1 /*
2 * Copyright (c) 2003, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2021, Red Hat Inc. All rights reserved.
4 * Copyright (c) 2021, Azul Systems, Inc. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27 #include "precompiled.hpp"
28 #include "asm/macroAssembler.hpp"
29 #include "asm/macroAssembler.inline.hpp"
30 #include "code/codeCache.hpp"
31 #include "code/compiledIC.hpp"
32 #include "code/debugInfoRec.hpp"
33 #include "code/icBuffer.hpp"
34 #include "code/vtableStubs.hpp"
35 #include "compiler/oopMap.hpp"
36 #include "gc/shared/barrierSetAssembler.hpp"
37 #include "interpreter/interpreter.hpp"
38 #include "interpreter/interp_masm.hpp"
39 #include "logging/log.hpp"
40 #include "memory/resourceArea.hpp"
41 #include "nativeInst_aarch64.hpp"
42 #include "oops/compiledICHolder.hpp"
43 #include "oops/klass.inline.hpp"
44 #include "oops/method.inline.hpp"
45 #include "prims/methodHandles.hpp"
46 #include "runtime/continuation.hpp"
47 #include "runtime/continuationEntry.inline.hpp"
48 #include "runtime/globals.hpp"
49 #include "runtime/jniHandles.hpp"
50 #include "runtime/safepointMechanism.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "runtime/signature.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "runtime/vframeArray.hpp"
55 #include "utilities/align.hpp"
56 #include "utilities/formatBuffer.hpp"
57 #include "vmreg_aarch64.inline.hpp"
58 #ifdef COMPILER1
59 #include "c1/c1_Runtime1.hpp"
60 #endif
61 #ifdef COMPILER2
62 #include "adfiles/ad_aarch64.hpp"
63 #include "opto/runtime.hpp"
64 #endif
65 #if INCLUDE_JVMCI
66 #include "jvmci/jvmciJavaClasses.hpp"
67 #endif
68
69 #define __ masm->
70
71 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
72
73 class SimpleRuntimeFrame {
74
75 public:
76
77 // Most of the runtime stubs have this simple frame layout.
78 // This class exists to make the layout shared in one place.
79 // Offsets are for compiler stack slots, which are jints.
80 enum layout {
81 // The frame sender code expects that rbp will be in the "natural" place and
82 // will override any oopMap setting for it. We must therefore force the layout
83 // so that it agrees with the frame sender code.
84 // we don't expect any arg reg save area so aarch64 asserts that
85 // frame::arg_reg_save_area_bytes == 0
86 rfp_off = 0,
87 rfp_off2,
88 return_off, return_off2,
89 framesize
90 };
91 };
92
93 // FIXME -- this is used by C1
94 class RegisterSaver {
95 const bool _save_vectors;
96 public:
97 RegisterSaver(bool save_vectors) : _save_vectors(save_vectors) {}
98
99 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
100 void restore_live_registers(MacroAssembler* masm);
101
102 // Offsets into the register save area
103 // Used by deoptimization when it is managing result register
104 // values on its own
105
106 int reg_offset_in_bytes(Register r);
107 int r0_offset_in_bytes() { return reg_offset_in_bytes(r0); }
108 int rscratch1_offset_in_bytes() { return reg_offset_in_bytes(rscratch1); }
109 int v0_offset_in_bytes();
110
111 // Total stack size in bytes for saving sve predicate registers.
112 int total_sve_predicate_in_bytes();
113
114 // Capture info about frame layout
115 // Note this is only correct when not saving full vectors.
116 enum layout {
117 fpu_state_off = 0,
118 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1,
119 // The frame sender code expects that rfp will be in
120 // the "natural" place and will override any oopMap
121 // setting for it. We must therefore force the layout
122 // so that it agrees with the frame sender code.
123 r0_off = fpu_state_off + FPUStateSizeInWords,
124 rfp_off = r0_off + (Register::number_of_registers - 2) * Register::max_slots_per_register,
125 return_off = rfp_off + Register::max_slots_per_register, // slot for return address
126 reg_save_size = return_off + Register::max_slots_per_register};
127
128 };
129
130 int RegisterSaver::reg_offset_in_bytes(Register r) {
131 // The integer registers are located above the floating point
132 // registers in the stack frame pushed by save_live_registers() so the
133 // offset depends on whether we are saving full vectors, and whether
134 // those vectors are NEON or SVE.
135
136 int slots_per_vect = FloatRegister::save_slots_per_register;
137
138 #if COMPILER2_OR_JVMCI
139 if (_save_vectors) {
140 slots_per_vect = FloatRegister::slots_per_neon_register;
141
142 #ifdef COMPILER2
143 if (Matcher::supports_scalable_vector()) {
144 slots_per_vect = Matcher::scalable_vector_reg_size(T_FLOAT);
145 }
146 #endif
147 }
148 #endif
149
150 int r0_offset = v0_offset_in_bytes() + (slots_per_vect * FloatRegister::number_of_registers) * BytesPerInt;
151 return r0_offset + r->encoding() * wordSize;
152 }
153
154 int RegisterSaver::v0_offset_in_bytes() {
155 // The floating point registers are located above the predicate registers if
156 // they are present in the stack frame pushed by save_live_registers(). So the
157 // offset depends on the saved total predicate vectors in the stack frame.
158 return (total_sve_predicate_in_bytes() / VMRegImpl::stack_slot_size) * BytesPerInt;
159 }
160
161 int RegisterSaver::total_sve_predicate_in_bytes() {
162 #ifdef COMPILER2
163 if (_save_vectors && Matcher::supports_scalable_vector()) {
164 return (Matcher::scalable_vector_reg_size(T_BYTE) >> LogBitsPerByte) *
165 PRegister::number_of_registers;
166 }
167 #endif
168 return 0;
169 }
170
171 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
172 bool use_sve = false;
173 int sve_vector_size_in_bytes = 0;
174 int sve_vector_size_in_slots = 0;
175 int sve_predicate_size_in_slots = 0;
176 int total_predicate_in_bytes = total_sve_predicate_in_bytes();
177 int total_predicate_in_slots = total_predicate_in_bytes / VMRegImpl::stack_slot_size;
178
179 #ifdef COMPILER2
180 use_sve = Matcher::supports_scalable_vector();
181 if (use_sve) {
182 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
183 sve_vector_size_in_slots = Matcher::scalable_vector_reg_size(T_FLOAT);
184 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
185 }
186 #endif
187
188 #if COMPILER2_OR_JVMCI
189 if (_save_vectors) {
190 int extra_save_slots_per_register = 0;
191 // Save upper half of vector registers
192 if (use_sve) {
193 extra_save_slots_per_register = sve_vector_size_in_slots - FloatRegister::save_slots_per_register;
194 } else {
195 extra_save_slots_per_register = FloatRegister::extra_save_slots_per_neon_register;
196 }
197 int extra_vector_bytes = extra_save_slots_per_register *
198 VMRegImpl::stack_slot_size *
199 FloatRegister::number_of_registers;
200 additional_frame_words += ((extra_vector_bytes + total_predicate_in_bytes) / wordSize);
201 }
202 #else
203 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
204 #endif
205
206 int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
207 reg_save_size * BytesPerInt, 16);
208 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
209 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
210 // The caller will allocate additional_frame_words
211 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
212 // CodeBlob frame size is in words.
213 int frame_size_in_words = frame_size_in_bytes / wordSize;
214 *total_frame_words = frame_size_in_words;
215
216 // Save Integer and Float registers.
217 __ enter();
218 __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
219
220 // Set an oopmap for the call site. This oopmap will map all
221 // oop-registers and debug-info registers as callee-saved. This
222 // will allow deoptimization at this safepoint to find all possible
223 // debug-info recordings, as well as let GC find all oops.
224
225 OopMapSet *oop_maps = new OopMapSet();
226 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
227
228 for (int i = 0; i < Register::number_of_registers; i++) {
229 Register r = as_Register(i);
230 if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
231 // SP offsets are in 4-byte words.
232 // Register slots are 8 bytes wide, 32 floating-point registers.
233 int sp_offset = Register::max_slots_per_register * i +
234 FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
235 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
236 }
237 }
238
239 for (int i = 0; i < FloatRegister::number_of_registers; i++) {
240 FloatRegister r = as_FloatRegister(i);
241 int sp_offset = 0;
242 if (_save_vectors) {
243 sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
244 (FloatRegister::slots_per_neon_register * i);
245 } else {
246 sp_offset = FloatRegister::save_slots_per_register * i;
247 }
248 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), r->as_VMReg());
249 }
250
251 return oop_map;
252 }
253
254 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
255 #ifdef COMPILER2
256 __ pop_CPU_state(_save_vectors, Matcher::supports_scalable_vector(),
257 Matcher::scalable_vector_reg_size(T_BYTE), total_sve_predicate_in_bytes());
258 #else
259 #if !INCLUDE_JVMCI
260 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
261 #endif
262 __ pop_CPU_state(_save_vectors);
263 #endif
264 __ ldp(rfp, lr, Address(__ post(sp, 2 * wordSize)));
265 __ authenticate_return_address();
266 }
267
268 // Is vector's size (in bytes) bigger than a size saved by default?
269 // 8 bytes vector registers are saved by default on AArch64.
270 // The SVE supported min vector size is 8 bytes and we need to save
271 // predicate registers when the vector size is 8 bytes as well.
272 bool SharedRuntime::is_wide_vector(int size) {
273 return size > 8 || (UseSVE > 0 && size >= 8);
274 }
275
276 // ---------------------------------------------------------------------------
277 // Read the array of BasicTypes from a signature, and compute where the
278 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
279 // quantities. Values less than VMRegImpl::stack0 are registers, those above
280 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
281 // as framesizes are fixed.
282 // VMRegImpl::stack0 refers to the first slot 0(sp).
283 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
284 // Register up to Register::number_of_registers are the 64-bit
285 // integer registers.
286
287 // Note: the INPUTS in sig_bt are in units of Java argument words,
288 // which are 64-bit. The OUTPUTS are in 32-bit units.
289
290 // The Java calling convention is a "shifted" version of the C ABI.
291 // By skipping the first C ABI register we can call non-static jni
292 // methods with small numbers of arguments without having to shuffle
293 // the arguments at all. Since we control the java ABI we ought to at
294 // least get some advantage out of it.
295
296 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
297 VMRegPair *regs,
298 int total_args_passed) {
299
300 // Create the mapping between argument positions and
301 // registers.
302 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
303 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7
304 };
305 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
306 j_farg0, j_farg1, j_farg2, j_farg3,
307 j_farg4, j_farg5, j_farg6, j_farg7
308 };
309
310
311 uint int_args = 0;
312 uint fp_args = 0;
313 uint stk_args = 0; // inc by 2 each time
314
315 for (int i = 0; i < total_args_passed; i++) {
316 switch (sig_bt[i]) {
317 case T_BOOLEAN:
318 case T_CHAR:
319 case T_BYTE:
320 case T_SHORT:
321 case T_INT:
322 if (int_args < Argument::n_int_register_parameters_j) {
323 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
324 } else {
325 regs[i].set1(VMRegImpl::stack2reg(stk_args));
326 stk_args += 2;
327 }
328 break;
329 case T_VOID:
330 // halves of T_LONG or T_DOUBLE
331 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
332 regs[i].set_bad();
333 break;
334 case T_LONG:
335 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
336 // fall through
337 case T_OBJECT:
338 case T_ARRAY:
339 case T_ADDRESS:
340 if (int_args < Argument::n_int_register_parameters_j) {
341 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
342 } else {
343 regs[i].set2(VMRegImpl::stack2reg(stk_args));
344 stk_args += 2;
345 }
346 break;
347 case T_FLOAT:
348 if (fp_args < Argument::n_float_register_parameters_j) {
349 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
350 } else {
351 regs[i].set1(VMRegImpl::stack2reg(stk_args));
352 stk_args += 2;
353 }
354 break;
355 case T_DOUBLE:
356 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
357 if (fp_args < Argument::n_float_register_parameters_j) {
358 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
359 } else {
360 regs[i].set2(VMRegImpl::stack2reg(stk_args));
361 stk_args += 2;
362 }
363 break;
364 default:
365 ShouldNotReachHere();
366 break;
367 }
368 }
369
370 return align_up(stk_args, 2);
371 }
372
373 // Patch the callers callsite with entry to compiled code if it exists.
374 static void patch_callers_callsite(MacroAssembler *masm) {
375 Label L;
376 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
377 __ cbz(rscratch1, L);
378
379 __ enter();
380 __ push_CPU_state();
381
382 // VM needs caller's callsite
383 // VM needs target method
384 // This needs to be a long call since we will relocate this adapter to
385 // the codeBuffer and it may not reach
386
387 #ifndef PRODUCT
388 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
389 #endif
390
391 __ mov(c_rarg0, rmethod);
392 __ mov(c_rarg1, lr);
393 __ authenticate_return_address(c_rarg1, rscratch1);
394 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
395 __ blr(rscratch1);
396
397 // Explicit isb required because fixup_callers_callsite may change the code
398 // stream.
399 __ safepoint_isb();
400
401 __ pop_CPU_state();
402 // restore sp
403 __ leave();
404 __ bind(L);
405 }
406
407 static void gen_c2i_adapter(MacroAssembler *masm,
408 int total_args_passed,
409 int comp_args_on_stack,
410 const BasicType *sig_bt,
411 const VMRegPair *regs,
412 Label& skip_fixup) {
413 // Before we get into the guts of the C2I adapter, see if we should be here
414 // at all. We've come from compiled code and are attempting to jump to the
415 // interpreter, which means the caller made a static call to get here
416 // (vcalls always get a compiled target if there is one). Check for a
417 // compiled target. If there is one, we need to patch the caller's call.
418 patch_callers_callsite(masm);
419
420 __ bind(skip_fixup);
421
422 int words_pushed = 0;
423
424 // Since all args are passed on the stack, total_args_passed *
425 // Interpreter::stackElementSize is the space we need.
426
427 int extraspace = total_args_passed * Interpreter::stackElementSize;
428
429 __ mov(r19_sender_sp, sp);
430
431 // stack is aligned, keep it that way
432 extraspace = align_up(extraspace, 2*wordSize);
433
434 if (extraspace)
435 __ sub(sp, sp, extraspace);
436
437 // Now write the args into the outgoing interpreter space
438 for (int i = 0; i < total_args_passed; i++) {
439 if (sig_bt[i] == T_VOID) {
440 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
441 continue;
442 }
443
444 // offset to start parameters
445 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
446 int next_off = st_off - Interpreter::stackElementSize;
447
448 // Say 4 args:
449 // i st_off
450 // 0 32 T_LONG
451 // 1 24 T_VOID
452 // 2 16 T_OBJECT
453 // 3 8 T_BOOL
454 // - 0 return address
455 //
456 // However to make thing extra confusing. Because we can fit a Java long/double in
457 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
458 // leaves one slot empty and only stores to a single slot. In this case the
459 // slot that is occupied is the T_VOID slot. See I said it was confusing.
460
461 VMReg r_1 = regs[i].first();
462 VMReg r_2 = regs[i].second();
463 if (!r_1->is_valid()) {
464 assert(!r_2->is_valid(), "");
465 continue;
466 }
467 if (r_1->is_stack()) {
468 // memory to memory use rscratch1
469 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
470 + extraspace
471 + words_pushed * wordSize);
472 if (!r_2->is_valid()) {
473 // sign extend??
474 __ ldrw(rscratch1, Address(sp, ld_off));
475 __ str(rscratch1, Address(sp, st_off));
476
477 } else {
478
479 __ ldr(rscratch1, Address(sp, ld_off));
480
481 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
482 // T_DOUBLE and T_LONG use two slots in the interpreter
483 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
484 // ld_off == LSW, ld_off+wordSize == MSW
485 // st_off == MSW, next_off == LSW
486 __ str(rscratch1, Address(sp, next_off));
487 #ifdef ASSERT
488 // Overwrite the unused slot with known junk
489 __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaaaull);
490 __ str(rscratch1, Address(sp, st_off));
491 #endif /* ASSERT */
492 } else {
493 __ str(rscratch1, Address(sp, st_off));
494 }
495 }
496 } else if (r_1->is_Register()) {
497 Register r = r_1->as_Register();
498 if (!r_2->is_valid()) {
499 // must be only an int (or less ) so move only 32bits to slot
500 // why not sign extend??
501 __ str(r, Address(sp, st_off));
502 } else {
503 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
504 // T_DOUBLE and T_LONG use two slots in the interpreter
505 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
506 // jlong/double in gpr
507 #ifdef ASSERT
508 // Overwrite the unused slot with known junk
509 __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaabull);
510 __ str(rscratch1, Address(sp, st_off));
511 #endif /* ASSERT */
512 __ str(r, Address(sp, next_off));
513 } else {
514 __ str(r, Address(sp, st_off));
515 }
516 }
517 } else {
518 assert(r_1->is_FloatRegister(), "");
519 if (!r_2->is_valid()) {
520 // only a float use just part of the slot
521 __ strs(r_1->as_FloatRegister(), Address(sp, st_off));
522 } else {
523 #ifdef ASSERT
524 // Overwrite the unused slot with known junk
525 __ mov(rscratch1, (uint64_t)0xdeadffffdeadaaacull);
526 __ str(rscratch1, Address(sp, st_off));
527 #endif /* ASSERT */
528 __ strd(r_1->as_FloatRegister(), Address(sp, next_off));
529 }
530 }
531 }
532
533 __ mov(esp, sp); // Interp expects args on caller's expression stack
534
535 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
536 __ br(rscratch1);
537 }
538
539
540 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
541 int total_args_passed,
542 int comp_args_on_stack,
543 const BasicType *sig_bt,
544 const VMRegPair *regs) {
545
546 // Note: r19_sender_sp contains the senderSP on entry. We must
547 // preserve it since we may do a i2c -> c2i transition if we lose a
548 // race where compiled code goes non-entrant while we get args
549 // ready.
550
551 // Adapters are frameless.
552
553 // An i2c adapter is frameless because the *caller* frame, which is
554 // interpreted, routinely repairs its own esp (from
555 // interpreter_frame_last_sp), even if a callee has modified the
556 // stack pointer. It also recalculates and aligns sp.
557
558 // A c2i adapter is frameless because the *callee* frame, which is
559 // interpreted, routinely repairs its caller's sp (from sender_sp,
560 // which is set up via the senderSP register).
561
562 // In other words, if *either* the caller or callee is interpreted, we can
563 // get the stack pointer repaired after a call.
564
565 // This is why c2i and i2c adapters cannot be indefinitely composed.
566 // In particular, if a c2i adapter were to somehow call an i2c adapter,
567 // both caller and callee would be compiled methods, and neither would
568 // clean up the stack pointer changes performed by the two adapters.
569 // If this happens, control eventually transfers back to the compiled
570 // caller, but with an uncorrected stack, causing delayed havoc.
571
572 if (VerifyAdapterCalls &&
573 (Interpreter::code() != nullptr || StubRoutines::final_stubs_code() != nullptr)) {
574 #if 0
575 // So, let's test for cascading c2i/i2c adapters right now.
576 // assert(Interpreter::contains($return_addr) ||
577 // StubRoutines::contains($return_addr),
578 // "i2c adapter must return to an interpreter frame");
579 __ block_comment("verify_i2c { ");
580 Label L_ok;
581 if (Interpreter::code() != nullptr) {
582 range_check(masm, rax, r11,
583 Interpreter::code()->code_start(), Interpreter::code()->code_end(),
584 L_ok);
585 }
586 if (StubRoutines::initial_stubs_code() != nullptr) {
587 range_check(masm, rax, r11,
588 StubRoutines::initial_stubs_code()->code_begin(),
589 StubRoutines::initial_stubs_code()->code_end(),
590 L_ok);
591 }
592 if (StubRoutines::final_stubs_code() != nullptr) {
593 range_check(masm, rax, r11,
594 StubRoutines::final_stubs_code()->code_begin(),
595 StubRoutines::final_stubs_code()->code_end(),
596 L_ok);
597 }
598 const char* msg = "i2c adapter must return to an interpreter frame";
599 __ block_comment(msg);
600 __ stop(msg);
601 __ bind(L_ok);
602 __ block_comment("} verify_i2ce ");
603 #endif
604 }
605
606 // Cut-out for having no stack args.
607 int comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
608 if (comp_args_on_stack) {
609 __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
610 __ andr(sp, rscratch1, -16);
611 }
612
613 // Will jump to the compiled code just as if compiled code was doing it.
614 // Pre-load the register-jump target early, to schedule it better.
615 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_offset())));
616
617 #if INCLUDE_JVMCI
618 if (EnableJVMCI) {
619 // check if this call should be routed towards a specific entry point
620 __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
621 Label no_alternative_target;
622 __ cbz(rscratch2, no_alternative_target);
623 __ mov(rscratch1, rscratch2);
624 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
625 __ bind(no_alternative_target);
626 }
627 #endif // INCLUDE_JVMCI
628
629 // Now generate the shuffle code.
630 for (int i = 0; i < total_args_passed; i++) {
631 if (sig_bt[i] == T_VOID) {
632 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
633 continue;
634 }
635
636 // Pick up 0, 1 or 2 words from SP+offset.
637
638 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
639 "scrambled load targets?");
640 // Load in argument order going down.
641 int ld_off = (total_args_passed - i - 1)*Interpreter::stackElementSize;
642 // Point to interpreter value (vs. tag)
643 int next_off = ld_off - Interpreter::stackElementSize;
644 //
645 //
646 //
647 VMReg r_1 = regs[i].first();
648 VMReg r_2 = regs[i].second();
649 if (!r_1->is_valid()) {
650 assert(!r_2->is_valid(), "");
651 continue;
652 }
653 if (r_1->is_stack()) {
654 // Convert stack slot to an SP offset (+ wordSize to account for return address )
655 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size;
656 if (!r_2->is_valid()) {
657 // sign extend???
658 __ ldrsw(rscratch2, Address(esp, ld_off));
659 __ str(rscratch2, Address(sp, st_off));
660 } else {
661 //
662 // We are using two optoregs. This can be either T_OBJECT,
663 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
664 // two slots but only uses one for thr T_LONG or T_DOUBLE case
665 // So we must adjust where to pick up the data to match the
666 // interpreter.
667 //
668 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
669 // are accessed as negative so LSW is at LOW address
670
671 // ld_off is MSW so get LSW
672 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
673 next_off : ld_off;
674 __ ldr(rscratch2, Address(esp, offset));
675 // st_off is LSW (i.e. reg.first())
676 __ str(rscratch2, Address(sp, st_off));
677 }
678 } else if (r_1->is_Register()) { // Register argument
679 Register r = r_1->as_Register();
680 if (r_2->is_valid()) {
681 //
682 // We are using two VMRegs. This can be either T_OBJECT,
683 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
684 // two slots but only uses one for thr T_LONG or T_DOUBLE case
685 // So we must adjust where to pick up the data to match the
686 // interpreter.
687
688 const int offset = (sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
689 next_off : ld_off;
690
691 // this can be a misaligned move
692 __ ldr(r, Address(esp, offset));
693 } else {
694 // sign extend and use a full word?
695 __ ldrw(r, Address(esp, ld_off));
696 }
697 } else {
698 if (!r_2->is_valid()) {
699 __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
700 } else {
701 __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
702 }
703 }
704 }
705
706 __ mov(rscratch2, rscratch1);
707 __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1
708 __ mov(rscratch1, rscratch2);
709
710 // 6243940 We might end up in handle_wrong_method if
711 // the callee is deoptimized as we race thru here. If that
712 // happens we don't want to take a safepoint because the
713 // caller frame will look interpreted and arguments are now
714 // "compiled" so it is much better to make this transition
715 // invisible to the stack walking code. Unfortunately if
716 // we try and find the callee by normal means a safepoint
717 // is possible. So we stash the desired callee in the thread
718 // and the vm will find there should this case occur.
719
720 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset()));
721
722 __ br(rscratch1);
723 }
724
725 // ---------------------------------------------------------------
726 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
727 int total_args_passed,
728 int comp_args_on_stack,
729 const BasicType *sig_bt,
730 const VMRegPair *regs,
731 AdapterFingerPrint* fingerprint) {
732 address i2c_entry = __ pc();
733
734 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
735
736 address c2i_unverified_entry = __ pc();
737 Label skip_fixup;
738
739 Label ok;
740
741 Register holder = rscratch2;
742 Register receiver = j_rarg0;
743 Register tmp = r10; // A call-clobbered register not used for arg passing
744
745 // -------------------------------------------------------------------------
746 // Generate a C2I adapter. On entry we know rmethod holds the Method* during calls
747 // to the interpreter. The args start out packed in the compiled layout. They
748 // need to be unpacked into the interpreter layout. This will almost always
749 // require some stack space. We grow the current (compiled) stack, then repack
750 // the args. We finally end in a jump to the generic interpreter entry point.
751 // On exit from the interpreter, the interpreter will restore our SP (lest the
752 // compiled code, which relies solely on SP and not FP, get sick).
753
754 {
755 __ block_comment("c2i_unverified_entry {");
756 __ load_klass(rscratch1, receiver);
757 __ ldr(tmp, Address(holder, CompiledICHolder::holder_klass_offset()));
758 __ cmp(rscratch1, tmp);
759 __ ldr(rmethod, Address(holder, CompiledICHolder::holder_metadata_offset()));
760 __ br(Assembler::EQ, ok);
761 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
762
763 __ bind(ok);
764 // Method might have been compiled since the call site was patched to
765 // interpreted; if that is the case treat it as a miss so we can get
766 // the call site corrected.
767 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
768 __ cbz(rscratch1, skip_fixup);
769 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
770 __ block_comment("} c2i_unverified_entry");
771 }
772
773 address c2i_entry = __ pc();
774
775 // Class initialization barrier for static methods
776 address c2i_no_clinit_check_entry = nullptr;
777 if (VM_Version::supports_fast_class_init_checks()) {
778 Label L_skip_barrier;
779
780 { // Bypass the barrier for non-static methods
781 __ ldrw(rscratch1, Address(rmethod, Method::access_flags_offset()));
782 __ andsw(zr, rscratch1, JVM_ACC_STATIC);
783 __ br(Assembler::EQ, L_skip_barrier); // non-static
784 }
785
786 __ load_method_holder(rscratch2, rmethod);
787 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
788 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
789
790 __ bind(L_skip_barrier);
791 c2i_no_clinit_check_entry = __ pc();
792 }
793
794 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
795 bs->c2i_entry_barrier(masm);
796
797 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
798
799 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
800 }
801
802 static int c_calling_convention_priv(const BasicType *sig_bt,
803 VMRegPair *regs,
804 VMRegPair *regs2,
805 int total_args_passed) {
806 assert(regs2 == nullptr, "not needed on AArch64");
807
808 // We return the amount of VMRegImpl stack slots we need to reserve for all
809 // the arguments NOT counting out_preserve_stack_slots.
810
811 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
812 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5, c_rarg6, c_rarg7
813 };
814 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
815 c_farg0, c_farg1, c_farg2, c_farg3,
816 c_farg4, c_farg5, c_farg6, c_farg7
817 };
818
819 uint int_args = 0;
820 uint fp_args = 0;
821 uint stk_args = 0; // inc by 2 each time
822
823 for (int i = 0; i < total_args_passed; i++) {
824 switch (sig_bt[i]) {
825 case T_BOOLEAN:
826 case T_CHAR:
827 case T_BYTE:
828 case T_SHORT:
829 case T_INT:
830 if (int_args < Argument::n_int_register_parameters_c) {
831 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
832 } else {
833 #ifdef __APPLE__
834 // Less-than word types are stored one after another.
835 // The code is unable to handle this so bailout.
836 return -1;
837 #endif
838 regs[i].set1(VMRegImpl::stack2reg(stk_args));
839 stk_args += 2;
840 }
841 break;
842 case T_LONG:
843 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
844 // fall through
845 case T_OBJECT:
846 case T_ARRAY:
847 case T_ADDRESS:
848 case T_METADATA:
849 if (int_args < Argument::n_int_register_parameters_c) {
850 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
851 } else {
852 regs[i].set2(VMRegImpl::stack2reg(stk_args));
853 stk_args += 2;
854 }
855 break;
856 case T_FLOAT:
857 if (fp_args < Argument::n_float_register_parameters_c) {
858 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
859 } else {
860 #ifdef __APPLE__
861 // Less-than word types are stored one after another.
862 // The code is unable to handle this so bailout.
863 return -1;
864 #endif
865 regs[i].set1(VMRegImpl::stack2reg(stk_args));
866 stk_args += 2;
867 }
868 break;
869 case T_DOUBLE:
870 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
871 if (fp_args < Argument::n_float_register_parameters_c) {
872 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
873 } else {
874 regs[i].set2(VMRegImpl::stack2reg(stk_args));
875 stk_args += 2;
876 }
877 break;
878 case T_VOID: // Halves of longs and doubles
879 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
880 regs[i].set_bad();
881 break;
882 default:
883 ShouldNotReachHere();
884 break;
885 }
886 }
887
888 return stk_args;
889 }
890
891 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
892 uint num_bits,
893 uint total_args_passed) {
894 Unimplemented();
895 return 0;
896 }
897
898 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
899 VMRegPair *regs,
900 VMRegPair *regs2,
901 int total_args_passed)
902 {
903 int result = c_calling_convention_priv(sig_bt, regs, regs2, total_args_passed);
904 guarantee(result >= 0, "Unsupported arguments configuration");
905 return result;
906 }
907
908
909 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
910 // We always ignore the frame_slots arg and just use the space just below frame pointer
911 // which by this time is free to use
912 switch (ret_type) {
913 case T_FLOAT:
914 __ strs(v0, Address(rfp, -wordSize));
915 break;
916 case T_DOUBLE:
917 __ strd(v0, Address(rfp, -wordSize));
918 break;
919 case T_VOID: break;
920 default: {
921 __ str(r0, Address(rfp, -wordSize));
922 }
923 }
924 }
925
926 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
927 // We always ignore the frame_slots arg and just use the space just below frame pointer
928 // which by this time is free to use
929 switch (ret_type) {
930 case T_FLOAT:
931 __ ldrs(v0, Address(rfp, -wordSize));
932 break;
933 case T_DOUBLE:
934 __ ldrd(v0, Address(rfp, -wordSize));
935 break;
936 case T_VOID: break;
937 default: {
938 __ ldr(r0, Address(rfp, -wordSize));
939 }
940 }
941 }
942 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
943 RegSet x;
944 for ( int i = first_arg ; i < arg_count ; i++ ) {
945 if (args[i].first()->is_Register()) {
946 x = x + args[i].first()->as_Register();
947 } else if (args[i].first()->is_FloatRegister()) {
948 __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize)));
949 }
950 }
951 __ push(x, sp);
952 }
953
954 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
955 RegSet x;
956 for ( int i = first_arg ; i < arg_count ; i++ ) {
957 if (args[i].first()->is_Register()) {
958 x = x + args[i].first()->as_Register();
959 } else {
960 ;
961 }
962 }
963 __ pop(x, sp);
964 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
965 if (args[i].first()->is_Register()) {
966 ;
967 } else if (args[i].first()->is_FloatRegister()) {
968 __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize)));
969 }
970 }
971 }
972
973 static void verify_oop_args(MacroAssembler* masm,
974 const methodHandle& method,
975 const BasicType* sig_bt,
976 const VMRegPair* regs) {
977 Register temp_reg = r19; // not part of any compiled calling seq
978 if (VerifyOops) {
979 for (int i = 0; i < method->size_of_parameters(); i++) {
980 if (sig_bt[i] == T_OBJECT ||
981 sig_bt[i] == T_ARRAY) {
982 VMReg r = regs[i].first();
983 assert(r->is_valid(), "bad oop arg");
984 if (r->is_stack()) {
985 __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
986 __ verify_oop(temp_reg);
987 } else {
988 __ verify_oop(r->as_Register());
989 }
990 }
991 }
992 }
993 }
994
995 // on exit, sp points to the ContinuationEntry
996 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
997 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
998 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
999 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1000
1001 stack_slots += (int)ContinuationEntry::size()/wordSize;
1002 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
1003
1004 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize)/ VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1005
1006 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1007 __ str(rscratch1, Address(sp, ContinuationEntry::parent_offset()));
1008 __ mov(rscratch1, sp); // we can't use sp as the source in str
1009 __ str(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1010
1011 return map;
1012 }
1013
1014 // on entry c_rarg1 points to the continuation
1015 // sp points to ContinuationEntry
1016 // c_rarg3 -- isVirtualThread
1017 static void fill_continuation_entry(MacroAssembler* masm) {
1018 #ifdef ASSERT
1019 __ movw(rscratch1, ContinuationEntry::cookie_value());
1020 __ strw(rscratch1, Address(sp, ContinuationEntry::cookie_offset()));
1021 #endif
1022
1023 __ str (c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
1024 __ strw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
1025 __ str (zr, Address(sp, ContinuationEntry::chunk_offset()));
1026 __ strw(zr, Address(sp, ContinuationEntry::argsize_offset()));
1027 __ strw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
1028
1029 __ ldr(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1030 __ str(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1031 __ ldr(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset()));
1032 __ str(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
1033
1034 __ str(zr, Address(rthread, JavaThread::cont_fastpath_offset()));
1035 __ str(zr, Address(rthread, JavaThread::held_monitor_count_offset()));
1036 }
1037
1038 // on entry, sp points to the ContinuationEntry
1039 // on exit, rfp points to the spilled rfp in the entry frame
1040 static void continuation_enter_cleanup(MacroAssembler* masm) {
1041 #ifndef PRODUCT
1042 Label OK;
1043 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1044 __ cmp(sp, rscratch1);
1045 __ br(Assembler::EQ, OK);
1046 __ stop("incorrect sp1");
1047 __ bind(OK);
1048 #endif
1049
1050 __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1051 __ str(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1052 __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
1053 __ str(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset()));
1054
1055 __ ldr(rscratch2, Address(sp, ContinuationEntry::parent_offset()));
1056 __ str(rscratch2, Address(rthread, JavaThread::cont_entry_offset()));
1057 __ add(rfp, sp, (int)ContinuationEntry::size());
1058 }
1059
1060 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1061 // On entry: c_rarg1 -- the continuation object
1062 // c_rarg2 -- isContinue
1063 // c_rarg3 -- isVirtualThread
1064 static void gen_continuation_enter(MacroAssembler* masm,
1065 const methodHandle& method,
1066 const BasicType* sig_bt,
1067 const VMRegPair* regs,
1068 int& exception_offset,
1069 OopMapSet*oop_maps,
1070 int& frame_complete,
1071 int& stack_slots,
1072 int& interpreted_entry_offset,
1073 int& compiled_entry_offset) {
1074 //verify_oop_args(masm, method, sig_bt, regs);
1075 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1076
1077 address start = __ pc();
1078
1079 Label call_thaw, exit;
1080
1081 // i2i entry used at interp_only_mode only
1082 interpreted_entry_offset = __ pc() - start;
1083 {
1084
1085 #ifdef ASSERT
1086 Label is_interp_only;
1087 __ ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
1088 __ cbnzw(rscratch1, is_interp_only);
1089 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1090 __ bind(is_interp_only);
1091 #endif
1092
1093 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1094 __ ldr(c_rarg1, Address(esp, Interpreter::stackElementSize*2));
1095 __ ldr(c_rarg2, Address(esp, Interpreter::stackElementSize*1));
1096 __ ldr(c_rarg3, Address(esp, Interpreter::stackElementSize*0));
1097 __ push_cont_fastpath(rthread);
1098
1099 __ enter();
1100 stack_slots = 2; // will be adjusted in setup
1101 OopMap* map = continuation_enter_setup(masm, stack_slots);
1102 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1103 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1104
1105 fill_continuation_entry(masm);
1106
1107 __ cbnz(c_rarg2, call_thaw);
1108
1109 const address tr_call = __ trampoline_call(resolve);
1110 if (tr_call == nullptr) {
1111 fatal("CodeCache is full at gen_continuation_enter");
1112 }
1113
1114 oop_maps->add_gc_map(__ pc() - start, map);
1115 __ post_call_nop();
1116
1117 __ b(exit);
1118
1119 CodeBuffer* cbuf = masm->code_section()->outer();
1120 address stub = CompiledStaticCall::emit_to_interp_stub(*cbuf, tr_call);
1121 if (stub == nullptr) {
1122 fatal("CodeCache is full at gen_continuation_enter");
1123 }
1124 }
1125
1126 // compiled entry
1127 __ align(CodeEntryAlignment);
1128 compiled_entry_offset = __ pc() - start;
1129
1130 __ enter();
1131 stack_slots = 2; // will be adjusted in setup
1132 OopMap* map = continuation_enter_setup(masm, stack_slots);
1133 frame_complete = __ pc() - start;
1134
1135 fill_continuation_entry(masm);
1136
1137 __ cbnz(c_rarg2, call_thaw);
1138
1139 const address tr_call = __ trampoline_call(resolve);
1140 if (tr_call == nullptr) {
1141 fatal("CodeCache is full at gen_continuation_enter");
1142 }
1143
1144 oop_maps->add_gc_map(__ pc() - start, map);
1145 __ post_call_nop();
1146
1147 __ b(exit);
1148
1149 __ bind(call_thaw);
1150
1151 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1152 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1153 ContinuationEntry::_return_pc_offset = __ pc() - start;
1154 __ post_call_nop();
1155
1156 __ bind(exit);
1157 continuation_enter_cleanup(masm);
1158 __ leave();
1159 __ ret(lr);
1160
1161 /// exception handling
1162
1163 exception_offset = __ pc() - start;
1164 {
1165 __ mov(r19, r0); // save return value contaning the exception oop in callee-saved R19
1166
1167 continuation_enter_cleanup(masm);
1168
1169 __ ldr(c_rarg1, Address(rfp, wordSize)); // return address
1170 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1);
1171
1172 // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc
1173
1174 __ mov(r1, r0); // the exception handler
1175 __ mov(r0, r19); // restore return value contaning the exception oop
1176 __ verify_oop(r0);
1177
1178 __ leave();
1179 __ mov(r3, lr);
1180 __ br(r1); // the exception handler
1181 }
1182
1183 CodeBuffer* cbuf = masm->code_section()->outer();
1184 address stub = CompiledStaticCall::emit_to_interp_stub(*cbuf, tr_call);
1185 if (stub == nullptr) {
1186 fatal("CodeCache is full at gen_continuation_enter");
1187 }
1188 }
1189
1190 static void gen_continuation_yield(MacroAssembler* masm,
1191 const methodHandle& method,
1192 const BasicType* sig_bt,
1193 const VMRegPair* regs,
1194 OopMapSet* oop_maps,
1195 int& frame_complete,
1196 int& stack_slots,
1197 int& compiled_entry_offset) {
1198 enum layout {
1199 rfp_off1,
1200 rfp_off2,
1201 lr_off,
1202 lr_off2,
1203 framesize // inclusive of return address
1204 };
1205 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1206 stack_slots = framesize / VMRegImpl::slots_per_word;
1207 assert(stack_slots == 2, "recheck layout");
1208
1209 address start = __ pc();
1210
1211 compiled_entry_offset = __ pc() - start;
1212 __ enter();
1213
1214 __ mov(c_rarg1, sp);
1215
1216 frame_complete = __ pc() - start;
1217 address the_pc = __ pc();
1218
1219 __ post_call_nop(); // this must be exactly after the pc value that is pushed into the frame info, we use this nop for fast CodeBlob lookup
1220
1221 __ mov(c_rarg0, rthread);
1222 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
1223 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1224 __ reset_last_Java_frame(true);
1225
1226 Label pinned;
1227
1228 __ cbnz(r0, pinned);
1229
1230 // We've succeeded, set sp to the ContinuationEntry
1231 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1232 __ mov(sp, rscratch1);
1233 continuation_enter_cleanup(masm);
1234
1235 __ bind(pinned); // pinned -- return to caller
1236
1237 // handle pending exception thrown by freeze
1238 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
1239 Label ok;
1240 __ cbz(rscratch1, ok);
1241 __ leave();
1242 __ lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
1243 __ br(rscratch1);
1244 __ bind(ok);
1245
1246 __ leave();
1247 __ ret(lr);
1248
1249 OopMap* map = new OopMap(framesize, 1);
1250 oop_maps->add_gc_map(the_pc - start, map);
1251 }
1252
1253 static void gen_special_dispatch(MacroAssembler* masm,
1254 const methodHandle& method,
1255 const BasicType* sig_bt,
1256 const VMRegPair* regs) {
1257 verify_oop_args(masm, method, sig_bt, regs);
1258 vmIntrinsics::ID iid = method->intrinsic_id();
1259
1260 // Now write the args into the outgoing interpreter space
1261 bool has_receiver = false;
1262 Register receiver_reg = noreg;
1263 int member_arg_pos = -1;
1264 Register member_reg = noreg;
1265 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1266 if (ref_kind != 0) {
1267 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1268 member_reg = r19; // known to be free at this point
1269 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1270 } else if (iid == vmIntrinsics::_invokeBasic) {
1271 has_receiver = true;
1272 } else if (iid == vmIntrinsics::_linkToNative) {
1273 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1274 member_reg = r19; // known to be free at this point
1275 } else {
1276 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1277 }
1278
1279 if (member_reg != noreg) {
1280 // Load the member_arg into register, if necessary.
1281 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1282 VMReg r = regs[member_arg_pos].first();
1283 if (r->is_stack()) {
1284 __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1285 } else {
1286 // no data motion is needed
1287 member_reg = r->as_Register();
1288 }
1289 }
1290
1291 if (has_receiver) {
1292 // Make sure the receiver is loaded into a register.
1293 assert(method->size_of_parameters() > 0, "oob");
1294 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1295 VMReg r = regs[0].first();
1296 assert(r->is_valid(), "bad receiver arg");
1297 if (r->is_stack()) {
1298 // Porting note: This assumes that compiled calling conventions always
1299 // pass the receiver oop in a register. If this is not true on some
1300 // platform, pick a temp and load the receiver from stack.
1301 fatal("receiver always in a register");
1302 receiver_reg = r2; // known to be free at this point
1303 __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1304 } else {
1305 // no data motion is needed
1306 receiver_reg = r->as_Register();
1307 }
1308 }
1309
1310 // Figure out which address we are really jumping to:
1311 MethodHandles::generate_method_handle_dispatch(masm, iid,
1312 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1313 }
1314
1315 // ---------------------------------------------------------------------------
1316 // Generate a native wrapper for a given method. The method takes arguments
1317 // in the Java compiled code convention, marshals them to the native
1318 // convention (handlizes oops, etc), transitions to native, makes the call,
1319 // returns to java state (possibly blocking), unhandlizes any result and
1320 // returns.
1321 //
1322 // Critical native functions are a shorthand for the use of
1323 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1324 // functions. The wrapper is expected to unpack the arguments before
1325 // passing them to the callee. Critical native functions leave the state _in_Java,
1326 // since they block out GC.
1327 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1328 // block and the check for pending exceptions it's impossible for them
1329 // to be thrown.
1330 //
1331 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1332 const methodHandle& method,
1333 int compile_id,
1334 BasicType* in_sig_bt,
1335 VMRegPair* in_regs,
1336 BasicType ret_type) {
1337 if (method->is_continuation_native_intrinsic()) {
1338 int exception_offset = -1;
1339 OopMapSet* oop_maps = new OopMapSet();
1340 int frame_complete = -1;
1341 int stack_slots = -1;
1342 int interpreted_entry_offset = -1;
1343 int vep_offset = -1;
1344 if (method->is_continuation_enter_intrinsic()) {
1345 gen_continuation_enter(masm,
1346 method,
1347 in_sig_bt,
1348 in_regs,
1349 exception_offset,
1350 oop_maps,
1351 frame_complete,
1352 stack_slots,
1353 interpreted_entry_offset,
1354 vep_offset);
1355 } else if (method->is_continuation_yield_intrinsic()) {
1356 gen_continuation_yield(masm,
1357 method,
1358 in_sig_bt,
1359 in_regs,
1360 oop_maps,
1361 frame_complete,
1362 stack_slots,
1363 vep_offset);
1364 } else {
1365 guarantee(false, "Unknown Continuation native intrinsic");
1366 }
1367
1368 #ifdef ASSERT
1369 if (method->is_continuation_enter_intrinsic()) {
1370 assert(interpreted_entry_offset != -1, "Must be set");
1371 assert(exception_offset != -1, "Must be set");
1372 } else {
1373 assert(interpreted_entry_offset == -1, "Must be unset");
1374 assert(exception_offset == -1, "Must be unset");
1375 }
1376 assert(frame_complete != -1, "Must be set");
1377 assert(stack_slots != -1, "Must be set");
1378 assert(vep_offset != -1, "Must be set");
1379 #endif
1380
1381 __ flush();
1382 nmethod* nm = nmethod::new_native_nmethod(method,
1383 compile_id,
1384 masm->code(),
1385 vep_offset,
1386 frame_complete,
1387 stack_slots,
1388 in_ByteSize(-1),
1389 in_ByteSize(-1),
1390 oop_maps,
1391 exception_offset);
1392 if (method->is_continuation_enter_intrinsic()) {
1393 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1394 } else if (method->is_continuation_yield_intrinsic()) {
1395 _cont_doYield_stub = nm;
1396 } else {
1397 guarantee(false, "Unknown Continuation native intrinsic");
1398 }
1399 return nm;
1400 }
1401
1402 if (method->is_method_handle_intrinsic()) {
1403 vmIntrinsics::ID iid = method->intrinsic_id();
1404 intptr_t start = (intptr_t)__ pc();
1405 int vep_offset = ((intptr_t)__ pc()) - start;
1406
1407 // First instruction must be a nop as it may need to be patched on deoptimisation
1408 __ nop();
1409 gen_special_dispatch(masm,
1410 method,
1411 in_sig_bt,
1412 in_regs);
1413 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1414 __ flush();
1415 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1416 return nmethod::new_native_nmethod(method,
1417 compile_id,
1418 masm->code(),
1419 vep_offset,
1420 frame_complete,
1421 stack_slots / VMRegImpl::slots_per_word,
1422 in_ByteSize(-1),
1423 in_ByteSize(-1),
1424 nullptr);
1425 }
1426 address native_func = method->native_function();
1427 assert(native_func != nullptr, "must have function");
1428
1429 // An OopMap for lock (and class if static)
1430 OopMapSet *oop_maps = new OopMapSet();
1431 intptr_t start = (intptr_t)__ pc();
1432
1433 // We have received a description of where all the java arg are located
1434 // on entry to the wrapper. We need to convert these args to where
1435 // the jni function will expect them. To figure out where they go
1436 // we convert the java signature to a C signature by inserting
1437 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1438
1439 const int total_in_args = method->size_of_parameters();
1440 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1441
1442 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1443 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1444 BasicType* in_elem_bt = nullptr;
1445
1446 int argc = 0;
1447 out_sig_bt[argc++] = T_ADDRESS;
1448 if (method->is_static()) {
1449 out_sig_bt[argc++] = T_OBJECT;
1450 }
1451
1452 for (int i = 0; i < total_in_args ; i++ ) {
1453 out_sig_bt[argc++] = in_sig_bt[i];
1454 }
1455
1456 // Now figure out where the args must be stored and how much stack space
1457 // they require.
1458 int out_arg_slots;
1459 out_arg_slots = c_calling_convention_priv(out_sig_bt, out_regs, nullptr, total_c_args);
1460
1461 if (out_arg_slots < 0) {
1462 return nullptr;
1463 }
1464
1465 // Compute framesize for the wrapper. We need to handlize all oops in
1466 // incoming registers
1467
1468 // Calculate the total number of stack slots we will need.
1469
1470 // First count the abi requirement plus all of the outgoing args
1471 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1472
1473 // Now the space for the inbound oop handle area
1474 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1475
1476 int oop_handle_offset = stack_slots;
1477 stack_slots += total_save_slots;
1478
1479 // Now any space we need for handlizing a klass if static method
1480
1481 int klass_slot_offset = 0;
1482 int klass_offset = -1;
1483 int lock_slot_offset = 0;
1484 bool is_static = false;
1485
1486 if (method->is_static()) {
1487 klass_slot_offset = stack_slots;
1488 stack_slots += VMRegImpl::slots_per_word;
1489 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1490 is_static = true;
1491 }
1492
1493 // Plus a lock if needed
1494
1495 if (method->is_synchronized()) {
1496 lock_slot_offset = stack_slots;
1497 stack_slots += VMRegImpl::slots_per_word;
1498 }
1499
1500 // Now a place (+2) to save return values or temp during shuffling
1501 // + 4 for return address (which we own) and saved rfp
1502 stack_slots += 6;
1503
1504 // Ok The space we have allocated will look like:
1505 //
1506 //
1507 // FP-> | |
1508 // |---------------------|
1509 // | 2 slots for moves |
1510 // |---------------------|
1511 // | lock box (if sync) |
1512 // |---------------------| <- lock_slot_offset
1513 // | klass (if static) |
1514 // |---------------------| <- klass_slot_offset
1515 // | oopHandle area |
1516 // |---------------------| <- oop_handle_offset (8 java arg registers)
1517 // | outbound memory |
1518 // | based arguments |
1519 // | |
1520 // |---------------------|
1521 // | |
1522 // SP-> | out_preserved_slots |
1523 //
1524 //
1525
1526
1527 // Now compute actual number of stack words we need rounding to make
1528 // stack properly aligned.
1529 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1530
1531 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1532
1533 // First thing make an ic check to see if we should even be here
1534
1535 // We are free to use all registers as temps without saving them and
1536 // restoring them except rfp. rfp is the only callee save register
1537 // as far as the interpreter and the compiler(s) are concerned.
1538
1539
1540 const Register ic_reg = rscratch2;
1541 const Register receiver = j_rarg0;
1542
1543 Label hit;
1544 Label exception_pending;
1545
1546 assert_different_registers(ic_reg, receiver, rscratch1);
1547 __ verify_oop(receiver);
1548 __ cmp_klass(receiver, ic_reg, rscratch1);
1549 __ br(Assembler::EQ, hit);
1550
1551 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1552
1553 // Verified entry point must be aligned
1554 __ align(8);
1555
1556 __ bind(hit);
1557
1558 int vep_offset = ((intptr_t)__ pc()) - start;
1559
1560 // If we have to make this method not-entrant we'll overwrite its
1561 // first instruction with a jump. For this action to be legal we
1562 // must ensure that this first instruction is a B, BL, NOP, BKPT,
1563 // SVC, HVC, or SMC. Make it a NOP.
1564 __ nop();
1565
1566 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1567 Label L_skip_barrier;
1568 __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass*
1569 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
1570 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1571
1572 __ bind(L_skip_barrier);
1573 }
1574
1575 // Generate stack overflow check
1576 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1577
1578 // Generate a new frame for the wrapper.
1579 __ enter();
1580 // -2 because return address is already present and so is saved rfp
1581 __ sub(sp, sp, stack_size - 2*wordSize);
1582
1583 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1584 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1585
1586 // Frame is now completed as far as size and linkage.
1587 int frame_complete = ((intptr_t)__ pc()) - start;
1588
1589 // We use r20 as the oop handle for the receiver/klass
1590 // It is callee save so it survives the call to native
1591
1592 const Register oop_handle_reg = r20;
1593
1594 //
1595 // We immediately shuffle the arguments so that any vm call we have to
1596 // make from here on out (sync slow path, jvmti, etc.) we will have
1597 // captured the oops from our caller and have a valid oopMap for
1598 // them.
1599
1600 // -----------------
1601 // The Grand Shuffle
1602
1603 // The Java calling convention is either equal (linux) or denser (win64) than the
1604 // c calling convention. However the because of the jni_env argument the c calling
1605 // convention always has at least one more (and two for static) arguments than Java.
1606 // Therefore if we move the args from java -> c backwards then we will never have
1607 // a register->register conflict and we don't have to build a dependency graph
1608 // and figure out how to break any cycles.
1609 //
1610
1611 // Record esp-based slot for receiver on stack for non-static methods
1612 int receiver_offset = -1;
1613
1614 // This is a trick. We double the stack slots so we can claim
1615 // the oops in the caller's frame. Since we are sure to have
1616 // more args than the caller doubling is enough to make
1617 // sure we can capture all the incoming oop args from the
1618 // caller.
1619 //
1620 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1621
1622 // Mark location of rfp (someday)
1623 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp));
1624
1625
1626 int float_args = 0;
1627 int int_args = 0;
1628
1629 #ifdef ASSERT
1630 bool reg_destroyed[Register::number_of_registers];
1631 bool freg_destroyed[FloatRegister::number_of_registers];
1632 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1633 reg_destroyed[r] = false;
1634 }
1635 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1636 freg_destroyed[f] = false;
1637 }
1638
1639 #endif /* ASSERT */
1640
1641 // For JNI natives the incoming and outgoing registers are offset upwards.
1642 GrowableArray<int> arg_order(2 * total_in_args);
1643 VMRegPair tmp_vmreg;
1644 tmp_vmreg.set2(r19->as_VMReg());
1645
1646 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1647 arg_order.push(i);
1648 arg_order.push(c_arg);
1649 }
1650
1651 int temploc = -1;
1652 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1653 int i = arg_order.at(ai);
1654 int c_arg = arg_order.at(ai + 1);
1655 __ block_comment(err_msg("move %d -> %d", i, c_arg));
1656 assert(c_arg != -1 && i != -1, "wrong order");
1657 #ifdef ASSERT
1658 if (in_regs[i].first()->is_Register()) {
1659 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1660 } else if (in_regs[i].first()->is_FloatRegister()) {
1661 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1662 }
1663 if (out_regs[c_arg].first()->is_Register()) {
1664 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1665 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1666 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1667 }
1668 #endif /* ASSERT */
1669 switch (in_sig_bt[i]) {
1670 case T_ARRAY:
1671 case T_OBJECT:
1672 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1673 ((i == 0) && (!is_static)),
1674 &receiver_offset);
1675 int_args++;
1676 break;
1677 case T_VOID:
1678 break;
1679
1680 case T_FLOAT:
1681 __ float_move(in_regs[i], out_regs[c_arg]);
1682 float_args++;
1683 break;
1684
1685 case T_DOUBLE:
1686 assert( i + 1 < total_in_args &&
1687 in_sig_bt[i + 1] == T_VOID &&
1688 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1689 __ double_move(in_regs[i], out_regs[c_arg]);
1690 float_args++;
1691 break;
1692
1693 case T_LONG :
1694 __ long_move(in_regs[i], out_regs[c_arg]);
1695 int_args++;
1696 break;
1697
1698 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1699
1700 default:
1701 __ move32_64(in_regs[i], out_regs[c_arg]);
1702 int_args++;
1703 }
1704 }
1705
1706 // point c_arg at the first arg that is already loaded in case we
1707 // need to spill before we call out
1708 int c_arg = total_c_args - total_in_args;
1709
1710 // Pre-load a static method's oop into c_rarg1.
1711 if (method->is_static()) {
1712
1713 // load oop into a register
1714 __ movoop(c_rarg1,
1715 JNIHandles::make_local(method->method_holder()->java_mirror()));
1716
1717 // Now handlize the static class mirror it's known not-null.
1718 __ str(c_rarg1, Address(sp, klass_offset));
1719 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1720
1721 // Now get the handle
1722 __ lea(c_rarg1, Address(sp, klass_offset));
1723 // and protect the arg if we must spill
1724 c_arg--;
1725 }
1726
1727 // Change state to native (we save the return address in the thread, since it might not
1728 // be pushed on the stack when we do a stack traversal).
1729 // We use the same pc/oopMap repeatedly when we call out
1730
1731 Label native_return;
1732 __ set_last_Java_frame(sp, noreg, native_return, rscratch1);
1733
1734 Label dtrace_method_entry, dtrace_method_entry_done;
1735 {
1736 uint64_t offset;
1737 __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset);
1738 __ ldrb(rscratch1, Address(rscratch1, offset));
1739 __ cbnzw(rscratch1, dtrace_method_entry);
1740 __ bind(dtrace_method_entry_done);
1741 }
1742
1743 // RedefineClasses() tracing support for obsolete method entry
1744 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1745 // protect the args we've loaded
1746 save_args(masm, total_c_args, c_arg, out_regs);
1747 __ mov_metadata(c_rarg1, method());
1748 __ call_VM_leaf(
1749 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1750 rthread, c_rarg1);
1751 restore_args(masm, total_c_args, c_arg, out_regs);
1752 }
1753
1754 // Lock a synchronized method
1755
1756 // Register definitions used by locking and unlocking
1757
1758 const Register swap_reg = r0;
1759 const Register obj_reg = r19; // Will contain the oop
1760 const Register lock_reg = r13; // Address of compiler lock object (BasicLock)
1761 const Register old_hdr = r13; // value of old header at unlock time
1762 const Register lock_tmp = r14; // Temporary used by lightweight_lock/unlock
1763 const Register tmp = lr;
1764
1765 Label slow_path_lock;
1766 Label lock_done;
1767
1768 if (method->is_synchronized()) {
1769 Label count;
1770 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1771
1772 // Get the handle (the 2nd argument)
1773 __ mov(oop_handle_reg, c_rarg1);
1774
1775 // Get address of the box
1776
1777 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1778
1779 // Load the oop from the handle
1780 __ ldr(obj_reg, Address(oop_handle_reg, 0));
1781
1782 if (LockingMode == LM_MONITOR) {
1783 __ b(slow_path_lock);
1784 } else if (LockingMode == LM_LEGACY) {
1785 // Load (object->mark() | 1) into swap_reg %r0
1786 __ ldr(rscratch1, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1787 __ orr(swap_reg, rscratch1, 1);
1788
1789 // Save (object->mark() | 1) into BasicLock's displaced header
1790 __ str(swap_reg, Address(lock_reg, mark_word_offset));
1791
1792 // src -> dest iff dest == r0 else r0 <- dest
1793 __ cmpxchg_obj_header(r0, lock_reg, obj_reg, rscratch1, count, /*fallthrough*/nullptr);
1794
1795 // Hmm should this move to the slow path code area???
1796
1797 // Test if the oopMark is an obvious stack pointer, i.e.,
1798 // 1) (mark & 3) == 0, and
1799 // 2) sp <= mark < mark + os::pagesize()
1800 // These 3 tests can be done by evaluating the following
1801 // expression: ((mark - sp) & (3 - os::vm_page_size())),
1802 // assuming both stack pointer and pagesize have their
1803 // least significant 2 bits clear.
1804 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg
1805
1806 __ sub(swap_reg, sp, swap_reg);
1807 __ neg(swap_reg, swap_reg);
1808 __ ands(swap_reg, swap_reg, 3 - (int)os::vm_page_size());
1809
1810 // Save the test result, for recursive case, the result is zero
1811 __ str(swap_reg, Address(lock_reg, mark_word_offset));
1812 __ br(Assembler::NE, slow_path_lock);
1813 } else {
1814 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
1815 __ lightweight_lock(obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1816 }
1817 __ bind(count);
1818 __ increment(Address(rthread, JavaThread::held_monitor_count_offset()));
1819
1820 // Slow path will re-enter here
1821 __ bind(lock_done);
1822 }
1823
1824
1825 // Finally just about ready to make the JNI call
1826
1827 // get JNIEnv* which is first argument to native
1828 __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset())));
1829
1830 // Now set thread in native
1831 __ mov(rscratch1, _thread_in_native);
1832 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
1833 __ stlrw(rscratch1, rscratch2);
1834
1835 __ rt_call(native_func);
1836
1837 __ bind(native_return);
1838
1839 intptr_t return_pc = (intptr_t) __ pc();
1840 oop_maps->add_gc_map(return_pc - start, map);
1841
1842 // Unpack native results.
1843 switch (ret_type) {
1844 case T_BOOLEAN: __ c2bool(r0); break;
1845 case T_CHAR : __ ubfx(r0, r0, 0, 16); break;
1846 case T_BYTE : __ sbfx(r0, r0, 0, 8); break;
1847 case T_SHORT : __ sbfx(r0, r0, 0, 16); break;
1848 case T_INT : __ sbfx(r0, r0, 0, 32); break;
1849 case T_DOUBLE :
1850 case T_FLOAT :
1851 // Result is in v0 we'll save as needed
1852 break;
1853 case T_ARRAY: // Really a handle
1854 case T_OBJECT: // Really a handle
1855 break; // can't de-handlize until after safepoint check
1856 case T_VOID: break;
1857 case T_LONG: break;
1858 default : ShouldNotReachHere();
1859 }
1860
1861 Label safepoint_in_progress, safepoint_in_progress_done;
1862 Label after_transition;
1863
1864 // Switch thread to "native transition" state before reading the synchronization state.
1865 // This additional state is necessary because reading and testing the synchronization
1866 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1867 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1868 // VM thread changes sync state to synchronizing and suspends threads for GC.
1869 // Thread A is resumed to finish this native method, but doesn't block here since it
1870 // didn't see any synchronization is progress, and escapes.
1871 __ mov(rscratch1, _thread_in_native_trans);
1872
1873 __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset()));
1874
1875 // Force this write out before the read below
1876 if (!UseSystemMemoryBarrier) {
1877 __ dmb(Assembler::ISH);
1878 }
1879
1880 __ verify_sve_vector_length();
1881
1882 // Check for safepoint operation in progress and/or pending suspend requests.
1883 {
1884 // We need an acquire here to ensure that any subsequent load of the
1885 // global SafepointSynchronize::_state flag is ordered after this load
1886 // of the thread-local polling word. We don't want this poll to
1887 // return false (i.e. not safepointing) and a later poll of the global
1888 // SafepointSynchronize::_state spuriously to return true.
1889 //
1890 // This is to avoid a race when we're in a native->Java transition
1891 // racing the code which wakes up from a safepoint.
1892
1893 __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */);
1894 __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset()));
1895 __ cbnzw(rscratch1, safepoint_in_progress);
1896 __ bind(safepoint_in_progress_done);
1897 }
1898
1899 // change thread state
1900 __ mov(rscratch1, _thread_in_Java);
1901 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
1902 __ stlrw(rscratch1, rscratch2);
1903 __ bind(after_transition);
1904
1905 Label reguard;
1906 Label reguard_done;
1907 __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset()));
1908 __ cmpw(rscratch1, StackOverflow::stack_guard_yellow_reserved_disabled);
1909 __ br(Assembler::EQ, reguard);
1910 __ bind(reguard_done);
1911
1912 // native result if any is live
1913
1914 // Unlock
1915 Label unlock_done;
1916 Label slow_path_unlock;
1917 if (method->is_synchronized()) {
1918
1919 // Get locked oop from the handle we passed to jni
1920 __ ldr(obj_reg, Address(oop_handle_reg, 0));
1921
1922 Label done, not_recursive;
1923
1924 if (LockingMode == LM_LEGACY) {
1925 // Simple recursive lock?
1926 __ ldr(rscratch1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1927 __ cbnz(rscratch1, not_recursive);
1928 __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
1929 __ b(done);
1930 }
1931
1932 __ bind(not_recursive);
1933
1934 // Must save r0 if if it is live now because cmpxchg must use it
1935 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1936 save_native_result(masm, ret_type, stack_slots);
1937 }
1938
1939 if (LockingMode == LM_MONITOR) {
1940 __ b(slow_path_unlock);
1941 } else if (LockingMode == LM_LEGACY) {
1942 // get address of the stack lock
1943 __ lea(r0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1944 // get old displaced header
1945 __ ldr(old_hdr, Address(r0, 0));
1946
1947 // Atomic swap old header if oop still contains the stack lock
1948 Label count;
1949 __ cmpxchg_obj_header(r0, old_hdr, obj_reg, rscratch1, count, &slow_path_unlock);
1950 __ bind(count);
1951 __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
1952 } else {
1953 assert(LockingMode == LM_LIGHTWEIGHT, "");
1954 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1955 __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
1956 }
1957
1958 // slow path re-enters here
1959 __ bind(unlock_done);
1960 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1961 restore_native_result(masm, ret_type, stack_slots);
1962 }
1963
1964 __ bind(done);
1965 }
1966
1967 Label dtrace_method_exit, dtrace_method_exit_done;
1968 {
1969 uint64_t offset;
1970 __ adrp(rscratch1, ExternalAddress((address)&DTraceMethodProbes), offset);
1971 __ ldrb(rscratch1, Address(rscratch1, offset));
1972 __ cbnzw(rscratch1, dtrace_method_exit);
1973 __ bind(dtrace_method_exit_done);
1974 }
1975
1976 __ reset_last_Java_frame(false);
1977
1978 // Unbox oop result, e.g. JNIHandles::resolve result.
1979 if (is_reference_type(ret_type)) {
1980 __ resolve_jobject(r0, r1, r2);
1981 }
1982
1983 if (CheckJNICalls) {
1984 // clear_pending_jni_exception_check
1985 __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset()));
1986 }
1987
1988 // reset handle block
1989 __ ldr(r2, Address(rthread, JavaThread::active_handles_offset()));
1990 __ str(zr, Address(r2, JNIHandleBlock::top_offset()));
1991
1992 __ leave();
1993
1994 // Any exception pending?
1995 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
1996 __ cbnz(rscratch1, exception_pending);
1997
1998 // We're done
1999 __ ret(lr);
2000
2001 // Unexpected paths are out of line and go here
2002
2003 // forward the exception
2004 __ bind(exception_pending);
2005
2006 // and forward the exception
2007 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2008
2009 // Slow path locking & unlocking
2010 if (method->is_synchronized()) {
2011
2012 __ block_comment("Slow path lock {");
2013 __ bind(slow_path_lock);
2014
2015 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2016 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2017
2018 // protect the args we've loaded
2019 save_args(masm, total_c_args, c_arg, out_regs);
2020
2021 __ mov(c_rarg0, obj_reg);
2022 __ mov(c_rarg1, lock_reg);
2023 __ mov(c_rarg2, rthread);
2024
2025 // Not a leaf but we have last_Java_frame setup as we want
2026 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
2027 restore_args(masm, total_c_args, c_arg, out_regs);
2028
2029 #ifdef ASSERT
2030 { Label L;
2031 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2032 __ cbz(rscratch1, L);
2033 __ stop("no pending exception allowed on exit from monitorenter");
2034 __ bind(L);
2035 }
2036 #endif
2037 __ b(lock_done);
2038
2039 __ block_comment("} Slow path lock");
2040
2041 __ block_comment("Slow path unlock {");
2042 __ bind(slow_path_unlock);
2043
2044 // If we haven't already saved the native result we must save it now as xmm registers
2045 // are still exposed.
2046
2047 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2048 save_native_result(masm, ret_type, stack_slots);
2049 }
2050
2051 __ mov(c_rarg2, rthread);
2052 __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2053 __ mov(c_rarg0, obj_reg);
2054
2055 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2056 // NOTE that obj_reg == r19 currently
2057 __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2058 __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2059
2060 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2061
2062 #ifdef ASSERT
2063 {
2064 Label L;
2065 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2066 __ cbz(rscratch1, L);
2067 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2068 __ bind(L);
2069 }
2070 #endif /* ASSERT */
2071
2072 __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2073
2074 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2075 restore_native_result(masm, ret_type, stack_slots);
2076 }
2077 __ b(unlock_done);
2078
2079 __ block_comment("} Slow path unlock");
2080
2081 } // synchronized
2082
2083 // SLOW PATH Reguard the stack if needed
2084
2085 __ bind(reguard);
2086 save_native_result(masm, ret_type, stack_slots);
2087 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2088 restore_native_result(masm, ret_type, stack_slots);
2089 // and continue
2090 __ b(reguard_done);
2091
2092 // SLOW PATH safepoint
2093 {
2094 __ block_comment("safepoint {");
2095 __ bind(safepoint_in_progress);
2096
2097 // Don't use call_VM as it will see a possible pending exception and forward it
2098 // and never return here preventing us from clearing _last_native_pc down below.
2099 //
2100 save_native_result(masm, ret_type, stack_slots);
2101 __ mov(c_rarg0, rthread);
2102 #ifndef PRODUCT
2103 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2104 #endif
2105 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)));
2106 __ blr(rscratch1);
2107
2108 // Restore any method result value
2109 restore_native_result(masm, ret_type, stack_slots);
2110
2111 __ b(safepoint_in_progress_done);
2112 __ block_comment("} safepoint");
2113 }
2114
2115 // SLOW PATH dtrace support
2116 {
2117 __ block_comment("dtrace entry {");
2118 __ bind(dtrace_method_entry);
2119
2120 // We have all of the arguments setup at this point. We must not touch any register
2121 // argument registers at this point (what if we save/restore them there are no oop?
2122
2123 save_args(masm, total_c_args, c_arg, out_regs);
2124 __ mov_metadata(c_rarg1, method());
2125 __ call_VM_leaf(
2126 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2127 rthread, c_rarg1);
2128 restore_args(masm, total_c_args, c_arg, out_regs);
2129 __ b(dtrace_method_entry_done);
2130 __ block_comment("} dtrace entry");
2131 }
2132
2133 {
2134 __ block_comment("dtrace exit {");
2135 __ bind(dtrace_method_exit);
2136 save_native_result(masm, ret_type, stack_slots);
2137 __ mov_metadata(c_rarg1, method());
2138 __ call_VM_leaf(
2139 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2140 rthread, c_rarg1);
2141 restore_native_result(masm, ret_type, stack_slots);
2142 __ b(dtrace_method_exit_done);
2143 __ block_comment("} dtrace exit");
2144 }
2145
2146
2147 __ flush();
2148
2149 nmethod *nm = nmethod::new_native_nmethod(method,
2150 compile_id,
2151 masm->code(),
2152 vep_offset,
2153 frame_complete,
2154 stack_slots / VMRegImpl::slots_per_word,
2155 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2156 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2157 oop_maps);
2158
2159 return nm;
2160 }
2161
2162 // this function returns the adjust size (in number of words) to a c2i adapter
2163 // activation for use during deoptimization
2164 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2165 assert(callee_locals >= callee_parameters,
2166 "test and remove; got more parms than locals");
2167 if (callee_locals < callee_parameters)
2168 return 0; // No adjustment for negative locals
2169 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2170 // diff is counted in stack words
2171 return align_up(diff, 2);
2172 }
2173
2174
2175 //------------------------------generate_deopt_blob----------------------------
2176 void SharedRuntime::generate_deopt_blob() {
2177 // Allocate space for the code
2178 ResourceMark rm;
2179 // Setup code generation tools
2180 int pad = 0;
2181 #if INCLUDE_JVMCI
2182 if (EnableJVMCI) {
2183 pad += 512; // Increase the buffer size when compiling for JVMCI
2184 }
2185 #endif
2186 CodeBuffer buffer("deopt_blob", 2048+pad, 1024);
2187 MacroAssembler* masm = new MacroAssembler(&buffer);
2188 int frame_size_in_words;
2189 OopMap* map = nullptr;
2190 OopMapSet *oop_maps = new OopMapSet();
2191 RegisterSaver reg_save(COMPILER2_OR_JVMCI != 0);
2192
2193 // -------------
2194 // This code enters when returning to a de-optimized nmethod. A return
2195 // address has been pushed on the stack, and return values are in
2196 // registers.
2197 // If we are doing a normal deopt then we were called from the patched
2198 // nmethod from the point we returned to the nmethod. So the return
2199 // address on the stack is wrong by NativeCall::instruction_size
2200 // We will adjust the value so it looks like we have the original return
2201 // address on the stack (like when we eagerly deoptimized).
2202 // In the case of an exception pending when deoptimizing, we enter
2203 // with a return address on the stack that points after the call we patched
2204 // into the exception handler. We have the following register state from,
2205 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp).
2206 // r0: exception oop
2207 // r19: exception handler
2208 // r3: throwing pc
2209 // So in this case we simply jam r3 into the useless return address and
2210 // the stack looks just like we want.
2211 //
2212 // At this point we need to de-opt. We save the argument return
2213 // registers. We call the first C routine, fetch_unroll_info(). This
2214 // routine captures the return values and returns a structure which
2215 // describes the current frame size and the sizes of all replacement frames.
2216 // The current frame is compiled code and may contain many inlined
2217 // functions, each with their own JVM state. We pop the current frame, then
2218 // push all the new frames. Then we call the C routine unpack_frames() to
2219 // populate these frames. Finally unpack_frames() returns us the new target
2220 // address. Notice that callee-save registers are BLOWN here; they have
2221 // already been captured in the vframeArray at the time the return PC was
2222 // patched.
2223 address start = __ pc();
2224 Label cont;
2225
2226 // Prolog for non exception case!
2227
2228 // Save everything in sight.
2229 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2230
2231 // Normal deoptimization. Save exec mode for unpack_frames.
2232 __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved
2233 __ b(cont);
2234
2235 int reexecute_offset = __ pc() - start;
2236 #if INCLUDE_JVMCI && !defined(COMPILER1)
2237 if (EnableJVMCI && UseJVMCICompiler) {
2238 // JVMCI does not use this kind of deoptimization
2239 __ should_not_reach_here();
2240 }
2241 #endif
2242
2243 // Reexecute case
2244 // return address is the pc describes what bci to do re-execute at
2245
2246 // No need to update map as each call to save_live_registers will produce identical oopmap
2247 (void) reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2248
2249 __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved
2250 __ b(cont);
2251
2252 #if INCLUDE_JVMCI
2253 Label after_fetch_unroll_info_call;
2254 int implicit_exception_uncommon_trap_offset = 0;
2255 int uncommon_trap_offset = 0;
2256
2257 if (EnableJVMCI) {
2258 implicit_exception_uncommon_trap_offset = __ pc() - start;
2259
2260 __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2261 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2262
2263 uncommon_trap_offset = __ pc() - start;
2264
2265 // Save everything in sight.
2266 reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2267 // fetch_unroll_info needs to call last_java_frame()
2268 Label retaddr;
2269 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2270
2271 __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2272 __ movw(rscratch1, -1);
2273 __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2274
2275 __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute);
2276 __ mov(c_rarg0, rthread);
2277 __ movw(c_rarg2, rcpool); // exec mode
2278 __ lea(rscratch1,
2279 RuntimeAddress(CAST_FROM_FN_PTR(address,
2280 Deoptimization::uncommon_trap)));
2281 __ blr(rscratch1);
2282 __ bind(retaddr);
2283 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2284
2285 __ reset_last_Java_frame(false);
2286
2287 __ b(after_fetch_unroll_info_call);
2288 } // EnableJVMCI
2289 #endif // INCLUDE_JVMCI
2290
2291 int exception_offset = __ pc() - start;
2292
2293 // Prolog for exception case
2294
2295 // all registers are dead at this entry point, except for r0, and
2296 // r3 which contain the exception oop and exception pc
2297 // respectively. Set them in TLS and fall thru to the
2298 // unpack_with_exception_in_tls entry point.
2299
2300 __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
2301 __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
2302
2303 int exception_in_tls_offset = __ pc() - start;
2304
2305 // new implementation because exception oop is now passed in JavaThread
2306
2307 // Prolog for exception case
2308 // All registers must be preserved because they might be used by LinearScan
2309 // Exceptiop oop and throwing PC are passed in JavaThread
2310 // tos: stack at point of call to method that threw the exception (i.e. only
2311 // args are on the stack, no return address)
2312
2313 // The return address pushed by save_live_registers will be patched
2314 // later with the throwing pc. The correct value is not available
2315 // now because loading it from memory would destroy registers.
2316
2317 // NB: The SP at this point must be the SP of the method that is
2318 // being deoptimized. Deoptimization assumes that the frame created
2319 // here by save_live_registers is immediately below the method's SP.
2320 // This is a somewhat fragile mechanism.
2321
2322 // Save everything in sight.
2323 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2324
2325 // Now it is safe to overwrite any register
2326
2327 // Deopt during an exception. Save exec mode for unpack_frames.
2328 __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved
2329
2330 // load throwing pc from JavaThread and patch it as the return address
2331 // of the current frame. Then clear the field in JavaThread
2332 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2333 __ protect_return_address(r3, rscratch1);
2334 __ str(r3, Address(rfp, wordSize));
2335 __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2336
2337 #ifdef ASSERT
2338 // verify that there is really an exception oop in JavaThread
2339 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2340 __ verify_oop(r0);
2341
2342 // verify that there is no pending exception
2343 Label no_pending_exception;
2344 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2345 __ cbz(rscratch1, no_pending_exception);
2346 __ stop("must not have pending exception here");
2347 __ bind(no_pending_exception);
2348 #endif
2349
2350 __ bind(cont);
2351
2352 // Call C code. Need thread and this frame, but NOT official VM entry
2353 // crud. We cannot block on this call, no GC can happen.
2354 //
2355 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2356
2357 // fetch_unroll_info needs to call last_java_frame().
2358
2359 Label retaddr;
2360 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2361 #ifdef ASSERT
2362 { Label L;
2363 __ ldr(rscratch1, Address(rthread, JavaThread::last_Java_fp_offset()));
2364 __ cbz(rscratch1, L);
2365 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2366 __ bind(L);
2367 }
2368 #endif // ASSERT
2369 __ mov(c_rarg0, rthread);
2370 __ mov(c_rarg1, rcpool);
2371 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2372 __ blr(rscratch1);
2373 __ bind(retaddr);
2374
2375 // Need to have an oopmap that tells fetch_unroll_info where to
2376 // find any register it might need.
2377 oop_maps->add_gc_map(__ pc() - start, map);
2378
2379 __ reset_last_Java_frame(false);
2380
2381 #if INCLUDE_JVMCI
2382 if (EnableJVMCI) {
2383 __ bind(after_fetch_unroll_info_call);
2384 }
2385 #endif
2386
2387 // Load UnrollBlock* into r5
2388 __ mov(r5, r0);
2389
2390 __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset()));
2391 Label noException;
2392 __ cmpw(rcpool, Deoptimization::Unpack_exception); // Was exception pending?
2393 __ br(Assembler::NE, noException);
2394 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2395 // QQQ this is useless it was null above
2396 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2397 __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
2398 __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2399
2400 __ verify_oop(r0);
2401
2402 // Overwrite the result registers with the exception results.
2403 __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2404 // I think this is useless
2405 // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2406
2407 __ bind(noException);
2408
2409 // Only register save data is on the stack.
2410 // Now restore the result registers. Everything else is either dead
2411 // or captured in the vframeArray.
2412
2413 // Restore fp result register
2414 __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2415 // Restore integer result register
2416 __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2417
2418 // Pop all of the register save area off the stack
2419 __ add(sp, sp, frame_size_in_words * wordSize);
2420
2421 // All of the register save area has been popped of the stack. Only the
2422 // return address remains.
2423
2424 // Pop all the frames we must move/replace.
2425 //
2426 // Frame picture (youngest to oldest)
2427 // 1: self-frame (no frame link)
2428 // 2: deopting frame (no frame link)
2429 // 3: caller of deopting frame (could be compiled/interpreted).
2430 //
2431 // Note: by leaving the return address of self-frame on the stack
2432 // and using the size of frame 2 to adjust the stack
2433 // when we are done the return to frame 3 will still be on the stack.
2434
2435 // Pop deoptimized frame
2436 __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2437 __ sub(r2, r2, 2 * wordSize);
2438 __ add(sp, sp, r2);
2439 __ ldp(rfp, lr, __ post(sp, 2 * wordSize));
2440 __ authenticate_return_address();
2441 // LR should now be the return address to the caller (3)
2442
2443 #ifdef ASSERT
2444 // Compilers generate code that bang the stack by as much as the
2445 // interpreter would need. So this stack banging should never
2446 // trigger a fault. Verify that it does not on non product builds.
2447 __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2448 __ bang_stack_size(r19, r2);
2449 #endif
2450 // Load address of array of frame pcs into r2
2451 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset()));
2452
2453 // Trash the old pc
2454 // __ addptr(sp, wordSize); FIXME ????
2455
2456 // Load address of array of frame sizes into r4
2457 __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset()));
2458
2459 // Load counter into r3
2460 __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset()));
2461
2462 // Now adjust the caller's stack to make up for the extra locals
2463 // but record the original sp so that we can save it in the skeletal interpreter
2464 // frame and the stack walking of interpreter_sender will get the unextended sp
2465 // value and not the "real" sp value.
2466
2467 const Register sender_sp = r6;
2468
2469 __ mov(sender_sp, sp);
2470 __ ldrw(r19, Address(r5,
2471 Deoptimization::UnrollBlock::
2472 caller_adjustment_offset()));
2473 __ sub(sp, sp, r19);
2474
2475 // Push interpreter frames in a loop
2476 __ mov(rscratch1, (uint64_t)0xDEADDEAD); // Make a recognizable pattern
2477 __ mov(rscratch2, rscratch1);
2478 Label loop;
2479 __ bind(loop);
2480 __ ldr(r19, Address(__ post(r4, wordSize))); // Load frame size
2481 __ sub(r19, r19, 2*wordSize); // We'll push pc and fp by hand
2482 __ ldr(lr, Address(__ post(r2, wordSize))); // Load pc
2483 __ enter(); // Save old & set new fp
2484 __ sub(sp, sp, r19); // Prolog
2485 // This value is corrected by layout_activation_impl
2486 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2487 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2488 __ mov(sender_sp, sp); // Pass sender_sp to next frame
2489 __ sub(r3, r3, 1); // Decrement counter
2490 __ cbnz(r3, loop);
2491
2492 // Re-push self-frame
2493 __ ldr(lr, Address(r2));
2494 __ enter();
2495
2496 // Allocate a full sized register save area. We subtract 2 because
2497 // enter() just pushed 2 words
2498 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2499
2500 // Restore frame locals after moving the frame
2501 __ strd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2502 __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2503
2504 // Call C code. Need thread but NOT official VM entry
2505 // crud. We cannot block on this call, no GC can happen. Call should
2506 // restore return values to their stack-slots with the new SP.
2507 //
2508 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2509
2510 // Use rfp because the frames look interpreted now
2511 // Don't need the precise return PC here, just precise enough to point into this code blob.
2512 address the_pc = __ pc();
2513 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
2514
2515 __ mov(c_rarg0, rthread);
2516 __ movw(c_rarg1, rcpool); // second arg: exec_mode
2517 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2518 __ blr(rscratch1);
2519
2520 // Set an oopmap for the call site
2521 // Use the same PC we used for the last java frame
2522 oop_maps->add_gc_map(the_pc - start,
2523 new OopMap( frame_size_in_words, 0 ));
2524
2525 // Clear fp AND pc
2526 __ reset_last_Java_frame(true);
2527
2528 // Collect return values
2529 __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2530 __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2531 // I think this is useless (throwing pc?)
2532 // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2533
2534 // Pop self-frame.
2535 __ leave(); // Epilog
2536
2537 // Jump to interpreter
2538 __ ret(lr);
2539
2540 // Make sure all code is generated
2541 masm->flush();
2542
2543 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2544 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2545 #if INCLUDE_JVMCI
2546 if (EnableJVMCI) {
2547 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2548 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2549 }
2550 #endif
2551 }
2552
2553 // Number of stack slots between incoming argument block and the start of
2554 // a new frame. The PROLOG must add this many slots to the stack. The
2555 // EPILOG must remove this many slots. aarch64 needs two slots for
2556 // return address and fp.
2557 // TODO think this is correct but check
2558 uint SharedRuntime::in_preserve_stack_slots() {
2559 return 4;
2560 }
2561
2562 uint SharedRuntime::out_preserve_stack_slots() {
2563 return 0;
2564 }
2565
2566 #ifdef COMPILER2
2567 //------------------------------generate_uncommon_trap_blob--------------------
2568 void SharedRuntime::generate_uncommon_trap_blob() {
2569 // Allocate space for the code
2570 ResourceMark rm;
2571 // Setup code generation tools
2572 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
2573 MacroAssembler* masm = new MacroAssembler(&buffer);
2574
2575 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned");
2576
2577 address start = __ pc();
2578
2579 // Push self-frame. We get here with a return address in LR
2580 // and sp should be 16 byte aligned
2581 // push rfp and retaddr by hand
2582 __ protect_return_address();
2583 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize)));
2584 // we don't expect an arg reg save area
2585 #ifndef PRODUCT
2586 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2587 #endif
2588 // compiler left unloaded_class_index in j_rarg0 move to where the
2589 // runtime expects it.
2590 if (c_rarg1 != j_rarg0) {
2591 __ movw(c_rarg1, j_rarg0);
2592 }
2593
2594 // we need to set the past SP to the stack pointer of the stub frame
2595 // and the pc to the address where this runtime call will return
2596 // although actually any pc in this code blob will do).
2597 Label retaddr;
2598 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2599
2600 // Call C code. Need thread but NOT official VM entry
2601 // crud. We cannot block on this call, no GC can happen. Call should
2602 // capture callee-saved registers as well as return values.
2603 // Thread is in rdi already.
2604 //
2605 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index);
2606 //
2607 // n.b. 2 gp args, 0 fp args, integral return type
2608
2609 __ mov(c_rarg0, rthread);
2610 __ movw(c_rarg2, (unsigned)Deoptimization::Unpack_uncommon_trap);
2611 __ lea(rscratch1,
2612 RuntimeAddress(CAST_FROM_FN_PTR(address,
2613 Deoptimization::uncommon_trap)));
2614 __ blr(rscratch1);
2615 __ bind(retaddr);
2616
2617 // Set an oopmap for the call site
2618 OopMapSet* oop_maps = new OopMapSet();
2619 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0);
2620
2621 // location of rfp is known implicitly by the frame sender code
2622
2623 oop_maps->add_gc_map(__ pc() - start, map);
2624
2625 __ reset_last_Java_frame(false);
2626
2627 // move UnrollBlock* into r4
2628 __ mov(r4, r0);
2629
2630 #ifdef ASSERT
2631 { Label L;
2632 __ ldrw(rscratch1, Address(r4, Deoptimization::UnrollBlock::unpack_kind_offset()));
2633 __ cmpw(rscratch1, (unsigned)Deoptimization::Unpack_uncommon_trap);
2634 __ br(Assembler::EQ, L);
2635 __ stop("SharedRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
2636 __ bind(L);
2637 }
2638 #endif
2639
2640 // Pop all the frames we must move/replace.
2641 //
2642 // Frame picture (youngest to oldest)
2643 // 1: self-frame (no frame link)
2644 // 2: deopting frame (no frame link)
2645 // 3: caller of deopting frame (could be compiled/interpreted).
2646
2647 // Pop self-frame. We have no frame, and must rely only on r0 and sp.
2648 __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog!
2649
2650 // Pop deoptimized frame (int)
2651 __ ldrw(r2, Address(r4,
2652 Deoptimization::UnrollBlock::
2653 size_of_deoptimized_frame_offset()));
2654 __ sub(r2, r2, 2 * wordSize);
2655 __ add(sp, sp, r2);
2656 __ ldp(rfp, lr, __ post(sp, 2 * wordSize));
2657 __ authenticate_return_address();
2658 // LR should now be the return address to the caller (3) frame
2659
2660 #ifdef ASSERT
2661 // Compilers generate code that bang the stack by as much as the
2662 // interpreter would need. So this stack banging should never
2663 // trigger a fault. Verify that it does not on non product builds.
2664 __ ldrw(r1, Address(r4,
2665 Deoptimization::UnrollBlock::
2666 total_frame_sizes_offset()));
2667 __ bang_stack_size(r1, r2);
2668 #endif
2669
2670 // Load address of array of frame pcs into r2 (address*)
2671 __ ldr(r2, Address(r4,
2672 Deoptimization::UnrollBlock::frame_pcs_offset()));
2673
2674 // Load address of array of frame sizes into r5 (intptr_t*)
2675 __ ldr(r5, Address(r4,
2676 Deoptimization::UnrollBlock::
2677 frame_sizes_offset()));
2678
2679 // Counter
2680 __ ldrw(r3, Address(r4,
2681 Deoptimization::UnrollBlock::
2682 number_of_frames_offset())); // (int)
2683
2684 // Now adjust the caller's stack to make up for the extra locals but
2685 // record the original sp so that we can save it in the skeletal
2686 // interpreter frame and the stack walking of interpreter_sender
2687 // will get the unextended sp value and not the "real" sp value.
2688
2689 const Register sender_sp = r8;
2690
2691 __ mov(sender_sp, sp);
2692 __ ldrw(r1, Address(r4,
2693 Deoptimization::UnrollBlock::
2694 caller_adjustment_offset())); // (int)
2695 __ sub(sp, sp, r1);
2696
2697 // Push interpreter frames in a loop
2698 Label loop;
2699 __ bind(loop);
2700 __ ldr(r1, Address(r5, 0)); // Load frame size
2701 __ sub(r1, r1, 2 * wordSize); // We'll push pc and rfp by hand
2702 __ ldr(lr, Address(r2, 0)); // Save return address
2703 __ enter(); // and old rfp & set new rfp
2704 __ sub(sp, sp, r1); // Prolog
2705 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2706 // This value is corrected by layout_activation_impl
2707 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2708 __ mov(sender_sp, sp); // Pass sender_sp to next frame
2709 __ add(r5, r5, wordSize); // Bump array pointer (sizes)
2710 __ add(r2, r2, wordSize); // Bump array pointer (pcs)
2711 __ subsw(r3, r3, 1); // Decrement counter
2712 __ br(Assembler::GT, loop);
2713 __ ldr(lr, Address(r2, 0)); // save final return address
2714 // Re-push self-frame
2715 __ enter(); // & old rfp & set new rfp
2716
2717 // Use rfp because the frames look interpreted now
2718 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP.
2719 // Don't need the precise return PC here, just precise enough to point into this code blob.
2720 address the_pc = __ pc();
2721 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
2722
2723 // Call C code. Need thread but NOT official VM entry
2724 // crud. We cannot block on this call, no GC can happen. Call should
2725 // restore return values to their stack-slots with the new SP.
2726 // Thread is in rdi already.
2727 //
2728 // BasicType unpack_frames(JavaThread* thread, int exec_mode);
2729 //
2730 // n.b. 2 gp args, 0 fp args, integral return type
2731
2732 // sp should already be aligned
2733 __ mov(c_rarg0, rthread);
2734 __ movw(c_rarg1, (unsigned)Deoptimization::Unpack_uncommon_trap);
2735 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2736 __ blr(rscratch1);
2737
2738 // Set an oopmap for the call site
2739 // Use the same PC we used for the last java frame
2740 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0));
2741
2742 // Clear fp AND pc
2743 __ reset_last_Java_frame(true);
2744
2745 // Pop self-frame.
2746 __ leave(); // Epilog
2747
2748 // Jump to interpreter
2749 __ ret(lr);
2750
2751 // Make sure all code is generated
2752 masm->flush();
2753
2754 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps,
2755 SimpleRuntimeFrame::framesize >> 1);
2756 }
2757 #endif // COMPILER2
2758
2759
2760 //------------------------------generate_handler_blob------
2761 //
2762 // Generate a special Compile2Runtime blob that saves all registers,
2763 // and setup oopmap.
2764 //
2765 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
2766 ResourceMark rm;
2767 OopMapSet *oop_maps = new OopMapSet();
2768 OopMap* map;
2769
2770 // Allocate space for the code. Setup code generation tools.
2771 CodeBuffer buffer("handler_blob", 2048, 1024);
2772 MacroAssembler* masm = new MacroAssembler(&buffer);
2773
2774 address start = __ pc();
2775 address call_pc = nullptr;
2776 int frame_size_in_words;
2777 bool cause_return = (poll_type == POLL_AT_RETURN);
2778 RegisterSaver reg_save(poll_type == POLL_AT_VECTOR_LOOP /* save_vectors */);
2779
2780 // When the signal occurred, the LR was either signed and stored on the stack (in which
2781 // case it will be restored from the stack before being used) or unsigned and not stored
2782 // on the stack. Stipping ensures we get the right value.
2783 __ strip_return_address();
2784
2785 // Save Integer and Float registers.
2786 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2787
2788 // The following is basically a call_VM. However, we need the precise
2789 // address of the call in order to generate an oopmap. Hence, we do all the
2790 // work ourselves.
2791
2792 Label retaddr;
2793 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2794
2795 // The return address must always be correct so that frame constructor never
2796 // sees an invalid pc.
2797
2798 if (!cause_return) {
2799 // overwrite the return address pushed by save_live_registers
2800 // Additionally, r20 is a callee-saved register so we can look at
2801 // it later to determine if someone changed the return address for
2802 // us!
2803 __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset()));
2804 __ protect_return_address(r20, rscratch1);
2805 __ str(r20, Address(rfp, wordSize));
2806 }
2807
2808 // Do the call
2809 __ mov(c_rarg0, rthread);
2810 __ lea(rscratch1, RuntimeAddress(call_ptr));
2811 __ blr(rscratch1);
2812 __ bind(retaddr);
2813
2814 // Set an oopmap for the call site. This oopmap will map all
2815 // oop-registers and debug-info registers as callee-saved. This
2816 // will allow deoptimization at this safepoint to find all possible
2817 // debug-info recordings, as well as let GC find all oops.
2818
2819 oop_maps->add_gc_map( __ pc() - start, map);
2820
2821 Label noException;
2822
2823 __ reset_last_Java_frame(false);
2824
2825 __ membar(Assembler::LoadLoad | Assembler::LoadStore);
2826
2827 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2828 __ cbz(rscratch1, noException);
2829
2830 // Exception pending
2831
2832 reg_save.restore_live_registers(masm);
2833
2834 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2835
2836 // No exception case
2837 __ bind(noException);
2838
2839 Label no_adjust, bail;
2840 if (!cause_return) {
2841 // If our stashed return pc was modified by the runtime we avoid touching it
2842 __ ldr(rscratch1, Address(rfp, wordSize));
2843 __ cmp(r20, rscratch1);
2844 __ br(Assembler::NE, no_adjust);
2845 __ authenticate_return_address(r20, rscratch1);
2846
2847 #ifdef ASSERT
2848 // Verify the correct encoding of the poll we're about to skip.
2849 // See NativeInstruction::is_ldrw_to_zr()
2850 __ ldrw(rscratch1, Address(r20));
2851 __ ubfx(rscratch2, rscratch1, 22, 10);
2852 __ cmpw(rscratch2, 0b1011100101);
2853 __ br(Assembler::NE, bail);
2854 __ ubfx(rscratch2, rscratch1, 0, 5);
2855 __ cmpw(rscratch2, 0b11111);
2856 __ br(Assembler::NE, bail);
2857 #endif
2858 // Adjust return pc forward to step over the safepoint poll instruction
2859 __ add(r20, r20, NativeInstruction::instruction_size);
2860 __ protect_return_address(r20, rscratch1);
2861 __ str(r20, Address(rfp, wordSize));
2862 }
2863
2864 __ bind(no_adjust);
2865 // Normal exit, restore registers and exit.
2866 reg_save.restore_live_registers(masm);
2867
2868 __ ret(lr);
2869
2870 #ifdef ASSERT
2871 __ bind(bail);
2872 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2873 #endif
2874
2875 // Make sure all code is generated
2876 masm->flush();
2877
2878 // Fill-out other meta info
2879 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2880 }
2881
2882 //
2883 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2884 //
2885 // Generate a stub that calls into vm to find out the proper destination
2886 // of a java call. All the argument registers are live at this point
2887 // but since this is generic code we don't know what they are and the caller
2888 // must do any gc of the args.
2889 //
2890 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
2891 assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2892
2893 // allocate space for the code
2894 ResourceMark rm;
2895
2896 CodeBuffer buffer(name, 1000, 512);
2897 MacroAssembler* masm = new MacroAssembler(&buffer);
2898
2899 int frame_size_in_words;
2900 RegisterSaver reg_save(false /* save_vectors */);
2901
2902 OopMapSet *oop_maps = new OopMapSet();
2903 OopMap* map = nullptr;
2904
2905 int start = __ offset();
2906
2907 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2908
2909 int frame_complete = __ offset();
2910
2911 {
2912 Label retaddr;
2913 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2914
2915 __ mov(c_rarg0, rthread);
2916 __ lea(rscratch1, RuntimeAddress(destination));
2917
2918 __ blr(rscratch1);
2919 __ bind(retaddr);
2920 }
2921
2922 // Set an oopmap for the call site.
2923 // We need this not only for callee-saved registers, but also for volatile
2924 // registers that the compiler might be keeping live across a safepoint.
2925
2926 oop_maps->add_gc_map( __ offset() - start, map);
2927
2928 // r0 contains the address we are going to jump to assuming no exception got installed
2929
2930 // clear last_Java_sp
2931 __ reset_last_Java_frame(false);
2932 // check for pending exceptions
2933 Label pending;
2934 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2935 __ cbnz(rscratch1, pending);
2936
2937 // get the returned Method*
2938 __ get_vm_result_2(rmethod, rthread);
2939 __ str(rmethod, Address(sp, reg_save.reg_offset_in_bytes(rmethod)));
2940
2941 // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch
2942 __ str(r0, Address(sp, reg_save.rscratch1_offset_in_bytes()));
2943 reg_save.restore_live_registers(masm);
2944
2945 // We are back to the original state on entry and ready to go.
2946
2947 __ br(rscratch1);
2948
2949 // Pending exception after the safepoint
2950
2951 __ bind(pending);
2952
2953 reg_save.restore_live_registers(masm);
2954
2955 // exception pending => remove activation and forward to exception handler
2956
2957 __ str(zr, Address(rthread, JavaThread::vm_result_offset()));
2958
2959 __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
2960 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2961
2962 // -------------
2963 // make sure all code is generated
2964 masm->flush();
2965
2966 // return the blob
2967 // frame_size_words or bytes??
2968 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2969 }
2970
2971 #ifdef COMPILER2
2972 // This is here instead of runtime_aarch64_64.cpp because it uses SimpleRuntimeFrame
2973 //
2974 //------------------------------generate_exception_blob---------------------------
2975 // creates exception blob at the end
2976 // Using exception blob, this code is jumped from a compiled method.
2977 // (see emit_exception_handler in x86_64.ad file)
2978 //
2979 // Given an exception pc at a call we call into the runtime for the
2980 // handler in this method. This handler might merely restore state
2981 // (i.e. callee save registers) unwind the frame and jump to the
2982 // exception handler for the nmethod if there is no Java level handler
2983 // for the nmethod.
2984 //
2985 // This code is entered with a jmp.
2986 //
2987 // Arguments:
2988 // r0: exception oop
2989 // r3: exception pc
2990 //
2991 // Results:
2992 // r0: exception oop
2993 // r3: exception pc in caller or ???
2994 // destination: exception handler of caller
2995 //
2996 // Note: the exception pc MUST be at a call (precise debug information)
2997 // Registers r0, r3, r2, r4, r5, r8-r11 are not callee saved.
2998 //
2999
3000 void OptoRuntime::generate_exception_blob() {
3001 assert(!OptoRuntime::is_callee_saved_register(R3_num), "");
3002 assert(!OptoRuntime::is_callee_saved_register(R0_num), "");
3003 assert(!OptoRuntime::is_callee_saved_register(R2_num), "");
3004
3005 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned");
3006
3007 // Allocate space for the code
3008 ResourceMark rm;
3009 // Setup code generation tools
3010 CodeBuffer buffer("exception_blob", 2048, 1024);
3011 MacroAssembler* masm = new MacroAssembler(&buffer);
3012
3013 // TODO check various assumptions made here
3014 //
3015 // make sure we do so before running this
3016
3017 address start = __ pc();
3018
3019 // push rfp and retaddr by hand
3020 // Exception pc is 'return address' for stack walker
3021 __ protect_return_address();
3022 __ stp(rfp, lr, Address(__ pre(sp, -2 * wordSize)));
3023 // there are no callee save registers and we don't expect an
3024 // arg reg save area
3025 #ifndef PRODUCT
3026 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
3027 #endif
3028 // Store exception in Thread object. We cannot pass any arguments to the
3029 // handle_exception call, since we do not want to make any assumption
3030 // about the size of the frame where the exception happened in.
3031 __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
3032 __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
3033
3034 // This call does all the hard work. It checks if an exception handler
3035 // exists in the method.
3036 // If so, it returns the handler address.
3037 // If not, it prepares for stack-unwinding, restoring the callee-save
3038 // registers of the frame being removed.
3039 //
3040 // address OptoRuntime::handle_exception_C(JavaThread* thread)
3041 //
3042 // n.b. 1 gp arg, 0 fp args, integral return type
3043
3044 // the stack should always be aligned
3045 address the_pc = __ pc();
3046 __ set_last_Java_frame(sp, noreg, the_pc, rscratch1);
3047 __ mov(c_rarg0, rthread);
3048 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C)));
3049 __ blr(rscratch1);
3050 // handle_exception_C is a special VM call which does not require an explicit
3051 // instruction sync afterwards.
3052
3053 // May jump to SVE compiled code
3054 __ reinitialize_ptrue();
3055
3056 // Set an oopmap for the call site. This oopmap will only be used if we
3057 // are unwinding the stack. Hence, all locations will be dead.
3058 // Callee-saved registers will be the same as the frame above (i.e.,
3059 // handle_exception_stub), since they were restored when we got the
3060 // exception.
3061
3062 OopMapSet* oop_maps = new OopMapSet();
3063
3064 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0));
3065
3066 __ reset_last_Java_frame(false);
3067
3068 // Restore callee-saved registers
3069
3070 // rfp is an implicitly saved callee saved register (i.e. the calling
3071 // convention will save restore it in prolog/epilog) Other than that
3072 // there are no callee save registers now that adapter frames are gone.
3073 // and we dont' expect an arg reg save area
3074 __ ldp(rfp, r3, Address(__ post(sp, 2 * wordSize)));
3075 __ authenticate_return_address(r3);
3076
3077 // r0: exception handler
3078
3079 // We have a handler in r0 (could be deopt blob).
3080 __ mov(r8, r0);
3081
3082 // Get the exception oop
3083 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
3084 // Get the exception pc in case we are deoptimized
3085 __ ldr(r4, Address(rthread, JavaThread::exception_pc_offset()));
3086 #ifdef ASSERT
3087 __ str(zr, Address(rthread, JavaThread::exception_handler_pc_offset()));
3088 __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
3089 #endif
3090 // Clear the exception oop so GC no longer processes it as a root.
3091 __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
3092
3093 // r0: exception oop
3094 // r8: exception handler
3095 // r4: exception pc
3096 // Jump to handler
3097
3098 __ br(r8);
3099
3100 // Make sure all code is generated
3101 masm->flush();
3102
3103 // Set exception blob
3104 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1);
3105 }
3106
3107 #endif // COMPILER2