Print this page
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/opto/library_call.cpp
+++ new/src/share/vm/opto/library_call.cpp
1 1 /*
2 2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
|
↓ open down ↓ |
17 lines elided |
↑ open up ↑ |
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/systemDictionary.hpp"
27 27 #include "classfile/vmSymbols.hpp"
28 +#include "compiler/compileBroker.hpp"
28 29 #include "compiler/compileLog.hpp"
29 30 #include "oops/objArrayKlass.hpp"
30 31 #include "opto/addnode.hpp"
31 32 #include "opto/callGenerator.hpp"
32 33 #include "opto/cfgnode.hpp"
33 34 #include "opto/idealKit.hpp"
34 35 #include "opto/mulnode.hpp"
35 36 #include "opto/parse.hpp"
36 37 #include "opto/runtime.hpp"
37 38 #include "opto/subnode.hpp"
38 39 #include "prims/nativeLookup.hpp"
39 40 #include "runtime/sharedRuntime.hpp"
40 41
41 42 class LibraryIntrinsic : public InlineCallGenerator {
42 43 // Extend the set of intrinsics known to the runtime:
43 44 public:
44 45 private:
45 46 bool _is_virtual;
46 47 vmIntrinsics::ID _intrinsic_id;
47 48
48 49 public:
49 50 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
50 51 : InlineCallGenerator(m),
51 52 _is_virtual(is_virtual),
52 53 _intrinsic_id(id)
53 54 {
54 55 }
55 56 virtual bool is_intrinsic() const { return true; }
56 57 virtual bool is_virtual() const { return _is_virtual; }
57 58 virtual JVMState* generate(JVMState* jvms);
58 59 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
59 60 };
60 61
61 62
62 63 // Local helper class for LibraryIntrinsic:
63 64 class LibraryCallKit : public GraphKit {
64 65 private:
65 66 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
66 67
67 68 public:
68 69 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
69 70 : GraphKit(caller),
70 71 _intrinsic(intrinsic)
71 72 {
72 73 }
73 74
74 75 ciMethod* caller() const { return jvms()->method(); }
75 76 int bci() const { return jvms()->bci(); }
76 77 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
77 78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
78 79 ciMethod* callee() const { return _intrinsic->method(); }
79 80 ciSignature* signature() const { return callee()->signature(); }
80 81 int arg_size() const { return callee()->arg_size(); }
81 82
82 83 bool try_to_inline();
83 84
84 85 // Helper functions to inline natives
85 86 void push_result(RegionNode* region, PhiNode* value);
86 87 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
87 88 Node* generate_slow_guard(Node* test, RegionNode* region);
88 89 Node* generate_fair_guard(Node* test, RegionNode* region);
89 90 Node* generate_negative_guard(Node* index, RegionNode* region,
90 91 // resulting CastII of index:
91 92 Node* *pos_index = NULL);
92 93 Node* generate_nonpositive_guard(Node* index, bool never_negative,
93 94 // resulting CastII of index:
94 95 Node* *pos_index = NULL);
95 96 Node* generate_limit_guard(Node* offset, Node* subseq_length,
96 97 Node* array_length,
97 98 RegionNode* region);
98 99 Node* generate_current_thread(Node* &tls_output);
99 100 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
100 101 bool disjoint_bases, const char* &name, bool dest_uninitialized);
101 102 Node* load_mirror_from_klass(Node* klass);
102 103 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
103 104 int nargs,
104 105 RegionNode* region, int null_path,
105 106 int offset);
106 107 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
107 108 RegionNode* region, int null_path) {
108 109 int offset = java_lang_Class::klass_offset_in_bytes();
109 110 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
110 111 region, null_path,
111 112 offset);
112 113 }
113 114 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
114 115 int nargs,
115 116 RegionNode* region, int null_path) {
116 117 int offset = java_lang_Class::array_klass_offset_in_bytes();
117 118 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
118 119 region, null_path,
119 120 offset);
120 121 }
121 122 Node* generate_access_flags_guard(Node* kls,
122 123 int modifier_mask, int modifier_bits,
123 124 RegionNode* region);
124 125 Node* generate_interface_guard(Node* kls, RegionNode* region);
125 126 Node* generate_array_guard(Node* kls, RegionNode* region) {
126 127 return generate_array_guard_common(kls, region, false, false);
127 128 }
128 129 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
129 130 return generate_array_guard_common(kls, region, false, true);
130 131 }
131 132 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
132 133 return generate_array_guard_common(kls, region, true, false);
133 134 }
134 135 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
135 136 return generate_array_guard_common(kls, region, true, true);
136 137 }
137 138 Node* generate_array_guard_common(Node* kls, RegionNode* region,
138 139 bool obj_array, bool not_array);
139 140 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
140 141 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
141 142 bool is_virtual = false, bool is_static = false);
142 143 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
143 144 return generate_method_call(method_id, false, true);
144 145 }
145 146 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
146 147 return generate_method_call(method_id, true, false);
147 148 }
148 149
149 150 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
150 151 bool inline_string_compareTo();
151 152 bool inline_string_indexOf();
152 153 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
153 154 bool inline_string_equals();
154 155 Node* pop_math_arg();
155 156 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
156 157 bool inline_math_native(vmIntrinsics::ID id);
157 158 bool inline_trig(vmIntrinsics::ID id);
158 159 bool inline_trans(vmIntrinsics::ID id);
159 160 bool inline_abs(vmIntrinsics::ID id);
160 161 bool inline_sqrt(vmIntrinsics::ID id);
161 162 bool inline_pow(vmIntrinsics::ID id);
162 163 bool inline_exp(vmIntrinsics::ID id);
163 164 bool inline_min_max(vmIntrinsics::ID id);
164 165 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
165 166 // This returns Type::AnyPtr, RawPtr, or OopPtr.
166 167 int classify_unsafe_addr(Node* &base, Node* &offset);
167 168 Node* make_unsafe_address(Node* base, Node* offset);
168 169 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
169 170 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
170 171 bool inline_unsafe_allocate();
171 172 bool inline_unsafe_copyMemory();
172 173 bool inline_native_currentThread();
173 174 bool inline_native_time_funcs(bool isNano);
174 175 bool inline_native_isInterrupted();
175 176 bool inline_native_Class_query(vmIntrinsics::ID id);
176 177 bool inline_native_subtype_check();
177 178
178 179 bool inline_native_newArray();
179 180 bool inline_native_getLength();
180 181 bool inline_array_copyOf(bool is_copyOfRange);
181 182 bool inline_array_equals();
182 183 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
183 184 bool inline_native_clone(bool is_virtual);
184 185 bool inline_native_Reflection_getCallerClass();
185 186 bool inline_native_AtomicLong_get();
186 187 bool inline_native_AtomicLong_attemptUpdate();
187 188 bool is_method_invoke_or_aux_frame(JVMState* jvms);
188 189 // Helper function for inlining native object hash method
189 190 bool inline_native_hashcode(bool is_virtual, bool is_static);
190 191 bool inline_native_getClass();
191 192
192 193 // Helper functions for inlining arraycopy
193 194 bool inline_arraycopy();
194 195 void generate_arraycopy(const TypePtr* adr_type,
195 196 BasicType basic_elem_type,
196 197 Node* src, Node* src_offset,
197 198 Node* dest, Node* dest_offset,
198 199 Node* copy_length,
199 200 bool disjoint_bases = false,
200 201 bool length_never_negative = false,
201 202 RegionNode* slow_region = NULL);
202 203 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
203 204 RegionNode* slow_region);
204 205 void generate_clear_array(const TypePtr* adr_type,
205 206 Node* dest,
206 207 BasicType basic_elem_type,
207 208 Node* slice_off,
208 209 Node* slice_len,
209 210 Node* slice_end);
210 211 bool generate_block_arraycopy(const TypePtr* adr_type,
211 212 BasicType basic_elem_type,
212 213 AllocateNode* alloc,
213 214 Node* src, Node* src_offset,
214 215 Node* dest, Node* dest_offset,
215 216 Node* dest_size, bool dest_uninitialized);
216 217 void generate_slow_arraycopy(const TypePtr* adr_type,
217 218 Node* src, Node* src_offset,
218 219 Node* dest, Node* dest_offset,
219 220 Node* copy_length, bool dest_uninitialized);
220 221 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
221 222 Node* dest_elem_klass,
222 223 Node* src, Node* src_offset,
223 224 Node* dest, Node* dest_offset,
224 225 Node* copy_length, bool dest_uninitialized);
225 226 Node* generate_generic_arraycopy(const TypePtr* adr_type,
226 227 Node* src, Node* src_offset,
227 228 Node* dest, Node* dest_offset,
228 229 Node* copy_length, bool dest_uninitialized);
229 230 void generate_unchecked_arraycopy(const TypePtr* adr_type,
230 231 BasicType basic_elem_type,
231 232 bool disjoint_bases,
232 233 Node* src, Node* src_offset,
233 234 Node* dest, Node* dest_offset,
234 235 Node* copy_length, bool dest_uninitialized);
235 236 bool inline_unsafe_CAS(BasicType type);
236 237 bool inline_unsafe_ordered_store(BasicType type);
237 238 bool inline_fp_conversions(vmIntrinsics::ID id);
238 239 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
239 240 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
240 241 bool inline_bitCount(vmIntrinsics::ID id);
241 242 bool inline_reverseBytes(vmIntrinsics::ID id);
242 243 };
243 244
244 245
245 246 //---------------------------make_vm_intrinsic----------------------------
246 247 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
247 248 vmIntrinsics::ID id = m->intrinsic_id();
248 249 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
249 250
250 251 if (DisableIntrinsic[0] != '\0'
251 252 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
252 253 // disabled by a user request on the command line:
253 254 // example: -XX:DisableIntrinsic=_hashCode,_getClass
254 255 return NULL;
255 256 }
256 257
257 258 if (!m->is_loaded()) {
258 259 // do not attempt to inline unloaded methods
259 260 return NULL;
260 261 }
261 262
262 263 // Only a few intrinsics implement a virtual dispatch.
263 264 // They are expensive calls which are also frequently overridden.
264 265 if (is_virtual) {
265 266 switch (id) {
266 267 case vmIntrinsics::_hashCode:
267 268 case vmIntrinsics::_clone:
268 269 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
269 270 break;
270 271 default:
271 272 return NULL;
272 273 }
273 274 }
274 275
275 276 // -XX:-InlineNatives disables nearly all intrinsics:
276 277 if (!InlineNatives) {
277 278 switch (id) {
278 279 case vmIntrinsics::_indexOf:
279 280 case vmIntrinsics::_compareTo:
280 281 case vmIntrinsics::_equals:
281 282 case vmIntrinsics::_equalsC:
282 283 break; // InlineNatives does not control String.compareTo
283 284 default:
284 285 return NULL;
285 286 }
286 287 }
287 288
288 289 switch (id) {
289 290 case vmIntrinsics::_compareTo:
290 291 if (!SpecialStringCompareTo) return NULL;
291 292 break;
292 293 case vmIntrinsics::_indexOf:
293 294 if (!SpecialStringIndexOf) return NULL;
294 295 break;
295 296 case vmIntrinsics::_equals:
296 297 if (!SpecialStringEquals) return NULL;
297 298 break;
298 299 case vmIntrinsics::_equalsC:
299 300 if (!SpecialArraysEquals) return NULL;
300 301 break;
301 302 case vmIntrinsics::_arraycopy:
302 303 if (!InlineArrayCopy) return NULL;
303 304 break;
304 305 case vmIntrinsics::_copyMemory:
305 306 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
306 307 if (!InlineArrayCopy) return NULL;
307 308 break;
308 309 case vmIntrinsics::_hashCode:
309 310 if (!InlineObjectHash) return NULL;
310 311 break;
311 312 case vmIntrinsics::_clone:
312 313 case vmIntrinsics::_copyOf:
313 314 case vmIntrinsics::_copyOfRange:
314 315 if (!InlineObjectCopy) return NULL;
315 316 // These also use the arraycopy intrinsic mechanism:
316 317 if (!InlineArrayCopy) return NULL;
317 318 break;
318 319 case vmIntrinsics::_checkIndex:
319 320 // We do not intrinsify this. The optimizer does fine with it.
320 321 return NULL;
321 322
322 323 case vmIntrinsics::_get_AtomicLong:
323 324 case vmIntrinsics::_attemptUpdate:
324 325 if (!InlineAtomicLong) return NULL;
325 326 break;
326 327
327 328 case vmIntrinsics::_getCallerClass:
328 329 if (!UseNewReflection) return NULL;
329 330 if (!InlineReflectionGetCallerClass) return NULL;
330 331 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
331 332 break;
332 333
333 334 case vmIntrinsics::_bitCount_i:
334 335 case vmIntrinsics::_bitCount_l:
335 336 if (!UsePopCountInstruction) return NULL;
336 337 break;
337 338
338 339 default:
339 340 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
340 341 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
341 342 break;
342 343 }
343 344
344 345 // -XX:-InlineClassNatives disables natives from the Class class.
345 346 // The flag applies to all reflective calls, notably Array.newArray
346 347 // (visible to Java programmers as Array.newInstance).
347 348 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
348 349 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
349 350 if (!InlineClassNatives) return NULL;
350 351 }
351 352
352 353 // -XX:-InlineThreadNatives disables natives from the Thread class.
353 354 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
354 355 if (!InlineThreadNatives) return NULL;
355 356 }
356 357
357 358 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
358 359 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
359 360 m->holder()->name() == ciSymbol::java_lang_Float() ||
360 361 m->holder()->name() == ciSymbol::java_lang_Double()) {
361 362 if (!InlineMathNatives) return NULL;
362 363 }
363 364
364 365 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
365 366 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
366 367 if (!InlineUnsafeOps) return NULL;
367 368 }
368 369
369 370 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
370 371 }
371 372
372 373 //----------------------register_library_intrinsics-----------------------
373 374 // Initialize this file's data structures, for each Compile instance.
374 375 void Compile::register_library_intrinsics() {
375 376 // Nothing to do here.
376 377 }
377 378
378 379 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
379 380 LibraryCallKit kit(jvms, this);
380 381 Compile* C = kit.C;
|
↓ open down ↓ |
343 lines elided |
↑ open up ↑ |
381 382 int nodes = C->unique();
382 383 #ifndef PRODUCT
383 384 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
384 385 char buf[1000];
385 386 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
386 387 tty->print_cr("Intrinsic %s", str);
387 388 }
388 389 #endif
389 390 if (kit.try_to_inline()) {
390 391 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
391 - tty->print("Inlining intrinsic %s%s at bci:%d in",
392 - vmIntrinsics::name_at(intrinsic_id()),
393 - (is_virtual() ? " (virtual)" : ""), kit.bci());
394 - kit.caller()->print_short_name(tty);
395 - tty->print_cr(" (%d bytes)", kit.caller()->code_size());
392 + CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
396 393 }
397 394 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
398 395 if (C->log()) {
399 396 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
400 397 vmIntrinsics::name_at(intrinsic_id()),
401 398 (is_virtual() ? " virtual='1'" : ""),
402 399 C->unique() - nodes);
403 400 }
404 401 return kit.transfer_exceptions_into_jvms();
405 402 }
406 403
407 404 if (PrintIntrinsics) {
408 405 tty->print("Did not inline intrinsic %s%s at bci:%d in",
409 406 vmIntrinsics::name_at(intrinsic_id()),
410 407 (is_virtual() ? " (virtual)" : ""), kit.bci());
411 408 kit.caller()->print_short_name(tty);
412 409 tty->print_cr(" (%d bytes)", kit.caller()->code_size());
413 410 }
414 411 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
415 412 return NULL;
416 413 }
417 414
418 415 bool LibraryCallKit::try_to_inline() {
419 416 // Handle symbolic names for otherwise undistinguished boolean switches:
420 417 const bool is_store = true;
421 418 const bool is_native_ptr = true;
422 419 const bool is_static = true;
423 420
424 421 switch (intrinsic_id()) {
425 422 case vmIntrinsics::_hashCode:
426 423 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
427 424 case vmIntrinsics::_identityHashCode:
428 425 return inline_native_hashcode(/*!virtual*/ false, is_static);
429 426 case vmIntrinsics::_getClass:
430 427 return inline_native_getClass();
431 428
432 429 case vmIntrinsics::_dsin:
433 430 case vmIntrinsics::_dcos:
434 431 case vmIntrinsics::_dtan:
435 432 case vmIntrinsics::_dabs:
436 433 case vmIntrinsics::_datan2:
437 434 case vmIntrinsics::_dsqrt:
438 435 case vmIntrinsics::_dexp:
439 436 case vmIntrinsics::_dlog:
440 437 case vmIntrinsics::_dlog10:
441 438 case vmIntrinsics::_dpow:
442 439 return inline_math_native(intrinsic_id());
443 440
444 441 case vmIntrinsics::_min:
445 442 case vmIntrinsics::_max:
446 443 return inline_min_max(intrinsic_id());
447 444
448 445 case vmIntrinsics::_arraycopy:
449 446 return inline_arraycopy();
450 447
451 448 case vmIntrinsics::_compareTo:
452 449 return inline_string_compareTo();
453 450 case vmIntrinsics::_indexOf:
454 451 return inline_string_indexOf();
455 452 case vmIntrinsics::_equals:
456 453 return inline_string_equals();
457 454
458 455 case vmIntrinsics::_getObject:
459 456 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
460 457 case vmIntrinsics::_getBoolean:
461 458 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
462 459 case vmIntrinsics::_getByte:
463 460 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
464 461 case vmIntrinsics::_getShort:
465 462 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
466 463 case vmIntrinsics::_getChar:
467 464 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
468 465 case vmIntrinsics::_getInt:
469 466 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
470 467 case vmIntrinsics::_getLong:
471 468 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
472 469 case vmIntrinsics::_getFloat:
473 470 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
474 471 case vmIntrinsics::_getDouble:
475 472 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
476 473
477 474 case vmIntrinsics::_putObject:
478 475 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
479 476 case vmIntrinsics::_putBoolean:
480 477 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
481 478 case vmIntrinsics::_putByte:
482 479 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
483 480 case vmIntrinsics::_putShort:
484 481 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
485 482 case vmIntrinsics::_putChar:
486 483 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
487 484 case vmIntrinsics::_putInt:
488 485 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
489 486 case vmIntrinsics::_putLong:
490 487 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
491 488 case vmIntrinsics::_putFloat:
492 489 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
493 490 case vmIntrinsics::_putDouble:
494 491 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
495 492
496 493 case vmIntrinsics::_getByte_raw:
497 494 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
498 495 case vmIntrinsics::_getShort_raw:
499 496 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
500 497 case vmIntrinsics::_getChar_raw:
501 498 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
502 499 case vmIntrinsics::_getInt_raw:
503 500 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
504 501 case vmIntrinsics::_getLong_raw:
505 502 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
506 503 case vmIntrinsics::_getFloat_raw:
507 504 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
508 505 case vmIntrinsics::_getDouble_raw:
509 506 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
510 507 case vmIntrinsics::_getAddress_raw:
511 508 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
512 509
513 510 case vmIntrinsics::_putByte_raw:
514 511 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
515 512 case vmIntrinsics::_putShort_raw:
516 513 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
517 514 case vmIntrinsics::_putChar_raw:
518 515 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
519 516 case vmIntrinsics::_putInt_raw:
520 517 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
521 518 case vmIntrinsics::_putLong_raw:
522 519 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
523 520 case vmIntrinsics::_putFloat_raw:
524 521 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
525 522 case vmIntrinsics::_putDouble_raw:
526 523 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
527 524 case vmIntrinsics::_putAddress_raw:
528 525 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
529 526
530 527 case vmIntrinsics::_getObjectVolatile:
531 528 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
532 529 case vmIntrinsics::_getBooleanVolatile:
533 530 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
534 531 case vmIntrinsics::_getByteVolatile:
535 532 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
536 533 case vmIntrinsics::_getShortVolatile:
537 534 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
538 535 case vmIntrinsics::_getCharVolatile:
539 536 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
540 537 case vmIntrinsics::_getIntVolatile:
541 538 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
542 539 case vmIntrinsics::_getLongVolatile:
543 540 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
544 541 case vmIntrinsics::_getFloatVolatile:
545 542 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
546 543 case vmIntrinsics::_getDoubleVolatile:
547 544 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
548 545
549 546 case vmIntrinsics::_putObjectVolatile:
550 547 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
551 548 case vmIntrinsics::_putBooleanVolatile:
552 549 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
553 550 case vmIntrinsics::_putByteVolatile:
554 551 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
555 552 case vmIntrinsics::_putShortVolatile:
556 553 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
557 554 case vmIntrinsics::_putCharVolatile:
558 555 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
559 556 case vmIntrinsics::_putIntVolatile:
560 557 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
561 558 case vmIntrinsics::_putLongVolatile:
562 559 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
563 560 case vmIntrinsics::_putFloatVolatile:
564 561 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
565 562 case vmIntrinsics::_putDoubleVolatile:
566 563 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
567 564
568 565 case vmIntrinsics::_prefetchRead:
569 566 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
570 567 case vmIntrinsics::_prefetchWrite:
571 568 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
572 569 case vmIntrinsics::_prefetchReadStatic:
573 570 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
574 571 case vmIntrinsics::_prefetchWriteStatic:
575 572 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
576 573
577 574 case vmIntrinsics::_compareAndSwapObject:
578 575 return inline_unsafe_CAS(T_OBJECT);
579 576 case vmIntrinsics::_compareAndSwapInt:
580 577 return inline_unsafe_CAS(T_INT);
581 578 case vmIntrinsics::_compareAndSwapLong:
582 579 return inline_unsafe_CAS(T_LONG);
583 580
584 581 case vmIntrinsics::_putOrderedObject:
585 582 return inline_unsafe_ordered_store(T_OBJECT);
586 583 case vmIntrinsics::_putOrderedInt:
587 584 return inline_unsafe_ordered_store(T_INT);
588 585 case vmIntrinsics::_putOrderedLong:
589 586 return inline_unsafe_ordered_store(T_LONG);
590 587
591 588 case vmIntrinsics::_currentThread:
592 589 return inline_native_currentThread();
593 590 case vmIntrinsics::_isInterrupted:
594 591 return inline_native_isInterrupted();
595 592
596 593 case vmIntrinsics::_currentTimeMillis:
597 594 return inline_native_time_funcs(false);
598 595 case vmIntrinsics::_nanoTime:
599 596 return inline_native_time_funcs(true);
600 597 case vmIntrinsics::_allocateInstance:
601 598 return inline_unsafe_allocate();
602 599 case vmIntrinsics::_copyMemory:
603 600 return inline_unsafe_copyMemory();
604 601 case vmIntrinsics::_newArray:
605 602 return inline_native_newArray();
606 603 case vmIntrinsics::_getLength:
607 604 return inline_native_getLength();
608 605 case vmIntrinsics::_copyOf:
609 606 return inline_array_copyOf(false);
610 607 case vmIntrinsics::_copyOfRange:
611 608 return inline_array_copyOf(true);
612 609 case vmIntrinsics::_equalsC:
613 610 return inline_array_equals();
614 611 case vmIntrinsics::_clone:
615 612 return inline_native_clone(intrinsic()->is_virtual());
616 613
617 614 case vmIntrinsics::_isAssignableFrom:
618 615 return inline_native_subtype_check();
619 616
620 617 case vmIntrinsics::_isInstance:
621 618 case vmIntrinsics::_getModifiers:
622 619 case vmIntrinsics::_isInterface:
623 620 case vmIntrinsics::_isArray:
624 621 case vmIntrinsics::_isPrimitive:
625 622 case vmIntrinsics::_getSuperclass:
626 623 case vmIntrinsics::_getComponentType:
627 624 case vmIntrinsics::_getClassAccessFlags:
628 625 return inline_native_Class_query(intrinsic_id());
629 626
630 627 case vmIntrinsics::_floatToRawIntBits:
631 628 case vmIntrinsics::_floatToIntBits:
632 629 case vmIntrinsics::_intBitsToFloat:
633 630 case vmIntrinsics::_doubleToRawLongBits:
634 631 case vmIntrinsics::_doubleToLongBits:
635 632 case vmIntrinsics::_longBitsToDouble:
636 633 return inline_fp_conversions(intrinsic_id());
637 634
638 635 case vmIntrinsics::_numberOfLeadingZeros_i:
639 636 case vmIntrinsics::_numberOfLeadingZeros_l:
640 637 return inline_numberOfLeadingZeros(intrinsic_id());
641 638
642 639 case vmIntrinsics::_numberOfTrailingZeros_i:
643 640 case vmIntrinsics::_numberOfTrailingZeros_l:
644 641 return inline_numberOfTrailingZeros(intrinsic_id());
645 642
646 643 case vmIntrinsics::_bitCount_i:
647 644 case vmIntrinsics::_bitCount_l:
648 645 return inline_bitCount(intrinsic_id());
649 646
650 647 case vmIntrinsics::_reverseBytes_i:
651 648 case vmIntrinsics::_reverseBytes_l:
652 649 case vmIntrinsics::_reverseBytes_s:
653 650 case vmIntrinsics::_reverseBytes_c:
654 651 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
655 652
656 653 case vmIntrinsics::_get_AtomicLong:
657 654 return inline_native_AtomicLong_get();
658 655 case vmIntrinsics::_attemptUpdate:
659 656 return inline_native_AtomicLong_attemptUpdate();
660 657
661 658 case vmIntrinsics::_getCallerClass:
662 659 return inline_native_Reflection_getCallerClass();
663 660
664 661 default:
665 662 // If you get here, it may be that someone has added a new intrinsic
666 663 // to the list in vmSymbols.hpp without implementing it here.
