Print this page
rev 2240 : [mq]: code-review-comments-tom
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/opto/compile.cpp
+++ new/src/share/vm/opto/compile.cpp
1 1 /*
2 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "asm/assembler.hpp"
27 27 #include "classfile/systemDictionary.hpp"
28 28 #include "code/exceptionHandlerTable.hpp"
29 29 #include "code/nmethod.hpp"
30 30 #include "compiler/compileLog.hpp"
31 31 #include "compiler/oopMap.hpp"
32 32 #include "opto/addnode.hpp"
33 33 #include "opto/block.hpp"
34 34 #include "opto/c2compiler.hpp"
35 35 #include "opto/callGenerator.hpp"
36 36 #include "opto/callnode.hpp"
37 37 #include "opto/cfgnode.hpp"
38 38 #include "opto/chaitin.hpp"
39 39 #include "opto/compile.hpp"
40 40 #include "opto/connode.hpp"
41 41 #include "opto/divnode.hpp"
42 42 #include "opto/escape.hpp"
43 43 #include "opto/idealGraphPrinter.hpp"
44 44 #include "opto/loopnode.hpp"
45 45 #include "opto/machnode.hpp"
46 46 #include "opto/macro.hpp"
47 47 #include "opto/matcher.hpp"
48 48 #include "opto/memnode.hpp"
49 49 #include "opto/mulnode.hpp"
50 50 #include "opto/node.hpp"
51 51 #include "opto/opcodes.hpp"
52 52 #include "opto/output.hpp"
53 53 #include "opto/parse.hpp"
54 54 #include "opto/phaseX.hpp"
55 55 #include "opto/rootnode.hpp"
56 56 #include "opto/runtime.hpp"
57 57 #include "opto/stringopts.hpp"
58 58 #include "opto/type.hpp"
59 59 #include "opto/vectornode.hpp"
60 60 #include "runtime/arguments.hpp"
61 61 #include "runtime/signature.hpp"
62 62 #include "runtime/stubRoutines.hpp"
63 63 #include "runtime/timer.hpp"
64 64 #include "utilities/copy.hpp"
65 65 #ifdef TARGET_ARCH_MODEL_x86_32
66 66 # include "adfiles/ad_x86_32.hpp"
67 67 #endif
68 68 #ifdef TARGET_ARCH_MODEL_x86_64
69 69 # include "adfiles/ad_x86_64.hpp"
70 70 #endif
71 71 #ifdef TARGET_ARCH_MODEL_sparc
72 72 # include "adfiles/ad_sparc.hpp"
73 73 #endif
74 74 #ifdef TARGET_ARCH_MODEL_zero
75 75 # include "adfiles/ad_zero.hpp"
76 76 #endif
77 77 #ifdef TARGET_ARCH_MODEL_arm
78 78 # include "adfiles/ad_arm.hpp"
79 79 #endif
80 80 #ifdef TARGET_ARCH_MODEL_ppc
81 81 # include "adfiles/ad_ppc.hpp"
82 82 #endif
83 83
84 84
85 85 // -------------------- Compile::mach_constant_base_node -----------------------
86 86 // Constant table base node singleton.
87 87 MachConstantBaseNode* Compile::mach_constant_base_node() {
88 88 if (_mach_constant_base_node == NULL) {
89 89 _mach_constant_base_node = new (C) MachConstantBaseNode();
90 90 _mach_constant_base_node->add_req(C->root());
91 91 }
92 92 return _mach_constant_base_node;
93 93 }
94 94
95 95
96 96 /// Support for intrinsics.
97 97
98 98 // Return the index at which m must be inserted (or already exists).
99 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
100 100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
101 101 #ifdef ASSERT
102 102 for (int i = 1; i < _intrinsics->length(); i++) {
103 103 CallGenerator* cg1 = _intrinsics->at(i-1);
104 104 CallGenerator* cg2 = _intrinsics->at(i);
105 105 assert(cg1->method() != cg2->method()
106 106 ? cg1->method() < cg2->method()
107 107 : cg1->is_virtual() < cg2->is_virtual(),
108 108 "compiler intrinsics list must stay sorted");
109 109 }
110 110 #endif
111 111 // Binary search sorted list, in decreasing intervals [lo, hi].
112 112 int lo = 0, hi = _intrinsics->length()-1;
113 113 while (lo <= hi) {
114 114 int mid = (uint)(hi + lo) / 2;
115 115 ciMethod* mid_m = _intrinsics->at(mid)->method();
116 116 if (m < mid_m) {
117 117 hi = mid-1;
118 118 } else if (m > mid_m) {
119 119 lo = mid+1;
120 120 } else {
121 121 // look at minor sort key
122 122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
123 123 if (is_virtual < mid_virt) {
124 124 hi = mid-1;
125 125 } else if (is_virtual > mid_virt) {
126 126 lo = mid+1;
127 127 } else {
128 128 return mid; // exact match
129 129 }
130 130 }
131 131 }
132 132 return lo; // inexact match
133 133 }
134 134
135 135 void Compile::register_intrinsic(CallGenerator* cg) {
136 136 if (_intrinsics == NULL) {
137 137 _intrinsics = new GrowableArray<CallGenerator*>(60);
138 138 }
139 139 // This code is stolen from ciObjectFactory::insert.
140 140 // Really, GrowableArray should have methods for
141 141 // insert_at, remove_at, and binary_search.
142 142 int len = _intrinsics->length();
143 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
144 144 if (index == len) {
145 145 _intrinsics->append(cg);
146 146 } else {
147 147 #ifdef ASSERT
148 148 CallGenerator* oldcg = _intrinsics->at(index);
149 149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
150 150 #endif
151 151 _intrinsics->append(_intrinsics->at(len-1));
152 152 int pos;
153 153 for (pos = len-2; pos >= index; pos--) {
154 154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
155 155 }
156 156 _intrinsics->at_put(index, cg);
157 157 }
158 158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
159 159 }
160 160
161 161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
162 162 assert(m->is_loaded(), "don't try this on unloaded methods");
163 163 if (_intrinsics != NULL) {
164 164 int index = intrinsic_insertion_index(m, is_virtual);
165 165 if (index < _intrinsics->length()
166 166 && _intrinsics->at(index)->method() == m
167 167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
168 168 return _intrinsics->at(index);
169 169 }
170 170 }
171 171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
172 172 if (m->intrinsic_id() != vmIntrinsics::_none &&
173 173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
174 174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
175 175 if (cg != NULL) {
176 176 // Save it for next time:
177 177 register_intrinsic(cg);
178 178 return cg;
179 179 } else {
180 180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
181 181 }
182 182 }
183 183 return NULL;
184 184 }
185 185
186 186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
187 187 // in library_call.cpp.
188 188
189 189
190 190 #ifndef PRODUCT
191 191 // statistics gathering...
192 192
193 193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
194 194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
195 195
196 196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
197 197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
198 198 int oflags = _intrinsic_hist_flags[id];
199 199 assert(flags != 0, "what happened?");
200 200 if (is_virtual) {
201 201 flags |= _intrinsic_virtual;
202 202 }
203 203 bool changed = (flags != oflags);
204 204 if ((flags & _intrinsic_worked) != 0) {
205 205 juint count = (_intrinsic_hist_count[id] += 1);
206 206 if (count == 1) {
207 207 changed = true; // first time
208 208 }
209 209 // increment the overall count also:
210 210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
211 211 }
212 212 if (changed) {
213 213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
214 214 // Something changed about the intrinsic's virtuality.
215 215 if ((flags & _intrinsic_virtual) != 0) {
216 216 // This is the first use of this intrinsic as a virtual call.
217 217 if (oflags != 0) {
218 218 // We already saw it as a non-virtual, so note both cases.
219 219 flags |= _intrinsic_both;
220 220 }
221 221 } else if ((oflags & _intrinsic_both) == 0) {
222 222 // This is the first use of this intrinsic as a non-virtual
223 223 flags |= _intrinsic_both;
224 224 }
225 225 }
226 226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
227 227 }
228 228 // update the overall flags also:
229 229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
230 230 return changed;
231 231 }
232 232
233 233 static char* format_flags(int flags, char* buf) {
234 234 buf[0] = 0;
235 235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
236 236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
237 237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
238 238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
239 239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
240 240 if (buf[0] == 0) strcat(buf, ",");
241 241 assert(buf[0] == ',', "must be");
242 242 return &buf[1];
243 243 }
244 244
245 245 void Compile::print_intrinsic_statistics() {
246 246 char flagsbuf[100];
247 247 ttyLocker ttyl;
248 248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
249 249 tty->print_cr("Compiler intrinsic usage:");
250 250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
251 251 if (total == 0) total = 1; // avoid div0 in case of no successes
252 252 #define PRINT_STAT_LINE(name, c, f) \
253 253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
254 254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
255 255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
256 256 int flags = _intrinsic_hist_flags[id];
257 257 juint count = _intrinsic_hist_count[id];
258 258 if ((flags | count) != 0) {
259 259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
260 260 }
261 261 }
262 262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
263 263 if (xtty != NULL) xtty->tail("statistics");
264 264 }
265 265
266 266 void Compile::print_statistics() {
267 267 { ttyLocker ttyl;
268 268 if (xtty != NULL) xtty->head("statistics type='opto'");
269 269 Parse::print_statistics();
270 270 PhaseCCP::print_statistics();
271 271 PhaseRegAlloc::print_statistics();
272 272 Scheduling::print_statistics();
273 273 PhasePeephole::print_statistics();
274 274 PhaseIdealLoop::print_statistics();
275 275 if (xtty != NULL) xtty->tail("statistics");
276 276 }
277 277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
278 278 // put this under its own <statistics> element.
279 279 print_intrinsic_statistics();
280 280 }
281 281 }
282 282 #endif //PRODUCT
283 283
284 284 // Support for bundling info
285 285 Bundle* Compile::node_bundling(const Node *n) {
286 286 assert(valid_bundle_info(n), "oob");
287 287 return &_node_bundling_base[n->_idx];
288 288 }
289 289
290 290 bool Compile::valid_bundle_info(const Node *n) {
291 291 return (_node_bundling_limit > n->_idx);
292 292 }
293 293
294 294
295 295 void Compile::gvn_replace_by(Node* n, Node* nn) {
296 296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
297 297 Node* use = n->last_out(i);
298 298 bool is_in_table = initial_gvn()->hash_delete(use);
299 299 uint uses_found = 0;
300 300 for (uint j = 0; j < use->len(); j++) {
301 301 if (use->in(j) == n) {
302 302 if (j < use->req())
303 303 use->set_req(j, nn);
304 304 else
305 305 use->set_prec(j, nn);
306 306 uses_found++;
307 307 }
308 308 }
309 309 if (is_in_table) {
310 310 // reinsert into table
311 311 initial_gvn()->hash_find_insert(use);
312 312 }
313 313 record_for_igvn(use);
314 314 i -= uses_found; // we deleted 1 or more copies of this edge
315 315 }
316 316 }
317 317
318 318
319 319
320 320
321 321 // Identify all nodes that are reachable from below, useful.
322 322 // Use breadth-first pass that records state in a Unique_Node_List,
323 323 // recursive traversal is slower.
324 324 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
325 325 int estimated_worklist_size = unique();
326 326 useful.map( estimated_worklist_size, NULL ); // preallocate space
327 327
328 328 // Initialize worklist
329 329 if (root() != NULL) { useful.push(root()); }
330 330 // If 'top' is cached, declare it useful to preserve cached node
331 331 if( cached_top_node() ) { useful.push(cached_top_node()); }
332 332
333 333 // Push all useful nodes onto the list, breadthfirst
334 334 for( uint next = 0; next < useful.size(); ++next ) {
335 335 assert( next < unique(), "Unique useful nodes < total nodes");
336 336 Node *n = useful.at(next);
337 337 uint max = n->len();
338 338 for( uint i = 0; i < max; ++i ) {
339 339 Node *m = n->in(i);
340 340 if( m == NULL ) continue;
341 341 useful.push(m);
342 342 }
343 343 }
344 344 }
345 345
346 346 // Disconnect all useless nodes by disconnecting those at the boundary.
347 347 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
348 348 uint next = 0;
349 349 while( next < useful.size() ) {
350 350 Node *n = useful.at(next++);
351 351 // Use raw traversal of out edges since this code removes out edges
352 352 int max = n->outcnt();
353 353 for (int j = 0; j < max; ++j ) {
354 354 Node* child = n->raw_out(j);
355 355 if( ! useful.member(child) ) {
356 356 assert( !child->is_top() || child != top(),
357 357 "If top is cached in Compile object it is in useful list");
358 358 // Only need to remove this out-edge to the useless node
359 359 n->raw_del_out(j);
360 360 --j;
361 361 --max;
362 362 }
363 363 }
364 364 if (n->outcnt() == 1 && n->has_special_unique_user()) {
365 365 record_for_igvn( n->unique_out() );
366 366 }
367 367 }
368 368 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
369 369 }
370 370
371 371 //------------------------------frame_size_in_words-----------------------------
372 372 // frame_slots in units of words
373 373 int Compile::frame_size_in_words() const {
374 374 // shift is 0 in LP32 and 1 in LP64
375 375 const int shift = (LogBytesPerWord - LogBytesPerInt);
376 376 int words = _frame_slots >> shift;
377 377 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
378 378 return words;
379 379 }
380 380
381 381 // ============================================================================
382 382 //------------------------------CompileWrapper---------------------------------
383 383 class CompileWrapper : public StackObj {
384 384 Compile *const _compile;
385 385 public:
386 386 CompileWrapper(Compile* compile);
387 387
388 388 ~CompileWrapper();
389 389 };
390 390
391 391 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
392 392 // the Compile* pointer is stored in the current ciEnv:
393 393 ciEnv* env = compile->env();
394 394 assert(env == ciEnv::current(), "must already be a ciEnv active");
395 395 assert(env->compiler_data() == NULL, "compile already active?");
396 396 env->set_compiler_data(compile);
397 397 assert(compile == Compile::current(), "sanity");
398 398
399 399 compile->set_type_dict(NULL);
400 400 compile->set_type_hwm(NULL);
401 401 compile->set_type_last_size(0);
402 402 compile->set_last_tf(NULL, NULL);
403 403 compile->set_indexSet_arena(NULL);
404 404 compile->set_indexSet_free_block_list(NULL);
405 405 compile->init_type_arena();
406 406 Type::Initialize(compile);
407 407 _compile->set_scratch_buffer_blob(NULL);
408 408 _compile->begin_method();
409 409 }
410 410 CompileWrapper::~CompileWrapper() {
411 411 _compile->end_method();
412 412 if (_compile->scratch_buffer_blob() != NULL)
413 413 BufferBlob::free(_compile->scratch_buffer_blob());
414 414 _compile->env()->set_compiler_data(NULL);
415 415 }
416 416
417 417
418 418 //----------------------------print_compile_messages---------------------------
419 419 void Compile::print_compile_messages() {
420 420 #ifndef PRODUCT
421 421 // Check if recompiling
422 422 if (_subsume_loads == false && PrintOpto) {
423 423 // Recompiling without allowing machine instructions to subsume loads
424 424 tty->print_cr("*********************************************************");
425 425 tty->print_cr("** Bailout: Recompile without subsuming loads **");
426 426 tty->print_cr("*********************************************************");
427 427 }
428 428 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
429 429 // Recompiling without escape analysis
430 430 tty->print_cr("*********************************************************");
431 431 tty->print_cr("** Bailout: Recompile without escape analysis **");
432 432 tty->print_cr("*********************************************************");
433 433 }
434 434 if (env()->break_at_compile()) {
435 435 // Open the debugger when compiling this method.
436 436 tty->print("### Breaking when compiling: ");
437 437 method()->print_short_name();
438 438 tty->cr();
439 439 BREAKPOINT;
440 440 }
441 441
442 442 if( PrintOpto ) {
443 443 if (is_osr_compilation()) {
444 444 tty->print("[OSR]%3d", _compile_id);
445 445 } else {
446 446 tty->print("%3d", _compile_id);
447 447 }
448 448 }
449 449 #endif
450 450 }
451 451
452 452
453 453 //-----------------------init_scratch_buffer_blob------------------------------
454 454 // Construct a temporary BufferBlob and cache it for this compile.
455 455 void Compile::init_scratch_buffer_blob(int const_size) {
456 456 // If there is already a scratch buffer blob allocated and the
457 457 // constant section is big enough, use it. Otherwise free the
458 458 // current and allocate a new one.
459 459 BufferBlob* blob = scratch_buffer_blob();
460 460 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
461 461 // Use the current blob.
462 462 } else {
463 463 if (blob != NULL) {
464 464 BufferBlob::free(blob);
465 465 }
466 466
467 467 ResourceMark rm;
468 468 _scratch_const_size = const_size;
469 469 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
470 470 blob = BufferBlob::create("Compile::scratch_buffer", size);
471 471 // Record the buffer blob for next time.
472 472 set_scratch_buffer_blob(blob);
473 473 // Have we run out of code space?
474 474 if (scratch_buffer_blob() == NULL) {
475 475 // Let CompilerBroker disable further compilations.
476 476 record_failure("Not enough space for scratch buffer in CodeCache");
477 477 return;
478 478 }
479 479 }
480 480
481 481 // Initialize the relocation buffers
482 482 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
483 483 set_scratch_locs_memory(locs_buf);
484 484 }
485 485
486 486
487 487 //-----------------------scratch_emit_size-------------------------------------
488 488 // Helper function that computes size by emitting code
489 489 uint Compile::scratch_emit_size(const Node* n) {
490 490 // Start scratch_emit_size section.
491 491 set_in_scratch_emit_size(true);
492 492
493 493 // Emit into a trash buffer and count bytes emitted.
494 494 // This is a pretty expensive way to compute a size,
495 495 // but it works well enough if seldom used.
496 496 // All common fixed-size instructions are given a size
497 497 // method by the AD file.
498 498 // Note that the scratch buffer blob and locs memory are
499 499 // allocated at the beginning of the compile task, and
500 500 // may be shared by several calls to scratch_emit_size.
501 501 // The allocation of the scratch buffer blob is particularly
502 502 // expensive, since it has to grab the code cache lock.
503 503 BufferBlob* blob = this->scratch_buffer_blob();
504 504 assert(blob != NULL, "Initialize BufferBlob at start");
505 505 assert(blob->size() > MAX_inst_size, "sanity");
506 506 relocInfo* locs_buf = scratch_locs_memory();
507 507 address blob_begin = blob->content_begin();
508 508 address blob_end = (address)locs_buf;
509 509 assert(blob->content_contains(blob_end), "sanity");
510 510 CodeBuffer buf(blob_begin, blob_end - blob_begin);
511 511 buf.initialize_consts_size(_scratch_const_size);
512 512 buf.initialize_stubs_size(MAX_stubs_size);
513 513 assert(locs_buf != NULL, "sanity");
514 514 int lsize = MAX_locs_size / 3;
515 515 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
516 516 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
517 517 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
518 518
519 519 // Do the emission.
520 520 n->emit(buf, this->regalloc());
521 521
522 522 // End scratch_emit_size section.
523 523 set_in_scratch_emit_size(false);
524 524
525 525 return buf.insts_size();
526 526 }
527 527
528 528
529 529 // ============================================================================
530 530 //------------------------------Compile standard-------------------------------
531 531 debug_only( int Compile::_debug_idx = 100000; )
532 532
533 533 // Compile a method. entry_bci is -1 for normal compilations and indicates
534 534 // the continuation bci for on stack replacement.
535 535
536 536
537 537 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
538 538 : Phase(Compiler),
539 539 _env(ci_env),
540 540 _log(ci_env->log()),
541 541 _compile_id(ci_env->compile_id()),
542 542 _save_argument_registers(false),
543 543 _stub_name(NULL),
544 544 _stub_function(NULL),
545 545 _stub_entry_point(NULL),
546 546 _method(target),
547 547 _entry_bci(osr_bci),
548 548 _initial_gvn(NULL),
549 549 _for_igvn(NULL),
550 550 _warm_calls(NULL),
551 551 _subsume_loads(subsume_loads),
552 552 _do_escape_analysis(do_escape_analysis),
553 553 _failure_reason(NULL),
554 554 _code_buffer("Compile::Fill_buffer"),
555 555 _orig_pc_slot(0),
556 556 _orig_pc_slot_offset_in_bytes(0),
557 557 _has_method_handle_invokes(false),
558 558 _mach_constant_base_node(NULL),
559 559 _node_bundling_limit(0),
560 560 _node_bundling_base(NULL),
561 561 _java_calls(0),
562 562 _inner_loops(0),
563 563 _scratch_const_size(-1),
564 564 _in_scratch_emit_size(false),
565 565 #ifndef PRODUCT
566 566 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
567 567 _printer(IdealGraphPrinter::printer()),
568 568 #endif
569 569 _congraph(NULL) {
570 570 C = this;
571 571
572 572 CompileWrapper cw(this);
573 573 #ifndef PRODUCT
574 574 if (TimeCompiler2) {
575 575 tty->print(" ");
576 576 target->holder()->name()->print();
577 577 tty->print(".");
578 578 target->print_short_name();
579 579 tty->print(" ");
580 580 }
581 581 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
582 582 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
583 583 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
584 584 if (!print_opto_assembly) {
585 585 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
586 586 if (print_assembly && !Disassembler::can_decode()) {
587 587 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
588 588 print_opto_assembly = true;
589 589 }
590 590 }
591 591 set_print_assembly(print_opto_assembly);
592 592 set_parsed_irreducible_loop(false);
593 593 #endif
594 594
595 595 if (ProfileTraps) {
596 596 // Make sure the method being compiled gets its own MDO,
597 597 // so we can at least track the decompile_count().
598 598 method()->ensure_method_data();
599 599 }
600 600
601 601 Init(::AliasLevel);
602 602
603 603
604 604 print_compile_messages();
605 605
606 606 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
607 607 _ilt = InlineTree::build_inline_tree_root();
608 608 else
609 609 _ilt = NULL;
610 610
611 611 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
612 612 assert(num_alias_types() >= AliasIdxRaw, "");
613 613
614 614 #define MINIMUM_NODE_HASH 1023
615 615 // Node list that Iterative GVN will start with
616 616 Unique_Node_List for_igvn(comp_arena());
617 617 set_for_igvn(&for_igvn);
618 618
619 619 // GVN that will be run immediately on new nodes
620 620 uint estimated_size = method()->code_size()*4+64;
621 621 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
↓ open down ↓ |
621 lines elided |
↑ open up ↑ |
622 622 PhaseGVN gvn(node_arena(), estimated_size);
623 623 set_initial_gvn(&gvn);
624 624
625 625 { // Scope for timing the parser
626 626 TracePhase t3("parse", &_t_parser, true);
627 627
628 628 // Put top into the hash table ASAP.
629 629 initial_gvn()->transform_no_reclaim(top());
630 630
631 631 // Set up tf(), start(), and find a CallGenerator.
632 - CallGenerator* cg;
632 + CallGenerator* cg = NULL;
633 633 if (is_osr_compilation()) {
634 634 const TypeTuple *domain = StartOSRNode::osr_domain();
635 635 const TypeTuple *range = TypeTuple::make_range(method()->signature());
636 636 init_tf(TypeFunc::make(domain, range));
637 637 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
638 638 initial_gvn()->set_type_bottom(s);
639 639 init_start(s);
640 640 cg = CallGenerator::for_osr(method(), entry_bci());
641 641 } else {
642 642 // Normal case.
643 643 init_tf(TypeFunc::make(method()));
644 644 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
645 645 initial_gvn()->set_type_bottom(s);
646 646 init_start(s);
647 - float past_uses = method()->interpreter_invocation_count();
648 - float expected_uses = past_uses;
649 - cg = CallGenerator::for_inline(method(), expected_uses);
647 + if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
648 + // With java.lang.ref.reference.get() we must go through the
649 + // intrinsic when G1 is enabled - even when get() is the root
650 + // method of the compile - so that, if necessary, the value in
651 + // the referent field of the reference object gets recorded by
652 + // the pre-barrier code.
653 + // Specifically, if G1 is enabled, the value in the referent
654 + // field is recorded by the G1 SATB pre barrier. This will
655 + // result in the referent being marked live and the reference
656 + // object removed from the list of discovered references during
657 + // reference processing.
658 + cg = find_intrinsic(method(), false);
659 + }
660 + if (cg == NULL) {
661 + float past_uses = method()->interpreter_invocation_count();
662 + float expected_uses = past_uses;
663 + cg = CallGenerator::for_inline(method(), expected_uses);
664 + }
650 665 }
651 666 if (failing()) return;
652 667 if (cg == NULL) {
653 668 record_method_not_compilable_all_tiers("cannot parse method");
654 669 return;
655 670 }
656 671 JVMState* jvms = build_start_state(start(), tf());
657 672 if ((jvms = cg->generate(jvms)) == NULL) {
658 673 record_method_not_compilable("method parse failed");
659 674 return;
660 675 }
661 676 GraphKit kit(jvms);
662 677
663 678 if (!kit.stopped()) {
664 679 // Accept return values, and transfer control we know not where.
665 680 // This is done by a special, unique ReturnNode bound to root.
666 681 return_values(kit.jvms());
667 682 }
668 683
669 684 if (kit.has_exceptions()) {
670 685 // Any exceptions that escape from this call must be rethrown
671 686 // to whatever caller is dynamically above us on the stack.
672 687 // This is done by a special, unique RethrowNode bound to root.
673 688 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
674 689 }
675 690
676 691 if (!failing() && has_stringbuilder()) {
677 692 {
678 693 // remove useless nodes to make the usage analysis simpler
679 694 ResourceMark rm;
680 695 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
681 696 }
682 697
683 698 {
684 699 ResourceMark rm;
685 700 print_method("Before StringOpts", 3);
686 701 PhaseStringOpts pso(initial_gvn(), &for_igvn);
687 702 print_method("After StringOpts", 3);
688 703 }
689 704
690 705 // now inline anything that we skipped the first time around
691 706 while (_late_inlines.length() > 0) {
692 707 CallGenerator* cg = _late_inlines.pop();
693 708 cg->do_late_inline();
694 709 }
695 710 }
696 711 assert(_late_inlines.length() == 0, "should have been processed");
697 712
698 713 print_method("Before RemoveUseless", 3);
699 714
700 715 // Remove clutter produced by parsing.
701 716 if (!failing()) {
702 717 ResourceMark rm;
703 718 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
704 719 }
705 720 }
706 721
707 722 // Note: Large methods are capped off in do_one_bytecode().
708 723 if (failing()) return;
709 724
710 725 // After parsing, node notes are no longer automagic.
711 726 // They must be propagated by register_new_node_with_optimizer(),
712 727 // clone(), or the like.
713 728 set_default_node_notes(NULL);
714 729
715 730 for (;;) {
716 731 int successes = Inline_Warm();
717 732 if (failing()) return;
718 733 if (successes == 0) break;
719 734 }
720 735
721 736 // Drain the list.
722 737 Finish_Warm();
723 738 #ifndef PRODUCT
724 739 if (_printer) {
725 740 _printer->print_inlining(this);
726 741 }
727 742 #endif
728 743
729 744 if (failing()) return;
730 745 NOT_PRODUCT( verify_graph_edges(); )
731 746
732 747 // Now optimize
733 748 Optimize();
734 749 if (failing()) return;
735 750 NOT_PRODUCT( verify_graph_edges(); )
736 751
737 752 #ifndef PRODUCT
738 753 if (PrintIdeal) {
739 754 ttyLocker ttyl; // keep the following output all in one block
740 755 // This output goes directly to the tty, not the compiler log.
741 756 // To enable tools to match it up with the compilation activity,
742 757 // be sure to tag this tty output with the compile ID.
743 758 if (xtty != NULL) {
744 759 xtty->head("ideal compile_id='%d'%s", compile_id(),
745 760 is_osr_compilation() ? " compile_kind='osr'" :
746 761 "");
747 762 }
748 763 root()->dump(9999);
749 764 if (xtty != NULL) {
750 765 xtty->tail("ideal");
751 766 }
752 767 }
753 768 #endif
754 769
755 770 // Now that we know the size of all the monitors we can add a fixed slot
756 771 // for the original deopt pc.
757 772
758 773 _orig_pc_slot = fixed_slots();
759 774 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
760 775 set_fixed_slots(next_slot);
761 776
762 777 // Now generate code
763 778 Code_Gen();
764 779 if (failing()) return;
765 780
766 781 // Check if we want to skip execution of all compiled code.
767 782 {
768 783 #ifndef PRODUCT
769 784 if (OptoNoExecute) {
770 785 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
771 786 return;
772 787 }
773 788 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
774 789 #endif
775 790
776 791 if (is_osr_compilation()) {
777 792 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
778 793 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
779 794 } else {
780 795 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
781 796 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
782 797 }
783 798
784 799 env()->register_method(_method, _entry_bci,
785 800 &_code_offsets,
786 801 _orig_pc_slot_offset_in_bytes,
787 802 code_buffer(),
788 803 frame_size_in_words(), _oop_map_set,
789 804 &_handler_table, &_inc_table,
790 805 compiler,
791 806 env()->comp_level(),
792 807 true, /*has_debug_info*/
793 808 has_unsafe_access()
794 809 );
795 810 }
796 811 }
797 812
798 813 //------------------------------Compile----------------------------------------
799 814 // Compile a runtime stub
800 815 Compile::Compile( ciEnv* ci_env,
801 816 TypeFunc_generator generator,
802 817 address stub_function,
803 818 const char *stub_name,
804 819 int is_fancy_jump,
805 820 bool pass_tls,
806 821 bool save_arg_registers,
807 822 bool return_pc )
808 823 : Phase(Compiler),
809 824 _env(ci_env),
810 825 _log(ci_env->log()),
811 826 _compile_id(-1),
812 827 _save_argument_registers(save_arg_registers),
813 828 _method(NULL),
814 829 _stub_name(stub_name),
815 830 _stub_function(stub_function),
816 831 _stub_entry_point(NULL),
817 832 _entry_bci(InvocationEntryBci),
818 833 _initial_gvn(NULL),
819 834 _for_igvn(NULL),
820 835 _warm_calls(NULL),
821 836 _orig_pc_slot(0),
822 837 _orig_pc_slot_offset_in_bytes(0),
823 838 _subsume_loads(true),
824 839 _do_escape_analysis(false),
825 840 _failure_reason(NULL),
826 841 _code_buffer("Compile::Fill_buffer"),
827 842 _has_method_handle_invokes(false),
828 843 _mach_constant_base_node(NULL),
829 844 _node_bundling_limit(0),
830 845 _node_bundling_base(NULL),
831 846 _java_calls(0),
832 847 _inner_loops(0),
833 848 #ifndef PRODUCT
834 849 _trace_opto_output(TraceOptoOutput),
835 850 _printer(NULL),
836 851 #endif
837 852 _congraph(NULL) {
838 853 C = this;
839 854
840 855 #ifndef PRODUCT
841 856 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
842 857 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
843 858 set_print_assembly(PrintFrameConverterAssembly);
844 859 set_parsed_irreducible_loop(false);
845 860 #endif
846 861 CompileWrapper cw(this);
847 862 Init(/*AliasLevel=*/ 0);
848 863 init_tf((*generator)());
849 864
850 865 {
851 866 // The following is a dummy for the sake of GraphKit::gen_stub
852 867 Unique_Node_List for_igvn(comp_arena());
853 868 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
854 869 PhaseGVN gvn(Thread::current()->resource_area(),255);
855 870 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
856 871 gvn.transform_no_reclaim(top());
857 872
858 873 GraphKit kit;
859 874 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
860 875 }
861 876
862 877 NOT_PRODUCT( verify_graph_edges(); )
863 878 Code_Gen();
864 879 if (failing()) return;
865 880
866 881
867 882 // Entry point will be accessed using compile->stub_entry_point();
868 883 if (code_buffer() == NULL) {
869 884 Matcher::soft_match_failure();
870 885 } else {
871 886 if (PrintAssembly && (WizardMode || Verbose))
872 887 tty->print_cr("### Stub::%s", stub_name);
873 888
874 889 if (!failing()) {
875 890 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
876 891
877 892 // Make the NMethod
878 893 // For now we mark the frame as never safe for profile stackwalking
879 894 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
880 895 code_buffer(),
881 896 CodeOffsets::frame_never_safe,
882 897 // _code_offsets.value(CodeOffsets::Frame_Complete),
883 898 frame_size_in_words(),
884 899 _oop_map_set,
885 900 save_arg_registers);
886 901 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
887 902
888 903 _stub_entry_point = rs->entry_point();
889 904 }
890 905 }
891 906 }
892 907
893 908 #ifndef PRODUCT
894 909 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
895 910 if(PrintOpto && Verbose) {
896 911 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
897 912 }
898 913 }
899 914 #endif
900 915
901 916 void Compile::print_codes() {
902 917 }
903 918
904 919 //------------------------------Init-------------------------------------------
905 920 // Prepare for a single compilation
906 921 void Compile::Init(int aliaslevel) {
907 922 _unique = 0;
908 923 _regalloc = NULL;
909 924
910 925 _tf = NULL; // filled in later
911 926 _top = NULL; // cached later
912 927 _matcher = NULL; // filled in later
913 928 _cfg = NULL; // filled in later
914 929
915 930 set_24_bit_selection_and_mode(Use24BitFP, false);
916 931
917 932 _node_note_array = NULL;
918 933 _default_node_notes = NULL;
919 934
920 935 _immutable_memory = NULL; // filled in at first inquiry
921 936
922 937 // Globally visible Nodes
923 938 // First set TOP to NULL to give safe behavior during creation of RootNode
924 939 set_cached_top_node(NULL);
925 940 set_root(new (this, 3) RootNode());
926 941 // Now that you have a Root to point to, create the real TOP
927 942 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
928 943 set_recent_alloc(NULL, NULL);
929 944
930 945 // Create Debug Information Recorder to record scopes, oopmaps, etc.
931 946 env()->set_oop_recorder(new OopRecorder(comp_arena()));
932 947 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
933 948 env()->set_dependencies(new Dependencies(env()));
934 949
935 950 _fixed_slots = 0;
936 951 set_has_split_ifs(false);
937 952 set_has_loops(has_method() && method()->has_loops()); // first approximation
938 953 set_has_stringbuilder(false);
939 954 _trap_can_recompile = false; // no traps emitted yet
940 955 _major_progress = true; // start out assuming good things will happen
941 956 set_has_unsafe_access(false);
942 957 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
943 958 set_decompile_count(0);
944 959
945 960 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
946 961 set_num_loop_opts(LoopOptsCount);
947 962 set_do_inlining(Inline);
948 963 set_max_inline_size(MaxInlineSize);
949 964 set_freq_inline_size(FreqInlineSize);
950 965 set_do_scheduling(OptoScheduling);
951 966 set_do_count_invocations(false);
952 967 set_do_method_data_update(false);
953 968
954 969 if (debug_info()->recording_non_safepoints()) {
955 970 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
956 971 (comp_arena(), 8, 0, NULL));
957 972 set_default_node_notes(Node_Notes::make(this));
958 973 }
959 974
960 975 // // -- Initialize types before each compile --
961 976 // // Update cached type information
962 977 // if( _method && _method->constants() )
963 978 // Type::update_loaded_types(_method, _method->constants());
964 979
965 980 // Init alias_type map.
966 981 if (!_do_escape_analysis && aliaslevel == 3)
967 982 aliaslevel = 2; // No unique types without escape analysis
968 983 _AliasLevel = aliaslevel;
969 984 const int grow_ats = 16;
970 985 _max_alias_types = grow_ats;
971 986 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
972 987 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
973 988 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
974 989 {
975 990 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
976 991 }
977 992 // Initialize the first few types.
978 993 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
979 994 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
980 995 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
981 996 _num_alias_types = AliasIdxRaw+1;
982 997 // Zero out the alias type cache.
983 998 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
984 999 // A NULL adr_type hits in the cache right away. Preload the right answer.
985 1000 probe_alias_cache(NULL)->_index = AliasIdxTop;
986 1001
987 1002 _intrinsics = NULL;
988 1003 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
989 1004 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
990 1005 register_library_intrinsics();
991 1006 }
992 1007
993 1008 //---------------------------init_start----------------------------------------
994 1009 // Install the StartNode on this compile object.
995 1010 void Compile::init_start(StartNode* s) {
996 1011 if (failing())
997 1012 return; // already failing
998 1013 assert(s == start(), "");
999 1014 }
1000 1015
1001 1016 StartNode* Compile::start() const {
1002 1017 assert(!failing(), "");
1003 1018 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1004 1019 Node* start = root()->fast_out(i);
1005 1020 if( start->is_Start() )
1006 1021 return start->as_Start();
1007 1022 }
1008 1023 ShouldNotReachHere();
1009 1024 return NULL;
1010 1025 }
1011 1026
1012 1027 //-------------------------------immutable_memory-------------------------------------
1013 1028 // Access immutable memory
1014 1029 Node* Compile::immutable_memory() {
1015 1030 if (_immutable_memory != NULL) {
1016 1031 return _immutable_memory;
1017 1032 }
1018 1033 StartNode* s = start();
1019 1034 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1020 1035 Node *p = s->fast_out(i);
1021 1036 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1022 1037 _immutable_memory = p;
1023 1038 return _immutable_memory;
1024 1039 }
1025 1040 }
1026 1041 ShouldNotReachHere();
1027 1042 return NULL;
1028 1043 }
1029 1044
1030 1045 //----------------------set_cached_top_node------------------------------------
1031 1046 // Install the cached top node, and make sure Node::is_top works correctly.
1032 1047 void Compile::set_cached_top_node(Node* tn) {
1033 1048 if (tn != NULL) verify_top(tn);
1034 1049 Node* old_top = _top;
1035 1050 _top = tn;
1036 1051 // Calling Node::setup_is_top allows the nodes the chance to adjust
1037 1052 // their _out arrays.
1038 1053 if (_top != NULL) _top->setup_is_top();
1039 1054 if (old_top != NULL) old_top->setup_is_top();
1040 1055 assert(_top == NULL || top()->is_top(), "");
1041 1056 }
1042 1057
1043 1058 #ifndef PRODUCT
1044 1059 void Compile::verify_top(Node* tn) const {
1045 1060 if (tn != NULL) {
1046 1061 assert(tn->is_Con(), "top node must be a constant");
1047 1062 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1048 1063 assert(tn->in(0) != NULL, "must have live top node");
1049 1064 }
1050 1065 }
1051 1066 #endif
1052 1067
1053 1068
1054 1069 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1055 1070
1056 1071 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1057 1072 guarantee(arr != NULL, "");
1058 1073 int num_blocks = arr->length();
1059 1074 if (grow_by < num_blocks) grow_by = num_blocks;
1060 1075 int num_notes = grow_by * _node_notes_block_size;
1061 1076 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1062 1077 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1063 1078 while (num_notes > 0) {
1064 1079 arr->append(notes);
1065 1080 notes += _node_notes_block_size;
1066 1081 num_notes -= _node_notes_block_size;
1067 1082 }
1068 1083 assert(num_notes == 0, "exact multiple, please");
1069 1084 }
1070 1085
1071 1086 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1072 1087 if (source == NULL || dest == NULL) return false;
1073 1088
1074 1089 if (dest->is_Con())
1075 1090 return false; // Do not push debug info onto constants.
1076 1091
1077 1092 #ifdef ASSERT
1078 1093 // Leave a bread crumb trail pointing to the original node:
1079 1094 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1080 1095 dest->set_debug_orig(source);
1081 1096 }
1082 1097 #endif
1083 1098
1084 1099 if (node_note_array() == NULL)
1085 1100 return false; // Not collecting any notes now.
1086 1101
1087 1102 // This is a copy onto a pre-existing node, which may already have notes.
1088 1103 // If both nodes have notes, do not overwrite any pre-existing notes.
1089 1104 Node_Notes* source_notes = node_notes_at(source->_idx);
1090 1105 if (source_notes == NULL || source_notes->is_clear()) return false;
1091 1106 Node_Notes* dest_notes = node_notes_at(dest->_idx);
1092 1107 if (dest_notes == NULL || dest_notes->is_clear()) {
1093 1108 return set_node_notes_at(dest->_idx, source_notes);
1094 1109 }
1095 1110
1096 1111 Node_Notes merged_notes = (*source_notes);
1097 1112 // The order of operations here ensures that dest notes will win...
1098 1113 merged_notes.update_from(dest_notes);
1099 1114 return set_node_notes_at(dest->_idx, &merged_notes);
1100 1115 }
1101 1116
1102 1117
1103 1118 //--------------------------allow_range_check_smearing-------------------------
1104 1119 // Gating condition for coalescing similar range checks.
1105 1120 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1106 1121 // single covering check that is at least as strong as any of them.
1107 1122 // If the optimization succeeds, the simplified (strengthened) range check
1108 1123 // will always succeed. If it fails, we will deopt, and then give up
1109 1124 // on the optimization.
1110 1125 bool Compile::allow_range_check_smearing() const {
1111 1126 // If this method has already thrown a range-check,
1112 1127 // assume it was because we already tried range smearing
1113 1128 // and it failed.
1114 1129 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1115 1130 return !already_trapped;
1116 1131 }
1117 1132
1118 1133
1119 1134 //------------------------------flatten_alias_type-----------------------------
1120 1135 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1121 1136 int offset = tj->offset();
1122 1137 TypePtr::PTR ptr = tj->ptr();
1123 1138
1124 1139 // Known instance (scalarizable allocation) alias only with itself.
1125 1140 bool is_known_inst = tj->isa_oopptr() != NULL &&
1126 1141 tj->is_oopptr()->is_known_instance();
1127 1142
1128 1143 // Process weird unsafe references.
1129 1144 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1130 1145 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1131 1146 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1132 1147 tj = TypeOopPtr::BOTTOM;
1133 1148 ptr = tj->ptr();
1134 1149 offset = tj->offset();
1135 1150 }
1136 1151
1137 1152 // Array pointers need some flattening
1138 1153 const TypeAryPtr *ta = tj->isa_aryptr();
1139 1154 if( ta && is_known_inst ) {
1140 1155 if ( offset != Type::OffsetBot &&
1141 1156 offset > arrayOopDesc::length_offset_in_bytes() ) {
1142 1157 offset = Type::OffsetBot; // Flatten constant access into array body only
1143 1158 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1144 1159 }
1145 1160 } else if( ta && _AliasLevel >= 2 ) {
1146 1161 // For arrays indexed by constant indices, we flatten the alias
1147 1162 // space to include all of the array body. Only the header, klass
1148 1163 // and array length can be accessed un-aliased.
1149 1164 if( offset != Type::OffsetBot ) {
1150 1165 if( ta->const_oop() ) { // methodDataOop or methodOop
1151 1166 offset = Type::OffsetBot; // Flatten constant access into array body
1152 1167 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1153 1168 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1154 1169 // range is OK as-is.
1155 1170 tj = ta = TypeAryPtr::RANGE;
1156 1171 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1157 1172 tj = TypeInstPtr::KLASS; // all klass loads look alike
1158 1173 ta = TypeAryPtr::RANGE; // generic ignored junk
1159 1174 ptr = TypePtr::BotPTR;
1160 1175 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1161 1176 tj = TypeInstPtr::MARK;
1162 1177 ta = TypeAryPtr::RANGE; // generic ignored junk
1163 1178 ptr = TypePtr::BotPTR;
1164 1179 } else { // Random constant offset into array body
1165 1180 offset = Type::OffsetBot; // Flatten constant access into array body
1166 1181 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1167 1182 }
1168 1183 }
1169 1184 // Arrays of fixed size alias with arrays of unknown size.
1170 1185 if (ta->size() != TypeInt::POS) {
1171 1186 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1172 1187 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1173 1188 }
1174 1189 // Arrays of known objects become arrays of unknown objects.
1175 1190 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1176 1191 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1177 1192 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1178 1193 }
1179 1194 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1180 1195 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1181 1196 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1182 1197 }
1183 1198 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1184 1199 // cannot be distinguished by bytecode alone.
1185 1200 if (ta->elem() == TypeInt::BOOL) {
1186 1201 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1187 1202 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1188 1203 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1189 1204 }
1190 1205 // During the 2nd round of IterGVN, NotNull castings are removed.
1191 1206 // Make sure the Bottom and NotNull variants alias the same.
1192 1207 // Also, make sure exact and non-exact variants alias the same.
1193 1208 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1194 1209 if (ta->const_oop()) {
1195 1210 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1196 1211 } else {
1197 1212 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1198 1213 }
1199 1214 }
1200 1215 }
1201 1216
1202 1217 // Oop pointers need some flattening
1203 1218 const TypeInstPtr *to = tj->isa_instptr();
1204 1219 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1205 1220 ciInstanceKlass *k = to->klass()->as_instance_klass();
1206 1221 if( ptr == TypePtr::Constant ) {
1207 1222 if (to->klass() != ciEnv::current()->Class_klass() ||
1208 1223 offset < k->size_helper() * wordSize) {
1209 1224 // No constant oop pointers (such as Strings); they alias with
1210 1225 // unknown strings.
1211 1226 assert(!is_known_inst, "not scalarizable allocation");
1212 1227 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1213 1228 }
1214 1229 } else if( is_known_inst ) {
1215 1230 tj = to; // Keep NotNull and klass_is_exact for instance type
1216 1231 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1217 1232 // During the 2nd round of IterGVN, NotNull castings are removed.
1218 1233 // Make sure the Bottom and NotNull variants alias the same.
1219 1234 // Also, make sure exact and non-exact variants alias the same.
1220 1235 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1221 1236 }
1222 1237 // Canonicalize the holder of this field
1223 1238 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1224 1239 // First handle header references such as a LoadKlassNode, even if the
1225 1240 // object's klass is unloaded at compile time (4965979).
1226 1241 if (!is_known_inst) { // Do it only for non-instance types
1227 1242 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1228 1243 }
1229 1244 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1230 1245 // Static fields are in the space above the normal instance
1231 1246 // fields in the java.lang.Class instance.
1232 1247 if (to->klass() != ciEnv::current()->Class_klass()) {
1233 1248 to = NULL;
1234 1249 tj = TypeOopPtr::BOTTOM;
1235 1250 offset = tj->offset();
1236 1251 }
1237 1252 } else {
1238 1253 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1239 1254 if (!k->equals(canonical_holder) || tj->offset() != offset) {
1240 1255 if( is_known_inst ) {
1241 1256 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1242 1257 } else {
1243 1258 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1244 1259 }
1245 1260 }
1246 1261 }
1247 1262 }
1248 1263
1249 1264 // Klass pointers to object array klasses need some flattening
1250 1265 const TypeKlassPtr *tk = tj->isa_klassptr();
1251 1266 if( tk ) {
1252 1267 // If we are referencing a field within a Klass, we need
1253 1268 // to assume the worst case of an Object. Both exact and
1254 1269 // inexact types must flatten to the same alias class.
1255 1270 // Since the flattened result for a klass is defined to be
1256 1271 // precisely java.lang.Object, use a constant ptr.
1257 1272 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1258 1273
1259 1274 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1260 1275 TypeKlassPtr::OBJECT->klass(),
1261 1276 offset);
1262 1277 }
1263 1278
1264 1279 ciKlass* klass = tk->klass();
1265 1280 if( klass->is_obj_array_klass() ) {
1266 1281 ciKlass* k = TypeAryPtr::OOPS->klass();
1267 1282 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
1268 1283 k = TypeInstPtr::BOTTOM->klass();
1269 1284 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1270 1285 }
1271 1286
1272 1287 // Check for precise loads from the primary supertype array and force them
1273 1288 // to the supertype cache alias index. Check for generic array loads from
1274 1289 // the primary supertype array and also force them to the supertype cache
1275 1290 // alias index. Since the same load can reach both, we need to merge
1276 1291 // these 2 disparate memories into the same alias class. Since the
1277 1292 // primary supertype array is read-only, there's no chance of confusion
1278 1293 // where we bypass an array load and an array store.
1279 1294 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1280 1295 if( offset == Type::OffsetBot ||
1281 1296 off2 < Klass::primary_super_limit()*wordSize ) {
1282 1297 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1283 1298 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1284 1299 }
1285 1300 }
1286 1301
1287 1302 // Flatten all Raw pointers together.
1288 1303 if (tj->base() == Type::RawPtr)
1289 1304 tj = TypeRawPtr::BOTTOM;
1290 1305
1291 1306 if (tj->base() == Type::AnyPtr)
1292 1307 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
1293 1308
1294 1309 // Flatten all to bottom for now
1295 1310 switch( _AliasLevel ) {
1296 1311 case 0:
1297 1312 tj = TypePtr::BOTTOM;
1298 1313 break;
1299 1314 case 1: // Flatten to: oop, static, field or array
1300 1315 switch (tj->base()) {
1301 1316 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
1302 1317 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
1303 1318 case Type::AryPtr: // do not distinguish arrays at all
1304 1319 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
1305 1320 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1306 1321 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
1307 1322 default: ShouldNotReachHere();
1308 1323 }
1309 1324 break;
1310 1325 case 2: // No collapsing at level 2; keep all splits
1311 1326 case 3: // No collapsing at level 3; keep all splits
1312 1327 break;
1313 1328 default:
1314 1329 Unimplemented();
1315 1330 }
1316 1331
1317 1332 offset = tj->offset();
1318 1333 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1319 1334
1320 1335 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1321 1336 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1322 1337 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1323 1338 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1324 1339 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1325 1340 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1326 1341 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
1327 1342 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1328 1343 assert( tj->ptr() != TypePtr::TopPTR &&
1329 1344 tj->ptr() != TypePtr::AnyNull &&
1330 1345 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1331 1346 // assert( tj->ptr() != TypePtr::Constant ||
1332 1347 // tj->base() == Type::RawPtr ||
1333 1348 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
1334 1349
1335 1350 return tj;
1336 1351 }
1337 1352
1338 1353 void Compile::AliasType::Init(int i, const TypePtr* at) {
1339 1354 _index = i;
1340 1355 _adr_type = at;
1341 1356 _field = NULL;
1342 1357 _is_rewritable = true; // default
1343 1358 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1344 1359 if (atoop != NULL && atoop->is_known_instance()) {
1345 1360 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1346 1361 _general_index = Compile::current()->get_alias_index(gt);
1347 1362 } else {
1348 1363 _general_index = 0;
1349 1364 }
1350 1365 }
1351 1366
1352 1367 //---------------------------------print_on------------------------------------
1353 1368 #ifndef PRODUCT
1354 1369 void Compile::AliasType::print_on(outputStream* st) {
1355 1370 if (index() < 10)
1356 1371 st->print("@ <%d> ", index());
1357 1372 else st->print("@ <%d>", index());
1358 1373 st->print(is_rewritable() ? " " : " RO");
1359 1374 int offset = adr_type()->offset();
1360 1375 if (offset == Type::OffsetBot)
1361 1376 st->print(" +any");
1362 1377 else st->print(" +%-3d", offset);
1363 1378 st->print(" in ");
1364 1379 adr_type()->dump_on(st);
1365 1380 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1366 1381 if (field() != NULL && tjp) {
1367 1382 if (tjp->klass() != field()->holder() ||
1368 1383 tjp->offset() != field()->offset_in_bytes()) {
1369 1384 st->print(" != ");
1370 1385 field()->print();
1371 1386 st->print(" ***");
1372 1387 }
1373 1388 }
1374 1389 }
1375 1390
1376 1391 void print_alias_types() {
1377 1392 Compile* C = Compile::current();
1378 1393 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1379 1394 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1380 1395 C->alias_type(idx)->print_on(tty);
1381 1396 tty->cr();
1382 1397 }
1383 1398 }
1384 1399 #endif
1385 1400
1386 1401
1387 1402 //----------------------------probe_alias_cache--------------------------------
1388 1403 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1389 1404 intptr_t key = (intptr_t) adr_type;
1390 1405 key ^= key >> logAliasCacheSize;
1391 1406 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1392 1407 }
1393 1408
1394 1409
1395 1410 //-----------------------------grow_alias_types--------------------------------
1396 1411 void Compile::grow_alias_types() {
1397 1412 const int old_ats = _max_alias_types; // how many before?
1398 1413 const int new_ats = old_ats; // how many more?
1399 1414 const int grow_ats = old_ats+new_ats; // how many now?
1400 1415 _max_alias_types = grow_ats;
1401 1416 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1402 1417 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1403 1418 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1404 1419 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
1405 1420 }
1406 1421
1407 1422
1408 1423 //--------------------------------find_alias_type------------------------------
1409 1424 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1410 1425 if (_AliasLevel == 0)
1411 1426 return alias_type(AliasIdxBot);
1412 1427
1413 1428 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1414 1429 if (ace->_adr_type == adr_type) {
1415 1430 return alias_type(ace->_index);
1416 1431 }
1417 1432
1418 1433 // Handle special cases.
1419 1434 if (adr_type == NULL) return alias_type(AliasIdxTop);
1420 1435 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
1421 1436
1422 1437 // Do it the slow way.
1423 1438 const TypePtr* flat = flatten_alias_type(adr_type);
1424 1439
1425 1440 #ifdef ASSERT
1426 1441 assert(flat == flatten_alias_type(flat), "idempotent");
1427 1442 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
1428 1443 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1429 1444 const TypeOopPtr* foop = flat->is_oopptr();
1430 1445 // Scalarizable allocations have exact klass always.
1431 1446 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1432 1447 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1433 1448 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1434 1449 }
1435 1450 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1436 1451 #endif
1437 1452
1438 1453 int idx = AliasIdxTop;
1439 1454 for (int i = 0; i < num_alias_types(); i++) {
1440 1455 if (alias_type(i)->adr_type() == flat) {
1441 1456 idx = i;
1442 1457 break;
1443 1458 }
1444 1459 }
1445 1460
1446 1461 if (idx == AliasIdxTop) {
1447 1462 if (no_create) return NULL;
1448 1463 // Grow the array if necessary.
1449 1464 if (_num_alias_types == _max_alias_types) grow_alias_types();
1450 1465 // Add a new alias type.
1451 1466 idx = _num_alias_types++;
1452 1467 _alias_types[idx]->Init(idx, flat);
1453 1468 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
1454 1469 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
1455 1470 if (flat->isa_instptr()) {
1456 1471 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1457 1472 && flat->is_instptr()->klass() == env()->Class_klass())
1458 1473 alias_type(idx)->set_rewritable(false);
1459 1474 }
1460 1475 if (flat->isa_klassptr()) {
1461 1476 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1462 1477 alias_type(idx)->set_rewritable(false);
1463 1478 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1464 1479 alias_type(idx)->set_rewritable(false);
1465 1480 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1466 1481 alias_type(idx)->set_rewritable(false);
1467 1482 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1468 1483 alias_type(idx)->set_rewritable(false);
1469 1484 }
1470 1485 // %%% (We would like to finalize JavaThread::threadObj_offset(),
1471 1486 // but the base pointer type is not distinctive enough to identify
1472 1487 // references into JavaThread.)
1473 1488
1474 1489 // Check for final fields.
1475 1490 const TypeInstPtr* tinst = flat->isa_instptr();
1476 1491 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1477 1492 ciField* field;
1478 1493 if (tinst->const_oop() != NULL &&
1479 1494 tinst->klass() == ciEnv::current()->Class_klass() &&
1480 1495 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1481 1496 // static field
1482 1497 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1483 1498 field = k->get_field_by_offset(tinst->offset(), true);
1484 1499 } else {
1485 1500 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1486 1501 field = k->get_field_by_offset(tinst->offset(), false);
1487 1502 }
1488 1503 assert(field == NULL ||
1489 1504 original_field == NULL ||
1490 1505 (field->holder() == original_field->holder() &&
1491 1506 field->offset() == original_field->offset() &&
1492 1507 field->is_static() == original_field->is_static()), "wrong field?");
1493 1508 // Set field() and is_rewritable() attributes.
1494 1509 if (field != NULL) alias_type(idx)->set_field(field);
1495 1510 }
1496 1511 }
1497 1512
1498 1513 // Fill the cache for next time.
1499 1514 ace->_adr_type = adr_type;
1500 1515 ace->_index = idx;
1501 1516 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
1502 1517
1503 1518 // Might as well try to fill the cache for the flattened version, too.
1504 1519 AliasCacheEntry* face = probe_alias_cache(flat);
1505 1520 if (face->_adr_type == NULL) {
1506 1521 face->_adr_type = flat;
1507 1522 face->_index = idx;
1508 1523 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1509 1524 }
1510 1525
1511 1526 return alias_type(idx);
1512 1527 }
1513 1528
1514 1529
1515 1530 Compile::AliasType* Compile::alias_type(ciField* field) {
1516 1531 const TypeOopPtr* t;
1517 1532 if (field->is_static())
1518 1533 t = TypeInstPtr::make(field->holder()->java_mirror());
1519 1534 else
1520 1535 t = TypeOopPtr::make_from_klass_raw(field->holder());
1521 1536 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1522 1537 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1523 1538 return atp;
1524 1539 }
1525 1540
1526 1541
1527 1542 //------------------------------have_alias_type--------------------------------
1528 1543 bool Compile::have_alias_type(const TypePtr* adr_type) {
1529 1544 AliasCacheEntry* ace = probe_alias_cache(adr_type);
1530 1545 if (ace->_adr_type == adr_type) {
1531 1546 return true;
1532 1547 }
1533 1548
1534 1549 // Handle special cases.
1535 1550 if (adr_type == NULL) return true;
1536 1551 if (adr_type == TypePtr::BOTTOM) return true;
1537 1552
1538 1553 return find_alias_type(adr_type, true, NULL) != NULL;
1539 1554 }
1540 1555
1541 1556 //-----------------------------must_alias--------------------------------------
1542 1557 // True if all values of the given address type are in the given alias category.
1543 1558 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1544 1559 if (alias_idx == AliasIdxBot) return true; // the universal category
1545 1560 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
1546 1561 if (alias_idx == AliasIdxTop) return false; // the empty category
1547 1562 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1548 1563
1549 1564 // the only remaining possible overlap is identity
1550 1565 int adr_idx = get_alias_index(adr_type);
1551 1566 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1552 1567 assert(adr_idx == alias_idx ||
1553 1568 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1554 1569 && adr_type != TypeOopPtr::BOTTOM),
1555 1570 "should not be testing for overlap with an unsafe pointer");
1556 1571 return adr_idx == alias_idx;
1557 1572 }
1558 1573
1559 1574 //------------------------------can_alias--------------------------------------
1560 1575 // True if any values of the given address type are in the given alias category.
1561 1576 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1562 1577 if (alias_idx == AliasIdxTop) return false; // the empty category
1563 1578 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
1564 1579 if (alias_idx == AliasIdxBot) return true; // the universal category
1565 1580 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
1566 1581
1567 1582 // the only remaining possible overlap is identity
1568 1583 int adr_idx = get_alias_index(adr_type);
1569 1584 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1570 1585 return adr_idx == alias_idx;
1571 1586 }
1572 1587
1573 1588
1574 1589
1575 1590 //---------------------------pop_warm_call-------------------------------------
1576 1591 WarmCallInfo* Compile::pop_warm_call() {
1577 1592 WarmCallInfo* wci = _warm_calls;
1578 1593 if (wci != NULL) _warm_calls = wci->remove_from(wci);
1579 1594 return wci;
1580 1595 }
1581 1596
1582 1597 //----------------------------Inline_Warm--------------------------------------
1583 1598 int Compile::Inline_Warm() {
1584 1599 // If there is room, try to inline some more warm call sites.
1585 1600 // %%% Do a graph index compaction pass when we think we're out of space?
1586 1601 if (!InlineWarmCalls) return 0;
1587 1602
1588 1603 int calls_made_hot = 0;
1589 1604 int room_to_grow = NodeCountInliningCutoff - unique();
1590 1605 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1591 1606 int amount_grown = 0;
1592 1607 WarmCallInfo* call;
1593 1608 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1594 1609 int est_size = (int)call->size();
1595 1610 if (est_size > (room_to_grow - amount_grown)) {
1596 1611 // This one won't fit anyway. Get rid of it.
1597 1612 call->make_cold();
1598 1613 continue;
1599 1614 }
1600 1615 call->make_hot();
1601 1616 calls_made_hot++;
1602 1617 amount_grown += est_size;
1603 1618 amount_to_grow -= est_size;
1604 1619 }
1605 1620
1606 1621 if (calls_made_hot > 0) set_major_progress();
1607 1622 return calls_made_hot;
1608 1623 }
1609 1624
1610 1625
1611 1626 //----------------------------Finish_Warm--------------------------------------
1612 1627 void Compile::Finish_Warm() {
1613 1628 if (!InlineWarmCalls) return;
1614 1629 if (failing()) return;
1615 1630 if (warm_calls() == NULL) return;
1616 1631
1617 1632 // Clean up loose ends, if we are out of space for inlining.
1618 1633 WarmCallInfo* call;
1619 1634 while ((call = pop_warm_call()) != NULL) {
1620 1635 call->make_cold();
1621 1636 }
1622 1637 }
1623 1638
1624 1639 //---------------------cleanup_loop_predicates-----------------------
1625 1640 // Remove the opaque nodes that protect the predicates so that all unused
1626 1641 // checks and uncommon_traps will be eliminated from the ideal graph
1627 1642 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1628 1643 if (predicate_count()==0) return;
1629 1644 for (int i = predicate_count(); i > 0; i--) {
1630 1645 Node * n = predicate_opaque1_node(i-1);
1631 1646 assert(n->Opcode() == Op_Opaque1, "must be");
1632 1647 igvn.replace_node(n, n->in(1));
1633 1648 }
1634 1649 assert(predicate_count()==0, "should be clean!");
1635 1650 igvn.optimize();
1636 1651 }
1637 1652
1638 1653 //------------------------------Optimize---------------------------------------
1639 1654 // Given a graph, optimize it.
1640 1655 void Compile::Optimize() {
1641 1656 TracePhase t1("optimizer", &_t_optimizer, true);
1642 1657
1643 1658 #ifndef PRODUCT
1644 1659 if (env()->break_at_compile()) {
1645 1660 BREAKPOINT;
1646 1661 }
1647 1662
1648 1663 #endif
1649 1664
1650 1665 ResourceMark rm;
1651 1666 int loop_opts_cnt;
1652 1667
1653 1668 NOT_PRODUCT( verify_graph_edges(); )
1654 1669
1655 1670 print_method("After Parsing");
1656 1671
1657 1672 {
1658 1673 // Iterative Global Value Numbering, including ideal transforms
1659 1674 // Initialize IterGVN with types and values from parse-time GVN
1660 1675 PhaseIterGVN igvn(initial_gvn());
1661 1676 {
1662 1677 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1663 1678 igvn.optimize();
1664 1679 }
1665 1680
1666 1681 print_method("Iter GVN 1", 2);
1667 1682
1668 1683 if (failing()) return;
1669 1684
1670 1685 // Perform escape analysis
1671 1686 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1672 1687 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
1673 1688 ConnectionGraph::do_analysis(this, &igvn);
1674 1689
1675 1690 if (failing()) return;
1676 1691
1677 1692 igvn.optimize();
1678 1693 print_method("Iter GVN 3", 2);
1679 1694
1680 1695 if (failing()) return;
1681 1696
1682 1697 }
1683 1698
1684 1699 // Loop transforms on the ideal graph. Range Check Elimination,
1685 1700 // peeling, unrolling, etc.
1686 1701
1687 1702 // Set loop opts counter
1688 1703 loop_opts_cnt = num_loop_opts();
1689 1704 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1690 1705 {
1691 1706 TracePhase t2("idealLoop", &_t_idealLoop, true);
1692 1707 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate);
1693 1708 loop_opts_cnt--;
1694 1709 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1695 1710 if (failing()) return;
1696 1711 }
1697 1712 // Loop opts pass if partial peeling occurred in previous pass
1698 1713 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1699 1714 TracePhase t3("idealLoop", &_t_idealLoop, true);
1700 1715 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
1701 1716 loop_opts_cnt--;
1702 1717 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1703 1718 if (failing()) return;
1704 1719 }
1705 1720 // Loop opts pass for loop-unrolling before CCP
1706 1721 if(major_progress() && (loop_opts_cnt > 0)) {
1707 1722 TracePhase t4("idealLoop", &_t_idealLoop, true);
1708 1723 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
1709 1724 loop_opts_cnt--;
1710 1725 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1711 1726 }
1712 1727 if (!failing()) {
1713 1728 // Verify that last round of loop opts produced a valid graph
1714 1729 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1715 1730 PhaseIdealLoop::verify(igvn);
1716 1731 }
1717 1732 }
1718 1733 if (failing()) return;
1719 1734
1720 1735 // Conditional Constant Propagation;
1721 1736 PhaseCCP ccp( &igvn );
1722 1737 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1723 1738 {
1724 1739 TracePhase t2("ccp", &_t_ccp, true);
1725 1740 ccp.do_transform();
1726 1741 }
1727 1742 print_method("PhaseCPP 1", 2);
1728 1743
1729 1744 assert( true, "Break here to ccp.dump_old2new_map()");
1730 1745
1731 1746 // Iterative Global Value Numbering, including ideal transforms
1732 1747 {
1733 1748 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1734 1749 igvn = ccp;
1735 1750 igvn.optimize();
1736 1751 }
1737 1752
1738 1753 print_method("Iter GVN 2", 2);
1739 1754
1740 1755 if (failing()) return;
1741 1756
1742 1757 // Loop transforms on the ideal graph. Range Check Elimination,
1743 1758 // peeling, unrolling, etc.
1744 1759 if(loop_opts_cnt > 0) {
1745 1760 debug_only( int cnt = 0; );
1746 1761 bool loop_predication = UseLoopPredicate;
1747 1762 while(major_progress() && (loop_opts_cnt > 0)) {
1748 1763 TracePhase t2("idealLoop", &_t_idealLoop, true);
1749 1764 assert( cnt++ < 40, "infinite cycle in loop optimization" );
1750 1765 PhaseIdealLoop ideal_loop( igvn, true, loop_predication);
1751 1766 loop_opts_cnt--;
1752 1767 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1753 1768 if (failing()) return;
1754 1769 // Perform loop predication optimization during first iteration after CCP.
1755 1770 // After that switch it off and cleanup unused loop predicates.
1756 1771 if (loop_predication) {
1757 1772 loop_predication = false;
1758 1773 cleanup_loop_predicates(igvn);
1759 1774 if (failing()) return;
1760 1775 }
1761 1776 }
1762 1777 }
1763 1778
1764 1779 {
1765 1780 // Verify that all previous optimizations produced a valid graph
1766 1781 // at least to this point, even if no loop optimizations were done.
1767 1782 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1768 1783 PhaseIdealLoop::verify(igvn);
1769 1784 }
1770 1785
1771 1786 {
1772 1787 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1773 1788 PhaseMacroExpand mex(igvn);
1774 1789 if (mex.expand_macro_nodes()) {
1775 1790 assert(failing(), "must bail out w/ explicit message");
1776 1791 return;
1777 1792 }
1778 1793 }
1779 1794
1780 1795 } // (End scope of igvn; run destructor if necessary for asserts.)
1781 1796
1782 1797 // A method with only infinite loops has no edges entering loops from root
1783 1798 {
1784 1799 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1785 1800 if (final_graph_reshaping()) {
1786 1801 assert(failing(), "must bail out w/ explicit message");
1787 1802 return;
1788 1803 }
1789 1804 }
1790 1805
1791 1806 print_method("Optimize finished", 2);
1792 1807 }
1793 1808
1794 1809
1795 1810 //------------------------------Code_Gen---------------------------------------
1796 1811 // Given a graph, generate code for it
1797 1812 void Compile::Code_Gen() {
1798 1813 if (failing()) return;
1799 1814
1800 1815 // Perform instruction selection. You might think we could reclaim Matcher
1801 1816 // memory PDQ, but actually the Matcher is used in generating spill code.
1802 1817 // Internals of the Matcher (including some VectorSets) must remain live
1803 1818 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1804 1819 // set a bit in reclaimed memory.
1805 1820
1806 1821 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1807 1822 // nodes. Mapping is only valid at the root of each matched subtree.
1808 1823 NOT_PRODUCT( verify_graph_edges(); )
1809 1824
1810 1825 Node_List proj_list;
1811 1826 Matcher m(proj_list);
1812 1827 _matcher = &m;
1813 1828 {
1814 1829 TracePhase t2("matcher", &_t_matcher, true);
1815 1830 m.match();
1816 1831 }
1817 1832 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1818 1833 // nodes. Mapping is only valid at the root of each matched subtree.
1819 1834 NOT_PRODUCT( verify_graph_edges(); )
1820 1835
1821 1836 // If you have too many nodes, or if matching has failed, bail out
1822 1837 check_node_count(0, "out of nodes matching instructions");
1823 1838 if (failing()) return;
1824 1839
1825 1840 // Build a proper-looking CFG
1826 1841 PhaseCFG cfg(node_arena(), root(), m);
1827 1842 _cfg = &cfg;
1828 1843 {
1829 1844 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1830 1845 cfg.Dominators();
1831 1846 if (failing()) return;
1832 1847
1833 1848 NOT_PRODUCT( verify_graph_edges(); )
1834 1849
1835 1850 cfg.Estimate_Block_Frequency();
1836 1851 cfg.GlobalCodeMotion(m,unique(),proj_list);
1837 1852
1838 1853 print_method("Global code motion", 2);
1839 1854
1840 1855 if (failing()) return;
1841 1856 NOT_PRODUCT( verify_graph_edges(); )
1842 1857
1843 1858 debug_only( cfg.verify(); )
1844 1859 }
1845 1860 NOT_PRODUCT( verify_graph_edges(); )
1846 1861
1847 1862 PhaseChaitin regalloc(unique(),cfg,m);
1848 1863 _regalloc = ®alloc;
1849 1864 {
1850 1865 TracePhase t2("regalloc", &_t_registerAllocation, true);
1851 1866 // Perform any platform dependent preallocation actions. This is used,
1852 1867 // for example, to avoid taking an implicit null pointer exception
1853 1868 // using the frame pointer on win95.
1854 1869 _regalloc->pd_preallocate_hook();
1855 1870
1856 1871 // Perform register allocation. After Chaitin, use-def chains are
1857 1872 // no longer accurate (at spill code) and so must be ignored.
1858 1873 // Node->LRG->reg mappings are still accurate.
1859 1874 _regalloc->Register_Allocate();
1860 1875
1861 1876 // Bail out if the allocator builds too many nodes
1862 1877 if (failing()) return;
1863 1878 }
1864 1879
1865 1880 // Prior to register allocation we kept empty basic blocks in case the
1866 1881 // the allocator needed a place to spill. After register allocation we
1867 1882 // are not adding any new instructions. If any basic block is empty, we
1868 1883 // can now safely remove it.
1869 1884 {
1870 1885 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1871 1886 cfg.remove_empty();
1872 1887 if (do_freq_based_layout()) {
1873 1888 PhaseBlockLayout layout(cfg);
1874 1889 } else {
1875 1890 cfg.set_loop_alignment();
1876 1891 }
1877 1892 cfg.fixup_flow();
1878 1893 }
1879 1894
1880 1895 // Perform any platform dependent postallocation verifications.
1881 1896 debug_only( _regalloc->pd_postallocate_verify_hook(); )
1882 1897
1883 1898 // Apply peephole optimizations
1884 1899 if( OptoPeephole ) {
1885 1900 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1886 1901 PhasePeephole peep( _regalloc, cfg);
1887 1902 peep.do_transform();
1888 1903 }
1889 1904
1890 1905 // Convert Nodes to instruction bits in a buffer
1891 1906 {
1892 1907 // %%%% workspace merge brought two timers together for one job
1893 1908 TracePhase t2a("output", &_t_output, true);
1894 1909 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1895 1910 Output();
1896 1911 }
1897 1912
1898 1913 print_method("Final Code");
1899 1914
1900 1915 // He's dead, Jim.
1901 1916 _cfg = (PhaseCFG*)0xdeadbeef;
1902 1917 _regalloc = (PhaseChaitin*)0xdeadbeef;
1903 1918 }
1904 1919
1905 1920
1906 1921 //------------------------------dump_asm---------------------------------------
1907 1922 // Dump formatted assembly
1908 1923 #ifndef PRODUCT
1909 1924 void Compile::dump_asm(int *pcs, uint pc_limit) {
1910 1925 bool cut_short = false;
1911 1926 tty->print_cr("#");
1912 1927 tty->print("# "); _tf->dump(); tty->cr();
1913 1928 tty->print_cr("#");
1914 1929
1915 1930 // For all blocks
1916 1931 int pc = 0x0; // Program counter
1917 1932 char starts_bundle = ' ';
1918 1933 _regalloc->dump_frame();
1919 1934
1920 1935 Node *n = NULL;
1921 1936 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1922 1937 if (VMThread::should_terminate()) { cut_short = true; break; }
1923 1938 Block *b = _cfg->_blocks[i];
1924 1939 if (b->is_connector() && !Verbose) continue;
1925 1940 n = b->_nodes[0];
1926 1941 if (pcs && n->_idx < pc_limit)
1927 1942 tty->print("%3.3x ", pcs[n->_idx]);
1928 1943 else
1929 1944 tty->print(" ");
1930 1945 b->dump_head( &_cfg->_bbs );
1931 1946 if (b->is_connector()) {
1932 1947 tty->print_cr(" # Empty connector block");
1933 1948 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1934 1949 tty->print_cr(" # Block is sole successor of call");
1935 1950 }
1936 1951
1937 1952 // For all instructions
1938 1953 Node *delay = NULL;
1939 1954 for( uint j = 0; j<b->_nodes.size(); j++ ) {
1940 1955 if (VMThread::should_terminate()) { cut_short = true; break; }
1941 1956 n = b->_nodes[j];
1942 1957 if (valid_bundle_info(n)) {
1943 1958 Bundle *bundle = node_bundling(n);
1944 1959 if (bundle->used_in_unconditional_delay()) {
1945 1960 delay = n;
1946 1961 continue;
1947 1962 }
1948 1963 if (bundle->starts_bundle())
1949 1964 starts_bundle = '+';
1950 1965 }
1951 1966
1952 1967 if (WizardMode) n->dump();
1953 1968
1954 1969 if( !n->is_Region() && // Dont print in the Assembly
1955 1970 !n->is_Phi() && // a few noisely useless nodes
1956 1971 !n->is_Proj() &&
1957 1972 !n->is_MachTemp() &&
1958 1973 !n->is_SafePointScalarObject() &&
1959 1974 !n->is_Catch() && // Would be nice to print exception table targets
1960 1975 !n->is_MergeMem() && // Not very interesting
1961 1976 !n->is_top() && // Debug info table constants
1962 1977 !(n->is_Con() && !n->is_Mach())// Debug info table constants
1963 1978 ) {
1964 1979 if (pcs && n->_idx < pc_limit)
1965 1980 tty->print("%3.3x", pcs[n->_idx]);
1966 1981 else
1967 1982 tty->print(" ");
1968 1983 tty->print(" %c ", starts_bundle);
1969 1984 starts_bundle = ' ';
1970 1985 tty->print("\t");
1971 1986 n->format(_regalloc, tty);
1972 1987 tty->cr();
1973 1988 }
1974 1989
1975 1990 // If we have an instruction with a delay slot, and have seen a delay,
1976 1991 // then back up and print it
1977 1992 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1978 1993 assert(delay != NULL, "no unconditional delay instruction");
1979 1994 if (WizardMode) delay->dump();
1980 1995
1981 1996 if (node_bundling(delay)->starts_bundle())
1982 1997 starts_bundle = '+';
1983 1998 if (pcs && n->_idx < pc_limit)
1984 1999 tty->print("%3.3x", pcs[n->_idx]);
1985 2000 else
1986 2001 tty->print(" ");
1987 2002 tty->print(" %c ", starts_bundle);
1988 2003 starts_bundle = ' ';
1989 2004 tty->print("\t");
1990 2005 delay->format(_regalloc, tty);
1991 2006 tty->print_cr("");
1992 2007 delay = NULL;
1993 2008 }
1994 2009
1995 2010 // Dump the exception table as well
1996 2011 if( n->is_Catch() && (Verbose || WizardMode) ) {
1997 2012 // Print the exception table for this offset
1998 2013 _handler_table.print_subtable_for(pc);
1999 2014 }
2000 2015 }
2001 2016
2002 2017 if (pcs && n->_idx < pc_limit)
2003 2018 tty->print_cr("%3.3x", pcs[n->_idx]);
2004 2019 else
2005 2020 tty->print_cr("");
2006 2021
2007 2022 assert(cut_short || delay == NULL, "no unconditional delay branch");
2008 2023
2009 2024 } // End of per-block dump
2010 2025 tty->print_cr("");
2011 2026
2012 2027 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
2013 2028 }
2014 2029 #endif
2015 2030
2016 2031 //------------------------------Final_Reshape_Counts---------------------------
2017 2032 // This class defines counters to help identify when a method
2018 2033 // may/must be executed using hardware with only 24-bit precision.
2019 2034 struct Final_Reshape_Counts : public StackObj {
2020 2035 int _call_count; // count non-inlined 'common' calls
2021 2036 int _float_count; // count float ops requiring 24-bit precision
2022 2037 int _double_count; // count double ops requiring more precision
2023 2038 int _java_call_count; // count non-inlined 'java' calls
2024 2039 int _inner_loop_count; // count loops which need alignment
2025 2040 VectorSet _visited; // Visitation flags
2026 2041 Node_List _tests; // Set of IfNodes & PCTableNodes
2027 2042
2028 2043 Final_Reshape_Counts() :
2029 2044 _call_count(0), _float_count(0), _double_count(0),
2030 2045 _java_call_count(0), _inner_loop_count(0),
2031 2046 _visited( Thread::current()->resource_area() ) { }
2032 2047
2033 2048 void inc_call_count () { _call_count ++; }
2034 2049 void inc_float_count () { _float_count ++; }
2035 2050 void inc_double_count() { _double_count++; }
2036 2051 void inc_java_call_count() { _java_call_count++; }
2037 2052 void inc_inner_loop_count() { _inner_loop_count++; }
2038 2053
2039 2054 int get_call_count () const { return _call_count ; }
2040 2055 int get_float_count () const { return _float_count ; }
2041 2056 int get_double_count() const { return _double_count; }
2042 2057 int get_java_call_count() const { return _java_call_count; }
2043 2058 int get_inner_loop_count() const { return _inner_loop_count; }
2044 2059 };
2045 2060
2046 2061 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2047 2062 ciInstanceKlass *k = tp->klass()->as_instance_klass();
2048 2063 // Make sure the offset goes inside the instance layout.
2049 2064 return k->contains_field_offset(tp->offset());
2050 2065 // Note that OffsetBot and OffsetTop are very negative.
2051 2066 }
2052 2067
2053 2068 //------------------------------final_graph_reshaping_impl----------------------
2054 2069 // Implement items 1-5 from final_graph_reshaping below.
2055 2070 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
2056 2071
2057 2072 if ( n->outcnt() == 0 ) return; // dead node
2058 2073 uint nop = n->Opcode();
2059 2074
2060 2075 // Check for 2-input instruction with "last use" on right input.
2061 2076 // Swap to left input. Implements item (2).
2062 2077 if( n->req() == 3 && // two-input instruction
2063 2078 n->in(1)->outcnt() > 1 && // left use is NOT a last use
2064 2079 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2065 2080 n->in(2)->outcnt() == 1 &&// right use IS a last use
2066 2081 !n->in(2)->is_Con() ) { // right use is not a constant
2067 2082 // Check for commutative opcode
2068 2083 switch( nop ) {
2069 2084 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
2070 2085 case Op_MaxI: case Op_MinI:
2071 2086 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
2072 2087 case Op_AndL: case Op_XorL: case Op_OrL:
2073 2088 case Op_AndI: case Op_XorI: case Op_OrI: {
2074 2089 // Move "last use" input to left by swapping inputs
2075 2090 n->swap_edges(1, 2);
2076 2091 break;
2077 2092 }
2078 2093 default:
2079 2094 break;
2080 2095 }
2081 2096 }
2082 2097
2083 2098 #ifdef ASSERT
2084 2099 if( n->is_Mem() ) {
2085 2100 Compile* C = Compile::current();
2086 2101 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
2087 2102 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2088 2103 // oop will be recorded in oop map if load crosses safepoint
2089 2104 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2090 2105 LoadNode::is_immutable_value(n->in(MemNode::Address))),
2091 2106 "raw memory operations should have control edge");
2092 2107 }
2093 2108 #endif
2094 2109 // Count FPU ops and common calls, implements item (3)
2095 2110 switch( nop ) {
2096 2111 // Count all float operations that may use FPU
2097 2112 case Op_AddF:
2098 2113 case Op_SubF:
2099 2114 case Op_MulF:
2100 2115 case Op_DivF:
2101 2116 case Op_NegF:
2102 2117 case Op_ModF:
2103 2118 case Op_ConvI2F:
2104 2119 case Op_ConF:
2105 2120 case Op_CmpF:
2106 2121 case Op_CmpF3:
2107 2122 // case Op_ConvL2F: // longs are split into 32-bit halves
2108 2123 frc.inc_float_count();
2109 2124 break;
2110 2125
2111 2126 case Op_ConvF2D:
2112 2127 case Op_ConvD2F:
2113 2128 frc.inc_float_count();
2114 2129 frc.inc_double_count();
2115 2130 break;
2116 2131
2117 2132 // Count all double operations that may use FPU
2118 2133 case Op_AddD:
2119 2134 case Op_SubD:
2120 2135 case Op_MulD:
2121 2136 case Op_DivD:
2122 2137 case Op_NegD:
2123 2138 case Op_ModD:
2124 2139 case Op_ConvI2D:
2125 2140 case Op_ConvD2I:
2126 2141 // case Op_ConvL2D: // handled by leaf call
2127 2142 // case Op_ConvD2L: // handled by leaf call
2128 2143 case Op_ConD:
2129 2144 case Op_CmpD:
2130 2145 case Op_CmpD3:
2131 2146 frc.inc_double_count();
2132 2147 break;
2133 2148 case Op_Opaque1: // Remove Opaque Nodes before matching
2134 2149 case Op_Opaque2: // Remove Opaque Nodes before matching
2135 2150 n->subsume_by(n->in(1));
2136 2151 break;
2137 2152 case Op_CallStaticJava:
2138 2153 case Op_CallJava:
2139 2154 case Op_CallDynamicJava:
2140 2155 frc.inc_java_call_count(); // Count java call site;
2141 2156 case Op_CallRuntime:
2142 2157 case Op_CallLeaf:
2143 2158 case Op_CallLeafNoFP: {
2144 2159 assert( n->is_Call(), "" );
2145 2160 CallNode *call = n->as_Call();
2146 2161 // Count call sites where the FP mode bit would have to be flipped.
2147 2162 // Do not count uncommon runtime calls:
2148 2163 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2149 2164 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2150 2165 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2151 2166 frc.inc_call_count(); // Count the call site
2152 2167 } else { // See if uncommon argument is shared
2153 2168 Node *n = call->in(TypeFunc::Parms);
2154 2169 int nop = n->Opcode();
2155 2170 // Clone shared simple arguments to uncommon calls, item (1).
2156 2171 if( n->outcnt() > 1 &&
2157 2172 !n->is_Proj() &&
2158 2173 nop != Op_CreateEx &&
2159 2174 nop != Op_CheckCastPP &&
2160 2175 nop != Op_DecodeN &&
2161 2176 !n->is_Mem() ) {
2162 2177 Node *x = n->clone();
2163 2178 call->set_req( TypeFunc::Parms, x );
2164 2179 }
2165 2180 }
2166 2181 break;
2167 2182 }
2168 2183
2169 2184 case Op_StoreD:
2170 2185 case Op_LoadD:
2171 2186 case Op_LoadD_unaligned:
2172 2187 frc.inc_double_count();
2173 2188 goto handle_mem;
2174 2189 case Op_StoreF:
2175 2190 case Op_LoadF:
2176 2191 frc.inc_float_count();
2177 2192 goto handle_mem;
2178 2193
2179 2194 case Op_StoreB:
2180 2195 case Op_StoreC:
2181 2196 case Op_StoreCM:
2182 2197 case Op_StorePConditional:
2183 2198 case Op_StoreI:
2184 2199 case Op_StoreL:
2185 2200 case Op_StoreIConditional:
2186 2201 case Op_StoreLConditional:
2187 2202 case Op_CompareAndSwapI:
2188 2203 case Op_CompareAndSwapL:
2189 2204 case Op_CompareAndSwapP:
2190 2205 case Op_CompareAndSwapN:
2191 2206 case Op_StoreP:
2192 2207 case Op_StoreN:
2193 2208 case Op_LoadB:
2194 2209 case Op_LoadUB:
2195 2210 case Op_LoadUS:
2196 2211 case Op_LoadI:
2197 2212 case Op_LoadUI2L:
2198 2213 case Op_LoadKlass:
2199 2214 case Op_LoadNKlass:
2200 2215 case Op_LoadL:
2201 2216 case Op_LoadL_unaligned:
2202 2217 case Op_LoadPLocked:
2203 2218 case Op_LoadLLocked:
2204 2219 case Op_LoadP:
2205 2220 case Op_LoadN:
2206 2221 case Op_LoadRange:
2207 2222 case Op_LoadS: {
2208 2223 handle_mem:
2209 2224 #ifdef ASSERT
2210 2225 if( VerifyOptoOopOffsets ) {
2211 2226 assert( n->is_Mem(), "" );
2212 2227 MemNode *mem = (MemNode*)n;
2213 2228 // Check to see if address types have grounded out somehow.
2214 2229 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2215 2230 assert( !tp || oop_offset_is_sane(tp), "" );
2216 2231 }
2217 2232 #endif
2218 2233 break;
2219 2234 }
2220 2235
2221 2236 case Op_AddP: { // Assert sane base pointers
2222 2237 Node *addp = n->in(AddPNode::Address);
2223 2238 assert( !addp->is_AddP() ||
2224 2239 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2225 2240 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2226 2241 "Base pointers must match" );
2227 2242 #ifdef _LP64
2228 2243 if (UseCompressedOops &&
2229 2244 addp->Opcode() == Op_ConP &&
2230 2245 addp == n->in(AddPNode::Base) &&
2231 2246 n->in(AddPNode::Offset)->is_Con()) {
2232 2247 // Use addressing with narrow klass to load with offset on x86.
2233 2248 // On sparc loading 32-bits constant and decoding it have less
2234 2249 // instructions (4) then load 64-bits constant (7).
2235 2250 // Do this transformation here since IGVN will convert ConN back to ConP.
2236 2251 const Type* t = addp->bottom_type();
2237 2252 if (t->isa_oopptr()) {
2238 2253 Node* nn = NULL;
2239 2254
2240 2255 // Look for existing ConN node of the same exact type.
2241 2256 Compile* C = Compile::current();
2242 2257 Node* r = C->root();
2243 2258 uint cnt = r->outcnt();
2244 2259 for (uint i = 0; i < cnt; i++) {
2245 2260 Node* m = r->raw_out(i);
2246 2261 if (m!= NULL && m->Opcode() == Op_ConN &&
2247 2262 m->bottom_type()->make_ptr() == t) {
2248 2263 nn = m;
2249 2264 break;
2250 2265 }
2251 2266 }
2252 2267 if (nn != NULL) {
2253 2268 // Decode a narrow oop to match address
2254 2269 // [R12 + narrow_oop_reg<<3 + offset]
2255 2270 nn = new (C, 2) DecodeNNode(nn, t);
2256 2271 n->set_req(AddPNode::Base, nn);
2257 2272 n->set_req(AddPNode::Address, nn);
2258 2273 if (addp->outcnt() == 0) {
2259 2274 addp->disconnect_inputs(NULL);
2260 2275 }
2261 2276 }
2262 2277 }
2263 2278 }
2264 2279 #endif
2265 2280 break;
2266 2281 }
2267 2282
2268 2283 #ifdef _LP64
2269 2284 case Op_CastPP:
2270 2285 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2271 2286 Compile* C = Compile::current();
2272 2287 Node* in1 = n->in(1);
2273 2288 const Type* t = n->bottom_type();
2274 2289 Node* new_in1 = in1->clone();
2275 2290 new_in1->as_DecodeN()->set_type(t);
2276 2291
2277 2292 if (!Matcher::narrow_oop_use_complex_address()) {
2278 2293 //
2279 2294 // x86, ARM and friends can handle 2 adds in addressing mode
2280 2295 // and Matcher can fold a DecodeN node into address by using
2281 2296 // a narrow oop directly and do implicit NULL check in address:
2282 2297 //
2283 2298 // [R12 + narrow_oop_reg<<3 + offset]
2284 2299 // NullCheck narrow_oop_reg
2285 2300 //
2286 2301 // On other platforms (Sparc) we have to keep new DecodeN node and
2287 2302 // use it to do implicit NULL check in address:
2288 2303 //
2289 2304 // decode_not_null narrow_oop_reg, base_reg
2290 2305 // [base_reg + offset]
2291 2306 // NullCheck base_reg
2292 2307 //
2293 2308 // Pin the new DecodeN node to non-null path on these platform (Sparc)
2294 2309 // to keep the information to which NULL check the new DecodeN node
2295 2310 // corresponds to use it as value in implicit_null_check().
2296 2311 //
2297 2312 new_in1->set_req(0, n->in(0));
2298 2313 }
2299 2314
2300 2315 n->subsume_by(new_in1);
2301 2316 if (in1->outcnt() == 0) {
2302 2317 in1->disconnect_inputs(NULL);
2303 2318 }
2304 2319 }
2305 2320 break;
2306 2321
2307 2322 case Op_CmpP:
2308 2323 // Do this transformation here to preserve CmpPNode::sub() and
2309 2324 // other TypePtr related Ideal optimizations (for example, ptr nullness).
2310 2325 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2311 2326 Node* in1 = n->in(1);
2312 2327 Node* in2 = n->in(2);
2313 2328 if (!in1->is_DecodeN()) {
2314 2329 in2 = in1;
2315 2330 in1 = n->in(2);
2316 2331 }
2317 2332 assert(in1->is_DecodeN(), "sanity");
2318 2333
2319 2334 Compile* C = Compile::current();
2320 2335 Node* new_in2 = NULL;
2321 2336 if (in2->is_DecodeN()) {
2322 2337 new_in2 = in2->in(1);
2323 2338 } else if (in2->Opcode() == Op_ConP) {
2324 2339 const Type* t = in2->bottom_type();
2325 2340 if (t == TypePtr::NULL_PTR) {
2326 2341 // Don't convert CmpP null check into CmpN if compressed
2327 2342 // oops implicit null check is not generated.
2328 2343 // This will allow to generate normal oop implicit null check.
2329 2344 if (Matcher::gen_narrow_oop_implicit_null_checks())
2330 2345 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2331 2346 //
2332 2347 // This transformation together with CastPP transformation above
2333 2348 // will generated code for implicit NULL checks for compressed oops.
2334 2349 //
2335 2350 // The original code after Optimize()
2336 2351 //
2337 2352 // LoadN memory, narrow_oop_reg
2338 2353 // decode narrow_oop_reg, base_reg
2339 2354 // CmpP base_reg, NULL
2340 2355 // CastPP base_reg // NotNull
2341 2356 // Load [base_reg + offset], val_reg
2342 2357 //
2343 2358 // after these transformations will be
2344 2359 //
2345 2360 // LoadN memory, narrow_oop_reg
2346 2361 // CmpN narrow_oop_reg, NULL
2347 2362 // decode_not_null narrow_oop_reg, base_reg
2348 2363 // Load [base_reg + offset], val_reg
2349 2364 //
2350 2365 // and the uncommon path (== NULL) will use narrow_oop_reg directly
2351 2366 // since narrow oops can be used in debug info now (see the code in
2352 2367 // final_graph_reshaping_walk()).
2353 2368 //
2354 2369 // At the end the code will be matched to
2355 2370 // on x86:
2356 2371 //
2357 2372 // Load_narrow_oop memory, narrow_oop_reg
2358 2373 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2359 2374 // NullCheck narrow_oop_reg
2360 2375 //
2361 2376 // and on sparc:
2362 2377 //
2363 2378 // Load_narrow_oop memory, narrow_oop_reg
2364 2379 // decode_not_null narrow_oop_reg, base_reg
2365 2380 // Load [base_reg + offset], val_reg
2366 2381 // NullCheck base_reg
2367 2382 //
2368 2383 } else if (t->isa_oopptr()) {
2369 2384 new_in2 = ConNode::make(C, t->make_narrowoop());
2370 2385 }
2371 2386 }
2372 2387 if (new_in2 != NULL) {
2373 2388 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2374 2389 n->subsume_by( cmpN );
2375 2390 if (in1->outcnt() == 0) {
2376 2391 in1->disconnect_inputs(NULL);
2377 2392 }
2378 2393 if (in2->outcnt() == 0) {
2379 2394 in2->disconnect_inputs(NULL);
2380 2395 }
2381 2396 }
2382 2397 }
2383 2398 break;
2384 2399
2385 2400 case Op_DecodeN:
2386 2401 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2387 2402 // DecodeN could be pinned when it can't be fold into
2388 2403 // an address expression, see the code for Op_CastPP above.
2389 2404 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2390 2405 break;
2391 2406
2392 2407 case Op_EncodeP: {
2393 2408 Node* in1 = n->in(1);
2394 2409 if (in1->is_DecodeN()) {
2395 2410 n->subsume_by(in1->in(1));
2396 2411 } else if (in1->Opcode() == Op_ConP) {
2397 2412 Compile* C = Compile::current();
2398 2413 const Type* t = in1->bottom_type();
2399 2414 if (t == TypePtr::NULL_PTR) {
2400 2415 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2401 2416 } else if (t->isa_oopptr()) {
2402 2417 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2403 2418 }
2404 2419 }
2405 2420 if (in1->outcnt() == 0) {
2406 2421 in1->disconnect_inputs(NULL);
2407 2422 }
2408 2423 break;
2409 2424 }
2410 2425
2411 2426 case Op_Proj: {
2412 2427 if (OptimizeStringConcat) {
2413 2428 ProjNode* p = n->as_Proj();
2414 2429 if (p->_is_io_use) {
2415 2430 // Separate projections were used for the exception path which
2416 2431 // are normally removed by a late inline. If it wasn't inlined
2417 2432 // then they will hang around and should just be replaced with
2418 2433 // the original one.
2419 2434 Node* proj = NULL;
2420 2435 // Replace with just one
2421 2436 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2422 2437 Node *use = i.get();
2423 2438 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2424 2439 proj = use;
2425 2440 break;
2426 2441 }
2427 2442 }
2428 2443 assert(p != NULL, "must be found");
2429 2444 p->subsume_by(proj);
2430 2445 }
2431 2446 }
2432 2447 break;
2433 2448 }
2434 2449
2435 2450 case Op_Phi:
2436 2451 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2437 2452 // The EncodeP optimization may create Phi with the same edges
2438 2453 // for all paths. It is not handled well by Register Allocator.
2439 2454 Node* unique_in = n->in(1);
2440 2455 assert(unique_in != NULL, "");
2441 2456 uint cnt = n->req();
2442 2457 for (uint i = 2; i < cnt; i++) {
2443 2458 Node* m = n->in(i);
2444 2459 assert(m != NULL, "");
2445 2460 if (unique_in != m)
2446 2461 unique_in = NULL;
2447 2462 }
2448 2463 if (unique_in != NULL) {
2449 2464 n->subsume_by(unique_in);
2450 2465 }
2451 2466 }
2452 2467 break;
2453 2468
2454 2469 #endif
2455 2470
2456 2471 case Op_ModI:
2457 2472 if (UseDivMod) {
2458 2473 // Check if a%b and a/b both exist
2459 2474 Node* d = n->find_similar(Op_DivI);
2460 2475 if (d) {
2461 2476 // Replace them with a fused divmod if supported
2462 2477 Compile* C = Compile::current();
2463 2478 if (Matcher::has_match_rule(Op_DivModI)) {
2464 2479 DivModINode* divmod = DivModINode::make(C, n);
2465 2480 d->subsume_by(divmod->div_proj());
2466 2481 n->subsume_by(divmod->mod_proj());
2467 2482 } else {
2468 2483 // replace a%b with a-((a/b)*b)
2469 2484 Node* mult = new (C, 3) MulINode(d, d->in(2));
2470 2485 Node* sub = new (C, 3) SubINode(d->in(1), mult);
2471 2486 n->subsume_by( sub );
2472 2487 }
2473 2488 }
2474 2489 }
2475 2490 break;
2476 2491
2477 2492 case Op_ModL:
2478 2493 if (UseDivMod) {
2479 2494 // Check if a%b and a/b both exist
2480 2495 Node* d = n->find_similar(Op_DivL);
2481 2496 if (d) {
2482 2497 // Replace them with a fused divmod if supported
2483 2498 Compile* C = Compile::current();
2484 2499 if (Matcher::has_match_rule(Op_DivModL)) {
2485 2500 DivModLNode* divmod = DivModLNode::make(C, n);
2486 2501 d->subsume_by(divmod->div_proj());
2487 2502 n->subsume_by(divmod->mod_proj());
2488 2503 } else {
2489 2504 // replace a%b with a-((a/b)*b)
2490 2505 Node* mult = new (C, 3) MulLNode(d, d->in(2));
2491 2506 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
2492 2507 n->subsume_by( sub );
2493 2508 }
2494 2509 }
2495 2510 }
2496 2511 break;
2497 2512
2498 2513 case Op_Load16B:
2499 2514 case Op_Load8B:
2500 2515 case Op_Load4B:
2501 2516 case Op_Load8S:
2502 2517 case Op_Load4S:
2503 2518 case Op_Load2S:
2504 2519 case Op_Load8C:
2505 2520 case Op_Load4C:
2506 2521 case Op_Load2C:
2507 2522 case Op_Load4I:
2508 2523 case Op_Load2I:
2509 2524 case Op_Load2L:
2510 2525 case Op_Load4F:
2511 2526 case Op_Load2F:
2512 2527 case Op_Load2D:
2513 2528 case Op_Store16B:
2514 2529 case Op_Store8B:
2515 2530 case Op_Store4B:
2516 2531 case Op_Store8C:
2517 2532 case Op_Store4C:
2518 2533 case Op_Store2C:
2519 2534 case Op_Store4I:
2520 2535 case Op_Store2I:
2521 2536 case Op_Store2L:
2522 2537 case Op_Store4F:
2523 2538 case Op_Store2F:
2524 2539 case Op_Store2D:
2525 2540 break;
2526 2541
2527 2542 case Op_PackB:
2528 2543 case Op_PackS:
2529 2544 case Op_PackC:
2530 2545 case Op_PackI:
2531 2546 case Op_PackF:
2532 2547 case Op_PackL:
2533 2548 case Op_PackD:
2534 2549 if (n->req()-1 > 2) {
2535 2550 // Replace many operand PackNodes with a binary tree for matching
2536 2551 PackNode* p = (PackNode*) n;
2537 2552 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2538 2553 n->subsume_by(btp);
2539 2554 }
2540 2555 break;
2541 2556 case Op_Loop:
2542 2557 case Op_CountedLoop:
2543 2558 if (n->as_Loop()->is_inner_loop()) {
2544 2559 frc.inc_inner_loop_count();
2545 2560 }
2546 2561 break;
2547 2562 default:
2548 2563 assert( !n->is_Call(), "" );
2549 2564 assert( !n->is_Mem(), "" );
2550 2565 break;
2551 2566 }
2552 2567
2553 2568 // Collect CFG split points
2554 2569 if (n->is_MultiBranch())
2555 2570 frc._tests.push(n);
2556 2571 }
2557 2572
2558 2573 //------------------------------final_graph_reshaping_walk---------------------
2559 2574 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2560 2575 // requires that the walk visits a node's inputs before visiting the node.
2561 2576 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2562 2577 ResourceArea *area = Thread::current()->resource_area();
2563 2578 Unique_Node_List sfpt(area);
2564 2579
2565 2580 frc._visited.set(root->_idx); // first, mark node as visited
2566 2581 uint cnt = root->req();
2567 2582 Node *n = root;
2568 2583 uint i = 0;
2569 2584 while (true) {
2570 2585 if (i < cnt) {
2571 2586 // Place all non-visited non-null inputs onto stack
2572 2587 Node* m = n->in(i);
2573 2588 ++i;
2574 2589 if (m != NULL && !frc._visited.test_set(m->_idx)) {
2575 2590 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2576 2591 sfpt.push(m);
2577 2592 cnt = m->req();
2578 2593 nstack.push(n, i); // put on stack parent and next input's index
2579 2594 n = m;
2580 2595 i = 0;
2581 2596 }
2582 2597 } else {
2583 2598 // Now do post-visit work
2584 2599 final_graph_reshaping_impl( n, frc );
2585 2600 if (nstack.is_empty())
2586 2601 break; // finished
2587 2602 n = nstack.node(); // Get node from stack
2588 2603 cnt = n->req();
2589 2604 i = nstack.index();
2590 2605 nstack.pop(); // Shift to the next node on stack
2591 2606 }
2592 2607 }
2593 2608
2594 2609 // Skip next transformation if compressed oops are not used.
2595 2610 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2596 2611 return;
2597 2612
2598 2613 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2599 2614 // It could be done for an uncommon traps or any safepoints/calls
2600 2615 // if the DecodeN node is referenced only in a debug info.
2601 2616 while (sfpt.size() > 0) {
2602 2617 n = sfpt.pop();
2603 2618 JVMState *jvms = n->as_SafePoint()->jvms();
2604 2619 assert(jvms != NULL, "sanity");
2605 2620 int start = jvms->debug_start();
2606 2621 int end = n->req();
2607 2622 bool is_uncommon = (n->is_CallStaticJava() &&
2608 2623 n->as_CallStaticJava()->uncommon_trap_request() != 0);
2609 2624 for (int j = start; j < end; j++) {
2610 2625 Node* in = n->in(j);
2611 2626 if (in->is_DecodeN()) {
2612 2627 bool safe_to_skip = true;
2613 2628 if (!is_uncommon ) {
2614 2629 // Is it safe to skip?
2615 2630 for (uint i = 0; i < in->outcnt(); i++) {
2616 2631 Node* u = in->raw_out(i);
2617 2632 if (!u->is_SafePoint() ||
2618 2633 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2619 2634 safe_to_skip = false;
2620 2635 }
2621 2636 }
2622 2637 }
2623 2638 if (safe_to_skip) {
2624 2639 n->set_req(j, in->in(1));
2625 2640 }
2626 2641 if (in->outcnt() == 0) {
2627 2642 in->disconnect_inputs(NULL);
2628 2643 }
2629 2644 }
2630 2645 }
2631 2646 }
2632 2647 }
2633 2648
2634 2649 //------------------------------final_graph_reshaping--------------------------
2635 2650 // Final Graph Reshaping.
2636 2651 //
2637 2652 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2638 2653 // and not commoned up and forced early. Must come after regular
2639 2654 // optimizations to avoid GVN undoing the cloning. Clone constant
2640 2655 // inputs to Loop Phis; these will be split by the allocator anyways.
2641 2656 // Remove Opaque nodes.
2642 2657 // (2) Move last-uses by commutative operations to the left input to encourage
2643 2658 // Intel update-in-place two-address operations and better register usage
2644 2659 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
2645 2660 // calls canonicalizing them back.
2646 2661 // (3) Count the number of double-precision FP ops, single-precision FP ops
2647 2662 // and call sites. On Intel, we can get correct rounding either by
2648 2663 // forcing singles to memory (requires extra stores and loads after each
2649 2664 // FP bytecode) or we can set a rounding mode bit (requires setting and
2650 2665 // clearing the mode bit around call sites). The mode bit is only used
2651 2666 // if the relative frequency of single FP ops to calls is low enough.
2652 2667 // This is a key transform for SPEC mpeg_audio.
2653 2668 // (4) Detect infinite loops; blobs of code reachable from above but not
2654 2669 // below. Several of the Code_Gen algorithms fail on such code shapes,
2655 2670 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
2656 2671 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
2657 2672 // Detection is by looking for IfNodes where only 1 projection is
2658 2673 // reachable from below or CatchNodes missing some targets.
2659 2674 // (5) Assert for insane oop offsets in debug mode.
2660 2675
2661 2676 bool Compile::final_graph_reshaping() {
2662 2677 // an infinite loop may have been eliminated by the optimizer,
2663 2678 // in which case the graph will be empty.
2664 2679 if (root()->req() == 1) {
2665 2680 record_method_not_compilable("trivial infinite loop");
2666 2681 return true;
2667 2682 }
2668 2683
2669 2684 Final_Reshape_Counts frc;
2670 2685
2671 2686 // Visit everybody reachable!
2672 2687 // Allocate stack of size C->unique()/2 to avoid frequent realloc
2673 2688 Node_Stack nstack(unique() >> 1);
2674 2689 final_graph_reshaping_walk(nstack, root(), frc);
2675 2690
2676 2691 // Check for unreachable (from below) code (i.e., infinite loops).
2677 2692 for( uint i = 0; i < frc._tests.size(); i++ ) {
2678 2693 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2679 2694 // Get number of CFG targets.
2680 2695 // Note that PCTables include exception targets after calls.
2681 2696 uint required_outcnt = n->required_outcnt();
2682 2697 if (n->outcnt() != required_outcnt) {
2683 2698 // Check for a few special cases. Rethrow Nodes never take the
2684 2699 // 'fall-thru' path, so expected kids is 1 less.
2685 2700 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2686 2701 if (n->in(0)->in(0)->is_Call()) {
2687 2702 CallNode *call = n->in(0)->in(0)->as_Call();
2688 2703 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2689 2704 required_outcnt--; // Rethrow always has 1 less kid
2690 2705 } else if (call->req() > TypeFunc::Parms &&
2691 2706 call->is_CallDynamicJava()) {
2692 2707 // Check for null receiver. In such case, the optimizer has
2693 2708 // detected that the virtual call will always result in a null
2694 2709 // pointer exception. The fall-through projection of this CatchNode
2695 2710 // will not be populated.
2696 2711 Node *arg0 = call->in(TypeFunc::Parms);
2697 2712 if (arg0->is_Type() &&
2698 2713 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2699 2714 required_outcnt--;
2700 2715 }
2701 2716 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2702 2717 call->req() > TypeFunc::Parms+1 &&
2703 2718 call->is_CallStaticJava()) {
2704 2719 // Check for negative array length. In such case, the optimizer has
2705 2720 // detected that the allocation attempt will always result in an
2706 2721 // exception. There is no fall-through projection of this CatchNode .
2707 2722 Node *arg1 = call->in(TypeFunc::Parms+1);
2708 2723 if (arg1->is_Type() &&
2709 2724 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2710 2725 required_outcnt--;
2711 2726 }
2712 2727 }
2713 2728 }
2714 2729 }
2715 2730 // Recheck with a better notion of 'required_outcnt'
2716 2731 if (n->outcnt() != required_outcnt) {
2717 2732 record_method_not_compilable("malformed control flow");
2718 2733 return true; // Not all targets reachable!
2719 2734 }
2720 2735 }
2721 2736 // Check that I actually visited all kids. Unreached kids
2722 2737 // must be infinite loops.
2723 2738 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2724 2739 if (!frc._visited.test(n->fast_out(j)->_idx)) {
2725 2740 record_method_not_compilable("infinite loop");
2726 2741 return true; // Found unvisited kid; must be unreach
2727 2742 }
2728 2743 }
2729 2744
2730 2745 // If original bytecodes contained a mixture of floats and doubles
2731 2746 // check if the optimizer has made it homogenous, item (3).
2732 2747 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2733 2748 frc.get_float_count() > 32 &&
2734 2749 frc.get_double_count() == 0 &&
2735 2750 (10 * frc.get_call_count() < frc.get_float_count()) ) {
2736 2751 set_24_bit_selection_and_mode( false, true );
2737 2752 }
2738 2753
2739 2754 set_java_calls(frc.get_java_call_count());
2740 2755 set_inner_loops(frc.get_inner_loop_count());
2741 2756
2742 2757 // No infinite loops, no reason to bail out.
2743 2758 return false;
2744 2759 }
2745 2760
2746 2761 //-----------------------------too_many_traps----------------------------------
2747 2762 // Report if there are too many traps at the current method and bci.
2748 2763 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2749 2764 bool Compile::too_many_traps(ciMethod* method,
2750 2765 int bci,
2751 2766 Deoptimization::DeoptReason reason) {
2752 2767 ciMethodData* md = method->method_data();
2753 2768 if (md->is_empty()) {
2754 2769 // Assume the trap has not occurred, or that it occurred only
2755 2770 // because of a transient condition during start-up in the interpreter.
2756 2771 return false;
2757 2772 }
2758 2773 if (md->has_trap_at(bci, reason) != 0) {
2759 2774 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2760 2775 // Also, if there are multiple reasons, or if there is no per-BCI record,
2761 2776 // assume the worst.
2762 2777 if (log())
2763 2778 log()->elem("observe trap='%s' count='%d'",
2764 2779 Deoptimization::trap_reason_name(reason),
2765 2780 md->trap_count(reason));
2766 2781 return true;
2767 2782 } else {
2768 2783 // Ignore method/bci and see if there have been too many globally.
2769 2784 return too_many_traps(reason, md);
2770 2785 }
2771 2786 }
2772 2787
2773 2788 // Less-accurate variant which does not require a method and bci.
2774 2789 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2775 2790 ciMethodData* logmd) {
2776 2791 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2777 2792 // Too many traps globally.
2778 2793 // Note that we use cumulative trap_count, not just md->trap_count.
2779 2794 if (log()) {
2780 2795 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2781 2796 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2782 2797 Deoptimization::trap_reason_name(reason),
2783 2798 mcount, trap_count(reason));
2784 2799 }
2785 2800 return true;
2786 2801 } else {
2787 2802 // The coast is clear.
2788 2803 return false;
2789 2804 }
2790 2805 }
2791 2806
2792 2807 //--------------------------too_many_recompiles--------------------------------
2793 2808 // Report if there are too many recompiles at the current method and bci.
2794 2809 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2795 2810 // Is not eager to return true, since this will cause the compiler to use
2796 2811 // Action_none for a trap point, to avoid too many recompilations.
2797 2812 bool Compile::too_many_recompiles(ciMethod* method,
2798 2813 int bci,
2799 2814 Deoptimization::DeoptReason reason) {
2800 2815 ciMethodData* md = method->method_data();
2801 2816 if (md->is_empty()) {
2802 2817 // Assume the trap has not occurred, or that it occurred only
2803 2818 // because of a transient condition during start-up in the interpreter.
2804 2819 return false;
2805 2820 }
2806 2821 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2807 2822 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2808 2823 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
2809 2824 Deoptimization::DeoptReason per_bc_reason
2810 2825 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2811 2826 if ((per_bc_reason == Deoptimization::Reason_none
2812 2827 || md->has_trap_at(bci, reason) != 0)
2813 2828 // The trap frequency measure we care about is the recompile count:
2814 2829 && md->trap_recompiled_at(bci)
2815 2830 && md->overflow_recompile_count() >= bc_cutoff) {
2816 2831 // Do not emit a trap here if it has already caused recompilations.
2817 2832 // Also, if there are multiple reasons, or if there is no per-BCI record,
2818 2833 // assume the worst.
2819 2834 if (log())
2820 2835 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2821 2836 Deoptimization::trap_reason_name(reason),
2822 2837 md->trap_count(reason),
2823 2838 md->overflow_recompile_count());
2824 2839 return true;
2825 2840 } else if (trap_count(reason) != 0
2826 2841 && decompile_count() >= m_cutoff) {
2827 2842 // Too many recompiles globally, and we have seen this sort of trap.
2828 2843 // Use cumulative decompile_count, not just md->decompile_count.
2829 2844 if (log())
2830 2845 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2831 2846 Deoptimization::trap_reason_name(reason),
2832 2847 md->trap_count(reason), trap_count(reason),
2833 2848 md->decompile_count(), decompile_count());
2834 2849 return true;
2835 2850 } else {
2836 2851 // The coast is clear.
2837 2852 return false;
2838 2853 }
2839 2854 }
2840 2855
2841 2856
2842 2857 #ifndef PRODUCT
2843 2858 //------------------------------verify_graph_edges---------------------------
2844 2859 // Walk the Graph and verify that there is a one-to-one correspondence
2845 2860 // between Use-Def edges and Def-Use edges in the graph.
2846 2861 void Compile::verify_graph_edges(bool no_dead_code) {
2847 2862 if (VerifyGraphEdges) {
2848 2863 ResourceArea *area = Thread::current()->resource_area();
2849 2864 Unique_Node_List visited(area);
2850 2865 // Call recursive graph walk to check edges
2851 2866 _root->verify_edges(visited);
2852 2867 if (no_dead_code) {
2853 2868 // Now make sure that no visited node is used by an unvisited node.
2854 2869 bool dead_nodes = 0;
2855 2870 Unique_Node_List checked(area);
2856 2871 while (visited.size() > 0) {
2857 2872 Node* n = visited.pop();
2858 2873 checked.push(n);
2859 2874 for (uint i = 0; i < n->outcnt(); i++) {
2860 2875 Node* use = n->raw_out(i);
2861 2876 if (checked.member(use)) continue; // already checked
2862 2877 if (visited.member(use)) continue; // already in the graph
2863 2878 if (use->is_Con()) continue; // a dead ConNode is OK
2864 2879 // At this point, we have found a dead node which is DU-reachable.
2865 2880 if (dead_nodes++ == 0)
2866 2881 tty->print_cr("*** Dead nodes reachable via DU edges:");
2867 2882 use->dump(2);
2868 2883 tty->print_cr("---");
2869 2884 checked.push(use); // No repeats; pretend it is now checked.
2870 2885 }
2871 2886 }
2872 2887 assert(dead_nodes == 0, "using nodes must be reachable from root");
2873 2888 }
2874 2889 }
2875 2890 }
2876 2891 #endif
2877 2892
2878 2893 // The Compile object keeps track of failure reasons separately from the ciEnv.
2879 2894 // This is required because there is not quite a 1-1 relation between the
2880 2895 // ciEnv and its compilation task and the Compile object. Note that one
2881 2896 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2882 2897 // to backtrack and retry without subsuming loads. Other than this backtracking
2883 2898 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2884 2899 // by the logic in C2Compiler.
2885 2900 void Compile::record_failure(const char* reason) {
2886 2901 if (log() != NULL) {
2887 2902 log()->elem("failure reason='%s' phase='compile'", reason);
2888 2903 }
2889 2904 if (_failure_reason == NULL) {
2890 2905 // Record the first failure reason.
2891 2906 _failure_reason = reason;
2892 2907 }
2893 2908 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2894 2909 C->print_method(_failure_reason);
2895 2910 }
2896 2911 _root = NULL; // flush the graph, too
2897 2912 }
2898 2913
2899 2914 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
2900 2915 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
2901 2916 {
2902 2917 if (dolog) {
2903 2918 C = Compile::current();
2904 2919 _log = C->log();
2905 2920 } else {
2906 2921 C = NULL;
2907 2922 _log = NULL;
2908 2923 }
2909 2924 if (_log != NULL) {
2910 2925 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
2911 2926 _log->stamp();
2912 2927 _log->end_head();
2913 2928 }
2914 2929 }
2915 2930
2916 2931 Compile::TracePhase::~TracePhase() {
2917 2932 if (_log != NULL) {
2918 2933 _log->done("phase nodes='%d'", C->unique());
2919 2934 }
2920 2935 }
2921 2936
2922 2937 //=============================================================================
2923 2938 // Two Constant's are equal when the type and the value are equal.
2924 2939 bool Compile::Constant::operator==(const Constant& other) {
2925 2940 if (type() != other.type() ) return false;
2926 2941 if (can_be_reused() != other.can_be_reused()) return false;
2927 2942 // For floating point values we compare the bit pattern.
2928 2943 switch (type()) {
2929 2944 case T_FLOAT: return (_value.i == other._value.i);
2930 2945 case T_LONG:
2931 2946 case T_DOUBLE: return (_value.j == other._value.j);
2932 2947 case T_OBJECT:
2933 2948 case T_ADDRESS: return (_value.l == other._value.l);
2934 2949 case T_VOID: return (_value.l == other._value.l); // jump-table entries
2935 2950 default: ShouldNotReachHere();
2936 2951 }
2937 2952 return false;
2938 2953 }
2939 2954
2940 2955 // Emit constants grouped in the following order:
2941 2956 static BasicType type_order[] = {
2942 2957 T_FLOAT, // 32-bit
2943 2958 T_OBJECT, // 32 or 64-bit
2944 2959 T_ADDRESS, // 32 or 64-bit
2945 2960 T_DOUBLE, // 64-bit
2946 2961 T_LONG, // 64-bit
2947 2962 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
2948 2963 T_ILLEGAL
2949 2964 };
2950 2965
2951 2966 static int type_to_size_in_bytes(BasicType t) {
2952 2967 switch (t) {
2953 2968 case T_LONG: return sizeof(jlong );
2954 2969 case T_FLOAT: return sizeof(jfloat );
2955 2970 case T_DOUBLE: return sizeof(jdouble);
2956 2971 // We use T_VOID as marker for jump-table entries (labels) which
2957 2972 // need an interal word relocation.
2958 2973 case T_VOID:
2959 2974 case T_ADDRESS:
2960 2975 case T_OBJECT: return sizeof(jobject);
2961 2976 }
2962 2977
2963 2978 ShouldNotReachHere();
2964 2979 return -1;
2965 2980 }
2966 2981
2967 2982 void Compile::ConstantTable::calculate_offsets_and_size() {
2968 2983 int size = 0;
2969 2984 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
2970 2985 BasicType type = type_order[t];
2971 2986
2972 2987 for (int i = 0; i < _constants.length(); i++) {
2973 2988 Constant con = _constants.at(i);
2974 2989 if (con.type() != type) continue; // Skip other types.
2975 2990
2976 2991 // Align size for type.
2977 2992 int typesize = type_to_size_in_bytes(con.type());
2978 2993 size = align_size_up(size, typesize);
2979 2994
2980 2995 // Set offset.
2981 2996 con.set_offset(size);
2982 2997 _constants.at_put(i, con);
2983 2998
2984 2999 // Add type size.
2985 3000 size = size + typesize;
2986 3001 }
2987 3002 }
2988 3003
2989 3004 // Align size up to the next section start (which is insts; see
2990 3005 // CodeBuffer::align_at_start).
2991 3006 assert(_size == -1, "already set?");
2992 3007 _size = align_size_up(size, CodeEntryAlignment);
2993 3008
2994 3009 if (Matcher::constant_table_absolute_addressing) {
2995 3010 set_table_base_offset(0); // No table base offset required
2996 3011 } else {
2997 3012 if (UseRDPCForConstantTableBase) {
2998 3013 // table base offset is set in MachConstantBaseNode::emit
2999 3014 } else {
3000 3015 // When RDPC is not used, the table base is set into the middle of
3001 3016 // the constant table.
3002 3017 int half_size = _size / 2;
3003 3018 assert(half_size * 2 == _size, "sanity");
3004 3019 set_table_base_offset(-half_size);
3005 3020 }
3006 3021 }
3007 3022 }
3008 3023
3009 3024 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3010 3025 MacroAssembler _masm(&cb);
3011 3026 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3012 3027 BasicType type = type_order[t];
3013 3028
3014 3029 for (int i = 0; i < _constants.length(); i++) {
3015 3030 Constant con = _constants.at(i);
3016 3031 if (con.type() != type) continue; // Skip other types.
3017 3032
3018 3033 address constant_addr;
3019 3034 switch (con.type()) {
3020 3035 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
3021 3036 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3022 3037 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3023 3038 case T_OBJECT: {
3024 3039 jobject obj = con.get_jobject();
3025 3040 int oop_index = _masm.oop_recorder()->find_index(obj);
3026 3041 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3027 3042 break;
3028 3043 }
3029 3044 case T_ADDRESS: {
3030 3045 address addr = (address) con.get_jobject();
3031 3046 constant_addr = _masm.address_constant(addr);
3032 3047 break;
3033 3048 }
3034 3049 // We use T_VOID as marker for jump-table entries (labels) which
3035 3050 // need an interal word relocation.
3036 3051 case T_VOID: {
3037 3052 // Write a dummy word. The real value is filled in later
3038 3053 // in fill_jump_table_in_constant_table.
3039 3054 address addr = (address) con.get_jobject();
3040 3055 constant_addr = _masm.address_constant(addr);
3041 3056 break;
3042 3057 }
3043 3058 default: ShouldNotReachHere();
3044 3059 }
3045 3060 assert(constant_addr != NULL, "consts section too small");
3046 3061 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3047 3062 }
3048 3063 }
3049 3064 }
3050 3065
3051 3066 int Compile::ConstantTable::find_offset(Constant& con) const {
3052 3067 int idx = _constants.find(con);
3053 3068 assert(idx != -1, "constant must be in constant table");
3054 3069 int offset = _constants.at(idx).offset();
3055 3070 assert(offset != -1, "constant table not emitted yet?");
3056 3071 return offset;
3057 3072 }
3058 3073
3059 3074 void Compile::ConstantTable::add(Constant& con) {
3060 3075 if (con.can_be_reused()) {
3061 3076 int idx = _constants.find(con);
3062 3077 if (idx != -1 && _constants.at(idx).can_be_reused()) {
3063 3078 return;
3064 3079 }
3065 3080 }
3066 3081 (void) _constants.append(con);
3067 3082 }
3068 3083
3069 3084 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
3070 3085 Constant con(type, value);
3071 3086 add(con);
3072 3087 return con;
3073 3088 }
3074 3089
3075 3090 Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
3076 3091 jvalue value;
3077 3092 BasicType type = oper->type()->basic_type();
3078 3093 switch (type) {
3079 3094 case T_LONG: value.j = oper->constantL(); break;
3080 3095 case T_FLOAT: value.f = oper->constantF(); break;
3081 3096 case T_DOUBLE: value.d = oper->constantD(); break;
3082 3097 case T_OBJECT:
3083 3098 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3084 3099 default: ShouldNotReachHere();
3085 3100 }
3086 3101 return add(type, value);
3087 3102 }
3088 3103
3089 3104 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
3090 3105 jvalue value;
3091 3106 // We can use the node pointer here to identify the right jump-table
3092 3107 // as this method is called from Compile::Fill_buffer right before
3093 3108 // the MachNodes are emitted and the jump-table is filled (means the
3094 3109 // MachNode pointers do not change anymore).
3095 3110 value.l = (jobject) n;
3096 3111 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused.
3097 3112 for (uint i = 0; i < n->outcnt(); i++) {
3098 3113 add(con);
3099 3114 }
3100 3115 return con;
3101 3116 }
3102 3117
3103 3118 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3104 3119 // If called from Compile::scratch_emit_size do nothing.
3105 3120 if (Compile::current()->in_scratch_emit_size()) return;
3106 3121
3107 3122 assert(labels.is_nonempty(), "must be");
3108 3123 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
3109 3124
3110 3125 // Since MachConstantNode::constant_offset() also contains
3111 3126 // table_base_offset() we need to subtract the table_base_offset()
3112 3127 // to get the plain offset into the constant table.
3113 3128 int offset = n->constant_offset() - table_base_offset();
3114 3129
3115 3130 MacroAssembler _masm(&cb);
3116 3131 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3117 3132
3118 3133 for (int i = 0; i < labels.length(); i++) {
3119 3134 address* constant_addr = &jump_table_base[i];
3120 3135 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
3121 3136 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3122 3137 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3123 3138 }
3124 3139 }
↓ open down ↓ |
2465 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX