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 - * Copyright 1999-2009 Sun Microsystems, Inc. All Rights Reserved.
2 + * Copyright 1999-2010 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 21 * have any questions.
22 22 *
23 23 */
24 24
25 25 #include "incls/_precompiled.incl"
26 26 #include "incls/_library_call.cpp.incl"
27 27
28 28 class LibraryIntrinsic : public InlineCallGenerator {
29 29 // Extend the set of intrinsics known to the runtime:
30 30 public:
31 31 private:
32 32 bool _is_virtual;
33 33 vmIntrinsics::ID _intrinsic_id;
34 34
35 35 public:
36 36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
37 37 : InlineCallGenerator(m),
38 38 _is_virtual(is_virtual),
39 39 _intrinsic_id(id)
40 40 {
41 41 }
42 42 virtual bool is_intrinsic() const { return true; }
43 43 virtual bool is_virtual() const { return _is_virtual; }
44 44 virtual JVMState* generate(JVMState* jvms);
45 45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
46 46 };
47 47
48 48
49 49 // Local helper class for LibraryIntrinsic:
50 50 class LibraryCallKit : public GraphKit {
51 51 private:
52 52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
53 53
54 54 public:
55 55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
56 56 : GraphKit(caller),
57 57 _intrinsic(intrinsic)
58 58 {
59 59 }
60 60
61 61 ciMethod* caller() const { return jvms()->method(); }
62 62 int bci() const { return jvms()->bci(); }
63 63 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
64 64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
65 65 ciMethod* callee() const { return _intrinsic->method(); }
66 66 ciSignature* signature() const { return callee()->signature(); }
67 67 int arg_size() const { return callee()->arg_size(); }
68 68
69 69 bool try_to_inline();
70 70
71 71 // Helper functions to inline natives
72 72 void push_result(RegionNode* region, PhiNode* value);
73 73 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
74 74 Node* generate_slow_guard(Node* test, RegionNode* region);
75 75 Node* generate_fair_guard(Node* test, RegionNode* region);
76 76 Node* generate_negative_guard(Node* index, RegionNode* region,
77 77 // resulting CastII of index:
78 78 Node* *pos_index = NULL);
79 79 Node* generate_nonpositive_guard(Node* index, bool never_negative,
80 80 // resulting CastII of index:
81 81 Node* *pos_index = NULL);
82 82 Node* generate_limit_guard(Node* offset, Node* subseq_length,
83 83 Node* array_length,
84 84 RegionNode* region);
85 85 Node* generate_current_thread(Node* &tls_output);
86 86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
87 87 bool disjoint_bases, const char* &name);
88 88 Node* load_mirror_from_klass(Node* klass);
89 89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
90 90 int nargs,
91 91 RegionNode* region, int null_path,
92 92 int offset);
93 93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
94 94 RegionNode* region, int null_path) {
95 95 int offset = java_lang_Class::klass_offset_in_bytes();
96 96 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
97 97 region, null_path,
98 98 offset);
99 99 }
100 100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
101 101 int nargs,
102 102 RegionNode* region, int null_path) {
103 103 int offset = java_lang_Class::array_klass_offset_in_bytes();
104 104 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
105 105 region, null_path,
106 106 offset);
107 107 }
108 108 Node* generate_access_flags_guard(Node* kls,
109 109 int modifier_mask, int modifier_bits,
110 110 RegionNode* region);
111 111 Node* generate_interface_guard(Node* kls, RegionNode* region);
112 112 Node* generate_array_guard(Node* kls, RegionNode* region) {
113 113 return generate_array_guard_common(kls, region, false, false);
114 114 }
115 115 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
116 116 return generate_array_guard_common(kls, region, false, true);
117 117 }
118 118 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
119 119 return generate_array_guard_common(kls, region, true, false);
120 120 }
121 121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
122 122 return generate_array_guard_common(kls, region, true, true);
123 123 }
124 124 Node* generate_array_guard_common(Node* kls, RegionNode* region,
125 125 bool obj_array, bool not_array);
126 126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
127 127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
128 128 bool is_virtual = false, bool is_static = false);
129 129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
130 130 return generate_method_call(method_id, false, true);
131 131 }
132 132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
133 133 return generate_method_call(method_id, true, false);
134 134 }
135 135
136 136 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
137 137 bool inline_string_compareTo();
138 138 bool inline_string_indexOf();
139 139 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
140 140 bool inline_string_equals();
141 141 Node* pop_math_arg();
142 142 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
143 143 bool inline_math_native(vmIntrinsics::ID id);
144 144 bool inline_trig(vmIntrinsics::ID id);
145 145 bool inline_trans(vmIntrinsics::ID id);
146 146 bool inline_abs(vmIntrinsics::ID id);
147 147 bool inline_sqrt(vmIntrinsics::ID id);
148 148 bool inline_pow(vmIntrinsics::ID id);
149 149 bool inline_exp(vmIntrinsics::ID id);
150 150 bool inline_min_max(vmIntrinsics::ID id);
151 151 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
152 152 // This returns Type::AnyPtr, RawPtr, or OopPtr.
153 153 int classify_unsafe_addr(Node* &base, Node* &offset);
154 154 Node* make_unsafe_address(Node* base, Node* offset);
155 155 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
156 156 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
157 157 bool inline_unsafe_allocate();
158 158 bool inline_unsafe_copyMemory();
159 159 bool inline_native_currentThread();
160 160 bool inline_native_time_funcs(bool isNano);
161 161 bool inline_native_isInterrupted();
162 162 bool inline_native_Class_query(vmIntrinsics::ID id);
163 163 bool inline_native_subtype_check();
164 164
165 165 bool inline_native_newArray();
166 166 bool inline_native_getLength();
167 167 bool inline_array_copyOf(bool is_copyOfRange);
168 168 bool inline_array_equals();
169 169 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
170 170 bool inline_native_clone(bool is_virtual);
171 171 bool inline_native_Reflection_getCallerClass();
172 172 bool inline_native_AtomicLong_get();
173 173 bool inline_native_AtomicLong_attemptUpdate();
174 174 bool is_method_invoke_or_aux_frame(JVMState* jvms);
175 175 // Helper function for inlining native object hash method
176 176 bool inline_native_hashcode(bool is_virtual, bool is_static);
177 177 bool inline_native_getClass();
178 178
179 179 // Helper functions for inlining arraycopy
180 180 bool inline_arraycopy();
181 181 void generate_arraycopy(const TypePtr* adr_type,
182 182 BasicType basic_elem_type,
183 183 Node* src, Node* src_offset,
184 184 Node* dest, Node* dest_offset,
185 185 Node* copy_length,
186 186 bool disjoint_bases = false,
187 187 bool length_never_negative = false,
188 188 RegionNode* slow_region = NULL);
189 189 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
190 190 RegionNode* slow_region);
191 191 void generate_clear_array(const TypePtr* adr_type,
192 192 Node* dest,
193 193 BasicType basic_elem_type,
194 194 Node* slice_off,
195 195 Node* slice_len,
196 196 Node* slice_end);
197 197 bool generate_block_arraycopy(const TypePtr* adr_type,
198 198 BasicType basic_elem_type,
199 199 AllocateNode* alloc,
200 200 Node* src, Node* src_offset,
201 201 Node* dest, Node* dest_offset,
202 202 Node* dest_size);
203 203 void generate_slow_arraycopy(const TypePtr* adr_type,
204 204 Node* src, Node* src_offset,
205 205 Node* dest, Node* dest_offset,
206 206 Node* copy_length);
207 207 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
208 208 Node* dest_elem_klass,
209 209 Node* src, Node* src_offset,
210 210 Node* dest, Node* dest_offset,
211 211 Node* copy_length);
212 212 Node* generate_generic_arraycopy(const TypePtr* adr_type,
213 213 Node* src, Node* src_offset,
214 214 Node* dest, Node* dest_offset,
215 215 Node* copy_length);
216 216 void generate_unchecked_arraycopy(const TypePtr* adr_type,
217 217 BasicType basic_elem_type,
218 218 bool disjoint_bases,
219 219 Node* src, Node* src_offset,
220 220 Node* dest, Node* dest_offset,
221 221 Node* copy_length);
222 222 bool inline_unsafe_CAS(BasicType type);
223 223 bool inline_unsafe_ordered_store(BasicType type);
224 224 bool inline_fp_conversions(vmIntrinsics::ID id);
225 225 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
226 226 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
227 227 bool inline_bitCount(vmIntrinsics::ID id);
228 228 bool inline_reverseBytes(vmIntrinsics::ID id);
229 229 };
230 230
231 231
232 232 //---------------------------make_vm_intrinsic----------------------------
233 233 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
234 234 vmIntrinsics::ID id = m->intrinsic_id();
235 235 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
236 236
237 237 if (DisableIntrinsic[0] != '\0'
238 238 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
239 239 // disabled by a user request on the command line:
240 240 // example: -XX:DisableIntrinsic=_hashCode,_getClass
241 241 return NULL;
242 242 }
243 243
244 244 if (!m->is_loaded()) {
245 245 // do not attempt to inline unloaded methods
246 246 return NULL;
247 247 }
248 248
249 249 // Only a few intrinsics implement a virtual dispatch.
250 250 // They are expensive calls which are also frequently overridden.
251 251 if (is_virtual) {
252 252 switch (id) {
253 253 case vmIntrinsics::_hashCode:
254 254 case vmIntrinsics::_clone:
255 255 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
256 256 break;
257 257 default:
258 258 return NULL;
259 259 }
260 260 }
261 261
262 262 // -XX:-InlineNatives disables nearly all intrinsics:
263 263 if (!InlineNatives) {
264 264 switch (id) {
265 265 case vmIntrinsics::_indexOf:
266 266 case vmIntrinsics::_compareTo:
267 267 case vmIntrinsics::_equals:
268 268 case vmIntrinsics::_equalsC:
269 269 break; // InlineNatives does not control String.compareTo
270 270 default:
271 271 return NULL;
272 272 }
273 273 }
274 274
275 275 switch (id) {
276 276 case vmIntrinsics::_compareTo:
277 277 if (!SpecialStringCompareTo) return NULL;
278 278 break;
279 279 case vmIntrinsics::_indexOf:
280 280 if (!SpecialStringIndexOf) return NULL;
281 281 break;
282 282 case vmIntrinsics::_equals:
283 283 if (!SpecialStringEquals) return NULL;
284 284 break;
285 285 case vmIntrinsics::_equalsC:
286 286 if (!SpecialArraysEquals) return NULL;
287 287 break;
288 288 case vmIntrinsics::_arraycopy:
289 289 if (!InlineArrayCopy) return NULL;
290 290 break;
291 291 case vmIntrinsics::_copyMemory:
292 292 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
293 293 if (!InlineArrayCopy) return NULL;
294 294 break;
295 295 case vmIntrinsics::_hashCode:
296 296 if (!InlineObjectHash) return NULL;
297 297 break;
298 298 case vmIntrinsics::_clone:
299 299 case vmIntrinsics::_copyOf:
300 300 case vmIntrinsics::_copyOfRange:
301 301 if (!InlineObjectCopy) return NULL;
302 302 // These also use the arraycopy intrinsic mechanism:
303 303 if (!InlineArrayCopy) return NULL;
304 304 break;
305 305 case vmIntrinsics::_checkIndex:
306 306 // We do not intrinsify this. The optimizer does fine with it.
307 307 return NULL;
308 308
309 309 case vmIntrinsics::_get_AtomicLong:
310 310 case vmIntrinsics::_attemptUpdate:
311 311 if (!InlineAtomicLong) return NULL;
312 312 break;
313 313
314 314 case vmIntrinsics::_getCallerClass:
315 315 if (!UseNewReflection) return NULL;
316 316 if (!InlineReflectionGetCallerClass) return NULL;
317 317 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
318 318 break;
319 319
320 320 case vmIntrinsics::_bitCount_i:
321 321 case vmIntrinsics::_bitCount_l:
322 322 if (!UsePopCountInstruction) return NULL;
323 323 break;
324 324
325 325 default:
326 326 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
327 327 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
328 328 break;
329 329 }
330 330
331 331 // -XX:-InlineClassNatives disables natives from the Class class.
332 332 // The flag applies to all reflective calls, notably Array.newArray
333 333 // (visible to Java programmers as Array.newInstance).
334 334 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
335 335 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
336 336 if (!InlineClassNatives) return NULL;
337 337 }
338 338
339 339 // -XX:-InlineThreadNatives disables natives from the Thread class.
340 340 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
341 341 if (!InlineThreadNatives) return NULL;
342 342 }
343 343
344 344 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
345 345 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
346 346 m->holder()->name() == ciSymbol::java_lang_Float() ||
347 347 m->holder()->name() == ciSymbol::java_lang_Double()) {
348 348 if (!InlineMathNatives) return NULL;
349 349 }
350 350
351 351 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
352 352 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
353 353 if (!InlineUnsafeOps) return NULL;
354 354 }
355 355
356 356 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
357 357 }
358 358
359 359 //----------------------register_library_intrinsics-----------------------
360 360 // Initialize this file's data structures, for each Compile instance.
361 361 void Compile::register_library_intrinsics() {
362 362 // Nothing to do here.
363 363 }
364 364
365 365 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
366 366 LibraryCallKit kit(jvms, this);
367 367 Compile* C = kit.C;
368 368 int nodes = C->unique();
369 369 #ifndef PRODUCT
370 370 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
371 371 char buf[1000];
372 372 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
373 373 tty->print_cr("Intrinsic %s", str);
374 374 }
375 375 #endif
376 376 if (kit.try_to_inline()) {
377 377 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
378 378 tty->print("Inlining intrinsic %s%s at bci:%d in",
379 379 vmIntrinsics::name_at(intrinsic_id()),
380 380 (is_virtual() ? " (virtual)" : ""), kit.bci());
381 381 kit.caller()->print_short_name(tty);
382 382 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
383 383 }
384 384 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
385 385 if (C->log()) {
386 386 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
387 387 vmIntrinsics::name_at(intrinsic_id()),
388 388 (is_virtual() ? " virtual='1'" : ""),
389 389 C->unique() - nodes);
390 390 }
391 391 return kit.transfer_exceptions_into_jvms();
392 392 }
393 393
394 394 if (PrintIntrinsics) {
395 395 tty->print("Did not inline intrinsic %s%s at bci:%d in",
396 396 vmIntrinsics::name_at(intrinsic_id()),
397 397 (is_virtual() ? " (virtual)" : ""), kit.bci());
398 398 kit.caller()->print_short_name(tty);
399 399 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
400 400 }
401 401 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
402 402 return NULL;
403 403 }
404 404
405 405 bool LibraryCallKit::try_to_inline() {
406 406 // Handle symbolic names for otherwise undistinguished boolean switches:
407 407 const bool is_store = true;
408 408 const bool is_native_ptr = true;
409 409 const bool is_static = true;
410 410
411 411 switch (intrinsic_id()) {
412 412 case vmIntrinsics::_hashCode:
413 413 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
414 414 case vmIntrinsics::_identityHashCode:
415 415 return inline_native_hashcode(/*!virtual*/ false, is_static);
416 416 case vmIntrinsics::_getClass:
417 417 return inline_native_getClass();
418 418
419 419 case vmIntrinsics::_dsin:
420 420 case vmIntrinsics::_dcos:
421 421 case vmIntrinsics::_dtan:
422 422 case vmIntrinsics::_dabs:
423 423 case vmIntrinsics::_datan2:
424 424 case vmIntrinsics::_dsqrt:
425 425 case vmIntrinsics::_dexp:
426 426 case vmIntrinsics::_dlog:
427 427 case vmIntrinsics::_dlog10:
428 428 case vmIntrinsics::_dpow:
429 429 return inline_math_native(intrinsic_id());
430 430
431 431 case vmIntrinsics::_min:
432 432 case vmIntrinsics::_max:
433 433 return inline_min_max(intrinsic_id());
434 434
435 435 case vmIntrinsics::_arraycopy:
436 436 return inline_arraycopy();
437 437
438 438 case vmIntrinsics::_compareTo:
439 439 return inline_string_compareTo();
440 440 case vmIntrinsics::_indexOf:
441 441 return inline_string_indexOf();
442 442 case vmIntrinsics::_equals:
443 443 return inline_string_equals();
444 444
445 445 case vmIntrinsics::_getObject:
446 446 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
447 447 case vmIntrinsics::_getBoolean:
448 448 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
449 449 case vmIntrinsics::_getByte:
450 450 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
451 451 case vmIntrinsics::_getShort:
452 452 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
453 453 case vmIntrinsics::_getChar:
454 454 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
455 455 case vmIntrinsics::_getInt:
456 456 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
457 457 case vmIntrinsics::_getLong:
458 458 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
459 459 case vmIntrinsics::_getFloat:
460 460 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
461 461 case vmIntrinsics::_getDouble:
462 462 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
463 463
464 464 case vmIntrinsics::_putObject:
465 465 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
466 466 case vmIntrinsics::_putBoolean:
467 467 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
468 468 case vmIntrinsics::_putByte:
469 469 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
470 470 case vmIntrinsics::_putShort:
471 471 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
472 472 case vmIntrinsics::_putChar:
473 473 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
474 474 case vmIntrinsics::_putInt:
475 475 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
476 476 case vmIntrinsics::_putLong:
477 477 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
478 478 case vmIntrinsics::_putFloat:
479 479 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
480 480 case vmIntrinsics::_putDouble:
481 481 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
482 482
483 483 case vmIntrinsics::_getByte_raw:
484 484 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
485 485 case vmIntrinsics::_getShort_raw:
486 486 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
487 487 case vmIntrinsics::_getChar_raw:
488 488 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
489 489 case vmIntrinsics::_getInt_raw:
490 490 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
491 491 case vmIntrinsics::_getLong_raw:
492 492 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
493 493 case vmIntrinsics::_getFloat_raw:
494 494 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
495 495 case vmIntrinsics::_getDouble_raw:
496 496 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
497 497 case vmIntrinsics::_getAddress_raw:
498 498 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
499 499
500 500 case vmIntrinsics::_putByte_raw:
501 501 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
502 502 case vmIntrinsics::_putShort_raw:
503 503 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
504 504 case vmIntrinsics::_putChar_raw:
505 505 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
506 506 case vmIntrinsics::_putInt_raw:
507 507 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
508 508 case vmIntrinsics::_putLong_raw:
509 509 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
510 510 case vmIntrinsics::_putFloat_raw:
511 511 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
512 512 case vmIntrinsics::_putDouble_raw:
513 513 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
514 514 case vmIntrinsics::_putAddress_raw:
515 515 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
516 516
517 517 case vmIntrinsics::_getObjectVolatile:
518 518 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
519 519 case vmIntrinsics::_getBooleanVolatile:
520 520 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
521 521 case vmIntrinsics::_getByteVolatile:
522 522 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
523 523 case vmIntrinsics::_getShortVolatile:
524 524 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
525 525 case vmIntrinsics::_getCharVolatile:
526 526 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
527 527 case vmIntrinsics::_getIntVolatile:
528 528 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
529 529 case vmIntrinsics::_getLongVolatile:
530 530 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
531 531 case vmIntrinsics::_getFloatVolatile:
532 532 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
533 533 case vmIntrinsics::_getDoubleVolatile:
534 534 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
535 535
536 536 case vmIntrinsics::_putObjectVolatile:
537 537 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
538 538 case vmIntrinsics::_putBooleanVolatile:
539 539 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
540 540 case vmIntrinsics::_putByteVolatile:
541 541 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
542 542 case vmIntrinsics::_putShortVolatile:
543 543 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
544 544 case vmIntrinsics::_putCharVolatile:
545 545 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
546 546 case vmIntrinsics::_putIntVolatile:
547 547 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
548 548 case vmIntrinsics::_putLongVolatile:
549 549 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
550 550 case vmIntrinsics::_putFloatVolatile:
551 551 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
552 552 case vmIntrinsics::_putDoubleVolatile:
553 553 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
554 554
555 555 case vmIntrinsics::_prefetchRead:
556 556 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
557 557 case vmIntrinsics::_prefetchWrite:
558 558 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
559 559 case vmIntrinsics::_prefetchReadStatic:
560 560 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
561 561 case vmIntrinsics::_prefetchWriteStatic:
562 562 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
563 563
564 564 case vmIntrinsics::_compareAndSwapObject:
565 565 return inline_unsafe_CAS(T_OBJECT);
566 566 case vmIntrinsics::_compareAndSwapInt:
567 567 return inline_unsafe_CAS(T_INT);
568 568 case vmIntrinsics::_compareAndSwapLong:
569 569 return inline_unsafe_CAS(T_LONG);
570 570
571 571 case vmIntrinsics::_putOrderedObject:
572 572 return inline_unsafe_ordered_store(T_OBJECT);
573 573 case vmIntrinsics::_putOrderedInt:
574 574 return inline_unsafe_ordered_store(T_INT);
575 575 case vmIntrinsics::_putOrderedLong:
576 576 return inline_unsafe_ordered_store(T_LONG);
577 577
578 578 case vmIntrinsics::_currentThread:
579 579 return inline_native_currentThread();
580 580 case vmIntrinsics::_isInterrupted:
581 581 return inline_native_isInterrupted();
582 582
583 583 case vmIntrinsics::_currentTimeMillis:
584 584 return inline_native_time_funcs(false);
585 585 case vmIntrinsics::_nanoTime:
586 586 return inline_native_time_funcs(true);
587 587 case vmIntrinsics::_allocateInstance:
588 588 return inline_unsafe_allocate();
589 589 case vmIntrinsics::_copyMemory:
590 590 return inline_unsafe_copyMemory();
591 591 case vmIntrinsics::_newArray:
592 592 return inline_native_newArray();
593 593 case vmIntrinsics::_getLength:
594 594 return inline_native_getLength();
595 595 case vmIntrinsics::_copyOf:
596 596 return inline_array_copyOf(false);
597 597 case vmIntrinsics::_copyOfRange:
598 598 return inline_array_copyOf(true);
599 599 case vmIntrinsics::_equalsC:
600 600 return inline_array_equals();
601 601 case vmIntrinsics::_clone:
602 602 return inline_native_clone(intrinsic()->is_virtual());
603 603
604 604 case vmIntrinsics::_isAssignableFrom:
605 605 return inline_native_subtype_check();
606 606
607 607 case vmIntrinsics::_isInstance:
608 608 case vmIntrinsics::_getModifiers:
609 609 case vmIntrinsics::_isInterface:
610 610 case vmIntrinsics::_isArray:
611 611 case vmIntrinsics::_isPrimitive:
612 612 case vmIntrinsics::_getSuperclass:
613 613 case vmIntrinsics::_getComponentType:
614 614 case vmIntrinsics::_getClassAccessFlags:
615 615 return inline_native_Class_query(intrinsic_id());
616 616
617 617 case vmIntrinsics::_floatToRawIntBits:
618 618 case vmIntrinsics::_floatToIntBits:
619 619 case vmIntrinsics::_intBitsToFloat:
620 620 case vmIntrinsics::_doubleToRawLongBits:
621 621 case vmIntrinsics::_doubleToLongBits:
622 622 case vmIntrinsics::_longBitsToDouble:
623 623 return inline_fp_conversions(intrinsic_id());
624 624
625 625 case vmIntrinsics::_numberOfLeadingZeros_i:
626 626 case vmIntrinsics::_numberOfLeadingZeros_l:
627 627 return inline_numberOfLeadingZeros(intrinsic_id());
628 628
629 629 case vmIntrinsics::_numberOfTrailingZeros_i:
630 630 case vmIntrinsics::_numberOfTrailingZeros_l:
631 631 return inline_numberOfTrailingZeros(intrinsic_id());
632 632
633 633 case vmIntrinsics::_bitCount_i:
634 634 case vmIntrinsics::_bitCount_l:
635 635 return inline_bitCount(intrinsic_id());
636 636
637 637 case vmIntrinsics::_reverseBytes_i:
638 638 case vmIntrinsics::_reverseBytes_l:
639 639 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
640 640
641 641 case vmIntrinsics::_get_AtomicLong:
642 642 return inline_native_AtomicLong_get();
643 643 case vmIntrinsics::_attemptUpdate:
644 644 return inline_native_AtomicLong_attemptUpdate();
645 645
646 646 case vmIntrinsics::_getCallerClass:
647 647 return inline_native_Reflection_getCallerClass();
648 648
649 649 default:
650 650 // If you get here, it may be that someone has added a new intrinsic
651 651 // to the list in vmSymbols.hpp without implementing it here.
652 652 #ifndef PRODUCT
653 653 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
654 654 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
655 655 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
656 656 }
657 657 #endif
658 658 return false;
659 659 }
660 660 }
661 661
662 662 //------------------------------push_result------------------------------
663 663 // Helper function for finishing intrinsics.
664 664 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
665 665 record_for_igvn(region);
666 666 set_control(_gvn.transform(region));
667 667 BasicType value_type = value->type()->basic_type();
668 668 push_node(value_type, _gvn.transform(value));
669 669 }
670 670
671 671 //------------------------------generate_guard---------------------------
672 672 // Helper function for generating guarded fast-slow graph structures.
673 673 // The given 'test', if true, guards a slow path. If the test fails
674 674 // then a fast path can be taken. (We generally hope it fails.)
675 675 // In all cases, GraphKit::control() is updated to the fast path.
676 676 // The returned value represents the control for the slow path.
677 677 // The return value is never 'top'; it is either a valid control
678 678 // or NULL if it is obvious that the slow path can never be taken.