667 664 #ifndef PRODUCT
668 665 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
669 666 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
670 667 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
671 668 }
672 669 #endif
673 670 return false;
674 671 }
675 672 }
676 673
677 674 //------------------------------push_result------------------------------
678 675 // Helper function for finishing intrinsics.
679 676 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
680 677 record_for_igvn(region);
681 678 set_control(_gvn.transform(region));
682 679 BasicType value_type = value->type()->basic_type();
683 680 push_node(value_type, _gvn.transform(value));
684 681 }
685 682
686 683 //------------------------------generate_guard---------------------------
687 684 // Helper function for generating guarded fast-slow graph structures.
688 685 // The given 'test', if true, guards a slow path. If the test fails
689 686 // then a fast path can be taken. (We generally hope it fails.)
690 687 // In all cases, GraphKit::control() is updated to the fast path.
691 688 // The returned value represents the control for the slow path.
692 689 // The return value is never 'top'; it is either a valid control
693 690 // or NULL if it is obvious that the slow path can never be taken.
694 691 // Also, if region and the slow control are not NULL, the slow edge
695 692 // is appended to the region.
696 693 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
697 694 if (stopped()) {
698 695 // Already short circuited.
699 696 return NULL;
700 697 }
701 698
702 699 // Build an if node and its projections.
703 700 // If test is true we take the slow path, which we assume is uncommon.
704 701 if (_gvn.type(test) == TypeInt::ZERO) {
705 702 // The slow branch is never taken. No need to build this guard.
706 703 return NULL;
707 704 }
708 705
709 706 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
710 707
711 708 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
712 709 if (if_slow == top()) {
713 710 // The slow branch is never taken. No need to build this guard.
714 711 return NULL;
715 712 }
716 713
717 714 if (region != NULL)
718 715 region->add_req(if_slow);
719 716
720 717 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
721 718 set_control(if_fast);
722 719
723 720 return if_slow;
724 721 }
725 722
726 723 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
727 724 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
728 725 }
729 726 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
730 727 return generate_guard(test, region, PROB_FAIR);
731 728 }
732 729
733 730 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
734 731 Node* *pos_index) {
735 732 if (stopped())
736 733 return NULL; // already stopped
737 734 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
738 735 return NULL; // index is already adequately typed
739 736 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
740 737 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
741 738 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
742 739 if (is_neg != NULL && pos_index != NULL) {
743 740 // Emulate effect of Parse::adjust_map_after_if.
744 741 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
745 742 ccast->set_req(0, control());
746 743 (*pos_index) = _gvn.transform(ccast);
747 744 }
748 745 return is_neg;
749 746 }
750 747
751 748 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
752 749 Node* *pos_index) {
753 750 if (stopped())
754 751 return NULL; // already stopped
755 752 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
756 753 return NULL; // index is already adequately typed
757 754 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
758 755 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
759 756 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
760 757 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
761 758 if (is_notp != NULL && pos_index != NULL) {
762 759 // Emulate effect of Parse::adjust_map_after_if.
763 760 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
764 761 ccast->set_req(0, control());
765 762 (*pos_index) = _gvn.transform(ccast);
766 763 }
767 764 return is_notp;
768 765 }
769 766
770 767 // Make sure that 'position' is a valid limit index, in [0..length].
771 768 // There are two equivalent plans for checking this:
772 769 // A. (offset + copyLength) unsigned<= arrayLength
773 770 // B. offset <= (arrayLength - copyLength)
774 771 // We require that all of the values above, except for the sum and
775 772 // difference, are already known to be non-negative.
776 773 // Plan A is robust in the face of overflow, if offset and copyLength
777 774 // are both hugely positive.
778 775 //
779 776 // Plan B is less direct and intuitive, but it does not overflow at
780 777 // all, since the difference of two non-negatives is always
781 778 // representable. Whenever Java methods must perform the equivalent
782 779 // check they generally use Plan B instead of Plan A.
783 780 // For the moment we use Plan A.
784 781 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
785 782 Node* subseq_length,
786 783 Node* array_length,
787 784 RegionNode* region) {
788 785 if (stopped())
789 786 return NULL; // already stopped
790 787 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
791 788 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
792 789 return NULL; // common case of whole-array copy
793 790 Node* last = subseq_length;
794 791 if (!zero_offset) // last += offset
795 792 last = _gvn.transform( new (C, 3) AddINode(last, offset));
796 793 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
797 794 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
798 795 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
799 796 return is_over;
800 797 }
801 798
802 799
803 800 //--------------------------generate_current_thread--------------------
804 801 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
805 802 ciKlass* thread_klass = env()->Thread_klass();
806 803 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
807 804 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
808 805 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
809 806 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
810 807 tls_output = thread;
811 808 return threadObj;
812 809 }
813 810
814 811
815 812 //------------------------------make_string_method_node------------------------
816 813 // Helper method for String intrinsic finctions.
817 814 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
818 815 const int value_offset = java_lang_String::value_offset_in_bytes();
819 816 const int count_offset = java_lang_String::count_offset_in_bytes();
820 817 const int offset_offset = java_lang_String::offset_offset_in_bytes();
821 818
822 819 Node* no_ctrl = NULL;
823 820
824 821 ciInstanceKlass* klass = env()->String_klass();
825 822 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
826 823
827 824 const TypeAryPtr* value_type =
828 825 TypeAryPtr::make(TypePtr::NotNull,
829 826 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
830 827 ciTypeArrayKlass::make(T_CHAR), true, 0);
831 828
832 829 // Get start addr of string and substring
833 830 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
834 831 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
835 832 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
836 833 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
837 834 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
838 835
839 836 // Pin loads from String::equals() argument since it could be NULL.
840 837 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
841 838 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
842 839 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
843 840 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
844 841 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
845 842 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
846 843
847 844 Node* result = NULL;
848 845 switch (opcode) {
849 846 case Op_StrIndexOf:
850 847 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
851 848 str1_start, cnt1, str2_start, cnt2);
852 849 break;
853 850 case Op_StrComp:
854 851 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
855 852 str1_start, cnt1, str2_start, cnt2);
856 853 break;
857 854 case Op_StrEquals:
858 855 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
859 856 str1_start, str2_start, cnt1);
860 857 break;
861 858 default:
862 859 ShouldNotReachHere();
863 860 return NULL;
864 861 }
865 862
866 863 // All these intrinsics have checks.
867 864 C->set_has_split_ifs(true); // Has chance for split-if optimization
868 865
869 866 return _gvn.transform(result);
870 867 }
871 868
872 869 //------------------------------inline_string_compareTo------------------------
873 870 bool LibraryCallKit::inline_string_compareTo() {
874 871
875 872 if (!Matcher::has_match_rule(Op_StrComp)) return false;
876 873
877 874 const int value_offset = java_lang_String::value_offset_in_bytes();
878 875 const int count_offset = java_lang_String::count_offset_in_bytes();
879 876 const int offset_offset = java_lang_String::offset_offset_in_bytes();
880 877
881 878 _sp += 2;
882 879 Node *argument = pop(); // pop non-receiver first: it was pushed second
883 880 Node *receiver = pop();
884 881
885 882 // Null check on self without removing any arguments. The argument
886 883 // null check technically happens in the wrong place, which can lead to
887 884 // invalid stack traces when string compare is inlined into a method
888 885 // which handles NullPointerExceptions.
889 886 _sp += 2;
890 887 receiver = do_null_check(receiver, T_OBJECT);
891 888 argument = do_null_check(argument, T_OBJECT);
892 889 _sp -= 2;
893 890 if (stopped()) {
894 891 return true;
895 892 }
896 893
897 894 ciInstanceKlass* klass = env()->String_klass();
898 895 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
899 896 Node* no_ctrl = NULL;
900 897
901 898 // Get counts for string and argument
902 899 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
903 900 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
904 901
905 902 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
906 903 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
907 904
908 905 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
909 906 push(compare);
910 907 return true;
911 908 }
912 909
913 910 //------------------------------inline_string_equals------------------------
914 911 bool LibraryCallKit::inline_string_equals() {
915 912
916 913 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
917 914
918 915 const int value_offset = java_lang_String::value_offset_in_bytes();
919 916 const int count_offset = java_lang_String::count_offset_in_bytes();
920 917 const int offset_offset = java_lang_String::offset_offset_in_bytes();
921 918
922 919 int nargs = 2;
923 920 _sp += nargs;
924 921 Node* argument = pop(); // pop non-receiver first: it was pushed second
925 922 Node* receiver = pop();
926 923
927 924 // Null check on self without removing any arguments. The argument
928 925 // null check technically happens in the wrong place, which can lead to
929 926 // invalid stack traces when string compare is inlined into a method
930 927 // which handles NullPointerExceptions.
931 928 _sp += nargs;
932 929 receiver = do_null_check(receiver, T_OBJECT);
933 930 //should not do null check for argument for String.equals(), because spec
934 931 //allows to specify NULL as argument.
935 932 _sp -= nargs;
936 933
937 934 if (stopped()) {
938 935 return true;
939 936 }
940 937
941 938 // paths (plus control) merge
942 939 RegionNode* region = new (C, 5) RegionNode(5);
943 940 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
944 941
945 942 // does source == target string?
946 943 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
947 944 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
948 945
949 946 Node* if_eq = generate_slow_guard(bol, NULL);
950 947 if (if_eq != NULL) {
951 948 // receiver == argument
952 949 phi->init_req(2, intcon(1));
953 950 region->init_req(2, if_eq);
954 951 }
955 952
956 953 // get String klass for instanceOf
957 954 ciInstanceKlass* klass = env()->String_klass();
958 955
959 956 if (!stopped()) {
960 957 _sp += nargs; // gen_instanceof might do an uncommon trap
961 958 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
962 959 _sp -= nargs;
963 960 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
964 961 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
965 962
966 963 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
967 964 //instanceOf == true, fallthrough
968 965
969 966 if (inst_false != NULL) {
970 967 phi->init_req(3, intcon(0));
971 968 region->init_req(3, inst_false);
972 969 }
973 970 }
974 971
975 972 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
976 973
977 974 Node* no_ctrl = NULL;
978 975 Node* receiver_cnt;
979 976 Node* argument_cnt;
980 977
981 978 if (!stopped()) {
982 979 // Properly cast the argument to String
983 980 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
984 981
985 982 // Get counts for string and argument
986 983 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
987 984 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
988 985
989 986 // Pin load from argument string since it could be NULL.
990 987 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
991 988 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
992 989
993 990 // Check for receiver count != argument count
994 991 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
995 992 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
996 993 Node* if_ne = generate_slow_guard(bol, NULL);
997 994 if (if_ne != NULL) {
998 995 phi->init_req(4, intcon(0));
999 996 region->init_req(4, if_ne);
1000 997 }
1001 998 }
1002 999
1003 1000 // Check for count == 0 is done by mach node StrEquals.
1004 1001
1005 1002 if (!stopped()) {
1006 1003 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1007 1004 phi->init_req(1, equals);
1008 1005 region->init_req(1, control());
1009 1006 }
1010 1007
1011 1008 // post merge
1012 1009 set_control(_gvn.transform(region));
1013 1010 record_for_igvn(region);
1014 1011
1015 1012 push(_gvn.transform(phi));
1016 1013
1017 1014 return true;
1018 1015 }
1019 1016
1020 1017 //------------------------------inline_array_equals----------------------------
1021 1018 bool LibraryCallKit::inline_array_equals() {
1022 1019
1023 1020 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1024 1021
1025 1022 _sp += 2;
1026 1023 Node *argument2 = pop();
1027 1024 Node *argument1 = pop();
1028 1025
1029 1026 Node* equals =
1030 1027 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1031 1028 argument1, argument2) );
1032 1029 push(equals);
1033 1030 return true;
1034 1031 }
1035 1032
1036 1033 // Java version of String.indexOf(constant string)
1037 1034 // class StringDecl {
1038 1035 // StringDecl(char[] ca) {
1039 1036 // offset = 0;
1040 1037 // count = ca.length;
1041 1038 // value = ca;
1042 1039 // }
1043 1040 // int offset;
1044 1041 // int count;
1045 1042 // char[] value;
1046 1043 // }
1047 1044 //
1048 1045 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1049 1046 // int targetOffset, int cache_i, int md2) {
1050 1047 // int cache = cache_i;
1051 1048 // int sourceOffset = string_object.offset;
1052 1049 // int sourceCount = string_object.count;
1053 1050 // int targetCount = target_object.length;
1054 1051 //
1055 1052 // int targetCountLess1 = targetCount - 1;
1056 1053 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1057 1054 //
1058 1055 // char[] source = string_object.value;
1059 1056 // char[] target = target_object;
1060 1057 // int lastChar = target[targetCountLess1];
1061 1058 //
1062 1059 // outer_loop:
1063 1060 // for (int i = sourceOffset; i < sourceEnd; ) {
1064 1061 // int src = source[i + targetCountLess1];
1065 1062 // if (src == lastChar) {
1066 1063 // // With random strings and a 4-character alphabet,
1067 1064 // // reverse matching at this point sets up 0.8% fewer
1068 1065 // // frames, but (paradoxically) makes 0.3% more probes.
1069 1066 // // Since those probes are nearer the lastChar probe,
1070 1067 // // there is may be a net D$ win with reverse matching.
1071 1068 // // But, reversing loop inhibits unroll of inner loop
1072 1069 // // for unknown reason. So, does running outer loop from
1073 1070 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1074 1071 // for (int j = 0; j < targetCountLess1; j++) {
1075 1072 // if (target[targetOffset + j] != source[i+j]) {
1076 1073 // if ((cache & (1 << source[i+j])) == 0) {
1077 1074 // if (md2 < j+1) {
1078 1075 // i += j+1;
1079 1076 // continue outer_loop;
1080 1077 // }
1081 1078 // }
1082 1079 // i += md2;
1083 1080 // continue outer_loop;
1084 1081 // }
1085 1082 // }
1086 1083 // return i - sourceOffset;
1087 1084 // }
1088 1085 // if ((cache & (1 << src)) == 0) {
1089 1086 // i += targetCountLess1;
1090 1087 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1091 1088 // i++;
1092 1089 // }
1093 1090 // return -1;
1094 1091 // }
1095 1092
1096 1093 //------------------------------string_indexOf------------------------
1097 1094 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1098 1095 jint cache_i, jint md2_i) {
1099 1096
1100 1097 Node* no_ctrl = NULL;
1101 1098 float likely = PROB_LIKELY(0.9);
1102 1099 float unlikely = PROB_UNLIKELY(0.9);
1103 1100
1104 1101 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1105 1102
1106 1103 const int value_offset = java_lang_String::value_offset_in_bytes();
1107 1104 const int count_offset = java_lang_String::count_offset_in_bytes();
1108 1105 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1109 1106
1110 1107 ciInstanceKlass* klass = env()->String_klass();
1111 1108 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1112 1109 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1113 1110
1114 1111 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1115 1112 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1116 1113 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1117 1114 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1118 1115 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1119 1116 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1120 1117
1121 1118 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1122 1119 jint target_length = target_array->length();
1123 1120 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1124 1121 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1125 1122
1126 1123 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1127 1124 #define __ kit.
1128 1125 Node* zero = __ ConI(0);
1129 1126 Node* one = __ ConI(1);
1130 1127 Node* cache = __ ConI(cache_i);
1131 1128 Node* md2 = __ ConI(md2_i);
1132 1129 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1133 1130 Node* targetCount = __ ConI(target_length);
1134 1131 Node* targetCountLess1 = __ ConI(target_length - 1);
1135 1132 Node* targetOffset = __ ConI(targetOffset_i);
1136 1133 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1137 1134
1138 1135 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1139 1136 Node* outer_loop = __ make_label(2 /* goto */);
1140 1137 Node* return_ = __ make_label(1);
1141 1138
1142 1139 __ set(rtn,__ ConI(-1));
1143 1140 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1144 1141 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1145 1142 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1146 1143 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1147 1144 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1148 1145 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1149 1146 Node* tpj = __ AddI(targetOffset, __ value(j));
1150 1147 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1151 1148 Node* ipj = __ AddI(__ value(i), __ value(j));
1152 1149 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1153 1150 __ if_then(targ, BoolTest::ne, src2); {
1154 1151 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1155 1152 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1156 1153 __ increment(i, __ AddI(__ value(j), one));
1157 1154 __ goto_(outer_loop);
1158 1155 } __ end_if(); __ dead(j);
1159 1156 }__ end_if(); __ dead(j);
1160 1157 __ increment(i, md2);
1161 1158 __ goto_(outer_loop);
1162 1159 }__ end_if();
1163 1160 __ increment(j, one);
1164 1161 }__ end_loop(); __ dead(j);
1165 1162 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1166 1163 __ goto_(return_);
1167 1164 }__ end_if();
1168 1165 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1169 1166 __ increment(i, targetCountLess1);
1170 1167 }__ end_if();
1171 1168 __ increment(i, one);
1172 1169 __ bind(outer_loop);
1173 1170 }__ end_loop(); __ dead(i);
1174 1171 __ bind(return_);
1175 1172
1176 1173 // Final sync IdealKit and GraphKit.
1177 1174 sync_kit(kit);
1178 1175 Node* result = __ value(rtn);
1179 1176 #undef __
1180 1177 C->set_has_loops(true);
1181 1178 return result;
1182 1179 }
1183 1180
1184 1181 //------------------------------inline_string_indexOf------------------------
1185 1182 bool LibraryCallKit::inline_string_indexOf() {
1186 1183
1187 1184 const int value_offset = java_lang_String::value_offset_in_bytes();
1188 1185 const int count_offset = java_lang_String::count_offset_in_bytes();
1189 1186 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1190 1187
1191 1188 _sp += 2;
1192 1189 Node *argument = pop(); // pop non-receiver first: it was pushed second
1193 1190 Node *receiver = pop();
1194 1191
1195 1192 Node* result;
1196 1193 // Disable the use of pcmpestri until it can be guaranteed that
1197 1194 // the load doesn't cross into the uncommited space.
1198 1195 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1199 1196 UseSSE42Intrinsics) {
1200 1197 // Generate SSE4.2 version of indexOf
1201 1198 // We currently only have match rules that use SSE4.2
1202 1199
1203 1200 // Null check on self without removing any arguments. The argument
1204 1201 // null check technically happens in the wrong place, which can lead to
1205 1202 // invalid stack traces when string compare is inlined into a method
1206 1203 // which handles NullPointerExceptions.
1207 1204 _sp += 2;
1208 1205 receiver = do_null_check(receiver, T_OBJECT);
1209 1206 argument = do_null_check(argument, T_OBJECT);
1210 1207 _sp -= 2;
1211 1208
1212 1209 if (stopped()) {
1213 1210 return true;
1214 1211 }
1215 1212
1216 1213 ciInstanceKlass* str_klass = env()->String_klass();
1217 1214 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1218 1215
1219 1216 // Make the merge point
1220 1217 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1221 1218 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1222 1219 Node* no_ctrl = NULL;
1223 1220
1224 1221 // Get counts for string and substr
1225 1222 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1226 1223 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1227 1224
1228 1225 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1229 1226 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1230 1227
1231 1228 // Check for substr count > string count
1232 1229 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1233 1230 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1234 1231 Node* if_gt = generate_slow_guard(bol, NULL);
1235 1232 if (if_gt != NULL) {
1236 1233 result_phi->init_req(2, intcon(-1));
1237 1234 result_rgn->init_req(2, if_gt);
1238 1235 }
1239 1236
1240 1237 if (!stopped()) {
1241 1238 // Check for substr count == 0
1242 1239 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1243 1240 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1244 1241 Node* if_zero = generate_slow_guard(bol, NULL);
1245 1242 if (if_zero != NULL) {
1246 1243 result_phi->init_req(3, intcon(0));
1247 1244 result_rgn->init_req(3, if_zero);
1248 1245 }
1249 1246 }
1250 1247
1251 1248 if (!stopped()) {
1252 1249 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1253 1250 result_phi->init_req(1, result);
1254 1251 result_rgn->init_req(1, control());
1255 1252 }
1256 1253 set_control(_gvn.transform(result_rgn));
1257 1254 record_for_igvn(result_rgn);
1258 1255 result = _gvn.transform(result_phi);
1259 1256
1260 1257 } else { // Use LibraryCallKit::string_indexOf
1261 1258 // don't intrinsify if argument isn't a constant string.
1262 1259 if (!argument->is_Con()) {
1263 1260 return false;
1264 1261 }
1265 1262 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1266 1263 if (str_type == NULL) {
1267 1264 return false;
1268 1265 }
1269 1266 ciInstanceKlass* klass = env()->String_klass();
1270 1267 ciObject* str_const = str_type->const_oop();
1271 1268 if (str_const == NULL || str_const->klass() != klass) {
1272 1269 return false;
1273 1270 }
1274 1271 ciInstance* str = str_const->as_instance();
1275 1272 assert(str != NULL, "must be instance");
1276 1273
1277 1274 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1278 1275 int o = str->field_value_by_offset(offset_offset).as_int();
1279 1276 int c = str->field_value_by_offset(count_offset).as_int();
1280 1277 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1281 1278
1282 1279 // constant strings have no offset and count == length which
1283 1280 // simplifies the resulting code somewhat so lets optimize for that.
1284 1281 if (o != 0 || c != pat->length()) {
1285 1282 return false;
1286 1283 }
1287 1284
1288 1285 // Null check on self without removing any arguments. The argument
1289 1286 // null check technically happens in the wrong place, which can lead to
1290 1287 // invalid stack traces when string compare is inlined into a method
1291 1288 // which handles NullPointerExceptions.
1292 1289 _sp += 2;
1293 1290 receiver = do_null_check(receiver, T_OBJECT);
1294 1291 // No null check on the argument is needed since it's a constant String oop.
1295 1292 _sp -= 2;
1296 1293 if (stopped()) {
1297 1294 return true;
1298 1295 }
1299 1296
1300 1297 // The null string as a pattern always returns 0 (match at beginning of string)
1301 1298 if (c == 0) {
1302 1299 push(intcon(0));
1303 1300 return true;
1304 1301 }
1305 1302
1306 1303 // Generate default indexOf
1307 1304 jchar lastChar = pat->char_at(o + (c - 1));
1308 1305 int cache = 0;
1309 1306 int i;
1310 1307 for (i = 0; i < c - 1; i++) {
1311 1308 assert(i < pat->length(), "out of range");
1312 1309 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1313 1310 }
1314 1311
1315 1312 int md2 = c;
1316 1313 for (i = 0; i < c - 1; i++) {
1317 1314 assert(i < pat->length(), "out of range");
1318 1315 if (pat->char_at(o + i) == lastChar) {
1319 1316 md2 = (c - 1) - i;
1320 1317 }
1321 1318 }
1322 1319
1323 1320 result = string_indexOf(receiver, pat, o, cache, md2);
1324 1321 }
1325 1322
1326 1323 push(result);
1327 1324 return true;
1328 1325 }
1329 1326
1330 1327 //--------------------------pop_math_arg--------------------------------
1331 1328 // Pop a double argument to a math function from the stack
1332 1329 // rounding it if necessary.
1333 1330 Node * LibraryCallKit::pop_math_arg() {
1334 1331 Node *arg = pop_pair();
1335 1332 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1336 1333 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1337 1334 return arg;
1338 1335 }
1339 1336
1340 1337 //------------------------------inline_trig----------------------------------
1341 1338 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1342 1339 // argument reduction which will turn into a fast/slow diamond.
1343 1340 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1344 1341 _sp += arg_size(); // restore stack pointer
1345 1342 Node* arg = pop_math_arg();
1346 1343 Node* trig = NULL;
1347 1344
1348 1345 switch (id) {
1349 1346 case vmIntrinsics::_dsin:
1350 1347 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1351 1348 break;
1352 1349 case vmIntrinsics::_dcos:
1353 1350 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1354 1351 break;
1355 1352 case vmIntrinsics::_dtan:
1356 1353 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1357 1354 break;
1358 1355 default:
1359 1356 assert(false, "bad intrinsic was passed in");
1360 1357 return false;
1361 1358 }
1362 1359
1363 1360 // Rounding required? Check for argument reduction!
1364 1361 if( Matcher::strict_fp_requires_explicit_rounding ) {
1365 1362
1366 1363 static const double pi_4 = 0.7853981633974483;
1367 1364 static const double neg_pi_4 = -0.7853981633974483;
1368 1365 // pi/2 in 80-bit extended precision
1369 1366 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1370 1367 // -pi/2 in 80-bit extended precision
1371 1368 // 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};
1372 1369 // Cutoff value for using this argument reduction technique
1373 1370 //static const double pi_2_minus_epsilon = 1.564660403643354;
1374 1371 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1375 1372
1376 1373 // Pseudocode for sin:
1377 1374 // if (x <= Math.PI / 4.0) {
1378 1375 // if (x >= -Math.PI / 4.0) return fsin(x);
1379 1376 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1380 1377 // } else {
1381 1378 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1382 1379 // }
1383 1380 // return StrictMath.sin(x);
1384 1381
1385 1382 // Pseudocode for cos:
1386 1383 // if (x <= Math.PI / 4.0) {
1387 1384 // if (x >= -Math.PI / 4.0) return fcos(x);
1388 1385 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1389 1386 // } else {
1390 1387 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1391 1388 // }
1392 1389 // return StrictMath.cos(x);
1393 1390
1394 1391 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1395 1392 // requires a special machine instruction to load it. Instead we'll try
1396 1393 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1397 1394 // probably do the math inside the SIN encoding.
