(" # Empty connector block");
1924 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1925 tty->print_cr(" # Block is sole successor of call");
1926 }
1928 // For all instructions
1929 Node *delay = NULL;
1930 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1931 if (VMThread::should_terminate()) { cut_short = true; break; }
1932 n = b->_nodes[j];
1933 if (valid_bundle_info(n)) {
1934 Bundle *bundle = node_bundling(n);
1935 if (bundle->used_in_unconditional_delay()) {
1936 delay = n;
1937 continue;
1938 }
1939 if (bundle->starts_bundle())
1940 starts_bundle = '+';
1941 }
1943 if (WizardMode) n->dump();
1945 if( !n->is_Region() && // Dont print in the Assembly
1946 !n->is_Phi() && // a few noisely useless nodes
1947 !n->is_Proj() &&
1948 !n->is_MachTemp() &&
1949 !n->is_SafePointScalarObject() &&
1950 !n->is_Catch() && // Would be nice to print exception table targets
1951 !n->is_MergeMem() && // Not very interesting
1952 !n->is_top() && // Debug info table constants
1953 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1954 ) {
1955 if (pcs && n->_idx < pc_limit)
1956 tty->print("%3.3x", pcs[n->_idx]);
1957 else
1958 tty->print(" ");
1959 tty->print(" %c ", starts_bundle);
1960 starts_bundle = ' ';
1961 tty->print("\t");
1962 n->format(_regalloc, tty);
1963 tty->cr();
1964 }
1966 // If we have an instruction with a delay slot, and have seen a delay,
1967 // then back up and print it
1968 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1969 assert(delay != NULL, "no unconditional delay instruction");
1970 if (WizardMode) delay->dump();
1972 if (node_bundling(delay)->starts_bundle())
1973 starts_bundle = '+';
1974 if (pcs && n->_idx < pc_limit)
1975 tty->print("%3.3x", pcs[n->_idx]);
1976 else
1977 tty->print(" ");
1978 tty->print(" %c ", starts_bundle);
1979 starts_bundle = ' ';
1980 tty->print("\t");
1981 delay->format(_regalloc, tty);
1982 tty->print_cr("");
1983 delay = NULL;
1984 }
1986 // Dump the exception table as well
1987 if( n->is_Catch() && (Verbose || WizardMode) ) {
1988 // Print the exception table for this offset
1989 _handler_table.print_subtable_for(pc);
1990 }
1991 }
1993 if (pcs && n->_idx < pc_limit)
1994 tty->print_cr("%3.3x", pcs[n->_idx]);
1995 else
1996 tty->print_cr("");
1998 assert(cut_short || delay == NULL, "no unconditional delay branch");
2000 } // End of per-block dump
2001 tty->print_cr("");
2003 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2004 }
2005 #endif
2007 //------------------------------Final_Reshape_Counts---------------------------
2008 // This class defines counters to help identify when a method
2009 // may/must be executed using hardware with only 24-bit precision.
2010 struct Final_Reshape_Counts : public StackObj {
2011 int _call_count; // count non-inlined 'common' calls
2012 int _float_count; // count float ops requiring 24-bit precision
2013 int _double_count; // count double ops requiring more precision
2014 int _java_call_count; // count non-inlined 'java' calls
2015 int _inner_loop_count; // count loops which need alignment
2016 VectorSet _visited; // Visitation flags
2017 Node_List _tests; // Set of IfNodes & PCTableNodes
2019 Final_Reshape_Counts() :
2020 _call_count(0), _float_count(0), _double_count(0),
2021 _java_call_count(0), _inner_loop_count(0),
2022 _visited( Thread::current()->resource_area() ) { }
2024 void inc_call_count () { _call_count ++; }
2025 void inc_float_count () { _float_count ++; }
2026 void inc_double_count() { _double_count++; }
2027 void inc_java_call_count() { _java_call_count++; }
2028 void inc_inner_loop_count() { _inner_loop_count++; }
2030 int get_call_count () const { return _call_count ; }
2031 int get_float_count () const { return _float_count ; }
2032 int get_double_count() const { return _double_count; }
2033 int get_java_call_count() const { return _java_call_count; }
2034 int get_inner_loop_count() const { return _inner_loop_count; }
2035 };
2037 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2038 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2039 // Make sure the offset goes inside the instance layout.
2040 return k->contains_field_offset(tp->offset());
2041 // Note that OffsetBot and OffsetTop are very negative.
2042 }
2044 //------------------------------final_graph_reshaping_impl----------------------
2045 // Implement items 1-5 from final_graph_reshaping below.
2046 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
2048 if ( n->outcnt() == 0 ) return; // dead node
2049 uint nop = n->Opcode();
2051 // Check for 2-input instruction with "last use" on right input.
2052 // Swap to left input. Implements item (2).
2053 if( n->req() == 3 && // two-input instruction
2054 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2055 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2056 n->in(2)->outcnt() == 1 &&// right use IS a last use
2057 !n->in(2)->is_Con() ) { // right use is not a constant
2058 // Check for commutative opcode
2059 switch( nop ) {
2060 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2061 case Op_MaxI: case Op_MinI:
2062 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2063 case Op_AndL: case Op_XorL: case Op_OrL:
2064 case Op_AndI: case Op_XorI: case Op_OrI: {
2065 // Move "last use" input to left by swapping inputs
2066 n->swap_edges(1, 2);
2067 break;
2068 }
2069 default:
2070 break;
2071 }
2072 }
2074 #ifdef ASSERT
2075 if( n->is_Mem() ) {
2076 Compile* C = Compile::current();
2077 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
2078 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2079 // oop will be recorded in oop map if load crosses safepoint
2080 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2081 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2082 "raw memory operations should have control edge");
2083 }
2084 #endif
2085 // Count FPU ops and common calls, implements item (3)
2086 switch( nop ) {
2087 // Count all float operations that may use FPU
2088 case Op_AddF:
2089 case Op_SubF:
2090 case Op_MulF:
2091 case Op_DivF:
2092 case Op_NegF:
2093 case Op_ModF:
2094 case Op_ConvI2F:
2095 case Op_ConF:
2096 case Op_CmpF:
2097 case Op_CmpF3:
2098 // case Op_ConvL2F: // longs are split into 32-bit halves
2099 frc.inc_float_count();
2100 break;
2102 case Op_ConvF2D:
2103 case Op_ConvD2F:
2104 frc.inc_float_count();
2105 frc.inc_double_count();
2106 break;
2108 // Count all double operations that may use FPU
2109 case Op_AddD:
2110 case Op_SubD:
2111 case Op_MulD:
2112 case Op_DivD:
2113 case Op_NegD:
2114 case Op_ModD:
2115 case Op_ConvI2D:
2116 case Op_ConvD2I:
2117 // case Op_ConvL2D: // handled by leaf call
2118 // case Op_ConvD2L: // handled by leaf call
2119 case Op_ConD:
2120 case Op_CmpD:
2121 case Op_CmpD3:
2122 frc.inc_double_count();
2123 break;
2124 case Op_Opaque1: // Remove Opaque Nodes before matching
2125 case Op_Opaque2: // Remove Opaque Nodes before matching
2126 n->subsume_by(n->in(1));
2127 break;
2128 case Op_CallStaticJava:
2129 case Op_CallJava:
2130 case Op_CallDynamicJava:
2131 frc.inc_java_call_count(); // Count java call site;
2132 case Op_CallRuntime:
2133 case Op_CallLeaf:
2134 case Op_CallLeafNoFP: {
2135 assert( n->is_Call(), "" );
2136 CallNode *call = n->as_Call();
2137 // Count call sites where the FP mode bit would have to be flipped.
2138 // Do not count uncommon runtime calls:
2139 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2140 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2141 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2142 frc.inc_call_count(); // Count the call site
2143 } else { // See if uncommon argument is shared
2144 Node *n = call->in(TypeFunc::Parms);
2145 int nop = n->Opcode();
2146 // Clone shared simple arguments to uncommon calls, item (1).
2147 if( n->outcnt() > 1 &&
2148 !n->is_Proj() &&
2149 nop != Op_CreateEx &&
2150 nop != Op_CheckCastPP &&
2151 nop != Op_DecodeN &&
2152 !n->is_Mem() ) {
2153 Node *x = n->clone();
2154 call->set_req( TypeFunc::Parms, x );
2155 }
2156 }
2157 break;
2158 }
2160 case Op_StoreD:
2161 case Op_LoadD:
2162 case Op_LoadD_unaligned:
2163 frc.inc_double_count();
2164 goto handle_mem;
2165 case Op_StoreF:
2166 case Op_LoadF:
2167 frc.inc_float_count();
2168 goto handle_mem;
2170 case Op_StoreB:
2171 case Op_StoreC:
2172 case Op_StoreCM:
2173 case Op_StorePConditional:
2174 case Op_StoreI:
2175 case Op_StoreL:
2176 case Op_StoreIConditional:
2177 case Op_StoreLConditional:
2178 case Op_CompareAndSwapI:
2179 case Op_CompareAndSwapL:
2180 case Op_CompareAndSwapP:
2181 case Op_CompareAndSwapN:
2182 case Op_StoreP:
2183 case Op_StoreN:
2184 case Op_LoadB:
2185 case Op_LoadUB:
2186 case Op_LoadUS:
2187 case Op_LoadI:
2188 case Op_LoadUI2L:
2189 case Op_LoadKlass:
2190 case Op_LoadNKlass:
2191 case Op_LoadL:
2192 case Op_LoadL_unaligned:
2193 case Op_LoadPLocked:
2194 case Op_LoadLLocked:
2195 case Op_LoadP:
2196 case Op_LoadN:
2197 case Op_LoadRange:
2198 case Op_LoadS: {
2199 handle_mem:
2200 #ifdef ASSERT
2201 if( VerifyOptoOopOffsets ) {
2202 assert( n->is_Mem(), "" );
2203 MemNode *mem = (MemNode*)n;
2204 // Check to see if address types have grounded out somehow.
2205 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2206 assert( !tp || oop_offset_is_sane(tp), "" );
2207 }
2208 #endif
2209 break;
2210 }
2212 case Op_AddP: { // Assert sane base pointers
2213 Node *addp = n->in(AddPNode::Address);
2214 assert( !addp->is_AddP() ||
2215 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2216 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2217 "Base pointers must match" );
2218 #ifdef _LP64
2219 if (UseCompressedOops &&
2220 addp->Opcode() == Op_ConP &&
2221 addp == n->in(AddPNode::Base) &&
2222 n->in(AddPNode::Offset)->is_Con()) {
2223 // Use addressing with narrow klass to load with offset on x86.
2224 // On sparc loading 32-bits constant and decoding it have less
2225 // instructions (4) then load 64-bits constant (7).
2226 // Do this transformation here since IGVN will convert ConN back to ConP.
2227 const Type* t = addp->bottom_type();
2228 if (t->isa_oopptr()) {
2229 Node* nn = NULL;
2231 // Look for existing ConN node of the same exact type.
2232 Compile* C = Compile::current();
2233 Node* r = C->root();
2234 uint cnt = r->outcnt();
2235 for (uint i = 0; i < cnt; i++) {
2236 Node* m = r->raw_out(i);
2237 if (m!= NULL && m->Opcode() == Op_ConN &&
2238 m->bottom_type()->make_ptr() == t) {
2239 nn = m;
2240 break;
2241 }
2242 }
2243 if (nn != NULL) {
2244 // Decode a narrow oop to match address
2245 // [R12 + narrow_oop_reg<<3 + offset]
2246 nn = new (C, 2) DecodeNNode(nn, t);
2247 n->set_req(AddPNode::Base, nn);
2248 n->set_req(AddPNode::Address, nn);
2249 if (addp->outcnt() == 0) {
2250 addp->disconnect_inputs(NULL);
2251 }
2252 }
2253 }
2254 }
2255 #endif
2256 break;
2257 }
2259 #ifdef _LP64
2260 case Op_CastPP:
2261 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2262 Compile* C = Compile::current();
2263 Node* in1 = n->in(1);
2264 const Type* t = n->bottom_type();
2265 Node* new_in1 = in1->clone();
2266 new_in1->as_DecodeN()->set_type(t);
2268 if (!Matcher::narrow_oop_use_complex_address()) {
2269 //
2270 // x86, ARM and friends can handle 2 adds in addressing mode
2271 // and Matcher can fold a DecodeN node into address by using
2272 // a narrow oop directly and do implicit NULL check in address:
2273 //
2274 // [R12 + narrow_oop_reg<<3 + offset]
2275 // NullCheck narrow_oop_reg
2276 //
2277 // On other platforms (Sparc) we have to keep new DecodeN node and
2278 // use it to do implicit NULL check in address:
2279 //
2280 // decode_not_null narrow_oop_reg, base_reg
2281 // [base_reg + offset]
2282 // NullCheck base_reg
2283 //
2284 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2285 // to keep the information to which NULL check the new DecodeN node
2286 // corresponds to use it as value in implicit_null_check().
2287 //
2288 new_in1->set_req(0, n->in(0));
2289 }
2291 n->subsume_by(new_in1);
2292 if (in1->outcnt() == 0) {
2293 in1->disconnect_inputs(NULL);
2294 }
2295 }
2296 break;
2298 case Op_CmpP:
2299 // Do this transformation here to preserve CmpPNode::sub() and
2300 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2301 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2302 Node* in1 = n->in(1);
2303 Node* in2 = n->in(2);
2304 if (!in1->is_DecodeN()) {
2305 in2 = in1;
2306 in1 = n->in(2);
2307 }
2308 assert(in1->is_DecodeN(), "sanity");
2310 Compile* C = Compile::current();
2311 Node* new_in2 = NULL;
2312 if (in2->is_DecodeN()) {
2313 new_in2 = in2->in(1);
2314 } else if (in2->Opcode() == Op_ConP) {
2315 const Type* t = in2->bottom_type();
2316 if (t == TypePtr::NULL_PTR) {
2317 // Don't convert CmpP null check into CmpN if compressed
2318 // oops implicit null check is not generated.
2319 // This will allow to generate normal oop implicit null check.
2320 if (Matcher::gen_narrow_oop_implicit_null_checks())
2321 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2322 //
2323 // This transformation together with CastPP transformation above
2324 // will generated code for implicit NULL checks for compressed oops.
2325 //
2326 // The original code after Optimize()
2327 //
2328 // LoadN memory, narrow_oop_reg
2329 // decode narrow_oop_reg, base_reg
2330 // CmpP base_reg, NULL
2331 // CastPP base_reg // NotNull
2332 // Load [base_reg + offset], val_reg
2333 //
2334 // after these transformations will be
2335 //
2336 // LoadN memory, narrow_oop_reg
2337 // CmpN narrow_oop_reg, NULL
2338 // decode_not_null narrow_oop_reg, base_reg
2339 // Load [base_reg + offset], val_reg
2340 //
2341 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2342 // since narrow oops can be used in debug info now (see the code in
2343 // final_graph_reshaping_walk()).
2344 //
2345 // At the end the code will be matched to
2346 // on x86:
2347 //
2348 // Load_narrow_oop memory, narrow_oop_reg
2349 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2350 // NullCheck narrow_oop_reg
2351 //
2352 // and on sparc:
2353 //
2354 // Load_narrow_oop memory, narrow_oop_reg
2355 // decode_not_null narrow_oop_reg, base_reg
2356 // Load [base_reg + offset], val_reg
2357 // NullCheck base_reg
2358 //
2359 } else if (t->isa_oopptr()) {
2360 new_in2 = ConNode::make(C, t->make_narrowoop());
2361 }
2362 }
2363 if (new_in2 != NULL) {
2364 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2365 n->subsume_by( cmpN );
2366 if (in1->outcnt() == 0) {
2367 in1->disconnect_inputs(NULL);
2368 }
2369 if (in2->outcnt() == 0) {
2370 in2->disconnect_inputs(NULL);
2371 }
2372 }
2373 }
2374 break;
2376 case Op_DecodeN:
2377 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2378 // DecodeN could be pinned when it can't be fold into
2379 // an address expression, see the code for Op_CastPP above.
2380 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2381 break;
2383 case Op_EncodeP: {
2384 Node* in1 = n->in(1);
2385 if (in1->is_DecodeN()) {
2386 n->subsume_by(in1->in(1));
2387 } else if (in1->Opcode() == Op_ConP) {
2388 Compile* C = Compile::current();
2389 const Type* t = in1->bottom_type();
2390 if (t == TypePtr::NULL_PTR) {
2391 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2392 } else if (t->isa_oopptr()) {
2393 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2394 }
2395 }
2396 if (in1->outcnt() == 0) {
2397 in1->disconnect_inputs(NULL);
2398 }
2399 break;
2400 }
2402 case Op_Proj: {
2403 if (OptimizeStringConcat) {
2404 ProjNode* p = n->as_Proj();
2405 if (p->_is_io_use) {
2406 // Separate projections were used for the exception path which
2407 // are normally removed by a late inline. If it wasn't inlined
2408 // then they will hang around and should just be replaced with
2409 // the original one.
2410 Node* proj = NULL;
2411 // Replace with just one
2412 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2413 Node *use = i.get();
2414 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2415 proj = use;
2416 break;
2417 }
2418 }
2419 assert(p != NULL, "must be found");
2420 p->subsume_by(proj);
2421 }
2422 }
2423 break;
2424 }
2426 case Op_Phi:
2427 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2428 // The EncodeP optimization may create Phi with the same edges
2429 // for all paths. It is not handled well by Register Allocator.
2430 Node* unique_in = n->in(1);
2431 assert(unique_in != NULL, "");
2432 uint cnt = n->req();
2433 for (uint i = 2; i < cnt; i++) {
2434 Node* m = n->in(i);
2435 assert(m != NULL, "");
2436 if (unique_in != m)
2437 unique_in = NULL;
2438 }
2439 if (unique_in != NULL) {
2440 n->subsume_by(unique_in);
2441 }
2442 }
2443 break;
2445 #endif
2447 case Op_ModI:
2448 if (UseDivMod) {
2449 // Check if a%b and a/b both exist
2450 Node* d = n->find_similar(Op_DivI);
2451 if (d) {
2452 // Replace them with a fused divmod if supported
2453 Compile* C = Compile::current();
2454 if (Matcher::has_match_rule(Op_DivModI)) {
2455 DivModINode* divmod = DivModINode::make(C, n);
2456 d->subsume_by(divmod->div_proj());
2457 n->subsume_by(divmod->mod_proj());
2458 } else {
2459 // replace a%b with a-((a/b)*b)
2460 Node* mult = new (C, 3) MulINode(d, d->in(2));
2461 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2462 n->subsume_by( sub );
2463 }
2464 }
2465 }
2466 break;
2468 case Op_ModL:
2469 if (UseDivMod) {
2470 // Check if a%b and a/b both exist
2471 Node* d = n->find_similar(Op_DivL);
2472 if (d) {
2473 // Replace them with a fused divmod if supported
2474 Compile* C = Compile::current();
2475 if (Matcher::has_match_rule(Op_DivModL)) {
2476 DivModLNode* divmod = DivModLNode::make(C, n);
2477 d->subsume_by(divmod->div_proj());
2478 n->subsume_by(divmod->mod_proj());
2479 } else {
2480 // replace a%b with a-((a/b)*b)
2481 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2482 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2483 n->subsume_by( sub );
2484 }
2485 }
2486 }
2487 break;
2489 case Op_Load16B:
2490 case Op_Load8B:
2491 case Op_Load4B:
2492 case Op_Load8S:
2493 case Op_Load4S:
2494 case Op_Load2S:
2495 case Op_Load8C:
2496 case Op_Load4C:
2497 case Op_Load2C:
2498 case Op_Load4I:
2499 case Op_Load2I:
2500 case Op_Load2L:
2501 case Op_Load4F:
2502 case Op_Load2F:
2503 case Op_Load2D:
2504 case Op_Store16B:
2505 case Op_Store8B:
2506 case Op_Store4B:
2507 case Op_Store8C:
2508 case Op_Store4C:
2509 case Op_Store2C:
2510 case Op_Store4I:
2511 case Op_Store2I:
2512 case Op_Store2L:
2513 case Op_Store4F:
2514 case Op_Store2F:
2515 case Op_Store2D:
2516 break;
2518 case Op_PackB:
2519 case Op_PackS:
2520 case Op_PackC:
2521 case Op_PackI:
2522 case Op_PackF:
2523 case Op_PackL:
2524 case Op_PackD:
2525 if (n->req()-1 > 2) {
2526 // Replace many operand PackNodes with a binary tree for matching
2527 PackNode* p = (PackNode*) n;
2528 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2529 n->subsume_by(btp);
2530 }
2531 break;
2532 case Op_Loop:
2533 case Op_CountedLoop:
2534 if (n->as_Loop()->is_inner_loop()) {
2535 frc.inc_inner_loop_count();
2536 }
2537 break;
2538 case Op_LShiftI:
2539 case Op_RShiftI:
2540 case Op_URShiftI:
2541 case Op_LShiftL:
2542 case Op_RShiftL:
2543 case Op_URShiftL:
2544 if (Matcher::need_masked_shift_count) {
2545 // The cpu's shift instructions don't restrict the count to the
2546 // lower 5/6 bits. We need to do the masking ourselves.
2547 Node* in2 = n->in(2);
2548 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2549 const TypeInt* t = in2->find_int_type();
2550 if (t != NULL && t->is_con()) {
2551 juint shift = t->get_con();
2552 if (shift > mask) { // Unsigned cmp
2553 Compile* C = Compile::current();
2554 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask)));
2555 }
2556 } else {
2557 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2558 Compile* C = Compile::current();
2559 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask)));
2560 n->set_req(2, shift);
2561 }
2562 }
2563 if (in2->outcnt() == 0) { // Remove dead node
2564 in2->disconnect_inputs(NULL);
2565 }
2566 }
2567 break;
2568 default:
2569 assert( !n->is_Call(), "" );
2570 assert( !n->is_Mem(), "" );
2571 break;
2572 }
2574 // Collect CFG split points
2575 if (n->is_MultiBranch())
2576 frc._tests.push(n);
2577 }
2579 //------------------------------final_graph_reshaping_walk---------------------
2580 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2581 // requires that the walk visits a node's inputs before visiting the node.
2582 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2583 ResourceArea *area = Thread::current()->resource_area();
2584 Unique_Node_List sfpt(area);
2586 frc._visited.set(root->_idx); // first, mark node as visited
2587 uint cnt = root->req();
2588 Node *n = root;
2589 uint i = 0;
2590 while (true) {
2591 if (i < cnt) {
2592 // Place all non-visited non-null inputs onto stack
2593 Node* m = n->in(i);
2594 ++i;
2595 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2596 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2597 sfpt.push(m);
2598 cnt = m->req();
2599 nstack.push(n, i); // put on stack parent and next input's index
2600 n = m;
2601 i = 0;
2602 }
2603 } else {
2604 // Now do post-visit work
2605 final_graph_reshaping_impl( n, frc );
2606 if (nstack.is_empty())
2607 break; // finished
2608 n = nstack.node(); // Get node from stack
2609 cnt = n->req();
2610 i = nstack.index();
2611 nstack.pop(); // Shift to the next node on stack
2612 }
2613 }
2615 // Skip next transformation if compressed oops are not used.
2616 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2617 return;
2619 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2620 // It could be done for an uncommon traps or any safepoints/calls
2621 // if the DecodeN node is referenced only in a debug info.
2622 while (sfpt.size() > 0) {
2623 n = sfpt.pop();
2624 JVMState *jvms = n->as_SafePoint()->jvms();
2625 assert(jvms != NULL, "sanity");
2626 int start = jvms->debug_start();
2627 int end = n->req();
2628 bool is_uncommon = (n->is_CallStaticJava() &&
2629 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2630 for (int j = start; j < end; j++) {
2631 Node* in = n->in(j);
2632 if (in->is_DecodeN()) {
2633 bool safe_to_skip = true;
2634 if (!is_uncommon ) {
2635 // Is it safe to skip?
2636 for (uint i = 0; i < in->outcnt(); i++) {
2637 Node* u = in->raw_out(i);
2638 if (!u->is_SafePoint() ||
2639 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2640 safe_to_skip = false;
2641 }
2642 }
2643 }
2644 if (safe_to_skip) {
2645 n->set_req(j, in->in(1));
2646 }
2647 if (in->outcnt() == 0) {
2648 in->disconnect_inputs(NULL);
2649 }
2650 }
2651 }
2652 }
2653 }
2655 //------------------------------final_graph_reshaping--------------------------
2656 // Final Graph Reshaping.
2657 //
2658 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2659 // and not commoned up and forced early. Must come after regular
2660 // optimizations to avoid GVN undoing the cloning. Clone constant
2661 // inputs to Loop Phis; these will be split by the allocator anyways.
2662 // Remove Opaque nodes.
2663 // (2) Move last-uses by commutative operations to the left input to encourage
2664 // Intel update-in-place two-address operations and better register usage
2665 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2666 // calls canonicalizing them back.
2667 // (3) Count the number of double-precision FP ops, single-precision FP ops
2668 // and call sites. On Intel, we can get correct rounding either by
2669 // forcing singles to memory (requires extra stores and loads after each
2670 // FP bytecode) or we can set a rounding mode bit (requires setting and
2671 // clearing the mode bit around call sites). The mode bit is only used
2672 // if the relative frequency of single FP ops to calls is low enough.
2673 // This is a key transform for SPEC mpeg_audio.
2674 // (4) Detect infinite loops; blobs of code reachable from above but not
2675 // below. Several of the Code_Gen algorithms fail on such code shapes,
2676 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2677 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2678 // Detection is by looking for IfNodes where only 1 projection is
2679 // reachable from below or CatchNodes missing some targets.
2680 // (5) Assert for insane oop offsets in debug mode.
2682 bool Compile::final_graph_reshaping() {
2683 // an infinite loop may have been eliminated by the optimizer,
2684 // in which case the graph will be empty.
2685 if (root()->req() == 1) {
2686 record_method_not_compilable("trivial infinite loop");
2687 return true;
2688 }
2690 Final_Reshape_Counts frc;
2692 // Visit everybody reachable!
2693 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2694 Node_Stack nstack(unique() >> 1);
2695 final_graph_reshaping_walk(nstack, root(), frc);
2697 // Check for unreachable (from below) code (i.e., infinite loops).
2698 for( uint i = 0; i < frc._tests.size(); i++ ) {
2699 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2700 // Get number of CFG targets.
2701 // Note that PCTables include exception targets after calls.
2702 uint required_outcnt = n->required_outcnt();
2703 if (n->outcnt() != required_outcnt) {
2704 // Check for a few special cases. Rethrow Nodes never take the
2705 // 'fall-thru' path, so expected kids is 1 less.
2706 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2707 if (n->in(0)->in(0)->is_Call()) {
2708 CallNode *call = n->in(0)->in(0)->as_Call();
2709 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2710 required_outcnt--; // Rethrow always has 1 less kid
2711 } else if (call->req() > TypeFunc::Parms &&
2712 call->is_CallDynamicJava()) {
2713 // Check for null receiver. In such case, the optimizer has
2714 // detected that the virtual call will always result in a null
2715 // pointer exception. The fall-through projection of this CatchNode
2716 // will not be populated.
2717 Node *arg0 = call->in(TypeFunc::Parms);
2718 if (arg0->is_Type() &&
2719 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2720 required_outcnt--;
2721 }
2722 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2723 call->req() > TypeFunc::Parms+1 &&
2724 call->is_CallStaticJava()) {
2725 // Check for negative array length. In such case, the optimizer has
2726 // detected that the allocation attempt will always result in an
2727 // exception. There is no fall-through projection of this CatchNode .
2728 Node *arg1 = call->in(TypeFunc::Parms+1);
2729 if (arg1->is_Type() &&
2730 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2731 required_outcnt--;
2732 }
2733 }
2734 }
2735 }
2736 // Recheck with a better notion of 'required_outcnt'
2737 if (n->outcnt() != required_outcnt) {
2738 record_method_not_compilable("malformed control flow");
2739 return true; // Not all targets reachable!
2740 }
2741 }
2742 // Check that I actually visited all kids. Unreached kids
2743 // must be infinite loops.
2744 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2745 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2746 record_method_not_compilable("infinite loop");
2747 return true; // Found unvisited kid; must be unreach
2748 }
2749 }
2751 // If original bytecodes contained a mixture of floats and doubles
2752 // check if the optimizer has made it homogenous, item (3).
2753 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2754 frc.get_float_count() > 32 &&
2755 frc.get_double_count() == 0 &&
2756 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2757 set_24_bit_selection_and_mode( false, true );
2758 }
2760 set_java_calls(frc.get_java_call_count());
2761 set_inner_loops(frc.get_inner_loop_count());
2763 // No infinite loops, no reason to bail out.
2764 return false;
2765 }
2767 //-----------------------------too_many_traps----------------------------------
2768 // Report if there are too many traps at the current method and bci.
2769 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2770 bool Compile::too_many_traps(ciMethod* method,
2771 int bci,
2772 Deoptimization::DeoptReason reason) {
2773 ciMethodData* md = method->method_data();
2774 if (md->is_empty()) {
2775 // Assume the trap has not occurred, or that it occurred only
2776 // because of a transient condition during start-up in the interpreter.
2777 return false;
2778 }
2779 if (md->has_trap_at(bci, reason) != 0) {
2780 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2781 // Also, if there are multiple reasons, or if there is no per-BCI record,
2782 // assume the worst.
2783 if (log())
2784 log()->elem("observe trap='%s' count='%d'",
2785 Deoptimization::trap_reason_name(reason),
2786 md->trap_count(reason));
2787 return true;
2788 } else {
2789 // Ignore method/bci and see if there have been too many globally.
2790 return too_many_traps(reason, md);
2791 }
2792 }
2794 // Less-accurate variant which does not require a method and bci.
2795 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2796 ciMethodData* logmd) {
2797 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2798 // Too many traps globally.
2799 // Note that we use cumulative trap_count, not just md->trap_count.
2800 if (log()) {
2801 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2802 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2803 Deoptimization::trap_reason_name(reason),
2804 mcount, trap_count(reason));
2805 }
2806 return true;
2807 } else {
2808 // The coast is clear.
2809 return false;
2810 }
2811 }
2813 //--------------------------too_many_recompiles--------------------------------
2814 // Report if there are too many recompiles at the current method and bci.
2815 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2816 // Is not eager to return true, since this will cause the compiler to use
2817 // Action_none for a trap point, to avoid too many recompilations.
2818 bool Compile::too_many_recompiles(ciMethod* method,
2819 int bci,
2820 Deoptimization::DeoptReason reason) {
2821 ciMethodData* md = method->method_data();
2822 if (md->is_empty()) {
2823 // Assume the trap has not occurred, or that it occurred only
2824 // because of a transient condition during start-up in the interpreter.
2825 return false;
2826 }
2827 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2828 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2829 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2830 Deoptimization::DeoptReason per_bc_reason
2831 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2832 if ((per_bc_reason == Deoptimization::Reason_none
2833 || md->has_trap_at(bci, reason) != 0)
2834 // The trap frequency measure we care about is the recompile count:
2835 && md->trap_recompiled_at(bci)
2836 && md->overflow_recompile_count() >= bc_cutoff) {
2837 // Do not emit a trap here if it has already caused recompilations.
2838 // Also, if there are multiple reasons, or if there is no per-BCI record,
2839 // assume the worst.
2840 if (log())
2841 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2842 Deoptimization::trap_reason_name(reason),
2843 md->trap_count(reason),
2844 md->overflow_recompile_count());
2845 return true;
2846 } else if (trap_count(reason) != 0
2847 && decompile_count() >= m_cutoff) {
2848 // Too many recompiles globally, and we have seen this sort of trap.
2849 // Use cumulative decompile_count, not just md->decompile_count.
2850 if (log())
2851 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2852 Deoptimization::trap_reason_name(reason),
2853 md->trap_count(reason), trap_count(reason),
2854 md->decompile_count(), decompile_count());
2855 return true;
2856 } else {
2857 // The coast is clear.
2858 return false;
2859 }
2860 }
2863 #ifndef PRODUCT
2864 //------------------------------verify_graph_edges---------------------------
2865 // Walk the Graph and verify that there is a one-to-one correspondence
2866 // between Use-Def edges and Def-Use edges in the graph.
2867 void Compile::verify_graph_edges(bool no_dead_code) {
2868 if (VerifyGraphEdges) {
2869 ResourceArea *area = Thread::current()->resource_area();
2870 Unique_Node_List visited(area);
2871 // Call recursive graph walk to check edges
2872 _root->verify_edges(visited);
2873 if (no_dead_code) {
2874 // Now make sure that no visited node is used by an unvisited node.
2875 bool dead_nodes = 0;
2876 Unique_Node_List checked(area);
2877 while (visited.size() > 0) {
2878 Node* n = visited.pop();
2879 checked.push(n);
2880 for (uint i = 0; i < n->outcnt(); i++) {
2881 Node* use = n->raw_out(i);
2882 if (checked.member(use)) continue; // already checked
2883 if (visited.member(use)) continue; // already in the graph
2884 if (use->is_Con()) continue; // a dead ConNode is OK
2885 // At this point, we have found a dead node which is DU-reachable.
2886 if (dead_nodes++ == 0)
2887 tty->print_cr("*** Dead nodes reachable via DU edges:");
2888 use->dump(2);
2889 tty->print_cr("---");
2890 checked.push(use); // No repeats; pretend it is now checked.
2891 }
2892 }
2893 assert(dead_nodes == 0, "using nodes must be reachable from root");
2894 }
2895 }
2896 }
2897 #endif
2899 // The Compile object keeps track of failure reasons separately from the ciEnv.
2900 // This is required because there is not quite a 1-1 relation between the
2901 // ciEnv and its compilation task and the Compile object. Note that one
2902 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2903 // to backtrack and retry without subsuming loads. Other than this backtracking
2904 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2905 // by the logic in C2Compiler.
2906 void Compile::record_failure(const char* reason) {
2907 if (log() != NULL) {
2908 log()->elem("failure reason='%s' phase='compile'", reason);
2909 }
2910 if (_failure_reason == NULL) {
2911 // Record the first failure reason.
2912 _failure_reason = reason;
2913 }
2914 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2915 C->print_method(_failure_reason);
2916 }
2917 _root = NULL; // flush the graph, too
2918 }
2920 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2921 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2922 {
2923 if (dolog) {
2924 C = Compile::current();
2925 _log = C->log();
2926 } else {
2927 C = NULL;
2928 _log = NULL;
2929 }
2930 if (_log != NULL) {
2931 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2932 _log->stamp();
2933 _log->end_head();
2934 }
2935 }
2937 Compile::TracePhase::~TracePhase() {
2938 if (_log != NULL) {
2939 _log->done("phase nodes='%d'", C->unique());
2940 }
2941 }
2943 //=============================================================================
2944 // Two Constant's are equal when the type and the value are equal.
2945 bool Compile::Constant::operator==(const Constant& other) {
2946 if (type() != other.type() ) return false;
2947 if (can_be_reused() != other.can_be_reused()) return false;
2948 // For floating point values we compare the bit pattern.
2949 switch (type()) {
2950 case T_FLOAT: return (_value.i == other._value.i);
2951 case T_LONG:
2952 case T_DOUBLE: return (_value.j == other._value.j);
2953 case T_OBJECT:
2954 case T_ADDRESS: return (_value.l == other._value.l);
2955 case T_VOID: return (_value.l == other._value.l); // jump-table entries
2956 default: ShouldNotReachHere();
2957 }
2958 return false;
2959 }
2961 // Emit constants grouped in the following order:
2962 static BasicType type_order[] = {
2963 T_FLOAT, // 32-bit
2964 T_OBJECT, // 32 or 64-bit
2965 T_ADDRESS, // 32 or 64-bit
2966 T_DOUBLE, // 64-bit
2967 T_LONG, // 64-bit
2968 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
2969 T_ILLEGAL
2970 };
2972 static int type_to_size_in_bytes(BasicType t) {
2973 switch (t) {
2974 case T_LONG: return sizeof(jlong );
2975 case T_FLOAT: return sizeof(jfloat );
2976 case T_DOUBLE: return sizeof(jdouble);
2977 // We use T_VOID as marker for jump-table entries (labels) which
2978 // need an interal word relocation.
2979 case T_VOID:
2980 case T_ADDRESS:
2981 case T_OBJECT: return sizeof(jobject);
2982 }
2984 ShouldNotReachHere();
2985 return -1;
2986 }
2988 void Compile::ConstantTable::calculate_offsets_and_size() {
2989 int size = 0;
2990 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
2991 BasicType type = type_order[t];
2993 for (int i = 0; i < _constants.length(); i++) {
2994 Constant con = _constants.at(i);
2995 if (con.type() != type) continue; // Skip other types.
2997 // Align size for type.
2998 int typesize = type_to_size_in_bytes(con.type());
2999 size = align_size_up(size, typesize);
3001 // Set offset.
3002 con.set_offset(size);
3003 _constants.at_put(i, con);
3005 // Add type size.
3006 size = size + typesize;
3007 }
3008 }
3010 // Align size up to the next section start (which is insts; see
3011 // CodeBuffer::align_at_start).
3012 assert(_size == -1, "already set?");
3013 _size = align_size_up(size, CodeEntryAlignment);
3015 if (Matcher::constant_table_absolute_addressing) {
3016 set_table_base_offset(0); // No table base offset required
3017 } else {
3018 if (UseRDPCForConstantTableBase) {
3019 // table base offset is set in MachConstantBaseNode::emit
3020 } else {
3021 // When RDPC is not used, the table base is set into the middle of
3022 // the constant table.
3023 int half_size = _size / 2;
3024 assert(half_size * 2 == _size, "sanity");
3025 set_table_base_offset(-half_size);
3026 }
3027 }
3028 }
3030 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3031 MacroAssembler _masm(&cb);
3032 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3033 BasicType type = type_order[t];
3035 for (int i = 0; i < _constants.length(); i++) {
3036 Constant con = _constants.at(i);
3037 if (con.type() != type) continue; // Skip other types.
3039 address constant_addr;
3040 switch (con.type()) {
3041 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3042 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3043 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3044 case T_OBJECT: {
3045 jobject obj = con.get_jobject();
3046 int oop_index = _masm.oop_recorder()->find_index(obj);
3047 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3048 break;
3049 }
3050 case T_ADDRESS: {
3051 address addr = (address) con.get_jobject();
3052 constant_addr = _masm.address_constant(addr);
3053 break;
3054 }
3055 // We use T_VOID as marker for jump-table entries (labels) which
3056 // need an interal word relocation.
3057 case T_VOID: {
3058 // Write a dummy word. The real value is filled in later
3059 // in fill_jump_table_in_constant_table.
3060 address addr = (address) con.get_jobject();
3061 constant_addr = _masm.address_constant(addr);
3062 break;
3063 }
3064 default: ShouldNotReachHere();
3065 }
3066 assert(constant_addr != NULL, "consts section too small");
3067 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3068 }
3069 }
3070 }
3072 int Compile::ConstantTable::find_offset(Constant& con) const {
3073 int idx = _constants.find(con);
3074 assert(idx != -1, "constant must be in constant table");
3075 int offset = _constants.at(idx).offset();
3076 assert(offset != -1, "constant table not emitted yet?");
3077 return offset;
3078 }
3080 void Compile::ConstantTable::add(Constant& con) {
3081 if (con.can_be_reused()) {
3082 int idx = _constants.find(con);
3083 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3084 return;
3085 }
3086 }
3087 (void) _constants.append(con);
3088 }
3090 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
3091 Constant con(type, value);
3092 add(con);
3093 return con;
3094 }
3096 Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
3097 jvalue value;
3098 BasicType type = oper->type()->basic_type();
3099 switch (type) {
3100 case T_LONG: value.j = oper->constantL(); break;
3101 case T_FLOAT: value.f = oper->constantF(); break;
3102 case T_DOUBLE: value.d = oper->constantD(); break;
3103 case T_OBJECT:
3104 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3105 default: ShouldNotReachHere();
3106 }
3107 return add(type, value);
3108 }
3110 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
3111 jvalue value;
3112 // We can use the node pointer here to identify the right jump-table
3113 // as this method is called from Compile::Fill_buffer right before
3114 // the MachNodes are emitted and the jump-table is filled (means the
3115 // MachNode pointers do not change anymore).
3116 value.l = (jobject) n;
3117 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused.
3118 for (uint i = 0; i < n->outcnt(); i++) {
3119 add(con);
3120 }
3121 return con;
3122 }
3124 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3125 // If called from Compile::scratch_emit_size do nothing.
3126 if (Compile::current()->in_scratch_emit_size()) return;
3128 assert(labels.is_nonempty(), "must be");
3129 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
3131 // Since MachConstantNode::constant_offset() also contains
3132 // table_base_offset() we need to subtract the table_base_offset()
3133 // to get the plain offset into the constant table.
3134 int offset = n->constant_offset() - table_base_offset();
3136 MacroAssembler _masm(&cb);
3137 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3139 for (int i = 0; i < labels.length(); i++) {
3140 address* constant_addr = &jump_table_base[i];
3141 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
3142 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3143 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3144 }
3145 }