679 679 // Also, if region and the slow control are not NULL, the slow edge
680 680 // is appended to the region.
681 681 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
682 682 if (stopped()) {
683 683 // Already short circuited.
684 684 return NULL;
685 685 }
686 686
687 687 // Build an if node and its projections.
688 688 // If test is true we take the slow path, which we assume is uncommon.
689 689 if (_gvn.type(test) == TypeInt::ZERO) {
690 690 // The slow branch is never taken. No need to build this guard.
691 691 return NULL;
692 692 }
693 693
694 694 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
695 695
696 696 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
697 697 if (if_slow == top()) {
698 698 // The slow branch is never taken. No need to build this guard.
699 699 return NULL;
700 700 }
701 701
702 702 if (region != NULL)
703 703 region->add_req(if_slow);
704 704
705 705 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
706 706 set_control(if_fast);
707 707
708 708 return if_slow;
709 709 }
710 710
711 711 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
712 712 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
713 713 }
714 714 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
715 715 return generate_guard(test, region, PROB_FAIR);
716 716 }
717 717
718 718 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
719 719 Node* *pos_index) {
720 720 if (stopped())
721 721 return NULL; // already stopped
722 722 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
723 723 return NULL; // index is already adequately typed
724 724 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
725 725 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
726 726 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
727 727 if (is_neg != NULL && pos_index != NULL) {
728 728 // Emulate effect of Parse::adjust_map_after_if.
729 729 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
730 730 ccast->set_req(0, control());
731 731 (*pos_index) = _gvn.transform(ccast);
732 732 }
733 733 return is_neg;
734 734 }
735 735
736 736 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
737 737 Node* *pos_index) {
738 738 if (stopped())
739 739 return NULL; // already stopped
740 740 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
741 741 return NULL; // index is already adequately typed
742 742 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
743 743 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
744 744 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
745 745 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
746 746 if (is_notp != NULL && pos_index != NULL) {
747 747 // Emulate effect of Parse::adjust_map_after_if.
748 748 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
749 749 ccast->set_req(0, control());
750 750 (*pos_index) = _gvn.transform(ccast);
751 751 }
752 752 return is_notp;
753 753 }
754 754
755 755 // Make sure that 'position' is a valid limit index, in [0..length].
756 756 // There are two equivalent plans for checking this:
757 757 // A. (offset + copyLength) unsigned<= arrayLength
758 758 // B. offset <= (arrayLength - copyLength)
759 759 // We require that all of the values above, except for the sum and
760 760 // difference, are already known to be non-negative.
761 761 // Plan A is robust in the face of overflow, if offset and copyLength
762 762 // are both hugely positive.
763 763 //
764 764 // Plan B is less direct and intuitive, but it does not overflow at
765 765 // all, since the difference of two non-negatives is always
766 766 // representable. Whenever Java methods must perform the equivalent
767 767 // check they generally use Plan B instead of Plan A.
768 768 // For the moment we use Plan A.
769 769 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
770 770 Node* subseq_length,
771 771 Node* array_length,
772 772 RegionNode* region) {
773 773 if (stopped())
774 774 return NULL; // already stopped
775 775 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
776 776 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
777 777 return NULL; // common case of whole-array copy
778 778 Node* last = subseq_length;
779 779 if (!zero_offset) // last += offset
780 780 last = _gvn.transform( new (C, 3) AddINode(last, offset));
781 781 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
782 782 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
783 783 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
784 784 return is_over;
785 785 }
786 786
787 787
788 788 //--------------------------generate_current_thread--------------------
789 789 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
790 790 ciKlass* thread_klass = env()->Thread_klass();
791 791 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
792 792 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
793 793 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
794 794 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
795 795 tls_output = thread;
796 796 return threadObj;
797 797 }
798 798
799 799
800 800 //------------------------------make_string_method_node------------------------
801 801 // Helper method for String intrinsic finctions.
802 802 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
803 803 const int value_offset = java_lang_String::value_offset_in_bytes();
804 804 const int count_offset = java_lang_String::count_offset_in_bytes();
805 805 const int offset_offset = java_lang_String::offset_offset_in_bytes();
806 806
807 807 Node* no_ctrl = NULL;
808 808
809 809 ciInstanceKlass* klass = env()->String_klass();
810 810 const TypeInstPtr* string_type =
811 811 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
812 812
813 813 const TypeAryPtr* value_type =
814 814 TypeAryPtr::make(TypePtr::NotNull,
815 815 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
816 816 ciTypeArrayKlass::make(T_CHAR), true, 0);
817 817
818 818 // Get start addr of string and substring
819 819 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
820 820 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
821 821 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
822 822 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
823 823 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
824 824
825 825 // Pin loads from String::equals() argument since it could be NULL.
826 826 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
827 827 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
828 828 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
829 829 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
830 830 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
831 831 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
832 832
833 833 Node* result = NULL;
834 834 switch (opcode) {
835 835 case Op_StrIndexOf:
836 836 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
837 837 str1_start, cnt1, str2_start, cnt2);
838 838 break;
839 839 case Op_StrComp:
840 840 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
841 841 str1_start, cnt1, str2_start, cnt2);
842 842 break;
843 843 case Op_StrEquals:
844 844 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
845 845 str1_start, str2_start, cnt1);
846 846 break;
847 847 default:
848 848 ShouldNotReachHere();
849 849 return NULL;
850 850 }
851 851
852 852 // All these intrinsics have checks.
853 853 C->set_has_split_ifs(true); // Has chance for split-if optimization
854 854
855 855 return _gvn.transform(result);
856 856 }
857 857
858 858 //------------------------------inline_string_compareTo------------------------
859 859 bool LibraryCallKit::inline_string_compareTo() {
860 860
861 861 if (!Matcher::has_match_rule(Op_StrComp)) return false;
862 862
863 863 const int value_offset = java_lang_String::value_offset_in_bytes();
864 864 const int count_offset = java_lang_String::count_offset_in_bytes();
865 865 const int offset_offset = java_lang_String::offset_offset_in_bytes();
866 866
867 867 _sp += 2;
868 868 Node *argument = pop(); // pop non-receiver first: it was pushed second
869 869 Node *receiver = pop();
870 870
871 871 // Null check on self without removing any arguments. The argument
872 872 // null check technically happens in the wrong place, which can lead to
873 873 // invalid stack traces when string compare is inlined into a method
874 874 // which handles NullPointerExceptions.
875 875 _sp += 2;
876 876 receiver = do_null_check(receiver, T_OBJECT);
877 877 argument = do_null_check(argument, T_OBJECT);
878 878 _sp -= 2;
879 879 if (stopped()) {
880 880 return true;
881 881 }
882 882
883 883 ciInstanceKlass* klass = env()->String_klass();
884 884 const TypeInstPtr* string_type =
885 885 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
886 886 Node* no_ctrl = NULL;
887 887
888 888 // Get counts for string and argument
889 889 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
890 890 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
891 891
892 892 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
893 893 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
894 894
895 895 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
896 896 push(compare);
897 897 return true;
898 898 }
899 899
900 900 //------------------------------inline_string_equals------------------------
901 901 bool LibraryCallKit::inline_string_equals() {
902 902
903 903 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
904 904
905 905 const int value_offset = java_lang_String::value_offset_in_bytes();
906 906 const int count_offset = java_lang_String::count_offset_in_bytes();
907 907 const int offset_offset = java_lang_String::offset_offset_in_bytes();
908 908
909 909 _sp += 2;
910 910 Node* argument = pop(); // pop non-receiver first: it was pushed second
911 911 Node* receiver = pop();
912 912
913 913 // Null check on self without removing any arguments. The argument
914 914 // null check technically happens in the wrong place, which can lead to
915 915 // invalid stack traces when string compare is inlined into a method
916 916 // which handles NullPointerExceptions.
917 917 _sp += 2;
918 918 receiver = do_null_check(receiver, T_OBJECT);
919 919 //should not do null check for argument for String.equals(), because spec
920 920 //allows to specify NULL as argument.
921 921 _sp -= 2;
922 922
923 923 if (stopped()) {
924 924 return true;
925 925 }
926 926
927 927 // paths (plus control) merge
928 928 RegionNode* region = new (C, 5) RegionNode(5);
929 929 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
930 930
931 931 // does source == target string?
932 932 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
933 933 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
934 934
935 935 Node* if_eq = generate_slow_guard(bol, NULL);
936 936 if (if_eq != NULL) {
937 937 // receiver == argument
938 938 phi->init_req(2, intcon(1));
939 939 region->init_req(2, if_eq);
940 940 }
941 941
942 942 // get String klass for instanceOf
943 943 ciInstanceKlass* klass = env()->String_klass();
944 944
945 945 if (!stopped()) {
946 946 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
947 947 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
948 948 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
949 949
950 950 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
951 951 //instanceOf == true, fallthrough
952 952
953 953 if (inst_false != NULL) {
954 954 phi->init_req(3, intcon(0));
955 955 region->init_req(3, inst_false);
956 956 }
957 957 }
958 958
959 959 const TypeInstPtr* string_type =
960 960 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
961 961
962 962 Node* no_ctrl = NULL;
963 963 Node* receiver_cnt;
964 964 Node* argument_cnt;
965 965
966 966 if (!stopped()) {
967 967 // Get counts for string and argument
968 968 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
969 969 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
970 970
971 971 // Pin load from argument string since it could be NULL.
972 972 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
973 973 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
974 974
975 975 // Check for receiver count != argument count
976 976 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
977 977 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
978 978 Node* if_ne = generate_slow_guard(bol, NULL);
979 979 if (if_ne != NULL) {
980 980 phi->init_req(4, intcon(0));
981 981 region->init_req(4, if_ne);
982 982 }
983 983 }
984 984
985 985 // Check for count == 0 is done by mach node StrEquals.
986 986
987 987 if (!stopped()) {
988 988 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
989 989 phi->init_req(1, equals);
990 990 region->init_req(1, control());
991 991 }
992 992
993 993 // post merge
994 994 set_control(_gvn.transform(region));
995 995 record_for_igvn(region);
996 996
997 997 push(_gvn.transform(phi));
998 998
999 999 return true;
1000 1000 }
1001 1001
1002 1002 //------------------------------inline_array_equals----------------------------
1003 1003 bool LibraryCallKit::inline_array_equals() {
1004 1004
1005 1005 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1006 1006
1007 1007 _sp += 2;
1008 1008 Node *argument2 = pop();
1009 1009 Node *argument1 = pop();
1010 1010
1011 1011 Node* equals =
1012 1012 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1013 1013 argument1, argument2) );
1014 1014 push(equals);
1015 1015 return true;
1016 1016 }
1017 1017
1018 1018 // Java version of String.indexOf(constant string)
1019 1019 // class StringDecl {
1020 1020 // StringDecl(char[] ca) {
1021 1021 // offset = 0;
1022 1022 // count = ca.length;
1023 1023 // value = ca;
1024 1024 // }
1025 1025 // int offset;
1026 1026 // int count;
1027 1027 // char[] value;
1028 1028 // }
1029 1029 //
1030 1030 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1031 1031 // int targetOffset, int cache_i, int md2) {
1032 1032 // int cache = cache_i;
1033 1033 // int sourceOffset = string_object.offset;
1034 1034 // int sourceCount = string_object.count;
1035 1035 // int targetCount = target_object.length;
1036 1036 //
1037 1037 // int targetCountLess1 = targetCount - 1;
1038 1038 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1039 1039 //
1040 1040 // char[] source = string_object.value;
1041 1041 // char[] target = target_object;
1042 1042 // int lastChar = target[targetCountLess1];
1043 1043 //
1044 1044 // outer_loop:
1045 1045 // for (int i = sourceOffset; i < sourceEnd; ) {
1046 1046 // int src = source[i + targetCountLess1];
1047 1047 // if (src == lastChar) {
1048 1048 // // With random strings and a 4-character alphabet,
1049 1049 // // reverse matching at this point sets up 0.8% fewer
1050 1050 // // frames, but (paradoxically) makes 0.3% more probes.
1051 1051 // // Since those probes are nearer the lastChar probe,
1052 1052 // // there is may be a net D$ win with reverse matching.
1053 1053 // // But, reversing loop inhibits unroll of inner loop
1054 1054 // // for unknown reason. So, does running outer loop from
1055 1055 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1056 1056 // for (int j = 0; j < targetCountLess1; j++) {
1057 1057 // if (target[targetOffset + j] != source[i+j]) {
1058 1058 // if ((cache & (1 << source[i+j])) == 0) {
1059 1059 // if (md2 < j+1) {
1060 1060 // i += j+1;
1061 1061 // continue outer_loop;
1062 1062 // }
1063 1063 // }
1064 1064 // i += md2;
1065 1065 // continue outer_loop;
1066 1066 // }
1067 1067 // }
1068 1068 // return i - sourceOffset;
1069 1069 // }
1070 1070 // if ((cache & (1 << src)) == 0) {
1071 1071 // i += targetCountLess1;
1072 1072 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1073 1073 // i++;
1074 1074 // }
1075 1075 // return -1;
1076 1076 // }
1077 1077
1078 1078 //------------------------------string_indexOf------------------------
1079 1079 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1080 1080 jint cache_i, jint md2_i) {
1081 1081
1082 1082 Node* no_ctrl = NULL;
1083 1083 float likely = PROB_LIKELY(0.9);
1084 1084 float unlikely = PROB_UNLIKELY(0.9);
1085 1085
1086 1086 const int value_offset = java_lang_String::value_offset_in_bytes();
1087 1087 const int count_offset = java_lang_String::count_offset_in_bytes();
1088 1088 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1089 1089
1090 1090 ciInstanceKlass* klass = env()->String_klass();
1091 1091 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1092 1092 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1093 1093
1094 1094 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1095 1095 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1096 1096 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1097 1097 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1098 1098 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1099 1099 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1100 1100
1101 1101 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
1102 1102 jint target_length = target_array->length();
1103 1103 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1104 1104 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1105 1105
1106 1106 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1107 1107 #define __ kit.
1108 1108 Node* zero = __ ConI(0);
1109 1109 Node* one = __ ConI(1);
1110 1110 Node* cache = __ ConI(cache_i);
1111 1111 Node* md2 = __ ConI(md2_i);
1112 1112 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1113 1113 Node* targetCount = __ ConI(target_length);
1114 1114 Node* targetCountLess1 = __ ConI(target_length - 1);
1115 1115 Node* targetOffset = __ ConI(targetOffset_i);
1116 1116 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1117 1117
1118 1118 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1119 1119 Node* outer_loop = __ make_label(2 /* goto */);
1120 1120 Node* return_ = __ make_label(1);
1121 1121
1122 1122 __ set(rtn,__ ConI(-1));
1123 1123 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
1124 1124 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1125 1125 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1126 1126 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1127 1127 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1128 1128 __ loop(j, zero, BoolTest::lt, targetCountLess1); {
1129 1129 Node* tpj = __ AddI(targetOffset, __ value(j));
1130 1130 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1131 1131 Node* ipj = __ AddI(__ value(i), __ value(j));
1132 1132 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1133 1133 __ if_then(targ, BoolTest::ne, src2); {
1134 1134 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1135 1135 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1136 1136 __ increment(i, __ AddI(__ value(j), one));
1137 1137 __ goto_(outer_loop);
1138 1138 } __ end_if(); __ dead(j);
1139 1139 }__ end_if(); __ dead(j);
1140 1140 __ increment(i, md2);
1141 1141 __ goto_(outer_loop);
1142 1142 }__ end_if();
1143 1143 __ increment(j, one);
1144 1144 }__ end_loop(); __ dead(j);
1145 1145 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1146 1146 __ goto_(return_);
1147 1147 }__ end_if();
1148 1148 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1149 1149 __ increment(i, targetCountLess1);
1150 1150 }__ end_if();
1151 1151 __ increment(i, one);
1152 1152 __ bind(outer_loop);
1153 1153 }__ end_loop(); __ dead(i);
1154 1154 __ bind(return_);
1155 1155
1156 1156 // Final sync IdealKit and GraphKit.
1157 1157 sync_kit(kit);
1158 1158 Node* result = __ value(rtn);
1159 1159 #undef __
1160 1160 C->set_has_loops(true);
1161 1161 return result;
1162 1162 }
1163 1163
1164 1164 //------------------------------inline_string_indexOf------------------------
1165 1165 bool LibraryCallKit::inline_string_indexOf() {
1166 1166
1167 1167 const int value_offset = java_lang_String::value_offset_in_bytes();
1168 1168 const int count_offset = java_lang_String::count_offset_in_bytes();
1169 1169 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1170 1170
1171 1171 _sp += 2;
1172 1172 Node *argument = pop(); // pop non-receiver first: it was pushed second
1173 1173 Node *receiver = pop();
1174 1174
1175 1175 Node* result;
1176 1176 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1177 1177 UseSSE42Intrinsics) {
1178 1178 // Generate SSE4.2 version of indexOf
1179 1179 // We currently only have match rules that use SSE4.2
1180 1180
1181 1181 // Null check on self without removing any arguments. The argument
1182 1182 // null check technically happens in the wrong place, which can lead to
1183 1183 // invalid stack traces when string compare is inlined into a method
1184 1184 // which handles NullPointerExceptions.
1185 1185 _sp += 2;
1186 1186 receiver = do_null_check(receiver, T_OBJECT);
1187 1187 argument = do_null_check(argument, T_OBJECT);
1188 1188 _sp -= 2;
1189 1189
1190 1190 if (stopped()) {
1191 1191 return true;
1192 1192 }
1193 1193
1194 1194 // Make the merge point
1195 1195 RegionNode* result_rgn = new (C, 3) RegionNode(3);
1196 1196 Node* result_phi = new (C, 3) PhiNode(result_rgn, TypeInt::INT);
1197 1197 Node* no_ctrl = NULL;
1198 1198
1199 1199 ciInstanceKlass* klass = env()->String_klass();
1200 1200 const TypeInstPtr* string_type =
1201 1201 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
1202 1202
1203 1203 // Get counts for string and substr
1204 1204 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1205 1205 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1206 1206
1207 1207 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1208 1208 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1209 1209
1210 1210 // Check for substr count > string count
1211 1211 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1212 1212 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1213 1213 Node* if_gt = generate_slow_guard(bol, NULL);
1214 1214 if (if_gt != NULL) {
1215 1215 result_phi->init_req(2, intcon(-1));
1216 1216 result_rgn->init_req(2, if_gt);
1217 1217 }
1218 1218
1219 1219 if (!stopped()) {
1220 1220 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1221 1221 result_phi->init_req(1, result);
1222 1222 result_rgn->init_req(1, control());
1223 1223 }
1224 1224 set_control(_gvn.transform(result_rgn));
1225 1225 record_for_igvn(result_rgn);
1226 1226 result = _gvn.transform(result_phi);
1227 1227
1228 1228 } else { //Use LibraryCallKit::string_indexOf
1229 1229 // don't intrinsify is argument isn't a constant string.
1230 1230 if (!argument->is_Con()) {
1231 1231 return false;
1232 1232 }
1233 1233 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1234 1234 if (str_type == NULL) {
1235 1235 return false;
1236 1236 }
1237 1237 ciInstanceKlass* klass = env()->String_klass();
1238 1238 ciObject* str_const = str_type->const_oop();
1239 1239 if (str_const == NULL || str_const->klass() != klass) {
1240 1240 return false;
1241 1241 }
1242 1242 ciInstance* str = str_const->as_instance();
1243 1243 assert(str != NULL, "must be instance");
1244 1244
1245 1245 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1246 1246 int o = str->field_value_by_offset(offset_offset).as_int();
1247 1247 int c = str->field_value_by_offset(count_offset).as_int();
1248 1248 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1249 1249
1250 1250 // constant strings have no offset and count == length which
1251 1251 // simplifies the resulting code somewhat so lets optimize for that.
1252 1252 if (o != 0 || c != pat->length()) {
1253 1253 return false;
1254 1254 }
1255 1255
1256 1256 // Null check on self without removing any arguments. The argument
1257 1257 // null check technically happens in the wrong place, which can lead to
1258 1258 // invalid stack traces when string compare is inlined into a method
1259 1259 // which handles NullPointerExceptions.
1260 1260 _sp += 2;
1261 1261 receiver = do_null_check(receiver, T_OBJECT);
1262 1262 // No null check on the argument is needed since it's a constant String oop.
1263 1263 _sp -= 2;
1264 1264 if (stopped()) {
1265 1265 return true;
1266 1266 }
1267 1267
1268 1268 // The null string as a pattern always returns 0 (match at beginning of string)
1269 1269 if (c == 0) {
1270 1270 push(intcon(0));
1271 1271 return true;
1272 1272 }
1273 1273
1274 1274 // Generate default indexOf
1275 1275 jchar lastChar = pat->char_at(o + (c - 1));
1276 1276 int cache = 0;
1277 1277 int i;
1278 1278 for (i = 0; i < c - 1; i++) {
1279 1279 assert(i < pat->length(), "out of range");
1280 1280 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1281 1281 }
1282 1282
1283 1283 int md2 = c;
1284 1284 for (i = 0; i < c - 1; i++) {
1285 1285 assert(i < pat->length(), "out of range");
1286 1286 if (pat->char_at(o + i) == lastChar) {
1287 1287 md2 = (c - 1) - i;
1288 1288 }
1289 1289 }
1290 1290
1291 1291 result = string_indexOf(receiver, pat, o, cache, md2);
1292 1292 }
1293 1293
1294 1294 push(result);
1295 1295 return true;
1296 1296 }
1297 1297
1298 1298 //--------------------------pop_math_arg--------------------------------
1299 1299 // Pop a double argument to a math function from the stack
1300 1300 // rounding it if necessary.
1301 1301 Node * LibraryCallKit::pop_math_arg() {
1302 1302 Node *arg = pop_pair();
1303 1303 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1304 1304 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1305 1305 return arg;
1306 1306 }
1307 1307
1308 1308 //------------------------------inline_trig----------------------------------
1309 1309 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1310 1310 // argument reduction which will turn into a fast/slow diamond.
1311 1311 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1312 1312 _sp += arg_size(); // restore stack pointer
1313 1313 Node* arg = pop_math_arg();
1314 1314 Node* trig = NULL;
1315 1315
1316 1316 switch (id) {
1317 1317 case vmIntrinsics::_dsin:
1318 1318 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1319 1319 break;
1320 1320 case vmIntrinsics::_dcos:
1321 1321 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1322 1322 break;
1323 1323 case vmIntrinsics::_dtan:
1324 1324 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1325 1325 break;
1326 1326 default:
1327 1327 assert(false, "bad intrinsic was passed in");
1328 1328 return false;
1329 1329 }
1330 1330
1331 1331 // Rounding required? Check for argument reduction!
1332 1332 if( Matcher::strict_fp_requires_explicit_rounding ) {
1333 1333
1334 1334 static const double pi_4 = 0.7853981633974483;
1335 1335 static const double neg_pi_4 = -0.7853981633974483;
1336 1336 // pi/2 in 80-bit extended precision
1337 1337 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1338 1338 // -pi/2 in 80-bit extended precision
1339 1339 // 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};
1340 1340 // Cutoff value for using this argument reduction technique
1341 1341 //static const double pi_2_minus_epsilon = 1.564660403643354;
1342 1342 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1343 1343
1344 1344 // Pseudocode for sin:
1345 1345 // if (x <= Math.PI / 4.0) {
1346 1346 // if (x >= -Math.PI / 4.0) return fsin(x);
1347 1347 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1348 1348 // } else {
1349 1349 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1350 1350 // }
1351 1351 // return StrictMath.sin(x);
1352 1352
1353 1353 // Pseudocode for cos:
1354 1354 // if (x <= Math.PI / 4.0) {
1355 1355 // if (x >= -Math.PI / 4.0) return fcos(x);
1356 1356 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1357 1357 // } else {
1358 1358 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1359 1359 // }
1360 1360 // return StrictMath.cos(x);
1361 1361
1362 1362 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1363 1363 // requires a special machine instruction to load it. Instead we'll try
1364 1364 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1365 1365 // probably do the math inside the SIN encoding.
1366 1366
1367 1367 // Make the merge point
1368 1368 RegionNode *r = new (C, 3) RegionNode(3);
1369 1369 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1370 1370
1371 1371 // Flatten arg so we need only 1 test
1372 1372 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1373 1373 // Node for PI/4 constant
1374 1374 Node *pi4 = makecon(TypeD::make(pi_4));
1375 1375 // Check PI/4 : abs(arg)
1376 1376 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1377 1377 // Check: If PI/4 < abs(arg) then go slow
1378 1378 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1379 1379 // Branch either way
1380 1380 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1381 1381 set_control(opt_iff(r,iff));
1382 1382
1383 1383 // Set fast path result
1384 1384 phi->init_req(2,trig);
1385 1385
1386 1386 // Slow path - non-blocking leaf call
1387 1387 Node* call = NULL;
1388 1388 switch (id) {
1389 1389 case vmIntrinsics::_dsin:
1390 1390 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1391 1391 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1392 1392 "Sin", NULL, arg, top());
1393 1393 break;
1394 1394 case vmIntrinsics::_dcos:
1395 1395 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1396 1396 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1397 1397 "Cos", NULL, arg, top());
1398 1398 break;
1399 1399 case vmIntrinsics::_dtan:
1400 1400 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1401 1401 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1402 1402 "Tan", NULL, arg, top());
1403 1403 break;
1404 1404 }
1405 1405 assert(control()->in(0) == call, "");
1406 1406 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1407 1407 r->init_req(1,control());
1408 1408 phi->init_req(1,slow_result);
1409 1409
1410 1410 // Post-merge
1411 1411 set_control(_gvn.transform(r));
1412 1412 record_for_igvn(r);
1413 1413 trig = _gvn.transform(phi);
1414 1414
1415 1415 C->set_has_split_ifs(true); // Has chance for split-if optimization
1416 1416 }
1417 1417 // Push result back on JVM stack
1418 1418 push_pair(trig);
1419 1419 return true;
1420 1420 }
1421 1421
1422 1422 //------------------------------inline_sqrt-------------------------------------
1423 1423 // Inline square root instruction, if possible.
1424 1424 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1425 1425 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1426 1426 _sp += arg_size(); // restore stack pointer
1427 1427 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1428 1428 return true;
1429 1429 }
1430 1430
1431 1431 //------------------------------inline_abs-------------------------------------
1432 1432 // Inline absolute value instruction, if possible.
1433 1433 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1434 1434 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1435 1435 _sp += arg_size(); // restore stack pointer
1436 1436 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1437 1437 return true;
1438 1438 }
1439 1439
1440 1440 //------------------------------inline_exp-------------------------------------
1441 1441 // Inline exp instructions, if possible. The Intel hardware only misses
1442 1442 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1443 1443 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1444 1444 assert(id == vmIntrinsics::_dexp, "Not exp");
1445 1445
1446 1446 // If this inlining ever returned NaN in the past, we do not intrinsify it
1447 1447 // every again. NaN results requires StrictMath.exp handling.
1448 1448 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1449 1449
1450 1450 // Do not intrinsify on older platforms which lack cmove.
1451 1451 if (ConditionalMoveLimit == 0) return false;
1452 1452
1453 1453 _sp += arg_size(); // restore stack pointer
1454 1454 Node *x = pop_math_arg();
1455 1455 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1456 1456
1457 1457 //-------------------
1458 1458 //result=(result.isNaN())? StrictMath::exp():result;
1459 1459 // Check: If isNaN() by checking result!=result? then go to Strict Math
1460 1460 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1461 1461 // Build the boolean node
1462 1462 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1463 1463
1464 1464 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1465 1465 // End the current control-flow path
1466 1466 push_pair(x);
1467 1467 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1468 1468 // to handle. Recompile without intrinsifying Math.exp
1469 1469 uncommon_trap(Deoptimization::Reason_intrinsic,
1470 1470 Deoptimization::Action_make_not_entrant);
1471 1471 }
1472 1472
1473 1473 C->set_has_split_ifs(true); // Has chance for split-if optimization
1474 1474
1475 1475 push_pair(result);
1476 1476
1477 1477 return true;
1478 1478 }
1479 1479
1480 1480 //------------------------------inline_pow-------------------------------------
1481 1481 // Inline power instructions, if possible.
1482 1482 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1483 1483 assert(id == vmIntrinsics::_dpow, "Not pow");
1484 1484
1485 1485 // If this inlining ever returned NaN in the past, we do not intrinsify it
1486 1486 // every again. NaN results requires StrictMath.pow handling.
1487 1487 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1488 1488
1489 1489 // Do not intrinsify on older platforms which lack cmove.
1490 1490 if (ConditionalMoveLimit == 0) return false;
1491 1491
1492 1492 // Pseudocode for pow
1493 1493 // if (x <= 0.0) {
1494 1494 // if ((double)((int)y)==y) { // if y is int
1495 1495 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1496 1496 // } else {
1497 1497 // result = NaN;
1498 1498 // }
1499 1499 // } else {
1500 1500 // result = DPow(x,y);
1501 1501 // }
1502 1502 // if (result != result)? {
1503 1503 // uncommon_trap();
1504 1504 // }
1505 1505 // return result;
1506 1506
1507 1507 _sp += arg_size(); // restore stack pointer
1508 1508 Node* y = pop_math_arg();
1509 1509 Node* x = pop_math_arg();
1510 1510
1511 1511 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1512 1512
1513 1513 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1514 1514 // inside of something) then skip the fancy tests and just check for
1515 1515 // NaN result.
1516 1516 Node *result = NULL;
1517 1517 if( jvms()->depth() >= 1 ) {
1518 1518 result = fast_result;
1519 1519 } else {
1520 1520
1521 1521 // Set the merge point for If node with condition of (x <= 0.0)
1522 1522 // There are four possible paths to region node and phi node
1523 1523 RegionNode *r = new (C, 4) RegionNode(4);
1524 1524 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1525 1525
1526 1526 // Build the first if node: if (x <= 0.0)
1527 1527 // Node for 0 constant
1528 1528 Node *zeronode = makecon(TypeD::ZERO);
1529 1529 // Check x:0
1530 1530 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1531 1531 // Check: If (x<=0) then go complex path
1532 1532 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1533 1533 // Branch either way
1534 1534 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1535 1535 Node *opt_test = _gvn.transform(if1);
1536 1536 //assert( opt_test->is_If(), "Expect an IfNode");
1537 1537 IfNode *opt_if1 = (IfNode*)opt_test;
1538 1538 // Fast path taken; set region slot 3
1539 1539 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1540 1540 r->init_req(3,fast_taken); // Capture fast-control
1541 1541
1542 1542 // Fast path not-taken, i.e. slow path
1543 1543 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1544 1544
1545 1545 // Set fast path result
1546 1546 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1547 1547 phi->init_req(3, fast_result);
1548 1548
1549 1549 // Complex path
1550 1550 // Build the second if node (if y is int)
1551 1551 // Node for (int)y
1552 1552 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1553 1553 // Node for (double)((int) y)
1554 1554 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1555 1555 // Check (double)((int) y) : y
1556 1556 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1557 1557 // Check if (y isn't int) then go to slow path
1558 1558
1559 1559 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1560 1560 // Branch either way
1561 1561 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1562 1562 Node *slow_path = opt_iff(r,if2); // Set region path 2
1563 1563
1564 1564 // Calculate DPow(abs(x), y)*(1 & (int)y)
1565 1565 // Node for constant 1
1566 1566 Node *conone = intcon(1);
1567 1567 // 1& (int)y
1568 1568 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1569 1569 // zero node
1570 1570 Node *conzero = intcon(0);
1571 1571 // Check (1&(int)y)==0?
1572 1572 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1573 1573 // Check if (1&(int)y)!=0?, if so the result is negative
1574 1574 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1575 1575 // abs(x)
1576 1576 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1577 1577 // abs(x)^y
1578 1578 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1579 1579 // -abs(x)^y
1580 1580 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1581 1581 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1582 1582 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1583 1583 // Set complex path fast result
1584 1584 phi->init_req(2, signresult);
1585 1585
1586 1586 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1587 1587 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1588 1588 r->init_req(1,slow_path);
1589 1589 phi->init_req(1,slow_result);
1590 1590
1591 1591 // Post merge
1592 1592 set_control(_gvn.transform(r));
1593 1593 record_for_igvn(r);
1594 1594 result=_gvn.transform(phi);
1595 1595 }
1596 1596
1597 1597 //-------------------
1598 1598 //result=(result.isNaN())? uncommon_trap():result;
1599 1599 // Check: If isNaN() by checking result!=result? then go to Strict Math
1600 1600 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1601 1601 // Build the boolean node
1602 1602 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1603 1603
1604 1604 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1605 1605 // End the current control-flow path
1606 1606 push_pair(x);
1607 1607 push_pair(y);
1608 1608 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1609 1609 // to handle. Recompile without intrinsifying Math.pow.
1610 1610 uncommon_trap(Deoptimization::Reason_intrinsic,
1611 1611 Deoptimization::Action_make_not_entrant);
1612 1612 }
1613 1613
1614 1614 C->set_has_split_ifs(true); // Has chance for split-if optimization
1615 1615
1616 1616 push_pair(result);
1617 1617
1618 1618 return true;
1619 1619 }
1620 1620
1621 1621 //------------------------------inline_trans-------------------------------------
1622 1622 // Inline transcendental instructions, if possible. The Intel hardware gets
1623 1623 // these right, no funny corner cases missed.
1624 1624 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1625 1625 _sp += arg_size(); // restore stack pointer
1626 1626 Node* arg = pop_math_arg();
1627 1627 Node* trans = NULL;
1628 1628
1629 1629 switch (id) {
1630 1630 case vmIntrinsics::_dlog:
1631 1631 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1632 1632 break;
1633 1633 case vmIntrinsics::_dlog10:
1634 1634 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1635 1635 break;
1636 1636 default:
1637 1637 assert(false, "bad intrinsic was passed in");
1638 1638 return false;
1639 1639 }
1640 1640
1641 1641 // Push result back on JVM stack
1642 1642 push_pair(trans);
1643 1643 return true;
1644 1644 }
1645 1645
1646 1646 //------------------------------runtime_math-----------------------------
1647 1647 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1648 1648 Node* a = NULL;
1649 1649 Node* b = NULL;
1650 1650
1651 1651 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1652 1652 "must be (DD)D or (D)D type");
1653 1653
1654 1654 // Inputs
1655 1655 _sp += arg_size(); // restore stack pointer
1656 1656 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1657 1657 b = pop_math_arg();
1658 1658 }
1659 1659 a = pop_math_arg();
1660 1660
1661 1661 const TypePtr* no_memory_effects = NULL;
1662 1662 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1663 1663 no_memory_effects,
1664 1664 a, top(), b, b ? top() : NULL);
1665 1665 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1666 1666 #ifdef ASSERT
1667 1667 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1668 1668 assert(value_top == top(), "second value must be top");
1669 1669 #endif
1670 1670
1671 1671 push_pair(value);
1672 1672 return true;
1673 1673 }
1674 1674
1675 1675 //------------------------------inline_math_native-----------------------------
1676 1676 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1677 1677 switch (id) {
1678 1678 // These intrinsics are not properly supported on all hardware
1679 1679 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1680 1680 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1681 1681 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1682 1682 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1683 1683 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1684 1684 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1685 1685
1686 1686 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1687 1687 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1688 1688 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1689 1689 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1690 1690
1691 1691 // These intrinsics are supported on all hardware
1692 1692 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1693 1693 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1694 1694
1695 1695 // These intrinsics don't work on X86. The ad implementation doesn't
1696 1696 // handle NaN's properly. Instead of returning infinity, the ad
1697 1697 // implementation returns a NaN on overflow. See bug: 6304089
1698 1698 // Once the ad implementations are fixed, change the code below
1699 1699 // to match the intrinsics above
1700 1700
1701 1701 case vmIntrinsics::_dexp: return
1702 1702 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1703 1703 case vmIntrinsics::_dpow: return
1704 1704 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1705 1705
1706 1706 // These intrinsics are not yet correctly implemented
1707 1707 case vmIntrinsics::_datan2:
1708 1708 return false;
1709 1709
1710 1710 default:
1711 1711 ShouldNotReachHere();
1712 1712 return false;
1713 1713 }
1714 1714 }
1715 1715
1716 1716 static bool is_simple_name(Node* n) {
1717 1717 return (n->req() == 1 // constant
1718 1718 || (n->is_Type() && n->as_Type()->type()->singleton())
1719 1719 || n->is_Proj() // parameter or return value
1720 1720 || n->is_Phi() // local of some sort
1721 1721 );
1722 1722 }
1723 1723
1724 1724 //----------------------------inline_min_max-----------------------------------
1725 1725 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1726 1726 push(generate_min_max(id, argument(0), argument(1)));
1727 1727
1728 1728 return true;
1729 1729 }
1730 1730
1731 1731 Node*
1732 1732 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1733 1733 // These are the candidate return value:
1734 1734 Node* xvalue = x0;
1735 1735 Node* yvalue = y0;
1736 1736
1737 1737 if (xvalue == yvalue) {
1738 1738 return xvalue;
1739 1739 }
1740 1740
1741 1741 bool want_max = (id == vmIntrinsics::_max);
1742 1742
1743 1743 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1744 1744 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1745 1745 if (txvalue == NULL || tyvalue == NULL) return top();
1746 1746 // This is not really necessary, but it is consistent with a
1747 1747 // hypothetical MaxINode::Value method:
1748 1748 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1749 1749
1750 1750 // %%% This folding logic should (ideally) be in a different place.
1751 1751 // Some should be inside IfNode, and there to be a more reliable
1752 1752 // transformation of ?: style patterns into cmoves. We also want
1753 1753 // more powerful optimizations around cmove and min/max.
1754 1754
1755 1755 // Try to find a dominating comparison of these guys.
1756 1756 // It can simplify the index computation for Arrays.copyOf
1757 1757 // and similar uses of System.arraycopy.
1758 1758 // First, compute the normalized version of CmpI(x, y).
1759 1759 int cmp_op = Op_CmpI;
1760 1760 Node* xkey = xvalue;
1761 1761 Node* ykey = yvalue;
1762 1762 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1763 1763 if (ideal_cmpxy->is_Cmp()) {
1764 1764 // E.g., if we have CmpI(length - offset, count),
1765 1765 // it might idealize to CmpI(length, count + offset)
1766 1766 cmp_op = ideal_cmpxy->Opcode();
1767 1767 xkey = ideal_cmpxy->in(1);
1768 1768 ykey = ideal_cmpxy->in(2);
1769 1769 }
1770 1770
1771 1771 // Start by locating any relevant comparisons.
1772 1772 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1773 1773 Node* cmpxy = NULL;
1774 1774 Node* cmpyx = NULL;
1775 1775 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1776 1776 Node* cmp = start_from->fast_out(k);
1777 1777 if (cmp->outcnt() > 0 && // must have prior uses
1778 1778 cmp->in(0) == NULL && // must be context-independent
1779 1779 cmp->Opcode() == cmp_op) { // right kind of compare
1780 1780 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1781 1781 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1782 1782 }
1783 1783 }
1784 1784
1785 1785 const int NCMPS = 2;
1786 1786 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1787 1787 int cmpn;
1788 1788 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1789 1789 if (cmps[cmpn] != NULL) break; // find a result
1790 1790 }
1791 1791 if (cmpn < NCMPS) {
1792 1792 // Look for a dominating test that tells us the min and max.
1793 1793 int depth = 0; // Limit search depth for speed
1794 1794 Node* dom = control();
1795 1795 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1796 1796 if (++depth >= 100) break;
1797 1797 Node* ifproj = dom;
1798 1798 if (!ifproj->is_Proj()) continue;
1799 1799 Node* iff = ifproj->in(0);
1800 1800 if (!iff->is_If()) continue;
1801 1801 Node* bol = iff->in(1);
1802 1802 if (!bol->is_Bool()) continue;
1803 1803 Node* cmp = bol->in(1);
1804 1804 if (cmp == NULL) continue;
1805 1805 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1806 1806 if (cmps[cmpn] == cmp) break;
1807 1807 if (cmpn == NCMPS) continue;
1808 1808 BoolTest::mask btest = bol->as_Bool()->_test._test;
1809 1809 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1810 1810 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1811 1811 // At this point, we know that 'x btest y' is true.
1812 1812 switch (btest) {
1813 1813 case BoolTest::eq:
1814 1814 // They are proven equal, so we can collapse the min/max.
1815 1815 // Either value is the answer. Choose the simpler.
1816 1816 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1817 1817 return yvalue;
1818 1818 return xvalue;
1819 1819 case BoolTest::lt: // x < y
1820 1820 case BoolTest::le: // x <= y
1821 1821 return (want_max ? yvalue : xvalue);
1822 1822 case BoolTest::gt: // x > y
1823 1823 case BoolTest::ge: // x >= y
1824 1824 return (want_max ? xvalue : yvalue);
1825 1825 }
1826 1826 }
1827 1827 }
1828 1828
1829 1829 // We failed to find a dominating test.
1830 1830 // Let's pick a test that might GVN with prior tests.
1831 1831 Node* best_bol = NULL;
1832 1832 BoolTest::mask best_btest = BoolTest::illegal;
1833 1833 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1834 1834 Node* cmp = cmps[cmpn];
1835 1835 if (cmp == NULL) continue;
1836 1836 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1837 1837 Node* bol = cmp->fast_out(j);
1838 1838 if (!bol->is_Bool()) continue;
1839 1839 BoolTest::mask btest = bol->as_Bool()->_test._test;
1840 1840 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1841 1841 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1842 1842 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1843 1843 best_bol = bol->as_Bool();
1844 1844 best_btest = btest;
1845 1845 }
1846 1846 }
1847 1847 }
1848 1848
1849 1849 Node* answer_if_true = NULL;
1850 1850 Node* answer_if_false = NULL;
1851 1851 switch (best_btest) {
1852 1852 default:
1853 1853 if (cmpxy == NULL)
1854 1854 cmpxy = ideal_cmpxy;
1855 1855 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1856 1856 // and fall through:
1857 1857 case BoolTest::lt: // x < y
1858 1858 case BoolTest::le: // x <= y
1859 1859 answer_if_true = (want_max ? yvalue : xvalue);
1860 1860 answer_if_false = (want_max ? xvalue : yvalue);
1861 1861 break;
1862 1862 case BoolTest::gt: // x > y
1863 1863 case BoolTest::ge: // x >= y
1864 1864 answer_if_true = (want_max ? xvalue : yvalue);
1865 1865 answer_if_false = (want_max ? yvalue : xvalue);
1866 1866 break;
1867 1867 }
1868 1868
1869 1869 jint hi, lo;
1870 1870 if (want_max) {
1871 1871 // We can sharpen the minimum.
1872 1872 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1873 1873 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1874 1874 } else {
1875 1875 // We can sharpen the maximum.
1876 1876 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1877 1877 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1878 1878 }
1879 1879
1880 1880 // Use a flow-free graph structure, to avoid creating excess control edges
1881 1881 // which could hinder other optimizations.
1882 1882 // Since Math.min/max is often used with arraycopy, we want
1883 1883 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1884 1884 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1885 1885 answer_if_false, answer_if_true,
1886 1886 TypeInt::make(lo, hi, widen));
1887 1887
1888 1888 return _gvn.transform(cmov);
1889 1889
1890 1890 /*
1891 1891 // This is not as desirable as it may seem, since Min and Max
1892 1892 // nodes do not have a full set of optimizations.
1893 1893 // And they would interfere, anyway, with 'if' optimizations
1894 1894 // and with CMoveI canonical forms.
1895 1895 switch (id) {
1896 1896 case vmIntrinsics::_min:
1897 1897 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1898 1898 case vmIntrinsics::_max:
1899 1899 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1900 1900 default:
1901 1901 ShouldNotReachHere();
1902 1902 }
1903 1903 */
1904 1904 }
1905 1905
1906 1906 inline int
1907 1907 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1908 1908 const TypePtr* base_type = TypePtr::NULL_PTR;
1909 1909 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1910 1910 if (base_type == NULL) {
1911 1911 // Unknown type.
1912 1912 return Type::AnyPtr;
1913 1913 } else if (base_type == TypePtr::NULL_PTR) {
1914 1914 // Since this is a NULL+long form, we have to switch to a rawptr.
1915 1915 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1916 1916 offset = MakeConX(0);
1917 1917 return Type::RawPtr;
1918 1918 } else if (base_type->base() == Type::RawPtr) {
1919 1919 return Type::RawPtr;
1920 1920 } else if (base_type->isa_oopptr()) {
1921 1921 // Base is never null => always a heap address.
1922 1922 if (base_type->ptr() == TypePtr::NotNull) {
1923 1923 return Type::OopPtr;
1924 1924 }
1925 1925 // Offset is small => always a heap address.
1926 1926 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1927 1927 if (offset_type != NULL &&
1928 1928 base_type->offset() == 0 && // (should always be?)
1929 1929 offset_type->_lo >= 0 &&
1930 1930 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1931 1931 return Type::OopPtr;
1932 1932 }
1933 1933 // Otherwise, it might either be oop+off or NULL+addr.
1934 1934 return Type::AnyPtr;
1935 1935 } else {
1936 1936 // No information:
1937 1937 return Type::AnyPtr;
1938 1938 }
1939 1939 }
1940 1940
1941 1941 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1942 1942 int kind = classify_unsafe_addr(base, offset);
1943 1943 if (kind == Type::RawPtr) {
1944 1944 return basic_plus_adr(top(), base, offset);
1945 1945 } else {
1946 1946 return basic_plus_adr(base, offset);
1947 1947 }
1948 1948 }
1949 1949
1950 1950 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1951 1951 // inline int Integer.numberOfLeadingZeros(int)
1952 1952 // inline int Long.numberOfLeadingZeros(long)
1953 1953 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1954 1954 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1955 1955 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1956 1956 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1957 1957 _sp += arg_size(); // restore stack pointer
1958 1958 switch (id) {
1959 1959 case vmIntrinsics::_numberOfLeadingZeros_i:
1960 1960 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1961 1961 break;
1962 1962 case vmIntrinsics::_numberOfLeadingZeros_l:
1963 1963 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1964 1964 break;
1965 1965 default:
1966 1966 ShouldNotReachHere();
1967 1967 }
1968 1968 return true;
1969 1969 }
1970 1970
1971 1971 //-------------------inline_numberOfTrailingZeros_int/long----------------------
1972 1972 // inline int Integer.numberOfTrailingZeros(int)
1973 1973 // inline int Long.numberOfTrailingZeros(long)
1974 1974 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
1975 1975 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
1976 1976 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
1977 1977 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
1978 1978 _sp += arg_size(); // restore stack pointer
1979 1979 switch (id) {
1980 1980 case vmIntrinsics::_numberOfTrailingZeros_i:
1981 1981 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
1982 1982 break;
1983 1983 case vmIntrinsics::_numberOfTrailingZeros_l:
1984 1984 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
1985 1985 break;
1986 1986 default:
1987 1987 ShouldNotReachHere();
1988 1988 }
1989 1989 return true;
1990 1990 }
1991 1991
1992 1992 //----------------------------inline_bitCount_int/long-----------------------
1993 1993 // inline int Integer.bitCount(int)
1994 1994 // inline int Long.bitCount(long)
1995 1995 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
1996 1996 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
1997 1997 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
1998 1998 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
1999 1999 _sp += arg_size(); // restore stack pointer
2000 2000 switch (id) {
2001 2001 case vmIntrinsics::_bitCount_i:
2002 2002 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2003 2003 break;
2004 2004 case vmIntrinsics::_bitCount_l:
2005 2005 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2006 2006 break;
2007 2007 default:
2008 2008 ShouldNotReachHere();
2009 2009 }
2010 2010 return true;
2011 2011 }
2012 2012
2013 2013 //----------------------------inline_reverseBytes_int/long-------------------
2014 2014 // inline Integer.reverseBytes(int)
2015 2015 // inline Long.reverseBytes(long)
2016 2016 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2017 2017 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
2018 2018 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2019 2019 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2020 2020 _sp += arg_size(); // restore stack pointer
2021 2021 switch (id) {
2022 2022 case vmIntrinsics::_reverseBytes_i:
2023 2023 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2024 2024 break;
2025 2025 case vmIntrinsics::_reverseBytes_l:
2026 2026 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2027 2027 break;
2028 2028 default:
2029 2029 ;
2030 2030 }
2031 2031 return true;
2032 2032 }
2033 2033
2034 2034 //----------------------------inline_unsafe_access----------------------------
2035 2035
2036 2036 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2037 2037
2038 2038 // Interpret Unsafe.fieldOffset cookies correctly:
2039 2039 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2040 2040
2041 2041 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2042 2042 if (callee()->is_static()) return false; // caller must have the capability!
2043 2043
2044 2044 #ifndef PRODUCT
2045 2045 {
2046 2046 ResourceMark rm;
2047 2047 // Check the signatures.
2048 2048 ciSignature* sig = signature();
2049 2049 #ifdef ASSERT
2050 2050 if (!is_store) {
2051 2051 // Object getObject(Object base, int/long offset), etc.
2052 2052 BasicType rtype = sig->return_type()->basic_type();
2053 2053 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2054 2054 rtype = T_ADDRESS; // it is really a C void*
2055 2055 assert(rtype == type, "getter must return the expected value");
2056 2056 if (!is_native_ptr) {
2057 2057 assert(sig->count() == 2, "oop getter has 2 arguments");
2058 2058 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2059 2059 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2060 2060 } else {
2061 2061 assert(sig->count() == 1, "native getter has 1 argument");
2062 2062 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2063 2063 }
2064 2064 } else {
2065 2065 // void putObject(Object base, int/long offset, Object x), etc.
2066 2066 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2067 2067 if (!is_native_ptr) {
2068 2068 assert(sig->count() == 3, "oop putter has 3 arguments");
2069 2069 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2070 2070 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2071 2071 } else {
2072 2072 assert(sig->count() == 2, "native putter has 2 arguments");
2073 2073 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2074 2074 }
2075 2075 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2076 2076 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2077 2077 vtype = T_ADDRESS; // it is really a C void*
2078 2078 assert(vtype == type, "putter must accept the expected value");
2079 2079 }
2080 2080 #endif // ASSERT
2081 2081 }
2082 2082 #endif //PRODUCT
2083 2083
2084 2084 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2085 2085
2086 2086 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2087 2087
2088 2088 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2089 2089 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2090 2090
2091 2091 debug_only(int saved_sp = _sp);
2092 2092 _sp += nargs;
2093 2093
2094 2094 Node* val;
2095 2095 debug_only(val = (Node*)(uintptr_t)-1);
2096 2096
2097 2097
2098 2098 if (is_store) {
2099 2099 // Get the value being stored. (Pop it first; it was pushed last.)
2100 2100 switch (type) {
2101 2101 case T_DOUBLE:
2102 2102 case T_LONG:
2103 2103 case T_ADDRESS:
2104 2104 val = pop_pair();
2105 2105 break;
2106 2106 default:
2107 2107 val = pop();
2108 2108 }
2109 2109 }
2110 2110
2111 2111 // Build address expression. See the code in inline_unsafe_prefetch.
2112 2112 Node *adr;
2113 2113 Node *heap_base_oop = top();
2114 2114 if (!is_native_ptr) {
2115 2115 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2116 2116 Node* offset = pop_pair();
2117 2117 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2118 2118 Node* base = pop();
2119 2119 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2120 2120 // to be plain byte offsets, which are also the same as those accepted
2121 2121 // by oopDesc::field_base.
2122 2122 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2123 2123 "fieldOffset must be byte-scaled");
2124 2124 // 32-bit machines ignore the high half!
2125 2125 offset = ConvL2X(offset);
2126 2126 adr = make_unsafe_address(base, offset);
2127 2127 heap_base_oop = base;
2128 2128 } else {
2129 2129 Node* ptr = pop_pair();
2130 2130 // Adjust Java long to machine word:
2131 2131 ptr = ConvL2X(ptr);
2132 2132 adr = make_unsafe_address(NULL, ptr);
2133 2133 }
2134 2134
2135 2135 // Pop receiver last: it was pushed first.
2136 2136 Node *receiver = pop();
2137 2137
2138 2138 assert(saved_sp == _sp, "must have correct argument count");
2139 2139
2140 2140 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2141 2141
2142 2142 // First guess at the value type.
2143 2143 const Type *value_type = Type::get_const_basic_type(type);
2144 2144
2145 2145 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2146 2146 // there was not enough information to nail it down.
2147 2147 Compile::AliasType* alias_type = C->alias_type(adr_type);
2148 2148 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2149 2149
2150 2150 // We will need memory barriers unless we can determine a unique
2151 2151 // alias category for this reference. (Note: If for some reason
2152 2152 // the barriers get omitted and the unsafe reference begins to "pollute"
2153 2153 // the alias analysis of the rest of the graph, either Compile::can_alias
2154 2154 // or Compile::must_alias will throw a diagnostic assert.)
2155 2155 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2156 2156
2157 2157 if (!is_store && type == T_OBJECT) {
2158 2158 // Attempt to infer a sharper value type from the offset and base type.
2159 2159 ciKlass* sharpened_klass = NULL;
2160 2160
2161 2161 // See if it is an instance field, with an object type.
2162 2162 if (alias_type->field() != NULL) {
2163 2163 assert(!is_native_ptr, "native pointer op cannot use a java address");
2164 2164 if (alias_type->field()->type()->is_klass()) {
2165 2165 sharpened_klass = alias_type->field()->type()->as_klass();
2166 2166 }
2167 2167 }
2168 2168
2169 2169 // See if it is a narrow oop array.
2170 2170 if (adr_type->isa_aryptr()) {
2171 2171 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2172 2172 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2173 2173 if (elem_type != NULL) {
2174 2174 sharpened_klass = elem_type->klass();
2175 2175 }
2176 2176 }
2177 2177 }
2178 2178
2179 2179 if (sharpened_klass != NULL) {
2180 2180 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2181 2181
2182 2182 // Sharpen the value type.
2183 2183 value_type = tjp;
2184 2184
2185 2185 #ifndef PRODUCT
2186 2186 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2187 2187 tty->print(" from base type: "); adr_type->dump();
2188 2188 tty->print(" sharpened value: "); value_type->dump();
2189 2189 }
2190 2190 #endif
2191 2191 }
2192 2192 }
2193 2193
2194 2194 // Null check on self without removing any arguments. The argument
2195 2195 // null check technically happens in the wrong place, which can lead to
2196 2196 // invalid stack traces when the primitive is inlined into a method
2197 2197 // which handles NullPointerExceptions.
2198 2198 _sp += nargs;
2199 2199 do_null_check(receiver, T_OBJECT);
2200 2200 _sp -= nargs;
2201 2201 if (stopped()) {
2202 2202 return true;
2203 2203 }
2204 2204 // Heap pointers get a null-check from the interpreter,
2205 2205 // as a courtesy. However, this is not guaranteed by Unsafe,
2206 2206 // and it is not possible to fully distinguish unintended nulls
2207 2207 // from intended ones in this API.
2208 2208
2209 2209 if (is_volatile) {
2210 2210 // We need to emit leading and trailing CPU membars (see below) in
2211 2211 // addition to memory membars when is_volatile. This is a little
2212 2212 // too strong, but avoids the need to insert per-alias-type
2213 2213 // volatile membars (for stores; compare Parse::do_put_xxx), which
2214 2214 // we cannot do effectively here because we probably only have a
2215 2215 // rough approximation of type.
2216 2216 need_mem_bar = true;
2217 2217 // For Stores, place a memory ordering barrier now.
2218 2218 if (is_store)
2219 2219 insert_mem_bar(Op_MemBarRelease);
2220 2220 }
2221 2221
2222 2222 // Memory barrier to prevent normal and 'unsafe' accesses from
2223 2223 // bypassing each other. Happens after null checks, so the
2224 2224 // exception paths do not take memory state from the memory barrier,
2225 2225 // so there's no problems making a strong assert about mixing users
2226 2226 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2227 2227 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2228 2228 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2229 2229
2230 2230 if (!is_store) {
2231 2231 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2232 2232 // load value and push onto stack
2233 2233 switch (type) {
2234 2234 case T_BOOLEAN:
2235 2235 case T_CHAR:
2236 2236 case T_BYTE:
2237 2237 case T_SHORT:
2238 2238 case T_INT:
2239 2239 case T_FLOAT:
2240 2240 case T_OBJECT:
2241 2241 push( p );
2242 2242 break;
2243 2243 case T_ADDRESS:
2244 2244 // Cast to an int type.
2245 2245 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2246 2246 p = ConvX2L(p);
2247 2247 push_pair(p);
2248 2248 break;
2249 2249 case T_DOUBLE:
2250 2250 case T_LONG:
2251 2251 push_pair( p );
2252 2252 break;
2253 2253 default: ShouldNotReachHere();
2254 2254 }
2255 2255 } else {
2256 2256 // place effect of store into memory
2257 2257 switch (type) {
2258 2258 case T_DOUBLE:
2259 2259 val = dstore_rounding(val);
2260 2260 break;
2261 2261 case T_ADDRESS:
2262 2262 // Repackage the long as a pointer.
2263 2263 val = ConvL2X(val);
2264 2264 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2265 2265 break;
2266 2266 }
2267 2267
2268 2268 if (type != T_OBJECT ) {
2269 2269 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2270 2270 } else {
2271 2271 // Possibly an oop being stored to Java heap or native memory
2272 2272 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2273 2273 // oop to Java heap.
2274 2274 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2275 2275 } else {
2276 2276 // We can't tell at compile time if we are storing in the Java heap or outside
2277 2277 // of it. So we need to emit code to conditionally do the proper type of
2278 2278 // store.
2279 2279
2280 2280 IdealKit ideal(gvn(), control(), merged_memory());
2281 2281 #define __ ideal.
2282 2282 // QQQ who knows what probability is here??
2283 2283 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2284 2284 // Sync IdealKit and graphKit.
2285 2285 set_all_memory( __ merged_memory());
2286 2286 set_control(__ ctrl());
2287 2287 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2288 2288 // Update IdealKit memory.
2289 2289 __ set_all_memory(merged_memory());
2290 2290 __ set_ctrl(control());
2291 2291 } __ else_(); {
2292 2292 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2293 2293 } __ end_if();
2294 2294 // Final sync IdealKit and GraphKit.
2295 2295 sync_kit(ideal);
2296 2296 #undef __
2297 2297 }
2298 2298 }
2299 2299 }
2300 2300
2301 2301 if (is_volatile) {
2302 2302 if (!is_store)
2303 2303 insert_mem_bar(Op_MemBarAcquire);
2304 2304 else
2305 2305 insert_mem_bar(Op_MemBarVolatile);
2306 2306 }
2307 2307
2308 2308 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2309 2309
2310 2310 return true;
2311 2311 }
2312 2312
2313 2313 //----------------------------inline_unsafe_prefetch----------------------------
2314 2314
2315 2315 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2316 2316 #ifndef PRODUCT
2317 2317 {
2318 2318 ResourceMark rm;
2319 2319 // Check the signatures.
2320 2320 ciSignature* sig = signature();
2321 2321 #ifdef ASSERT
2322 2322 // Object getObject(Object base, int/long offset), etc.
2323 2323 BasicType rtype = sig->return_type()->basic_type();
2324 2324 if (!is_native_ptr) {
2325 2325 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2326 2326 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2327 2327 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2328 2328 } else {
2329 2329 assert(sig->count() == 1, "native prefetch has 1 argument");
2330 2330 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2331 2331 }
2332 2332 #endif // ASSERT
2333 2333 }
2334 2334 #endif // !PRODUCT
2335 2335
2336 2336 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2337 2337
2338 2338 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2339 2339 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2340 2340
2341 2341 debug_only(int saved_sp = _sp);
2342 2342 _sp += nargs;
2343 2343
2344 2344 // Build address expression. See the code in inline_unsafe_access.
2345 2345 Node *adr;
2346 2346 if (!is_native_ptr) {
2347 2347 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2348 2348 Node* offset = pop_pair();
2349 2349 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2350 2350 Node* base = pop();
2351 2351 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2352 2352 // to be plain byte offsets, which are also the same as those accepted
2353 2353 // by oopDesc::field_base.
2354 2354 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2355 2355 "fieldOffset must be byte-scaled");
2356 2356 // 32-bit machines ignore the high half!
2357 2357 offset = ConvL2X(offset);
2358 2358 adr = make_unsafe_address(base, offset);
2359 2359 } else {
2360 2360 Node* ptr = pop_pair();
2361 2361 // Adjust Java long to machine word:
2362 2362 ptr = ConvL2X(ptr);
2363 2363 adr = make_unsafe_address(NULL, ptr);
2364 2364 }
2365 2365
2366 2366 if (is_static) {
2367 2367 assert(saved_sp == _sp, "must have correct argument count");
2368 2368 } else {
2369 2369 // Pop receiver last: it was pushed first.
2370 2370 Node *receiver = pop();
2371 2371 assert(saved_sp == _sp, "must have correct argument count");
2372 2372
2373 2373 // Null check on self without removing any arguments. The argument
2374 2374 // null check technically happens in the wrong place, which can lead to
2375 2375 // invalid stack traces when the primitive is inlined into a method
2376 2376 // which handles NullPointerExceptions.
2377 2377 _sp += nargs;
2378 2378 do_null_check(receiver, T_OBJECT);
2379 2379 _sp -= nargs;
2380 2380 if (stopped()) {
2381 2381 return true;
2382 2382 }
2383 2383 }
2384 2384
2385 2385 // Generate the read or write prefetch
2386 2386 Node *prefetch;
2387 2387 if (is_store) {
2388 2388 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2389 2389 } else {
2390 2390 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2391 2391 }
2392 2392 prefetch->init_req(0, control());
2393 2393 set_i_o(_gvn.transform(prefetch));
2394 2394
2395 2395 return true;
2396 2396 }
2397 2397
2398 2398 //----------------------------inline_unsafe_CAS----------------------------
2399 2399
2400 2400 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2401 2401 // This basic scheme here is the same as inline_unsafe_access, but
2402 2402 // differs in enough details that combining them would make the code
2403 2403 // overly confusing. (This is a true fact! I originally combined
2404 2404 // them, but even I was confused by it!) As much code/comments as
2405 2405 // possible are retained from inline_unsafe_access though to make
2406 2406 // the correspondences clearer. - dl
2407 2407
2408 2408 if (callee()->is_static()) return false; // caller must have the capability!
2409 2409
2410 2410 #ifndef PRODUCT
2411 2411 {
2412 2412 ResourceMark rm;
2413 2413 // Check the signatures.
2414 2414 ciSignature* sig = signature();
2415 2415 #ifdef ASSERT
2416 2416 BasicType rtype = sig->return_type()->basic_type();
2417 2417 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2418 2418 assert(sig->count() == 4, "CAS has 4 arguments");
2419 2419 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2420 2420 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2421 2421 #endif // ASSERT
2422 2422 }
2423 2423 #endif //PRODUCT
2424 2424
2425 2425 // number of stack slots per value argument (1 or 2)
2426 2426 int type_words = type2size[type];
2427 2427
2428 2428 // Cannot inline wide CAS on machines that don't support it natively
2429 2429 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2430 2430 return false;
2431 2431
2432 2432 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2433 2433
2434 2434 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2435 2435 int nargs = 1 + 1 + 2 + type_words + type_words;
2436 2436
2437 2437 // pop arguments: newval, oldval, offset, base, and receiver
2438 2438 debug_only(int saved_sp = _sp);
2439 2439 _sp += nargs;
2440 2440 Node* newval = (type_words == 1) ? pop() : pop_pair();
2441 2441 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2442 2442 Node *offset = pop_pair();
2443 2443 Node *base = pop();
2444 2444 Node *receiver = pop();
2445 2445 assert(saved_sp == _sp, "must have correct argument count");
2446 2446
2447 2447 // Null check receiver.
2448 2448 _sp += nargs;
2449 2449 do_null_check(receiver, T_OBJECT);
2450 2450 _sp -= nargs;
2451 2451 if (stopped()) {
2452 2452 return true;
2453 2453 }
2454 2454
2455 2455 // Build field offset expression.
2456 2456 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2457 2457 // to be plain byte offsets, which are also the same as those accepted
2458 2458 // by oopDesc::field_base.
2459 2459 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2460 2460 // 32-bit machines ignore the high half of long offsets
2461 2461 offset = ConvL2X(offset);
2462 2462 Node* adr = make_unsafe_address(base, offset);
2463 2463 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2464 2464
2465 2465 // (Unlike inline_unsafe_access, there seems no point in trying
2466 2466 // to refine types. Just use the coarse types here.
2467 2467 const Type *value_type = Type::get_const_basic_type(type);
2468 2468 Compile::AliasType* alias_type = C->alias_type(adr_type);
2469 2469 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2470 2470 int alias_idx = C->get_alias_index(adr_type);
2471 2471
2472 2472 // Memory-model-wise, a CAS acts like a little synchronized block,
2473 2473 // so needs barriers on each side. These don't translate into
2474 2474 // actual barriers on most machines, but we still need rest of
2475 2475 // compiler to respect ordering.
2476 2476
2477 2477 insert_mem_bar(Op_MemBarRelease);
2478 2478 insert_mem_bar(Op_MemBarCPUOrder);
2479 2479
2480 2480 // 4984716: MemBars must be inserted before this
2481 2481 // memory node in order to avoid a false
2482 2482 // dependency which will confuse the scheduler.
2483 2483 Node *mem = memory(alias_idx);
2484 2484
2485 2485 // For now, we handle only those cases that actually exist: ints,
2486 2486 // longs, and Object. Adding others should be straightforward.
2487 2487 Node* cas;
2488 2488 switch(type) {
2489 2489 case T_INT:
2490 2490 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2491 2491 break;
2492 2492 case T_LONG:
2493 2493 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2494 2494 break;
2495 2495 case T_OBJECT:
2496 2496 // reference stores need a store barrier.
2497 2497 // (They don't if CAS fails, but it isn't worth checking.)
2498 2498 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2499 2499 #ifdef _LP64
2500 2500 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2501 2501 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2502 2502 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2503 2503 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2504 2504 newval_enc, oldval_enc));
2505 2505 } else
2506 2506 #endif
2507 2507 {
2508 2508 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2509 2509 }
2510 2510 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2511 2511 break;
2512 2512 default:
2513 2513 ShouldNotReachHere();
2514 2514 break;
2515 2515 }
2516 2516
2517 2517 // SCMemProjNodes represent the memory state of CAS. Their main
2518 2518 // role is to prevent CAS nodes from being optimized away when their
2519 2519 // results aren't used.
2520 2520 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2521 2521 set_memory(proj, alias_idx);
2522 2522
2523 2523 // Add the trailing membar surrounding the access
2524 2524 insert_mem_bar(Op_MemBarCPUOrder);
2525 2525 insert_mem_bar(Op_MemBarAcquire);
2526 2526
2527 2527 push(cas);
2528 2528 return true;
2529 2529 }
2530 2530
2531 2531 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2532 2532 // This is another variant of inline_unsafe_access, differing in
2533 2533 // that it always issues store-store ("release") barrier and ensures
2534 2534 // store-atomicity (which only matters for "long").
2535 2535
2536 2536 if (callee()->is_static()) return false; // caller must have the capability!
2537 2537
2538 2538 #ifndef PRODUCT
2539 2539 {
2540 2540 ResourceMark rm;
2541 2541 // Check the signatures.
2542 2542 ciSignature* sig = signature();
2543 2543 #ifdef ASSERT
2544 2544 BasicType rtype = sig->return_type()->basic_type();
2545 2545 assert(rtype == T_VOID, "must return void");
2546 2546 assert(sig->count() == 3, "has 3 arguments");
2547 2547 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2548 2548 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2549 2549 #endif // ASSERT
2550 2550 }
2551 2551 #endif //PRODUCT
2552 2552
2553 2553 // number of stack slots per value argument (1 or 2)
2554 2554 int type_words = type2size[type];
2555 2555
2556 2556 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2557 2557
2558 2558 // Argument words: "this" plus oop plus offset plus value;
2559 2559 int nargs = 1 + 1 + 2 + type_words;
2560 2560
2561 2561 // pop arguments: val, offset, base, and receiver
2562 2562 debug_only(int saved_sp = _sp);
2563 2563 _sp += nargs;
2564 2564 Node* val = (type_words == 1) ? pop() : pop_pair();
2565 2565 Node *offset = pop_pair();
2566 2566 Node *base = pop();
2567 2567 Node *receiver = pop();
2568 2568 assert(saved_sp == _sp, "must have correct argument count");
2569 2569
2570 2570 // Null check receiver.
2571 2571 _sp += nargs;
2572 2572 do_null_check(receiver, T_OBJECT);
2573 2573 _sp -= nargs;
2574 2574 if (stopped()) {
2575 2575 return true;
2576 2576 }
2577 2577
2578 2578 // Build field offset expression.
2579 2579 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2580 2580 // 32-bit machines ignore the high half of long offsets
2581 2581 offset = ConvL2X(offset);
2582 2582 Node* adr = make_unsafe_address(base, offset);
2583 2583 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2584 2584 const Type *value_type = Type::get_const_basic_type(type);
2585 2585 Compile::AliasType* alias_type = C->alias_type(adr_type);
2586 2586
2587 2587 insert_mem_bar(Op_MemBarRelease);
2588 2588 insert_mem_bar(Op_MemBarCPUOrder);
2589 2589 // Ensure that the store is atomic for longs:
2590 2590 bool require_atomic_access = true;
2591 2591 Node* store;
2592 2592 if (type == T_OBJECT) // reference stores need a store barrier.
2593 2593 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2594 2594 else {
2595 2595 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2596 2596 }
2597 2597 insert_mem_bar(Op_MemBarCPUOrder);
2598 2598 return true;
2599 2599 }
2600 2600
2601 2601 bool LibraryCallKit::inline_unsafe_allocate() {
2602 2602 if (callee()->is_static()) return false; // caller must have the capability!
2603 2603 int nargs = 1 + 1;
2604 2604 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2605 2605 null_check_receiver(callee()); // check then ignore argument(0)
2606 2606 _sp += nargs; // set original stack for use by uncommon_trap
2607 2607 Node* cls = do_null_check(argument(1), T_OBJECT);
2608 2608 _sp -= nargs;
2609 2609 if (stopped()) return true;
2610 2610
2611 2611 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2612 2612 _sp += nargs; // set original stack for use by uncommon_trap
2613 2613 kls = do_null_check(kls, T_OBJECT);
2614 2614 _sp -= nargs;
2615 2615 if (stopped()) return true; // argument was like int.class
2616 2616
2617 2617 // Note: The argument might still be an illegal value like
2618 2618 // Serializable.class or Object[].class. The runtime will handle it.
2619 2619 // But we must make an explicit check for initialization.
2620 2620 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2621 2621 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2622 2622 Node* bits = intcon(instanceKlass::fully_initialized);
2623 2623 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2624 2624 // The 'test' is non-zero if we need to take a slow path.
2625 2625
2626 2626 Node* obj = new_instance(kls, test);
2627 2627 push(obj);
2628 2628
2629 2629 return true;
2630 2630 }
2631 2631
2632 2632 //------------------------inline_native_time_funcs--------------
2633 2633 // inline code for System.currentTimeMillis() and System.nanoTime()
2634 2634 // these have the same type and signature
2635 2635 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2636 2636 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2637 2637 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2638 2638 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2639 2639 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2640 2640 const TypePtr* no_memory_effects = NULL;
2641 2641 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2642 2642 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2643 2643 #ifdef ASSERT
2644 2644 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2645 2645 assert(value_top == top(), "second value must be top");
2646 2646 #endif
2647 2647 push_pair(value);
2648 2648 return true;
2649 2649 }
2650 2650
2651 2651 //------------------------inline_native_currentThread------------------
2652 2652 bool LibraryCallKit::inline_native_currentThread() {
2653 2653 Node* junk = NULL;
2654 2654 push(generate_current_thread(junk));
2655 2655 return true;
2656 2656 }
2657 2657
2658 2658 //------------------------inline_native_isInterrupted------------------
2659 2659 bool LibraryCallKit::inline_native_isInterrupted() {
2660 2660 const int nargs = 1+1; // receiver + boolean
2661 2661 assert(nargs == arg_size(), "sanity");
2662 2662 // Add a fast path to t.isInterrupted(clear_int):
2663 2663 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2664 2664 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2665 2665 // So, in the common case that the interrupt bit is false,
2666 2666 // we avoid making a call into the VM. Even if the interrupt bit
2667 2667 // is true, if the clear_int argument is false, we avoid the VM call.
2668 2668 // However, if the receiver is not currentThread, we must call the VM,
2669 2669 // because there must be some locking done around the operation.
2670 2670
2671 2671 // We only go to the fast case code if we pass two guards.
2672 2672 // Paths which do not pass are accumulated in the slow_region.
2673 2673 RegionNode* slow_region = new (C, 1) RegionNode(1);
2674 2674 record_for_igvn(slow_region);
2675 2675 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2676 2676 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2677 2677 enum { no_int_result_path = 1,
2678 2678 no_clear_result_path = 2,
2679 2679 slow_result_path = 3
2680 2680 };
2681 2681
2682 2682 // (a) Receiving thread must be the current thread.
2683 2683 Node* rec_thr = argument(0);
2684 2684 Node* tls_ptr = NULL;
2685 2685 Node* cur_thr = generate_current_thread(tls_ptr);
2686 2686 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2687 2687 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2688 2688
2689 2689 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2690 2690 if (!known_current_thread)
2691 2691 generate_slow_guard(bol_thr, slow_region);
2692 2692
2693 2693 // (b) Interrupt bit on TLS must be false.
2694 2694 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2695 2695 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2696 2696 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2697 2697 // Set the control input on the field _interrupted read to prevent it floating up.
2698 2698 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2699 2699 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2700 2700 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2701 2701
2702 2702 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2703 2703
2704 2704 // First fast path: if (!TLS._interrupted) return false;
2705 2705 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2706 2706 result_rgn->init_req(no_int_result_path, false_bit);
2707 2707 result_val->init_req(no_int_result_path, intcon(0));
2708 2708
2709 2709 // drop through to next case
2710 2710 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2711 2711
2712 2712 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2713 2713 Node* clr_arg = argument(1);
2714 2714 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2715 2715 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2716 2716 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2717 2717
2718 2718 // Second fast path: ... else if (!clear_int) return true;
2719 2719 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2720 2720 result_rgn->init_req(no_clear_result_path, false_arg);
2721 2721 result_val->init_req(no_clear_result_path, intcon(1));
2722 2722
2723 2723 // drop through to next case
2724 2724 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2725 2725
2726 2726 // (d) Otherwise, go to the slow path.
2727 2727 slow_region->add_req(control());
2728 2728 set_control( _gvn.transform(slow_region) );
2729 2729
2730 2730 if (stopped()) {
2731 2731 // There is no slow path.
2732 2732 result_rgn->init_req(slow_result_path, top());
2733 2733 result_val->init_req(slow_result_path, top());
2734 2734 } else {
2735 2735 // non-virtual because it is a private non-static
2736 2736 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2737 2737
2738 2738 Node* slow_val = set_results_for_java_call(slow_call);
2739 2739 // this->control() comes from set_results_for_java_call
2740 2740
2741 2741 // If we know that the result of the slow call will be true, tell the optimizer!
