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 "classfile/symbolTable.hpp"
31 #include "code/codeCache.hpp"
32 #include "code/compiledIC.hpp"
33 #include "code/debugInfoRec.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/klass.inline.hpp"
43 #include "oops/method.inline.hpp"
44 #include "prims/methodHandles.hpp"
45 #include "runtime/continuation.hpp"
46 #include "runtime/continuationEntry.inline.hpp"
47 #include "runtime/globals.hpp"
48 #include "runtime/jniHandles.hpp"
49 #include "runtime/safepointMechanism.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "runtime/signature.hpp"
52 #include "runtime/stubRoutines.hpp"
53 #include "runtime/timerTrace.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 #ifdef PRODUCT
72 #define BLOCK_COMMENT(str) /* nothing */
73 #else
74 #define BLOCK_COMMENT(str) __ block_comment(str)
75 #endif
76
77 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
78
79 // FIXME -- this is used by C1
80 class RegisterSaver {
81 const bool _save_vectors;
82 public:
83 RegisterSaver(bool save_vectors) : _save_vectors(save_vectors) {}
84
85 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
86 void restore_live_registers(MacroAssembler* masm);
87
88 // Offsets into the register save area
89 // Used by deoptimization when it is managing result register
90 // values on its own
91
92 int reg_offset_in_bytes(Register r);
93 int r0_offset_in_bytes() { return reg_offset_in_bytes(r0); }
94 int rscratch1_offset_in_bytes() { return reg_offset_in_bytes(rscratch1); }
95 int v0_offset_in_bytes();
96
97 // Total stack size in bytes for saving sve predicate registers.
98 int total_sve_predicate_in_bytes();
99
100 // Capture info about frame layout
101 // Note this is only correct when not saving full vectors.
102 enum layout {
103 fpu_state_off = 0,
104 fpu_state_end = fpu_state_off + FPUStateSizeInWords - 1,
105 // The frame sender code expects that rfp will be in
106 // the "natural" place and will override any oopMap
107 // setting for it. We must therefore force the layout
108 // so that it agrees with the frame sender code.
109 r0_off = fpu_state_off + FPUStateSizeInWords,
110 rfp_off = r0_off + (Register::number_of_registers - 2) * Register::max_slots_per_register,
111 return_off = rfp_off + Register::max_slots_per_register, // slot for return address
112 reg_save_size = return_off + Register::max_slots_per_register};
113
114 };
115
116 int RegisterSaver::reg_offset_in_bytes(Register r) {
117 // The integer registers are located above the floating point
118 // registers in the stack frame pushed by save_live_registers() so the
119 // offset depends on whether we are saving full vectors, and whether
120 // those vectors are NEON or SVE.
121
122 int slots_per_vect = FloatRegister::save_slots_per_register;
123
124 #if COMPILER2_OR_JVMCI
125 if (_save_vectors) {
126 slots_per_vect = FloatRegister::slots_per_neon_register;
127
128 #ifdef COMPILER2
129 if (Matcher::supports_scalable_vector()) {
130 slots_per_vect = Matcher::scalable_vector_reg_size(T_FLOAT);
131 }
132 #endif
133 }
134 #endif
135
136 int r0_offset = v0_offset_in_bytes() + (slots_per_vect * FloatRegister::number_of_registers) * BytesPerInt;
137 return r0_offset + r->encoding() * wordSize;
138 }
139
140 int RegisterSaver::v0_offset_in_bytes() {
141 // The floating point registers are located above the predicate registers if
142 // they are present in the stack frame pushed by save_live_registers(). So the
143 // offset depends on the saved total predicate vectors in the stack frame.
144 return (total_sve_predicate_in_bytes() / VMRegImpl::stack_slot_size) * BytesPerInt;
145 }
146
147 int RegisterSaver::total_sve_predicate_in_bytes() {
148 #ifdef COMPILER2
149 if (_save_vectors && Matcher::supports_scalable_vector()) {
150 return (Matcher::scalable_vector_reg_size(T_BYTE) >> LogBitsPerByte) *
151 PRegister::number_of_registers;
152 }
153 #endif
154 return 0;
155 }
156
157 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
158 bool use_sve = false;
159 int sve_vector_size_in_bytes = 0;
160 int sve_vector_size_in_slots = 0;
161 int sve_predicate_size_in_slots = 0;
162 int total_predicate_in_bytes = total_sve_predicate_in_bytes();
163 int total_predicate_in_slots = total_predicate_in_bytes / VMRegImpl::stack_slot_size;
164
165 #ifdef COMPILER2
166 use_sve = Matcher::supports_scalable_vector();
167 if (use_sve) {
168 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
169 sve_vector_size_in_slots = Matcher::scalable_vector_reg_size(T_FLOAT);
170 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
171 }
172 #endif
173
174 #if COMPILER2_OR_JVMCI
175 if (_save_vectors) {
176 int extra_save_slots_per_register = 0;
177 // Save upper half of vector registers
178 if (use_sve) {
179 extra_save_slots_per_register = sve_vector_size_in_slots - FloatRegister::save_slots_per_register;
180 } else {
181 extra_save_slots_per_register = FloatRegister::extra_save_slots_per_neon_register;
182 }
183 int extra_vector_bytes = extra_save_slots_per_register *
184 VMRegImpl::stack_slot_size *
185 FloatRegister::number_of_registers;
186 additional_frame_words += ((extra_vector_bytes + total_predicate_in_bytes) / wordSize);
187 }
188 #else
189 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
190 #endif
191
192 int frame_size_in_bytes = align_up(additional_frame_words * wordSize +
193 reg_save_size * BytesPerInt, 16);
194 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
195 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
196 // The caller will allocate additional_frame_words
197 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
198 // CodeBlob frame size is in words.
199 int frame_size_in_words = frame_size_in_bytes / wordSize;
200 *total_frame_words = frame_size_in_words;
201
202 // Save Integer and Float registers.
203 __ enter();
204 __ push_CPU_state(_save_vectors, use_sve, sve_vector_size_in_bytes, total_predicate_in_bytes);
205
206 // Set an oopmap for the call site. This oopmap will map all
207 // oop-registers and debug-info registers as callee-saved. This
208 // will allow deoptimization at this safepoint to find all possible
209 // debug-info recordings, as well as let GC find all oops.
210
211 OopMapSet *oop_maps = new OopMapSet();
212 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
213
214 for (int i = 0; i < Register::number_of_registers; i++) {
215 Register r = as_Register(i);
216 if (i <= rfp->encoding() && r != rscratch1 && r != rscratch2) {
217 // SP offsets are in 4-byte words.
218 // Register slots are 8 bytes wide, 32 floating-point registers.
219 int sp_offset = Register::max_slots_per_register * i +
220 FloatRegister::save_slots_per_register * FloatRegister::number_of_registers;
221 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset + additional_frame_slots), r->as_VMReg());
222 }
223 }
224
225 for (int i = 0; i < FloatRegister::number_of_registers; i++) {
226 FloatRegister r = as_FloatRegister(i);
227 int sp_offset = 0;
228 if (_save_vectors) {
229 sp_offset = use_sve ? (total_predicate_in_slots + sve_vector_size_in_slots * i) :
230 (FloatRegister::slots_per_neon_register * i);
231 } else {
232 sp_offset = FloatRegister::save_slots_per_register * i;
233 }
234 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset), r->as_VMReg());
235 }
236
237 return oop_map;
238 }
239
240 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
241 #ifdef COMPILER2
242 __ pop_CPU_state(_save_vectors, Matcher::supports_scalable_vector(),
243 Matcher::scalable_vector_reg_size(T_BYTE), total_sve_predicate_in_bytes());
244 #else
245 #if !INCLUDE_JVMCI
246 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
247 #endif
248 __ pop_CPU_state(_save_vectors);
249 #endif
250 __ ldp(rfp, lr, Address(__ post(sp, 2 * wordSize)));
251 __ authenticate_return_address();
252 }
253
254 // Is vector's size (in bytes) bigger than a size saved by default?
255 // 8 bytes vector registers are saved by default on AArch64.
256 // The SVE supported min vector size is 8 bytes and we need to save
257 // predicate registers when the vector size is 8 bytes as well.
258 bool SharedRuntime::is_wide_vector(int size) {
259 return size > 8 || (UseSVE > 0 && size >= 8);
260 }
261
262 // ---------------------------------------------------------------------------
263 // Read the array of BasicTypes from a signature, and compute where the
264 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
265 // quantities. Values less than VMRegImpl::stack0 are registers, those above
266 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
267 // as framesizes are fixed.
268 // VMRegImpl::stack0 refers to the first slot 0(sp).
269 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
270 // Register up to Register::number_of_registers are the 64-bit
271 // integer registers.
272
273 // Note: the INPUTS in sig_bt are in units of Java argument words,
274 // which are 64-bit. The OUTPUTS are in 32-bit units.
275
276 // The Java calling convention is a "shifted" version of the C ABI.
277 // By skipping the first C ABI register we can call non-static jni
278 // methods with small numbers of arguments without having to shuffle
279 // the arguments at all. Since we control the java ABI we ought to at
280 // least get some advantage out of it.
281
282 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
283 VMRegPair *regs,
284 int total_args_passed) {
285
286 // Create the mapping between argument positions and
287 // registers.
288 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
289 j_rarg0, j_rarg1, j_rarg2, j_rarg3, j_rarg4, j_rarg5, j_rarg6, j_rarg7
290 };
291 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
292 j_farg0, j_farg1, j_farg2, j_farg3,
293 j_farg4, j_farg5, j_farg6, j_farg7
294 };
295
296
297 uint int_args = 0;
298 uint fp_args = 0;
299 uint stk_args = 0;
300
301 for (int i = 0; i < total_args_passed; i++) {
302 switch (sig_bt[i]) {
303 case T_BOOLEAN:
304 case T_CHAR:
305 case T_BYTE:
306 case T_SHORT:
307 case T_INT:
308 if (int_args < Argument::n_int_register_parameters_j) {
309 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
310 } else {
311 stk_args = align_up(stk_args, 2);
312 regs[i].set1(VMRegImpl::stack2reg(stk_args));
313 stk_args += 1;
314 }
315 break;
316 case T_VOID:
317 // halves of T_LONG or T_DOUBLE
318 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
319 regs[i].set_bad();
320 break;
321 case T_LONG:
322 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
323 // fall through
324 case T_OBJECT:
325 case T_ARRAY:
326 case T_ADDRESS:
327 if (int_args < Argument::n_int_register_parameters_j) {
328 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
329 } else {
330 stk_args = align_up(stk_args, 2);
331 regs[i].set2(VMRegImpl::stack2reg(stk_args));
332 stk_args += 2;
333 }
334 break;
335 case T_FLOAT:
336 if (fp_args < Argument::n_float_register_parameters_j) {
337 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
338 } else {
339 stk_args = align_up(stk_args, 2);
340 regs[i].set1(VMRegImpl::stack2reg(stk_args));
341 stk_args += 1;
342 }
343 break;
344 case T_DOUBLE:
345 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
346 if (fp_args < Argument::n_float_register_parameters_j) {
347 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
348 } else {
349 stk_args = align_up(stk_args, 2);
350 regs[i].set2(VMRegImpl::stack2reg(stk_args));
351 stk_args += 2;
352 }
353 break;
354 default:
355 ShouldNotReachHere();
356 break;
357 }
358 }
359
360 return stk_args;
361 }
362
363
364 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
365 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
366
367 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
368
369 // Create the mapping between argument positions and registers.
370
371 static const Register INT_ArgReg[java_return_convention_max_int] = {
372 r0 /* j_rarg7 */, j_rarg6, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
373 };
374
375 static const FloatRegister FP_ArgReg[java_return_convention_max_float] = {
376 j_farg0, j_farg1, j_farg2, j_farg3, j_farg4, j_farg5, j_farg6, j_farg7
377 };
378
379 uint int_args = 0;
380 uint fp_args = 0;
381
382 for (int i = 0; i < total_args_passed; i++) {
383 switch (sig_bt[i]) {
384 case T_BOOLEAN:
385 case T_CHAR:
386 case T_BYTE:
387 case T_SHORT:
388 case T_INT:
389 if (int_args < SharedRuntime::java_return_convention_max_int) {
390 regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
391 int_args ++;
392 } else {
393 return -1;
394 }
395 break;
396 case T_VOID:
397 // halves of T_LONG or T_DOUBLE
398 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
399 regs[i].set_bad();
400 break;
401 case T_LONG:
402 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
403 // fall through
404 case T_OBJECT:
405 case T_ARRAY:
406 case T_ADDRESS:
407 // Should T_METADATA be added to java_calling_convention as well ?
408 case T_METADATA:
409 if (int_args < SharedRuntime::java_return_convention_max_int) {
410 regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
411 int_args ++;
412 } else {
413 return -1;
414 }
415 break;
416 case T_FLOAT:
417 if (fp_args < SharedRuntime::java_return_convention_max_float) {
418 regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
419 fp_args ++;
420 } else {
421 return -1;
422 }
423 break;
424 case T_DOUBLE:
425 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
426 if (fp_args < SharedRuntime::java_return_convention_max_float) {
427 regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
428 fp_args ++;
429 } else {
430 return -1;
431 }
432 break;
433 default:
434 ShouldNotReachHere();
435 break;
436 }
437 }
438
439 return int_args + fp_args;
440 }
441
442 // Patch the callers callsite with entry to compiled code if it exists.
443 static void patch_callers_callsite(MacroAssembler *masm) {
444 Label L;
445 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
446 __ cbz(rscratch1, L);
447
448 __ enter();
449 __ push_CPU_state();
450
451 // VM needs caller's callsite
452 // VM needs target method
453 // This needs to be a long call since we will relocate this adapter to
454 // the codeBuffer and it may not reach
455
456 #ifndef PRODUCT
457 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
458 #endif
459
460 __ mov(c_rarg0, rmethod);
461 __ mov(c_rarg1, lr);
462 __ authenticate_return_address(c_rarg1);
463 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
464 __ blr(rscratch1);
465
466 // Explicit isb required because fixup_callers_callsite may change the code
467 // stream.
468 __ safepoint_isb();
469
470 __ pop_CPU_state();
471 // restore sp
472 __ leave();
473 __ bind(L);
474 }
475
476 // For each inline type argument, sig includes the list of fields of
477 // the inline type. This utility function computes the number of
478 // arguments for the call if inline types are passed by reference (the
479 // calling convention the interpreter expects).
