Mercurial > hg > graal-compiler
annotate src/share/vm/opto/compile.cpp @ 584:49a36a80b0c7
6802499: EA: assert(false,"unknown node on this path")
Summary: Add missing checks for SCMemProj node in Escape analysis code.
Reviewed-by: never
author | kvn |
---|---|
date | Thu, 19 Feb 2009 17:38:53 -0800 |
parents | 3b5ac9e7e6ea |
children | 0fbdb4381b99 98cb887364d3 |
rev | line source |
---|---|
0 | 1 /* |
196 | 2 * Copyright 1997-2008 Sun Microsystems, Inc. All Rights Reserved. |
0 | 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 _compile->end_method(); | |
317 if (_compile->scratch_buffer_blob() != NULL) | |
318 BufferBlob::free(_compile->scratch_buffer_blob()); | |
319 _compile->env()->set_compiler_data(NULL); | |
320 } | |
321 | |
322 | |
323 //----------------------------print_compile_messages--------------------------- | |
324 void Compile::print_compile_messages() { | |
325 #ifndef PRODUCT | |
326 // Check if recompiling | |
327 if (_subsume_loads == false && PrintOpto) { | |
328 // Recompiling without allowing machine instructions to subsume loads | |
329 tty->print_cr("*********************************************************"); | |
330 tty->print_cr("** Bailout: Recompile without subsuming loads **"); | |
331 tty->print_cr("*********************************************************"); | |
332 } | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
333 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
334 // Recompiling without escape analysis |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
335 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
336 tty->print_cr("** Bailout: Recompile without escape analysis **"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
337 tty->print_cr("*********************************************************"); |
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
338 } |
0 | 339 if (env()->break_at_compile()) { |
340 // Open the debugger when compiing this method. | |
341 tty->print("### Breaking when compiling: "); | |
342 method()->print_short_name(); | |
343 tty->cr(); | |
344 BREAKPOINT; | |
345 } | |
346 | |
347 if( PrintOpto ) { | |
348 if (is_osr_compilation()) { | |
349 tty->print("[OSR]%3d", _compile_id); | |
350 } else { | |
351 tty->print("%3d", _compile_id); | |
352 } | |
353 } | |
354 #endif | |
355 } | |
356 | |
357 | |
358 void Compile::init_scratch_buffer_blob() { | |
359 if( scratch_buffer_blob() != NULL ) return; | |
360 | |
361 // Construct a temporary CodeBuffer to have it construct a BufferBlob | |
362 // Cache this BufferBlob for this compile. | |
363 ResourceMark rm; | |
364 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); | |
365 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); | |
366 // Record the buffer blob for next time. | |
367 set_scratch_buffer_blob(blob); | |
163 | 368 // Have we run out of code space? |
369 if (scratch_buffer_blob() == NULL) { | |
370 // Let CompilerBroker disable further compilations. | |
371 record_failure("Not enough space for scratch buffer in CodeCache"); | |
372 return; | |
373 } | |
0 | 374 |
375 // Initialize the relocation buffers | |
376 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; | |
377 set_scratch_locs_memory(locs_buf); | |
378 } | |
379 | |
380 | |
381 //-----------------------scratch_emit_size------------------------------------- | |
382 // Helper function that computes size by emitting code | |
383 uint Compile::scratch_emit_size(const Node* n) { | |
384 // Emit into a trash buffer and count bytes emitted. | |
385 // This is a pretty expensive way to compute a size, | |
386 // but it works well enough if seldom used. | |
387 // All common fixed-size instructions are given a size | |
388 // method by the AD file. | |
389 // Note that the scratch buffer blob and locs memory are | |
390 // allocated at the beginning of the compile task, and | |
391 // may be shared by several calls to scratch_emit_size. | |
392 // The allocation of the scratch buffer blob is particularly | |
393 // expensive, since it has to grab the code cache lock. | |
394 BufferBlob* blob = this->scratch_buffer_blob(); | |
395 assert(blob != NULL, "Initialize BufferBlob at start"); | |
396 assert(blob->size() > MAX_inst_size, "sanity"); | |
397 relocInfo* locs_buf = scratch_locs_memory(); | |
398 address blob_begin = blob->instructions_begin(); | |
399 address blob_end = (address)locs_buf; | |
400 assert(blob->instructions_contains(blob_end), "sanity"); | |
401 CodeBuffer buf(blob_begin, blob_end - blob_begin); | |
402 buf.initialize_consts_size(MAX_const_size); | |
403 buf.initialize_stubs_size(MAX_stubs_size); | |
404 assert(locs_buf != NULL, "sanity"); | |
405 int lsize = MAX_locs_size / 2; | |
406 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); | |
407 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); | |
408 n->emit(buf, this->regalloc()); | |
409 return buf.code_size(); | |
410 } | |
411 | |
412 | |
413 // ============================================================================ | |
414 //------------------------------Compile standard------------------------------- | |
415 debug_only( int Compile::_debug_idx = 100000; ) | |
416 | |
417 // Compile a method. entry_bci is -1 for normal compilations and indicates | |
418 // the continuation bci for on stack replacement. | |
419 | |
420 | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
421 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) |
0 | 422 : Phase(Compiler), |
423 _env(ci_env), | |
424 _log(ci_env->log()), | |
425 _compile_id(ci_env->compile_id()), | |
426 _save_argument_registers(false), | |
427 _stub_name(NULL), | |
428 _stub_function(NULL), | |
429 _stub_entry_point(NULL), | |
430 _method(target), | |
431 _entry_bci(osr_bci), | |
432 _initial_gvn(NULL), | |
433 _for_igvn(NULL), | |
434 _warm_calls(NULL), | |
435 _subsume_loads(subsume_loads), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
436 _do_escape_analysis(do_escape_analysis), |
0 | 437 _failure_reason(NULL), |
438 _code_buffer("Compile::Fill_buffer"), | |
439 _orig_pc_slot(0), | |
440 _orig_pc_slot_offset_in_bytes(0), | |
441 _node_bundling_limit(0), | |
442 _node_bundling_base(NULL), | |
443 #ifndef PRODUCT | |
444 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), | |
445 _printer(IdealGraphPrinter::printer()), | |
446 #endif | |
447 _congraph(NULL) { | |
448 C = this; | |
449 | |
450 CompileWrapper cw(this); | |
451 #ifndef PRODUCT | |
452 if (TimeCompiler2) { | |
453 tty->print(" "); | |
454 target->holder()->name()->print(); | |
455 tty->print("."); | |
456 target->print_short_name(); | |
457 tty->print(" "); | |
458 } | |
459 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); | |
460 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); | |
100
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
461 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
462 if (!print_opto_assembly) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
463 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
464 if (print_assembly && !Disassembler::can_decode()) { |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
465 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
466 print_opto_assembly = true; |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
467 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
468 } |
c7c777385a15
6667042: PrintAssembly option does not work without special plugin
jrose
parents:
65
diff
changeset
|
469 set_print_assembly(print_opto_assembly); |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
333
diff
changeset
|
470 set_parsed_irreducible_loop(false); |
0 | 471 #endif |
472 | |
473 if (ProfileTraps) { | |
474 // Make sure the method being compiled gets its own MDO, | |
475 // so we can at least track the decompile_count(). | |
476 method()->build_method_data(); | |
477 } | |
478 | |
479 Init(::AliasLevel); | |
480 | |
481 | |
482 print_compile_messages(); | |
483 | |
484 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) | |
485 _ilt = InlineTree::build_inline_tree_root(); | |
486 else | |
487 _ilt = NULL; | |
488 | |
489 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice | |
490 assert(num_alias_types() >= AliasIdxRaw, ""); | |
491 | |
492 #define MINIMUM_NODE_HASH 1023 | |
493 // Node list that Iterative GVN will start with | |
494 Unique_Node_List for_igvn(comp_arena()); | |
495 set_for_igvn(&for_igvn); | |
496 | |
497 // GVN that will be run immediately on new nodes | |
498 uint estimated_size = method()->code_size()*4+64; | |
499 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); | |
500 PhaseGVN gvn(node_arena(), estimated_size); | |
501 set_initial_gvn(&gvn); | |
502 | |
503 { // Scope for timing the parser | |
504 TracePhase t3("parse", &_t_parser, true); | |
505 | |
506 // Put top into the hash table ASAP. | |
507 initial_gvn()->transform_no_reclaim(top()); | |
508 | |
509 // Set up tf(), start(), and find a CallGenerator. | |
510 CallGenerator* cg; | |
511 if (is_osr_compilation()) { | |
512 const TypeTuple *domain = StartOSRNode::osr_domain(); | |
513 const TypeTuple *range = TypeTuple::make_range(method()->signature()); | |
514 init_tf(TypeFunc::make(domain, range)); | |
515 StartNode* s = new (this, 2) StartOSRNode(root(), domain); | |
516 initial_gvn()->set_type_bottom(s); | |
517 init_start(s); | |
518 cg = CallGenerator::for_osr(method(), entry_bci()); | |
519 } else { | |
520 // Normal case. | |
521 init_tf(TypeFunc::make(method())); | |
522 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); | |
523 initial_gvn()->set_type_bottom(s); | |
524 init_start(s); | |
525 float past_uses = method()->interpreter_invocation_count(); | |
526 float expected_uses = past_uses; | |
527 cg = CallGenerator::for_inline(method(), expected_uses); | |
528 } | |
529 if (failing()) return; | |
530 if (cg == NULL) { | |
531 record_method_not_compilable_all_tiers("cannot parse method"); | |
532 return; | |
533 } | |
534 JVMState* jvms = build_start_state(start(), tf()); | |
535 if ((jvms = cg->generate(jvms)) == NULL) { | |
536 record_method_not_compilable("method parse failed"); | |
537 return; | |
538 } | |
539 GraphKit kit(jvms); | |
540 | |
541 if (!kit.stopped()) { | |
542 // Accept return values, and transfer control we know not where. | |
543 // This is done by a special, unique ReturnNode bound to root. | |
544 return_values(kit.jvms()); | |
545 } | |
546 | |
547 if (kit.has_exceptions()) { | |
548 // Any exceptions that escape from this call must be rethrown | |
549 // to whatever caller is dynamically above us on the stack. | |
550 // This is done by a special, unique RethrowNode bound to root. | |
551 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); | |
552 } | |
553 | |
417 | 554 print_method("Before RemoveUseless", 3); |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
333
diff
changeset
|
555 |
0 | 556 // Remove clutter produced by parsing. |
557 if (!failing()) { | |
558 ResourceMark rm; | |
559 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); | |
560 } | |
561 } | |
562 | |
563 // Note: Large methods are capped off in do_one_bytecode(). | |
564 if (failing()) return; | |
565 | |
566 // After parsing, node notes are no longer automagic. | |
567 // They must be propagated by register_new_node_with_optimizer(), | |
568 // clone(), or the like. | |
569 set_default_node_notes(NULL); | |
570 | |
571 for (;;) { | |
572 int successes = Inline_Warm(); | |
573 if (failing()) return; | |
574 if (successes == 0) break; | |
575 } | |
576 | |
577 // Drain the list. | |
578 Finish_Warm(); | |
579 #ifndef PRODUCT | |
580 if (_printer) { | |
581 _printer->print_inlining(this); | |
582 } | |
583 #endif | |
584 | |
585 if (failing()) return; | |
586 NOT_PRODUCT( verify_graph_edges(); ) | |
587 | |
588 // Perform escape analysis | |
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
589 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
590 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
591 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
592 PhaseGVN* igvn = initial_gvn(); |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
593 Node* oop_null = igvn->zerocon(T_OBJECT); |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
594 Node* noop_null = igvn->zerocon(T_NARROWOOP); |
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
595 |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
596 _congraph = new(comp_arena()) ConnectionGraph(this); |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
597 bool has_non_escaping_obj = _congraph->compute_escape(); |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
598 |
0 | 599 #ifndef PRODUCT |
600 if (PrintEscapeAnalysis) { | |
601 _congraph->dump(); | |
602 } | |
603 #endif | |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
604 // Cleanup. |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
605 if (oop_null->outcnt() == 0) |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
606 igvn->hash_delete(oop_null); |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
607 if (noop_null->outcnt() == 0) |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
608 igvn->hash_delete(noop_null); |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
609 |
244
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
610 if (!has_non_escaping_obj) { |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
611 _congraph = NULL; |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
612 } |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
613 |
524eca34ea76
6684714: Optimize EA Connection Graph build performance
kvn
parents:
223
diff
changeset
|
614 if (failing()) return; |
0 | 615 } |
616 // Now optimize | |
617 Optimize(); | |
618 if (failing()) return; | |
619 NOT_PRODUCT( verify_graph_edges(); ) | |
620 | |
621 #ifndef PRODUCT | |
622 if (PrintIdeal) { | |
623 ttyLocker ttyl; // keep the following output all in one block | |
624 // This output goes directly to the tty, not the compiler log. | |
625 // To enable tools to match it up with the compilation activity, | |
626 // be sure to tag this tty output with the compile ID. | |
627 if (xtty != NULL) { | |
628 xtty->head("ideal compile_id='%d'%s", compile_id(), | |
629 is_osr_compilation() ? " compile_kind='osr'" : | |
630 ""); | |
631 } | |
632 root()->dump(9999); | |
633 if (xtty != NULL) { | |
634 xtty->tail("ideal"); | |
635 } | |
636 } | |
637 #endif | |
638 | |
639 // Now that we know the size of all the monitors we can add a fixed slot | |
640 // for the original deopt pc. | |
641 | |
642 _orig_pc_slot = fixed_slots(); | |
643 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); | |
644 set_fixed_slots(next_slot); | |
645 | |
646 // Now generate code | |
647 Code_Gen(); | |
648 if (failing()) return; | |
649 | |
650 // Check if we want to skip execution of all compiled code. | |
651 { | |
652 #ifndef PRODUCT | |
653 if (OptoNoExecute) { | |
654 record_method_not_compilable("+OptoNoExecute"); // Flag as failed | |
655 return; | |
656 } | |
657 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); | |
658 #endif | |
659 | |
660 if (is_osr_compilation()) { | |
661 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); | |
662 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); | |
663 } else { | |
664 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); | |
665 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); | |
666 } | |
667 | |
668 env()->register_method(_method, _entry_bci, | |
669 &_code_offsets, | |
670 _orig_pc_slot_offset_in_bytes, | |
671 code_buffer(), | |
672 frame_size_in_words(), _oop_map_set, | |
673 &_handler_table, &_inc_table, | |
674 compiler, | |
675 env()->comp_level(), | |
676 true, /*has_debug_info*/ | |
677 has_unsafe_access() | |
678 ); | |
679 } | |
680 } | |
681 | |
682 //------------------------------Compile---------------------------------------- | |
683 // Compile a runtime stub | |
684 Compile::Compile( ciEnv* ci_env, | |
685 TypeFunc_generator generator, | |
686 address stub_function, | |
687 const char *stub_name, | |
688 int is_fancy_jump, | |
689 bool pass_tls, | |
690 bool save_arg_registers, | |
691 bool return_pc ) | |
692 : Phase(Compiler), | |
693 _env(ci_env), | |
694 _log(ci_env->log()), | |
695 _compile_id(-1), | |
696 _save_argument_registers(save_arg_registers), | |
697 _method(NULL), | |
698 _stub_name(stub_name), | |
699 _stub_function(stub_function), | |
700 _stub_entry_point(NULL), | |
701 _entry_bci(InvocationEntryBci), | |
702 _initial_gvn(NULL), | |
703 _for_igvn(NULL), | |
704 _warm_calls(NULL), | |
705 _orig_pc_slot(0), | |
706 _orig_pc_slot_offset_in_bytes(0), | |
707 _subsume_loads(true), | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
708 _do_escape_analysis(false), |
0 | 709 _failure_reason(NULL), |
710 _code_buffer("Compile::Fill_buffer"), | |
711 _node_bundling_limit(0), | |
712 _node_bundling_base(NULL), | |
713 #ifndef PRODUCT | |
714 _trace_opto_output(TraceOptoOutput), | |
715 _printer(NULL), | |
716 #endif | |
717 _congraph(NULL) { | |
718 C = this; | |
719 | |
720 #ifndef PRODUCT | |
721 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); | |
722 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); | |
723 set_print_assembly(PrintFrameConverterAssembly); | |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
333
diff
changeset
|
724 set_parsed_irreducible_loop(false); |
0 | 725 #endif |
726 CompileWrapper cw(this); | |
727 Init(/*AliasLevel=*/ 0); | |
728 init_tf((*generator)()); | |
729 | |
730 { | |
731 // The following is a dummy for the sake of GraphKit::gen_stub | |
732 Unique_Node_List for_igvn(comp_arena()); | |
733 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this | |
734 PhaseGVN gvn(Thread::current()->resource_area(),255); | |
735 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively | |
736 gvn.transform_no_reclaim(top()); | |
737 | |
738 GraphKit kit; | |
739 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); | |
740 } | |
741 | |
742 NOT_PRODUCT( verify_graph_edges(); ) | |
743 Code_Gen(); | |
744 if (failing()) return; | |
745 | |
746 | |
747 // Entry point will be accessed using compile->stub_entry_point(); | |
748 if (code_buffer() == NULL) { | |
749 Matcher::soft_match_failure(); | |
750 } else { | |
751 if (PrintAssembly && (WizardMode || Verbose)) | |
752 tty->print_cr("### Stub::%s", stub_name); | |
753 | |
754 if (!failing()) { | |
755 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); | |
756 | |
757 // Make the NMethod | |
758 // For now we mark the frame as never safe for profile stackwalking | |
759 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, | |
760 code_buffer(), | |
761 CodeOffsets::frame_never_safe, | |
762 // _code_offsets.value(CodeOffsets::Frame_Complete), | |
763 frame_size_in_words(), | |
764 _oop_map_set, | |
765 save_arg_registers); | |
766 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); | |
767 | |
768 _stub_entry_point = rs->entry_point(); | |
769 } | |
770 } | |
771 } | |
772 | |
773 #ifndef PRODUCT | |
774 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { | |
775 if(PrintOpto && Verbose) { | |
776 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); | |
777 } | |
778 } | |
779 #endif | |
780 | |
781 void Compile::print_codes() { | |
782 } | |
783 | |
784 //------------------------------Init------------------------------------------- | |
785 // Prepare for a single compilation | |
786 void Compile::Init(int aliaslevel) { | |
787 _unique = 0; | |
788 _regalloc = NULL; | |
789 | |
790 _tf = NULL; // filled in later | |
791 _top = NULL; // cached later | |
792 _matcher = NULL; // filled in later | |
793 _cfg = NULL; // filled in later | |
794 | |
795 set_24_bit_selection_and_mode(Use24BitFP, false); | |
796 | |
797 _node_note_array = NULL; | |
798 _default_node_notes = NULL; | |
799 | |
800 _immutable_memory = NULL; // filled in at first inquiry | |
801 | |
802 // Globally visible Nodes | |
803 // First set TOP to NULL to give safe behavior during creation of RootNode | |
804 set_cached_top_node(NULL); | |
805 set_root(new (this, 3) RootNode()); | |
806 // Now that you have a Root to point to, create the real TOP | |
807 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); | |
808 set_recent_alloc(NULL, NULL); | |
809 | |
810 // Create Debug Information Recorder to record scopes, oopmaps, etc. | |
811 env()->set_oop_recorder(new OopRecorder(comp_arena())); | |
812 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); | |
813 env()->set_dependencies(new Dependencies(env())); | |
814 | |
815 _fixed_slots = 0; | |
816 set_has_split_ifs(false); | |
817 set_has_loops(has_method() && method()->has_loops()); // first approximation | |
818 _deopt_happens = true; // start out assuming the worst | |
819 _trap_can_recompile = false; // no traps emitted yet | |
820 _major_progress = true; // start out assuming good things will happen | |
821 set_has_unsafe_access(false); | |
822 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); | |
823 set_decompile_count(0); | |
824 | |
418 | 825 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); |
0 | 826 // Compilation level related initialization |
827 if (env()->comp_level() == CompLevel_fast_compile) { | |
828 set_num_loop_opts(Tier1LoopOptsCount); | |
829 set_do_inlining(Tier1Inline != 0); | |
830 set_max_inline_size(Tier1MaxInlineSize); | |
831 set_freq_inline_size(Tier1FreqInlineSize); | |
832 set_do_scheduling(false); | |
833 set_do_count_invocations(Tier1CountInvocations); | |
834 set_do_method_data_update(Tier1UpdateMethodData); | |
835 } else { | |
836 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); | |
837 set_num_loop_opts(LoopOptsCount); | |
838 set_do_inlining(Inline); | |
839 set_max_inline_size(MaxInlineSize); | |
840 set_freq_inline_size(FreqInlineSize); | |
841 set_do_scheduling(OptoScheduling); | |
842 set_do_count_invocations(false); | |
843 set_do_method_data_update(false); | |
844 } | |
845 | |
846 if (debug_info()->recording_non_safepoints()) { | |
847 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> | |
848 (comp_arena(), 8, 0, NULL)); | |
849 set_default_node_notes(Node_Notes::make(this)); | |
850 } | |
851 | |
852 // // -- Initialize types before each compile -- | |
853 // // Update cached type information | |
854 // if( _method && _method->constants() ) | |
855 // Type::update_loaded_types(_method, _method->constants()); | |
856 | |
857 // Init alias_type map. | |
38
b789bcaf2dd9
6667610: (Escape Analysis) retry compilation without EA if it fails
kvn
parents:
0
diff
changeset
|
858 if (!_do_escape_analysis && aliaslevel == 3) |
0 | 859 aliaslevel = 2; // No unique types without escape analysis |
860 _AliasLevel = aliaslevel; | |
861 const int grow_ats = 16; | |
862 _max_alias_types = grow_ats; | |
863 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); | |
864 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); | |
865 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); | |
866 { | |
867 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; | |
868 } | |
869 // Initialize the first few types. | |
870 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); | |
871 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); | |
872 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); | |
873 _num_alias_types = AliasIdxRaw+1; | |
874 // Zero out the alias type cache. | |
875 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); | |
876 // A NULL adr_type hits in the cache right away. Preload the right answer. | |
877 probe_alias_cache(NULL)->_index = AliasIdxTop; | |
878 | |
879 _intrinsics = NULL; | |
880 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); | |
881 register_library_intrinsics(); | |
882 } | |
883 | |
884 //---------------------------init_start---------------------------------------- | |
885 // Install the StartNode on this compile object. | |
886 void Compile::init_start(StartNode* s) { | |
887 if (failing()) | |
888 return; // already failing | |
889 assert(s == start(), ""); | |
890 } | |
891 | |
892 StartNode* Compile::start() const { | |
893 assert(!failing(), ""); | |
894 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { | |
895 Node* start = root()->fast_out(i); | |
896 if( start->is_Start() ) | |
897 return start->as_Start(); | |
898 } | |
899 ShouldNotReachHere(); | |
900 return NULL; | |
901 } | |
902 | |
903 //-------------------------------immutable_memory------------------------------------- | |
904 // Access immutable memory | |
905 Node* Compile::immutable_memory() { | |
906 if (_immutable_memory != NULL) { | |
907 return _immutable_memory; | |
908 } | |
909 StartNode* s = start(); | |
910 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { | |
911 Node *p = s->fast_out(i); | |
912 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { | |
913 _immutable_memory = p; | |
914 return _immutable_memory; | |
915 } | |
916 } | |
917 ShouldNotReachHere(); | |
918 return NULL; | |
919 } | |
920 | |
921 //----------------------set_cached_top_node------------------------------------ | |
922 // Install the cached top node, and make sure Node::is_top works correctly. | |
923 void Compile::set_cached_top_node(Node* tn) { | |
924 if (tn != NULL) verify_top(tn); | |
925 Node* old_top = _top; | |
926 _top = tn; | |
927 // Calling Node::setup_is_top allows the nodes the chance to adjust | |
928 // their _out arrays. | |
929 if (_top != NULL) _top->setup_is_top(); | |
930 if (old_top != NULL) old_top->setup_is_top(); | |
931 assert(_top == NULL || top()->is_top(), ""); | |
932 } | |
933 | |
934 #ifndef PRODUCT | |
935 void Compile::verify_top(Node* tn) const { | |
936 if (tn != NULL) { | |
937 assert(tn->is_Con(), "top node must be a constant"); | |
938 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); | |
939 assert(tn->in(0) != NULL, "must have live top node"); | |
940 } | |
941 } | |
942 #endif | |
943 | |
944 | |
945 ///-------------------Managing Per-Node Debug & Profile Info------------------- | |
946 | |
947 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { | |
948 guarantee(arr != NULL, ""); | |
949 int num_blocks = arr->length(); | |
950 if (grow_by < num_blocks) grow_by = num_blocks; | |
951 int num_notes = grow_by * _node_notes_block_size; | |
952 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); | |
953 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); | |
954 while (num_notes > 0) { | |
955 arr->append(notes); | |
956 notes += _node_notes_block_size; | |
957 num_notes -= _node_notes_block_size; | |
958 } | |
959 assert(num_notes == 0, "exact multiple, please"); | |
960 } | |
961 | |
962 bool Compile::copy_node_notes_to(Node* dest, Node* source) { | |
963 if (source == NULL || dest == NULL) return false; | |
964 | |
965 if (dest->is_Con()) | |
966 return false; // Do not push debug info onto constants. | |
967 | |
968 #ifdef ASSERT | |
969 // Leave a bread crumb trail pointing to the original node: | |
970 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { | |
971 dest->set_debug_orig(source); | |
972 } | |
973 #endif | |
974 | |
975 if (node_note_array() == NULL) | |
976 return false; // Not collecting any notes now. | |
977 | |
978 // This is a copy onto a pre-existing node, which may already have notes. | |
979 // If both nodes have notes, do not overwrite any pre-existing notes. | |
980 Node_Notes* source_notes = node_notes_at(source->_idx); | |
981 if (source_notes == NULL || source_notes->is_clear()) return false; | |
982 Node_Notes* dest_notes = node_notes_at(dest->_idx); | |
983 if (dest_notes == NULL || dest_notes->is_clear()) { | |
984 return set_node_notes_at(dest->_idx, source_notes); | |
985 } | |
986 | |
987 Node_Notes merged_notes = (*source_notes); | |
988 // The order of operations here ensures that dest notes will win... | |
989 merged_notes.update_from(dest_notes); | |
990 return set_node_notes_at(dest->_idx, &merged_notes); | |
991 } | |
992 | |
993 | |
994 //--------------------------allow_range_check_smearing------------------------- | |
995 // Gating condition for coalescing similar range checks. | |
996 // Sometimes we try 'speculatively' replacing a series of a range checks by a | |
997 // single covering check that is at least as strong as any of them. | |
998 // If the optimization succeeds, the simplified (strengthened) range check | |
999 // will always succeed. If it fails, we will deopt, and then give up | |
1000 // on the optimization. | |
1001 bool Compile::allow_range_check_smearing() const { | |
1002 // If this method has already thrown a range-check, | |
1003 // assume it was because we already tried range smearing | |
1004 // and it failed. | |
1005 uint already_trapped = trap_count(Deoptimization::Reason_range_check); | |
1006 return !already_trapped; | |
1007 } | |
1008 | |
1009 | |
1010 //------------------------------flatten_alias_type----------------------------- | |
1011 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { | |
1012 int offset = tj->offset(); | |
1013 TypePtr::PTR ptr = tj->ptr(); | |
1014 | |
247 | 1015 // Known instance (scalarizable allocation) alias only with itself. |
1016 bool is_known_inst = tj->isa_oopptr() != NULL && | |
1017 tj->is_oopptr()->is_known_instance(); | |
1018 | |
0 | 1019 // Process weird unsafe references. |
1020 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { | |
1021 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); | |
247 | 1022 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); |
0 | 1023 tj = TypeOopPtr::BOTTOM; |
1024 ptr = tj->ptr(); | |
1025 offset = tj->offset(); | |
1026 } | |
1027 | |
1028 // Array pointers need some flattening | |
1029 const TypeAryPtr *ta = tj->isa_aryptr(); | |
247 | 1030 if( ta && is_known_inst ) { |
1031 if ( offset != Type::OffsetBot && | |
1032 offset > arrayOopDesc::length_offset_in_bytes() ) { | |
1033 offset = Type::OffsetBot; // Flatten constant access into array body only | |
1034 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); | |
1035 } | |
1036 } else if( ta && _AliasLevel >= 2 ) { | |
0 | 1037 // For arrays indexed by constant indices, we flatten the alias |
1038 // space to include all of the array body. Only the header, klass | |
1039 // and array length can be accessed un-aliased. | |
1040 if( offset != Type::OffsetBot ) { | |
1041 if( ta->const_oop() ) { // methodDataOop or methodOop | |
1042 offset = Type::OffsetBot; // Flatten constant access into array body | |
247 | 1043 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); |
0 | 1044 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { |
1045 // range is OK as-is. | |
1046 tj = ta = TypeAryPtr::RANGE; | |
1047 } else if( offset == oopDesc::klass_offset_in_bytes() ) { | |
1048 tj = TypeInstPtr::KLASS; // all klass loads look alike | |
1049 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1050 ptr = TypePtr::BotPTR; | |
1051 } else if( offset == oopDesc::mark_offset_in_bytes() ) { | |
1052 tj = TypeInstPtr::MARK; | |
1053 ta = TypeAryPtr::RANGE; // generic ignored junk | |
1054 ptr = TypePtr::BotPTR; | |
1055 } else { // Random constant offset into array body | |
1056 offset = Type::OffsetBot; // Flatten constant access into array body | |
247 | 1057 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); |
0 | 1058 } |
1059 } | |
1060 // Arrays of fixed size alias with arrays of unknown size. | |
1061 if (ta->size() != TypeInt::POS) { | |
1062 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); | |
247 | 1063 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); |
0 | 1064 } |
1065 // Arrays of known objects become arrays of unknown objects. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1066 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1067 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); |
247 | 1068 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1069 } |
0 | 1070 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { |
1071 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); | |
247 | 1072 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); |
0 | 1073 } |
1074 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so | |
1075 // cannot be distinguished by bytecode alone. | |
1076 if (ta->elem() == TypeInt::BOOL) { | |
1077 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); | |
1078 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); | |
247 | 1079 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); |
0 | 1080 } |
1081 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1082 // Make sure the Bottom and NotNull variants alias the same. | |
1083 // Also, make sure exact and non-exact variants alias the same. | |
1084 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { | |
1085 if (ta->const_oop()) { | |
1086 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); | |
1087 } else { | |
1088 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); | |
1089 } | |
1090 } | |
1091 } | |
1092 | |
1093 // Oop pointers need some flattening | |
1094 const TypeInstPtr *to = tj->isa_instptr(); | |
1095 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { | |
1096 if( ptr == TypePtr::Constant ) { | |
1097 // No constant oop pointers (such as Strings); they alias with | |
1098 // unknown strings. | |
247 | 1099 assert(!is_known_inst, "not scalarizable allocation"); |
0 | 1100 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
247 | 1101 } else if( is_known_inst ) { |
163 | 1102 tj = to; // Keep NotNull and klass_is_exact for instance type |
0 | 1103 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { |
1104 // During the 2nd round of IterGVN, NotNull castings are removed. | |
1105 // Make sure the Bottom and NotNull variants alias the same. | |
1106 // Also, make sure exact and non-exact variants alias the same. | |
247 | 1107 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); |
0 | 1108 } |
1109 // Canonicalize the holder of this field | |
1110 ciInstanceKlass *k = to->klass()->as_instance_klass(); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1111 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { |
0 | 1112 // First handle header references such as a LoadKlassNode, even if the |
1113 // object's klass is unloaded at compile time (4965979). | |
247 | 1114 if (!is_known_inst) { // Do it only for non-instance types |
1115 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); | |
1116 } | |
0 | 1117 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { |
1118 to = NULL; | |
1119 tj = TypeOopPtr::BOTTOM; | |
1120 offset = tj->offset(); | |
1121 } else { | |
1122 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); | |
1123 if (!k->equals(canonical_holder) || tj->offset() != offset) { | |
247 | 1124 if( is_known_inst ) { |
1125 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); | |
1126 } else { | |
1127 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); | |
1128 } | |
0 | 1129 } |
1130 } | |
1131 } | |
1132 | |
1133 // Klass pointers to object array klasses need some flattening | |
1134 const TypeKlassPtr *tk = tj->isa_klassptr(); | |
1135 if( tk ) { | |
1136 // If we are referencing a field within a Klass, we need | |
1137 // to assume the worst case of an Object. Both exact and | |
1138 // inexact types must flatten to the same alias class. | |
1139 // Since the flattened result for a klass is defined to be | |
1140 // precisely java.lang.Object, use a constant ptr. | |
1141 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { | |
1142 | |
1143 tj = tk = TypeKlassPtr::make(TypePtr::Constant, | |
1144 TypeKlassPtr::OBJECT->klass(), | |
1145 offset); | |
1146 } | |
1147 | |
1148 ciKlass* klass = tk->klass(); | |
1149 if( klass->is_obj_array_klass() ) { | |
1150 ciKlass* k = TypeAryPtr::OOPS->klass(); | |
1151 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs | |
1152 k = TypeInstPtr::BOTTOM->klass(); | |
1153 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); | |
1154 } | |
1155 | |
1156 // Check for precise loads from the primary supertype array and force them | |
1157 // to the supertype cache alias index. Check for generic array loads from | |
1158 // the primary supertype array and also force them to the supertype cache | |
1159 // alias index. Since the same load can reach both, we need to merge | |
1160 // these 2 disparate memories into the same alias class. Since the | |
1161 // primary supertype array is read-only, there's no chance of confusion | |
1162 // where we bypass an array load and an array store. | |
1163 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); | |
1164 if( offset == Type::OffsetBot || | |
1165 off2 < Klass::primary_super_limit()*wordSize ) { | |
1166 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); | |
1167 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); | |
1168 } | |
1169 } | |
1170 | |
1171 // Flatten all Raw pointers together. | |
1172 if (tj->base() == Type::RawPtr) | |
1173 tj = TypeRawPtr::BOTTOM; | |
1174 | |
1175 if (tj->base() == Type::AnyPtr) | |
1176 tj = TypePtr::BOTTOM; // An error, which the caller must check for. | |
1177 | |
1178 // Flatten all to bottom for now | |
1179 switch( _AliasLevel ) { | |
1180 case 0: | |
1181 tj = TypePtr::BOTTOM; | |
1182 break; | |
1183 case 1: // Flatten to: oop, static, field or array | |
1184 switch (tj->base()) { | |
1185 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; | |
1186 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; | |
1187 case Type::AryPtr: // do not distinguish arrays at all | |
1188 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; | |
1189 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; | |
1190 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it | |
1191 default: ShouldNotReachHere(); | |
1192 } | |
1193 break; | |
1194 case 2: // No collasping at level 2; keep all splits | |
1195 case 3: // No collasping at level 3; keep all splits | |
1196 break; | |
1197 default: | |
1198 Unimplemented(); | |
1199 } | |
1200 | |
1201 offset = tj->offset(); | |
1202 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); | |
1203 | |
1204 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || | |
1205 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || | |
1206 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || | |
1207 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || | |
1208 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1209 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || | |
1210 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , | |
1211 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); | |
1212 assert( tj->ptr() != TypePtr::TopPTR && | |
1213 tj->ptr() != TypePtr::AnyNull && | |
1214 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); | |
1215 // assert( tj->ptr() != TypePtr::Constant || | |
1216 // tj->base() == Type::RawPtr || | |
1217 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); | |
1218 | |
1219 return tj; | |
1220 } | |
1221 | |
1222 void Compile::AliasType::Init(int i, const TypePtr* at) { | |
1223 _index = i; | |
1224 _adr_type = at; | |
1225 _field = NULL; | |
1226 _is_rewritable = true; // default | |
1227 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; | |
223 | 1228 if (atoop != NULL && atoop->is_known_instance()) { |
1229 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); | |
0 | 1230 _general_index = Compile::current()->get_alias_index(gt); |
1231 } else { | |
1232 _general_index = 0; | |
1233 } | |
1234 } | |
1235 | |
1236 //---------------------------------print_on------------------------------------ | |
1237 #ifndef PRODUCT | |
1238 void Compile::AliasType::print_on(outputStream* st) { | |
1239 if (index() < 10) | |
1240 st->print("@ <%d> ", index()); | |
1241 else st->print("@ <%d>", index()); | |
1242 st->print(is_rewritable() ? " " : " RO"); | |
1243 int offset = adr_type()->offset(); | |
1244 if (offset == Type::OffsetBot) | |
1245 st->print(" +any"); | |
1246 else st->print(" +%-3d", offset); | |
1247 st->print(" in "); | |
1248 adr_type()->dump_on(st); | |
1249 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); | |
1250 if (field() != NULL && tjp) { | |
1251 if (tjp->klass() != field()->holder() || | |
1252 tjp->offset() != field()->offset_in_bytes()) { | |
1253 st->print(" != "); | |
1254 field()->print(); | |
1255 st->print(" ***"); | |
1256 } | |
1257 } | |
1258 } | |
1259 | |
1260 void print_alias_types() { | |
1261 Compile* C = Compile::current(); | |
1262 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); | |
1263 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { | |
1264 C->alias_type(idx)->print_on(tty); | |
1265 tty->cr(); | |
1266 } | |
1267 } | |
1268 #endif | |
1269 | |
1270 | |
1271 //----------------------------probe_alias_cache-------------------------------- | |
1272 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { | |
1273 intptr_t key = (intptr_t) adr_type; | |
1274 key ^= key >> logAliasCacheSize; | |
1275 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; | |
1276 } | |
1277 | |
1278 | |
1279 //-----------------------------grow_alias_types-------------------------------- | |
1280 void Compile::grow_alias_types() { | |
1281 const int old_ats = _max_alias_types; // how many before? | |
1282 const int new_ats = old_ats; // how many more? | |
1283 const int grow_ats = old_ats+new_ats; // how many now? | |
1284 _max_alias_types = grow_ats; | |
1285 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); | |
1286 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); | |
1287 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); | |
1288 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; | |
1289 } | |
1290 | |
1291 | |
1292 //--------------------------------find_alias_type------------------------------ | |
1293 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { | |
1294 if (_AliasLevel == 0) | |
1295 return alias_type(AliasIdxBot); | |
1296 | |
1297 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1298 if (ace->_adr_type == adr_type) { | |
1299 return alias_type(ace->_index); | |
1300 } | |
1301 | |
1302 // Handle special cases. | |
1303 if (adr_type == NULL) return alias_type(AliasIdxTop); | |
1304 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); | |
1305 | |
1306 // Do it the slow way. | |
1307 const TypePtr* flat = flatten_alias_type(adr_type); | |
1308 | |
1309 #ifdef ASSERT | |
1310 assert(flat == flatten_alias_type(flat), "idempotent"); | |
1311 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); | |
1312 if (flat->isa_oopptr() && !flat->isa_klassptr()) { | |
1313 const TypeOopPtr* foop = flat->is_oopptr(); | |
247 | 1314 // Scalarizable allocations have exact klass always. |
1315 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); | |
1316 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); | |
0 | 1317 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); |
1318 } | |
1319 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); | |
1320 #endif | |
1321 | |
1322 int idx = AliasIdxTop; | |
1323 for (int i = 0; i < num_alias_types(); i++) { | |
1324 if (alias_type(i)->adr_type() == flat) { | |
1325 idx = i; | |
1326 break; | |
1327 } | |
1328 } | |
1329 | |
1330 if (idx == AliasIdxTop) { | |
1331 if (no_create) return NULL; | |
1332 // Grow the array if necessary. | |
1333 if (_num_alias_types == _max_alias_types) grow_alias_types(); | |
1334 // Add a new alias type. | |
1335 idx = _num_alias_types++; | |
1336 _alias_types[idx]->Init(idx, flat); | |
1337 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); | |
1338 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); | |
1339 if (flat->isa_instptr()) { | |
1340 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() | |
1341 && flat->is_instptr()->klass() == env()->Class_klass()) | |
1342 alias_type(idx)->set_rewritable(false); | |
1343 } | |
1344 if (flat->isa_klassptr()) { | |
1345 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1346 alias_type(idx)->set_rewritable(false); | |
1347 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1348 alias_type(idx)->set_rewritable(false); | |
1349 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1350 alias_type(idx)->set_rewritable(false); | |
1351 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) | |
1352 alias_type(idx)->set_rewritable(false); | |
1353 } | |
1354 // %%% (We would like to finalize JavaThread::threadObj_offset(), | |
1355 // but the base pointer type is not distinctive enough to identify | |
1356 // references into JavaThread.) | |
1357 | |
1358 // Check for final instance fields. | |
1359 const TypeInstPtr* tinst = flat->isa_instptr(); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1360 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { |
0 | 1361 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); |
1362 ciField* field = k->get_field_by_offset(tinst->offset(), false); | |
1363 // Set field() and is_rewritable() attributes. | |
1364 if (field != NULL) alias_type(idx)->set_field(field); | |
1365 } | |
1366 const TypeKlassPtr* tklass = flat->isa_klassptr(); | |
1367 // Check for final static fields. | |
1368 if (tklass && tklass->klass()->is_instance_klass()) { | |
1369 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); | |
1370 ciField* field = k->get_field_by_offset(tklass->offset(), true); | |
1371 // Set field() and is_rewritable() attributes. | |
1372 if (field != NULL) alias_type(idx)->set_field(field); | |
1373 } | |
1374 } | |
1375 | |
1376 // Fill the cache for next time. | |
1377 ace->_adr_type = adr_type; | |
1378 ace->_index = idx; | |
1379 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); | |
1380 | |
1381 // Might as well try to fill the cache for the flattened version, too. | |
1382 AliasCacheEntry* face = probe_alias_cache(flat); | |
1383 if (face->_adr_type == NULL) { | |
1384 face->_adr_type = flat; | |
1385 face->_index = idx; | |
1386 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); | |
1387 } | |
1388 | |
1389 return alias_type(idx); | |
1390 } | |
1391 | |
1392 | |
1393 Compile::AliasType* Compile::alias_type(ciField* field) { | |
1394 const TypeOopPtr* t; | |
1395 if (field->is_static()) | |
1396 t = TypeKlassPtr::make(field->holder()); | |
1397 else | |
1398 t = TypeOopPtr::make_from_klass_raw(field->holder()); | |
1399 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); | |
1400 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); | |
1401 return atp; | |
1402 } | |
1403 | |
1404 | |
1405 //------------------------------have_alias_type-------------------------------- | |
1406 bool Compile::have_alias_type(const TypePtr* adr_type) { | |
1407 AliasCacheEntry* ace = probe_alias_cache(adr_type); | |
1408 if (ace->_adr_type == adr_type) { | |
1409 return true; | |
1410 } | |
1411 | |
1412 // Handle special cases. | |
1413 if (adr_type == NULL) return true; | |
1414 if (adr_type == TypePtr::BOTTOM) return true; | |
1415 | |
1416 return find_alias_type(adr_type, true) != NULL; | |
1417 } | |
1418 | |
1419 //-----------------------------must_alias-------------------------------------- | |
1420 // True if all values of the given address type are in the given alias category. | |
1421 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { | |
1422 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1423 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP | |
1424 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1425 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins | |
1426 | |
1427 // the only remaining possible overlap is identity | |
1428 int adr_idx = get_alias_index(adr_type); | |
1429 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1430 assert(adr_idx == alias_idx || | |
1431 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM | |
1432 && adr_type != TypeOopPtr::BOTTOM), | |
1433 "should not be testing for overlap with an unsafe pointer"); | |
1434 return adr_idx == alias_idx; | |
1435 } | |
1436 | |
1437 //------------------------------can_alias-------------------------------------- | |
1438 // True if any values of the given address type are in the given alias category. | |
1439 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { | |
1440 if (alias_idx == AliasIdxTop) return false; // the empty category | |
1441 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP | |
1442 if (alias_idx == AliasIdxBot) return true; // the universal category | |
1443 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins | |
1444 | |
1445 // the only remaining possible overlap is identity | |
1446 int adr_idx = get_alias_index(adr_type); | |
1447 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); | |
1448 return adr_idx == alias_idx; | |
1449 } | |
1450 | |
1451 | |
1452 | |
1453 //---------------------------pop_warm_call------------------------------------- | |
1454 WarmCallInfo* Compile::pop_warm_call() { | |
1455 WarmCallInfo* wci = _warm_calls; | |
1456 if (wci != NULL) _warm_calls = wci->remove_from(wci); | |
1457 return wci; | |
1458 } | |
1459 | |
1460 //----------------------------Inline_Warm-------------------------------------- | |
1461 int Compile::Inline_Warm() { | |
1462 // If there is room, try to inline some more warm call sites. | |
1463 // %%% Do a graph index compaction pass when we think we're out of space? | |
1464 if (!InlineWarmCalls) return 0; | |
1465 | |
1466 int calls_made_hot = 0; | |
1467 int room_to_grow = NodeCountInliningCutoff - unique(); | |
1468 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); | |
1469 int amount_grown = 0; | |
1470 WarmCallInfo* call; | |
1471 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { | |
1472 int est_size = (int)call->size(); | |
1473 if (est_size > (room_to_grow - amount_grown)) { | |
1474 // This one won't fit anyway. Get rid of it. | |
1475 call->make_cold(); | |
1476 continue; | |
1477 } | |
1478 call->make_hot(); | |
1479 calls_made_hot++; | |
1480 amount_grown += est_size; | |
1481 amount_to_grow -= est_size; | |
1482 } | |
1483 | |
1484 if (calls_made_hot > 0) set_major_progress(); | |
1485 return calls_made_hot; | |
1486 } | |
1487 | |
1488 | |
1489 //----------------------------Finish_Warm-------------------------------------- | |
1490 void Compile::Finish_Warm() { | |
1491 if (!InlineWarmCalls) return; | |
1492 if (failing()) return; | |
1493 if (warm_calls() == NULL) return; | |
1494 | |
1495 // Clean up loose ends, if we are out of space for inlining. | |
1496 WarmCallInfo* call; | |
1497 while ((call = pop_warm_call()) != NULL) { | |
1498 call->make_cold(); | |
1499 } | |
1500 } | |
1501 | |
1502 | |
1503 //------------------------------Optimize--------------------------------------- | |
1504 // Given a graph, optimize it. | |
1505 void Compile::Optimize() { | |
1506 TracePhase t1("optimizer", &_t_optimizer, true); | |
1507 | |
1508 #ifndef PRODUCT | |
1509 if (env()->break_at_compile()) { | |
1510 BREAKPOINT; | |
1511 } | |
1512 | |
1513 #endif | |
1514 | |
1515 ResourceMark rm; | |
1516 int loop_opts_cnt; | |
1517 | |
1518 NOT_PRODUCT( verify_graph_edges(); ) | |
1519 | |
222 | 1520 print_method("After Parsing"); |
0 | 1521 |
1522 { | |
1523 // Iterative Global Value Numbering, including ideal transforms | |
1524 // Initialize IterGVN with types and values from parse-time GVN | |
1525 PhaseIterGVN igvn(initial_gvn()); | |
1526 { | |
1527 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) | |
1528 igvn.optimize(); | |
1529 } | |
1530 | |
1531 print_method("Iter GVN 1", 2); | |
1532 | |
1533 if (failing()) return; | |
1534 | |
1535 // Loop transforms on the ideal graph. Range Check Elimination, | |
1536 // peeling, unrolling, etc. | |
1537 | |
1538 // Set loop opts counter | |
1539 loop_opts_cnt = num_loop_opts(); | |
1540 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { | |
1541 { | |
1542 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
1543 PhaseIdealLoop ideal_loop( igvn, NULL, true ); | |
1544 loop_opts_cnt--; | |
1545 if (major_progress()) print_method("PhaseIdealLoop 1", 2); | |
1546 if (failing()) return; | |
1547 } | |
1548 // Loop opts pass if partial peeling occurred in previous pass | |
1549 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { | |
1550 TracePhase t3("idealLoop", &_t_idealLoop, true); | |
1551 PhaseIdealLoop ideal_loop( igvn, NULL, false ); | |
1552 loop_opts_cnt--; | |
1553 if (major_progress()) print_method("PhaseIdealLoop 2", 2); | |
1554 if (failing()) return; | |
1555 } | |
1556 // Loop opts pass for loop-unrolling before CCP | |
1557 if(major_progress() && (loop_opts_cnt > 0)) { | |
1558 TracePhase t4("idealLoop", &_t_idealLoop, true); | |
1559 PhaseIdealLoop ideal_loop( igvn, NULL, false ); | |
1560 loop_opts_cnt--; | |
1561 if (major_progress()) print_method("PhaseIdealLoop 3", 2); | |
1562 } | |
1563 } | |
1564 if (failing()) return; | |
1565 | |
1566 // Conditional Constant Propagation; | |
1567 PhaseCCP ccp( &igvn ); | |
1568 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); | |
1569 { | |
1570 TracePhase t2("ccp", &_t_ccp, true); | |
1571 ccp.do_transform(); | |
1572 } | |
1573 print_method("PhaseCPP 1", 2); | |
1574 | |
1575 assert( true, "Break here to ccp.dump_old2new_map()"); | |
1576 | |
1577 // Iterative Global Value Numbering, including ideal transforms | |
1578 { | |
1579 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) | |
1580 igvn = ccp; | |
1581 igvn.optimize(); | |
1582 } | |
1583 | |
1584 print_method("Iter GVN 2", 2); | |
1585 | |
1586 if (failing()) return; | |
1587 | |
1588 // Loop transforms on the ideal graph. Range Check Elimination, | |
1589 // peeling, unrolling, etc. | |
1590 if(loop_opts_cnt > 0) { | |
1591 debug_only( int cnt = 0; ); | |
1592 while(major_progress() && (loop_opts_cnt > 0)) { | |
1593 TracePhase t2("idealLoop", &_t_idealLoop, true); | |
1594 assert( cnt++ < 40, "infinite cycle in loop optimization" ); | |
1595 PhaseIdealLoop ideal_loop( igvn, NULL, true ); | |
1596 loop_opts_cnt--; | |
1597 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); | |
1598 if (failing()) return; | |
1599 } | |
1600 } | |
1601 { | |
1602 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) | |
1603 PhaseMacroExpand mex(igvn); | |
1604 if (mex.expand_macro_nodes()) { | |
1605 assert(failing(), "must bail out w/ explicit message"); | |
1606 return; | |
1607 } | |
1608 } | |
1609 | |
1610 } // (End scope of igvn; run destructor if necessary for asserts.) | |
1611 | |
1612 // A method with only infinite loops has no edges entering loops from root | |
1613 { | |
1614 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) | |
1615 if (final_graph_reshaping()) { | |
1616 assert(failing(), "must bail out w/ explicit message"); | |
1617 return; | |
1618 } | |
1619 } | |
1620 | |
1621 print_method("Optimize finished", 2); | |
1622 } | |
1623 | |
1624 | |
1625 //------------------------------Code_Gen--------------------------------------- | |
1626 // Given a graph, generate code for it | |
1627 void Compile::Code_Gen() { | |
1628 if (failing()) return; | |
1629 | |
1630 // Perform instruction selection. You might think we could reclaim Matcher | |
1631 // memory PDQ, but actually the Matcher is used in generating spill code. | |
1632 // Internals of the Matcher (including some VectorSets) must remain live | |
1633 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage | |
1634 // set a bit in reclaimed memory. | |
1635 | |
1636 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
1637 // nodes. Mapping is only valid at the root of each matched subtree. | |
1638 NOT_PRODUCT( verify_graph_edges(); ) | |
1639 | |
1640 Node_List proj_list; | |
1641 Matcher m(proj_list); | |
1642 _matcher = &m; | |
1643 { | |
1644 TracePhase t2("matcher", &_t_matcher, true); | |
1645 m.match(); | |
1646 } | |
1647 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine | |
1648 // nodes. Mapping is only valid at the root of each matched subtree. | |
1649 NOT_PRODUCT( verify_graph_edges(); ) | |
1650 | |
1651 // If you have too many nodes, or if matching has failed, bail out | |
1652 check_node_count(0, "out of nodes matching instructions"); | |
1653 if (failing()) return; | |
1654 | |
1655 // Build a proper-looking CFG | |
1656 PhaseCFG cfg(node_arena(), root(), m); | |
1657 _cfg = &cfg; | |
1658 { | |
1659 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) | |
1660 cfg.Dominators(); | |
1661 if (failing()) return; | |
1662 | |
1663 NOT_PRODUCT( verify_graph_edges(); ) | |
1664 | |
1665 cfg.Estimate_Block_Frequency(); | |
1666 cfg.GlobalCodeMotion(m,unique(),proj_list); | |
1667 | |
1668 print_method("Global code motion", 2); | |
1669 | |
1670 if (failing()) return; | |
1671 NOT_PRODUCT( verify_graph_edges(); ) | |
1672 | |
1673 debug_only( cfg.verify(); ) | |
1674 } | |
1675 NOT_PRODUCT( verify_graph_edges(); ) | |
1676 | |
1677 PhaseChaitin regalloc(unique(),cfg,m); | |
1678 _regalloc = ®alloc; | |
1679 { | |
1680 TracePhase t2("regalloc", &_t_registerAllocation, true); | |
1681 // Perform any platform dependent preallocation actions. This is used, | |
1682 // for example, to avoid taking an implicit null pointer exception | |
1683 // using the frame pointer on win95. | |
1684 _regalloc->pd_preallocate_hook(); | |
1685 | |
1686 // Perform register allocation. After Chaitin, use-def chains are | |
1687 // no longer accurate (at spill code) and so must be ignored. | |
1688 // Node->LRG->reg mappings are still accurate. | |
1689 _regalloc->Register_Allocate(); | |
1690 | |
1691 // Bail out if the allocator builds too many nodes | |
1692 if (failing()) return; | |
1693 } | |
1694 | |
1695 // Prior to register allocation we kept empty basic blocks in case the | |
1696 // the allocator needed a place to spill. After register allocation we | |
1697 // are not adding any new instructions. If any basic block is empty, we | |
1698 // can now safely remove it. | |
1699 { | |
418 | 1700 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) |
1701 cfg.remove_empty(); | |
1702 if (do_freq_based_layout()) { | |
1703 PhaseBlockLayout layout(cfg); | |
1704 } else { | |
1705 cfg.set_loop_alignment(); | |
1706 } | |
1707 cfg.fixup_flow(); | |
0 | 1708 } |
1709 | |
1710 // Perform any platform dependent postallocation verifications. | |
1711 debug_only( _regalloc->pd_postallocate_verify_hook(); ) | |
1712 | |
1713 // Apply peephole optimizations | |
1714 if( OptoPeephole ) { | |
1715 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) | |
1716 PhasePeephole peep( _regalloc, cfg); | |
1717 peep.do_transform(); | |
1718 } | |
1719 | |
1720 // Convert Nodes to instruction bits in a buffer | |
1721 { | |
1722 // %%%% workspace merge brought two timers together for one job | |
1723 TracePhase t2a("output", &_t_output, true); | |
1724 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) | |
1725 Output(); | |
1726 } | |
1727 | |
222 | 1728 print_method("Final Code"); |
0 | 1729 |
1730 // He's dead, Jim. | |
1731 _cfg = (PhaseCFG*)0xdeadbeef; | |
1732 _regalloc = (PhaseChaitin*)0xdeadbeef; | |
1733 } | |
1734 | |
1735 | |
1736 //------------------------------dump_asm--------------------------------------- | |
1737 // Dump formatted assembly | |
1738 #ifndef PRODUCT | |
1739 void Compile::dump_asm(int *pcs, uint pc_limit) { | |
1740 bool cut_short = false; | |
1741 tty->print_cr("#"); | |
1742 tty->print("# "); _tf->dump(); tty->cr(); | |
1743 tty->print_cr("#"); | |
1744 | |
1745 // For all blocks | |
1746 int pc = 0x0; // Program counter | |
1747 char starts_bundle = ' '; | |
1748 _regalloc->dump_frame(); | |
1749 | |
1750 Node *n = NULL; | |
1751 for( uint i=0; i<_cfg->_num_blocks; i++ ) { | |
1752 if (VMThread::should_terminate()) { cut_short = true; break; } | |
1753 Block *b = _cfg->_blocks[i]; | |
1754 if (b->is_connector() && !Verbose) continue; | |
1755 n = b->_nodes[0]; | |
1756 if (pcs && n->_idx < pc_limit) | |
1757 tty->print("%3.3x ", pcs[n->_idx]); | |
1758 else | |
1759 tty->print(" "); | |
1760 b->dump_head( &_cfg->_bbs ); | |
1761 if (b->is_connector()) { | |
1762 tty->print_cr(" # Empty connector block"); | |
1763 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { | |
1764 tty->print_cr(" # Block is sole successor of call"); | |
1765 } | |
1766 | |
1767 // For all instructions | |
1768 Node *delay = NULL; | |
1769 for( uint j = 0; j<b->_nodes.size(); j++ ) { | |
1770 if (VMThread::should_terminate()) { cut_short = true; break; } | |
1771 n = b->_nodes[j]; | |
1772 if (valid_bundle_info(n)) { | |
1773 Bundle *bundle = node_bundling(n); | |
1774 if (bundle->used_in_unconditional_delay()) { | |
1775 delay = n; | |
1776 continue; | |
1777 } | |
1778 if (bundle->starts_bundle()) | |
1779 starts_bundle = '+'; | |
1780 } | |
1781 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1782 if (WizardMode) n->dump(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1783 |
0 | 1784 if( !n->is_Region() && // Dont print in the Assembly |
1785 !n->is_Phi() && // a few noisely useless nodes | |
1786 !n->is_Proj() && | |
1787 !n->is_MachTemp() && | |
1788 !n->is_Catch() && // Would be nice to print exception table targets | |
1789 !n->is_MergeMem() && // Not very interesting | |
1790 !n->is_top() && // Debug info table constants | |
1791 !(n->is_Con() && !n->is_Mach())// Debug info table constants | |
1792 ) { | |
1793 if (pcs && n->_idx < pc_limit) | |
1794 tty->print("%3.3x", pcs[n->_idx]); | |
1795 else | |
1796 tty->print(" "); | |
1797 tty->print(" %c ", starts_bundle); | |
1798 starts_bundle = ' '; | |
1799 tty->print("\t"); | |
1800 n->format(_regalloc, tty); | |
1801 tty->cr(); | |
1802 } | |
1803 | |
1804 // If we have an instruction with a delay slot, and have seen a delay, | |
1805 // then back up and print it | |
1806 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { | |
1807 assert(delay != NULL, "no unconditional delay instruction"); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1808 if (WizardMode) delay->dump(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1809 |
0 | 1810 if (node_bundling(delay)->starts_bundle()) |
1811 starts_bundle = '+'; | |
1812 if (pcs && n->_idx < pc_limit) | |
1813 tty->print("%3.3x", pcs[n->_idx]); | |
1814 else | |
1815 tty->print(" "); | |
1816 tty->print(" %c ", starts_bundle); | |
1817 starts_bundle = ' '; | |
1818 tty->print("\t"); | |
1819 delay->format(_regalloc, tty); | |
1820 tty->print_cr(""); | |
1821 delay = NULL; | |
1822 } | |
1823 | |
1824 // Dump the exception table as well | |
1825 if( n->is_Catch() && (Verbose || WizardMode) ) { | |
1826 // Print the exception table for this offset | |
1827 _handler_table.print_subtable_for(pc); | |
1828 } | |
1829 } | |
1830 | |
1831 if (pcs && n->_idx < pc_limit) | |
1832 tty->print_cr("%3.3x", pcs[n->_idx]); | |
1833 else | |
1834 tty->print_cr(""); | |
1835 | |
1836 assert(cut_short || delay == NULL, "no unconditional delay branch"); | |
1837 | |
1838 } // End of per-block dump | |
1839 tty->print_cr(""); | |
1840 | |
1841 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); | |
1842 } | |
1843 #endif | |
1844 | |
1845 //------------------------------Final_Reshape_Counts--------------------------- | |
1846 // This class defines counters to help identify when a method | |
1847 // may/must be executed using hardware with only 24-bit precision. | |
1848 struct Final_Reshape_Counts : public StackObj { | |
1849 int _call_count; // count non-inlined 'common' calls | |
1850 int _float_count; // count float ops requiring 24-bit precision | |
1851 int _double_count; // count double ops requiring more precision | |
1852 int _java_call_count; // count non-inlined 'java' calls | |
1853 VectorSet _visited; // Visitation flags | |
1854 Node_List _tests; // Set of IfNodes & PCTableNodes | |
1855 | |
1856 Final_Reshape_Counts() : | |
1857 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), | |
1858 _visited( Thread::current()->resource_area() ) { } | |
1859 | |
1860 void inc_call_count () { _call_count ++; } | |
1861 void inc_float_count () { _float_count ++; } | |
1862 void inc_double_count() { _double_count++; } | |
1863 void inc_java_call_count() { _java_call_count++; } | |
1864 | |
1865 int get_call_count () const { return _call_count ; } | |
1866 int get_float_count () const { return _float_count ; } | |
1867 int get_double_count() const { return _double_count; } | |
1868 int get_java_call_count() const { return _java_call_count; } | |
1869 }; | |
1870 | |
1871 static bool oop_offset_is_sane(const TypeInstPtr* tp) { | |
1872 ciInstanceKlass *k = tp->klass()->as_instance_klass(); | |
1873 // Make sure the offset goes inside the instance layout. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
1874 return k->contains_field_offset(tp->offset()); |
0 | 1875 // Note that OffsetBot and OffsetTop are very negative. |
1876 } | |
1877 | |
1878 //------------------------------final_graph_reshaping_impl---------------------- | |
1879 // Implement items 1-5 from final_graph_reshaping below. | |
1880 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { | |
1881 | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1882 if ( n->outcnt() == 0 ) return; // dead node |
0 | 1883 uint nop = n->Opcode(); |
1884 | |
1885 // Check for 2-input instruction with "last use" on right input. | |
1886 // Swap to left input. Implements item (2). | |
1887 if( n->req() == 3 && // two-input instruction | |
1888 n->in(1)->outcnt() > 1 && // left use is NOT a last use | |
1889 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop | |
1890 n->in(2)->outcnt() == 1 &&// right use IS a last use | |
1891 !n->in(2)->is_Con() ) { // right use is not a constant | |
1892 // Check for commutative opcode | |
1893 switch( nop ) { | |
1894 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: | |
1895 case Op_MaxI: case Op_MinI: | |
1896 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: | |
1897 case Op_AndL: case Op_XorL: case Op_OrL: | |
1898 case Op_AndI: case Op_XorI: case Op_OrI: { | |
1899 // Move "last use" input to left by swapping inputs | |
1900 n->swap_edges(1, 2); | |
1901 break; | |
1902 } | |
1903 default: | |
1904 break; | |
1905 } | |
1906 } | |
1907 | |
1908 // Count FPU ops and common calls, implements item (3) | |
1909 switch( nop ) { | |
1910 // Count all float operations that may use FPU | |
1911 case Op_AddF: | |
1912 case Op_SubF: | |
1913 case Op_MulF: | |
1914 case Op_DivF: | |
1915 case Op_NegF: | |
1916 case Op_ModF: | |
1917 case Op_ConvI2F: | |
1918 case Op_ConF: | |
1919 case Op_CmpF: | |
1920 case Op_CmpF3: | |
1921 // case Op_ConvL2F: // longs are split into 32-bit halves | |
1922 fpu.inc_float_count(); | |
1923 break; | |
1924 | |
1925 case Op_ConvF2D: | |
1926 case Op_ConvD2F: | |
1927 fpu.inc_float_count(); | |
1928 fpu.inc_double_count(); | |
1929 break; | |
1930 | |
1931 // Count all double operations that may use FPU | |
1932 case Op_AddD: | |
1933 case Op_SubD: | |
1934 case Op_MulD: | |
1935 case Op_DivD: | |
1936 case Op_NegD: | |
1937 case Op_ModD: | |
1938 case Op_ConvI2D: | |
1939 case Op_ConvD2I: | |
1940 // case Op_ConvL2D: // handled by leaf call | |
1941 // case Op_ConvD2L: // handled by leaf call | |
1942 case Op_ConD: | |
1943 case Op_CmpD: | |
1944 case Op_CmpD3: | |
1945 fpu.inc_double_count(); | |
1946 break; | |
1947 case Op_Opaque1: // Remove Opaque Nodes before matching | |
1948 case Op_Opaque2: // Remove Opaque Nodes before matching | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1949 n->subsume_by(n->in(1)); |
0 | 1950 break; |
1951 case Op_CallStaticJava: | |
1952 case Op_CallJava: | |
1953 case Op_CallDynamicJava: | |
1954 fpu.inc_java_call_count(); // Count java call site; | |
1955 case Op_CallRuntime: | |
1956 case Op_CallLeaf: | |
1957 case Op_CallLeafNoFP: { | |
1958 assert( n->is_Call(), "" ); | |
1959 CallNode *call = n->as_Call(); | |
1960 // Count call sites where the FP mode bit would have to be flipped. | |
1961 // Do not count uncommon runtime calls: | |
1962 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, | |
1963 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... | |
1964 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { | |
1965 fpu.inc_call_count(); // Count the call site | |
1966 } else { // See if uncommon argument is shared | |
1967 Node *n = call->in(TypeFunc::Parms); | |
1968 int nop = n->Opcode(); | |
1969 // Clone shared simple arguments to uncommon calls, item (1). | |
1970 if( n->outcnt() > 1 && | |
1971 !n->is_Proj() && | |
1972 nop != Op_CreateEx && | |
1973 nop != Op_CheckCastPP && | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
1974 nop != Op_DecodeN && |
0 | 1975 !n->is_Mem() ) { |
1976 Node *x = n->clone(); | |
1977 call->set_req( TypeFunc::Parms, x ); | |
1978 } | |
1979 } | |
1980 break; | |
1981 } | |
1982 | |
1983 case Op_StoreD: | |
1984 case Op_LoadD: | |
1985 case Op_LoadD_unaligned: | |
1986 fpu.inc_double_count(); | |
1987 goto handle_mem; | |
1988 case Op_StoreF: | |
1989 case Op_LoadF: | |
1990 fpu.inc_float_count(); | |
1991 goto handle_mem; | |
1992 | |
1993 case Op_StoreB: | |
1994 case Op_StoreC: | |
1995 case Op_StoreCM: | |
1996 case Op_StorePConditional: | |
1997 case Op_StoreI: | |
1998 case Op_StoreL: | |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
418
diff
changeset
|
1999 case Op_StoreIConditional: |
0 | 2000 case Op_StoreLConditional: |
2001 case Op_CompareAndSwapI: | |
2002 case Op_CompareAndSwapL: | |
2003 case Op_CompareAndSwapP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2004 case Op_CompareAndSwapN: |
0 | 2005 case Op_StoreP: |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2006 case Op_StoreN: |
0 | 2007 case Op_LoadB: |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
492
diff
changeset
|
2008 case Op_LoadUS: |
0 | 2009 case Op_LoadI: |
2010 case Op_LoadKlass: | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2011 case Op_LoadNKlass: |
0 | 2012 case Op_LoadL: |
2013 case Op_LoadL_unaligned: | |
2014 case Op_LoadPLocked: | |
2015 case Op_LoadLLocked: | |
2016 case Op_LoadP: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
100
diff
changeset
|
2017 case Op_LoadN: |
0 | 2018 case Op_LoadRange: |
2019 case Op_LoadS: { | |
2020 handle_mem: | |
2021 #ifdef ASSERT | |
2022 if( VerifyOptoOopOffsets ) { | |
2023 assert( n->is_Mem(), "" ); | |
2024 MemNode *mem = (MemNode*)n; | |
2025 // Check to see if address types have grounded out somehow. | |
2026 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); | |
2027 assert( !tp || oop_offset_is_sane(tp), "" ); | |
2028 } | |
2029 #endif | |
2030 break; | |
2031 } | |
2032 | |
2033 case Op_AddP: { // Assert sane base pointers | |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2034 Node *addp = n->in(AddPNode::Address); |
0 | 2035 assert( !addp->is_AddP() || |
2036 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation | |
2037 addp->in(AddPNode::Base) == n->in(AddPNode::Base), | |
2038 "Base pointers must match" ); | |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2039 #ifdef _LP64 |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2040 if (UseCompressedOops && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2041 addp->Opcode() == Op_ConP && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2042 addp == n->in(AddPNode::Base) && |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2043 n->in(AddPNode::Offset)->is_Con()) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2044 // Use addressing with narrow klass to load with offset on x86. |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2045 // On sparc loading 32-bits constant and decoding it have less |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2046 // instructions (4) then load 64-bits constant (7). |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2047 // Do this transformation here since IGVN will convert ConN back to ConP. |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2048 const Type* t = addp->bottom_type(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2049 if (t->isa_oopptr()) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2050 Node* nn = NULL; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2051 |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2052 // Look for existing ConN node of the same exact type. |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2053 Compile* C = Compile::current(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2054 Node* r = C->root(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2055 uint cnt = r->outcnt(); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2056 for (uint i = 0; i < cnt; i++) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2057 Node* m = r->raw_out(i); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2058 if (m!= NULL && m->Opcode() == Op_ConN && |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
182
diff
changeset
|
2059 m->bottom_type()->make_ptr() == t) { |
182
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2060 nn = m; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2061 break; |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2062 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2063 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2064 if (nn != NULL) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2065 // Decode a narrow oop to match address |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2066 // [R12 + narrow_oop_reg<<3 + offset] |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2067 nn = new (C, 2) DecodeNNode(nn, t); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2068 n->set_req(AddPNode::Base, nn); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2069 n->set_req(AddPNode::Address, nn); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2070 if (addp->outcnt() == 0) { |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2071 addp->disconnect_inputs(NULL); |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2072 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2073 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2074 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2075 } |
44abbb0d4c18
6709093: Compressed Oops: reduce size of compiled methods
kvn
parents:
168
diff
changeset
|
2076 #endif |
0 | 2077 break; |
2078 } | |
2079 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2080 #ifdef _LP64 |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2081 case Op_CastPP: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2082 if (n->in(1)->is_DecodeN() && UseImplicitNullCheckForNarrowOop) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2083 Compile* C = Compile::current(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2084 Node* in1 = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2085 const Type* t = n->bottom_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2086 Node* new_in1 = in1->clone(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2087 new_in1->as_DecodeN()->set_type(t); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2088 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2089 if (!Matcher::clone_shift_expressions) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2090 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2091 // x86, ARM and friends can handle 2 adds in addressing mode |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2092 // and Matcher can fold a DecodeN node into address by using |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2093 // a narrow oop directly and do implicit NULL check in address: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2094 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2095 // [R12 + narrow_oop_reg<<3 + offset] |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2096 // NullCheck narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2097 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2098 // On other platforms (Sparc) we have to keep new DecodeN node and |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2099 // use it to do implicit NULL check in address: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2100 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2101 // decode_not_null narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2102 // [base_reg + offset] |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2103 // NullCheck base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2104 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2105 // Pin the new DecodeN node to non-null path on these patforms (Sparc) |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2106 // to keep the information to which NULL check the new DecodeN node |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2107 // corresponds to use it as value in implicit_null_check(). |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2108 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2109 new_in1->set_req(0, n->in(0)); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2110 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2111 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2112 n->subsume_by(new_in1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2113 if (in1->outcnt() == 0) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2114 in1->disconnect_inputs(NULL); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2115 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2116 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2117 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2118 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2119 case Op_CmpP: |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2120 // Do this transformation here to preserve CmpPNode::sub() and |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2121 // other TypePtr related Ideal optimizations (for example, ptr nullness). |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2122 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2123 Node* in1 = n->in(1); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2124 Node* in2 = n->in(2); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2125 if (!in1->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2126 in2 = in1; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2127 in1 = n->in(2); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2128 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2129 assert(in1->is_DecodeN(), "sanity"); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2130 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2131 Compile* C = Compile::current(); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2132 Node* new_in2 = NULL; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2133 if (in2->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2134 new_in2 = in2->in(1); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2135 } else if (in2->Opcode() == Op_ConP) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2136 const Type* t = in2->bottom_type(); |
325
93befa083681
6741004: UseLargePages + UseCompressedOops breaks implicit null checking guard page
coleenp
parents:
293
diff
changeset
|
2137 if (t == TypePtr::NULL_PTR && UseImplicitNullCheckForNarrowOop) { |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2138 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2139 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2140 // This transformation together with CastPP transformation above |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2141 // will generated code for implicit NULL checks for compressed oops. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2142 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2143 // The original code after Optimize() |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2144 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2145 // LoadN memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2146 // decode narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2147 // CmpP base_reg, NULL |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2148 // CastPP base_reg // NotNull |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2149 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2150 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2151 // after these transformations will be |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2152 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2153 // LoadN memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2154 // CmpN narrow_oop_reg, NULL |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2155 // decode_not_null narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2156 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2157 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2158 // and the uncommon path (== NULL) will use narrow_oop_reg directly |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2159 // since narrow oops can be used in debug info now (see the code in |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2160 // final_graph_reshaping_walk()). |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2161 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2162 // At the end the code will be matched to |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2163 // on x86: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2164 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2165 // Load_narrow_oop memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2166 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2167 // NullCheck narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2168 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2169 // and on sparc: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2170 // |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2171 // Load_narrow_oop memory, narrow_oop_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2172 // decode_not_null narrow_oop_reg, base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2173 // Load [base_reg + offset], val_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2174 // NullCheck base_reg |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2175 // |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2176 } else if (t->isa_oopptr()) { |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2177 new_in2 = ConNode::make(C, t->make_narrowoop()); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2178 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2179 } |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2180 if (new_in2 != NULL) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2181 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2182 n->subsume_by( cmpN ); |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2183 if (in1->outcnt() == 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2184 in1->disconnect_inputs(NULL); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2185 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2186 if (in2->outcnt() == 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2187 in2->disconnect_inputs(NULL); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2188 } |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2189 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2190 } |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
2191 break; |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2192 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2193 case Op_DecodeN: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2194 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); |
492
5496e074077f
6787050: assert(n->in(0) == 0L,"no control") with UseCompressedOops on sparcv9
kvn
parents:
490
diff
changeset
|
2195 // DecodeN could be pinned on Sparc where it can't be fold into |
5496e074077f
6787050: assert(n->in(0) == 0L,"no control") with UseCompressedOops on sparcv9
kvn
parents:
490
diff
changeset
|
2196 // an address expression, see the code for Op_CastPP above. |
5496e074077f
6787050: assert(n->in(0) == 0L,"no control") with UseCompressedOops on sparcv9
kvn
parents:
490
diff
changeset
|
2197 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc"); |
368
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2198 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2199 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2200 case Op_EncodeP: { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2201 Node* in1 = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2202 if (in1->is_DecodeN()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2203 n->subsume_by(in1->in(1)); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2204 } else if (in1->Opcode() == Op_ConP) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2205 Compile* C = Compile::current(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2206 const Type* t = in1->bottom_type(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2207 if (t == TypePtr::NULL_PTR) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2208 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2209 } else if (t->isa_oopptr()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2210 n->subsume_by(ConNode::make(C, t->make_narrowoop())); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2211 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2212 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2213 if (in1->outcnt() == 0) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2214 in1->disconnect_inputs(NULL); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2215 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2216 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2217 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2218 |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2219 case Op_Phi: |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2220 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2221 // The EncodeP optimization may create Phi with the same edges |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2222 // for all paths. It is not handled well by Register Allocator. |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2223 Node* unique_in = n->in(1); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2224 assert(unique_in != NULL, ""); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2225 uint cnt = n->req(); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2226 for (uint i = 2; i < cnt; i++) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2227 Node* m = n->in(i); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2228 assert(m != NULL, ""); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2229 if (unique_in != m) |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2230 unique_in = NULL; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2231 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2232 if (unique_in != NULL) { |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2233 n->subsume_by(unique_in); |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2234 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2235 } |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2236 break; |
36ccc817fca4
6747051: Improve code and implicit null check generation for compressed oops
kvn
parents:
367
diff
changeset
|
2237 |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2238 #endif |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2239 |
0 | 2240 case Op_ModI: |
2241 if (UseDivMod) { | |
2242 // Check if a%b and a/b both exist | |
2243 Node* d = n->find_similar(Op_DivI); | |
2244 if (d) { | |
2245 // Replace them with a fused divmod if supported | |
2246 Compile* C = Compile::current(); | |
2247 if (Matcher::has_match_rule(Op_DivModI)) { | |
2248 DivModINode* divmod = DivModINode::make(C, n); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2249 d->subsume_by(divmod->div_proj()); |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2250 n->subsume_by(divmod->mod_proj()); |
0 | 2251 } else { |
2252 // replace a%b with a-((a/b)*b) | |
2253 Node* mult = new (C, 3) MulINode(d, d->in(2)); | |
2254 Node* sub = new (C, 3) SubINode(d->in(1), mult); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2255 n->subsume_by( sub ); |
0 | 2256 } |
2257 } | |
2258 } | |
2259 break; | |
2260 | |
2261 case Op_ModL: | |
2262 if (UseDivMod) { | |
2263 // Check if a%b and a/b both exist | |
2264 Node* d = n->find_similar(Op_DivL); | |
2265 if (d) { | |
2266 // Replace them with a fused divmod if supported | |
2267 Compile* C = Compile::current(); | |
2268 if (Matcher::has_match_rule(Op_DivModL)) { | |
2269 DivModLNode* divmod = DivModLNode::make(C, n); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2270 d->subsume_by(divmod->div_proj()); |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2271 n->subsume_by(divmod->mod_proj()); |
0 | 2272 } else { |
2273 // replace a%b with a-((a/b)*b) | |
2274 Node* mult = new (C, 3) MulLNode(d, d->in(2)); | |
2275 Node* sub = new (C, 3) SubLNode(d->in(1), mult); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2276 n->subsume_by( sub ); |
0 | 2277 } |
2278 } | |
2279 } | |
2280 break; | |
2281 | |
2282 case Op_Load16B: | |
2283 case Op_Load8B: | |
2284 case Op_Load4B: | |
2285 case Op_Load8S: | |
2286 case Op_Load4S: | |
2287 case Op_Load2S: | |
2288 case Op_Load8C: | |
2289 case Op_Load4C: | |
2290 case Op_Load2C: | |
2291 case Op_Load4I: | |
2292 case Op_Load2I: | |
2293 case Op_Load2L: | |
2294 case Op_Load4F: | |
2295 case Op_Load2F: | |
2296 case Op_Load2D: | |
2297 case Op_Store16B: | |
2298 case Op_Store8B: | |
2299 case Op_Store4B: | |
2300 case Op_Store8C: | |
2301 case Op_Store4C: | |
2302 case Op_Store2C: | |
2303 case Op_Store4I: | |
2304 case Op_Store2I: | |
2305 case Op_Store2L: | |
2306 case Op_Store4F: | |
2307 case Op_Store2F: | |
2308 case Op_Store2D: | |
2309 break; | |
2310 | |
2311 case Op_PackB: | |
2312 case Op_PackS: | |
2313 case Op_PackC: | |
2314 case Op_PackI: | |
2315 case Op_PackF: | |
2316 case Op_PackL: | |
2317 case Op_PackD: | |
2318 if (n->req()-1 > 2) { | |
2319 // Replace many operand PackNodes with a binary tree for matching | |
2320 PackNode* p = (PackNode*) n; | |
2321 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); | |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2322 n->subsume_by(btp); |
0 | 2323 } |
2324 break; | |
2325 default: | |
2326 assert( !n->is_Call(), "" ); | |
2327 assert( !n->is_Mem(), "" ); | |
2328 break; | |
2329 } | |
127 | 2330 |
2331 // Collect CFG split points | |
2332 if (n->is_MultiBranch()) | |
2333 fpu._tests.push(n); | |
0 | 2334 } |
2335 | |
2336 //------------------------------final_graph_reshaping_walk--------------------- | |
2337 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), | |
2338 // requires that the walk visits a node's inputs before visiting the node. | |
2339 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2340 ResourceArea *area = Thread::current()->resource_area(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2341 Unique_Node_List sfpt(area); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2342 |
0 | 2343 fpu._visited.set(root->_idx); // first, mark node as visited |
2344 uint cnt = root->req(); | |
2345 Node *n = root; | |
2346 uint i = 0; | |
2347 while (true) { | |
2348 if (i < cnt) { | |
2349 // Place all non-visited non-null inputs onto stack | |
2350 Node* m = n->in(i); | |
2351 ++i; | |
2352 if (m != NULL && !fpu._visited.test_set(m->_idx)) { | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2353 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2354 sfpt.push(m); |
0 | 2355 cnt = m->req(); |
2356 nstack.push(n, i); // put on stack parent and next input's index | |
2357 n = m; | |
2358 i = 0; | |
2359 } | |
2360 } else { | |
2361 // Now do post-visit work | |
2362 final_graph_reshaping_impl( n, fpu ); | |
2363 if (nstack.is_empty()) | |
2364 break; // finished | |
2365 n = nstack.node(); // Get node from stack | |
2366 cnt = n->req(); | |
2367 i = nstack.index(); | |
2368 nstack.pop(); // Shift to the next node on stack | |
2369 } | |
2370 } | |
331
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2371 |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2372 // Go over safepoints nodes to skip DecodeN nodes for debug edges. |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2373 // It could be done for an uncommon traps or any safepoints/calls |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2374 // if the DecodeN node is referenced only in a debug info. |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2375 while (sfpt.size() > 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2376 n = sfpt.pop(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2377 JVMState *jvms = n->as_SafePoint()->jvms(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2378 assert(jvms != NULL, "sanity"); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2379 int start = jvms->debug_start(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2380 int end = n->req(); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2381 bool is_uncommon = (n->is_CallStaticJava() && |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2382 n->as_CallStaticJava()->uncommon_trap_request() != 0); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2383 for (int j = start; j < end; j++) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2384 Node* in = n->in(j); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2385 if (in->is_DecodeN()) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2386 bool safe_to_skip = true; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2387 if (!is_uncommon ) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2388 // Is it safe to skip? |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2389 for (uint i = 0; i < in->outcnt(); i++) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2390 Node* u = in->raw_out(i); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2391 if (!u->is_SafePoint() || |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2392 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2393 safe_to_skip = false; |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2394 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2395 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2396 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2397 if (safe_to_skip) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2398 n->set_req(j, in->in(1)); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2399 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2400 if (in->outcnt() == 0) { |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2401 in->disconnect_inputs(NULL); |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2402 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2403 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2404 } |
cecd8eb4e0ca
6706829: Compressed Oops: add debug info for narrow oops
kvn
parents:
293
diff
changeset
|
2405 } |
0 | 2406 } |
2407 | |
2408 //------------------------------final_graph_reshaping-------------------------- | |
2409 // Final Graph Reshaping. | |
2410 // | |
2411 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late | |
2412 // and not commoned up and forced early. Must come after regular | |
2413 // optimizations to avoid GVN undoing the cloning. Clone constant | |
2414 // inputs to Loop Phis; these will be split by the allocator anyways. | |
2415 // Remove Opaque nodes. | |
2416 // (2) Move last-uses by commutative operations to the left input to encourage | |
2417 // Intel update-in-place two-address operations and better register usage | |
2418 // on RISCs. Must come after regular optimizations to avoid GVN Ideal | |
2419 // calls canonicalizing them back. | |
2420 // (3) Count the number of double-precision FP ops, single-precision FP ops | |
2421 // and call sites. On Intel, we can get correct rounding either by | |
2422 // forcing singles to memory (requires extra stores and loads after each | |
2423 // FP bytecode) or we can set a rounding mode bit (requires setting and | |
2424 // clearing the mode bit around call sites). The mode bit is only used | |
2425 // if the relative frequency of single FP ops to calls is low enough. | |
2426 // This is a key transform for SPEC mpeg_audio. | |
2427 // (4) Detect infinite loops; blobs of code reachable from above but not | |
2428 // below. Several of the Code_Gen algorithms fail on such code shapes, | |
2429 // so we simply bail out. Happens a lot in ZKM.jar, but also happens | |
2430 // from time to time in other codes (such as -Xcomp finalizer loops, etc). | |
2431 // Detection is by looking for IfNodes where only 1 projection is | |
2432 // reachable from below or CatchNodes missing some targets. | |
2433 // (5) Assert for insane oop offsets in debug mode. | |
2434 | |
2435 bool Compile::final_graph_reshaping() { | |
2436 // an infinite loop may have been eliminated by the optimizer, | |
2437 // in which case the graph will be empty. | |
2438 if (root()->req() == 1) { | |
2439 record_method_not_compilable("trivial infinite loop"); | |
2440 return true; | |
2441 } | |
2442 | |
2443 Final_Reshape_Counts fpu; | |
2444 | |
2445 // Visit everybody reachable! | |
2446 // Allocate stack of size C->unique()/2 to avoid frequent realloc | |
2447 Node_Stack nstack(unique() >> 1); | |
2448 final_graph_reshaping_walk(nstack, root(), fpu); | |
2449 | |
2450 // Check for unreachable (from below) code (i.e., infinite loops). | |
2451 for( uint i = 0; i < fpu._tests.size(); i++ ) { | |
127 | 2452 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch(); |
2453 // Get number of CFG targets. | |
0 | 2454 // Note that PCTables include exception targets after calls. |
127 | 2455 uint required_outcnt = n->required_outcnt(); |
2456 if (n->outcnt() != required_outcnt) { | |
0 | 2457 // Check for a few special cases. Rethrow Nodes never take the |
2458 // 'fall-thru' path, so expected kids is 1 less. | |
2459 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { | |
2460 if (n->in(0)->in(0)->is_Call()) { | |
2461 CallNode *call = n->in(0)->in(0)->as_Call(); | |
2462 if (call->entry_point() == OptoRuntime::rethrow_stub()) { | |
127 | 2463 required_outcnt--; // Rethrow always has 1 less kid |
0 | 2464 } else if (call->req() > TypeFunc::Parms && |
2465 call->is_CallDynamicJava()) { | |
2466 // Check for null receiver. In such case, the optimizer has | |
2467 // detected that the virtual call will always result in a null | |
2468 // pointer exception. The fall-through projection of this CatchNode | |
2469 // will not be populated. | |
2470 Node *arg0 = call->in(TypeFunc::Parms); | |
2471 if (arg0->is_Type() && | |
2472 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { | |
127 | 2473 required_outcnt--; |
0 | 2474 } |
2475 } else if (call->entry_point() == OptoRuntime::new_array_Java() && | |
2476 call->req() > TypeFunc::Parms+1 && | |
2477 call->is_CallStaticJava()) { | |
2478 // Check for negative array length. In such case, the optimizer has | |
2479 // detected that the allocation attempt will always result in an | |
2480 // exception. There is no fall-through projection of this CatchNode . | |
2481 Node *arg1 = call->in(TypeFunc::Parms+1); | |
2482 if (arg1->is_Type() && | |
2483 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { | |
127 | 2484 required_outcnt--; |
0 | 2485 } |
2486 } | |
2487 } | |
2488 } | |
127 | 2489 // Recheck with a better notion of 'required_outcnt' |
2490 if (n->outcnt() != required_outcnt) { | |
0 | 2491 record_method_not_compilable("malformed control flow"); |
2492 return true; // Not all targets reachable! | |
2493 } | |
2494 } | |
2495 // Check that I actually visited all kids. Unreached kids | |
2496 // must be infinite loops. | |
2497 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) | |
2498 if (!fpu._visited.test(n->fast_out(j)->_idx)) { | |
2499 record_method_not_compilable("infinite loop"); | |
2500 return true; // Found unvisited kid; must be unreach | |
2501 } | |
2502 } | |
2503 | |
2504 // If original bytecodes contained a mixture of floats and doubles | |
2505 // check if the optimizer has made it homogenous, item (3). | |
2506 if( Use24BitFPMode && Use24BitFP && | |
2507 fpu.get_float_count() > 32 && | |
2508 fpu.get_double_count() == 0 && | |
2509 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { | |
2510 set_24_bit_selection_and_mode( false, true ); | |
2511 } | |
2512 | |
2513 set_has_java_calls(fpu.get_java_call_count() > 0); | |
2514 | |
2515 // No infinite loops, no reason to bail out. | |
2516 return false; | |
2517 } | |
2518 | |
2519 //-----------------------------too_many_traps---------------------------------- | |
2520 // Report if there are too many traps at the current method and bci. | |
2521 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. | |
2522 bool Compile::too_many_traps(ciMethod* method, | |
2523 int bci, | |
2524 Deoptimization::DeoptReason reason) { | |
2525 ciMethodData* md = method->method_data(); | |
2526 if (md->is_empty()) { | |
2527 // Assume the trap has not occurred, or that it occurred only | |
2528 // because of a transient condition during start-up in the interpreter. | |
2529 return false; | |
2530 } | |
2531 if (md->has_trap_at(bci, reason) != 0) { | |
2532 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. | |
2533 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
2534 // assume the worst. | |
2535 if (log()) | |
2536 log()->elem("observe trap='%s' count='%d'", | |
2537 Deoptimization::trap_reason_name(reason), | |
2538 md->trap_count(reason)); | |
2539 return true; | |
2540 } else { | |
2541 // Ignore method/bci and see if there have been too many globally. | |
2542 return too_many_traps(reason, md); | |
2543 } | |
2544 } | |
2545 | |
2546 // Less-accurate variant which does not require a method and bci. | |
2547 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, | |
2548 ciMethodData* logmd) { | |
2549 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { | |
2550 // Too many traps globally. | |
2551 // Note that we use cumulative trap_count, not just md->trap_count. | |
2552 if (log()) { | |
2553 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); | |
2554 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", | |
2555 Deoptimization::trap_reason_name(reason), | |
2556 mcount, trap_count(reason)); | |
2557 } | |
2558 return true; | |
2559 } else { | |
2560 // The coast is clear. | |
2561 return false; | |
2562 } | |
2563 } | |
2564 | |
2565 //--------------------------too_many_recompiles-------------------------------- | |
2566 // Report if there are too many recompiles at the current method and bci. | |
2567 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. | |
2568 // Is not eager to return true, since this will cause the compiler to use | |
2569 // Action_none for a trap point, to avoid too many recompilations. | |
2570 bool Compile::too_many_recompiles(ciMethod* method, | |
2571 int bci, | |
2572 Deoptimization::DeoptReason reason) { | |
2573 ciMethodData* md = method->method_data(); | |
2574 if (md->is_empty()) { | |
2575 // Assume the trap has not occurred, or that it occurred only | |
2576 // because of a transient condition during start-up in the interpreter. | |
2577 return false; | |
2578 } | |
2579 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. | |
2580 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; | |
2581 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero | |
2582 Deoptimization::DeoptReason per_bc_reason | |
2583 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); | |
2584 if ((per_bc_reason == Deoptimization::Reason_none | |
2585 || md->has_trap_at(bci, reason) != 0) | |
2586 // The trap frequency measure we care about is the recompile count: | |
2587 && md->trap_recompiled_at(bci) | |
2588 && md->overflow_recompile_count() >= bc_cutoff) { | |
2589 // Do not emit a trap here if it has already caused recompilations. | |
2590 // Also, if there are multiple reasons, or if there is no per-BCI record, | |
2591 // assume the worst. | |
2592 if (log()) | |
2593 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", | |
2594 Deoptimization::trap_reason_name(reason), | |
2595 md->trap_count(reason), | |
2596 md->overflow_recompile_count()); | |
2597 return true; | |
2598 } else if (trap_count(reason) != 0 | |
2599 && decompile_count() >= m_cutoff) { | |
2600 // Too many recompiles globally, and we have seen this sort of trap. | |
2601 // Use cumulative decompile_count, not just md->decompile_count. | |
2602 if (log()) | |
2603 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", | |
2604 Deoptimization::trap_reason_name(reason), | |
2605 md->trap_count(reason), trap_count(reason), | |
2606 md->decompile_count(), decompile_count()); | |
2607 return true; | |
2608 } else { | |
2609 // The coast is clear. | |
2610 return false; | |
2611 } | |
2612 } | |
2613 | |
2614 | |
2615 #ifndef PRODUCT | |
2616 //------------------------------verify_graph_edges--------------------------- | |
2617 // Walk the Graph and verify that there is a one-to-one correspondence | |
2618 // between Use-Def edges and Def-Use edges in the graph. | |
2619 void Compile::verify_graph_edges(bool no_dead_code) { | |
2620 if (VerifyGraphEdges) { | |
2621 ResourceArea *area = Thread::current()->resource_area(); | |
2622 Unique_Node_List visited(area); | |
2623 // Call recursive graph walk to check edges | |
2624 _root->verify_edges(visited); | |
2625 if (no_dead_code) { | |
2626 // Now make sure that no visited node is used by an unvisited node. | |
2627 bool dead_nodes = 0; | |
2628 Unique_Node_List checked(area); | |
2629 while (visited.size() > 0) { | |
2630 Node* n = visited.pop(); | |
2631 checked.push(n); | |
2632 for (uint i = 0; i < n->outcnt(); i++) { | |
2633 Node* use = n->raw_out(i); | |
2634 if (checked.member(use)) continue; // already checked | |
2635 if (visited.member(use)) continue; // already in the graph | |
2636 if (use->is_Con()) continue; // a dead ConNode is OK | |
2637 // At this point, we have found a dead node which is DU-reachable. | |
2638 if (dead_nodes++ == 0) | |
2639 tty->print_cr("*** Dead nodes reachable via DU edges:"); | |
2640 use->dump(2); | |
2641 tty->print_cr("---"); | |
2642 checked.push(use); // No repeats; pretend it is now checked. | |
2643 } | |
2644 } | |
2645 assert(dead_nodes == 0, "using nodes must be reachable from root"); | |
2646 } | |
2647 } | |
2648 } | |
2649 #endif | |
2650 | |
2651 // The Compile object keeps track of failure reasons separately from the ciEnv. | |
2652 // This is required because there is not quite a 1-1 relation between the | |
2653 // ciEnv and its compilation task and the Compile object. Note that one | |
2654 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides | |
2655 // to backtrack and retry without subsuming loads. Other than this backtracking | |
2656 // behavior, the Compile's failure reason is quietly copied up to the ciEnv | |
2657 // by the logic in C2Compiler. | |
2658 void Compile::record_failure(const char* reason) { | |
2659 if (log() != NULL) { | |
2660 log()->elem("failure reason='%s' phase='compile'", reason); | |
2661 } | |
2662 if (_failure_reason == NULL) { | |
2663 // Record the first failure reason. | |
2664 _failure_reason = reason; | |
2665 } | |
222 | 2666 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { |
2667 C->print_method(_failure_reason); | |
2668 } | |
0 | 2669 _root = NULL; // flush the graph, too |
2670 } | |
2671 | |
2672 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) | |
2673 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) | |
2674 { | |
2675 if (dolog) { | |
2676 C = Compile::current(); | |
2677 _log = C->log(); | |
2678 } else { | |
2679 C = NULL; | |
2680 _log = NULL; | |
2681 } | |
2682 if (_log != NULL) { | |
2683 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); | |
2684 _log->stamp(); | |
2685 _log->end_head(); | |
2686 } | |
2687 } | |
2688 | |
2689 Compile::TracePhase::~TracePhase() { | |
2690 if (_log != NULL) { | |
2691 _log->done("phase nodes='%d'", C->unique()); | |
2692 } | |
2693 } |