2742 2742 if (known_current_thread) slow_val = intcon(1);
2743 2743
2744 2744 Node* fast_io = slow_call->in(TypeFunc::I_O);
2745 2745 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2746 2746 // These two phis are pre-filled with copies of of the fast IO and Memory
2747 2747 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2748 2748 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2749 2749
2750 2750 result_rgn->init_req(slow_result_path, control());
2751 2751 io_phi ->init_req(slow_result_path, i_o());
2752 2752 mem_phi ->init_req(slow_result_path, reset_memory());
2753 2753 result_val->init_req(slow_result_path, slow_val);
2754 2754
2755 2755 set_all_memory( _gvn.transform(mem_phi) );
2756 2756 set_i_o( _gvn.transform(io_phi) );
2757 2757 }
2758 2758
2759 2759 push_result(result_rgn, result_val);
2760 2760 C->set_has_split_ifs(true); // Has chance for split-if optimization
2761 2761
2762 2762 return true;
2763 2763 }
2764 2764
2765 2765 //---------------------------load_mirror_from_klass----------------------------
2766 2766 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2767 2767 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2768 2768 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2769 2769 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2770 2770 }
2771 2771
2772 2772 //-----------------------load_klass_from_mirror_common-------------------------
2773 2773 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2774 2774 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2775 2775 // and branch to the given path on the region.
2776 2776 // If never_see_null, take an uncommon trap on null, so we can optimistically
2777 2777 // compile for the non-null case.
2778 2778 // If the region is NULL, force never_see_null = true.
2779 2779 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2780 2780 bool never_see_null,
2781 2781 int nargs,
2782 2782 RegionNode* region,
2783 2783 int null_path,
2784 2784 int offset) {
2785 2785 if (region == NULL) never_see_null = true;
2786 2786 Node* p = basic_plus_adr(mirror, offset);
2787 2787 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2788 2788 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2789 2789 _sp += nargs; // any deopt will start just before call to enclosing method
2790 2790 Node* null_ctl = top();
2791 2791 kls = null_check_oop(kls, &null_ctl, never_see_null);
2792 2792 if (region != NULL) {
2793 2793 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2794 2794 region->init_req(null_path, null_ctl);
2795 2795 } else {
2796 2796 assert(null_ctl == top(), "no loose ends");
2797 2797 }
2798 2798 _sp -= nargs;
2799 2799 return kls;
2800 2800 }
2801 2801
2802 2802 //--------------------(inline_native_Class_query helpers)---------------------
2803 2803 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2804 2804 // Fall through if (mods & mask) == bits, take the guard otherwise.
2805 2805 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2806 2806 // Branch around if the given klass has the given modifier bit set.
2807 2807 // Like generate_guard, adds a new path onto the region.
2808 2808 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2809 2809 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2810 2810 Node* mask = intcon(modifier_mask);
2811 2811 Node* bits = intcon(modifier_bits);
2812 2812 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2813 2813 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2814 2814 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2815 2815 return generate_fair_guard(bol, region);
2816 2816 }
2817 2817 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2818 2818 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2819 2819 }
2820 2820
2821 2821 //-------------------------inline_native_Class_query-------------------
2822 2822 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2823 2823 int nargs = 1+0; // just the Class mirror, in most cases
2824 2824 const Type* return_type = TypeInt::BOOL;
2825 2825 Node* prim_return_value = top(); // what happens if it's a primitive class?
2826 2826 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2827 2827 bool expect_prim = false; // most of these guys expect to work on refs
2828 2828
2829 2829 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2830 2830
2831 2831 switch (id) {
2832 2832 case vmIntrinsics::_isInstance:
2833 2833 nargs = 1+1; // the Class mirror, plus the object getting queried about
2834 2834 // nothing is an instance of a primitive type
2835 2835 prim_return_value = intcon(0);
2836 2836 break;
2837 2837 case vmIntrinsics::_getModifiers:
2838 2838 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2839 2839 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2840 2840 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2841 2841 break;
2842 2842 case vmIntrinsics::_isInterface:
2843 2843 prim_return_value = intcon(0);
2844 2844 break;
2845 2845 case vmIntrinsics::_isArray:
2846 2846 prim_return_value = intcon(0);
2847 2847 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2848 2848 break;
2849 2849 case vmIntrinsics::_isPrimitive:
2850 2850 prim_return_value = intcon(1);
2851 2851 expect_prim = true; // obviously
2852 2852 break;
2853 2853 case vmIntrinsics::_getSuperclass:
2854 2854 prim_return_value = null();
2855 2855 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2856 2856 break;
2857 2857 case vmIntrinsics::_getComponentType:
2858 2858 prim_return_value = null();
2859 2859 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2860 2860 break;
2861 2861 case vmIntrinsics::_getClassAccessFlags:
2862 2862 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2863 2863 return_type = TypeInt::INT; // not bool! 6297094
2864 2864 break;
2865 2865 default:
2866 2866 ShouldNotReachHere();
2867 2867 }
2868 2868
2869 2869 Node* mirror = argument(0);
2870 2870 Node* obj = (nargs <= 1)? top(): argument(1);
2871 2871
2872 2872 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2873 2873 if (mirror_con == NULL) return false; // cannot happen?
2874 2874
2875 2875 #ifndef PRODUCT
2876 2876 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2877 2877 ciType* k = mirror_con->java_mirror_type();
2878 2878 if (k) {
2879 2879 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2880 2880 k->print_name();
2881 2881 tty->cr();
2882 2882 }
2883 2883 }
2884 2884 #endif
2885 2885
2886 2886 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2887 2887 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2888 2888 record_for_igvn(region);
2889 2889 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2890 2890
2891 2891 // The mirror will never be null of Reflection.getClassAccessFlags, however
2892 2892 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2893 2893 // if it is. See bug 4774291.
2894 2894
2895 2895 // For Reflection.getClassAccessFlags(), the null check occurs in
2896 2896 // the wrong place; see inline_unsafe_access(), above, for a similar
2897 2897 // situation.
2898 2898 _sp += nargs; // set original stack for use by uncommon_trap
2899 2899 mirror = do_null_check(mirror, T_OBJECT);
2900 2900 _sp -= nargs;
2901 2901 // If mirror or obj is dead, only null-path is taken.
2902 2902 if (stopped()) return true;
2903 2903
2904 2904 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2905 2905
2906 2906 // Now load the mirror's klass metaobject, and null-check it.
2907 2907 // Side-effects region with the control path if the klass is null.
2908 2908 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2909 2909 region, _prim_path);
2910 2910 // If kls is null, we have a primitive mirror.
2911 2911 phi->init_req(_prim_path, prim_return_value);
2912 2912 if (stopped()) { push_result(region, phi); return true; }
2913 2913
2914 2914 Node* p; // handy temp
2915 2915 Node* null_ctl;
2916 2916
2917 2917 // Now that we have the non-null klass, we can perform the real query.
2918 2918 // For constant classes, the query will constant-fold in LoadNode::Value.
2919 2919 Node* query_value = top();
2920 2920 switch (id) {
2921 2921 case vmIntrinsics::_isInstance:
2922 2922 // nothing is an instance of a primitive type
2923 2923 query_value = gen_instanceof(obj, kls);
2924 2924 break;
2925 2925
2926 2926 case vmIntrinsics::_getModifiers:
2927 2927 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2928 2928 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2929 2929 break;
2930 2930
2931 2931 case vmIntrinsics::_isInterface:
2932 2932 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2933 2933 if (generate_interface_guard(kls, region) != NULL)
2934 2934 // A guard was added. If the guard is taken, it was an interface.
2935 2935 phi->add_req(intcon(1));
2936 2936 // If we fall through, it's a plain class.
2937 2937 query_value = intcon(0);
2938 2938 break;
2939 2939
2940 2940 case vmIntrinsics::_isArray:
2941 2941 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2942 2942 if (generate_array_guard(kls, region) != NULL)
2943 2943 // A guard was added. If the guard is taken, it was an array.
2944 2944 phi->add_req(intcon(1));
2945 2945 // If we fall through, it's a plain class.
2946 2946 query_value = intcon(0);
2947 2947 break;
2948 2948
2949 2949 case vmIntrinsics::_isPrimitive:
2950 2950 query_value = intcon(0); // "normal" path produces false
2951 2951 break;
2952 2952
2953 2953 case vmIntrinsics::_getSuperclass:
2954 2954 // The rules here are somewhat unfortunate, but we can still do better
2955 2955 // with random logic than with a JNI call.
2956 2956 // Interfaces store null or Object as _super, but must report null.
2957 2957 // Arrays store an intermediate super as _super, but must report Object.
2958 2958 // Other types can report the actual _super.
2959 2959 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2960 2960 if (generate_interface_guard(kls, region) != NULL)
2961 2961 // A guard was added. If the guard is taken, it was an interface.
2962 2962 phi->add_req(null());
2963 2963 if (generate_array_guard(kls, region) != NULL)
2964 2964 // A guard was added. If the guard is taken, it was an array.
2965 2965 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
2966 2966 // If we fall through, it's a plain class. Get its _super.
2967 2967 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
2968 2968 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
2969 2969 null_ctl = top();
2970 2970 kls = null_check_oop(kls, &null_ctl);
2971 2971 if (null_ctl != top()) {
2972 2972 // If the guard is taken, Object.superClass is null (both klass and mirror).
2973 2973 region->add_req(null_ctl);
2974 2974 phi ->add_req(null());
2975 2975 }
2976 2976 if (!stopped()) {
2977 2977 query_value = load_mirror_from_klass(kls);
2978 2978 }
2979 2979 break;
2980 2980
2981 2981 case vmIntrinsics::_getComponentType:
2982 2982 if (generate_array_guard(kls, region) != NULL) {
2983 2983 // Be sure to pin the oop load to the guard edge just created:
2984 2984 Node* is_array_ctrl = region->in(region->req()-1);
2985 2985 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
2986 2986 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
2987 2987 phi->add_req(cmo);
2988 2988 }
2989 2989 query_value = null(); // non-array case is null
2990 2990 break;
2991 2991
2992 2992 case vmIntrinsics::_getClassAccessFlags:
2993 2993 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2994 2994 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2995 2995 break;
2996 2996
2997 2997 default:
2998 2998 ShouldNotReachHere();
2999 2999 }
3000 3000
3001 3001 // Fall-through is the normal case of a query to a real class.
3002 3002 phi->init_req(1, query_value);
3003 3003 region->init_req(1, control());
3004 3004
3005 3005 push_result(region, phi);
3006 3006 C->set_has_split_ifs(true); // Has chance for split-if optimization
3007 3007
3008 3008 return true;
3009 3009 }
3010 3010
3011 3011 //--------------------------inline_native_subtype_check------------------------
3012 3012 // This intrinsic takes the JNI calls out of the heart of
3013 3013 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3014 3014 bool LibraryCallKit::inline_native_subtype_check() {
3015 3015 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3016 3016
3017 3017 // Pull both arguments off the stack.
3018 3018 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3019 3019 args[0] = argument(0);
3020 3020 args[1] = argument(1);
3021 3021 Node* klasses[2]; // corresponding Klasses: superk, subk
3022 3022 klasses[0] = klasses[1] = top();
3023 3023
3024 3024 enum {
3025 3025 // A full decision tree on {superc is prim, subc is prim}:
3026 3026 _prim_0_path = 1, // {P,N} => false
3027 3027 // {P,P} & superc!=subc => false
3028 3028 _prim_same_path, // {P,P} & superc==subc => true
3029 3029 _prim_1_path, // {N,P} => false
3030 3030 _ref_subtype_path, // {N,N} & subtype check wins => true
3031 3031 _both_ref_path, // {N,N} & subtype check loses => false
3032 3032 PATH_LIMIT
3033 3033 };
3034 3034
3035 3035 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3036 3036 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3037 3037 record_for_igvn(region);
3038 3038
3039 3039 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3040 3040 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3041 3041 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3042 3042
3043 3043 // First null-check both mirrors and load each mirror's klass metaobject.
3044 3044 int which_arg;
3045 3045 for (which_arg = 0; which_arg <= 1; which_arg++) {
3046 3046 Node* arg = args[which_arg];
3047 3047 _sp += nargs; // set original stack for use by uncommon_trap
3048 3048 arg = do_null_check(arg, T_OBJECT);
3049 3049 _sp -= nargs;
3050 3050 if (stopped()) break;
3051 3051 args[which_arg] = _gvn.transform(arg);
3052 3052
3053 3053 Node* p = basic_plus_adr(arg, class_klass_offset);
3054 3054 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3055 3055 klasses[which_arg] = _gvn.transform(kls);
3056 3056 }
3057 3057
3058 3058 // Having loaded both klasses, test each for null.
3059 3059 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3060 3060 for (which_arg = 0; which_arg <= 1; which_arg++) {
3061 3061 Node* kls = klasses[which_arg];
3062 3062 Node* null_ctl = top();
3063 3063 _sp += nargs; // set original stack for use by uncommon_trap
3064 3064 kls = null_check_oop(kls, &null_ctl, never_see_null);
3065 3065 _sp -= nargs;
3066 3066 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3067 3067 region->init_req(prim_path, null_ctl);
3068 3068 if (stopped()) break;
3069 3069 klasses[which_arg] = kls;
3070 3070 }
3071 3071
3072 3072 if (!stopped()) {
3073 3073 // now we have two reference types, in klasses[0..1]
3074 3074 Node* subk = klasses[1]; // the argument to isAssignableFrom
3075 3075 Node* superk = klasses[0]; // the receiver
3076 3076 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3077 3077 // now we have a successful reference subtype check
3078 3078 region->set_req(_ref_subtype_path, control());
3079 3079 }
3080 3080
3081 3081 // If both operands are primitive (both klasses null), then
3082 3082 // we must return true when they are identical primitives.
3083 3083 // It is convenient to test this after the first null klass check.
3084 3084 set_control(region->in(_prim_0_path)); // go back to first null check
3085 3085 if (!stopped()) {
3086 3086 // Since superc is primitive, make a guard for the superc==subc case.
3087 3087 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3088 3088 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3089 3089 generate_guard(bol_eq, region, PROB_FAIR);
3090 3090 if (region->req() == PATH_LIMIT+1) {
3091 3091 // A guard was added. If the added guard is taken, superc==subc.
3092 3092 region->swap_edges(PATH_LIMIT, _prim_same_path);
3093 3093 region->del_req(PATH_LIMIT);
3094 3094 }
3095 3095 region->set_req(_prim_0_path, control()); // Not equal after all.
3096 3096 }
3097 3097
3098 3098 // these are the only paths that produce 'true':
3099 3099 phi->set_req(_prim_same_path, intcon(1));
3100 3100 phi->set_req(_ref_subtype_path, intcon(1));
3101 3101
3102 3102 // pull together the cases:
3103 3103 assert(region->req() == PATH_LIMIT, "sane region");
3104 3104 for (uint i = 1; i < region->req(); i++) {
3105 3105 Node* ctl = region->in(i);
3106 3106 if (ctl == NULL || ctl == top()) {
3107 3107 region->set_req(i, top());
3108 3108 phi ->set_req(i, top());
3109 3109 } else if (phi->in(i) == NULL) {
3110 3110 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3111 3111 }
3112 3112 }
3113 3113
3114 3114 set_control(_gvn.transform(region));
3115 3115 push(_gvn.transform(phi));
3116 3116
3117 3117 return true;
3118 3118 }
3119 3119
3120 3120 //---------------------generate_array_guard_common------------------------
3121 3121 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3122 3122 bool obj_array, bool not_array) {
3123 3123 // If obj_array/non_array==false/false:
3124 3124 // Branch around if the given klass is in fact an array (either obj or prim).
3125 3125 // If obj_array/non_array==false/true:
3126 3126 // Branch around if the given klass is not an array klass of any kind.
3127 3127 // If obj_array/non_array==true/true:
3128 3128 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3129 3129 // If obj_array/non_array==true/false:
3130 3130 // Branch around if the kls is an oop array (Object[] or subtype)
3131 3131 //
3132 3132 // Like generate_guard, adds a new path onto the region.
3133 3133 jint layout_con = 0;
3134 3134 Node* layout_val = get_layout_helper(kls, layout_con);
3135 3135 if (layout_val == NULL) {
3136 3136 bool query = (obj_array
3137 3137 ? Klass::layout_helper_is_objArray(layout_con)
3138 3138 : Klass::layout_helper_is_javaArray(layout_con));
3139 3139 if (query == not_array) {
3140 3140 return NULL; // never a branch
3141 3141 } else { // always a branch
3142 3142 Node* always_branch = control();
3143 3143 if (region != NULL)
3144 3144 region->add_req(always_branch);
3145 3145 set_control(top());
3146 3146 return always_branch;
3147 3147 }
3148 3148 }
3149 3149 // Now test the correct condition.
3150 3150 jint nval = (obj_array
3151 3151 ? ((jint)Klass::_lh_array_tag_type_value
3152 3152 << Klass::_lh_array_tag_shift)
3153 3153 : Klass::_lh_neutral_value);
3154 3154 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3155 3155 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3156 3156 // invert the test if we are looking for a non-array
3157 3157 if (not_array) btest = BoolTest(btest).negate();
3158 3158 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3159 3159 return generate_fair_guard(bol, region);
3160 3160 }
3161 3161
3162 3162
3163 3163 //-----------------------inline_native_newArray--------------------------
3164 3164 bool LibraryCallKit::inline_native_newArray() {
3165 3165 int nargs = 2;
3166 3166 Node* mirror = argument(0);
3167 3167 Node* count_val = argument(1);
3168 3168
3169 3169 _sp += nargs; // set original stack for use by uncommon_trap
3170 3170 mirror = do_null_check(mirror, T_OBJECT);
3171 3171 _sp -= nargs;
3172 3172 // If mirror or obj is dead, only null-path is taken.
3173 3173 if (stopped()) return true;
3174 3174
3175 3175 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3176 3176 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3177 3177 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3178 3178 TypeInstPtr::NOTNULL);
3179 3179 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3180 3180 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3181 3181 TypePtr::BOTTOM);
3182 3182
3183 3183 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3184 3184 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3185 3185 nargs,
3186 3186 result_reg, _slow_path);
3187 3187 Node* normal_ctl = control();
3188 3188 Node* no_array_ctl = result_reg->in(_slow_path);
3189 3189
3190 3190 // Generate code for the slow case. We make a call to newArray().
3191 3191 set_control(no_array_ctl);
3192 3192 if (!stopped()) {
3193 3193 // Either the input type is void.class, or else the
3194 3194 // array klass has not yet been cached. Either the
3195 3195 // ensuing call will throw an exception, or else it
3196 3196 // will cache the array klass for next time.
3197 3197 PreserveJVMState pjvms(this);
3198 3198 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3199 3199 Node* slow_result = set_results_for_java_call(slow_call);
3200 3200 // this->control() comes from set_results_for_java_call
3201 3201 result_reg->set_req(_slow_path, control());
3202 3202 result_val->set_req(_slow_path, slow_result);
3203 3203 result_io ->set_req(_slow_path, i_o());
3204 3204 result_mem->set_req(_slow_path, reset_memory());
3205 3205 }
3206 3206
3207 3207 set_control(normal_ctl);
3208 3208 if (!stopped()) {
3209 3209 // Normal case: The array type has been cached in the java.lang.Class.
3210 3210 // The following call works fine even if the array type is polymorphic.
3211 3211 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3212 3212 Node* obj = new_array(klass_node, count_val, nargs);
3213 3213 result_reg->init_req(_normal_path, control());
3214 3214 result_val->init_req(_normal_path, obj);
3215 3215 result_io ->init_req(_normal_path, i_o());
3216 3216 result_mem->init_req(_normal_path, reset_memory());
3217 3217 }
3218 3218
3219 3219 // Return the combined state.
3220 3220 set_i_o( _gvn.transform(result_io) );
3221 3221 set_all_memory( _gvn.transform(result_mem) );
3222 3222 push_result(result_reg, result_val);
3223 3223 C->set_has_split_ifs(true); // Has chance for split-if optimization
3224 3224
3225 3225 return true;
3226 3226 }
3227 3227
3228 3228 //----------------------inline_native_getLength--------------------------
3229 3229 bool LibraryCallKit::inline_native_getLength() {
3230 3230 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3231 3231
3232 3232 int nargs = 1;
3233 3233 Node* array = argument(0);
3234 3234
3235 3235 _sp += nargs; // set original stack for use by uncommon_trap
3236 3236 array = do_null_check(array, T_OBJECT);
3237 3237 _sp -= nargs;
3238 3238
3239 3239 // If array is dead, only null-path is taken.
3240 3240 if (stopped()) return true;
3241 3241
3242 3242 // Deoptimize if it is a non-array.
3243 3243 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3244 3244
3245 3245 if (non_array != NULL) {
3246 3246 PreserveJVMState pjvms(this);
3247 3247 set_control(non_array);
3248 3248 _sp += nargs; // push the arguments back on the stack
3249 3249 uncommon_trap(Deoptimization::Reason_intrinsic,
3250 3250 Deoptimization::Action_maybe_recompile);
3251 3251 }
3252 3252
3253 3253 // If control is dead, only non-array-path is taken.
3254 3254 if (stopped()) return true;
3255 3255
3256 3256 // The works fine even if the array type is polymorphic.
3257 3257 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3258 3258 push( load_array_length(array) );
3259 3259
3260 3260 C->set_has_split_ifs(true); // Has chance for split-if optimization
3261 3261
3262 3262 return true;
3263 3263 }
3264 3264
3265 3265 //------------------------inline_array_copyOf----------------------------
3266 3266 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3267 3267 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3268 3268
3269 3269 // Restore the stack and pop off the arguments.
3270 3270 int nargs = 3 + (is_copyOfRange? 1: 0);
3271 3271 Node* original = argument(0);
3272 3272 Node* start = is_copyOfRange? argument(1): intcon(0);
3273 3273 Node* end = is_copyOfRange? argument(2): argument(1);
3274 3274 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3275 3275
3276 3276 Node* newcopy;
3277 3277
3278 3278 //set the original stack and the reexecute bit for the interpreter to reexecute
3279 3279 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3280 3280 { PreserveReexecuteState preexecs(this);
3281 3281 _sp += nargs;
3282 3282 jvms()->set_should_reexecute(true);
3283 3283
3284 3284 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3285 3285 original = do_null_check(original, T_OBJECT);
3286 3286
3287 3287 // Check if a null path was taken unconditionally.
3288 3288 if (stopped()) return true;
3289 3289
3290 3290 Node* orig_length = load_array_length(original);
3291 3291
3292 3292 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3293 3293 NULL, 0);
3294 3294 klass_node = do_null_check(klass_node, T_OBJECT);
3295 3295
3296 3296 RegionNode* bailout = new (C, 1) RegionNode(1);
3297 3297 record_for_igvn(bailout);
3298 3298
3299 3299 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3300 3300 // Bail out if that is so.
3301 3301 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3302 3302 if (not_objArray != NULL) {
3303 3303 // Improve the klass node's type from the new optimistic assumption:
3304 3304 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3305 3305 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3306 3306 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3307 3307 cast->init_req(0, control());
3308 3308 klass_node = _gvn.transform(cast);
3309 3309 }
3310 3310
3311 3311 // Bail out if either start or end is negative.
3312 3312 generate_negative_guard(start, bailout, &start);
3313 3313 generate_negative_guard(end, bailout, &end);
3314 3314
3315 3315 Node* length = end;
3316 3316 if (_gvn.type(start) != TypeInt::ZERO) {
3317 3317 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3318 3318 }
3319 3319
3320 3320 // Bail out if length is negative.
3321 3321 // ...Not needed, since the new_array will throw the right exception.
3322 3322 //generate_negative_guard(length, bailout, &length);
3323 3323
3324 3324 if (bailout->req() > 1) {
3325 3325 PreserveJVMState pjvms(this);
3326 3326 set_control( _gvn.transform(bailout) );
3327 3327 uncommon_trap(Deoptimization::Reason_intrinsic,
3328 3328 Deoptimization::Action_maybe_recompile);
3329 3329 }
3330 3330
3331 3331 if (!stopped()) {
3332 3332
3333 3333 // How many elements will we copy from the original?
3334 3334 // The answer is MinI(orig_length - start, length).
3335 3335 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3336 3336 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3337 3337
3338 3338 const bool raw_mem_only = true;
3339 3339 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3340 3340
3341 3341 // Generate a direct call to the right arraycopy function(s).
3342 3342 // We know the copy is disjoint but we might not know if the
3343 3343 // oop stores need checking.
3344 3344 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3345 3345 // This will fail a store-check if x contains any non-nulls.
3346 3346 bool disjoint_bases = true;
3347 3347 bool length_never_negative = true;
3348 3348 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3349 3349 original, start, newcopy, intcon(0), moved,
3350 3350 disjoint_bases, length_never_negative);
3351 3351 }
3352 3352 } //original reexecute and sp are set back here
3353 3353
3354 3354 if(!stopped()) {
3355 3355 push(newcopy);
3356 3356 }
3357 3357
3358 3358 C->set_has_split_ifs(true); // Has chance for split-if optimization
3359 3359
3360 3360 return true;
3361 3361 }
3362 3362
3363 3363
3364 3364 //----------------------generate_virtual_guard---------------------------
3365 3365 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3366 3366 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3367 3367 RegionNode* slow_region) {
3368 3368 ciMethod* method = callee();
3369 3369 int vtable_index = method->vtable_index();
3370 3370 // Get the methodOop out of the appropriate vtable entry.
3371 3371 int entry_offset = (instanceKlass::vtable_start_offset() +
3372 3372 vtable_index*vtableEntry::size()) * wordSize +
3373 3373 vtableEntry::method_offset_in_bytes();
3374 3374 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3375 3375 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3376 3376
3377 3377 // Compare the target method with the expected method (e.g., Object.hashCode).