1398 1395
1399 1396 // Make the merge point
1400 1397 RegionNode *r = new (C, 3) RegionNode(3);
1401 1398 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1402 1399
1403 1400 // Flatten arg so we need only 1 test
1404 1401 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1405 1402 // Node for PI/4 constant
1406 1403 Node *pi4 = makecon(TypeD::make(pi_4));
1407 1404 // Check PI/4 : abs(arg)
1408 1405 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1409 1406 // Check: If PI/4 < abs(arg) then go slow
1410 1407 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1411 1408 // Branch either way
1412 1409 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1413 1410 set_control(opt_iff(r,iff));
1414 1411
1415 1412 // Set fast path result
1416 1413 phi->init_req(2,trig);
1417 1414
1418 1415 // Slow path - non-blocking leaf call
1419 1416 Node* call = NULL;
1420 1417 switch (id) {
1421 1418 case vmIntrinsics::_dsin:
1422 1419 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1423 1420 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1424 1421 "Sin", NULL, arg, top());
1425 1422 break;
1426 1423 case vmIntrinsics::_dcos:
1427 1424 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1428 1425 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1429 1426 "Cos", NULL, arg, top());
1430 1427 break;
1431 1428 case vmIntrinsics::_dtan:
1432 1429 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1433 1430 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1434 1431 "Tan", NULL, arg, top());
1435 1432 break;
1436 1433 }
1437 1434 assert(control()->in(0) == call, "");
1438 1435 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1439 1436 r->init_req(1,control());
1440 1437 phi->init_req(1,slow_result);
1441 1438
1442 1439 // Post-merge
1443 1440 set_control(_gvn.transform(r));
1444 1441 record_for_igvn(r);
1445 1442 trig = _gvn.transform(phi);
1446 1443
1447 1444 C->set_has_split_ifs(true); // Has chance for split-if optimization
1448 1445 }
1449 1446 // Push result back on JVM stack
1450 1447 push_pair(trig);
1451 1448 return true;
1452 1449 }
1453 1450
1454 1451 //------------------------------inline_sqrt-------------------------------------
1455 1452 // Inline square root instruction, if possible.
1456 1453 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1457 1454 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1458 1455 _sp += arg_size(); // restore stack pointer
1459 1456 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1460 1457 return true;
1461 1458 }
1462 1459
1463 1460 //------------------------------inline_abs-------------------------------------
1464 1461 // Inline absolute value instruction, if possible.
1465 1462 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1466 1463 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1467 1464 _sp += arg_size(); // restore stack pointer
1468 1465 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1469 1466 return true;
1470 1467 }
1471 1468
1472 1469 //------------------------------inline_exp-------------------------------------
1473 1470 // Inline exp instructions, if possible. The Intel hardware only misses
1474 1471 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1475 1472 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1476 1473 assert(id == vmIntrinsics::_dexp, "Not exp");
1477 1474
1478 1475 // If this inlining ever returned NaN in the past, we do not intrinsify it
1479 1476 // every again. NaN results requires StrictMath.exp handling.
1480 1477 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1481 1478
1482 1479 // Do not intrinsify on older platforms which lack cmove.
1483 1480 if (ConditionalMoveLimit == 0) return false;
1484 1481
1485 1482 _sp += arg_size(); // restore stack pointer
1486 1483 Node *x = pop_math_arg();
1487 1484 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1488 1485
1489 1486 //-------------------
1490 1487 //result=(result.isNaN())? StrictMath::exp():result;
1491 1488 // Check: If isNaN() by checking result!=result? then go to Strict Math
1492 1489 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1493 1490 // Build the boolean node
1494 1491 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1495 1492
1496 1493 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1497 1494 // End the current control-flow path
1498 1495 push_pair(x);
1499 1496 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1500 1497 // to handle. Recompile without intrinsifying Math.exp
1501 1498 uncommon_trap(Deoptimization::Reason_intrinsic,
1502 1499 Deoptimization::Action_make_not_entrant);
1503 1500 }
1504 1501
1505 1502 C->set_has_split_ifs(true); // Has chance for split-if optimization
1506 1503
1507 1504 push_pair(result);
1508 1505
1509 1506 return true;
1510 1507 }
1511 1508
1512 1509 //------------------------------inline_pow-------------------------------------
1513 1510 // Inline power instructions, if possible.
1514 1511 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1515 1512 assert(id == vmIntrinsics::_dpow, "Not pow");
1516 1513
1517 1514 // If this inlining ever returned NaN in the past, we do not intrinsify it
1518 1515 // every again. NaN results requires StrictMath.pow handling.
1519 1516 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1520 1517
1521 1518 // Do not intrinsify on older platforms which lack cmove.
1522 1519 if (ConditionalMoveLimit == 0) return false;
1523 1520
1524 1521 // Pseudocode for pow
1525 1522 // if (x <= 0.0) {
1526 1523 // if ((double)((int)y)==y) { // if y is int
1527 1524 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1528 1525 // } else {
1529 1526 // result = NaN;
1530 1527 // }
1531 1528 // } else {
1532 1529 // result = DPow(x,y);
1533 1530 // }
1534 1531 // if (result != result)? {
1535 1532 // uncommon_trap();
1536 1533 // }
1537 1534 // return result;
1538 1535
1539 1536 _sp += arg_size(); // restore stack pointer
1540 1537 Node* y = pop_math_arg();
1541 1538 Node* x = pop_math_arg();
1542 1539
1543 1540 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1544 1541
1545 1542 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1546 1543 // inside of something) then skip the fancy tests and just check for
1547 1544 // NaN result.
1548 1545 Node *result = NULL;
1549 1546 if( jvms()->depth() >= 1 ) {
1550 1547 result = fast_result;
1551 1548 } else {
1552 1549
1553 1550 // Set the merge point for If node with condition of (x <= 0.0)
1554 1551 // There are four possible paths to region node and phi node
1555 1552 RegionNode *r = new (C, 4) RegionNode(4);
1556 1553 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1557 1554
1558 1555 // Build the first if node: if (x <= 0.0)
1559 1556 // Node for 0 constant
1560 1557 Node *zeronode = makecon(TypeD::ZERO);
1561 1558 // Check x:0
1562 1559 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1563 1560 // Check: If (x<=0) then go complex path
1564 1561 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1565 1562 // Branch either way
1566 1563 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1567 1564 Node *opt_test = _gvn.transform(if1);
1568 1565 //assert( opt_test->is_If(), "Expect an IfNode");
1569 1566 IfNode *opt_if1 = (IfNode*)opt_test;
1570 1567 // Fast path taken; set region slot 3
1571 1568 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1572 1569 r->init_req(3,fast_taken); // Capture fast-control
1573 1570
1574 1571 // Fast path not-taken, i.e. slow path
1575 1572 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1576 1573
1577 1574 // Set fast path result
1578 1575 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1579 1576 phi->init_req(3, fast_result);
1580 1577
1581 1578 // Complex path
1582 1579 // Build the second if node (if y is int)
1583 1580 // Node for (int)y
1584 1581 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1585 1582 // Node for (double)((int) y)
1586 1583 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1587 1584 // Check (double)((int) y) : y
1588 1585 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1589 1586 // Check if (y isn't int) then go to slow path
1590 1587
1591 1588 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1592 1589 // Branch either way
1593 1590 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1594 1591 Node *slow_path = opt_iff(r,if2); // Set region path 2
1595 1592
1596 1593 // Calculate DPow(abs(x), y)*(1 & (int)y)
1597 1594 // Node for constant 1
1598 1595 Node *conone = intcon(1);
1599 1596 // 1& (int)y
1600 1597 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1601 1598 // zero node
1602 1599 Node *conzero = intcon(0);
1603 1600 // Check (1&(int)y)==0?
1604 1601 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1605 1602 // Check if (1&(int)y)!=0?, if so the result is negative
1606 1603 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1607 1604 // abs(x)
1608 1605 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1609 1606 // abs(x)^y
1610 1607 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1611 1608 // -abs(x)^y
1612 1609 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1613 1610 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1614 1611 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1615 1612 // Set complex path fast result
1616 1613 phi->init_req(2, signresult);
1617 1614
1618 1615 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1619 1616 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1620 1617 r->init_req(1,slow_path);
1621 1618 phi->init_req(1,slow_result);
1622 1619
1623 1620 // Post merge
1624 1621 set_control(_gvn.transform(r));
1625 1622 record_for_igvn(r);
1626 1623 result=_gvn.transform(phi);
1627 1624 }
1628 1625
1629 1626 //-------------------
1630 1627 //result=(result.isNaN())? uncommon_trap():result;
1631 1628 // Check: If isNaN() by checking result!=result? then go to Strict Math
1632 1629 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1633 1630 // Build the boolean node
1634 1631 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1635 1632
1636 1633 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1637 1634 // End the current control-flow path
1638 1635 push_pair(x);
1639 1636 push_pair(y);
1640 1637 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1641 1638 // to handle. Recompile without intrinsifying Math.pow.
1642 1639 uncommon_trap(Deoptimization::Reason_intrinsic,
1643 1640 Deoptimization::Action_make_not_entrant);
1644 1641 }
1645 1642
1646 1643 C->set_has_split_ifs(true); // Has chance for split-if optimization
1647 1644
1648 1645 push_pair(result);
1649 1646
1650 1647 return true;
1651 1648 }
1652 1649
1653 1650 //------------------------------inline_trans-------------------------------------
1654 1651 // Inline transcendental instructions, if possible. The Intel hardware gets
1655 1652 // these right, no funny corner cases missed.
1656 1653 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1657 1654 _sp += arg_size(); // restore stack pointer
1658 1655 Node* arg = pop_math_arg();
1659 1656 Node* trans = NULL;
1660 1657
1661 1658 switch (id) {
1662 1659 case vmIntrinsics::_dlog:
1663 1660 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1664 1661 break;
1665 1662 case vmIntrinsics::_dlog10:
1666 1663 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1667 1664 break;
1668 1665 default:
1669 1666 assert(false, "bad intrinsic was passed in");
1670 1667 return false;
1671 1668 }
1672 1669
1673 1670 // Push result back on JVM stack
1674 1671 push_pair(trans);
1675 1672 return true;
1676 1673 }
1677 1674
1678 1675 //------------------------------runtime_math-----------------------------
1679 1676 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1680 1677 Node* a = NULL;
1681 1678 Node* b = NULL;
1682 1679
1683 1680 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1684 1681 "must be (DD)D or (D)D type");
1685 1682
1686 1683 // Inputs
1687 1684 _sp += arg_size(); // restore stack pointer
1688 1685 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1689 1686 b = pop_math_arg();
1690 1687 }
1691 1688 a = pop_math_arg();
1692 1689
1693 1690 const TypePtr* no_memory_effects = NULL;
1694 1691 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1695 1692 no_memory_effects,
1696 1693 a, top(), b, b ? top() : NULL);
1697 1694 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1698 1695 #ifdef ASSERT
1699 1696 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1700 1697 assert(value_top == top(), "second value must be top");
1701 1698 #endif
1702 1699
1703 1700 push_pair(value);
1704 1701 return true;
1705 1702 }
1706 1703
1707 1704 //------------------------------inline_math_native-----------------------------
1708 1705 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1709 1706 switch (id) {
1710 1707 // These intrinsics are not properly supported on all hardware
1711 1708 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1712 1709 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1713 1710 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1714 1711 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1715 1712 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1716 1713 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1717 1714
1718 1715 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1719 1716 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1720 1717 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1721 1718 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1722 1719
1723 1720 // These intrinsics are supported on all hardware
1724 1721 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1725 1722 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1726 1723
1727 1724 // These intrinsics don't work on X86. The ad implementation doesn't
1728 1725 // handle NaN's properly. Instead of returning infinity, the ad
1729 1726 // implementation returns a NaN on overflow. See bug: 6304089
1730 1727 // Once the ad implementations are fixed, change the code below
1731 1728 // to match the intrinsics above
1732 1729
1733 1730 case vmIntrinsics::_dexp: return
1734 1731 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1735 1732 case vmIntrinsics::_dpow: return
1736 1733 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1737 1734
1738 1735 // These intrinsics are not yet correctly implemented
1739 1736 case vmIntrinsics::_datan2:
1740 1737 return false;
1741 1738
1742 1739 default:
1743 1740 ShouldNotReachHere();
1744 1741 return false;
1745 1742 }
1746 1743 }
1747 1744
1748 1745 static bool is_simple_name(Node* n) {
1749 1746 return (n->req() == 1 // constant
1750 1747 || (n->is_Type() && n->as_Type()->type()->singleton())
1751 1748 || n->is_Proj() // parameter or return value
1752 1749 || n->is_Phi() // local of some sort
1753 1750 );
1754 1751 }
1755 1752
1756 1753 //----------------------------inline_min_max-----------------------------------
1757 1754 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1758 1755 push(generate_min_max(id, argument(0), argument(1)));
1759 1756
1760 1757 return true;
1761 1758 }
1762 1759
1763 1760 Node*
1764 1761 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1765 1762 // These are the candidate return value:
1766 1763 Node* xvalue = x0;
1767 1764 Node* yvalue = y0;
1768 1765
1769 1766 if (xvalue == yvalue) {
1770 1767 return xvalue;
1771 1768 }
1772 1769
1773 1770 bool want_max = (id == vmIntrinsics::_max);
1774 1771
1775 1772 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1776 1773 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1777 1774 if (txvalue == NULL || tyvalue == NULL) return top();
1778 1775 // This is not really necessary, but it is consistent with a
1779 1776 // hypothetical MaxINode::Value method:
1780 1777 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1781 1778
1782 1779 // %%% This folding logic should (ideally) be in a different place.
1783 1780 // Some should be inside IfNode, and there to be a more reliable
1784 1781 // transformation of ?: style patterns into cmoves. We also want
1785 1782 // more powerful optimizations around cmove and min/max.
1786 1783
1787 1784 // Try to find a dominating comparison of these guys.
1788 1785 // It can simplify the index computation for Arrays.copyOf
1789 1786 // and similar uses of System.arraycopy.
1790 1787 // First, compute the normalized version of CmpI(x, y).
1791 1788 int cmp_op = Op_CmpI;
1792 1789 Node* xkey = xvalue;
1793 1790 Node* ykey = yvalue;
1794 1791 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1795 1792 if (ideal_cmpxy->is_Cmp()) {
1796 1793 // E.g., if we have CmpI(length - offset, count),
1797 1794 // it might idealize to CmpI(length, count + offset)
1798 1795 cmp_op = ideal_cmpxy->Opcode();
1799 1796 xkey = ideal_cmpxy->in(1);
1800 1797 ykey = ideal_cmpxy->in(2);
1801 1798 }
1802 1799
1803 1800 // Start by locating any relevant comparisons.
1804 1801 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1805 1802 Node* cmpxy = NULL;
1806 1803 Node* cmpyx = NULL;
1807 1804 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1808 1805 Node* cmp = start_from->fast_out(k);
1809 1806 if (cmp->outcnt() > 0 && // must have prior uses
1810 1807 cmp->in(0) == NULL && // must be context-independent
1811 1808 cmp->Opcode() == cmp_op) { // right kind of compare
1812 1809 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1813 1810 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1814 1811 }
1815 1812 }
1816 1813
1817 1814 const int NCMPS = 2;
1818 1815 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1819 1816 int cmpn;
1820 1817 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1821 1818 if (cmps[cmpn] != NULL) break; // find a result
1822 1819 }
1823 1820 if (cmpn < NCMPS) {
1824 1821 // Look for a dominating test that tells us the min and max.
1825 1822 int depth = 0; // Limit search depth for speed
1826 1823 Node* dom = control();
1827 1824 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1828 1825 if (++depth >= 100) break;
1829 1826 Node* ifproj = dom;
1830 1827 if (!ifproj->is_Proj()) continue;
1831 1828 Node* iff = ifproj->in(0);
1832 1829 if (!iff->is_If()) continue;
1833 1830 Node* bol = iff->in(1);
1834 1831 if (!bol->is_Bool()) continue;
1835 1832 Node* cmp = bol->in(1);
1836 1833 if (cmp == NULL) continue;
1837 1834 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1838 1835 if (cmps[cmpn] == cmp) break;
1839 1836 if (cmpn == NCMPS) continue;
1840 1837 BoolTest::mask btest = bol->as_Bool()->_test._test;
1841 1838 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1842 1839 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1843 1840 // At this point, we know that 'x btest y' is true.
1844 1841 switch (btest) {
1845 1842 case BoolTest::eq:
1846 1843 // They are proven equal, so we can collapse the min/max.
1847 1844 // Either value is the answer. Choose the simpler.
1848 1845 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1849 1846 return yvalue;
1850 1847 return xvalue;
1851 1848 case BoolTest::lt: // x < y
1852 1849 case BoolTest::le: // x <= y
1853 1850 return (want_max ? yvalue : xvalue);
1854 1851 case BoolTest::gt: // x > y
1855 1852 case BoolTest::ge: // x >= y
1856 1853 return (want_max ? xvalue : yvalue);
1857 1854 }
1858 1855 }
1859 1856 }
1860 1857
1861 1858 // We failed to find a dominating test.
1862 1859 // Let's pick a test that might GVN with prior tests.
1863 1860 Node* best_bol = NULL;
1864 1861 BoolTest::mask best_btest = BoolTest::illegal;
1865 1862 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1866 1863 Node* cmp = cmps[cmpn];
1867 1864 if (cmp == NULL) continue;
1868 1865 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1869 1866 Node* bol = cmp->fast_out(j);
1870 1867 if (!bol->is_Bool()) continue;
1871 1868 BoolTest::mask btest = bol->as_Bool()->_test._test;
1872 1869 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1873 1870 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1874 1871 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1875 1872 best_bol = bol->as_Bool();
1876 1873 best_btest = btest;
1877 1874 }
1878 1875 }
1879 1876 }
1880 1877
1881 1878 Node* answer_if_true = NULL;
1882 1879 Node* answer_if_false = NULL;
1883 1880 switch (best_btest) {
1884 1881 default:
1885 1882 if (cmpxy == NULL)
1886 1883 cmpxy = ideal_cmpxy;
1887 1884 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1888 1885 // and fall through:
1889 1886 case BoolTest::lt: // x < y
1890 1887 case BoolTest::le: // x <= y
1891 1888 answer_if_true = (want_max ? yvalue : xvalue);
1892 1889 answer_if_false = (want_max ? xvalue : yvalue);
1893 1890 break;
1894 1891 case BoolTest::gt: // x > y
1895 1892 case BoolTest::ge: // x >= y
1896 1893 answer_if_true = (want_max ? xvalue : yvalue);
1897 1894 answer_if_false = (want_max ? yvalue : xvalue);
1898 1895 break;
1899 1896 }
1900 1897
1901 1898 jint hi, lo;
1902 1899 if (want_max) {
1903 1900 // We can sharpen the minimum.
1904 1901 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1905 1902 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1906 1903 } else {
1907 1904 // We can sharpen the maximum.
1908 1905 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1909 1906 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1910 1907 }
1911 1908
1912 1909 // Use a flow-free graph structure, to avoid creating excess control edges
1913 1910 // which could hinder other optimizations.
1914 1911 // Since Math.min/max is often used with arraycopy, we want
1915 1912 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1916 1913 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1917 1914 answer_if_false, answer_if_true,
1918 1915 TypeInt::make(lo, hi, widen));
1919 1916
1920 1917 return _gvn.transform(cmov);
1921 1918
1922 1919 /*
1923 1920 // This is not as desirable as it may seem, since Min and Max
1924 1921 // nodes do not have a full set of optimizations.
1925 1922 // And they would interfere, anyway, with 'if' optimizations
1926 1923 // and with CMoveI canonical forms.
1927 1924 switch (id) {
1928 1925 case vmIntrinsics::_min:
1929 1926 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1930 1927 case vmIntrinsics::_max:
1931 1928 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1932 1929 default:
1933 1930 ShouldNotReachHere();
1934 1931 }
1935 1932 */
1936 1933 }
1937 1934
1938 1935 inline int
1939 1936 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1940 1937 const TypePtr* base_type = TypePtr::NULL_PTR;
1941 1938 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1942 1939 if (base_type == NULL) {
1943 1940 // Unknown type.
1944 1941 return Type::AnyPtr;
1945 1942 } else if (base_type == TypePtr::NULL_PTR) {
1946 1943 // Since this is a NULL+long form, we have to switch to a rawptr.
1947 1944 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1948 1945 offset = MakeConX(0);
1949 1946 return Type::RawPtr;
1950 1947 } else if (base_type->base() == Type::RawPtr) {
1951 1948 return Type::RawPtr;
1952 1949 } else if (base_type->isa_oopptr()) {
1953 1950 // Base is never null => always a heap address.
1954 1951 if (base_type->ptr() == TypePtr::NotNull) {
1955 1952 return Type::OopPtr;
1956 1953 }
1957 1954 // Offset is small => always a heap address.
1958 1955 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1959 1956 if (offset_type != NULL &&
1960 1957 base_type->offset() == 0 && // (should always be?)
1961 1958 offset_type->_lo >= 0 &&
1962 1959 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1963 1960 return Type::OopPtr;
1964 1961 }
1965 1962 // Otherwise, it might either be oop+off or NULL+addr.
1966 1963 return Type::AnyPtr;
1967 1964 } else {
1968 1965 // No information:
1969 1966 return Type::AnyPtr;
1970 1967 }
1971 1968 }
1972 1969
1973 1970 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1974 1971 int kind = classify_unsafe_addr(base, offset);
1975 1972 if (kind == Type::RawPtr) {
1976 1973 return basic_plus_adr(top(), base, offset);
1977 1974 } else {
1978 1975 return basic_plus_adr(base, offset);
1979 1976 }
1980 1977 }
1981 1978
1982 1979 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1983 1980 // inline int Integer.numberOfLeadingZeros(int)
1984 1981 // inline int Long.numberOfLeadingZeros(long)
1985 1982 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1986 1983 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1987 1984 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1988 1985 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1989 1986 _sp += arg_size(); // restore stack pointer
1990 1987 switch (id) {
1991 1988 case vmIntrinsics::_numberOfLeadingZeros_i:
1992 1989 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1993 1990 break;
1994 1991 case vmIntrinsics::_numberOfLeadingZeros_l:
1995 1992 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1996 1993 break;
1997 1994 default:
1998 1995 ShouldNotReachHere();
1999 1996 }
2000 1997 return true;
2001 1998 }
2002 1999
2003 2000 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2004 2001 // inline int Integer.numberOfTrailingZeros(int)
2005 2002 // inline int Long.numberOfTrailingZeros(long)
2006 2003 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2007 2004 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2008 2005 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2009 2006 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2010 2007 _sp += arg_size(); // restore stack pointer
2011 2008 switch (id) {
2012 2009 case vmIntrinsics::_numberOfTrailingZeros_i:
2013 2010 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2014 2011 break;
2015 2012 case vmIntrinsics::_numberOfTrailingZeros_l:
2016 2013 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2017 2014 break;
2018 2015 default:
2019 2016 ShouldNotReachHere();
2020 2017 }
2021 2018 return true;
2022 2019 }
2023 2020
2024 2021 //----------------------------inline_bitCount_int/long-----------------------
2025 2022 // inline int Integer.bitCount(int)
2026 2023 // inline int Long.bitCount(long)
2027 2024 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2028 2025 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2029 2026 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2030 2027 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2031 2028 _sp += arg_size(); // restore stack pointer
2032 2029 switch (id) {
2033 2030 case vmIntrinsics::_bitCount_i:
2034 2031 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2035 2032 break;
2036 2033 case vmIntrinsics::_bitCount_l:
2037 2034 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2038 2035 break;
2039 2036 default:
2040 2037 ShouldNotReachHere();
2041 2038 }
2042 2039 return true;
2043 2040 }
2044 2041
2045 2042 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2046 2043 // inline Integer.reverseBytes(int)
2047 2044 // inline Long.reverseBytes(long)
2048 2045 // inline Character.reverseBytes(char)
2049 2046 // inline Short.reverseBytes(short)
2050 2047 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2051 2048 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2052 2049 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2053 2050 "not reverse Bytes");
2054 2051 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2055 2052 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2056 2053 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2057 2054 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2058 2055 _sp += arg_size(); // restore stack pointer
2059 2056 switch (id) {
2060 2057 case vmIntrinsics::_reverseBytes_i:
2061 2058 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2062 2059 break;
2063 2060 case vmIntrinsics::_reverseBytes_l:
2064 2061 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2065 2062 break;
2066 2063 case vmIntrinsics::_reverseBytes_c:
2067 2064 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2068 2065 break;
2069 2066 case vmIntrinsics::_reverseBytes_s:
2070 2067 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2071 2068 break;
2072 2069 default:
2073 2070 ;
2074 2071 }
2075 2072 return true;
2076 2073 }
2077 2074
2078 2075 //----------------------------inline_unsafe_access----------------------------
2079 2076
2080 2077 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2081 2078
2082 2079 // Interpret Unsafe.fieldOffset cookies correctly:
2083 2080 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2084 2081
2085 2082 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2086 2083 if (callee()->is_static()) return false; // caller must have the capability!
2087 2084
2088 2085 #ifndef PRODUCT
2089 2086 {
2090 2087 ResourceMark rm;
2091 2088 // Check the signatures.
2092 2089 ciSignature* sig = signature();
2093 2090 #ifdef ASSERT
2094 2091 if (!is_store) {
2095 2092 // Object getObject(Object base, int/long offset), etc.
2096 2093 BasicType rtype = sig->return_type()->basic_type();
2097 2094 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2098 2095 rtype = T_ADDRESS; // it is really a C void*
2099 2096 assert(rtype == type, "getter must return the expected value");
2100 2097 if (!is_native_ptr) {
2101 2098 assert(sig->count() == 2, "oop getter has 2 arguments");
2102 2099 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2103 2100 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2104 2101 } else {
2105 2102 assert(sig->count() == 1, "native getter has 1 argument");
2106 2103 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2107 2104 }
2108 2105 } else {
2109 2106 // void putObject(Object base, int/long offset, Object x), etc.
2110 2107 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2111 2108 if (!is_native_ptr) {
2112 2109 assert(sig->count() == 3, "oop putter has 3 arguments");
2113 2110 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2114 2111 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2115 2112 } else {
2116 2113 assert(sig->count() == 2, "native putter has 2 arguments");
2117 2114 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2118 2115 }
2119 2116 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2120 2117 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2121 2118 vtype = T_ADDRESS; // it is really a C void*
2122 2119 assert(vtype == type, "putter must accept the expected value");
2123 2120 }
2124 2121 #endif // ASSERT
2125 2122 }
2126 2123 #endif //PRODUCT
2127 2124
2128 2125 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2129 2126
2130 2127 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2131 2128
2132 2129 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2133 2130 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2134 2131
2135 2132 debug_only(int saved_sp = _sp);
2136 2133 _sp += nargs;
2137 2134
2138 2135 Node* val;
2139 2136 debug_only(val = (Node*)(uintptr_t)-1);
2140 2137
2141 2138
2142 2139 if (is_store) {
2143 2140 // Get the value being stored. (Pop it first; it was pushed last.)
2144 2141 switch (type) {
2145 2142 case T_DOUBLE:
2146 2143 case T_LONG:
2147 2144 case T_ADDRESS:
2148 2145 val = pop_pair();
2149 2146 break;
2150 2147 default:
2151 2148 val = pop();
2152 2149 }
2153 2150 }
2154 2151
2155 2152 // Build address expression. See the code in inline_unsafe_prefetch.
2156 2153 Node *adr;
2157 2154 Node *heap_base_oop = top();
2158 2155 if (!is_native_ptr) {
2159 2156 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2160 2157 Node* offset = pop_pair();
2161 2158 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2162 2159 Node* base = pop();
2163 2160 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2164 2161 // to be plain byte offsets, which are also the same as those accepted
2165 2162 // by oopDesc::field_base.
2166 2163 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2167 2164 "fieldOffset must be byte-scaled");
2168 2165 // 32-bit machines ignore the high half!
2169 2166 offset = ConvL2X(offset);
2170 2167 adr = make_unsafe_address(base, offset);
2171 2168 heap_base_oop = base;
2172 2169 } else {
2173 2170 Node* ptr = pop_pair();
2174 2171 // Adjust Java long to machine word:
2175 2172 ptr = ConvL2X(ptr);
2176 2173 adr = make_unsafe_address(NULL, ptr);
2177 2174 }
2178 2175
2179 2176 // Pop receiver last: it was pushed first.
2180 2177 Node *receiver = pop();
2181 2178
2182 2179 assert(saved_sp == _sp, "must have correct argument count");
2183 2180
2184 2181 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2185 2182
2186 2183 // First guess at the value type.
2187 2184 const Type *value_type = Type::get_const_basic_type(type);
2188 2185
2189 2186 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2190 2187 // there was not enough information to nail it down.
2191 2188 Compile::AliasType* alias_type = C->alias_type(adr_type);
2192 2189 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2193 2190
2194 2191 // We will need memory barriers unless we can determine a unique
2195 2192 // alias category for this reference. (Note: If for some reason
2196 2193 // the barriers get omitted and the unsafe reference begins to "pollute"
2197 2194 // the alias analysis of the rest of the graph, either Compile::can_alias
2198 2195 // or Compile::must_alias will throw a diagnostic assert.)
2199 2196 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2200 2197
2201 2198 if (!is_store && type == T_OBJECT) {
2202 2199 // Attempt to infer a sharper value type from the offset and base type.
2203 2200 ciKlass* sharpened_klass = NULL;
2204 2201
2205 2202 // See if it is an instance field, with an object type.
2206 2203 if (alias_type->field() != NULL) {
2207 2204 assert(!is_native_ptr, "native pointer op cannot use a java address");
2208 2205 if (alias_type->field()->type()->is_klass()) {
2209 2206 sharpened_klass = alias_type->field()->type()->as_klass();
2210 2207 }
2211 2208 }
2212 2209
2213 2210 // See if it is a narrow oop array.
2214 2211 if (adr_type->isa_aryptr()) {
2215 2212 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2216 2213 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2217 2214 if (elem_type != NULL) {
2218 2215 sharpened_klass = elem_type->klass();
2219 2216 }
2220 2217 }
2221 2218 }
2222 2219
2223 2220 if (sharpened_klass != NULL) {
2224 2221 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2225 2222
2226 2223 // Sharpen the value type.
2227 2224 value_type = tjp;
2228 2225
2229 2226 #ifndef PRODUCT
2230 2227 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2231 2228 tty->print(" from base type: "); adr_type->dump();
2232 2229 tty->print(" sharpened value: "); value_type->dump();
2233 2230 }
2234 2231 #endif
2235 2232 }
2236 2233 }
2237 2234
2238 2235 // Null check on self without removing any arguments. The argument
2239 2236 // null check technically happens in the wrong place, which can lead to
2240 2237 // invalid stack traces when the primitive is inlined into a method
2241 2238 // which handles NullPointerExceptions.
2242 2239 _sp += nargs;
2243 2240 do_null_check(receiver, T_OBJECT);
2244 2241 _sp -= nargs;
2245 2242 if (stopped()) {
2246 2243 return true;
2247 2244 }
2248 2245 // Heap pointers get a null-check from the interpreter,
2249 2246 // as a courtesy. However, this is not guaranteed by Unsafe,
2250 2247 // and it is not possible to fully distinguish unintended nulls
2251 2248 // from intended ones in this API.
2252 2249
2253 2250 if (is_volatile) {
2254 2251 // We need to emit leading and trailing CPU membars (see below) in
2255 2252 // addition to memory membars when is_volatile. This is a little
2256 2253 // too strong, but avoids the need to insert per-alias-type
2257 2254 // volatile membars (for stores; compare Parse::do_put_xxx), which
2258 2255 // we cannot do effectively here because we probably only have a
2259 2256 // rough approximation of type.
2260 2257 need_mem_bar = true;
2261 2258 // For Stores, place a memory ordering barrier now.
2262 2259 if (is_store)
2263 2260 insert_mem_bar(Op_MemBarRelease);
2264 2261 }
2265 2262
2266 2263 // Memory barrier to prevent normal and 'unsafe' accesses from
2267 2264 // bypassing each other. Happens after null checks, so the
2268 2265 // exception paths do not take memory state from the memory barrier,
2269 2266 // so there's no problems making a strong assert about mixing users
2270 2267 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2271 2268 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2272 2269 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2273 2270
2274 2271 if (!is_store) {
2275 2272 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2276 2273 // load value and push onto stack
2277 2274 switch (type) {
2278 2275 case T_BOOLEAN:
2279 2276 case T_CHAR:
2280 2277 case T_BYTE:
2281 2278 case T_SHORT:
2282 2279 case T_INT:
2283 2280 case T_FLOAT:
2284 2281 case T_OBJECT:
2285 2282 push( p );
2286 2283 break;
2287 2284 case T_ADDRESS:
2288 2285 // Cast to an int type.
2289 2286 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2290 2287 p = ConvX2L(p);
2291 2288 push_pair(p);
2292 2289 break;
2293 2290 case T_DOUBLE:
2294 2291 case T_LONG:
2295 2292 push_pair( p );
2296 2293 break;
2297 2294 default: ShouldNotReachHere();
2298 2295 }
2299 2296 } else {
2300 2297 // place effect of store into memory
2301 2298 switch (type) {
2302 2299 case T_DOUBLE:
2303 2300 val = dstore_rounding(val);
2304 2301 break;
2305 2302 case T_ADDRESS:
2306 2303 // Repackage the long as a pointer.
2307 2304 val = ConvL2X(val);
2308 2305 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2309 2306 break;
2310 2307 }
2311 2308
2312 2309 if (type != T_OBJECT ) {
2313 2310 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2314 2311 } else {
2315 2312 // Possibly an oop being stored to Java heap or native memory
2316 2313 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2317 2314 // oop to Java heap.
2318 2315 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2319 2316 } else {
2320 2317 // We can't tell at compile time if we are storing in the Java heap or outside
2321 2318 // of it. So we need to emit code to conditionally do the proper type of
2322 2319 // store.
2323 2320
2324 2321 IdealKit ideal(gvn(), control(), merged_memory());
2325 2322 #define __ ideal.
2326 2323 // QQQ who knows what probability is here??
2327 2324 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2328 2325 // Sync IdealKit and graphKit.
2329 2326 set_all_memory( __ merged_memory());
2330 2327 set_control(__ ctrl());
2331 2328 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2332 2329 // Update IdealKit memory.
2333 2330 __ set_all_memory(merged_memory());
2334 2331 __ set_ctrl(control());
2335 2332 } __ else_(); {
2336 2333 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2337 2334 } __ end_if();
2338 2335 // Final sync IdealKit and GraphKit.
2339 2336 sync_kit(ideal);
2340 2337 #undef __
2341 2338 }
2342 2339 }
2343 2340 }
2344 2341
2345 2342 if (is_volatile) {
2346 2343 if (!is_store)
2347 2344 insert_mem_bar(Op_MemBarAcquire);
2348 2345 else
2349 2346 insert_mem_bar(Op_MemBarVolatile);
2350 2347 }
2351 2348
2352 2349 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2353 2350
2354 2351 return true;
2355 2352 }
2356 2353
2357 2354 //----------------------------inline_unsafe_prefetch----------------------------
2358 2355
2359 2356 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2360 2357 #ifndef PRODUCT
2361 2358 {
2362 2359 ResourceMark rm;
2363 2360 // Check the signatures.
2364 2361 ciSignature* sig = signature();
2365 2362 #ifdef ASSERT
2366 2363 // Object getObject(Object base, int/long offset), etc.
2367 2364 BasicType rtype = sig->return_type()->basic_type();
2368 2365 if (!is_native_ptr) {
2369 2366 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2370 2367 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2371 2368 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2372 2369 } else {
2373 2370 assert(sig->count() == 1, "native prefetch has 1 argument");
2374 2371 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2375 2372 }
2376 2373 #endif // ASSERT
2377 2374 }
2378 2375 #endif // !PRODUCT
2379 2376
2380 2377 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2381 2378
2382 2379 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2383 2380 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2384 2381
2385 2382 debug_only(int saved_sp = _sp);
2386 2383 _sp += nargs;
2387 2384
2388 2385 // Build address expression. See the code in inline_unsafe_access.
2389 2386 Node *adr;
2390 2387 if (!is_native_ptr) {
2391 2388 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2392 2389 Node* offset = pop_pair();
2393 2390 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2394 2391 Node* base = pop();
2395 2392 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2396 2393 // to be plain byte offsets, which are also the same as those accepted
2397 2394 // by oopDesc::field_base.
2398 2395 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2399 2396 "fieldOffset must be byte-scaled");
2400 2397 // 32-bit machines ignore the high half!
2401 2398 offset = ConvL2X(offset);
2402 2399 adr = make_unsafe_address(base, offset);
2403 2400 } else {
2404 2401 Node* ptr = pop_pair();
2405 2402 // Adjust Java long to machine word:
2406 2403 ptr = ConvL2X(ptr);
2407 2404 adr = make_unsafe_address(NULL, ptr);
2408 2405 }
2409 2406
2410 2407 if (is_static) {
2411 2408 assert(saved_sp == _sp, "must have correct argument count");
2412 2409 } else {
2413 2410 // Pop receiver last: it was pushed first.
2414 2411 Node *receiver = pop();
2415 2412 assert(saved_sp == _sp, "must have correct argument count");
2416 2413
2417 2414 // Null check on self without removing any arguments. The argument
2418 2415 // null check technically happens in the wrong place, which can lead to
2419 2416 // invalid stack traces when the primitive is inlined into a method
2420 2417 // which handles NullPointerExceptions.
2421 2418 _sp += nargs;
2422 2419 do_null_check(receiver, T_OBJECT);
2423 2420 _sp -= nargs;
2424 2421 if (stopped()) {
2425 2422 return true;
2426 2423 }
2427 2424 }
2428 2425
2429 2426 // Generate the read or write prefetch
2430 2427 Node *prefetch;
2431 2428 if (is_store) {
2432 2429 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2433 2430 } else {
2434 2431 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2435 2432 }
2436 2433 prefetch->init_req(0, control());
2437 2434 set_i_o(_gvn.transform(prefetch));
2438 2435
2439 2436 return true;
2440 2437 }
2441 2438
2442 2439 //----------------------------inline_unsafe_CAS----------------------------
2443 2440
2444 2441 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2445 2442 // This basic scheme here is the same as inline_unsafe_access, but
2446 2443 // differs in enough details that combining them would make the code
2447 2444 // overly confusing. (This is a true fact! I originally combined
2448 2445 // them, but even I was confused by it!) As much code/comments as
2449 2446 // possible are retained from inline_unsafe_access though to make
2450 2447 // the correspondences clearer. - dl
2451 2448
2452 2449 if (callee()->is_static()) return false; // caller must have the capability!
2453 2450
2454 2451 #ifndef PRODUCT
2455 2452 {
2456 2453 ResourceMark rm;
2457 2454 // Check the signatures.
2458 2455 ciSignature* sig = signature();
2459 2456 #ifdef ASSERT
2460 2457 BasicType rtype = sig->return_type()->basic_type();
2461 2458 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2462 2459 assert(sig->count() == 4, "CAS has 4 arguments");
2463 2460 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2464 2461 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2465 2462 #endif // ASSERT
2466 2463 }
2467 2464 #endif //PRODUCT
2468 2465
2469 2466 // number of stack slots per value argument (1 or 2)
2470 2467 int type_words = type2size[type];
2471 2468
2472 2469 // Cannot inline wide CAS on machines that don't support it natively
2473 2470 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2474 2471 return false;
2475 2472
2476 2473 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2477 2474
2478 2475 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2479 2476 int nargs = 1 + 1 + 2 + type_words + type_words;
2480 2477
2481 2478 // pop arguments: newval, oldval, offset, base, and receiver
2482 2479 debug_only(int saved_sp = _sp);
2483 2480 _sp += nargs;
2484 2481 Node* newval = (type_words == 1) ? pop() : pop_pair();
2485 2482 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2486 2483 Node *offset = pop_pair();
2487 2484 Node *base = pop();
2488 2485 Node *receiver = pop();
2489 2486 assert(saved_sp == _sp, "must have correct argument count");
2490 2487
2491 2488 // Null check receiver.
2492 2489 _sp += nargs;
2493 2490 do_null_check(receiver, T_OBJECT);
2494 2491 _sp -= nargs;
2495 2492 if (stopped()) {
2496 2493 return true;
2497 2494 }
2498 2495
2499 2496 // Build field offset expression.
2500 2497 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2501 2498 // to be plain byte offsets, which are also the same as those accepted
2502 2499 // by oopDesc::field_base.
2503 2500 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2504 2501 // 32-bit machines ignore the high half of long offsets
2505 2502 offset = ConvL2X(offset);
2506 2503 Node* adr = make_unsafe_address(base, offset);
2507 2504 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2508 2505
2509 2506 // (Unlike inline_unsafe_access, there seems no point in trying
2510 2507 // to refine types. Just use the coarse types here.
2511 2508 const Type *value_type = Type::get_const_basic_type(type);
2512 2509 Compile::AliasType* alias_type = C->alias_type(adr_type);
2513 2510 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2514 2511 int alias_idx = C->get_alias_index(adr_type);
2515 2512
2516 2513 // Memory-model-wise, a CAS acts like a little synchronized block,
2517 2514 // so needs barriers on each side. These don't translate into
2518 2515 // actual barriers on most machines, but we still need rest of
2519 2516 // compiler to respect ordering.
2520 2517
2521 2518 insert_mem_bar(Op_MemBarRelease);
2522 2519 insert_mem_bar(Op_MemBarCPUOrder);
2523 2520
2524 2521 // 4984716: MemBars must be inserted before this
2525 2522 // memory node in order to avoid a false
2526 2523 // dependency which will confuse the scheduler.
2527 2524 Node *mem = memory(alias_idx);
2528 2525
2529 2526 // For now, we handle only those cases that actually exist: ints,
2530 2527 // longs, and Object. Adding others should be straightforward.
2531 2528 Node* cas;
2532 2529 switch(type) {
2533 2530 case T_INT:
2534 2531 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2535 2532 break;
2536 2533 case T_LONG:
2537 2534 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2538 2535 break;
2539 2536 case T_OBJECT:
2540 2537 // reference stores need a store barrier.
2541 2538 // (They don't if CAS fails, but it isn't worth checking.)
2542 2539 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2543 2540 #ifdef _LP64
2544 2541 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2545 2542 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2546 2543 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2547 2544 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2548 2545 newval_enc, oldval_enc));
2549 2546 } else
2550 2547 #endif
2551 2548 {
2552 2549 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2553 2550 }
2554 2551 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2555 2552 break;
2556 2553 default:
2557 2554 ShouldNotReachHere();
2558 2555 break;
2559 2556 }
2560 2557
2561 2558 // SCMemProjNodes represent the memory state of CAS. Their main
2562 2559 // role is to prevent CAS nodes from being optimized away when their
2563 2560 // results aren't used.
2564 2561 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2565 2562 set_memory(proj, alias_idx);
2566 2563
2567 2564 // Add the trailing membar surrounding the access
2568 2565 insert_mem_bar(Op_MemBarCPUOrder);
2569 2566 insert_mem_bar(Op_MemBarAcquire);
2570 2567
2571 2568 push(cas);
2572 2569 return true;
2573 2570 }
2574 2571
2575 2572 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2576 2573 // This is another variant of inline_unsafe_access, differing in
2577 2574 // that it always issues store-store ("release") barrier and ensures
2578 2575 // store-atomicity (which only matters for "long").
2579 2576
2580 2577 if (callee()->is_static()) return false; // caller must have the capability!
2581 2578
2582 2579 #ifndef PRODUCT
2583 2580 {
2584 2581 ResourceMark rm;
2585 2582 // Check the signatures.
2586 2583 ciSignature* sig = signature();
2587 2584 #ifdef ASSERT
2588 2585 BasicType rtype = sig->return_type()->basic_type();
2589 2586 assert(rtype == T_VOID, "must return void");
2590 2587 assert(sig->count() == 3, "has 3 arguments");
2591 2588 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2592 2589 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2593 2590 #endif // ASSERT
2594 2591 }
2595 2592 #endif //PRODUCT
2596 2593
2597 2594 // number of stack slots per value argument (1 or 2)
2598 2595 int type_words = type2size[type];
2599 2596
2600 2597 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2601 2598
2602 2599 // Argument words: "this" plus oop plus offset plus value;
2603 2600 int nargs = 1 + 1 + 2 + type_words;
2604 2601
2605 2602 // pop arguments: val, offset, base, and receiver
2606 2603 debug_only(int saved_sp = _sp);
2607 2604 _sp += nargs;
2608 2605 Node* val = (type_words == 1) ? pop() : pop_pair();
2609 2606 Node *offset = pop_pair();
2610 2607 Node *base = pop();
2611 2608 Node *receiver = pop();
2612 2609 assert(saved_sp == _sp, "must have correct argument count");
2613 2610
2614 2611 // Null check receiver.
2615 2612 _sp += nargs;
2616 2613 do_null_check(receiver, T_OBJECT);
2617 2614 _sp -= nargs;
2618 2615 if (stopped()) {
2619 2616 return true;
2620 2617 }
2621 2618
2622 2619 // Build field offset expression.
2623 2620 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2624 2621 // 32-bit machines ignore the high half of long offsets
2625 2622 offset = ConvL2X(offset);
2626 2623 Node* adr = make_unsafe_address(base, offset);
2627 2624 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2628 2625 const Type *value_type = Type::get_const_basic_type(type);
2629 2626 Compile::AliasType* alias_type = C->alias_type(adr_type);
2630 2627
2631 2628 insert_mem_bar(Op_MemBarRelease);
2632 2629 insert_mem_bar(Op_MemBarCPUOrder);
2633 2630 // Ensure that the store is atomic for longs:
2634 2631 bool require_atomic_access = true;
2635 2632 Node* store;
2636 2633 if (type == T_OBJECT) // reference stores need a store barrier.
2637 2634 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2638 2635 else {
2639 2636 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2640 2637 }
2641 2638 insert_mem_bar(Op_MemBarCPUOrder);
2642 2639 return true;
2643 2640 }
2644 2641
2645 2642 bool LibraryCallKit::inline_unsafe_allocate() {
2646 2643 if (callee()->is_static()) return false; // caller must have the capability!
2647 2644 int nargs = 1 + 1;
2648 2645 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2649 2646 null_check_receiver(callee()); // check then ignore argument(0)
2650 2647 _sp += nargs; // set original stack for use by uncommon_trap
2651 2648 Node* cls = do_null_check(argument(1), T_OBJECT);
2652 2649 _sp -= nargs;
2653 2650 if (stopped()) return true;
2654 2651
2655 2652 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2656 2653 _sp += nargs; // set original stack for use by uncommon_trap
2657 2654 kls = do_null_check(kls, T_OBJECT);
2658 2655 _sp -= nargs;
2659 2656 if (stopped()) return true; // argument was like int.class
2660 2657
2661 2658 // Note: The argument might still be an illegal value like
2662 2659 // Serializable.class or Object[].class. The runtime will handle it.
2663 2660 // But we must make an explicit check for initialization.
2664 2661 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2665 2662 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2666 2663 Node* bits = intcon(instanceKlass::fully_initialized);
2667 2664 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2668 2665 // The 'test' is non-zero if we need to take a slow path.
2669 2666
2670 2667 Node* obj = new_instance(kls, test);
2671 2668 push(obj);
2672 2669
2673 2670 return true;
2674 2671 }
2675 2672
2676 2673 //------------------------inline_native_time_funcs--------------
2677 2674 // inline code for System.currentTimeMillis() and System.nanoTime()
2678 2675 // these have the same type and signature
2679 2676 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2680 2677 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2681 2678 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2682 2679 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2683 2680 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2684 2681 const TypePtr* no_memory_effects = NULL;
2685 2682 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2686 2683 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2687 2684 #ifdef ASSERT
2688 2685 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2689 2686 assert(value_top == top(), "second value must be top");
2690 2687 #endif
2691 2688 push_pair(value);
2692 2689 return true;
2693 2690 }
2694 2691
2695 2692 //------------------------inline_native_currentThread------------------
2696 2693 bool LibraryCallKit::inline_native_currentThread() {
2697 2694 Node* junk = NULL;
2698 2695 push(generate_current_thread(junk));
2699 2696 return true;
2700 2697 }
2701 2698
2702 2699 //------------------------inline_native_isInterrupted------------------
2703 2700 bool LibraryCallKit::inline_native_isInterrupted() {
2704 2701 const int nargs = 1+1; // receiver + boolean
2705 2702 assert(nargs == arg_size(), "sanity");
2706 2703 // Add a fast path to t.isInterrupted(clear_int):
2707 2704 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2708 2705 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2709 2706 // So, in the common case that the interrupt bit is false,
2710 2707 // we avoid making a call into the VM. Even if the interrupt bit
2711 2708 // is true, if the clear_int argument is false, we avoid the VM call.
2712 2709 // However, if the receiver is not currentThread, we must call the VM,
2713 2710 // because there must be some locking done around the operation.
2714 2711
2715 2712 // We only go to the fast case code if we pass two guards.
2716 2713 // Paths which do not pass are accumulated in the slow_region.
2717 2714 RegionNode* slow_region = new (C, 1) RegionNode(1);
2718 2715 record_for_igvn(slow_region);
2719 2716 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2720 2717 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2721 2718 enum { no_int_result_path = 1,
2722 2719 no_clear_result_path = 2,
2723 2720 slow_result_path = 3
2724 2721 };
2725 2722
2726 2723 // (a) Receiving thread must be the current thread.
2727 2724 Node* rec_thr = argument(0);
2728 2725 Node* tls_ptr = NULL;
2729 2726 Node* cur_thr = generate_current_thread(tls_ptr);
2730 2727 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2731 2728 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2732 2729
2733 2730 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2734 2731 if (!known_current_thread)
2735 2732 generate_slow_guard(bol_thr, slow_region);
2736 2733
2737 2734 // (b) Interrupt bit on TLS must be false.
2738 2735 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2739 2736 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2740 2737 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2741 2738 // Set the control input on the field _interrupted read to prevent it floating up.
2742 2739 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2743 2740 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2744 2741 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2745 2742
2746 2743 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2747 2744
2748 2745 // First fast path: if (!TLS._interrupted) return false;
2749 2746 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2750 2747 result_rgn->init_req(no_int_result_path, false_bit);
2751 2748 result_val->init_req(no_int_result_path, intcon(0));
2752 2749
2753 2750 // drop through to next case
2754 2751 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2755 2752
2756 2753 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2757 2754 Node* clr_arg = argument(1);
2758 2755 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2759 2756 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2760 2757 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2761 2758
2762 2759 // Second fast path: ... else if (!clear_int) return true;
2763 2760 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2764 2761 result_rgn->init_req(no_clear_result_path, false_arg);
2765 2762 result_val->init_req(no_clear_result_path, intcon(1));
2766 2763
2767 2764 // drop through to next case
2768 2765 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2769 2766
2770 2767 // (d) Otherwise, go to the slow path.
2771 2768 slow_region->add_req(control());
2772 2769 set_control( _gvn.transform(slow_region) );
2773 2770
2774 2771 if (stopped()) {
2775 2772 // There is no slow path.
2776 2773 result_rgn->init_req(slow_result_path, top());
2777 2774 result_val->init_req(slow_result_path, top());
2778 2775 } else {
2779 2776 // non-virtual because it is a private non-static
2780 2777 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2781 2778
2782 2779 Node* slow_val = set_results_for_java_call(slow_call);
2783 2780 // this->control() comes from set_results_for_java_call
2784 2781
2785 2782 // If we know that the result of the slow call will be true, tell the optimizer!
2786 2783 if (known_current_thread) slow_val = intcon(1);
2787 2784
2788 2785 Node* fast_io = slow_call->in(TypeFunc::I_O);
2789 2786 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2790 2787 // These two phis are pre-filled with copies of of the fast IO and Memory
2791 2788 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2792 2789 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2793 2790
2794 2791 result_rgn->init_req(slow_result_path, control());
2795 2792 io_phi ->init_req(slow_result_path, i_o());
2796 2793 mem_phi ->init_req(slow_result_path, reset_memory());
2797 2794 result_val->init_req(slow_result_path, slow_val);
2798 2795
2799 2796 set_all_memory( _gvn.transform(mem_phi) );
2800 2797 set_i_o( _gvn.transform(io_phi) );
2801 2798 }
2802 2799
2803 2800 push_result(result_rgn, result_val);
2804 2801 C->set_has_split_ifs(true); // Has chance for split-if optimization
2805 2802
2806 2803 return true;
2807 2804 }
2808 2805
2809 2806 //---------------------------load_mirror_from_klass----------------------------
2810 2807 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2811 2808 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2812 2809 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2813 2810 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2814 2811 }
2815 2812
2816 2813 //-----------------------load_klass_from_mirror_common-------------------------
2817 2814 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2818 2815 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2819 2816 // and branch to the given path on the region.
2820 2817 // If never_see_null, take an uncommon trap on null, so we can optimistically
2821 2818 // compile for the non-null case.
2822 2819 // If the region is NULL, force never_see_null = true.
2823 2820 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2824 2821 bool never_see_null,
2825 2822 int nargs,
2826 2823 RegionNode* region,
2827 2824 int null_path,
2828 2825 int offset) {
2829 2826 if (region == NULL) never_see_null = true;
2830 2827 Node* p = basic_plus_adr(mirror, offset);
2831 2828 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2832 2829 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2833 2830 _sp += nargs; // any deopt will start just before call to enclosing method
2834 2831 Node* null_ctl = top();
2835 2832 kls = null_check_oop(kls, &null_ctl, never_see_null);
2836 2833 if (region != NULL) {
2837 2834 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2838 2835 region->init_req(null_path, null_ctl);
2839 2836 } else {
2840 2837 assert(null_ctl == top(), "no loose ends");
2841 2838 }
2842 2839 _sp -= nargs;
2843 2840 return kls;
2844 2841 }
2845 2842
2846 2843 //--------------------(inline_native_Class_query helpers)---------------------
2847 2844 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2848 2845 // Fall through if (mods & mask) == bits, take the guard otherwise.
2849 2846 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2850 2847 // Branch around if the given klass has the given modifier bit set.
2851 2848 // Like generate_guard, adds a new path onto the region.
2852 2849 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2853 2850 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2854 2851 Node* mask = intcon(modifier_mask);
2855 2852 Node* bits = intcon(modifier_bits);
2856 2853 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2857 2854 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2858 2855 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2859 2856 return generate_fair_guard(bol, region);
2860 2857 }
2861 2858 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2862 2859 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2863 2860 }
2864 2861
2865 2862 //-------------------------inline_native_Class_query-------------------
2866 2863 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2867 2864 int nargs = 1+0; // just the Class mirror, in most cases
2868 2865 const Type* return_type = TypeInt::BOOL;
2869 2866 Node* prim_return_value = top(); // what happens if it's a primitive class?
2870 2867 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2871 2868 bool expect_prim = false; // most of these guys expect to work on refs
2872 2869
2873 2870 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2874 2871
2875 2872 switch (id) {
2876 2873 case vmIntrinsics::_isInstance:
2877 2874 nargs = 1+1; // the Class mirror, plus the object getting queried about
2878 2875 // nothing is an instance of a primitive type
2879 2876 prim_return_value = intcon(0);
2880 2877 break;
2881 2878 case vmIntrinsics::_getModifiers:
2882 2879 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2883 2880 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2884 2881 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2885 2882 break;
2886 2883 case vmIntrinsics::_isInterface:
2887 2884 prim_return_value = intcon(0);
2888 2885 break;
2889 2886 case vmIntrinsics::_isArray:
2890 2887 prim_return_value = intcon(0);
2891 2888 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2892 2889 break;
2893 2890 case vmIntrinsics::_isPrimitive:
2894 2891 prim_return_value = intcon(1);
2895 2892 expect_prim = true; // obviously
2896 2893 break;
2897 2894 case vmIntrinsics::_getSuperclass:
2898 2895 prim_return_value = null();
2899 2896 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2900 2897 break;
2901 2898 case vmIntrinsics::_getComponentType:
2902 2899 prim_return_value = null();
2903 2900 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2904 2901 break;
2905 2902 case vmIntrinsics::_getClassAccessFlags:
2906 2903 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2907 2904 return_type = TypeInt::INT; // not bool! 6297094
2908 2905 break;
2909 2906 default:
2910 2907 ShouldNotReachHere();
2911 2908 }
2912 2909
2913 2910 Node* mirror = argument(0);
2914 2911 Node* obj = (nargs <= 1)? top(): argument(1);
2915 2912
2916 2913 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2917 2914 if (mirror_con == NULL) return false; // cannot happen?
2918 2915
2919 2916 #ifndef PRODUCT
2920 2917 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2921 2918 ciType* k = mirror_con->java_mirror_type();
2922 2919 if (k) {
2923 2920 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2924 2921 k->print_name();
2925 2922 tty->cr();
2926 2923 }
2927 2924 }
2928 2925 #endif
2929 2926
2930 2927 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2931 2928 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2932 2929 record_for_igvn(region);
2933 2930 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2934 2931
2935 2932 // The mirror will never be null of Reflection.getClassAccessFlags, however
2936 2933 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2937 2934 // if it is. See bug 4774291.
2938 2935
2939 2936 // For Reflection.getClassAccessFlags(), the null check occurs in
2940 2937 // the wrong place; see inline_unsafe_access(), above, for a similar
2941 2938 // situation.
2942 2939 _sp += nargs; // set original stack for use by uncommon_trap
2943 2940 mirror = do_null_check(mirror, T_OBJECT);
2944 2941 _sp -= nargs;
2945 2942 // If mirror or obj is dead, only null-path is taken.
2946 2943 if (stopped()) return true;
2947 2944
2948 2945 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2949 2946
2950 2947 // Now load the mirror's klass metaobject, and null-check it.
2951 2948 // Side-effects region with the control path if the klass is null.
2952 2949 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2953 2950 region, _prim_path);
2954 2951 // If kls is null, we have a primitive mirror.
2955 2952 phi->init_req(_prim_path, prim_return_value);
2956 2953 if (stopped()) { push_result(region, phi); return true; }
2957 2954
2958 2955 Node* p; // handy temp
2959 2956 Node* null_ctl;
2960 2957
2961 2958 // Now that we have the non-null klass, we can perform the real query.
2962 2959 // For constant classes, the query will constant-fold in LoadNode::Value.
2963 2960 Node* query_value = top();
2964 2961 switch (id) {
2965 2962 case vmIntrinsics::_isInstance:
2966 2963 // nothing is an instance of a primitive type
2967 2964 _sp += nargs; // gen_instanceof might do an uncommon trap
2968 2965 query_value = gen_instanceof(obj, kls);
2969 2966 _sp -= nargs;
2970 2967 break;
2971 2968
2972 2969 case vmIntrinsics::_getModifiers:
2973 2970 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2974 2971 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2975 2972 break;
2976 2973
2977 2974 case vmIntrinsics::_isInterface:
2978 2975 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2979 2976 if (generate_interface_guard(kls, region) != NULL)
2980 2977 // A guard was added. If the guard is taken, it was an interface.
2981 2978 phi->add_req(intcon(1));
2982 2979 // If we fall through, it's a plain class.
2983 2980 query_value = intcon(0);
2984 2981 break;
2985 2982
2986 2983 case vmIntrinsics::_isArray:
2987 2984 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2988 2985 if (generate_array_guard(kls, region) != NULL)
2989 2986 // A guard was added. If the guard is taken, it was an array.
2990 2987 phi->add_req(intcon(1));
2991 2988 // If we fall through, it's a plain class.
2992 2989 query_value = intcon(0);
2993 2990 break;
2994 2991
2995 2992 case vmIntrinsics::_isPrimitive:
2996 2993 query_value = intcon(0); // "normal" path produces false
2997 2994 break;
2998 2995
2999 2996 case vmIntrinsics::_getSuperclass:
3000 2997 // The rules here are somewhat unfortunate, but we can still do better
3001 2998 // with random logic than with a JNI call.
3002 2999 // Interfaces store null or Object as _super, but must report null.
3003 3000 // Arrays store an intermediate super as _super, but must report Object.
3004 3001 // Other types can report the actual _super.
3005 3002 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3006 3003 if (generate_interface_guard(kls, region) != NULL)
3007 3004 // A guard was added. If the guard is taken, it was an interface.
3008 3005 phi->add_req(null());
3009 3006 if (generate_array_guard(kls, region) != NULL)
3010 3007 // A guard was added. If the guard is taken, it was an array.
3011 3008 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3012 3009 // If we fall through, it's a plain class. Get its _super.
3013 3010 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
3014 3011 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3015 3012 null_ctl = top();
3016 3013 kls = null_check_oop(kls, &null_ctl);
3017 3014 if (null_ctl != top()) {
3018 3015 // If the guard is taken, Object.superClass is null (both klass and mirror).
3019 3016 region->add_req(null_ctl);
3020 3017 phi ->add_req(null());
3021 3018 }
3022 3019 if (!stopped()) {
3023 3020 query_value = load_mirror_from_klass(kls);
3024 3021 }
3025 3022 break;
3026 3023
3027 3024 case vmIntrinsics::_getComponentType:
3028 3025 if (generate_array_guard(kls, region) != NULL) {
3029 3026 // Be sure to pin the oop load to the guard edge just created:
3030 3027 Node* is_array_ctrl = region->in(region->req()-1);
3031 3028 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3032 3029 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3033 3030 phi->add_req(cmo);
3034 3031 }
3035 3032 query_value = null(); // non-array case is null
3036 3033 break;
3037 3034
3038 3035 case vmIntrinsics::_getClassAccessFlags:
3039 3036 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3040 3037 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3041 3038 break;
3042 3039
3043 3040 default:
3044 3041 ShouldNotReachHere();
3045 3042 }
3046 3043
3047 3044 // Fall-through is the normal case of a query to a real class.
3048 3045 phi->init_req(1, query_value);
3049 3046 region->init_req(1, control());
3050 3047
3051 3048 push_result(region, phi);
3052 3049 C->set_has_split_ifs(true); // Has chance for split-if optimization
3053 3050
3054 3051 return true;
3055 3052 }
3056 3053
3057 3054 //--------------------------inline_native_subtype_check------------------------
3058 3055 // This intrinsic takes the JNI calls out of the heart of
3059 3056 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3060 3057 bool LibraryCallKit::inline_native_subtype_check() {
3061 3058 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3062 3059
3063 3060 // Pull both arguments off the stack.
3064 3061 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3065 3062 args[0] = argument(0);
3066 3063 args[1] = argument(1);
3067 3064 Node* klasses[2]; // corresponding Klasses: superk, subk
3068 3065 klasses[0] = klasses[1] = top();
3069 3066
3070 3067 enum {
3071 3068 // A full decision tree on {superc is prim, subc is prim}:
3072 3069 _prim_0_path = 1, // {P,N} => false
3073 3070 // {P,P} & superc!=subc => false
3074 3071 _prim_same_path, // {P,P} & superc==subc => true
3075 3072 _prim_1_path, // {N,P} => false
3076 3073 _ref_subtype_path, // {N,N} & subtype check wins => true
3077 3074 _both_ref_path, // {N,N} & subtype check loses => false
3078 3075 PATH_LIMIT
3079 3076 };
3080 3077
3081 3078 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3082 3079 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3083 3080 record_for_igvn(region);
3084 3081
3085 3082 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3086 3083 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3087 3084 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3088 3085
3089 3086 // First null-check both mirrors and load each mirror's klass metaobject.
3090 3087 int which_arg;
3091 3088 for (which_arg = 0; which_arg <= 1; which_arg++) {
3092 3089 Node* arg = args[which_arg];
3093 3090 _sp += nargs; // set original stack for use by uncommon_trap
3094 3091 arg = do_null_check(arg, T_OBJECT);
3095 3092 _sp -= nargs;
3096 3093 if (stopped()) break;
3097 3094 args[which_arg] = _gvn.transform(arg);
3098 3095
3099 3096 Node* p = basic_plus_adr(arg, class_klass_offset);
3100 3097 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3101 3098 klasses[which_arg] = _gvn.transform(kls);
3102 3099 }
3103 3100
3104 3101 // Having loaded both klasses, test each for null.
3105 3102 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3106 3103 for (which_arg = 0; which_arg <= 1; which_arg++) {
3107 3104 Node* kls = klasses[which_arg];
3108 3105 Node* null_ctl = top();
3109 3106 _sp += nargs; // set original stack for use by uncommon_trap
3110 3107 kls = null_check_oop(kls, &null_ctl, never_see_null);
3111 3108 _sp -= nargs;
3112 3109 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3113 3110 region->init_req(prim_path, null_ctl);
3114 3111 if (stopped()) break;
3115 3112 klasses[which_arg] = kls;
3116 3113 }
3117 3114
3118 3115 if (!stopped()) {
3119 3116 // now we have two reference types, in klasses[0..1]
3120 3117 Node* subk = klasses[1]; // the argument to isAssignableFrom
3121 3118 Node* superk = klasses[0]; // the receiver
3122 3119 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3123 3120 // now we have a successful reference subtype check
3124 3121 region->set_req(_ref_subtype_path, control());
3125 3122 }
3126 3123
3127 3124 // If both operands are primitive (both klasses null), then
3128 3125 // we must return true when they are identical primitives.
3129 3126 // It is convenient to test this after the first null klass check.
3130 3127 set_control(region->in(_prim_0_path)); // go back to first null check
3131 3128 if (!stopped()) {
3132 3129 // Since superc is primitive, make a guard for the superc==subc case.
3133 3130 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3134 3131 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3135 3132 generate_guard(bol_eq, region, PROB_FAIR);
3136 3133 if (region->req() == PATH_LIMIT+1) {
3137 3134 // A guard was added. If the added guard is taken, superc==subc.
3138 3135 region->swap_edges(PATH_LIMIT, _prim_same_path);
3139 3136 region->del_req(PATH_LIMIT);
3140 3137 }
3141 3138 region->set_req(_prim_0_path, control()); // Not equal after all.
3142 3139 }
3143 3140
3144 3141 // these are the only paths that produce 'true':
3145 3142 phi->set_req(_prim_same_path, intcon(1));
3146 3143 phi->set_req(_ref_subtype_path, intcon(1));
3147 3144
3148 3145 // pull together the cases:
3149 3146 assert(region->req() == PATH_LIMIT, "sane region");
3150 3147 for (uint i = 1; i < region->req(); i++) {
3151 3148 Node* ctl = region->in(i);
3152 3149 if (ctl == NULL || ctl == top()) {
3153 3150 region->set_req(i, top());
3154 3151 phi ->set_req(i, top());
3155 3152 } else if (phi->in(i) == NULL) {
3156 3153 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3157 3154 }
3158 3155 }
3159 3156
3160 3157 set_control(_gvn.transform(region));
3161 3158 push(_gvn.transform(phi));
3162 3159
3163 3160 return true;
3164 3161 }
3165 3162
3166 3163 //---------------------generate_array_guard_common------------------------
3167 3164 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3168 3165 bool obj_array, bool not_array) {
3169 3166 // If obj_array/non_array==false/false:
3170 3167 // Branch around if the given klass is in fact an array (either obj or prim).
3171 3168 // If obj_array/non_array==false/true:
3172 3169 // Branch around if the given klass is not an array klass of any kind.
3173 3170 // If obj_array/non_array==true/true:
3174 3171 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3175 3172 // If obj_array/non_array==true/false:
3176 3173 // Branch around if the kls is an oop array (Object[] or subtype)
3177 3174 //
3178 3175 // Like generate_guard, adds a new path onto the region.
3179 3176 jint layout_con = 0;
3180 3177 Node* layout_val = get_layout_helper(kls, layout_con);
3181 3178 if (layout_val == NULL) {
3182 3179 bool query = (obj_array
3183 3180 ? Klass::layout_helper_is_objArray(layout_con)
3184 3181 : Klass::layout_helper_is_javaArray(layout_con));
3185 3182 if (query == not_array) {
3186 3183 return NULL; // never a branch
3187 3184 } else { // always a branch
3188 3185 Node* always_branch = control();
3189 3186 if (region != NULL)
3190 3187 region->add_req(always_branch);
3191 3188 set_control(top());
3192 3189 return always_branch;
3193 3190 }
3194 3191 }
3195 3192 // Now test the correct condition.
3196 3193 jint nval = (obj_array
3197 3194 ? ((jint)Klass::_lh_array_tag_type_value
3198 3195 << Klass::_lh_array_tag_shift)
3199 3196 : Klass::_lh_neutral_value);
3200 3197 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3201 3198 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3202 3199 // invert the test if we are looking for a non-array
3203 3200 if (not_array) btest = BoolTest(btest).negate();
3204 3201 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3205 3202 return generate_fair_guard(bol, region);
3206 3203 }
3207 3204
3208 3205
3209 3206 //-----------------------inline_native_newArray--------------------------
3210 3207 bool LibraryCallKit::inline_native_newArray() {
3211 3208 int nargs = 2;
3212 3209 Node* mirror = argument(0);
3213 3210 Node* count_val = argument(1);
3214 3211
3215 3212 _sp += nargs; // set original stack for use by uncommon_trap
3216 3213 mirror = do_null_check(mirror, T_OBJECT);
3217 3214 _sp -= nargs;
3218 3215 // If mirror or obj is dead, only null-path is taken.
3219 3216 if (stopped()) return true;
3220 3217
3221 3218 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3222 3219 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3223 3220 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3224 3221 TypeInstPtr::NOTNULL);
3225 3222 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3226 3223 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3227 3224 TypePtr::BOTTOM);
3228 3225
3229 3226 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3230 3227 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3231 3228 nargs,
3232 3229 result_reg, _slow_path);
3233 3230 Node* normal_ctl = control();
3234 3231 Node* no_array_ctl = result_reg->in(_slow_path);
3235 3232
3236 3233 // Generate code for the slow case. We make a call to newArray().
3237 3234 set_control(no_array_ctl);
3238 3235 if (!stopped()) {
3239 3236 // Either the input type is void.class, or else the
3240 3237 // array klass has not yet been cached. Either the
3241 3238 // ensuing call will throw an exception, or else it
3242 3239 // will cache the array klass for next time.
3243 3240 PreserveJVMState pjvms(this);
3244 3241 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3245 3242 Node* slow_result = set_results_for_java_call(slow_call);
3246 3243 // this->control() comes from set_results_for_java_call
3247 3244 result_reg->set_req(_slow_path, control());
3248 3245 result_val->set_req(_slow_path, slow_result);
3249 3246 result_io ->set_req(_slow_path, i_o());
3250 3247 result_mem->set_req(_slow_path, reset_memory());
3251 3248 }
3252 3249
3253 3250 set_control(normal_ctl);
3254 3251 if (!stopped()) {
3255 3252 // Normal case: The array type has been cached in the java.lang.Class.
3256 3253 // The following call works fine even if the array type is polymorphic.
3257 3254 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3258 3255 Node* obj = new_array(klass_node, count_val, nargs);
3259 3256 result_reg->init_req(_normal_path, control());
3260 3257 result_val->init_req(_normal_path, obj);
3261 3258 result_io ->init_req(_normal_path, i_o());
3262 3259 result_mem->init_req(_normal_path, reset_memory());
3263 3260 }
3264 3261
3265 3262 // Return the combined state.
3266 3263 set_i_o( _gvn.transform(result_io) );
3267 3264 set_all_memory( _gvn.transform(result_mem) );
3268 3265 push_result(result_reg, result_val);
3269 3266 C->set_has_split_ifs(true); // Has chance for split-if optimization
3270 3267
3271 3268 return true;
3272 3269 }
3273 3270
3274 3271 //----------------------inline_native_getLength--------------------------
3275 3272 bool LibraryCallKit::inline_native_getLength() {
3276 3273 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3277 3274
3278 3275 int nargs = 1;
3279 3276 Node* array = argument(0);
3280 3277
3281 3278 _sp += nargs; // set original stack for use by uncommon_trap
3282 3279 array = do_null_check(array, T_OBJECT);
3283 3280 _sp -= nargs;
3284 3281
3285 3282 // If array is dead, only null-path is taken.
3286 3283 if (stopped()) return true;
3287 3284
3288 3285 // Deoptimize if it is a non-array.
3289 3286 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3290 3287
3291 3288 if (non_array != NULL) {
3292 3289 PreserveJVMState pjvms(this);
3293 3290 set_control(non_array);
3294 3291 _sp += nargs; // push the arguments back on the stack
3295 3292 uncommon_trap(Deoptimization::Reason_intrinsic,
3296 3293 Deoptimization::Action_maybe_recompile);
3297 3294 }
3298 3295
3299 3296 // If control is dead, only non-array-path is taken.
3300 3297 if (stopped()) return true;
3301 3298
3302 3299 // The works fine even if the array type is polymorphic.
3303 3300 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3304 3301 push( load_array_length(array) );
3305 3302
3306 3303 C->set_has_split_ifs(true); // Has chance for split-if optimization
3307 3304
3308 3305 return true;
3309 3306 }
3310 3307
3311 3308 //------------------------inline_array_copyOf----------------------------
3312 3309 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3313 3310 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3314 3311
3315 3312 // Restore the stack and pop off the arguments.
3316 3313 int nargs = 3 + (is_copyOfRange? 1: 0);
3317 3314 Node* original = argument(0);
3318 3315 Node* start = is_copyOfRange? argument(1): intcon(0);
3319 3316 Node* end = is_copyOfRange? argument(2): argument(1);
3320 3317 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3321 3318
3322 3319 Node* newcopy;
3323 3320
3324 3321 //set the original stack and the reexecute bit for the interpreter to reexecute
3325 3322 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3326 3323 { PreserveReexecuteState preexecs(this);
3327 3324 _sp += nargs;
3328 3325 jvms()->set_should_reexecute(true);
3329 3326
3330 3327 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3331 3328 original = do_null_check(original, T_OBJECT);
3332 3329
3333 3330 // Check if a null path was taken unconditionally.
3334 3331 if (stopped()) return true;
3335 3332
3336 3333 Node* orig_length = load_array_length(original);
3337 3334
3338 3335 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3339 3336 NULL, 0);
3340 3337 klass_node = do_null_check(klass_node, T_OBJECT);
3341 3338
3342 3339 RegionNode* bailout = new (C, 1) RegionNode(1);
3343 3340 record_for_igvn(bailout);
3344 3341
3345 3342 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3346 3343 // Bail out if that is so.
3347 3344 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3348 3345 if (not_objArray != NULL) {
3349 3346 // Improve the klass node's type from the new optimistic assumption:
3350 3347 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3351 3348 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3352 3349 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3353 3350 cast->init_req(0, control());
3354 3351 klass_node = _gvn.transform(cast);
3355 3352 }
3356 3353
3357 3354 // Bail out if either start or end is negative.
3358 3355 generate_negative_guard(start, bailout, &start);
3359 3356 generate_negative_guard(end, bailout, &end);
3360 3357
3361 3358 Node* length = end;
3362 3359 if (_gvn.type(start) != TypeInt::ZERO) {
3363 3360 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3364 3361 }
3365 3362
3366 3363 // Bail out if length is negative.
3367 3364 // ...Not needed, since the new_array will throw the right exception.
3368 3365 //generate_negative_guard(length, bailout, &length);
3369 3366
3370 3367 if (bailout->req() > 1) {
3371 3368 PreserveJVMState pjvms(this);
3372 3369 set_control( _gvn.transform(bailout) );
3373 3370 uncommon_trap(Deoptimization::Reason_intrinsic,
3374 3371 Deoptimization::Action_maybe_recompile);
3375 3372 }
3376 3373
3377 3374 if (!stopped()) {
3378 3375
3379 3376 // How many elements will we copy from the original?
3380 3377 // The answer is MinI(orig_length - start, length).
3381 3378 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3382 3379 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3383 3380
3384 3381 const bool raw_mem_only = true;
3385 3382 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3386 3383
3387 3384 // Generate a direct call to the right arraycopy function(s).
3388 3385 // We know the copy is disjoint but we might not know if the
3389 3386 // oop stores need checking.
3390 3387 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3391 3388 // This will fail a store-check if x contains any non-nulls.
3392 3389 bool disjoint_bases = true;
3393 3390 bool length_never_negative = true;
3394 3391 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3395 3392 original, start, newcopy, intcon(0), moved,
3396 3393 disjoint_bases, length_never_negative);
3397 3394 }
3398 3395 } //original reexecute and sp are set back here
3399 3396
3400 3397 if(!stopped()) {
3401 3398 push(newcopy);
3402 3399 }
3403 3400
3404 3401 C->set_has_split_ifs(true); // Has chance for split-if optimization
3405 3402
3406 3403 return true;
3407 3404 }
3408 3405
3409 3406
3410 3407 //----------------------generate_virtual_guard---------------------------
3411 3408 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3412 3409 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3413 3410 RegionNode* slow_region) {
3414 3411 ciMethod* method = callee();
3415 3412 int vtable_index = method->vtable_index();
3416 3413 // Get the methodOop out of the appropriate vtable entry.
3417 3414 int entry_offset = (instanceKlass::vtable_start_offset() +
3418 3415 vtable_index*vtableEntry::size()) * wordSize +
3419 3416 vtableEntry::method_offset_in_bytes();
3420 3417 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3421 3418 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3422 3419
3423 3420 // Compare the target method with the expected method (e.g., Object.hashCode).
3424 3421 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3425 3422
3426 3423 Node* native_call = makecon(native_call_addr);
3427 3424 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3428 3425 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3429 3426
3430 3427 return generate_slow_guard(test_native, slow_region);
3431 3428 }
3432 3429
3433 3430 //-----------------------generate_method_call----------------------------
3434 3431 // Use generate_method_call to make a slow-call to the real
3435 3432 // method if the fast path fails. An alternative would be to
3436 3433 // use a stub like OptoRuntime::slow_arraycopy_Java.
3437 3434 // This only works for expanding the current library call,
3438 3435 // not another intrinsic. (E.g., don't use this for making an
3439 3436 // arraycopy call inside of the copyOf intrinsic.)
3440 3437 CallJavaNode*
3441 3438 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3442 3439 // When compiling the intrinsic method itself, do not use this technique.
3443 3440 guarantee(callee() != C->method(), "cannot make slow-call to self");
3444 3441
3445 3442 ciMethod* method = callee();
3446 3443 // ensure the JVMS we have will be correct for this call
3447 3444 guarantee(method_id == method->intrinsic_id(), "must match");
3448 3445
3449 3446 const TypeFunc* tf = TypeFunc::make(method);
3450 3447 int tfdc = tf->domain()->cnt();
3451 3448 CallJavaNode* slow_call;
3452 3449 if (is_static) {
3453 3450 assert(!is_virtual, "");
3454 3451 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3455 3452 SharedRuntime::get_resolve_static_call_stub(),
3456 3453 method, bci());
3457 3454 } else if (is_virtual) {
3458 3455 null_check_receiver(method);
3459 3456 int vtable_index = methodOopDesc::invalid_vtable_index;
3460 3457 if (UseInlineCaches) {
3461 3458 // Suppress the vtable call
3462 3459 } else {
3463 3460 // hashCode and clone are not a miranda methods,
3464 3461 // so the vtable index is fixed.
3465 3462 // No need to use the linkResolver to get it.
3466 3463 vtable_index = method->vtable_index();
3467 3464 }
3468 3465 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3469 3466 SharedRuntime::get_resolve_virtual_call_stub(),
3470 3467 method, vtable_index, bci());
3471 3468 } else { // neither virtual nor static: opt_virtual
3472 3469 null_check_receiver(method);
3473 3470 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3474 3471 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3475 3472 method, bci());
3476 3473 slow_call->set_optimized_virtual(true);
3477 3474 }
3478 3475 set_arguments_for_java_call(slow_call);
3479 3476 set_edges_for_java_call(slow_call);
3480 3477 return slow_call;
3481 3478 }
3482 3479
3483 3480
3484 3481 //------------------------------inline_native_hashcode--------------------
3485 3482 // Build special case code for calls to hashCode on an object.
3486 3483 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3487 3484 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3488 3485 assert(!(is_virtual && is_static), "either virtual, special, or static");
3489 3486
3490 3487 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3491 3488
3492 3489 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3493 3490 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3494 3491 TypeInt::INT);
3495 3492 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3496 3493 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3497 3494 TypePtr::BOTTOM);
3498 3495 Node* obj = NULL;
3499 3496 if (!is_static) {
3500 3497 // Check for hashing null object
3501 3498 obj = null_check_receiver(callee());
3502 3499 if (stopped()) return true; // unconditionally null
3503 3500 result_reg->init_req(_null_path, top());
3504 3501 result_val->init_req(_null_path, top());
3505 3502 } else {
3506 3503 // Do a null check, and return zero if null.
3507 3504 // System.identityHashCode(null) == 0
3508 3505 obj = argument(0);
3509 3506 Node* null_ctl = top();
3510 3507 obj = null_check_oop(obj, &null_ctl);
3511 3508 result_reg->init_req(_null_path, null_ctl);
3512 3509 result_val->init_req(_null_path, _gvn.intcon(0));
3513 3510 }
3514 3511
3515 3512 // Unconditionally null? Then return right away.
3516 3513 if (stopped()) {
3517 3514 set_control( result_reg->in(_null_path) );
3518 3515 if (!stopped())
3519 3516 push( result_val ->in(_null_path) );
3520 3517 return true;
3521 3518 }
3522 3519
3523 3520 // After null check, get the object's klass.
3524 3521 Node* obj_klass = load_object_klass(obj);
3525 3522
3526 3523 // This call may be virtual (invokevirtual) or bound (invokespecial).
3527 3524 // For each case we generate slightly different code.
3528 3525
3529 3526 // We only go to the fast case code if we pass a number of guards. The
3530 3527 // paths which do not pass are accumulated in the slow_region.
3531 3528 RegionNode* slow_region = new (C, 1) RegionNode(1);
3532 3529 record_for_igvn(slow_region);
3533 3530
3534 3531 // If this is a virtual call, we generate a funny guard. We pull out
3535 3532 // the vtable entry corresponding to hashCode() from the target object.
3536 3533 // If the target method which we are calling happens to be the native
3537 3534 // Object hashCode() method, we pass the guard. We do not need this
3538 3535 // guard for non-virtual calls -- the caller is known to be the native
3539 3536 // Object hashCode().
3540 3537 if (is_virtual) {
3541 3538 generate_virtual_guard(obj_klass, slow_region);
3542 3539 }
3543 3540
3544 3541 // Get the header out of the object, use LoadMarkNode when available
3545 3542 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3546 3543 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3547 3544
3548 3545 // Test the header to see if it is unlocked.
3549 3546 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3550 3547 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3551 3548 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3552 3549 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3553 3550 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3554 3551
3555 3552 generate_slow_guard(test_unlocked, slow_region);
3556 3553
3557 3554 // Get the hash value and check to see that it has been properly assigned.
3558 3555 // We depend on hash_mask being at most 32 bits and avoid the use of
3559 3556 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3560 3557 // vm: see markOop.hpp.
3561 3558 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3562 3559 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3563 3560 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3564 3561 // This hack lets the hash bits live anywhere in the mark object now, as long
3565 3562 // as the shift drops the relevant bits into the low 32 bits. Note that
3566 3563 // Java spec says that HashCode is an int so there's no point in capturing
3567 3564 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3568 3565 hshifted_header = ConvX2I(hshifted_header);
3569 3566 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3570 3567
3571 3568 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3572 3569 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3573 3570 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3574 3571
3575 3572 generate_slow_guard(test_assigned, slow_region);
3576 3573
3577 3574 Node* init_mem = reset_memory();
3578 3575 // fill in the rest of the null path:
3579 3576 result_io ->init_req(_null_path, i_o());
3580 3577 result_mem->init_req(_null_path, init_mem);
3581 3578
3582 3579 result_val->init_req(_fast_path, hash_val);
3583 3580 result_reg->init_req(_fast_path, control());
3584 3581 result_io ->init_req(_fast_path, i_o());
3585 3582 result_mem->init_req(_fast_path, init_mem);
3586 3583
3587 3584 // Generate code for the slow case. We make a call to hashCode().
3588 3585 set_control(_gvn.transform(slow_region));
3589 3586 if (!stopped()) {
3590 3587 // No need for PreserveJVMState, because we're using up the present state.
3591 3588 set_all_memory(init_mem);
3592 3589 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3593 3590 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3594 3591 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3595 3592 Node* slow_result = set_results_for_java_call(slow_call);
3596 3593 // this->control() comes from set_results_for_java_call
3597 3594 result_reg->init_req(_slow_path, control());
3598 3595 result_val->init_req(_slow_path, slow_result);
3599 3596 result_io ->set_req(_slow_path, i_o());
3600 3597 result_mem ->set_req(_slow_path, reset_memory());
3601 3598 }
3602 3599
3603 3600 // Return the combined state.
3604 3601 set_i_o( _gvn.transform(result_io) );
3605 3602 set_all_memory( _gvn.transform(result_mem) );
3606 3603 push_result(result_reg, result_val);
3607 3604
3608 3605 return true;
3609 3606 }
3610 3607
3611 3608 //---------------------------inline_native_getClass----------------------------
3612 3609 // Build special case code for calls to getClass on an object.
3613 3610 bool LibraryCallKit::inline_native_getClass() {
3614 3611 Node* obj = null_check_receiver(callee());
3615 3612 if (stopped()) return true;
3616 3613 push( load_mirror_from_klass(load_object_klass(obj)) );
3617 3614 return true;
3618 3615 }
3619 3616
3620 3617 //-----------------inline_native_Reflection_getCallerClass---------------------
3621 3618 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3622 3619 //
3623 3620 // NOTE that this code must perform the same logic as
3624 3621 // vframeStream::security_get_caller_frame in that it must skip
3625 3622 // Method.invoke() and auxiliary frames.
3626 3623
3627 3624
3628 3625
3629 3626
3630 3627 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3631 3628 ciMethod* method = callee();
3632 3629
3633 3630 #ifndef PRODUCT
3634 3631 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3635 3632 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3636 3633 }
3637 3634 #endif
3638 3635
3639 3636 debug_only(int saved_sp = _sp);
3640 3637
3641 3638 // Argument words: (int depth)
3642 3639 int nargs = 1;
3643 3640
3644 3641 _sp += nargs;
3645 3642 Node* caller_depth_node = pop();
3646 3643
3647 3644 assert(saved_sp == _sp, "must have correct argument count");
3648 3645
3649 3646 // The depth value must be a constant in order for the runtime call
3650 3647 // to be eliminated.
3651 3648 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3652 3649 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3653 3650 #ifndef PRODUCT
3654 3651 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3655 3652 tty->print_cr(" Bailing out because caller depth was not a constant");
3656 3653 }
3657 3654 #endif
3658 3655 return false;
3659 3656 }
3660 3657 // Note that the JVM state at this point does not include the
3661 3658 // getCallerClass() frame which we are trying to inline. The
3662 3659 // semantics of getCallerClass(), however, are that the "first"
3663 3660 // frame is the getCallerClass() frame, so we subtract one from the
3664 3661 // requested depth before continuing. We don't inline requests of
3665 3662 // getCallerClass(0).
3666 3663 int caller_depth = caller_depth_type->get_con() - 1;
3667 3664 if (caller_depth < 0) {
3668 3665 #ifndef PRODUCT
3669 3666 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3670 3667 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3671 3668 }
3672 3669 #endif
3673 3670 return false;
3674 3671 }
3675 3672
3676 3673 if (!jvms()->has_method()) {
3677 3674 #ifndef PRODUCT
3678 3675 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3679 3676 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3680 3677 }
3681 3678 #endif
3682 3679 return false;
3683 3680 }
3684 3681 int _depth = jvms()->depth(); // cache call chain depth
3685 3682
3686 3683 // Walk back up the JVM state to find the caller at the required
3687 3684 // depth. NOTE that this code must perform the same logic as
3688 3685 // vframeStream::security_get_caller_frame in that it must skip
3689 3686 // Method.invoke() and auxiliary frames. Note also that depth is
3690 3687 // 1-based (1 is the bottom of the inlining).
3691 3688 int inlining_depth = _depth;
3692 3689 JVMState* caller_jvms = NULL;
3693 3690
3694 3691 if (inlining_depth > 0) {
3695 3692 caller_jvms = jvms();
3696 3693 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3697 3694 do {
3698 3695 // The following if-tests should be performed in this order
3699 3696 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3700 3697 // Skip a Method.invoke() or auxiliary frame
3701 3698 } else if (caller_depth > 0) {
3702 3699 // Skip real frame
3703 3700 --caller_depth;
3704 3701 } else {
3705 3702 // We're done: reached desired caller after skipping.
3706 3703 break;
3707 3704 }
3708 3705 caller_jvms = caller_jvms->caller();
3709 3706 --inlining_depth;
3710 3707 } while (inlining_depth > 0);
3711 3708 }
3712 3709
3713 3710 if (inlining_depth == 0) {
3714 3711 #ifndef PRODUCT
3715 3712 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3716 3713 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3717 3714 tty->print_cr(" JVM state at this point:");
3718 3715 for (int i = _depth; i >= 1; i--) {
3719 3716 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3720 3717 }
3721 3718 }
3722 3719 #endif
3723 3720 return false; // Reached end of inlining
3724 3721 }
3725 3722
3726 3723 // Acquire method holder as java.lang.Class
3727 3724 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3728 3725 ciInstance* caller_mirror = caller_klass->java_mirror();
3729 3726 // Push this as a constant
3730 3727 push(makecon(TypeInstPtr::make(caller_mirror)));
3731 3728 #ifndef PRODUCT
3732 3729 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3733 3730 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);
3734 3731 tty->print_cr(" JVM state at this point:");
3735 3732 for (int i = _depth; i >= 1; i--) {
3736 3733 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3737 3734 }
3738 3735 }
3739 3736 #endif
3740 3737 return true;
3741 3738 }
3742 3739
3743 3740 // Helper routine for above
3744 3741 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3745 3742 ciMethod* method = jvms->method();
3746 3743
3747 3744 // Is this the Method.invoke method itself?
3748 3745 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3749 3746 return true;
3750 3747
3751 3748 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3752 3749 ciKlass* k = method->holder();
3753 3750 if (k->is_instance_klass()) {
3754 3751 ciInstanceKlass* ik = k->as_instance_klass();
3755 3752 for (; ik != NULL; ik = ik->super()) {
3756 3753 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3757 3754 ik == env()->find_system_klass(ik->name())) {
3758 3755 return true;
3759 3756 }
3760 3757 }
3761 3758 }
3762 3759 else if (method->is_method_handle_adapter()) {
3763 3760 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3764 3761 return true;
3765 3762 }
3766 3763
3767 3764 return false;
3768 3765 }
3769 3766
3770 3767 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3771 3768 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3772 3769 // computing it since there is no lookup field by name function in the
3773 3770 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3774 3771 // Using a static variable here is safe even if we have multiple compilation
3775 3772 // threads because the offset is constant. At worst the same offset will be
3776 3773 // computed and stored multiple
3777 3774
3778 3775 bool LibraryCallKit::inline_native_AtomicLong_get() {
3779 3776 // Restore the stack and pop off the argument
3780 3777 _sp+=1;
3781 3778 Node *obj = pop();
3782 3779
3783 3780 // get the offset of the "value" field. Since the CI interfaces
3784 3781 // does not provide a way to look up a field by name, we scan the bytecodes
3785 3782 // to get the field index. We expect the first 2 instructions of the method
3786 3783 // to be:
3787 3784 // 0 aload_0
3788 3785 // 1 getfield "value"
3789 3786 ciMethod* method = callee();
3790 3787 if (value_field_offset == -1)
3791 3788 {
3792 3789 ciField* value_field;
3793 3790 ciBytecodeStream iter(method);
3794 3791 Bytecodes::Code bc = iter.next();
3795 3792
3796 3793 if ((bc != Bytecodes::_aload_0) &&
3797 3794 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3798 3795 return false;
3799 3796 bc = iter.next();
3800 3797 if (bc != Bytecodes::_getfield)
3801 3798 return false;
3802 3799 bool ignore;
3803 3800 value_field = iter.get_field(ignore);
3804 3801 value_field_offset = value_field->offset_in_bytes();
3805 3802 }
3806 3803
3807 3804 // Null check without removing any arguments.
3808 3805 _sp++;
3809 3806 obj = do_null_check(obj, T_OBJECT);
3810 3807 _sp--;
3811 3808 // Check for locking null object
3812 3809 if (stopped()) return true;
3813 3810
3814 3811 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3815 3812 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3816 3813 int alias_idx = C->get_alias_index(adr_type);
3817 3814
3818 3815 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3819 3816
3820 3817 push_pair(result);
3821 3818
3822 3819 return true;
3823 3820 }
3824 3821
3825 3822 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3826 3823 // Restore the stack and pop off the arguments
3827 3824 _sp+=5;
3828 3825 Node *newVal = pop_pair();
3829 3826 Node *oldVal = pop_pair();
3830 3827 Node *obj = pop();
3831 3828
3832 3829 // we need the offset of the "value" field which was computed when
3833 3830 // inlining the get() method. Give up if we don't have it.
3834 3831 if (value_field_offset == -1)
3835 3832 return false;
3836 3833
3837 3834 // Null check without removing any arguments.
3838 3835 _sp+=5;
3839 3836 obj = do_null_check(obj, T_OBJECT);
3840 3837 _sp-=5;
3841 3838 // Check for locking null object
3842 3839 if (stopped()) return true;
3843 3840
3844 3841 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3845 3842 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3846 3843 int alias_idx = C->get_alias_index(adr_type);
3847 3844
3848 3845 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3849 3846 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3850 3847 set_memory(store_proj, alias_idx);
3851 3848 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3852 3849
3853 3850 Node *result;
3854 3851 // CMove node is not used to be able fold a possible check code
3855 3852 // after attemptUpdate() call. This code could be transformed
3856 3853 // into CMove node by loop optimizations.
3857 3854 {
3858 3855 RegionNode *r = new (C, 3) RegionNode(3);
3859 3856 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3860 3857
3861 3858 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3862 3859 Node *iftrue = opt_iff(r, iff);
3863 3860 r->init_req(1, iftrue);
3864 3861 result->init_req(1, intcon(1));
3865 3862 result->init_req(2, intcon(0));
3866 3863
3867 3864 set_control(_gvn.transform(r));
3868 3865 record_for_igvn(r);
3869 3866
3870 3867 C->set_has_split_ifs(true); // Has chance for split-if optimization
3871 3868 }
3872 3869
3873 3870 push(_gvn.transform(result));
3874 3871 return true;
3875 3872 }
3876 3873
3877 3874 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3878 3875 // restore the arguments
3879 3876 _sp += arg_size();
3880 3877
3881 3878 switch (id) {
3882 3879 case vmIntrinsics::_floatToRawIntBits:
3883 3880 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3884 3881 break;
3885 3882
3886 3883 case vmIntrinsics::_intBitsToFloat:
3887 3884 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3888 3885 break;
3889 3886
3890 3887 case vmIntrinsics::_doubleToRawLongBits:
3891 3888 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3892 3889 break;
3893 3890
3894 3891 case vmIntrinsics::_longBitsToDouble:
3895 3892 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3896 3893 break;
3897 3894
3898 3895 case vmIntrinsics::_doubleToLongBits: {
3899 3896 Node* value = pop_pair();
3900 3897
3901 3898 // two paths (plus control) merge in a wood
3902 3899 RegionNode *r = new (C, 3) RegionNode(3);
3903 3900 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3904 3901
3905 3902 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3906 3903 // Build the boolean node
3907 3904 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3908 3905
3909 3906 // Branch either way.
3910 3907 // NaN case is less traveled, which makes all the difference.
3911 3908 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3912 3909 Node *opt_isnan = _gvn.transform(ifisnan);
3913 3910 assert( opt_isnan->is_If(), "Expect an IfNode");
3914 3911 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3915 3912 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3916 3913
3917 3914 set_control(iftrue);
3918 3915
3919 3916 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3920 3917 Node *slow_result = longcon(nan_bits); // return NaN
3921 3918 phi->init_req(1, _gvn.transform( slow_result ));
3922 3919 r->init_req(1, iftrue);
3923 3920
3924 3921 // Else fall through
3925 3922 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3926 3923 set_control(iffalse);
3927 3924
3928 3925 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3929 3926 r->init_req(2, iffalse);
3930 3927
3931 3928 // Post merge
3932 3929 set_control(_gvn.transform(r));
3933 3930 record_for_igvn(r);
3934 3931
3935 3932 Node* result = _gvn.transform(phi);
3936 3933 assert(result->bottom_type()->isa_long(), "must be");
3937 3934 push_pair(result);
3938 3935
3939 3936 C->set_has_split_ifs(true); // Has chance for split-if optimization
3940 3937
3941 3938 break;
3942 3939 }
3943 3940
3944 3941 case vmIntrinsics::_floatToIntBits: {
3945 3942 Node* value = pop();
3946 3943
3947 3944 // two paths (plus control) merge in a wood
3948 3945 RegionNode *r = new (C, 3) RegionNode(3);
3949 3946 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3950 3947
3951 3948 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3952 3949 // Build the boolean node
3953 3950 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3954 3951
3955 3952 // Branch either way.
3956 3953 // NaN case is less traveled, which makes all the difference.
3957 3954 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3958 3955 Node *opt_isnan = _gvn.transform(ifisnan);
3959 3956 assert( opt_isnan->is_If(), "Expect an IfNode");
3960 3957 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3961 3958 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3962 3959
3963 3960 set_control(iftrue);
3964 3961
3965 3962 static const jint nan_bits = 0x7fc00000;
3966 3963 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3967 3964 phi->init_req(1, _gvn.transform( slow_result ));
3968 3965 r->init_req(1, iftrue);
3969 3966
3970 3967 // Else fall through
3971 3968 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3972 3969 set_control(iffalse);
3973 3970
3974 3971 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3975 3972 r->init_req(2, iffalse);
3976 3973
3977 3974 // Post merge
3978 3975 set_control(_gvn.transform(r));
3979 3976 record_for_igvn(r);
3980 3977
3981 3978 Node* result = _gvn.transform(phi);
3982 3979 assert(result->bottom_type()->isa_int(), "must be");
3983 3980 push(result);
3984 3981
3985 3982 C->set_has_split_ifs(true); // Has chance for split-if optimization
3986 3983
3987 3984 break;
3988 3985 }
3989 3986
3990 3987 default:
3991 3988 ShouldNotReachHere();
3992 3989 }
3993 3990
3994 3991 return true;
3995 3992 }
3996 3993
3997 3994 #ifdef _LP64
3998 3995 #define XTOP ,top() /*additional argument*/
3999 3996 #else //_LP64
4000 3997 #define XTOP /*no additional argument*/
4001 3998 #endif //_LP64
4002 3999
4003 4000 //----------------------inline_unsafe_copyMemory-------------------------
4004 4001 bool LibraryCallKit::inline_unsafe_copyMemory() {
4005 4002 if (callee()->is_static()) return false; // caller must have the capability!
4006 4003 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4007 4004 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4008 4005 null_check_receiver(callee()); // check then ignore argument(0)
4009 4006 if (stopped()) return true;
4010 4007
4011 4008 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4012 4009
4013 4010 Node* src_ptr = argument(1);
4014 4011 Node* src_off = ConvL2X(argument(2));
4015 4012 assert(argument(3)->is_top(), "2nd half of long");
4016 4013 Node* dst_ptr = argument(4);
4017 4014 Node* dst_off = ConvL2X(argument(5));
4018 4015 assert(argument(6)->is_top(), "2nd half of long");
4019 4016 Node* size = ConvL2X(argument(7));
4020 4017 assert(argument(8)->is_top(), "2nd half of long");
4021 4018
4022 4019 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4023 4020 "fieldOffset must be byte-scaled");
4024 4021
4025 4022 Node* src = make_unsafe_address(src_ptr, src_off);
4026 4023 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4027 4024
4028 4025 // Conservatively insert a memory barrier on all memory slices.
4029 4026 // Do not let writes of the copy source or destination float below the copy.
4030 4027 insert_mem_bar(Op_MemBarCPUOrder);
4031 4028
4032 4029 // Call it. Note that the length argument is not scaled.
4033 4030 make_runtime_call(RC_LEAF|RC_NO_FP,
4034 4031 OptoRuntime::fast_arraycopy_Type(),
4035 4032 StubRoutines::unsafe_arraycopy(),
4036 4033 "unsafe_arraycopy",
4037 4034 TypeRawPtr::BOTTOM,
4038 4035 src, dst, size XTOP);
4039 4036
4040 4037 // Do not let reads of the copy destination float above the copy.
4041 4038 insert_mem_bar(Op_MemBarCPUOrder);
4042 4039
4043 4040 return true;
4044 4041 }
4045 4042
4046 4043 //------------------------clone_coping-----------------------------------
4047 4044 // Helper function for inline_native_clone.
4048 4045 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4049 4046 assert(obj_size != NULL, "");
4050 4047 Node* raw_obj = alloc_obj->in(1);
4051 4048 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4052 4049
4053 4050 if (ReduceBulkZeroing) {
4054 4051 // We will be completely responsible for initializing this object -
4055 4052 // mark Initialize node as complete.
4056 4053 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4057 4054 // The object was just allocated - there should be no any stores!
4058 4055 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4059 4056 }
4060 4057
4061 4058 // Copy the fastest available way.
4062 4059 // TODO: generate fields copies for small objects instead.
4063 4060 Node* src = obj;
4064 4061 Node* dest = alloc_obj;
4065 4062 Node* size = _gvn.transform(obj_size);
4066 4063
4067 4064 // Exclude the header but include array length to copy by 8 bytes words.
4068 4065 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4069 4066 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4070 4067 instanceOopDesc::base_offset_in_bytes();
4071 4068 // base_off:
4072 4069 // 8 - 32-bit VM
4073 4070 // 12 - 64-bit VM, compressed oops
4074 4071 // 16 - 64-bit VM, normal oops
4075 4072 if (base_off % BytesPerLong != 0) {
4076 4073 assert(UseCompressedOops, "");
4077 4074 if (is_array) {
4078 4075 // Exclude length to copy by 8 bytes words.
4079 4076 base_off += sizeof(int);
4080 4077 } else {
4081 4078 // Include klass to copy by 8 bytes words.
4082 4079 base_off = instanceOopDesc::klass_offset_in_bytes();
4083 4080 }
4084 4081 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4085 4082 }
4086 4083 src = basic_plus_adr(src, base_off);
4087 4084 dest = basic_plus_adr(dest, base_off);
4088 4085
4089 4086 // Compute the length also, if needed:
4090 4087 Node* countx = size;
4091 4088 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4092 4089 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4093 4090
4094 4091 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4095 4092 bool disjoint_bases = true;
4096 4093 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4097 4094 src, NULL, dest, NULL, countx,
4098 4095 /*dest_uninitialized*/true);
4099 4096
4100 4097 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4101 4098 if (card_mark) {
4102 4099 assert(!is_array, "");
4103 4100 // Put in store barrier for any and all oops we are sticking
4104 4101 // into this object. (We could avoid this if we could prove
4105 4102 // that the object type contains no oop fields at all.)
4106 4103 Node* no_particular_value = NULL;
4107 4104 Node* no_particular_field = NULL;
4108 4105 int raw_adr_idx = Compile::AliasIdxRaw;
4109 4106 post_barrier(control(),
4110 4107 memory(raw_adr_type),
4111 4108 alloc_obj,
4112 4109 no_particular_field,
4113 4110 raw_adr_idx,
4114 4111 no_particular_value,
4115 4112 T_OBJECT,
4116 4113 false);
4117 4114 }
4118 4115
4119 4116 // Do not let reads from the cloned object float above the arraycopy.
4120 4117 insert_mem_bar(Op_MemBarCPUOrder);
4121 4118 }
4122 4119
4123 4120 //------------------------inline_native_clone----------------------------
4124 4121 // Here are the simple edge cases:
4125 4122 // null receiver => normal trap
4126 4123 // virtual and clone was overridden => slow path to out-of-line clone
4127 4124 // not cloneable or finalizer => slow path to out-of-line Object.clone
4128 4125 //
4129 4126 // The general case has two steps, allocation and copying.
4130 4127 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4131 4128 //
4132 4129 // Copying also has two cases, oop arrays and everything else.
4133 4130 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4134 4131 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4135 4132 //
4136 4133 // These steps fold up nicely if and when the cloned object's klass
4137 4134 // can be sharply typed as an object array, a type array, or an instance.
4138 4135 //
4139 4136 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4140 4137 int nargs = 1;
4141 4138 PhiNode* result_val;
4142 4139
4143 4140 //set the original stack and the reexecute bit for the interpreter to reexecute
4144 4141 //the bytecode that invokes Object.clone if deoptimization happens
4145 4142 { PreserveReexecuteState preexecs(this);
4146 4143 jvms()->set_should_reexecute(true);
4147 4144
4148 4145 //null_check_receiver will adjust _sp (push and pop)
4149 4146 Node* obj = null_check_receiver(callee());
4150 4147 if (stopped()) return true;
4151 4148
4152 4149 _sp += nargs;
4153 4150
4154 4151 Node* obj_klass = load_object_klass(obj);
4155 4152 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4156 4153 const TypeOopPtr* toop = ((tklass != NULL)
4157 4154 ? tklass->as_instance_type()
4158 4155 : TypeInstPtr::NOTNULL);
4159 4156
4160 4157 // Conservatively insert a memory barrier on all memory slices.
4161 4158 // Do not let writes into the original float below the clone.
4162 4159 insert_mem_bar(Op_MemBarCPUOrder);
4163 4160
4164 4161 // paths into result_reg:
4165 4162 enum {
4166 4163 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4167 4164 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4168 4165 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4169 4166 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4170 4167 PATH_LIMIT
4171 4168 };
4172 4169 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4173 4170 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4174 4171 TypeInstPtr::NOTNULL);
4175 4172 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4176 4173 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4177 4174 TypePtr::BOTTOM);
4178 4175 record_for_igvn(result_reg);
4179 4176
4180 4177 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4181 4178 int raw_adr_idx = Compile::AliasIdxRaw;
4182 4179 const bool raw_mem_only = true;
4183 4180
4184 4181
4185 4182 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4186 4183 if (array_ctl != NULL) {
4187 4184 // It's an array.
4188 4185 PreserveJVMState pjvms(this);
4189 4186 set_control(array_ctl);
4190 4187 Node* obj_length = load_array_length(obj);
4191 4188 Node* obj_size = NULL;
4192 4189 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4193 4190 raw_mem_only, &obj_size);
4194 4191
4195 4192 if (!use_ReduceInitialCardMarks()) {
4196 4193 // If it is an oop array, it requires very special treatment,
4197 4194 // because card marking is required on each card of the array.
4198 4195 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4199 4196 if (is_obja != NULL) {
4200 4197 PreserveJVMState pjvms2(this);
4201 4198 set_control(is_obja);
4202 4199 // Generate a direct call to the right arraycopy function(s).
4203 4200 bool disjoint_bases = true;
4204 4201 bool length_never_negative = true;
4205 4202 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4206 4203 obj, intcon(0), alloc_obj, intcon(0),
4207 4204 obj_length,
4208 4205 disjoint_bases, length_never_negative);
4209 4206 result_reg->init_req(_objArray_path, control());
4210 4207 result_val->init_req(_objArray_path, alloc_obj);
4211 4208 result_i_o ->set_req(_objArray_path, i_o());
4212 4209 result_mem ->set_req(_objArray_path, reset_memory());
4213 4210 }
4214 4211 }
4215 4212 // Otherwise, there are no card marks to worry about.
4216 4213 // (We can dispense with card marks if we know the allocation
4217 4214 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4218 4215 // causes the non-eden paths to take compensating steps to
4219 4216 // simulate a fresh allocation, so that no further
4220 4217 // card marks are required in compiled code to initialize
4221 4218 // the object.)
4222 4219
4223 4220 if (!stopped()) {
4224 4221 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4225 4222
4226 4223 // Present the results of the copy.
4227 4224 result_reg->init_req(_array_path, control());
4228 4225 result_val->init_req(_array_path, alloc_obj);
4229 4226 result_i_o ->set_req(_array_path, i_o());
4230 4227 result_mem ->set_req(_array_path, reset_memory());
4231 4228 }
4232 4229 }
4233 4230
4234 4231 // We only go to the instance fast case code if we pass a number of guards.
4235 4232 // The paths which do not pass are accumulated in the slow_region.
4236 4233 RegionNode* slow_region = new (C, 1) RegionNode(1);
4237 4234 record_for_igvn(slow_region);
4238 4235 if (!stopped()) {
4239 4236 // It's an instance (we did array above). Make the slow-path tests.
4240 4237 // If this is a virtual call, we generate a funny guard. We grab
4241 4238 // the vtable entry corresponding to clone() from the target object.
4242 4239 // If the target method which we are calling happens to be the
4243 4240 // Object clone() method, we pass the guard. We do not need this
4244 4241 // guard for non-virtual calls; the caller is known to be the native
4245 4242 // Object clone().
4246 4243 if (is_virtual) {
4247 4244 generate_virtual_guard(obj_klass, slow_region);
4248 4245 }
4249 4246
4250 4247 // The object must be cloneable and must not have a finalizer.
4251 4248 // Both of these conditions may be checked in a single test.
4252 4249 // We could optimize the cloneable test further, but we don't care.
4253 4250 generate_access_flags_guard(obj_klass,
4254 4251 // Test both conditions:
4255 4252 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4256 4253 // Must be cloneable but not finalizer:
4257 4254 JVM_ACC_IS_CLONEABLE,
4258 4255 slow_region);
4259 4256 }
4260 4257
4261 4258 if (!stopped()) {
4262 4259 // It's an instance, and it passed the slow-path tests.
4263 4260 PreserveJVMState pjvms(this);
4264 4261 Node* obj_size = NULL;
4265 4262 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4266 4263
4267 4264 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4268 4265
4269 4266 // Present the results of the slow call.
4270 4267 result_reg->init_req(_instance_path, control());
4271 4268 result_val->init_req(_instance_path, alloc_obj);
4272 4269 result_i_o ->set_req(_instance_path, i_o());
4273 4270 result_mem ->set_req(_instance_path, reset_memory());
4274 4271 }
4275 4272
4276 4273 // Generate code for the slow case. We make a call to clone().
4277 4274 set_control(_gvn.transform(slow_region));
4278 4275 if (!stopped()) {
4279 4276 PreserveJVMState pjvms(this);
4280 4277 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4281 4278 Node* slow_result = set_results_for_java_call(slow_call);
4282 4279 // this->control() comes from set_results_for_java_call
4283 4280 result_reg->init_req(_slow_path, control());
4284 4281 result_val->init_req(_slow_path, slow_result);
4285 4282 result_i_o ->set_req(_slow_path, i_o());
4286 4283 result_mem ->set_req(_slow_path, reset_memory());
4287 4284 }
4288 4285
4289 4286 // Return the combined state.
4290 4287 set_control( _gvn.transform(result_reg) );
4291 4288 set_i_o( _gvn.transform(result_i_o) );
4292 4289 set_all_memory( _gvn.transform(result_mem) );
4293 4290 } //original reexecute and sp are set back here
4294 4291
4295 4292 push(_gvn.transform(result_val));
4296 4293
4297 4294 return true;
4298 4295 }
4299 4296
4300 4297
4301 4298 // constants for computing the copy function
4302 4299 enum {
4303 4300 COPYFUNC_UNALIGNED = 0,
4304 4301 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4305 4302 COPYFUNC_CONJOINT = 0,
4306 4303 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4307 4304 };
4308 4305
4309 4306 // Note: The condition "disjoint" applies also for overlapping copies
4310 4307 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4311 4308 static address
4312 4309 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name, bool dest_uninitialized) {
4313 4310 int selector =
4314 4311 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4315 4312 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4316 4313
4317 4314 #define RETURN_STUB(xxx_arraycopy) { \
4318 4315 name = #xxx_arraycopy; \
4319 4316 return StubRoutines::xxx_arraycopy(); }
4320 4317
4321 4318 #define RETURN_STUB_PARM(xxx_arraycopy, parm) { \
4322 4319 name = #xxx_arraycopy; \
4323 4320 return StubRoutines::xxx_arraycopy(parm); }
4324 4321
4325 4322 switch (t) {
4326 4323 case T_BYTE:
4327 4324 case T_BOOLEAN:
4328 4325 switch (selector) {
4329 4326 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4330 4327 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4331 4328 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4332 4329 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4333 4330 }
4334 4331 case T_CHAR:
4335 4332 case T_SHORT:
4336 4333 switch (selector) {
4337 4334 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4338 4335 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4339 4336 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4340 4337 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4341 4338 }
4342 4339 case T_INT:
4343 4340 case T_FLOAT:
4344 4341 switch (selector) {
4345 4342 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4346 4343 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4347 4344 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4348 4345 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4349 4346 }
4350 4347 case T_DOUBLE:
4351 4348 case T_LONG:
4352 4349 switch (selector) {
4353 4350 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4354 4351 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4355 4352 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4356 4353 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4357 4354 }
4358 4355 case T_ARRAY:
4359 4356 case T_OBJECT:
4360 4357 switch (selector) {
4361 4358 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_arraycopy, dest_uninitialized);
4362 4359 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_arraycopy, dest_uninitialized);
4363 4360 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_disjoint_arraycopy, dest_uninitialized);
4364 4361 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_disjoint_arraycopy, dest_uninitialized);
4365 4362 }
4366 4363 default:
4367 4364 ShouldNotReachHere();
4368 4365 return NULL;
4369 4366 }
4370 4367
4371 4368 #undef RETURN_STUB
4372 4369 #undef RETURN_STUB_PARM
4373 4370 }
4374 4371
4375 4372 //------------------------------basictype2arraycopy----------------------------
4376 4373 address LibraryCallKit::basictype2arraycopy(BasicType t,
4377 4374 Node* src_offset,
4378 4375 Node* dest_offset,
4379 4376 bool disjoint_bases,
4380 4377 const char* &name,
4381 4378 bool dest_uninitialized) {
4382 4379 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4383 4380 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4384 4381
4385 4382 bool aligned = false;
4386 4383 bool disjoint = disjoint_bases;
4387 4384
4388 4385 // if the offsets are the same, we can treat the memory regions as
4389 4386 // disjoint, because either the memory regions are in different arrays,
4390 4387 // or they are identical (which we can treat as disjoint.) We can also
4391 4388 // treat a copy with a destination index less that the source index
4392 4389 // as disjoint since a low->high copy will work correctly in this case.
4393 4390 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4394 4391 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4395 4392 // both indices are constants
4396 4393 int s_offs = src_offset_inttype->get_con();
4397 4394 int d_offs = dest_offset_inttype->get_con();
4398 4395 int element_size = type2aelembytes(t);
4399 4396 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4400 4397 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4401 4398 if (s_offs >= d_offs) disjoint = true;
4402 4399 } else if (src_offset == dest_offset && src_offset != NULL) {
4403 4400 // This can occur if the offsets are identical non-constants.
4404 4401 disjoint = true;
4405 4402 }
4406 4403
4407 4404 return select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4408 4405 }
4409 4406
4410 4407
4411 4408 //------------------------------inline_arraycopy-----------------------
4412 4409 bool LibraryCallKit::inline_arraycopy() {
4413 4410 // Restore the stack and pop off the arguments.
4414 4411 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4415 4412 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4416 4413
4417 4414 Node *src = argument(0);
4418 4415 Node *src_offset = argument(1);
4419 4416 Node *dest = argument(2);
4420 4417 Node *dest_offset = argument(3);
4421 4418 Node *length = argument(4);
4422 4419
4423 4420 // Compile time checks. If any of these checks cannot be verified at compile time,
4424 4421 // we do not make a fast path for this call. Instead, we let the call remain as it
4425 4422 // is. The checks we choose to mandate at compile time are:
4426 4423 //
4427 4424 // (1) src and dest are arrays.
4428 4425 const Type* src_type = src->Value(&_gvn);
4429 4426 const Type* dest_type = dest->Value(&_gvn);
4430 4427 const TypeAryPtr* top_src = src_type->isa_aryptr();
4431 4428 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4432 4429 if (top_src == NULL || top_src->klass() == NULL ||
4433 4430 top_dest == NULL || top_dest->klass() == NULL) {
4434 4431 // Conservatively insert a memory barrier on all memory slices.
4435 4432 // Do not let writes into the source float below the arraycopy.
4436 4433 insert_mem_bar(Op_MemBarCPUOrder);
4437 4434
4438 4435 // Call StubRoutines::generic_arraycopy stub.
4439 4436 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4440 4437 src, src_offset, dest, dest_offset, length);
4441 4438
4442 4439 // Do not let reads from the destination float above the arraycopy.
4443 4440 // Since we cannot type the arrays, we don't know which slices
4444 4441 // might be affected. We could restrict this barrier only to those
4445 4442 // memory slices which pertain to array elements--but don't bother.
4446 4443 if (!InsertMemBarAfterArraycopy)
4447 4444 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4448 4445 insert_mem_bar(Op_MemBarCPUOrder);
4449 4446 return true;
4450 4447 }
4451 4448
4452 4449 // (2) src and dest arrays must have elements of the same BasicType
4453 4450 // Figure out the size and type of the elements we will be copying.
4454 4451 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4455 4452 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4456 4453 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4457 4454 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4458 4455
4459 4456 if (src_elem != dest_elem || dest_elem == T_VOID) {
4460 4457 // The component types are not the same or are not recognized. Punt.
4461 4458 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4462 4459 generate_slow_arraycopy(TypePtr::BOTTOM,
4463 4460 src, src_offset, dest, dest_offset, length,
4464 4461 /*dest_uninitialized*/false);
4465 4462 return true;
4466 4463 }
4467 4464
4468 4465 //---------------------------------------------------------------------------
4469 4466 // We will make a fast path for this call to arraycopy.
4470 4467
4471 4468 // We have the following tests left to perform:
4472 4469 //
4473 4470 // (3) src and dest must not be null.
4474 4471 // (4) src_offset must not be negative.
4475 4472 // (5) dest_offset must not be negative.
4476 4473 // (6) length must not be negative.
4477 4474 // (7) src_offset + length must not exceed length of src.
4478 4475 // (8) dest_offset + length must not exceed length of dest.
4479 4476 // (9) each element of an oop array must be assignable
4480 4477
4481 4478 RegionNode* slow_region = new (C, 1) RegionNode(1);
4482 4479 record_for_igvn(slow_region);
4483 4480
4484 4481 // (3) operands must not be null
4485 4482 // We currently perform our null checks with the do_null_check routine.
4486 4483 // This means that the null exceptions will be reported in the caller
4487 4484 // rather than (correctly) reported inside of the native arraycopy call.
4488 4485 // This should be corrected, given time. We do our null check with the
4489 4486 // stack pointer restored.
4490 4487 _sp += nargs;
4491 4488 src = do_null_check(src, T_ARRAY);
4492 4489 dest = do_null_check(dest, T_ARRAY);
4493 4490 _sp -= nargs;
4494 4491
4495 4492 // (4) src_offset must not be negative.
4496 4493 generate_negative_guard(src_offset, slow_region);
4497 4494
4498 4495 // (5) dest_offset must not be negative.
4499 4496 generate_negative_guard(dest_offset, slow_region);
4500 4497
4501 4498 // (6) length must not be negative (moved to generate_arraycopy()).
4502 4499 // generate_negative_guard(length, slow_region);
4503 4500
4504 4501 // (7) src_offset + length must not exceed length of src.
4505 4502 generate_limit_guard(src_offset, length,
4506 4503 load_array_length(src),
4507 4504 slow_region);
4508 4505
4509 4506 // (8) dest_offset + length must not exceed length of dest.
4510 4507 generate_limit_guard(dest_offset, length,
4511 4508 load_array_length(dest),
4512 4509 slow_region);
4513 4510
4514 4511 // (9) each element of an oop array must be assignable
4515 4512 // The generate_arraycopy subroutine checks this.
4516 4513
4517 4514 // This is where the memory effects are placed:
4518 4515 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4519 4516 generate_arraycopy(adr_type, dest_elem,
4520 4517 src, src_offset, dest, dest_offset, length,
4521 4518 false, false, slow_region);
4522 4519
4523 4520 return true;
4524 4521 }
4525 4522
4526 4523 //-----------------------------generate_arraycopy----------------------
4527 4524 // Generate an optimized call to arraycopy.
4528 4525 // Caller must guard against non-arrays.
4529 4526 // Caller must determine a common array basic-type for both arrays.
4530 4527 // Caller must validate offsets against array bounds.
4531 4528 // The slow_region has already collected guard failure paths
4532 4529 // (such as out of bounds length or non-conformable array types).
4533 4530 // The generated code has this shape, in general:
4534 4531 //
4535 4532 // if (length == 0) return // via zero_path
4536 4533 // slowval = -1
4537 4534 // if (types unknown) {
4538 4535 // slowval = call generic copy loop
4539 4536 // if (slowval == 0) return // via checked_path
4540 4537 // } else if (indexes in bounds) {
4541 4538 // if ((is object array) && !(array type check)) {
4542 4539 // slowval = call checked copy loop
4543 4540 // if (slowval == 0) return // via checked_path
4544 4541 // } else {
4545 4542 // call bulk copy loop
4546 4543 // return // via fast_path
4547 4544 // }
4548 4545 // }
4549 4546 // // adjust params for remaining work:
4550 4547 // if (slowval != -1) {
4551 4548 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4552 4549 // }
4553 4550 // slow_region:
4554 4551 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4555 4552 // return // via slow_call_path
4556 4553 //
4557 4554 // This routine is used from several intrinsics: System.arraycopy,
4558 4555 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4559 4556 //
4560 4557 void
4561 4558 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4562 4559 BasicType basic_elem_type,
4563 4560 Node* src, Node* src_offset,
4564 4561 Node* dest, Node* dest_offset,
4565 4562 Node* copy_length,
4566 4563 bool disjoint_bases,
4567 4564 bool length_never_negative,
4568 4565 RegionNode* slow_region) {
4569 4566
4570 4567 if (slow_region == NULL) {
4571 4568 slow_region = new(C,1) RegionNode(1);
4572 4569 record_for_igvn(slow_region);
4573 4570 }
4574 4571
4575 4572 Node* original_dest = dest;
4576 4573 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4577 4574 bool dest_uninitialized = false;
4578 4575
4579 4576 // See if this is the initialization of a newly-allocated array.
4580 4577 // If so, we will take responsibility here for initializing it to zero.
4581 4578 // (Note: Because tightly_coupled_allocation performs checks on the
4582 4579 // out-edges of the dest, we need to avoid making derived pointers
4583 4580 // from it until we have checked its uses.)
4584 4581 if (ReduceBulkZeroing
4585 4582 && !ZeroTLAB // pointless if already zeroed
4586 4583 && basic_elem_type != T_CONFLICT // avoid corner case
4587 4584 && !_gvn.eqv_uncast(src, dest)
4588 4585 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4589 4586 != NULL)
4590 4587 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4591 4588 && alloc->maybe_set_complete(&_gvn)) {
4592 4589 // "You break it, you buy it."
4593 4590 InitializeNode* init = alloc->initialization();
4594 4591 assert(init->is_complete(), "we just did this");
4595 4592 assert(dest->is_CheckCastPP(), "sanity");
4596 4593 assert(dest->in(0)->in(0) == init, "dest pinned");
4597 4594 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4598 4595 // From this point on, every exit path is responsible for
4599 4596 // initializing any non-copied parts of the object to zero.
4600 4597 // Also, if this flag is set we make sure that arraycopy interacts properly
4601 4598 // with G1, eliding pre-barriers. See CR 6627983.
4602 4599 dest_uninitialized = true;
4603 4600 } else {
4604 4601 // No zeroing elimination here.
4605 4602 alloc = NULL;
4606 4603 //original_dest = dest;
4607 4604 //dest_uninitialized = false;
4608 4605 }
4609 4606
4610 4607 // Results are placed here:
4611 4608 enum { fast_path = 1, // normal void-returning assembly stub
4612 4609 checked_path = 2, // special assembly stub with cleanup
4613 4610 slow_call_path = 3, // something went wrong; call the VM
4614 4611 zero_path = 4, // bypass when length of copy is zero
4615 4612 bcopy_path = 5, // copy primitive array by 64-bit blocks
4616 4613 PATH_LIMIT = 6
4617 4614 };
4618 4615 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4619 4616 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4620 4617 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4621 4618 record_for_igvn(result_region);
4622 4619 _gvn.set_type_bottom(result_i_o);
4623 4620 _gvn.set_type_bottom(result_memory);
4624 4621 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4625 4622
4626 4623 // The slow_control path:
4627 4624 Node* slow_control;
4628 4625 Node* slow_i_o = i_o();
4629 4626 Node* slow_mem = memory(adr_type);
4630 4627 debug_only(slow_control = (Node*) badAddress);
4631 4628
4632 4629 // Checked control path:
4633 4630 Node* checked_control = top();
4634 4631 Node* checked_mem = NULL;
4635 4632 Node* checked_i_o = NULL;
4636 4633 Node* checked_value = NULL;
4637 4634
4638 4635 if (basic_elem_type == T_CONFLICT) {
4639 4636 assert(!dest_uninitialized, "");
4640 4637 Node* cv = generate_generic_arraycopy(adr_type,
4641 4638 src, src_offset, dest, dest_offset,
4642 4639 copy_length, dest_uninitialized);
4643 4640 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4644 4641 checked_control = control();
4645 4642 checked_i_o = i_o();
4646 4643 checked_mem = memory(adr_type);
4647 4644 checked_value = cv;
4648 4645 set_control(top()); // no fast path
4649 4646 }
4650 4647
4651 4648 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4652 4649 if (not_pos != NULL) {
4653 4650 PreserveJVMState pjvms(this);
4654 4651 set_control(not_pos);
4655 4652
4656 4653 // (6) length must not be negative.
4657 4654 if (!length_never_negative) {
4658 4655 generate_negative_guard(copy_length, slow_region);
4659 4656 }
4660 4657
4661 4658 // copy_length is 0.
4662 4659 if (!stopped() && dest_uninitialized) {
4663 4660 Node* dest_length = alloc->in(AllocateNode::ALength);
4664 4661 if (_gvn.eqv_uncast(copy_length, dest_length)
4665 4662 || _gvn.find_int_con(dest_length, 1) <= 0) {
4666 4663 // There is no zeroing to do. No need for a secondary raw memory barrier.
4667 4664 } else {
4668 4665 // Clear the whole thing since there are no source elements to copy.
4669 4666 generate_clear_array(adr_type, dest, basic_elem_type,
4670 4667 intcon(0), NULL,
4671 4668 alloc->in(AllocateNode::AllocSize));
4672 4669 // Use a secondary InitializeNode as raw memory barrier.
4673 4670 // Currently it is needed only on this path since other
4674 4671 // paths have stub or runtime calls as raw memory barriers.
4675 4672 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4676 4673 Compile::AliasIdxRaw,
4677 4674 top())->as_Initialize();
4678 4675 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4679 4676 }
4680 4677 }
4681 4678
4682 4679 // Present the results of the fast call.
4683 4680 result_region->init_req(zero_path, control());
4684 4681 result_i_o ->init_req(zero_path, i_o());
4685 4682 result_memory->init_req(zero_path, memory(adr_type));
4686 4683 }
4687 4684
4688 4685 if (!stopped() && dest_uninitialized) {
4689 4686 // We have to initialize the *uncopied* part of the array to zero.
4690 4687 // The copy destination is the slice dest[off..off+len]. The other slices
4691 4688 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4692 4689 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4693 4690 Node* dest_length = alloc->in(AllocateNode::ALength);
4694 4691 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4695 4692 copy_length) );
4696 4693
4697 4694 // If there is a head section that needs zeroing, do it now.
4698 4695 if (find_int_con(dest_offset, -1) != 0) {
4699 4696 generate_clear_array(adr_type, dest, basic_elem_type,
4700 4697 intcon(0), dest_offset,
4701 4698 NULL);
4702 4699 }
4703 4700
4704 4701 // Next, perform a dynamic check on the tail length.
4705 4702 // It is often zero, and we can win big if we prove this.
4706 4703 // There are two wins: Avoid generating the ClearArray
4707 4704 // with its attendant messy index arithmetic, and upgrade
4708 4705 // the copy to a more hardware-friendly word size of 64 bits.
4709 4706 Node* tail_ctl = NULL;
4710 4707 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4711 4708 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4712 4709 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4713 4710 tail_ctl = generate_slow_guard(bol_lt, NULL);
4714 4711 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4715 4712 }
4716 4713
4717 4714 // At this point, let's assume there is no tail.
4718 4715 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4719 4716 // There is no tail. Try an upgrade to a 64-bit copy.
4720 4717 bool didit = false;
4721 4718 { PreserveJVMState pjvms(this);
4722 4719 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4723 4720 src, src_offset, dest, dest_offset,
4724 4721 dest_size, dest_uninitialized);
4725 4722 if (didit) {
4726 4723 // Present the results of the block-copying fast call.
4727 4724 result_region->init_req(bcopy_path, control());
4728 4725 result_i_o ->init_req(bcopy_path, i_o());
4729 4726 result_memory->init_req(bcopy_path, memory(adr_type));
4730 4727 }
4731 4728 }
4732 4729 if (didit)
4733 4730 set_control(top()); // no regular fast path
4734 4731 }
4735 4732
4736 4733 // Clear the tail, if any.
4737 4734 if (tail_ctl != NULL) {
4738 4735 Node* notail_ctl = stopped() ? NULL : control();
4739 4736 set_control(tail_ctl);
4740 4737 if (notail_ctl == NULL) {
4741 4738 generate_clear_array(adr_type, dest, basic_elem_type,
4742 4739 dest_tail, NULL,
4743 4740 dest_size);
4744 4741 } else {
4745 4742 // Make a local merge.
4746 4743 Node* done_ctl = new(C,3) RegionNode(3);
4747 4744 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4748 4745 done_ctl->init_req(1, notail_ctl);
4749 4746 done_mem->init_req(1, memory(adr_type));
4750 4747 generate_clear_array(adr_type, dest, basic_elem_type,
4751 4748 dest_tail, NULL,
4752 4749 dest_size);
4753 4750 done_ctl->init_req(2, control());
4754 4751 done_mem->init_req(2, memory(adr_type));
4755 4752 set_control( _gvn.transform(done_ctl) );
4756 4753 set_memory( _gvn.transform(done_mem), adr_type );
4757 4754 }
4758 4755 }
4759 4756 }
4760 4757
4761 4758 BasicType copy_type = basic_elem_type;
4762 4759 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4763 4760 if (!stopped() && copy_type == T_OBJECT) {
4764 4761 // If src and dest have compatible element types, we can copy bits.
4765 4762 // Types S[] and D[] are compatible if D is a supertype of S.
4766 4763 //
4767 4764 // If they are not, we will use checked_oop_disjoint_arraycopy,
4768 4765 // which performs a fast optimistic per-oop check, and backs off
4769 4766 // further to JVM_ArrayCopy on the first per-oop check that fails.
4770 4767 // (Actually, we don't move raw bits only; the GC requires card marks.)
4771 4768
4772 4769 // Get the klassOop for both src and dest
4773 4770 Node* src_klass = load_object_klass(src);
4774 4771 Node* dest_klass = load_object_klass(dest);
4775 4772
4776 4773 // Generate the subtype check.
4777 4774 // This might fold up statically, or then again it might not.
4778 4775 //
4779 4776 // Non-static example: Copying List<String>.elements to a new String[].
4780 4777 // The backing store for a List<String> is always an Object[],
4781 4778 // but its elements are always type String, if the generic types
4782 4779 // are correct at the source level.
4783 4780 //
4784 4781 // Test S[] against D[], not S against D, because (probably)
4785 4782 // the secondary supertype cache is less busy for S[] than S.
4786 4783 // This usually only matters when D is an interface.
4787 4784 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4788 4785 // Plug failing path into checked_oop_disjoint_arraycopy
4789 4786 if (not_subtype_ctrl != top()) {
4790 4787 PreserveJVMState pjvms(this);
4791 4788 set_control(not_subtype_ctrl);
4792 4789 // (At this point we can assume disjoint_bases, since types differ.)
4793 4790 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4794 4791 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4795 4792 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4796 4793 Node* dest_elem_klass = _gvn.transform(n1);
4797 4794 Node* cv = generate_checkcast_arraycopy(adr_type,
4798 4795 dest_elem_klass,
4799 4796 src, src_offset, dest, dest_offset,
4800 4797 ConvI2X(copy_length), dest_uninitialized);
4801 4798 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4802 4799 checked_control = control();
4803 4800 checked_i_o = i_o();
4804 4801 checked_mem = memory(adr_type);
4805 4802 checked_value = cv;
4806 4803 }
4807 4804 // At this point we know we do not need type checks on oop stores.
4808 4805
4809 4806 // Let's see if we need card marks:
4810 4807 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4811 4808 // If we do not need card marks, copy using the jint or jlong stub.
4812 4809 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4813 4810 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4814 4811 "sizes agree");
4815 4812 }
4816 4813 }
4817 4814
4818 4815 if (!stopped()) {
4819 4816 // Generate the fast path, if possible.
4820 4817 PreserveJVMState pjvms(this);
4821 4818 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4822 4819 src, src_offset, dest, dest_offset,
4823 4820 ConvI2X(copy_length), dest_uninitialized);
4824 4821
4825 4822 // Present the results of the fast call.
4826 4823 result_region->init_req(fast_path, control());
4827 4824 result_i_o ->init_req(fast_path, i_o());
4828 4825 result_memory->init_req(fast_path, memory(adr_type));
4829 4826 }
4830 4827
4831 4828 // Here are all the slow paths up to this point, in one bundle:
4832 4829 slow_control = top();
4833 4830 if (slow_region != NULL)
4834 4831 slow_control = _gvn.transform(slow_region);
4835 4832 debug_only(slow_region = (RegionNode*)badAddress);
4836 4833
4837 4834 set_control(checked_control);
4838 4835 if (!stopped()) {
4839 4836 // Clean up after the checked call.
4840 4837 // The returned value is either 0 or -1^K,
4841 4838 // where K = number of partially transferred array elements.
4842 4839 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4843 4840 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4844 4841 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4845 4842
4846 4843 // If it is 0, we are done, so transfer to the end.
4847 4844 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4848 4845 result_region->init_req(checked_path, checks_done);
4849 4846 result_i_o ->init_req(checked_path, checked_i_o);
4850 4847 result_memory->init_req(checked_path, checked_mem);
4851 4848
4852 4849 // If it is not zero, merge into the slow call.
4853 4850 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4854 4851 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4855 4852 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4856 4853 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4857 4854 record_for_igvn(slow_reg2);
4858 4855 slow_reg2 ->init_req(1, slow_control);
4859 4856 slow_i_o2 ->init_req(1, slow_i_o);
4860 4857 slow_mem2 ->init_req(1, slow_mem);
4861 4858 slow_reg2 ->init_req(2, control());
4862 4859 slow_i_o2 ->init_req(2, checked_i_o);
4863 4860 slow_mem2 ->init_req(2, checked_mem);
4864 4861
4865 4862 slow_control = _gvn.transform(slow_reg2);
4866 4863 slow_i_o = _gvn.transform(slow_i_o2);
4867 4864 slow_mem = _gvn.transform(slow_mem2);
4868 4865
4869 4866 if (alloc != NULL) {
4870 4867 // We'll restart from the very beginning, after zeroing the whole thing.
4871 4868 // This can cause double writes, but that's OK since dest is brand new.
4872 4869 // So we ignore the low 31 bits of the value returned from the stub.
4873 4870 } else {
4874 4871 // We must continue the copy exactly where it failed, or else
4875 4872 // another thread might see the wrong number of writes to dest.
4876 4873 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4877 4874 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4878 4875 slow_offset->init_req(1, intcon(0));
4879 4876 slow_offset->init_req(2, checked_offset);
4880 4877 slow_offset = _gvn.transform(slow_offset);
4881 4878
4882 4879 // Adjust the arguments by the conditionally incoming offset.
4883 4880 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4884 4881 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4885 4882 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4886 4883
4887 4884 // Tweak the node variables to adjust the code produced below:
4888 4885 src_offset = src_off_plus;
4889 4886 dest_offset = dest_off_plus;
4890 4887 copy_length = length_minus;
4891 4888 }
4892 4889 }
4893 4890
4894 4891 set_control(slow_control);
4895 4892 if (!stopped()) {
4896 4893 // Generate the slow path, if needed.
4897 4894 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4898 4895
4899 4896 set_memory(slow_mem, adr_type);
4900 4897 set_i_o(slow_i_o);
4901 4898
4902 4899 if (dest_uninitialized) {
4903 4900 generate_clear_array(adr_type, dest, basic_elem_type,
4904 4901 intcon(0), NULL,
4905 4902 alloc->in(AllocateNode::AllocSize));
4906 4903 }
4907 4904
4908 4905 generate_slow_arraycopy(adr_type,
4909 4906 src, src_offset, dest, dest_offset,
4910 4907 copy_length, /*dest_uninitialized*/false);
4911 4908
4912 4909 result_region->init_req(slow_call_path, control());
4913 4910 result_i_o ->init_req(slow_call_path, i_o());
4914 4911 result_memory->init_req(slow_call_path, memory(adr_type));
4915 4912 }
4916 4913
4917 4914 // Remove unused edges.
4918 4915 for (uint i = 1; i < result_region->req(); i++) {
4919 4916 if (result_region->in(i) == NULL)
4920 4917 result_region->init_req(i, top());
4921 4918 }
4922 4919
4923 4920 // Finished; return the combined state.
4924 4921 set_control( _gvn.transform(result_region) );
4925 4922 set_i_o( _gvn.transform(result_i_o) );
4926 4923 set_memory( _gvn.transform(result_memory), adr_type );
4927 4924
4928 4925 // The memory edges above are precise in order to model effects around
4929 4926 // array copies accurately to allow value numbering of field loads around
4930 4927 // arraycopy. Such field loads, both before and after, are common in Java
4931 4928 // collections and similar classes involving header/array data structures.
4932 4929 //
4933 4930 // But with low number of register or when some registers are used or killed
4934 4931 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4935 4932 // The next memory barrier is added to avoid it. If the arraycopy can be
4936 4933 // optimized away (which it can, sometimes) then we can manually remove
4937 4934 // the membar also.
4938 4935 //
4939 4936 // Do not let reads from the cloned object float above the arraycopy.
4940 4937 if (InsertMemBarAfterArraycopy || alloc != NULL)
4941 4938 insert_mem_bar(Op_MemBarCPUOrder);
4942 4939 }
4943 4940
4944 4941
4945 4942 // Helper function which determines if an arraycopy immediately follows
4946 4943 // an allocation, with no intervening tests or other escapes for the object.
4947 4944 AllocateArrayNode*
4948 4945 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4949 4946 RegionNode* slow_region) {
4950 4947 if (stopped()) return NULL; // no fast path
4951 4948 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4952 4949
4953 4950 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4954 4951 if (alloc == NULL) return NULL;
4955 4952
4956 4953 Node* rawmem = memory(Compile::AliasIdxRaw);
4957 4954 // Is the allocation's memory state untouched?
4958 4955 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4959 4956 // Bail out if there have been raw-memory effects since the allocation.
4960 4957 // (Example: There might have been a call or safepoint.)
4961 4958 return NULL;
4962 4959 }
4963 4960 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4964 4961 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4965 4962 return NULL;
4966 4963 }
4967 4964
4968 4965 // There must be no unexpected observers of this allocation.
4969 4966 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4970 4967 Node* obs = ptr->fast_out(i);
4971 4968 if (obs != this->map()) {
4972 4969 return NULL;
4973 4970 }
4974 4971 }
4975 4972
4976 4973 // This arraycopy must unconditionally follow the allocation of the ptr.
4977 4974 Node* alloc_ctl = ptr->in(0);
4978 4975 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4979 4976
4980 4977 Node* ctl = control();
4981 4978 while (ctl != alloc_ctl) {
4982 4979 // There may be guards which feed into the slow_region.
4983 4980 // Any other control flow means that we might not get a chance
4984 4981 // to finish initializing the allocated object.
4985 4982 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4986 4983 IfNode* iff = ctl->in(0)->as_If();
4987 4984 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4988 4985 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4989 4986 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4990 4987 ctl = iff->in(0); // This test feeds the known slow_region.
4991 4988 continue;
4992 4989 }
4993 4990 // One more try: Various low-level checks bottom out in
4994 4991 // uncommon traps. If the debug-info of the trap omits
4995 4992 // any reference to the allocation, as we've already
4996 4993 // observed, then there can be no objection to the trap.
4997 4994 bool found_trap = false;
4998 4995 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4999 4996 Node* obs = not_ctl->fast_out(j);
5000 4997 if (obs->in(0) == not_ctl && obs->is_Call() &&
5001 4998 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5002 4999 found_trap = true; break;
5003 5000 }
5004 5001 }
5005 5002 if (found_trap) {
5006 5003 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5007 5004 continue;
5008 5005 }
5009 5006 }
5010 5007 return NULL;
5011 5008 }
5012 5009
5013 5010 // If we get this far, we have an allocation which immediately
5014 5011 // precedes the arraycopy, and we can take over zeroing the new object.
5015 5012 // The arraycopy will finish the initialization, and provide
5016 5013 // a new control state to which we will anchor the destination pointer.
5017 5014
5018 5015 return alloc;
5019 5016 }
5020 5017
5021 5018 // Helper for initialization of arrays, creating a ClearArray.
5022 5019 // It writes zero bits in [start..end), within the body of an array object.
5023 5020 // The memory effects are all chained onto the 'adr_type' alias category.
5024 5021 //
5025 5022 // Since the object is otherwise uninitialized, we are free
5026 5023 // to put a little "slop" around the edges of the cleared area,
5027 5024 // as long as it does not go back into the array's header,
5028 5025 // or beyond the array end within the heap.
5029 5026 //
5030 5027 // The lower edge can be rounded down to the nearest jint and the
5031 5028 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5032 5029 //
5033 5030 // Arguments:
5034 5031 // adr_type memory slice where writes are generated
5035 5032 // dest oop of the destination array
5036 5033 // basic_elem_type element type of the destination
5037 5034 // slice_idx array index of first element to store
5038 5035 // slice_len number of elements to store (or NULL)
5039 5036 // dest_size total size in bytes of the array object
5040 5037 //
5041 5038 // Exactly one of slice_len or dest_size must be non-NULL.
5042 5039 // If dest_size is non-NULL, zeroing extends to the end of the object.
5043 5040 // If slice_len is non-NULL, the slice_idx value must be a constant.
5044 5041 void
5045 5042 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5046 5043 Node* dest,
5047 5044 BasicType basic_elem_type,
5048 5045 Node* slice_idx,
5049 5046 Node* slice_len,
5050 5047 Node* dest_size) {
5051 5048 // one or the other but not both of slice_len and dest_size:
5052 5049 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5053 5050 if (slice_len == NULL) slice_len = top();
5054 5051 if (dest_size == NULL) dest_size = top();
5055 5052
5056 5053 // operate on this memory slice:
5057 5054 Node* mem = memory(adr_type); // memory slice to operate on
5058 5055
5059 5056 // scaling and rounding of indexes:
5060 5057 int scale = exact_log2(type2aelembytes(basic_elem_type));
5061 5058 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5062 5059 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5063 5060 int bump_bit = (-1 << scale) & BytesPerInt;
5064 5061
5065 5062 // determine constant starts and ends
5066 5063 const intptr_t BIG_NEG = -128;
5067 5064 assert(BIG_NEG + 2*abase < 0, "neg enough");
5068 5065 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5069 5066 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5070 5067 if (slice_len_con == 0) {
5071 5068 return; // nothing to do here
5072 5069 }
5073 5070 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5074 5071 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5075 5072 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5076 5073 assert(end_con < 0, "not two cons");
5077 5074 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5078 5075 BytesPerLong);
5079 5076 }
5080 5077
5081 5078 if (start_con >= 0 && end_con >= 0) {
5082 5079 // Constant start and end. Simple.
5083 5080 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5084 5081 start_con, end_con, &_gvn);
5085 5082 } else if (start_con >= 0 && dest_size != top()) {
5086 5083 // Constant start, pre-rounded end after the tail of the array.
5087 5084 Node* end = dest_size;
5088 5085 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5089 5086 start_con, end, &_gvn);
5090 5087 } else if (start_con >= 0 && slice_len != top()) {
5091 5088 // Constant start, non-constant end. End needs rounding up.
5092 5089 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5093 5090 intptr_t end_base = abase + (slice_idx_con << scale);
5094 5091 int end_round = (-1 << scale) & (BytesPerLong - 1);
5095 5092 Node* end = ConvI2X(slice_len);
5096 5093 if (scale != 0)
5097 5094 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5098 5095 end_base += end_round;
5099 5096 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5100 5097 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5101 5098 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5102 5099 start_con, end, &_gvn);
5103 5100 } else if (start_con < 0 && dest_size != top()) {
5104 5101 // Non-constant start, pre-rounded end after the tail of the array.
5105 5102 // This is almost certainly a "round-to-end" operation.
5106 5103 Node* start = slice_idx;
5107 5104 start = ConvI2X(start);
5108 5105 if (scale != 0)
5109 5106 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5110 5107 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5111 5108 if ((bump_bit | clear_low) != 0) {
5112 5109 int to_clear = (bump_bit | clear_low);
5113 5110 // Align up mod 8, then store a jint zero unconditionally
5114 5111 // just before the mod-8 boundary.
5115 5112 if (((abase + bump_bit) & ~to_clear) - bump_bit
5116 5113 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5117 5114 bump_bit = 0;
5118 5115 assert((abase & to_clear) == 0, "array base must be long-aligned");
5119 5116 } else {
5120 5117 // Bump 'start' up to (or past) the next jint boundary:
5121 5118 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5122 5119 assert((abase & clear_low) == 0, "array base must be int-aligned");
5123 5120 }
5124 5121 // Round bumped 'start' down to jlong boundary in body of array.
5125 5122 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5126 5123 if (bump_bit != 0) {
5127 5124 // Store a zero to the immediately preceding jint:
5128 5125 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5129 5126 Node* p1 = basic_plus_adr(dest, x1);
5130 5127 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5131 5128 mem = _gvn.transform(mem);
5132 5129 }
5133 5130 }
5134 5131 Node* end = dest_size; // pre-rounded
5135 5132 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5136 5133 start, end, &_gvn);
5137 5134 } else {
5138 5135 // Non-constant start, unrounded non-constant end.
5139 5136 // (Nobody zeroes a random midsection of an array using this routine.)
5140 5137 ShouldNotReachHere(); // fix caller
5141 5138 }
5142 5139
5143 5140 // Done.
5144 5141 set_memory(mem, adr_type);
5145 5142 }
5146 5143
5147 5144
5148 5145 bool
5149 5146 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5150 5147 BasicType basic_elem_type,
5151 5148 AllocateNode* alloc,
5152 5149 Node* src, Node* src_offset,
5153 5150 Node* dest, Node* dest_offset,
5154 5151 Node* dest_size, bool dest_uninitialized) {
5155 5152 // See if there is an advantage from block transfer.
5156 5153 int scale = exact_log2(type2aelembytes(basic_elem_type));
5157 5154 if (scale >= LogBytesPerLong)
5158 5155 return false; // it is already a block transfer
5159 5156
5160 5157 // Look at the alignment of the starting offsets.
5161 5158 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5162 5159 const intptr_t BIG_NEG = -128;
5163 5160 assert(BIG_NEG + 2*abase < 0, "neg enough");
5164 5161
5165 5162 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5166 5163 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5167 5164 if (src_off < 0 || dest_off < 0)
5168 5165 // At present, we can only understand constants.
5169 5166 return false;
5170 5167
5171 5168 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5172 5169 // Non-aligned; too bad.
5173 5170 // One more chance: Pick off an initial 32-bit word.
5174 5171 // This is a common case, since abase can be odd mod 8.
5175 5172 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5176 5173 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5177 5174 Node* sptr = basic_plus_adr(src, src_off);
5178 5175 Node* dptr = basic_plus_adr(dest, dest_off);
5179 5176 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5180 5177 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5181 5178 src_off += BytesPerInt;
5182 5179 dest_off += BytesPerInt;
5183 5180 } else {
5184 5181 return false;
5185 5182 }
5186 5183 }
5187 5184 assert(src_off % BytesPerLong == 0, "");
5188 5185 assert(dest_off % BytesPerLong == 0, "");
5189 5186
5190 5187 // Do this copy by giant steps.
5191 5188 Node* sptr = basic_plus_adr(src, src_off);
5192 5189 Node* dptr = basic_plus_adr(dest, dest_off);
5193 5190 Node* countx = dest_size;
5194 5191 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5195 5192 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5196 5193
5197 5194 bool disjoint_bases = true; // since alloc != NULL
5198 5195 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5199 5196 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5200 5197
5201 5198 return true;
5202 5199 }
5203 5200
5204 5201
5205 5202 // Helper function; generates code for the slow case.
5206 5203 // We make a call to a runtime method which emulates the native method,
5207 5204 // but without the native wrapper overhead.
5208 5205 void
5209 5206 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5210 5207 Node* src, Node* src_offset,
5211 5208 Node* dest, Node* dest_offset,
5212 5209 Node* copy_length, bool dest_uninitialized) {
5213 5210 assert(!dest_uninitialized, "Invariant");
5214 5211 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5215 5212 OptoRuntime::slow_arraycopy_Type(),
5216 5213 OptoRuntime::slow_arraycopy_Java(),
5217 5214 "slow_arraycopy", adr_type,
5218 5215 src, src_offset, dest, dest_offset,
5219 5216 copy_length);
5220 5217
5221 5218 // Handle exceptions thrown by this fellow:
5222 5219 make_slow_call_ex(call, env()->Throwable_klass(), false);
5223 5220 }
5224 5221
5225 5222 // Helper function; generates code for cases requiring runtime checks.
5226 5223 Node*
5227 5224 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5228 5225 Node* dest_elem_klass,
5229 5226 Node* src, Node* src_offset,
5230 5227 Node* dest, Node* dest_offset,
5231 5228 Node* copy_length, bool dest_uninitialized) {
5232 5229 if (stopped()) return NULL;
5233 5230
5234 5231 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5235 5232 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5236 5233 return NULL;
5237 5234 }
5238 5235
5239 5236 // Pick out the parameters required to perform a store-check
5240 5237 // for the target array. This is an optimistic check. It will
5241 5238 // look in each non-null element's class, at the desired klass's
5242 5239 // super_check_offset, for the desired klass.
5243 5240 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5244 5241 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5245 5242 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5246 5243 Node* check_offset = ConvI2X(_gvn.transform(n3));
5247 5244 Node* check_value = dest_elem_klass;
5248 5245
5249 5246 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5250 5247 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5251 5248
5252 5249 // (We know the arrays are never conjoint, because their types differ.)
5253 5250 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5254 5251 OptoRuntime::checkcast_arraycopy_Type(),
5255 5252 copyfunc_addr, "checkcast_arraycopy", adr_type,
5256 5253 // five arguments, of which two are
5257 5254 // intptr_t (jlong in LP64)
5258 5255 src_start, dest_start,
5259 5256 copy_length XTOP,
5260 5257 check_offset XTOP,
5261 5258 check_value);
5262 5259
5263 5260 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5264 5261 }
5265 5262
5266 5263
5267 5264 // Helper function; generates code for cases requiring runtime checks.
5268 5265 Node*
5269 5266 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5270 5267 Node* src, Node* src_offset,
5271 5268 Node* dest, Node* dest_offset,
5272 5269 Node* copy_length, bool dest_uninitialized) {
5273 5270 assert(!dest_uninitialized, "Invariant");
5274 5271 if (stopped()) return NULL;
5275 5272 address copyfunc_addr = StubRoutines::generic_arraycopy();
5276 5273 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5277 5274 return NULL;
5278 5275 }
5279 5276
5280 5277 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5281 5278 OptoRuntime::generic_arraycopy_Type(),
5282 5279 copyfunc_addr, "generic_arraycopy", adr_type,
5283 5280 src, src_offset, dest, dest_offset, copy_length);
5284 5281
5285 5282 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5286 5283 }
5287 5284
5288 5285 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5289 5286 void
5290 5287 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5291 5288 BasicType basic_elem_type,
5292 5289 bool disjoint_bases,
5293 5290 Node* src, Node* src_offset,
5294 5291 Node* dest, Node* dest_offset,
5295 5292 Node* copy_length, bool dest_uninitialized) {
5296 5293 if (stopped()) return; // nothing to do
5297 5294
5298 5295 Node* src_start = src;
5299 5296 Node* dest_start = dest;
5300 5297 if (src_offset != NULL || dest_offset != NULL) {
5301 5298 assert(src_offset != NULL && dest_offset != NULL, "");
5302 5299 src_start = array_element_address(src, src_offset, basic_elem_type);
5303 5300 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5304 5301 }
5305 5302
5306 5303 // Figure out which arraycopy runtime method to call.
5307 5304 const char* copyfunc_name = "arraycopy";
5308 5305 address copyfunc_addr =
5309 5306 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5310 5307 disjoint_bases, copyfunc_name, dest_uninitialized);
5311 5308
5312 5309 // Call it. Note that the count_ix value is not scaled to a byte-size.
5313 5310 make_runtime_call(RC_LEAF|RC_NO_FP,
5314 5311 OptoRuntime::fast_arraycopy_Type(),
5315 5312 copyfunc_addr, copyfunc_name, adr_type,
5316 5313 src_start, dest_start, copy_length XTOP);
5317 5314 }
|
↓ open down ↓ |
4912 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX