Mercurial > hg > truffle
annotate src/share/vm/opto/compile.cpp @ 100:c7c777385a15
6667042: PrintAssembly option does not work without special plugin
Summary: remove old private plugin interface, simplify, rework old plugin to use unchanged Gnu sources
Reviewed-by: kvn, rasbold
author | jrose |
---|---|
date | Wed, 02 Apr 2008 12:09:59 -0700 |
parents | 99269dbf4ba8 |
children | ba764ed4b6f2 |
rev | line source |
---|---|
0 | 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 } | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
336 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
337 // Recompiling without escape analysis |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
338 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
339 tty->print_cr("** Bailout: Recompile without escape analysis **"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
340 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
341 } |
0 | 342 if (env()->break_at_compile()) { |
343 // Open the debugger when compiing this method. | |
344 tty->print("### Breaking when compiling: "); | |
345 method()->print_short_name(); | |
346 tty->cr(); | |
347 BREAKPOINT; | |
348 } | |
349 | |
350 if( PrintOpto ) { | |
351 if (is_osr_compilation()) { | |
352 tty->print("[OSR]%3d", _compile_id); | |
353 } else { | |
354 tty->print("%3d", _compile_id); | |
355 } | |
356 } | |
357 #endif | |
358 } | |
359 | |
360 | |
361 void Compile::init_scratch_buffer_blob() { | |
362 if( scratch_buffer_blob() != NULL ) return; | |
363 | |
364 // Construct a temporary CodeBuffer to have it construct a BufferBlob | |
365 // Cache this BufferBlob for this compile. | |
366 ResourceMark rm; | |
367 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); | |
368 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); | |
369 // Record the buffer blob for next time. | |
370 set_scratch_buffer_blob(blob); | |
371 guarantee(scratch_buffer_blob() != NULL, "Need BufferBlob for code generation"); | |
372 | |
373 // Initialize the relocation buffers | |
374 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; | |
375 set_scratch_locs_memory(locs_buf); | |
376 } | |
377 | |
378 | |
379 //-----------------------scratch_emit_size------------------------------------- | |
380 // Helper function that computes size by emitting code | |
381 uint Compile::scratch_emit_size(const Node* n) { | |
382 // Emit into a trash buffer and count bytes emitted. | |
383 // This is a pretty expensive way to compute a size, | |
384 // but it works well enough if seldom used. | |
385 // All common fixed-size instructions are given a size | |
386 // method by the AD file. | |
387 // Note that the scratch buffer blob and locs memory are | |
388 // allocated at the beginning of the compile task, and | |
389 // may be shared by several calls to scratch_emit_size. | |
390 // The allocation of the scratch buffer blob is particularly | |
391 // expensive, since it has to grab the code cache lock. | |
392 BufferBlob* blob = this->scratch_buffer_blob(); | |
393 assert(blob != NULL, "Initialize BufferBlob at start"); | |
394 assert(blob->size() > MAX_inst_size, "sanity"); | |
395 relocInfo* locs_buf = scratch_locs_memory(); | |
396 address blob_begin = blob->instructions_begin(); | |
397 address blob_end = (address)locs_buf; | |
398 assert(blob->instructions_contains(blob_end), "sanity"); | |
399 CodeBuffer buf(blob_begin, blob_end - blob_begin); | |
400 buf.initialize_consts_size(MAX_const_size); | |
401 buf.initialize_stubs_size(MAX_stubs_size); | |
402 assert(locs_buf != NULL, "sanity"); | |
403 int lsize = MAX_locs_size / 2; | |
404 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); | |
405 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); | |
406 n->emit(buf, this->regalloc()); | |
407 return buf.code_size(); | |
408 } | |
409 | |
410 | |
411 // ============================================================================ | |
412 //------------------------------Compile standard------------------------------- | |
413 debug_only( int Compile::_debug_idx = 100000; ) | |
414 | |
415 // Compile a method. entry_bci is -1 for normal compilations and indicates | |
416 // the continuation bci for on stack replacement. | |
417 | |
418 | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
419 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) |
0 | 420 : Phase(Compiler), |
421 _env(ci_env), | |
422 _log(ci_env->log()), | |
423 _compile_id(ci_env->compile_id()), | |
424 _save_argument_registers(false), | |
425 _stub_name(NULL), | |
426 _stub_function(NULL), | |
427 _stub_entry_point(NULL), | |
428 _method(target), | |
429 _entry_bci(osr_bci), | |
430 _initial_gvn(NULL), | |
431 _for_igvn(NULL), | |
432 _warm_calls(NULL), | |
433 _subsume_loads(subsume_loads), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
434 _do_escape_analysis(do_escape_analysis), |
0 | 435 _failure_reason(NULL), |
436 _code_buffer("Compile::Fill_buffer"), | |
437 _orig_pc_slot(0), | |
438 _orig_pc_slot_offset_in_bytes(0), | |
439 _node_bundling_limit(0), | |
440 _node_bundling_base(NULL), | |
441 #ifndef PRODUCT | |
442 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), | |
443 _printer(IdealGraphPrinter::printer()), | |
444 #endif | |
445 _congraph(NULL) { | |
446 C = this; | |
447 | |
448 CompileWrapper cw(this); | |
449 #ifndef PRODUCT | |
450 if (TimeCompiler2) { | |
451 tty->print(" "); | |
452 target->holder()->name()->print(); | |
453 tty->print("."); | |
454 target->print_short_name(); | |
455 tty->print(" "); | |
456 } | |
457 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); | |
458 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); | |
100
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
459 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
460 if (!print_opto_assembly) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
461 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
462 if (print_assembly && !Disassembler::can_decode()) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
463 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
464 print_opto_assembly = true; |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
465 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
466 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
467 set_print_assembly(print_opto_assembly); |
0 | 468 #endif |
469 | |
470 if (ProfileTraps) { | |
471 // Make sure the method being compiled gets its own MDO, | |
472 // so we can at least track the decompile_count(). | |
473 method()->build_method_data(); | |
474 } | |
475 | |
476 Init(::AliasLevel); | |
477 | |
478 | |
479 print_compile_messages(); | |
480 | |
481 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) | |
482 _ilt = InlineTree::build_inline_tree_root(); | |
483 else | |
484 _ilt = NULL; | |
485 | |
486 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice | |
487 assert(num_alias_types() >= AliasIdxRaw, ""); | |
488 | |
489 #define MINIMUM_NODE_HASH 1023 | |
490 // Node list that Iterative GVN will start with | |
491 Unique_Node_List for_igvn(comp_arena()); | |
492 set_for_igvn(&for_igvn); | |
493 | |
494 // GVN that will be run immediately on new nodes | |
495 uint estimated_size = method()->code_size()*4+64; | |
496 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); | |
497 PhaseGVN gvn(node_arena(), estimated_size); | |
498 set_initial_gvn(&gvn); | |
499 | |
500 { // Scope for timing the parser | |
501 TracePhase t3("parse", &_t_parser, true); | |
502 | |
503 // Put top into the hash table ASAP. | |
504 initial_gvn()->transform_no_reclaim(top()); | |
505 | |
506 // Set up tf(), start(), and find a CallGenerator. | |
507 CallGenerator* cg; | |
508 if (is_osr_compilation()) { | |
509 const TypeTuple *domain = StartOSRNode::osr_domain(); | |
510 const TypeTuple *range = TypeTuple::make_range(method()->signature()); | |
511 init_tf(TypeFunc::make(domain, range)); | |
512 StartNode* s = new (this, 2) StartOSRNode(root(), domain); | |
513 initial_gvn()->set_type_bottom(s); | |
514 init_start(s); | |
515 cg = CallGenerator::for_osr(method(), entry_bci()); | |
516 } else { | |
517 // Normal case. | |
518 init_tf(TypeFunc::make(method())); | |
519 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); | |
520 initial_gvn()->set_type_bottom(s); | |
521 init_start(s); | |
522 float past_uses = method()->interpreter_invocation_count(); | |
523 float expected_uses = past_uses; | |
524 cg = CallGenerator::for_inline(method(), expected_uses); | |
525 } | |
526 if (failing()) return; | |
527 if (cg == NULL) { | |
528 record_method_not_compilable_all_tiers("cannot parse method"); | |
529 return; | |
530 } | |
531 JVMState* jvms = build_start_state(start(), tf()); | |
532 if ((jvms = cg->generate(jvms)) == NULL) { | |
533 record_method_not_compilable("method parse failed"); | |
534 return; | |
535 } | |
536 GraphKit kit(jvms); | |
537 | |
538 if (!kit.stopped()) { | |
539 // Accept return values, and transfer control we know not where. | |
540 // This is done by a special, unique ReturnNode bound to root. | |
541 return_values(kit.jvms()); | |
542 } | |
543 | |
544 if (kit.has_exceptions()) { | |
545 // Any exceptions that escape from this call must be rethrown | |
546 // to whatever caller is dynamically above us on the stack. | |
547 // This is done by a special, unique RethrowNode bound to root. | |
548 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); | |
549 } | |
550 | |
551 // Remove clutter produced by parsing. | |
552 if (!failing()) { | |
553 ResourceMark rm; | |
554 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); | |
555 } | |
556 } | |
557 | |
558 // Note: Large methods are capped off in do_one_bytecode(). | |
559 if (failing()) return; | |
560 | |
561 // After parsing, node notes are no longer automagic. | |
562 // They must be propagated by register_new_node_with_optimizer(), | |
563 // clone(), or the like. | |
564 set_default_node_notes(NULL); | |
565 | |
566 for (;;) { | |
567 int successes = Inline_Warm(); | |
568 if (failing()) return; | |
569 if (successes == 0) break; | |
570 } | |
571 | |
572 // Drain the list. | |
573 Finish_Warm(); | |
574 #ifndef PRODUCT | |
575 if (_printer) { | |
576 _printer->print_inlining(this); | |
577 } | |
578 #endif | |
579 | |
580 if (failing()) return; | |
581 NOT_PRODUCT( verify_graph_edges(); ) | |
582 | |
583 // Perform escape analysis | |
65 | 584 if (_do_escape_analysis) |
585 _congraph = new ConnectionGraph(this); | |
0 | 586 if (_congraph != NULL) { |
587 NOT_PRODUCT( TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, TimeCompiler); ) | |
588 _congraph->compute_escape(); | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
589 if (failing()) return; |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
590 |
0 | 591 #ifndef PRODUCT |
592 if (PrintEscapeAnalysis) { | |
593 _congraph->dump(); | |
594 } | |
595 #endif | |
596 } | |
597 // Now optimize | |
598 Optimize(); | |
599 if (failing()) return; | |
600 NOT_PRODUCT( verify_graph_edges(); ) | |
601 | |
602 #ifndef PRODUCT | |
603 if (PrintIdeal) { | |
604 ttyLocker ttyl; // keep the following output all in one block | |
605 // This output goes directly to the tty, not the compiler log. | |
606 // To enable tools to match it up with the compilation activity, | |
607 // be sure to tag this tty output with the compile ID. | |
608 if (xtty != NULL) { | |
609 xtty->head("ideal compile_id='%d'%s", compile_id(), | |
610 is_osr_compilation() ? " compile_kind='osr'" : | |
611 ""); | |
612 } | |
613 root()->dump(9999); | |
614 if (xtty != NULL) { | |
615 xtty->tail("ideal"); | |
616 } | |
617 } | |
618 #endif | |
619 | |
620 // Now that we know the size of all the monitors we can add a fixed slot | |
621 // for the original deopt pc. | |
622 | |
623 _orig_pc_slot = fixed_slots(); | |
624 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); | |
625 set_fixed_slots(next_slot); | |
626 | |
627 // Now generate code | |
628 Code_Gen(); | |
629 if (failing()) return; | |
630 | |
631 // Check if we want to skip execution of all compiled code. | |
632 { | |
633 #ifndef PRODUCT | |
634 if (OptoNoExecute) { | |
635 record_method_not_compilable("+OptoNoExecute"); // Flag as failed | |
636 return; | |
637 } | |
638 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); | |
639 #endif | |
640 | |
641 if (is_osr_compilation()) { | |
642 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); | |
643 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); | |
644 } else { | |
645 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); | |
646 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); | |
647 } | |
648 | |
649 env()->register_method(_method, _entry_bci, | |
650 &_code_offsets, | |
651 _orig_pc_slot_offset_in_bytes, | |
652 code_buffer(), | |
653 frame_size_in_words(), _oop_map_set, | |
654 &_handler_table, &_inc_table, | |
655 compiler, | |
656 env()->comp_level(), | |
657 true, /*has_debug_info*/ | |
658 has_unsafe_access() | |
659 ); | |
660 } | |
661 } | |
662 | |
663 //------------------------------Compile---------------------------------------- | |
664 // Compile a runtime stub | |
665 Compile::Compile( ciEnv* ci_env, | |
666 TypeFunc_generator generator, | |
667 address stub_function, | |
668 const char *stub_name, | |
669 int is_fancy_jump, | |
670 bool pass_tls, | |
671 bool save_arg_registers, | |
672 bool return_pc ) | |
673 : Phase(Compiler), | |
674 _env(ci_env), | |
675 _log(ci_env->log()), | |
676 _compile_id(-1), | |
677 _save_argument_registers(save_arg_registers), | |
678 _method(NULL), | |
679 _stub_name(stub_name), | |
680 _stub_function(stub_function), | |
681 _stub_entry_point(NULL), | |
682 _entry_bci(InvocationEntryBci), | |
683 _initial_gvn(NULL), | |
684 _for_igvn(NULL), | |
685 _warm_calls(NULL), | |
686 _orig_pc_slot(0), | |
687 _orig_pc_slot_offset_in_bytes(0), | |
688 _subsume_loads(true), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
689 _do_escape_analysis(false), |
0 | 690 _failure_reason(NULL), |
691 _code_buffer("Compile::Fill_buffer"), | |
692 _node_bundling_limit(0), | |
693 _node_bundling_base(NULL), | |
694 #ifndef PRODUCT | |
695 _trace_opto_output(TraceOptoOutput), | |
696 _printer(NULL), | |
697 #endif | |
698 _congraph(NULL) { | |
699 C = this; | |
700 | |
701 #ifndef PRODUCT | |
702 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); | |
703 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); | |
704 set_print_assembly(PrintFrameConverterAssembly); | |
705 #endif | |
706 CompileWrapper cw(this); | |
707 Init(/*AliasLevel=*/ 0); | |
708 init_tf((*generator)()); | |
709 | |
710 { | |
711 // The following is a dummy for the sake of GraphKit::gen_stub | |
712 Unique_Node_List for_igvn(comp_arena()); | |
713 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this | |
714 PhaseGVN gvn(Thread::current()->resource_area(),255); | |
715 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively | |
716 gvn.transform_no_reclaim(top()); | |
717 | |
718 GraphKit kit; | |
719 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); | |
720 } | |
721 | |
722 NOT_PRODUCT( verify_graph_edges(); ) | |
723 Code_Gen(); | |
724 if (failing()) return; | |
725 | |
726 | |
727 // Entry point will be accessed using compile->stub_entry_point(); | |
728 if (code_buffer() == NULL) { | |
729 Matcher::soft_match_failure(); | |
730 } else { | |
731 if (PrintAssembly && (WizardMode || Verbose)) | |
732 tty->print_cr("### Stub::%s", stub_name); | |
733 | |
734 if (!failing()) { | |
735 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); | |
736 | |
737 // Make the NMethod | |
738 // For now we mark the frame as never safe for profile stackwalking | |
739 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, | |
740 code_buffer(), | |
741 CodeOffsets::frame_never_safe, | |
742 // _code_offsets.value(CodeOffsets::Frame_Complete), | |
743 frame_size_in_words(), | |
744 _oop_map_set, | |
745 save_arg_registers); | |
746 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); | |
747 | |
748 _stub_entry_point = rs->entry_point(); | |
749 } | |
750 } | |
751 } | |
752 | |
753 #ifndef PRODUCT | |
754 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { | |
755 if(PrintOpto && Verbose) { | |
756 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); | |
757 } | |
758 } | |
759 #endif | |
760 | |
761 void Compile::print_codes() { | |
762 } | |
763 | |
764 //------------------------------Init------------------------------------------- | |
765 // Prepare for a single compilation | |
766 void Compile::Init(int aliaslevel) { | |
767 _unique = 0; | |
768 _regalloc = NULL; | |
769 | |
770 _tf = NULL; // filled in later | |
771 _top = NULL; // cached later | |
772 _matcher = NULL; // filled in later | |
773 _cfg = NULL; // filled in later | |
774 | |
775 set_24_bit_selection_and_mode(Use24BitFP, false); | |
776 | |
777 _node_note_array = NULL; | |
778 _default_node_notes = NULL; | |
779 | |
780 _immutable_memory = NULL; // filled in at first inquiry | |
781 | |
782 // Globally visible Nodes | |
783 // First set TOP to NULL to give safe behavior during creation of RootNode | |
784 set_cached_top_node(NULL); | |
785 set_root(new (this, 3) RootNode()); | |
786 // Now that you have a Root to point to, create the real TOP | |
787 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); | |
788 set_recent_alloc(NULL, NULL); | |
789 | |
790 // Create Debug Information Recorder to record scopes, oopmaps, etc. | |
791 env()->set_oop_recorder(new OopRecorder(comp_arena())); | |
792 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); | |
793 env()->set_dependencies(new Dependencies(env())); | |
794 | |
795 _fixed_slots = 0; | |
796 set_has_split_ifs(false); | |
797 set_has_loops(has_method() && method()->has_loops()); // first approximation | |
798 _deopt_happens = true; // start out assuming the worst | |
799 _trap_can_recompile = false; // no traps emitted yet | |
800 _major_progress = true; // start out assuming good things will happen | |
801 set_has_unsafe_access(false); | |
802 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); | |
803 set_decompile_count(0); | |
804 | |
805 // Compilation level related initialization | |
806 if (env()->comp_level() == CompLevel_fast_compile) { | |
807 set_num_loop_opts(Tier1LoopOptsCount); | |
808 set_do_inlining(Tier1Inline != 0); | |
809 set_max_inline_size(Tier1MaxInlineSize); | |
810 set_freq_inline_size(Tier1FreqInlineSize); | |
811 set_do_scheduling(false); | |
812 set_do_count_invocations(Tier1CountInvocations); | |
813 set_do_method_data_update(Tier1UpdateMethodData); | |
814 } else { | |
815 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); | |
816 set_num_loop_opts(LoopOptsCount); | |
817 set_do_inlining(Inline); | |
818 set_max_inline_size(MaxInlineSize); | |
819 set_freq_inline_size(FreqInlineSize); | |
820 set_do_scheduling(OptoScheduling); | |
821 set_do_count_invocations(false); | |
822 set_do_method_data_update(false); | |
823 } | |
824 | |
825 if (debug_info()->recording_non_safepoints()) { | |
826 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> | |
827 (comp_arena(), 8, 0, NULL)); | |
828 set_default_node_notes(Node_Notes::make(this)); | |
829 } | |
830 | |
831 // // -- Initialize types before each compile -- | |
832 // // Update cached type information | |
833 // if( _method && _method->constants() ) | |
834 // Type::update_loaded_types(_method, _method->constants()); | |
835 | |
836 // Init alias_type map. | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
837 if (!_do_escape_analysis && aliaslevel == 3) |
0 | 838 aliaslevel = 2; // No unique types without escape analysis |
839 _AliasLevel = aliaslevel; | |
840 const int grow_ats = 16; | |
841 _max_alias_types = grow_ats; | |
842 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); | |
843 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); | |
844 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); | |
845 { | |
846 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; | |
847 } | |
848 // Initialize the first few types. | |
849 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); | |
850 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); | |
851 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); | |
852 _num_alias_types = AliasIdxRaw+1; | |
853 // Zero out the alias type cache. | |
854 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); | |
855 // A NULL adr_type hits in the cache right away. Preload the right answer. | |
856 probe_alias_cache(NULL)->_index = AliasIdxTop; | |
857 | |
858 _intrinsics = NULL; | |
859 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); | |
860 register_library_intrinsics(); | |
861 } | |
862 | |
863 //---------------------------init_start---------------------------------------- | |
864 // Install the StartNode on this compile object. | |
865 void Compile::init_start(StartNode* s) { | |
866 if (failing()) | |
867 return; // already failing | |
868 assert(s == start(), ""); | |
869 } | |
870 | |
871 StartNode* Compile::start() const { | |
872 assert(!failing(), ""); | |
873 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { | |
874 Node* start = root()->fast_out(i); | |
875 if( start->is_Start() ) | |
876 return start->as_Start(); | |
877 } | |
878 ShouldNotReachHere(); | |
879 return NULL; | |
880 } | |
881 | |
882 //-------------------------------immutable_memory------------------------------------- | |
883 // Access immutable memory | |
884 Node* Compile::immutable_memory() { | |
885 if (_immutable_memory != NULL) { | |
886 return _immutable_memory; | |
887 } | |
888 StartNode* s = start(); | |
889 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { | |
890 Node *p = s->fast_out(i); | |
891 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { | |
892 _immutable_memory = p; | |
893 return _immutable_memory; | |
894 } | |
895 } | |
896 ShouldNotReachHere(); | |
897 return NULL; | |
898 } | |
899 | |
900 //----------------------set_cached_top_node------------------------------------ | |
901 // Install the cached top node, and make sure Node::is_top works correctly. | |
902 void Compile::set_cached_top_node(Node* tn) { | |
903 if (tn != NULL) verify_top(tn); | |
904 Node* old_top = _top; | |
905 _top = tn; | |
906 // Calling Node::setup_is_top allows the nodes the chance to adjust | |
907 // their _out arrays. | |
908 if (_top != NULL) _top->setup_is_top(); | |
909 if (old_top != NULL) old_top->setup_is_top(); | |
910 assert(_top == NULL || top()->is_top(), ""); | |
911 } | |
912 | |
913 #ifndef PRODUCT | |
914 void Compile::verify_top(Node* tn) const { | |
915 if (tn != NULL) { | |
916 assert(tn->is_Con(), "top node must be a constant"); | |
917 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); | |
918 assert(tn->in(0) != NULL, "must have live top node"); | |
919 } | |
920 } | |
921 #endif | |
922 | |
923 | |
924 ///-------------------Managing Per-Node Debug & Profile Info------------------- | |
925 | |
926 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { | |
927 guarantee(arr != NULL, ""); | |
928 int num_blocks = arr->length(); | |
929 if (grow_by < num_blocks) grow_by = num_blocks; | |
930 int num_notes = grow_by * _node_notes_block_size; | |
931 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); | |
932 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); | |
933 while (num_notes > 0) { | |
934 arr->append(notes); | |
935 notes += _node_notes_block_size; | |
936 num_notes -= _node_notes_block_size; | |
937 } | |
938 assert(num_notes == 0, "exact multiple, please"); | |
939 } | |
940 | |
941 bool Compile::copy_node_notes_to(Node* dest, Node* source) { | |
942 if (source == NULL || dest == NULL) return false; | |
943 | |
944 if (dest->is_Con()) | |
945 return false; // Do not push debug info onto constants. | |
946 | |
947 #ifdef ASSERT | |
948 // Leave a bread crumb trail pointing to the original node: | |
949 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { | |
950 dest->set_debug_orig(source); | |
951 } | |
952 #endif | |
953 | |
954 if (node_note_array() == NULL) | |
955 return false; // Not collecting any notes now. | |
956 | |
957 // This is a copy onto a pre-existing node, which may already have notes. | |
958 // If both nodes have notes, do not overwrite any pre-existing notes. | |
959 Node_Notes* source_notes = node_notes_at(source->_idx); | |
960 if (source_notes == NULL || source_notes->is_clear()) return false; | |
961 Node_Notes* dest_notes = node_notes_at(dest->_idx); | |
962 if (dest_notes == NULL || dest_notes->is_clear()) { | |
963 return set_node_notes_at(dest->_idx, source_notes); | |
964 } | |
965 | |
966 Node_Notes merged_notes = (*source_notes); | |
967 // The order of operations here ensures that dest notes will win... | |
968 merged_notes.update_from(dest_notes); | |
969 return set_node_notes_at(dest->_idx, &merged_notes); | |
970 } | |
971 | |
972 | |
973 //--------------------------allow_range_check_smearing------------------------- | |
974 // Gating condition for coalescing similar range checks. | |
975 // Sometimes we try 'speculatively' replacing a series of a range checks by a | |
976 // single covering check that is at least as strong as any of them. | |
977 // If the optimization succeeds, the simplified (strengthened) range check | |
978 // will always succeed. If it fails, we will deopt, and then give up | |
979 // on the optimization. | |
980 bool Compile::allow_range_check_smearing() const { | |
981 // If this method has already thrown a range-check, | |
982 // assume it was because we already tried range smearing | |
983 // and it failed. | |
984 uint already_trapped = trap_count(Deoptimization::Reason_range_check); | |
985 return !already_trapped; | |
986 } | |
987 | |
988 | |
989 //------------------------------flatten_alias_type----------------------------- | |
990 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { | |
991 int offset = tj->offset(); | |
992 TypePtr::PTR ptr = tj->ptr(); | |
993 | |
994 // Process weird unsafe references. | |
995 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { | |
996 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); | |
997 tj = TypeOopPtr::BOTTOM; | |
998 ptr = tj->ptr(); | |
999 offset = tj->offset(); | |
1000 } | |
1001 | |
1002 // Array pointers need some flattening | |
1003 const TypeAryPtr *ta = tj->isa_aryptr(); | |
1004 if( ta && _AliasLevel >= 2 ) { | |
1005 // For arrays indexed by constant indices, we flatten the alias | |
1006 // space to include all of the array body. Only the header, klass | |
1007 // and array length can be accessed un-aliased. | |
1008 if( offset != Type::OffsetBot ) { | |
1009 if( ta->const_oop() ) { // methodDataOop or methodOop | |
1010 offset = Type::OffsetBot; // Flatten constant access into array body | |
1011 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); | |
1012 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { | |
1013 // range is OK as-is. | |
1014 tj = ta = TypeAryPtr::RANGE; | |
1015 } else if( offset == oopDesc::klass_offset_in_bytes() ) { | |
1016 tj = TypeInstPtr::KLASS; // all klass loads look alike | |
1017 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1018 ptr = TypePtr::BotPTR; | |
1019 } else if( offset == oopDesc::mark_offset_in_bytes() ) { | |
1020 tj = TypeInstPtr::MARK; | |
1021 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1022 ptr = TypePtr::BotPTR; | |
1023 } else { // Random constant offset into array body | |
1024 offset = Type::OffsetBot; // Flatten constant access into array body | |
1025 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); | |
1026 } | |
1027 } | |
1028 // Arrays of fixed size alias with arrays of unknown size. | |
1029 if (ta->size() != TypeInt::POS) { | |
1030 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); | |
1031 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset, ta->instance_id()); | |
1032 } | |
1033 // Arrays of known objects become arrays of unknown objects. | |
1034 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { | |
1035 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); | |
1036 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id()); | |
1037 } | |
1038 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so | |
1039 // cannot be distinguished by bytecode alone. | |
1040 if (ta->elem() == TypeInt::BOOL) { | |
1041 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); | |
1042 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); | |
1043 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset, ta->instance_id()); | |
1044 } | |
1045 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1046 // Make sure the Bottom and NotNull variants alias the same. | |
1047 // Also, make sure exact and non-exact variants alias the same. | |
1048 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { | |
1049 if (ta->const_oop()) { | |
1050 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); | |
1051 } else { | |
1052 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); | |
1053 } | |
1054 } | |
1055 } | |
1056 | |
1057 // Oop pointers need some flattening | |
1058 const TypeInstPtr *to = tj->isa_instptr(); | |
1059 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { | |
1060 if( ptr == TypePtr::Constant ) { | |
1061 // No constant oop pointers (such as Strings); they alias with | |
1062 // unknown strings. | |
1063 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); | |
1064 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { | |
1065 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1066 // Make sure the Bottom and NotNull variants alias the same. | |
1067 // Also, make sure exact and non-exact variants alias the same. | |
1068 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset, to->instance_id()); | |
1069 } | |
1070 // Canonicalize the holder of this field | |
1071 ciInstanceKlass *k = to->klass()->as_instance_klass(); | |
1072 if (offset >= 0 && offset < oopDesc::header_size() * wordSize) { | |
1073 // First handle header references such as a LoadKlassNode, even if the | |
1074 // object's klass is unloaded at compile time (4965979). | |
1075 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset, to->instance_id()); | |
1076 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { | |
1077 to = NULL; | |
1078 tj = TypeOopPtr::BOTTOM; | |
1079 offset = tj->offset(); | |
1080 } else { | |
1081 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); | |
1082 if (!k->equals(canonical_holder) || tj->offset() != offset) { | |
1083 tj = to = TypeInstPtr::make(TypePtr::BotPTR, canonical_holder, false, NULL, offset, to->instance_id()); | |
1084 } | |
1085 } | |
1086 } | |
1087 | |
1088 // Klass pointers to object array klasses need some flattening | |
1089 const TypeKlassPtr *tk = tj->isa_klassptr(); | |
1090 if( tk ) { | |
1091 // If we are referencing a field within a Klass, we need | |
1092 // to assume the worst case of an Object. Both exact and | |
1093 // inexact types must flatten to the same alias class. | |
1094 // Since the flattened result for a klass is defined to be | |
1095 // precisely java.lang.Object, use a constant ptr. | |
1096 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { | |
1097 | |
1098 tj = tk = TypeKlassPtr::make(TypePtr::Constant, | |
1099 TypeKlassPtr::OBJECT->klass(), | |
1100 offset); | |
1101 } | |
1102 | |
1103 ciKlass* klass = tk->klass(); | |
1104 if( klass->is_obj_array_klass() ) { | |
1105 ciKlass* k = TypeAryPtr::OOPS->klass(); | |
1106 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs | |
1107 k = TypeInstPtr::BOTTOM->klass(); | |
1108 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); | |
1109 } | |
1110 | |
1111 // Check for precise loads from the primary supertype array and force them | |
1112 // to the supertype cache alias index. Check for generic array loads from | |
1113 // the primary supertype array and also force them to the supertype cache | |
1114 // alias index. Since the same load can reach both, we need to merge | |
1115 // these 2 disparate memories into the same alias class. Since the | |
1116 // primary supertype array is read-only, there's no chance of confusion | |
1117 // where we bypass an array load and an array store. | |
1118 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); | |
1119 if( offset == Type::OffsetBot || | |
1120 off2 < Klass::primary_super_limit()*wordSize ) { | |
1121 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); | |
1122 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); | |
1123 } | |
1124 } | |
1125 | |
1126 // Flatten all Raw pointers together. | |
1127 if (tj->base() == Type::RawPtr) | |
1128 tj = TypeRawPtr::BOTTOM; | |
1129 | |
1130 if (tj->base() == Type::AnyPtr) | |
1131 tj = TypePtr::BOTTOM; // An error, which the caller must check for. | |
1132 | |
1133 // Flatten all to bottom for now | |
1134 switch( _AliasLevel ) { | |
1135 case 0: | |
1136 tj = TypePtr::BOTTOM; | |
1137 break; | |
1138 case 1: // Flatten to: oop, static, field or array | |
1139 switch (tj->base()) { | |
1140 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; | |
1141 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; | |
1142 case Type::AryPtr: // do not distinguish arrays at all | |
1143 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; | |
1144 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; | |
1145 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it | |
1146 default: ShouldNotReachHere(); | |
1147 } | |
1148 break; | |
1149 case 2: // No collasping at level 2; keep all splits | |
1150 case 3: // No collasping at level 3; keep all splits | |
1151 break; | |
1152 default: | |
1153 Unimplemented(); | |
1154 } | |
1155 | |
1156 offset = tj->offset(); | |
1157 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); | |
1158 | |
1159 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || | |
1160 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || | |
1161 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || | |
1162 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || | |
1163 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1164 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1165 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , | |
1166 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); | |
1167 assert( tj->ptr() != TypePtr::TopPTR && | |
1168 tj->ptr() != TypePtr::AnyNull && | |
1169 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); | |
1170 // assert( tj->ptr() != TypePtr::Constant || | |
1171 // tj->base() == Type::RawPtr || | |
1172 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); | |
1173 | |
1174 return tj; | |
1175 } | |
1176 | |
1177 void Compile::AliasType::Init(int i, const TypePtr* at) { | |
1178 _index = i; | |
1179 _adr_type = at; | |
1180 _field = NULL; | |
1181 _is_rewritable = true; // default | |
1182 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; | |
1183 if (atoop != NULL && atoop->is_instance()) { | |
1184 const TypeOopPtr *gt = atoop->cast_to_instance(TypeOopPtr::UNKNOWN_INSTANCE); | |
1185 _general_index = Compile::current()->get_alias_index(gt); | |
1186 } else { | |
1187 _general_index = 0; | |
1188 } | |
1189 } | |
1190 | |
1191 //---------------------------------print_on------------------------------------ | |
1192 #ifndef PRODUCT | |
1193 void Compile::AliasType::print_on(outputStream* st) { | |
1194 if (index() < 10) | |
1195 st->print("@ <%d> ", index()); | |
1196 else st->print("@ <%d>", index()); | |
1197 st->print(is_rewritable() ? " " : " RO"); | |
1198 int offset = adr_type()->offset(); | |
1199 if (offset == Type::OffsetBot) | |
1200 st->print(" +any"); | |
1201 else st->print(" +%-3d", offset); | |
1202 st->print(" in "); | |
1203 adr_type()->dump_on(st); | |
1204 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); | |
1205 if (field() != NULL && tjp) { | |
1206 if (tjp->klass() != field()->holder() || | |
1207 tjp->offset() != field()->offset_in_bytes()) { | |
1208 st->print(" != "); | |
1209 field()->print(); | |
1210 st->print(" ***"); | |
1211 } | |
1212 } | |
1213 } | |
1214 | |
1215 void print_alias_types() { | |
1216 Compile* C = Compile::current(); | |
1217 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); | |
1218 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { | |
1219 C->alias_type(idx)->print_on(tty); | |
1220 tty->cr(); | |
1221 } | |
1222 } | |
1223 #endif | |
1224 | |
1225 | |
1226 //----------------------------probe_alias_cache-------------------------------- | |
1227 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { | |
1228 intptr_t key = (intptr_t) adr_type; | |
1229 key ^= key >> logAliasCacheSize; | |
1230 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; | |
1231 } | |
1232 | |
1233 | |
1234 //-----------------------------grow_alias_types-------------------------------- | |
1235 void Compile::grow_alias_types() { | |
1236 const int old_ats = _max_alias_types; // how many before? | |
1237 const int new_ats = old_ats; // how many more? | |
1238 const int grow_ats = old_ats+new_ats; // how many now? | |
1239 _max_alias_types = grow_ats; | |
1240 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); | |
1241 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); | |
1242 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); | |
1243 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; | |
1244 } | |
1245 | |
1246 | |
1247 //--------------------------------find_alias_type------------------------------ | |
1248 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { | |
1249 if (_AliasLevel == 0) | |
1250 return alias_type(AliasIdxBot); | |
1251 | |
1252 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1253 if (ace->_adr_type == adr_type) { | |
1254 return alias_type(ace->_index); | |
1255 } | |
1256 | |
1257 // Handle special cases. | |
1258 if (adr_type == NULL) return alias_type(AliasIdxTop); | |
1259 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); | |
1260 | |
1261 // Do it the slow way. | |
1262 const TypePtr* flat = flatten_alias_type(adr_type); | |
1263 | |
1264 #ifdef ASSERT | |
1265 assert(flat == flatten_alias_type(flat), "idempotent"); | |
1266 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); | |
1267 if (flat->isa_oopptr() && !flat->isa_klassptr()) { | |
1268 const TypeOopPtr* foop = flat->is_oopptr(); | |
1269 const TypePtr* xoop = foop->cast_to_exactness(!foop->klass_is_exact())->is_ptr(); | |
1270 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); | |
1271 } | |
1272 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); | |
1273 #endif | |
1274 | |
1275 int idx = AliasIdxTop; | |
1276 for (int i = 0; i < num_alias_types(); i++) { | |
1277 if (alias_type(i)->adr_type() == flat) { | |
1278 idx = i; | |
1279 break; | |
1280 } | |
1281 } | |
1282 | |
1283 if (idx == AliasIdxTop) { | |
1284 if (no_create) return NULL; | |
1285 // Grow the array if necessary. | |
1286 if (_num_alias_types == _max_alias_types) grow_alias_types(); | |
1287 // Add a new alias type. | |
1288 idx = _num_alias_types++; | |
1289 _alias_types[idx]->Init(idx, flat); | |
1290 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); | |
1291 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); | |
1292 if (flat->isa_instptr()) { | |
1293 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() | |
1294 && flat->is_instptr()->klass() == env()->Class_klass()) | |
1295 alias_type(idx)->set_rewritable(false); | |
1296 } | |
1297 if (flat->isa_klassptr()) { | |
1298 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1299 alias_type(idx)->set_rewritable(false); | |
1300 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1301 alias_type(idx)->set_rewritable(false); | |
1302 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1303 alias_type(idx)->set_rewritable(false); | |
1304 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1305 alias_type(idx)->set_rewritable(false); | |
1306 } | |
1307 // %%% (We would like to finalize JavaThread::threadObj_offset(), | |
1308 // but the base pointer type is not distinctive enough to identify | |
1309 // references into JavaThread.) | |
1310 | |
1311 // Check for final instance fields. | |
1312 const TypeInstPtr* tinst = flat->isa_instptr(); | |
1313 if (tinst && tinst->offset() >= oopDesc::header_size() * wordSize) { | |
1314 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); | |
1315 ciField* field = k->get_field_by_offset(tinst->offset(), false); | |
1316 // Set field() and is_rewritable() attributes. | |
1317 if (field != NULL) alias_type(idx)->set_field(field); | |
1318 } | |
1319 const TypeKlassPtr* tklass = flat->isa_klassptr(); | |
1320 // Check for final static fields. | |
1321 if (tklass && tklass->klass()->is_instance_klass()) { | |
1322 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); | |
1323 ciField* field = k->get_field_by_offset(tklass->offset(), true); | |
1324 // Set field() and is_rewritable() attributes. | |
1325 if (field != NULL) alias_type(idx)->set_field(field); | |
1326 } | |
1327 } | |
1328 | |
1329 // Fill the cache for next time. | |
1330 ace->_adr_type = adr_type; | |
1331 ace->_index = idx; | |
1332 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); | |
1333 | |
1334 // Might as well try to fill the cache for the flattened version, too. | |
1335 AliasCacheEntry* face = probe_alias_cache(flat); | |
1336 if (face->_adr_type == NULL) { | |
1337 face->_adr_type = flat; | |
1338 face->_index = idx; | |
1339 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); | |
1340 } | |
1341 | |
1342 return alias_type(idx); | |
1343 } | |
1344 | |
1345 | |
1346 Compile::AliasType* Compile::alias_type(ciField* field) { | |
1347 const TypeOopPtr* t; | |
1348 if (field->is_static()) | |
1349 t = TypeKlassPtr::make(field->holder()); | |
1350 else | |
1351 t = TypeOopPtr::make_from_klass_raw(field->holder()); | |
1352 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); | |
1353 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); | |
1354 return atp; | |
1355 } | |
1356 | |
1357 | |
1358 //------------------------------have_alias_type-------------------------------- | |
1359 bool Compile::have_alias_type(const TypePtr* adr_type) { | |
1360 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1361 if (ace->_adr_type == adr_type) { | |
1362 return true; | |
1363 } | |
1364 | |
1365 // Handle special cases. | |
1366 if (adr_type == NULL) return true; | |
1367 if (adr_type == TypePtr::BOTTOM) return true; | |
1368 | |
1369 return find_alias_type(adr_type, true) != NULL; | |
1370 } | |
1371 | |
1372 //-----------------------------must_alias-------------------------------------- | |
1373 // True if all values of the given address type are in the given alias category. | |
1374 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { | |
1375 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1376 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP | |
1377 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1378 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins | |
1379 | |
1380 // the only remaining possible overlap is identity | |
1381 int adr_idx = get_alias_index(adr_type); | |
1382 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1383 assert(adr_idx == alias_idx || | |
1384 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM | |
1385 && adr_type != TypeOopPtr::BOTTOM), | |
1386 "should not be testing for overlap with an unsafe pointer"); | |
1387 return adr_idx == alias_idx; | |
1388 } | |
1389 | |
1390 //------------------------------can_alias-------------------------------------- | |
1391 // True if any values of the given address type are in the given alias category. | |
1392 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { | |
1393 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1394 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP | |
1395 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1396 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins | |
1397 | |
1398 // the only remaining possible overlap is identity | |
1399 int adr_idx = get_alias_index(adr_type); | |
1400 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1401 return adr_idx == alias_idx; | |
1402 } | |
1403 | |
1404 | |
1405 | |
1406 //---------------------------pop_warm_call------------------------------------- | |
1407 WarmCallInfo* Compile::pop_warm_call() { | |
1408 WarmCallInfo* wci = _warm_calls; | |
1409 if (wci != NULL) _warm_calls = wci->remove_from(wci); | |
1410 return wci; | |
1411 } | |
1412 | |
1413 //----------------------------Inline_Warm-------------------------------------- | |
1414 int Compile::Inline_Warm() { | |
1415 // If there is room, try to inline some more warm call sites. | |
1416 // %%% Do a graph index compaction pass when we think we're out of space? | |
1417 if (!InlineWarmCalls) return 0; | |
1418 | |
1419 int calls_made_hot = 0; | |
1420 int room_to_grow = NodeCountInliningCutoff - unique(); | |
1421 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); | |
1422 int amount_grown = 0; | |
1423 WarmCallInfo* call; | |
1424 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { | |
1425 int est_size = (int)call->size(); | |
1426 if (est_size > (room_to_grow - amount_grown)) { | |
1427 // This one won't fit anyway. Get rid of it. | |
1428 call->make_cold(); | |
1429 continue; | |
1430 } | |
1431 call->make_hot(); | |
1432 calls_made_hot++; | |
1433 amount_grown += est_size; | |
1434 amount_to_grow -= est_size; | |
1435 } | |
1436 | |
1437 if (calls_made_hot > 0) set_major_progress(); | |
1438 return calls_made_hot; | |
1439 } | |
1440 | |
1441 | |
1442 //----------------------------Finish_Warm-------------------------------------- | |
1443 void Compile::Finish_Warm() { | |
1444 if (!InlineWarmCalls) return; | |
1445 if (failing()) return; | |
1446 if (warm_calls() == NULL) return; | |
1447 | |
1448 // Clean up loose ends, if we are out of space for inlining. | |
1449 WarmCallInfo* call; | |
1450 while ((call = pop_warm_call()) != NULL) { | |
1451 call->make_cold(); | |
1452 } | |
1453 } | |
1454 | |
1455 | |
1456 //------------------------------Optimize--------------------------------------- | |
1457 // Given a graph, optimize it. | |
1458 void Compile::Optimize() { | |
1459 TracePhase t1("optimizer", &_t_optimizer, true); | |
1460 | |
1461 #ifndef PRODUCT | |
1462 if (env()->break_at_compile()) { | |
1463 BREAKPOINT; | |
1464 } | |
1465 | |
1466 #endif | |
1467 | |
1468 ResourceMark rm; | |
1469 int loop_opts_cnt; | |
1470 | |
1471 NOT_PRODUCT( verify_graph_edges(); ) | |
1472 | |
1473 print_method("Start"); | |
1474 | |
1475 { | |
1476 // Iterative Global Value Numbering, including ideal transforms | |
1477 // Initialize IterGVN with types and values from parse-time GVN | |
1478 PhaseIterGVN igvn(initial_gvn()); | |
1479 { | |
1480 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) | |
1481 igvn.optimize(); | |
1482 } | |
1483 | |
1484 print_method("Iter GVN 1", 2); | |
1485 | |
1486 if (failing()) return; | |
1487 | |
1488 // get rid of the connection graph since it's information is not | |
1489 // updated by optimizations | |
1490 _congraph = NULL; | |
1491 | |
1492 | |
1493 // Loop transforms on the ideal graph. Range Check Elimination, | |
1494 // peeling, unrolling, etc. | |
1495 | |
1496 // Set loop opts counter | |
1497 loop_opts_cnt = num_loop_opts(); | |
1498 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { | |
1499 { | |
1500 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
1501 PhaseIdealLoop ideal_loop( igvn, NULL, true ); | |
1502 loop_opts_cnt--; | |
1503 if (major_progress()) print_method("PhaseIdealLoop 1", 2); | |
1504 if (failing()) return; | |
1505 } | |
1506 // Loop opts pass if partial peeling occurred in previous pass | |
1507 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { | |
1508 TracePhase t3("idealLoop", &_t_idealLoop, true); | |
1509 PhaseIdealLoop ideal_loop( igvn, NULL, false ); | |
1510 loop_opts_cnt--; | |
1511 if (major_progress()) print_method("PhaseIdealLoop 2", 2); | |
1512 if (failing()) return; | |
1513 } | |
1514 // Loop opts pass for loop-unrolling before CCP | |
1515 if(major_progress() && (loop_opts_cnt > 0)) { | |
1516 TracePhase t4("idealLoop", &_t_idealLoop, true); | |
1517 PhaseIdealLoop ideal_loop( igvn, NULL, false ); | |
1518 loop_opts_cnt--; | |
1519 if (major_progress()) print_method("PhaseIdealLoop 3", 2); | |
1520 } | |
1521 } | |
1522 if (failing()) return; | |
1523 | |
1524 // Conditional Constant Propagation; | |
1525 PhaseCCP ccp( &igvn ); | |
1526 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); | |
1527 { | |
1528 TracePhase t2("ccp", &_t_ccp, true); | |
1529 ccp.do_transform(); | |
1530 } | |
1531 print_method("PhaseCPP 1", 2); | |
1532 | |
1533 assert( true, "Break here to ccp.dump_old2new_map()"); | |
1534 | |
1535 // Iterative Global Value Numbering, including ideal transforms | |
1536 { | |
1537 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) | |
1538 igvn = ccp; | |
1539 igvn.optimize(); | |
1540 } | |
1541 | |
1542 print_method("Iter GVN 2", 2); | |
1543 | |
1544 if (failing()) return; | |
1545 | |
1546 // Loop transforms on the ideal graph. Range Check Elimination, | |
1547 // peeling, unrolling, etc. | |
1548 if(loop_opts_cnt > 0) { | |
1549 debug_only( int cnt = 0; ); | |
1550 while(major_progress() && (loop_opts_cnt > 0)) { | |
1551 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
1552 assert( cnt++ < 40, "infinite cycle in loop optimization" ); | |
1553 PhaseIdealLoop ideal_loop( igvn, NULL, true ); | |
1554 loop_opts_cnt--; | |
1555 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); | |
1556 if (failing()) return; | |
1557 } | |
1558 } | |
1559 { | |
1560 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) | |
1561 PhaseMacroExpand mex(igvn); | |
1562 if (mex.expand_macro_nodes()) { | |
1563 assert(failing(), "must bail out w/ explicit message"); | |
1564 return; | |
1565 } | |
1566 } | |
1567 | |
1568 } // (End scope of igvn; run destructor if necessary for asserts.) | |
1569 | |
1570 // A method with only infinite loops has no edges entering loops from root | |
1571 { | |
1572 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) | |
1573 if (final_graph_reshaping()) { | |
1574 assert(failing(), "must bail out w/ explicit message"); | |
1575 return; | |
1576 } | |
1577 } | |
1578 | |
1579 print_method("Optimize finished", 2); | |
1580 } | |
1581 | |
1582 | |
1583 //------------------------------Code_Gen--------------------------------------- | |
1584 // Given a graph, generate code for it | |
1585 void Compile::Code_Gen() { | |
1586 if (failing()) return; | |
1587 | |
1588 // Perform instruction selection. You might think we could reclaim Matcher | |
1589 // memory PDQ, but actually the Matcher is used in generating spill code. | |
1590 // Internals of the Matcher (including some VectorSets) must remain live | |
1591 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage | |
1592 // set a bit in reclaimed memory. | |
1593 | |
1594 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
1595 // nodes. Mapping is only valid at the root of each matched subtree. | |
1596 NOT_PRODUCT( verify_graph_edges(); ) | |
1597 | |
1598 Node_List proj_list; | |
1599 Matcher m(proj_list); | |
1600 _matcher = &m; | |
1601 { | |
1602 TracePhase t2("matcher", &_t_matcher, true); | |
1603 m.match(); | |
1604 } | |
1605 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
1606 // nodes. Mapping is only valid at the root of each matched subtree. | |
1607 NOT_PRODUCT( verify_graph_edges(); ) | |
1608 | |
1609 // If you have too many nodes, or if matching has failed, bail out | |
1610 check_node_count(0, "out of nodes matching instructions"); | |
1611 if (failing()) return; | |
1612 | |
1613 // Build a proper-looking CFG | |
1614 PhaseCFG cfg(node_arena(), root(), m); | |
1615 _cfg = &cfg; | |
1616 { | |
1617 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) | |
1618 cfg.Dominators(); | |
1619 if (failing()) return; | |
1620 | |
1621 NOT_PRODUCT( verify_graph_edges(); ) | |
1622 | |
1623 cfg.Estimate_Block_Frequency(); | |
1624 cfg.GlobalCodeMotion(m,unique(),proj_list); | |
1625 | |
1626 print_method("Global code motion", 2); | |
1627 | |
1628 if (failing()) return; | |
1629 NOT_PRODUCT( verify_graph_edges(); ) | |
1630 | |
1631 debug_only( cfg.verify(); ) | |
1632 } | |
1633 NOT_PRODUCT( verify_graph_edges(); ) | |
1634 | |
1635 PhaseChaitin regalloc(unique(),cfg,m); | |
1636 _regalloc = ®alloc; | |
1637 { | |
1638 TracePhase t2("regalloc", &_t_registerAllocation, true); | |
1639 // Perform any platform dependent preallocation actions. This is used, | |
1640 // for example, to avoid taking an implicit null pointer exception | |
1641 // using the frame pointer on win95. | |
1642 _regalloc->pd_preallocate_hook(); | |
1643 | |
1644 // Perform register allocation. After Chaitin, use-def chains are | |
1645 // no longer accurate (at spill code) and so must be ignored. | |
1646 // Node->LRG->reg mappings are still accurate. | |
1647 _regalloc->Register_Allocate(); | |
1648 | |
1649 // Bail out if the allocator builds too many nodes | |
1650 if (failing()) return; | |
1651 } | |
1652 | |
1653 // Prior to register allocation we kept empty basic blocks in case the | |
1654 // the allocator needed a place to spill. After register allocation we | |
1655 // are not adding any new instructions. If any basic block is empty, we | |
1656 // can now safely remove it. | |
1657 { | |
1658 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); ) | |
1659 cfg.RemoveEmpty(); | |
1660 } | |
1661 | |
1662 // Perform any platform dependent postallocation verifications. | |
1663 debug_only( _regalloc->pd_postallocate_verify_hook(); ) | |
1664 | |
1665 // Apply peephole optimizations | |
1666 if( OptoPeephole ) { | |
1667 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) | |
1668 PhasePeephole peep( _regalloc, cfg); | |
1669 peep.do_transform(); | |
1670 } | |
1671 | |
1672 // Convert Nodes to instruction bits in a buffer | |
1673 { | |
1674 // %%%% workspace merge brought two timers together for one job | |
1675 TracePhase t2a("output", &_t_output, true); | |
1676 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) | |
1677 Output(); | |
1678 } | |
1679 | |
1680 print_method("End"); | |
1681 | |
1682 // He's dead, Jim. | |
1683 _cfg = (PhaseCFG*)0xdeadbeef; | |
1684 _regalloc = (PhaseChaitin*)0xdeadbeef; | |
1685 } | |
1686 | |
1687 | |
1688 //------------------------------dump_asm--------------------------------------- | |
1689 // Dump formatted assembly | |
1690 #ifndef PRODUCT | |
1691 void Compile::dump_asm(int *pcs, uint pc_limit) { | |
1692 bool cut_short = false; | |
1693 tty->print_cr("#"); | |
1694 tty->print("# "); _tf->dump(); tty->cr(); | |
1695 tty->print_cr("#"); | |
1696 | |
1697 // For all blocks | |
1698 int pc = 0x0; // Program counter | |
1699 char starts_bundle = ' '; | |
1700 _regalloc->dump_frame(); | |
1701 | |
1702 Node *n = NULL; | |
1703 for( uint i=0; i<_cfg->_num_blocks; i++ ) { | |
1704 if (VMThread::should_terminate()) { cut_short = true; break; } | |
1705 Block *b = _cfg->_blocks[i]; | |
1706 if (b->is_connector() && !Verbose) continue; | |
1707 n = b->_nodes[0]; | |
1708 if (pcs && n->_idx < pc_limit) | |
1709 tty->print("%3.3x ", pcs[n->_idx]); | |
1710 else | |
1711 tty->print(" "); | |
1712 b->dump_head( &_cfg->_bbs ); | |
1713 if (b->is_connector()) { | |
1714 tty->print_cr(" # Empty connector block"); | |
1715 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { | |
1716 tty->print_cr(" # Block is sole successor of call"); | |
1717 } | |
1718 | |
1719 // For all instructions | |
1720 Node *delay = NULL; | |
1721 for( uint j = 0; j<b->_nodes.size(); j++ ) { | |
1722 if (VMThread::should_terminate()) { cut_short = true; break; } | |
1723 n = b->_nodes[j]; | |
1724 if (valid_bundle_info(n)) { | |
1725 Bundle *bundle = node_bundling(n); | |
1726 if (bundle->used_in_unconditional_delay()) { | |
1727 delay = n; | |
1728 continue; | |
1729 } | |
1730 if (bundle->starts_bundle()) | |
1731 starts_bundle = '+'; | |
1732 } | |
1733 | |
1734 if( !n->is_Region() && // Dont print in the Assembly | |
1735 !n->is_Phi() && // a few noisely useless nodes | |
1736 !n->is_Proj() && | |
1737 !n->is_MachTemp() && | |
1738 !n->is_Catch() && // Would be nice to print exception table targets | |
1739 !n->is_MergeMem() && // Not very interesting | |
1740 !n->is_top() && // Debug info table constants | |
1741 !(n->is_Con() && !n->is_Mach())// Debug info table constants | |
1742 ) { | |
1743 if (pcs && n->_idx < pc_limit) | |
1744 tty->print("%3.3x", pcs[n->_idx]); | |
1745 else | |
1746 tty->print(" "); | |
1747 tty->print(" %c ", starts_bundle); | |
1748 starts_bundle = ' '; | |
1749 tty->print("\t"); | |
1750 n->format(_regalloc, tty); | |
1751 tty->cr(); | |
1752 } | |
1753 | |
1754 // If we have an instruction with a delay slot, and have seen a delay, | |
1755 // then back up and print it | |
1756 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { | |
1757 assert(delay != NULL, "no unconditional delay instruction"); | |
1758 if (node_bundling(delay)->starts_bundle()) | |
1759 starts_bundle = '+'; | |
1760 if (pcs && n->_idx < pc_limit) | |
1761 tty->print("%3.3x", pcs[n->_idx]); | |
1762 else | |
1763 tty->print(" "); | |
1764 tty->print(" %c ", starts_bundle); | |
1765 starts_bundle = ' '; | |
1766 tty->print("\t"); | |
1767 delay->format(_regalloc, tty); | |
1768 tty->print_cr(""); | |
1769 delay = NULL; | |
1770 } | |
1771 | |
1772 // Dump the exception table as well | |
1773 if( n->is_Catch() && (Verbose || WizardMode) ) { | |
1774 // Print the exception table for this offset | |
1775 _handler_table.print_subtable_for(pc); | |
1776 } | |
1777 } | |
1778 | |
1779 if (pcs && n->_idx < pc_limit) | |
1780 tty->print_cr("%3.3x", pcs[n->_idx]); | |
1781 else | |
1782 tty->print_cr(""); | |
1783 | |
1784 assert(cut_short || delay == NULL, "no unconditional delay branch"); | |
1785 | |
1786 } // End of per-block dump | |
1787 tty->print_cr(""); | |
1788 | |
1789 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); | |
1790 } | |
1791 #endif | |
1792 | |
1793 //------------------------------Final_Reshape_Counts--------------------------- | |
1794 // This class defines counters to help identify when a method | |
1795 // may/must be executed using hardware with only 24-bit precision. | |
1796 struct Final_Reshape_Counts : public StackObj { | |
1797 int _call_count; // count non-inlined 'common' calls | |
1798 int _float_count; // count float ops requiring 24-bit precision | |
1799 int _double_count; // count double ops requiring more precision | |
1800 int _java_call_count; // count non-inlined 'java' calls | |
1801 VectorSet _visited; // Visitation flags | |
1802 Node_List _tests; // Set of IfNodes & PCTableNodes | |
1803 | |
1804 Final_Reshape_Counts() : | |
1805 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), | |
1806 _visited( Thread::current()->resource_area() ) { } | |
1807 | |
1808 void inc_call_count () { _call_count ++; } | |
1809 void inc_float_count () { _float_count ++; } | |
1810 void inc_double_count() { _double_count++; } | |
1811 void inc_java_call_count() { _java_call_count++; } | |
1812 | |
1813 int get_call_count () const { return _call_count ; } | |
1814 int get_float_count () const { return _float_count ; } | |
1815 int get_double_count() const { return _double_count; } | |
1816 int get_java_call_count() const { return _java_call_count; } | |
1817 }; | |
1818 | |
1819 static bool oop_offset_is_sane(const TypeInstPtr* tp) { | |
1820 ciInstanceKlass *k = tp->klass()->as_instance_klass(); | |
1821 // Make sure the offset goes inside the instance layout. | |
1822 return (uint)tp->offset() < (uint)(oopDesc::header_size() + k->nonstatic_field_size())*wordSize; | |
1823 // Note that OffsetBot and OffsetTop are very negative. | |
1824 } | |
1825 | |
1826 //------------------------------final_graph_reshaping_impl---------------------- | |
1827 // Implement items 1-5 from final_graph_reshaping below. | |
1828 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { | |
1829 | |
1830 uint nop = n->Opcode(); | |
1831 | |
1832 // Check for 2-input instruction with "last use" on right input. | |
1833 // Swap to left input. Implements item (2). | |
1834 if( n->req() == 3 && // two-input instruction | |
1835 n->in(1)->outcnt() > 1 && // left use is NOT a last use | |
1836 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop | |
1837 n->in(2)->outcnt() == 1 &&// right use IS a last use | |
1838 !n->in(2)->is_Con() ) { // right use is not a constant | |
1839 // Check for commutative opcode | |
1840 switch( nop ) { | |
1841 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: | |
1842 case Op_MaxI: case Op_MinI: | |
1843 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: | |
1844 case Op_AndL: case Op_XorL: case Op_OrL: | |
1845 case Op_AndI: case Op_XorI: case Op_OrI: { | |
1846 // Move "last use" input to left by swapping inputs | |
1847 n->swap_edges(1, 2); | |
1848 break; | |
1849 } | |
1850 default: | |
1851 break; | |
1852 } | |
1853 } | |
1854 | |
1855 // Count FPU ops and common calls, implements item (3) | |
1856 switch( nop ) { | |
1857 // Count all float operations that may use FPU | |
1858 case Op_AddF: | |
1859 case Op_SubF: | |
1860 case Op_MulF: | |
1861 case Op_DivF: | |
1862 case Op_NegF: | |
1863 case Op_ModF: | |
1864 case Op_ConvI2F: | |
1865 case Op_ConF: | |
1866 case Op_CmpF: | |
1867 case Op_CmpF3: | |
1868 // case Op_ConvL2F: // longs are split into 32-bit halves | |
1869 fpu.inc_float_count(); | |
1870 break; | |
1871 | |
1872 case Op_ConvF2D: | |
1873 case Op_ConvD2F: | |
1874 fpu.inc_float_count(); | |
1875 fpu.inc_double_count(); | |
1876 break; | |
1877 | |
1878 // Count all double operations that may use FPU | |
1879 case Op_AddD: | |
1880 case Op_SubD: | |
1881 case Op_MulD: | |
1882 case Op_DivD: | |
1883 case Op_NegD: | |
1884 case Op_ModD: | |
1885 case Op_ConvI2D: | |
1886 case Op_ConvD2I: | |
1887 // case Op_ConvL2D: // handled by leaf call | |
1888 // case Op_ConvD2L: // handled by leaf call | |
1889 case Op_ConD: | |
1890 case Op_CmpD: | |
1891 case Op_CmpD3: | |
1892 fpu.inc_double_count(); | |
1893 break; | |
1894 case Op_Opaque1: // Remove Opaque Nodes before matching | |
1895 case Op_Opaque2: // Remove Opaque Nodes before matching | |
1896 n->replace_by(n->in(1)); | |
1897 break; | |
1898 case Op_CallStaticJava: | |
1899 case Op_CallJava: | |
1900 case Op_CallDynamicJava: | |
1901 fpu.inc_java_call_count(); // Count java call site; | |
1902 case Op_CallRuntime: | |
1903 case Op_CallLeaf: | |
1904 case Op_CallLeafNoFP: { | |
1905 assert( n->is_Call(), "" ); | |
1906 CallNode *call = n->as_Call(); | |
1907 // Count call sites where the FP mode bit would have to be flipped. | |
1908 // Do not count uncommon runtime calls: | |
1909 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, | |
1910 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... | |
1911 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { | |
1912 fpu.inc_call_count(); // Count the call site | |
1913 } else { // See if uncommon argument is shared | |
1914 Node *n = call->in(TypeFunc::Parms); | |
1915 int nop = n->Opcode(); | |
1916 // Clone shared simple arguments to uncommon calls, item (1). | |
1917 if( n->outcnt() > 1 && | |
1918 !n->is_Proj() && | |
1919 nop != Op_CreateEx && | |
1920 nop != Op_CheckCastPP && | |
1921 !n->is_Mem() ) { | |
1922 Node *x = n->clone(); | |
1923 call->set_req( TypeFunc::Parms, x ); | |
1924 } | |
1925 } | |
1926 break; | |
1927 } | |
1928 | |
1929 case Op_StoreD: | |
1930 case Op_LoadD: | |
1931 case Op_LoadD_unaligned: | |
1932 fpu.inc_double_count(); | |
1933 goto handle_mem; | |
1934 case Op_StoreF: | |
1935 case Op_LoadF: | |
1936 fpu.inc_float_count(); | |
1937 goto handle_mem; | |
1938 | |
1939 case Op_StoreB: | |
1940 case Op_StoreC: | |
1941 case Op_StoreCM: | |
1942 case Op_StorePConditional: | |
1943 case Op_StoreI: | |
1944 case Op_StoreL: | |
1945 case Op_StoreLConditional: | |
1946 case Op_CompareAndSwapI: | |
1947 case Op_CompareAndSwapL: | |
1948 case Op_CompareAndSwapP: | |
1949 case Op_StoreP: | |
1950 case Op_LoadB: | |
1951 case Op_LoadC: | |
1952 case Op_LoadI: | |
1953 case Op_LoadKlass: | |
1954 case Op_LoadL: | |
1955 case Op_LoadL_unaligned: | |
1956 case Op_LoadPLocked: | |
1957 case Op_LoadLLocked: | |
1958 case Op_LoadP: | |
1959 case Op_LoadRange: | |
1960 case Op_LoadS: { | |
1961 handle_mem: | |
1962 #ifdef ASSERT | |
1963 if( VerifyOptoOopOffsets ) { | |
1964 assert( n->is_Mem(), "" ); | |
1965 MemNode *mem = (MemNode*)n; | |
1966 // Check to see if address types have grounded out somehow. | |
1967 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); | |
1968 assert( !tp || oop_offset_is_sane(tp), "" ); | |
1969 } | |
1970 #endif | |
1971 break; | |
1972 } | |
1973 case Op_If: | |
1974 case Op_CountedLoopEnd: | |
1975 fpu._tests.push(n); // Collect CFG split points | |
1976 break; | |
1977 | |
1978 case Op_AddP: { // Assert sane base pointers | |
1979 const Node *addp = n->in(AddPNode::Address); | |
1980 assert( !addp->is_AddP() || | |
1981 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation | |
1982 addp->in(AddPNode::Base) == n->in(AddPNode::Base), | |
1983 "Base pointers must match" ); | |
1984 break; | |
1985 } | |
1986 | |
1987 case Op_ModI: | |
1988 if (UseDivMod) { | |
1989 // Check if a%b and a/b both exist | |
1990 Node* d = n->find_similar(Op_DivI); | |
1991 if (d) { | |
1992 // Replace them with a fused divmod if supported | |
1993 Compile* C = Compile::current(); | |
1994 if (Matcher::has_match_rule(Op_DivModI)) { | |
1995 DivModINode* divmod = DivModINode::make(C, n); | |
1996 d->replace_by(divmod->div_proj()); | |
1997 n->replace_by(divmod->mod_proj()); | |
1998 } else { | |
1999 // replace a%b with a-((a/b)*b) | |
2000 Node* mult = new (C, 3) MulINode(d, d->in(2)); | |
2001 Node* sub = new (C, 3) SubINode(d->in(1), mult); | |
2002 n->replace_by( sub ); | |
2003 } | |
2004 } | |
2005 } | |
2006 break; | |
2007 | |
2008 case Op_ModL: | |
2009 if (UseDivMod) { | |
2010 // Check if a%b and a/b both exist | |
2011 Node* d = n->find_similar(Op_DivL); | |
2012 if (d) { | |
2013 // Replace them with a fused divmod if supported | |
2014 Compile* C = Compile::current(); | |
2015 if (Matcher::has_match_rule(Op_DivModL)) { | |
2016 DivModLNode* divmod = DivModLNode::make(C, n); | |
2017 d->replace_by(divmod->div_proj()); | |
2018 n->replace_by(divmod->mod_proj()); | |
2019 } else { | |
2020 // replace a%b with a-((a/b)*b) | |
2021 Node* mult = new (C, 3) MulLNode(d, d->in(2)); | |
2022 Node* sub = new (C, 3) SubLNode(d->in(1), mult); | |
2023 n->replace_by( sub ); | |
2024 } | |
2025 } | |
2026 } | |
2027 break; | |
2028 | |
2029 case Op_Load16B: | |
2030 case Op_Load8B: | |
2031 case Op_Load4B: | |
2032 case Op_Load8S: | |
2033 case Op_Load4S: | |
2034 case Op_Load2S: | |
2035 case Op_Load8C: | |
2036 case Op_Load4C: | |
2037 case Op_Load2C: | |
2038 case Op_Load4I: | |
2039 case Op_Load2I: | |
2040 case Op_Load2L: | |
2041 case Op_Load4F: | |
2042 case Op_Load2F: | |
2043 case Op_Load2D: | |
2044 case Op_Store16B: | |
2045 case Op_Store8B: | |
2046 case Op_Store4B: | |
2047 case Op_Store8C: | |
2048 case Op_Store4C: | |
2049 case Op_Store2C: | |
2050 case Op_Store4I: | |
2051 case Op_Store2I: | |
2052 case Op_Store2L: | |
2053 case Op_Store4F: | |
2054 case Op_Store2F: | |
2055 case Op_Store2D: | |
2056 break; | |
2057 | |
2058 case Op_PackB: | |
2059 case Op_PackS: | |
2060 case Op_PackC: | |
2061 case Op_PackI: | |
2062 case Op_PackF: | |
2063 case Op_PackL: | |
2064 case Op_PackD: | |
2065 if (n->req()-1 > 2) { | |
2066 // Replace many operand PackNodes with a binary tree for matching | |
2067 PackNode* p = (PackNode*) n; | |
2068 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); | |
2069 n->replace_by(btp); | |
2070 } | |
2071 break; | |
2072 default: | |
2073 assert( !n->is_Call(), "" ); | |
2074 assert( !n->is_Mem(), "" ); | |
2075 if( n->is_If() || n->is_PCTable() ) | |
2076 fpu._tests.push(n); // Collect CFG split points | |
2077 break; | |
2078 } | |
2079 } | |
2080 | |
2081 //------------------------------final_graph_reshaping_walk--------------------- | |
2082 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), | |
2083 // requires that the walk visits a node's inputs before visiting the node. | |
2084 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { | |
2085 fpu._visited.set(root->_idx); // first, mark node as visited | |
2086 uint cnt = root->req(); | |
2087 Node *n = root; | |
2088 uint i = 0; | |
2089 while (true) { | |
2090 if (i < cnt) { | |
2091 // Place all non-visited non-null inputs onto stack | |
2092 Node* m = n->in(i); | |
2093 ++i; | |
2094 if (m != NULL && !fpu._visited.test_set(m->_idx)) { | |
2095 cnt = m->req(); | |
2096 nstack.push(n, i); // put on stack parent and next input's index | |
2097 n = m; | |
2098 i = 0; | |
2099 } | |
2100 } else { | |
2101 // Now do post-visit work | |
2102 final_graph_reshaping_impl( n, fpu ); | |
2103 if (nstack.is_empty()) | |
2104 break; // finished | |
2105 n = nstack.node(); // Get node from stack | |
2106 cnt = n->req(); | |
2107 i = nstack.index(); | |
2108 nstack.pop(); // Shift to the next node on stack | |
2109 } | |
2110 } | |
2111 } | |
2112 | |
2113 //------------------------------final_graph_reshaping-------------------------- | |
2114 // Final Graph Reshaping. | |
2115 // | |
2116 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late | |
2117 // and not commoned up and forced early. Must come after regular | |
2118 // optimizations to avoid GVN undoing the cloning. Clone constant | |
2119 // inputs to Loop Phis; these will be split by the allocator anyways. | |
2120 // Remove Opaque nodes. | |
2121 // (2) Move last-uses by commutative operations to the left input to encourage | |
2122 // Intel update-in-place two-address operations and better register usage | |
2123 // on RISCs. Must come after regular optimizations to avoid GVN Ideal | |
2124 // calls canonicalizing them back. | |
2125 // (3) Count the number of double-precision FP ops, single-precision FP ops | |
2126 // and call sites. On Intel, we can get correct rounding either by | |
2127 // forcing singles to memory (requires extra stores and loads after each | |
2128 // FP bytecode) or we can set a rounding mode bit (requires setting and | |
2129 // clearing the mode bit around call sites). The mode bit is only used | |
2130 // if the relative frequency of single FP ops to calls is low enough. | |
2131 // This is a key transform for SPEC mpeg_audio. | |
2132 // (4) Detect infinite loops; blobs of code reachable from above but not | |
2133 // below. Several of the Code_Gen algorithms fail on such code shapes, | |
2134 // so we simply bail out. Happens a lot in ZKM.jar, but also happens | |
2135 // from time to time in other codes (such as -Xcomp finalizer loops, etc). | |
2136 // Detection is by looking for IfNodes where only 1 projection is | |
2137 // reachable from below or CatchNodes missing some targets. | |
2138 // (5) Assert for insane oop offsets in debug mode. | |
2139 | |
2140 bool Compile::final_graph_reshaping() { | |
2141 // an infinite loop may have been eliminated by the optimizer, | |
2142 // in which case the graph will be empty. | |
2143 if (root()->req() == 1) { | |
2144 record_method_not_compilable("trivial infinite loop"); | |
2145 return true; | |
2146 } | |
2147 | |
2148 Final_Reshape_Counts fpu; | |
2149 | |
2150 // Visit everybody reachable! | |
2151 // Allocate stack of size C->unique()/2 to avoid frequent realloc | |
2152 Node_Stack nstack(unique() >> 1); | |
2153 final_graph_reshaping_walk(nstack, root(), fpu); | |
2154 | |
2155 // Check for unreachable (from below) code (i.e., infinite loops). | |
2156 for( uint i = 0; i < fpu._tests.size(); i++ ) { | |
2157 Node *n = fpu._tests[i]; | |
2158 assert( n->is_PCTable() || n->is_If(), "either PCTables or IfNodes" ); | |
2159 // Get number of CFG targets; 2 for IfNodes or _size for PCTables. | |
2160 // Note that PCTables include exception targets after calls. | |
2161 uint expected_kids = n->is_PCTable() ? n->as_PCTable()->_size : 2; | |
2162 if (n->outcnt() != expected_kids) { | |
2163 // Check for a few special cases. Rethrow Nodes never take the | |
2164 // 'fall-thru' path, so expected kids is 1 less. | |
2165 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { | |
2166 if (n->in(0)->in(0)->is_Call()) { | |
2167 CallNode *call = n->in(0)->in(0)->as_Call(); | |
2168 if (call->entry_point() == OptoRuntime::rethrow_stub()) { | |
2169 expected_kids--; // Rethrow always has 1 less kid | |
2170 } else if (call->req() > TypeFunc::Parms && | |
2171 call->is_CallDynamicJava()) { | |
2172 // Check for null receiver. In such case, the optimizer has | |
2173 // detected that the virtual call will always result in a null | |
2174 // pointer exception. The fall-through projection of this CatchNode | |
2175 // will not be populated. | |
2176 Node *arg0 = call->in(TypeFunc::Parms); | |
2177 if (arg0->is_Type() && | |
2178 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { | |
2179 expected_kids--; | |
2180 } | |
2181 } else if (call->entry_point() == OptoRuntime::new_array_Java() && | |
2182 call->req() > TypeFunc::Parms+1 && | |
2183 call->is_CallStaticJava()) { | |
2184 // Check for negative array length. In such case, the optimizer has | |
2185 // detected that the allocation attempt will always result in an | |
2186 // exception. There is no fall-through projection of this CatchNode . | |
2187 Node *arg1 = call->in(TypeFunc::Parms+1); | |
2188 if (arg1->is_Type() && | |
2189 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { | |
2190 expected_kids--; | |
2191 } | |
2192 } | |
2193 } | |
2194 } | |
2195 // Recheck with a better notion of 'expected_kids' | |
2196 if (n->outcnt() != expected_kids) { | |
2197 record_method_not_compilable("malformed control flow"); | |
2198 return true; // Not all targets reachable! | |
2199 } | |
2200 } | |
2201 // Check that I actually visited all kids. Unreached kids | |
2202 // must be infinite loops. | |
2203 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) | |
2204 if (!fpu._visited.test(n->fast_out(j)->_idx)) { | |
2205 record_method_not_compilable("infinite loop"); | |
2206 return true; // Found unvisited kid; must be unreach | |
2207 } | |
2208 } | |
2209 | |
2210 // If original bytecodes contained a mixture of floats and doubles | |
2211 // check if the optimizer has made it homogenous, item (3). | |
2212 if( Use24BitFPMode && Use24BitFP && | |
2213 fpu.get_float_count() > 32 && | |
2214 fpu.get_double_count() == 0 && | |
2215 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { | |
2216 set_24_bit_selection_and_mode( false, true ); | |
2217 } | |
2218 | |
2219 set_has_java_calls(fpu.get_java_call_count() > 0); | |
2220 | |
2221 // No infinite loops, no reason to bail out. | |
2222 return false; | |
2223 } | |
2224 | |
2225 //-----------------------------too_many_traps---------------------------------- | |
2226 // Report if there are too many traps at the current method and bci. | |
2227 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. | |
2228 bool Compile::too_many_traps(ciMethod* method, | |
2229 int bci, | |
2230 Deoptimization::DeoptReason reason) { | |
2231 ciMethodData* md = method->method_data(); | |
2232 if (md->is_empty()) { | |
2233 // Assume the trap has not occurred, or that it occurred only | |
2234 // because of a transient condition during start-up in the interpreter. | |
2235 return false; | |
2236 } | |
2237 if (md->has_trap_at(bci, reason) != 0) { | |
2238 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. | |
2239 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
2240 // assume the worst. | |
2241 if (log()) | |
2242 log()->elem("observe trap='%s' count='%d'", | |
2243 Deoptimization::trap_reason_name(reason), | |
2244 md->trap_count(reason)); | |
2245 return true; | |
2246 } else { | |
2247 // Ignore method/bci and see if there have been too many globally. | |
2248 return too_many_traps(reason, md); | |
2249 } | |
2250 } | |
2251 | |
2252 // Less-accurate variant which does not require a method and bci. | |
2253 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, | |
2254 ciMethodData* logmd) { | |
2255 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { | |
2256 // Too many traps globally. | |
2257 // Note that we use cumulative trap_count, not just md->trap_count. | |
2258 if (log()) { | |
2259 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); | |
2260 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", | |
2261 Deoptimization::trap_reason_name(reason), | |
2262 mcount, trap_count(reason)); | |
2263 } | |
2264 return true; | |
2265 } else { | |
2266 // The coast is clear. | |
2267 return false; | |
2268 } | |
2269 } | |
2270 | |
2271 //--------------------------too_many_recompiles-------------------------------- | |
2272 // Report if there are too many recompiles at the current method and bci. | |
2273 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. | |
2274 // Is not eager to return true, since this will cause the compiler to use | |
2275 // Action_none for a trap point, to avoid too many recompilations. | |
2276 bool Compile::too_many_recompiles(ciMethod* method, | |
2277 int bci, | |
2278 Deoptimization::DeoptReason reason) { | |
2279 ciMethodData* md = method->method_data(); | |
2280 if (md->is_empty()) { | |
2281 // Assume the trap has not occurred, or that it occurred only | |
2282 // because of a transient condition during start-up in the interpreter. | |
2283 return false; | |
2284 } | |
2285 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. | |
2286 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; | |
2287 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero | |
2288 Deoptimization::DeoptReason per_bc_reason | |
2289 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); | |
2290 if ((per_bc_reason == Deoptimization::Reason_none | |
2291 || md->has_trap_at(bci, reason) != 0) | |
2292 // The trap frequency measure we care about is the recompile count: | |
2293 && md->trap_recompiled_at(bci) | |
2294 && md->overflow_recompile_count() >= bc_cutoff) { | |
2295 // Do not emit a trap here if it has already caused recompilations. | |
2296 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
2297 // assume the worst. | |
2298 if (log()) | |
2299 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", | |
2300 Deoptimization::trap_reason_name(reason), | |
2301 md->trap_count(reason), | |
2302 md->overflow_recompile_count()); | |
2303 return true; | |
2304 } else if (trap_count(reason) != 0 | |
2305 && decompile_count() >= m_cutoff) { | |
2306 // Too many recompiles globally, and we have seen this sort of trap. | |
2307 // Use cumulative decompile_count, not just md->decompile_count. | |
2308 if (log()) | |
2309 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", | |
2310 Deoptimization::trap_reason_name(reason), | |
2311 md->trap_count(reason), trap_count(reason), | |
2312 md->decompile_count(), decompile_count()); | |
2313 return true; | |
2314 } else { | |
2315 // The coast is clear. | |
2316 return false; | |
2317 } | |
2318 } | |
2319 | |
2320 | |
2321 #ifndef PRODUCT | |
2322 //------------------------------verify_graph_edges--------------------------- | |
2323 // Walk the Graph and verify that there is a one-to-one correspondence | |
2324 // between Use-Def edges and Def-Use edges in the graph. | |
2325 void Compile::verify_graph_edges(bool no_dead_code) { | |
2326 if (VerifyGraphEdges) { | |
2327 ResourceArea *area = Thread::current()->resource_area(); | |
2328 Unique_Node_List visited(area); | |
2329 // Call recursive graph walk to check edges | |
2330 _root->verify_edges(visited); | |
2331 if (no_dead_code) { | |
2332 // Now make sure that no visited node is used by an unvisited node. | |
2333 bool dead_nodes = 0; | |
2334 Unique_Node_List checked(area); | |
2335 while (visited.size() > 0) { | |
2336 Node* n = visited.pop(); | |
2337 checked.push(n); | |
2338 for (uint i = 0; i < n->outcnt(); i++) { | |
2339 Node* use = n->raw_out(i); | |
2340 if (checked.member(use)) continue; // already checked | |
2341 if (visited.member(use)) continue; // already in the graph | |
2342 if (use->is_Con()) continue; // a dead ConNode is OK | |
2343 // At this point, we have found a dead node which is DU-reachable. | |
2344 if (dead_nodes++ == 0) | |
2345 tty->print_cr("*** Dead nodes reachable via DU edges:"); | |
2346 use->dump(2); | |
2347 tty->print_cr("---"); | |
2348 checked.push(use); // No repeats; pretend it is now checked. | |
2349 } | |
2350 } | |
2351 assert(dead_nodes == 0, "using nodes must be reachable from root"); | |
2352 } | |
2353 } | |
2354 } | |
2355 #endif | |
2356 | |
2357 // The Compile object keeps track of failure reasons separately from the ciEnv. | |
2358 // This is required because there is not quite a 1-1 relation between the | |
2359 // ciEnv and its compilation task and the Compile object. Note that one | |
2360 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides | |
2361 // to backtrack and retry without subsuming loads. Other than this backtracking | |
2362 // behavior, the Compile's failure reason is quietly copied up to the ciEnv | |
2363 // by the logic in C2Compiler. | |
2364 void Compile::record_failure(const char* reason) { | |
2365 if (log() != NULL) { | |
2366 log()->elem("failure reason='%s' phase='compile'", reason); | |
2367 } | |
2368 if (_failure_reason == NULL) { | |
2369 // Record the first failure reason. | |
2370 _failure_reason = reason; | |
2371 } | |
2372 _root = NULL; // flush the graph, too | |
2373 } | |
2374 | |
2375 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) | |
2376 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) | |
2377 { | |
2378 if (dolog) { | |
2379 C = Compile::current(); | |
2380 _log = C->log(); | |
2381 } else { | |
2382 C = NULL; | |
2383 _log = NULL; | |
2384 } | |
2385 if (_log != NULL) { | |
2386 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); | |
2387 _log->stamp(); | |
2388 _log->end_head(); | |
2389 } | |
2390 } | |
2391 | |
2392 Compile::TracePhase::~TracePhase() { | |
2393 if (_log != NULL) { | |
2394 _log->done("phase nodes='%d'", C->unique()); | |
2395 } | |
2396 } |