480 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
481 int total_args_passed = 0;
482 if (InlineTypePassFieldsAsArgs) {
483 for (int i = 0; i < sig_extended->length(); i++) {
484 BasicType bt = sig_extended->at(i)._bt;
485 if (bt == T_METADATA) {
486 // In sig_extended, an inline type argument starts with:
487 // T_METADATA, followed by the types of the fields of the
488 // inline type and T_VOID to mark the end of the value
489 // type. Inline types are flattened so, for instance, in the
490 // case of an inline type with an int field and an inline type
491 // field that itself has 2 fields, an int and a long:
492 // T_METADATA T_INT T_METADATA T_INT T_LONG T_VOID (second
493 // slot for the T_LONG) T_VOID (inner inline type) T_VOID
494 // (outer inline type)
495 total_args_passed++;
496 int vt = 1;
497 do {
498 i++;
499 BasicType bt = sig_extended->at(i)._bt;
500 BasicType prev_bt = sig_extended->at(i-1)._bt;
501 if (bt == T_METADATA) {
502 vt++;
503 } else if (bt == T_VOID &&
504 prev_bt != T_LONG &&
505 prev_bt != T_DOUBLE) {
506 vt--;
507 }
508 } while (vt != 0);
509 } else {
510 total_args_passed++;
511 }
512 }
513 } else {
514 total_args_passed = sig_extended->length();
515 }
516
517 return total_args_passed;
518 }
519
520
521 static void gen_c2i_adapter_helper(MacroAssembler* masm,
522 BasicType bt,
523 BasicType prev_bt,
524 size_t size_in_bytes,
525 const VMRegPair& reg_pair,
526 const Address& to,
527 Register tmp1,
528 Register tmp2,
529 Register tmp3,
530 int extraspace,
531 bool is_oop) {
532 if (bt == T_VOID) {
533 assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
534 return;
535 }
536
537 // Say 4 args:
538 // i st_off
539 // 0 32 T_LONG
540 // 1 24 T_VOID
541 // 2 16 T_OBJECT
542 // 3 8 T_BOOL
543 // - 0 return address
544 //
545 // However to make thing extra confusing. Because we can fit a Java long/double in
546 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
547 // leaves one slot empty and only stores to a single slot. In this case the
548 // slot that is occupied is the T_VOID slot. See I said it was confusing.
549
550 bool wide = (size_in_bytes == wordSize);
551 VMReg r_1 = reg_pair.first();
552 VMReg r_2 = reg_pair.second();
553 assert(r_2->is_valid() == wide, "invalid size");
554 if (!r_1->is_valid()) {
555 assert(!r_2->is_valid(), "");
556 return;
557 }
558
559 if (!r_1->is_FloatRegister()) {
560 Register val = r25;
561 if (r_1->is_stack()) {
562 // memory to memory use r25 (scratch registers is used by store_heap_oop)
563 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
564 __ load_sized_value(val, Address(sp, ld_off), size_in_bytes, /* is_signed */ false);
565 } else {
566 val = r_1->as_Register();
567 }
568 assert_different_registers(to.base(), val, tmp1, tmp2, tmp3);
569 if (is_oop) {
570 __ store_heap_oop(to, val, tmp1, tmp2, tmp3, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
571 } else {
572 __ store_sized_value(to, val, size_in_bytes);
573 }
574 } else {
575 if (wide) {
576 __ strd(r_1->as_FloatRegister(), to);
577 } else {
578 // only a float use just part of the slot
579 __ strs(r_1->as_FloatRegister(), to);
580 }
581 }
582 }
583
584 static void gen_c2i_adapter(MacroAssembler *masm,
585 const GrowableArray<SigEntry>* sig_extended,
586 const VMRegPair *regs,
587 bool requires_clinit_barrier,
588 address& c2i_no_clinit_check_entry,
589 Label& skip_fixup,
590 address start,
591 OopMapSet* oop_maps,
592 int& frame_complete,
593 int& frame_size_in_words,
594 bool alloc_inline_receiver) {
595 if (requires_clinit_barrier && VM_Version::supports_fast_class_init_checks()) {
596 Label L_skip_barrier;
597
598 { // Bypass the barrier for non-static methods
599 __ ldrw(rscratch1, Address(rmethod, Method::access_flags_offset()));
600 __ andsw(zr, rscratch1, JVM_ACC_STATIC);
601 __ br(Assembler::EQ, L_skip_barrier); // non-static
602 }
603
604 __ load_method_holder(rscratch2, rmethod);
605 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
606 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
607
608 __ bind(L_skip_barrier);
609 c2i_no_clinit_check_entry = __ pc();
610 }
611
612 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
613 bs->c2i_entry_barrier(masm);
614
615 // Before we get into the guts of the C2I adapter, see if we should be here
616 // at all. We've come from compiled code and are attempting to jump to the
617 // interpreter, which means the caller made a static call to get here
618 // (vcalls always get a compiled target if there is one). Check for a
619 // compiled target. If there is one, we need to patch the caller's call.
620 patch_callers_callsite(masm);
621
622 __ bind(skip_fixup);
623
624 // Name some registers to be used in the following code. We can use
625 // anything except r0-r7 which are arguments in the Java calling
626 // convention, rmethod (r12), and r13 which holds the outgoing sender
627 // SP for the interpreter.
628 Register buf_array = r10; // Array of buffered inline types
629 Register buf_oop = r11; // Buffered inline type oop
630 Register tmp1 = r15;
631 Register tmp2 = r16;
632 Register tmp3 = r17;
633
634 if (InlineTypePassFieldsAsArgs) {
635 // Is there an inline type argument?
636 bool has_inline_argument = false;
637 for (int i = 0; i < sig_extended->length() && !has_inline_argument; i++) {
638 has_inline_argument = (sig_extended->at(i)._bt == T_METADATA);
639 }
640 if (has_inline_argument) {
641 // There is at least an inline type argument: we're coming from
642 // compiled code so we have no buffers to back the inline types
643 // Allocate the buffers here with a runtime call.
644 RegisterSaver reg_save(false /* save_vectors */);
645 OopMap* map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
646
647 frame_complete = __ offset();
648 address the_pc = __ pc();
649
650 Label retaddr;
651 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
652
653 __ mov(c_rarg0, rthread);
654 __ mov(c_rarg1, rmethod);
655 __ mov(c_rarg2, (int64_t)alloc_inline_receiver);
656
657 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::allocate_inline_types)));
658 __ blr(rscratch1);
659 __ bind(retaddr);
660
661 oop_maps->add_gc_map(__ pc() - start, map);
662 __ reset_last_Java_frame(false);
663
664 reg_save.restore_live_registers(masm);
665
666 Label no_exception;
667 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
668 __ cbz(rscratch1, no_exception);
669
670 __ str(zr, Address(rthread, JavaThread::vm_result_offset()));
671 __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
672 __ b(RuntimeAddress(StubRoutines::forward_exception_entry()));
673
674 __ bind(no_exception);
675
676 // We get an array of objects from the runtime call
677 __ get_vm_result(buf_array, rthread);
678 __ get_vm_result_2(rmethod, rthread); // TODO: required to keep the callee Method live?
679 }
680 }
681
682 // Since all args are passed on the stack, total_args_passed *
683 // Interpreter::stackElementSize is the space we need.
684
685 int total_args_passed = compute_total_args_passed_int(sig_extended);
686 int extraspace = total_args_passed * Interpreter::stackElementSize;
687
688 // stack is aligned, keep it that way
689 extraspace = align_up(extraspace, StackAlignmentInBytes);
690
691 // set senderSP value
692 __ mov(r19_sender_sp, sp);
693
694 __ sub(sp, sp, extraspace);
695
696 // Now write the args into the outgoing interpreter space
697
698 // next_arg_comp is the next argument from the compiler point of
699 // view (inline type fields are passed in registers/on the stack). In
700 // sig_extended, an inline type argument starts with: T_METADATA,
701 // followed by the types of the fields of the inline type and T_VOID
702 // to mark the end of the inline type. ignored counts the number of
703 // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
704 // used to get the buffer for that argument from the pool of buffers
705 // we allocated above and want to pass to the
706 // interpreter. next_arg_int is the next argument from the
707 // interpreter point of view (inline types are passed by reference).
708 for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
709 next_arg_comp < sig_extended->length(); next_arg_comp++) {
710 assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
711 assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
712 BasicType bt = sig_extended->at(next_arg_comp)._bt;
713 int st_off = (total_args_passed - next_arg_int - 1) * Interpreter::stackElementSize;
714 if (!InlineTypePassFieldsAsArgs || bt != T_METADATA) {
715 int next_off = st_off - Interpreter::stackElementSize;
716 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
717 const VMRegPair reg_pair = regs[next_arg_comp-ignored];
718 size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
719 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
720 size_in_bytes, reg_pair, Address(sp, offset), tmp1, tmp2, tmp3, extraspace, false);
721 next_arg_int++;
722 #ifdef ASSERT
723 if (bt == T_LONG || bt == T_DOUBLE) {
724 // Overwrite the unused slot with known junk
725 __ mov(rscratch1, CONST64(0xdeadffffdeadaaaa));
726 __ str(rscratch1, Address(sp, st_off));
727 }
728 #endif /* ASSERT */
729 } else {
730 ignored++;
731 // get the buffer from the just allocated pool of buffers
732 int index = arrayOopDesc::base_offset_in_bytes(T_OBJECT) + next_vt_arg * type2aelembytes(T_OBJECT);
733 __ load_heap_oop(buf_oop, Address(buf_array, index), tmp1, tmp2);
734 next_vt_arg++; next_arg_int++;
735 int vt = 1;
736 // write fields we get from compiled code in registers/stack
737 // slots to the buffer: we know we are done with that inline type
738 // argument when we hit the T_VOID that acts as an end of inline
739 // type delimiter for this inline type. Inline types are flattened
740 // so we might encounter embedded inline types. Each entry in
741 // sig_extended contains a field offset in the buffer.
742 Label L_null;
743 do {
744 next_arg_comp++;
745 BasicType bt = sig_extended->at(next_arg_comp)._bt;
746 BasicType prev_bt = sig_extended->at(next_arg_comp - 1)._bt;
747 if (bt == T_METADATA) {
748 vt++;
749 ignored++;
750 } else if (bt == T_VOID && prev_bt != T_LONG && prev_bt != T_DOUBLE) {
751 vt--;
752 ignored++;
753 } else {
754 int off = sig_extended->at(next_arg_comp)._offset;
755 if (off == -1) {
756 // Nullable inline type argument, emit null check
757 VMReg reg = regs[next_arg_comp-ignored].first();
758 Label L_notNull;
759 if (reg->is_stack()) {
760 int ld_off = reg->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
761 __ ldrb(tmp1, Address(sp, ld_off));
762 __ cbnz(tmp1, L_notNull);
763 } else {
764 __ cbnz(reg->as_Register(), L_notNull);
765 }
766 __ str(zr, Address(sp, st_off));
767 __ b(L_null);
768 __ bind(L_notNull);
769 continue;
770 }
771 assert(off > 0, "offset in object should be positive");
772 size_t size_in_bytes = is_java_primitive(bt) ? type2aelembytes(bt) : wordSize;
773 bool is_oop = is_reference_type(bt);
774 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp-1)._bt : T_ILLEGAL,
775 size_in_bytes, regs[next_arg_comp-ignored], Address(buf_oop, off), tmp1, tmp2, tmp3, extraspace, is_oop);
776 }
777 } while (vt != 0);
778 // pass the buffer to the interpreter
779 __ str(buf_oop, Address(sp, st_off));
780 __ bind(L_null);
781 }
782 }
783
784 __ mov(esp, sp); // Interp expects args on caller's expression stack
785
786 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::interpreter_entry_offset())));
787 __ br(rscratch1);
788 }
789
790 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm, int comp_args_on_stack, const GrowableArray<SigEntry>* sig, const VMRegPair *regs) {
791
792
793 // Note: r19_sender_sp contains the senderSP on entry. We must
794 // preserve it since we may do a i2c -> c2i transition if we lose a
795 // race where compiled code goes non-entrant while we get args
796 // ready.
797
798 // Adapters are frameless.
799
800 // An i2c adapter is frameless because the *caller* frame, which is
801 // interpreted, routinely repairs its own esp (from
802 // interpreter_frame_last_sp), even if a callee has modified the
803 // stack pointer. It also recalculates and aligns sp.
804
805 // A c2i adapter is frameless because the *callee* frame, which is
806 // interpreted, routinely repairs its caller's sp (from sender_sp,
807 // which is set up via the senderSP register).
808
809 // In other words, if *either* the caller or callee is interpreted, we can
810 // get the stack pointer repaired after a call.
811
812 // This is why c2i and i2c adapters cannot be indefinitely composed.
813 // In particular, if a c2i adapter were to somehow call an i2c adapter,
814 // both caller and callee would be compiled methods, and neither would
815 // clean up the stack pointer changes performed by the two adapters.
816 // If this happens, control eventually transfers back to the compiled
817 // caller, but with an uncorrected stack, causing delayed havoc.
818
819 if (VerifyAdapterCalls &&
820 (Interpreter::code() != nullptr || StubRoutines::final_stubs_code() != nullptr)) {
821 #if 0
822 // So, let's test for cascading c2i/i2c adapters right now.
823 // assert(Interpreter::contains($return_addr) ||
824 // StubRoutines::contains($return_addr),
825 // "i2c adapter must return to an interpreter frame");
826 __ block_comment("verify_i2c { ");
827 Label L_ok;
828 if (Interpreter::code() != nullptr) {
829 range_check(masm, rax, r11,
830 Interpreter::code()->code_start(), Interpreter::code()->code_end(),
831 L_ok);
832 }
833 if (StubRoutines::initial_stubs_code() != nullptr) {
834 range_check(masm, rax, r11,
835 StubRoutines::initial_stubs_code()->code_begin(),
836 StubRoutines::initial_stubs_code()->code_end(),
837 L_ok);
838 }
839 if (StubRoutines::final_stubs_code() != nullptr) {
840 range_check(masm, rax, r11,
841 StubRoutines::final_stubs_code()->code_begin(),
842 StubRoutines::final_stubs_code()->code_end(),
843 L_ok);
844 }
845 const char* msg = "i2c adapter must return to an interpreter frame";
846 __ block_comment(msg);
847 __ stop(msg);
848 __ bind(L_ok);
849 __ block_comment("} verify_i2ce ");
850 #endif
851 }
852
853 // Cut-out for having no stack args.
854 int comp_words_on_stack = 0;
855 if (comp_args_on_stack) {
856 comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
857 __ sub(rscratch1, sp, comp_words_on_stack * wordSize);
858 __ andr(sp, rscratch1, -16);
859 }
860
861 // Will jump to the compiled code just as if compiled code was doing it.
862 // Pre-load the register-jump target early, to schedule it better.
863 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::from_compiled_inline_offset())));
864
865 #if INCLUDE_JVMCI
866 if (EnableJVMCI) {
867 // check if this call should be routed towards a specific entry point
868 __ ldr(rscratch2, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
869 Label no_alternative_target;
870 __ cbz(rscratch2, no_alternative_target);
871 __ mov(rscratch1, rscratch2);
872 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
873 __ bind(no_alternative_target);
874 }
875 #endif // INCLUDE_JVMCI
876
877 int total_args_passed = sig->length();
878
879 // Now generate the shuffle code.
880 for (int i = 0; i < total_args_passed; i++) {
881 BasicType bt = sig->at(i)._bt;
882 if (bt == T_VOID) {
883 assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
884 continue;
885 }
886
887 // Pick up 0, 1 or 2 words from SP+offset.
888 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(), "scrambled load targets?");
889
890 // Load in argument order going down.
891 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
892 // Point to interpreter value (vs. tag)
893 int next_off = ld_off - Interpreter::stackElementSize;
894 //
895 //
896 //
897 VMReg r_1 = regs[i].first();
898 VMReg r_2 = regs[i].second();
899 if (!r_1->is_valid()) {
900 assert(!r_2->is_valid(), "");
901 continue;
902 }
903 if (r_1->is_stack()) {
904 // Convert stack slot to an SP offset (+ wordSize to account for return address )
905 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
906 if (!r_2->is_valid()) {
907 // sign extend???
908 __ ldrsw(rscratch2, Address(esp, ld_off));
909 __ str(rscratch2, Address(sp, st_off));
910 } else {
911 //
912 // We are using two optoregs. This can be either T_OBJECT,
913 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
914 // two slots but only uses one for thr T_LONG or T_DOUBLE case
915 // So we must adjust where to pick up the data to match the
916 // interpreter.
917 //
918 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
919 // are accessed as negative so LSW is at LOW address
920
921 // ld_off is MSW so get LSW
922 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
923 __ ldr(rscratch2, Address(esp, offset));
924 // st_off is LSW (i.e. reg.first())
925 __ str(rscratch2, Address(sp, st_off));
926 }
927 } else if (r_1->is_Register()) { // Register argument
928 Register r = r_1->as_Register();
929 if (r_2->is_valid()) {
930 //
931 // We are using two VMRegs. This can be either T_OBJECT,
932 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
933 // two slots but only uses one for thr T_LONG or T_DOUBLE case
934 // So we must adjust where to pick up the data to match the
935 // interpreter.
936
937 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
938
939 // this can be a misaligned move
940 __ ldr(r, Address(esp, offset));
941 } else {
942 // sign extend and use a full word?
943 __ ldrw(r, Address(esp, ld_off));
944 }
945 } else {
946 if (!r_2->is_valid()) {
947 __ ldrs(r_1->as_FloatRegister(), Address(esp, ld_off));
948 } else {
949 __ ldrd(r_1->as_FloatRegister(), Address(esp, next_off));
950 }
951 }
952 }
953
954
955 __ mov(rscratch2, rscratch1);
956 __ push_cont_fastpath(rthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about; kills rscratch1
957 __ mov(rscratch1, rscratch2);
958
959 // 6243940 We might end up in handle_wrong_method if
960 // the callee is deoptimized as we race thru here. If that
961 // happens we don't want to take a safepoint because the
962 // caller frame will look interpreted and arguments are now
963 // "compiled" so it is much better to make this transition
964 // invisible to the stack walking code. Unfortunately if
965 // we try and find the callee by normal means a safepoint
966 // is possible. So we stash the desired callee in the thread
967 // and the vm will find there should this case occur.
968
969 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset()));
970 __ br(rscratch1);
971 }
972
973 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
974 Register data = rscratch2;
975 __ ic_check(1 /* end_alignment */);
976 __ ldr(rmethod, Address(data, CompiledICData::speculated_method_offset()));
977
978 // Method might have been compiled since the call site was patched to
979 // interpreted; if that is the case treat it as a miss so we can get
980 // the call site corrected.
981 __ ldr(rscratch1, Address(rmethod, in_bytes(Method::code_offset())));
982 __ cbz(rscratch1, skip_fixup);
983 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
984 }
985
986 // ---------------------------------------------------------------
987 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
988 int comp_args_on_stack,
989 const GrowableArray<SigEntry>* sig,
990 const VMRegPair* regs,
991 const GrowableArray<SigEntry>* sig_cc,
992 const VMRegPair* regs_cc,
993 const GrowableArray<SigEntry>* sig_cc_ro,
994 const VMRegPair* regs_cc_ro,
995 AdapterFingerPrint* fingerprint,
996 AdapterBlob*& new_adapter,
997 bool allocate_code_blob) {
998
999 address i2c_entry = __ pc();
1000 gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
1001
1002 // -------------------------------------------------------------------------
1003 // Generate a C2I adapter. On entry we know rmethod holds the Method* during calls
1004 // to the interpreter. The args start out packed in the compiled layout. They
1005 // need to be unpacked into the interpreter layout. This will almost always
1006 // require some stack space. We grow the current (compiled) stack, then repack
1007 // the args. We finally end in a jump to the generic interpreter entry point.
1008 // On exit from the interpreter, the interpreter will restore our SP (lest the
1009 // compiled code, which relies solely on SP and not FP, get sick).
1010
1011 address c2i_unverified_entry = __ pc();
1012 address c2i_unverified_inline_entry = __ pc();
1013 Label skip_fixup;
1014
1015 gen_inline_cache_check(masm, skip_fixup);
1016
1017 OopMapSet* oop_maps = new OopMapSet();
1018 int frame_complete = CodeOffsets::frame_never_safe;
1019 int frame_size_in_words = 0;
1020
1021 // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
1022 address c2i_no_clinit_check_entry = nullptr;
1023 address c2i_inline_ro_entry = __ pc();
1024 if (regs_cc != regs_cc_ro) {
1025 // No class init barrier needed because method is guaranteed to be non-static
1026 gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, c2i_no_clinit_check_entry,
1027 skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1028 skip_fixup.reset();
1029 }
1030
1031 // Scalarized c2i adapter
1032 address c2i_entry = __ pc();
1033 address c2i_inline_entry = __ pc();
1034 gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1035 skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
1036
1037 // Non-scalarized c2i adapter
1038 if (regs != regs_cc) {
1039 c2i_unverified_inline_entry = __ pc();
1040 Label inline_entry_skip_fixup;
1041 gen_inline_cache_check(masm, inline_entry_skip_fixup);
1042
1043 c2i_inline_entry = __ pc();
1044 gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, c2i_no_clinit_check_entry,
1045 inline_entry_skip_fixup, i2c_entry, oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
1046 }
1047
1048
1049 // The c2i adapter might safepoint and trigger a GC. The caller must make sure that
1050 // the GC knows about the location of oop argument locations passed to the c2i adapter.
1051 if (allocate_code_blob) {
1052 bool caller_must_gc_arguments = (regs != regs_cc);
1053 new_adapter = AdapterBlob::create(masm->code(), frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
1054 }
1055
1056 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_inline_entry, c2i_inline_ro_entry, c2i_unverified_entry, c2i_unverified_inline_entry, c2i_no_clinit_check_entry);
1057 }
1058
1059 static int c_calling_convention_priv(const BasicType *sig_bt,
1060 VMRegPair *regs,
1061 int total_args_passed) {
1062
1063 // We return the amount of VMRegImpl stack slots we need to reserve for all
1064 // the arguments NOT counting out_preserve_stack_slots.
1065
1066 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
1067 c_rarg0, c_rarg1, c_rarg2, c_rarg3, c_rarg4, c_rarg5, c_rarg6, c_rarg7
1068 };
1069 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
1070 c_farg0, c_farg1, c_farg2, c_farg3,
1071 c_farg4, c_farg5, c_farg6, c_farg7
1072 };
1073
1074 uint int_args = 0;
1075 uint fp_args = 0;
1076 uint stk_args = 0; // inc by 2 each time
1077
1078 for (int i = 0; i < total_args_passed; i++) {
1079 switch (sig_bt[i]) {
1080 case T_BOOLEAN:
1081 case T_CHAR:
1082 case T_BYTE:
1083 case T_SHORT:
1084 case T_INT:
1085 if (int_args < Argument::n_int_register_parameters_c) {
1086 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
1087 } else {
1088 #ifdef __APPLE__
1089 // Less-than word types are stored one after another.
1090 // The code is unable to handle this so bailout.
1091 return -1;
1092 #endif
1093 regs[i].set1(VMRegImpl::stack2reg(stk_args));
1094 stk_args += 2;
1095 }
1096 break;
1097 case T_LONG:
1098 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1099 // fall through
1100 case T_OBJECT:
1101 case T_ARRAY:
1102 case T_ADDRESS:
1103 case T_METADATA:
1104 if (int_args < Argument::n_int_register_parameters_c) {
1105 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
1106 } else {
1107 regs[i].set2(VMRegImpl::stack2reg(stk_args));
1108 stk_args += 2;
1109 }
1110 break;
1111 case T_FLOAT:
1112 if (fp_args < Argument::n_float_register_parameters_c) {
1113 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
1114 } else {
1115 #ifdef __APPLE__
1116 // Less-than word types are stored one after another.
1117 // The code is unable to handle this so bailout.
1118 return -1;
1119 #endif
1120 regs[i].set1(VMRegImpl::stack2reg(stk_args));
1121 stk_args += 2;
1122 }
1123 break;
1124 case T_DOUBLE:
1125 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
1126 if (fp_args < Argument::n_float_register_parameters_c) {
1127 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
1128 } else {
1129 regs[i].set2(VMRegImpl::stack2reg(stk_args));
1130 stk_args += 2;
1131 }
1132 break;
1133 case T_VOID: // Halves of longs and doubles
1134 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
1135 regs[i].set_bad();
1136 break;
1137 default:
1138 ShouldNotReachHere();
1139 break;
1140 }
1141 }
1142
1143 return stk_args;
1144 }
1145
1146 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
1147 uint num_bits,
1148 uint total_args_passed) {
1149 Unimplemented();
1150 return 0;
1151 }
1152
1153 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1154 VMRegPair *regs,
1155 int total_args_passed)
1156 {
1157 int result = c_calling_convention_priv(sig_bt, regs, total_args_passed);
1158 guarantee(result >= 0, "Unsupported arguments configuration");
1159 return result;
1160 }
1161
1162
1163 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1164 // We always ignore the frame_slots arg and just use the space just below frame pointer
1165 // which by this time is free to use
1166 switch (ret_type) {
1167 case T_FLOAT:
1168 __ strs(v0, Address(rfp, -wordSize));
1169 break;
1170 case T_DOUBLE:
1171 __ strd(v0, Address(rfp, -wordSize));
1172 break;
1173 case T_VOID: break;
1174 default: {
1175 __ str(r0, Address(rfp, -wordSize));
1176 }
1177 }
1178 }
1179
1180 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1181 // We always ignore the frame_slots arg and just use the space just below frame pointer
1182 // which by this time is free to use
1183 switch (ret_type) {
1184 case T_FLOAT:
1185 __ ldrs(v0, Address(rfp, -wordSize));
1186 break;
1187 case T_DOUBLE:
1188 __ ldrd(v0, Address(rfp, -wordSize));
1189 break;
1190 case T_VOID: break;
1191 default: {
1192 __ ldr(r0, Address(rfp, -wordSize));
1193 }
1194 }
1195 }
1196 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1197 RegSet x;
1198 for ( int i = first_arg ; i < arg_count ; i++ ) {
1199 if (args[i].first()->is_Register()) {
1200 x = x + args[i].first()->as_Register();
1201 } else if (args[i].first()->is_FloatRegister()) {
1202 __ strd(args[i].first()->as_FloatRegister(), Address(__ pre(sp, -2 * wordSize)));
1203 }
1204 }
1205 __ push(x, sp);
1206 }
1207
1208 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
1209 RegSet x;
1210 for ( int i = first_arg ; i < arg_count ; i++ ) {
1211 if (args[i].first()->is_Register()) {
1212 x = x + args[i].first()->as_Register();
1213 } else {
1214 ;
1215 }
1216 }
1217 __ pop(x, sp);
1218 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
1219 if (args[i].first()->is_Register()) {
1220 ;
1221 } else if (args[i].first()->is_FloatRegister()) {
1222 __ ldrd(args[i].first()->as_FloatRegister(), Address(__ post(sp, 2 * wordSize)));
1223 }
1224 }
1225 }
1226
1227 static void verify_oop_args(MacroAssembler* masm,
1228 const methodHandle& method,
1229 const BasicType* sig_bt,
1230 const VMRegPair* regs) {
1231 Register temp_reg = r19; // not part of any compiled calling seq
1232 if (VerifyOops) {
1233 for (int i = 0; i < method->size_of_parameters(); i++) {
1234 if (sig_bt[i] == T_OBJECT ||
1235 sig_bt[i] == T_ARRAY) {
1236 VMReg r = regs[i].first();
1237 assert(r->is_valid(), "bad oop arg");
1238 if (r->is_stack()) {
1239 __ ldr(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1240 __ verify_oop(temp_reg);
1241 } else {
1242 __ verify_oop(r->as_Register());
1243 }
1244 }
1245 }
1246 }
1247 }
1248
1249 // on exit, sp points to the ContinuationEntry
1250 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
1251 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1252 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1253 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1254
1255 stack_slots += (int)ContinuationEntry::size()/wordSize;
1256 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
1257
1258 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize)/ VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1259
1260 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1261 __ str(rscratch1, Address(sp, ContinuationEntry::parent_offset()));
1262 __ mov(rscratch1, sp); // we can't use sp as the source in str
1263 __ str(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1264
1265 return map;
1266 }
1267
1268 // on entry c_rarg1 points to the continuation
1269 // sp points to ContinuationEntry
1270 // c_rarg3 -- isVirtualThread
1271 static void fill_continuation_entry(MacroAssembler* masm) {
1272 #ifdef ASSERT
1273 __ movw(rscratch1, ContinuationEntry::cookie_value());
1274 __ strw(rscratch1, Address(sp, ContinuationEntry::cookie_offset()));
1275 #endif
1276
1277 __ str (c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
1278 __ strw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
1279 __ str (zr, Address(sp, ContinuationEntry::chunk_offset()));
1280 __ strw(zr, Address(sp, ContinuationEntry::argsize_offset()));
1281 __ strw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
1282
1283 __ ldr(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1284 __ str(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1285 __ ldr(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset()));
1286 __ str(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
1287
1288 __ str(zr, Address(rthread, JavaThread::cont_fastpath_offset()));
1289 __ str(zr, Address(rthread, JavaThread::held_monitor_count_offset()));
1290 }
1291
1292 // on entry, sp points to the ContinuationEntry
1293 // on exit, rfp points to the spilled rfp in the entry frame
1294 static void continuation_enter_cleanup(MacroAssembler* masm) {
1295 #ifndef PRODUCT
1296 Label OK;
1297 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1298 __ cmp(sp, rscratch1);
1299 __ br(Assembler::EQ, OK);
1300 __ stop("incorrect sp1");
1301 __ bind(OK);
1302 #endif
1303 __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1304 __ str(rscratch1, Address(rthread, JavaThread::cont_fastpath_offset()));
1305
1306 if (CheckJNICalls) {
1307 // Check if this is a virtual thread continuation
1308 Label L_skip_vthread_code;
1309 __ ldrw(rscratch1, Address(sp, ContinuationEntry::flags_offset()));
1310 __ cbzw(rscratch1, L_skip_vthread_code);
1311
1312 // If the held monitor count is > 0 and this vthread is terminating then
1313 // it failed to release a JNI monitor. So we issue the same log message
1314 // that JavaThread::exit does.
1315 __ ldr(rscratch1, Address(rthread, JavaThread::jni_monitor_count_offset()));
1316 __ cbz(rscratch1, L_skip_vthread_code);
1317
1318 // Save return value potentially containing the exception oop in callee-saved R19.
1319 __ mov(r19, r0);
1320 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1321 // Restore potential return value.
1322 __ mov(r0, r19);
1323
1324 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1325 // on termination. The held count is implicitly zeroed below when we restore from
1326 // the parent held count (which has to be zero).
1327 __ str(zr, Address(rthread, JavaThread::jni_monitor_count_offset()));
1328
1329 __ bind(L_skip_vthread_code);
1330 }
1331 #ifdef ASSERT
1332 else {
1333 // Check if this is a virtual thread continuation
1334 Label L_skip_vthread_code;
1335 __ ldrw(rscratch1, Address(sp, ContinuationEntry::flags_offset()));
1336 __ cbzw(rscratch1, L_skip_vthread_code);
1337
1338 // See comment just above. If not checking JNI calls the JNI count is only
1339 // needed for assertion checking.
1340 __ str(zr, Address(rthread, JavaThread::jni_monitor_count_offset()));
1341
1342 __ bind(L_skip_vthread_code);
1343 }
1344 #endif
1345
1346 __ ldr(rscratch1, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
1347 __ str(rscratch1, Address(rthread, JavaThread::held_monitor_count_offset()));
1348
1349 __ ldr(rscratch2, Address(sp, ContinuationEntry::parent_offset()));
1350 __ str(rscratch2, Address(rthread, JavaThread::cont_entry_offset()));
1351 __ add(rfp, sp, (int)ContinuationEntry::size());
1352 }
1353
1354 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1355 // On entry: c_rarg1 -- the continuation object
1356 // c_rarg2 -- isContinue
1357 // c_rarg3 -- isVirtualThread
1358 static void gen_continuation_enter(MacroAssembler* masm,
1359 const methodHandle& method,
1360 const BasicType* sig_bt,
1361 const VMRegPair* regs,
1362 int& exception_offset,
1363 OopMapSet*oop_maps,
1364 int& frame_complete,
1365 int& stack_slots,
1366 int& interpreted_entry_offset,
1367 int& compiled_entry_offset) {
1368 //verify_oop_args(masm, method, sig_bt, regs);
1369 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1370
1371 address start = __ pc();
1372
1373 Label call_thaw, exit;
1374
1375 // i2i entry used at interp_only_mode only
1376 interpreted_entry_offset = __ pc() - start;
1377 {
1378
1379 #ifdef ASSERT
1380 Label is_interp_only;
1381 __ ldrw(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
1382 __ cbnzw(rscratch1, is_interp_only);
1383 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1384 __ bind(is_interp_only);
1385 #endif
1386
1387 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1388 __ ldr(c_rarg1, Address(esp, Interpreter::stackElementSize*2));
1389 __ ldr(c_rarg2, Address(esp, Interpreter::stackElementSize*1));
1390 __ ldr(c_rarg3, Address(esp, Interpreter::stackElementSize*0));
1391 __ push_cont_fastpath(rthread);
1392
1393 __ enter();
1394 stack_slots = 2; // will be adjusted in setup
1395 OopMap* map = continuation_enter_setup(masm, stack_slots);
1396 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1397 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1398
1399 fill_continuation_entry(masm);
1400
1401 __ cbnz(c_rarg2, call_thaw);
1402
1403 const address tr_call = __ trampoline_call(resolve);
1404 if (tr_call == nullptr) {
1405 fatal("CodeCache is full at gen_continuation_enter");
1406 }
1407
1408 oop_maps->add_gc_map(__ pc() - start, map);
1409 __ post_call_nop();
1410
1411 __ b(exit);
1412
1413 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1414 if (stub == nullptr) {
1415 fatal("CodeCache is full at gen_continuation_enter");
1416 }
1417 }
1418
1419 // compiled entry
1420 __ align(CodeEntryAlignment);
1421 compiled_entry_offset = __ pc() - start;
1422
1423 __ enter();
1424 stack_slots = 2; // will be adjusted in setup
1425 OopMap* map = continuation_enter_setup(masm, stack_slots);
1426 frame_complete = __ pc() - start;
1427
1428 fill_continuation_entry(masm);
1429
1430 __ cbnz(c_rarg2, call_thaw);
1431
1432 const address tr_call = __ trampoline_call(resolve);
1433 if (tr_call == nullptr) {
1434 fatal("CodeCache is full at gen_continuation_enter");
1435 }
1436
1437 oop_maps->add_gc_map(__ pc() - start, map);
1438 __ post_call_nop();
1439
1440 __ b(exit);
1441
1442 __ bind(call_thaw);
1443
1444 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1445 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1446 ContinuationEntry::_return_pc_offset = __ pc() - start;
1447 __ post_call_nop();
1448
1449 __ bind(exit);
1450 continuation_enter_cleanup(masm);
1451 __ leave();
1452 __ ret(lr);
1453
1454 /// exception handling
1455
1456 exception_offset = __ pc() - start;
1457 {
1458 __ mov(r19, r0); // save return value contaning the exception oop in callee-saved R19
1459
1460 continuation_enter_cleanup(masm);
1461
1462 __ ldr(c_rarg1, Address(rfp, wordSize)); // return address
1463 __ authenticate_return_address(c_rarg1);
1464 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1);
1465
1466 // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc
1467
1468 __ mov(r1, r0); // the exception handler
1469 __ mov(r0, r19); // restore return value contaning the exception oop
1470 __ verify_oop(r0);
1471
1472 __ leave();
1473 __ mov(r3, lr);
1474 __ br(r1); // the exception handler
1475 }
1476
1477 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1478 if (stub == nullptr) {
1479 fatal("CodeCache is full at gen_continuation_enter");
1480 }
1481 }
1482
1483 static void gen_continuation_yield(MacroAssembler* masm,
1484 const methodHandle& method,
1485 const BasicType* sig_bt,
1486 const VMRegPair* regs,
1487 OopMapSet* oop_maps,
1488 int& frame_complete,
1489 int& stack_slots,
1490 int& compiled_entry_offset) {
1491 enum layout {
1492 rfp_off1,
1493 rfp_off2,
1494 lr_off,
1495 lr_off2,
1496 framesize // inclusive of return address
1497 };
1498 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1499 stack_slots = framesize / VMRegImpl::slots_per_word;
1500 assert(stack_slots == 2, "recheck layout");
1501
1502 address start = __ pc();
1503
1504 compiled_entry_offset = __ pc() - start;
1505 __ enter();
1506
1507 __ mov(c_rarg1, sp);
1508
1509 frame_complete = __ pc() - start;
1510 address the_pc = __ pc();
1511
1512 __ 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
1513
1514 __ mov(c_rarg0, rthread);
1515 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
1516 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1517 __ reset_last_Java_frame(true);
1518
1519 Label pinned;
1520
1521 __ cbnz(r0, pinned);
1522
1523 // We've succeeded, set sp to the ContinuationEntry
1524 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
1525 __ mov(sp, rscratch1);
1526 continuation_enter_cleanup(masm);
1527
1528 __ bind(pinned); // pinned -- return to caller
1529
1530 // handle pending exception thrown by freeze
1531 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
1532 Label ok;
1533 __ cbz(rscratch1, ok);
1534 __ leave();
1535 __ lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
1536 __ br(rscratch1);
1537 __ bind(ok);
1538
1539 __ leave();
1540 __ ret(lr);
1541
1542 OopMap* map = new OopMap(framesize, 1);
1543 oop_maps->add_gc_map(the_pc - start, map);
1544 }
1545
1546 static void gen_special_dispatch(MacroAssembler* masm,
1547 const methodHandle& method,
1548 const BasicType* sig_bt,
1549 const VMRegPair* regs) {
1550 verify_oop_args(masm, method, sig_bt, regs);
1551 vmIntrinsics::ID iid = method->intrinsic_id();
1552
1553 // Now write the args into the outgoing interpreter space
1554 bool has_receiver = false;
1555 Register receiver_reg = noreg;
1556 int member_arg_pos = -1;
1557 Register member_reg = noreg;
1558 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1559 if (ref_kind != 0) {
1560 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1561 member_reg = r19; // known to be free at this point
1562 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1563 } else if (iid == vmIntrinsics::_invokeBasic) {
1564 has_receiver = true;
1565 } else if (iid == vmIntrinsics::_linkToNative) {
1566 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1567 member_reg = r19; // known to be free at this point
1568 } else {
1569 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1570 }
1571
1572 if (member_reg != noreg) {
1573 // Load the member_arg into register, if necessary.
1574 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1575 VMReg r = regs[member_arg_pos].first();
1576 if (r->is_stack()) {
1577 __ ldr(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1578 } else {
1579 // no data motion is needed
1580 member_reg = r->as_Register();
1581 }
1582 }
1583
1584 if (has_receiver) {
1585 // Make sure the receiver is loaded into a register.
1586 assert(method->size_of_parameters() > 0, "oob");
1587 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1588 VMReg r = regs[0].first();
1589 assert(r->is_valid(), "bad receiver arg");
1590 if (r->is_stack()) {
1591 // Porting note: This assumes that compiled calling conventions always
1592 // pass the receiver oop in a register. If this is not true on some
1593 // platform, pick a temp and load the receiver from stack.
1594 fatal("receiver always in a register");
1595 receiver_reg = r2; // known to be free at this point
1596 __ ldr(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1597 } else {
1598 // no data motion is needed
1599 receiver_reg = r->as_Register();
1600 }
1601 }
1602
1603 // Figure out which address we are really jumping to:
1604 MethodHandles::generate_method_handle_dispatch(masm, iid,
1605 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1606 }
1607
1608 // ---------------------------------------------------------------------------
1609 // Generate a native wrapper for a given method. The method takes arguments
1610 // in the Java compiled code convention, marshals them to the native
1611 // convention (handlizes oops, etc), transitions to native, makes the call,
1612 // returns to java state (possibly blocking), unhandlizes any result and
1613 // returns.
1614 //
1615 // Critical native functions are a shorthand for the use of
1616 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1617 // functions. The wrapper is expected to unpack the arguments before
1618 // passing them to the callee. Critical native functions leave the state _in_Java,
1619 // since they block out GC.
1620 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1621 // block and the check for pending exceptions it's impossible for them
1622 // to be thrown.
1623 //
1624 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1625 const methodHandle& method,
1626 int compile_id,
1627 BasicType* in_sig_bt,
1628 VMRegPair* in_regs,
1629 BasicType ret_type) {
1630 if (method->is_continuation_native_intrinsic()) {
1631 int exception_offset = -1;
1632 OopMapSet* oop_maps = new OopMapSet();
1633 int frame_complete = -1;
1634 int stack_slots = -1;
1635 int interpreted_entry_offset = -1;
1636 int vep_offset = -1;
1637 if (method->is_continuation_enter_intrinsic()) {
1638 gen_continuation_enter(masm,
1639 method,
1640 in_sig_bt,
1641 in_regs,
1642 exception_offset,
1643 oop_maps,
1644 frame_complete,
1645 stack_slots,
1646 interpreted_entry_offset,
1647 vep_offset);
1648 } else if (method->is_continuation_yield_intrinsic()) {
1649 gen_continuation_yield(masm,
1650 method,
1651 in_sig_bt,
1652 in_regs,
1653 oop_maps,
1654 frame_complete,
1655 stack_slots,
1656 vep_offset);
1657 } else {
1658 guarantee(false, "Unknown Continuation native intrinsic");
1659 }
1660
1661 #ifdef ASSERT
1662 if (method->is_continuation_enter_intrinsic()) {
1663 assert(interpreted_entry_offset != -1, "Must be set");
1664 assert(exception_offset != -1, "Must be set");
1665 } else {
1666 assert(interpreted_entry_offset == -1, "Must be unset");
1667 assert(exception_offset == -1, "Must be unset");
1668 }
1669 assert(frame_complete != -1, "Must be set");
1670 assert(stack_slots != -1, "Must be set");
1671 assert(vep_offset != -1, "Must be set");
1672 #endif
1673
1674 __ flush();
1675 nmethod* nm = nmethod::new_native_nmethod(method,
1676 compile_id,
1677 masm->code(),
1678 vep_offset,
1679 frame_complete,
1680 stack_slots,
1681 in_ByteSize(-1),
1682 in_ByteSize(-1),
1683 oop_maps,
1684 exception_offset);
1685 if (nm == nullptr) return nm;
1686 if (method->is_continuation_enter_intrinsic()) {
1687 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1688 } else if (method->is_continuation_yield_intrinsic()) {
1689 _cont_doYield_stub = nm;
1690 } else {
1691 guarantee(false, "Unknown Continuation native intrinsic");
1692 }
1693 return nm;
1694 }
1695
1696 if (method->is_method_handle_intrinsic()) {
1697 vmIntrinsics::ID iid = method->intrinsic_id();
1698 intptr_t start = (intptr_t)__ pc();
1699 int vep_offset = ((intptr_t)__ pc()) - start;
1700
1701 // First instruction must be a nop as it may need to be patched on deoptimisation
1702 __ nop();
1703 gen_special_dispatch(masm,
1704 method,
1705 in_sig_bt,
1706 in_regs);
1707 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1708 __ flush();
1709 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1710 return nmethod::new_native_nmethod(method,
1711 compile_id,
1712 masm->code(),
1713 vep_offset,
1714 frame_complete,
1715 stack_slots / VMRegImpl::slots_per_word,
1716 in_ByteSize(-1),
1717 in_ByteSize(-1),
1718 nullptr);
1719 }
1720 address native_func = method->native_function();
1721 assert(native_func != nullptr, "must have function");
1722
1723 // An OopMap for lock (and class if static)
1724 OopMapSet *oop_maps = new OopMapSet();
1725 intptr_t start = (intptr_t)__ pc();
1726
1727 // We have received a description of where all the java arg are located
1728 // on entry to the wrapper. We need to convert these args to where
1729 // the jni function will expect them. To figure out where they go
1730 // we convert the java signature to a C signature by inserting
1731 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1732
1733 const int total_in_args = method->size_of_parameters();
1734 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1735
1736 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1737 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1738 BasicType* in_elem_bt = nullptr;
1739
1740 int argc = 0;
1741 out_sig_bt[argc++] = T_ADDRESS;
1742 if (method->is_static()) {
1743 out_sig_bt[argc++] = T_OBJECT;
1744 }
1745
1746 for (int i = 0; i < total_in_args ; i++ ) {
1747 out_sig_bt[argc++] = in_sig_bt[i];
1748 }
1749
1750 // Now figure out where the args must be stored and how much stack space
1751 // they require.
1752 int out_arg_slots;
1753 out_arg_slots = c_calling_convention_priv(out_sig_bt, out_regs, total_c_args);
1754
1755 if (out_arg_slots < 0) {
1756 return nullptr;
1757 }
1758
1759 // Compute framesize for the wrapper. We need to handlize all oops in
1760 // incoming registers
1761
1762 // Calculate the total number of stack slots we will need.
1763
1764 // First count the abi requirement plus all of the outgoing args
1765 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1766
1767 // Now the space for the inbound oop handle area
1768 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1769
1770 int oop_handle_offset = stack_slots;
1771 stack_slots += total_save_slots;
1772
1773 // Now any space we need for handlizing a klass if static method
1774
1775 int klass_slot_offset = 0;
1776 int klass_offset = -1;
1777 int lock_slot_offset = 0;
1778 bool is_static = false;
1779
1780 if (method->is_static()) {
1781 klass_slot_offset = stack_slots;
1782 stack_slots += VMRegImpl::slots_per_word;
1783 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1784 is_static = true;
1785 }
1786
1787 // Plus a lock if needed
1788
1789 if (method->is_synchronized()) {
1790 lock_slot_offset = stack_slots;
1791 stack_slots += VMRegImpl::slots_per_word;
1792 }
1793
1794 // Now a place (+2) to save return values or temp during shuffling
1795 // + 4 for return address (which we own) and saved rfp
1796 stack_slots += 6;
1797
1798 // Ok The space we have allocated will look like:
1799 //
1800 //
1801 // FP-> | |
1802 // |---------------------|
1803 // | 2 slots for moves |
1804 // |---------------------|
1805 // | lock box (if sync) |
1806 // |---------------------| <- lock_slot_offset
1807 // | klass (if static) |
1808 // |---------------------| <- klass_slot_offset
1809 // | oopHandle area |
1810 // |---------------------| <- oop_handle_offset (8 java arg registers)
1811 // | outbound memory |
1812 // | based arguments |
1813 // | |
1814 // |---------------------|
1815 // | |
1816 // SP-> | out_preserved_slots |
1817 //
1818 //
1819
1820
1821 // Now compute actual number of stack words we need rounding to make
1822 // stack properly aligned.
1823 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1824
1825 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1826
1827 // First thing make an ic check to see if we should even be here
1828
1829 // We are free to use all registers as temps without saving them and
1830 // restoring them except rfp. rfp is the only callee save register
1831 // as far as the interpreter and the compiler(s) are concerned.
1832
1833 const Register receiver = j_rarg0;
1834
1835 Label exception_pending;
1836
1837 assert_different_registers(receiver, rscratch1);
1838 __ verify_oop(receiver);
1839 __ ic_check(8 /* end_alignment */);
1840
1841 // Verified entry point must be aligned
1842 int vep_offset = ((intptr_t)__ pc()) - start;
1843
1844 // If we have to make this method not-entrant we'll overwrite its
1845 // first instruction with a jump. For this action to be legal we
1846 // must ensure that this first instruction is a B, BL, NOP, BKPT,
1847 // SVC, HVC, or SMC. Make it a NOP.
1848 __ nop();
1849
1850 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1851 Label L_skip_barrier;
1852 __ mov_metadata(rscratch2, method->method_holder()); // InstanceKlass*
1853 __ clinit_barrier(rscratch2, rscratch1, &L_skip_barrier);
1854 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1855
1856 __ bind(L_skip_barrier);
1857 }
1858
1859 // Generate stack overflow check
1860 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1861
1862 // Generate a new frame for the wrapper.
1863 __ enter();
1864 // -2 because return address is already present and so is saved rfp
1865 __ sub(sp, sp, stack_size - 2*wordSize);
1866
1867 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1868 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1869
1870 // Frame is now completed as far as size and linkage.
1871 int frame_complete = ((intptr_t)__ pc()) - start;
1872
1873 // We use r20 as the oop handle for the receiver/klass
1874 // It is callee save so it survives the call to native
1875
1876 const Register oop_handle_reg = r20;
1877
1878 //
1879 // We immediately shuffle the arguments so that any vm call we have to
1880 // make from here on out (sync slow path, jvmti, etc.) we will have
1881 // captured the oops from our caller and have a valid oopMap for
1882 // them.
1883
1884 // -----------------
1885 // The Grand Shuffle
1886
1887 // The Java calling convention is either equal (linux) or denser (win64) than the
1888 // c calling convention. However the because of the jni_env argument the c calling
1889 // convention always has at least one more (and two for static) arguments than Java.
1890 // Therefore if we move the args from java -> c backwards then we will never have
1891 // a register->register conflict and we don't have to build a dependency graph
1892 // and figure out how to break any cycles.
1893 //
1894
1895 // Record esp-based slot for receiver on stack for non-static methods
1896 int receiver_offset = -1;
1897
1898 // This is a trick. We double the stack slots so we can claim
1899 // the oops in the caller's frame. Since we are sure to have
1900 // more args than the caller doubling is enough to make
1901 // sure we can capture all the incoming oop args from the
1902 // caller.
1903 //
1904 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1905
1906 // Mark location of rfp (someday)
1907 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, vmreg(rfp));
1908
1909
1910 int float_args = 0;
1911 int int_args = 0;
1912
1913 #ifdef ASSERT
1914 bool reg_destroyed[Register::number_of_registers];
1915 bool freg_destroyed[FloatRegister::number_of_registers];
1916 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1917 reg_destroyed[r] = false;
1918 }
1919 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1920 freg_destroyed[f] = false;
1921 }
1922
1923 #endif /* ASSERT */
1924
1925 // For JNI natives the incoming and outgoing registers are offset upwards.
1926 GrowableArray<int> arg_order(2 * total_in_args);
1927 VMRegPair tmp_vmreg;
1928 tmp_vmreg.set2(r19->as_VMReg());
1929
1930 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1931 arg_order.push(i);
1932 arg_order.push(c_arg);
1933 }
1934
1935 int temploc = -1;
1936 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1937 int i = arg_order.at(ai);
1938 int c_arg = arg_order.at(ai + 1);
1939 __ block_comment(err_msg("move %d -> %d", i, c_arg));
1940 assert(c_arg != -1 && i != -1, "wrong order");
1941 #ifdef ASSERT
1942 if (in_regs[i].first()->is_Register()) {
1943 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1944 } else if (in_regs[i].first()->is_FloatRegister()) {
1945 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1946 }
1947 if (out_regs[c_arg].first()->is_Register()) {
1948 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1949 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1950 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1951 }
1952 #endif /* ASSERT */
1953 switch (in_sig_bt[i]) {
1954 case T_ARRAY:
1955 case T_OBJECT:
1956 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1957 ((i == 0) && (!is_static)),
1958 &receiver_offset);
1959 int_args++;
1960 break;
1961 case T_VOID:
1962 break;
1963
1964 case T_FLOAT:
1965 __ float_move(in_regs[i], out_regs[c_arg]);
1966 float_args++;
1967 break;
1968
1969 case T_DOUBLE:
1970 assert( i + 1 < total_in_args &&
1971 in_sig_bt[i + 1] == T_VOID &&
1972 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1973 __ double_move(in_regs[i], out_regs[c_arg]);
1974 float_args++;
1975 break;
1976
1977 case T_LONG :
1978 __ long_move(in_regs[i], out_regs[c_arg]);
1979 int_args++;
1980 break;
1981
1982 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1983
1984 default:
1985 __ move32_64(in_regs[i], out_regs[c_arg]);
1986 int_args++;
1987 }
1988 }
1989
1990 // point c_arg at the first arg that is already loaded in case we
1991 // need to spill before we call out
1992 int c_arg = total_c_args - total_in_args;
1993
1994 // Pre-load a static method's oop into c_rarg1.
1995 if (method->is_static()) {
1996
1997 // load oop into a register
1998 __ movoop(c_rarg1,
1999 JNIHandles::make_local(method->method_holder()->java_mirror()));
2000
2001 // Now handlize the static class mirror it's known not-null.
2002 __ str(c_rarg1, Address(sp, klass_offset));
2003 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2004
2005 // Now get the handle
2006 __ lea(c_rarg1, Address(sp, klass_offset));
2007 // and protect the arg if we must spill
2008 c_arg--;
2009 }
2010
2011 // Change state to native (we save the return address in the thread, since it might not
2012 // be pushed on the stack when we do a stack traversal).
2013 // We use the same pc/oopMap repeatedly when we call out
2014
2015 Label native_return;
2016 __ set_last_Java_frame(sp, noreg, native_return, rscratch1);
2017
2018 Label dtrace_method_entry, dtrace_method_entry_done;
2019 if (DTraceMethodProbes) {
2020 __ b(dtrace_method_entry);
2021 __ bind(dtrace_method_entry_done);
2022 }
2023
2024 // RedefineClasses() tracing support for obsolete method entry
2025 if (log_is_enabled(Trace, redefine, class, obsolete)) {
2026 // protect the args we've loaded
2027 save_args(masm, total_c_args, c_arg, out_regs);
2028 __ mov_metadata(c_rarg1, method());
2029 __ call_VM_leaf(
2030 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
2031 rthread, c_rarg1);
2032 restore_args(masm, total_c_args, c_arg, out_regs);
2033 }
2034
2035 // Lock a synchronized method
2036
2037 // Register definitions used by locking and unlocking
2038
2039 const Register swap_reg = r0;
2040 const Register obj_reg = r19; // Will contain the oop
2041 const Register lock_reg = r13; // Address of compiler lock object (BasicLock)
2042 const Register old_hdr = r13; // value of old header at unlock time
2043 const Register lock_tmp = r14; // Temporary used by lightweight_lock/unlock
2044 const Register tmp = lr;
2045
2046 Label slow_path_lock;
2047 Label lock_done;
2048
2049 if (method->is_synchronized()) {
2050 Label count;
2051 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
2052
2053 // Get the handle (the 2nd argument)
2054 __ mov(oop_handle_reg, c_rarg1);
2055
2056 // Get address of the box
2057
2058 __ lea(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2059
2060 // Load the oop from the handle
2061 __ ldr(obj_reg, Address(oop_handle_reg, 0));
2062
2063 if (LockingMode == LM_MONITOR) {
2064 __ b(slow_path_lock);
2065 } else if (LockingMode == LM_LEGACY) {
2066 // Load (object->mark() | 1) into swap_reg %r0
2067 __ ldr(rscratch1, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2068 __ orr(swap_reg, rscratch1, 1);
2069 if (EnableValhalla) {
2070 // Mask inline_type bit such that we go to the slow path if object is an inline type
2071 __ andr(swap_reg, swap_reg, ~((int) markWord::inline_type_bit_in_place));
2072 }
2073
2074 // Save (object->mark() | 1) into BasicLock's displaced header
2075 __ str(swap_reg, Address(lock_reg, mark_word_offset));
2076
2077 // src -> dest iff dest == r0 else r0 <- dest
2078 __ cmpxchg_obj_header(r0, lock_reg, obj_reg, rscratch1, count, /*fallthrough*/nullptr);
2079
2080 // Hmm should this move to the slow path code area???
2081
2082 // Test if the oopMark is an obvious stack pointer, i.e.,
2083 // 1) (mark & 3) == 0, and
2084 // 2) sp <= mark < mark + os::pagesize()
2085 // These 3 tests can be done by evaluating the following
2086 // expression: ((mark - sp) & (3 - os::vm_page_size())),
2087 // assuming both stack pointer and pagesize have their
2088 // least significant 2 bits clear.
2089 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg
2090
2091 __ sub(swap_reg, sp, swap_reg);
2092 __ neg(swap_reg, swap_reg);
2093 __ ands(swap_reg, swap_reg, 3 - (int)os::vm_page_size());
2094
2095 // Save the test result, for recursive case, the result is zero
2096 __ str(swap_reg, Address(lock_reg, mark_word_offset));
2097 __ br(Assembler::NE, slow_path_lock);
2098 } else {
2099 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
2100 __ lightweight_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
2101 }
2102 __ bind(count);
2103 __ increment(Address(rthread, JavaThread::held_monitor_count_offset()));
2104
2105 // Slow path will re-enter here
2106 __ bind(lock_done);
2107 }
2108
2109
2110 // Finally just about ready to make the JNI call
2111
2112 // get JNIEnv* which is first argument to native
2113 __ lea(c_rarg0, Address(rthread, in_bytes(JavaThread::jni_environment_offset())));
2114
2115 // Now set thread in native
2116 __ mov(rscratch1, _thread_in_native);
2117 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2118 __ stlrw(rscratch1, rscratch2);
2119
2120 __ rt_call(native_func);
2121
2122 __ bind(native_return);
2123
2124 intptr_t return_pc = (intptr_t) __ pc();
2125 oop_maps->add_gc_map(return_pc - start, map);
2126
2127 // Verify or restore cpu control state after JNI call
2128 __ restore_cpu_control_state_after_jni(rscratch1, rscratch2);
2129
2130 // Unpack native results.
2131 switch (ret_type) {
2132 case T_BOOLEAN: __ c2bool(r0); break;
2133 case T_CHAR : __ ubfx(r0, r0, 0, 16); break;
2134 case T_BYTE : __ sbfx(r0, r0, 0, 8); break;
2135 case T_SHORT : __ sbfx(r0, r0, 0, 16); break;
2136 case T_INT : __ sbfx(r0, r0, 0, 32); break;
2137 case T_DOUBLE :
2138 case T_FLOAT :
2139 // Result is in v0 we'll save as needed
2140 break;
2141 case T_ARRAY: // Really a handle
2142 case T_OBJECT: // Really a handle
2143 break; // can't de-handlize until after safepoint check
2144 case T_VOID: break;
2145 case T_LONG: break;
2146 default : ShouldNotReachHere();
2147 }
2148
2149 Label safepoint_in_progress, safepoint_in_progress_done;
2150 Label after_transition;
2151
2152 // Switch thread to "native transition" state before reading the synchronization state.
2153 // This additional state is necessary because reading and testing the synchronization
2154 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2155 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2156 // VM thread changes sync state to synchronizing and suspends threads for GC.
2157 // Thread A is resumed to finish this native method, but doesn't block here since it
2158 // didn't see any synchronization is progress, and escapes.
2159 __ mov(rscratch1, _thread_in_native_trans);
2160
2161 __ strw(rscratch1, Address(rthread, JavaThread::thread_state_offset()));
2162
2163 // Force this write out before the read below
2164 if (!UseSystemMemoryBarrier) {
2165 __ dmb(Assembler::ISH);
2166 }
2167
2168 __ verify_sve_vector_length();
2169
2170 // Check for safepoint operation in progress and/or pending suspend requests.
2171 {
2172 // No need for acquire as Java threads always disarm themselves.
2173 __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* acquire */, false /* in_nmethod */);
2174 __ ldrw(rscratch1, Address(rthread, JavaThread::suspend_flags_offset()));
2175 __ cbnzw(rscratch1, safepoint_in_progress);
2176 __ bind(safepoint_in_progress_done);
2177 }
2178
2179 // change thread state
2180 __ mov(rscratch1, _thread_in_Java);
2181 __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
2182 __ stlrw(rscratch1, rscratch2);
2183 __ bind(after_transition);
2184
2185 Label reguard;
2186 Label reguard_done;
2187 __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset()));
2188 __ cmpw(rscratch1, StackOverflow::stack_guard_yellow_reserved_disabled);
2189 __ br(Assembler::EQ, reguard);
2190 __ bind(reguard_done);
2191
2192 // native result if any is live
2193
2194 // Unlock
2195 Label unlock_done;
2196 Label slow_path_unlock;
2197 if (method->is_synchronized()) {
2198
2199 // Get locked oop from the handle we passed to jni
2200 __ ldr(obj_reg, Address(oop_handle_reg, 0));
2201
2202 Label done, not_recursive;
2203
2204 if (LockingMode == LM_LEGACY) {
2205 // Simple recursive lock?
2206 __ ldr(rscratch1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2207 __ cbnz(rscratch1, not_recursive);
2208 __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
2209 __ b(done);
2210 }
2211
2212 __ bind(not_recursive);
2213
2214 // Must save r0 if if it is live now because cmpxchg must use it
2215 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2216 save_native_result(masm, ret_type, stack_slots);
2217 }
2218
2219 if (LockingMode == LM_MONITOR) {
2220 __ b(slow_path_unlock);
2221 } else if (LockingMode == LM_LEGACY) {
2222 // get address of the stack lock
2223 __ lea(r0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2224 // get old displaced header
2225 __ ldr(old_hdr, Address(r0, 0));
2226
2227 // Atomic swap old header if oop still contains the stack lock
2228 Label count;
2229 __ cmpxchg_obj_header(r0, old_hdr, obj_reg, rscratch1, count, &slow_path_unlock);
2230 __ bind(count);
2231 __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
2232 } else {
2233 assert(LockingMode == LM_LIGHTWEIGHT, "");
2234 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
2235 __ decrement(Address(rthread, JavaThread::held_monitor_count_offset()));
2236 }
2237
2238 // slow path re-enters here
2239 __ bind(unlock_done);
2240 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2241 restore_native_result(masm, ret_type, stack_slots);
2242 }
2243
2244 __ bind(done);
2245 }
2246
2247 Label dtrace_method_exit, dtrace_method_exit_done;
2248 if (DTraceMethodProbes) {
2249 __ b(dtrace_method_exit);
2250 __ bind(dtrace_method_exit_done);
2251 }
2252
2253 __ reset_last_Java_frame(false);
2254
2255 // Unbox oop result, e.g. JNIHandles::resolve result.
2256 if (is_reference_type(ret_type)) {
2257 __ resolve_jobject(r0, r1, r2);
2258 }
2259
2260 if (CheckJNICalls) {
2261 // clear_pending_jni_exception_check
2262 __ str(zr, Address(rthread, JavaThread::pending_jni_exception_check_fn_offset()));
2263 }
2264
2265 // reset handle block
2266 __ ldr(r2, Address(rthread, JavaThread::active_handles_offset()));
2267 __ str(zr, Address(r2, JNIHandleBlock::top_offset()));
2268
2269 __ leave();
2270
2271 // Any exception pending?
2272 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2273 __ cbnz(rscratch1, exception_pending);
2274
2275 // We're done
2276 __ ret(lr);
2277
2278 // Unexpected paths are out of line and go here
2279
2280 // forward the exception
2281 __ bind(exception_pending);
2282
2283 // and forward the exception
2284 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2285
2286 // Slow path locking & unlocking
2287 if (method->is_synchronized()) {
2288
2289 __ block_comment("Slow path lock {");
2290 __ bind(slow_path_lock);
2291
2292 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2293 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2294
2295 // protect the args we've loaded
2296 save_args(masm, total_c_args, c_arg, out_regs);
2297
2298 __ mov(c_rarg0, obj_reg);
2299 __ mov(c_rarg1, lock_reg);
2300 __ mov(c_rarg2, rthread);
2301
2302 // Not a leaf but we have last_Java_frame setup as we want
2303 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
2304 restore_args(masm, total_c_args, c_arg, out_regs);
2305
2306 #ifdef ASSERT
2307 { Label L;
2308 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2309 __ cbz(rscratch1, L);
2310 __ stop("no pending exception allowed on exit from monitorenter");
2311 __ bind(L);
2312 }
2313 #endif
2314 __ b(lock_done);
2315
2316 __ block_comment("} Slow path lock");
2317
2318 __ block_comment("Slow path unlock {");
2319 __ bind(slow_path_unlock);
2320
2321 // If we haven't already saved the native result we must save it now as xmm registers
2322 // are still exposed.
2323
2324 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2325 save_native_result(masm, ret_type, stack_slots);
2326 }
2327
2328 __ mov(c_rarg2, rthread);
2329 __ lea(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
2330 __ mov(c_rarg0, obj_reg);
2331
2332 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2333 // NOTE that obj_reg == r19 currently
2334 __ ldr(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2335 __ str(zr, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2336
2337 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2338
2339 #ifdef ASSERT
2340 {
2341 Label L;
2342 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2343 __ cbz(rscratch1, L);
2344 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2345 __ bind(L);
2346 }
2347 #endif /* ASSERT */
2348
2349 __ str(r19, Address(rthread, in_bytes(Thread::pending_exception_offset())));
2350
2351 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2352 restore_native_result(masm, ret_type, stack_slots);
2353 }
2354 __ b(unlock_done);
2355
2356 __ block_comment("} Slow path unlock");
2357
2358 } // synchronized
2359
2360 // SLOW PATH Reguard the stack if needed
2361
2362 __ bind(reguard);
2363 save_native_result(masm, ret_type, stack_slots);
2364 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2365 restore_native_result(masm, ret_type, stack_slots);
2366 // and continue
2367 __ b(reguard_done);
2368
2369 // SLOW PATH safepoint
2370 {
2371 __ block_comment("safepoint {");
2372 __ bind(safepoint_in_progress);
2373
2374 // Don't use call_VM as it will see a possible pending exception and forward it
2375 // and never return here preventing us from clearing _last_native_pc down below.
2376 //
2377 save_native_result(masm, ret_type, stack_slots);
2378 __ mov(c_rarg0, rthread);
2379 #ifndef PRODUCT
2380 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2381 #endif
2382 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans)));
2383 __ blr(rscratch1);
2384
2385 // Restore any method result value
2386 restore_native_result(masm, ret_type, stack_slots);
2387
2388 __ b(safepoint_in_progress_done);
2389 __ block_comment("} safepoint");
2390 }
2391
2392 // SLOW PATH dtrace support
2393 if (DTraceMethodProbes) {
2394 {
2395 __ block_comment("dtrace entry {");
2396 __ bind(dtrace_method_entry);
2397
2398 // We have all of the arguments setup at this point. We must not touch any register
2399 // argument registers at this point (what if we save/restore them there are no oop?
2400
2401 save_args(masm, total_c_args, c_arg, out_regs);
2402 __ mov_metadata(c_rarg1, method());
2403 __ call_VM_leaf(
2404 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2405 rthread, c_rarg1);
2406 restore_args(masm, total_c_args, c_arg, out_regs);
2407 __ b(dtrace_method_entry_done);
2408 __ block_comment("} dtrace entry");
2409 }
2410
2411 {
2412 __ block_comment("dtrace exit {");
2413 __ bind(dtrace_method_exit);
2414 save_native_result(masm, ret_type, stack_slots);
2415 __ mov_metadata(c_rarg1, method());
2416 __ call_VM_leaf(
2417 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2418 rthread, c_rarg1);
2419 restore_native_result(masm, ret_type, stack_slots);
2420 __ b(dtrace_method_exit_done);
2421 __ block_comment("} dtrace exit");
2422 }
2423 }
2424
2425 __ flush();
2426
2427 nmethod *nm = nmethod::new_native_nmethod(method,
2428 compile_id,
2429 masm->code(),
2430 vep_offset,
2431 frame_complete,
2432 stack_slots / VMRegImpl::slots_per_word,
2433 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2434 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2435 oop_maps);
2436
2437 return nm;
2438 }
2439
2440 // this function returns the adjust size (in number of words) to a c2i adapter
2441 // activation for use during deoptimization
2442 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2443 assert(callee_locals >= callee_parameters,
2444 "test and remove; got more parms than locals");
2445 if (callee_locals < callee_parameters)
2446 return 0; // No adjustment for negative locals
2447 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2448 // diff is counted in stack words
2449 return align_up(diff, 2);
2450 }
2451
2452
2453 //------------------------------generate_deopt_blob----------------------------
2454 void SharedRuntime::generate_deopt_blob() {
2455 // Allocate space for the code
2456 ResourceMark rm;
2457 // Setup code generation tools
2458 int pad = 0;
2459 #if INCLUDE_JVMCI
2460 if (EnableJVMCI) {
2461 pad += 512; // Increase the buffer size when compiling for JVMCI
2462 }
2463 #endif
2464 const char* name = SharedRuntime::stub_name(SharedStubId::deopt_id);
2465 CodeBuffer buffer(name, 2048+pad, 1024);
2466 MacroAssembler* masm = new MacroAssembler(&buffer);
2467 int frame_size_in_words;
2468 OopMap* map = nullptr;
2469 OopMapSet *oop_maps = new OopMapSet();
2470 RegisterSaver reg_save(COMPILER2_OR_JVMCI != 0);
2471
2472 // -------------
2473 // This code enters when returning to a de-optimized nmethod. A return
2474 // address has been pushed on the stack, and return values are in
2475 // registers.
2476 // If we are doing a normal deopt then we were called from the patched
2477 // nmethod from the point we returned to the nmethod. So the return
2478 // address on the stack is wrong by NativeCall::instruction_size
2479 // We will adjust the value so it looks like we have the original return
2480 // address on the stack (like when we eagerly deoptimized).
2481 // In the case of an exception pending when deoptimizing, we enter
2482 // with a return address on the stack that points after the call we patched
2483 // into the exception handler. We have the following register state from,
2484 // e.g., the forward exception stub (see stubGenerator_x86_64.cpp).
2485 // r0: exception oop
2486 // r19: exception handler
2487 // r3: throwing pc
2488 // So in this case we simply jam r3 into the useless return address and
2489 // the stack looks just like we want.
2490 //
2491 // At this point we need to de-opt. We save the argument return
2492 // registers. We call the first C routine, fetch_unroll_info(). This
2493 // routine captures the return values and returns a structure which
2494 // describes the current frame size and the sizes of all replacement frames.
2495 // The current frame is compiled code and may contain many inlined
2496 // functions, each with their own JVM state. We pop the current frame, then
2497 // push all the new frames. Then we call the C routine unpack_frames() to
2498 // populate these frames. Finally unpack_frames() returns us the new target
2499 // address. Notice that callee-save registers are BLOWN here; they have
2500 // already been captured in the vframeArray at the time the return PC was
2501 // patched.
2502 address start = __ pc();
2503 Label cont;
2504
2505 // Prolog for non exception case!
2506
2507 // Save everything in sight.
2508 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2509
2510 // Normal deoptimization. Save exec mode for unpack_frames.
2511 __ movw(rcpool, Deoptimization::Unpack_deopt); // callee-saved
2512 __ b(cont);
2513
2514 int reexecute_offset = __ pc() - start;
2515 #if INCLUDE_JVMCI && !defined(COMPILER1)
2516 if (UseJVMCICompiler) {
2517 // JVMCI does not use this kind of deoptimization
2518 __ should_not_reach_here();
2519 }
2520 #endif
2521
2522 // Reexecute case
2523 // return address is the pc describes what bci to do re-execute at
2524
2525 // No need to update map as each call to save_live_registers will produce identical oopmap
2526 (void) reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2527
2528 __ movw(rcpool, Deoptimization::Unpack_reexecute); // callee-saved
2529 __ b(cont);
2530
2531 #if INCLUDE_JVMCI
2532 Label after_fetch_unroll_info_call;
2533 int implicit_exception_uncommon_trap_offset = 0;
2534 int uncommon_trap_offset = 0;
2535
2536 if (EnableJVMCI) {
2537 implicit_exception_uncommon_trap_offset = __ pc() - start;
2538
2539 __ ldr(lr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2540 __ str(zr, Address(rthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2541
2542 uncommon_trap_offset = __ pc() - start;
2543
2544 // Save everything in sight.
2545 reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2546 // fetch_unroll_info needs to call last_java_frame()
2547 Label retaddr;
2548 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2549
2550 __ ldrw(c_rarg1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2551 __ movw(rscratch1, -1);
2552 __ strw(rscratch1, Address(rthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2553
2554 __ movw(rcpool, (int32_t)Deoptimization::Unpack_reexecute);
2555 __ mov(c_rarg0, rthread);
2556 __ movw(c_rarg2, rcpool); // exec mode
2557 __ lea(rscratch1,
2558 RuntimeAddress(CAST_FROM_FN_PTR(address,
2559 Deoptimization::uncommon_trap)));
2560 __ blr(rscratch1);
2561 __ bind(retaddr);
2562 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2563
2564 __ reset_last_Java_frame(false);
2565
2566 __ b(after_fetch_unroll_info_call);
2567 } // EnableJVMCI
2568 #endif // INCLUDE_JVMCI
2569
2570 int exception_offset = __ pc() - start;
2571
2572 // Prolog for exception case
2573
2574 // all registers are dead at this entry point, except for r0, and
2575 // r3 which contain the exception oop and exception pc
2576 // respectively. Set them in TLS and fall thru to the
2577 // unpack_with_exception_in_tls entry point.
2578
2579 __ str(r3, Address(rthread, JavaThread::exception_pc_offset()));
2580 __ str(r0, Address(rthread, JavaThread::exception_oop_offset()));
2581
2582 int exception_in_tls_offset = __ pc() - start;
2583
2584 // new implementation because exception oop is now passed in JavaThread
2585
2586 // Prolog for exception case
2587 // All registers must be preserved because they might be used by LinearScan
2588 // Exceptiop oop and throwing PC are passed in JavaThread
2589 // tos: stack at point of call to method that threw the exception (i.e. only
2590 // args are on the stack, no return address)
2591
2592 // The return address pushed by save_live_registers will be patched
2593 // later with the throwing pc. The correct value is not available
2594 // now because loading it from memory would destroy registers.
2595
2596 // NB: The SP at this point must be the SP of the method that is
2597 // being deoptimized. Deoptimization assumes that the frame created
2598 // here by save_live_registers is immediately below the method's SP.
2599 // This is a somewhat fragile mechanism.
2600
2601 // Save everything in sight.
2602 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2603
2604 // Now it is safe to overwrite any register
2605
2606 // Deopt during an exception. Save exec mode for unpack_frames.
2607 __ mov(rcpool, Deoptimization::Unpack_exception); // callee-saved
2608
2609 // load throwing pc from JavaThread and patch it as the return address
2610 // of the current frame. Then clear the field in JavaThread
2611 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2612 __ protect_return_address(r3);
2613 __ str(r3, Address(rfp, wordSize));
2614 __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2615
2616 #ifdef ASSERT
2617 // verify that there is really an exception oop in JavaThread
2618 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2619 __ verify_oop(r0);
2620
2621 // verify that there is no pending exception
2622 Label no_pending_exception;
2623 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2624 __ cbz(rscratch1, no_pending_exception);
2625 __ stop("must not have pending exception here");
2626 __ bind(no_pending_exception);
2627 #endif
2628
2629 __ bind(cont);
2630
2631 // Call C code. Need thread and this frame, but NOT official VM entry
2632 // crud. We cannot block on this call, no GC can happen.
2633 //
2634 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2635
2636 // fetch_unroll_info needs to call last_java_frame().
2637
2638 Label retaddr;
2639 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2640 #ifdef ASSERT
2641 { Label L;
2642 __ ldr(rscratch1, Address(rthread, JavaThread::last_Java_fp_offset()));
2643 __ cbz(rscratch1, L);
2644 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2645 __ bind(L);
2646 }
2647 #endif // ASSERT
2648 __ mov(c_rarg0, rthread);
2649 __ mov(c_rarg1, rcpool);
2650 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2651 __ blr(rscratch1);
2652 __ bind(retaddr);
2653
2654 // Need to have an oopmap that tells fetch_unroll_info where to
2655 // find any register it might need.
2656 oop_maps->add_gc_map(__ pc() - start, map);
2657
2658 __ reset_last_Java_frame(false);
2659
2660 #if INCLUDE_JVMCI
2661 if (EnableJVMCI) {
2662 __ bind(after_fetch_unroll_info_call);
2663 }
2664 #endif
2665
2666 // Load UnrollBlock* into r5
2667 __ mov(r5, r0);
2668
2669 __ ldrw(rcpool, Address(r5, Deoptimization::UnrollBlock::unpack_kind_offset()));
2670 Label noException;
2671 __ cmpw(rcpool, Deoptimization::Unpack_exception); // Was exception pending?
2672 __ br(Assembler::NE, noException);
2673 __ ldr(r0, Address(rthread, JavaThread::exception_oop_offset()));
2674 // QQQ this is useless it was null above
2675 __ ldr(r3, Address(rthread, JavaThread::exception_pc_offset()));
2676 __ str(zr, Address(rthread, JavaThread::exception_oop_offset()));
2677 __ str(zr, Address(rthread, JavaThread::exception_pc_offset()));
2678
2679 __ verify_oop(r0);
2680
2681 // Overwrite the result registers with the exception results.
2682 __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2683 // I think this is useless
2684 // __ str(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2685
2686 __ bind(noException);
2687
2688 // Only register save data is on the stack.
2689 // Now restore the result registers. Everything else is either dead
2690 // or captured in the vframeArray.
2691
2692 // Restore fp result register
2693 __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2694 // Restore integer result register
2695 __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2696
2697 // Pop all of the register save area off the stack
2698 __ add(sp, sp, frame_size_in_words * wordSize);
2699
2700 // All of the register save area has been popped of the stack. Only the
2701 // return address remains.
2702
2703 // Pop all the frames we must move/replace.
2704 //
2705 // Frame picture (youngest to oldest)
2706 // 1: self-frame (no frame link)
2707 // 2: deopting frame (no frame link)
2708 // 3: caller of deopting frame (could be compiled/interpreted).
2709 //
2710 // Note: by leaving the return address of self-frame on the stack
2711 // and using the size of frame 2 to adjust the stack
2712 // when we are done the return to frame 3 will still be on the stack.
2713
2714 // Pop deoptimized frame
2715 __ ldrw(r2, Address(r5, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2716 __ sub(r2, r2, 2 * wordSize);
2717 __ add(sp, sp, r2);
2718 __ ldp(rfp, zr, __ post(sp, 2 * wordSize));
2719
2720 #ifdef ASSERT
2721 // Compilers generate code that bang the stack by as much as the
2722 // interpreter would need. So this stack banging should never
2723 // trigger a fault. Verify that it does not on non product builds.
2724 __ ldrw(r19, Address(r5, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2725 __ bang_stack_size(r19, r2);
2726 #endif
2727 // Load address of array of frame pcs into r2
2728 __ ldr(r2, Address(r5, Deoptimization::UnrollBlock::frame_pcs_offset()));
2729
2730 // Trash the old pc
2731 // __ addptr(sp, wordSize); FIXME ????
2732
2733 // Load address of array of frame sizes into r4
2734 __ ldr(r4, Address(r5, Deoptimization::UnrollBlock::frame_sizes_offset()));
2735
2736 // Load counter into r3
2737 __ ldrw(r3, Address(r5, Deoptimization::UnrollBlock::number_of_frames_offset()));
2738
2739 // Now adjust the caller's stack to make up for the extra locals
2740 // but record the original sp so that we can save it in the skeletal interpreter
2741 // frame and the stack walking of interpreter_sender will get the unextended sp
2742 // value and not the "real" sp value.
2743
2744 const Register sender_sp = r6;
2745
2746 __ mov(sender_sp, sp);
2747 __ ldrw(r19, Address(r5,
2748 Deoptimization::UnrollBlock::
2749 caller_adjustment_offset()));
2750 __ sub(sp, sp, r19);
2751
2752 // Push interpreter frames in a loop
2753 __ mov(rscratch1, (uint64_t)0xDEADDEAD); // Make a recognizable pattern
2754 __ mov(rscratch2, rscratch1);
2755 Label loop;
2756 __ bind(loop);
2757 __ ldr(r19, Address(__ post(r4, wordSize))); // Load frame size
2758 __ sub(r19, r19, 2*wordSize); // We'll push pc and fp by hand
2759 __ ldr(lr, Address(__ post(r2, wordSize))); // Load pc
2760 __ enter(); // Save old & set new fp
2761 __ sub(sp, sp, r19); // Prolog
2762 // This value is corrected by layout_activation_impl
2763 __ str(zr, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
2764 __ str(sender_sp, Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2765 __ mov(sender_sp, sp); // Pass sender_sp to next frame
2766 __ sub(r3, r3, 1); // Decrement counter
2767 __ cbnz(r3, loop);
2768
2769 // Re-push self-frame
2770 __ ldr(lr, Address(r2));
2771 __ enter();
2772
2773 // Allocate a full sized register save area. We subtract 2 because
2774 // enter() just pushed 2 words
2775 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2776
2777 // Restore frame locals after moving the frame
2778 __ strd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2779 __ str(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2780
2781 // Call C code. Need thread but NOT official VM entry
2782 // crud. We cannot block on this call, no GC can happen. Call should
2783 // restore return values to their stack-slots with the new SP.
2784 //
2785 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2786
2787 // Use rfp because the frames look interpreted now
2788 // Don't need the precise return PC here, just precise enough to point into this code blob.
2789 address the_pc = __ pc();
2790 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
2791
2792 __ mov(c_rarg0, rthread);
2793 __ movw(c_rarg1, rcpool); // second arg: exec_mode
2794 __ lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2795 __ blr(rscratch1);
2796
2797 // Set an oopmap for the call site
2798 // Use the same PC we used for the last java frame
2799 oop_maps->add_gc_map(the_pc - start,
2800 new OopMap( frame_size_in_words, 0 ));
2801
2802 // Clear fp AND pc
2803 __ reset_last_Java_frame(true);
2804
2805 // Collect return values
2806 __ ldrd(v0, Address(sp, reg_save.v0_offset_in_bytes()));
2807 __ ldr(r0, Address(sp, reg_save.r0_offset_in_bytes()));
2808 // I think this is useless (throwing pc?)
2809 // __ ldr(r3, Address(sp, RegisterSaver::r3_offset_in_bytes()));
2810
2811 // Pop self-frame.
2812 __ leave(); // Epilog
2813
2814 // Jump to interpreter
2815 __ ret(lr);
2816
2817 // Make sure all code is generated
2818 masm->flush();
2819
2820 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2821 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2822 #if INCLUDE_JVMCI
2823 if (EnableJVMCI) {
2824 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2825 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2826 }
2827 #endif
2828 }
2829
2830 // Number of stack slots between incoming argument block and the start of
2831 // a new frame. The PROLOG must add this many slots to the stack. The
2832 // EPILOG must remove this many slots. aarch64 needs two slots for
2833 // return address and fp.
2834 // TODO think this is correct but check
2835 uint SharedRuntime::in_preserve_stack_slots() {
2836 return 4;
2837 }
2838
2839 uint SharedRuntime::out_preserve_stack_slots() {
2840 return 0;
2841 }
2842
2843
2844 //------------------------------generate_handler_blob------
2845 //
2846 // Generate a special Compile2Runtime blob that saves all registers,
2847 // and setup oopmap.
2848 //
2849 SafepointBlob* SharedRuntime::generate_handler_blob(SharedStubId id, address call_ptr) {
2850 assert(is_polling_page_id(id), "expected a polling page stub id");
2851
2852 ResourceMark rm;
2853 OopMapSet *oop_maps = new OopMapSet();
2854 OopMap* map;
2855
2856 // Allocate space for the code. Setup code generation tools.
2857 const char* name = SharedRuntime::stub_name(id);
2858 CodeBuffer buffer(name, 2048, 1024);
2859 MacroAssembler* masm = new MacroAssembler(&buffer);
2860
2861 address start = __ pc();
2862 address call_pc = nullptr;
2863 int frame_size_in_words;
2864 bool cause_return = (id == SharedStubId::polling_page_return_handler_id);
2865 RegisterSaver reg_save(id == SharedStubId::polling_page_vectors_safepoint_handler_id /* save_vectors */);
2866
2867 // When the signal occurred, the LR was either signed and stored on the stack (in which
2868 // case it will be restored from the stack before being used) or unsigned and not stored
2869 // on the stack. Stipping ensures we get the right value.
2870 __ strip_return_address();
2871
2872 // Save Integer and Float registers.
2873 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2874
2875 // The following is basically a call_VM. However, we need the precise
2876 // address of the call in order to generate an oopmap. Hence, we do all the
2877 // work ourselves.
2878
2879 Label retaddr;
2880 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
2881
2882 // The return address must always be correct so that frame constructor never
2883 // sees an invalid pc.
2884
2885 if (!cause_return) {
2886 // overwrite the return address pushed by save_live_registers
2887 // Additionally, r20 is a callee-saved register so we can look at
2888 // it later to determine if someone changed the return address for
2889 // us!
2890 __ ldr(r20, Address(rthread, JavaThread::saved_exception_pc_offset()));
2891 __ protect_return_address(r20);
2892 __ str(r20, Address(rfp, wordSize));
2893 }
2894
2895 // Do the call
2896 __ mov(c_rarg0, rthread);
2897 __ lea(rscratch1, RuntimeAddress(call_ptr));
2898 __ blr(rscratch1);
2899 __ bind(retaddr);
2900
2901 // Set an oopmap for the call site. This oopmap will map all
2902 // oop-registers and debug-info registers as callee-saved. This
2903 // will allow deoptimization at this safepoint to find all possible
2904 // debug-info recordings, as well as let GC find all oops.
2905
2906 oop_maps->add_gc_map( __ pc() - start, map);
2907
2908 Label noException;
2909
2910 __ reset_last_Java_frame(false);
2911
2912 __ membar(Assembler::LoadLoad | Assembler::LoadStore);
2913
2914 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
2915 __ cbz(rscratch1, noException);
2916
2917 // Exception pending
2918
2919 reg_save.restore_live_registers(masm);
2920
2921 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2922
2923 // No exception case
2924 __ bind(noException);
2925
2926 Label no_adjust, bail;
2927 if (!cause_return) {
2928 // If our stashed return pc was modified by the runtime we avoid touching it
2929 __ ldr(rscratch1, Address(rfp, wordSize));
2930 __ cmp(r20, rscratch1);
2931 __ br(Assembler::NE, no_adjust);
2932 __ authenticate_return_address(r20);
2933
2934 #ifdef ASSERT
2935 // Verify the correct encoding of the poll we're about to skip.
2936 // See NativeInstruction::is_ldrw_to_zr()
2937 __ ldrw(rscratch1, Address(r20));
2938 __ ubfx(rscratch2, rscratch1, 22, 10);
2939 __ cmpw(rscratch2, 0b1011100101);
2940 __ br(Assembler::NE, bail);
2941 __ ubfx(rscratch2, rscratch1, 0, 5);
2942 __ cmpw(rscratch2, 0b11111);
2943 __ br(Assembler::NE, bail);
2944 #endif
2945 // Adjust return pc forward to step over the safepoint poll instruction
2946 __ add(r20, r20, NativeInstruction::instruction_size);
2947 __ protect_return_address(r20);
2948 __ str(r20, Address(rfp, wordSize));
2949 }
2950
2951 __ bind(no_adjust);
2952 // Normal exit, restore registers and exit.
2953 reg_save.restore_live_registers(masm);
2954
2955 __ ret(lr);
2956
2957 #ifdef ASSERT
2958 __ bind(bail);
2959 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2960 #endif
2961
2962 // Make sure all code is generated
2963 masm->flush();
2964
2965 // Fill-out other meta info
2966 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2967 }
2968
2969 //
2970 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2971 //
2972 // Generate a stub that calls into vm to find out the proper destination
2973 // of a java call. All the argument registers are live at this point
2974 // but since this is generic code we don't know what they are and the caller
2975 // must do any gc of the args.
2976 //
2977 RuntimeStub* SharedRuntime::generate_resolve_blob(SharedStubId id, address destination) {
2978 assert (StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2979 assert(is_resolve_id(id), "expected a resolve stub id");
2980
2981 // allocate space for the code
2982 ResourceMark rm;
2983
2984 const char* name = SharedRuntime::stub_name(id);
2985 CodeBuffer buffer(name, 1000, 512);
2986 MacroAssembler* masm = new MacroAssembler(&buffer);
2987
2988 int frame_size_in_words;
2989 RegisterSaver reg_save(false /* save_vectors */);
2990
2991 OopMapSet *oop_maps = new OopMapSet();
2992 OopMap* map = nullptr;
2993
2994 int start = __ offset();
2995
2996 map = reg_save.save_live_registers(masm, 0, &frame_size_in_words);
2997
2998 int frame_complete = __ offset();
2999
3000 {
3001 Label retaddr;
3002 __ set_last_Java_frame(sp, noreg, retaddr, rscratch1);
3003
3004 __ mov(c_rarg0, rthread);
3005 __ lea(rscratch1, RuntimeAddress(destination));
3006
3007 __ blr(rscratch1);
3008 __ bind(retaddr);
3009 }
3010
3011 // Set an oopmap for the call site.
3012 // We need this not only for callee-saved registers, but also for volatile
3013 // registers that the compiler might be keeping live across a safepoint.
3014
3015 oop_maps->add_gc_map( __ offset() - start, map);
3016
3017 // r0 contains the address we are going to jump to assuming no exception got installed
3018
3019 // clear last_Java_sp
3020 __ reset_last_Java_frame(false);
3021 // check for pending exceptions
3022 Label pending;
3023 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
3024 __ cbnz(rscratch1, pending);
3025
3026 // get the returned Method*
3027 __ get_vm_result_2(rmethod, rthread);
3028 __ str(rmethod, Address(sp, reg_save.reg_offset_in_bytes(rmethod)));
3029
3030 // r0 is where we want to jump, overwrite rscratch1 which is saved and scratch
3031 __ str(r0, Address(sp, reg_save.rscratch1_offset_in_bytes()));
3032 reg_save.restore_live_registers(masm);
3033
3034 // We are back to the original state on entry and ready to go.
3035
3036 __ br(rscratch1);
3037
3038 // Pending exception after the safepoint
3039
3040 __ bind(pending);
3041
3042 reg_save.restore_live_registers(masm);
3043
3044 // exception pending => remove activation and forward to exception handler
3045
3046 __ str(zr, Address(rthread, JavaThread::vm_result_offset()));
3047
3048 __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
3049 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3050
3051 // -------------
3052 // make sure all code is generated
3053 masm->flush();
3054
3055 // return the blob
3056 // frame_size_words or bytes??
3057 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
3058 }
3059
3060 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
3061 BufferBlob* buf = BufferBlob::create("inline types pack/unpack", 16 * K);
3062 CodeBuffer buffer(buf);
3063 short buffer_locs[20];
3064 buffer.insts()->initialize_shared_locs((relocInfo*)buffer_locs,
3065 sizeof(buffer_locs)/sizeof(relocInfo));
3066
3067 MacroAssembler _masm(&buffer);
3068 MacroAssembler* masm = &_masm;
3069
3070 const Array<SigEntry>* sig_vk = vk->extended_sig();
3071 const Array<VMRegPair>* regs = vk->return_regs();
3072
3073 int pack_fields_jobject_off = __ offset();
3074 // Resolve pre-allocated buffer from JNI handle.
3075 // We cannot do this in generate_call_stub() because it requires GC code to be initialized.
3076 Register Rresult = r14; // See StubGenerator::generate_call_stub().
3077 __ ldr(r0, Address(Rresult));
3078 __ resolve_jobject(r0 /* value */,
3079 rthread /* thread */,
3080 r12 /* tmp */);
3081 __ str(r0, Address(Rresult));
3082
3083 int pack_fields_off = __ offset();
3084
3085 int j = 1;
3086 for (int i = 0; i < sig_vk->length(); i++) {
3087 BasicType bt = sig_vk->at(i)._bt;
3088 if (bt == T_METADATA) {
3089 continue;
3090 }
3091 if (bt == T_VOID) {
3092 if (sig_vk->at(i-1)._bt == T_LONG ||
3093 sig_vk->at(i-1)._bt == T_DOUBLE) {
3094 j++;
3095 }
3096 continue;
3097 }
3098 int off = sig_vk->at(i)._offset;
3099 VMRegPair pair = regs->at(j);
3100 VMReg r_1 = pair.first();
3101 VMReg r_2 = pair.second();
3102 Address to(r0, off);
3103 if (bt == T_FLOAT) {
3104 __ strs(r_1->as_FloatRegister(), to);
3105 } else if (bt == T_DOUBLE) {
3106 __ strd(r_1->as_FloatRegister(), to);
3107 } else {
3108 Register val = r_1->as_Register();
3109 assert_different_registers(to.base(), val, r15, r16, r17);
3110 if (is_reference_type(bt)) {
3111 __ store_heap_oop(to, val, r15, r16, r17, IN_HEAP | ACCESS_WRITE | IS_DEST_UNINITIALIZED);
3112 } else {
3113 __ store_sized_value(to, r_1->as_Register(), type2aelembytes(bt));
3114 }
3115 }
3116 j++;
3117 }
3118 assert(j == regs->length(), "missed a field?");
3119
3120 __ ret(lr);
3121
3122 int unpack_fields_off = __ offset();
3123
3124 Label skip;
3125 __ cbz(r0, skip);
3126
3127 j = 1;
3128 for (int i = 0; i < sig_vk->length(); i++) {
3129 BasicType bt = sig_vk->at(i)._bt;
3130 if (bt == T_METADATA) {
3131 continue;
3132 }
3133 if (bt == T_VOID) {
3134 if (sig_vk->at(i-1)._bt == T_LONG ||
3135 sig_vk->at(i-1)._bt == T_DOUBLE) {
3136 j++;
3137 }
3138 continue;
3139 }
3140 int off = sig_vk->at(i)._offset;
3141 assert(off > 0, "offset in object should be positive");
3142 VMRegPair pair = regs->at(j);
3143 VMReg r_1 = pair.first();
3144 VMReg r_2 = pair.second();
3145 Address from(r0, off);
3146 if (bt == T_FLOAT) {
3147 __ ldrs(r_1->as_FloatRegister(), from);
3148 } else if (bt == T_DOUBLE) {
3149 __ ldrd(r_1->as_FloatRegister(), from);
3150 } else if (bt == T_OBJECT || bt == T_ARRAY) {
3151 assert_different_registers(r0, r_1->as_Register());
3152 __ load_heap_oop(r_1->as_Register(), from, rscratch1, rscratch2);
3153 } else {
3154 assert(is_java_primitive(bt), "unexpected basic type");
3155 assert_different_registers(r0, r_1->as_Register());
3156
3157 size_t size_in_bytes = type2aelembytes(bt);
3158 __ load_sized_value(r_1->as_Register(), from, size_in_bytes, bt != T_CHAR && bt != T_BOOLEAN);
3159 }
3160 j++;
3161 }
3162 assert(j == regs->length(), "missed a field?");
3163
3164 __ bind(skip);
3165
3166 __ ret(lr);
3167
3168 __ flush();
3169
3170 return BufferedInlineTypeBlob::create(&buffer, pack_fields_off, pack_fields_jobject_off, unpack_fields_off);
3171 }
3172
3173 // Continuation point for throwing of implicit exceptions that are
3174 // not handled in the current activation. Fabricates an exception
3175 // oop and initiates normal exception dispatching in this
3176 // frame. Since we need to preserve callee-saved values (currently
3177 // only for C2, but done for C1 as well) we need a callee-saved oop
3178 // map and therefore have to make these stubs into RuntimeStubs
3179 // rather than BufferBlobs. If the compiler needs all registers to
3180 // be preserved between the fault point and the exception handler
3181 // then it must assume responsibility for that in
3182 // AbstractCompiler::continuation_for_implicit_null_exception or
3183 // continuation_for_implicit_division_by_zero_exception. All other
3184 // implicit exceptions (e.g., NullPointerException or
3185 // AbstractMethodError on entry) are either at call sites or
3186 // otherwise assume that stack unwinding will be initiated, so
3187 // caller saved registers were assumed volatile in the compiler.
3188
3189 RuntimeStub* SharedRuntime::generate_throw_exception(SharedStubId id, address runtime_entry) {
3190 assert(is_throw_id(id), "expected a throw stub id");
3191
3192 const char* name = SharedRuntime::stub_name(id);
3193
3194 // Information about frame layout at time of blocking runtime call.
3195 // Note that we only have to preserve callee-saved registers since
3196 // the compilers are responsible for supplying a continuation point
3197 // if they expect all registers to be preserved.
3198 // n.b. aarch64 asserts that frame::arg_reg_save_area_bytes == 0
3199 enum layout {
3200 rfp_off = 0,
3201 rfp_off2,
3202 return_off,
3203 return_off2,
3204 framesize // inclusive of return address
3205 };
3206
3207 int insts_size = 512;
3208 int locs_size = 64;
3209
3210 ResourceMark rm;
3211 const char* timer_msg = "SharedRuntime generate_throw_exception";
3212 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3213
3214 CodeBuffer code(name, insts_size, locs_size);
3215 OopMapSet* oop_maps = new OopMapSet();
3216 MacroAssembler* masm = new MacroAssembler(&code);
3217
3218 address start = __ pc();
3219
3220 // This is an inlined and slightly modified version of call_VM
3221 // which has the ability to fetch the return PC out of
3222 // thread-local storage and also sets up last_Java_sp slightly
3223 // differently than the real call_VM
3224
3225 __ enter(); // Save FP and LR before call
3226
3227 assert(is_even(framesize/2), "sp not 16-byte aligned");
3228
3229 // lr and fp are already in place
3230 __ sub(sp, rfp, ((uint64_t)framesize-4) << LogBytesPerInt); // prolog
3231
3232 int frame_complete = __ pc() - start;
3233
3234 // Set up last_Java_sp and last_Java_fp
3235 address the_pc = __ pc();
3236 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3237
3238 __ mov(c_rarg0, rthread);
3239 BLOCK_COMMENT("call runtime_entry");
3240 __ mov(rscratch1, runtime_entry);
3241 __ blr(rscratch1);
3242
3243 // Generate oop map
3244 OopMap* map = new OopMap(framesize, 0);
3245
3246 oop_maps->add_gc_map(the_pc - start, map);
3247
3248 __ reset_last_Java_frame(true);
3249
3250 // Reinitialize the ptrue predicate register, in case the external runtime
3251 // call clobbers ptrue reg, as we may return to SVE compiled code.
3252 __ reinitialize_ptrue();
3253
3254 __ leave();
3255
3256 // check for pending exceptions
3257 #ifdef ASSERT
3258 Label L;
3259 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
3260 __ cbnz(rscratch1, L);
3261 __ should_not_reach_here();
3262 __ bind(L);
3263 #endif // ASSERT
3264 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3265
3266 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3267 RuntimeStub* stub =
3268 RuntimeStub::new_runtime_stub(name,
3269 &code,
3270 frame_complete,
3271 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3272 oop_maps, false);
3273 return stub;
3274 }
3275
3276 #if INCLUDE_JFR
3277
3278 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
3279 __ set_last_Java_frame(sp, rfp, the_pc, rscratch1);
3280 __ mov(c_rarg0, thread);
3281 }
3282
3283 // The handle is dereferenced through a load barrier.
3284 static void jfr_epilogue(MacroAssembler* masm) {
3285 __ reset_last_Java_frame(true);
3286 }
3287
3288 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3289 // It returns a jobject handle to the event writer.
3290 // The handle is dereferenced and the return value is the event writer oop.
3291 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3292 enum layout {
3293 rbp_off,
3294 rbpH_off,
3295 return_off,
3296 return_off2,
3297 framesize // inclusive of return address
3298 };
3299
3300 int insts_size = 1024;
3301 int locs_size = 64;
3302 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
3303 CodeBuffer code(name, insts_size, locs_size);
3304 OopMapSet* oop_maps = new OopMapSet();
3305 MacroAssembler* masm = new MacroAssembler(&code);
3306
3307 address start = __ pc();
3308 __ enter();
3309 int frame_complete = __ pc() - start;
3310 address the_pc = __ pc();
3311 jfr_prologue(the_pc, masm, rthread);
3312 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
3313 jfr_epilogue(masm);
3314 __ resolve_global_jobject(r0, rscratch1, rscratch2);
3315 __ leave();
3316 __ ret(lr);
3317
3318 OopMap* map = new OopMap(framesize, 1); // rfp
3319 oop_maps->add_gc_map(the_pc - start, map);
3320
3321 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3322 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3323 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3324 oop_maps, false);
3325 return stub;
3326 }
3327
3328 // For c2: call to return a leased buffer.
3329 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3330 enum layout {
3331 rbp_off,
3332 rbpH_off,
3333 return_off,
3334 return_off2,
3335 framesize // inclusive of return address
3336 };
3337
3338 int insts_size = 1024;
3339 int locs_size = 64;
3340
3341 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_return_lease_id);
3342 CodeBuffer code(name, insts_size, locs_size);
3343 OopMapSet* oop_maps = new OopMapSet();
3344 MacroAssembler* masm = new MacroAssembler(&code);
3345
3346 address start = __ pc();
3347 __ enter();
3348 int frame_complete = __ pc() - start;
3349 address the_pc = __ pc();
3350 jfr_prologue(the_pc, masm, rthread);
3351 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
3352 jfr_epilogue(masm);
3353
3354 __ leave();
3355 __ ret(lr);
3356
3357 OopMap* map = new OopMap(framesize, 1); // rfp
3358 oop_maps->add_gc_map(the_pc - start, map);
3359
3360 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3361 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3362 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3363 oop_maps, false);
3364 return stub;
3365 }
3366
3367 #endif // INCLUDE_JFR