3378 3378 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3379 3379
3380 3380 Node* native_call = makecon(native_call_addr);
3381 3381 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3382 3382 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3383 3383
3384 3384 return generate_slow_guard(test_native, slow_region);
3385 3385 }
3386 3386
3387 3387 //-----------------------generate_method_call----------------------------
3388 3388 // Use generate_method_call to make a slow-call to the real
3389 3389 // method if the fast path fails. An alternative would be to
3390 3390 // use a stub like OptoRuntime::slow_arraycopy_Java.
3391 3391 // This only works for expanding the current library call,
3392 3392 // not another intrinsic. (E.g., don't use this for making an
3393 3393 // arraycopy call inside of the copyOf intrinsic.)
3394 3394 CallJavaNode*
3395 3395 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3396 3396 // When compiling the intrinsic method itself, do not use this technique.
3397 3397 guarantee(callee() != C->method(), "cannot make slow-call to self");
3398 3398
3399 3399 ciMethod* method = callee();
3400 3400 // ensure the JVMS we have will be correct for this call
3401 3401 guarantee(method_id == method->intrinsic_id(), "must match");
3402 3402
3403 3403 const TypeFunc* tf = TypeFunc::make(method);
3404 3404 int tfdc = tf->domain()->cnt();
3405 3405 CallJavaNode* slow_call;
3406 3406 if (is_static) {
3407 3407 assert(!is_virtual, "");
3408 3408 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3409 3409 SharedRuntime::get_resolve_static_call_stub(),
3410 3410 method, bci());
3411 3411 } else if (is_virtual) {
3412 3412 null_check_receiver(method);
3413 3413 int vtable_index = methodOopDesc::invalid_vtable_index;
3414 3414 if (UseInlineCaches) {
3415 3415 // Suppress the vtable call
3416 3416 } else {
3417 3417 // hashCode and clone are not a miranda methods,
3418 3418 // so the vtable index is fixed.
3419 3419 // No need to use the linkResolver to get it.
3420 3420 vtable_index = method->vtable_index();
3421 3421 }
3422 3422 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3423 3423 SharedRuntime::get_resolve_virtual_call_stub(),
3424 3424 method, vtable_index, bci());
3425 3425 } else { // neither virtual nor static: opt_virtual
3426 3426 null_check_receiver(method);
3427 3427 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3428 3428 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3429 3429 method, bci());
3430 3430 slow_call->set_optimized_virtual(true);
3431 3431 }
3432 3432 set_arguments_for_java_call(slow_call);
3433 3433 set_edges_for_java_call(slow_call);
3434 3434 return slow_call;
3435 3435 }
3436 3436
3437 3437
3438 3438 //------------------------------inline_native_hashcode--------------------
3439 3439 // Build special case code for calls to hashCode on an object.
3440 3440 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3441 3441 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3442 3442 assert(!(is_virtual && is_static), "either virtual, special, or static");
3443 3443
3444 3444 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3445 3445
3446 3446 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3447 3447 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3448 3448 TypeInt::INT);
3449 3449 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3450 3450 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3451 3451 TypePtr::BOTTOM);
3452 3452 Node* obj = NULL;
3453 3453 if (!is_static) {
3454 3454 // Check for hashing null object
3455 3455 obj = null_check_receiver(callee());
3456 3456 if (stopped()) return true; // unconditionally null
3457 3457 result_reg->init_req(_null_path, top());
3458 3458 result_val->init_req(_null_path, top());
3459 3459 } else {
3460 3460 // Do a null check, and return zero if null.
3461 3461 // System.identityHashCode(null) == 0
3462 3462 obj = argument(0);
3463 3463 Node* null_ctl = top();
3464 3464 obj = null_check_oop(obj, &null_ctl);
3465 3465 result_reg->init_req(_null_path, null_ctl);
3466 3466 result_val->init_req(_null_path, _gvn.intcon(0));
3467 3467 }
3468 3468
3469 3469 // Unconditionally null? Then return right away.
3470 3470 if (stopped()) {
3471 3471 set_control( result_reg->in(_null_path) );
3472 3472 if (!stopped())
3473 3473 push( result_val ->in(_null_path) );
3474 3474 return true;
3475 3475 }
3476 3476
3477 3477 // After null check, get the object's klass.
3478 3478 Node* obj_klass = load_object_klass(obj);
3479 3479
3480 3480 // This call may be virtual (invokevirtual) or bound (invokespecial).
3481 3481 // For each case we generate slightly different code.
3482 3482
3483 3483 // We only go to the fast case code if we pass a number of guards. The
3484 3484 // paths which do not pass are accumulated in the slow_region.
3485 3485 RegionNode* slow_region = new (C, 1) RegionNode(1);
3486 3486 record_for_igvn(slow_region);
3487 3487
3488 3488 // If this is a virtual call, we generate a funny guard. We pull out
3489 3489 // the vtable entry corresponding to hashCode() from the target object.
3490 3490 // If the target method which we are calling happens to be the native
3491 3491 // Object hashCode() method, we pass the guard. We do not need this
3492 3492 // guard for non-virtual calls -- the caller is known to be the native
3493 3493 // Object hashCode().
3494 3494 if (is_virtual) {
3495 3495 generate_virtual_guard(obj_klass, slow_region);
3496 3496 }
3497 3497
3498 3498 // Get the header out of the object, use LoadMarkNode when available
3499 3499 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3500 3500 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
3501 3501 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
3502 3502
3503 3503 // Test the header to see if it is unlocked.
3504 3504 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3505 3505 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3506 3506 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3507 3507 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3508 3508 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3509 3509
3510 3510 generate_slow_guard(test_unlocked, slow_region);
3511 3511
3512 3512 // Get the hash value and check to see that it has been properly assigned.
3513 3513 // We depend on hash_mask being at most 32 bits and avoid the use of
3514 3514 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3515 3515 // vm: see markOop.hpp.
3516 3516 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3517 3517 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3518 3518 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3519 3519 // This hack lets the hash bits live anywhere in the mark object now, as long
3520 3520 // as the shift drops the relevant bits into the low 32 bits. Note that
3521 3521 // Java spec says that HashCode is an int so there's no point in capturing
3522 3522 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3523 3523 hshifted_header = ConvX2I(hshifted_header);
3524 3524 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3525 3525
3526 3526 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3527 3527 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3528 3528 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3529 3529
3530 3530 generate_slow_guard(test_assigned, slow_region);
3531 3531
3532 3532 Node* init_mem = reset_memory();
3533 3533 // fill in the rest of the null path:
3534 3534 result_io ->init_req(_null_path, i_o());
3535 3535 result_mem->init_req(_null_path, init_mem);
3536 3536
3537 3537 result_val->init_req(_fast_path, hash_val);
3538 3538 result_reg->init_req(_fast_path, control());
3539 3539 result_io ->init_req(_fast_path, i_o());
3540 3540 result_mem->init_req(_fast_path, init_mem);
3541 3541
3542 3542 // Generate code for the slow case. We make a call to hashCode().
3543 3543 set_control(_gvn.transform(slow_region));
3544 3544 if (!stopped()) {
3545 3545 // No need for PreserveJVMState, because we're using up the present state.
3546 3546 set_all_memory(init_mem);
3547 3547 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3548 3548 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3549 3549 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3550 3550 Node* slow_result = set_results_for_java_call(slow_call);
3551 3551 // this->control() comes from set_results_for_java_call
3552 3552 result_reg->init_req(_slow_path, control());
3553 3553 result_val->init_req(_slow_path, slow_result);
3554 3554 result_io ->set_req(_slow_path, i_o());
3555 3555 result_mem ->set_req(_slow_path, reset_memory());
3556 3556 }
3557 3557
3558 3558 // Return the combined state.
3559 3559 set_i_o( _gvn.transform(result_io) );
3560 3560 set_all_memory( _gvn.transform(result_mem) );
3561 3561 push_result(result_reg, result_val);
3562 3562
3563 3563 return true;
3564 3564 }
3565 3565
3566 3566 //---------------------------inline_native_getClass----------------------------
3567 3567 // Build special case code for calls to getClass on an object.
3568 3568 bool LibraryCallKit::inline_native_getClass() {
3569 3569 Node* obj = null_check_receiver(callee());
3570 3570 if (stopped()) return true;
3571 3571 push( load_mirror_from_klass(load_object_klass(obj)) );
3572 3572 return true;
3573 3573 }
3574 3574
3575 3575 //-----------------inline_native_Reflection_getCallerClass---------------------
3576 3576 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3577 3577 //
3578 3578 // NOTE that this code must perform the same logic as
3579 3579 // vframeStream::security_get_caller_frame in that it must skip
3580 3580 // Method.invoke() and auxiliary frames.
3581 3581
3582 3582
3583 3583
3584 3584
3585 3585 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3586 3586 ciMethod* method = callee();
3587 3587
3588 3588 #ifndef PRODUCT
3589 3589 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3590 3590 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3591 3591 }
3592 3592 #endif
3593 3593
3594 3594 debug_only(int saved_sp = _sp);
3595 3595
3596 3596 // Argument words: (int depth)
3597 3597 int nargs = 1;
3598 3598
3599 3599 _sp += nargs;
3600 3600 Node* caller_depth_node = pop();
3601 3601
3602 3602 assert(saved_sp == _sp, "must have correct argument count");
3603 3603
3604 3604 // The depth value must be a constant in order for the runtime call
3605 3605 // to be eliminated.
3606 3606 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3607 3607 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3608 3608 #ifndef PRODUCT
3609 3609 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3610 3610 tty->print_cr(" Bailing out because caller depth was not a constant");
3611 3611 }
3612 3612 #endif
3613 3613 return false;
3614 3614 }
3615 3615 // Note that the JVM state at this point does not include the
3616 3616 // getCallerClass() frame which we are trying to inline. The
3617 3617 // semantics of getCallerClass(), however, are that the "first"
3618 3618 // frame is the getCallerClass() frame, so we subtract one from the
3619 3619 // requested depth before continuing. We don't inline requests of
3620 3620 // getCallerClass(0).
3621 3621 int caller_depth = caller_depth_type->get_con() - 1;
3622 3622 if (caller_depth < 0) {
3623 3623 #ifndef PRODUCT
3624 3624 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3625 3625 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3626 3626 }
3627 3627 #endif
3628 3628 return false;
3629 3629 }
3630 3630
3631 3631 if (!jvms()->has_method()) {
3632 3632 #ifndef PRODUCT
3633 3633 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3634 3634 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3635 3635 }
3636 3636 #endif
3637 3637 return false;
3638 3638 }
3639 3639 int _depth = jvms()->depth(); // cache call chain depth
3640 3640
3641 3641 // Walk back up the JVM state to find the caller at the required
3642 3642 // depth. NOTE that this code must perform the same logic as
3643 3643 // vframeStream::security_get_caller_frame in that it must skip
3644 3644 // Method.invoke() and auxiliary frames. Note also that depth is
3645 3645 // 1-based (1 is the bottom of the inlining).
3646 3646 int inlining_depth = _depth;
3647 3647 JVMState* caller_jvms = NULL;
3648 3648
3649 3649 if (inlining_depth > 0) {
3650 3650 caller_jvms = jvms();
3651 3651 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3652 3652 do {
3653 3653 // The following if-tests should be performed in this order
3654 3654 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3655 3655 // Skip a Method.invoke() or auxiliary frame
3656 3656 } else if (caller_depth > 0) {
3657 3657 // Skip real frame
3658 3658 --caller_depth;
3659 3659 } else {
3660 3660 // We're done: reached desired caller after skipping.
3661 3661 break;
3662 3662 }
3663 3663 caller_jvms = caller_jvms->caller();
3664 3664 --inlining_depth;
3665 3665 } while (inlining_depth > 0);
3666 3666 }
3667 3667
3668 3668 if (inlining_depth == 0) {
3669 3669 #ifndef PRODUCT
3670 3670 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3671 3671 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3672 3672 tty->print_cr(" JVM state at this point:");
3673 3673 for (int i = _depth; i >= 1; i--) {
3674 3674 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3675 3675 }
3676 3676 }
3677 3677 #endif
3678 3678 return false; // Reached end of inlining
3679 3679 }
3680 3680
3681 3681 // Acquire method holder as java.lang.Class
3682 3682 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3683 3683 ciInstance* caller_mirror = caller_klass->java_mirror();
3684 3684 // Push this as a constant
3685 3685 push(makecon(TypeInstPtr::make(caller_mirror)));
3686 3686 #ifndef PRODUCT
3687 3687 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3688 3688 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);
3689 3689 tty->print_cr(" JVM state at this point:");
↓ open down ↓ |
3677 lines elided |
↑ open up ↑ |
3690 3690 for (int i = _depth; i >= 1; i--) {
3691 3691 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3692 3692 }
3693 3693 }
3694 3694 #endif
3695 3695 return true;
3696 3696 }
3697 3697
3698 3698 // Helper routine for above
3699 3699 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3700 + ciMethod* method = jvms->method();
3701 +
3700 3702 // Is this the Method.invoke method itself?
3701 - if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
3703 + if (method->intrinsic_id() == vmIntrinsics::_invoke)
3702 3704 return true;
3703 3705
3704 3706 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3705 - ciKlass* k = jvms->method()->holder();
3707 + ciKlass* k = method->holder();
3706 3708 if (k->is_instance_klass()) {
3707 3709 ciInstanceKlass* ik = k->as_instance_klass();
3708 3710 for (; ik != NULL; ik = ik->super()) {
3709 3711 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3710 3712 ik == env()->find_system_klass(ik->name())) {
3711 3713 return true;
3712 3714 }
3713 3715 }
3714 3716 }
3717 + else if (method->is_method_handle_adapter()) {
3718 + // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3719 + return true;
3720 + }
3715 3721
3716 3722 return false;
3717 3723 }
3718 3724
3719 3725 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3720 3726 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3721 3727 // computing it since there is no lookup field by name function in the
3722 3728 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3723 3729 // Using a static variable here is safe even if we have multiple compilation
3724 3730 // threads because the offset is constant. At worst the same offset will be
3725 3731 // computed and stored multiple
3726 3732
3727 3733 bool LibraryCallKit::inline_native_AtomicLong_get() {
3728 3734 // Restore the stack and pop off the argument
3729 3735 _sp+=1;
3730 3736 Node *obj = pop();
3731 3737
3732 3738 // get the offset of the "value" field. Since the CI interfaces
3733 3739 // does not provide a way to look up a field by name, we scan the bytecodes
3734 3740 // to get the field index. We expect the first 2 instructions of the method
3735 3741 // to be:
3736 3742 // 0 aload_0
3737 3743 // 1 getfield "value"
3738 3744 ciMethod* method = callee();
3739 3745 if (value_field_offset == -1)
3740 3746 {
3741 3747 ciField* value_field;
3742 3748 ciBytecodeStream iter(method);
3743 3749 Bytecodes::Code bc = iter.next();
3744 3750
3745 3751 if ((bc != Bytecodes::_aload_0) &&
3746 3752 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3747 3753 return false;
3748 3754 bc = iter.next();
3749 3755 if (bc != Bytecodes::_getfield)
3750 3756 return false;
3751 3757 bool ignore;
3752 3758 value_field = iter.get_field(ignore);
3753 3759 value_field_offset = value_field->offset_in_bytes();
3754 3760 }
3755 3761
3756 3762 // Null check without removing any arguments.
3757 3763 _sp++;
3758 3764 obj = do_null_check(obj, T_OBJECT);
3759 3765 _sp--;
3760 3766 // Check for locking null object
3761 3767 if (stopped()) return true;
3762 3768
3763 3769 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3764 3770 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3765 3771 int alias_idx = C->get_alias_index(adr_type);
3766 3772
3767 3773 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3768 3774
3769 3775 push_pair(result);
3770 3776
3771 3777 return true;
3772 3778 }
3773 3779
3774 3780 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3775 3781 // Restore the stack and pop off the arguments
3776 3782 _sp+=5;
3777 3783 Node *newVal = pop_pair();
3778 3784 Node *oldVal = pop_pair();
3779 3785 Node *obj = pop();
3780 3786
3781 3787 // we need the offset of the "value" field which was computed when
3782 3788 // inlining the get() method. Give up if we don't have it.
3783 3789 if (value_field_offset == -1)
3784 3790 return false;
3785 3791
3786 3792 // Null check without removing any arguments.
3787 3793 _sp+=5;
3788 3794 obj = do_null_check(obj, T_OBJECT);
3789 3795 _sp-=5;
3790 3796 // Check for locking null object
3791 3797 if (stopped()) return true;
3792 3798
3793 3799 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3794 3800 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3795 3801 int alias_idx = C->get_alias_index(adr_type);
3796 3802
3797 3803 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3798 3804 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3799 3805 set_memory(store_proj, alias_idx);
3800 3806 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3801 3807
3802 3808 Node *result;
3803 3809 // CMove node is not used to be able fold a possible check code
3804 3810 // after attemptUpdate() call. This code could be transformed
3805 3811 // into CMove node by loop optimizations.
3806 3812 {
3807 3813 RegionNode *r = new (C, 3) RegionNode(3);
3808 3814 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3809 3815
3810 3816 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3811 3817 Node *iftrue = opt_iff(r, iff);
3812 3818 r->init_req(1, iftrue);
3813 3819 result->init_req(1, intcon(1));
3814 3820 result->init_req(2, intcon(0));
3815 3821
3816 3822 set_control(_gvn.transform(r));
3817 3823 record_for_igvn(r);
3818 3824
3819 3825 C->set_has_split_ifs(true); // Has chance for split-if optimization
3820 3826 }
3821 3827
3822 3828 push(_gvn.transform(result));
3823 3829 return true;
3824 3830 }
3825 3831
3826 3832 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3827 3833 // restore the arguments
3828 3834 _sp += arg_size();
3829 3835
3830 3836 switch (id) {
3831 3837 case vmIntrinsics::_floatToRawIntBits:
3832 3838 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3833 3839 break;
3834 3840
3835 3841 case vmIntrinsics::_intBitsToFloat:
3836 3842 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3837 3843 break;
3838 3844
3839 3845 case vmIntrinsics::_doubleToRawLongBits:
3840 3846 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3841 3847 break;
3842 3848
3843 3849 case vmIntrinsics::_longBitsToDouble:
3844 3850 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3845 3851 break;
3846 3852
3847 3853 case vmIntrinsics::_doubleToLongBits: {
3848 3854 Node* value = pop_pair();
3849 3855
3850 3856 // two paths (plus control) merge in a wood
3851 3857 RegionNode *r = new (C, 3) RegionNode(3);
3852 3858 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3853 3859
3854 3860 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3855 3861 // Build the boolean node
3856 3862 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3857 3863
3858 3864 // Branch either way.
3859 3865 // NaN case is less traveled, which makes all the difference.
3860 3866 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3861 3867 Node *opt_isnan = _gvn.transform(ifisnan);
3862 3868 assert( opt_isnan->is_If(), "Expect an IfNode");
3863 3869 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3864 3870 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3865 3871
3866 3872 set_control(iftrue);
3867 3873
3868 3874 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3869 3875 Node *slow_result = longcon(nan_bits); // return NaN
3870 3876 phi->init_req(1, _gvn.transform( slow_result ));
3871 3877 r->init_req(1, iftrue);
3872 3878
3873 3879 // Else fall through
3874 3880 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3875 3881 set_control(iffalse);
3876 3882
3877 3883 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3878 3884 r->init_req(2, iffalse);
3879 3885
3880 3886 // Post merge
3881 3887 set_control(_gvn.transform(r));
3882 3888 record_for_igvn(r);
3883 3889
3884 3890 Node* result = _gvn.transform(phi);
3885 3891 assert(result->bottom_type()->isa_long(), "must be");
3886 3892 push_pair(result);
3887 3893
3888 3894 C->set_has_split_ifs(true); // Has chance for split-if optimization
3889 3895
3890 3896 break;
3891 3897 }
3892 3898
3893 3899 case vmIntrinsics::_floatToIntBits: {
3894 3900 Node* value = pop();
3895 3901
3896 3902 // two paths (plus control) merge in a wood
3897 3903 RegionNode *r = new (C, 3) RegionNode(3);
3898 3904 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3899 3905
3900 3906 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3901 3907 // Build the boolean node
3902 3908 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3903 3909
3904 3910 // Branch either way.
3905 3911 // NaN case is less traveled, which makes all the difference.
3906 3912 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3907 3913 Node *opt_isnan = _gvn.transform(ifisnan);
3908 3914 assert( opt_isnan->is_If(), "Expect an IfNode");
3909 3915 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3910 3916 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3911 3917
3912 3918 set_control(iftrue);
3913 3919
3914 3920 static const jint nan_bits = 0x7fc00000;
3915 3921 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3916 3922 phi->init_req(1, _gvn.transform( slow_result ));
3917 3923 r->init_req(1, iftrue);
3918 3924
3919 3925 // Else fall through
3920 3926 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3921 3927 set_control(iffalse);
3922 3928
3923 3929 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3924 3930 r->init_req(2, iffalse);
3925 3931
3926 3932 // Post merge
3927 3933 set_control(_gvn.transform(r));
3928 3934 record_for_igvn(r);
3929 3935
3930 3936 Node* result = _gvn.transform(phi);
3931 3937 assert(result->bottom_type()->isa_int(), "must be");
3932 3938 push(result);
3933 3939
3934 3940 C->set_has_split_ifs(true); // Has chance for split-if optimization
3935 3941
3936 3942 break;
3937 3943 }
3938 3944
3939 3945 default:
3940 3946 ShouldNotReachHere();
3941 3947 }
3942 3948
3943 3949 return true;
3944 3950 }
3945 3951
3946 3952 #ifdef _LP64
3947 3953 #define XTOP ,top() /*additional argument*/
3948 3954 #else //_LP64
3949 3955 #define XTOP /*no additional argument*/
3950 3956 #endif //_LP64
3951 3957
3952 3958 //----------------------inline_unsafe_copyMemory-------------------------
3953 3959 bool LibraryCallKit::inline_unsafe_copyMemory() {
3954 3960 if (callee()->is_static()) return false; // caller must have the capability!
3955 3961 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
3956 3962 assert(signature()->size() == nargs-1, "copy has 5 arguments");
3957 3963 null_check_receiver(callee()); // check then ignore argument(0)
3958 3964 if (stopped()) return true;
3959 3965
3960 3966 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3961 3967
3962 3968 Node* src_ptr = argument(1);
3963 3969 Node* src_off = ConvL2X(argument(2));
3964 3970 assert(argument(3)->is_top(), "2nd half of long");
3965 3971 Node* dst_ptr = argument(4);
3966 3972 Node* dst_off = ConvL2X(argument(5));
3967 3973 assert(argument(6)->is_top(), "2nd half of long");
3968 3974 Node* size = ConvL2X(argument(7));
3969 3975 assert(argument(8)->is_top(), "2nd half of long");
3970 3976
3971 3977 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
3972 3978 "fieldOffset must be byte-scaled");
3973 3979
3974 3980 Node* src = make_unsafe_address(src_ptr, src_off);
3975 3981 Node* dst = make_unsafe_address(dst_ptr, dst_off);
3976 3982
3977 3983 // Conservatively insert a memory barrier on all memory slices.
3978 3984 // Do not let writes of the copy source or destination float below the copy.
3979 3985 insert_mem_bar(Op_MemBarCPUOrder);
3980 3986
3981 3987 // Call it. Note that the length argument is not scaled.
3982 3988 make_runtime_call(RC_LEAF|RC_NO_FP,
3983 3989 OptoRuntime::fast_arraycopy_Type(),
3984 3990 StubRoutines::unsafe_arraycopy(),
3985 3991 "unsafe_arraycopy",
3986 3992 TypeRawPtr::BOTTOM,
3987 3993 src, dst, size XTOP);
3988 3994
3989 3995 // Do not let reads of the copy destination float above the copy.
3990 3996 insert_mem_bar(Op_MemBarCPUOrder);
3991 3997
3992 3998 return true;
3993 3999 }
3994 4000
3995 4001 //------------------------clone_coping-----------------------------------
3996 4002 // Helper function for inline_native_clone.
3997 4003 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
3998 4004 assert(obj_size != NULL, "");
3999 4005 Node* raw_obj = alloc_obj->in(1);
4000 4006 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4001 4007
4002 4008 if (ReduceBulkZeroing) {
4003 4009 // We will be completely responsible for initializing this object -
4004 4010 // mark Initialize node as complete.
4005 4011 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4006 4012 // The object was just allocated - there should be no any stores!
4007 4013 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4008 4014 }
4009 4015
4010 4016 // Copy the fastest available way.
4011 4017 // TODO: generate fields copies for small objects instead.
4012 4018 Node* src = obj;
4013 4019 Node* dest = alloc_obj;
4014 4020 Node* size = _gvn.transform(obj_size);
4015 4021
4016 4022 // Exclude the header but include array length to copy by 8 bytes words.
4017 4023 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4018 4024 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4019 4025 instanceOopDesc::base_offset_in_bytes();
4020 4026 // base_off:
4021 4027 // 8 - 32-bit VM
4022 4028 // 12 - 64-bit VM, compressed oops
4023 4029 // 16 - 64-bit VM, normal oops
4024 4030 if (base_off % BytesPerLong != 0) {
4025 4031 assert(UseCompressedOops, "");
4026 4032 if (is_array) {
4027 4033 // Exclude length to copy by 8 bytes words.
4028 4034 base_off += sizeof(int);
4029 4035 } else {
4030 4036 // Include klass to copy by 8 bytes words.
4031 4037 base_off = instanceOopDesc::klass_offset_in_bytes();
4032 4038 }
4033 4039 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4034 4040 }
4035 4041 src = basic_plus_adr(src, base_off);
4036 4042 dest = basic_plus_adr(dest, base_off);
4037 4043
4038 4044 // Compute the length also, if needed:
4039 4045 Node* countx = size;
4040 4046 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4041 4047 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4042 4048
4043 4049 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4044 4050 bool disjoint_bases = true;
4045 4051 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4046 4052 src, NULL, dest, NULL, countx);
4047 4053
4048 4054 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4049 4055 if (card_mark) {
4050 4056 assert(!is_array, "");
4051 4057 // Put in store barrier for any and all oops we are sticking
4052 4058 // into this object. (We could avoid this if we could prove
4053 4059 // that the object type contains no oop fields at all.)
4054 4060 Node* no_particular_value = NULL;
4055 4061 Node* no_particular_field = NULL;
4056 4062 int raw_adr_idx = Compile::AliasIdxRaw;
4057 4063 post_barrier(control(),
4058 4064 memory(raw_adr_type),
4059 4065 alloc_obj,
4060 4066 no_particular_field,
4061 4067 raw_adr_idx,
4062 4068 no_particular_value,
4063 4069 T_OBJECT,
4064 4070 false);
4065 4071 }
4066 4072
4067 4073 // Do not let reads from the cloned object float above the arraycopy.
4068 4074 insert_mem_bar(Op_MemBarCPUOrder);
4069 4075 }
4070 4076
4071 4077 //------------------------inline_native_clone----------------------------
4072 4078 // Here are the simple edge cases:
4073 4079 // null receiver => normal trap
4074 4080 // virtual and clone was overridden => slow path to out-of-line clone
4075 4081 // not cloneable or finalizer => slow path to out-of-line Object.clone
4076 4082 //
4077 4083 // The general case has two steps, allocation and copying.
4078 4084 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4079 4085 //
4080 4086 // Copying also has two cases, oop arrays and everything else.
4081 4087 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4082 4088 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4083 4089 //
4084 4090 // These steps fold up nicely if and when the cloned object's klass
4085 4091 // can be sharply typed as an object array, a type array, or an instance.
4086 4092 //
4087 4093 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4088 4094 int nargs = 1;
4089 4095 PhiNode* result_val;
4090 4096
4091 4097 //set the original stack and the reexecute bit for the interpreter to reexecute
4092 4098 //the bytecode that invokes Object.clone if deoptimization happens
4093 4099 { PreserveReexecuteState preexecs(this);
4094 4100 jvms()->set_should_reexecute(true);
4095 4101
4096 4102 //null_check_receiver will adjust _sp (push and pop)
4097 4103 Node* obj = null_check_receiver(callee());
4098 4104 if (stopped()) return true;
4099 4105
4100 4106 _sp += nargs;
4101 4107
4102 4108 Node* obj_klass = load_object_klass(obj);
4103 4109 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4104 4110 const TypeOopPtr* toop = ((tklass != NULL)
4105 4111 ? tklass->as_instance_type()
4106 4112 : TypeInstPtr::NOTNULL);
4107 4113
4108 4114 // Conservatively insert a memory barrier on all memory slices.
4109 4115 // Do not let writes into the original float below the clone.
4110 4116 insert_mem_bar(Op_MemBarCPUOrder);
4111 4117
4112 4118 // paths into result_reg:
4113 4119 enum {
4114 4120 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4115 4121 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4116 4122 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4117 4123 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4118 4124 PATH_LIMIT
4119 4125 };
4120 4126 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4121 4127 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4122 4128 TypeInstPtr::NOTNULL);
4123 4129 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4124 4130 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4125 4131 TypePtr::BOTTOM);
4126 4132 record_for_igvn(result_reg);
4127 4133
4128 4134 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4129 4135 int raw_adr_idx = Compile::AliasIdxRaw;
4130 4136 const bool raw_mem_only = true;
4131 4137
4132 4138
4133 4139 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4134 4140 if (array_ctl != NULL) {
4135 4141 // It's an array.
4136 4142 PreserveJVMState pjvms(this);
4137 4143 set_control(array_ctl);
4138 4144 Node* obj_length = load_array_length(obj);
4139 4145 Node* obj_size = NULL;
4140 4146 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4141 4147 raw_mem_only, &obj_size);
4142 4148
4143 4149 if (!use_ReduceInitialCardMarks()) {
4144 4150 // If it is an oop array, it requires very special treatment,
4145 4151 // because card marking is required on each card of the array.
4146 4152 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4147 4153 if (is_obja != NULL) {
4148 4154 PreserveJVMState pjvms2(this);
4149 4155 set_control(is_obja);
4150 4156 // Generate a direct call to the right arraycopy function(s).
4151 4157 bool disjoint_bases = true;
4152 4158 bool length_never_negative = true;
4153 4159 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4154 4160 obj, intcon(0), alloc_obj, intcon(0),
4155 4161 obj_length,
4156 4162 disjoint_bases, length_never_negative);
4157 4163 result_reg->init_req(_objArray_path, control());
4158 4164 result_val->init_req(_objArray_path, alloc_obj);
4159 4165 result_i_o ->set_req(_objArray_path, i_o());
4160 4166 result_mem ->set_req(_objArray_path, reset_memory());
4161 4167 }
4162 4168 }
4163 4169 // Otherwise, there are no card marks to worry about.
4164 4170 // (We can dispense with card marks if we know the allocation
4165 4171 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4166 4172 // causes the non-eden paths to take compensating steps to
4167 4173 // simulate a fresh allocation, so that no further
4168 4174 // card marks are required in compiled code to initialize
4169 4175 // the object.)
4170 4176
4171 4177 if (!stopped()) {
4172 4178 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4173 4179
4174 4180 // Present the results of the copy.
4175 4181 result_reg->init_req(_array_path, control());
4176 4182 result_val->init_req(_array_path, alloc_obj);
4177 4183 result_i_o ->set_req(_array_path, i_o());
4178 4184 result_mem ->set_req(_array_path, reset_memory());
4179 4185 }
4180 4186 }
4181 4187
4182 4188 // We only go to the instance fast case code if we pass a number of guards.
4183 4189 // The paths which do not pass are accumulated in the slow_region.
4184 4190 RegionNode* slow_region = new (C, 1) RegionNode(1);
4185 4191 record_for_igvn(slow_region);
4186 4192 if (!stopped()) {
4187 4193 // It's an instance (we did array above). Make the slow-path tests.
4188 4194 // If this is a virtual call, we generate a funny guard. We grab
4189 4195 // the vtable entry corresponding to clone() from the target object.
4190 4196 // If the target method which we are calling happens to be the
4191 4197 // Object clone() method, we pass the guard. We do not need this
4192 4198 // guard for non-virtual calls; the caller is known to be the native
4193 4199 // Object clone().
4194 4200 if (is_virtual) {
4195 4201 generate_virtual_guard(obj_klass, slow_region);
4196 4202 }
4197 4203
4198 4204 // The object must be cloneable and must not have a finalizer.
4199 4205 // Both of these conditions may be checked in a single test.
4200 4206 // We could optimize the cloneable test further, but we don't care.
4201 4207 generate_access_flags_guard(obj_klass,
4202 4208 // Test both conditions:
4203 4209 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4204 4210 // Must be cloneable but not finalizer:
4205 4211 JVM_ACC_IS_CLONEABLE,
4206 4212 slow_region);
4207 4213 }
4208 4214
4209 4215 if (!stopped()) {
4210 4216 // It's an instance, and it passed the slow-path tests.
4211 4217 PreserveJVMState pjvms(this);
4212 4218 Node* obj_size = NULL;
4213 4219 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4214 4220
4215 4221 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4216 4222
4217 4223 // Present the results of the slow call.
4218 4224 result_reg->init_req(_instance_path, control());
4219 4225 result_val->init_req(_instance_path, alloc_obj);
4220 4226 result_i_o ->set_req(_instance_path, i_o());
4221 4227 result_mem ->set_req(_instance_path, reset_memory());
4222 4228 }
4223 4229
4224 4230 // Generate code for the slow case. We make a call to clone().
4225 4231 set_control(_gvn.transform(slow_region));
4226 4232 if (!stopped()) {
4227 4233 PreserveJVMState pjvms(this);
4228 4234 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4229 4235 Node* slow_result = set_results_for_java_call(slow_call);
4230 4236 // this->control() comes from set_results_for_java_call
4231 4237 result_reg->init_req(_slow_path, control());
4232 4238 result_val->init_req(_slow_path, slow_result);
4233 4239 result_i_o ->set_req(_slow_path, i_o());
4234 4240 result_mem ->set_req(_slow_path, reset_memory());
4235 4241 }
4236 4242
4237 4243 // Return the combined state.
4238 4244 set_control( _gvn.transform(result_reg) );
4239 4245 set_i_o( _gvn.transform(result_i_o) );
4240 4246 set_all_memory( _gvn.transform(result_mem) );
4241 4247 } //original reexecute and sp are set back here
4242 4248
4243 4249 push(_gvn.transform(result_val));
4244 4250
4245 4251 return true;
4246 4252 }
4247 4253
4248 4254
4249 4255 // constants for computing the copy function
4250 4256 enum {
4251 4257 COPYFUNC_UNALIGNED = 0,
4252 4258 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4253 4259 COPYFUNC_CONJOINT = 0,
4254 4260 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4255 4261 };
4256 4262
4257 4263 // Note: The condition "disjoint" applies also for overlapping copies
4258 4264 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4259 4265 static address
4260 4266 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
4261 4267 int selector =
4262 4268 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4263 4269 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4264 4270
4265 4271 #define RETURN_STUB(xxx_arraycopy) { \
4266 4272 name = #xxx_arraycopy; \
4267 4273 return StubRoutines::xxx_arraycopy(); }
4268 4274
4269 4275 switch (t) {
4270 4276 case T_BYTE:
4271 4277 case T_BOOLEAN:
4272 4278 switch (selector) {
4273 4279 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4274 4280 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4275 4281 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4276 4282 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4277 4283 }
4278 4284 case T_CHAR:
4279 4285 case T_SHORT:
4280 4286 switch (selector) {
4281 4287 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4282 4288 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4283 4289 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4284 4290 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4285 4291 }
4286 4292 case T_INT:
4287 4293 case T_FLOAT:
4288 4294 switch (selector) {
4289 4295 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4290 4296 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4291 4297 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4292 4298 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4293 4299 }
4294 4300 case T_DOUBLE:
4295 4301 case T_LONG:
4296 4302 switch (selector) {
4297 4303 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4298 4304 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4299 4305 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4300 4306 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4301 4307 }
4302 4308 case T_ARRAY:
4303 4309 case T_OBJECT:
4304 4310 switch (selector) {
4305 4311 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
4306 4312 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
4307 4313 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
4308 4314 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
4309 4315 }
4310 4316 default:
4311 4317 ShouldNotReachHere();
4312 4318 return NULL;
4313 4319 }
4314 4320
4315 4321 #undef RETURN_STUB
4316 4322 }
4317 4323
4318 4324 //------------------------------basictype2arraycopy----------------------------
4319 4325 address LibraryCallKit::basictype2arraycopy(BasicType t,
4320 4326 Node* src_offset,
4321 4327 Node* dest_offset,
4322 4328 bool disjoint_bases,
4323 4329 const char* &name) {
4324 4330 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4325 4331 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4326 4332
4327 4333 bool aligned = false;
4328 4334 bool disjoint = disjoint_bases;
4329 4335
4330 4336 // if the offsets are the same, we can treat the memory regions as
4331 4337 // disjoint, because either the memory regions are in different arrays,
4332 4338 // or they are identical (which we can treat as disjoint.) We can also
4333 4339 // treat a copy with a destination index less that the source index
4334 4340 // as disjoint since a low->high copy will work correctly in this case.
4335 4341 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4336 4342 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4337 4343 // both indices are constants
4338 4344 int s_offs = src_offset_inttype->get_con();
4339 4345 int d_offs = dest_offset_inttype->get_con();
4340 4346 int element_size = type2aelembytes(t);
4341 4347 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4342 4348 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4343 4349 if (s_offs >= d_offs) disjoint = true;
4344 4350 } else if (src_offset == dest_offset && src_offset != NULL) {
4345 4351 // This can occur if the offsets are identical non-constants.
4346 4352 disjoint = true;
4347 4353 }
4348 4354
4349 4355 return select_arraycopy_function(t, aligned, disjoint, name);
4350 4356 }
4351 4357
4352 4358
4353 4359 //------------------------------inline_arraycopy-----------------------
4354 4360 bool LibraryCallKit::inline_arraycopy() {
4355 4361 // Restore the stack and pop off the arguments.
4356 4362 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4357 4363 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4358 4364
4359 4365 Node *src = argument(0);
4360 4366 Node *src_offset = argument(1);
4361 4367 Node *dest = argument(2);
4362 4368 Node *dest_offset = argument(3);
4363 4369 Node *length = argument(4);
4364 4370
4365 4371 // Compile time checks. If any of these checks cannot be verified at compile time,
4366 4372 // we do not make a fast path for this call. Instead, we let the call remain as it
4367 4373 // is. The checks we choose to mandate at compile time are:
4368 4374 //
4369 4375 // (1) src and dest are arrays.
4370 4376 const Type* src_type = src->Value(&_gvn);
4371 4377 const Type* dest_type = dest->Value(&_gvn);
4372 4378 const TypeAryPtr* top_src = src_type->isa_aryptr();
4373 4379 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4374 4380 if (top_src == NULL || top_src->klass() == NULL ||
4375 4381 top_dest == NULL || top_dest->klass() == NULL) {
4376 4382 // Conservatively insert a memory barrier on all memory slices.
4377 4383 // Do not let writes into the source float below the arraycopy.
4378 4384 insert_mem_bar(Op_MemBarCPUOrder);
4379 4385
4380 4386 // Call StubRoutines::generic_arraycopy stub.
4381 4387 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4382 4388 src, src_offset, dest, dest_offset, length);
4383 4389
4384 4390 // Do not let reads from the destination float above the arraycopy.
4385 4391 // Since we cannot type the arrays, we don't know which slices
4386 4392 // might be affected. We could restrict this barrier only to those
4387 4393 // memory slices which pertain to array elements--but don't bother.
4388 4394 if (!InsertMemBarAfterArraycopy)
4389 4395 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4390 4396 insert_mem_bar(Op_MemBarCPUOrder);
4391 4397 return true;
4392 4398 }
4393 4399
4394 4400 // (2) src and dest arrays must have elements of the same BasicType
4395 4401 // Figure out the size and type of the elements we will be copying.
4396 4402 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4397 4403 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4398 4404 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4399 4405 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4400 4406
4401 4407 if (src_elem != dest_elem || dest_elem == T_VOID) {
4402 4408 // The component types are not the same or are not recognized. Punt.
4403 4409 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4404 4410 generate_slow_arraycopy(TypePtr::BOTTOM,
4405 4411 src, src_offset, dest, dest_offset, length);
4406 4412 return true;
4407 4413 }
4408 4414
4409 4415 //---------------------------------------------------------------------------
4410 4416 // We will make a fast path for this call to arraycopy.
4411 4417
4412 4418 // We have the following tests left to perform:
4413 4419 //
4414 4420 // (3) src and dest must not be null.
4415 4421 // (4) src_offset must not be negative.
4416 4422 // (5) dest_offset must not be negative.
4417 4423 // (6) length must not be negative.
4418 4424 // (7) src_offset + length must not exceed length of src.
4419 4425 // (8) dest_offset + length must not exceed length of dest.
4420 4426 // (9) each element of an oop array must be assignable
4421 4427
4422 4428 RegionNode* slow_region = new (C, 1) RegionNode(1);
4423 4429 record_for_igvn(slow_region);
4424 4430
4425 4431 // (3) operands must not be null
4426 4432 // We currently perform our null checks with the do_null_check routine.
4427 4433 // This means that the null exceptions will be reported in the caller
4428 4434 // rather than (correctly) reported inside of the native arraycopy call.
4429 4435 // This should be corrected, given time. We do our null check with the
4430 4436 // stack pointer restored.
4431 4437 _sp += nargs;
4432 4438 src = do_null_check(src, T_ARRAY);
4433 4439 dest = do_null_check(dest, T_ARRAY);
4434 4440 _sp -= nargs;
4435 4441
4436 4442 // (4) src_offset must not be negative.
4437 4443 generate_negative_guard(src_offset, slow_region);
4438 4444
4439 4445 // (5) dest_offset must not be negative.
4440 4446 generate_negative_guard(dest_offset, slow_region);
4441 4447
4442 4448 // (6) length must not be negative (moved to generate_arraycopy()).
4443 4449 // generate_negative_guard(length, slow_region);
4444 4450
4445 4451 // (7) src_offset + length must not exceed length of src.
4446 4452 generate_limit_guard(src_offset, length,
4447 4453 load_array_length(src),
4448 4454 slow_region);
4449 4455
4450 4456 // (8) dest_offset + length must not exceed length of dest.
4451 4457 generate_limit_guard(dest_offset, length,
4452 4458 load_array_length(dest),
4453 4459 slow_region);
4454 4460
4455 4461 // (9) each element of an oop array must be assignable
4456 4462 // The generate_arraycopy subroutine checks this.
4457 4463
4458 4464 // This is where the memory effects are placed:
4459 4465 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4460 4466 generate_arraycopy(adr_type, dest_elem,
4461 4467 src, src_offset, dest, dest_offset, length,
4462 4468 false, false, slow_region);
4463 4469
4464 4470 return true;
4465 4471 }
4466 4472
4467 4473 //-----------------------------generate_arraycopy----------------------
4468 4474 // Generate an optimized call to arraycopy.
4469 4475 // Caller must guard against non-arrays.
4470 4476 // Caller must determine a common array basic-type for both arrays.
4471 4477 // Caller must validate offsets against array bounds.
4472 4478 // The slow_region has already collected guard failure paths
4473 4479 // (such as out of bounds length or non-conformable array types).
4474 4480 // The generated code has this shape, in general:
4475 4481 //
4476 4482 // if (length == 0) return // via zero_path
4477 4483 // slowval = -1
4478 4484 // if (types unknown) {
4479 4485 // slowval = call generic copy loop
4480 4486 // if (slowval == 0) return // via checked_path
4481 4487 // } else if (indexes in bounds) {
4482 4488 // if ((is object array) && !(array type check)) {
4483 4489 // slowval = call checked copy loop
4484 4490 // if (slowval == 0) return // via checked_path
4485 4491 // } else {
4486 4492 // call bulk copy loop
4487 4493 // return // via fast_path
4488 4494 // }
4489 4495 // }
4490 4496 // // adjust params for remaining work:
4491 4497 // if (slowval != -1) {
4492 4498 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4493 4499 // }
4494 4500 // slow_region:
4495 4501 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4496 4502 // return // via slow_call_path
4497 4503 //
4498 4504 // This routine is used from several intrinsics: System.arraycopy,
4499 4505 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4500 4506 //
4501 4507 void
4502 4508 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4503 4509 BasicType basic_elem_type,
4504 4510 Node* src, Node* src_offset,
4505 4511 Node* dest, Node* dest_offset,
4506 4512 Node* copy_length,
4507 4513 bool disjoint_bases,
4508 4514 bool length_never_negative,
4509 4515 RegionNode* slow_region) {
4510 4516
4511 4517 if (slow_region == NULL) {
4512 4518 slow_region = new(C,1) RegionNode(1);
4513 4519 record_for_igvn(slow_region);
4514 4520 }
4515 4521
4516 4522 Node* original_dest = dest;
4517 4523 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4518 4524 bool must_clear_dest = false;
4519 4525
4520 4526 // See if this is the initialization of a newly-allocated array.
4521 4527 // If so, we will take responsibility here for initializing it to zero.
4522 4528 // (Note: Because tightly_coupled_allocation performs checks on the
4523 4529 // out-edges of the dest, we need to avoid making derived pointers
4524 4530 // from it until we have checked its uses.)
4525 4531 if (ReduceBulkZeroing
4526 4532 && !ZeroTLAB // pointless if already zeroed
4527 4533 && basic_elem_type != T_CONFLICT // avoid corner case
4528 4534 && !_gvn.eqv_uncast(src, dest)
4529 4535 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4530 4536 != NULL)
4531 4537 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4532 4538 && alloc->maybe_set_complete(&_gvn)) {
4533 4539 // "You break it, you buy it."
4534 4540 InitializeNode* init = alloc->initialization();
4535 4541 assert(init->is_complete(), "we just did this");
4536 4542 assert(dest->is_CheckCastPP(), "sanity");
4537 4543 assert(dest->in(0)->in(0) == init, "dest pinned");
4538 4544 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4539 4545 // From this point on, every exit path is responsible for
4540 4546 // initializing any non-copied parts of the object to zero.
4541 4547 must_clear_dest = true;
4542 4548 } else {
4543 4549 // No zeroing elimination here.
4544 4550 alloc = NULL;
4545 4551 //original_dest = dest;
4546 4552 //must_clear_dest = false;
4547 4553 }
4548 4554
4549 4555 // Results are placed here:
4550 4556 enum { fast_path = 1, // normal void-returning assembly stub
4551 4557 checked_path = 2, // special assembly stub with cleanup
4552 4558 slow_call_path = 3, // something went wrong; call the VM
4553 4559 zero_path = 4, // bypass when length of copy is zero
4554 4560 bcopy_path = 5, // copy primitive array by 64-bit blocks
4555 4561 PATH_LIMIT = 6
4556 4562 };
4557 4563 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4558 4564 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4559 4565 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4560 4566 record_for_igvn(result_region);
4561 4567 _gvn.set_type_bottom(result_i_o);
4562 4568 _gvn.set_type_bottom(result_memory);
4563 4569 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4564 4570
4565 4571 // The slow_control path:
4566 4572 Node* slow_control;
4567 4573 Node* slow_i_o = i_o();
4568 4574 Node* slow_mem = memory(adr_type);
4569 4575 debug_only(slow_control = (Node*) badAddress);
4570 4576
4571 4577 // Checked control path:
4572 4578 Node* checked_control = top();
4573 4579 Node* checked_mem = NULL;
4574 4580 Node* checked_i_o = NULL;
4575 4581 Node* checked_value = NULL;
4576 4582
4577 4583 if (basic_elem_type == T_CONFLICT) {
4578 4584 assert(!must_clear_dest, "");
4579 4585 Node* cv = generate_generic_arraycopy(adr_type,
4580 4586 src, src_offset, dest, dest_offset,
4581 4587 copy_length);
4582 4588 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4583 4589 checked_control = control();
4584 4590 checked_i_o = i_o();
4585 4591 checked_mem = memory(adr_type);
4586 4592 checked_value = cv;
4587 4593 set_control(top()); // no fast path
4588 4594 }
4589 4595
4590 4596 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4591 4597 if (not_pos != NULL) {
4592 4598 PreserveJVMState pjvms(this);
4593 4599 set_control(not_pos);
4594 4600
4595 4601 // (6) length must not be negative.
4596 4602 if (!length_never_negative) {
4597 4603 generate_negative_guard(copy_length, slow_region);
4598 4604 }
4599 4605
4600 4606 // copy_length is 0.
4601 4607 if (!stopped() && must_clear_dest) {
4602 4608 Node* dest_length = alloc->in(AllocateNode::ALength);
4603 4609 if (_gvn.eqv_uncast(copy_length, dest_length)
4604 4610 || _gvn.find_int_con(dest_length, 1) <= 0) {
4605 4611 // There is no zeroing to do. No need for a secondary raw memory barrier.
4606 4612 } else {
4607 4613 // Clear the whole thing since there are no source elements to copy.
4608 4614 generate_clear_array(adr_type, dest, basic_elem_type,
4609 4615 intcon(0), NULL,
4610 4616 alloc->in(AllocateNode::AllocSize));
4611 4617 // Use a secondary InitializeNode as raw memory barrier.
4612 4618 // Currently it is needed only on this path since other
4613 4619 // paths have stub or runtime calls as raw memory barriers.
4614 4620 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4615 4621 Compile::AliasIdxRaw,
4616 4622 top())->as_Initialize();
4617 4623 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4618 4624 }
4619 4625 }
4620 4626
4621 4627 // Present the results of the fast call.
4622 4628 result_region->init_req(zero_path, control());
4623 4629 result_i_o ->init_req(zero_path, i_o());
4624 4630 result_memory->init_req(zero_path, memory(adr_type));
4625 4631 }
4626 4632
4627 4633 if (!stopped() && must_clear_dest) {
4628 4634 // We have to initialize the *uncopied* part of the array to zero.
4629 4635 // The copy destination is the slice dest[off..off+len]. The other slices
4630 4636 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4631 4637 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4632 4638 Node* dest_length = alloc->in(AllocateNode::ALength);
4633 4639 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4634 4640 copy_length) );
4635 4641
4636 4642 // If there is a head section that needs zeroing, do it now.
4637 4643 if (find_int_con(dest_offset, -1) != 0) {
4638 4644 generate_clear_array(adr_type, dest, basic_elem_type,
4639 4645 intcon(0), dest_offset,
4640 4646 NULL);
4641 4647 }
4642 4648
4643 4649 // Next, perform a dynamic check on the tail length.
4644 4650 // It is often zero, and we can win big if we prove this.
4645 4651 // There are two wins: Avoid generating the ClearArray
4646 4652 // with its attendant messy index arithmetic, and upgrade
4647 4653 // the copy to a more hardware-friendly word size of 64 bits.
4648 4654 Node* tail_ctl = NULL;
4649 4655 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4650 4656 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4651 4657 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4652 4658 tail_ctl = generate_slow_guard(bol_lt, NULL);
4653 4659 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4654 4660 }
4655 4661
4656 4662 // At this point, let's assume there is no tail.
4657 4663 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4658 4664 // There is no tail. Try an upgrade to a 64-bit copy.
4659 4665 bool didit = false;
4660 4666 { PreserveJVMState pjvms(this);
4661 4667 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4662 4668 src, src_offset, dest, dest_offset,
4663 4669 dest_size);
4664 4670 if (didit) {
4665 4671 // Present the results of the block-copying fast call.
4666 4672 result_region->init_req(bcopy_path, control());
4667 4673 result_i_o ->init_req(bcopy_path, i_o());
4668 4674 result_memory->init_req(bcopy_path, memory(adr_type));
4669 4675 }
4670 4676 }
4671 4677 if (didit)
4672 4678 set_control(top()); // no regular fast path
4673 4679 }
4674 4680
4675 4681 // Clear the tail, if any.
4676 4682 if (tail_ctl != NULL) {
4677 4683 Node* notail_ctl = stopped() ? NULL : control();
4678 4684 set_control(tail_ctl);
4679 4685 if (notail_ctl == NULL) {
4680 4686 generate_clear_array(adr_type, dest, basic_elem_type,
4681 4687 dest_tail, NULL,
4682 4688 dest_size);
4683 4689 } else {
4684 4690 // Make a local merge.
4685 4691 Node* done_ctl = new(C,3) RegionNode(3);
4686 4692 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4687 4693 done_ctl->init_req(1, notail_ctl);
4688 4694 done_mem->init_req(1, memory(adr_type));
4689 4695 generate_clear_array(adr_type, dest, basic_elem_type,
4690 4696 dest_tail, NULL,
4691 4697 dest_size);
4692 4698 done_ctl->init_req(2, control());
4693 4699 done_mem->init_req(2, memory(adr_type));
4694 4700 set_control( _gvn.transform(done_ctl) );
4695 4701 set_memory( _gvn.transform(done_mem), adr_type );
4696 4702 }
4697 4703 }
4698 4704 }
4699 4705
4700 4706 BasicType copy_type = basic_elem_type;
4701 4707 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4702 4708 if (!stopped() && copy_type == T_OBJECT) {
4703 4709 // If src and dest have compatible element types, we can copy bits.
4704 4710 // Types S[] and D[] are compatible if D is a supertype of S.
4705 4711 //
4706 4712 // If they are not, we will use checked_oop_disjoint_arraycopy,
4707 4713 // which performs a fast optimistic per-oop check, and backs off
4708 4714 // further to JVM_ArrayCopy on the first per-oop check that fails.
4709 4715 // (Actually, we don't move raw bits only; the GC requires card marks.)
4710 4716
4711 4717 // Get the klassOop for both src and dest
4712 4718 Node* src_klass = load_object_klass(src);
4713 4719 Node* dest_klass = load_object_klass(dest);
4714 4720
4715 4721 // Generate the subtype check.
4716 4722 // This might fold up statically, or then again it might not.
4717 4723 //
4718 4724 // Non-static example: Copying List<String>.elements to a new String[].
4719 4725 // The backing store for a List<String> is always an Object[],
4720 4726 // but its elements are always type String, if the generic types
4721 4727 // are correct at the source level.
4722 4728 //
4723 4729 // Test S[] against D[], not S against D, because (probably)
4724 4730 // the secondary supertype cache is less busy for S[] than S.
4725 4731 // This usually only matters when D is an interface.
4726 4732 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4727 4733 // Plug failing path into checked_oop_disjoint_arraycopy
4728 4734 if (not_subtype_ctrl != top()) {
4729 4735 PreserveJVMState pjvms(this);
4730 4736 set_control(not_subtype_ctrl);
4731 4737 // (At this point we can assume disjoint_bases, since types differ.)
4732 4738 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4733 4739 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4734 4740 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4735 4741 Node* dest_elem_klass = _gvn.transform(n1);
4736 4742 Node* cv = generate_checkcast_arraycopy(adr_type,
4737 4743 dest_elem_klass,
4738 4744 src, src_offset, dest, dest_offset,
4739 4745 copy_length);
4740 4746 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4741 4747 checked_control = control();
4742 4748 checked_i_o = i_o();
4743 4749 checked_mem = memory(adr_type);
4744 4750 checked_value = cv;
4745 4751 }
4746 4752 // At this point we know we do not need type checks on oop stores.
4747 4753
4748 4754 // Let's see if we need card marks:
4749 4755 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4750 4756 // If we do not need card marks, copy using the jint or jlong stub.
4751 4757 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4752 4758 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4753 4759 "sizes agree");
4754 4760 }
4755 4761 }
4756 4762
4757 4763 if (!stopped()) {
4758 4764 // Generate the fast path, if possible.
4759 4765 PreserveJVMState pjvms(this);
4760 4766 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4761 4767 src, src_offset, dest, dest_offset,
4762 4768 ConvI2X(copy_length));
4763 4769
4764 4770 // Present the results of the fast call.
4765 4771 result_region->init_req(fast_path, control());
4766 4772 result_i_o ->init_req(fast_path, i_o());
4767 4773 result_memory->init_req(fast_path, memory(adr_type));
4768 4774 }
4769 4775
4770 4776 // Here are all the slow paths up to this point, in one bundle:
4771 4777 slow_control = top();
4772 4778 if (slow_region != NULL)
4773 4779 slow_control = _gvn.transform(slow_region);
4774 4780 debug_only(slow_region = (RegionNode*)badAddress);
4775 4781
4776 4782 set_control(checked_control);
4777 4783 if (!stopped()) {
4778 4784 // Clean up after the checked call.
4779 4785 // The returned value is either 0 or -1^K,
4780 4786 // where K = number of partially transferred array elements.
4781 4787 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4782 4788 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4783 4789 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4784 4790
4785 4791 // If it is 0, we are done, so transfer to the end.
4786 4792 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4787 4793 result_region->init_req(checked_path, checks_done);
4788 4794 result_i_o ->init_req(checked_path, checked_i_o);
4789 4795 result_memory->init_req(checked_path, checked_mem);
4790 4796
4791 4797 // If it is not zero, merge into the slow call.
4792 4798 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4793 4799 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4794 4800 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4795 4801 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4796 4802 record_for_igvn(slow_reg2);
4797 4803 slow_reg2 ->init_req(1, slow_control);
4798 4804 slow_i_o2 ->init_req(1, slow_i_o);
4799 4805 slow_mem2 ->init_req(1, slow_mem);
4800 4806 slow_reg2 ->init_req(2, control());
4801 4807 slow_i_o2 ->init_req(2, checked_i_o);
4802 4808 slow_mem2 ->init_req(2, checked_mem);
4803 4809
4804 4810 slow_control = _gvn.transform(slow_reg2);
4805 4811 slow_i_o = _gvn.transform(slow_i_o2);
4806 4812 slow_mem = _gvn.transform(slow_mem2);
4807 4813
4808 4814 if (alloc != NULL) {
4809 4815 // We'll restart from the very beginning, after zeroing the whole thing.
4810 4816 // This can cause double writes, but that's OK since dest is brand new.
4811 4817 // So we ignore the low 31 bits of the value returned from the stub.
4812 4818 } else {
4813 4819 // We must continue the copy exactly where it failed, or else
4814 4820 // another thread might see the wrong number of writes to dest.
4815 4821 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4816 4822 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4817 4823 slow_offset->init_req(1, intcon(0));
4818 4824 slow_offset->init_req(2, checked_offset);
4819 4825 slow_offset = _gvn.transform(slow_offset);
4820 4826
4821 4827 // Adjust the arguments by the conditionally incoming offset.
4822 4828 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4823 4829 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4824 4830 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4825 4831
4826 4832 // Tweak the node variables to adjust the code produced below:
4827 4833 src_offset = src_off_plus;
4828 4834 dest_offset = dest_off_plus;
4829 4835 copy_length = length_minus;
4830 4836 }
4831 4837 }
4832 4838
4833 4839 set_control(slow_control);
4834 4840 if (!stopped()) {
4835 4841 // Generate the slow path, if needed.
4836 4842 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4837 4843
4838 4844 set_memory(slow_mem, adr_type);
4839 4845 set_i_o(slow_i_o);
4840 4846
4841 4847 if (must_clear_dest) {
4842 4848 generate_clear_array(adr_type, dest, basic_elem_type,
4843 4849 intcon(0), NULL,
4844 4850 alloc->in(AllocateNode::AllocSize));
4845 4851 }
4846 4852
4847 4853 generate_slow_arraycopy(adr_type,
4848 4854 src, src_offset, dest, dest_offset,
4849 4855 copy_length);
4850 4856
4851 4857 result_region->init_req(slow_call_path, control());
4852 4858 result_i_o ->init_req(slow_call_path, i_o());
4853 4859 result_memory->init_req(slow_call_path, memory(adr_type));
4854 4860 }
4855 4861
4856 4862 // Remove unused edges.
4857 4863 for (uint i = 1; i < result_region->req(); i++) {
4858 4864 if (result_region->in(i) == NULL)
4859 4865 result_region->init_req(i, top());
4860 4866 }
4861 4867
4862 4868 // Finished; return the combined state.
4863 4869 set_control( _gvn.transform(result_region) );
4864 4870 set_i_o( _gvn.transform(result_i_o) );
4865 4871 set_memory( _gvn.transform(result_memory), adr_type );
4866 4872
4867 4873 // The memory edges above are precise in order to model effects around
4868 4874 // array copies accurately to allow value numbering of field loads around
4869 4875 // arraycopy. Such field loads, both before and after, are common in Java
4870 4876 // collections and similar classes involving header/array data structures.
4871 4877 //
4872 4878 // But with low number of register or when some registers are used or killed
4873 4879 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4874 4880 // The next memory barrier is added to avoid it. If the arraycopy can be
4875 4881 // optimized away (which it can, sometimes) then we can manually remove
4876 4882 // the membar also.
4877 4883 //
4878 4884 // Do not let reads from the cloned object float above the arraycopy.
4879 4885 if (InsertMemBarAfterArraycopy || alloc != NULL)
4880 4886 insert_mem_bar(Op_MemBarCPUOrder);
4881 4887 }
4882 4888
4883 4889
4884 4890 // Helper function which determines if an arraycopy immediately follows
4885 4891 // an allocation, with no intervening tests or other escapes for the object.
4886 4892 AllocateArrayNode*
4887 4893 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4888 4894 RegionNode* slow_region) {
4889 4895 if (stopped()) return NULL; // no fast path
4890 4896 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4891 4897
4892 4898 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4893 4899 if (alloc == NULL) return NULL;
4894 4900
4895 4901 Node* rawmem = memory(Compile::AliasIdxRaw);
4896 4902 // Is the allocation's memory state untouched?
4897 4903 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4898 4904 // Bail out if there have been raw-memory effects since the allocation.
4899 4905 // (Example: There might have been a call or safepoint.)
4900 4906 return NULL;
4901 4907 }
4902 4908 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4903 4909 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4904 4910 return NULL;
4905 4911 }
4906 4912
4907 4913 // There must be no unexpected observers of this allocation.
4908 4914 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4909 4915 Node* obs = ptr->fast_out(i);
4910 4916 if (obs != this->map()) {
4911 4917 return NULL;
4912 4918 }
4913 4919 }
4914 4920
4915 4921 // This arraycopy must unconditionally follow the allocation of the ptr.
4916 4922 Node* alloc_ctl = ptr->in(0);
4917 4923 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4918 4924
4919 4925 Node* ctl = control();
4920 4926 while (ctl != alloc_ctl) {
4921 4927 // There may be guards which feed into the slow_region.
4922 4928 // Any other control flow means that we might not get a chance
4923 4929 // to finish initializing the allocated object.
4924 4930 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4925 4931 IfNode* iff = ctl->in(0)->as_If();
4926 4932 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4927 4933 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4928 4934 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4929 4935 ctl = iff->in(0); // This test feeds the known slow_region.
4930 4936 continue;
4931 4937 }
4932 4938 // One more try: Various low-level checks bottom out in
4933 4939 // uncommon traps. If the debug-info of the trap omits
4934 4940 // any reference to the allocation, as we've already
4935 4941 // observed, then there can be no objection to the trap.
4936 4942 bool found_trap = false;
4937 4943 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4938 4944 Node* obs = not_ctl->fast_out(j);
4939 4945 if (obs->in(0) == not_ctl && obs->is_Call() &&
4940 4946 (obs->as_Call()->entry_point() ==
4941 4947 SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
4942 4948 found_trap = true; break;
4943 4949 }
4944 4950 }
4945 4951 if (found_trap) {
4946 4952 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4947 4953 continue;
4948 4954 }
4949 4955 }
4950 4956 return NULL;
4951 4957 }
4952 4958
4953 4959 // If we get this far, we have an allocation which immediately
4954 4960 // precedes the arraycopy, and we can take over zeroing the new object.
4955 4961 // The arraycopy will finish the initialization, and provide
4956 4962 // a new control state to which we will anchor the destination pointer.
4957 4963
4958 4964 return alloc;
4959 4965 }
4960 4966
4961 4967 // Helper for initialization of arrays, creating a ClearArray.
4962 4968 // It writes zero bits in [start..end), within the body of an array object.
4963 4969 // The memory effects are all chained onto the 'adr_type' alias category.
4964 4970 //
4965 4971 // Since the object is otherwise uninitialized, we are free
4966 4972 // to put a little "slop" around the edges of the cleared area,
4967 4973 // as long as it does not go back into the array's header,
4968 4974 // or beyond the array end within the heap.
4969 4975 //
4970 4976 // The lower edge can be rounded down to the nearest jint and the
4971 4977 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
4972 4978 //
4973 4979 // Arguments:
4974 4980 // adr_type memory slice where writes are generated
4975 4981 // dest oop of the destination array
4976 4982 // basic_elem_type element type of the destination
4977 4983 // slice_idx array index of first element to store
4978 4984 // slice_len number of elements to store (or NULL)
4979 4985 // dest_size total size in bytes of the array object
4980 4986 //
4981 4987 // Exactly one of slice_len or dest_size must be non-NULL.
4982 4988 // If dest_size is non-NULL, zeroing extends to the end of the object.
4983 4989 // If slice_len is non-NULL, the slice_idx value must be a constant.
4984 4990 void
4985 4991 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
4986 4992 Node* dest,
4987 4993 BasicType basic_elem_type,
4988 4994 Node* slice_idx,
4989 4995 Node* slice_len,
4990 4996 Node* dest_size) {
4991 4997 // one or the other but not both of slice_len and dest_size:
4992 4998 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
4993 4999 if (slice_len == NULL) slice_len = top();
4994 5000 if (dest_size == NULL) dest_size = top();
4995 5001
4996 5002 // operate on this memory slice:
4997 5003 Node* mem = memory(adr_type); // memory slice to operate on
4998 5004
4999 5005 // scaling and rounding of indexes:
5000 5006 int scale = exact_log2(type2aelembytes(basic_elem_type));
5001 5007 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5002 5008 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5003 5009 int bump_bit = (-1 << scale) & BytesPerInt;
5004 5010
5005 5011 // determine constant starts and ends
5006 5012 const intptr_t BIG_NEG = -128;
5007 5013 assert(BIG_NEG + 2*abase < 0, "neg enough");
5008 5014 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5009 5015 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5010 5016 if (slice_len_con == 0) {
5011 5017 return; // nothing to do here
5012 5018 }
5013 5019 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5014 5020 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5015 5021 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5016 5022 assert(end_con < 0, "not two cons");
5017 5023 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5018 5024 BytesPerLong);
5019 5025 }
5020 5026
5021 5027 if (start_con >= 0 && end_con >= 0) {
5022 5028 // Constant start and end. Simple.
5023 5029 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5024 5030 start_con, end_con, &_gvn);
5025 5031 } else if (start_con >= 0 && dest_size != top()) {
5026 5032 // Constant start, pre-rounded end after the tail of the array.
5027 5033 Node* end = dest_size;
5028 5034 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5029 5035 start_con, end, &_gvn);
5030 5036 } else if (start_con >= 0 && slice_len != top()) {
5031 5037 // Constant start, non-constant end. End needs rounding up.
5032 5038 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5033 5039 intptr_t end_base = abase + (slice_idx_con << scale);
5034 5040 int end_round = (-1 << scale) & (BytesPerLong - 1);
5035 5041 Node* end = ConvI2X(slice_len);
5036 5042 if (scale != 0)
5037 5043 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5038 5044 end_base += end_round;
5039 5045 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5040 5046 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5041 5047 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5042 5048 start_con, end, &_gvn);
5043 5049 } else if (start_con < 0 && dest_size != top()) {
5044 5050 // Non-constant start, pre-rounded end after the tail of the array.
5045 5051 // This is almost certainly a "round-to-end" operation.
5046 5052 Node* start = slice_idx;
5047 5053 start = ConvI2X(start);
5048 5054 if (scale != 0)
5049 5055 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5050 5056 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5051 5057 if ((bump_bit | clear_low) != 0) {
5052 5058 int to_clear = (bump_bit | clear_low);
5053 5059 // Align up mod 8, then store a jint zero unconditionally
5054 5060 // just before the mod-8 boundary.
5055 5061 if (((abase + bump_bit) & ~to_clear) - bump_bit
5056 5062 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5057 5063 bump_bit = 0;
5058 5064 assert((abase & to_clear) == 0, "array base must be long-aligned");
5059 5065 } else {
5060 5066 // Bump 'start' up to (or past) the next jint boundary:
5061 5067 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5062 5068 assert((abase & clear_low) == 0, "array base must be int-aligned");
5063 5069 }
5064 5070 // Round bumped 'start' down to jlong boundary in body of array.
5065 5071 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5066 5072 if (bump_bit != 0) {
5067 5073 // Store a zero to the immediately preceding jint:
5068 5074 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5069 5075 Node* p1 = basic_plus_adr(dest, x1);
5070 5076 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5071 5077 mem = _gvn.transform(mem);
5072 5078 }
5073 5079 }
5074 5080 Node* end = dest_size; // pre-rounded
5075 5081 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5076 5082 start, end, &_gvn);
5077 5083 } else {
5078 5084 // Non-constant start, unrounded non-constant end.
5079 5085 // (Nobody zeroes a random midsection of an array using this routine.)
5080 5086 ShouldNotReachHere(); // fix caller
5081 5087 }
5082 5088
5083 5089 // Done.
5084 5090 set_memory(mem, adr_type);
5085 5091 }
5086 5092
5087 5093
5088 5094 bool
5089 5095 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5090 5096 BasicType basic_elem_type,
5091 5097 AllocateNode* alloc,
5092 5098 Node* src, Node* src_offset,
5093 5099 Node* dest, Node* dest_offset,
5094 5100 Node* dest_size) {
5095 5101 // See if there is an advantage from block transfer.
5096 5102 int scale = exact_log2(type2aelembytes(basic_elem_type));
5097 5103 if (scale >= LogBytesPerLong)
5098 5104 return false; // it is already a block transfer
5099 5105
5100 5106 // Look at the alignment of the starting offsets.
5101 5107 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5102 5108 const intptr_t BIG_NEG = -128;
5103 5109 assert(BIG_NEG + 2*abase < 0, "neg enough");
5104 5110
5105 5111 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5106 5112 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5107 5113 if (src_off < 0 || dest_off < 0)
5108 5114 // At present, we can only understand constants.
5109 5115 return false;
5110 5116
5111 5117 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5112 5118 // Non-aligned; too bad.
5113 5119 // One more chance: Pick off an initial 32-bit word.
5114 5120 // This is a common case, since abase can be odd mod 8.
5115 5121 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5116 5122 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5117 5123 Node* sptr = basic_plus_adr(src, src_off);
5118 5124 Node* dptr = basic_plus_adr(dest, dest_off);
5119 5125 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5120 5126 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5121 5127 src_off += BytesPerInt;
5122 5128 dest_off += BytesPerInt;
5123 5129 } else {
5124 5130 return false;
5125 5131 }
5126 5132 }
5127 5133 assert(src_off % BytesPerLong == 0, "");
5128 5134 assert(dest_off % BytesPerLong == 0, "");
5129 5135
5130 5136 // Do this copy by giant steps.
5131 5137 Node* sptr = basic_plus_adr(src, src_off);
5132 5138 Node* dptr = basic_plus_adr(dest, dest_off);
5133 5139 Node* countx = dest_size;
5134 5140 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5135 5141 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5136 5142
5137 5143 bool disjoint_bases = true; // since alloc != NULL
5138 5144 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5139 5145 sptr, NULL, dptr, NULL, countx);
5140 5146
5141 5147 return true;
5142 5148 }
5143 5149
5144 5150
5145 5151 // Helper function; generates code for the slow case.
5146 5152 // We make a call to a runtime method which emulates the native method,
5147 5153 // but without the native wrapper overhead.
5148 5154 void
5149 5155 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5150 5156 Node* src, Node* src_offset,
5151 5157 Node* dest, Node* dest_offset,
5152 5158 Node* copy_length) {
5153 5159 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5154 5160 OptoRuntime::slow_arraycopy_Type(),
5155 5161 OptoRuntime::slow_arraycopy_Java(),
5156 5162 "slow_arraycopy", adr_type,
5157 5163 src, src_offset, dest, dest_offset,
5158 5164 copy_length);
5159 5165
5160 5166 // Handle exceptions thrown by this fellow:
5161 5167 make_slow_call_ex(call, env()->Throwable_klass(), false);
5162 5168 }
5163 5169
5164 5170 // Helper function; generates code for cases requiring runtime checks.
5165 5171 Node*
5166 5172 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5167 5173 Node* dest_elem_klass,
5168 5174 Node* src, Node* src_offset,
5169 5175 Node* dest, Node* dest_offset,
5170 5176 Node* copy_length) {
5171 5177 if (stopped()) return NULL;
5172 5178
5173 5179 address copyfunc_addr = StubRoutines::checkcast_arraycopy();
5174 5180 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5175 5181 return NULL;
5176 5182 }
5177 5183
5178 5184 // Pick out the parameters required to perform a store-check
5179 5185 // for the target array. This is an optimistic check. It will
5180 5186 // look in each non-null element's class, at the desired klass's
5181 5187 // super_check_offset, for the desired klass.
5182 5188 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5183 5189 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5184 5190 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
5185 5191 Node* check_offset = _gvn.transform(n3);
5186 5192 Node* check_value = dest_elem_klass;
5187 5193
5188 5194 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5189 5195 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5190 5196
5191 5197 // (We know the arrays are never conjoint, because their types differ.)
5192 5198 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5193 5199 OptoRuntime::checkcast_arraycopy_Type(),
5194 5200 copyfunc_addr, "checkcast_arraycopy", adr_type,
5195 5201 // five arguments, of which two are
5196 5202 // intptr_t (jlong in LP64)
5197 5203 src_start, dest_start,
5198 5204 copy_length XTOP,
5199 5205 check_offset XTOP,
5200 5206 check_value);
5201 5207
5202 5208 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5203 5209 }
5204 5210
5205 5211
5206 5212 // Helper function; generates code for cases requiring runtime checks.
5207 5213 Node*
5208 5214 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5209 5215 Node* src, Node* src_offset,
5210 5216 Node* dest, Node* dest_offset,
5211 5217 Node* copy_length) {
5212 5218 if (stopped()) return NULL;
5213 5219
5214 5220 address copyfunc_addr = StubRoutines::generic_arraycopy();
5215 5221 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5216 5222 return NULL;
5217 5223 }
5218 5224
5219 5225 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5220 5226 OptoRuntime::generic_arraycopy_Type(),
5221 5227 copyfunc_addr, "generic_arraycopy", adr_type,
5222 5228 src, src_offset, dest, dest_offset, copy_length);
5223 5229
5224 5230 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5225 5231 }
5226 5232
5227 5233 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5228 5234 void
5229 5235 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5230 5236 BasicType basic_elem_type,
5231 5237 bool disjoint_bases,
5232 5238 Node* src, Node* src_offset,
5233 5239 Node* dest, Node* dest_offset,
5234 5240 Node* copy_length) {
5235 5241 if (stopped()) return; // nothing to do
5236 5242
5237 5243 Node* src_start = src;
5238 5244 Node* dest_start = dest;
5239 5245 if (src_offset != NULL || dest_offset != NULL) {
5240 5246 assert(src_offset != NULL && dest_offset != NULL, "");
5241 5247 src_start = array_element_address(src, src_offset, basic_elem_type);
5242 5248 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5243 5249 }
5244 5250
5245 5251 // Figure out which arraycopy runtime method to call.
5246 5252 const char* copyfunc_name = "arraycopy";
5247 5253 address copyfunc_addr =
5248 5254 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5249 5255 disjoint_bases, copyfunc_name);
5250 5256
5251 5257 // Call it. Note that the count_ix value is not scaled to a byte-size.
5252 5258 make_runtime_call(RC_LEAF|RC_NO_FP,
5253 5259 OptoRuntime::fast_arraycopy_Type(),
5254 5260 copyfunc_addr, copyfunc_name, adr_type,
5255 5261 src_start, dest_start, copy_length XTOP);
5256 5262 }
↓ open down ↓ |
1532 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX