|
1 /* |
|
2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or |
|
21 * have any questions. |
|
22 * |
|
23 */ |
|
24 |
|
25 #include "incls/_precompiled.incl" |
|
26 #include "incls/_compile.cpp.incl" |
|
27 |
|
28 /// Support for intrinsics. |
|
29 |
|
30 // Return the index at which m must be inserted (or already exists). |
|
31 // The sort order is by the address of the ciMethod, with is_virtual as minor key. |
|
32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { |
|
33 #ifdef ASSERT |
|
34 for (int i = 1; i < _intrinsics->length(); i++) { |
|
35 CallGenerator* cg1 = _intrinsics->at(i-1); |
|
36 CallGenerator* cg2 = _intrinsics->at(i); |
|
37 assert(cg1->method() != cg2->method() |
|
38 ? cg1->method() < cg2->method() |
|
39 : cg1->is_virtual() < cg2->is_virtual(), |
|
40 "compiler intrinsics list must stay sorted"); |
|
41 } |
|
42 #endif |
|
43 // Binary search sorted list, in decreasing intervals [lo, hi]. |
|
44 int lo = 0, hi = _intrinsics->length()-1; |
|
45 while (lo <= hi) { |
|
46 int mid = (uint)(hi + lo) / 2; |
|
47 ciMethod* mid_m = _intrinsics->at(mid)->method(); |
|
48 if (m < mid_m) { |
|
49 hi = mid-1; |
|
50 } else if (m > mid_m) { |
|
51 lo = mid+1; |
|
52 } else { |
|
53 // look at minor sort key |
|
54 bool mid_virt = _intrinsics->at(mid)->is_virtual(); |
|
55 if (is_virtual < mid_virt) { |
|
56 hi = mid-1; |
|
57 } else if (is_virtual > mid_virt) { |
|
58 lo = mid+1; |
|
59 } else { |
|
60 return mid; // exact match |
|
61 } |
|
62 } |
|
63 } |
|
64 return lo; // inexact match |
|
65 } |
|
66 |
|
67 void Compile::register_intrinsic(CallGenerator* cg) { |
|
68 if (_intrinsics == NULL) { |
|
69 _intrinsics = new GrowableArray<CallGenerator*>(60); |
|
70 } |
|
71 // This code is stolen from ciObjectFactory::insert. |
|
72 // Really, GrowableArray should have methods for |
|
73 // insert_at, remove_at, and binary_search. |
|
74 int len = _intrinsics->length(); |
|
75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); |
|
76 if (index == len) { |
|
77 _intrinsics->append(cg); |
|
78 } else { |
|
79 #ifdef ASSERT |
|
80 CallGenerator* oldcg = _intrinsics->at(index); |
|
81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); |
|
82 #endif |
|
83 _intrinsics->append(_intrinsics->at(len-1)); |
|
84 int pos; |
|
85 for (pos = len-2; pos >= index; pos--) { |
|
86 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); |
|
87 } |
|
88 _intrinsics->at_put(index, cg); |
|
89 } |
|
90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); |
|
91 } |
|
92 |
|
93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { |
|
94 assert(m->is_loaded(), "don't try this on unloaded methods"); |
|
95 if (_intrinsics != NULL) { |
|
96 int index = intrinsic_insertion_index(m, is_virtual); |
|
97 if (index < _intrinsics->length() |
|
98 && _intrinsics->at(index)->method() == m |
|
99 && _intrinsics->at(index)->is_virtual() == is_virtual) { |
|
100 return _intrinsics->at(index); |
|
101 } |
|
102 } |
|
103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. |
|
104 if (m->intrinsic_id() != vmIntrinsics::_none) { |
|
105 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); |
|
106 if (cg != NULL) { |
|
107 // Save it for next time: |
|
108 register_intrinsic(cg); |
|
109 return cg; |
|
110 } else { |
|
111 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); |
|
112 } |
|
113 } |
|
114 return NULL; |
|
115 } |
|
116 |
|
117 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined |
|
118 // in library_call.cpp. |
|
119 |
|
120 |
|
121 #ifndef PRODUCT |
|
122 // statistics gathering... |
|
123 |
|
124 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; |
|
125 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; |
|
126 |
|
127 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { |
|
128 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); |
|
129 int oflags = _intrinsic_hist_flags[id]; |
|
130 assert(flags != 0, "what happened?"); |
|
131 if (is_virtual) { |
|
132 flags |= _intrinsic_virtual; |
|
133 } |
|
134 bool changed = (flags != oflags); |
|
135 if ((flags & _intrinsic_worked) != 0) { |
|
136 juint count = (_intrinsic_hist_count[id] += 1); |
|
137 if (count == 1) { |
|
138 changed = true; // first time |
|
139 } |
|
140 // increment the overall count also: |
|
141 _intrinsic_hist_count[vmIntrinsics::_none] += 1; |
|
142 } |
|
143 if (changed) { |
|
144 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { |
|
145 // Something changed about the intrinsic's virtuality. |
|
146 if ((flags & _intrinsic_virtual) != 0) { |
|
147 // This is the first use of this intrinsic as a virtual call. |
|
148 if (oflags != 0) { |
|
149 // We already saw it as a non-virtual, so note both cases. |
|
150 flags |= _intrinsic_both; |
|
151 } |
|
152 } else if ((oflags & _intrinsic_both) == 0) { |
|
153 // This is the first use of this intrinsic as a non-virtual |
|
154 flags |= _intrinsic_both; |
|
155 } |
|
156 } |
|
157 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); |
|
158 } |
|
159 // update the overall flags also: |
|
160 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; |
|
161 return changed; |
|
162 } |
|
163 |
|
164 static char* format_flags(int flags, char* buf) { |
|
165 buf[0] = 0; |
|
166 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); |
|
167 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); |
|
168 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); |
|
169 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); |
|
170 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); |
|
171 if (buf[0] == 0) strcat(buf, ","); |
|
172 assert(buf[0] == ',', "must be"); |
|
173 return &buf[1]; |
|
174 } |
|
175 |
|
176 void Compile::print_intrinsic_statistics() { |
|
177 char flagsbuf[100]; |
|
178 ttyLocker ttyl; |
|
179 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); |
|
180 tty->print_cr("Compiler intrinsic usage:"); |
|
181 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; |
|
182 if (total == 0) total = 1; // avoid div0 in case of no successes |
|
183 #define PRINT_STAT_LINE(name, c, f) \ |
|
184 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); |
|
185 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { |
|
186 vmIntrinsics::ID id = (vmIntrinsics::ID) index; |
|
187 int flags = _intrinsic_hist_flags[id]; |
|
188 juint count = _intrinsic_hist_count[id]; |
|
189 if ((flags | count) != 0) { |
|
190 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); |
|
191 } |
|
192 } |
|
193 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); |
|
194 if (xtty != NULL) xtty->tail("statistics"); |
|
195 } |
|
196 |
|
197 void Compile::print_statistics() { |
|
198 { ttyLocker ttyl; |
|
199 if (xtty != NULL) xtty->head("statistics type='opto'"); |
|
200 Parse::print_statistics(); |
|
201 PhaseCCP::print_statistics(); |
|
202 PhaseRegAlloc::print_statistics(); |
|
203 Scheduling::print_statistics(); |
|
204 PhasePeephole::print_statistics(); |
|
205 PhaseIdealLoop::print_statistics(); |
|
206 if (xtty != NULL) xtty->tail("statistics"); |
|
207 } |
|
208 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { |
|
209 // put this under its own <statistics> element. |
|
210 print_intrinsic_statistics(); |
|
211 } |
|
212 } |
|
213 #endif //PRODUCT |
|
214 |
|
215 // Support for bundling info |
|
216 Bundle* Compile::node_bundling(const Node *n) { |
|
217 assert(valid_bundle_info(n), "oob"); |
|
218 return &_node_bundling_base[n->_idx]; |
|
219 } |
|
220 |
|
221 bool Compile::valid_bundle_info(const Node *n) { |
|
222 return (_node_bundling_limit > n->_idx); |
|
223 } |
|
224 |
|
225 |
|
226 // Identify all nodes that are reachable from below, useful. |
|
227 // Use breadth-first pass that records state in a Unique_Node_List, |
|
228 // recursive traversal is slower. |
|
229 void Compile::identify_useful_nodes(Unique_Node_List &useful) { |
|
230 int estimated_worklist_size = unique(); |
|
231 useful.map( estimated_worklist_size, NULL ); // preallocate space |
|
232 |
|
233 // Initialize worklist |
|
234 if (root() != NULL) { useful.push(root()); } |
|
235 // If 'top' is cached, declare it useful to preserve cached node |
|
236 if( cached_top_node() ) { useful.push(cached_top_node()); } |
|
237 |
|
238 // Push all useful nodes onto the list, breadthfirst |
|
239 for( uint next = 0; next < useful.size(); ++next ) { |
|
240 assert( next < unique(), "Unique useful nodes < total nodes"); |
|
241 Node *n = useful.at(next); |
|
242 uint max = n->len(); |
|
243 for( uint i = 0; i < max; ++i ) { |
|
244 Node *m = n->in(i); |
|
245 if( m == NULL ) continue; |
|
246 useful.push(m); |
|
247 } |
|
248 } |
|
249 } |
|
250 |
|
251 // Disconnect all useless nodes by disconnecting those at the boundary. |
|
252 void Compile::remove_useless_nodes(Unique_Node_List &useful) { |
|
253 uint next = 0; |
|
254 while( next < useful.size() ) { |
|
255 Node *n = useful.at(next++); |
|
256 // Use raw traversal of out edges since this code removes out edges |
|
257 int max = n->outcnt(); |
|
258 for (int j = 0; j < max; ++j ) { |
|
259 Node* child = n->raw_out(j); |
|
260 if( ! useful.member(child) ) { |
|
261 assert( !child->is_top() || child != top(), |
|
262 "If top is cached in Compile object it is in useful list"); |
|
263 // Only need to remove this out-edge to the useless node |
|
264 n->raw_del_out(j); |
|
265 --j; |
|
266 --max; |
|
267 } |
|
268 } |
|
269 if (n->outcnt() == 1 && n->has_special_unique_user()) { |
|
270 record_for_igvn( n->unique_out() ); |
|
271 } |
|
272 } |
|
273 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) |
|
274 } |
|
275 |
|
276 //------------------------------frame_size_in_words----------------------------- |
|
277 // frame_slots in units of words |
|
278 int Compile::frame_size_in_words() const { |
|
279 // shift is 0 in LP32 and 1 in LP64 |
|
280 const int shift = (LogBytesPerWord - LogBytesPerInt); |
|
281 int words = _frame_slots >> shift; |
|
282 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); |
|
283 return words; |
|
284 } |
|
285 |
|
286 // ============================================================================ |
|
287 //------------------------------CompileWrapper--------------------------------- |
|
288 class CompileWrapper : public StackObj { |
|
289 Compile *const _compile; |
|
290 public: |
|
291 CompileWrapper(Compile* compile); |
|
292 |
|
293 ~CompileWrapper(); |
|
294 }; |
|
295 |
|
296 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { |
|
297 // the Compile* pointer is stored in the current ciEnv: |
|
298 ciEnv* env = compile->env(); |
|
299 assert(env == ciEnv::current(), "must already be a ciEnv active"); |
|
300 assert(env->compiler_data() == NULL, "compile already active?"); |
|
301 env->set_compiler_data(compile); |
|
302 assert(compile == Compile::current(), "sanity"); |
|
303 |
|
304 compile->set_type_dict(NULL); |
|
305 compile->set_type_hwm(NULL); |
|
306 compile->set_type_last_size(0); |
|
307 compile->set_last_tf(NULL, NULL); |
|
308 compile->set_indexSet_arena(NULL); |
|
309 compile->set_indexSet_free_block_list(NULL); |
|
310 compile->init_type_arena(); |
|
311 Type::Initialize(compile); |
|
312 _compile->set_scratch_buffer_blob(NULL); |
|
313 _compile->begin_method(); |
|
314 } |
|
315 CompileWrapper::~CompileWrapper() { |
|
316 if (_compile->failing()) { |
|
317 _compile->print_method("Failed"); |
|
318 } |
|
319 _compile->end_method(); |
|
320 if (_compile->scratch_buffer_blob() != NULL) |
|
321 BufferBlob::free(_compile->scratch_buffer_blob()); |
|
322 _compile->env()->set_compiler_data(NULL); |
|
323 } |
|
324 |
|
325 |
|
326 //----------------------------print_compile_messages--------------------------- |
|
327 void Compile::print_compile_messages() { |
|
328 #ifndef PRODUCT |
|
329 // Check if recompiling |
|
330 if (_subsume_loads == false && PrintOpto) { |
|
331 // Recompiling without allowing machine instructions to subsume loads |
|
332 tty->print_cr("*********************************************************"); |
|
333 tty->print_cr("** Bailout: Recompile without subsuming loads **"); |
|
334 tty->print_cr("*********************************************************"); |
|
335 } |
|
336 if (env()->break_at_compile()) { |
|
337 // Open the debugger when compiing this method. |
|
338 tty->print("### Breaking when compiling: "); |
|
339 method()->print_short_name(); |
|
340 tty->cr(); |
|
341 BREAKPOINT; |
|
342 } |
|
343 |
|
344 if( PrintOpto ) { |
|
345 if (is_osr_compilation()) { |
|
346 tty->print("[OSR]%3d", _compile_id); |
|
347 } else { |
|
348 tty->print("%3d", _compile_id); |
|
349 } |
|
350 } |
|
351 #endif |
|
352 } |
|
353 |
|
354 |
|
355 void Compile::init_scratch_buffer_blob() { |
|
356 if( scratch_buffer_blob() != NULL ) return; |
|
357 |
|
358 // Construct a temporary CodeBuffer to have it construct a BufferBlob |
|
359 // Cache this BufferBlob for this compile. |
|
360 ResourceMark rm; |
|
361 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); |
|
362 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); |
|
363 // Record the buffer blob for next time. |
|
364 set_scratch_buffer_blob(blob); |
|
365 guarantee(scratch_buffer_blob() != NULL, "Need BufferBlob for code generation"); |
|
366 |
|
367 // Initialize the relocation buffers |
|
368 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; |
|
369 set_scratch_locs_memory(locs_buf); |
|
370 } |
|
371 |
|
372 |
|
373 //-----------------------scratch_emit_size------------------------------------- |
|
374 // Helper function that computes size by emitting code |
|
375 uint Compile::scratch_emit_size(const Node* n) { |
|
376 // Emit into a trash buffer and count bytes emitted. |
|
377 // This is a pretty expensive way to compute a size, |
|
378 // but it works well enough if seldom used. |
|
379 // All common fixed-size instructions are given a size |
|
380 // method by the AD file. |
|
381 // Note that the scratch buffer blob and locs memory are |
|
382 // allocated at the beginning of the compile task, and |
|
383 // may be shared by several calls to scratch_emit_size. |
|
384 // The allocation of the scratch buffer blob is particularly |
|
385 // expensive, since it has to grab the code cache lock. |
|
386 BufferBlob* blob = this->scratch_buffer_blob(); |
|
387 assert(blob != NULL, "Initialize BufferBlob at start"); |
|
388 assert(blob->size() > MAX_inst_size, "sanity"); |
|
389 relocInfo* locs_buf = scratch_locs_memory(); |
|
390 address blob_begin = blob->instructions_begin(); |
|
391 address blob_end = (address)locs_buf; |
|
392 assert(blob->instructions_contains(blob_end), "sanity"); |
|
393 CodeBuffer buf(blob_begin, blob_end - blob_begin); |
|
394 buf.initialize_consts_size(MAX_const_size); |
|
395 buf.initialize_stubs_size(MAX_stubs_size); |
|
396 assert(locs_buf != NULL, "sanity"); |
|
397 int lsize = MAX_locs_size / 2; |
|
398 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); |
|
399 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); |
|
400 n->emit(buf, this->regalloc()); |
|
401 return buf.code_size(); |
|
402 } |
|
403 |
|
404 void Compile::record_for_escape_analysis(Node* n) { |
|
405 if (_congraph != NULL) |
|
406 _congraph->record_for_escape_analysis(n); |
|
407 } |
|
408 |
|
409 |
|
410 // ============================================================================ |
|
411 //------------------------------Compile standard------------------------------- |
|
412 debug_only( int Compile::_debug_idx = 100000; ) |
|
413 |
|
414 // Compile a method. entry_bci is -1 for normal compilations and indicates |
|
415 // the continuation bci for on stack replacement. |
|
416 |
|
417 |
|
418 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads ) |
|
419 : Phase(Compiler), |
|
420 _env(ci_env), |
|
421 _log(ci_env->log()), |
|
422 _compile_id(ci_env->compile_id()), |
|
423 _save_argument_registers(false), |
|
424 _stub_name(NULL), |
|
425 _stub_function(NULL), |
|
426 _stub_entry_point(NULL), |
|
427 _method(target), |
|
428 _entry_bci(osr_bci), |
|
429 _initial_gvn(NULL), |
|
430 _for_igvn(NULL), |
|
431 _warm_calls(NULL), |
|
432 _subsume_loads(subsume_loads), |
|
433 _failure_reason(NULL), |
|
434 _code_buffer("Compile::Fill_buffer"), |
|
435 _orig_pc_slot(0), |
|
436 _orig_pc_slot_offset_in_bytes(0), |
|
437 _node_bundling_limit(0), |
|
438 _node_bundling_base(NULL), |
|
439 #ifndef PRODUCT |
|
440 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), |
|
441 _printer(IdealGraphPrinter::printer()), |
|
442 #endif |
|
443 _congraph(NULL) { |
|
444 C = this; |
|
445 |
|
446 CompileWrapper cw(this); |
|
447 #ifndef PRODUCT |
|
448 if (TimeCompiler2) { |
|
449 tty->print(" "); |
|
450 target->holder()->name()->print(); |
|
451 tty->print("."); |
|
452 target->print_short_name(); |
|
453 tty->print(" "); |
|
454 } |
|
455 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); |
|
456 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); |
|
457 set_print_assembly(PrintOptoAssembly || _method->should_print_assembly()); |
|
458 #endif |
|
459 |
|
460 if (ProfileTraps) { |
|
461 // Make sure the method being compiled gets its own MDO, |
|
462 // so we can at least track the decompile_count(). |
|
463 method()->build_method_data(); |
|
464 } |
|
465 |
|
466 Init(::AliasLevel); |
|
467 |
|
468 |
|
469 print_compile_messages(); |
|
470 |
|
471 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) |
|
472 _ilt = InlineTree::build_inline_tree_root(); |
|
473 else |
|
474 _ilt = NULL; |
|
475 |
|
476 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice |
|
477 assert(num_alias_types() >= AliasIdxRaw, ""); |
|
478 |
|
479 #define MINIMUM_NODE_HASH 1023 |
|
480 // Node list that Iterative GVN will start with |
|
481 Unique_Node_List for_igvn(comp_arena()); |
|
482 set_for_igvn(&for_igvn); |
|
483 |
|
484 // GVN that will be run immediately on new nodes |
|
485 uint estimated_size = method()->code_size()*4+64; |
|
486 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); |
|
487 PhaseGVN gvn(node_arena(), estimated_size); |
|
488 set_initial_gvn(&gvn); |
|
489 |
|
490 if (DoEscapeAnalysis) |
|
491 _congraph = new ConnectionGraph(this); |
|
492 |
|
493 { // Scope for timing the parser |
|
494 TracePhase t3("parse", &_t_parser, true); |
|
495 |
|
496 // Put top into the hash table ASAP. |
|
497 initial_gvn()->transform_no_reclaim(top()); |
|
498 |
|
499 // Set up tf(), start(), and find a CallGenerator. |
|
500 CallGenerator* cg; |
|
501 if (is_osr_compilation()) { |
|
502 const TypeTuple *domain = StartOSRNode::osr_domain(); |
|
503 const TypeTuple *range = TypeTuple::make_range(method()->signature()); |
|
504 init_tf(TypeFunc::make(domain, range)); |
|
505 StartNode* s = new (this, 2) StartOSRNode(root(), domain); |
|
506 initial_gvn()->set_type_bottom(s); |
|
507 init_start(s); |
|
508 cg = CallGenerator::for_osr(method(), entry_bci()); |
|
509 } else { |
|
510 // Normal case. |
|
511 init_tf(TypeFunc::make(method())); |
|
512 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); |
|
513 initial_gvn()->set_type_bottom(s); |
|
514 init_start(s); |
|
515 float past_uses = method()->interpreter_invocation_count(); |
|
516 float expected_uses = past_uses; |
|
517 cg = CallGenerator::for_inline(method(), expected_uses); |
|
518 } |
|
519 if (failing()) return; |
|
520 if (cg == NULL) { |
|
521 record_method_not_compilable_all_tiers("cannot parse method"); |
|
522 return; |
|
523 } |
|
524 JVMState* jvms = build_start_state(start(), tf()); |
|
525 if ((jvms = cg->generate(jvms)) == NULL) { |
|
526 record_method_not_compilable("method parse failed"); |
|
527 return; |
|
528 } |
|
529 GraphKit kit(jvms); |
|
530 |
|
531 if (!kit.stopped()) { |
|
532 // Accept return values, and transfer control we know not where. |
|
533 // This is done by a special, unique ReturnNode bound to root. |
|
534 return_values(kit.jvms()); |
|
535 } |
|
536 |
|
537 if (kit.has_exceptions()) { |
|
538 // Any exceptions that escape from this call must be rethrown |
|
539 // to whatever caller is dynamically above us on the stack. |
|
540 // This is done by a special, unique RethrowNode bound to root. |
|
541 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); |
|
542 } |
|
543 |
|
544 // Remove clutter produced by parsing. |
|
545 if (!failing()) { |
|
546 ResourceMark rm; |
|
547 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); |
|
548 } |
|
549 } |
|
550 |
|
551 // Note: Large methods are capped off in do_one_bytecode(). |
|
552 if (failing()) return; |
|
553 |
|
554 // After parsing, node notes are no longer automagic. |
|
555 // They must be propagated by register_new_node_with_optimizer(), |
|
556 // clone(), or the like. |
|
557 set_default_node_notes(NULL); |
|
558 |
|
559 for (;;) { |
|
560 int successes = Inline_Warm(); |
|
561 if (failing()) return; |
|
562 if (successes == 0) break; |
|
563 } |
|
564 |
|
565 // Drain the list. |
|
566 Finish_Warm(); |
|
567 #ifndef PRODUCT |
|
568 if (_printer) { |
|
569 _printer->print_inlining(this); |
|
570 } |
|
571 #endif |
|
572 |
|
573 if (failing()) return; |
|
574 NOT_PRODUCT( verify_graph_edges(); ) |
|
575 |
|
576 // Perform escape analysis |
|
577 if (_congraph != NULL) { |
|
578 NOT_PRODUCT( TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, TimeCompiler); ) |
|
579 _congraph->compute_escape(); |
|
580 #ifndef PRODUCT |
|
581 if (PrintEscapeAnalysis) { |
|
582 _congraph->dump(); |
|
583 } |
|
584 #endif |
|
585 } |
|
586 // Now optimize |
|
587 Optimize(); |
|
588 if (failing()) return; |
|
589 NOT_PRODUCT( verify_graph_edges(); ) |
|
590 |
|
591 #ifndef PRODUCT |
|
592 if (PrintIdeal) { |
|
593 ttyLocker ttyl; // keep the following output all in one block |
|
594 // This output goes directly to the tty, not the compiler log. |
|
595 // To enable tools to match it up with the compilation activity, |
|
596 // be sure to tag this tty output with the compile ID. |
|
597 if (xtty != NULL) { |
|
598 xtty->head("ideal compile_id='%d'%s", compile_id(), |
|
599 is_osr_compilation() ? " compile_kind='osr'" : |
|
600 ""); |
|
601 } |
|
602 root()->dump(9999); |
|
603 if (xtty != NULL) { |
|
604 xtty->tail("ideal"); |
|
605 } |
|
606 } |
|
607 #endif |
|
608 |
|
609 // Now that we know the size of all the monitors we can add a fixed slot |
|
610 // for the original deopt pc. |
|
611 |
|
612 _orig_pc_slot = fixed_slots(); |
|
613 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); |
|
614 set_fixed_slots(next_slot); |
|
615 |
|
616 // Now generate code |
|
617 Code_Gen(); |
|
618 if (failing()) return; |
|
619 |
|
620 // Check if we want to skip execution of all compiled code. |
|
621 { |
|
622 #ifndef PRODUCT |
|
623 if (OptoNoExecute) { |
|
624 record_method_not_compilable("+OptoNoExecute"); // Flag as failed |
|
625 return; |
|
626 } |
|
627 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); |
|
628 #endif |
|
629 |
|
630 if (is_osr_compilation()) { |
|
631 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); |
|
632 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); |
|
633 } else { |
|
634 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); |
|
635 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); |
|
636 } |
|
637 |
|
638 env()->register_method(_method, _entry_bci, |
|
639 &_code_offsets, |
|
640 _orig_pc_slot_offset_in_bytes, |
|
641 code_buffer(), |
|
642 frame_size_in_words(), _oop_map_set, |
|
643 &_handler_table, &_inc_table, |
|
644 compiler, |
|
645 env()->comp_level(), |
|
646 true, /*has_debug_info*/ |
|
647 has_unsafe_access() |
|
648 ); |
|
649 } |
|
650 } |
|
651 |
|
652 //------------------------------Compile---------------------------------------- |
|
653 // Compile a runtime stub |
|
654 Compile::Compile( ciEnv* ci_env, |
|
655 TypeFunc_generator generator, |
|
656 address stub_function, |
|
657 const char *stub_name, |
|
658 int is_fancy_jump, |
|
659 bool pass_tls, |
|
660 bool save_arg_registers, |
|
661 bool return_pc ) |
|
662 : Phase(Compiler), |
|
663 _env(ci_env), |
|
664 _log(ci_env->log()), |
|
665 _compile_id(-1), |
|
666 _save_argument_registers(save_arg_registers), |
|
667 _method(NULL), |
|
668 _stub_name(stub_name), |
|
669 _stub_function(stub_function), |
|
670 _stub_entry_point(NULL), |
|
671 _entry_bci(InvocationEntryBci), |
|
672 _initial_gvn(NULL), |
|
673 _for_igvn(NULL), |
|
674 _warm_calls(NULL), |
|
675 _orig_pc_slot(0), |
|
676 _orig_pc_slot_offset_in_bytes(0), |
|
677 _subsume_loads(true), |
|
678 _failure_reason(NULL), |
|
679 _code_buffer("Compile::Fill_buffer"), |
|
680 _node_bundling_limit(0), |
|
681 _node_bundling_base(NULL), |
|
682 #ifndef PRODUCT |
|
683 _trace_opto_output(TraceOptoOutput), |
|
684 _printer(NULL), |
|
685 #endif |
|
686 _congraph(NULL) { |
|
687 C = this; |
|
688 |
|
689 #ifndef PRODUCT |
|
690 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); |
|
691 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); |
|
692 set_print_assembly(PrintFrameConverterAssembly); |
|
693 #endif |
|
694 CompileWrapper cw(this); |
|
695 Init(/*AliasLevel=*/ 0); |
|
696 init_tf((*generator)()); |
|
697 |
|
698 { |
|
699 // The following is a dummy for the sake of GraphKit::gen_stub |
|
700 Unique_Node_List for_igvn(comp_arena()); |
|
701 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this |
|
702 PhaseGVN gvn(Thread::current()->resource_area(),255); |
|
703 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively |
|
704 gvn.transform_no_reclaim(top()); |
|
705 |
|
706 GraphKit kit; |
|
707 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); |
|
708 } |
|
709 |
|
710 NOT_PRODUCT( verify_graph_edges(); ) |
|
711 Code_Gen(); |
|
712 if (failing()) return; |
|
713 |
|
714 |
|
715 // Entry point will be accessed using compile->stub_entry_point(); |
|
716 if (code_buffer() == NULL) { |
|
717 Matcher::soft_match_failure(); |
|
718 } else { |
|
719 if (PrintAssembly && (WizardMode || Verbose)) |
|
720 tty->print_cr("### Stub::%s", stub_name); |
|
721 |
|
722 if (!failing()) { |
|
723 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); |
|
724 |
|
725 // Make the NMethod |
|
726 // For now we mark the frame as never safe for profile stackwalking |
|
727 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, |
|
728 code_buffer(), |
|
729 CodeOffsets::frame_never_safe, |
|
730 // _code_offsets.value(CodeOffsets::Frame_Complete), |
|
731 frame_size_in_words(), |
|
732 _oop_map_set, |
|
733 save_arg_registers); |
|
734 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); |
|
735 |
|
736 _stub_entry_point = rs->entry_point(); |
|
737 } |
|
738 } |
|
739 } |
|
740 |
|
741 #ifndef PRODUCT |
|
742 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { |
|
743 if(PrintOpto && Verbose) { |
|
744 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); |
|
745 } |
|
746 } |
|
747 #endif |
|
748 |
|
749 void Compile::print_codes() { |
|
750 } |
|
751 |
|
752 //------------------------------Init------------------------------------------- |
|
753 // Prepare for a single compilation |
|
754 void Compile::Init(int aliaslevel) { |
|
755 _unique = 0; |
|
756 _regalloc = NULL; |
|
757 |
|
758 _tf = NULL; // filled in later |
|
759 _top = NULL; // cached later |
|
760 _matcher = NULL; // filled in later |
|
761 _cfg = NULL; // filled in later |
|
762 |
|
763 set_24_bit_selection_and_mode(Use24BitFP, false); |
|
764 |
|
765 _node_note_array = NULL; |
|
766 _default_node_notes = NULL; |
|
767 |
|
768 _immutable_memory = NULL; // filled in at first inquiry |
|
769 |
|
770 // Globally visible Nodes |
|
771 // First set TOP to NULL to give safe behavior during creation of RootNode |
|
772 set_cached_top_node(NULL); |
|
773 set_root(new (this, 3) RootNode()); |
|
774 // Now that you have a Root to point to, create the real TOP |
|
775 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); |
|
776 set_recent_alloc(NULL, NULL); |
|
777 |
|
778 // Create Debug Information Recorder to record scopes, oopmaps, etc. |
|
779 env()->set_oop_recorder(new OopRecorder(comp_arena())); |
|
780 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); |
|
781 env()->set_dependencies(new Dependencies(env())); |
|
782 |
|
783 _fixed_slots = 0; |
|
784 set_has_split_ifs(false); |
|
785 set_has_loops(has_method() && method()->has_loops()); // first approximation |
|
786 _deopt_happens = true; // start out assuming the worst |
|
787 _trap_can_recompile = false; // no traps emitted yet |
|
788 _major_progress = true; // start out assuming good things will happen |
|
789 set_has_unsafe_access(false); |
|
790 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); |
|
791 set_decompile_count(0); |
|
792 |
|
793 // Compilation level related initialization |
|
794 if (env()->comp_level() == CompLevel_fast_compile) { |
|
795 set_num_loop_opts(Tier1LoopOptsCount); |
|
796 set_do_inlining(Tier1Inline != 0); |
|
797 set_max_inline_size(Tier1MaxInlineSize); |
|
798 set_freq_inline_size(Tier1FreqInlineSize); |
|
799 set_do_scheduling(false); |
|
800 set_do_count_invocations(Tier1CountInvocations); |
|
801 set_do_method_data_update(Tier1UpdateMethodData); |
|
802 } else { |
|
803 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); |
|
804 set_num_loop_opts(LoopOptsCount); |
|
805 set_do_inlining(Inline); |
|
806 set_max_inline_size(MaxInlineSize); |
|
807 set_freq_inline_size(FreqInlineSize); |
|
808 set_do_scheduling(OptoScheduling); |
|
809 set_do_count_invocations(false); |
|
810 set_do_method_data_update(false); |
|
811 } |
|
812 |
|
813 if (debug_info()->recording_non_safepoints()) { |
|
814 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> |
|
815 (comp_arena(), 8, 0, NULL)); |
|
816 set_default_node_notes(Node_Notes::make(this)); |
|
817 } |
|
818 |
|
819 // // -- Initialize types before each compile -- |
|
820 // // Update cached type information |
|
821 // if( _method && _method->constants() ) |
|
822 // Type::update_loaded_types(_method, _method->constants()); |
|
823 |
|
824 // Init alias_type map. |
|
825 if (!DoEscapeAnalysis && aliaslevel == 3) |
|
826 aliaslevel = 2; // No unique types without escape analysis |
|
827 _AliasLevel = aliaslevel; |
|
828 const int grow_ats = 16; |
|
829 _max_alias_types = grow_ats; |
|
830 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); |
|
831 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); |
|
832 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); |
|
833 { |
|
834 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; |
|
835 } |
|
836 // Initialize the first few types. |
|
837 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); |
|
838 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); |
|
839 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); |
|
840 _num_alias_types = AliasIdxRaw+1; |
|
841 // Zero out the alias type cache. |
|
842 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); |
|
843 // A NULL adr_type hits in the cache right away. Preload the right answer. |
|
844 probe_alias_cache(NULL)->_index = AliasIdxTop; |
|
845 |
|
846 _intrinsics = NULL; |
|
847 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); |
|
848 register_library_intrinsics(); |
|
849 } |
|
850 |
|
851 //---------------------------init_start---------------------------------------- |
|
852 // Install the StartNode on this compile object. |
|
853 void Compile::init_start(StartNode* s) { |
|
854 if (failing()) |
|
855 return; // already failing |
|
856 assert(s == start(), ""); |
|
857 } |
|
858 |
|
859 StartNode* Compile::start() const { |
|
860 assert(!failing(), ""); |
|
861 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { |
|
862 Node* start = root()->fast_out(i); |
|
863 if( start->is_Start() ) |
|
864 return start->as_Start(); |
|
865 } |
|
866 ShouldNotReachHere(); |
|
867 return NULL; |
|
868 } |
|
869 |
|
870 //-------------------------------immutable_memory------------------------------------- |
|
871 // Access immutable memory |
|
872 Node* Compile::immutable_memory() { |
|
873 if (_immutable_memory != NULL) { |
|
874 return _immutable_memory; |
|
875 } |
|
876 StartNode* s = start(); |
|
877 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { |
|
878 Node *p = s->fast_out(i); |
|
879 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { |
|
880 _immutable_memory = p; |
|
881 return _immutable_memory; |
|
882 } |
|
883 } |
|
884 ShouldNotReachHere(); |
|
885 return NULL; |
|
886 } |
|
887 |
|
888 //----------------------set_cached_top_node------------------------------------ |
|
889 // Install the cached top node, and make sure Node::is_top works correctly. |
|
890 void Compile::set_cached_top_node(Node* tn) { |
|
891 if (tn != NULL) verify_top(tn); |
|
892 Node* old_top = _top; |
|
893 _top = tn; |
|
894 // Calling Node::setup_is_top allows the nodes the chance to adjust |
|
895 // their _out arrays. |
|
896 if (_top != NULL) _top->setup_is_top(); |
|
897 if (old_top != NULL) old_top->setup_is_top(); |
|
898 assert(_top == NULL || top()->is_top(), ""); |
|
899 } |
|
900 |
|
901 #ifndef PRODUCT |
|
902 void Compile::verify_top(Node* tn) const { |
|
903 if (tn != NULL) { |
|
904 assert(tn->is_Con(), "top node must be a constant"); |
|
905 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); |
|
906 assert(tn->in(0) != NULL, "must have live top node"); |
|
907 } |
|
908 } |
|
909 #endif |
|
910 |
|
911 |
|
912 ///-------------------Managing Per-Node Debug & Profile Info------------------- |
|
913 |
|
914 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { |
|
915 guarantee(arr != NULL, ""); |
|
916 int num_blocks = arr->length(); |
|
917 if (grow_by < num_blocks) grow_by = num_blocks; |
|
918 int num_notes = grow_by * _node_notes_block_size; |
|
919 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); |
|
920 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); |
|
921 while (num_notes > 0) { |
|
922 arr->append(notes); |
|
923 notes += _node_notes_block_size; |
|
924 num_notes -= _node_notes_block_size; |
|
925 } |
|
926 assert(num_notes == 0, "exact multiple, please"); |
|
927 } |
|
928 |
|
929 bool Compile::copy_node_notes_to(Node* dest, Node* source) { |
|
930 if (source == NULL || dest == NULL) return false; |
|
931 |
|
932 if (dest->is_Con()) |
|
933 return false; // Do not push debug info onto constants. |
|
934 |
|
935 #ifdef ASSERT |
|
936 // Leave a bread crumb trail pointing to the original node: |
|
937 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { |
|
938 dest->set_debug_orig(source); |
|
939 } |
|
940 #endif |
|
941 |
|
942 if (node_note_array() == NULL) |
|
943 return false; // Not collecting any notes now. |
|
944 |
|
945 // This is a copy onto a pre-existing node, which may already have notes. |
|
946 // If both nodes have notes, do not overwrite any pre-existing notes. |
|
947 Node_Notes* source_notes = node_notes_at(source->_idx); |
|
948 if (source_notes == NULL || source_notes->is_clear()) return false; |
|
949 Node_Notes* dest_notes = node_notes_at(dest->_idx); |
|
950 if (dest_notes == NULL || dest_notes->is_clear()) { |
|
951 return set_node_notes_at(dest->_idx, source_notes); |
|
952 } |
|
953 |
|
954 Node_Notes merged_notes = (*source_notes); |
|
955 // The order of operations here ensures that dest notes will win... |
|
956 merged_notes.update_from(dest_notes); |
|
957 return set_node_notes_at(dest->_idx, &merged_notes); |
|
958 } |
|
959 |
|
960 |
|
961 //--------------------------allow_range_check_smearing------------------------- |
|
962 // Gating condition for coalescing similar range checks. |
|
963 // Sometimes we try 'speculatively' replacing a series of a range checks by a |
|
964 // single covering check that is at least as strong as any of them. |
|
965 // If the optimization succeeds, the simplified (strengthened) range check |
|
966 // will always succeed. If it fails, we will deopt, and then give up |
|
967 // on the optimization. |
|
968 bool Compile::allow_range_check_smearing() const { |
|
969 // If this method has already thrown a range-check, |
|
970 // assume it was because we already tried range smearing |
|
971 // and it failed. |
|
972 uint already_trapped = trap_count(Deoptimization::Reason_range_check); |
|
973 return !already_trapped; |
|
974 } |
|
975 |
|
976 |
|
977 //------------------------------flatten_alias_type----------------------------- |
|
978 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { |
|
979 int offset = tj->offset(); |
|
980 TypePtr::PTR ptr = tj->ptr(); |
|
981 |
|
982 // Process weird unsafe references. |
|
983 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { |
|
984 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); |
|
985 tj = TypeOopPtr::BOTTOM; |
|
986 ptr = tj->ptr(); |
|
987 offset = tj->offset(); |
|
988 } |
|
989 |
|
990 // Array pointers need some flattening |
|
991 const TypeAryPtr *ta = tj->isa_aryptr(); |
|
992 if( ta && _AliasLevel >= 2 ) { |
|
993 // For arrays indexed by constant indices, we flatten the alias |
|
994 // space to include all of the array body. Only the header, klass |
|
995 // and array length can be accessed un-aliased. |
|
996 if( offset != Type::OffsetBot ) { |
|
997 if( ta->const_oop() ) { // methodDataOop or methodOop |
|
998 offset = Type::OffsetBot; // Flatten constant access into array body |
|
999 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); |
|
1000 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { |
|
1001 // range is OK as-is. |
|
1002 tj = ta = TypeAryPtr::RANGE; |
|
1003 } else if( offset == oopDesc::klass_offset_in_bytes() ) { |
|
1004 tj = TypeInstPtr::KLASS; // all klass loads look alike |
|
1005 ta = TypeAryPtr::RANGE; // generic ignored junk |
|
1006 ptr = TypePtr::BotPTR; |
|
1007 } else if( offset == oopDesc::mark_offset_in_bytes() ) { |
|
1008 tj = TypeInstPtr::MARK; |
|
1009 ta = TypeAryPtr::RANGE; // generic ignored junk |
|
1010 ptr = TypePtr::BotPTR; |
|
1011 } else { // Random constant offset into array body |
|
1012 offset = Type::OffsetBot; // Flatten constant access into array body |
|
1013 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); |
|
1014 } |
|
1015 } |
|
1016 // Arrays of fixed size alias with arrays of unknown size. |
|
1017 if (ta->size() != TypeInt::POS) { |
|
1018 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); |
|
1019 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset, ta->instance_id()); |
|
1020 } |
|
1021 // Arrays of known objects become arrays of unknown objects. |
|
1022 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { |
|
1023 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); |
|
1024 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id()); |
|
1025 } |
|
1026 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so |
|
1027 // cannot be distinguished by bytecode alone. |
|
1028 if (ta->elem() == TypeInt::BOOL) { |
|
1029 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); |
|
1030 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); |
|
1031 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset, ta->instance_id()); |
|
1032 } |
|
1033 // During the 2nd round of IterGVN, NotNull castings are removed. |
|
1034 // Make sure the Bottom and NotNull variants alias the same. |
|
1035 // Also, make sure exact and non-exact variants alias the same. |
|
1036 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { |
|
1037 if (ta->const_oop()) { |
|
1038 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); |
|
1039 } else { |
|
1040 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); |
|
1041 } |
|
1042 } |
|
1043 } |
|
1044 |
|
1045 // Oop pointers need some flattening |
|
1046 const TypeInstPtr *to = tj->isa_instptr(); |
|
1047 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { |
|
1048 if( ptr == TypePtr::Constant ) { |
|
1049 // No constant oop pointers (such as Strings); they alias with |
|
1050 // unknown strings. |
|
1051 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
|
1052 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { |
|
1053 // During the 2nd round of IterGVN, NotNull castings are removed. |
|
1054 // Make sure the Bottom and NotNull variants alias the same. |
|
1055 // Also, make sure exact and non-exact variants alias the same. |
|
1056 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset, to->instance_id()); |
|
1057 } |
|
1058 // Canonicalize the holder of this field |
|
1059 ciInstanceKlass *k = to->klass()->as_instance_klass(); |
|
1060 if (offset >= 0 && offset < oopDesc::header_size() * wordSize) { |
|
1061 // First handle header references such as a LoadKlassNode, even if the |
|
1062 // object's klass is unloaded at compile time (4965979). |
|
1063 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset, to->instance_id()); |
|
1064 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { |
|
1065 to = NULL; |
|
1066 tj = TypeOopPtr::BOTTOM; |
|
1067 offset = tj->offset(); |
|
1068 } else { |
|
1069 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); |
|
1070 if (!k->equals(canonical_holder) || tj->offset() != offset) { |
|
1071 tj = to = TypeInstPtr::make(TypePtr::BotPTR, canonical_holder, false, NULL, offset, to->instance_id()); |
|
1072 } |
|
1073 } |
|
1074 } |
|
1075 |
|
1076 // Klass pointers to object array klasses need some flattening |
|
1077 const TypeKlassPtr *tk = tj->isa_klassptr(); |
|
1078 if( tk ) { |
|
1079 // If we are referencing a field within a Klass, we need |
|
1080 // to assume the worst case of an Object. Both exact and |
|
1081 // inexact types must flatten to the same alias class. |
|
1082 // Since the flattened result for a klass is defined to be |
|
1083 // precisely java.lang.Object, use a constant ptr. |
|
1084 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { |
|
1085 |
|
1086 tj = tk = TypeKlassPtr::make(TypePtr::Constant, |
|
1087 TypeKlassPtr::OBJECT->klass(), |
|
1088 offset); |
|
1089 } |
|
1090 |
|
1091 ciKlass* klass = tk->klass(); |
|
1092 if( klass->is_obj_array_klass() ) { |
|
1093 ciKlass* k = TypeAryPtr::OOPS->klass(); |
|
1094 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs |
|
1095 k = TypeInstPtr::BOTTOM->klass(); |
|
1096 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); |
|
1097 } |
|
1098 |
|
1099 // Check for precise loads from the primary supertype array and force them |
|
1100 // to the supertype cache alias index. Check for generic array loads from |
|
1101 // the primary supertype array and also force them to the supertype cache |
|
1102 // alias index. Since the same load can reach both, we need to merge |
|
1103 // these 2 disparate memories into the same alias class. Since the |
|
1104 // primary supertype array is read-only, there's no chance of confusion |
|
1105 // where we bypass an array load and an array store. |
|
1106 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); |
|
1107 if( offset == Type::OffsetBot || |
|
1108 off2 < Klass::primary_super_limit()*wordSize ) { |
|
1109 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); |
|
1110 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); |
|
1111 } |
|
1112 } |
|
1113 |
|
1114 // Flatten all Raw pointers together. |
|
1115 if (tj->base() == Type::RawPtr) |
|
1116 tj = TypeRawPtr::BOTTOM; |
|
1117 |
|
1118 if (tj->base() == Type::AnyPtr) |
|
1119 tj = TypePtr::BOTTOM; // An error, which the caller must check for. |
|
1120 |
|
1121 // Flatten all to bottom for now |
|
1122 switch( _AliasLevel ) { |
|
1123 case 0: |
|
1124 tj = TypePtr::BOTTOM; |
|
1125 break; |
|
1126 case 1: // Flatten to: oop, static, field or array |
|
1127 switch (tj->base()) { |
|
1128 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; |
|
1129 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; |
|
1130 case Type::AryPtr: // do not distinguish arrays at all |
|
1131 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; |
|
1132 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; |
|
1133 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it |
|
1134 default: ShouldNotReachHere(); |
|
1135 } |
|
1136 break; |
|
1137 case 2: // No collasping at level 2; keep all splits |
|
1138 case 3: // No collasping at level 3; keep all splits |
|
1139 break; |
|
1140 default: |
|
1141 Unimplemented(); |
|
1142 } |
|
1143 |
|
1144 offset = tj->offset(); |
|
1145 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); |
|
1146 |
|
1147 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || |
|
1148 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || |
|
1149 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || |
|
1150 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || |
|
1151 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || |
|
1152 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || |
|
1153 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , |
|
1154 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); |
|
1155 assert( tj->ptr() != TypePtr::TopPTR && |
|
1156 tj->ptr() != TypePtr::AnyNull && |
|
1157 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); |
|
1158 // assert( tj->ptr() != TypePtr::Constant || |
|
1159 // tj->base() == Type::RawPtr || |
|
1160 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); |
|
1161 |
|
1162 return tj; |
|
1163 } |
|
1164 |
|
1165 void Compile::AliasType::Init(int i, const TypePtr* at) { |
|
1166 _index = i; |
|
1167 _adr_type = at; |
|
1168 _field = NULL; |
|
1169 _is_rewritable = true; // default |
|
1170 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; |
|
1171 if (atoop != NULL && atoop->is_instance()) { |
|
1172 const TypeOopPtr *gt = atoop->cast_to_instance(TypeOopPtr::UNKNOWN_INSTANCE); |
|
1173 _general_index = Compile::current()->get_alias_index(gt); |
|
1174 } else { |
|
1175 _general_index = 0; |
|
1176 } |
|
1177 } |
|
1178 |
|
1179 //---------------------------------print_on------------------------------------ |
|
1180 #ifndef PRODUCT |
|
1181 void Compile::AliasType::print_on(outputStream* st) { |
|
1182 if (index() < 10) |
|
1183 st->print("@ <%d> ", index()); |
|
1184 else st->print("@ <%d>", index()); |
|
1185 st->print(is_rewritable() ? " " : " RO"); |
|
1186 int offset = adr_type()->offset(); |
|
1187 if (offset == Type::OffsetBot) |
|
1188 st->print(" +any"); |
|
1189 else st->print(" +%-3d", offset); |
|
1190 st->print(" in "); |
|
1191 adr_type()->dump_on(st); |
|
1192 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); |
|
1193 if (field() != NULL && tjp) { |
|
1194 if (tjp->klass() != field()->holder() || |
|
1195 tjp->offset() != field()->offset_in_bytes()) { |
|
1196 st->print(" != "); |
|
1197 field()->print(); |
|
1198 st->print(" ***"); |
|
1199 } |
|
1200 } |
|
1201 } |
|
1202 |
|
1203 void print_alias_types() { |
|
1204 Compile* C = Compile::current(); |
|
1205 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); |
|
1206 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { |
|
1207 C->alias_type(idx)->print_on(tty); |
|
1208 tty->cr(); |
|
1209 } |
|
1210 } |
|
1211 #endif |
|
1212 |
|
1213 |
|
1214 //----------------------------probe_alias_cache-------------------------------- |
|
1215 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { |
|
1216 intptr_t key = (intptr_t) adr_type; |
|
1217 key ^= key >> logAliasCacheSize; |
|
1218 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; |
|
1219 } |
|
1220 |
|
1221 |
|
1222 //-----------------------------grow_alias_types-------------------------------- |
|
1223 void Compile::grow_alias_types() { |
|
1224 const int old_ats = _max_alias_types; // how many before? |
|
1225 const int new_ats = old_ats; // how many more? |
|
1226 const int grow_ats = old_ats+new_ats; // how many now? |
|
1227 _max_alias_types = grow_ats; |
|
1228 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); |
|
1229 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); |
|
1230 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); |
|
1231 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; |
|
1232 } |
|
1233 |
|
1234 |
|
1235 //--------------------------------find_alias_type------------------------------ |
|
1236 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { |
|
1237 if (_AliasLevel == 0) |
|
1238 return alias_type(AliasIdxBot); |
|
1239 |
|
1240 AliasCacheEntry* ace = probe_alias_cache(adr_type); |
|
1241 if (ace->_adr_type == adr_type) { |
|
1242 return alias_type(ace->_index); |
|
1243 } |
|
1244 |
|
1245 // Handle special cases. |
|
1246 if (adr_type == NULL) return alias_type(AliasIdxTop); |
|
1247 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); |
|
1248 |
|
1249 // Do it the slow way. |
|
1250 const TypePtr* flat = flatten_alias_type(adr_type); |
|
1251 |
|
1252 #ifdef ASSERT |
|
1253 assert(flat == flatten_alias_type(flat), "idempotent"); |
|
1254 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); |
|
1255 if (flat->isa_oopptr() && !flat->isa_klassptr()) { |
|
1256 const TypeOopPtr* foop = flat->is_oopptr(); |
|
1257 const TypePtr* xoop = foop->cast_to_exactness(!foop->klass_is_exact())->is_ptr(); |
|
1258 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); |
|
1259 } |
|
1260 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); |
|
1261 #endif |
|
1262 |
|
1263 int idx = AliasIdxTop; |
|
1264 for (int i = 0; i < num_alias_types(); i++) { |
|
1265 if (alias_type(i)->adr_type() == flat) { |
|
1266 idx = i; |
|
1267 break; |
|
1268 } |
|
1269 } |
|
1270 |
|
1271 if (idx == AliasIdxTop) { |
|
1272 if (no_create) return NULL; |
|
1273 // Grow the array if necessary. |
|
1274 if (_num_alias_types == _max_alias_types) grow_alias_types(); |
|
1275 // Add a new alias type. |
|
1276 idx = _num_alias_types++; |
|
1277 _alias_types[idx]->Init(idx, flat); |
|
1278 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); |
|
1279 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); |
|
1280 if (flat->isa_instptr()) { |
|
1281 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() |
|
1282 && flat->is_instptr()->klass() == env()->Class_klass()) |
|
1283 alias_type(idx)->set_rewritable(false); |
|
1284 } |
|
1285 if (flat->isa_klassptr()) { |
|
1286 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) |
|
1287 alias_type(idx)->set_rewritable(false); |
|
1288 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) |
|
1289 alias_type(idx)->set_rewritable(false); |
|
1290 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) |
|
1291 alias_type(idx)->set_rewritable(false); |
|
1292 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) |
|
1293 alias_type(idx)->set_rewritable(false); |
|
1294 } |
|
1295 // %%% (We would like to finalize JavaThread::threadObj_offset(), |
|
1296 // but the base pointer type is not distinctive enough to identify |
|
1297 // references into JavaThread.) |
|
1298 |
|
1299 // Check for final instance fields. |
|
1300 const TypeInstPtr* tinst = flat->isa_instptr(); |
|
1301 if (tinst && tinst->offset() >= oopDesc::header_size() * wordSize) { |
|
1302 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); |
|
1303 ciField* field = k->get_field_by_offset(tinst->offset(), false); |
|
1304 // Set field() and is_rewritable() attributes. |
|
1305 if (field != NULL) alias_type(idx)->set_field(field); |
|
1306 } |
|
1307 const TypeKlassPtr* tklass = flat->isa_klassptr(); |
|
1308 // Check for final static fields. |
|
1309 if (tklass && tklass->klass()->is_instance_klass()) { |
|
1310 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); |
|
1311 ciField* field = k->get_field_by_offset(tklass->offset(), true); |
|
1312 // Set field() and is_rewritable() attributes. |
|
1313 if (field != NULL) alias_type(idx)->set_field(field); |
|
1314 } |
|
1315 } |
|
1316 |
|
1317 // Fill the cache for next time. |
|
1318 ace->_adr_type = adr_type; |
|
1319 ace->_index = idx; |
|
1320 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); |
|
1321 |
|
1322 // Might as well try to fill the cache for the flattened version, too. |
|
1323 AliasCacheEntry* face = probe_alias_cache(flat); |
|
1324 if (face->_adr_type == NULL) { |
|
1325 face->_adr_type = flat; |
|
1326 face->_index = idx; |
|
1327 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); |
|
1328 } |
|
1329 |
|
1330 return alias_type(idx); |
|
1331 } |
|
1332 |
|
1333 |
|
1334 Compile::AliasType* Compile::alias_type(ciField* field) { |
|
1335 const TypeOopPtr* t; |
|
1336 if (field->is_static()) |
|
1337 t = TypeKlassPtr::make(field->holder()); |
|
1338 else |
|
1339 t = TypeOopPtr::make_from_klass_raw(field->holder()); |
|
1340 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); |
|
1341 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); |
|
1342 return atp; |
|
1343 } |
|
1344 |
|
1345 |
|
1346 //------------------------------have_alias_type-------------------------------- |
|
1347 bool Compile::have_alias_type(const TypePtr* adr_type) { |
|
1348 AliasCacheEntry* ace = probe_alias_cache(adr_type); |
|
1349 if (ace->_adr_type == adr_type) { |
|
1350 return true; |
|
1351 } |
|
1352 |
|
1353 // Handle special cases. |
|
1354 if (adr_type == NULL) return true; |
|
1355 if (adr_type == TypePtr::BOTTOM) return true; |
|
1356 |
|
1357 return find_alias_type(adr_type, true) != NULL; |
|
1358 } |
|
1359 |
|
1360 //-----------------------------must_alias-------------------------------------- |
|
1361 // True if all values of the given address type are in the given alias category. |
|
1362 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { |
|
1363 if (alias_idx == AliasIdxBot) return true; // the universal category |
|
1364 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP |
|
1365 if (alias_idx == AliasIdxTop) return false; // the empty category |
|
1366 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins |
|
1367 |
|
1368 // the only remaining possible overlap is identity |
|
1369 int adr_idx = get_alias_index(adr_type); |
|
1370 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); |
|
1371 assert(adr_idx == alias_idx || |
|
1372 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM |
|
1373 && adr_type != TypeOopPtr::BOTTOM), |
|
1374 "should not be testing for overlap with an unsafe pointer"); |
|
1375 return adr_idx == alias_idx; |
|
1376 } |
|
1377 |
|
1378 //------------------------------can_alias-------------------------------------- |
|
1379 // True if any values of the given address type are in the given alias category. |
|
1380 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { |
|
1381 if (alias_idx == AliasIdxTop) return false; // the empty category |
|
1382 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP |
|
1383 if (alias_idx == AliasIdxBot) return true; // the universal category |
|
1384 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins |
|
1385 |
|
1386 // the only remaining possible overlap is identity |
|
1387 int adr_idx = get_alias_index(adr_type); |
|
1388 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); |
|
1389 return adr_idx == alias_idx; |
|
1390 } |
|
1391 |
|
1392 |
|
1393 |
|
1394 //---------------------------pop_warm_call------------------------------------- |
|
1395 WarmCallInfo* Compile::pop_warm_call() { |
|
1396 WarmCallInfo* wci = _warm_calls; |
|
1397 if (wci != NULL) _warm_calls = wci->remove_from(wci); |
|
1398 return wci; |
|
1399 } |
|
1400 |
|
1401 //----------------------------Inline_Warm-------------------------------------- |
|
1402 int Compile::Inline_Warm() { |
|
1403 // If there is room, try to inline some more warm call sites. |
|
1404 // %%% Do a graph index compaction pass when we think we're out of space? |
|
1405 if (!InlineWarmCalls) return 0; |
|
1406 |
|
1407 int calls_made_hot = 0; |
|
1408 int room_to_grow = NodeCountInliningCutoff - unique(); |
|
1409 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); |
|
1410 int amount_grown = 0; |
|
1411 WarmCallInfo* call; |
|
1412 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { |
|
1413 int est_size = (int)call->size(); |
|
1414 if (est_size > (room_to_grow - amount_grown)) { |
|
1415 // This one won't fit anyway. Get rid of it. |
|
1416 call->make_cold(); |
|
1417 continue; |
|
1418 } |
|
1419 call->make_hot(); |
|
1420 calls_made_hot++; |
|
1421 amount_grown += est_size; |
|
1422 amount_to_grow -= est_size; |
|
1423 } |
|
1424 |
|
1425 if (calls_made_hot > 0) set_major_progress(); |
|
1426 return calls_made_hot; |
|
1427 } |
|
1428 |
|
1429 |
|
1430 //----------------------------Finish_Warm-------------------------------------- |
|
1431 void Compile::Finish_Warm() { |
|
1432 if (!InlineWarmCalls) return; |
|
1433 if (failing()) return; |
|
1434 if (warm_calls() == NULL) return; |
|
1435 |
|
1436 // Clean up loose ends, if we are out of space for inlining. |
|
1437 WarmCallInfo* call; |
|
1438 while ((call = pop_warm_call()) != NULL) { |
|
1439 call->make_cold(); |
|
1440 } |
|
1441 } |
|
1442 |
|
1443 |
|
1444 //------------------------------Optimize--------------------------------------- |
|
1445 // Given a graph, optimize it. |
|
1446 void Compile::Optimize() { |
|
1447 TracePhase t1("optimizer", &_t_optimizer, true); |
|
1448 |
|
1449 #ifndef PRODUCT |
|
1450 if (env()->break_at_compile()) { |
|
1451 BREAKPOINT; |
|
1452 } |
|
1453 |
|
1454 #endif |
|
1455 |
|
1456 ResourceMark rm; |
|
1457 int loop_opts_cnt; |
|
1458 |
|
1459 NOT_PRODUCT( verify_graph_edges(); ) |
|
1460 |
|
1461 print_method("Start"); |
|
1462 |
|
1463 { |
|
1464 // Iterative Global Value Numbering, including ideal transforms |
|
1465 // Initialize IterGVN with types and values from parse-time GVN |
|
1466 PhaseIterGVN igvn(initial_gvn()); |
|
1467 { |
|
1468 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) |
|
1469 igvn.optimize(); |
|
1470 } |
|
1471 |
|
1472 print_method("Iter GVN 1", 2); |
|
1473 |
|
1474 if (failing()) return; |
|
1475 |
|
1476 // get rid of the connection graph since it's information is not |
|
1477 // updated by optimizations |
|
1478 _congraph = NULL; |
|
1479 |
|
1480 |
|
1481 // Loop transforms on the ideal graph. Range Check Elimination, |
|
1482 // peeling, unrolling, etc. |
|
1483 |
|
1484 // Set loop opts counter |
|
1485 loop_opts_cnt = num_loop_opts(); |
|
1486 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { |
|
1487 { |
|
1488 TracePhase t2("idealLoop", &_t_idealLoop, true); |
|
1489 PhaseIdealLoop ideal_loop( igvn, NULL, true ); |
|
1490 loop_opts_cnt--; |
|
1491 if (major_progress()) print_method("PhaseIdealLoop 1", 2); |
|
1492 if (failing()) return; |
|
1493 } |
|
1494 // Loop opts pass if partial peeling occurred in previous pass |
|
1495 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { |
|
1496 TracePhase t3("idealLoop", &_t_idealLoop, true); |
|
1497 PhaseIdealLoop ideal_loop( igvn, NULL, false ); |
|
1498 loop_opts_cnt--; |
|
1499 if (major_progress()) print_method("PhaseIdealLoop 2", 2); |
|
1500 if (failing()) return; |
|
1501 } |
|
1502 // Loop opts pass for loop-unrolling before CCP |
|
1503 if(major_progress() && (loop_opts_cnt > 0)) { |
|
1504 TracePhase t4("idealLoop", &_t_idealLoop, true); |
|
1505 PhaseIdealLoop ideal_loop( igvn, NULL, false ); |
|
1506 loop_opts_cnt--; |
|
1507 if (major_progress()) print_method("PhaseIdealLoop 3", 2); |
|
1508 } |
|
1509 } |
|
1510 if (failing()) return; |
|
1511 |
|
1512 // Conditional Constant Propagation; |
|
1513 PhaseCCP ccp( &igvn ); |
|
1514 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); |
|
1515 { |
|
1516 TracePhase t2("ccp", &_t_ccp, true); |
|
1517 ccp.do_transform(); |
|
1518 } |
|
1519 print_method("PhaseCPP 1", 2); |
|
1520 |
|
1521 assert( true, "Break here to ccp.dump_old2new_map()"); |
|
1522 |
|
1523 // Iterative Global Value Numbering, including ideal transforms |
|
1524 { |
|
1525 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) |
|
1526 igvn = ccp; |
|
1527 igvn.optimize(); |
|
1528 } |
|
1529 |
|
1530 print_method("Iter GVN 2", 2); |
|
1531 |
|
1532 if (failing()) return; |
|
1533 |
|
1534 // Loop transforms on the ideal graph. Range Check Elimination, |
|
1535 // peeling, unrolling, etc. |
|
1536 if(loop_opts_cnt > 0) { |
|
1537 debug_only( int cnt = 0; ); |
|
1538 while(major_progress() && (loop_opts_cnt > 0)) { |
|
1539 TracePhase t2("idealLoop", &_t_idealLoop, true); |
|
1540 assert( cnt++ < 40, "infinite cycle in loop optimization" ); |
|
1541 PhaseIdealLoop ideal_loop( igvn, NULL, true ); |
|
1542 loop_opts_cnt--; |
|
1543 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); |
|
1544 if (failing()) return; |
|
1545 } |
|
1546 } |
|
1547 { |
|
1548 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) |
|
1549 PhaseMacroExpand mex(igvn); |
|
1550 if (mex.expand_macro_nodes()) { |
|
1551 assert(failing(), "must bail out w/ explicit message"); |
|
1552 return; |
|
1553 } |
|
1554 } |
|
1555 |
|
1556 } // (End scope of igvn; run destructor if necessary for asserts.) |
|
1557 |
|
1558 // A method with only infinite loops has no edges entering loops from root |
|
1559 { |
|
1560 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) |
|
1561 if (final_graph_reshaping()) { |
|
1562 assert(failing(), "must bail out w/ explicit message"); |
|
1563 return; |
|
1564 } |
|
1565 } |
|
1566 |
|
1567 print_method("Optimize finished", 2); |
|
1568 } |
|
1569 |
|
1570 |
|
1571 //------------------------------Code_Gen--------------------------------------- |
|
1572 // Given a graph, generate code for it |
|
1573 void Compile::Code_Gen() { |
|
1574 if (failing()) return; |
|
1575 |
|
1576 // Perform instruction selection. You might think we could reclaim Matcher |
|
1577 // memory PDQ, but actually the Matcher is used in generating spill code. |
|
1578 // Internals of the Matcher (including some VectorSets) must remain live |
|
1579 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage |
|
1580 // set a bit in reclaimed memory. |
|
1581 |
|
1582 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine |
|
1583 // nodes. Mapping is only valid at the root of each matched subtree. |
|
1584 NOT_PRODUCT( verify_graph_edges(); ) |
|
1585 |
|
1586 Node_List proj_list; |
|
1587 Matcher m(proj_list); |
|
1588 _matcher = &m; |
|
1589 { |
|
1590 TracePhase t2("matcher", &_t_matcher, true); |
|
1591 m.match(); |
|
1592 } |
|
1593 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine |
|
1594 // nodes. Mapping is only valid at the root of each matched subtree. |
|
1595 NOT_PRODUCT( verify_graph_edges(); ) |
|
1596 |
|
1597 // If you have too many nodes, or if matching has failed, bail out |
|
1598 check_node_count(0, "out of nodes matching instructions"); |
|
1599 if (failing()) return; |
|
1600 |
|
1601 // Build a proper-looking CFG |
|
1602 PhaseCFG cfg(node_arena(), root(), m); |
|
1603 _cfg = &cfg; |
|
1604 { |
|
1605 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) |
|
1606 cfg.Dominators(); |
|
1607 if (failing()) return; |
|
1608 |
|
1609 NOT_PRODUCT( verify_graph_edges(); ) |
|
1610 |
|
1611 cfg.Estimate_Block_Frequency(); |
|
1612 cfg.GlobalCodeMotion(m,unique(),proj_list); |
|
1613 |
|
1614 print_method("Global code motion", 2); |
|
1615 |
|
1616 if (failing()) return; |
|
1617 NOT_PRODUCT( verify_graph_edges(); ) |
|
1618 |
|
1619 debug_only( cfg.verify(); ) |
|
1620 } |
|
1621 NOT_PRODUCT( verify_graph_edges(); ) |
|
1622 |
|
1623 PhaseChaitin regalloc(unique(),cfg,m); |
|
1624 _regalloc = ®alloc; |
|
1625 { |
|
1626 TracePhase t2("regalloc", &_t_registerAllocation, true); |
|
1627 // Perform any platform dependent preallocation actions. This is used, |
|
1628 // for example, to avoid taking an implicit null pointer exception |
|
1629 // using the frame pointer on win95. |
|
1630 _regalloc->pd_preallocate_hook(); |
|
1631 |
|
1632 // Perform register allocation. After Chaitin, use-def chains are |
|
1633 // no longer accurate (at spill code) and so must be ignored. |
|
1634 // Node->LRG->reg mappings are still accurate. |
|
1635 _regalloc->Register_Allocate(); |
|
1636 |
|
1637 // Bail out if the allocator builds too many nodes |
|
1638 if (failing()) return; |
|
1639 } |
|
1640 |
|
1641 // Prior to register allocation we kept empty basic blocks in case the |
|
1642 // the allocator needed a place to spill. After register allocation we |
|
1643 // are not adding any new instructions. If any basic block is empty, we |
|
1644 // can now safely remove it. |
|
1645 { |
|
1646 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); ) |
|
1647 cfg.RemoveEmpty(); |
|
1648 } |
|
1649 |
|
1650 // Perform any platform dependent postallocation verifications. |
|
1651 debug_only( _regalloc->pd_postallocate_verify_hook(); ) |
|
1652 |
|
1653 // Apply peephole optimizations |
|
1654 if( OptoPeephole ) { |
|
1655 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) |
|
1656 PhasePeephole peep( _regalloc, cfg); |
|
1657 peep.do_transform(); |
|
1658 } |
|
1659 |
|
1660 // Convert Nodes to instruction bits in a buffer |
|
1661 { |
|
1662 // %%%% workspace merge brought two timers together for one job |
|
1663 TracePhase t2a("output", &_t_output, true); |
|
1664 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) |
|
1665 Output(); |
|
1666 } |
|
1667 |
|
1668 print_method("End"); |
|
1669 |
|
1670 // He's dead, Jim. |
|
1671 _cfg = (PhaseCFG*)0xdeadbeef; |
|
1672 _regalloc = (PhaseChaitin*)0xdeadbeef; |
|
1673 } |
|
1674 |
|
1675 |
|
1676 //------------------------------dump_asm--------------------------------------- |
|
1677 // Dump formatted assembly |
|
1678 #ifndef PRODUCT |
|
1679 void Compile::dump_asm(int *pcs, uint pc_limit) { |
|
1680 bool cut_short = false; |
|
1681 tty->print_cr("#"); |
|
1682 tty->print("# "); _tf->dump(); tty->cr(); |
|
1683 tty->print_cr("#"); |
|
1684 |
|
1685 // For all blocks |
|
1686 int pc = 0x0; // Program counter |
|
1687 char starts_bundle = ' '; |
|
1688 _regalloc->dump_frame(); |
|
1689 |
|
1690 Node *n = NULL; |
|
1691 for( uint i=0; i<_cfg->_num_blocks; i++ ) { |
|
1692 if (VMThread::should_terminate()) { cut_short = true; break; } |
|
1693 Block *b = _cfg->_blocks[i]; |
|
1694 if (b->is_connector() && !Verbose) continue; |
|
1695 n = b->_nodes[0]; |
|
1696 if (pcs && n->_idx < pc_limit) |
|
1697 tty->print("%3.3x ", pcs[n->_idx]); |
|
1698 else |
|
1699 tty->print(" "); |
|
1700 b->dump_head( &_cfg->_bbs ); |
|
1701 if (b->is_connector()) { |
|
1702 tty->print_cr(" # Empty connector block"); |
|
1703 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { |
|
1704 tty->print_cr(" # Block is sole successor of call"); |
|
1705 } |
|
1706 |
|
1707 // For all instructions |
|
1708 Node *delay = NULL; |
|
1709 for( uint j = 0; j<b->_nodes.size(); j++ ) { |
|
1710 if (VMThread::should_terminate()) { cut_short = true; break; } |
|
1711 n = b->_nodes[j]; |
|
1712 if (valid_bundle_info(n)) { |
|
1713 Bundle *bundle = node_bundling(n); |
|
1714 if (bundle->used_in_unconditional_delay()) { |
|
1715 delay = n; |
|
1716 continue; |
|
1717 } |
|
1718 if (bundle->starts_bundle()) |
|
1719 starts_bundle = '+'; |
|
1720 } |
|
1721 |
|
1722 if( !n->is_Region() && // Dont print in the Assembly |
|
1723 !n->is_Phi() && // a few noisely useless nodes |
|
1724 !n->is_Proj() && |
|
1725 !n->is_MachTemp() && |
|
1726 !n->is_Catch() && // Would be nice to print exception table targets |
|
1727 !n->is_MergeMem() && // Not very interesting |
|
1728 !n->is_top() && // Debug info table constants |
|
1729 !(n->is_Con() && !n->is_Mach())// Debug info table constants |
|
1730 ) { |
|
1731 if (pcs && n->_idx < pc_limit) |
|
1732 tty->print("%3.3x", pcs[n->_idx]); |
|
1733 else |
|
1734 tty->print(" "); |
|
1735 tty->print(" %c ", starts_bundle); |
|
1736 starts_bundle = ' '; |
|
1737 tty->print("\t"); |
|
1738 n->format(_regalloc, tty); |
|
1739 tty->cr(); |
|
1740 } |
|
1741 |
|
1742 // If we have an instruction with a delay slot, and have seen a delay, |
|
1743 // then back up and print it |
|
1744 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { |
|
1745 assert(delay != NULL, "no unconditional delay instruction"); |
|
1746 if (node_bundling(delay)->starts_bundle()) |
|
1747 starts_bundle = '+'; |
|
1748 if (pcs && n->_idx < pc_limit) |
|
1749 tty->print("%3.3x", pcs[n->_idx]); |
|
1750 else |
|
1751 tty->print(" "); |
|
1752 tty->print(" %c ", starts_bundle); |
|
1753 starts_bundle = ' '; |
|
1754 tty->print("\t"); |
|
1755 delay->format(_regalloc, tty); |
|
1756 tty->print_cr(""); |
|
1757 delay = NULL; |
|
1758 } |
|
1759 |
|
1760 // Dump the exception table as well |
|
1761 if( n->is_Catch() && (Verbose || WizardMode) ) { |
|
1762 // Print the exception table for this offset |
|
1763 _handler_table.print_subtable_for(pc); |
|
1764 } |
|
1765 } |
|
1766 |
|
1767 if (pcs && n->_idx < pc_limit) |
|
1768 tty->print_cr("%3.3x", pcs[n->_idx]); |
|
1769 else |
|
1770 tty->print_cr(""); |
|
1771 |
|
1772 assert(cut_short || delay == NULL, "no unconditional delay branch"); |
|
1773 |
|
1774 } // End of per-block dump |
|
1775 tty->print_cr(""); |
|
1776 |
|
1777 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); |
|
1778 } |
|
1779 #endif |
|
1780 |
|
1781 //------------------------------Final_Reshape_Counts--------------------------- |
|
1782 // This class defines counters to help identify when a method |
|
1783 // may/must be executed using hardware with only 24-bit precision. |
|
1784 struct Final_Reshape_Counts : public StackObj { |
|
1785 int _call_count; // count non-inlined 'common' calls |
|
1786 int _float_count; // count float ops requiring 24-bit precision |
|
1787 int _double_count; // count double ops requiring more precision |
|
1788 int _java_call_count; // count non-inlined 'java' calls |
|
1789 VectorSet _visited; // Visitation flags |
|
1790 Node_List _tests; // Set of IfNodes & PCTableNodes |
|
1791 |
|
1792 Final_Reshape_Counts() : |
|
1793 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), |
|
1794 _visited( Thread::current()->resource_area() ) { } |
|
1795 |
|
1796 void inc_call_count () { _call_count ++; } |
|
1797 void inc_float_count () { _float_count ++; } |
|
1798 void inc_double_count() { _double_count++; } |
|
1799 void inc_java_call_count() { _java_call_count++; } |
|
1800 |
|
1801 int get_call_count () const { return _call_count ; } |
|
1802 int get_float_count () const { return _float_count ; } |
|
1803 int get_double_count() const { return _double_count; } |
|
1804 int get_java_call_count() const { return _java_call_count; } |
|
1805 }; |
|
1806 |
|
1807 static bool oop_offset_is_sane(const TypeInstPtr* tp) { |
|
1808 ciInstanceKlass *k = tp->klass()->as_instance_klass(); |
|
1809 // Make sure the offset goes inside the instance layout. |
|
1810 return (uint)tp->offset() < (uint)(oopDesc::header_size() + k->nonstatic_field_size())*wordSize; |
|
1811 // Note that OffsetBot and OffsetTop are very negative. |
|
1812 } |
|
1813 |
|
1814 //------------------------------final_graph_reshaping_impl---------------------- |
|
1815 // Implement items 1-5 from final_graph_reshaping below. |
|
1816 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { |
|
1817 |
|
1818 uint nop = n->Opcode(); |
|
1819 |
|
1820 // Check for 2-input instruction with "last use" on right input. |
|
1821 // Swap to left input. Implements item (2). |
|
1822 if( n->req() == 3 && // two-input instruction |
|
1823 n->in(1)->outcnt() > 1 && // left use is NOT a last use |
|
1824 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop |
|
1825 n->in(2)->outcnt() == 1 &&// right use IS a last use |
|
1826 !n->in(2)->is_Con() ) { // right use is not a constant |
|
1827 // Check for commutative opcode |
|
1828 switch( nop ) { |
|
1829 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: |
|
1830 case Op_MaxI: case Op_MinI: |
|
1831 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: |
|
1832 case Op_AndL: case Op_XorL: case Op_OrL: |
|
1833 case Op_AndI: case Op_XorI: case Op_OrI: { |
|
1834 // Move "last use" input to left by swapping inputs |
|
1835 n->swap_edges(1, 2); |
|
1836 break; |
|
1837 } |
|
1838 default: |
|
1839 break; |
|
1840 } |
|
1841 } |
|
1842 |
|
1843 // Count FPU ops and common calls, implements item (3) |
|
1844 switch( nop ) { |
|
1845 // Count all float operations that may use FPU |
|
1846 case Op_AddF: |
|
1847 case Op_SubF: |
|
1848 case Op_MulF: |
|
1849 case Op_DivF: |
|
1850 case Op_NegF: |
|
1851 case Op_ModF: |
|
1852 case Op_ConvI2F: |
|
1853 case Op_ConF: |
|
1854 case Op_CmpF: |
|
1855 case Op_CmpF3: |
|
1856 // case Op_ConvL2F: // longs are split into 32-bit halves |
|
1857 fpu.inc_float_count(); |
|
1858 break; |
|
1859 |
|
1860 case Op_ConvF2D: |
|
1861 case Op_ConvD2F: |
|
1862 fpu.inc_float_count(); |
|
1863 fpu.inc_double_count(); |
|
1864 break; |
|
1865 |
|
1866 // Count all double operations that may use FPU |
|
1867 case Op_AddD: |
|
1868 case Op_SubD: |
|
1869 case Op_MulD: |
|
1870 case Op_DivD: |
|
1871 case Op_NegD: |
|
1872 case Op_ModD: |
|
1873 case Op_ConvI2D: |
|
1874 case Op_ConvD2I: |
|
1875 // case Op_ConvL2D: // handled by leaf call |
|
1876 // case Op_ConvD2L: // handled by leaf call |
|
1877 case Op_ConD: |
|
1878 case Op_CmpD: |
|
1879 case Op_CmpD3: |
|
1880 fpu.inc_double_count(); |
|
1881 break; |
|
1882 case Op_Opaque1: // Remove Opaque Nodes before matching |
|
1883 case Op_Opaque2: // Remove Opaque Nodes before matching |
|
1884 n->replace_by(n->in(1)); |
|
1885 break; |
|
1886 case Op_CallStaticJava: |
|
1887 case Op_CallJava: |
|
1888 case Op_CallDynamicJava: |
|
1889 fpu.inc_java_call_count(); // Count java call site; |
|
1890 case Op_CallRuntime: |
|
1891 case Op_CallLeaf: |
|
1892 case Op_CallLeafNoFP: { |
|
1893 assert( n->is_Call(), "" ); |
|
1894 CallNode *call = n->as_Call(); |
|
1895 // Count call sites where the FP mode bit would have to be flipped. |
|
1896 // Do not count uncommon runtime calls: |
|
1897 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, |
|
1898 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... |
|
1899 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { |
|
1900 fpu.inc_call_count(); // Count the call site |
|
1901 } else { // See if uncommon argument is shared |
|
1902 Node *n = call->in(TypeFunc::Parms); |
|
1903 int nop = n->Opcode(); |
|
1904 // Clone shared simple arguments to uncommon calls, item (1). |
|
1905 if( n->outcnt() > 1 && |
|
1906 !n->is_Proj() && |
|
1907 nop != Op_CreateEx && |
|
1908 nop != Op_CheckCastPP && |
|
1909 !n->is_Mem() ) { |
|
1910 Node *x = n->clone(); |
|
1911 call->set_req( TypeFunc::Parms, x ); |
|
1912 } |
|
1913 } |
|
1914 break; |
|
1915 } |
|
1916 |
|
1917 case Op_StoreD: |
|
1918 case Op_LoadD: |
|
1919 case Op_LoadD_unaligned: |
|
1920 fpu.inc_double_count(); |
|
1921 goto handle_mem; |
|
1922 case Op_StoreF: |
|
1923 case Op_LoadF: |
|
1924 fpu.inc_float_count(); |
|
1925 goto handle_mem; |
|
1926 |
|
1927 case Op_StoreB: |
|
1928 case Op_StoreC: |
|
1929 case Op_StoreCM: |
|
1930 case Op_StorePConditional: |
|
1931 case Op_StoreI: |
|
1932 case Op_StoreL: |
|
1933 case Op_StoreLConditional: |
|
1934 case Op_CompareAndSwapI: |
|
1935 case Op_CompareAndSwapL: |
|
1936 case Op_CompareAndSwapP: |
|
1937 case Op_StoreP: |
|
1938 case Op_LoadB: |
|
1939 case Op_LoadC: |
|
1940 case Op_LoadI: |
|
1941 case Op_LoadKlass: |
|
1942 case Op_LoadL: |
|
1943 case Op_LoadL_unaligned: |
|
1944 case Op_LoadPLocked: |
|
1945 case Op_LoadLLocked: |
|
1946 case Op_LoadP: |
|
1947 case Op_LoadRange: |
|
1948 case Op_LoadS: { |
|
1949 handle_mem: |
|
1950 #ifdef ASSERT |
|
1951 if( VerifyOptoOopOffsets ) { |
|
1952 assert( n->is_Mem(), "" ); |
|
1953 MemNode *mem = (MemNode*)n; |
|
1954 // Check to see if address types have grounded out somehow. |
|
1955 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); |
|
1956 assert( !tp || oop_offset_is_sane(tp), "" ); |
|
1957 } |
|
1958 #endif |
|
1959 break; |
|
1960 } |
|
1961 case Op_If: |
|
1962 case Op_CountedLoopEnd: |
|
1963 fpu._tests.push(n); // Collect CFG split points |
|
1964 break; |
|
1965 |
|
1966 case Op_AddP: { // Assert sane base pointers |
|
1967 const Node *addp = n->in(AddPNode::Address); |
|
1968 assert( !addp->is_AddP() || |
|
1969 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation |
|
1970 addp->in(AddPNode::Base) == n->in(AddPNode::Base), |
|
1971 "Base pointers must match" ); |
|
1972 break; |
|
1973 } |
|
1974 |
|
1975 case Op_ModI: |
|
1976 if (UseDivMod) { |
|
1977 // Check if a%b and a/b both exist |
|
1978 Node* d = n->find_similar(Op_DivI); |
|
1979 if (d) { |
|
1980 // Replace them with a fused divmod if supported |
|
1981 Compile* C = Compile::current(); |
|
1982 if (Matcher::has_match_rule(Op_DivModI)) { |
|
1983 DivModINode* divmod = DivModINode::make(C, n); |
|
1984 d->replace_by(divmod->div_proj()); |
|
1985 n->replace_by(divmod->mod_proj()); |
|
1986 } else { |
|
1987 // replace a%b with a-((a/b)*b) |
|
1988 Node* mult = new (C, 3) MulINode(d, d->in(2)); |
|
1989 Node* sub = new (C, 3) SubINode(d->in(1), mult); |
|
1990 n->replace_by( sub ); |
|
1991 } |
|
1992 } |
|
1993 } |
|
1994 break; |
|
1995 |
|
1996 case Op_ModL: |
|
1997 if (UseDivMod) { |
|
1998 // Check if a%b and a/b both exist |
|
1999 Node* d = n->find_similar(Op_DivL); |
|
2000 if (d) { |
|
2001 // Replace them with a fused divmod if supported |
|
2002 Compile* C = Compile::current(); |
|
2003 if (Matcher::has_match_rule(Op_DivModL)) { |
|
2004 DivModLNode* divmod = DivModLNode::make(C, n); |
|
2005 d->replace_by(divmod->div_proj()); |
|
2006 n->replace_by(divmod->mod_proj()); |
|
2007 } else { |
|
2008 // replace a%b with a-((a/b)*b) |
|
2009 Node* mult = new (C, 3) MulLNode(d, d->in(2)); |
|
2010 Node* sub = new (C, 3) SubLNode(d->in(1), mult); |
|
2011 n->replace_by( sub ); |
|
2012 } |
|
2013 } |
|
2014 } |
|
2015 break; |
|
2016 |
|
2017 case Op_Load16B: |
|
2018 case Op_Load8B: |
|
2019 case Op_Load4B: |
|
2020 case Op_Load8S: |
|
2021 case Op_Load4S: |
|
2022 case Op_Load2S: |
|
2023 case Op_Load8C: |
|
2024 case Op_Load4C: |
|
2025 case Op_Load2C: |
|
2026 case Op_Load4I: |
|
2027 case Op_Load2I: |
|
2028 case Op_Load2L: |
|
2029 case Op_Load4F: |
|
2030 case Op_Load2F: |
|
2031 case Op_Load2D: |
|
2032 case Op_Store16B: |
|
2033 case Op_Store8B: |
|
2034 case Op_Store4B: |
|
2035 case Op_Store8C: |
|
2036 case Op_Store4C: |
|
2037 case Op_Store2C: |
|
2038 case Op_Store4I: |
|
2039 case Op_Store2I: |
|
2040 case Op_Store2L: |
|
2041 case Op_Store4F: |
|
2042 case Op_Store2F: |
|
2043 case Op_Store2D: |
|
2044 break; |
|
2045 |
|
2046 case Op_PackB: |
|
2047 case Op_PackS: |
|
2048 case Op_PackC: |
|
2049 case Op_PackI: |
|
2050 case Op_PackF: |
|
2051 case Op_PackL: |
|
2052 case Op_PackD: |
|
2053 if (n->req()-1 > 2) { |
|
2054 // Replace many operand PackNodes with a binary tree for matching |
|
2055 PackNode* p = (PackNode*) n; |
|
2056 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); |
|
2057 n->replace_by(btp); |
|
2058 } |
|
2059 break; |
|
2060 default: |
|
2061 assert( !n->is_Call(), "" ); |
|
2062 assert( !n->is_Mem(), "" ); |
|
2063 if( n->is_If() || n->is_PCTable() ) |
|
2064 fpu._tests.push(n); // Collect CFG split points |
|
2065 break; |
|
2066 } |
|
2067 } |
|
2068 |
|
2069 //------------------------------final_graph_reshaping_walk--------------------- |
|
2070 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), |
|
2071 // requires that the walk visits a node's inputs before visiting the node. |
|
2072 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { |
|
2073 fpu._visited.set(root->_idx); // first, mark node as visited |
|
2074 uint cnt = root->req(); |
|
2075 Node *n = root; |
|
2076 uint i = 0; |
|
2077 while (true) { |
|
2078 if (i < cnt) { |
|
2079 // Place all non-visited non-null inputs onto stack |
|
2080 Node* m = n->in(i); |
|
2081 ++i; |
|
2082 if (m != NULL && !fpu._visited.test_set(m->_idx)) { |
|
2083 cnt = m->req(); |
|
2084 nstack.push(n, i); // put on stack parent and next input's index |
|
2085 n = m; |
|
2086 i = 0; |
|
2087 } |
|
2088 } else { |
|
2089 // Now do post-visit work |
|
2090 final_graph_reshaping_impl( n, fpu ); |
|
2091 if (nstack.is_empty()) |
|
2092 break; // finished |
|
2093 n = nstack.node(); // Get node from stack |
|
2094 cnt = n->req(); |
|
2095 i = nstack.index(); |
|
2096 nstack.pop(); // Shift to the next node on stack |
|
2097 } |
|
2098 } |
|
2099 } |
|
2100 |
|
2101 //------------------------------final_graph_reshaping-------------------------- |
|
2102 // Final Graph Reshaping. |
|
2103 // |
|
2104 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late |
|
2105 // and not commoned up and forced early. Must come after regular |
|
2106 // optimizations to avoid GVN undoing the cloning. Clone constant |
|
2107 // inputs to Loop Phis; these will be split by the allocator anyways. |
|
2108 // Remove Opaque nodes. |
|
2109 // (2) Move last-uses by commutative operations to the left input to encourage |
|
2110 // Intel update-in-place two-address operations and better register usage |
|
2111 // on RISCs. Must come after regular optimizations to avoid GVN Ideal |
|
2112 // calls canonicalizing them back. |
|
2113 // (3) Count the number of double-precision FP ops, single-precision FP ops |
|
2114 // and call sites. On Intel, we can get correct rounding either by |
|
2115 // forcing singles to memory (requires extra stores and loads after each |
|
2116 // FP bytecode) or we can set a rounding mode bit (requires setting and |
|
2117 // clearing the mode bit around call sites). The mode bit is only used |
|
2118 // if the relative frequency of single FP ops to calls is low enough. |
|
2119 // This is a key transform for SPEC mpeg_audio. |
|
2120 // (4) Detect infinite loops; blobs of code reachable from above but not |
|
2121 // below. Several of the Code_Gen algorithms fail on such code shapes, |
|
2122 // so we simply bail out. Happens a lot in ZKM.jar, but also happens |
|
2123 // from time to time in other codes (such as -Xcomp finalizer loops, etc). |
|
2124 // Detection is by looking for IfNodes where only 1 projection is |
|
2125 // reachable from below or CatchNodes missing some targets. |
|
2126 // (5) Assert for insane oop offsets in debug mode. |
|
2127 |
|
2128 bool Compile::final_graph_reshaping() { |
|
2129 // an infinite loop may have been eliminated by the optimizer, |
|
2130 // in which case the graph will be empty. |
|
2131 if (root()->req() == 1) { |
|
2132 record_method_not_compilable("trivial infinite loop"); |
|
2133 return true; |
|
2134 } |
|
2135 |
|
2136 Final_Reshape_Counts fpu; |
|
2137 |
|
2138 // Visit everybody reachable! |
|
2139 // Allocate stack of size C->unique()/2 to avoid frequent realloc |
|
2140 Node_Stack nstack(unique() >> 1); |
|
2141 final_graph_reshaping_walk(nstack, root(), fpu); |
|
2142 |
|
2143 // Check for unreachable (from below) code (i.e., infinite loops). |
|
2144 for( uint i = 0; i < fpu._tests.size(); i++ ) { |
|
2145 Node *n = fpu._tests[i]; |
|
2146 assert( n->is_PCTable() || n->is_If(), "either PCTables or IfNodes" ); |
|
2147 // Get number of CFG targets; 2 for IfNodes or _size for PCTables. |
|
2148 // Note that PCTables include exception targets after calls. |
|
2149 uint expected_kids = n->is_PCTable() ? n->as_PCTable()->_size : 2; |
|
2150 if (n->outcnt() != expected_kids) { |
|
2151 // Check for a few special cases. Rethrow Nodes never take the |
|
2152 // 'fall-thru' path, so expected kids is 1 less. |
|
2153 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { |
|
2154 if (n->in(0)->in(0)->is_Call()) { |
|
2155 CallNode *call = n->in(0)->in(0)->as_Call(); |
|
2156 if (call->entry_point() == OptoRuntime::rethrow_stub()) { |
|
2157 expected_kids--; // Rethrow always has 1 less kid |
|
2158 } else if (call->req() > TypeFunc::Parms && |
|
2159 call->is_CallDynamicJava()) { |
|
2160 // Check for null receiver. In such case, the optimizer has |
|
2161 // detected that the virtual call will always result in a null |
|
2162 // pointer exception. The fall-through projection of this CatchNode |
|
2163 // will not be populated. |
|
2164 Node *arg0 = call->in(TypeFunc::Parms); |
|
2165 if (arg0->is_Type() && |
|
2166 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { |
|
2167 expected_kids--; |
|
2168 } |
|
2169 } else if (call->entry_point() == OptoRuntime::new_array_Java() && |
|
2170 call->req() > TypeFunc::Parms+1 && |
|
2171 call->is_CallStaticJava()) { |
|
2172 // Check for negative array length. In such case, the optimizer has |
|
2173 // detected that the allocation attempt will always result in an |
|
2174 // exception. There is no fall-through projection of this CatchNode . |
|
2175 Node *arg1 = call->in(TypeFunc::Parms+1); |
|
2176 if (arg1->is_Type() && |
|
2177 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { |
|
2178 expected_kids--; |
|
2179 } |
|
2180 } |
|
2181 } |
|
2182 } |
|
2183 // Recheck with a better notion of 'expected_kids' |
|
2184 if (n->outcnt() != expected_kids) { |
|
2185 record_method_not_compilable("malformed control flow"); |
|
2186 return true; // Not all targets reachable! |
|
2187 } |
|
2188 } |
|
2189 // Check that I actually visited all kids. Unreached kids |
|
2190 // must be infinite loops. |
|
2191 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) |
|
2192 if (!fpu._visited.test(n->fast_out(j)->_idx)) { |
|
2193 record_method_not_compilable("infinite loop"); |
|
2194 return true; // Found unvisited kid; must be unreach |
|
2195 } |
|
2196 } |
|
2197 |
|
2198 // If original bytecodes contained a mixture of floats and doubles |
|
2199 // check if the optimizer has made it homogenous, item (3). |
|
2200 if( Use24BitFPMode && Use24BitFP && |
|
2201 fpu.get_float_count() > 32 && |
|
2202 fpu.get_double_count() == 0 && |
|
2203 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { |
|
2204 set_24_bit_selection_and_mode( false, true ); |
|
2205 } |
|
2206 |
|
2207 set_has_java_calls(fpu.get_java_call_count() > 0); |
|
2208 |
|
2209 // No infinite loops, no reason to bail out. |
|
2210 return false; |
|
2211 } |
|
2212 |
|
2213 //-----------------------------too_many_traps---------------------------------- |
|
2214 // Report if there are too many traps at the current method and bci. |
|
2215 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. |
|
2216 bool Compile::too_many_traps(ciMethod* method, |
|
2217 int bci, |
|
2218 Deoptimization::DeoptReason reason) { |
|
2219 ciMethodData* md = method->method_data(); |
|
2220 if (md->is_empty()) { |
|
2221 // Assume the trap has not occurred, or that it occurred only |
|
2222 // because of a transient condition during start-up in the interpreter. |
|
2223 return false; |
|
2224 } |
|
2225 if (md->has_trap_at(bci, reason) != 0) { |
|
2226 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. |
|
2227 // Also, if there are multiple reasons, or if there is no per-BCI record, |
|
2228 // assume the worst. |
|
2229 if (log()) |
|
2230 log()->elem("observe trap='%s' count='%d'", |
|
2231 Deoptimization::trap_reason_name(reason), |
|
2232 md->trap_count(reason)); |
|
2233 return true; |
|
2234 } else { |
|
2235 // Ignore method/bci and see if there have been too many globally. |
|
2236 return too_many_traps(reason, md); |
|
2237 } |
|
2238 } |
|
2239 |
|
2240 // Less-accurate variant which does not require a method and bci. |
|
2241 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, |
|
2242 ciMethodData* logmd) { |
|
2243 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { |
|
2244 // Too many traps globally. |
|
2245 // Note that we use cumulative trap_count, not just md->trap_count. |
|
2246 if (log()) { |
|
2247 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); |
|
2248 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", |
|
2249 Deoptimization::trap_reason_name(reason), |
|
2250 mcount, trap_count(reason)); |
|
2251 } |
|
2252 return true; |
|
2253 } else { |
|
2254 // The coast is clear. |
|
2255 return false; |
|
2256 } |
|
2257 } |
|
2258 |
|
2259 //--------------------------too_many_recompiles-------------------------------- |
|
2260 // Report if there are too many recompiles at the current method and bci. |
|
2261 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. |
|
2262 // Is not eager to return true, since this will cause the compiler to use |
|
2263 // Action_none for a trap point, to avoid too many recompilations. |
|
2264 bool Compile::too_many_recompiles(ciMethod* method, |
|
2265 int bci, |
|
2266 Deoptimization::DeoptReason reason) { |
|
2267 ciMethodData* md = method->method_data(); |
|
2268 if (md->is_empty()) { |
|
2269 // Assume the trap has not occurred, or that it occurred only |
|
2270 // because of a transient condition during start-up in the interpreter. |
|
2271 return false; |
|
2272 } |
|
2273 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. |
|
2274 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; |
|
2275 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero |
|
2276 Deoptimization::DeoptReason per_bc_reason |
|
2277 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); |
|
2278 if ((per_bc_reason == Deoptimization::Reason_none |
|
2279 || md->has_trap_at(bci, reason) != 0) |
|
2280 // The trap frequency measure we care about is the recompile count: |
|
2281 && md->trap_recompiled_at(bci) |
|
2282 && md->overflow_recompile_count() >= bc_cutoff) { |
|
2283 // Do not emit a trap here if it has already caused recompilations. |
|
2284 // Also, if there are multiple reasons, or if there is no per-BCI record, |
|
2285 // assume the worst. |
|
2286 if (log()) |
|
2287 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", |
|
2288 Deoptimization::trap_reason_name(reason), |
|
2289 md->trap_count(reason), |
|
2290 md->overflow_recompile_count()); |
|
2291 return true; |
|
2292 } else if (trap_count(reason) != 0 |
|
2293 && decompile_count() >= m_cutoff) { |
|
2294 // Too many recompiles globally, and we have seen this sort of trap. |
|
2295 // Use cumulative decompile_count, not just md->decompile_count. |
|
2296 if (log()) |
|
2297 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", |
|
2298 Deoptimization::trap_reason_name(reason), |
|
2299 md->trap_count(reason), trap_count(reason), |
|
2300 md->decompile_count(), decompile_count()); |
|
2301 return true; |
|
2302 } else { |
|
2303 // The coast is clear. |
|
2304 return false; |
|
2305 } |
|
2306 } |
|
2307 |
|
2308 |
|
2309 #ifndef PRODUCT |
|
2310 //------------------------------verify_graph_edges--------------------------- |
|
2311 // Walk the Graph and verify that there is a one-to-one correspondence |
|
2312 // between Use-Def edges and Def-Use edges in the graph. |
|
2313 void Compile::verify_graph_edges(bool no_dead_code) { |
|
2314 if (VerifyGraphEdges) { |
|
2315 ResourceArea *area = Thread::current()->resource_area(); |
|
2316 Unique_Node_List visited(area); |
|
2317 // Call recursive graph walk to check edges |
|
2318 _root->verify_edges(visited); |
|
2319 if (no_dead_code) { |
|
2320 // Now make sure that no visited node is used by an unvisited node. |
|
2321 bool dead_nodes = 0; |
|
2322 Unique_Node_List checked(area); |
|
2323 while (visited.size() > 0) { |
|
2324 Node* n = visited.pop(); |
|
2325 checked.push(n); |
|
2326 for (uint i = 0; i < n->outcnt(); i++) { |
|
2327 Node* use = n->raw_out(i); |
|
2328 if (checked.member(use)) continue; // already checked |
|
2329 if (visited.member(use)) continue; // already in the graph |
|
2330 if (use->is_Con()) continue; // a dead ConNode is OK |
|
2331 // At this point, we have found a dead node which is DU-reachable. |
|
2332 if (dead_nodes++ == 0) |
|
2333 tty->print_cr("*** Dead nodes reachable via DU edges:"); |
|
2334 use->dump(2); |
|
2335 tty->print_cr("---"); |
|
2336 checked.push(use); // No repeats; pretend it is now checked. |
|
2337 } |
|
2338 } |
|
2339 assert(dead_nodes == 0, "using nodes must be reachable from root"); |
|
2340 } |
|
2341 } |
|
2342 } |
|
2343 #endif |
|
2344 |
|
2345 // The Compile object keeps track of failure reasons separately from the ciEnv. |
|
2346 // This is required because there is not quite a 1-1 relation between the |
|
2347 // ciEnv and its compilation task and the Compile object. Note that one |
|
2348 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides |
|
2349 // to backtrack and retry without subsuming loads. Other than this backtracking |
|
2350 // behavior, the Compile's failure reason is quietly copied up to the ciEnv |
|
2351 // by the logic in C2Compiler. |
|
2352 void Compile::record_failure(const char* reason) { |
|
2353 if (log() != NULL) { |
|
2354 log()->elem("failure reason='%s' phase='compile'", reason); |
|
2355 } |
|
2356 if (_failure_reason == NULL) { |
|
2357 // Record the first failure reason. |
|
2358 _failure_reason = reason; |
|
2359 } |
|
2360 _root = NULL; // flush the graph, too |
|
2361 } |
|
2362 |
|
2363 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) |
|
2364 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) |
|
2365 { |
|
2366 if (dolog) { |
|
2367 C = Compile::current(); |
|
2368 _log = C->log(); |
|
2369 } else { |
|
2370 C = NULL; |
|
2371 _log = NULL; |
|
2372 } |
|
2373 if (_log != NULL) { |
|
2374 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); |
|
2375 _log->stamp(); |
|
2376 _log->end_head(); |
|
2377 } |
|
2378 } |
|
2379 |
|
2380 Compile::TracePhase::~TracePhase() { |
|
2381 if (_log != NULL) { |
|
2382 _log->done("phase nodes='%d'", C->unique()); |
|
2383 } |
|
2384 } |