Print this page
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/cpu/sparc/vm/cppInterpreter_sparc.cpp
+++ new/src/cpu/sparc/vm/cppInterpreter_sparc.cpp
1 1 /*
2 2 * Copyright 2007 Sun Microsystems, Inc. All Rights Reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 21 * have any questions.
22 22 *
23 23 */
24 24
25 25 #include "incls/_precompiled.incl"
26 26 #include "incls/_cppInterpreter_sparc.cpp.incl"
27 27
28 28 #ifdef CC_INTERP
29 29
30 30 // Routine exists to make tracebacks look decent in debugger
31 31 // while "shadow" interpreter frames are on stack. It is also
32 32 // used to distinguish interpreter frames.
33 33
34 34 extern "C" void RecursiveInterpreterActivation(interpreterState istate) {
35 35 ShouldNotReachHere();
36 36 }
37 37
38 38 bool CppInterpreter::contains(address pc) {
39 39 return ( _code->contains(pc) ||
40 40 ( pc == (CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset)));
41 41 }
42 42
43 43 #define STATE(field_name) Lstate, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
44 44 #define __ _masm->
45 45
46 46 Label frame_manager_entry;
47 47 Label fast_accessor_slow_entry_path; // fast accessor methods need to be able to jmp to unsynchronized
48 48 // c++ interpreter entry point this holds that entry point label.
49 49
50 50 static address unctrap_frame_manager_entry = NULL;
51 51
52 52 static address interpreter_return_address = NULL;
53 53 static address deopt_frame_manager_return_atos = NULL;
54 54 static address deopt_frame_manager_return_btos = NULL;
55 55 static address deopt_frame_manager_return_itos = NULL;
56 56 static address deopt_frame_manager_return_ltos = NULL;
57 57 static address deopt_frame_manager_return_ftos = NULL;
58 58 static address deopt_frame_manager_return_dtos = NULL;
59 59 static address deopt_frame_manager_return_vtos = NULL;
60 60
61 61 const Register prevState = G1_scratch;
62 62
63 63 void InterpreterGenerator::save_native_result(void) {
64 64 // result potentially in O0/O1: save it across calls
65 65 __ stf(FloatRegisterImpl::D, F0, STATE(_native_fresult));
66 66 #ifdef _LP64
67 67 __ stx(O0, STATE(_native_lresult));
68 68 #else
69 69 __ std(O0, STATE(_native_lresult));
70 70 #endif
71 71 }
72 72
73 73 void InterpreterGenerator::restore_native_result(void) {
74 74
75 75 // Restore any method result value
76 76 __ ldf(FloatRegisterImpl::D, STATE(_native_fresult), F0);
77 77 #ifdef _LP64
78 78 __ ldx(STATE(_native_lresult), O0);
79 79 #else
80 80 __ ldd(STATE(_native_lresult), O0);
81 81 #endif
82 82 }
83 83
84 84 // A result handler converts/unboxes a native call result into
85 85 // a java interpreter/compiler result. The current frame is an
86 86 // interpreter frame. The activation frame unwind code must be
87 87 // consistent with that of TemplateTable::_return(...). In the
88 88 // case of native methods, the caller's SP was not modified.
89 89 address CppInterpreterGenerator::generate_result_handler_for(BasicType type) {
90 90 address entry = __ pc();
91 91 Register Itos_i = Otos_i ->after_save();
92 92 Register Itos_l = Otos_l ->after_save();
93 93 Register Itos_l1 = Otos_l1->after_save();
94 94 Register Itos_l2 = Otos_l2->after_save();
95 95 switch (type) {
96 96 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, Itos_i); break; // !0 => true; 0 => false
97 97 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, Itos_i); break; // cannot use and3, 0xFFFF too big as immediate value!
98 98 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, Itos_i); break;
99 99 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, Itos_i); break;
100 100 case T_LONG :
101 101 #ifndef _LP64
102 102 __ mov(O1, Itos_l2); // move other half of long
103 103 #endif // ifdef or no ifdef, fall through to the T_INT case
104 104 case T_INT : __ mov(O0, Itos_i); break;
105 105 case T_VOID : /* nothing to do */ break;
106 106 case T_FLOAT : assert(F0 == Ftos_f, "fix this code" ); break;
107 107 case T_DOUBLE : assert(F0 == Ftos_d, "fix this code" ); break;
108 108 case T_OBJECT :
109 109 __ ld_ptr(STATE(_oop_temp), Itos_i);
110 110 __ verify_oop(Itos_i);
111 111 break;
112 112 default : ShouldNotReachHere();
113 113 }
114 114 __ ret(); // return from interpreter activation
115 115 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
116 116 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
117 117 return entry;
118 118 }
119 119
120 120 // tosca based result to c++ interpreter stack based result.
121 121 // Result goes to address in L1_scratch
122 122
123 123 address CppInterpreterGenerator::generate_tosca_to_stack_converter(BasicType type) {
124 124 // A result is in the native abi result register from a native method call.
125 125 // We need to return this result to the interpreter by pushing the result on the interpreter's
126 126 // stack. This is relatively simple the destination is in L1_scratch
127 127 // i.e. L1_scratch is the first free element on the stack. If we "push" a return value we must
128 128 // adjust L1_scratch
129 129 address entry = __ pc();
130 130 switch (type) {
131 131 case T_BOOLEAN:
132 132 // !0 => true; 0 => false
133 133 __ subcc(G0, O0, G0);
134 134 __ addc(G0, 0, O0);
135 135 __ st(O0, L1_scratch, 0);
136 136 __ sub(L1_scratch, wordSize, L1_scratch);
137 137 break;
138 138
139 139 // cannot use and3, 0xFFFF too big as immediate value!
140 140 case T_CHAR :
141 141 __ sll(O0, 16, O0);
142 142 __ srl(O0, 16, O0);
143 143 __ st(O0, L1_scratch, 0);
144 144 __ sub(L1_scratch, wordSize, L1_scratch);
145 145 break;
146 146
147 147 case T_BYTE :
148 148 __ sll(O0, 24, O0);
149 149 __ sra(O0, 24, O0);
150 150 __ st(O0, L1_scratch, 0);
151 151 __ sub(L1_scratch, wordSize, L1_scratch);
|
↓ open down ↓ |
151 lines elided |
↑ open up ↑ |
152 152 break;
153 153
154 154 case T_SHORT :
155 155 __ sll(O0, 16, O0);
156 156 __ sra(O0, 16, O0);
157 157 __ st(O0, L1_scratch, 0);
158 158 __ sub(L1_scratch, wordSize, L1_scratch);
159 159 break;
160 160 case T_LONG :
161 161 #ifndef _LP64
162 -#if !defined(_LP64) && defined(COMPILER2)
162 +#if defined(COMPILER2)
163 163 // All return values are where we want them, except for Longs. C2 returns
164 164 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
165 165 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
166 166 // build even if we are returning from interpreted we just do a little
167 167 // stupid shuffing.
168 168 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
169 169 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
170 170 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
171 171 __ stx(G1, L1_scratch, -wordSize);
172 172 #else
173 173 // native result is in O0, O1
174 174 __ st(O1, L1_scratch, 0); // Low order
175 175 __ st(O0, L1_scratch, -wordSize); // High order
176 -#endif /* !_LP64 && COMPILER2 */
176 +#endif /* COMPILER2 */
177 177 #else
178 - __ stx(O0, L1_scratch, 0);
179 -__ breakpoint_trap();
178 + __ stx(O0, L1_scratch, -wordSize);
180 179 #endif
181 180 __ sub(L1_scratch, 2*wordSize, L1_scratch);
182 181 break;
183 182
184 183 case T_INT :
185 184 __ st(O0, L1_scratch, 0);
186 185 __ sub(L1_scratch, wordSize, L1_scratch);
187 186 break;
188 187
189 188 case T_VOID : /* nothing to do */
190 189 break;
191 190
192 191 case T_FLOAT :
193 192 __ stf(FloatRegisterImpl::S, F0, L1_scratch, 0);
194 193 __ sub(L1_scratch, wordSize, L1_scratch);
195 194 break;
196 195
197 196 case T_DOUBLE :
198 197 // Every stack slot is aligned on 64 bit, However is this
199 198 // the correct stack slot on 64bit?? QQQ
200 199 __ stf(FloatRegisterImpl::D, F0, L1_scratch, -wordSize);
201 200 __ sub(L1_scratch, 2*wordSize, L1_scratch);
202 201 break;
203 202 case T_OBJECT :
204 203 __ verify_oop(O0);
205 204 __ st_ptr(O0, L1_scratch, 0);
206 205 __ sub(L1_scratch, wordSize, L1_scratch);
207 206 break;
208 207 default : ShouldNotReachHere();
209 208 }
210 209 __ retl(); // return from interpreter activation
211 210 __ delayed()->nop(); // schedule this better
212 211 NOT_PRODUCT(__ emit_long(0);) // marker for disassembly
213 212 return entry;
214 213 }
215 214
216 215 address CppInterpreterGenerator::generate_stack_to_stack_converter(BasicType type) {
217 216 // A result is in the java expression stack of the interpreted method that has just
218 217 // returned. Place this result on the java expression stack of the caller.
219 218 //
220 219 // The current interpreter activation in Lstate is for the method just returning its
221 220 // result. So we know that the result of this method is on the top of the current
222 221 // execution stack (which is pre-pushed) and will be return to the top of the caller
223 222 // stack. The top of the callers stack is the bottom of the locals of the current
224 223 // activation.
225 224 // Because of the way activation are managed by the frame manager the value of esp is
226 225 // below both the stack top of the current activation and naturally the stack top
227 226 // of the calling activation. This enable this routine to leave the return address
228 227 // to the frame manager on the stack and do a vanilla return.
229 228 //
|
↓ open down ↓ |
40 lines elided |
↑ open up ↑ |
230 229 // On entry: O0 - points to source (callee stack top)
231 230 // O1 - points to destination (caller stack top [i.e. free location])
232 231 // destroys O2, O3
233 232 //
234 233
235 234 address entry = __ pc();
236 235 switch (type) {
237 236 case T_VOID: break;
238 237 break;
239 238 case T_FLOAT :
240 - __ breakpoint_trap(Assembler::zero);
241 239 case T_BOOLEAN:
242 240 case T_CHAR :
243 241 case T_BYTE :
244 242 case T_SHORT :
245 243 case T_INT :
246 244 // 1 word result
247 245 __ ld(O0, 0, O2);
248 246 __ st(O2, O1, 0);
249 247 __ sub(O1, wordSize, O1);
250 248 break;
251 249 case T_DOUBLE :
252 250 case T_LONG :
253 251 // return top two words on current expression stack to caller's expression stack
254 252 // The caller's expression stack is adjacent to the current frame manager's intepretState
255 253 // except we allocated one extra word for this intepretState so we won't overwrite it
256 254 // when we return a two word result.
257 255 #ifdef _LP64
258 -__ breakpoint_trap();
259 - // Hmm now that longs are in one entry should "_ptr" really be "x"?
260 256 __ ld_ptr(O0, 0, O2);
261 - __ ld_ptr(O0, wordSize, O3);
262 - __ st_ptr(O3, O1, 0);
263 257 __ st_ptr(O2, O1, -wordSize);
264 258 #else
265 259 __ ld(O0, 0, O2);
266 260 __ ld(O0, wordSize, O3);
267 261 __ st(O3, O1, 0);
268 262 __ st(O2, O1, -wordSize);
269 263 #endif
270 264 __ sub(O1, 2*wordSize, O1);
271 265 break;
272 266 case T_OBJECT :
273 267 __ ld_ptr(O0, 0, O2);
274 268 __ verify_oop(O2); // verify it
275 269 __ st_ptr(O2, O1, 0);
276 270 __ sub(O1, wordSize, O1);
277 271 break;
278 272 default : ShouldNotReachHere();
279 273 }
280 274 __ retl();
281 275 __ delayed()->nop(); // QQ schedule this better
282 276 return entry;
283 277 }
284 278
285 279 address CppInterpreterGenerator::generate_stack_to_native_abi_converter(BasicType type) {
286 280 // A result is in the java expression stack of the interpreted method that has just
287 281 // returned. Place this result in the native abi that the caller expects.
288 282 // We are in a new frame registers we set must be in caller (i.e. callstub) frame.
289 283 //
290 284 // Similar to generate_stack_to_stack_converter above. Called at a similar time from the
291 285 // frame manager execept in this situation the caller is native code (c1/c2/call_stub)
292 286 // and so rather than return result onto caller's java expression stack we return the
293 287 // result in the expected location based on the native abi.
294 288 // On entry: O0 - source (stack top)
295 289 // On exit result in expected output register
296 290 // QQQ schedule this better
297 291
298 292 address entry = __ pc();
299 293 switch (type) {
300 294 case T_VOID: break;
301 295 break;
302 296 case T_FLOAT :
303 297 __ ldf(FloatRegisterImpl::S, O0, 0, F0);
304 298 break;
305 299 case T_BOOLEAN:
306 300 case T_CHAR :
307 301 case T_BYTE :
308 302 case T_SHORT :
309 303 case T_INT :
310 304 // 1 word result
311 305 __ ld(O0, 0, O0->after_save());
|
↓ open down ↓ |
39 lines elided |
↑ open up ↑ |
312 306 break;
313 307 case T_DOUBLE :
314 308 __ ldf(FloatRegisterImpl::D, O0, 0, F0);
315 309 break;
316 310 case T_LONG :
317 311 // return top two words on current expression stack to caller's expression stack
318 312 // The caller's expression stack is adjacent to the current frame manager's interpretState
319 313 // except we allocated one extra word for this intepretState so we won't overwrite it
320 314 // when we return a two word result.
321 315 #ifdef _LP64
322 -__ breakpoint_trap();
323 - // Hmm now that longs are in one entry should "_ptr" really be "x"?
324 316 __ ld_ptr(O0, 0, O0->after_save());
325 - __ ld_ptr(O0, wordSize, O1->after_save());
326 317 #else
327 318 __ ld(O0, wordSize, O1->after_save());
328 319 __ ld(O0, 0, O0->after_save());
329 320 #endif
330 321 #if defined(COMPILER2) && !defined(_LP64)
331 322 // C2 expects long results in G1 we can't tell if we're returning to interpreted
332 323 // or compiled so just be safe use G1 and O0/O1
333 324
334 325 // Shift bits into high (msb) of G1
335 326 __ sllx(Otos_l1->after_save(), 32, G1);
336 327 // Zero extend low bits
337 328 __ srl (Otos_l2->after_save(), 0, Otos_l2->after_save());
338 329 __ or3 (Otos_l2->after_save(), G1, G1);
339 330 #endif /* COMPILER2 */
340 331 break;
341 332 case T_OBJECT :
342 333 __ ld_ptr(O0, 0, O0->after_save());
343 334 __ verify_oop(O0->after_save()); // verify it
344 335 break;
345 336 default : ShouldNotReachHere();
346 337 }
347 338 __ retl();
348 339 __ delayed()->nop();
349 340 return entry;
350 341 }
351 342
352 343 address CppInterpreter::return_entry(TosState state, int length) {
353 344 // make it look good in the debugger
354 345 return CAST_FROM_FN_PTR(address, RecursiveInterpreterActivation) + frame::pc_return_offset;
355 346 }
356 347
357 348 address CppInterpreter::deopt_entry(TosState state, int length) {
358 349 address ret = NULL;
359 350 if (length != 0) {
360 351 switch (state) {
361 352 case atos: ret = deopt_frame_manager_return_atos; break;
362 353 case btos: ret = deopt_frame_manager_return_btos; break;
363 354 case ctos:
364 355 case stos:
365 356 case itos: ret = deopt_frame_manager_return_itos; break;
366 357 case ltos: ret = deopt_frame_manager_return_ltos; break;
367 358 case ftos: ret = deopt_frame_manager_return_ftos; break;
368 359 case dtos: ret = deopt_frame_manager_return_dtos; break;
369 360 case vtos: ret = deopt_frame_manager_return_vtos; break;
370 361 }
371 362 } else {
372 363 ret = unctrap_frame_manager_entry; // re-execute the bytecode ( e.g. uncommon trap)
373 364 }
374 365 assert(ret != NULL, "Not initialized");
375 366 return ret;
376 367 }
377 368
378 369 //
379 370 // Helpers for commoning out cases in the various type of method entries.
380 371 //
381 372
382 373 // increment invocation count & check for overflow
383 374 //
384 375 // Note: checking for negative value instead of overflow
385 376 // so we have a 'sticky' overflow test
386 377 //
387 378 // Lmethod: method
388 379 // ??: invocation counter
389 380 //
390 381 void InterpreterGenerator::generate_counter_incr(Label* overflow, Label* profile_method, Label* profile_method_continue) {
391 382 // Update standard invocation counters
392 383 __ increment_invocation_counter(O0, G3_scratch);
393 384 if (ProfileInterpreter) { // %%% Merge this into methodDataOop
394 385 __ ld_ptr(STATE(_method), G3_scratch);
395 386 Address interpreter_invocation_counter(G3_scratch, 0, in_bytes(methodOopDesc::interpreter_invocation_counter_offset()));
396 387 __ ld(interpreter_invocation_counter, G3_scratch);
397 388 __ inc(G3_scratch);
398 389 __ st(G3_scratch, interpreter_invocation_counter);
399 390 }
400 391
401 392 Address invocation_limit(G3_scratch, (address)&InvocationCounter::InterpreterInvocationLimit);
402 393 __ sethi(invocation_limit);
403 394 __ ld(invocation_limit, G3_scratch);
404 395 __ cmp(O0, G3_scratch);
405 396 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pn, *overflow);
406 397 __ delayed()->nop();
407 398
408 399 }
409 400
410 401 address InterpreterGenerator::generate_empty_entry(void) {
411 402
412 403 // A method that does nothing but return...
413 404
414 405 address entry = __ pc();
415 406 Label slow_path;
416 407
417 408 __ verify_oop(G5_method);
418 409
419 410 // do nothing for empty methods (do not even increment invocation counter)
420 411 if ( UseFastEmptyMethods) {
421 412 // If we need a safepoint check, generate full interpreter entry.
422 413 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
423 414 __ load_contents(sync_state, G3_scratch);
424 415 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
425 416 __ br(Assembler::notEqual, false, Assembler::pn, frame_manager_entry);
426 417 __ delayed()->nop();
427 418
428 419 // Code: _return
429 420 __ retl();
430 421 __ delayed()->mov(O5_savedSP, SP);
431 422 return entry;
432 423 }
433 424 return NULL;
434 425 }
435 426
436 427 // Call an accessor method (assuming it is resolved, otherwise drop into
437 428 // vanilla (slow path) entry
438 429
439 430 // Generates code to elide accessor methods
440 431 // Uses G3_scratch and G1_scratch as scratch
441 432 address InterpreterGenerator::generate_accessor_entry(void) {
442 433
443 434 // Code: _aload_0, _(i|a)getfield, _(i|a)return or any rewrites thereof;
444 435 // parameter size = 1
445 436 // Note: We can only use this code if the getfield has been resolved
446 437 // and if we don't have a null-pointer exception => check for
447 438 // these conditions first and use slow path if necessary.
448 439 address entry = __ pc();
449 440 Label slow_path;
450 441
451 442 if ( UseFastAccessorMethods) {
452 443 // Check if we need to reach a safepoint and generate full interpreter
453 444 // frame if so.
454 445 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
455 446 __ load_contents(sync_state, G3_scratch);
456 447 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
457 448 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
458 449 __ delayed()->nop();
459 450
460 451 // Check if local 0 != NULL
461 452 __ ld_ptr(Gargs, G0, Otos_i ); // get local 0
462 453 __ tst(Otos_i); // check if local 0 == NULL and go the slow path
463 454 __ brx(Assembler::zero, false, Assembler::pn, slow_path);
464 455 __ delayed()->nop();
465 456
466 457
467 458 // read first instruction word and extract bytecode @ 1 and index @ 2
468 459 // get first 4 bytes of the bytecodes (big endian!)
469 460 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::const_offset())), G1_scratch);
470 461 __ ld(Address(G1_scratch, 0, in_bytes(constMethodOopDesc::codes_offset())), G1_scratch);
471 462
472 463 // move index @ 2 far left then to the right most two bytes.
473 464 __ sll(G1_scratch, 2*BitsPerByte, G1_scratch);
474 465 __ srl(G1_scratch, 2*BitsPerByte - exact_log2(in_words(
475 466 ConstantPoolCacheEntry::size()) * BytesPerWord), G1_scratch);
476 467
477 468 // get constant pool cache
478 469 __ ld_ptr(G5_method, in_bytes(methodOopDesc::constants_offset()), G3_scratch);
479 470 __ ld_ptr(G3_scratch, constantPoolOopDesc::cache_offset_in_bytes(), G3_scratch);
480 471
481 472 // get specific constant pool cache entry
482 473 __ add(G3_scratch, G1_scratch, G3_scratch);
483 474
484 475 // Check the constant Pool cache entry to see if it has been resolved.
485 476 // If not, need the slow path.
486 477 ByteSize cp_base_offset = constantPoolCacheOopDesc::base_offset();
487 478 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::indices_offset()), G1_scratch);
488 479 __ srl(G1_scratch, 2*BitsPerByte, G1_scratch);
489 480 __ and3(G1_scratch, 0xFF, G1_scratch);
490 481 __ cmp(G1_scratch, Bytecodes::_getfield);
491 482 __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
492 483 __ delayed()->nop();
493 484
494 485 // Get the type and return field offset from the constant pool cache
495 486 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::flags_offset()), G1_scratch);
496 487 __ ld_ptr(G3_scratch, in_bytes(cp_base_offset + ConstantPoolCacheEntry::f2_offset()), G3_scratch);
497 488
498 489 Label xreturn_path;
499 490 // Need to differentiate between igetfield, agetfield, bgetfield etc.
500 491 // because they are different sizes.
501 492 // Get the type from the constant pool cache
502 493 __ srl(G1_scratch, ConstantPoolCacheEntry::tosBits, G1_scratch);
503 494 // Make sure we don't need to mask G1_scratch for tosBits after the above shift
504 495 ConstantPoolCacheEntry::verify_tosBits();
505 496 __ cmp(G1_scratch, atos );
506 497 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
507 498 __ delayed()->ld_ptr(Otos_i, G3_scratch, Otos_i);
508 499 __ cmp(G1_scratch, itos);
509 500 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
510 501 __ delayed()->ld(Otos_i, G3_scratch, Otos_i);
511 502 __ cmp(G1_scratch, stos);
512 503 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
513 504 __ delayed()->ldsh(Otos_i, G3_scratch, Otos_i);
514 505 __ cmp(G1_scratch, ctos);
515 506 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
516 507 __ delayed()->lduh(Otos_i, G3_scratch, Otos_i);
517 508 #ifdef ASSERT
518 509 __ cmp(G1_scratch, btos);
519 510 __ br(Assembler::equal, true, Assembler::pt, xreturn_path);
520 511 __ delayed()->ldsb(Otos_i, G3_scratch, Otos_i);
521 512 __ should_not_reach_here();
522 513 #endif
523 514 __ ldsb(Otos_i, G3_scratch, Otos_i);
524 515 __ bind(xreturn_path);
525 516
526 517 // _ireturn/_areturn
527 518 __ retl(); // return from leaf routine
528 519 __ delayed()->mov(O5_savedSP, SP);
529 520
530 521 // Generate regular method entry
531 522 __ bind(slow_path);
532 523 __ ba(false, fast_accessor_slow_entry_path);
533 524 __ delayed()->nop();
534 525 return entry;
535 526 }
536 527 return NULL;
537 528 }
538 529
539 530 //
540 531 // Interpreter stub for calling a native method. (C++ interpreter)
541 532 // This sets up a somewhat different looking stack for calling the native method
542 533 // than the typical interpreter frame setup.
543 534 //
544 535
545 536 address InterpreterGenerator::generate_native_entry(bool synchronized) {
546 537 address entry = __ pc();
547 538
548 539 // the following temporary registers are used during frame creation
549 540 const Register Gtmp1 = G3_scratch ;
550 541 const Register Gtmp2 = G1_scratch;
551 542 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
552 543
553 544 bool inc_counter = UseCompiler || CountCompiledCalls;
554 545
555 546 // make sure registers are different!
556 547 assert_different_registers(G2_thread, G5_method, Gargs, Gtmp1, Gtmp2);
557 548
558 549 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
559 550
560 551 Label Lentry;
561 552 __ bind(Lentry);
562 553
563 554 __ verify_oop(G5_method);
564 555
565 556 const Register Glocals_size = G3;
566 557 assert_different_registers(Glocals_size, G4_scratch, Gframe_size);
567 558
568 559 // make sure method is native & not abstract
569 560 // rethink these assertions - they can be simplified and shared (gri 2/25/2000)
570 561 #ifdef ASSERT
571 562 __ ld(access_flags, Gtmp1);
572 563 {
573 564 Label L;
574 565 __ btst(JVM_ACC_NATIVE, Gtmp1);
575 566 __ br(Assembler::notZero, false, Assembler::pt, L);
576 567 __ delayed()->nop();
577 568 __ stop("tried to execute non-native method as native");
578 569 __ bind(L);
579 570 }
580 571 { Label L;
581 572 __ btst(JVM_ACC_ABSTRACT, Gtmp1);
582 573 __ br(Assembler::zero, false, Assembler::pt, L);
583 574 __ delayed()->nop();
584 575 __ stop("tried to execute abstract method as non-abstract");
585 576 __ bind(L);
586 577 }
587 578 #endif // ASSERT
588 579
589 580 __ lduh(size_of_parameters, Gtmp1);
590 581 __ sll(Gtmp1, LogBytesPerWord, Gtmp2); // parameter size in bytes
591 582 __ add(Gargs, Gtmp2, Gargs); // points to first local + BytesPerWord
592 583 // NEW
593 584 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
594 585 // generate the code to allocate the interpreter stack frame
595 586 // NEW FRAME ALLOCATED HERE
596 587 // save callers original sp
597 588 // __ mov(SP, I5_savedSP->after_restore());
598 589
599 590 generate_compute_interpreter_state(Lstate, G0, true);
600 591
601 592 // At this point Lstate points to new interpreter state
602 593 //
603 594
604 595 const Address do_not_unlock_if_synchronized(G2_thread, 0,
605 596 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
606 597 // Since at this point in the method invocation the exception handler
607 598 // would try to exit the monitor of synchronized methods which hasn't
608 599 // been entered yet, we set the thread local variable
609 600 // _do_not_unlock_if_synchronized to true. If any exception was thrown by
610 601 // runtime, exception handling i.e. unlock_if_synchronized_method will
611 602 // check this thread local flag.
612 603 // This flag has two effects, one is to force an unwind in the topmost
613 604 // interpreter frame and not perform an unlock while doing so.
614 605
615 606 __ movbool(true, G3_scratch);
616 607 __ stbool(G3_scratch, do_not_unlock_if_synchronized);
617 608
618 609
619 610 // increment invocation counter and check for overflow
620 611 //
621 612 // Note: checking for negative value instead of overflow
622 613 // so we have a 'sticky' overflow test (may be of
623 614 // importance as soon as we have true MT/MP)
624 615 Label invocation_counter_overflow;
625 616 if (inc_counter) {
626 617 generate_counter_incr(&invocation_counter_overflow, NULL, NULL);
627 618 }
628 619 Label Lcontinue;
629 620 __ bind(Lcontinue);
630 621
631 622 bang_stack_shadow_pages(true);
632 623 // reset the _do_not_unlock_if_synchronized flag
633 624 __ stbool(G0, do_not_unlock_if_synchronized);
634 625
635 626 // check for synchronized methods
636 627 // Must happen AFTER invocation_counter check, so method is not locked
637 628 // if counter overflows.
638 629
639 630 if (synchronized) {
640 631 lock_method();
641 632 // Don't see how G2_thread is preserved here...
642 633 // __ verify_thread(); QQQ destroys L0,L1 can't use
643 634 } else {
644 635 #ifdef ASSERT
645 636 { Label ok;
646 637 __ ld_ptr(STATE(_method), G5_method);
647 638 __ ld(access_flags, O0);
648 639 __ btst(JVM_ACC_SYNCHRONIZED, O0);
649 640 __ br( Assembler::zero, false, Assembler::pt, ok);
650 641 __ delayed()->nop();
651 642 __ stop("method needs synchronization");
652 643 __ bind(ok);
653 644 }
654 645 #endif // ASSERT
655 646 }
656 647
657 648 // start execution
658 649
659 650 // __ verify_thread(); kills L1,L2 can't use at the moment
660 651
661 652 // jvmti/jvmpi support
662 653 __ notify_method_entry();
663 654
664 655 // native call
665 656
666 657 // (note that O0 is never an oop--at most it is a handle)
667 658 // It is important not to smash any handles created by this call,
668 659 // until any oop handle in O0 is dereferenced.
669 660
670 661 // (note that the space for outgoing params is preallocated)
671 662
672 663 // get signature handler
673 664
674 665 Label pending_exception_present;
675 666
676 667 { Label L;
677 668 __ ld_ptr(STATE(_method), G5_method);
678 669 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
679 670 __ tst(G3_scratch);
680 671 __ brx(Assembler::notZero, false, Assembler::pt, L);
681 672 __ delayed()->nop();
682 673 __ call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::prepare_native_call), G5_method, false);
683 674 __ ld_ptr(STATE(_method), G5_method);
684 675
685 676 Address exception_addr(G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
686 677 __ ld_ptr(exception_addr, G3_scratch);
687 678 __ br_notnull(G3_scratch, false, Assembler::pn, pending_exception_present);
688 679 __ delayed()->nop();
689 680 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::signature_handler_offset())), G3_scratch);
690 681 __ bind(L);
691 682 }
692 683
693 684 // Push a new frame so that the args will really be stored in
694 685 // Copy a few locals across so the new frame has the variables
695 686 // we need but these values will be dead at the jni call and
696 687 // therefore not gc volatile like the values in the current
697 688 // frame (Lstate in particular)
698 689
699 690 // Flush the state pointer to the register save area
700 691 // Which is the only register we need for a stack walk.
701 692 __ st_ptr(Lstate, SP, (Lstate->sp_offset_in_saved_window() * wordSize) + STACK_BIAS);
702 693
703 694 __ mov(Lstate, O1); // Need to pass the state pointer across the frame
704 695
705 696 // Calculate current frame size
706 697 __ sub(SP, FP, O3); // Calculate negative of current frame size
707 698 __ save(SP, O3, SP); // Allocate an identical sized frame
708 699
709 700 __ mov(I1, Lstate); // In the "natural" register.
710 701
711 702 // Note I7 has leftover trash. Slow signature handler will fill it in
712 703 // should we get there. Normal jni call will set reasonable last_Java_pc
713 704 // below (and fix I7 so the stack trace doesn't have a meaningless frame
714 705 // in it).
715 706
716 707
717 708 // call signature handler
718 709 __ ld_ptr(STATE(_method), Lmethod);
719 710 __ ld_ptr(STATE(_locals), Llocals);
720 711
721 712 __ callr(G3_scratch, 0);
722 713 __ delayed()->nop();
723 714 __ ld_ptr(STATE(_thread), G2_thread); // restore thread (shouldn't be needed)
724 715
725 716 { Label not_static;
726 717
727 718 __ ld_ptr(STATE(_method), G5_method);
728 719 __ ld(access_flags, O0);
729 720 __ btst(JVM_ACC_STATIC, O0);
730 721 __ br( Assembler::zero, false, Assembler::pt, not_static);
731 722 __ delayed()->
732 723 // get native function entry point(O0 is a good temp until the very end)
733 724 ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc::native_function_offset())), O0);
734 725 // for static methods insert the mirror argument
735 726 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
736 727
737 728 __ ld_ptr(Address(G5_method, 0, in_bytes(methodOopDesc:: constants_offset())), O1);
738 729 __ ld_ptr(Address(O1, 0, constantPoolOopDesc::pool_holder_offset_in_bytes()), O1);
739 730 __ ld_ptr(O1, mirror_offset, O1);
740 731 // where the mirror handle body is allocated:
741 732 #ifdef ASSERT
742 733 if (!PrintSignatureHandlers) // do not dirty the output with this
743 734 { Label L;
744 735 __ tst(O1);
745 736 __ brx(Assembler::notZero, false, Assembler::pt, L);
746 737 __ delayed()->nop();
747 738 __ stop("mirror is missing");
748 739 __ bind(L);
749 740 }
750 741 #endif // ASSERT
751 742 __ st_ptr(O1, STATE(_oop_temp));
752 743 __ add(STATE(_oop_temp), O1); // this is really an LEA not an add
753 744 __ bind(not_static);
754 745 }
755 746
756 747 // At this point, arguments have been copied off of stack into
757 748 // their JNI positions, which are O1..O5 and SP[68..].
758 749 // Oops are boxed in-place on the stack, with handles copied to arguments.
759 750 // The result handler is in Lscratch. O0 will shortly hold the JNIEnv*.
760 751
761 752 #ifdef ASSERT
762 753 { Label L;
763 754 __ tst(O0);
764 755 __ brx(Assembler::notZero, false, Assembler::pt, L);
765 756 __ delayed()->nop();
766 757 __ stop("native entry point is missing");
767 758 __ bind(L);
768 759 }
769 760 #endif // ASSERT
770 761
771 762 //
772 763 // setup the java frame anchor
773 764 //
774 765 // The scavenge function only needs to know that the PC of this frame is
775 766 // in the interpreter method entry code, it doesn't need to know the exact
776 767 // PC and hence we can use O7 which points to the return address from the
777 768 // previous call in the code stream (signature handler function)
778 769 //
779 770 // The other trick is we set last_Java_sp to FP instead of the usual SP because
780 771 // we have pushed the extra frame in order to protect the volatile register(s)
781 772 // in that frame when we return from the jni call
782 773 //
783 774
784 775
785 776 __ set_last_Java_frame(FP, O7);
786 777 __ mov(O7, I7); // make dummy interpreter frame look like one above,
787 778 // not meaningless information that'll confuse me.
788 779
789 780 // flush the windows now. We don't care about the current (protection) frame
790 781 // only the outer frames
791 782
792 783 __ flush_windows();
793 784
794 785 // mark windows as flushed
795 786 Address flags(G2_thread,
796 787 0,
797 788 in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset()));
798 789 __ set(JavaFrameAnchor::flushed, G3_scratch);
799 790 __ st(G3_scratch, flags);
800 791
801 792 // Transition from _thread_in_Java to _thread_in_native. We are already safepoint ready.
802 793
803 794 Address thread_state(G2_thread, 0, in_bytes(JavaThread::thread_state_offset()));
804 795 #ifdef ASSERT
805 796 { Label L;
806 797 __ ld(thread_state, G3_scratch);
807 798 __ cmp(G3_scratch, _thread_in_Java);
808 799 __ br(Assembler::equal, false, Assembler::pt, L);
809 800 __ delayed()->nop();
810 801 __ stop("Wrong thread state in native stub");
811 802 __ bind(L);
812 803 }
813 804 #endif // ASSERT
814 805 __ set(_thread_in_native, G3_scratch);
815 806 __ st(G3_scratch, thread_state);
816 807
817 808 // Call the jni method, using the delay slot to set the JNIEnv* argument.
818 809 __ callr(O0, 0);
819 810 __ delayed()->
820 811 add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
821 812 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
822 813
823 814 // must we block?
824 815
825 816 // Block, if necessary, before resuming in _thread_in_Java state.
826 817 // In order for GC to work, don't clear the last_Java_sp until after blocking.
827 818 { Label no_block;
828 819 Address sync_state(G3_scratch, SafepointSynchronize::address_of_state());
829 820
830 821 // Switch thread to "native transition" state before reading the synchronization state.
831 822 // This additional state is necessary because reading and testing the synchronization
832 823 // state is not atomic w.r.t. GC, as this scenario demonstrates:
833 824 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
834 825 // VM thread changes sync state to synchronizing and suspends threads for GC.
835 826 // Thread A is resumed to finish this native method, but doesn't block here since it
836 827 // didn't see any synchronization is progress, and escapes.
837 828 __ set(_thread_in_native_trans, G3_scratch);
838 829 __ st(G3_scratch, thread_state);
839 830 if(os::is_MP()) {
840 831 // Write serialization page so VM thread can do a pseudo remote membar.
841 832 // We use the current thread pointer to calculate a thread specific
842 833 // offset to write to within the page. This minimizes bus traffic
843 834 // due to cache line collision.
844 835 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
845 836 }
846 837 __ load_contents(sync_state, G3_scratch);
847 838 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
848 839
849 840
850 841 Label L;
851 842 Address suspend_state(G2_thread, 0, in_bytes(JavaThread::suspend_flags_offset()));
852 843 __ br(Assembler::notEqual, false, Assembler::pn, L);
853 844 __ delayed()->
854 845 ld(suspend_state, G3_scratch);
855 846 __ cmp(G3_scratch, 0);
856 847 __ br(Assembler::equal, false, Assembler::pt, no_block);
857 848 __ delayed()->nop();
858 849 __ bind(L);
859 850
860 851 // Block. Save any potential method result value before the operation and
861 852 // use a leaf call to leave the last_Java_frame setup undisturbed.
862 853 save_native_result();
863 854 __ call_VM_leaf(noreg,
864 855 CAST_FROM_FN_PTR(address, JavaThread::check_safepoint_and_suspend_for_native_trans),
865 856 G2_thread);
866 857 __ ld_ptr(STATE(_thread), G2_thread); // restore thread
867 858 // Restore any method result value
868 859 restore_native_result();
869 860 __ bind(no_block);
870 861 }
871 862
872 863 // Clear the frame anchor now
873 864
874 865 __ reset_last_Java_frame();
875 866
876 867 // Move the result handler address
877 868 __ mov(Lscratch, G3_scratch);
878 869 // return possible result to the outer frame
879 870 #ifndef __LP64
880 871 __ mov(O0, I0);
881 872 __ restore(O1, G0, O1);
882 873 #else
883 874 __ restore(O0, G0, O0);
884 875 #endif /* __LP64 */
885 876
886 877 // Move result handler to expected register
887 878 __ mov(G3_scratch, Lscratch);
888 879
889 880
890 881 // thread state is thread_in_native_trans. Any safepoint blocking has
891 882 // happened in the trampoline we are ready to switch to thread_in_Java.
892 883
893 884 __ set(_thread_in_Java, G3_scratch);
894 885 __ st(G3_scratch, thread_state);
895 886
896 887 // If we have an oop result store it where it will be safe for any further gc
897 888 // until we return now that we've released the handle it might be protected by
898 889
899 890 {
900 891 Label no_oop, store_result;
901 892
902 893 __ set((intptr_t)AbstractInterpreter::result_handler(T_OBJECT), G3_scratch);
903 894 __ cmp(G3_scratch, Lscratch);
904 895 __ brx(Assembler::notEqual, false, Assembler::pt, no_oop);
905 896 __ delayed()->nop();
906 897 __ addcc(G0, O0, O0);
907 898 __ brx(Assembler::notZero, true, Assembler::pt, store_result); // if result is not NULL:
908 899 __ delayed()->ld_ptr(O0, 0, O0); // unbox it
909 900 __ mov(G0, O0);
910 901
911 902 __ bind(store_result);
912 903 // Store it where gc will look for it and result handler expects it.
913 904 __ st_ptr(O0, STATE(_oop_temp));
914 905
915 906 __ bind(no_oop);
916 907
917 908 }
918 909
919 910 // reset handle block
920 911 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), G3_scratch);
921 912 __ st_ptr(G0, G3_scratch, JNIHandleBlock::top_offset_in_bytes());
922 913
923 914
924 915 // handle exceptions (exception handling will handle unlocking!)
925 916 { Label L;
926 917 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
927 918
928 919 __ ld_ptr(exception_addr, Gtemp);
929 920 __ tst(Gtemp);
930 921 __ brx(Assembler::equal, false, Assembler::pt, L);
931 922 __ delayed()->nop();
932 923 __ bind(pending_exception_present);
933 924 // With c++ interpreter we just leave it pending caller will do the correct thing. However...
934 925 // Like x86 we ignore the result of the native call and leave the method locked. This
935 926 // seems wrong to leave things locked.
936 927
937 928 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
938 929 __ delayed()->restore(I5_savedSP, G0, SP); // remove interpreter frame
939 930
940 931 __ bind(L);
941 932 }
942 933
943 934 // jvmdi/jvmpi support (preserves thread register)
944 935 __ notify_method_exit(true, ilgl, InterpreterMacroAssembler::NotifyJVMTI);
945 936
946 937 if (synchronized) {
947 938 // save and restore any potential method result value around the unlocking operation
948 939 save_native_result();
949 940
950 941 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
951 942 // Get the initial monitor we allocated
952 943 __ sub(Lstate, entry_size, O1); // initial monitor
953 944 __ unlock_object(O1);
954 945 restore_native_result();
955 946 }
956 947
957 948 #if defined(COMPILER2) && !defined(_LP64)
958 949
959 950 // C2 expects long results in G1 we can't tell if we're returning to interpreted
960 951 // or compiled so just be safe.
961 952
962 953 __ sllx(O0, 32, G1); // Shift bits into high G1
963 954 __ srl (O1, 0, O1); // Zero extend O1
964 955 __ or3 (O1, G1, G1); // OR 64 bits into G1
965 956
966 957 #endif /* COMPILER2 && !_LP64 */
967 958
968 959 #ifdef ASSERT
969 960 {
970 961 Label ok;
971 962 __ cmp(I5_savedSP, FP);
972 963 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, ok);
973 964 __ delayed()->nop();
974 965 __ stop("bad I5_savedSP value");
975 966 __ should_not_reach_here();
976 967 __ bind(ok);
977 968 }
978 969 #endif
979 970 // Calls result handler which POPS FRAME
980 971 if (TraceJumps) {
981 972 // Move target to register that is recordable
982 973 __ mov(Lscratch, G3_scratch);
983 974 __ JMP(G3_scratch, 0);
984 975 } else {
985 976 __ jmp(Lscratch, 0);
986 977 }
987 978 __ delayed()->nop();
988 979
989 980 if (inc_counter) {
990 981 // handle invocation counter overflow
991 982 __ bind(invocation_counter_overflow);
992 983 generate_counter_overflow(Lcontinue);
993 984 }
994 985
995 986
996 987 return entry;
997 988 }
998 989
999 990 void CppInterpreterGenerator::generate_compute_interpreter_state(const Register state,
1000 991 const Register prev_state,
1001 992 bool native) {
1002 993
1003 994 // On entry
1004 995 // G5_method - caller's method
1005 996 // Gargs - points to initial parameters (i.e. locals[0])
1006 997 // G2_thread - valid? (C1 only??)
1007 998 // "prev_state" - contains any previous frame manager state which we must save a link
1008 999 //
1009 1000 // On return
1010 1001 // "state" is a pointer to the newly allocated state object. We must allocate and initialize
1011 1002 // a new interpretState object and the method expression stack.
1012 1003
1013 1004 assert_different_registers(state, prev_state);
1014 1005 assert_different_registers(prev_state, G3_scratch);
1015 1006 const Register Gtmp = G3_scratch;
1016 1007 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1017 1008 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1018 1009 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1019 1010 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1020 1011 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1021 1012
1022 1013 // slop factor is two extra slots on the expression stack so that
1023 1014 // we always have room to store a result when returning from a call without parameters
1024 1015 // that returns a result.
1025 1016
1026 1017 const int slop_factor = 2*wordSize;
1027 1018
1028 1019 const int fixed_size = ((sizeof(BytecodeInterpreter) + slop_factor) >> LogBytesPerWord) + // what is the slop factor?
1029 1020 frame::memory_parameter_word_sp_offset + // register save area + param window
1030 1021 (native ? frame::interpreter_frame_extra_outgoing_argument_words : 0); // JNI, class
1031 1022
1032 1023 // XXX G5_method valid
1033 1024
1034 1025 // Now compute new frame size
1035 1026
1036 1027 if (native) {
1037 1028 __ lduh( size_of_parameters, Gtmp );
1038 1029 __ calc_mem_param_words(Gtmp, Gtmp); // space for native call parameters passed on the stack in words
1039 1030 } else {
1040 1031 __ lduh(max_stack, Gtmp); // Full size expression stack
1041 1032 }
1042 1033 __ add(Gtmp, fixed_size, Gtmp); // plus the fixed portion
1043 1034
1044 1035 __ neg(Gtmp); // negative space for stack/parameters in words
1045 1036 __ and3(Gtmp, -WordsPerLong, Gtmp); // make multiple of 2 (SP must be 2-word aligned)
1046 1037 __ sll(Gtmp, LogBytesPerWord, Gtmp); // negative space for frame in bytes
1047 1038
1048 1039 // Need to do stack size check here before we fault on large frames
1049 1040
1050 1041 Label stack_ok;
1051 1042
1052 1043 const int max_pages = StackShadowPages > (StackRedPages+StackYellowPages) ? StackShadowPages :
1053 1044 (StackRedPages+StackYellowPages);
1054 1045
1055 1046
1056 1047 __ ld_ptr(G2_thread, in_bytes(Thread::stack_base_offset()), O0);
1057 1048 __ ld_ptr(G2_thread, in_bytes(Thread::stack_size_offset()), O1);
1058 1049 // compute stack bottom
1059 1050 __ sub(O0, O1, O0);
1060 1051
1061 1052 // Avoid touching the guard pages
1062 1053 // Also a fudge for frame size of BytecodeInterpreter::run
1063 1054 // It varies from 1k->4k depending on build type
1064 1055 const int fudge = 6 * K;
1065 1056
1066 1057 __ set(fudge + (max_pages * os::vm_page_size()), O1);
1067 1058
1068 1059 __ add(O0, O1, O0);
1069 1060 __ sub(O0, Gtmp, O0);
1070 1061 __ cmp(SP, O0);
1071 1062 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, stack_ok);
1072 1063 __ delayed()->nop();
1073 1064
1074 1065 // throw exception return address becomes throwing pc
1075 1066
1076 1067 __ call_VM(Oexception, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_StackOverflowError));
1077 1068 __ stop("never reached");
1078 1069
1079 1070 __ bind(stack_ok);
1080 1071
1081 1072 __ save(SP, Gtmp, SP); // setup new frame and register window
1082 1073
1083 1074 // New window I7 call_stub or previous activation
1084 1075 // O6 - register save area, BytecodeInterpreter just below it, args/locals just above that
1085 1076 //
1086 1077 __ sub(FP, sizeof(BytecodeInterpreter), state); // Point to new Interpreter state
1087 1078 __ add(state, STACK_BIAS, state ); // Account for 64bit bias
1088 1079
1089 1080 #define XXX_STATE(field_name) state, in_bytes(byte_offset_of(BytecodeInterpreter, field_name))
1090 1081
1091 1082 // Initialize a new Interpreter state
1092 1083 // orig_sp - caller's original sp
1093 1084 // G2_thread - thread
1094 1085 // Gargs - &locals[0] (unbiased?)
1095 1086 // G5_method - method
1096 1087 // SP (biased) - accounts for full size java stack, BytecodeInterpreter object, register save area, and register parameter save window
1097 1088
1098 1089
1099 1090 __ set(0xdead0004, O1);
1100 1091
1101 1092
1102 1093 __ st_ptr(Gargs, XXX_STATE(_locals));
1103 1094 __ st_ptr(G0, XXX_STATE(_oop_temp));
1104 1095
1105 1096 __ st_ptr(state, XXX_STATE(_self_link)); // point to self
1106 1097 __ st_ptr(prev_state->after_save(), XXX_STATE(_prev_link)); // Chain interpreter states
1107 1098 __ st_ptr(G2_thread, XXX_STATE(_thread)); // Store javathread
1108 1099
1109 1100 if (native) {
1110 1101 __ st_ptr(G0, XXX_STATE(_bcp));
1111 1102 } else {
1112 1103 __ ld_ptr(G5_method, in_bytes(methodOopDesc::const_offset()), O2); // get constMethodOop
1113 1104 __ add(O2, in_bytes(constMethodOopDesc::codes_offset()), O2); // get bcp
1114 1105 __ st_ptr(O2, XXX_STATE(_bcp));
1115 1106 }
1116 1107
1117 1108 __ st_ptr(G0, XXX_STATE(_mdx));
1118 1109 __ st_ptr(G5_method, XXX_STATE(_method));
1119 1110
1120 1111 __ set((int) BytecodeInterpreter::method_entry, O1);
1121 1112 __ st(O1, XXX_STATE(_msg));
1122 1113
1123 1114 __ ld_ptr(constants, O3);
1124 1115 __ ld_ptr(O3, constantPoolOopDesc::cache_offset_in_bytes(), O2);
1125 1116 __ st_ptr(O2, XXX_STATE(_constants));
1126 1117
1127 1118 __ st_ptr(G0, XXX_STATE(_result._to_call._callee));
1128 1119
1129 1120 // Monitor base is just start of BytecodeInterpreter object;
1130 1121 __ mov(state, O2);
1131 1122 __ st_ptr(O2, XXX_STATE(_monitor_base));
1132 1123
1133 1124 // Do we need a monitor for synchonized method?
1134 1125 {
1135 1126 __ ld(access_flags, O1);
1136 1127 Label done;
1137 1128 Label got_obj;
1138 1129 __ btst(JVM_ACC_SYNCHRONIZED, O1);
1139 1130 __ br( Assembler::zero, false, Assembler::pt, done);
1140 1131
1141 1132 const int mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1142 1133 __ delayed()->btst(JVM_ACC_STATIC, O1);
1143 1134 __ ld_ptr(XXX_STATE(_locals), O1);
1144 1135 __ br( Assembler::zero, true, Assembler::pt, got_obj);
1145 1136 __ delayed()->ld_ptr(O1, 0, O1); // get receiver for not-static case
1146 1137 __ ld_ptr(constants, O1);
1147 1138 __ ld_ptr( O1, constantPoolOopDesc::pool_holder_offset_in_bytes(), O1);
1148 1139 // lock the mirror, not the klassOop
1149 1140 __ ld_ptr( O1, mirror_offset, O1);
1150 1141
1151 1142 __ bind(got_obj);
1152 1143
1153 1144 #ifdef ASSERT
1154 1145 __ tst(O1);
1155 1146 __ breakpoint_trap(Assembler::zero);
1156 1147 #endif // ASSERT
1157 1148
1158 1149 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1159 1150 __ sub(SP, entry_size, SP); // account for initial monitor
1160 1151 __ sub(O2, entry_size, O2); // initial monitor
1161 1152 __ st_ptr(O1, O2, BasicObjectLock::obj_offset_in_bytes()); // and allocate it for interpreter use
1162 1153 __ bind(done);
1163 1154 }
1164 1155
1165 1156 // Remember initial frame bottom
1166 1157
1167 1158 __ st_ptr(SP, XXX_STATE(_frame_bottom));
1168 1159
1169 1160 __ st_ptr(O2, XXX_STATE(_stack_base));
1170 1161
1171 1162 __ sub(O2, wordSize, O2); // prepush
1172 1163 __ st_ptr(O2, XXX_STATE(_stack)); // PREPUSH
1173 1164
1174 1165 __ lduh(max_stack, O3); // Full size expression stack
1175 1166 __ sll(O3, LogBytesPerWord, O3);
1176 1167 __ sub(O2, O3, O3);
1177 1168 // __ sub(O3, wordSize, O3); // so prepush doesn't look out of bounds
1178 1169 __ st_ptr(O3, XXX_STATE(_stack_limit));
1179 1170
1180 1171 if (!native) {
1181 1172 //
1182 1173 // Code to initialize locals
1183 1174 //
1184 1175 Register init_value = noreg; // will be G0 if we must clear locals
1185 1176 // Now zero locals
1186 1177 if (true /* zerolocals */ || ClearInterpreterLocals) {
1187 1178 // explicitly initialize locals
1188 1179 init_value = G0;
1189 1180 } else {
1190 1181 #ifdef ASSERT
1191 1182 // initialize locals to a garbage pattern for better debugging
1192 1183 init_value = O3;
1193 1184 __ set( 0x0F0F0F0F, init_value );
1194 1185 #endif // ASSERT
1195 1186 }
1196 1187 if (init_value != noreg) {
1197 1188 Label clear_loop;
1198 1189
1199 1190 // NOTE: If you change the frame layout, this code will need to
1200 1191 // be updated!
1201 1192 __ lduh( size_of_locals, O2 );
1202 1193 __ lduh( size_of_parameters, O1 );
1203 1194 __ sll( O2, LogBytesPerWord, O2);
1204 1195 __ sll( O1, LogBytesPerWord, O1 );
1205 1196 __ ld_ptr(XXX_STATE(_locals), L2_scratch);
1206 1197 __ sub( L2_scratch, O2, O2 );
1207 1198 __ sub( L2_scratch, O1, O1 );
1208 1199
1209 1200 __ bind( clear_loop );
1210 1201 __ inc( O2, wordSize );
1211 1202
1212 1203 __ cmp( O2, O1 );
1213 1204 __ br( Assembler::lessEqualUnsigned, true, Assembler::pt, clear_loop );
1214 1205 __ delayed()->st_ptr( init_value, O2, 0 );
1215 1206 }
1216 1207 }
1217 1208 }
1218 1209 // Find preallocated monitor and lock method (C++ interpreter)
1219 1210 //
1220 1211 void InterpreterGenerator::lock_method(void) {
1221 1212 // Lock the current method.
1222 1213 // Destroys registers L2_scratch, L3_scratch, O0
1223 1214 //
1224 1215 // Find everything relative to Lstate
1225 1216
1226 1217 #ifdef ASSERT
1227 1218 __ ld_ptr(STATE(_method), L2_scratch);
1228 1219 __ ld(L2_scratch, in_bytes(methodOopDesc::access_flags_offset()), O0);
1229 1220
1230 1221 { Label ok;
1231 1222 __ btst(JVM_ACC_SYNCHRONIZED, O0);
1232 1223 __ br( Assembler::notZero, false, Assembler::pt, ok);
1233 1224 __ delayed()->nop();
1234 1225 __ stop("method doesn't need synchronization");
1235 1226 __ bind(ok);
1236 1227 }
1237 1228 #endif // ASSERT
1238 1229
1239 1230 // monitor is already allocated at stack base
1240 1231 // and the lockee is already present
1241 1232 __ ld_ptr(STATE(_stack_base), L2_scratch);
1242 1233 __ ld_ptr(L2_scratch, BasicObjectLock::obj_offset_in_bytes(), O0); // get object
1243 1234 __ lock_object(L2_scratch, O0);
1244 1235
1245 1236 }
1246 1237
1247 1238 // Generate code for handling resuming a deopted method
1248 1239 void CppInterpreterGenerator::generate_deopt_handling() {
1249 1240
1250 1241 Label return_from_deopt_common;
1251 1242
1252 1243 // deopt needs to jump to here to enter the interpreter (return a result)
1253 1244 deopt_frame_manager_return_atos = __ pc();
1254 1245
1255 1246 // O0/O1 live
1256 1247 __ ba(false, return_from_deopt_common);
1257 1248 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_OBJECT), L3_scratch); // Result stub address array index
1258 1249
1259 1250
1260 1251 // deopt needs to jump to here to enter the interpreter (return a result)
1261 1252 deopt_frame_manager_return_btos = __ pc();
1262 1253
1263 1254 // O0/O1 live
1264 1255 __ ba(false, return_from_deopt_common);
1265 1256 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_BOOLEAN), L3_scratch); // Result stub address array index
1266 1257
1267 1258 // deopt needs to jump to here to enter the interpreter (return a result)
1268 1259 deopt_frame_manager_return_itos = __ pc();
1269 1260
1270 1261 // O0/O1 live
1271 1262 __ ba(false, return_from_deopt_common);
1272 1263 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_INT), L3_scratch); // Result stub address array index
1273 1264
1274 1265 // deopt needs to jump to here to enter the interpreter (return a result)
1275 1266
1276 1267 deopt_frame_manager_return_ltos = __ pc();
1277 1268 #if !defined(_LP64) && defined(COMPILER2)
1278 1269 // All return values are where we want them, except for Longs. C2 returns
1279 1270 // longs in G1 in the 32-bit build whereas the interpreter wants them in O0/O1.
1280 1271 // Since the interpreter will return longs in G1 and O0/O1 in the 32bit
1281 1272 // build even if we are returning from interpreted we just do a little
1282 1273 // stupid shuffing.
1283 1274 // Note: I tried to make c2 return longs in O0/O1 and G1 so we wouldn't have to
1284 1275 // do this here. Unfortunately if we did a rethrow we'd see an machepilog node
1285 1276 // first which would move g1 -> O0/O1 and destroy the exception we were throwing.
1286 1277
1287 1278 __ srl (G1, 0,O1);
1288 1279 __ srlx(G1,32,O0);
1289 1280 #endif /* !_LP64 && COMPILER2 */
1290 1281 // O0/O1 live
1291 1282 __ ba(false, return_from_deopt_common);
1292 1283 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_LONG), L3_scratch); // Result stub address array index
1293 1284
1294 1285 // deopt needs to jump to here to enter the interpreter (return a result)
1295 1286
1296 1287 deopt_frame_manager_return_ftos = __ pc();
1297 1288 // O0/O1 live
1298 1289 __ ba(false, return_from_deopt_common);
1299 1290 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_FLOAT), L3_scratch); // Result stub address array index
1300 1291
1301 1292 // deopt needs to jump to here to enter the interpreter (return a result)
1302 1293 deopt_frame_manager_return_dtos = __ pc();
1303 1294
1304 1295 // O0/O1 live
1305 1296 __ ba(false, return_from_deopt_common);
1306 1297 __ delayed()->set(AbstractInterpreter::BasicType_as_index(T_DOUBLE), L3_scratch); // Result stub address array index
1307 1298
1308 1299 // deopt needs to jump to here to enter the interpreter (return a result)
1309 1300 deopt_frame_manager_return_vtos = __ pc();
1310 1301
1311 1302 // O0/O1 live
1312 1303 __ set(AbstractInterpreter::BasicType_as_index(T_VOID), L3_scratch);
1313 1304
1314 1305 // Deopt return common
1315 1306 // an index is present that lets us move any possible result being
1316 1307 // return to the interpreter's stack
1317 1308 //
1318 1309 __ bind(return_from_deopt_common);
1319 1310
1320 1311 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1321 1312 // stack is in the state that the calling convention left it.
1322 1313 // Copy the result from native abi result and place it on java expression stack.
1323 1314
1324 1315 // Current interpreter state is present in Lstate
1325 1316
1326 1317 // Get current pre-pushed top of interpreter stack
1327 1318 // Any result (if any) is in native abi
1328 1319 // result type index is in L3_scratch
1329 1320
1330 1321 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1331 1322
1332 1323 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1333 1324 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1334 1325 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1335 1326 __ jmpl(Lscratch, G0, O7); // and convert it
1336 1327 __ delayed()->nop();
1337 1328
1338 1329 // L1_scratch points to top of stack (prepushed)
1339 1330 __ st_ptr(L1_scratch, STATE(_stack));
1340 1331 }
1341 1332
1342 1333 // Generate the code to handle a more_monitors message from the c++ interpreter
1343 1334 void CppInterpreterGenerator::generate_more_monitors() {
1344 1335
1345 1336 Label entry, loop;
1346 1337 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
1347 1338 // 1. compute new pointers // esp: old expression stack top
1348 1339 __ delayed()->ld_ptr(STATE(_stack_base), L4_scratch); // current expression stack bottom
1349 1340 __ sub(L4_scratch, entry_size, L4_scratch);
1350 1341 __ st_ptr(L4_scratch, STATE(_stack_base));
1351 1342
1352 1343 __ sub(SP, entry_size, SP); // Grow stack
1353 1344 __ st_ptr(SP, STATE(_frame_bottom));
1354 1345
1355 1346 __ ld_ptr(STATE(_stack_limit), L2_scratch);
1356 1347 __ sub(L2_scratch, entry_size, L2_scratch);
1357 1348 __ st_ptr(L2_scratch, STATE(_stack_limit));
1358 1349
1359 1350 __ ld_ptr(STATE(_stack), L1_scratch); // Get current stack top
1360 1351 __ sub(L1_scratch, entry_size, L1_scratch);
1361 1352 __ st_ptr(L1_scratch, STATE(_stack));
1362 1353 __ ba(false, entry);
1363 1354 __ delayed()->add(L1_scratch, wordSize, L1_scratch); // first real entry (undo prepush)
1364 1355
1365 1356 // 2. move expression stack
|
↓ open down ↓ |
1030 lines elided |
↑ open up ↑ |
1366 1357
1367 1358 __ bind(loop);
1368 1359 __ st_ptr(L3_scratch, Address(L1_scratch, 0));
1369 1360 __ add(L1_scratch, wordSize, L1_scratch);
1370 1361 __ bind(entry);
1371 1362 __ cmp(L1_scratch, L4_scratch);
1372 1363 __ br(Assembler::notEqual, false, Assembler::pt, loop);
1373 1364 __ delayed()->ld_ptr(L1_scratch, entry_size, L3_scratch);
1374 1365
1375 1366 // now zero the slot so we can find it.
1376 - __ st(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1367 + __ st_ptr(G0, L4_scratch, BasicObjectLock::obj_offset_in_bytes());
1377 1368
1378 1369 }
1379 1370
1380 1371 // Initial entry to C++ interpreter from the call_stub.
1381 1372 // This entry point is called the frame manager since it handles the generation
1382 1373 // of interpreter activation frames via requests directly from the vm (via call_stub)
1383 1374 // and via requests from the interpreter. The requests from the call_stub happen
1384 1375 // directly thru the entry point. Requests from the interpreter happen via returning
1385 1376 // from the interpreter and examining the message the interpreter has returned to
1386 1377 // the frame manager. The frame manager can take the following requests:
1387 1378
1388 1379 // NO_REQUEST - error, should never happen.
1389 1380 // MORE_MONITORS - need a new monitor. Shuffle the expression stack on down and
1390 1381 // allocate a new monitor.
1391 1382 // CALL_METHOD - setup a new activation to call a new method. Very similar to what
1392 1383 // happens during entry during the entry via the call stub.
1393 1384 // RETURN_FROM_METHOD - remove an activation. Return to interpreter or call stub.
1394 1385 //
1395 1386 // Arguments:
1396 1387 //
1397 1388 // ebx: methodOop
1398 1389 // ecx: receiver - unused (retrieved from stack as needed)
1399 1390 // esi: previous frame manager state (NULL from the call_stub/c1/c2)
1400 1391 //
1401 1392 //
1402 1393 // Stack layout at entry
1403 1394 //
1404 1395 // [ return address ] <--- esp
1405 1396 // [ parameter n ]
1406 1397 // ...
1407 1398 // [ parameter 1 ]
1408 1399 // [ expression stack ]
1409 1400 //
1410 1401 //
1411 1402 // We are free to blow any registers we like because the call_stub which brought us here
1412 1403 // initially has preserved the callee save registers already.
1413 1404 //
1414 1405 //
1415 1406
1416 1407 static address interpreter_frame_manager = NULL;
1417 1408
1418 1409 #ifdef ASSERT
1419 1410 #define VALIDATE_STATE(scratch, marker) \
1420 1411 { \
1421 1412 Label skip; \
1422 1413 __ ld_ptr(STATE(_self_link), scratch); \
1423 1414 __ cmp(Lstate, scratch); \
1424 1415 __ brx(Assembler::equal, false, Assembler::pt, skip); \
1425 1416 __ delayed()->nop(); \
1426 1417 __ breakpoint_trap(); \
1427 1418 __ emit_long(marker); \
1428 1419 __ bind(skip); \
1429 1420 }
1430 1421 #else
1431 1422 #define VALIDATE_STATE(scratch, marker)
1432 1423 #endif /* ASSERT */
1433 1424
1434 1425 void CppInterpreterGenerator::adjust_callers_stack(Register args) {
1435 1426 //
1436 1427 // Adjust caller's stack so that all the locals can be contiguous with
1437 1428 // the parameters.
1438 1429 // Worries about stack overflow make this a pain.
1439 1430 //
1440 1431 // Destroys args, G3_scratch, G3_scratch
1441 1432 // In/Out O5_savedSP (sender's original SP)
1442 1433 //
1443 1434 // assert_different_registers(state, prev_state);
1444 1435 const Register Gtmp = G3_scratch;
1445 1436 const Register tmp = O2;
1446 1437 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1447 1438 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1448 1439
1449 1440 __ lduh(size_of_parameters, tmp);
1450 1441 __ sll(tmp, LogBytesPerWord, Gtmp); // parameter size in bytes
1451 1442 __ add(args, Gtmp, Gargs); // points to first local + BytesPerWord
1452 1443 // NEW
1453 1444 __ add(Gargs, -wordSize, Gargs); // points to first local[0]
1454 1445 // determine extra space for non-argument locals & adjust caller's SP
1455 1446 // Gtmp1: parameter size in words
1456 1447 __ lduh(size_of_locals, Gtmp);
1457 1448 __ compute_extra_locals_size_in_bytes(tmp, Gtmp, Gtmp);
1458 1449
1459 1450 #if 1
1460 1451 // c2i adapters place the final interpreter argument in the register save area for O0/I0
1461 1452 // the call_stub will place the final interpreter argument at
1462 1453 // frame::memory_parameter_word_sp_offset. This is mostly not noticable for either asm
1463 1454 // or c++ interpreter. However with the c++ interpreter when we do a recursive call
1464 1455 // and try to make it look good in the debugger we will store the argument to
1465 1456 // RecursiveInterpreterActivation in the register argument save area. Without allocating
1466 1457 // extra space for the compiler this will overwrite locals in the local array of the
1467 1458 // interpreter.
1468 1459 // QQQ still needed with frameless adapters???
1469 1460
1470 1461 const int c2i_adjust_words = frame::memory_parameter_word_sp_offset - frame::callee_register_argument_save_area_sp_offset;
1471 1462
1472 1463 __ add(Gtmp, c2i_adjust_words*wordSize, Gtmp);
1473 1464 #endif // 1
1474 1465
1475 1466
1476 1467 __ sub(SP, Gtmp, SP); // just caller's frame for the additional space we need.
1477 1468 }
1478 1469
1479 1470 address InterpreterGenerator::generate_normal_entry(bool synchronized) {
1480 1471
1481 1472 // G5_method: methodOop
1482 1473 // G2_thread: thread (unused)
1483 1474 // Gargs: bottom of args (sender_sp)
1484 1475 // O5: sender's sp
1485 1476
1486 1477 // A single frame manager is plenty as we don't specialize for synchronized. We could and
1487 1478 // the code is pretty much ready. Would need to change the test below and for good measure
1488 1479 // modify generate_interpreter_state to only do the (pre) sync stuff stuff for synchronized
1489 1480 // routines. Not clear this is worth it yet.
1490 1481
1491 1482 if (interpreter_frame_manager) {
1492 1483 return interpreter_frame_manager;
1493 1484 }
1494 1485
1495 1486 __ bind(frame_manager_entry);
1496 1487
1497 1488 // the following temporary registers are used during frame creation
1498 1489 const Register Gtmp1 = G3_scratch;
1499 1490 // const Register Lmirror = L1; // native mirror (native calls only)
1500 1491
1501 1492 const Address constants (G5_method, 0, in_bytes(methodOopDesc::constants_offset()));
1502 1493 const Address access_flags (G5_method, 0, in_bytes(methodOopDesc::access_flags_offset()));
1503 1494 const Address size_of_parameters(G5_method, 0, in_bytes(methodOopDesc::size_of_parameters_offset()));
1504 1495 const Address max_stack (G5_method, 0, in_bytes(methodOopDesc::max_stack_offset()));
1505 1496 const Address size_of_locals (G5_method, 0, in_bytes(methodOopDesc::size_of_locals_offset()));
1506 1497
1507 1498 address entry_point = __ pc();
1508 1499 __ mov(G0, prevState); // no current activation
1509 1500
1510 1501
1511 1502 Label re_dispatch;
1512 1503
1513 1504 __ bind(re_dispatch);
1514 1505
1515 1506 // Interpreter needs to have locals completely contiguous. In order to do that
1516 1507 // We must adjust the caller's stack pointer for any locals beyond just the
1517 1508 // parameters
1518 1509 adjust_callers_stack(Gargs);
1519 1510
1520 1511 // O5_savedSP still contains sender's sp
1521 1512
1522 1513 // NEW FRAME
1523 1514
1524 1515 generate_compute_interpreter_state(Lstate, prevState, false);
1525 1516
1526 1517 // At this point a new interpreter frame and state object are created and initialized
1527 1518 // Lstate has the pointer to the new activation
1528 1519 // Any stack banging or limit check should already be done.
1529 1520
1530 1521 Label call_interpreter;
1531 1522
1532 1523 __ bind(call_interpreter);
1533 1524
1534 1525
1535 1526 #if 1
1536 1527 __ set(0xdead002, Lmirror);
1537 1528 __ set(0xdead002, L2_scratch);
1538 1529 __ set(0xdead003, L3_scratch);
1539 1530 __ set(0xdead004, L4_scratch);
1540 1531 __ set(0xdead005, Lscratch);
1541 1532 __ set(0xdead006, Lscratch2);
1542 1533 __ set(0xdead007, L7_scratch);
1543 1534
1544 1535 __ set(0xdeaf002, O2);
1545 1536 __ set(0xdeaf003, O3);
1546 1537 __ set(0xdeaf004, O4);
1547 1538 __ set(0xdeaf005, O5);
1548 1539 #endif
1549 1540
1550 1541 // Call interpreter (stack bang complete) enter here if message is
1551 1542 // set and we know stack size is valid
1552 1543
1553 1544 Label call_interpreter_2;
1554 1545
1555 1546 __ bind(call_interpreter_2);
1556 1547
1557 1548 #ifdef ASSERT
1558 1549 {
1559 1550 Label skip;
1560 1551 __ ld_ptr(STATE(_frame_bottom), G3_scratch);
1561 1552 __ cmp(G3_scratch, SP);
1562 1553 __ brx(Assembler::equal, false, Assembler::pt, skip);
1563 1554 __ delayed()->nop();
1564 1555 __ stop("SP not restored to frame bottom");
1565 1556 __ bind(skip);
1566 1557 }
1567 1558 #endif
1568 1559
1569 1560 VALIDATE_STATE(G3_scratch, 4);
1570 1561 __ set_last_Java_frame(SP, noreg);
1571 1562 __ mov(Lstate, O0); // (arg) pointer to current state
1572 1563
1573 1564 __ call(CAST_FROM_FN_PTR(address,
1574 1565 JvmtiExport::can_post_interpreter_events() ?
1575 1566 BytecodeInterpreter::runWithChecks
1576 1567 : BytecodeInterpreter::run),
1577 1568 relocInfo::runtime_call_type);
1578 1569
1579 1570 __ delayed()->nop();
1580 1571
1581 1572 __ ld_ptr(STATE(_thread), G2_thread);
1582 1573 __ reset_last_Java_frame();
1583 1574
1584 1575 // examine msg from interpreter to determine next action
1585 1576 __ ld_ptr(STATE(_thread), G2_thread); // restore G2_thread
1586 1577
1587 1578 __ ld(STATE(_msg), L1_scratch); // Get new message
1588 1579
1589 1580 Label call_method;
1590 1581 Label return_from_interpreted_method;
1591 1582 Label throw_exception;
1592 1583 Label do_OSR;
1593 1584 Label bad_msg;
1594 1585 Label resume_interpreter;
1595 1586
1596 1587 __ cmp(L1_scratch, (int)BytecodeInterpreter::call_method);
1597 1588 __ br(Assembler::equal, false, Assembler::pt, call_method);
1598 1589 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::return_from_method);
1599 1590 __ br(Assembler::equal, false, Assembler::pt, return_from_interpreted_method);
1600 1591 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::throwing_exception);
1601 1592 __ br(Assembler::equal, false, Assembler::pt, throw_exception);
1602 1593 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::do_osr);
1603 1594 __ br(Assembler::equal, false, Assembler::pt, do_OSR);
1604 1595 __ delayed()->cmp(L1_scratch, (int)BytecodeInterpreter::more_monitors);
1605 1596 __ br(Assembler::notEqual, false, Assembler::pt, bad_msg);
1606 1597
1607 1598 // Allocate more monitor space, shuffle expression stack....
1608 1599
1609 1600 generate_more_monitors();
1610 1601
1611 1602 // new monitor slot allocated, resume the interpreter.
1612 1603
1613 1604 __ set((int)BytecodeInterpreter::got_monitors, L1_scratch);
1614 1605 VALIDATE_STATE(G3_scratch, 5);
1615 1606 __ ba(false, call_interpreter);
1616 1607 __ delayed()->st(L1_scratch, STATE(_msg));
1617 1608
1618 1609 // uncommon trap needs to jump to here to enter the interpreter (re-execute current bytecode)
1619 1610 unctrap_frame_manager_entry = __ pc();
1620 1611
1621 1612 // QQQ what message do we send
1622 1613
1623 1614 __ ba(false, call_interpreter);
1624 1615 __ delayed()->ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1625 1616
1626 1617 //=============================================================================
1627 1618 // Returning from a compiled method into a deopted method. The bytecode at the
1628 1619 // bcp has completed. The result of the bytecode is in the native abi (the tosca
1629 1620 // for the template based interpreter). Any stack space that was used by the
1630 1621 // bytecode that has completed has been removed (e.g. parameters for an invoke)
1631 1622 // so all that we have to do is place any pending result on the expression stack
1632 1623 // and resume execution on the next bytecode.
1633 1624
1634 1625 generate_deopt_handling();
1635 1626
1636 1627 // ready to resume the interpreter
1637 1628
1638 1629 __ set((int)BytecodeInterpreter::deopt_resume, L1_scratch);
1639 1630 __ ba(false, call_interpreter);
1640 1631 __ delayed()->st(L1_scratch, STATE(_msg));
1641 1632
1642 1633 // Current frame has caught an exception we need to dispatch to the
1643 1634 // handler. We can get here because a native interpreter frame caught
1644 1635 // an exception in which case there is no handler and we must rethrow
1645 1636 // If it is a vanilla interpreted frame the we simply drop into the
1646 1637 // interpreter and let it do the lookup.
1647 1638
1648 1639 Interpreter::_rethrow_exception_entry = __ pc();
1649 1640
1650 1641 Label return_with_exception;
1651 1642 Label unwind_and_forward;
1652 1643
1653 1644 // O0: exception
1654 1645 // O7: throwing pc
1655 1646
1656 1647 // We want exception in the thread no matter what we ultimately decide about frame type.
1657 1648
1658 1649 Address exception_addr (G2_thread, 0, in_bytes(Thread::pending_exception_offset()));
1659 1650 __ verify_thread();
1660 1651 __ st_ptr(O0, exception_addr);
1661 1652
1662 1653 // get the methodOop
1663 1654 __ ld_ptr(STATE(_method), G5_method);
1664 1655
1665 1656 // if this current frame vanilla or native?
1666 1657
1667 1658 __ ld(access_flags, Gtmp1);
1668 1659 __ btst(JVM_ACC_NATIVE, Gtmp1);
1669 1660 __ br(Assembler::zero, false, Assembler::pt, return_with_exception); // vanilla interpreted frame handle directly
1670 1661 __ delayed()->nop();
1671 1662
1672 1663 // We drop thru to unwind a native interpreted frame with a pending exception
1673 1664 // We jump here for the initial interpreter frame with exception pending
1674 1665 // We unwind the current acivation and forward it to our caller.
1675 1666
1676 1667 __ bind(unwind_and_forward);
1677 1668
1678 1669 // Unwind frame and jump to forward exception. unwinding will place throwing pc in O7
1679 1670 // as expected by forward_exception.
1680 1671
1681 1672 __ restore(FP, G0, SP); // unwind interpreter state frame
1682 1673 __ br(Assembler::always, false, Assembler::pt, StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
1683 1674 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1684 1675
1685 1676 // Return point from a call which returns a result in the native abi
1686 1677 // (c1/c2/jni-native). This result must be processed onto the java
1687 1678 // expression stack.
1688 1679 //
1689 1680 // A pending exception may be present in which case there is no result present
1690 1681
1691 1682 address return_from_native_method = __ pc();
1692 1683
1693 1684 VALIDATE_STATE(G3_scratch, 6);
1694 1685
1695 1686 // Result if any is in native abi result (O0..O1/F0..F1). The java expression
1696 1687 // stack is in the state that the calling convention left it.
1697 1688 // Copy the result from native abi result and place it on java expression stack.
1698 1689
1699 1690 // Current interpreter state is present in Lstate
1700 1691
1701 1692 // Exception pending?
1702 1693
1703 1694 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1704 1695 __ ld_ptr(exception_addr, Lscratch); // get any pending exception
1705 1696 __ tst(Lscratch); // exception pending?
|
↓ open down ↓ |
319 lines elided |
↑ open up ↑ |
1706 1697 __ brx(Assembler::notZero, false, Assembler::pt, return_with_exception);
1707 1698 __ delayed()->nop();
1708 1699
1709 1700 // Process the native abi result to java expression stack
1710 1701
1711 1702 __ ld_ptr(STATE(_result._to_call._callee), L4_scratch); // called method
1712 1703 __ ld_ptr(STATE(_stack), L1_scratch); // get top of java expr stack
1713 1704 __ lduh(L4_scratch, in_bytes(methodOopDesc::size_of_parameters_offset()), L2_scratch); // get parameter size
1714 1705 __ sll(L2_scratch, LogBytesPerWord, L2_scratch ); // parameter size in bytes
1715 1706 __ add(L1_scratch, L2_scratch, L1_scratch); // stack destination for result
1716 - __ ld_ptr(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1707 + __ ld(L4_scratch, in_bytes(methodOopDesc::result_index_offset()), L3_scratch); // called method result type index
1717 1708
1718 1709 // tosca is really just native abi
1719 1710 __ set((intptr_t)CppInterpreter::_tosca_to_stack, L4_scratch);
1720 1711 __ sll(L3_scratch, LogBytesPerWord, L3_scratch);
1721 1712 __ ld_ptr(L4_scratch, L3_scratch, Lscratch); // get typed result converter address
1722 1713 __ jmpl(Lscratch, G0, O7); // and convert it
1723 1714 __ delayed()->nop();
1724 1715
1725 1716 // L1_scratch points to top of stack (prepushed)
1726 1717
1727 1718 __ ba(false, resume_interpreter);
1728 1719 __ delayed()->mov(L1_scratch, O1);
1729 1720
1730 1721 // An exception is being caught on return to a vanilla interpreter frame.
1731 1722 // Empty the stack and resume interpreter
1732 1723
1733 1724 __ bind(return_with_exception);
1734 1725
1735 1726 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1736 1727 __ ld_ptr(STATE(_stack_base), O1); // empty java expression stack
1737 1728 __ ba(false, resume_interpreter);
1738 1729 __ delayed()->sub(O1, wordSize, O1); // account for prepush
1739 1730
1740 1731 // Return from interpreted method we return result appropriate to the caller (i.e. "recursive"
1741 1732 // interpreter call, or native) and unwind this interpreter activation.
1742 1733 // All monitors should be unlocked.
1743 1734
1744 1735 __ bind(return_from_interpreted_method);
1745 1736
1746 1737 VALIDATE_STATE(G3_scratch, 7);
1747 1738
1748 1739 Label return_to_initial_caller;
1749 1740
|
↓ open down ↓ |
23 lines elided |
↑ open up ↑ |
1750 1741 // Interpreted result is on the top of the completed activation expression stack.
1751 1742 // We must return it to the top of the callers stack if caller was interpreted
1752 1743 // otherwise we convert to native abi result and return to call_stub/c1/c2
1753 1744 // The caller's expression stack was truncated by the call however the current activation
1754 1745 // has enough stuff on the stack that we have usable space there no matter what. The
1755 1746 // other thing that makes it easy is that the top of the caller's stack is stored in STATE(_locals)
1756 1747 // for the current activation
1757 1748
1758 1749 __ ld_ptr(STATE(_prev_link), L1_scratch);
1759 1750 __ ld_ptr(STATE(_method), L2_scratch); // get method just executed
1760 - __ ld_ptr(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1751 + __ ld(L2_scratch, in_bytes(methodOopDesc::result_index_offset()), L2_scratch);
1761 1752 __ tst(L1_scratch);
1762 1753 __ brx(Assembler::zero, false, Assembler::pt, return_to_initial_caller);
1763 1754 __ delayed()->sll(L2_scratch, LogBytesPerWord, L2_scratch);
1764 1755
1765 1756 // Copy result to callers java stack
1766 1757
1767 1758 __ set((intptr_t)CppInterpreter::_stack_to_stack, L4_scratch);
1768 1759 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1769 1760 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1770 1761 __ ld_ptr(STATE(_locals), O1); // stack destination
1771 1762
1772 1763 // O0 - will be source, O1 - will be destination (preserved)
1773 1764 __ jmpl(Lscratch, G0, O7); // and convert it
1774 1765 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1775 1766
1776 1767 // O1 == &locals[0]
1777 1768
1778 1769 // Result is now on caller's stack. Just unwind current activation and resume
1779 1770
1780 1771 Label unwind_recursive_activation;
1781 1772
1782 1773
1783 1774 __ bind(unwind_recursive_activation);
1784 1775
1785 1776 // O1 == &locals[0] (really callers stacktop) for activation now returning
1786 1777 // returning to interpreter method from "recursive" interpreter call
1787 1778 // result converter left O1 pointing to top of the( prepushed) java stack for method we are returning
1788 1779 // to. Now all we must do is unwind the state from the completed call
1789 1780
1790 1781 // Must restore stack
1791 1782 VALIDATE_STATE(G3_scratch, 8);
1792 1783
1793 1784 // Return to interpreter method after a method call (interpreted/native/c1/c2) has completed.
1794 1785 // Result if any is already on the caller's stack. All we must do now is remove the now dead
1795 1786 // frame and tell interpreter to resume.
1796 1787
1797 1788
1798 1789 __ mov(O1, I1); // pass back new stack top across activation
1799 1790 // POP FRAME HERE ==================================
1800 1791 __ restore(FP, G0, SP); // unwind interpreter state frame
1801 1792 __ ld_ptr(STATE(_frame_bottom), SP); // restore to full stack frame
1802 1793
1803 1794
1804 1795 // Resume the interpreter. The current frame contains the current interpreter
1805 1796 // state object.
1806 1797 //
1807 1798 // O1 == new java stack pointer
1808 1799
1809 1800 __ bind(resume_interpreter);
1810 1801 VALIDATE_STATE(G3_scratch, 10);
1811 1802
1812 1803 // A frame we have already used before so no need to bang stack so use call_interpreter_2 entry
1813 1804
1814 1805 __ set((int)BytecodeInterpreter::method_resume, L1_scratch);
1815 1806 __ st(L1_scratch, STATE(_msg));
1816 1807 __ ba(false, call_interpreter_2);
1817 1808 __ delayed()->st_ptr(O1, STATE(_stack));
1818 1809
1819 1810
1820 1811 // Fast accessor methods share this entry point.
1821 1812 // This works because frame manager is in the same codelet
1822 1813 // This can either be an entry via call_stub/c1/c2 or a recursive interpreter call
1823 1814 // we need to do a little register fixup here once we distinguish the two of them
1824 1815 if (UseFastAccessorMethods && !synchronized) {
1825 1816 // Call stub_return address still in O7
1826 1817 __ bind(fast_accessor_slow_entry_path);
1827 1818 __ set((intptr_t)return_from_native_method - 8, Gtmp1);
1828 1819 __ cmp(Gtmp1, O7); // returning to interpreter?
1829 1820 __ brx(Assembler::equal, true, Assembler::pt, re_dispatch); // yep
1830 1821 __ delayed()->nop();
1831 1822 __ ba(false, re_dispatch);
1832 1823 __ delayed()->mov(G0, prevState); // initial entry
1833 1824
1834 1825 }
1835 1826
1836 1827 // interpreter returning to native code (call_stub/c1/c2)
1837 1828 // convert result and unwind initial activation
1838 1829 // L2_scratch - scaled result type index
1839 1830
1840 1831 __ bind(return_to_initial_caller);
1841 1832
1842 1833 __ set((intptr_t)CppInterpreter::_stack_to_native_abi, L4_scratch);
1843 1834 __ ld_ptr(L4_scratch, L2_scratch, Lscratch); // get typed result converter address
1844 1835 __ ld_ptr(STATE(_stack), O0); // current top (prepushed)
1845 1836 __ jmpl(Lscratch, G0, O7); // and convert it
1846 1837 __ delayed()->add(O0, wordSize, O0); // get source (top of current expr stack)
1847 1838
1848 1839 Label unwind_initial_activation;
1849 1840 __ bind(unwind_initial_activation);
1850 1841
1851 1842 // RETURN TO CALL_STUB/C1/C2 code (result if any in I0..I1/(F0/..F1)
1852 1843 // we can return here with an exception that wasn't handled by interpreted code
1853 1844 // how does c1/c2 see it on return?
1854 1845
1855 1846 // compute resulting sp before/after args popped depending upon calling convention
1856 1847 // __ ld_ptr(STATE(_saved_sp), Gtmp1);
1857 1848 //
1858 1849 // POP FRAME HERE ==================================
1859 1850 __ restore(FP, G0, SP);
1860 1851 __ retl();
1861 1852 __ delayed()->mov(I5_savedSP->after_restore(), SP);
1862 1853
1863 1854 // OSR request, unwind the current frame and transfer to the OSR entry
1864 1855 // and enter OSR nmethod
1865 1856
1866 1857 __ bind(do_OSR);
1867 1858 Label remove_initial_frame;
1868 1859 __ ld_ptr(STATE(_prev_link), L1_scratch);
1869 1860 __ ld_ptr(STATE(_result._osr._osr_buf), G1_scratch);
1870 1861
1871 1862 // We are going to pop this frame. Is there another interpreter frame underneath
1872 1863 // it or is it callstub/compiled?
1873 1864
1874 1865 __ tst(L1_scratch);
1875 1866 __ brx(Assembler::zero, false, Assembler::pt, remove_initial_frame);
1876 1867 __ delayed()->ld_ptr(STATE(_result._osr._osr_entry), G3_scratch);
1877 1868
1878 1869 // Frame underneath is an interpreter frame simply unwind
1879 1870 // POP FRAME HERE ==================================
1880 1871 __ restore(FP, G0, SP); // unwind interpreter state frame
1881 1872 __ mov(I5_savedSP->after_restore(), SP);
1882 1873
1883 1874 // Since we are now calling native need to change our "return address" from the
1884 1875 // dummy RecursiveInterpreterActivation to a return from native
1885 1876
1886 1877 __ set((intptr_t)return_from_native_method - 8, O7);
1887 1878
1888 1879 __ jmpl(G3_scratch, G0, G0);
1889 1880 __ delayed()->mov(G1_scratch, O0);
1890 1881
1891 1882 __ bind(remove_initial_frame);
1892 1883
1893 1884 // POP FRAME HERE ==================================
1894 1885 __ restore(FP, G0, SP);
1895 1886 __ mov(I5_savedSP->after_restore(), SP);
1896 1887 __ jmpl(G3_scratch, G0, G0);
1897 1888 __ delayed()->mov(G1_scratch, O0);
1898 1889
1899 1890 // Call a new method. All we do is (temporarily) trim the expression stack
1900 1891 // push a return address to bring us back to here and leap to the new entry.
1901 1892 // At this point we have a topmost frame that was allocated by the frame manager
1902 1893 // which contains the current method interpreted state. We trim this frame
1903 1894 // of excess java expression stack entries and then recurse.
1904 1895
1905 1896 __ bind(call_method);
1906 1897
1907 1898 // stack points to next free location and not top element on expression stack
1908 1899 // method expects sp to be pointing to topmost element
1909 1900
1910 1901 __ ld_ptr(STATE(_thread), G2_thread);
1911 1902 __ ld_ptr(STATE(_result._to_call._callee), G5_method);
1912 1903
1913 1904
1914 1905 // SP already takes in to account the 2 extra words we use for slop
1915 1906 // when we call a "static long no_params()" method. So if
|
↓ open down ↓ |
145 lines elided |
↑ open up ↑ |
1916 1907 // we trim back sp by the amount of unused java expression stack
1917 1908 // there will be automagically the 2 extra words we need.
1918 1909 // We also have to worry about keeping SP aligned.
1919 1910
1920 1911 __ ld_ptr(STATE(_stack), Gargs);
1921 1912 __ ld_ptr(STATE(_stack_limit), L1_scratch);
1922 1913
1923 1914 // compute the unused java stack size
1924 1915 __ sub(Gargs, L1_scratch, L2_scratch); // compute unused space
1925 1916
1926 - // Round down the unused space to that stack is always aligned
1927 - // by making the unused space a multiple of the size of a long.
1917 + // Round down the unused space to that stack is always 16-byte aligned
1918 + // by making the unused space a multiple of the size of two longs. (vs)
1928 1919
1929 - __ and3(L2_scratch, -BytesPerLong, L2_scratch);
1920 + __ and3(L2_scratch, -2*BytesPerLong, L2_scratch);
1930 1921
1931 1922 // Now trim the stack
1932 1923 __ add(SP, L2_scratch, SP);
1933 1924
1934 1925
1935 1926 // Now point to the final argument (account for prepush)
1936 1927 __ add(Gargs, wordSize, Gargs);
1937 1928 #ifdef ASSERT
1938 1929 // Make sure we have space for the window
1939 1930 __ sub(Gargs, SP, L1_scratch);
1940 1931 __ cmp(L1_scratch, 16*wordSize);
1941 1932 {
1942 1933 Label skip;
1943 1934 __ brx(Assembler::greaterEqual, false, Assembler::pt, skip);
1944 1935 __ delayed()->nop();
1945 1936 __ stop("killed stack");
1946 1937 __ bind(skip);
1947 1938 }
1948 1939 #endif // ASSERT
1949 1940
1950 1941 // Create a new frame where we can store values that make it look like the interpreter
1951 1942 // really recursed.
1952 1943
1953 1944 // prepare to recurse or call specialized entry
1954 1945
1955 1946 // First link the registers we need
1956 1947
1957 1948 // make the pc look good in debugger
1958 1949 __ set(CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation), O7);
1959 1950 // argument too
1960 1951 __ mov(Lstate, I0);
1961 1952
1962 1953 // Record our sending SP
1963 1954 __ mov(SP, O5_savedSP);
1964 1955
1965 1956 __ ld_ptr(STATE(_result._to_call._callee_entry_point), L2_scratch);
1966 1957 __ set((intptr_t) entry_point, L1_scratch);
1967 1958 __ cmp(L1_scratch, L2_scratch);
1968 1959 __ brx(Assembler::equal, false, Assembler::pt, re_dispatch);
1969 1960 __ delayed()->mov(Lstate, prevState); // link activations
1970 1961
1971 1962 // method uses specialized entry, push a return so we look like call stub setup
1972 1963 // this path will handle fact that result is returned in registers and not
1973 1964 // on the java stack.
1974 1965
1975 1966 __ set((intptr_t)return_from_native_method - 8, O7);
1976 1967 __ jmpl(L2_scratch, G0, G0); // Do specialized entry
1977 1968 __ delayed()->nop();
1978 1969
1979 1970 //
1980 1971 // Bad Message from interpreter
1981 1972 //
1982 1973 __ bind(bad_msg);
1983 1974 __ stop("Bad message from interpreter");
1984 1975
1985 1976 // Interpreted method "returned" with an exception pass it on...
1986 1977 // Pass result, unwind activation and continue/return to interpreter/call_stub
1987 1978 // We handle result (if any) differently based on return to interpreter or call_stub
1988 1979
1989 1980 __ bind(throw_exception);
1990 1981 __ ld_ptr(STATE(_prev_link), L1_scratch);
1991 1982 __ tst(L1_scratch);
1992 1983 __ brx(Assembler::zero, false, Assembler::pt, unwind_and_forward);
1993 1984 __ delayed()->nop();
1994 1985
1995 1986 __ ld_ptr(STATE(_locals), O1); // get result of popping callee's args
1996 1987 __ ba(false, unwind_recursive_activation);
1997 1988 __ delayed()->nop();
1998 1989
1999 1990 interpreter_frame_manager = entry_point;
2000 1991 return entry_point;
2001 1992 }
2002 1993
2003 1994 InterpreterGenerator::InterpreterGenerator(StubQueue* code)
2004 1995 : CppInterpreterGenerator(code) {
2005 1996 generate_all(); // down here so it can be "virtual"
2006 1997 }
2007 1998
2008 1999
2009 2000 static int size_activation_helper(int callee_extra_locals, int max_stack, int monitor_size) {
2010 2001
2011 2002 // Figure out the size of an interpreter frame (in words) given that we have a fully allocated
2012 2003 // expression stack, the callee will have callee_extra_locals (so we can account for
2013 2004 // frame extension) and monitor_size for monitors. Basically we need to calculate
2014 2005 // this exactly like generate_fixed_frame/generate_compute_interpreter_state.
2015 2006 //
2016 2007 //
2017 2008 // The big complicating thing here is that we must ensure that the stack stays properly
2018 2009 // aligned. This would be even uglier if monitor size wasn't modulo what the stack
2019 2010 // needs to be aligned for). We are given that the sp (fp) is already aligned by
2020 2011 // the caller so we must ensure that it is properly aligned for our callee.
2021 2012 //
2022 2013 // Ths c++ interpreter always makes sure that we have a enough extra space on the
2023 2014 // stack at all times to deal with the "stack long no_params()" method issue. This
2024 2015 // is "slop_factor" here.
2025 2016 const int slop_factor = 2;
2026 2017
2027 2018 const int fixed_size = sizeof(BytecodeInterpreter)/wordSize + // interpreter state object
2028 2019 frame::memory_parameter_word_sp_offset; // register save area + param window
2029 2020 return (round_to(max_stack +
2030 2021 slop_factor +
2031 2022 fixed_size +
2032 2023 monitor_size +
2033 2024 (callee_extra_locals * Interpreter::stackElementWords()), WordsPerLong));
2034 2025
2035 2026 }
2036 2027
2037 2028 int AbstractInterpreter::size_top_interpreter_activation(methodOop method) {
2038 2029
2039 2030 // See call_stub code
2040 2031 int call_stub_size = round_to(7 + frame::memory_parameter_word_sp_offset,
2041 2032 WordsPerLong); // 7 + register save area
2042 2033
2043 2034 // Save space for one monitor to get into the interpreted method in case
2044 2035 // the method is synchronized
2045 2036 int monitor_size = method->is_synchronized() ?
2046 2037 1*frame::interpreter_frame_monitor_size() : 0;
2047 2038 return size_activation_helper(method->max_locals(), method->max_stack(),
2048 2039 monitor_size) + call_stub_size;
2049 2040 }
2050 2041
2051 2042 void BytecodeInterpreter::layout_interpreterState(interpreterState to_fill,
2052 2043 frame* caller,
2053 2044 frame* current,
2054 2045 methodOop method,
2055 2046 intptr_t* locals,
2056 2047 intptr_t* stack,
2057 2048 intptr_t* stack_base,
2058 2049 intptr_t* monitor_base,
2059 2050 intptr_t* frame_bottom,
2060 2051 bool is_top_frame
2061 2052 )
2062 2053 {
2063 2054 // What about any vtable?
2064 2055 //
2065 2056 to_fill->_thread = JavaThread::current();
2066 2057 // This gets filled in later but make it something recognizable for now
2067 2058 to_fill->_bcp = method->code_base();
2068 2059 to_fill->_locals = locals;
2069 2060 to_fill->_constants = method->constants()->cache();
2070 2061 to_fill->_method = method;
2071 2062 to_fill->_mdx = NULL;
2072 2063 to_fill->_stack = stack;
2073 2064 if (is_top_frame && JavaThread::current()->popframe_forcing_deopt_reexecution() ) {
2074 2065 to_fill->_msg = deopt_resume2;
2075 2066 } else {
2076 2067 to_fill->_msg = method_resume;
2077 2068 }
2078 2069 to_fill->_result._to_call._bcp_advance = 0;
2079 2070 to_fill->_result._to_call._callee_entry_point = NULL; // doesn't matter to anyone
2080 2071 to_fill->_result._to_call._callee = NULL; // doesn't matter to anyone
2081 2072 to_fill->_prev_link = NULL;
2082 2073
2083 2074 // Fill in the registers for the frame
2084 2075
2085 2076 // Need to install _sender_sp. Actually not too hard in C++!
2086 2077 // When the skeletal frames are layed out we fill in a value
2087 2078 // for _sender_sp. That value is only correct for the oldest
2088 2079 // skeletal frame constructed (because there is only a single
2089 2080 // entry for "caller_adjustment". While the skeletal frames
2090 2081 // exist that is good enough. We correct that calculation
2091 2082 // here and get all the frames correct.
2092 2083
2093 2084 // to_fill->_sender_sp = locals - (method->size_of_parameters() - 1);
2094 2085
2095 2086 *current->register_addr(Lstate) = (intptr_t) to_fill;
2096 2087 // skeletal already places a useful value here and this doesn't account
2097 2088 // for alignment so don't bother.
2098 2089 // *current->register_addr(I5_savedSP) = (intptr_t) locals - (method->size_of_parameters() - 1);
2099 2090
2100 2091 if (caller->is_interpreted_frame()) {
2101 2092 interpreterState prev = caller->get_interpreterState();
2102 2093 to_fill->_prev_link = prev;
2103 2094 // Make the prev callee look proper
2104 2095 prev->_result._to_call._callee = method;
2105 2096 if (*prev->_bcp == Bytecodes::_invokeinterface) {
2106 2097 prev->_result._to_call._bcp_advance = 5;
2107 2098 } else {
2108 2099 prev->_result._to_call._bcp_advance = 3;
2109 2100 }
2110 2101 }
2111 2102 to_fill->_oop_temp = NULL;
2112 2103 to_fill->_stack_base = stack_base;
2113 2104 // Need +1 here because stack_base points to the word just above the first expr stack entry
2114 2105 // and stack_limit is supposed to point to the word just below the last expr stack entry.
2115 2106 // See generate_compute_interpreter_state.
2116 2107 to_fill->_stack_limit = stack_base - (method->max_stack() + 1);
2117 2108 to_fill->_monitor_base = (BasicObjectLock*) monitor_base;
2118 2109
2119 2110 // sparc specific
2120 2111 to_fill->_frame_bottom = frame_bottom;
2121 2112 to_fill->_self_link = to_fill;
2122 2113 #ifdef ASSERT
2123 2114 to_fill->_native_fresult = 123456.789;
2124 2115 to_fill->_native_lresult = CONST64(0xdeadcafedeafcafe);
2125 2116 #endif
2126 2117 }
2127 2118
2128 2119 void BytecodeInterpreter::pd_layout_interpreterState(interpreterState istate, address last_Java_pc, intptr_t* last_Java_fp) {
2129 2120 istate->_last_Java_pc = (intptr_t*) last_Java_pc;
2130 2121 }
2131 2122
2132 2123
2133 2124 int AbstractInterpreter::layout_activation(methodOop method,
2134 2125 int tempcount, // Number of slots on java expression stack in use
2135 2126 int popframe_extra_args,
2136 2127 int moncount, // Number of active monitors
2137 2128 int callee_param_size,
2138 2129 int callee_locals_size,
2139 2130 frame* caller,
2140 2131 frame* interpreter_frame,
2141 2132 bool is_top_frame) {
2142 2133
2143 2134 assert(popframe_extra_args == 0, "NEED TO FIX");
2144 2135 // NOTE this code must exactly mimic what InterpreterGenerator::generate_compute_interpreter_state()
2145 2136 // does as far as allocating an interpreter frame.
2146 2137 // If interpreter_frame!=NULL, set up the method, locals, and monitors.
2147 2138 // The frame interpreter_frame, if not NULL, is guaranteed to be the right size,
2148 2139 // as determined by a previous call to this method.
2149 2140 // It is also guaranteed to be walkable even though it is in a skeletal state
2150 2141 // NOTE: return size is in words not bytes
2151 2142 // NOTE: tempcount is the current size of the java expression stack. For top most
2152 2143 // frames we will allocate a full sized expression stack and not the curback
2153 2144 // version that non-top frames have.
2154 2145
2155 2146 // Calculate the amount our frame will be adjust by the callee. For top frame
2156 2147 // this is zero.
2157 2148
2158 2149 // NOTE: ia64 seems to do this wrong (or at least backwards) in that it
2159 2150 // calculates the extra locals based on itself. Not what the callee does
2160 2151 // to it. So it ignores last_frame_adjust value. Seems suspicious as far
2161 2152 // as getting sender_sp correct.
2162 2153
2163 2154 int extra_locals_size = callee_locals_size - callee_param_size;
2164 2155 int monitor_size = (sizeof(BasicObjectLock) * moncount) / wordSize;
2165 2156 int full_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2166 2157 int short_frame_words = size_activation_helper(extra_locals_size, method->max_stack(), monitor_size);
2167 2158 int frame_words = is_top_frame ? full_frame_words : short_frame_words;
2168 2159
|
↓ open down ↓ |
229 lines elided |
↑ open up ↑ |
2169 2160
2170 2161 /*
2171 2162 if we actually have a frame to layout we must now fill in all the pieces. This means both
2172 2163 the interpreterState and the registers.
2173 2164 */
2174 2165 if (interpreter_frame != NULL) {
2175 2166
2176 2167 // MUCHO HACK
2177 2168
2178 2169 intptr_t* frame_bottom = interpreter_frame->sp() - (full_frame_words - frame_words);
2170 + // 'interpreter_frame->sp()' is unbiased while 'frame_bottom' must be a biased value in 64bit mode. (vs)
2171 + assert(((intptr_t)frame_bottom & 0xf) == 0, "SP biased in layout_activation");
2172 + frame_bottom = (intptr_t*)((intptr_t)frame_bottom - STACK_BIAS);
2179 2173
2180 2174 /* Now fillin the interpreterState object */
2181 2175
2182 2176 interpreterState cur_state = (interpreterState) ((intptr_t)interpreter_frame->fp() - sizeof(BytecodeInterpreter));
2183 2177
2184 2178
2185 2179 intptr_t* locals;
2186 2180
2187 2181 // Calculate the postion of locals[0]. This is painful because of
2188 2182 // stack alignment (same as ia64). The problem is that we can
2189 2183 // not compute the location of locals from fp(). fp() will account
2190 2184 // for the extra locals but it also accounts for aligning the stack
2191 2185 // and we can't determine if the locals[0] was misaligned but max_locals
2192 2186 // was enough to have the
2193 2187 // calculate postion of locals. fp already accounts for extra locals.
2194 2188 // +2 for the static long no_params() issue.
2195 2189
2196 2190 if (caller->is_interpreted_frame()) {
2197 2191 // locals must agree with the caller because it will be used to set the
2198 2192 // caller's tos when we return.
2199 2193 interpreterState prev = caller->get_interpreterState();
2200 2194 // stack() is prepushed.
2201 2195 locals = prev->stack() + method->size_of_parameters();
2202 2196 } else {
2203 2197 // Lay out locals block in the caller adjacent to the register window save area.
2204 2198 //
2205 2199 // Compiled frames do not allocate a varargs area which is why this if
2206 2200 // statement is needed.
2207 2201 //
2208 2202 intptr_t* fp = interpreter_frame->fp();
2209 2203 int local_words = method->max_locals() * Interpreter::stackElementWords();
2210 2204
2211 2205 if (caller->is_compiled_frame()) {
2212 2206 locals = fp + frame::register_save_words + local_words - 1;
2213 2207 } else {
2214 2208 locals = fp + frame::memory_parameter_word_sp_offset + local_words - 1;
2215 2209 }
2216 2210
2217 2211 }
2218 2212 // END MUCHO HACK
2219 2213
2220 2214 intptr_t* monitor_base = (intptr_t*) cur_state;
2221 2215 intptr_t* stack_base = monitor_base - monitor_size;
2222 2216 /* +1 because stack is always prepushed */
2223 2217 intptr_t* stack = stack_base - (tempcount + 1);
2224 2218
2225 2219
2226 2220 BytecodeInterpreter::layout_interpreterState(cur_state,
2227 2221 caller,
2228 2222 interpreter_frame,
2229 2223 method,
2230 2224 locals,
2231 2225 stack,
2232 2226 stack_base,
2233 2227 monitor_base,
2234 2228 frame_bottom,
2235 2229 is_top_frame);
2236 2230
2237 2231 BytecodeInterpreter::pd_layout_interpreterState(cur_state, interpreter_return_address, interpreter_frame->fp());
2238 2232
2239 2233 }
2240 2234 return frame_words;
2241 2235 }
2242 2236
2243 2237 #endif // CC_INTERP
|
↓ open down ↓ |
55 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX