Print this page
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/opto/library_call.cpp
+++ new/src/share/vm/opto/library_call.cpp
1 1 /*
2 2 * Copyright (c) 1999, 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 "classfile/systemDictionary.hpp"
27 27 #include "classfile/vmSymbols.hpp"
28 28 #include "compiler/compileLog.hpp"
29 29 #include "oops/objArrayKlass.hpp"
30 30 #include "opto/addnode.hpp"
31 31 #include "opto/callGenerator.hpp"
32 32 #include "opto/cfgnode.hpp"
33 33 #include "opto/idealKit.hpp"
34 34 #include "opto/mulnode.hpp"
35 35 #include "opto/parse.hpp"
36 36 #include "opto/runtime.hpp"
37 37 #include "opto/subnode.hpp"
38 38 #include "prims/nativeLookup.hpp"
39 39 #include "runtime/sharedRuntime.hpp"
40 40
41 41 class LibraryIntrinsic : public InlineCallGenerator {
42 42 // Extend the set of intrinsics known to the runtime:
43 43 public:
44 44 private:
45 45 bool _is_virtual;
46 46 vmIntrinsics::ID _intrinsic_id;
47 47
48 48 public:
49 49 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
50 50 : InlineCallGenerator(m),
51 51 _is_virtual(is_virtual),
52 52 _intrinsic_id(id)
53 53 {
54 54 }
55 55 virtual bool is_intrinsic() const { return true; }
56 56 virtual bool is_virtual() const { return _is_virtual; }
57 57 virtual JVMState* generate(JVMState* jvms);
58 58 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
59 59 };
60 60
61 61
62 62 // Local helper class for LibraryIntrinsic:
63 63 class LibraryCallKit : public GraphKit {
64 64 private:
65 65 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
66 66
67 67 public:
68 68 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
69 69 : GraphKit(caller),
70 70 _intrinsic(intrinsic)
71 71 {
72 72 }
73 73
74 74 ciMethod* caller() const { return jvms()->method(); }
75 75 int bci() const { return jvms()->bci(); }
76 76 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
77 77 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
78 78 ciMethod* callee() const { return _intrinsic->method(); }
79 79 ciSignature* signature() const { return callee()->signature(); }
80 80 int arg_size() const { return callee()->arg_size(); }
81 81
82 82 bool try_to_inline();
83 83
84 84 // Helper functions to inline natives
85 85 void push_result(RegionNode* region, PhiNode* value);
86 86 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
87 87 Node* generate_slow_guard(Node* test, RegionNode* region);
88 88 Node* generate_fair_guard(Node* test, RegionNode* region);
89 89 Node* generate_negative_guard(Node* index, RegionNode* region,
90 90 // resulting CastII of index:
91 91 Node* *pos_index = NULL);
92 92 Node* generate_nonpositive_guard(Node* index, bool never_negative,
93 93 // resulting CastII of index:
94 94 Node* *pos_index = NULL);
95 95 Node* generate_limit_guard(Node* offset, Node* subseq_length,
96 96 Node* array_length,
97 97 RegionNode* region);
98 98 Node* generate_current_thread(Node* &tls_output);
99 99 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
100 100 bool disjoint_bases, const char* &name, bool dest_uninitialized);
101 101 Node* load_mirror_from_klass(Node* klass);
102 102 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
103 103 int nargs,
104 104 RegionNode* region, int null_path,
105 105 int offset);
106 106 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
107 107 RegionNode* region, int null_path) {
108 108 int offset = java_lang_Class::klass_offset_in_bytes();
109 109 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
110 110 region, null_path,
111 111 offset);
112 112 }
113 113 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
114 114 int nargs,
115 115 RegionNode* region, int null_path) {
116 116 int offset = java_lang_Class::array_klass_offset_in_bytes();
117 117 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
118 118 region, null_path,
119 119 offset);
120 120 }
121 121 Node* generate_access_flags_guard(Node* kls,
122 122 int modifier_mask, int modifier_bits,
123 123 RegionNode* region);
124 124 Node* generate_interface_guard(Node* kls, RegionNode* region);
125 125 Node* generate_array_guard(Node* kls, RegionNode* region) {
126 126 return generate_array_guard_common(kls, region, false, false);
127 127 }
128 128 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
129 129 return generate_array_guard_common(kls, region, false, true);
130 130 }
131 131 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
132 132 return generate_array_guard_common(kls, region, true, false);
133 133 }
134 134 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
135 135 return generate_array_guard_common(kls, region, true, true);
136 136 }
137 137 Node* generate_array_guard_common(Node* kls, RegionNode* region,
138 138 bool obj_array, bool not_array);
139 139 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
140 140 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
141 141 bool is_virtual = false, bool is_static = false);
142 142 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
143 143 return generate_method_call(method_id, false, true);
144 144 }
145 145 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
146 146 return generate_method_call(method_id, true, false);
147 147 }
148 148
149 149 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
150 150 bool inline_string_compareTo();
151 151 bool inline_string_indexOf();
152 152 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
153 153 bool inline_string_equals();
154 154 Node* pop_math_arg();
155 155 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
156 156 bool inline_math_native(vmIntrinsics::ID id);
157 157 bool inline_trig(vmIntrinsics::ID id);
158 158 bool inline_trans(vmIntrinsics::ID id);
159 159 bool inline_abs(vmIntrinsics::ID id);
160 160 bool inline_sqrt(vmIntrinsics::ID id);
161 161 bool inline_pow(vmIntrinsics::ID id);
162 162 bool inline_exp(vmIntrinsics::ID id);
163 163 bool inline_min_max(vmIntrinsics::ID id);
164 164 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
165 165 // This returns Type::AnyPtr, RawPtr, or OopPtr.
166 166 int classify_unsafe_addr(Node* &base, Node* &offset);
167 167 Node* make_unsafe_address(Node* base, Node* offset);
168 168 // Helper for inline_unsafe_access.
169 169 // Generates the guards that check whether the result of
170 170 // Unsafe.getObject should be recorded in an SATB log buffer.
171 171 void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val);
172 172 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
173 173 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
174 174 bool inline_unsafe_allocate();
175 175 bool inline_unsafe_copyMemory();
176 176 bool inline_native_currentThread();
177 177 bool inline_native_time_funcs(bool isNano);
178 178 bool inline_native_isInterrupted();
179 179 bool inline_native_Class_query(vmIntrinsics::ID id);
180 180 bool inline_native_subtype_check();
181 181
182 182 bool inline_native_newArray();
183 183 bool inline_native_getLength();
184 184 bool inline_array_copyOf(bool is_copyOfRange);
185 185 bool inline_array_equals();
186 186 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
187 187 bool inline_native_clone(bool is_virtual);
188 188 bool inline_native_Reflection_getCallerClass();
189 189 bool inline_native_AtomicLong_get();
190 190 bool inline_native_AtomicLong_attemptUpdate();
191 191 bool is_method_invoke_or_aux_frame(JVMState* jvms);
192 192 // Helper function for inlining native object hash method
193 193 bool inline_native_hashcode(bool is_virtual, bool is_static);
194 194 bool inline_native_getClass();
195 195
196 196 // Helper functions for inlining arraycopy
197 197 bool inline_arraycopy();
198 198 void generate_arraycopy(const TypePtr* adr_type,
199 199 BasicType basic_elem_type,
200 200 Node* src, Node* src_offset,
201 201 Node* dest, Node* dest_offset,
202 202 Node* copy_length,
203 203 bool disjoint_bases = false,
204 204 bool length_never_negative = false,
205 205 RegionNode* slow_region = NULL);
206 206 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
207 207 RegionNode* slow_region);
208 208 void generate_clear_array(const TypePtr* adr_type,
209 209 Node* dest,
210 210 BasicType basic_elem_type,
211 211 Node* slice_off,
212 212 Node* slice_len,
213 213 Node* slice_end);
214 214 bool generate_block_arraycopy(const TypePtr* adr_type,
215 215 BasicType basic_elem_type,
216 216 AllocateNode* alloc,
217 217 Node* src, Node* src_offset,
218 218 Node* dest, Node* dest_offset,
219 219 Node* dest_size, bool dest_uninitialized);
220 220 void generate_slow_arraycopy(const TypePtr* adr_type,
221 221 Node* src, Node* src_offset,
222 222 Node* dest, Node* dest_offset,
223 223 Node* copy_length, bool dest_uninitialized);
224 224 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
225 225 Node* dest_elem_klass,
226 226 Node* src, Node* src_offset,
227 227 Node* dest, Node* dest_offset,
228 228 Node* copy_length, bool dest_uninitialized);
229 229 Node* generate_generic_arraycopy(const TypePtr* adr_type,
230 230 Node* src, Node* src_offset,
231 231 Node* dest, Node* dest_offset,
232 232 Node* copy_length, bool dest_uninitialized);
233 233 void generate_unchecked_arraycopy(const TypePtr* adr_type,
234 234 BasicType basic_elem_type,
235 235 bool disjoint_bases,
236 236 Node* src, Node* src_offset,
237 237 Node* dest, Node* dest_offset,
238 238 Node* copy_length, bool dest_uninitialized);
239 239 bool inline_unsafe_CAS(BasicType type);
240 240 bool inline_unsafe_ordered_store(BasicType type);
241 241 bool inline_fp_conversions(vmIntrinsics::ID id);
242 242 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
243 243 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
244 244 bool inline_bitCount(vmIntrinsics::ID id);
245 245 bool inline_reverseBytes(vmIntrinsics::ID id);
246 246
247 247 bool inline_reference_get();
248 248 };
249 249
250 250
251 251 //---------------------------make_vm_intrinsic----------------------------
252 252 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
253 253 vmIntrinsics::ID id = m->intrinsic_id();
254 254 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
255 255
256 256 if (DisableIntrinsic[0] != '\0'
257 257 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
258 258 // disabled by a user request on the command line:
259 259 // example: -XX:DisableIntrinsic=_hashCode,_getClass
260 260 return NULL;
261 261 }
262 262
263 263 if (!m->is_loaded()) {
264 264 // do not attempt to inline unloaded methods
265 265 return NULL;
266 266 }
267 267
268 268 // Only a few intrinsics implement a virtual dispatch.
269 269 // They are expensive calls which are also frequently overridden.
270 270 if (is_virtual) {
271 271 switch (id) {
272 272 case vmIntrinsics::_hashCode:
273 273 case vmIntrinsics::_clone:
274 274 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
275 275 break;
276 276 default:
277 277 return NULL;
278 278 }
279 279 }
280 280
281 281 // -XX:-InlineNatives disables nearly all intrinsics:
282 282 if (!InlineNatives) {
283 283 switch (id) {
284 284 case vmIntrinsics::_indexOf:
285 285 case vmIntrinsics::_compareTo:
286 286 case vmIntrinsics::_equals:
287 287 case vmIntrinsics::_equalsC:
288 288 break; // InlineNatives does not control String.compareTo
289 289 default:
290 290 return NULL;
291 291 }
292 292 }
293 293
294 294 switch (id) {
295 295 case vmIntrinsics::_compareTo:
296 296 if (!SpecialStringCompareTo) return NULL;
297 297 break;
298 298 case vmIntrinsics::_indexOf:
299 299 if (!SpecialStringIndexOf) return NULL;
300 300 break;
301 301 case vmIntrinsics::_equals:
302 302 if (!SpecialStringEquals) return NULL;
303 303 break;
304 304 case vmIntrinsics::_equalsC:
305 305 if (!SpecialArraysEquals) return NULL;
306 306 break;
307 307 case vmIntrinsics::_arraycopy:
308 308 if (!InlineArrayCopy) return NULL;
309 309 break;
310 310 case vmIntrinsics::_copyMemory:
311 311 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
312 312 if (!InlineArrayCopy) return NULL;
313 313 break;
314 314 case vmIntrinsics::_hashCode:
315 315 if (!InlineObjectHash) return NULL;
316 316 break;
317 317 case vmIntrinsics::_clone:
318 318 case vmIntrinsics::_copyOf:
319 319 case vmIntrinsics::_copyOfRange:
320 320 if (!InlineObjectCopy) return NULL;
321 321 // These also use the arraycopy intrinsic mechanism:
322 322 if (!InlineArrayCopy) return NULL;
323 323 break;
324 324 case vmIntrinsics::_checkIndex:
325 325 // We do not intrinsify this. The optimizer does fine with it.
326 326 return NULL;
327 327
328 328 case vmIntrinsics::_get_AtomicLong:
329 329 case vmIntrinsics::_attemptUpdate:
330 330 if (!InlineAtomicLong) return NULL;
331 331 break;
332 332
333 333 case vmIntrinsics::_getCallerClass:
334 334 if (!UseNewReflection) return NULL;
335 335 if (!InlineReflectionGetCallerClass) return NULL;
336 336 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
337 337 break;
338 338
339 339 case vmIntrinsics::_bitCount_i:
340 340 case vmIntrinsics::_bitCount_l:
341 341 if (!UsePopCountInstruction) return NULL;
342 342 break;
343 343
344 344 case vmIntrinsics::_Reference_get:
345 345 // It is only when G1 is enabled that we absolutely
346 346 // need to use the intrinsic version of Reference.get()
347 347 // so that the value in the referent field, if necessary,
348 348 // can be registered by the pre-barrier code.
349 349 if (!UseG1GC) return NULL;
350 350 break;
351 351
352 352 default:
353 353 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
354 354 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
355 355 break;
356 356 }
357 357
358 358 // -XX:-InlineClassNatives disables natives from the Class class.
359 359 // The flag applies to all reflective calls, notably Array.newArray
360 360 // (visible to Java programmers as Array.newInstance).
361 361 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
362 362 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
363 363 if (!InlineClassNatives) return NULL;
364 364 }
365 365
366 366 // -XX:-InlineThreadNatives disables natives from the Thread class.
367 367 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
368 368 if (!InlineThreadNatives) return NULL;
369 369 }
370 370
371 371 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
372 372 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
373 373 m->holder()->name() == ciSymbol::java_lang_Float() ||
374 374 m->holder()->name() == ciSymbol::java_lang_Double()) {
375 375 if (!InlineMathNatives) return NULL;
376 376 }
377 377
378 378 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
379 379 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
380 380 if (!InlineUnsafeOps) return NULL;
381 381 }
382 382
383 383 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
384 384 }
385 385
386 386 //----------------------register_library_intrinsics-----------------------
387 387 // Initialize this file's data structures, for each Compile instance.
388 388 void Compile::register_library_intrinsics() {
389 389 // Nothing to do here.
390 390 }
391 391
392 392 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
393 393 LibraryCallKit kit(jvms, this);
394 394 Compile* C = kit.C;
395 395 int nodes = C->unique();
396 396 #ifndef PRODUCT
397 397 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
398 398 char buf[1000];
399 399 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
400 400 tty->print_cr("Intrinsic %s", str);
401 401 }
402 402 #endif
403 403
404 404 if (kit.try_to_inline()) {
405 405 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
406 406 if (jvms->has_method()) {
407 407 // Not a root compile.
408 408 tty->print("Inlining intrinsic %s%s at bci:%d in",
409 409 vmIntrinsics::name_at(intrinsic_id()),
410 410 (is_virtual() ? " (virtual)" : ""), kit.bci());
411 411 kit.caller()->print_short_name(tty);
412 412 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
413 413 } else {
414 414 // Root compile
415 415 tty->print_cr("Generating intrinsic %s%s at bci:%d",
416 416 vmIntrinsics::name_at(intrinsic_id()),
417 417 (is_virtual() ? " (virtual)" : ""), kit.bci());
418 418 }
419 419 }
420 420 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
421 421 if (C->log()) {
422 422 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
423 423 vmIntrinsics::name_at(intrinsic_id()),
424 424 (is_virtual() ? " virtual='1'" : ""),
425 425 C->unique() - nodes);
426 426 }
427 427 return kit.transfer_exceptions_into_jvms();
428 428 }
429 429
430 430 if (PrintIntrinsics) {
431 431 if (jvms->has_method()) {
432 432 // Not a root compile.
433 433 tty->print("Did not inline intrinsic %s%s at bci:%d in",
434 434 vmIntrinsics::name_at(intrinsic_id()),
435 435 (is_virtual() ? " (virtual)" : ""), kit.bci());
436 436 kit.caller()->print_short_name(tty);
437 437 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
438 438 } else {
439 439 // Root compile
440 440 tty->print("Did not generate intrinsic %s%s at bci:%d in",
441 441 vmIntrinsics::name_at(intrinsic_id()),
442 442 (is_virtual() ? " (virtual)" : ""), kit.bci());
443 443 }
444 444 }
445 445 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
446 446 return NULL;
447 447 }
448 448
449 449 bool LibraryCallKit::try_to_inline() {
450 450 // Handle symbolic names for otherwise undistinguished boolean switches:
451 451 const bool is_store = true;
452 452 const bool is_native_ptr = true;
453 453 const bool is_static = true;
454 454
455 455 if (!jvms()->has_method()) {
456 456 // Root JVMState has a null method.
457 457 assert(map()->memory()->Opcode() == Op_Parm, "");
458 458 // Insert the memory aliasing node
459 459 set_all_memory(reset_memory());
460 460 }
461 461 assert(merged_memory(), "");
462 462
463 463 switch (intrinsic_id()) {
464 464 case vmIntrinsics::_hashCode:
465 465 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
466 466 case vmIntrinsics::_identityHashCode:
467 467 return inline_native_hashcode(/*!virtual*/ false, is_static);
468 468 case vmIntrinsics::_getClass:
469 469 return inline_native_getClass();
470 470
471 471 case vmIntrinsics::_dsin:
472 472 case vmIntrinsics::_dcos:
473 473 case vmIntrinsics::_dtan:
474 474 case vmIntrinsics::_dabs:
475 475 case vmIntrinsics::_datan2:
476 476 case vmIntrinsics::_dsqrt:
477 477 case vmIntrinsics::_dexp:
478 478 case vmIntrinsics::_dlog:
479 479 case vmIntrinsics::_dlog10:
480 480 case vmIntrinsics::_dpow:
481 481 return inline_math_native(intrinsic_id());
482 482
483 483 case vmIntrinsics::_min:
484 484 case vmIntrinsics::_max:
485 485 return inline_min_max(intrinsic_id());
486 486
487 487 case vmIntrinsics::_arraycopy:
488 488 return inline_arraycopy();
489 489
490 490 case vmIntrinsics::_compareTo:
491 491 return inline_string_compareTo();
492 492 case vmIntrinsics::_indexOf:
493 493 return inline_string_indexOf();
494 494 case vmIntrinsics::_equals:
495 495 return inline_string_equals();
496 496
497 497 case vmIntrinsics::_getObject:
498 498 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
499 499 case vmIntrinsics::_getBoolean:
500 500 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
501 501 case vmIntrinsics::_getByte:
502 502 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
503 503 case vmIntrinsics::_getShort:
504 504 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
505 505 case vmIntrinsics::_getChar:
506 506 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
507 507 case vmIntrinsics::_getInt:
508 508 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
509 509 case vmIntrinsics::_getLong:
510 510 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
511 511 case vmIntrinsics::_getFloat:
512 512 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
513 513 case vmIntrinsics::_getDouble:
514 514 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
515 515
516 516 case vmIntrinsics::_putObject:
517 517 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
518 518 case vmIntrinsics::_putBoolean:
519 519 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
520 520 case vmIntrinsics::_putByte:
521 521 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
522 522 case vmIntrinsics::_putShort:
523 523 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
524 524 case vmIntrinsics::_putChar:
525 525 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
526 526 case vmIntrinsics::_putInt:
527 527 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
528 528 case vmIntrinsics::_putLong:
529 529 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
530 530 case vmIntrinsics::_putFloat:
531 531 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
532 532 case vmIntrinsics::_putDouble:
533 533 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
534 534
535 535 case vmIntrinsics::_getByte_raw:
536 536 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
537 537 case vmIntrinsics::_getShort_raw:
538 538 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
539 539 case vmIntrinsics::_getChar_raw:
540 540 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
541 541 case vmIntrinsics::_getInt_raw:
542 542 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
543 543 case vmIntrinsics::_getLong_raw:
544 544 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
545 545 case vmIntrinsics::_getFloat_raw:
546 546 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
547 547 case vmIntrinsics::_getDouble_raw:
548 548 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
549 549 case vmIntrinsics::_getAddress_raw:
550 550 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
551 551
552 552 case vmIntrinsics::_putByte_raw:
553 553 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
554 554 case vmIntrinsics::_putShort_raw:
555 555 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
556 556 case vmIntrinsics::_putChar_raw:
557 557 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
558 558 case vmIntrinsics::_putInt_raw:
559 559 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
560 560 case vmIntrinsics::_putLong_raw:
561 561 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
562 562 case vmIntrinsics::_putFloat_raw:
563 563 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
564 564 case vmIntrinsics::_putDouble_raw:
565 565 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
566 566 case vmIntrinsics::_putAddress_raw:
567 567 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
568 568
569 569 case vmIntrinsics::_getObjectVolatile:
570 570 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
571 571 case vmIntrinsics::_getBooleanVolatile:
572 572 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
573 573 case vmIntrinsics::_getByteVolatile:
574 574 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
575 575 case vmIntrinsics::_getShortVolatile:
576 576 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
577 577 case vmIntrinsics::_getCharVolatile:
578 578 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
579 579 case vmIntrinsics::_getIntVolatile:
580 580 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
581 581 case vmIntrinsics::_getLongVolatile:
582 582 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
583 583 case vmIntrinsics::_getFloatVolatile:
584 584 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
585 585 case vmIntrinsics::_getDoubleVolatile:
586 586 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
587 587
588 588 case vmIntrinsics::_putObjectVolatile:
589 589 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
590 590 case vmIntrinsics::_putBooleanVolatile:
591 591 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
592 592 case vmIntrinsics::_putByteVolatile:
593 593 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
594 594 case vmIntrinsics::_putShortVolatile:
595 595 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
596 596 case vmIntrinsics::_putCharVolatile:
597 597 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
598 598 case vmIntrinsics::_putIntVolatile:
599 599 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
600 600 case vmIntrinsics::_putLongVolatile:
601 601 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
602 602 case vmIntrinsics::_putFloatVolatile:
603 603 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
604 604 case vmIntrinsics::_putDoubleVolatile:
605 605 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
606 606
607 607 case vmIntrinsics::_prefetchRead:
608 608 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
609 609 case vmIntrinsics::_prefetchWrite:
610 610 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
611 611 case vmIntrinsics::_prefetchReadStatic:
612 612 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
613 613 case vmIntrinsics::_prefetchWriteStatic:
614 614 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
615 615
616 616 case vmIntrinsics::_compareAndSwapObject:
617 617 return inline_unsafe_CAS(T_OBJECT);
618 618 case vmIntrinsics::_compareAndSwapInt:
619 619 return inline_unsafe_CAS(T_INT);
620 620 case vmIntrinsics::_compareAndSwapLong:
621 621 return inline_unsafe_CAS(T_LONG);
622 622
623 623 case vmIntrinsics::_putOrderedObject:
624 624 return inline_unsafe_ordered_store(T_OBJECT);
625 625 case vmIntrinsics::_putOrderedInt:
626 626 return inline_unsafe_ordered_store(T_INT);
627 627 case vmIntrinsics::_putOrderedLong:
628 628 return inline_unsafe_ordered_store(T_LONG);
629 629
630 630 case vmIntrinsics::_currentThread:
631 631 return inline_native_currentThread();
632 632 case vmIntrinsics::_isInterrupted:
633 633 return inline_native_isInterrupted();
634 634
635 635 case vmIntrinsics::_currentTimeMillis:
636 636 return inline_native_time_funcs(false);
637 637 case vmIntrinsics::_nanoTime:
638 638 return inline_native_time_funcs(true);
639 639 case vmIntrinsics::_allocateInstance:
640 640 return inline_unsafe_allocate();
641 641 case vmIntrinsics::_copyMemory:
642 642 return inline_unsafe_copyMemory();
643 643 case vmIntrinsics::_newArray:
644 644 return inline_native_newArray();
645 645 case vmIntrinsics::_getLength:
646 646 return inline_native_getLength();
647 647 case vmIntrinsics::_copyOf:
648 648 return inline_array_copyOf(false);
649 649 case vmIntrinsics::_copyOfRange:
650 650 return inline_array_copyOf(true);
651 651 case vmIntrinsics::_equalsC:
652 652 return inline_array_equals();
653 653 case vmIntrinsics::_clone:
654 654 return inline_native_clone(intrinsic()->is_virtual());
655 655
656 656 case vmIntrinsics::_isAssignableFrom:
657 657 return inline_native_subtype_check();
658 658
659 659 case vmIntrinsics::_isInstance:
660 660 case vmIntrinsics::_getModifiers:
661 661 case vmIntrinsics::_isInterface:
662 662 case vmIntrinsics::_isArray:
663 663 case vmIntrinsics::_isPrimitive:
664 664 case vmIntrinsics::_getSuperclass:
665 665 case vmIntrinsics::_getComponentType:
666 666 case vmIntrinsics::_getClassAccessFlags:
667 667 return inline_native_Class_query(intrinsic_id());
668 668
669 669 case vmIntrinsics::_floatToRawIntBits:
670 670 case vmIntrinsics::_floatToIntBits:
671 671 case vmIntrinsics::_intBitsToFloat:
672 672 case vmIntrinsics::_doubleToRawLongBits:
673 673 case vmIntrinsics::_doubleToLongBits:
674 674 case vmIntrinsics::_longBitsToDouble:
675 675 return inline_fp_conversions(intrinsic_id());
676 676
677 677 case vmIntrinsics::_numberOfLeadingZeros_i:
678 678 case vmIntrinsics::_numberOfLeadingZeros_l:
679 679 return inline_numberOfLeadingZeros(intrinsic_id());
680 680
681 681 case vmIntrinsics::_numberOfTrailingZeros_i:
682 682 case vmIntrinsics::_numberOfTrailingZeros_l:
683 683 return inline_numberOfTrailingZeros(intrinsic_id());
684 684
685 685 case vmIntrinsics::_bitCount_i:
686 686 case vmIntrinsics::_bitCount_l:
687 687 return inline_bitCount(intrinsic_id());
688 688
689 689 case vmIntrinsics::_reverseBytes_i:
690 690 case vmIntrinsics::_reverseBytes_l:
691 691 case vmIntrinsics::_reverseBytes_s:
692 692 case vmIntrinsics::_reverseBytes_c:
693 693 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
694 694
695 695 case vmIntrinsics::_get_AtomicLong:
696 696 return inline_native_AtomicLong_get();
697 697 case vmIntrinsics::_attemptUpdate:
698 698 return inline_native_AtomicLong_attemptUpdate();
699 699
700 700 case vmIntrinsics::_getCallerClass:
701 701 return inline_native_Reflection_getCallerClass();
702 702
703 703 case vmIntrinsics::_Reference_get:
704 704 return inline_reference_get();
705 705
706 706 default:
707 707 // If you get here, it may be that someone has added a new intrinsic
708 708 // to the list in vmSymbols.hpp without implementing it here.
709 709 #ifndef PRODUCT
710 710 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
711 711 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
712 712 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
713 713 }
714 714 #endif
715 715 return false;
716 716 }
717 717 }
718 718
719 719 //------------------------------push_result------------------------------
720 720 // Helper function for finishing intrinsics.
721 721 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
722 722 record_for_igvn(region);
723 723 set_control(_gvn.transform(region));
724 724 BasicType value_type = value->type()->basic_type();
725 725 push_node(value_type, _gvn.transform(value));
726 726 }
727 727
728 728 //------------------------------generate_guard---------------------------
729 729 // Helper function for generating guarded fast-slow graph structures.
730 730 // The given 'test', if true, guards a slow path. If the test fails
731 731 // then a fast path can be taken. (We generally hope it fails.)
732 732 // In all cases, GraphKit::control() is updated to the fast path.
733 733 // The returned value represents the control for the slow path.
734 734 // The return value is never 'top'; it is either a valid control
735 735 // or NULL if it is obvious that the slow path can never be taken.
736 736 // Also, if region and the slow control are not NULL, the slow edge
737 737 // is appended to the region.
738 738 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
739 739 if (stopped()) {
740 740 // Already short circuited.
741 741 return NULL;
742 742 }
743 743
744 744 // Build an if node and its projections.
745 745 // If test is true we take the slow path, which we assume is uncommon.
746 746 if (_gvn.type(test) == TypeInt::ZERO) {
747 747 // The slow branch is never taken. No need to build this guard.
748 748 return NULL;
749 749 }
750 750
751 751 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
752 752
753 753 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
754 754 if (if_slow == top()) {
755 755 // The slow branch is never taken. No need to build this guard.
756 756 return NULL;
757 757 }
758 758
759 759 if (region != NULL)
760 760 region->add_req(if_slow);
761 761
762 762 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
763 763 set_control(if_fast);
764 764
765 765 return if_slow;
766 766 }
767 767
768 768 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
769 769 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
770 770 }
771 771 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
772 772 return generate_guard(test, region, PROB_FAIR);
773 773 }
774 774
775 775 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
776 776 Node* *pos_index) {
777 777 if (stopped())
778 778 return NULL; // already stopped
779 779 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
780 780 return NULL; // index is already adequately typed
781 781 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
782 782 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
783 783 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
784 784 if (is_neg != NULL && pos_index != NULL) {
785 785 // Emulate effect of Parse::adjust_map_after_if.
786 786 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
787 787 ccast->set_req(0, control());
788 788 (*pos_index) = _gvn.transform(ccast);
789 789 }
790 790 return is_neg;
791 791 }
792 792
793 793 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
794 794 Node* *pos_index) {
795 795 if (stopped())
796 796 return NULL; // already stopped
797 797 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
798 798 return NULL; // index is already adequately typed
799 799 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
800 800 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
801 801 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
802 802 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
803 803 if (is_notp != NULL && pos_index != NULL) {
804 804 // Emulate effect of Parse::adjust_map_after_if.
805 805 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
806 806 ccast->set_req(0, control());
807 807 (*pos_index) = _gvn.transform(ccast);
808 808 }
809 809 return is_notp;
810 810 }
811 811
812 812 // Make sure that 'position' is a valid limit index, in [0..length].
813 813 // There are two equivalent plans for checking this:
814 814 // A. (offset + copyLength) unsigned<= arrayLength
815 815 // B. offset <= (arrayLength - copyLength)
816 816 // We require that all of the values above, except for the sum and
817 817 // difference, are already known to be non-negative.
818 818 // Plan A is robust in the face of overflow, if offset and copyLength
819 819 // are both hugely positive.
820 820 //
821 821 // Plan B is less direct and intuitive, but it does not overflow at
822 822 // all, since the difference of two non-negatives is always
823 823 // representable. Whenever Java methods must perform the equivalent
824 824 // check they generally use Plan B instead of Plan A.
825 825 // For the moment we use Plan A.
826 826 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
827 827 Node* subseq_length,
828 828 Node* array_length,
829 829 RegionNode* region) {
830 830 if (stopped())
831 831 return NULL; // already stopped
832 832 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
833 833 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
834 834 return NULL; // common case of whole-array copy
835 835 Node* last = subseq_length;
836 836 if (!zero_offset) // last += offset
837 837 last = _gvn.transform( new (C, 3) AddINode(last, offset));
838 838 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
839 839 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
840 840 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
841 841 return is_over;
842 842 }
843 843
844 844
845 845 //--------------------------generate_current_thread--------------------
846 846 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
847 847 ciKlass* thread_klass = env()->Thread_klass();
848 848 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
849 849 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
850 850 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
851 851 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
852 852 tls_output = thread;
853 853 return threadObj;
854 854 }
855 855
856 856
857 857 //------------------------------make_string_method_node------------------------
858 858 // Helper method for String intrinsic finctions.
859 859 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
860 860 const int value_offset = java_lang_String::value_offset_in_bytes();
861 861 const int count_offset = java_lang_String::count_offset_in_bytes();
862 862 const int offset_offset = java_lang_String::offset_offset_in_bytes();
863 863
864 864 Node* no_ctrl = NULL;
865 865
866 866 ciInstanceKlass* klass = env()->String_klass();
867 867 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
868 868
869 869 const TypeAryPtr* value_type =
870 870 TypeAryPtr::make(TypePtr::NotNull,
871 871 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
872 872 ciTypeArrayKlass::make(T_CHAR), true, 0);
873 873
874 874 // Get start addr of string and substring
875 875 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
876 876 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
877 877 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
878 878 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
879 879 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
880 880
881 881 // Pin loads from String::equals() argument since it could be NULL.
882 882 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
883 883 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
884 884 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
885 885 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
886 886 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
887 887 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
888 888
889 889 Node* result = NULL;
890 890 switch (opcode) {
891 891 case Op_StrIndexOf:
892 892 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
893 893 str1_start, cnt1, str2_start, cnt2);
894 894 break;
895 895 case Op_StrComp:
896 896 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
897 897 str1_start, cnt1, str2_start, cnt2);
898 898 break;
899 899 case Op_StrEquals:
900 900 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
901 901 str1_start, str2_start, cnt1);
902 902 break;
903 903 default:
904 904 ShouldNotReachHere();
905 905 return NULL;
906 906 }
907 907
908 908 // All these intrinsics have checks.
909 909 C->set_has_split_ifs(true); // Has chance for split-if optimization
910 910
911 911 return _gvn.transform(result);
912 912 }
913 913
914 914 //------------------------------inline_string_compareTo------------------------
915 915 bool LibraryCallKit::inline_string_compareTo() {
916 916
917 917 if (!Matcher::has_match_rule(Op_StrComp)) return false;
918 918
919 919 const int value_offset = java_lang_String::value_offset_in_bytes();
920 920 const int count_offset = java_lang_String::count_offset_in_bytes();
921 921 const int offset_offset = java_lang_String::offset_offset_in_bytes();
922 922
923 923 _sp += 2;
924 924 Node *argument = pop(); // pop non-receiver first: it was pushed second
925 925 Node *receiver = pop();
926 926
927 927 // Null check on self without removing any arguments. The argument
928 928 // null check technically happens in the wrong place, which can lead to
929 929 // invalid stack traces when string compare is inlined into a method
930 930 // which handles NullPointerExceptions.
931 931 _sp += 2;
932 932 receiver = do_null_check(receiver, T_OBJECT);
933 933 argument = do_null_check(argument, T_OBJECT);
934 934 _sp -= 2;
935 935 if (stopped()) {
936 936 return true;
937 937 }
938 938
939 939 ciInstanceKlass* klass = env()->String_klass();
940 940 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
941 941 Node* no_ctrl = NULL;
942 942
943 943 // Get counts for string and argument
944 944 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
945 945 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
946 946
947 947 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
948 948 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
949 949
950 950 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
951 951 push(compare);
952 952 return true;
953 953 }
954 954
955 955 //------------------------------inline_string_equals------------------------
956 956 bool LibraryCallKit::inline_string_equals() {
957 957
958 958 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
959 959
960 960 const int value_offset = java_lang_String::value_offset_in_bytes();
961 961 const int count_offset = java_lang_String::count_offset_in_bytes();
962 962 const int offset_offset = java_lang_String::offset_offset_in_bytes();
963 963
964 964 int nargs = 2;
965 965 _sp += nargs;
966 966 Node* argument = pop(); // pop non-receiver first: it was pushed second
967 967 Node* receiver = pop();
968 968
969 969 // Null check on self without removing any arguments. The argument
970 970 // null check technically happens in the wrong place, which can lead to
971 971 // invalid stack traces when string compare is inlined into a method
972 972 // which handles NullPointerExceptions.
973 973 _sp += nargs;
974 974 receiver = do_null_check(receiver, T_OBJECT);
975 975 //should not do null check for argument for String.equals(), because spec
976 976 //allows to specify NULL as argument.
977 977 _sp -= nargs;
978 978
979 979 if (stopped()) {
980 980 return true;
981 981 }
982 982
983 983 // paths (plus control) merge
984 984 RegionNode* region = new (C, 5) RegionNode(5);
985 985 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
986 986
987 987 // does source == target string?
988 988 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
989 989 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
990 990
991 991 Node* if_eq = generate_slow_guard(bol, NULL);
992 992 if (if_eq != NULL) {
993 993 // receiver == argument
994 994 phi->init_req(2, intcon(1));
995 995 region->init_req(2, if_eq);
996 996 }
997 997
998 998 // get String klass for instanceOf
999 999 ciInstanceKlass* klass = env()->String_klass();
1000 1000
1001 1001 if (!stopped()) {
1002 1002 _sp += nargs; // gen_instanceof might do an uncommon trap
1003 1003 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1004 1004 _sp -= nargs;
1005 1005 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
1006 1006 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
1007 1007
1008 1008 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1009 1009 //instanceOf == true, fallthrough
1010 1010
1011 1011 if (inst_false != NULL) {
1012 1012 phi->init_req(3, intcon(0));
1013 1013 region->init_req(3, inst_false);
1014 1014 }
1015 1015 }
1016 1016
1017 1017 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1018 1018
1019 1019 Node* no_ctrl = NULL;
1020 1020 Node* receiver_cnt;
1021 1021 Node* argument_cnt;
1022 1022
1023 1023 if (!stopped()) {
1024 1024 // Properly cast the argument to String
1025 1025 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
1026 1026
1027 1027 // Get counts for string and argument
1028 1028 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
1029 1029 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1030 1030
1031 1031 // Pin load from argument string since it could be NULL.
1032 1032 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
1033 1033 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1034 1034
1035 1035 // Check for receiver count != argument count
1036 1036 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
1037 1037 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
1038 1038 Node* if_ne = generate_slow_guard(bol, NULL);
1039 1039 if (if_ne != NULL) {
1040 1040 phi->init_req(4, intcon(0));
1041 1041 region->init_req(4, if_ne);
1042 1042 }
1043 1043 }
1044 1044
1045 1045 // Check for count == 0 is done by mach node StrEquals.
1046 1046
1047 1047 if (!stopped()) {
1048 1048 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1049 1049 phi->init_req(1, equals);
1050 1050 region->init_req(1, control());
1051 1051 }
1052 1052
1053 1053 // post merge
1054 1054 set_control(_gvn.transform(region));
1055 1055 record_for_igvn(region);
1056 1056
1057 1057 push(_gvn.transform(phi));
1058 1058
1059 1059 return true;
1060 1060 }
1061 1061
1062 1062 //------------------------------inline_array_equals----------------------------
1063 1063 bool LibraryCallKit::inline_array_equals() {
1064 1064
1065 1065 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1066 1066
1067 1067 _sp += 2;
1068 1068 Node *argument2 = pop();
1069 1069 Node *argument1 = pop();
1070 1070
1071 1071 Node* equals =
1072 1072 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1073 1073 argument1, argument2) );
1074 1074 push(equals);
1075 1075 return true;
1076 1076 }
1077 1077
1078 1078 // Java version of String.indexOf(constant string)
1079 1079 // class StringDecl {
1080 1080 // StringDecl(char[] ca) {
1081 1081 // offset = 0;
1082 1082 // count = ca.length;
1083 1083 // value = ca;
1084 1084 // }
1085 1085 // int offset;
1086 1086 // int count;
1087 1087 // char[] value;
1088 1088 // }
1089 1089 //
1090 1090 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1091 1091 // int targetOffset, int cache_i, int md2) {
1092 1092 // int cache = cache_i;
1093 1093 // int sourceOffset = string_object.offset;
1094 1094 // int sourceCount = string_object.count;
1095 1095 // int targetCount = target_object.length;
1096 1096 //
1097 1097 // int targetCountLess1 = targetCount - 1;
1098 1098 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1099 1099 //
1100 1100 // char[] source = string_object.value;
1101 1101 // char[] target = target_object;
1102 1102 // int lastChar = target[targetCountLess1];
1103 1103 //
1104 1104 // outer_loop:
1105 1105 // for (int i = sourceOffset; i < sourceEnd; ) {
1106 1106 // int src = source[i + targetCountLess1];
1107 1107 // if (src == lastChar) {
1108 1108 // // With random strings and a 4-character alphabet,
1109 1109 // // reverse matching at this point sets up 0.8% fewer
1110 1110 // // frames, but (paradoxically) makes 0.3% more probes.
1111 1111 // // Since those probes are nearer the lastChar probe,
1112 1112 // // there is may be a net D$ win with reverse matching.
1113 1113 // // But, reversing loop inhibits unroll of inner loop
1114 1114 // // for unknown reason. So, does running outer loop from
1115 1115 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1116 1116 // for (int j = 0; j < targetCountLess1; j++) {
1117 1117 // if (target[targetOffset + j] != source[i+j]) {
1118 1118 // if ((cache & (1 << source[i+j])) == 0) {
1119 1119 // if (md2 < j+1) {
1120 1120 // i += j+1;
1121 1121 // continue outer_loop;
1122 1122 // }
1123 1123 // }
1124 1124 // i += md2;
1125 1125 // continue outer_loop;
1126 1126 // }
1127 1127 // }
1128 1128 // return i - sourceOffset;
1129 1129 // }
1130 1130 // if ((cache & (1 << src)) == 0) {
1131 1131 // i += targetCountLess1;
1132 1132 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1133 1133 // i++;
1134 1134 // }
1135 1135 // return -1;
1136 1136 // }
1137 1137
1138 1138 //------------------------------string_indexOf------------------------
1139 1139 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1140 1140 jint cache_i, jint md2_i) {
1141 1141
1142 1142 Node* no_ctrl = NULL;
1143 1143 float likely = PROB_LIKELY(0.9);
1144 1144 float unlikely = PROB_UNLIKELY(0.9);
1145 1145
1146 1146 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1147 1147
1148 1148 const int value_offset = java_lang_String::value_offset_in_bytes();
1149 1149 const int count_offset = java_lang_String::count_offset_in_bytes();
1150 1150 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1151 1151
1152 1152 ciInstanceKlass* klass = env()->String_klass();
1153 1153 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1154 1154 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1155 1155
1156 1156 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1157 1157 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1158 1158 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1159 1159 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1160 1160 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1161 1161 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1162 1162
1163 1163 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1164 1164 jint target_length = target_array->length();
1165 1165 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1166 1166 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1167 1167
1168 1168 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1169 1169 #define __ kit.
1170 1170 Node* zero = __ ConI(0);
1171 1171 Node* one = __ ConI(1);
1172 1172 Node* cache = __ ConI(cache_i);
1173 1173 Node* md2 = __ ConI(md2_i);
1174 1174 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1175 1175 Node* targetCount = __ ConI(target_length);
1176 1176 Node* targetCountLess1 = __ ConI(target_length - 1);
1177 1177 Node* targetOffset = __ ConI(targetOffset_i);
1178 1178 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1179 1179
1180 1180 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1181 1181 Node* outer_loop = __ make_label(2 /* goto */);
1182 1182 Node* return_ = __ make_label(1);
1183 1183
1184 1184 __ set(rtn,__ ConI(-1));
1185 1185 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1186 1186 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1187 1187 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1188 1188 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1189 1189 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1190 1190 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1191 1191 Node* tpj = __ AddI(targetOffset, __ value(j));
1192 1192 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1193 1193 Node* ipj = __ AddI(__ value(i), __ value(j));
1194 1194 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1195 1195 __ if_then(targ, BoolTest::ne, src2); {
1196 1196 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1197 1197 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1198 1198 __ increment(i, __ AddI(__ value(j), one));
1199 1199 __ goto_(outer_loop);
1200 1200 } __ end_if(); __ dead(j);
1201 1201 }__ end_if(); __ dead(j);
1202 1202 __ increment(i, md2);
1203 1203 __ goto_(outer_loop);
1204 1204 }__ end_if();
1205 1205 __ increment(j, one);
1206 1206 }__ end_loop(); __ dead(j);
1207 1207 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1208 1208 __ goto_(return_);
1209 1209 }__ end_if();
1210 1210 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1211 1211 __ increment(i, targetCountLess1);
1212 1212 }__ end_if();
1213 1213 __ increment(i, one);
1214 1214 __ bind(outer_loop);
1215 1215 }__ end_loop(); __ dead(i);
1216 1216 __ bind(return_);
1217 1217
1218 1218 // Final sync IdealKit and GraphKit.
1219 1219 sync_kit(kit);
1220 1220 Node* result = __ value(rtn);
1221 1221 #undef __
1222 1222 C->set_has_loops(true);
1223 1223 return result;
1224 1224 }
1225 1225
1226 1226 //------------------------------inline_string_indexOf------------------------
1227 1227 bool LibraryCallKit::inline_string_indexOf() {
1228 1228
1229 1229 const int value_offset = java_lang_String::value_offset_in_bytes();
1230 1230 const int count_offset = java_lang_String::count_offset_in_bytes();
1231 1231 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1232 1232
1233 1233 _sp += 2;
1234 1234 Node *argument = pop(); // pop non-receiver first: it was pushed second
1235 1235 Node *receiver = pop();
1236 1236
1237 1237 Node* result;
1238 1238 // Disable the use of pcmpestri until it can be guaranteed that
1239 1239 // the load doesn't cross into the uncommited space.
1240 1240 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1241 1241 UseSSE42Intrinsics) {
1242 1242 // Generate SSE4.2 version of indexOf
1243 1243 // We currently only have match rules that use SSE4.2
1244 1244
1245 1245 // Null check on self without removing any arguments. The argument
1246 1246 // null check technically happens in the wrong place, which can lead to
1247 1247 // invalid stack traces when string compare is inlined into a method
1248 1248 // which handles NullPointerExceptions.
1249 1249 _sp += 2;
1250 1250 receiver = do_null_check(receiver, T_OBJECT);
1251 1251 argument = do_null_check(argument, T_OBJECT);
1252 1252 _sp -= 2;
1253 1253
1254 1254 if (stopped()) {
1255 1255 return true;
1256 1256 }
1257 1257
1258 1258 ciInstanceKlass* str_klass = env()->String_klass();
1259 1259 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1260 1260
1261 1261 // Make the merge point
1262 1262 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1263 1263 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1264 1264 Node* no_ctrl = NULL;
1265 1265
1266 1266 // Get counts for string and substr
1267 1267 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1268 1268 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1269 1269
1270 1270 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1271 1271 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1272 1272
1273 1273 // Check for substr count > string count
1274 1274 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1275 1275 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1276 1276 Node* if_gt = generate_slow_guard(bol, NULL);
1277 1277 if (if_gt != NULL) {
1278 1278 result_phi->init_req(2, intcon(-1));
1279 1279 result_rgn->init_req(2, if_gt);
1280 1280 }
1281 1281
1282 1282 if (!stopped()) {
1283 1283 // Check for substr count == 0
1284 1284 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1285 1285 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1286 1286 Node* if_zero = generate_slow_guard(bol, NULL);
1287 1287 if (if_zero != NULL) {
1288 1288 result_phi->init_req(3, intcon(0));
1289 1289 result_rgn->init_req(3, if_zero);
1290 1290 }
1291 1291 }
1292 1292
1293 1293 if (!stopped()) {
1294 1294 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1295 1295 result_phi->init_req(1, result);
1296 1296 result_rgn->init_req(1, control());
1297 1297 }
1298 1298 set_control(_gvn.transform(result_rgn));
1299 1299 record_for_igvn(result_rgn);
1300 1300 result = _gvn.transform(result_phi);
1301 1301
1302 1302 } else { // Use LibraryCallKit::string_indexOf
1303 1303 // don't intrinsify if argument isn't a constant string.
1304 1304 if (!argument->is_Con()) {
1305 1305 return false;
1306 1306 }
1307 1307 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1308 1308 if (str_type == NULL) {
1309 1309 return false;
1310 1310 }
1311 1311 ciInstanceKlass* klass = env()->String_klass();
1312 1312 ciObject* str_const = str_type->const_oop();
1313 1313 if (str_const == NULL || str_const->klass() != klass) {
1314 1314 return false;
1315 1315 }
1316 1316 ciInstance* str = str_const->as_instance();
1317 1317 assert(str != NULL, "must be instance");
1318 1318
1319 1319 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1320 1320 int o = str->field_value_by_offset(offset_offset).as_int();
1321 1321 int c = str->field_value_by_offset(count_offset).as_int();
1322 1322 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1323 1323
1324 1324 // constant strings have no offset and count == length which
1325 1325 // simplifies the resulting code somewhat so lets optimize for that.
1326 1326 if (o != 0 || c != pat->length()) {
1327 1327 return false;
1328 1328 }
1329 1329
1330 1330 // Null check on self without removing any arguments. The argument
1331 1331 // null check technically happens in the wrong place, which can lead to
1332 1332 // invalid stack traces when string compare is inlined into a method
1333 1333 // which handles NullPointerExceptions.
1334 1334 _sp += 2;
1335 1335 receiver = do_null_check(receiver, T_OBJECT);
1336 1336 // No null check on the argument is needed since it's a constant String oop.
1337 1337 _sp -= 2;
1338 1338 if (stopped()) {
1339 1339 return true;
1340 1340 }
1341 1341
1342 1342 // The null string as a pattern always returns 0 (match at beginning of string)
1343 1343 if (c == 0) {
1344 1344 push(intcon(0));
1345 1345 return true;
1346 1346 }
1347 1347
1348 1348 // Generate default indexOf
1349 1349 jchar lastChar = pat->char_at(o + (c - 1));
1350 1350 int cache = 0;
1351 1351 int i;
1352 1352 for (i = 0; i < c - 1; i++) {
1353 1353 assert(i < pat->length(), "out of range");
1354 1354 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1355 1355 }
1356 1356
1357 1357 int md2 = c;
1358 1358 for (i = 0; i < c - 1; i++) {
1359 1359 assert(i < pat->length(), "out of range");
1360 1360 if (pat->char_at(o + i) == lastChar) {
1361 1361 md2 = (c - 1) - i;
1362 1362 }
1363 1363 }
1364 1364
1365 1365 result = string_indexOf(receiver, pat, o, cache, md2);
1366 1366 }
1367 1367
1368 1368 push(result);
1369 1369 return true;
1370 1370 }
1371 1371
1372 1372 //--------------------------pop_math_arg--------------------------------
1373 1373 // Pop a double argument to a math function from the stack
1374 1374 // rounding it if necessary.
1375 1375 Node * LibraryCallKit::pop_math_arg() {
1376 1376 Node *arg = pop_pair();
1377 1377 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1378 1378 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1379 1379 return arg;
1380 1380 }
1381 1381
1382 1382 //------------------------------inline_trig----------------------------------
1383 1383 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1384 1384 // argument reduction which will turn into a fast/slow diamond.
1385 1385 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1386 1386 _sp += arg_size(); // restore stack pointer
1387 1387 Node* arg = pop_math_arg();
1388 1388 Node* trig = NULL;
1389 1389
1390 1390 switch (id) {
1391 1391 case vmIntrinsics::_dsin:
1392 1392 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1393 1393 break;
1394 1394 case vmIntrinsics::_dcos:
1395 1395 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1396 1396 break;
1397 1397 case vmIntrinsics::_dtan:
1398 1398 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1399 1399 break;
1400 1400 default:
1401 1401 assert(false, "bad intrinsic was passed in");
1402 1402 return false;
1403 1403 }
1404 1404
1405 1405 // Rounding required? Check for argument reduction!
1406 1406 if( Matcher::strict_fp_requires_explicit_rounding ) {
1407 1407
1408 1408 static const double pi_4 = 0.7853981633974483;
1409 1409 static const double neg_pi_4 = -0.7853981633974483;
1410 1410 // pi/2 in 80-bit extended precision
1411 1411 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1412 1412 // -pi/2 in 80-bit extended precision
1413 1413 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1414 1414 // Cutoff value for using this argument reduction technique
1415 1415 //static const double pi_2_minus_epsilon = 1.564660403643354;
1416 1416 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1417 1417
1418 1418 // Pseudocode for sin:
1419 1419 // if (x <= Math.PI / 4.0) {
1420 1420 // if (x >= -Math.PI / 4.0) return fsin(x);
1421 1421 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1422 1422 // } else {
1423 1423 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1424 1424 // }
1425 1425 // return StrictMath.sin(x);
1426 1426
1427 1427 // Pseudocode for cos:
1428 1428 // if (x <= Math.PI / 4.0) {
1429 1429 // if (x >= -Math.PI / 4.0) return fcos(x);
1430 1430 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1431 1431 // } else {
1432 1432 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1433 1433 // }
1434 1434 // return StrictMath.cos(x);
1435 1435
1436 1436 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1437 1437 // requires a special machine instruction to load it. Instead we'll try
1438 1438 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1439 1439 // probably do the math inside the SIN encoding.
1440 1440
1441 1441 // Make the merge point
1442 1442 RegionNode *r = new (C, 3) RegionNode(3);
1443 1443 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1444 1444
1445 1445 // Flatten arg so we need only 1 test
1446 1446 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1447 1447 // Node for PI/4 constant
1448 1448 Node *pi4 = makecon(TypeD::make(pi_4));
1449 1449 // Check PI/4 : abs(arg)
1450 1450 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1451 1451 // Check: If PI/4 < abs(arg) then go slow
1452 1452 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1453 1453 // Branch either way
1454 1454 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1455 1455 set_control(opt_iff(r,iff));
1456 1456
1457 1457 // Set fast path result
1458 1458 phi->init_req(2,trig);
1459 1459
1460 1460 // Slow path - non-blocking leaf call
1461 1461 Node* call = NULL;
1462 1462 switch (id) {
1463 1463 case vmIntrinsics::_dsin:
1464 1464 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1465 1465 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1466 1466 "Sin", NULL, arg, top());
1467 1467 break;
1468 1468 case vmIntrinsics::_dcos:
1469 1469 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1470 1470 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1471 1471 "Cos", NULL, arg, top());
1472 1472 break;
1473 1473 case vmIntrinsics::_dtan:
1474 1474 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1475 1475 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1476 1476 "Tan", NULL, arg, top());
1477 1477 break;
1478 1478 }
1479 1479 assert(control()->in(0) == call, "");
1480 1480 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1481 1481 r->init_req(1,control());
1482 1482 phi->init_req(1,slow_result);
1483 1483
1484 1484 // Post-merge
1485 1485 set_control(_gvn.transform(r));
1486 1486 record_for_igvn(r);
1487 1487 trig = _gvn.transform(phi);
1488 1488
1489 1489 C->set_has_split_ifs(true); // Has chance for split-if optimization
1490 1490 }
1491 1491 // Push result back on JVM stack
1492 1492 push_pair(trig);
1493 1493 return true;
1494 1494 }
1495 1495
1496 1496 //------------------------------inline_sqrt-------------------------------------
1497 1497 // Inline square root instruction, if possible.
1498 1498 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1499 1499 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1500 1500 _sp += arg_size(); // restore stack pointer
1501 1501 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1502 1502 return true;
1503 1503 }
1504 1504
1505 1505 //------------------------------inline_abs-------------------------------------
1506 1506 // Inline absolute value instruction, if possible.
1507 1507 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1508 1508 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1509 1509 _sp += arg_size(); // restore stack pointer
1510 1510 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1511 1511 return true;
1512 1512 }
1513 1513
1514 1514 //------------------------------inline_exp-------------------------------------
1515 1515 // Inline exp instructions, if possible. The Intel hardware only misses
1516 1516 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1517 1517 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1518 1518 assert(id == vmIntrinsics::_dexp, "Not exp");
1519 1519
1520 1520 // If this inlining ever returned NaN in the past, we do not intrinsify it
1521 1521 // every again. NaN results requires StrictMath.exp handling.
1522 1522 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1523 1523
1524 1524 // Do not intrinsify on older platforms which lack cmove.
1525 1525 if (ConditionalMoveLimit == 0) return false;
1526 1526
1527 1527 _sp += arg_size(); // restore stack pointer
1528 1528 Node *x = pop_math_arg();
1529 1529 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1530 1530
1531 1531 //-------------------
1532 1532 //result=(result.isNaN())? StrictMath::exp():result;
1533 1533 // Check: If isNaN() by checking result!=result? then go to Strict Math
1534 1534 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1535 1535 // Build the boolean node
1536 1536 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1537 1537
1538 1538 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1539 1539 // End the current control-flow path
1540 1540 push_pair(x);
1541 1541 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1542 1542 // to handle. Recompile without intrinsifying Math.exp
1543 1543 uncommon_trap(Deoptimization::Reason_intrinsic,
1544 1544 Deoptimization::Action_make_not_entrant);
1545 1545 }
1546 1546
1547 1547 C->set_has_split_ifs(true); // Has chance for split-if optimization
1548 1548
1549 1549 push_pair(result);
1550 1550
1551 1551 return true;
1552 1552 }
1553 1553
1554 1554 //------------------------------inline_pow-------------------------------------
1555 1555 // Inline power instructions, if possible.
1556 1556 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1557 1557 assert(id == vmIntrinsics::_dpow, "Not pow");
1558 1558
1559 1559 // If this inlining ever returned NaN in the past, we do not intrinsify it
1560 1560 // every again. NaN results requires StrictMath.pow handling.
1561 1561 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1562 1562
1563 1563 // Do not intrinsify on older platforms which lack cmove.
1564 1564 if (ConditionalMoveLimit == 0) return false;
1565 1565
1566 1566 // Pseudocode for pow
1567 1567 // if (x <= 0.0) {
1568 1568 // if ((double)((int)y)==y) { // if y is int
1569 1569 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1570 1570 // } else {
1571 1571 // result = NaN;
1572 1572 // }
1573 1573 // } else {
1574 1574 // result = DPow(x,y);
1575 1575 // }
1576 1576 // if (result != result)? {
1577 1577 // uncommon_trap();
1578 1578 // }
1579 1579 // return result;
1580 1580
1581 1581 _sp += arg_size(); // restore stack pointer
1582 1582 Node* y = pop_math_arg();
1583 1583 Node* x = pop_math_arg();
1584 1584
1585 1585 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1586 1586
1587 1587 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1588 1588 // inside of something) then skip the fancy tests and just check for
1589 1589 // NaN result.
1590 1590 Node *result = NULL;
1591 1591 if( jvms()->depth() >= 1 ) {
1592 1592 result = fast_result;
1593 1593 } else {
1594 1594
1595 1595 // Set the merge point for If node with condition of (x <= 0.0)
1596 1596 // There are four possible paths to region node and phi node
1597 1597 RegionNode *r = new (C, 4) RegionNode(4);
1598 1598 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1599 1599
1600 1600 // Build the first if node: if (x <= 0.0)
1601 1601 // Node for 0 constant
1602 1602 Node *zeronode = makecon(TypeD::ZERO);
1603 1603 // Check x:0
1604 1604 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1605 1605 // Check: If (x<=0) then go complex path
1606 1606 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1607 1607 // Branch either way
1608 1608 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1609 1609 Node *opt_test = _gvn.transform(if1);
1610 1610 //assert( opt_test->is_If(), "Expect an IfNode");
1611 1611 IfNode *opt_if1 = (IfNode*)opt_test;
1612 1612 // Fast path taken; set region slot 3
1613 1613 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1614 1614 r->init_req(3,fast_taken); // Capture fast-control
1615 1615
1616 1616 // Fast path not-taken, i.e. slow path
1617 1617 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1618 1618
1619 1619 // Set fast path result
1620 1620 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1621 1621 phi->init_req(3, fast_result);
1622 1622
1623 1623 // Complex path
1624 1624 // Build the second if node (if y is int)
1625 1625 // Node for (int)y
1626 1626 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1627 1627 // Node for (double)((int) y)
1628 1628 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1629 1629 // Check (double)((int) y) : y
1630 1630 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1631 1631 // Check if (y isn't int) then go to slow path
1632 1632
1633 1633 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1634 1634 // Branch either way
1635 1635 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1636 1636 Node *slow_path = opt_iff(r,if2); // Set region path 2
1637 1637
1638 1638 // Calculate DPow(abs(x), y)*(1 & (int)y)
1639 1639 // Node for constant 1
1640 1640 Node *conone = intcon(1);
1641 1641 // 1& (int)y
1642 1642 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1643 1643 // zero node
1644 1644 Node *conzero = intcon(0);
1645 1645 // Check (1&(int)y)==0?
1646 1646 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1647 1647 // Check if (1&(int)y)!=0?, if so the result is negative
1648 1648 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1649 1649 // abs(x)
1650 1650 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1651 1651 // abs(x)^y
1652 1652 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1653 1653 // -abs(x)^y
1654 1654 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1655 1655 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1656 1656 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1657 1657 // Set complex path fast result
1658 1658 phi->init_req(2, signresult);
1659 1659
1660 1660 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1661 1661 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1662 1662 r->init_req(1,slow_path);
1663 1663 phi->init_req(1,slow_result);
1664 1664
1665 1665 // Post merge
1666 1666 set_control(_gvn.transform(r));
1667 1667 record_for_igvn(r);
1668 1668 result=_gvn.transform(phi);
1669 1669 }
1670 1670
1671 1671 //-------------------
1672 1672 //result=(result.isNaN())? uncommon_trap():result;
1673 1673 // Check: If isNaN() by checking result!=result? then go to Strict Math
1674 1674 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1675 1675 // Build the boolean node
1676 1676 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1677 1677
1678 1678 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1679 1679 // End the current control-flow path
1680 1680 push_pair(x);
1681 1681 push_pair(y);
1682 1682 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1683 1683 // to handle. Recompile without intrinsifying Math.pow.
1684 1684 uncommon_trap(Deoptimization::Reason_intrinsic,
1685 1685 Deoptimization::Action_make_not_entrant);
1686 1686 }
1687 1687
1688 1688 C->set_has_split_ifs(true); // Has chance for split-if optimization
1689 1689
1690 1690 push_pair(result);
1691 1691
1692 1692 return true;
1693 1693 }
1694 1694
1695 1695 //------------------------------inline_trans-------------------------------------
1696 1696 // Inline transcendental instructions, if possible. The Intel hardware gets
1697 1697 // these right, no funny corner cases missed.
1698 1698 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1699 1699 _sp += arg_size(); // restore stack pointer
1700 1700 Node* arg = pop_math_arg();
1701 1701 Node* trans = NULL;
1702 1702
1703 1703 switch (id) {
1704 1704 case vmIntrinsics::_dlog:
1705 1705 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1706 1706 break;
1707 1707 case vmIntrinsics::_dlog10:
1708 1708 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1709 1709 break;
1710 1710 default:
1711 1711 assert(false, "bad intrinsic was passed in");
1712 1712 return false;
1713 1713 }
1714 1714
1715 1715 // Push result back on JVM stack
1716 1716 push_pair(trans);
1717 1717 return true;
1718 1718 }
1719 1719
1720 1720 //------------------------------runtime_math-----------------------------
1721 1721 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1722 1722 Node* a = NULL;
1723 1723 Node* b = NULL;
1724 1724
1725 1725 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1726 1726 "must be (DD)D or (D)D type");
1727 1727
1728 1728 // Inputs
1729 1729 _sp += arg_size(); // restore stack pointer
1730 1730 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1731 1731 b = pop_math_arg();
1732 1732 }
1733 1733 a = pop_math_arg();
1734 1734
1735 1735 const TypePtr* no_memory_effects = NULL;
1736 1736 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1737 1737 no_memory_effects,
1738 1738 a, top(), b, b ? top() : NULL);
1739 1739 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1740 1740 #ifdef ASSERT
1741 1741 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1742 1742 assert(value_top == top(), "second value must be top");
1743 1743 #endif
1744 1744
1745 1745 push_pair(value);
1746 1746 return true;
1747 1747 }
1748 1748
1749 1749 //------------------------------inline_math_native-----------------------------
1750 1750 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1751 1751 switch (id) {
1752 1752 // These intrinsics are not properly supported on all hardware
1753 1753 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1754 1754 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1755 1755 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1756 1756 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1757 1757 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1758 1758 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1759 1759
1760 1760 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1761 1761 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1762 1762 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1763 1763 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1764 1764
1765 1765 // These intrinsics are supported on all hardware
1766 1766 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1767 1767 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1768 1768
1769 1769 // These intrinsics don't work on X86. The ad implementation doesn't
1770 1770 // handle NaN's properly. Instead of returning infinity, the ad
1771 1771 // implementation returns a NaN on overflow. See bug: 6304089
1772 1772 // Once the ad implementations are fixed, change the code below
1773 1773 // to match the intrinsics above
1774 1774
1775 1775 case vmIntrinsics::_dexp: return
1776 1776 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1777 1777 case vmIntrinsics::_dpow: return
1778 1778 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1779 1779
1780 1780 // These intrinsics are not yet correctly implemented
1781 1781 case vmIntrinsics::_datan2:
1782 1782 return false;
1783 1783
1784 1784 default:
1785 1785 ShouldNotReachHere();
1786 1786 return false;
1787 1787 }
1788 1788 }
1789 1789
1790 1790 static bool is_simple_name(Node* n) {
1791 1791 return (n->req() == 1 // constant
1792 1792 || (n->is_Type() && n->as_Type()->type()->singleton())
1793 1793 || n->is_Proj() // parameter or return value
1794 1794 || n->is_Phi() // local of some sort
1795 1795 );
1796 1796 }
1797 1797
1798 1798 //----------------------------inline_min_max-----------------------------------
1799 1799 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1800 1800 push(generate_min_max(id, argument(0), argument(1)));
1801 1801
1802 1802 return true;
1803 1803 }
1804 1804
1805 1805 Node*
1806 1806 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1807 1807 // These are the candidate return value:
1808 1808 Node* xvalue = x0;
1809 1809 Node* yvalue = y0;
1810 1810
1811 1811 if (xvalue == yvalue) {
1812 1812 return xvalue;
1813 1813 }
1814 1814
1815 1815 bool want_max = (id == vmIntrinsics::_max);
1816 1816
1817 1817 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1818 1818 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1819 1819 if (txvalue == NULL || tyvalue == NULL) return top();
1820 1820 // This is not really necessary, but it is consistent with a
1821 1821 // hypothetical MaxINode::Value method:
1822 1822 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1823 1823
1824 1824 // %%% This folding logic should (ideally) be in a different place.
1825 1825 // Some should be inside IfNode, and there to be a more reliable
1826 1826 // transformation of ?: style patterns into cmoves. We also want
1827 1827 // more powerful optimizations around cmove and min/max.
1828 1828
1829 1829 // Try to find a dominating comparison of these guys.
1830 1830 // It can simplify the index computation for Arrays.copyOf
1831 1831 // and similar uses of System.arraycopy.
1832 1832 // First, compute the normalized version of CmpI(x, y).
1833 1833 int cmp_op = Op_CmpI;
1834 1834 Node* xkey = xvalue;
1835 1835 Node* ykey = yvalue;
1836 1836 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1837 1837 if (ideal_cmpxy->is_Cmp()) {
1838 1838 // E.g., if we have CmpI(length - offset, count),
1839 1839 // it might idealize to CmpI(length, count + offset)
1840 1840 cmp_op = ideal_cmpxy->Opcode();
1841 1841 xkey = ideal_cmpxy->in(1);
1842 1842 ykey = ideal_cmpxy->in(2);
1843 1843 }
1844 1844
1845 1845 // Start by locating any relevant comparisons.
1846 1846 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1847 1847 Node* cmpxy = NULL;
1848 1848 Node* cmpyx = NULL;
1849 1849 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1850 1850 Node* cmp = start_from->fast_out(k);
1851 1851 if (cmp->outcnt() > 0 && // must have prior uses
1852 1852 cmp->in(0) == NULL && // must be context-independent
1853 1853 cmp->Opcode() == cmp_op) { // right kind of compare
1854 1854 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1855 1855 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1856 1856 }
1857 1857 }
1858 1858
1859 1859 const int NCMPS = 2;
1860 1860 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1861 1861 int cmpn;
1862 1862 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1863 1863 if (cmps[cmpn] != NULL) break; // find a result
1864 1864 }
1865 1865 if (cmpn < NCMPS) {
1866 1866 // Look for a dominating test that tells us the min and max.
1867 1867 int depth = 0; // Limit search depth for speed
1868 1868 Node* dom = control();
1869 1869 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1870 1870 if (++depth >= 100) break;
1871 1871 Node* ifproj = dom;
1872 1872 if (!ifproj->is_Proj()) continue;
1873 1873 Node* iff = ifproj->in(0);
1874 1874 if (!iff->is_If()) continue;
1875 1875 Node* bol = iff->in(1);
1876 1876 if (!bol->is_Bool()) continue;
1877 1877 Node* cmp = bol->in(1);
1878 1878 if (cmp == NULL) continue;
1879 1879 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1880 1880 if (cmps[cmpn] == cmp) break;
1881 1881 if (cmpn == NCMPS) continue;
1882 1882 BoolTest::mask btest = bol->as_Bool()->_test._test;
1883 1883 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1884 1884 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1885 1885 // At this point, we know that 'x btest y' is true.
1886 1886 switch (btest) {
1887 1887 case BoolTest::eq:
1888 1888 // They are proven equal, so we can collapse the min/max.
1889 1889 // Either value is the answer. Choose the simpler.
1890 1890 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1891 1891 return yvalue;
1892 1892 return xvalue;
1893 1893 case BoolTest::lt: // x < y
1894 1894 case BoolTest::le: // x <= y
1895 1895 return (want_max ? yvalue : xvalue);
1896 1896 case BoolTest::gt: // x > y
1897 1897 case BoolTest::ge: // x >= y
1898 1898 return (want_max ? xvalue : yvalue);
1899 1899 }
1900 1900 }
1901 1901 }
1902 1902
1903 1903 // We failed to find a dominating test.
1904 1904 // Let's pick a test that might GVN with prior tests.
1905 1905 Node* best_bol = NULL;
1906 1906 BoolTest::mask best_btest = BoolTest::illegal;
1907 1907 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1908 1908 Node* cmp = cmps[cmpn];
1909 1909 if (cmp == NULL) continue;
1910 1910 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1911 1911 Node* bol = cmp->fast_out(j);
1912 1912 if (!bol->is_Bool()) continue;
1913 1913 BoolTest::mask btest = bol->as_Bool()->_test._test;
1914 1914 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1915 1915 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1916 1916 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1917 1917 best_bol = bol->as_Bool();
1918 1918 best_btest = btest;
1919 1919 }
1920 1920 }
1921 1921 }
1922 1922
1923 1923 Node* answer_if_true = NULL;
1924 1924 Node* answer_if_false = NULL;
1925 1925 switch (best_btest) {
1926 1926 default:
1927 1927 if (cmpxy == NULL)
1928 1928 cmpxy = ideal_cmpxy;
1929 1929 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1930 1930 // and fall through:
1931 1931 case BoolTest::lt: // x < y
1932 1932 case BoolTest::le: // x <= y
1933 1933 answer_if_true = (want_max ? yvalue : xvalue);
1934 1934 answer_if_false = (want_max ? xvalue : yvalue);
1935 1935 break;
1936 1936 case BoolTest::gt: // x > y
1937 1937 case BoolTest::ge: // x >= y
1938 1938 answer_if_true = (want_max ? xvalue : yvalue);
1939 1939 answer_if_false = (want_max ? yvalue : xvalue);
1940 1940 break;
1941 1941 }
1942 1942
1943 1943 jint hi, lo;
1944 1944 if (want_max) {
1945 1945 // We can sharpen the minimum.
1946 1946 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1947 1947 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1948 1948 } else {
1949 1949 // We can sharpen the maximum.
1950 1950 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1951 1951 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1952 1952 }
1953 1953
1954 1954 // Use a flow-free graph structure, to avoid creating excess control edges
1955 1955 // which could hinder other optimizations.
1956 1956 // Since Math.min/max is often used with arraycopy, we want
1957 1957 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1958 1958 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1959 1959 answer_if_false, answer_if_true,
1960 1960 TypeInt::make(lo, hi, widen));
1961 1961
1962 1962 return _gvn.transform(cmov);
1963 1963
1964 1964 /*
1965 1965 // This is not as desirable as it may seem, since Min and Max
1966 1966 // nodes do not have a full set of optimizations.
1967 1967 // And they would interfere, anyway, with 'if' optimizations
1968 1968 // and with CMoveI canonical forms.
1969 1969 switch (id) {
1970 1970 case vmIntrinsics::_min:
1971 1971 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1972 1972 case vmIntrinsics::_max:
1973 1973 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1974 1974 default:
1975 1975 ShouldNotReachHere();
1976 1976 }
1977 1977 */
1978 1978 }
1979 1979
1980 1980 inline int
1981 1981 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1982 1982 const TypePtr* base_type = TypePtr::NULL_PTR;
1983 1983 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1984 1984 if (base_type == NULL) {
1985 1985 // Unknown type.
1986 1986 return Type::AnyPtr;
1987 1987 } else if (base_type == TypePtr::NULL_PTR) {
1988 1988 // Since this is a NULL+long form, we have to switch to a rawptr.
1989 1989 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1990 1990 offset = MakeConX(0);
1991 1991 return Type::RawPtr;
1992 1992 } else if (base_type->base() == Type::RawPtr) {
1993 1993 return Type::RawPtr;
1994 1994 } else if (base_type->isa_oopptr()) {
1995 1995 // Base is never null => always a heap address.
1996 1996 if (base_type->ptr() == TypePtr::NotNull) {
1997 1997 return Type::OopPtr;
1998 1998 }
1999 1999 // Offset is small => always a heap address.
2000 2000 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2001 2001 if (offset_type != NULL &&
2002 2002 base_type->offset() == 0 && // (should always be?)
2003 2003 offset_type->_lo >= 0 &&
2004 2004 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2005 2005 return Type::OopPtr;
2006 2006 }
2007 2007 // Otherwise, it might either be oop+off or NULL+addr.
2008 2008 return Type::AnyPtr;
2009 2009 } else {
2010 2010 // No information:
2011 2011 return Type::AnyPtr;
2012 2012 }
2013 2013 }
2014 2014
2015 2015 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2016 2016 int kind = classify_unsafe_addr(base, offset);
2017 2017 if (kind == Type::RawPtr) {
2018 2018 return basic_plus_adr(top(), base, offset);
2019 2019 } else {
2020 2020 return basic_plus_adr(base, offset);
2021 2021 }
2022 2022 }
2023 2023
2024 2024 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2025 2025 // inline int Integer.numberOfLeadingZeros(int)
2026 2026 // inline int Long.numberOfLeadingZeros(long)
2027 2027 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2028 2028 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2029 2029 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2030 2030 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2031 2031 _sp += arg_size(); // restore stack pointer
2032 2032 switch (id) {
2033 2033 case vmIntrinsics::_numberOfLeadingZeros_i:
2034 2034 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
2035 2035 break;
2036 2036 case vmIntrinsics::_numberOfLeadingZeros_l:
2037 2037 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
2038 2038 break;
2039 2039 default:
2040 2040 ShouldNotReachHere();
2041 2041 }
2042 2042 return true;
2043 2043 }
2044 2044
2045 2045 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2046 2046 // inline int Integer.numberOfTrailingZeros(int)
2047 2047 // inline int Long.numberOfTrailingZeros(long)
2048 2048 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2049 2049 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2050 2050 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2051 2051 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2052 2052 _sp += arg_size(); // restore stack pointer
2053 2053 switch (id) {
2054 2054 case vmIntrinsics::_numberOfTrailingZeros_i:
2055 2055 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2056 2056 break;
2057 2057 case vmIntrinsics::_numberOfTrailingZeros_l:
2058 2058 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2059 2059 break;
2060 2060 default:
2061 2061 ShouldNotReachHere();
2062 2062 }
2063 2063 return true;
2064 2064 }
2065 2065
2066 2066 //----------------------------inline_bitCount_int/long-----------------------
2067 2067 // inline int Integer.bitCount(int)
2068 2068 // inline int Long.bitCount(long)
2069 2069 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2070 2070 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2071 2071 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2072 2072 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2073 2073 _sp += arg_size(); // restore stack pointer
2074 2074 switch (id) {
2075 2075 case vmIntrinsics::_bitCount_i:
2076 2076 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2077 2077 break;
2078 2078 case vmIntrinsics::_bitCount_l:
2079 2079 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2080 2080 break;
2081 2081 default:
2082 2082 ShouldNotReachHere();
2083 2083 }
2084 2084 return true;
2085 2085 }
2086 2086
2087 2087 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2088 2088 // inline Integer.reverseBytes(int)
2089 2089 // inline Long.reverseBytes(long)
2090 2090 // inline Character.reverseBytes(char)
2091 2091 // inline Short.reverseBytes(short)
2092 2092 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2093 2093 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2094 2094 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2095 2095 "not reverse Bytes");
2096 2096 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2097 2097 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2098 2098 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2099 2099 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2100 2100 _sp += arg_size(); // restore stack pointer
2101 2101 switch (id) {
2102 2102 case vmIntrinsics::_reverseBytes_i:
2103 2103 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2104 2104 break;
2105 2105 case vmIntrinsics::_reverseBytes_l:
2106 2106 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2107 2107 break;
2108 2108 case vmIntrinsics::_reverseBytes_c:
2109 2109 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2110 2110 break;
2111 2111 case vmIntrinsics::_reverseBytes_s:
2112 2112 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2113 2113 break;
2114 2114 default:
2115 2115 ;
2116 2116 }
2117 2117 return true;
2118 2118 }
2119 2119
2120 2120 //----------------------------inline_unsafe_access----------------------------
2121 2121
2122 2122 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2123 2123
2124 2124 // Helper that guards and inserts a G1 pre-barrier.
2125 2125 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2126 2126 assert(UseG1GC, "should not call this otherwise");
2127 2127
2128 2128 // We could be accessing the referent field of a reference object. If so, when G1
2129 2129 // is enabled, we need to log the value in the referent field in an SATB buffer.
2130 2130 // This routine performs some compile time filters and generates suitable
2131 2131 // runtime filters that guard the pre-barrier code.
2132 2132
2133 2133 // Some compile time checks.
2134 2134
2135 2135 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2136 2136 const TypeX* otype = offset->find_intptr_t_type();
2137 2137 if (otype != NULL && otype->is_con() &&
2138 2138 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2139 2139 // Constant offset but not the reference_offset so just return
2140 2140 return;
2141 2141 }
2142 2142
2143 2143 // We only need to generate the runtime guards for instances.
2144 2144 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2145 2145 if (btype != NULL) {
2146 2146 if (btype->isa_aryptr()) {
2147 2147 // Array type so nothing to do
2148 2148 return;
2149 2149 }
2150 2150
2151 2151 const TypeInstPtr* itype = btype->isa_instptr();
2152 2152 if (itype != NULL) {
2153 2153 // Can the klass of base_oop be statically determined
2154 2154 // to be _not_ a sub-class of Reference?
2155 2155 ciKlass* klass = itype->klass();
2156 2156 if (klass->is_subtype_of(env()->Reference_klass()) &&
2157 2157 !env()->Reference_klass()->is_subtype_of(klass)) {
2158 2158 return;
2159 2159 }
2160 2160 }
2161 2161 }
2162 2162
2163 2163 // The compile time filters did not reject base_oop/offset so
2164 2164 // we need to generate the following runtime filters
2165 2165 //
2166 2166 // if (offset == java_lang_ref_Reference::_reference_offset) {
2167 2167 // if (base != null) {
2168 2168 // if (klass(base)->reference_type() != REF_NONE)) {
2169 2169 // pre_barrier(_, pre_val, ...);
2170 2170 // }
2171 2171 // }
2172 2172 // }
↓ open down ↓ |
2172 lines elided |
↑ open up ↑ |
2173 2173
2174 2174 float likely = PROB_LIKELY(0.999);
2175 2175 float unlikely = PROB_UNLIKELY(0.999);
2176 2176
2177 2177 IdealKit ideal(gvn(), control(), merged_memory());
2178 2178 #define __ ideal.
2179 2179
2180 2180 const int reference_type_offset = instanceKlass::reference_type_offset_in_bytes() +
2181 2181 sizeof(oopDesc);
2182 2182
2183 - Node* referent_off = __ ConI(java_lang_ref_Reference::referent_offset);
2183 + Node* referent_off;
2184 +
2185 + if (offset->bottom_type()->isa_int() != NULL) {
2186 + referent_off = __ makecon(TypeInt::make(java_lang_ref_Reference::referent_offset));
2187 + } else {
2188 + assert(offset->bottom_type()->isa_long() != NULL, "what else?");
2189 + referent_off = __ makecon(TypeLong::make(java_lang_ref_Reference::referent_offset));
2190 + }
2184 2191
2185 2192 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2186 2193 __ if_then(base_oop, BoolTest::ne, null(), likely); {
2187 2194
2188 2195 // Update graphKit memory and control from IdealKit.
2189 2196 set_all_memory(__ merged_memory());
2190 2197 set_control(__ ctrl());
2191 2198
2192 2199 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2193 2200 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2194 2201
2195 2202 // Update IdealKit memory and control from graphKit.
2196 2203 __ set_all_memory(merged_memory());
2197 2204 __ set_ctrl(control());
2198 2205
2199 2206 Node* one = __ ConI(1);
2200 2207
2201 2208 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2202 2209
2203 2210 // Update graphKit from IdeakKit.
2204 2211 set_all_memory(__ merged_memory());
2205 2212 set_control(__ ctrl());
2206 2213
2207 2214 // Use the pre-barrier to record the value in the referent field
2208 2215 pre_barrier(false /* do_load */,
2209 2216 __ ctrl(),
2210 2217 NULL /* obj */, NULL /* adr */, -1 /* alias_idx */, NULL /* val */, NULL /* val_type */,
2211 2218 pre_val /* pre_val */,
2212 2219 T_OBJECT);
2213 2220
2214 2221 // Update IdealKit from graphKit.
2215 2222 __ set_all_memory(merged_memory());
2216 2223 __ set_ctrl(control());
2217 2224
2218 2225 } __ end_if(); // _ref_type != ref_none
2219 2226 } __ end_if(); // base != NULL
2220 2227 } __ end_if(); // offset == referent_offset
2221 2228
2222 2229 // Final sync IdealKit and GraphKit.
2223 2230 sync_kit(ideal);
2224 2231 #undef __
2225 2232 }
2226 2233
2227 2234
2228 2235 // Interpret Unsafe.fieldOffset cookies correctly:
2229 2236 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2230 2237
2231 2238 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2232 2239 if (callee()->is_static()) return false; // caller must have the capability!
2233 2240
2234 2241 #ifndef PRODUCT
2235 2242 {
2236 2243 ResourceMark rm;
2237 2244 // Check the signatures.
2238 2245 ciSignature* sig = signature();
2239 2246 #ifdef ASSERT
2240 2247 if (!is_store) {
2241 2248 // Object getObject(Object base, int/long offset), etc.
2242 2249 BasicType rtype = sig->return_type()->basic_type();
2243 2250 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2244 2251 rtype = T_ADDRESS; // it is really a C void*
2245 2252 assert(rtype == type, "getter must return the expected value");
2246 2253 if (!is_native_ptr) {
2247 2254 assert(sig->count() == 2, "oop getter has 2 arguments");
2248 2255 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2249 2256 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2250 2257 } else {
2251 2258 assert(sig->count() == 1, "native getter has 1 argument");
2252 2259 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2253 2260 }
2254 2261 } else {
2255 2262 // void putObject(Object base, int/long offset, Object x), etc.
2256 2263 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2257 2264 if (!is_native_ptr) {
2258 2265 assert(sig->count() == 3, "oop putter has 3 arguments");
2259 2266 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2260 2267 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2261 2268 } else {
2262 2269 assert(sig->count() == 2, "native putter has 2 arguments");
2263 2270 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2264 2271 }
2265 2272 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2266 2273 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2267 2274 vtype = T_ADDRESS; // it is really a C void*
2268 2275 assert(vtype == type, "putter must accept the expected value");
2269 2276 }
2270 2277 #endif // ASSERT
2271 2278 }
2272 2279 #endif //PRODUCT
2273 2280
2274 2281 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2275 2282
2276 2283 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2277 2284
2278 2285 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2279 2286 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2280 2287
2281 2288 debug_only(int saved_sp = _sp);
2282 2289 _sp += nargs;
2283 2290
2284 2291 Node* val;
2285 2292 debug_only(val = (Node*)(uintptr_t)-1);
2286 2293
2287 2294
2288 2295 if (is_store) {
2289 2296 // Get the value being stored. (Pop it first; it was pushed last.)
2290 2297 switch (type) {
2291 2298 case T_DOUBLE:
2292 2299 case T_LONG:
2293 2300 case T_ADDRESS:
2294 2301 val = pop_pair();
2295 2302 break;
2296 2303 default:
2297 2304 val = pop();
2298 2305 }
2299 2306 }
2300 2307
2301 2308 // Build address expression. See the code in inline_unsafe_prefetch.
2302 2309 Node *adr;
2303 2310 Node *heap_base_oop = top();
2304 2311 Node* offset = top();
2305 2312
2306 2313 if (!is_native_ptr) {
2307 2314 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2308 2315 offset = pop_pair();
2309 2316 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2310 2317 Node* base = pop();
2311 2318 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2312 2319 // to be plain byte offsets, which are also the same as those accepted
2313 2320 // by oopDesc::field_base.
2314 2321 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2315 2322 "fieldOffset must be byte-scaled");
2316 2323 // 32-bit machines ignore the high half!
2317 2324 offset = ConvL2X(offset);
2318 2325 adr = make_unsafe_address(base, offset);
2319 2326 heap_base_oop = base;
2320 2327 } else {
2321 2328 Node* ptr = pop_pair();
2322 2329 // Adjust Java long to machine word:
2323 2330 ptr = ConvL2X(ptr);
2324 2331 adr = make_unsafe_address(NULL, ptr);
2325 2332 }
2326 2333
2327 2334 // Pop receiver last: it was pushed first.
2328 2335 Node *receiver = pop();
2329 2336
2330 2337 assert(saved_sp == _sp, "must have correct argument count");
2331 2338
2332 2339 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2333 2340
2334 2341 // First guess at the value type.
2335 2342 const Type *value_type = Type::get_const_basic_type(type);
2336 2343
2337 2344 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2338 2345 // there was not enough information to nail it down.
2339 2346 Compile::AliasType* alias_type = C->alias_type(adr_type);
2340 2347 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2341 2348
2342 2349 // We will need memory barriers unless we can determine a unique
2343 2350 // alias category for this reference. (Note: If for some reason
2344 2351 // the barriers get omitted and the unsafe reference begins to "pollute"
2345 2352 // the alias analysis of the rest of the graph, either Compile::can_alias
2346 2353 // or Compile::must_alias will throw a diagnostic assert.)
2347 2354 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2348 2355
2349 2356 // If we are reading the value of the referent field of a Reference
2350 2357 // object (either by using Unsafe directly or through reflection)
2351 2358 // then, if G1 is enabled, we need to record the referent in an
2352 2359 // SATB log buffer using the pre-barrier mechanism.
2353 2360 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2354 2361 offset != top() && heap_base_oop != top();
2355 2362
2356 2363 if (!is_store && type == T_OBJECT) {
2357 2364 // Attempt to infer a sharper value type from the offset and base type.
2358 2365 ciKlass* sharpened_klass = NULL;
2359 2366
2360 2367 // See if it is an instance field, with an object type.
2361 2368 if (alias_type->field() != NULL) {
2362 2369 assert(!is_native_ptr, "native pointer op cannot use a java address");
2363 2370 if (alias_type->field()->type()->is_klass()) {
2364 2371 sharpened_klass = alias_type->field()->type()->as_klass();
2365 2372 }
2366 2373 }
2367 2374
2368 2375 // See if it is a narrow oop array.
2369 2376 if (adr_type->isa_aryptr()) {
2370 2377 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2371 2378 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2372 2379 if (elem_type != NULL) {
2373 2380 sharpened_klass = elem_type->klass();
2374 2381 }
2375 2382 }
2376 2383 }
2377 2384
2378 2385 if (sharpened_klass != NULL) {
2379 2386 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2380 2387
2381 2388 // Sharpen the value type.
2382 2389 value_type = tjp;
2383 2390
2384 2391 #ifndef PRODUCT
2385 2392 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2386 2393 tty->print(" from base type: "); adr_type->dump();
2387 2394 tty->print(" sharpened value: "); value_type->dump();
2388 2395 }
2389 2396 #endif
2390 2397 }
2391 2398 }
2392 2399
2393 2400 // Null check on self without removing any arguments. The argument
2394 2401 // null check technically happens in the wrong place, which can lead to
2395 2402 // invalid stack traces when the primitive is inlined into a method
2396 2403 // which handles NullPointerExceptions.
2397 2404 _sp += nargs;
2398 2405 do_null_check(receiver, T_OBJECT);
2399 2406 _sp -= nargs;
2400 2407 if (stopped()) {
2401 2408 return true;
2402 2409 }
2403 2410 // Heap pointers get a null-check from the interpreter,
2404 2411 // as a courtesy. However, this is not guaranteed by Unsafe,
2405 2412 // and it is not possible to fully distinguish unintended nulls
2406 2413 // from intended ones in this API.
2407 2414
2408 2415 if (is_volatile) {
2409 2416 // We need to emit leading and trailing CPU membars (see below) in
2410 2417 // addition to memory membars when is_volatile. This is a little
2411 2418 // too strong, but avoids the need to insert per-alias-type
2412 2419 // volatile membars (for stores; compare Parse::do_put_xxx), which
2413 2420 // we cannot do effectively here because we probably only have a
2414 2421 // rough approximation of type.
2415 2422 need_mem_bar = true;
2416 2423 // For Stores, place a memory ordering barrier now.
2417 2424 if (is_store)
2418 2425 insert_mem_bar(Op_MemBarRelease);
2419 2426 }
2420 2427
2421 2428 // Memory barrier to prevent normal and 'unsafe' accesses from
2422 2429 // bypassing each other. Happens after null checks, so the
2423 2430 // exception paths do not take memory state from the memory barrier,
2424 2431 // so there's no problems making a strong assert about mixing users
2425 2432 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2426 2433 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2427 2434 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2428 2435
2429 2436 if (!is_store) {
2430 2437 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2431 2438 // load value and push onto stack
2432 2439 switch (type) {
2433 2440 case T_BOOLEAN:
2434 2441 case T_CHAR:
2435 2442 case T_BYTE:
2436 2443 case T_SHORT:
2437 2444 case T_INT:
2438 2445 case T_FLOAT:
2439 2446 push(p);
2440 2447 break;
2441 2448 case T_OBJECT:
2442 2449 if (need_read_barrier) {
2443 2450 insert_g1_pre_barrier(heap_base_oop, offset, p);
2444 2451 }
2445 2452 push(p);
2446 2453 break;
2447 2454 case T_ADDRESS:
2448 2455 // Cast to an int type.
2449 2456 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2450 2457 p = ConvX2L(p);
2451 2458 push_pair(p);
2452 2459 break;
2453 2460 case T_DOUBLE:
2454 2461 case T_LONG:
2455 2462 push_pair( p );
2456 2463 break;
2457 2464 default: ShouldNotReachHere();
2458 2465 }
2459 2466 } else {
2460 2467 // place effect of store into memory
2461 2468 switch (type) {
2462 2469 case T_DOUBLE:
2463 2470 val = dstore_rounding(val);
2464 2471 break;
2465 2472 case T_ADDRESS:
2466 2473 // Repackage the long as a pointer.
2467 2474 val = ConvL2X(val);
2468 2475 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2469 2476 break;
2470 2477 }
2471 2478
2472 2479 if (type != T_OBJECT ) {
2473 2480 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2474 2481 } else {
2475 2482 // Possibly an oop being stored to Java heap or native memory
2476 2483 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2477 2484 // oop to Java heap.
2478 2485 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2479 2486 } else {
2480 2487 // We can't tell at compile time if we are storing in the Java heap or outside
2481 2488 // of it. So we need to emit code to conditionally do the proper type of
2482 2489 // store.
2483 2490
2484 2491 IdealKit ideal(gvn(), control(), merged_memory());
2485 2492 #define __ ideal.
2486 2493 // QQQ who knows what probability is here??
2487 2494 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2488 2495 // Sync IdealKit and graphKit.
2489 2496 set_all_memory( __ merged_memory());
2490 2497 set_control(__ ctrl());
2491 2498 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2492 2499 // Update IdealKit memory.
2493 2500 __ set_all_memory(merged_memory());
2494 2501 __ set_ctrl(control());
2495 2502 } __ else_(); {
2496 2503 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2497 2504 } __ end_if();
2498 2505 // Final sync IdealKit and GraphKit.
2499 2506 sync_kit(ideal);
2500 2507 #undef __
2501 2508 }
2502 2509 }
2503 2510 }
2504 2511
2505 2512 if (is_volatile) {
2506 2513 if (!is_store)
2507 2514 insert_mem_bar(Op_MemBarAcquire);
2508 2515 else
2509 2516 insert_mem_bar(Op_MemBarVolatile);
2510 2517 }
2511 2518
2512 2519 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2513 2520
2514 2521 return true;
2515 2522 }
2516 2523
2517 2524 //----------------------------inline_unsafe_prefetch----------------------------
2518 2525
2519 2526 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2520 2527 #ifndef PRODUCT
2521 2528 {
2522 2529 ResourceMark rm;
2523 2530 // Check the signatures.
2524 2531 ciSignature* sig = signature();
2525 2532 #ifdef ASSERT
2526 2533 // Object getObject(Object base, int/long offset), etc.
2527 2534 BasicType rtype = sig->return_type()->basic_type();
2528 2535 if (!is_native_ptr) {
2529 2536 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2530 2537 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2531 2538 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2532 2539 } else {
2533 2540 assert(sig->count() == 1, "native prefetch has 1 argument");
2534 2541 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2535 2542 }
2536 2543 #endif // ASSERT
2537 2544 }
2538 2545 #endif // !PRODUCT
2539 2546
2540 2547 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2541 2548
2542 2549 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2543 2550 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2544 2551
2545 2552 debug_only(int saved_sp = _sp);
2546 2553 _sp += nargs;
2547 2554
2548 2555 // Build address expression. See the code in inline_unsafe_access.
2549 2556 Node *adr;
2550 2557 if (!is_native_ptr) {
2551 2558 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2552 2559 Node* offset = pop_pair();
2553 2560 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2554 2561 Node* base = pop();
2555 2562 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2556 2563 // to be plain byte offsets, which are also the same as those accepted
2557 2564 // by oopDesc::field_base.
2558 2565 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2559 2566 "fieldOffset must be byte-scaled");
2560 2567 // 32-bit machines ignore the high half!
2561 2568 offset = ConvL2X(offset);
2562 2569 adr = make_unsafe_address(base, offset);
2563 2570 } else {
2564 2571 Node* ptr = pop_pair();
2565 2572 // Adjust Java long to machine word:
2566 2573 ptr = ConvL2X(ptr);
2567 2574 adr = make_unsafe_address(NULL, ptr);
2568 2575 }
2569 2576
2570 2577 if (is_static) {
2571 2578 assert(saved_sp == _sp, "must have correct argument count");
2572 2579 } else {
2573 2580 // Pop receiver last: it was pushed first.
2574 2581 Node *receiver = pop();
2575 2582 assert(saved_sp == _sp, "must have correct argument count");
2576 2583
2577 2584 // Null check on self without removing any arguments. The argument
2578 2585 // null check technically happens in the wrong place, which can lead to
2579 2586 // invalid stack traces when the primitive is inlined into a method
2580 2587 // which handles NullPointerExceptions.
2581 2588 _sp += nargs;
2582 2589 do_null_check(receiver, T_OBJECT);
2583 2590 _sp -= nargs;
2584 2591 if (stopped()) {
2585 2592 return true;
2586 2593 }
2587 2594 }
2588 2595
2589 2596 // Generate the read or write prefetch
2590 2597 Node *prefetch;
2591 2598 if (is_store) {
2592 2599 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2593 2600 } else {
2594 2601 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2595 2602 }
2596 2603 prefetch->init_req(0, control());
2597 2604 set_i_o(_gvn.transform(prefetch));
2598 2605
2599 2606 return true;
2600 2607 }
2601 2608
2602 2609 //----------------------------inline_unsafe_CAS----------------------------
2603 2610
2604 2611 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2605 2612 // This basic scheme here is the same as inline_unsafe_access, but
2606 2613 // differs in enough details that combining them would make the code
2607 2614 // overly confusing. (This is a true fact! I originally combined
2608 2615 // them, but even I was confused by it!) As much code/comments as
2609 2616 // possible are retained from inline_unsafe_access though to make
2610 2617 // the correspondences clearer. - dl
2611 2618
2612 2619 if (callee()->is_static()) return false; // caller must have the capability!
2613 2620
2614 2621 #ifndef PRODUCT
2615 2622 {
2616 2623 ResourceMark rm;
2617 2624 // Check the signatures.
2618 2625 ciSignature* sig = signature();
2619 2626 #ifdef ASSERT
2620 2627 BasicType rtype = sig->return_type()->basic_type();
2621 2628 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2622 2629 assert(sig->count() == 4, "CAS has 4 arguments");
2623 2630 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2624 2631 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2625 2632 #endif // ASSERT
2626 2633 }
2627 2634 #endif //PRODUCT
2628 2635
2629 2636 // number of stack slots per value argument (1 or 2)
2630 2637 int type_words = type2size[type];
2631 2638
2632 2639 // Cannot inline wide CAS on machines that don't support it natively
2633 2640 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2634 2641 return false;
2635 2642
2636 2643 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2637 2644
2638 2645 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2639 2646 int nargs = 1 + 1 + 2 + type_words + type_words;
2640 2647
2641 2648 // pop arguments: newval, oldval, offset, base, and receiver
2642 2649 debug_only(int saved_sp = _sp);
2643 2650 _sp += nargs;
2644 2651 Node* newval = (type_words == 1) ? pop() : pop_pair();
2645 2652 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2646 2653 Node *offset = pop_pair();
2647 2654 Node *base = pop();
2648 2655 Node *receiver = pop();
2649 2656 assert(saved_sp == _sp, "must have correct argument count");
2650 2657
2651 2658 // Null check receiver.
2652 2659 _sp += nargs;
2653 2660 do_null_check(receiver, T_OBJECT);
2654 2661 _sp -= nargs;
2655 2662 if (stopped()) {
2656 2663 return true;
2657 2664 }
2658 2665
2659 2666 // Build field offset expression.
2660 2667 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2661 2668 // to be plain byte offsets, which are also the same as those accepted
2662 2669 // by oopDesc::field_base.
2663 2670 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2664 2671 // 32-bit machines ignore the high half of long offsets
2665 2672 offset = ConvL2X(offset);
2666 2673 Node* adr = make_unsafe_address(base, offset);
2667 2674 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2668 2675
2669 2676 // (Unlike inline_unsafe_access, there seems no point in trying
2670 2677 // to refine types. Just use the coarse types here.
2671 2678 const Type *value_type = Type::get_const_basic_type(type);
2672 2679 Compile::AliasType* alias_type = C->alias_type(adr_type);
2673 2680 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2674 2681 int alias_idx = C->get_alias_index(adr_type);
2675 2682
2676 2683 // Memory-model-wise, a CAS acts like a little synchronized block,
2677 2684 // so needs barriers on each side. These don't translate into
2678 2685 // actual barriers on most machines, but we still need rest of
2679 2686 // compiler to respect ordering.
2680 2687
2681 2688 insert_mem_bar(Op_MemBarRelease);
2682 2689 insert_mem_bar(Op_MemBarCPUOrder);
2683 2690
2684 2691 // 4984716: MemBars must be inserted before this
2685 2692 // memory node in order to avoid a false
2686 2693 // dependency which will confuse the scheduler.
2687 2694 Node *mem = memory(alias_idx);
2688 2695
2689 2696 // For now, we handle only those cases that actually exist: ints,
2690 2697 // longs, and Object. Adding others should be straightforward.
2691 2698 Node* cas;
2692 2699 switch(type) {
2693 2700 case T_INT:
2694 2701 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2695 2702 break;
2696 2703 case T_LONG:
2697 2704 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2698 2705 break;
2699 2706 case T_OBJECT:
2700 2707 // reference stores need a store barrier.
2701 2708 // (They don't if CAS fails, but it isn't worth checking.)
2702 2709 pre_barrier(true /* do_load*/,
2703 2710 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2704 2711 NULL /* pre_val*/,
2705 2712 T_OBJECT);
2706 2713 #ifdef _LP64
2707 2714 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2708 2715 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2709 2716 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2710 2717 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2711 2718 newval_enc, oldval_enc));
2712 2719 } else
2713 2720 #endif
2714 2721 {
2715 2722 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2716 2723 }
2717 2724 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2718 2725 break;
2719 2726 default:
2720 2727 ShouldNotReachHere();
2721 2728 break;
2722 2729 }
2723 2730
2724 2731 // SCMemProjNodes represent the memory state of CAS. Their main
2725 2732 // role is to prevent CAS nodes from being optimized away when their
2726 2733 // results aren't used.
2727 2734 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2728 2735 set_memory(proj, alias_idx);
2729 2736
2730 2737 // Add the trailing membar surrounding the access
2731 2738 insert_mem_bar(Op_MemBarCPUOrder);
2732 2739 insert_mem_bar(Op_MemBarAcquire);
2733 2740
2734 2741 push(cas);
2735 2742 return true;
2736 2743 }
2737 2744
2738 2745 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2739 2746 // This is another variant of inline_unsafe_access, differing in
2740 2747 // that it always issues store-store ("release") barrier and ensures
2741 2748 // store-atomicity (which only matters for "long").
2742 2749
2743 2750 if (callee()->is_static()) return false; // caller must have the capability!
2744 2751
2745 2752 #ifndef PRODUCT
2746 2753 {
2747 2754 ResourceMark rm;
2748 2755 // Check the signatures.
2749 2756 ciSignature* sig = signature();
2750 2757 #ifdef ASSERT
2751 2758 BasicType rtype = sig->return_type()->basic_type();
2752 2759 assert(rtype == T_VOID, "must return void");
2753 2760 assert(sig->count() == 3, "has 3 arguments");
2754 2761 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2755 2762 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2756 2763 #endif // ASSERT
2757 2764 }
2758 2765 #endif //PRODUCT
2759 2766
2760 2767 // number of stack slots per value argument (1 or 2)
2761 2768 int type_words = type2size[type];
2762 2769
2763 2770 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2764 2771
2765 2772 // Argument words: "this" plus oop plus offset plus value;
2766 2773 int nargs = 1 + 1 + 2 + type_words;
2767 2774
2768 2775 // pop arguments: val, offset, base, and receiver
2769 2776 debug_only(int saved_sp = _sp);
2770 2777 _sp += nargs;
2771 2778 Node* val = (type_words == 1) ? pop() : pop_pair();
2772 2779 Node *offset = pop_pair();
2773 2780 Node *base = pop();
2774 2781 Node *receiver = pop();
2775 2782 assert(saved_sp == _sp, "must have correct argument count");
2776 2783
2777 2784 // Null check receiver.
2778 2785 _sp += nargs;
2779 2786 do_null_check(receiver, T_OBJECT);
2780 2787 _sp -= nargs;
2781 2788 if (stopped()) {
2782 2789 return true;
2783 2790 }
2784 2791
2785 2792 // Build field offset expression.
2786 2793 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2787 2794 // 32-bit machines ignore the high half of long offsets
2788 2795 offset = ConvL2X(offset);
2789 2796 Node* adr = make_unsafe_address(base, offset);
2790 2797 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2791 2798 const Type *value_type = Type::get_const_basic_type(type);
2792 2799 Compile::AliasType* alias_type = C->alias_type(adr_type);
2793 2800
2794 2801 insert_mem_bar(Op_MemBarRelease);
2795 2802 insert_mem_bar(Op_MemBarCPUOrder);
2796 2803 // Ensure that the store is atomic for longs:
2797 2804 bool require_atomic_access = true;
2798 2805 Node* store;
2799 2806 if (type == T_OBJECT) // reference stores need a store barrier.
2800 2807 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2801 2808 else {
2802 2809 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2803 2810 }
2804 2811 insert_mem_bar(Op_MemBarCPUOrder);
2805 2812 return true;
2806 2813 }
2807 2814
2808 2815 bool LibraryCallKit::inline_unsafe_allocate() {
2809 2816 if (callee()->is_static()) return false; // caller must have the capability!
2810 2817 int nargs = 1 + 1;
2811 2818 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2812 2819 null_check_receiver(callee()); // check then ignore argument(0)
2813 2820 _sp += nargs; // set original stack for use by uncommon_trap
2814 2821 Node* cls = do_null_check(argument(1), T_OBJECT);
2815 2822 _sp -= nargs;
2816 2823 if (stopped()) return true;
2817 2824
2818 2825 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2819 2826 _sp += nargs; // set original stack for use by uncommon_trap
2820 2827 kls = do_null_check(kls, T_OBJECT);
2821 2828 _sp -= nargs;
2822 2829 if (stopped()) return true; // argument was like int.class
2823 2830
2824 2831 // Note: The argument might still be an illegal value like
2825 2832 // Serializable.class or Object[].class. The runtime will handle it.
2826 2833 // But we must make an explicit check for initialization.
2827 2834 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2828 2835 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2829 2836 Node* bits = intcon(instanceKlass::fully_initialized);
2830 2837 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2831 2838 // The 'test' is non-zero if we need to take a slow path.
2832 2839
2833 2840 Node* obj = new_instance(kls, test);
2834 2841 push(obj);
2835 2842
2836 2843 return true;
2837 2844 }
2838 2845
2839 2846 //------------------------inline_native_time_funcs--------------
2840 2847 // inline code for System.currentTimeMillis() and System.nanoTime()
2841 2848 // these have the same type and signature
2842 2849 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2843 2850 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2844 2851 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2845 2852 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2846 2853 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2847 2854 const TypePtr* no_memory_effects = NULL;
2848 2855 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2849 2856 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2850 2857 #ifdef ASSERT
2851 2858 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2852 2859 assert(value_top == top(), "second value must be top");
2853 2860 #endif
2854 2861 push_pair(value);
2855 2862 return true;
2856 2863 }
2857 2864
2858 2865 //------------------------inline_native_currentThread------------------
2859 2866 bool LibraryCallKit::inline_native_currentThread() {
2860 2867 Node* junk = NULL;
2861 2868 push(generate_current_thread(junk));
2862 2869 return true;
2863 2870 }
2864 2871
2865 2872 //------------------------inline_native_isInterrupted------------------
2866 2873 bool LibraryCallKit::inline_native_isInterrupted() {
2867 2874 const int nargs = 1+1; // receiver + boolean
2868 2875 assert(nargs == arg_size(), "sanity");
2869 2876 // Add a fast path to t.isInterrupted(clear_int):
2870 2877 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2871 2878 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2872 2879 // So, in the common case that the interrupt bit is false,
2873 2880 // we avoid making a call into the VM. Even if the interrupt bit
2874 2881 // is true, if the clear_int argument is false, we avoid the VM call.
2875 2882 // However, if the receiver is not currentThread, we must call the VM,
2876 2883 // because there must be some locking done around the operation.
2877 2884
2878 2885 // We only go to the fast case code if we pass two guards.
2879 2886 // Paths which do not pass are accumulated in the slow_region.
2880 2887 RegionNode* slow_region = new (C, 1) RegionNode(1);
2881 2888 record_for_igvn(slow_region);
2882 2889 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2883 2890 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2884 2891 enum { no_int_result_path = 1,
2885 2892 no_clear_result_path = 2,
2886 2893 slow_result_path = 3
2887 2894 };
2888 2895
2889 2896 // (a) Receiving thread must be the current thread.
2890 2897 Node* rec_thr = argument(0);
2891 2898 Node* tls_ptr = NULL;
2892 2899 Node* cur_thr = generate_current_thread(tls_ptr);
2893 2900 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2894 2901 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2895 2902
2896 2903 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2897 2904 if (!known_current_thread)
2898 2905 generate_slow_guard(bol_thr, slow_region);
2899 2906
2900 2907 // (b) Interrupt bit on TLS must be false.
2901 2908 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2902 2909 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2903 2910 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2904 2911 // Set the control input on the field _interrupted read to prevent it floating up.
2905 2912 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2906 2913 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2907 2914 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2908 2915
2909 2916 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2910 2917
2911 2918 // First fast path: if (!TLS._interrupted) return false;
2912 2919 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2913 2920 result_rgn->init_req(no_int_result_path, false_bit);
2914 2921 result_val->init_req(no_int_result_path, intcon(0));
2915 2922
2916 2923 // drop through to next case
2917 2924 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2918 2925
2919 2926 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2920 2927 Node* clr_arg = argument(1);
2921 2928 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2922 2929 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2923 2930 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2924 2931
2925 2932 // Second fast path: ... else if (!clear_int) return true;
2926 2933 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2927 2934 result_rgn->init_req(no_clear_result_path, false_arg);
2928 2935 result_val->init_req(no_clear_result_path, intcon(1));
2929 2936
2930 2937 // drop through to next case
2931 2938 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2932 2939
2933 2940 // (d) Otherwise, go to the slow path.
2934 2941 slow_region->add_req(control());
2935 2942 set_control( _gvn.transform(slow_region) );
2936 2943
2937 2944 if (stopped()) {
2938 2945 // There is no slow path.
2939 2946 result_rgn->init_req(slow_result_path, top());
2940 2947 result_val->init_req(slow_result_path, top());
2941 2948 } else {
2942 2949 // non-virtual because it is a private non-static
2943 2950 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2944 2951
2945 2952 Node* slow_val = set_results_for_java_call(slow_call);
2946 2953 // this->control() comes from set_results_for_java_call
2947 2954
2948 2955 // If we know that the result of the slow call will be true, tell the optimizer!
2949 2956 if (known_current_thread) slow_val = intcon(1);
2950 2957
2951 2958 Node* fast_io = slow_call->in(TypeFunc::I_O);
2952 2959 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2953 2960 // These two phis are pre-filled with copies of of the fast IO and Memory
2954 2961 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2955 2962 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2956 2963
2957 2964 result_rgn->init_req(slow_result_path, control());
2958 2965 io_phi ->init_req(slow_result_path, i_o());
2959 2966 mem_phi ->init_req(slow_result_path, reset_memory());
2960 2967 result_val->init_req(slow_result_path, slow_val);
2961 2968
2962 2969 set_all_memory( _gvn.transform(mem_phi) );
2963 2970 set_i_o( _gvn.transform(io_phi) );
2964 2971 }
2965 2972
2966 2973 push_result(result_rgn, result_val);
2967 2974 C->set_has_split_ifs(true); // Has chance for split-if optimization
2968 2975
2969 2976 return true;
2970 2977 }
2971 2978
2972 2979 //---------------------------load_mirror_from_klass----------------------------
2973 2980 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2974 2981 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2975 2982 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2976 2983 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2977 2984 }
2978 2985
2979 2986 //-----------------------load_klass_from_mirror_common-------------------------
2980 2987 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2981 2988 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2982 2989 // and branch to the given path on the region.
2983 2990 // If never_see_null, take an uncommon trap on null, so we can optimistically
2984 2991 // compile for the non-null case.
2985 2992 // If the region is NULL, force never_see_null = true.
2986 2993 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2987 2994 bool never_see_null,
2988 2995 int nargs,
2989 2996 RegionNode* region,
2990 2997 int null_path,
2991 2998 int offset) {
2992 2999 if (region == NULL) never_see_null = true;
2993 3000 Node* p = basic_plus_adr(mirror, offset);
2994 3001 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2995 3002 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2996 3003 _sp += nargs; // any deopt will start just before call to enclosing method
2997 3004 Node* null_ctl = top();
2998 3005 kls = null_check_oop(kls, &null_ctl, never_see_null);
2999 3006 if (region != NULL) {
3000 3007 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3001 3008 region->init_req(null_path, null_ctl);
3002 3009 } else {
3003 3010 assert(null_ctl == top(), "no loose ends");
3004 3011 }
3005 3012 _sp -= nargs;
3006 3013 return kls;
3007 3014 }
3008 3015
3009 3016 //--------------------(inline_native_Class_query helpers)---------------------
3010 3017 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3011 3018 // Fall through if (mods & mask) == bits, take the guard otherwise.
3012 3019 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3013 3020 // Branch around if the given klass has the given modifier bit set.
3014 3021 // Like generate_guard, adds a new path onto the region.
3015 3022 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3016 3023 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3017 3024 Node* mask = intcon(modifier_mask);
3018 3025 Node* bits = intcon(modifier_bits);
3019 3026 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
3020 3027 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
3021 3028 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
3022 3029 return generate_fair_guard(bol, region);
3023 3030 }
3024 3031 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3025 3032 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3026 3033 }
3027 3034
3028 3035 //-------------------------inline_native_Class_query-------------------
3029 3036 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3030 3037 int nargs = 1+0; // just the Class mirror, in most cases
3031 3038 const Type* return_type = TypeInt::BOOL;
3032 3039 Node* prim_return_value = top(); // what happens if it's a primitive class?
3033 3040 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3034 3041 bool expect_prim = false; // most of these guys expect to work on refs
3035 3042
3036 3043 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3037 3044
3038 3045 switch (id) {
3039 3046 case vmIntrinsics::_isInstance:
3040 3047 nargs = 1+1; // the Class mirror, plus the object getting queried about
3041 3048 // nothing is an instance of a primitive type
3042 3049 prim_return_value = intcon(0);
3043 3050 break;
3044 3051 case vmIntrinsics::_getModifiers:
3045 3052 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3046 3053 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3047 3054 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3048 3055 break;
3049 3056 case vmIntrinsics::_isInterface:
3050 3057 prim_return_value = intcon(0);
3051 3058 break;
3052 3059 case vmIntrinsics::_isArray:
3053 3060 prim_return_value = intcon(0);
3054 3061 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3055 3062 break;
3056 3063 case vmIntrinsics::_isPrimitive:
3057 3064 prim_return_value = intcon(1);
3058 3065 expect_prim = true; // obviously
3059 3066 break;
3060 3067 case vmIntrinsics::_getSuperclass:
3061 3068 prim_return_value = null();
3062 3069 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3063 3070 break;
3064 3071 case vmIntrinsics::_getComponentType:
3065 3072 prim_return_value = null();
3066 3073 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3067 3074 break;
3068 3075 case vmIntrinsics::_getClassAccessFlags:
3069 3076 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3070 3077 return_type = TypeInt::INT; // not bool! 6297094
3071 3078 break;
3072 3079 default:
3073 3080 ShouldNotReachHere();
3074 3081 }
3075 3082
3076 3083 Node* mirror = argument(0);
3077 3084 Node* obj = (nargs <= 1)? top(): argument(1);
3078 3085
3079 3086 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3080 3087 if (mirror_con == NULL) return false; // cannot happen?
3081 3088
3082 3089 #ifndef PRODUCT
3083 3090 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3084 3091 ciType* k = mirror_con->java_mirror_type();
3085 3092 if (k) {
3086 3093 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3087 3094 k->print_name();
3088 3095 tty->cr();
3089 3096 }
3090 3097 }
3091 3098 #endif
3092 3099
3093 3100 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3094 3101 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3095 3102 record_for_igvn(region);
3096 3103 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
3097 3104
3098 3105 // The mirror will never be null of Reflection.getClassAccessFlags, however
3099 3106 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3100 3107 // if it is. See bug 4774291.
3101 3108
3102 3109 // For Reflection.getClassAccessFlags(), the null check occurs in
3103 3110 // the wrong place; see inline_unsafe_access(), above, for a similar
3104 3111 // situation.
3105 3112 _sp += nargs; // set original stack for use by uncommon_trap
3106 3113 mirror = do_null_check(mirror, T_OBJECT);
3107 3114 _sp -= nargs;
3108 3115 // If mirror or obj is dead, only null-path is taken.
3109 3116 if (stopped()) return true;
3110 3117
3111 3118 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3112 3119
3113 3120 // Now load the mirror's klass metaobject, and null-check it.
3114 3121 // Side-effects region with the control path if the klass is null.
3115 3122 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3116 3123 region, _prim_path);
3117 3124 // If kls is null, we have a primitive mirror.
3118 3125 phi->init_req(_prim_path, prim_return_value);
3119 3126 if (stopped()) { push_result(region, phi); return true; }
3120 3127
3121 3128 Node* p; // handy temp
3122 3129 Node* null_ctl;
3123 3130
3124 3131 // Now that we have the non-null klass, we can perform the real query.
3125 3132 // For constant classes, the query will constant-fold in LoadNode::Value.
3126 3133 Node* query_value = top();
3127 3134 switch (id) {
3128 3135 case vmIntrinsics::_isInstance:
3129 3136 // nothing is an instance of a primitive type
3130 3137 _sp += nargs; // gen_instanceof might do an uncommon trap
3131 3138 query_value = gen_instanceof(obj, kls);
3132 3139 _sp -= nargs;
3133 3140 break;
3134 3141
3135 3142 case vmIntrinsics::_getModifiers:
3136 3143 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
3137 3144 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3138 3145 break;
3139 3146
3140 3147 case vmIntrinsics::_isInterface:
3141 3148 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3142 3149 if (generate_interface_guard(kls, region) != NULL)
3143 3150 // A guard was added. If the guard is taken, it was an interface.
3144 3151 phi->add_req(intcon(1));
3145 3152 // If we fall through, it's a plain class.
3146 3153 query_value = intcon(0);
3147 3154 break;
3148 3155
3149 3156 case vmIntrinsics::_isArray:
3150 3157 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3151 3158 if (generate_array_guard(kls, region) != NULL)
3152 3159 // A guard was added. If the guard is taken, it was an array.
3153 3160 phi->add_req(intcon(1));
3154 3161 // If we fall through, it's a plain class.
3155 3162 query_value = intcon(0);
3156 3163 break;
3157 3164
3158 3165 case vmIntrinsics::_isPrimitive:
3159 3166 query_value = intcon(0); // "normal" path produces false
3160 3167 break;
3161 3168
3162 3169 case vmIntrinsics::_getSuperclass:
3163 3170 // The rules here are somewhat unfortunate, but we can still do better
3164 3171 // with random logic than with a JNI call.
3165 3172 // Interfaces store null or Object as _super, but must report null.
3166 3173 // Arrays store an intermediate super as _super, but must report Object.
3167 3174 // Other types can report the actual _super.
3168 3175 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3169 3176 if (generate_interface_guard(kls, region) != NULL)
3170 3177 // A guard was added. If the guard is taken, it was an interface.
3171 3178 phi->add_req(null());
3172 3179 if (generate_array_guard(kls, region) != NULL)
3173 3180 // A guard was added. If the guard is taken, it was an array.
3174 3181 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3175 3182 // If we fall through, it's a plain class. Get its _super.
3176 3183 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
3177 3184 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3178 3185 null_ctl = top();
3179 3186 kls = null_check_oop(kls, &null_ctl);
3180 3187 if (null_ctl != top()) {
3181 3188 // If the guard is taken, Object.superClass is null (both klass and mirror).
3182 3189 region->add_req(null_ctl);
3183 3190 phi ->add_req(null());
3184 3191 }
3185 3192 if (!stopped()) {
3186 3193 query_value = load_mirror_from_klass(kls);
3187 3194 }
3188 3195 break;
3189 3196
3190 3197 case vmIntrinsics::_getComponentType:
3191 3198 if (generate_array_guard(kls, region) != NULL) {
3192 3199 // Be sure to pin the oop load to the guard edge just created:
3193 3200 Node* is_array_ctrl = region->in(region->req()-1);
3194 3201 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3195 3202 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3196 3203 phi->add_req(cmo);
3197 3204 }
3198 3205 query_value = null(); // non-array case is null
3199 3206 break;
3200 3207
3201 3208 case vmIntrinsics::_getClassAccessFlags:
3202 3209 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3203 3210 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3204 3211 break;
3205 3212
3206 3213 default:
3207 3214 ShouldNotReachHere();
3208 3215 }
3209 3216
3210 3217 // Fall-through is the normal case of a query to a real class.
3211 3218 phi->init_req(1, query_value);
3212 3219 region->init_req(1, control());
3213 3220
3214 3221 push_result(region, phi);
3215 3222 C->set_has_split_ifs(true); // Has chance for split-if optimization
3216 3223
3217 3224 return true;
3218 3225 }
3219 3226
3220 3227 //--------------------------inline_native_subtype_check------------------------
3221 3228 // This intrinsic takes the JNI calls out of the heart of
3222 3229 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3223 3230 bool LibraryCallKit::inline_native_subtype_check() {
3224 3231 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3225 3232
3226 3233 // Pull both arguments off the stack.
3227 3234 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3228 3235 args[0] = argument(0);
3229 3236 args[1] = argument(1);
3230 3237 Node* klasses[2]; // corresponding Klasses: superk, subk
3231 3238 klasses[0] = klasses[1] = top();
3232 3239
3233 3240 enum {
3234 3241 // A full decision tree on {superc is prim, subc is prim}:
3235 3242 _prim_0_path = 1, // {P,N} => false
3236 3243 // {P,P} & superc!=subc => false
3237 3244 _prim_same_path, // {P,P} & superc==subc => true
3238 3245 _prim_1_path, // {N,P} => false
3239 3246 _ref_subtype_path, // {N,N} & subtype check wins => true
3240 3247 _both_ref_path, // {N,N} & subtype check loses => false
3241 3248 PATH_LIMIT
3242 3249 };
3243 3250
3244 3251 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3245 3252 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3246 3253 record_for_igvn(region);
3247 3254
3248 3255 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3249 3256 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3250 3257 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3251 3258
3252 3259 // First null-check both mirrors and load each mirror's klass metaobject.
3253 3260 int which_arg;
3254 3261 for (which_arg = 0; which_arg <= 1; which_arg++) {
3255 3262 Node* arg = args[which_arg];
3256 3263 _sp += nargs; // set original stack for use by uncommon_trap
3257 3264 arg = do_null_check(arg, T_OBJECT);
3258 3265 _sp -= nargs;
3259 3266 if (stopped()) break;
3260 3267 args[which_arg] = _gvn.transform(arg);
3261 3268
3262 3269 Node* p = basic_plus_adr(arg, class_klass_offset);
3263 3270 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3264 3271 klasses[which_arg] = _gvn.transform(kls);
3265 3272 }
3266 3273
3267 3274 // Having loaded both klasses, test each for null.
3268 3275 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3269 3276 for (which_arg = 0; which_arg <= 1; which_arg++) {
3270 3277 Node* kls = klasses[which_arg];
3271 3278 Node* null_ctl = top();
3272 3279 _sp += nargs; // set original stack for use by uncommon_trap
3273 3280 kls = null_check_oop(kls, &null_ctl, never_see_null);
3274 3281 _sp -= nargs;
3275 3282 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3276 3283 region->init_req(prim_path, null_ctl);
3277 3284 if (stopped()) break;
3278 3285 klasses[which_arg] = kls;
3279 3286 }
3280 3287
3281 3288 if (!stopped()) {
3282 3289 // now we have two reference types, in klasses[0..1]
3283 3290 Node* subk = klasses[1]; // the argument to isAssignableFrom
3284 3291 Node* superk = klasses[0]; // the receiver
3285 3292 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3286 3293 // now we have a successful reference subtype check
3287 3294 region->set_req(_ref_subtype_path, control());
3288 3295 }
3289 3296
3290 3297 // If both operands are primitive (both klasses null), then
3291 3298 // we must return true when they are identical primitives.
3292 3299 // It is convenient to test this after the first null klass check.
3293 3300 set_control(region->in(_prim_0_path)); // go back to first null check
3294 3301 if (!stopped()) {
3295 3302 // Since superc is primitive, make a guard for the superc==subc case.
3296 3303 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3297 3304 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3298 3305 generate_guard(bol_eq, region, PROB_FAIR);
3299 3306 if (region->req() == PATH_LIMIT+1) {
3300 3307 // A guard was added. If the added guard is taken, superc==subc.
3301 3308 region->swap_edges(PATH_LIMIT, _prim_same_path);
3302 3309 region->del_req(PATH_LIMIT);
3303 3310 }
3304 3311 region->set_req(_prim_0_path, control()); // Not equal after all.
3305 3312 }
3306 3313
3307 3314 // these are the only paths that produce 'true':
3308 3315 phi->set_req(_prim_same_path, intcon(1));
3309 3316 phi->set_req(_ref_subtype_path, intcon(1));
3310 3317
3311 3318 // pull together the cases:
3312 3319 assert(region->req() == PATH_LIMIT, "sane region");
3313 3320 for (uint i = 1; i < region->req(); i++) {
3314 3321 Node* ctl = region->in(i);
3315 3322 if (ctl == NULL || ctl == top()) {
3316 3323 region->set_req(i, top());
3317 3324 phi ->set_req(i, top());
3318 3325 } else if (phi->in(i) == NULL) {
3319 3326 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3320 3327 }
3321 3328 }
3322 3329
3323 3330 set_control(_gvn.transform(region));
3324 3331 push(_gvn.transform(phi));
3325 3332
3326 3333 return true;
3327 3334 }
3328 3335
3329 3336 //---------------------generate_array_guard_common------------------------
3330 3337 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3331 3338 bool obj_array, bool not_array) {
3332 3339 // If obj_array/non_array==false/false:
3333 3340 // Branch around if the given klass is in fact an array (either obj or prim).
3334 3341 // If obj_array/non_array==false/true:
3335 3342 // Branch around if the given klass is not an array klass of any kind.
3336 3343 // If obj_array/non_array==true/true:
3337 3344 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3338 3345 // If obj_array/non_array==true/false:
3339 3346 // Branch around if the kls is an oop array (Object[] or subtype)
3340 3347 //
3341 3348 // Like generate_guard, adds a new path onto the region.
3342 3349 jint layout_con = 0;
3343 3350 Node* layout_val = get_layout_helper(kls, layout_con);
3344 3351 if (layout_val == NULL) {
3345 3352 bool query = (obj_array
3346 3353 ? Klass::layout_helper_is_objArray(layout_con)
3347 3354 : Klass::layout_helper_is_javaArray(layout_con));
3348 3355 if (query == not_array) {
3349 3356 return NULL; // never a branch
3350 3357 } else { // always a branch
3351 3358 Node* always_branch = control();
3352 3359 if (region != NULL)
3353 3360 region->add_req(always_branch);
3354 3361 set_control(top());
3355 3362 return always_branch;
3356 3363 }
3357 3364 }
3358 3365 // Now test the correct condition.
3359 3366 jint nval = (obj_array
3360 3367 ? ((jint)Klass::_lh_array_tag_type_value
3361 3368 << Klass::_lh_array_tag_shift)
3362 3369 : Klass::_lh_neutral_value);
3363 3370 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3364 3371 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3365 3372 // invert the test if we are looking for a non-array
3366 3373 if (not_array) btest = BoolTest(btest).negate();
3367 3374 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3368 3375 return generate_fair_guard(bol, region);
3369 3376 }
3370 3377
3371 3378
3372 3379 //-----------------------inline_native_newArray--------------------------
3373 3380 bool LibraryCallKit::inline_native_newArray() {
3374 3381 int nargs = 2;
3375 3382 Node* mirror = argument(0);
3376 3383 Node* count_val = argument(1);
3377 3384
3378 3385 _sp += nargs; // set original stack for use by uncommon_trap
3379 3386 mirror = do_null_check(mirror, T_OBJECT);
3380 3387 _sp -= nargs;
3381 3388 // If mirror or obj is dead, only null-path is taken.
3382 3389 if (stopped()) return true;
3383 3390
3384 3391 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3385 3392 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3386 3393 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3387 3394 TypeInstPtr::NOTNULL);
3388 3395 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3389 3396 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3390 3397 TypePtr::BOTTOM);
3391 3398
3392 3399 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3393 3400 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3394 3401 nargs,
3395 3402 result_reg, _slow_path);
3396 3403 Node* normal_ctl = control();
3397 3404 Node* no_array_ctl = result_reg->in(_slow_path);
3398 3405
3399 3406 // Generate code for the slow case. We make a call to newArray().
3400 3407 set_control(no_array_ctl);
3401 3408 if (!stopped()) {
3402 3409 // Either the input type is void.class, or else the
3403 3410 // array klass has not yet been cached. Either the
3404 3411 // ensuing call will throw an exception, or else it
3405 3412 // will cache the array klass for next time.
3406 3413 PreserveJVMState pjvms(this);
3407 3414 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3408 3415 Node* slow_result = set_results_for_java_call(slow_call);
3409 3416 // this->control() comes from set_results_for_java_call
3410 3417 result_reg->set_req(_slow_path, control());
3411 3418 result_val->set_req(_slow_path, slow_result);
3412 3419 result_io ->set_req(_slow_path, i_o());
3413 3420 result_mem->set_req(_slow_path, reset_memory());
3414 3421 }
3415 3422
3416 3423 set_control(normal_ctl);
3417 3424 if (!stopped()) {
3418 3425 // Normal case: The array type has been cached in the java.lang.Class.
3419 3426 // The following call works fine even if the array type is polymorphic.
3420 3427 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3421 3428 Node* obj = new_array(klass_node, count_val, nargs);
3422 3429 result_reg->init_req(_normal_path, control());
3423 3430 result_val->init_req(_normal_path, obj);
3424 3431 result_io ->init_req(_normal_path, i_o());
3425 3432 result_mem->init_req(_normal_path, reset_memory());
3426 3433 }
3427 3434
3428 3435 // Return the combined state.
3429 3436 set_i_o( _gvn.transform(result_io) );
3430 3437 set_all_memory( _gvn.transform(result_mem) );
3431 3438 push_result(result_reg, result_val);
3432 3439 C->set_has_split_ifs(true); // Has chance for split-if optimization
3433 3440
3434 3441 return true;
3435 3442 }
3436 3443
3437 3444 //----------------------inline_native_getLength--------------------------
3438 3445 bool LibraryCallKit::inline_native_getLength() {
3439 3446 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3440 3447
3441 3448 int nargs = 1;
3442 3449 Node* array = argument(0);
3443 3450
3444 3451 _sp += nargs; // set original stack for use by uncommon_trap
3445 3452 array = do_null_check(array, T_OBJECT);
3446 3453 _sp -= nargs;
3447 3454
3448 3455 // If array is dead, only null-path is taken.
3449 3456 if (stopped()) return true;
3450 3457
3451 3458 // Deoptimize if it is a non-array.
3452 3459 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3453 3460
3454 3461 if (non_array != NULL) {
3455 3462 PreserveJVMState pjvms(this);
3456 3463 set_control(non_array);
3457 3464 _sp += nargs; // push the arguments back on the stack
3458 3465 uncommon_trap(Deoptimization::Reason_intrinsic,
3459 3466 Deoptimization::Action_maybe_recompile);
3460 3467 }
3461 3468
3462 3469 // If control is dead, only non-array-path is taken.
3463 3470 if (stopped()) return true;
3464 3471
3465 3472 // The works fine even if the array type is polymorphic.
3466 3473 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3467 3474 push( load_array_length(array) );
3468 3475
3469 3476 C->set_has_split_ifs(true); // Has chance for split-if optimization
3470 3477
3471 3478 return true;
3472 3479 }
3473 3480
3474 3481 //------------------------inline_array_copyOf----------------------------
3475 3482 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3476 3483 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3477 3484
3478 3485 // Restore the stack and pop off the arguments.
3479 3486 int nargs = 3 + (is_copyOfRange? 1: 0);
3480 3487 Node* original = argument(0);
3481 3488 Node* start = is_copyOfRange? argument(1): intcon(0);
3482 3489 Node* end = is_copyOfRange? argument(2): argument(1);
3483 3490 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3484 3491
3485 3492 Node* newcopy;
3486 3493
3487 3494 //set the original stack and the reexecute bit for the interpreter to reexecute
3488 3495 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3489 3496 { PreserveReexecuteState preexecs(this);
3490 3497 _sp += nargs;
3491 3498 jvms()->set_should_reexecute(true);
3492 3499
3493 3500 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3494 3501 original = do_null_check(original, T_OBJECT);
3495 3502
3496 3503 // Check if a null path was taken unconditionally.
3497 3504 if (stopped()) return true;
3498 3505
3499 3506 Node* orig_length = load_array_length(original);
3500 3507
3501 3508 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3502 3509 NULL, 0);
3503 3510 klass_node = do_null_check(klass_node, T_OBJECT);
3504 3511
3505 3512 RegionNode* bailout = new (C, 1) RegionNode(1);
3506 3513 record_for_igvn(bailout);
3507 3514
3508 3515 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3509 3516 // Bail out if that is so.
3510 3517 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3511 3518 if (not_objArray != NULL) {
3512 3519 // Improve the klass node's type from the new optimistic assumption:
3513 3520 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3514 3521 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3515 3522 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3516 3523 cast->init_req(0, control());
3517 3524 klass_node = _gvn.transform(cast);
3518 3525 }
3519 3526
3520 3527 // Bail out if either start or end is negative.
3521 3528 generate_negative_guard(start, bailout, &start);
3522 3529 generate_negative_guard(end, bailout, &end);
3523 3530
3524 3531 Node* length = end;
3525 3532 if (_gvn.type(start) != TypeInt::ZERO) {
3526 3533 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3527 3534 }
3528 3535
3529 3536 // Bail out if length is negative.
3530 3537 // ...Not needed, since the new_array will throw the right exception.
3531 3538 //generate_negative_guard(length, bailout, &length);
3532 3539
3533 3540 if (bailout->req() > 1) {
3534 3541 PreserveJVMState pjvms(this);
3535 3542 set_control( _gvn.transform(bailout) );
3536 3543 uncommon_trap(Deoptimization::Reason_intrinsic,
3537 3544 Deoptimization::Action_maybe_recompile);
3538 3545 }
3539 3546
3540 3547 if (!stopped()) {
3541 3548
3542 3549 // How many elements will we copy from the original?
3543 3550 // The answer is MinI(orig_length - start, length).
3544 3551 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3545 3552 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3546 3553
3547 3554 const bool raw_mem_only = true;
3548 3555 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3549 3556
3550 3557 // Generate a direct call to the right arraycopy function(s).
3551 3558 // We know the copy is disjoint but we might not know if the
3552 3559 // oop stores need checking.
3553 3560 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3554 3561 // This will fail a store-check if x contains any non-nulls.
3555 3562 bool disjoint_bases = true;
3556 3563 bool length_never_negative = true;
3557 3564 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3558 3565 original, start, newcopy, intcon(0), moved,
3559 3566 disjoint_bases, length_never_negative);
3560 3567 }
3561 3568 } //original reexecute and sp are set back here
3562 3569
3563 3570 if(!stopped()) {
3564 3571 push(newcopy);
3565 3572 }
3566 3573
3567 3574 C->set_has_split_ifs(true); // Has chance for split-if optimization
3568 3575
3569 3576 return true;
3570 3577 }
3571 3578
3572 3579
3573 3580 //----------------------generate_virtual_guard---------------------------
3574 3581 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3575 3582 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3576 3583 RegionNode* slow_region) {
3577 3584 ciMethod* method = callee();
3578 3585 int vtable_index = method->vtable_index();
3579 3586 // Get the methodOop out of the appropriate vtable entry.
3580 3587 int entry_offset = (instanceKlass::vtable_start_offset() +
3581 3588 vtable_index*vtableEntry::size()) * wordSize +
3582 3589 vtableEntry::method_offset_in_bytes();
3583 3590 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3584 3591 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3585 3592
3586 3593 // Compare the target method with the expected method (e.g., Object.hashCode).
3587 3594 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3588 3595
3589 3596 Node* native_call = makecon(native_call_addr);
3590 3597 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3591 3598 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3592 3599
3593 3600 return generate_slow_guard(test_native, slow_region);
3594 3601 }
3595 3602
3596 3603 //-----------------------generate_method_call----------------------------
3597 3604 // Use generate_method_call to make a slow-call to the real
3598 3605 // method if the fast path fails. An alternative would be to
3599 3606 // use a stub like OptoRuntime::slow_arraycopy_Java.
3600 3607 // This only works for expanding the current library call,
3601 3608 // not another intrinsic. (E.g., don't use this for making an
3602 3609 // arraycopy call inside of the copyOf intrinsic.)
3603 3610 CallJavaNode*
3604 3611 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3605 3612 // When compiling the intrinsic method itself, do not use this technique.
3606 3613 guarantee(callee() != C->method(), "cannot make slow-call to self");
3607 3614
3608 3615 ciMethod* method = callee();
3609 3616 // ensure the JVMS we have will be correct for this call
3610 3617 guarantee(method_id == method->intrinsic_id(), "must match");
3611 3618
3612 3619 const TypeFunc* tf = TypeFunc::make(method);
3613 3620 int tfdc = tf->domain()->cnt();
3614 3621 CallJavaNode* slow_call;
3615 3622 if (is_static) {
3616 3623 assert(!is_virtual, "");
3617 3624 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3618 3625 SharedRuntime::get_resolve_static_call_stub(),
3619 3626 method, bci());
3620 3627 } else if (is_virtual) {
3621 3628 null_check_receiver(method);
3622 3629 int vtable_index = methodOopDesc::invalid_vtable_index;
3623 3630 if (UseInlineCaches) {
3624 3631 // Suppress the vtable call
3625 3632 } else {
3626 3633 // hashCode and clone are not a miranda methods,
3627 3634 // so the vtable index is fixed.
3628 3635 // No need to use the linkResolver to get it.
3629 3636 vtable_index = method->vtable_index();
3630 3637 }
3631 3638 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3632 3639 SharedRuntime::get_resolve_virtual_call_stub(),
3633 3640 method, vtable_index, bci());
3634 3641 } else { // neither virtual nor static: opt_virtual
3635 3642 null_check_receiver(method);
3636 3643 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3637 3644 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3638 3645 method, bci());
3639 3646 slow_call->set_optimized_virtual(true);
3640 3647 }
3641 3648 set_arguments_for_java_call(slow_call);
3642 3649 set_edges_for_java_call(slow_call);
3643 3650 return slow_call;
3644 3651 }
3645 3652
3646 3653
3647 3654 //------------------------------inline_native_hashcode--------------------
3648 3655 // Build special case code for calls to hashCode on an object.
3649 3656 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3650 3657 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3651 3658 assert(!(is_virtual && is_static), "either virtual, special, or static");
3652 3659
3653 3660 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3654 3661
3655 3662 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3656 3663 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3657 3664 TypeInt::INT);
3658 3665 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3659 3666 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3660 3667 TypePtr::BOTTOM);
3661 3668 Node* obj = NULL;
3662 3669 if (!is_static) {
3663 3670 // Check for hashing null object
3664 3671 obj = null_check_receiver(callee());
3665 3672 if (stopped()) return true; // unconditionally null
3666 3673 result_reg->init_req(_null_path, top());
3667 3674 result_val->init_req(_null_path, top());
3668 3675 } else {
3669 3676 // Do a null check, and return zero if null.
3670 3677 // System.identityHashCode(null) == 0
3671 3678 obj = argument(0);
3672 3679 Node* null_ctl = top();
3673 3680 obj = null_check_oop(obj, &null_ctl);
3674 3681 result_reg->init_req(_null_path, null_ctl);
3675 3682 result_val->init_req(_null_path, _gvn.intcon(0));
3676 3683 }
3677 3684
3678 3685 // Unconditionally null? Then return right away.
3679 3686 if (stopped()) {
3680 3687 set_control( result_reg->in(_null_path) );
3681 3688 if (!stopped())
3682 3689 push( result_val ->in(_null_path) );
3683 3690 return true;
3684 3691 }
3685 3692
3686 3693 // After null check, get the object's klass.
3687 3694 Node* obj_klass = load_object_klass(obj);
3688 3695
3689 3696 // This call may be virtual (invokevirtual) or bound (invokespecial).
3690 3697 // For each case we generate slightly different code.
3691 3698
3692 3699 // We only go to the fast case code if we pass a number of guards. The
3693 3700 // paths which do not pass are accumulated in the slow_region.
3694 3701 RegionNode* slow_region = new (C, 1) RegionNode(1);
3695 3702 record_for_igvn(slow_region);
3696 3703
3697 3704 // If this is a virtual call, we generate a funny guard. We pull out
3698 3705 // the vtable entry corresponding to hashCode() from the target object.
3699 3706 // If the target method which we are calling happens to be the native
3700 3707 // Object hashCode() method, we pass the guard. We do not need this
3701 3708 // guard for non-virtual calls -- the caller is known to be the native
3702 3709 // Object hashCode().
3703 3710 if (is_virtual) {
3704 3711 generate_virtual_guard(obj_klass, slow_region);
3705 3712 }
3706 3713
3707 3714 // Get the header out of the object, use LoadMarkNode when available
3708 3715 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3709 3716 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3710 3717
3711 3718 // Test the header to see if it is unlocked.
3712 3719 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3713 3720 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3714 3721 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3715 3722 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3716 3723 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3717 3724
3718 3725 generate_slow_guard(test_unlocked, slow_region);
3719 3726
3720 3727 // Get the hash value and check to see that it has been properly assigned.
3721 3728 // We depend on hash_mask being at most 32 bits and avoid the use of
3722 3729 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3723 3730 // vm: see markOop.hpp.
3724 3731 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3725 3732 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3726 3733 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3727 3734 // This hack lets the hash bits live anywhere in the mark object now, as long
3728 3735 // as the shift drops the relevant bits into the low 32 bits. Note that
3729 3736 // Java spec says that HashCode is an int so there's no point in capturing
3730 3737 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3731 3738 hshifted_header = ConvX2I(hshifted_header);
3732 3739 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3733 3740
3734 3741 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3735 3742 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3736 3743 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3737 3744
3738 3745 generate_slow_guard(test_assigned, slow_region);
3739 3746
3740 3747 Node* init_mem = reset_memory();
3741 3748 // fill in the rest of the null path:
3742 3749 result_io ->init_req(_null_path, i_o());
3743 3750 result_mem->init_req(_null_path, init_mem);
3744 3751
3745 3752 result_val->init_req(_fast_path, hash_val);
3746 3753 result_reg->init_req(_fast_path, control());
3747 3754 result_io ->init_req(_fast_path, i_o());
3748 3755 result_mem->init_req(_fast_path, init_mem);
3749 3756
3750 3757 // Generate code for the slow case. We make a call to hashCode().
3751 3758 set_control(_gvn.transform(slow_region));
3752 3759 if (!stopped()) {
3753 3760 // No need for PreserveJVMState, because we're using up the present state.
3754 3761 set_all_memory(init_mem);
3755 3762 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3756 3763 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3757 3764 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3758 3765 Node* slow_result = set_results_for_java_call(slow_call);
3759 3766 // this->control() comes from set_results_for_java_call
3760 3767 result_reg->init_req(_slow_path, control());
3761 3768 result_val->init_req(_slow_path, slow_result);
3762 3769 result_io ->set_req(_slow_path, i_o());
3763 3770 result_mem ->set_req(_slow_path, reset_memory());
3764 3771 }
3765 3772
3766 3773 // Return the combined state.
3767 3774 set_i_o( _gvn.transform(result_io) );
3768 3775 set_all_memory( _gvn.transform(result_mem) );
3769 3776 push_result(result_reg, result_val);
3770 3777
3771 3778 return true;
3772 3779 }
3773 3780
3774 3781 //---------------------------inline_native_getClass----------------------------
3775 3782 // Build special case code for calls to getClass on an object.
3776 3783 bool LibraryCallKit::inline_native_getClass() {
3777 3784 Node* obj = null_check_receiver(callee());
3778 3785 if (stopped()) return true;
3779 3786 push( load_mirror_from_klass(load_object_klass(obj)) );
3780 3787 return true;
3781 3788 }
3782 3789
3783 3790 //-----------------inline_native_Reflection_getCallerClass---------------------
3784 3791 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3785 3792 //
3786 3793 // NOTE that this code must perform the same logic as
3787 3794 // vframeStream::security_get_caller_frame in that it must skip
3788 3795 // Method.invoke() and auxiliary frames.
3789 3796
3790 3797
3791 3798
3792 3799
3793 3800 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3794 3801 ciMethod* method = callee();
3795 3802
3796 3803 #ifndef PRODUCT
3797 3804 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3798 3805 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3799 3806 }
3800 3807 #endif
3801 3808
3802 3809 debug_only(int saved_sp = _sp);
3803 3810
3804 3811 // Argument words: (int depth)
3805 3812 int nargs = 1;
3806 3813
3807 3814 _sp += nargs;
3808 3815 Node* caller_depth_node = pop();
3809 3816
3810 3817 assert(saved_sp == _sp, "must have correct argument count");
3811 3818
3812 3819 // The depth value must be a constant in order for the runtime call
3813 3820 // to be eliminated.
3814 3821 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3815 3822 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3816 3823 #ifndef PRODUCT
3817 3824 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3818 3825 tty->print_cr(" Bailing out because caller depth was not a constant");
3819 3826 }
3820 3827 #endif
3821 3828 return false;
3822 3829 }
3823 3830 // Note that the JVM state at this point does not include the
3824 3831 // getCallerClass() frame which we are trying to inline. The
3825 3832 // semantics of getCallerClass(), however, are that the "first"
3826 3833 // frame is the getCallerClass() frame, so we subtract one from the
3827 3834 // requested depth before continuing. We don't inline requests of
3828 3835 // getCallerClass(0).
3829 3836 int caller_depth = caller_depth_type->get_con() - 1;
3830 3837 if (caller_depth < 0) {
3831 3838 #ifndef PRODUCT
3832 3839 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3833 3840 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3834 3841 }
3835 3842 #endif
3836 3843 return false;
3837 3844 }
3838 3845
3839 3846 if (!jvms()->has_method()) {
3840 3847 #ifndef PRODUCT
3841 3848 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3842 3849 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3843 3850 }
3844 3851 #endif
3845 3852 return false;
3846 3853 }
3847 3854 int _depth = jvms()->depth(); // cache call chain depth
3848 3855
3849 3856 // Walk back up the JVM state to find the caller at the required
3850 3857 // depth. NOTE that this code must perform the same logic as
3851 3858 // vframeStream::security_get_caller_frame in that it must skip
3852 3859 // Method.invoke() and auxiliary frames. Note also that depth is
3853 3860 // 1-based (1 is the bottom of the inlining).
3854 3861 int inlining_depth = _depth;
3855 3862 JVMState* caller_jvms = NULL;
3856 3863
3857 3864 if (inlining_depth > 0) {
3858 3865 caller_jvms = jvms();
3859 3866 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3860 3867 do {
3861 3868 // The following if-tests should be performed in this order
3862 3869 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3863 3870 // Skip a Method.invoke() or auxiliary frame
3864 3871 } else if (caller_depth > 0) {
3865 3872 // Skip real frame
3866 3873 --caller_depth;
3867 3874 } else {
3868 3875 // We're done: reached desired caller after skipping.
3869 3876 break;
3870 3877 }
3871 3878 caller_jvms = caller_jvms->caller();
3872 3879 --inlining_depth;
3873 3880 } while (inlining_depth > 0);
3874 3881 }
3875 3882
3876 3883 if (inlining_depth == 0) {
3877 3884 #ifndef PRODUCT
3878 3885 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3879 3886 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3880 3887 tty->print_cr(" JVM state at this point:");
3881 3888 for (int i = _depth; i >= 1; i--) {
3882 3889 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3883 3890 }
3884 3891 }
3885 3892 #endif
3886 3893 return false; // Reached end of inlining
3887 3894 }
3888 3895
3889 3896 // Acquire method holder as java.lang.Class
3890 3897 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3891 3898 ciInstance* caller_mirror = caller_klass->java_mirror();
3892 3899 // Push this as a constant
3893 3900 push(makecon(TypeInstPtr::make(caller_mirror)));
3894 3901 #ifndef PRODUCT
3895 3902 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3896 3903 tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth);
3897 3904 tty->print_cr(" JVM state at this point:");
3898 3905 for (int i = _depth; i >= 1; i--) {
3899 3906 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3900 3907 }
3901 3908 }
3902 3909 #endif
3903 3910 return true;
3904 3911 }
3905 3912
3906 3913 // Helper routine for above
3907 3914 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3908 3915 ciMethod* method = jvms->method();
3909 3916
3910 3917 // Is this the Method.invoke method itself?
3911 3918 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3912 3919 return true;
3913 3920
3914 3921 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3915 3922 ciKlass* k = method->holder();
3916 3923 if (k->is_instance_klass()) {
3917 3924 ciInstanceKlass* ik = k->as_instance_klass();
3918 3925 for (; ik != NULL; ik = ik->super()) {
3919 3926 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3920 3927 ik == env()->find_system_klass(ik->name())) {
3921 3928 return true;
3922 3929 }
3923 3930 }
3924 3931 }
3925 3932 else if (method->is_method_handle_adapter()) {
3926 3933 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3927 3934 return true;
3928 3935 }
3929 3936
3930 3937 return false;
3931 3938 }
3932 3939
3933 3940 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3934 3941 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3935 3942 // computing it since there is no lookup field by name function in the
3936 3943 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3937 3944 // Using a static variable here is safe even if we have multiple compilation
3938 3945 // threads because the offset is constant. At worst the same offset will be
3939 3946 // computed and stored multiple
3940 3947
3941 3948 bool LibraryCallKit::inline_native_AtomicLong_get() {
3942 3949 // Restore the stack and pop off the argument
3943 3950 _sp+=1;
3944 3951 Node *obj = pop();
3945 3952
3946 3953 // get the offset of the "value" field. Since the CI interfaces
3947 3954 // does not provide a way to look up a field by name, we scan the bytecodes
3948 3955 // to get the field index. We expect the first 2 instructions of the method
3949 3956 // to be:
3950 3957 // 0 aload_0
3951 3958 // 1 getfield "value"
3952 3959 ciMethod* method = callee();
3953 3960 if (value_field_offset == -1)
3954 3961 {
3955 3962 ciField* value_field;
3956 3963 ciBytecodeStream iter(method);
3957 3964 Bytecodes::Code bc = iter.next();
3958 3965
3959 3966 if ((bc != Bytecodes::_aload_0) &&
3960 3967 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3961 3968 return false;
3962 3969 bc = iter.next();
3963 3970 if (bc != Bytecodes::_getfield)
3964 3971 return false;
3965 3972 bool ignore;
3966 3973 value_field = iter.get_field(ignore);
3967 3974 value_field_offset = value_field->offset_in_bytes();
3968 3975 }
3969 3976
3970 3977 // Null check without removing any arguments.
3971 3978 _sp++;
3972 3979 obj = do_null_check(obj, T_OBJECT);
3973 3980 _sp--;
3974 3981 // Check for locking null object
3975 3982 if (stopped()) return true;
3976 3983
3977 3984 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3978 3985 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3979 3986 int alias_idx = C->get_alias_index(adr_type);
3980 3987
3981 3988 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3982 3989
3983 3990 push_pair(result);
3984 3991
3985 3992 return true;
3986 3993 }
3987 3994
3988 3995 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3989 3996 // Restore the stack and pop off the arguments
3990 3997 _sp+=5;
3991 3998 Node *newVal = pop_pair();
3992 3999 Node *oldVal = pop_pair();
3993 4000 Node *obj = pop();
3994 4001
3995 4002 // we need the offset of the "value" field which was computed when
3996 4003 // inlining the get() method. Give up if we don't have it.
3997 4004 if (value_field_offset == -1)
3998 4005 return false;
3999 4006
4000 4007 // Null check without removing any arguments.
4001 4008 _sp+=5;
4002 4009 obj = do_null_check(obj, T_OBJECT);
4003 4010 _sp-=5;
4004 4011 // Check for locking null object
4005 4012 if (stopped()) return true;
4006 4013
4007 4014 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
4008 4015 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
4009 4016 int alias_idx = C->get_alias_index(adr_type);
4010 4017
4011 4018 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
4012 4019 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
4013 4020 set_memory(store_proj, alias_idx);
4014 4021 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
4015 4022
4016 4023 Node *result;
4017 4024 // CMove node is not used to be able fold a possible check code
4018 4025 // after attemptUpdate() call. This code could be transformed
4019 4026 // into CMove node by loop optimizations.
4020 4027 {
4021 4028 RegionNode *r = new (C, 3) RegionNode(3);
4022 4029 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
4023 4030
4024 4031 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4025 4032 Node *iftrue = opt_iff(r, iff);
4026 4033 r->init_req(1, iftrue);
4027 4034 result->init_req(1, intcon(1));
4028 4035 result->init_req(2, intcon(0));
4029 4036
4030 4037 set_control(_gvn.transform(r));
4031 4038 record_for_igvn(r);
4032 4039
4033 4040 C->set_has_split_ifs(true); // Has chance for split-if optimization
4034 4041 }
4035 4042
4036 4043 push(_gvn.transform(result));
4037 4044 return true;
4038 4045 }
4039 4046
4040 4047 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4041 4048 // restore the arguments
4042 4049 _sp += arg_size();
4043 4050
4044 4051 switch (id) {
4045 4052 case vmIntrinsics::_floatToRawIntBits:
4046 4053 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
4047 4054 break;
4048 4055
4049 4056 case vmIntrinsics::_intBitsToFloat:
4050 4057 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
4051 4058 break;
4052 4059
4053 4060 case vmIntrinsics::_doubleToRawLongBits:
4054 4061 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
4055 4062 break;
4056 4063
4057 4064 case vmIntrinsics::_longBitsToDouble:
4058 4065 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
4059 4066 break;
4060 4067
4061 4068 case vmIntrinsics::_doubleToLongBits: {
4062 4069 Node* value = pop_pair();
4063 4070
4064 4071 // two paths (plus control) merge in a wood
4065 4072 RegionNode *r = new (C, 3) RegionNode(3);
4066 4073 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
4067 4074
4068 4075 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
4069 4076 // Build the boolean node
4070 4077 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4071 4078
4072 4079 // Branch either way.
4073 4080 // NaN case is less traveled, which makes all the difference.
4074 4081 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4075 4082 Node *opt_isnan = _gvn.transform(ifisnan);
4076 4083 assert( opt_isnan->is_If(), "Expect an IfNode");
4077 4084 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4078 4085 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4079 4086
4080 4087 set_control(iftrue);
4081 4088
4082 4089 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4083 4090 Node *slow_result = longcon(nan_bits); // return NaN
4084 4091 phi->init_req(1, _gvn.transform( slow_result ));
4085 4092 r->init_req(1, iftrue);
4086 4093
4087 4094 // Else fall through
4088 4095 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4089 4096 set_control(iffalse);
4090 4097
4091 4098 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
4092 4099 r->init_req(2, iffalse);
4093 4100
4094 4101 // Post merge
4095 4102 set_control(_gvn.transform(r));
4096 4103 record_for_igvn(r);
4097 4104
4098 4105 Node* result = _gvn.transform(phi);
4099 4106 assert(result->bottom_type()->isa_long(), "must be");
4100 4107 push_pair(result);
4101 4108
4102 4109 C->set_has_split_ifs(true); // Has chance for split-if optimization
4103 4110
4104 4111 break;
4105 4112 }
4106 4113
4107 4114 case vmIntrinsics::_floatToIntBits: {
4108 4115 Node* value = pop();
4109 4116
4110 4117 // two paths (plus control) merge in a wood
4111 4118 RegionNode *r = new (C, 3) RegionNode(3);
4112 4119 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
4113 4120
4114 4121 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
4115 4122 // Build the boolean node
4116 4123 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4117 4124
4118 4125 // Branch either way.
4119 4126 // NaN case is less traveled, which makes all the difference.
4120 4127 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4121 4128 Node *opt_isnan = _gvn.transform(ifisnan);
4122 4129 assert( opt_isnan->is_If(), "Expect an IfNode");
4123 4130 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4124 4131 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4125 4132
4126 4133 set_control(iftrue);
4127 4134
4128 4135 static const jint nan_bits = 0x7fc00000;
4129 4136 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4130 4137 phi->init_req(1, _gvn.transform( slow_result ));
4131 4138 r->init_req(1, iftrue);
4132 4139
4133 4140 // Else fall through
4134 4141 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4135 4142 set_control(iffalse);
4136 4143
4137 4144 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
4138 4145 r->init_req(2, iffalse);
4139 4146
4140 4147 // Post merge
4141 4148 set_control(_gvn.transform(r));
4142 4149 record_for_igvn(r);
4143 4150
4144 4151 Node* result = _gvn.transform(phi);
4145 4152 assert(result->bottom_type()->isa_int(), "must be");
4146 4153 push(result);
4147 4154
4148 4155 C->set_has_split_ifs(true); // Has chance for split-if optimization
4149 4156
4150 4157 break;
4151 4158 }
4152 4159
4153 4160 default:
4154 4161 ShouldNotReachHere();
4155 4162 }
4156 4163
4157 4164 return true;
4158 4165 }
4159 4166
4160 4167 #ifdef _LP64
4161 4168 #define XTOP ,top() /*additional argument*/
4162 4169 #else //_LP64
4163 4170 #define XTOP /*no additional argument*/
4164 4171 #endif //_LP64
4165 4172
4166 4173 //----------------------inline_unsafe_copyMemory-------------------------
4167 4174 bool LibraryCallKit::inline_unsafe_copyMemory() {
4168 4175 if (callee()->is_static()) return false; // caller must have the capability!
4169 4176 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4170 4177 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4171 4178 null_check_receiver(callee()); // check then ignore argument(0)
4172 4179 if (stopped()) return true;
4173 4180
4174 4181 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4175 4182
4176 4183 Node* src_ptr = argument(1);
4177 4184 Node* src_off = ConvL2X(argument(2));
4178 4185 assert(argument(3)->is_top(), "2nd half of long");
4179 4186 Node* dst_ptr = argument(4);
4180 4187 Node* dst_off = ConvL2X(argument(5));
4181 4188 assert(argument(6)->is_top(), "2nd half of long");
4182 4189 Node* size = ConvL2X(argument(7));
4183 4190 assert(argument(8)->is_top(), "2nd half of long");
4184 4191
4185 4192 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4186 4193 "fieldOffset must be byte-scaled");
4187 4194
4188 4195 Node* src = make_unsafe_address(src_ptr, src_off);
4189 4196 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4190 4197
4191 4198 // Conservatively insert a memory barrier on all memory slices.
4192 4199 // Do not let writes of the copy source or destination float below the copy.
4193 4200 insert_mem_bar(Op_MemBarCPUOrder);
4194 4201
4195 4202 // Call it. Note that the length argument is not scaled.
4196 4203 make_runtime_call(RC_LEAF|RC_NO_FP,
4197 4204 OptoRuntime::fast_arraycopy_Type(),
4198 4205 StubRoutines::unsafe_arraycopy(),
4199 4206 "unsafe_arraycopy",
4200 4207 TypeRawPtr::BOTTOM,
4201 4208 src, dst, size XTOP);
4202 4209
4203 4210 // Do not let reads of the copy destination float above the copy.
4204 4211 insert_mem_bar(Op_MemBarCPUOrder);
4205 4212
4206 4213 return true;
4207 4214 }
4208 4215
4209 4216 //------------------------clone_coping-----------------------------------
4210 4217 // Helper function for inline_native_clone.
4211 4218 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4212 4219 assert(obj_size != NULL, "");
4213 4220 Node* raw_obj = alloc_obj->in(1);
4214 4221 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4215 4222
4216 4223 if (ReduceBulkZeroing) {
4217 4224 // We will be completely responsible for initializing this object -
4218 4225 // mark Initialize node as complete.
4219 4226 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4220 4227 // The object was just allocated - there should be no any stores!
4221 4228 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4222 4229 }
4223 4230
4224 4231 // Copy the fastest available way.
4225 4232 // TODO: generate fields copies for small objects instead.
4226 4233 Node* src = obj;
4227 4234 Node* dest = alloc_obj;
4228 4235 Node* size = _gvn.transform(obj_size);
4229 4236
4230 4237 // Exclude the header but include array length to copy by 8 bytes words.
4231 4238 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4232 4239 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4233 4240 instanceOopDesc::base_offset_in_bytes();
4234 4241 // base_off:
4235 4242 // 8 - 32-bit VM
4236 4243 // 12 - 64-bit VM, compressed oops
4237 4244 // 16 - 64-bit VM, normal oops
4238 4245 if (base_off % BytesPerLong != 0) {
4239 4246 assert(UseCompressedOops, "");
4240 4247 if (is_array) {
4241 4248 // Exclude length to copy by 8 bytes words.
4242 4249 base_off += sizeof(int);
4243 4250 } else {
4244 4251 // Include klass to copy by 8 bytes words.
4245 4252 base_off = instanceOopDesc::klass_offset_in_bytes();
4246 4253 }
4247 4254 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4248 4255 }
4249 4256 src = basic_plus_adr(src, base_off);
4250 4257 dest = basic_plus_adr(dest, base_off);
4251 4258
4252 4259 // Compute the length also, if needed:
4253 4260 Node* countx = size;
4254 4261 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4255 4262 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4256 4263
4257 4264 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4258 4265 bool disjoint_bases = true;
4259 4266 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4260 4267 src, NULL, dest, NULL, countx,
4261 4268 /*dest_uninitialized*/true);
4262 4269
4263 4270 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4264 4271 if (card_mark) {
4265 4272 assert(!is_array, "");
4266 4273 // Put in store barrier for any and all oops we are sticking
4267 4274 // into this object. (We could avoid this if we could prove
4268 4275 // that the object type contains no oop fields at all.)
4269 4276 Node* no_particular_value = NULL;
4270 4277 Node* no_particular_field = NULL;
4271 4278 int raw_adr_idx = Compile::AliasIdxRaw;
4272 4279 post_barrier(control(),
4273 4280 memory(raw_adr_type),
4274 4281 alloc_obj,
4275 4282 no_particular_field,
4276 4283 raw_adr_idx,
4277 4284 no_particular_value,
4278 4285 T_OBJECT,
4279 4286 false);
4280 4287 }
4281 4288
4282 4289 // Do not let reads from the cloned object float above the arraycopy.
4283 4290 insert_mem_bar(Op_MemBarCPUOrder);
4284 4291 }
4285 4292
4286 4293 //------------------------inline_native_clone----------------------------
4287 4294 // Here are the simple edge cases:
4288 4295 // null receiver => normal trap
4289 4296 // virtual and clone was overridden => slow path to out-of-line clone
4290 4297 // not cloneable or finalizer => slow path to out-of-line Object.clone
4291 4298 //
4292 4299 // The general case has two steps, allocation and copying.
4293 4300 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4294 4301 //
4295 4302 // Copying also has two cases, oop arrays and everything else.
4296 4303 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4297 4304 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4298 4305 //
4299 4306 // These steps fold up nicely if and when the cloned object's klass
4300 4307 // can be sharply typed as an object array, a type array, or an instance.
4301 4308 //
4302 4309 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4303 4310 int nargs = 1;
4304 4311 PhiNode* result_val;
4305 4312
4306 4313 //set the original stack and the reexecute bit for the interpreter to reexecute
4307 4314 //the bytecode that invokes Object.clone if deoptimization happens
4308 4315 { PreserveReexecuteState preexecs(this);
4309 4316 jvms()->set_should_reexecute(true);
4310 4317
4311 4318 //null_check_receiver will adjust _sp (push and pop)
4312 4319 Node* obj = null_check_receiver(callee());
4313 4320 if (stopped()) return true;
4314 4321
4315 4322 _sp += nargs;
4316 4323
4317 4324 Node* obj_klass = load_object_klass(obj);
4318 4325 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4319 4326 const TypeOopPtr* toop = ((tklass != NULL)
4320 4327 ? tklass->as_instance_type()
4321 4328 : TypeInstPtr::NOTNULL);
4322 4329
4323 4330 // Conservatively insert a memory barrier on all memory slices.
4324 4331 // Do not let writes into the original float below the clone.
4325 4332 insert_mem_bar(Op_MemBarCPUOrder);
4326 4333
4327 4334 // paths into result_reg:
4328 4335 enum {
4329 4336 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4330 4337 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4331 4338 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4332 4339 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4333 4340 PATH_LIMIT
4334 4341 };
4335 4342 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4336 4343 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4337 4344 TypeInstPtr::NOTNULL);
4338 4345 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4339 4346 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4340 4347 TypePtr::BOTTOM);
4341 4348 record_for_igvn(result_reg);
4342 4349
4343 4350 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4344 4351 int raw_adr_idx = Compile::AliasIdxRaw;
4345 4352 const bool raw_mem_only = true;
4346 4353
4347 4354
4348 4355 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4349 4356 if (array_ctl != NULL) {
4350 4357 // It's an array.
4351 4358 PreserveJVMState pjvms(this);
4352 4359 set_control(array_ctl);
4353 4360 Node* obj_length = load_array_length(obj);
4354 4361 Node* obj_size = NULL;
4355 4362 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4356 4363 raw_mem_only, &obj_size);
4357 4364
4358 4365 if (!use_ReduceInitialCardMarks()) {
4359 4366 // If it is an oop array, it requires very special treatment,
4360 4367 // because card marking is required on each card of the array.
4361 4368 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4362 4369 if (is_obja != NULL) {
4363 4370 PreserveJVMState pjvms2(this);
4364 4371 set_control(is_obja);
4365 4372 // Generate a direct call to the right arraycopy function(s).
4366 4373 bool disjoint_bases = true;
4367 4374 bool length_never_negative = true;
4368 4375 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4369 4376 obj, intcon(0), alloc_obj, intcon(0),
4370 4377 obj_length,
4371 4378 disjoint_bases, length_never_negative);
4372 4379 result_reg->init_req(_objArray_path, control());
4373 4380 result_val->init_req(_objArray_path, alloc_obj);
4374 4381 result_i_o ->set_req(_objArray_path, i_o());
4375 4382 result_mem ->set_req(_objArray_path, reset_memory());
4376 4383 }
4377 4384 }
4378 4385 // Otherwise, there are no card marks to worry about.
4379 4386 // (We can dispense with card marks if we know the allocation
4380 4387 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4381 4388 // causes the non-eden paths to take compensating steps to
4382 4389 // simulate a fresh allocation, so that no further
4383 4390 // card marks are required in compiled code to initialize
4384 4391 // the object.)
4385 4392
4386 4393 if (!stopped()) {
4387 4394 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4388 4395
4389 4396 // Present the results of the copy.
4390 4397 result_reg->init_req(_array_path, control());
4391 4398 result_val->init_req(_array_path, alloc_obj);
4392 4399 result_i_o ->set_req(_array_path, i_o());
4393 4400 result_mem ->set_req(_array_path, reset_memory());
4394 4401 }
4395 4402 }
4396 4403
4397 4404 // We only go to the instance fast case code if we pass a number of guards.
4398 4405 // The paths which do not pass are accumulated in the slow_region.
4399 4406 RegionNode* slow_region = new (C, 1) RegionNode(1);
4400 4407 record_for_igvn(slow_region);
4401 4408 if (!stopped()) {
4402 4409 // It's an instance (we did array above). Make the slow-path tests.
4403 4410 // If this is a virtual call, we generate a funny guard. We grab
4404 4411 // the vtable entry corresponding to clone() from the target object.
4405 4412 // If the target method which we are calling happens to be the
4406 4413 // Object clone() method, we pass the guard. We do not need this
4407 4414 // guard for non-virtual calls; the caller is known to be the native
4408 4415 // Object clone().
4409 4416 if (is_virtual) {
4410 4417 generate_virtual_guard(obj_klass, slow_region);
4411 4418 }
4412 4419
4413 4420 // The object must be cloneable and must not have a finalizer.
4414 4421 // Both of these conditions may be checked in a single test.
4415 4422 // We could optimize the cloneable test further, but we don't care.
4416 4423 generate_access_flags_guard(obj_klass,
4417 4424 // Test both conditions:
4418 4425 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4419 4426 // Must be cloneable but not finalizer:
4420 4427 JVM_ACC_IS_CLONEABLE,
4421 4428 slow_region);
4422 4429 }
4423 4430
4424 4431 if (!stopped()) {
4425 4432 // It's an instance, and it passed the slow-path tests.
4426 4433 PreserveJVMState pjvms(this);
4427 4434 Node* obj_size = NULL;
4428 4435 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4429 4436
4430 4437 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4431 4438
4432 4439 // Present the results of the slow call.
4433 4440 result_reg->init_req(_instance_path, control());
4434 4441 result_val->init_req(_instance_path, alloc_obj);
4435 4442 result_i_o ->set_req(_instance_path, i_o());
4436 4443 result_mem ->set_req(_instance_path, reset_memory());
4437 4444 }
4438 4445
4439 4446 // Generate code for the slow case. We make a call to clone().
4440 4447 set_control(_gvn.transform(slow_region));
4441 4448 if (!stopped()) {
4442 4449 PreserveJVMState pjvms(this);
4443 4450 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4444 4451 Node* slow_result = set_results_for_java_call(slow_call);
4445 4452 // this->control() comes from set_results_for_java_call
4446 4453 result_reg->init_req(_slow_path, control());
4447 4454 result_val->init_req(_slow_path, slow_result);
4448 4455 result_i_o ->set_req(_slow_path, i_o());
4449 4456 result_mem ->set_req(_slow_path, reset_memory());
4450 4457 }
4451 4458
4452 4459 // Return the combined state.
4453 4460 set_control( _gvn.transform(result_reg) );
4454 4461 set_i_o( _gvn.transform(result_i_o) );
4455 4462 set_all_memory( _gvn.transform(result_mem) );
4456 4463 } //original reexecute and sp are set back here
4457 4464
4458 4465 push(_gvn.transform(result_val));
4459 4466
4460 4467 return true;
4461 4468 }
4462 4469
4463 4470
4464 4471 // constants for computing the copy function
4465 4472 enum {
4466 4473 COPYFUNC_UNALIGNED = 0,
4467 4474 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4468 4475 COPYFUNC_CONJOINT = 0,
4469 4476 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4470 4477 };
4471 4478
4472 4479 // Note: The condition "disjoint" applies also for overlapping copies
4473 4480 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4474 4481 static address
4475 4482 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name, bool dest_uninitialized) {
4476 4483 int selector =
4477 4484 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4478 4485 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4479 4486
4480 4487 #define RETURN_STUB(xxx_arraycopy) { \
4481 4488 name = #xxx_arraycopy; \
4482 4489 return StubRoutines::xxx_arraycopy(); }
4483 4490
4484 4491 #define RETURN_STUB_PARM(xxx_arraycopy, parm) { \
4485 4492 name = #xxx_arraycopy; \
4486 4493 return StubRoutines::xxx_arraycopy(parm); }
4487 4494
4488 4495 switch (t) {
4489 4496 case T_BYTE:
4490 4497 case T_BOOLEAN:
4491 4498 switch (selector) {
4492 4499 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4493 4500 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4494 4501 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4495 4502 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4496 4503 }
4497 4504 case T_CHAR:
4498 4505 case T_SHORT:
4499 4506 switch (selector) {
4500 4507 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4501 4508 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4502 4509 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4503 4510 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4504 4511 }
4505 4512 case T_INT:
4506 4513 case T_FLOAT:
4507 4514 switch (selector) {
4508 4515 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4509 4516 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4510 4517 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4511 4518 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4512 4519 }
4513 4520 case T_DOUBLE:
4514 4521 case T_LONG:
4515 4522 switch (selector) {
4516 4523 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4517 4524 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4518 4525 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4519 4526 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4520 4527 }
4521 4528 case T_ARRAY:
4522 4529 case T_OBJECT:
4523 4530 switch (selector) {
4524 4531 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_arraycopy, dest_uninitialized);
4525 4532 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_arraycopy, dest_uninitialized);
4526 4533 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_disjoint_arraycopy, dest_uninitialized);
4527 4534 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_disjoint_arraycopy, dest_uninitialized);
4528 4535 }
4529 4536 default:
4530 4537 ShouldNotReachHere();
4531 4538 return NULL;
4532 4539 }
4533 4540
4534 4541 #undef RETURN_STUB
4535 4542 #undef RETURN_STUB_PARM
4536 4543 }
4537 4544
4538 4545 //------------------------------basictype2arraycopy----------------------------
4539 4546 address LibraryCallKit::basictype2arraycopy(BasicType t,
4540 4547 Node* src_offset,
4541 4548 Node* dest_offset,
4542 4549 bool disjoint_bases,
4543 4550 const char* &name,
4544 4551 bool dest_uninitialized) {
4545 4552 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4546 4553 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4547 4554
4548 4555 bool aligned = false;
4549 4556 bool disjoint = disjoint_bases;
4550 4557
4551 4558 // if the offsets are the same, we can treat the memory regions as
4552 4559 // disjoint, because either the memory regions are in different arrays,
4553 4560 // or they are identical (which we can treat as disjoint.) We can also
4554 4561 // treat a copy with a destination index less that the source index
4555 4562 // as disjoint since a low->high copy will work correctly in this case.
4556 4563 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4557 4564 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4558 4565 // both indices are constants
4559 4566 int s_offs = src_offset_inttype->get_con();
4560 4567 int d_offs = dest_offset_inttype->get_con();
4561 4568 int element_size = type2aelembytes(t);
4562 4569 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4563 4570 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4564 4571 if (s_offs >= d_offs) disjoint = true;
4565 4572 } else if (src_offset == dest_offset && src_offset != NULL) {
4566 4573 // This can occur if the offsets are identical non-constants.
4567 4574 disjoint = true;
4568 4575 }
4569 4576
4570 4577 return select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4571 4578 }
4572 4579
4573 4580
4574 4581 //------------------------------inline_arraycopy-----------------------
4575 4582 bool LibraryCallKit::inline_arraycopy() {
4576 4583 // Restore the stack and pop off the arguments.
4577 4584 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4578 4585 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4579 4586
4580 4587 Node *src = argument(0);
4581 4588 Node *src_offset = argument(1);
4582 4589 Node *dest = argument(2);
4583 4590 Node *dest_offset = argument(3);
4584 4591 Node *length = argument(4);
4585 4592
4586 4593 // Compile time checks. If any of these checks cannot be verified at compile time,
4587 4594 // we do not make a fast path for this call. Instead, we let the call remain as it
4588 4595 // is. The checks we choose to mandate at compile time are:
4589 4596 //
4590 4597 // (1) src and dest are arrays.
4591 4598 const Type* src_type = src->Value(&_gvn);
4592 4599 const Type* dest_type = dest->Value(&_gvn);
4593 4600 const TypeAryPtr* top_src = src_type->isa_aryptr();
4594 4601 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4595 4602 if (top_src == NULL || top_src->klass() == NULL ||
4596 4603 top_dest == NULL || top_dest->klass() == NULL) {
4597 4604 // Conservatively insert a memory barrier on all memory slices.
4598 4605 // Do not let writes into the source float below the arraycopy.
4599 4606 insert_mem_bar(Op_MemBarCPUOrder);
4600 4607
4601 4608 // Call StubRoutines::generic_arraycopy stub.
4602 4609 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4603 4610 src, src_offset, dest, dest_offset, length);
4604 4611
4605 4612 // Do not let reads from the destination float above the arraycopy.
4606 4613 // Since we cannot type the arrays, we don't know which slices
4607 4614 // might be affected. We could restrict this barrier only to those
4608 4615 // memory slices which pertain to array elements--but don't bother.
4609 4616 if (!InsertMemBarAfterArraycopy)
4610 4617 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4611 4618 insert_mem_bar(Op_MemBarCPUOrder);
4612 4619 return true;
4613 4620 }
4614 4621
4615 4622 // (2) src and dest arrays must have elements of the same BasicType
4616 4623 // Figure out the size and type of the elements we will be copying.
4617 4624 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4618 4625 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4619 4626 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4620 4627 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4621 4628
4622 4629 if (src_elem != dest_elem || dest_elem == T_VOID) {
4623 4630 // The component types are not the same or are not recognized. Punt.
4624 4631 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4625 4632 generate_slow_arraycopy(TypePtr::BOTTOM,
4626 4633 src, src_offset, dest, dest_offset, length,
4627 4634 /*dest_uninitialized*/false);
4628 4635 return true;
4629 4636 }
4630 4637
4631 4638 //---------------------------------------------------------------------------
4632 4639 // We will make a fast path for this call to arraycopy.
4633 4640
4634 4641 // We have the following tests left to perform:
4635 4642 //
4636 4643 // (3) src and dest must not be null.
4637 4644 // (4) src_offset must not be negative.
4638 4645 // (5) dest_offset must not be negative.
4639 4646 // (6) length must not be negative.
4640 4647 // (7) src_offset + length must not exceed length of src.
4641 4648 // (8) dest_offset + length must not exceed length of dest.
4642 4649 // (9) each element of an oop array must be assignable
4643 4650
4644 4651 RegionNode* slow_region = new (C, 1) RegionNode(1);
4645 4652 record_for_igvn(slow_region);
4646 4653
4647 4654 // (3) operands must not be null
4648 4655 // We currently perform our null checks with the do_null_check routine.
4649 4656 // This means that the null exceptions will be reported in the caller
4650 4657 // rather than (correctly) reported inside of the native arraycopy call.
4651 4658 // This should be corrected, given time. We do our null check with the
4652 4659 // stack pointer restored.
4653 4660 _sp += nargs;
4654 4661 src = do_null_check(src, T_ARRAY);
4655 4662 dest = do_null_check(dest, T_ARRAY);
4656 4663 _sp -= nargs;
4657 4664
4658 4665 // (4) src_offset must not be negative.
4659 4666 generate_negative_guard(src_offset, slow_region);
4660 4667
4661 4668 // (5) dest_offset must not be negative.
4662 4669 generate_negative_guard(dest_offset, slow_region);
4663 4670
4664 4671 // (6) length must not be negative (moved to generate_arraycopy()).
4665 4672 // generate_negative_guard(length, slow_region);
4666 4673
4667 4674 // (7) src_offset + length must not exceed length of src.
4668 4675 generate_limit_guard(src_offset, length,
4669 4676 load_array_length(src),
4670 4677 slow_region);
4671 4678
4672 4679 // (8) dest_offset + length must not exceed length of dest.
4673 4680 generate_limit_guard(dest_offset, length,
4674 4681 load_array_length(dest),
4675 4682 slow_region);
4676 4683
4677 4684 // (9) each element of an oop array must be assignable
4678 4685 // The generate_arraycopy subroutine checks this.
4679 4686
4680 4687 // This is where the memory effects are placed:
4681 4688 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4682 4689 generate_arraycopy(adr_type, dest_elem,
4683 4690 src, src_offset, dest, dest_offset, length,
4684 4691 false, false, slow_region);
4685 4692
4686 4693 return true;
4687 4694 }
4688 4695
4689 4696 //-----------------------------generate_arraycopy----------------------
4690 4697 // Generate an optimized call to arraycopy.
4691 4698 // Caller must guard against non-arrays.
4692 4699 // Caller must determine a common array basic-type for both arrays.
4693 4700 // Caller must validate offsets against array bounds.
4694 4701 // The slow_region has already collected guard failure paths
4695 4702 // (such as out of bounds length or non-conformable array types).
4696 4703 // The generated code has this shape, in general:
4697 4704 //
4698 4705 // if (length == 0) return // via zero_path
4699 4706 // slowval = -1
4700 4707 // if (types unknown) {
4701 4708 // slowval = call generic copy loop
4702 4709 // if (slowval == 0) return // via checked_path
4703 4710 // } else if (indexes in bounds) {
4704 4711 // if ((is object array) && !(array type check)) {
4705 4712 // slowval = call checked copy loop
4706 4713 // if (slowval == 0) return // via checked_path
4707 4714 // } else {
4708 4715 // call bulk copy loop
4709 4716 // return // via fast_path
4710 4717 // }
4711 4718 // }
4712 4719 // // adjust params for remaining work:
4713 4720 // if (slowval != -1) {
4714 4721 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4715 4722 // }
4716 4723 // slow_region:
4717 4724 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4718 4725 // return // via slow_call_path
4719 4726 //
4720 4727 // This routine is used from several intrinsics: System.arraycopy,
4721 4728 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4722 4729 //
4723 4730 void
4724 4731 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4725 4732 BasicType basic_elem_type,
4726 4733 Node* src, Node* src_offset,
4727 4734 Node* dest, Node* dest_offset,
4728 4735 Node* copy_length,
4729 4736 bool disjoint_bases,
4730 4737 bool length_never_negative,
4731 4738 RegionNode* slow_region) {
4732 4739
4733 4740 if (slow_region == NULL) {
4734 4741 slow_region = new(C,1) RegionNode(1);
4735 4742 record_for_igvn(slow_region);
4736 4743 }
4737 4744
4738 4745 Node* original_dest = dest;
4739 4746 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4740 4747 bool dest_uninitialized = false;
4741 4748
4742 4749 // See if this is the initialization of a newly-allocated array.
4743 4750 // If so, we will take responsibility here for initializing it to zero.
4744 4751 // (Note: Because tightly_coupled_allocation performs checks on the
4745 4752 // out-edges of the dest, we need to avoid making derived pointers
4746 4753 // from it until we have checked its uses.)
4747 4754 if (ReduceBulkZeroing
4748 4755 && !ZeroTLAB // pointless if already zeroed
4749 4756 && basic_elem_type != T_CONFLICT // avoid corner case
4750 4757 && !_gvn.eqv_uncast(src, dest)
4751 4758 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4752 4759 != NULL)
4753 4760 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4754 4761 && alloc->maybe_set_complete(&_gvn)) {
4755 4762 // "You break it, you buy it."
4756 4763 InitializeNode* init = alloc->initialization();
4757 4764 assert(init->is_complete(), "we just did this");
4758 4765 assert(dest->is_CheckCastPP(), "sanity");
4759 4766 assert(dest->in(0)->in(0) == init, "dest pinned");
4760 4767 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4761 4768 // From this point on, every exit path is responsible for
4762 4769 // initializing any non-copied parts of the object to zero.
4763 4770 // Also, if this flag is set we make sure that arraycopy interacts properly
4764 4771 // with G1, eliding pre-barriers. See CR 6627983.
4765 4772 dest_uninitialized = true;
4766 4773 } else {
4767 4774 // No zeroing elimination here.
4768 4775 alloc = NULL;
4769 4776 //original_dest = dest;
4770 4777 //dest_uninitialized = false;
4771 4778 }
4772 4779
4773 4780 // Results are placed here:
4774 4781 enum { fast_path = 1, // normal void-returning assembly stub
4775 4782 checked_path = 2, // special assembly stub with cleanup
4776 4783 slow_call_path = 3, // something went wrong; call the VM
4777 4784 zero_path = 4, // bypass when length of copy is zero
4778 4785 bcopy_path = 5, // copy primitive array by 64-bit blocks
4779 4786 PATH_LIMIT = 6
4780 4787 };
4781 4788 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4782 4789 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4783 4790 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4784 4791 record_for_igvn(result_region);
4785 4792 _gvn.set_type_bottom(result_i_o);
4786 4793 _gvn.set_type_bottom(result_memory);
4787 4794 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4788 4795
4789 4796 // The slow_control path:
4790 4797 Node* slow_control;
4791 4798 Node* slow_i_o = i_o();
4792 4799 Node* slow_mem = memory(adr_type);
4793 4800 debug_only(slow_control = (Node*) badAddress);
4794 4801
4795 4802 // Checked control path:
4796 4803 Node* checked_control = top();
4797 4804 Node* checked_mem = NULL;
4798 4805 Node* checked_i_o = NULL;
4799 4806 Node* checked_value = NULL;
4800 4807
4801 4808 if (basic_elem_type == T_CONFLICT) {
4802 4809 assert(!dest_uninitialized, "");
4803 4810 Node* cv = generate_generic_arraycopy(adr_type,
4804 4811 src, src_offset, dest, dest_offset,
4805 4812 copy_length, dest_uninitialized);
4806 4813 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4807 4814 checked_control = control();
4808 4815 checked_i_o = i_o();
4809 4816 checked_mem = memory(adr_type);
4810 4817 checked_value = cv;
4811 4818 set_control(top()); // no fast path
4812 4819 }
4813 4820
4814 4821 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4815 4822 if (not_pos != NULL) {
4816 4823 PreserveJVMState pjvms(this);
4817 4824 set_control(not_pos);
4818 4825
4819 4826 // (6) length must not be negative.
4820 4827 if (!length_never_negative) {
4821 4828 generate_negative_guard(copy_length, slow_region);
4822 4829 }
4823 4830
4824 4831 // copy_length is 0.
4825 4832 if (!stopped() && dest_uninitialized) {
4826 4833 Node* dest_length = alloc->in(AllocateNode::ALength);
4827 4834 if (_gvn.eqv_uncast(copy_length, dest_length)
4828 4835 || _gvn.find_int_con(dest_length, 1) <= 0) {
4829 4836 // There is no zeroing to do. No need for a secondary raw memory barrier.
4830 4837 } else {
4831 4838 // Clear the whole thing since there are no source elements to copy.
4832 4839 generate_clear_array(adr_type, dest, basic_elem_type,
4833 4840 intcon(0), NULL,
4834 4841 alloc->in(AllocateNode::AllocSize));
4835 4842 // Use a secondary InitializeNode as raw memory barrier.
4836 4843 // Currently it is needed only on this path since other
4837 4844 // paths have stub or runtime calls as raw memory barriers.
4838 4845 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4839 4846 Compile::AliasIdxRaw,
4840 4847 top())->as_Initialize();
4841 4848 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4842 4849 }
4843 4850 }
4844 4851
4845 4852 // Present the results of the fast call.
4846 4853 result_region->init_req(zero_path, control());
4847 4854 result_i_o ->init_req(zero_path, i_o());
4848 4855 result_memory->init_req(zero_path, memory(adr_type));
4849 4856 }
4850 4857
4851 4858 if (!stopped() && dest_uninitialized) {
4852 4859 // We have to initialize the *uncopied* part of the array to zero.
4853 4860 // The copy destination is the slice dest[off..off+len]. The other slices
4854 4861 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4855 4862 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4856 4863 Node* dest_length = alloc->in(AllocateNode::ALength);
4857 4864 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4858 4865 copy_length) );
4859 4866
4860 4867 // If there is a head section that needs zeroing, do it now.
4861 4868 if (find_int_con(dest_offset, -1) != 0) {
4862 4869 generate_clear_array(adr_type, dest, basic_elem_type,
4863 4870 intcon(0), dest_offset,
4864 4871 NULL);
4865 4872 }
4866 4873
4867 4874 // Next, perform a dynamic check on the tail length.
4868 4875 // It is often zero, and we can win big if we prove this.
4869 4876 // There are two wins: Avoid generating the ClearArray
4870 4877 // with its attendant messy index arithmetic, and upgrade
4871 4878 // the copy to a more hardware-friendly word size of 64 bits.
4872 4879 Node* tail_ctl = NULL;
4873 4880 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4874 4881 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4875 4882 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4876 4883 tail_ctl = generate_slow_guard(bol_lt, NULL);
4877 4884 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4878 4885 }
4879 4886
4880 4887 // At this point, let's assume there is no tail.
4881 4888 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4882 4889 // There is no tail. Try an upgrade to a 64-bit copy.
4883 4890 bool didit = false;
4884 4891 { PreserveJVMState pjvms(this);
4885 4892 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4886 4893 src, src_offset, dest, dest_offset,
4887 4894 dest_size, dest_uninitialized);
4888 4895 if (didit) {
4889 4896 // Present the results of the block-copying fast call.
4890 4897 result_region->init_req(bcopy_path, control());
4891 4898 result_i_o ->init_req(bcopy_path, i_o());
4892 4899 result_memory->init_req(bcopy_path, memory(adr_type));
4893 4900 }
4894 4901 }
4895 4902 if (didit)
4896 4903 set_control(top()); // no regular fast path
4897 4904 }
4898 4905
4899 4906 // Clear the tail, if any.
4900 4907 if (tail_ctl != NULL) {
4901 4908 Node* notail_ctl = stopped() ? NULL : control();
4902 4909 set_control(tail_ctl);
4903 4910 if (notail_ctl == NULL) {
4904 4911 generate_clear_array(adr_type, dest, basic_elem_type,
4905 4912 dest_tail, NULL,
4906 4913 dest_size);
4907 4914 } else {
4908 4915 // Make a local merge.
4909 4916 Node* done_ctl = new(C,3) RegionNode(3);
4910 4917 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4911 4918 done_ctl->init_req(1, notail_ctl);
4912 4919 done_mem->init_req(1, memory(adr_type));
4913 4920 generate_clear_array(adr_type, dest, basic_elem_type,
4914 4921 dest_tail, NULL,
4915 4922 dest_size);
4916 4923 done_ctl->init_req(2, control());
4917 4924 done_mem->init_req(2, memory(adr_type));
4918 4925 set_control( _gvn.transform(done_ctl) );
4919 4926 set_memory( _gvn.transform(done_mem), adr_type );
4920 4927 }
4921 4928 }
4922 4929 }
4923 4930
4924 4931 BasicType copy_type = basic_elem_type;
4925 4932 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4926 4933 if (!stopped() && copy_type == T_OBJECT) {
4927 4934 // If src and dest have compatible element types, we can copy bits.
4928 4935 // Types S[] and D[] are compatible if D is a supertype of S.
4929 4936 //
4930 4937 // If they are not, we will use checked_oop_disjoint_arraycopy,
4931 4938 // which performs a fast optimistic per-oop check, and backs off
4932 4939 // further to JVM_ArrayCopy on the first per-oop check that fails.
4933 4940 // (Actually, we don't move raw bits only; the GC requires card marks.)
4934 4941
4935 4942 // Get the klassOop for both src and dest
4936 4943 Node* src_klass = load_object_klass(src);
4937 4944 Node* dest_klass = load_object_klass(dest);
4938 4945
4939 4946 // Generate the subtype check.
4940 4947 // This might fold up statically, or then again it might not.
4941 4948 //
4942 4949 // Non-static example: Copying List<String>.elements to a new String[].
4943 4950 // The backing store for a List<String> is always an Object[],
4944 4951 // but its elements are always type String, if the generic types
4945 4952 // are correct at the source level.
4946 4953 //
4947 4954 // Test S[] against D[], not S against D, because (probably)
4948 4955 // the secondary supertype cache is less busy for S[] than S.
4949 4956 // This usually only matters when D is an interface.
4950 4957 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4951 4958 // Plug failing path into checked_oop_disjoint_arraycopy
4952 4959 if (not_subtype_ctrl != top()) {
4953 4960 PreserveJVMState pjvms(this);
4954 4961 set_control(not_subtype_ctrl);
4955 4962 // (At this point we can assume disjoint_bases, since types differ.)
4956 4963 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4957 4964 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4958 4965 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4959 4966 Node* dest_elem_klass = _gvn.transform(n1);
4960 4967 Node* cv = generate_checkcast_arraycopy(adr_type,
4961 4968 dest_elem_klass,
4962 4969 src, src_offset, dest, dest_offset,
4963 4970 ConvI2X(copy_length), dest_uninitialized);
4964 4971 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4965 4972 checked_control = control();
4966 4973 checked_i_o = i_o();
4967 4974 checked_mem = memory(adr_type);
4968 4975 checked_value = cv;
4969 4976 }
4970 4977 // At this point we know we do not need type checks on oop stores.
4971 4978
4972 4979 // Let's see if we need card marks:
4973 4980 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4974 4981 // If we do not need card marks, copy using the jint or jlong stub.
4975 4982 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4976 4983 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4977 4984 "sizes agree");
4978 4985 }
4979 4986 }
4980 4987
4981 4988 if (!stopped()) {
4982 4989 // Generate the fast path, if possible.
4983 4990 PreserveJVMState pjvms(this);
4984 4991 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4985 4992 src, src_offset, dest, dest_offset,
4986 4993 ConvI2X(copy_length), dest_uninitialized);
4987 4994
4988 4995 // Present the results of the fast call.
4989 4996 result_region->init_req(fast_path, control());
4990 4997 result_i_o ->init_req(fast_path, i_o());
4991 4998 result_memory->init_req(fast_path, memory(adr_type));
4992 4999 }
4993 5000
4994 5001 // Here are all the slow paths up to this point, in one bundle:
4995 5002 slow_control = top();
4996 5003 if (slow_region != NULL)
4997 5004 slow_control = _gvn.transform(slow_region);
4998 5005 debug_only(slow_region = (RegionNode*)badAddress);
4999 5006
5000 5007 set_control(checked_control);
5001 5008 if (!stopped()) {
5002 5009 // Clean up after the checked call.
5003 5010 // The returned value is either 0 or -1^K,
5004 5011 // where K = number of partially transferred array elements.
5005 5012 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
5006 5013 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
5007 5014 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5008 5015
5009 5016 // If it is 0, we are done, so transfer to the end.
5010 5017 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
5011 5018 result_region->init_req(checked_path, checks_done);
5012 5019 result_i_o ->init_req(checked_path, checked_i_o);
5013 5020 result_memory->init_req(checked_path, checked_mem);
5014 5021
5015 5022 // If it is not zero, merge into the slow call.
5016 5023 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
5017 5024 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
5018 5025 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
5019 5026 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5020 5027 record_for_igvn(slow_reg2);
5021 5028 slow_reg2 ->init_req(1, slow_control);
5022 5029 slow_i_o2 ->init_req(1, slow_i_o);
5023 5030 slow_mem2 ->init_req(1, slow_mem);
5024 5031 slow_reg2 ->init_req(2, control());
5025 5032 slow_i_o2 ->init_req(2, checked_i_o);
5026 5033 slow_mem2 ->init_req(2, checked_mem);
5027 5034
5028 5035 slow_control = _gvn.transform(slow_reg2);
5029 5036 slow_i_o = _gvn.transform(slow_i_o2);
5030 5037 slow_mem = _gvn.transform(slow_mem2);
5031 5038
5032 5039 if (alloc != NULL) {
5033 5040 // We'll restart from the very beginning, after zeroing the whole thing.
5034 5041 // This can cause double writes, but that's OK since dest is brand new.
5035 5042 // So we ignore the low 31 bits of the value returned from the stub.
5036 5043 } else {
5037 5044 // We must continue the copy exactly where it failed, or else
5038 5045 // another thread might see the wrong number of writes to dest.
5039 5046 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
5040 5047 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
5041 5048 slow_offset->init_req(1, intcon(0));
5042 5049 slow_offset->init_req(2, checked_offset);
5043 5050 slow_offset = _gvn.transform(slow_offset);
5044 5051
5045 5052 // Adjust the arguments by the conditionally incoming offset.
5046 5053 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
5047 5054 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
5048 5055 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
5049 5056
5050 5057 // Tweak the node variables to adjust the code produced below:
5051 5058 src_offset = src_off_plus;
5052 5059 dest_offset = dest_off_plus;
5053 5060 copy_length = length_minus;
5054 5061 }
5055 5062 }
5056 5063
5057 5064 set_control(slow_control);
5058 5065 if (!stopped()) {
5059 5066 // Generate the slow path, if needed.
5060 5067 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5061 5068
5062 5069 set_memory(slow_mem, adr_type);
5063 5070 set_i_o(slow_i_o);
5064 5071
5065 5072 if (dest_uninitialized) {
5066 5073 generate_clear_array(adr_type, dest, basic_elem_type,
5067 5074 intcon(0), NULL,
5068 5075 alloc->in(AllocateNode::AllocSize));
5069 5076 }
5070 5077
5071 5078 generate_slow_arraycopy(adr_type,
5072 5079 src, src_offset, dest, dest_offset,
5073 5080 copy_length, /*dest_uninitialized*/false);
5074 5081
5075 5082 result_region->init_req(slow_call_path, control());
5076 5083 result_i_o ->init_req(slow_call_path, i_o());
5077 5084 result_memory->init_req(slow_call_path, memory(adr_type));
5078 5085 }
5079 5086
5080 5087 // Remove unused edges.
5081 5088 for (uint i = 1; i < result_region->req(); i++) {
5082 5089 if (result_region->in(i) == NULL)
5083 5090 result_region->init_req(i, top());
5084 5091 }
5085 5092
5086 5093 // Finished; return the combined state.
5087 5094 set_control( _gvn.transform(result_region) );
5088 5095 set_i_o( _gvn.transform(result_i_o) );
5089 5096 set_memory( _gvn.transform(result_memory), adr_type );
5090 5097
5091 5098 // The memory edges above are precise in order to model effects around
5092 5099 // array copies accurately to allow value numbering of field loads around
5093 5100 // arraycopy. Such field loads, both before and after, are common in Java
5094 5101 // collections and similar classes involving header/array data structures.
5095 5102 //
5096 5103 // But with low number of register or when some registers are used or killed
5097 5104 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5098 5105 // The next memory barrier is added to avoid it. If the arraycopy can be
5099 5106 // optimized away (which it can, sometimes) then we can manually remove
5100 5107 // the membar also.
5101 5108 //
5102 5109 // Do not let reads from the cloned object float above the arraycopy.
5103 5110 if (InsertMemBarAfterArraycopy || alloc != NULL)
5104 5111 insert_mem_bar(Op_MemBarCPUOrder);
5105 5112 }
5106 5113
5107 5114
5108 5115 // Helper function which determines if an arraycopy immediately follows
5109 5116 // an allocation, with no intervening tests or other escapes for the object.
5110 5117 AllocateArrayNode*
5111 5118 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5112 5119 RegionNode* slow_region) {
5113 5120 if (stopped()) return NULL; // no fast path
5114 5121 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5115 5122
5116 5123 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5117 5124 if (alloc == NULL) return NULL;
5118 5125
5119 5126 Node* rawmem = memory(Compile::AliasIdxRaw);
5120 5127 // Is the allocation's memory state untouched?
5121 5128 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5122 5129 // Bail out if there have been raw-memory effects since the allocation.
5123 5130 // (Example: There might have been a call or safepoint.)
5124 5131 return NULL;
5125 5132 }
5126 5133 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5127 5134 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5128 5135 return NULL;
5129 5136 }
5130 5137
5131 5138 // There must be no unexpected observers of this allocation.
5132 5139 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5133 5140 Node* obs = ptr->fast_out(i);
5134 5141 if (obs != this->map()) {
5135 5142 return NULL;
5136 5143 }
5137 5144 }
5138 5145
5139 5146 // This arraycopy must unconditionally follow the allocation of the ptr.
5140 5147 Node* alloc_ctl = ptr->in(0);
5141 5148 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5142 5149
5143 5150 Node* ctl = control();
5144 5151 while (ctl != alloc_ctl) {
5145 5152 // There may be guards which feed into the slow_region.
5146 5153 // Any other control flow means that we might not get a chance
5147 5154 // to finish initializing the allocated object.
5148 5155 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5149 5156 IfNode* iff = ctl->in(0)->as_If();
5150 5157 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5151 5158 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5152 5159 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5153 5160 ctl = iff->in(0); // This test feeds the known slow_region.
5154 5161 continue;
5155 5162 }
5156 5163 // One more try: Various low-level checks bottom out in
5157 5164 // uncommon traps. If the debug-info of the trap omits
5158 5165 // any reference to the allocation, as we've already
5159 5166 // observed, then there can be no objection to the trap.
5160 5167 bool found_trap = false;
5161 5168 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5162 5169 Node* obs = not_ctl->fast_out(j);
5163 5170 if (obs->in(0) == not_ctl && obs->is_Call() &&
5164 5171 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5165 5172 found_trap = true; break;
5166 5173 }
5167 5174 }
5168 5175 if (found_trap) {
5169 5176 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5170 5177 continue;
5171 5178 }
5172 5179 }
5173 5180 return NULL;
5174 5181 }
5175 5182
5176 5183 // If we get this far, we have an allocation which immediately
5177 5184 // precedes the arraycopy, and we can take over zeroing the new object.
5178 5185 // The arraycopy will finish the initialization, and provide
5179 5186 // a new control state to which we will anchor the destination pointer.
5180 5187
5181 5188 return alloc;
5182 5189 }
5183 5190
5184 5191 // Helper for initialization of arrays, creating a ClearArray.
5185 5192 // It writes zero bits in [start..end), within the body of an array object.
5186 5193 // The memory effects are all chained onto the 'adr_type' alias category.
5187 5194 //
5188 5195 // Since the object is otherwise uninitialized, we are free
5189 5196 // to put a little "slop" around the edges of the cleared area,
5190 5197 // as long as it does not go back into the array's header,
5191 5198 // or beyond the array end within the heap.
5192 5199 //
5193 5200 // The lower edge can be rounded down to the nearest jint and the
5194 5201 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5195 5202 //
5196 5203 // Arguments:
5197 5204 // adr_type memory slice where writes are generated
5198 5205 // dest oop of the destination array
5199 5206 // basic_elem_type element type of the destination
5200 5207 // slice_idx array index of first element to store
5201 5208 // slice_len number of elements to store (or NULL)
5202 5209 // dest_size total size in bytes of the array object
5203 5210 //
5204 5211 // Exactly one of slice_len or dest_size must be non-NULL.
5205 5212 // If dest_size is non-NULL, zeroing extends to the end of the object.
5206 5213 // If slice_len is non-NULL, the slice_idx value must be a constant.
5207 5214 void
5208 5215 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5209 5216 Node* dest,
5210 5217 BasicType basic_elem_type,
5211 5218 Node* slice_idx,
5212 5219 Node* slice_len,
5213 5220 Node* dest_size) {
5214 5221 // one or the other but not both of slice_len and dest_size:
5215 5222 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5216 5223 if (slice_len == NULL) slice_len = top();
5217 5224 if (dest_size == NULL) dest_size = top();
5218 5225
5219 5226 // operate on this memory slice:
5220 5227 Node* mem = memory(adr_type); // memory slice to operate on
5221 5228
5222 5229 // scaling and rounding of indexes:
5223 5230 int scale = exact_log2(type2aelembytes(basic_elem_type));
5224 5231 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5225 5232 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5226 5233 int bump_bit = (-1 << scale) & BytesPerInt;
5227 5234
5228 5235 // determine constant starts and ends
5229 5236 const intptr_t BIG_NEG = -128;
5230 5237 assert(BIG_NEG + 2*abase < 0, "neg enough");
5231 5238 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5232 5239 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5233 5240 if (slice_len_con == 0) {
5234 5241 return; // nothing to do here
5235 5242 }
5236 5243 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5237 5244 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5238 5245 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5239 5246 assert(end_con < 0, "not two cons");
5240 5247 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5241 5248 BytesPerLong);
5242 5249 }
5243 5250
5244 5251 if (start_con >= 0 && end_con >= 0) {
5245 5252 // Constant start and end. Simple.
5246 5253 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5247 5254 start_con, end_con, &_gvn);
5248 5255 } else if (start_con >= 0 && dest_size != top()) {
5249 5256 // Constant start, pre-rounded end after the tail of the array.
5250 5257 Node* end = dest_size;
5251 5258 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5252 5259 start_con, end, &_gvn);
5253 5260 } else if (start_con >= 0 && slice_len != top()) {
5254 5261 // Constant start, non-constant end. End needs rounding up.
5255 5262 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5256 5263 intptr_t end_base = abase + (slice_idx_con << scale);
5257 5264 int end_round = (-1 << scale) & (BytesPerLong - 1);
5258 5265 Node* end = ConvI2X(slice_len);
5259 5266 if (scale != 0)
5260 5267 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5261 5268 end_base += end_round;
5262 5269 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5263 5270 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5264 5271 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5265 5272 start_con, end, &_gvn);
5266 5273 } else if (start_con < 0 && dest_size != top()) {
5267 5274 // Non-constant start, pre-rounded end after the tail of the array.
5268 5275 // This is almost certainly a "round-to-end" operation.
5269 5276 Node* start = slice_idx;
5270 5277 start = ConvI2X(start);
5271 5278 if (scale != 0)
5272 5279 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5273 5280 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5274 5281 if ((bump_bit | clear_low) != 0) {
5275 5282 int to_clear = (bump_bit | clear_low);
5276 5283 // Align up mod 8, then store a jint zero unconditionally
5277 5284 // just before the mod-8 boundary.
5278 5285 if (((abase + bump_bit) & ~to_clear) - bump_bit
5279 5286 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5280 5287 bump_bit = 0;
5281 5288 assert((abase & to_clear) == 0, "array base must be long-aligned");
5282 5289 } else {
5283 5290 // Bump 'start' up to (or past) the next jint boundary:
5284 5291 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5285 5292 assert((abase & clear_low) == 0, "array base must be int-aligned");
5286 5293 }
5287 5294 // Round bumped 'start' down to jlong boundary in body of array.
5288 5295 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5289 5296 if (bump_bit != 0) {
5290 5297 // Store a zero to the immediately preceding jint:
5291 5298 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5292 5299 Node* p1 = basic_plus_adr(dest, x1);
5293 5300 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5294 5301 mem = _gvn.transform(mem);
5295 5302 }
5296 5303 }
5297 5304 Node* end = dest_size; // pre-rounded
5298 5305 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5299 5306 start, end, &_gvn);
5300 5307 } else {
5301 5308 // Non-constant start, unrounded non-constant end.
5302 5309 // (Nobody zeroes a random midsection of an array using this routine.)
5303 5310 ShouldNotReachHere(); // fix caller
5304 5311 }
5305 5312
5306 5313 // Done.
5307 5314 set_memory(mem, adr_type);
5308 5315 }
5309 5316
5310 5317
5311 5318 bool
5312 5319 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5313 5320 BasicType basic_elem_type,
5314 5321 AllocateNode* alloc,
5315 5322 Node* src, Node* src_offset,
5316 5323 Node* dest, Node* dest_offset,
5317 5324 Node* dest_size, bool dest_uninitialized) {
5318 5325 // See if there is an advantage from block transfer.
5319 5326 int scale = exact_log2(type2aelembytes(basic_elem_type));
5320 5327 if (scale >= LogBytesPerLong)
5321 5328 return false; // it is already a block transfer
5322 5329
5323 5330 // Look at the alignment of the starting offsets.
5324 5331 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5325 5332 const intptr_t BIG_NEG = -128;
5326 5333 assert(BIG_NEG + 2*abase < 0, "neg enough");
5327 5334
5328 5335 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5329 5336 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5330 5337 if (src_off < 0 || dest_off < 0)
5331 5338 // At present, we can only understand constants.
5332 5339 return false;
5333 5340
5334 5341 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5335 5342 // Non-aligned; too bad.
5336 5343 // One more chance: Pick off an initial 32-bit word.
5337 5344 // This is a common case, since abase can be odd mod 8.
5338 5345 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5339 5346 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5340 5347 Node* sptr = basic_plus_adr(src, src_off);
5341 5348 Node* dptr = basic_plus_adr(dest, dest_off);
5342 5349 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5343 5350 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5344 5351 src_off += BytesPerInt;
5345 5352 dest_off += BytesPerInt;
5346 5353 } else {
5347 5354 return false;
5348 5355 }
5349 5356 }
5350 5357 assert(src_off % BytesPerLong == 0, "");
5351 5358 assert(dest_off % BytesPerLong == 0, "");
5352 5359
5353 5360 // Do this copy by giant steps.
5354 5361 Node* sptr = basic_plus_adr(src, src_off);
5355 5362 Node* dptr = basic_plus_adr(dest, dest_off);
5356 5363 Node* countx = dest_size;
5357 5364 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5358 5365 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5359 5366
5360 5367 bool disjoint_bases = true; // since alloc != NULL
5361 5368 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5362 5369 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5363 5370
5364 5371 return true;
5365 5372 }
5366 5373
5367 5374
5368 5375 // Helper function; generates code for the slow case.
5369 5376 // We make a call to a runtime method which emulates the native method,
5370 5377 // but without the native wrapper overhead.
5371 5378 void
5372 5379 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5373 5380 Node* src, Node* src_offset,
5374 5381 Node* dest, Node* dest_offset,
5375 5382 Node* copy_length, bool dest_uninitialized) {
5376 5383 assert(!dest_uninitialized, "Invariant");
5377 5384 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5378 5385 OptoRuntime::slow_arraycopy_Type(),
5379 5386 OptoRuntime::slow_arraycopy_Java(),
5380 5387 "slow_arraycopy", adr_type,
5381 5388 src, src_offset, dest, dest_offset,
5382 5389 copy_length);
5383 5390
5384 5391 // Handle exceptions thrown by this fellow:
5385 5392 make_slow_call_ex(call, env()->Throwable_klass(), false);
5386 5393 }
5387 5394
5388 5395 // Helper function; generates code for cases requiring runtime checks.
5389 5396 Node*
5390 5397 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5391 5398 Node* dest_elem_klass,
5392 5399 Node* src, Node* src_offset,
5393 5400 Node* dest, Node* dest_offset,
5394 5401 Node* copy_length, bool dest_uninitialized) {
5395 5402 if (stopped()) return NULL;
5396 5403
5397 5404 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5398 5405 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5399 5406 return NULL;
5400 5407 }
5401 5408
5402 5409 // Pick out the parameters required to perform a store-check
5403 5410 // for the target array. This is an optimistic check. It will
5404 5411 // look in each non-null element's class, at the desired klass's
5405 5412 // super_check_offset, for the desired klass.
5406 5413 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5407 5414 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5408 5415 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5409 5416 Node* check_offset = ConvI2X(_gvn.transform(n3));
5410 5417 Node* check_value = dest_elem_klass;
5411 5418
5412 5419 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5413 5420 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5414 5421
5415 5422 // (We know the arrays are never conjoint, because their types differ.)
5416 5423 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5417 5424 OptoRuntime::checkcast_arraycopy_Type(),
5418 5425 copyfunc_addr, "checkcast_arraycopy", adr_type,
5419 5426 // five arguments, of which two are
5420 5427 // intptr_t (jlong in LP64)
5421 5428 src_start, dest_start,
5422 5429 copy_length XTOP,
5423 5430 check_offset XTOP,
5424 5431 check_value);
5425 5432
5426 5433 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5427 5434 }
5428 5435
5429 5436
5430 5437 // Helper function; generates code for cases requiring runtime checks.
5431 5438 Node*
5432 5439 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5433 5440 Node* src, Node* src_offset,
5434 5441 Node* dest, Node* dest_offset,
5435 5442 Node* copy_length, bool dest_uninitialized) {
5436 5443 assert(!dest_uninitialized, "Invariant");
5437 5444 if (stopped()) return NULL;
5438 5445 address copyfunc_addr = StubRoutines::generic_arraycopy();
5439 5446 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5440 5447 return NULL;
5441 5448 }
5442 5449
5443 5450 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5444 5451 OptoRuntime::generic_arraycopy_Type(),
5445 5452 copyfunc_addr, "generic_arraycopy", adr_type,
5446 5453 src, src_offset, dest, dest_offset, copy_length);
5447 5454
5448 5455 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5449 5456 }
5450 5457
5451 5458 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5452 5459 void
5453 5460 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5454 5461 BasicType basic_elem_type,
5455 5462 bool disjoint_bases,
5456 5463 Node* src, Node* src_offset,
5457 5464 Node* dest, Node* dest_offset,
5458 5465 Node* copy_length, bool dest_uninitialized) {
5459 5466 if (stopped()) return; // nothing to do
5460 5467
5461 5468 Node* src_start = src;
5462 5469 Node* dest_start = dest;
5463 5470 if (src_offset != NULL || dest_offset != NULL) {
5464 5471 assert(src_offset != NULL && dest_offset != NULL, "");
5465 5472 src_start = array_element_address(src, src_offset, basic_elem_type);
5466 5473 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5467 5474 }
5468 5475
5469 5476 // Figure out which arraycopy runtime method to call.
5470 5477 const char* copyfunc_name = "arraycopy";
5471 5478 address copyfunc_addr =
5472 5479 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5473 5480 disjoint_bases, copyfunc_name, dest_uninitialized);
5474 5481
5475 5482 // Call it. Note that the count_ix value is not scaled to a byte-size.
5476 5483 make_runtime_call(RC_LEAF|RC_NO_FP,
5477 5484 OptoRuntime::fast_arraycopy_Type(),
5478 5485 copyfunc_addr, copyfunc_name, adr_type,
5479 5486 src_start, dest_start, copy_length XTOP);
5480 5487 }
5481 5488
5482 5489 //----------------------------inline_reference_get----------------------------
5483 5490
5484 5491 bool LibraryCallKit::inline_reference_get() {
5485 5492 const int nargs = 1; // self
5486 5493
5487 5494 guarantee(java_lang_ref_Reference::referent_offset > 0,
5488 5495 "should have already been set");
5489 5496
5490 5497 int referent_offset = java_lang_ref_Reference::referent_offset;
5491 5498
5492 5499 // Restore the stack and pop off the argument
5493 5500 _sp += nargs;
5494 5501 Node *reference_obj = pop();
5495 5502
5496 5503 // Null check on self without removing any arguments.
5497 5504 _sp += nargs;
5498 5505 reference_obj = do_null_check(reference_obj, T_OBJECT);
5499 5506 _sp -= nargs;;
5500 5507
5501 5508 if (stopped()) return true;
5502 5509
5503 5510 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5504 5511
5505 5512 ciInstanceKlass* klass = env()->Object_klass();
5506 5513 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5507 5514
5508 5515 Node* no_ctrl = NULL;
5509 5516 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5510 5517
5511 5518 // Use the pre-barrier to record the value in the referent field
5512 5519 pre_barrier(false /* do_load */,
5513 5520 control(),
5514 5521 NULL /* obj */, NULL /* adr */, -1 /* alias_idx */, NULL /* val */, NULL /* val_type */,
5515 5522 result /* pre_val */,
5516 5523 T_OBJECT);
5517 5524
5518 5525 push(result);
5519 5526 return true;
5520 5527 }
5521 5528
↓ open down ↓ |
3328 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX