1 /*
2 * Copyright (c) 2003, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2020, Red Hat Inc. All rights reserved.
4 * Copyright (c) 2020, 2023, Huawei Technologies Co., Ltd. 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 "asm/macroAssembler.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "classfile/symbolTable.hpp"
30 #include "code/compiledIC.hpp"
31 #include "code/debugInfoRec.hpp"
32 #include "code/vtableStubs.hpp"
33 #include "compiler/oopMap.hpp"
34 #include "gc/shared/barrierSetAssembler.hpp"
35 #include "interpreter/interp_masm.hpp"
36 #include "interpreter/interpreter.hpp"
37 #include "logging/log.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "nativeInst_riscv.hpp"
40 #include "oops/klass.inline.hpp"
41 #include "oops/method.inline.hpp"
42 #include "prims/methodHandles.hpp"
43 #include "runtime/continuation.hpp"
44 #include "runtime/continuationEntry.inline.hpp"
45 #include "runtime/globals.hpp"
46 #include "runtime/jniHandles.hpp"
47 #include "runtime/safepointMechanism.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "runtime/signature.hpp"
50 #include "runtime/stubRoutines.hpp"
51 #include "runtime/timerTrace.hpp"
52 #include "runtime/vframeArray.hpp"
53 #include "utilities/align.hpp"
54 #include "utilities/formatBuffer.hpp"
55 #include "vmreg_riscv.inline.hpp"
56 #ifdef COMPILER1
57 #include "c1/c1_Runtime1.hpp"
58 #endif
59 #ifdef COMPILER2
60 #include "adfiles/ad_riscv.hpp"
61 #include "opto/runtime.hpp"
62 #endif
63
64 #define __ masm->
65
66 #ifdef PRODUCT
67 #define BLOCK_COMMENT(str) /* nothing */
68 #else
69 #define BLOCK_COMMENT(str) __ block_comment(str)
70 #endif
71
72 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
73
74 class RegisterSaver {
75 const bool _save_vectors;
76 public:
77 RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
78 ~RegisterSaver() {}
79 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
80 void restore_live_registers(MacroAssembler* masm);
81
82 // Offsets into the register save area
83 // Used by deoptimization when it is managing result register
84 // values on its own
85 // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
86 // |---v0---|<---SP
87 // |---v1---|save vectors only in generate_handler_blob
88 // |-- .. --|
89 // |---v31--|-----
90 // |---f0---|
91 // |---f1---|
92 // | .. |
93 // |---f31--|
94 // |---reserved slot for stack alignment---|
95 // |---x5---|
96 // | x6 |
97 // |---.. --|
98 // |---x31--|
99 // |---fp---|
100 // |---ra---|
101 int v0_offset_in_bytes(void) { return 0; }
102 int f0_offset_in_bytes(void) {
103 int f0_offset = 0;
104 #ifdef COMPILER2
105 if (_save_vectors) {
106 f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
107 BytesPerInt;
108 }
109 #endif
110 return f0_offset;
111 }
112 int reserved_slot_offset_in_bytes(void) {
113 return f0_offset_in_bytes() +
114 FloatRegister::max_slots_per_register *
115 FloatRegister::number_of_registers *
116 BytesPerInt;
117 }
118
119 int reg_offset_in_bytes(Register r) {
120 assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
121 return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
122 }
123
124 int freg_offset_in_bytes(FloatRegister f) {
125 return f0_offset_in_bytes() + f->encoding() * wordSize;
126 }
127
128 int ra_offset_in_bytes(void) {
129 return reserved_slot_offset_in_bytes() +
130 (Register::number_of_registers - 3) *
131 Register::max_slots_per_register *
132 BytesPerInt;
133 }
134 };
135
136 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
137 int vector_size_in_bytes = 0;
138 int vector_size_in_slots = 0;
139 #ifdef COMPILER2
140 if (_save_vectors) {
141 vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
142 vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
143 }
144 #endif
145
146 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
147 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
148 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
149 // The caller will allocate additional_frame_words
150 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
151 // CodeBlob frame size is in words.
152 int frame_size_in_words = frame_size_in_bytes / wordSize;
153 *total_frame_words = frame_size_in_words;
154
155 // Save Integer, Float and Vector registers.
156 __ enter();
157 __ push_CPU_state(_save_vectors, vector_size_in_bytes);
158
159 // Set an oopmap for the call site. This oopmap will map all
160 // oop-registers and debug-info registers as callee-saved. This
161 // will allow deoptimization at this safepoint to find all possible
162 // debug-info recordings, as well as let GC find all oops.
163
164 OopMapSet *oop_maps = new OopMapSet();
165 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
166 assert_cond(oop_maps != nullptr && oop_map != nullptr);
167
168 int sp_offset_in_slots = 0;
169 int step_in_slots = 0;
170 if (_save_vectors) {
171 step_in_slots = vector_size_in_slots;
172 for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
173 VectorRegister r = as_VectorRegister(i);
174 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
175 }
176 }
177
178 step_in_slots = FloatRegister::max_slots_per_register;
179 for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
180 FloatRegister r = as_FloatRegister(i);
181 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
182 }
183
184 step_in_slots = Register::max_slots_per_register;
185 // skip the slot reserved for alignment, see MacroAssembler::push_reg;
186 // also skip x5 ~ x6 on the stack because they are caller-saved registers.
187 sp_offset_in_slots += Register::max_slots_per_register * 3;
188 // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
189 for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
190 Register r = as_Register(i);
191 if (r != xthread) {
192 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
193 }
194 }
195
196 return oop_map;
197 }
198
199 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
200 #ifdef COMPILER2
201 __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
202 #else
203 assert(!_save_vectors, "vectors are generated only by C2");
204 __ pop_CPU_state(_save_vectors);
205 #endif // COMPILER2
206 __ leave();
207 }
208
209 // Is vector's size (in bytes) bigger than a size saved by default?
210 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
211 bool SharedRuntime::is_wide_vector(int size) {
212 return UseRVV && size > 0;
213 }
214
215 // ---------------------------------------------------------------------------
216 // Read the array of BasicTypes from a signature, and compute where the
217 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
218 // quantities. Values less than VMRegImpl::stack0 are registers, those above
219 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
220 // as framesizes are fixed.
221 // VMRegImpl::stack0 refers to the first slot 0(sp).
222 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
223 // Register up to Register::number_of_registers) are the 64-bit
224 // integer registers.
225
226 // Note: the INPUTS in sig_bt are in units of Java argument words,
227 // which are 64-bit. The OUTPUTS are in 32-bit units.
228
229 // The Java calling convention is a "shifted" version of the C ABI.
230 // By skipping the first C ABI register we can call non-static jni
231 // methods with small numbers of arguments without having to shuffle
232 // the arguments at all. Since we control the java ABI we ought to at
233 // least get some advantage out of it.
234
235 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
236 VMRegPair *regs,
237 int total_args_passed) {
238 // Create the mapping between argument positions and
239 // registers.
240 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
241 j_rarg0, j_rarg1, j_rarg2, j_rarg3,
242 j_rarg4, j_rarg5, j_rarg6, j_rarg7
243 };
244 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
245 j_farg0, j_farg1, j_farg2, j_farg3,
246 j_farg4, j_farg5, j_farg6, j_farg7
247 };
248
249 uint int_args = 0;
250 uint fp_args = 0;
251 uint stk_args = 0;
252
253 for (int i = 0; i < total_args_passed; i++) {
254 switch (sig_bt[i]) {
255 case T_BOOLEAN: // fall through
256 case T_CHAR: // fall through
257 case T_BYTE: // fall through
258 case T_SHORT: // fall through
259 case T_INT:
260 if (int_args < Argument::n_int_register_parameters_j) {
261 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
262 } else {
263 stk_args = align_up(stk_args, 2);
264 regs[i].set1(VMRegImpl::stack2reg(stk_args));
265 stk_args += 1;
266 }
267 break;
268 case T_VOID:
269 // halves of T_LONG or T_DOUBLE
270 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
271 regs[i].set_bad();
272 break;
273 case T_LONG: // fall through
274 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
275 case T_OBJECT: // fall through
276 case T_ARRAY: // fall through
277 case T_ADDRESS:
278 if (int_args < Argument::n_int_register_parameters_j) {
279 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
280 } else {
281 stk_args = align_up(stk_args, 2);
282 regs[i].set2(VMRegImpl::stack2reg(stk_args));
283 stk_args += 2;
284 }
285 break;
286 case T_FLOAT:
287 if (fp_args < Argument::n_float_register_parameters_j) {
288 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
289 } else {
290 stk_args = align_up(stk_args, 2);
291 regs[i].set1(VMRegImpl::stack2reg(stk_args));
292 stk_args += 1;
293 }
294 break;
295 case T_DOUBLE:
296 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
297 if (fp_args < Argument::n_float_register_parameters_j) {
298 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
299 } else {
300 stk_args = align_up(stk_args, 2);
301 regs[i].set2(VMRegImpl::stack2reg(stk_args));
302 stk_args += 2;
303 }
304 break;
305 default:
306 ShouldNotReachHere();
307 }
308 }
309
310 return stk_args;
311 }
312
313 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
314 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
315
316 int SharedRuntime::java_return_convention(const BasicType *sig_bt,
317 VMRegPair *regs,
318 int total_args_passed) {
319 // Create the mapping between argument positions and registers.
320
321 static const Register INT_ArgReg[java_return_convention_max_int] = {
322 x10 /* j_rarg7 */, j_rarg6, j_rarg5, j_rarg4, j_rarg3, j_rarg2, j_rarg1, j_rarg0
323 };
324
325 static const FloatRegister FP_ArgReg[java_return_convention_max_float] = {
326 j_farg0, j_farg1, j_farg2, j_farg3, j_farg4, j_farg5, j_farg6, j_farg7
327 };
328
329 uint int_args = 0;
330 uint fp_args = 0;
331
332 for (int i = 0; i < total_args_passed; i++) {
333 switch (sig_bt[i]) {
334 case T_BOOLEAN:
335 case T_CHAR:
336 case T_BYTE:
337 case T_SHORT:
338 case T_INT:
339 if (int_args < SharedRuntime::java_return_convention_max_int) {
340 regs[i].set1(INT_ArgReg[int_args]->as_VMReg());
341 int_args ++;
342 } else {
343 return -1;
344 }
345 break;
346 case T_VOID:
347 // halves of T_LONG or T_DOUBLE
348 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
349 regs[i].set_bad();
350 break;
351 case T_LONG:
352 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
353 // fall through
354 case T_OBJECT:
355 case T_ARRAY:
356 case T_ADDRESS:
357 // Should T_METADATA be added to java_calling_convention as well ?
358 case T_METADATA:
359 if (int_args < SharedRuntime::java_return_convention_max_int) {
360 regs[i].set2(INT_ArgReg[int_args]->as_VMReg());
361 int_args ++;
362 } else {
363 return -1;
364 }
365 break;
366 case T_FLOAT:
367 if (fp_args < SharedRuntime::java_return_convention_max_float) {
368 regs[i].set1(FP_ArgReg[fp_args]->as_VMReg());
369 fp_args ++;
370 } else {
371 return -1;
372 }
373 break;
374 case T_DOUBLE:
375 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
376 if (fp_args < SharedRuntime::java_return_convention_max_float) {
377 regs[i].set2(FP_ArgReg[fp_args]->as_VMReg());
378 fp_args ++;
379 } else {
380 return -1;
381 }
382 break;
383 default:
384 ShouldNotReachHere();
385 break;
386 }
387 }
388
389 return int_args + fp_args;
390 }
391
392 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
393 Unimplemented();
394 return nullptr;
395 }
396
397 // Patch the callers callsite with entry to compiled code if it exists.
398 static void patch_callers_callsite(MacroAssembler *masm) {
399 Label L;
400 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
401 __ beqz(t0, L);
402
403 __ enter();
404 __ push_CPU_state();
405
406 // VM needs caller's callsite
407 // VM needs target method
408 // This needs to be a long call since we will relocate this adapter to
409 // the codeBuffer and it may not reach
410
411 #ifndef PRODUCT
412 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
413 #endif
414
415 __ mv(c_rarg0, xmethod);
416 __ mv(c_rarg1, ra);
417 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
418
419 __ pop_CPU_state();
420 // restore sp
421 __ leave();
422 __ bind(L);
423 }
424
425 // For each inline type argument, sig includes the list of fields of
426 // the inline type. This utility function computes the number of
427 // arguments for the call if inline types are passed by reference (the
428 // calling convention the interpreter expects).
429 static int compute_total_args_passed_int(const GrowableArray<SigEntry>* sig_extended) {
430 int total_args_passed = 0;
431 assert(!InlineTypePassFieldsAsArgs, "");
432 total_args_passed = sig_extended->length();
433 return total_args_passed;
434 }
435
436 static void gen_c2i_adapter_helper(MacroAssembler* masm,
437 BasicType bt,
438 BasicType prev_bt,
439 size_t size_in_bytes,
440 const VMRegPair& reg_pair,
441 const Address& to,
442 int extraspace) {
443 if (bt == T_VOID) {
444 assert(prev_bt == T_LONG || prev_bt == T_DOUBLE, "missing half");
445 return;
446 }
447
448 // Say 4 args:
449 // i st_off
450 // 0 32 T_LONG
451 // 1 24 T_VOID
452 // 2 16 T_OBJECT
453 // 3 8 T_BOOL
454 // - 0 return address
455 //
456 // However to make thing extra confusing. Because we can fit a Java long/double in
457 // a single slot on a 64 bit vm and it would be silly to break them up, the interpreter
458 // leaves one slot empty and only stores to a single slot. In this case the
459 // slot that is occupied is the T_VOID slot. See I said it was confusing.
460
461 bool wide = (size_in_bytes == wordSize);
462
463 VMReg r_1 = reg_pair.first();
464 VMReg r_2 = reg_pair.second();
465 assert(r_2->is_valid() == wide, "invalid size");
466 if (!r_1->is_valid()) {
467 assert(!r_2->is_valid(), "");
468 return;
469 }
470
471 if (!r_1->is_FloatRegister()) {
472 Register val = t1;
473 if (r_1->is_stack()) {
474 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
475 __ load_sized_value(val, Address(sp, ld_off), size_in_bytes, /* is_signed */ false);
476 } else {
477 val = r_1->as_Register();
478 }
479 __ store_sized_value(to, val, size_in_bytes);
480 } else {
481 if (wide) {
482 __ fsd(r_1->as_FloatRegister(), to);
483 } else {
484 // only a float use just part of the slot
485 __ fsw(r_1->as_FloatRegister(), to);
486 }
487 }
488 }
489
490 static void gen_c2i_adapter(MacroAssembler *masm,
491 const GrowableArray<SigEntry>* sig_extended,
492 const VMRegPair *regs,
493 bool requires_clinit_barrier,
494 address& c2i_no_clinit_check_entry,
495 Label& skip_fixup,
496 address start,
497 OopMapSet* oop_maps,
498 int& frame_complete,
499 int& frame_size_in_words,
500 bool alloc_inline_receiver) {
501 if (requires_clinit_barrier) {
502 assert(VM_Version::supports_fast_class_init_checks(), "sanity");
503 Label L_skip_barrier;
504
505 { // Bypass the barrier for non-static methods
506 __ lhu(t0, Address(xmethod, Method::access_flags_offset()));
507 __ test_bit(t0, t0, exact_log2(JVM_ACC_STATIC));
508 __ beqz(t0, L_skip_barrier); // non-static
509 }
510
511 __ load_method_holder(t1, xmethod);
512 __ clinit_barrier(t1, t0, &L_skip_barrier);
513 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
514
515 __ bind(L_skip_barrier);
516 c2i_no_clinit_check_entry = __ pc();
517 }
518
519 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
520 bs->c2i_entry_barrier(masm);
521
522 // Before we get into the guts of the C2I adapter, see if we should be here
523 // at all. We've come from compiled code and are attempting to jump to the
524 // interpreter, which means the caller made a static call to get here
525 // (vcalls always get a compiled target if there is one). Check for a
526 // compiled target. If there is one, we need to patch the caller's call.
527 patch_callers_callsite(masm);
528
529 __ bind(skip_fixup);
530
531 // Since all args are passed on the stack, total_args_passed *
532 // Interpreter::stackElementSize is the space we need.
533
534 int total_args_passed = compute_total_args_passed_int(sig_extended);
535 int extraspace = total_args_passed * Interpreter::stackElementSize;
536
537 __ mv(x19_sender_sp, sp);
538
539 // stack is aligned, keep it that way
540 extraspace = align_up(extraspace, StackAlignmentInBytes);
541
542 if (extraspace) {
543 __ sub(sp, sp, extraspace);
544 }
545
546 // Now write the args into the outgoing interpreter space
547
548 // next_arg_comp is the next argument from the compiler point of
549 // view (inline type fields are passed in registers/on the stack). In
550 // sig_extended, an inline type argument starts with: T_METADATA,
551 // followed by the types of the fields of the inline type and T_VOID
552 // to mark the end of the inline type. ignored counts the number of
553 // T_METADATA/T_VOID. next_vt_arg is the next inline type argument:
554 // used to get the buffer for that argument from the pool of buffers
555 // we allocated above and want to pass to the
556 // interpreter. next_arg_int is the next argument from the
557 // interpreter point of view (inline types are passed by reference).
558 for (int next_arg_comp = 0, ignored = 0, next_vt_arg = 0, next_arg_int = 0;
559 next_arg_comp < sig_extended->length(); next_arg_comp++) {
560 assert(ignored <= next_arg_comp, "shouldn't skip over more slots than there are arguments");
561 assert(next_arg_int <= total_args_passed, "more arguments for the interpreter than expected?");
562 BasicType bt = sig_extended->at(next_arg_comp)._bt;
563 assert(!InlineTypePassFieldsAsArgs, "");
564
565 int st_off = (total_args_passed - next_arg_int - 1) * Interpreter::stackElementSize;
566 int next_off = st_off - Interpreter::stackElementSize;
567 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : st_off;
568 const VMRegPair reg_pair = regs[next_arg_comp-ignored];
569 size_t size_in_bytes = reg_pair.second()->is_valid() ? 8 : 4;
570 gen_c2i_adapter_helper(masm, bt, next_arg_comp > 0 ? sig_extended->at(next_arg_comp - 1)._bt : T_ILLEGAL,
571 size_in_bytes, reg_pair, Address(sp, offset), extraspace);
572 next_arg_int++;
573
574 #ifdef ASSERT
575 if (bt == T_LONG || bt == T_DOUBLE) {
576 // Overwrite the unused slot with known junk
577 __ mv(t0, CONST64(0xdeadffffdeadaaaa));
578 __ sd(t0, Address(sp, st_off));
579 }
580 #endif /* ASSERT */
581 }
582
583 __ mv(esp, sp); // Interp expects args on caller's expression stack
584
585 __ ld(t1, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
586 __ jr(t1);
587 }
588
589 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
590 int comp_args_on_stack,
591 const GrowableArray<SigEntry>* sig,
592 const VMRegPair *regs) {
593 // Note: x19_sender_sp contains the senderSP on entry. We must
594 // preserve it since we may do a i2c -> c2i transition if we lose a
595 // race where compiled code goes non-entrant while we get args
596 // ready.
597
598 // Cut-out for having no stack args.
599 int comp_words_on_stack = 0;
600 if (comp_args_on_stack != 0) {
601 comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
602 __ sub(t0, sp, comp_words_on_stack * wordSize);
603 __ andi(sp, t0, -16);
604 }
605
606 // Will jump to the compiled code just as if compiled code was doing it.
607 // Pre-load the register-jump target early, to schedule it better.
608 __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_inline_offset())));
609
610 int total_args_passed = sig->length();
611
612 // Now generate the shuffle code.
613 for (int i = 0; i < total_args_passed; i++) {
614 BasicType bt = sig->at(i)._bt;
615 if (bt == T_VOID) {
616 assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
617 continue;
618 }
619
620 // Pick up 0, 1 or 2 words from SP+offset.
621
622 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
623 "scrambled load targets?");
624 // Load in argument order going down.
625 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
626 // Point to interpreter value (vs. tag)
627 int next_off = ld_off - Interpreter::stackElementSize;
628
629 VMReg r_1 = regs[i].first();
630 VMReg r_2 = regs[i].second();
631 if (!r_1->is_valid()) {
632 assert(!r_2->is_valid(), "");
633 continue;
634 }
635 if (r_1->is_stack()) {
636 // Convert stack slot to an SP offset (+ wordSize to account for return address )
637 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
638 if (!r_2->is_valid()) {
639 __ lw(t0, Address(esp, ld_off));
640 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
641 } else {
642 //
643 // We are using two optoregs. This can be either T_OBJECT,
644 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
645 // two slots but only uses one for thr T_LONG or T_DOUBLE case
646 // So we must adjust where to pick up the data to match the
647 // interpreter.
648 //
649 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
650 // are accessed as negative so LSW is at LOW address
651
652 // ld_off is MSW so get LSW
653 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
654 __ ld(t0, Address(esp, offset));
655 // st_off is LSW (i.e. reg.first())
656 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
657 }
658 } else if (r_1->is_Register()) { // Register argument
659 Register r = r_1->as_Register();
660 if (r_2->is_valid()) {
661 //
662 // We are using two VMRegs. This can be either T_OBJECT,
663 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
664 // two slots but only uses one for thr T_LONG or T_DOUBLE case
665 // So we must adjust where to pick up the data to match the
666 // interpreter.
667
668 const int offset = (bt == T_LONG || bt == T_DOUBLE) ? next_off : ld_off;
669
670 // this can be a misaligned move
671 __ ld(r, Address(esp, offset));
672 } else {
673 // sign extend and use a full word?
674 __ lw(r, Address(esp, ld_off));
675 }
676 } else {
677 if (!r_2->is_valid()) {
678 __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
679 } else {
680 __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
681 }
682 }
683 }
684
685 __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
686
687 // 6243940 We might end up in handle_wrong_method if
688 // the callee is deoptimized as we race thru here. If that
689 // happens we don't want to take a safepoint because the
690 // caller frame will look interpreted and arguments are now
691 // "compiled" so it is much better to make this transition
692 // invisible to the stack walking code. Unfortunately if
693 // we try and find the callee by normal means a safepoint
694 // is possible. So we stash the desired callee in the thread
695 // and the vm will find there should this case occur.
696
697 __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
698
699 __ jr(t1);
700 }
701
702 static void gen_inline_cache_check(MacroAssembler *masm, Label& skip_fixup) {
703 Register data = t0;
704
705 __ ic_check();
706 __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
707
708 // Method might have been compiled since the call site was patched to
709 // interpreted; if that is the case treat it as a miss so we can get
710 // the call site corrected.
711 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
712 __ beqz(t0, skip_fixup);
713 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
714 }
715
716 // ---------------------------------------------------------------
717
718 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
719 int comp_args_on_stack,
720 const GrowableArray<SigEntry>* sig,
721 const VMRegPair* regs,
722 const GrowableArray<SigEntry>* sig_cc,
723 const VMRegPair* regs_cc,
724 const GrowableArray<SigEntry>* sig_cc_ro,
725 const VMRegPair* regs_cc_ro,
726 address entry_address[AdapterBlob::ENTRY_COUNT],
727 AdapterBlob*& new_adapter,
728 bool allocate_code_blob) {
729
730 entry_address[AdapterBlob::I2C] = __ pc();
731 gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
732
733 // -------------------------------------------------------------------------
734 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls
735 // to the interpreter. The args start out packed in the compiled layout. They
736 // need to be unpacked into the interpreter layout. This will almost always
737 // require some stack space. We grow the current (compiled) stack, then repack
738 // the args. We finally end in a jump to the generic interpreter entry point.
739 // On exit from the interpreter, the interpreter will restore our SP (lest the
740 // compiled code, which relies solely on SP and not FP, get sick).
741 entry_address[AdapterBlob::C2I_Unverified] = __ pc();
742 entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
743
744 Label skip_fixup;
745 gen_inline_cache_check(masm, skip_fixup);
746
747 OopMapSet* oop_maps = new OopMapSet();
748 int frame_complete = CodeOffsets::frame_never_safe;
749 int frame_size_in_words = 0;
750
751 // Scalarized c2i adapter with non-scalarized receiver (i.e., don't pack receiver)
752 entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
753 entry_address[AdapterBlob::C2I_Inline_RO] = __ pc();
754 if (regs_cc != regs_cc_ro) {
755 // No class init barrier needed because method is guaranteed to be non-static
756 gen_c2i_adapter(masm, sig_cc_ro, regs_cc_ro, /* requires_clinit_barrier = */ false, entry_address[AdapterBlob::C2I_No_Clinit_Check],
757 skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
758 skip_fixup.reset();
759 }
760
761 // Scalarized c2i adapter
762 entry_address[AdapterBlob::C2I] = __ pc();
763 entry_address[AdapterBlob::C2I_Inline] = __ pc();
764 gen_c2i_adapter(masm, sig_cc, regs_cc, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
765 skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ true);
766
767 // Non-scalarized c2i adapter
768 if (regs != regs_cc) {
769 entry_address[AdapterBlob::C2I_Unverified_Inline] = __ pc();
770 Label inline_entry_skip_fixup;
771 gen_inline_cache_check(masm, inline_entry_skip_fixup);
772
773 entry_address[AdapterBlob::C2I_Inline] = __ pc();
774 gen_c2i_adapter(masm, sig, regs, /* requires_clinit_barrier = */ true, entry_address[AdapterBlob::C2I_No_Clinit_Check],
775 inline_entry_skip_fixup, entry_address[AdapterBlob::I2C], oop_maps, frame_complete, frame_size_in_words, /* alloc_inline_receiver = */ false);
776 }
777
778 // The c2i adapters might safepoint and trigger a GC. The caller must make sure that
779 // the GC knows about the location of oop argument locations passed to the c2i adapter.
780 if (allocate_code_blob) {
781 bool caller_must_gc_arguments = (regs != regs_cc);
782 int entry_offset[AdapterHandlerEntry::ENTRIES_COUNT];
783 assert(AdapterHandlerEntry::ENTRIES_COUNT == 7, "sanity");
784 AdapterHandlerLibrary::address_to_offset(entry_address, entry_offset);
785 new_adapter = AdapterBlob::create(masm->code(), entry_offset, frame_complete, frame_size_in_words, oop_maps, caller_must_gc_arguments);
786 }
787 }
788
789 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
790 uint num_bits,
791 uint total_args_passed) {
792 assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
793 assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
794
795 // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
796 static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
797 v8, v9, v10, v11, v12, v13, v14, v15,
798 v16, v17, v18, v19, v20, v21, v22, v23
799 };
800
801 const int next_reg_val = 3;
802 for (uint i = 0; i < total_args_passed; i++) {
803 VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
804 regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
805 }
806 return 0;
807 }
808
809 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
810 VMRegPair *regs,
811 int total_args_passed) {
812
813 // We return the amount of VMRegImpl stack slots we need to reserve for all
814 // the arguments NOT counting out_preserve_stack_slots.
815
816 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
817 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
818 c_rarg4, c_rarg5, c_rarg6, c_rarg7
819 };
820 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
821 c_farg0, c_farg1, c_farg2, c_farg3,
822 c_farg4, c_farg5, c_farg6, c_farg7
823 };
824
825 uint int_args = 0;
826 uint fp_args = 0;
827 uint stk_args = 0; // inc by 2 each time
828
829 for (int i = 0; i < total_args_passed; i++) {
830 switch (sig_bt[i]) {
831 case T_BOOLEAN: // fall through
832 case T_CHAR: // fall through
833 case T_BYTE: // fall through
834 case T_SHORT: // fall through
835 case T_INT:
836 if (int_args < Argument::n_int_register_parameters_c) {
837 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
838 } else {
839 regs[i].set1(VMRegImpl::stack2reg(stk_args));
840 stk_args += 2;
841 }
842 break;
843 case T_LONG: // fall through
844 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
845 case T_OBJECT: // fall through
846 case T_ARRAY: // fall through
847 case T_ADDRESS: // fall through
848 case T_METADATA:
849 if (int_args < Argument::n_int_register_parameters_c) {
850 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
851 } else {
852 regs[i].set2(VMRegImpl::stack2reg(stk_args));
853 stk_args += 2;
854 }
855 break;
856 case T_FLOAT:
857 if (fp_args < Argument::n_float_register_parameters_c) {
858 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
859 } else if (int_args < Argument::n_int_register_parameters_c) {
860 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
861 } else {
862 regs[i].set1(VMRegImpl::stack2reg(stk_args));
863 stk_args += 2;
864 }
865 break;
866 case T_DOUBLE:
867 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
868 if (fp_args < Argument::n_float_register_parameters_c) {
869 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
870 } else if (int_args < Argument::n_int_register_parameters_c) {
871 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
872 } else {
873 regs[i].set2(VMRegImpl::stack2reg(stk_args));
874 stk_args += 2;
875 }
876 break;
877 case T_VOID: // Halves of longs and doubles
878 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
879 regs[i].set_bad();
880 break;
881 default:
882 ShouldNotReachHere();
883 }
884 }
885
886 return stk_args;
887 }
888
889 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
890 // We always ignore the frame_slots arg and just use the space just below frame pointer
891 // which by this time is free to use
892 switch (ret_type) {
893 case T_FLOAT:
894 __ fsw(f10, Address(fp, -3 * wordSize));
895 break;
896 case T_DOUBLE:
897 __ fsd(f10, Address(fp, -3 * wordSize));
898 break;
899 case T_VOID: break;
900 default: {
901 __ sd(x10, Address(fp, -3 * wordSize));
902 }
903 }
904 }
905
906 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
907 // We always ignore the frame_slots arg and just use the space just below frame pointer
908 // which by this time is free to use
909 switch (ret_type) {
910 case T_FLOAT:
911 __ flw(f10, Address(fp, -3 * wordSize));
912 break;
913 case T_DOUBLE:
914 __ fld(f10, Address(fp, -3 * wordSize));
915 break;
916 case T_VOID: break;
917 default: {
918 __ ld(x10, Address(fp, -3 * wordSize));
919 }
920 }
921 }
922
923 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
924 RegSet x;
925 for ( int i = first_arg ; i < arg_count ; i++ ) {
926 if (args[i].first()->is_Register()) {
927 x = x + args[i].first()->as_Register();
928 } else if (args[i].first()->is_FloatRegister()) {
929 __ subi(sp, sp, 2 * wordSize);
930 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
931 }
932 }
933 __ push_reg(x, sp);
934 }
935
936 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
937 RegSet x;
938 for ( int i = first_arg ; i < arg_count ; i++ ) {
939 if (args[i].first()->is_Register()) {
940 x = x + args[i].first()->as_Register();
941 } else {
942 ;
943 }
944 }
945 __ pop_reg(x, sp);
946 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
947 if (args[i].first()->is_Register()) {
948 ;
949 } else if (args[i].first()->is_FloatRegister()) {
950 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
951 __ addi(sp, sp, 2 * wordSize);
952 }
953 }
954 }
955
956 static void verify_oop_args(MacroAssembler* masm,
957 const methodHandle& method,
958 const BasicType* sig_bt,
959 const VMRegPair* regs) {
960 const Register temp_reg = x9; // not part of any compiled calling seq
961 if (VerifyOops) {
962 for (int i = 0; i < method->size_of_parameters(); i++) {
963 if (sig_bt[i] == T_OBJECT ||
964 sig_bt[i] == T_ARRAY) {
965 VMReg r = regs[i].first();
966 assert(r->is_valid(), "bad oop arg");
967 if (r->is_stack()) {
968 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
969 __ verify_oop(temp_reg);
970 } else {
971 __ verify_oop(r->as_Register());
972 }
973 }
974 }
975 }
976 }
977
978 // on exit, sp points to the ContinuationEntry
979 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
980 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
981 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
982 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
983
984 stack_slots += (int)ContinuationEntry::size() / wordSize;
985 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
986
987 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
988
989 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
990 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
991 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
992
993 return map;
994 }
995
996 // on entry c_rarg1 points to the continuation
997 // sp points to ContinuationEntry
998 // c_rarg3 -- isVirtualThread
999 static void fill_continuation_entry(MacroAssembler* masm) {
1000 #ifdef ASSERT
1001 __ mv(t0, ContinuationEntry::cookie_value());
1002 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
1003 #endif
1004
1005 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
1006 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
1007 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
1008 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
1009 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
1010
1011 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
1012 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1013
1014 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
1015 }
1016
1017 // on entry, sp points to the ContinuationEntry
1018 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
1019 static void continuation_enter_cleanup(MacroAssembler* masm) {
1020 #ifndef PRODUCT
1021 Label OK;
1022 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
1023 __ beq(sp, t0, OK);
1024 __ stop("incorrect sp");
1025 __ bind(OK);
1026 #endif
1027
1028 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
1029 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
1030 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
1031 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
1032 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
1033 }
1034
1035 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1036 // On entry: c_rarg1 -- the continuation object
1037 // c_rarg2 -- isContinue
1038 // c_rarg3 -- isVirtualThread
1039 static void gen_continuation_enter(MacroAssembler* masm,
1040 const methodHandle& method,
1041 const BasicType* sig_bt,
1042 const VMRegPair* regs,
1043 int& exception_offset,
1044 OopMapSet*oop_maps,
1045 int& frame_complete,
1046 int& stack_slots,
1047 int& interpreted_entry_offset,
1048 int& compiled_entry_offset) {
1049 // verify_oop_args(masm, method, sig_bt, regs);
1050 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
1051
1052 address start = __ pc();
1053
1054 Label call_thaw, exit;
1055
1056 // i2i entry used at interp_only_mode only
1057 interpreted_entry_offset = __ pc() - start;
1058 {
1059 #ifdef ASSERT
1060 Label is_interp_only;
1061 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
1062 __ bnez(t0, is_interp_only);
1063 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1064 __ bind(is_interp_only);
1065 #endif
1066
1067 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1068 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
1069 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
1070 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
1071 __ push_cont_fastpath(xthread);
1072
1073 __ enter();
1074 stack_slots = 2; // will be adjusted in setup
1075 OopMap* map = continuation_enter_setup(masm, stack_slots);
1076 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1077 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1078
1079 fill_continuation_entry(masm);
1080
1081 __ bnez(c_rarg2, call_thaw);
1082
1083 address call_pc;
1084 {
1085 Assembler::IncompressibleScope scope(masm);
1086 // Make sure the call is patchable
1087 __ align(NativeInstruction::instruction_size);
1088
1089 call_pc = __ reloc_call(resolve);
1090 if (call_pc == nullptr) {
1091 fatal("CodeCache is full at gen_continuation_enter");
1092 }
1093
1094 oop_maps->add_gc_map(__ pc() - start, map);
1095 __ post_call_nop();
1096 }
1097 __ j(exit);
1098
1099 address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1100 if (stub == nullptr) {
1101 fatal("CodeCache is full at gen_continuation_enter");
1102 }
1103 }
1104
1105 // compiled entry
1106 __ align(CodeEntryAlignment);
1107 compiled_entry_offset = __ pc() - start;
1108
1109 __ enter();
1110 stack_slots = 2; // will be adjusted in setup
1111 OopMap* map = continuation_enter_setup(masm, stack_slots);
1112 frame_complete = __ pc() - start;
1113
1114 fill_continuation_entry(masm);
1115
1116 __ bnez(c_rarg2, call_thaw);
1117
1118 address call_pc;
1119 {
1120 Assembler::IncompressibleScope scope(masm);
1121 // Make sure the call is patchable
1122 __ align(NativeInstruction::instruction_size);
1123
1124 call_pc = __ reloc_call(resolve);
1125 if (call_pc == nullptr) {
1126 fatal("CodeCache is full at gen_continuation_enter");
1127 }
1128
1129 oop_maps->add_gc_map(__ pc() - start, map);
1130 __ post_call_nop();
1131 }
1132
1133 __ j(exit);
1134
1135 __ bind(call_thaw);
1136
1137 // Post call nops must be natural aligned due to cmodx rules.
1138 {
1139 Assembler::IncompressibleScope scope(masm);
1140 __ align(NativeInstruction::instruction_size);
1141
1142 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1143 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1144 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1145 ContinuationEntry::_return_pc_offset = __ pc() - start;
1146 __ post_call_nop();
1147 }
1148
1149 __ bind(exit);
1150 ContinuationEntry::_cleanup_offset = __ pc() - start;
1151 continuation_enter_cleanup(masm);
1152 __ leave();
1153 __ ret();
1154
1155 // exception handling
1156 exception_offset = __ pc() - start;
1157 {
1158 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1159
1160 continuation_enter_cleanup(masm);
1161
1162 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1163 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1164
1165 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1166
1167 __ mv(x11, x10); // the exception handler
1168 __ mv(x10, x9); // restore return value contaning the exception oop
1169 __ verify_oop(x10);
1170
1171 __ leave();
1172 __ mv(x13, ra);
1173 __ jr(x11); // the exception handler
1174 }
1175
1176 address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1177 if (stub == nullptr) {
1178 fatal("CodeCache is full at gen_continuation_enter");
1179 }
1180 }
1181
1182 static void gen_continuation_yield(MacroAssembler* masm,
1183 const methodHandle& method,
1184 const BasicType* sig_bt,
1185 const VMRegPair* regs,
1186 OopMapSet* oop_maps,
1187 int& frame_complete,
1188 int& stack_slots,
1189 int& compiled_entry_offset) {
1190 enum layout {
1191 fp_off,
1192 fp_off2,
1193 return_off,
1194 return_off2,
1195 framesize // inclusive of return address
1196 };
1197 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1198
1199 stack_slots = framesize / VMRegImpl::slots_per_word;
1200 assert(stack_slots == 2, "recheck layout");
1201
1202 address start = __ pc();
1203
1204 compiled_entry_offset = __ pc() - start;
1205 __ enter();
1206
1207 __ mv(c_rarg1, sp);
1208
1209 // Post call nops must be natural aligned due to cmodx rules.
1210 __ align(NativeInstruction::instruction_size);
1211
1212 frame_complete = __ pc() - start;
1213 address the_pc = __ pc();
1214
1215 {
1216 Assembler::IncompressibleScope scope(masm);
1217 __ 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
1218 }
1219
1220 __ mv(c_rarg0, xthread);
1221 __ set_last_Java_frame(sp, fp, the_pc, t0);
1222 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1223 __ reset_last_Java_frame(true);
1224
1225 Label pinned;
1226
1227 __ bnez(x10, pinned);
1228
1229 // We've succeeded, set sp to the ContinuationEntry
1230 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1231 continuation_enter_cleanup(masm);
1232
1233 __ bind(pinned); // pinned -- return to caller
1234
1235 // handle pending exception thrown by freeze
1236 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1237 Label ok;
1238 __ beqz(t0, ok);
1239 __ leave();
1240 __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1241 __ bind(ok);
1242
1243 __ leave();
1244 __ ret();
1245
1246 OopMap* map = new OopMap(framesize, 1);
1247 oop_maps->add_gc_map(the_pc - start, map);
1248 }
1249
1250 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1251 ::continuation_enter_cleanup(masm);
1252 }
1253
1254 static void gen_special_dispatch(MacroAssembler* masm,
1255 const methodHandle& method,
1256 const BasicType* sig_bt,
1257 const VMRegPair* regs) {
1258 verify_oop_args(masm, method, sig_bt, regs);
1259 vmIntrinsics::ID iid = method->intrinsic_id();
1260
1261 // Now write the args into the outgoing interpreter space
1262 bool has_receiver = false;
1263 Register receiver_reg = noreg;
1264 int member_arg_pos = -1;
1265 Register member_reg = noreg;
1266 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1267 if (ref_kind != 0) {
1268 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1269 member_reg = x9; // known to be free at this point
1270 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1271 } else if (iid == vmIntrinsics::_invokeBasic) {
1272 has_receiver = true;
1273 } else if (iid == vmIntrinsics::_linkToNative) {
1274 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1275 member_reg = x9; // known to be free at this point
1276 } else {
1277 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1278 }
1279
1280 if (member_reg != noreg) {
1281 // Load the member_arg into register, if necessary.
1282 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1283 VMReg r = regs[member_arg_pos].first();
1284 if (r->is_stack()) {
1285 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1286 } else {
1287 // no data motion is needed
1288 member_reg = r->as_Register();
1289 }
1290 }
1291
1292 if (has_receiver) {
1293 // Make sure the receiver is loaded into a register.
1294 assert(method->size_of_parameters() > 0, "oob");
1295 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1296 VMReg r = regs[0].first();
1297 assert(r->is_valid(), "bad receiver arg");
1298 if (r->is_stack()) {
1299 // Porting note: This assumes that compiled calling conventions always
1300 // pass the receiver oop in a register. If this is not true on some
1301 // platform, pick a temp and load the receiver from stack.
1302 fatal("receiver always in a register");
1303 receiver_reg = x12; // known to be free at this point
1304 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1305 } else {
1306 // no data motion is needed
1307 receiver_reg = r->as_Register();
1308 }
1309 }
1310
1311 // Figure out which address we are really jumping to:
1312 MethodHandles::generate_method_handle_dispatch(masm, iid,
1313 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1314 }
1315
1316 // ---------------------------------------------------------------------------
1317 // Generate a native wrapper for a given method. The method takes arguments
1318 // in the Java compiled code convention, marshals them to the native
1319 // convention (handlizes oops, etc), transitions to native, makes the call,
1320 // returns to java state (possibly blocking), unhandlizes any result and
1321 // returns.
1322 //
1323 // Critical native functions are a shorthand for the use of
1324 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1325 // functions. The wrapper is expected to unpack the arguments before
1326 // passing them to the callee and perform checks before and after the
1327 // native call to ensure that they GCLocker
1328 // lock_critical/unlock_critical semantics are followed. Some other
1329 // parts of JNI setup are skipped like the tear down of the JNI handle
1330 // block and the check for pending exceptions it's impossible for them
1331 // to be thrown.
1332 //
1333 // They are roughly structured like this:
1334 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1335 // tranistion to thread_in_native
1336 // unpack array arguments and call native entry point
1337 // check for safepoint in progress
1338 // check if any thread suspend flags are set
1339 // call into JVM and possible unlock the JNI critical
1340 // if a GC was suppressed while in the critical native.
1341 // transition back to thread_in_Java
1342 // return to caller
1343 //
1344 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1345 const methodHandle& method,
1346 int compile_id,
1347 BasicType* in_sig_bt,
1348 VMRegPair* in_regs,
1349 BasicType ret_type) {
1350 if (method->is_continuation_native_intrinsic()) {
1351 int exception_offset = -1;
1352 OopMapSet* oop_maps = new OopMapSet();
1353 int frame_complete = -1;
1354 int stack_slots = -1;
1355 int interpreted_entry_offset = -1;
1356 int vep_offset = -1;
1357 if (method->is_continuation_enter_intrinsic()) {
1358 gen_continuation_enter(masm,
1359 method,
1360 in_sig_bt,
1361 in_regs,
1362 exception_offset,
1363 oop_maps,
1364 frame_complete,
1365 stack_slots,
1366 interpreted_entry_offset,
1367 vep_offset);
1368 } else if (method->is_continuation_yield_intrinsic()) {
1369 gen_continuation_yield(masm,
1370 method,
1371 in_sig_bt,
1372 in_regs,
1373 oop_maps,
1374 frame_complete,
1375 stack_slots,
1376 vep_offset);
1377 } else {
1378 guarantee(false, "Unknown Continuation native intrinsic");
1379 }
1380
1381 #ifdef ASSERT
1382 if (method->is_continuation_enter_intrinsic()) {
1383 assert(interpreted_entry_offset != -1, "Must be set");
1384 assert(exception_offset != -1, "Must be set");
1385 } else {
1386 assert(interpreted_entry_offset == -1, "Must be unset");
1387 assert(exception_offset == -1, "Must be unset");
1388 }
1389 assert(frame_complete != -1, "Must be set");
1390 assert(stack_slots != -1, "Must be set");
1391 assert(vep_offset != -1, "Must be set");
1392 #endif
1393
1394 __ flush();
1395 nmethod* nm = nmethod::new_native_nmethod(method,
1396 compile_id,
1397 masm->code(),
1398 vep_offset,
1399 frame_complete,
1400 stack_slots,
1401 in_ByteSize(-1),
1402 in_ByteSize(-1),
1403 oop_maps,
1404 exception_offset);
1405 if (nm == nullptr) return nm;
1406 if (method->is_continuation_enter_intrinsic()) {
1407 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1408 } else if (method->is_continuation_yield_intrinsic()) {
1409 _cont_doYield_stub = nm;
1410 } else {
1411 guarantee(false, "Unknown Continuation native intrinsic");
1412 }
1413 return nm;
1414 }
1415
1416 if (method->is_method_handle_intrinsic()) {
1417 vmIntrinsics::ID iid = method->intrinsic_id();
1418 intptr_t start = (intptr_t)__ pc();
1419 int vep_offset = ((intptr_t)__ pc()) - start;
1420
1421 // First instruction must be a nop as it may need to be patched on deoptimisation
1422 {
1423 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1424 MacroAssembler::assert_alignment(__ pc());
1425 __ nop(); // 4 bytes
1426 }
1427 gen_special_dispatch(masm,
1428 method,
1429 in_sig_bt,
1430 in_regs);
1431 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1432 __ flush();
1433 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1434 return nmethod::new_native_nmethod(method,
1435 compile_id,
1436 masm->code(),
1437 vep_offset,
1438 frame_complete,
1439 stack_slots / VMRegImpl::slots_per_word,
1440 in_ByteSize(-1),
1441 in_ByteSize(-1),
1442 (OopMapSet*)nullptr);
1443 }
1444 address native_func = method->native_function();
1445 assert(native_func != nullptr, "must have function");
1446
1447 // An OopMap for lock (and class if static)
1448 OopMapSet *oop_maps = new OopMapSet();
1449 assert_cond(oop_maps != nullptr);
1450 intptr_t start = (intptr_t)__ pc();
1451
1452 // We have received a description of where all the java arg are located
1453 // on entry to the wrapper. We need to convert these args to where
1454 // the jni function will expect them. To figure out where they go
1455 // we convert the java signature to a C signature by inserting
1456 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1457
1458 const int total_in_args = method->size_of_parameters();
1459 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1460
1461 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1462 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1463
1464 int argc = 0;
1465 out_sig_bt[argc++] = T_ADDRESS;
1466 if (method->is_static()) {
1467 out_sig_bt[argc++] = T_OBJECT;
1468 }
1469
1470 for (int i = 0; i < total_in_args ; i++) {
1471 out_sig_bt[argc++] = in_sig_bt[i];
1472 }
1473
1474 // Now figure out where the args must be stored and how much stack space
1475 // they require.
1476 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1477
1478 // Compute framesize for the wrapper. We need to handlize all oops in
1479 // incoming registers
1480
1481 // Calculate the total number of stack slots we will need.
1482
1483 // First count the abi requirement plus all of the outgoing args
1484 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1485
1486 // Now the space for the inbound oop handle area
1487 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1488
1489 int oop_handle_offset = stack_slots;
1490 stack_slots += total_save_slots;
1491
1492 // Now any space we need for handlizing a klass if static method
1493
1494 int klass_slot_offset = 0;
1495 int klass_offset = -1;
1496 int lock_slot_offset = 0;
1497 bool is_static = false;
1498
1499 if (method->is_static()) {
1500 klass_slot_offset = stack_slots;
1501 stack_slots += VMRegImpl::slots_per_word;
1502 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1503 is_static = true;
1504 }
1505
1506 // Plus a lock if needed
1507
1508 if (method->is_synchronized()) {
1509 lock_slot_offset = stack_slots;
1510 stack_slots += VMRegImpl::slots_per_word;
1511 }
1512
1513 // Now a place (+2) to save return values or temp during shuffling
1514 // + 4 for return address (which we own) and saved fp
1515 stack_slots += 6;
1516
1517 // Ok The space we have allocated will look like:
1518 //
1519 //
1520 // FP-> | |
1521 // | 2 slots (ra) |
1522 // | 2 slots (fp) |
1523 // |---------------------|
1524 // | 2 slots for moves |
1525 // |---------------------|
1526 // | lock box (if sync) |
1527 // |---------------------| <- lock_slot_offset
1528 // | klass (if static) |
1529 // |---------------------| <- klass_slot_offset
1530 // | oopHandle area |
1531 // |---------------------| <- oop_handle_offset (8 java arg registers)
1532 // | outbound memory |
1533 // | based arguments |
1534 // | |
1535 // |---------------------|
1536 // | |
1537 // SP-> | out_preserved_slots |
1538 //
1539 //
1540
1541
1542 // Now compute actual number of stack words we need rounding to make
1543 // stack properly aligned.
1544 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1545
1546 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1547
1548 // First thing make an ic check to see if we should even be here
1549
1550 // We are free to use all registers as temps without saving them and
1551 // restoring them except fp. fp is the only callee save register
1552 // as far as the interpreter and the compiler(s) are concerned.
1553
1554 const Register receiver = j_rarg0;
1555
1556 __ verify_oop(receiver);
1557 assert_different_registers(receiver, t0, t1);
1558
1559 __ ic_check();
1560
1561 int vep_offset = ((intptr_t)__ pc()) - start;
1562
1563 // If we have to make this method not-entrant we'll overwrite its
1564 // first instruction with a jump.
1565 {
1566 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1567 MacroAssembler::assert_alignment(__ pc());
1568 __ nop(); // 4 bytes
1569 }
1570
1571 if (method->needs_clinit_barrier()) {
1572 assert(VM_Version::supports_fast_class_init_checks(), "sanity");
1573 Label L_skip_barrier;
1574 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1575 __ clinit_barrier(t1, t0, &L_skip_barrier);
1576 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1577
1578 __ bind(L_skip_barrier);
1579 }
1580
1581 // Generate stack overflow check
1582 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1583
1584 // Generate a new frame for the wrapper.
1585 __ enter();
1586 // -2 because return address is already present and so is saved fp
1587 __ sub(sp, sp, stack_size - 2 * wordSize);
1588
1589 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1590 assert_cond(bs != nullptr);
1591 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1592
1593 // Frame is now completed as far as size and linkage.
1594 int frame_complete = ((intptr_t)__ pc()) - start;
1595
1596 // We use x18 as the oop handle for the receiver/klass
1597 // It is callee save so it survives the call to native
1598
1599 const Register oop_handle_reg = x18;
1600
1601 //
1602 // We immediately shuffle the arguments so that any vm call we have to
1603 // make from here on out (sync slow path, jvmti, etc.) we will have
1604 // captured the oops from our caller and have a valid oopMap for
1605 // them.
1606
1607 // -----------------
1608 // The Grand Shuffle
1609
1610 // The Java calling convention is either equal (linux) or denser (win64) than the
1611 // c calling convention. However the because of the jni_env argument the c calling
1612 // convention always has at least one more (and two for static) arguments than Java.
1613 // Therefore if we move the args from java -> c backwards then we will never have
1614 // a register->register conflict and we don't have to build a dependency graph
1615 // and figure out how to break any cycles.
1616 //
1617
1618 // Record esp-based slot for receiver on stack for non-static methods
1619 int receiver_offset = -1;
1620
1621 // This is a trick. We double the stack slots so we can claim
1622 // the oops in the caller's frame. Since we are sure to have
1623 // more args than the caller doubling is enough to make
1624 // sure we can capture all the incoming oop args from the
1625 // caller.
1626 //
1627 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1628 assert_cond(map != nullptr);
1629
1630 int float_args = 0;
1631 int int_args = 0;
1632
1633 #ifdef ASSERT
1634 bool reg_destroyed[Register::number_of_registers];
1635 bool freg_destroyed[FloatRegister::number_of_registers];
1636 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1637 reg_destroyed[r] = false;
1638 }
1639 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1640 freg_destroyed[f] = false;
1641 }
1642
1643 #endif /* ASSERT */
1644
1645 // For JNI natives the incoming and outgoing registers are offset upwards.
1646 GrowableArray<int> arg_order(2 * total_in_args);
1647
1648 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1649 arg_order.push(i);
1650 arg_order.push(c_arg);
1651 }
1652
1653 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1654 int i = arg_order.at(ai);
1655 int c_arg = arg_order.at(ai + 1);
1656 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1657 assert(c_arg != -1 && i != -1, "wrong order");
1658 #ifdef ASSERT
1659 if (in_regs[i].first()->is_Register()) {
1660 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1661 } else if (in_regs[i].first()->is_FloatRegister()) {
1662 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1663 }
1664 if (out_regs[c_arg].first()->is_Register()) {
1665 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1666 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1667 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1668 }
1669 #endif /* ASSERT */
1670 switch (in_sig_bt[i]) {
1671 case T_ARRAY:
1672 case T_OBJECT:
1673 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1674 ((i == 0) && (!is_static)),
1675 &receiver_offset);
1676 int_args++;
1677 break;
1678 case T_VOID:
1679 break;
1680
1681 case T_FLOAT:
1682 __ float_move(in_regs[i], out_regs[c_arg]);
1683 float_args++;
1684 break;
1685
1686 case T_DOUBLE:
1687 assert( i + 1 < total_in_args &&
1688 in_sig_bt[i + 1] == T_VOID &&
1689 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1690 __ double_move(in_regs[i], out_regs[c_arg]);
1691 float_args++;
1692 break;
1693
1694 case T_LONG :
1695 __ long_move(in_regs[i], out_regs[c_arg]);
1696 int_args++;
1697 break;
1698
1699 case T_ADDRESS:
1700 assert(false, "found T_ADDRESS in java args");
1701 break;
1702
1703 default:
1704 __ move32_64(in_regs[i], out_regs[c_arg]);
1705 int_args++;
1706 }
1707 }
1708
1709 // point c_arg at the first arg that is already loaded in case we
1710 // need to spill before we call out
1711 int c_arg = total_c_args - total_in_args;
1712
1713 // Pre-load a static method's oop into c_rarg1.
1714 if (method->is_static()) {
1715
1716 // load oop into a register
1717 __ movoop(c_rarg1,
1718 JNIHandles::make_local(method->method_holder()->java_mirror()));
1719
1720 // Now handlize the static class mirror it's known not-null.
1721 __ sd(c_rarg1, Address(sp, klass_offset));
1722 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1723
1724 // Now get the handle
1725 __ la(c_rarg1, Address(sp, klass_offset));
1726 // and protect the arg if we must spill
1727 c_arg--;
1728 }
1729
1730 // Change state to native (we save the return address in the thread, since it might not
1731 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1732 // points into the right code segment. It does not have to be the correct return pc.
1733 // We use the same pc/oopMap repeatedly when we call out.
1734
1735 Label native_return;
1736 if (method->is_object_wait0()) {
1737 // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1738 __ set_last_Java_frame(sp, noreg, native_return, t0);
1739 } else {
1740 intptr_t the_pc = (intptr_t) __ pc();
1741 oop_maps->add_gc_map(the_pc - start, map);
1742
1743 __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1744 }
1745
1746 Label dtrace_method_entry, dtrace_method_entry_done;
1747 if (DTraceMethodProbes) {
1748 __ j(dtrace_method_entry);
1749 __ bind(dtrace_method_entry_done);
1750 }
1751
1752 // RedefineClasses() tracing support for obsolete method entry
1753 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1754 // protect the args we've loaded
1755 save_args(masm, total_c_args, c_arg, out_regs);
1756 __ mov_metadata(c_rarg1, method());
1757 __ call_VM_leaf(
1758 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1759 xthread, c_rarg1);
1760 restore_args(masm, total_c_args, c_arg, out_regs);
1761 }
1762
1763 // Lock a synchronized method
1764
1765 // Register definitions used by locking and unlocking
1766
1767 const Register swap_reg = x10;
1768 const Register obj_reg = x9; // Will contain the oop
1769 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1770 const Register old_hdr = x30; // value of old header at unlock time
1771 const Register lock_tmp = x31; // Temporary used by fast_lock/unlock
1772 const Register tmp = ra;
1773
1774 Label slow_path_lock;
1775 Label lock_done;
1776
1777 if (method->is_synchronized()) {
1778 // Get the handle (the 2nd argument)
1779 __ mv(oop_handle_reg, c_rarg1);
1780
1781 // Get address of the box
1782
1783 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1784
1785 // Load the oop from the handle
1786 __ ld(obj_reg, Address(oop_handle_reg, 0));
1787
1788 __ fast_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1789
1790 // Slow path will re-enter here
1791 __ bind(lock_done);
1792 }
1793
1794
1795 // Finally just about ready to make the JNI call
1796
1797 // get JNIEnv* which is first argument to native
1798 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1799
1800 // Now set thread in native
1801 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1802 __ mv(t0, _thread_in_native);
1803 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1804 __ sw(t0, Address(t1));
1805
1806 // Clobbers t1
1807 __ rt_call(native_func);
1808
1809 // Verify or restore cpu control state after JNI call
1810 __ restore_cpu_control_state_after_jni(t0);
1811
1812 // Unpack native results.
1813 if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1814 __ cast_primitive_type(ret_type, x10);
1815 }
1816
1817 Label safepoint_in_progress, safepoint_in_progress_done;
1818
1819 // Switch thread to "native transition" state before reading the synchronization state.
1820 // This additional state is necessary because reading and testing the synchronization
1821 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1822 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1823 // VM thread changes sync state to synchronizing and suspends threads for GC.
1824 // Thread A is resumed to finish this native method, but doesn't block here since it
1825 // didn't see any synchronization is progress, and escapes.
1826 __ mv(t0, _thread_in_native_trans);
1827
1828 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1829
1830 // Force this write out before the read below
1831 if (!UseSystemMemoryBarrier) {
1832 __ membar(MacroAssembler::AnyAny);
1833 }
1834
1835 // check for safepoint operation in progress and/or pending suspend requests
1836 {
1837 __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
1838 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1839 __ bnez(t0, safepoint_in_progress);
1840 __ bind(safepoint_in_progress_done);
1841 }
1842
1843 // change thread state
1844 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1845 __ mv(t0, _thread_in_Java);
1846 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1847 __ sw(t0, Address(t1));
1848
1849 if (method->is_object_wait0()) {
1850 // Check preemption for Object.wait()
1851 __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1852 __ beqz(t1, native_return);
1853 __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1854 __ jr(t1);
1855 __ bind(native_return);
1856
1857 intptr_t the_pc = (intptr_t) __ pc();
1858 oop_maps->add_gc_map(the_pc - start, map);
1859 }
1860
1861 Label reguard;
1862 Label reguard_done;
1863 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1864 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1865 __ beq(t0, t1, reguard);
1866 __ bind(reguard_done);
1867
1868 // native result if any is live
1869
1870 // Unlock
1871 Label unlock_done;
1872 Label slow_path_unlock;
1873 if (method->is_synchronized()) {
1874
1875 // Get locked oop from the handle we passed to jni
1876 __ ld(obj_reg, Address(oop_handle_reg, 0));
1877
1878 // Must save x10 if if it is live now because cmpxchg must use it
1879 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1880 save_native_result(masm, ret_type, stack_slots);
1881 }
1882
1883 __ fast_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1884
1885 // slow path re-enters here
1886 __ bind(unlock_done);
1887 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1888 restore_native_result(masm, ret_type, stack_slots);
1889 }
1890 }
1891
1892 Label dtrace_method_exit, dtrace_method_exit_done;
1893 if (DTraceMethodProbes) {
1894 __ j(dtrace_method_exit);
1895 __ bind(dtrace_method_exit_done);
1896 }
1897
1898 __ reset_last_Java_frame(false);
1899
1900 // Unbox oop result, e.g. JNIHandles::resolve result.
1901 if (is_reference_type(ret_type)) {
1902 __ resolve_jobject(x10, x11, x12);
1903 }
1904
1905 if (CheckJNICalls) {
1906 // clear_pending_jni_exception_check
1907 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1908 }
1909
1910 // reset handle block
1911 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1912 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1913
1914 __ leave();
1915
1916 #if INCLUDE_JFR
1917 // We need to do a poll test after unwind in case the sampler
1918 // managed to sample the native frame after returning to Java.
1919 Label L_return;
1920 __ ld(t0, Address(xthread, JavaThread::polling_word_offset()));
1921 address poll_test_pc = __ pc();
1922 __ relocate(relocInfo::poll_return_type);
1923 __ test_bit(t0, t0, log2i_exact(SafepointMechanism::poll_bit()));
1924 __ beqz(t0, L_return);
1925 assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
1926 "polling page return stub not created yet");
1927 address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
1928 __ la(t0, InternalAddress(poll_test_pc));
1929 __ sd(t0, Address(xthread, JavaThread::saved_exception_pc_offset()));
1930 __ far_jump(RuntimeAddress(stub));
1931 __ bind(L_return);
1932 #endif // INCLUDE_JFR
1933
1934 // Any exception pending?
1935 Label exception_pending;
1936 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1937 __ bnez(t0, exception_pending);
1938
1939 // We're done
1940 __ ret();
1941
1942 // Unexpected paths are out of line and go here
1943
1944 // forward the exception
1945 __ bind(exception_pending);
1946
1947 // and forward the exception
1948 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1949
1950 // Slow path locking & unlocking
1951 if (method->is_synchronized()) {
1952
1953 __ block_comment("Slow path lock {");
1954 __ bind(slow_path_lock);
1955
1956 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1957 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1958
1959 // protect the args we've loaded
1960 save_args(masm, total_c_args, c_arg, out_regs);
1961
1962 __ mv(c_rarg0, obj_reg);
1963 __ mv(c_rarg1, lock_reg);
1964 __ mv(c_rarg2, xthread);
1965
1966 // Not a leaf but we have last_Java_frame setup as we want.
1967 // We don't want to unmount in case of contention since that would complicate preserving
1968 // the arguments that had already been marshalled into the native convention. So we force
1969 // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1970 // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1971 __ push_cont_fastpath();
1972 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1973 __ pop_cont_fastpath();
1974 restore_args(masm, total_c_args, c_arg, out_regs);
1975
1976 #ifdef ASSERT
1977 { Label L;
1978 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1979 __ beqz(t0, L);
1980 __ stop("no pending exception allowed on exit from monitorenter");
1981 __ bind(L);
1982 }
1983 #endif
1984 __ j(lock_done);
1985
1986 __ block_comment("} Slow path lock");
1987
1988 __ block_comment("Slow path unlock {");
1989 __ bind(slow_path_unlock);
1990
1991 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1992 save_native_result(masm, ret_type, stack_slots);
1993 }
1994
1995 __ mv(c_rarg2, xthread);
1996 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1997 __ mv(c_rarg0, obj_reg);
1998
1999 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2000 // NOTE that obj_reg == x9 currently
2001 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
2002 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
2003
2004 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
2005
2006 #ifdef ASSERT
2007 {
2008 Label L;
2009 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
2010 __ beqz(t0, L);
2011 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2012 __ bind(L);
2013 }
2014 #endif /* ASSERT */
2015
2016 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
2017
2018 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
2019 restore_native_result(masm, ret_type, stack_slots);
2020 }
2021 __ j(unlock_done);
2022
2023 __ block_comment("} Slow path unlock");
2024
2025 } // synchronized
2026
2027 // SLOW PATH Reguard the stack if needed
2028
2029 __ bind(reguard);
2030 save_native_result(masm, ret_type, stack_slots);
2031 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2032 restore_native_result(masm, ret_type, stack_slots);
2033 // and continue
2034 __ j(reguard_done);
2035
2036 // SLOW PATH safepoint
2037 {
2038 __ block_comment("safepoint {");
2039 __ bind(safepoint_in_progress);
2040
2041 // Don't use call_VM as it will see a possible pending exception and forward it
2042 // and never return here preventing us from clearing _last_native_pc down below.
2043 //
2044 save_native_result(masm, ret_type, stack_slots);
2045 __ mv(c_rarg0, xthread);
2046 #ifndef PRODUCT
2047 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2048 #endif
2049 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
2050
2051 // Restore any method result value
2052 restore_native_result(masm, ret_type, stack_slots);
2053
2054 __ j(safepoint_in_progress_done);
2055 __ block_comment("} safepoint");
2056 }
2057
2058 // SLOW PATH dtrace support
2059 if (DTraceMethodProbes) {
2060 {
2061 __ block_comment("dtrace entry {");
2062 __ bind(dtrace_method_entry);
2063
2064 // We have all of the arguments setup at this point. We must not touch any register
2065 // argument registers at this point (what if we save/restore them there are no oop?
2066
2067 save_args(masm, total_c_args, c_arg, out_regs);
2068 __ mov_metadata(c_rarg1, method());
2069 __ call_VM_leaf(
2070 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2071 xthread, c_rarg1);
2072 restore_args(masm, total_c_args, c_arg, out_regs);
2073 __ j(dtrace_method_entry_done);
2074 __ block_comment("} dtrace entry");
2075 }
2076
2077 {
2078 __ block_comment("dtrace exit {");
2079 __ bind(dtrace_method_exit);
2080 save_native_result(masm, ret_type, stack_slots);
2081 __ mov_metadata(c_rarg1, method());
2082 __ call_VM_leaf(
2083 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2084 xthread, c_rarg1);
2085 restore_native_result(masm, ret_type, stack_slots);
2086 __ j(dtrace_method_exit_done);
2087 __ block_comment("} dtrace exit");
2088 }
2089 }
2090
2091 __ flush();
2092
2093 nmethod *nm = nmethod::new_native_nmethod(method,
2094 compile_id,
2095 masm->code(),
2096 vep_offset,
2097 frame_complete,
2098 stack_slots / VMRegImpl::slots_per_word,
2099 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2100 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2101 oop_maps);
2102 assert(nm != nullptr, "create native nmethod fail!");
2103 return nm;
2104 }
2105
2106 // this function returns the adjust size (in number of words) to a c2i adapter
2107 // activation for use during deoptimization
2108 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2109 assert(callee_locals >= callee_parameters,
2110 "test and remove; got more parms than locals");
2111 if (callee_locals < callee_parameters) {
2112 return 0; // No adjustment for negative locals
2113 }
2114 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2115 // diff is counted in stack words
2116 return align_up(diff, 2);
2117 }
2118
2119 //------------------------------generate_deopt_blob----------------------------
2120 void SharedRuntime::generate_deopt_blob() {
2121 // Allocate space for the code
2122 ResourceMark rm;
2123 // Setup code generation tools
2124 int pad = 0;
2125 const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2126 CodeBuffer buffer(name, 2048 + pad, 1024);
2127 MacroAssembler* masm = new MacroAssembler(&buffer);
2128 int frame_size_in_words = -1;
2129 OopMap* map = nullptr;
2130 OopMapSet *oop_maps = new OopMapSet();
2131 assert_cond(masm != nullptr && oop_maps != nullptr);
2132 RegisterSaver reg_saver(COMPILER2_PRESENT(true) NOT_COMPILER2(false));
2133
2134 // -------------
2135 // This code enters when returning to a de-optimized nmethod. A return
2136 // address has been pushed on the stack, and return values are in
2137 // registers.
2138 // If we are doing a normal deopt then we were called from the patched
2139 // nmethod from the point we returned to the nmethod. So the return
2140 // address on the stack is wrong by NativeCall::instruction_size
2141 // We will adjust the value so it looks like we have the original return
2142 // address on the stack (like when we eagerly deoptimized).
2143 // In the case of an exception pending when deoptimizing, we enter
2144 // with a return address on the stack that points after the call we patched
2145 // into the exception handler. We have the following register state from,
2146 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2147 // x10: exception oop
2148 // x9: exception handler
2149 // x13: throwing pc
2150 // So in this case we simply jam x13 into the useless return address and
2151 // the stack looks just like we want.
2152 //
2153 // At this point we need to de-opt. We save the argument return
2154 // registers. We call the first C routine, fetch_unroll_info(). This
2155 // routine captures the return values and returns a structure which
2156 // describes the current frame size and the sizes of all replacement frames.
2157 // The current frame is compiled code and may contain many inlined
2158 // functions, each with their own JVM state. We pop the current frame, then
2159 // push all the new frames. Then we call the C routine unpack_frames() to
2160 // populate these frames. Finally unpack_frames() returns us the new target
2161 // address. Notice that callee-save registers are BLOWN here; they have
2162 // already been captured in the vframeArray at the time the return PC was
2163 // patched.
2164 address start = __ pc();
2165 Label cont;
2166
2167 // Prolog for non exception case!
2168
2169 // Save everything in sight.
2170 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2171
2172 // Normal deoptimization. Save exec mode for unpack_frames.
2173 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2174 __ j(cont);
2175
2176 int reexecute_offset = __ pc() - start;
2177 // Reexecute case
2178 // return address is the pc describes what bci to do re-execute at
2179
2180 // No need to update map as each call to save_live_registers will produce identical oopmap
2181 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2182
2183 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2184 __ j(cont);
2185
2186 int exception_offset = __ pc() - start;
2187
2188 // Prolog for exception case
2189
2190 // all registers are dead at this entry point, except for x10, and
2191 // x13 which contain the exception oop and exception pc
2192 // respectively. Set them in TLS and fall thru to the
2193 // unpack_with_exception_in_tls entry point.
2194
2195 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2196 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2197
2198 int exception_in_tls_offset = __ pc() - start;
2199
2200 // new implementation because exception oop is now passed in JavaThread
2201
2202 // Prolog for exception case
2203 // All registers must be preserved because they might be used by LinearScan
2204 // Exceptiop oop and throwing PC are passed in JavaThread
2205 // tos: stack at point of call to method that threw the exception (i.e. only
2206 // args are on the stack, no return address)
2207
2208 // The return address pushed by save_live_registers will be patched
2209 // later with the throwing pc. The correct value is not available
2210 // now because loading it from memory would destroy registers.
2211
2212 // NB: The SP at this point must be the SP of the method that is
2213 // being deoptimized. Deoptimization assumes that the frame created
2214 // here by save_live_registers is immediately below the method's SP.
2215 // This is a somewhat fragile mechanism.
2216
2217 // Save everything in sight.
2218 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2219
2220 // Now it is safe to overwrite any register
2221
2222 // Deopt during an exception. Save exec mode for unpack_frames.
2223 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2224
2225 // load throwing pc from JavaThread and patch it as the return address
2226 // of the current frame. Then clear the field in JavaThread
2227
2228 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2229 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2230 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2231
2232 #ifdef ASSERT
2233 // verify that there is really an exception oop in JavaThread
2234 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2235 __ verify_oop(x10);
2236
2237 // verify that there is no pending exception
2238 Label no_pending_exception;
2239 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2240 __ beqz(t0, no_pending_exception);
2241 __ stop("must not have pending exception here");
2242 __ bind(no_pending_exception);
2243 #endif
2244
2245 __ bind(cont);
2246
2247 // Call C code. Need thread and this frame, but NOT official VM entry
2248 // crud. We cannot block on this call, no GC can happen.
2249 //
2250 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2251
2252 // fetch_unroll_info needs to call last_java_frame().
2253
2254 Label retaddr;
2255 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2256 #ifdef ASSERT
2257 {
2258 Label L;
2259 __ ld(t0, Address(xthread,
2260 JavaThread::last_Java_fp_offset()));
2261 __ beqz(t0, L);
2262 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2263 __ bind(L);
2264 }
2265 #endif // ASSERT
2266 __ mv(c_rarg0, xthread);
2267 __ mv(c_rarg1, xcpool);
2268 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2269 __ bind(retaddr);
2270
2271 // Need to have an oopmap that tells fetch_unroll_info where to
2272 // find any register it might need.
2273 oop_maps->add_gc_map(__ pc() - start, map);
2274
2275 __ reset_last_Java_frame(false);
2276
2277 // Load UnrollBlock* into x15
2278 __ mv(x15, x10);
2279
2280 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2281 Label noException;
2282 __ mv(t0, Deoptimization::Unpack_exception);
2283 __ bne(xcpool, t0, noException); // Was exception pending?
2284 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2285 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2286 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2287 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2288
2289 __ verify_oop(x10);
2290
2291 // Overwrite the result registers with the exception results.
2292 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2293
2294 __ bind(noException);
2295
2296 // Only register save data is on the stack.
2297 // Now restore the result registers. Everything else is either dead
2298 // or captured in the vframeArray.
2299
2300 // Restore fp result register
2301 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2302 // Restore integer result register
2303 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2304
2305 // Pop all of the register save area off the stack
2306 __ add(sp, sp, frame_size_in_words * wordSize);
2307
2308 // All of the register save area has been popped of the stack. Only the
2309 // return address remains.
2310
2311 // Pop all the frames we must move/replace.
2312 //
2313 // Frame picture (youngest to oldest)
2314 // 1: self-frame (no frame link)
2315 // 2: deopting frame (no frame link)
2316 // 3: caller of deopting frame (could be compiled/interpreted).
2317 //
2318 // Note: by leaving the return address of self-frame on the stack
2319 // and using the size of frame 2 to adjust the stack
2320 // when we are done the return to frame 3 will still be on the stack.
2321
2322 // Pop deoptimized frame
2323 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2324 __ subi(x12, x12, 2 * wordSize);
2325 __ add(sp, sp, x12);
2326 __ ld(fp, Address(sp, 0));
2327 __ ld(ra, Address(sp, wordSize));
2328 __ addi(sp, sp, 2 * wordSize);
2329 // RA should now be the return address to the caller (3)
2330
2331 #ifdef ASSERT
2332 // Compilers generate code that bang the stack by as much as the
2333 // interpreter would need. So this stack banging should never
2334 // trigger a fault. Verify that it does not on non product builds.
2335 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2336 __ bang_stack_size(x9, x12);
2337 #endif
2338 // Load address of array of frame pcs into x12
2339 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2340
2341 // Load address of array of frame sizes into x14
2342 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2343
2344 // Load counter into x13
2345 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2346
2347 // Now adjust the caller's stack to make up for the extra locals
2348 // but record the original sp so that we can save it in the skeletal interpreter
2349 // frame and the stack walking of interpreter_sender will get the unextended sp
2350 // value and not the "real" sp value.
2351
2352 const Register sender_sp = x16;
2353
2354 __ mv(sender_sp, sp);
2355 __ lwu(x9, Address(x15,
2356 Deoptimization::UnrollBlock::
2357 caller_adjustment_offset()));
2358 __ sub(sp, sp, x9);
2359
2360 // Push interpreter frames in a loop
2361 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2362 __ mv(t1, t0);
2363 Label loop;
2364 __ bind(loop);
2365 __ ld(x9, Address(x14, 0)); // Load frame size
2366 __ addi(x14, x14, wordSize);
2367 __ subi(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2368 __ ld(ra, Address(x12, 0)); // Load pc
2369 __ addi(x12, x12, wordSize);
2370 __ enter(); // Save old & set new fp
2371 __ sub(sp, sp, x9); // Prolog
2372 // This value is corrected by layout_activation_impl
2373 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2374 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2375 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2376 __ subi(x13, x13, 1); // Decrement counter
2377 __ bnez(x13, loop);
2378
2379 // Re-push self-frame
2380 __ ld(ra, Address(x12));
2381 __ enter();
2382
2383 // Allocate a full sized register save area. We subtract 2 because
2384 // enter() just pushed 2 words
2385 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2386
2387 // Restore frame locals after moving the frame
2388 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2389 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2390
2391 // Call C code. Need thread but NOT official VM entry
2392 // crud. We cannot block on this call, no GC can happen. Call should
2393 // restore return values to their stack-slots with the new SP.
2394 //
2395 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2396
2397 // Use fp because the frames look interpreted now
2398 // Don't need the precise return PC here, just precise enough to point into this code blob.
2399 address the_pc = __ pc();
2400 __ set_last_Java_frame(sp, fp, the_pc, t0);
2401
2402 __ mv(c_rarg0, xthread);
2403 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2404 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2405
2406 // Set an oopmap for the call site
2407 // Use the same PC we used for the last java frame
2408 oop_maps->add_gc_map(the_pc - start,
2409 new OopMap(frame_size_in_words, 0));
2410
2411 // Clear fp AND pc
2412 __ reset_last_Java_frame(true);
2413
2414 // Collect return values
2415 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2416 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2417
2418 // Pop self-frame.
2419 __ leave(); // Epilog
2420
2421 // Jump to interpreter
2422 __ ret();
2423
2424 // Make sure all code is generated
2425 masm->flush();
2426
2427 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2428 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2429 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2430 }
2431
2432 // Number of stack slots between incoming argument block and the start of
2433 // a new frame. The PROLOG must add this many slots to the stack. The
2434 // EPILOG must remove this many slots.
2435 // RISCV needs two words for RA (return address) and FP (frame pointer).
2436 uint SharedRuntime::in_preserve_stack_slots() {
2437 return 2 * VMRegImpl::slots_per_word + (VerifyStackAtCalls ? 0 : 2) ;
2438 }
2439
2440 uint SharedRuntime::out_preserve_stack_slots() {
2441 return 0;
2442 }
2443
2444 VMReg SharedRuntime::thread_register() {
2445 return xthread->as_VMReg();
2446 }
2447
2448 //------------------------------generate_handler_blob------
2449 //
2450 // Generate a special Compile2Runtime blob that saves all registers,
2451 // and setup oopmap.
2452 //
2453 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2454 assert(is_polling_page_id(id), "expected a polling page stub id");
2455
2456 ResourceMark rm;
2457 OopMapSet *oop_maps = new OopMapSet();
2458 assert_cond(oop_maps != nullptr);
2459 OopMap* map = nullptr;
2460
2461 // Allocate space for the code. Setup code generation tools.
2462 const char* name = SharedRuntime::stub_name(id);
2463 CodeBuffer buffer(name, 2048, 1024);
2464 MacroAssembler* masm = new MacroAssembler(&buffer);
2465 assert_cond(masm != nullptr);
2466
2467 address start = __ pc();
2468 address call_pc = nullptr;
2469 int frame_size_in_words = -1;
2470 bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2471 RegisterSaver reg_saver(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2472
2473 // Save Integer and Float registers.
2474 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2475
2476 // The following is basically a call_VM. However, we need the precise
2477 // address of the call in order to generate an oopmap. Hence, we do all the
2478 // work ourselves.
2479
2480 Label retaddr;
2481 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2482
2483 // The return address must always be correct so that frame constructor never
2484 // sees an invalid pc.
2485
2486 if (!cause_return) {
2487 // overwrite the return address pushed by save_live_registers
2488 // Additionally, x18 is a callee-saved register so we can look at
2489 // it later to determine if someone changed the return address for
2490 // us!
2491 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2492 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2493 }
2494
2495 // Do the call
2496 __ mv(c_rarg0, xthread);
2497 __ rt_call(call_ptr);
2498 __ bind(retaddr);
2499
2500 // Set an oopmap for the call site. This oopmap will map all
2501 // oop-registers and debug-info registers as callee-saved. This
2502 // will allow deoptimization at this safepoint to find all possible
2503 // debug-info recordings, as well as let GC find all oops.
2504
2505 oop_maps->add_gc_map( __ pc() - start, map);
2506
2507 Label noException;
2508
2509 __ reset_last_Java_frame(false);
2510
2511 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2512
2513 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2514 __ beqz(t0, noException);
2515
2516 // Exception pending
2517
2518 reg_saver.restore_live_registers(masm);
2519
2520 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2521
2522 // No exception case
2523 __ bind(noException);
2524
2525 Label no_adjust, bail;
2526 if (!cause_return) {
2527 // If our stashed return pc was modified by the runtime we avoid touching it
2528 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2529 __ bne(x18, t0, no_adjust);
2530
2531 #ifdef ASSERT
2532 // Verify the correct encoding of the poll we're about to skip.
2533 // See NativeInstruction::is_lwu_to_zr()
2534 __ lwu(t0, Address(x18));
2535 __ andi(t1, t0, 0b1111111);
2536 __ mv(t2, 0b0000011);
2537 __ bne(t1, t2, bail); // 0-6:0b0000011
2538 __ srli(t1, t0, 7);
2539 __ andi(t1, t1, 0b11111);
2540 __ bnez(t1, bail); // 7-11:0b00000
2541 __ srli(t1, t0, 12);
2542 __ andi(t1, t1, 0b111);
2543 __ mv(t2, 0b110);
2544 __ bne(t1, t2, bail); // 12-14:0b110
2545 #endif
2546
2547 // Adjust return pc forward to step over the safepoint poll instruction
2548 __ addi(x18, x18, NativeInstruction::instruction_size);
2549 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2550 }
2551
2552 __ bind(no_adjust);
2553 // Normal exit, restore registers and exit.
2554
2555 reg_saver.restore_live_registers(masm);
2556 __ ret();
2557
2558 #ifdef ASSERT
2559 __ bind(bail);
2560 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2561 #endif
2562
2563 // Make sure all code is generated
2564 masm->flush();
2565
2566 // Fill-out other meta info
2567 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2568 }
2569
2570 //
2571 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2572 //
2573 // Generate a stub that calls into vm to find out the proper destination
2574 // of a java call. All the argument registers are live at this point
2575 // but since this is generic code we don't know what they are and the caller
2576 // must do any gc of the args.
2577 //
2578 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2579 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2580 assert(is_resolve_id(id), "expected a resolve stub id");
2581
2582 // allocate space for the code
2583 ResourceMark rm;
2584
2585 const char* name = SharedRuntime::stub_name(id);
2586 CodeBuffer buffer(name, 1000, 512);
2587 MacroAssembler* masm = new MacroAssembler(&buffer);
2588 assert_cond(masm != nullptr);
2589
2590 int frame_size_in_words = -1;
2591 RegisterSaver reg_saver(false /* save_vectors */);
2592
2593 OopMapSet *oop_maps = new OopMapSet();
2594 assert_cond(oop_maps != nullptr);
2595 OopMap* map = nullptr;
2596
2597 int start = __ offset();
2598
2599 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2600
2601 int frame_complete = __ offset();
2602
2603 {
2604 Label retaddr;
2605 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2606
2607 __ mv(c_rarg0, xthread);
2608 __ rt_call(destination);
2609 __ bind(retaddr);
2610 }
2611
2612 // Set an oopmap for the call site.
2613 // We need this not only for callee-saved registers, but also for volatile
2614 // registers that the compiler might be keeping live across a safepoint.
2615
2616 oop_maps->add_gc_map( __ offset() - start, map);
2617
2618 // x10 contains the address we are going to jump to assuming no exception got installed
2619
2620 // clear last_Java_sp
2621 __ reset_last_Java_frame(false);
2622 // check for pending exceptions
2623 Label pending;
2624 __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2625 __ bnez(t1, pending);
2626
2627 // get the returned Method*
2628 __ get_vm_result_metadata(xmethod, xthread);
2629 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2630
2631 // x10 is where we want to jump, overwrite t1 which is saved and temporary
2632 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2633 reg_saver.restore_live_registers(masm);
2634
2635 // We are back to the original state on entry and ready to go.
2636 __ jr(t1);
2637
2638 // Pending exception after the safepoint
2639
2640 __ bind(pending);
2641
2642 reg_saver.restore_live_registers(masm);
2643
2644 // exception pending => remove activation and forward to exception handler
2645
2646 __ sd(zr, Address(xthread, JavaThread::vm_result_oop_offset()));
2647
2648 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2649 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2650
2651 // -------------
2652 // make sure all code is generated
2653 masm->flush();
2654
2655 // return the blob
2656 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2657 }
2658
2659 // Continuation point for throwing of implicit exceptions that are
2660 // not handled in the current activation. Fabricates an exception
2661 // oop and initiates normal exception dispatching in this
2662 // frame. Since we need to preserve callee-saved values (currently
2663 // only for C2, but done for C1 as well) we need a callee-saved oop
2664 // map and therefore have to make these stubs into RuntimeStubs
2665 // rather than BufferBlobs. If the compiler needs all registers to
2666 // be preserved between the fault point and the exception handler
2667 // then it must assume responsibility for that in
2668 // AbstractCompiler::continuation_for_implicit_null_exception or
2669 // continuation_for_implicit_division_by_zero_exception. All other
2670 // implicit exceptions (e.g., NullPointerException or
2671 // AbstractMethodError on entry) are either at call sites or
2672 // otherwise assume that stack unwinding will be initiated, so
2673 // caller saved registers were assumed volatile in the compiler.
2674
2675 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2676 assert(is_throw_id(id), "expected a throw stub id");
2677
2678 const char* name = SharedRuntime::stub_name(id);
2679
2680 // Information about frame layout at time of blocking runtime call.
2681 // Note that we only have to preserve callee-saved registers since
2682 // the compilers are responsible for supplying a continuation point
2683 // if they expect all registers to be preserved.
2684 // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2685 assert_cond(runtime_entry != nullptr);
2686 enum layout {
2687 fp_off = 0,
2688 fp_off2,
2689 return_off,
2690 return_off2,
2691 framesize // inclusive of return address
2692 };
2693
2694 const int insts_size = 1024;
2695 const int locs_size = 64;
2696
2697 ResourceMark rm;
2698 const char* timer_msg = "SharedRuntime generate_throw_exception";
2699 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2700
2701 CodeBuffer code(name, insts_size, locs_size);
2702 OopMapSet* oop_maps = new OopMapSet();
2703 MacroAssembler* masm = new MacroAssembler(&code);
2704 assert_cond(oop_maps != nullptr && masm != nullptr);
2705
2706 address start = __ pc();
2707
2708 // This is an inlined and slightly modified version of call_VM
2709 // which has the ability to fetch the return PC out of
2710 // thread-local storage and also sets up last_Java_sp slightly
2711 // differently than the real call_VM
2712
2713 __ enter(); // Save FP and RA before call
2714
2715 assert(is_even(framesize / 2), "sp not 16-byte aligned");
2716
2717 // ra and fp are already in place
2718 __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2719
2720 int frame_complete = __ pc() - start;
2721
2722 // Set up last_Java_sp and last_Java_fp
2723 address the_pc = __ pc();
2724 __ set_last_Java_frame(sp, fp, the_pc, t0);
2725
2726 // Call runtime
2727 __ mv(c_rarg0, xthread);
2728 BLOCK_COMMENT("call runtime_entry");
2729 __ rt_call(runtime_entry);
2730
2731 // Generate oop map
2732 OopMap* map = new OopMap(framesize, 0);
2733 assert_cond(map != nullptr);
2734
2735 oop_maps->add_gc_map(the_pc - start, map);
2736
2737 __ reset_last_Java_frame(true);
2738
2739 __ leave();
2740
2741 // check for pending exceptions
2742 #ifdef ASSERT
2743 Label L;
2744 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2745 __ bnez(t0, L);
2746 __ should_not_reach_here();
2747 __ bind(L);
2748 #endif // ASSERT
2749 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2750
2751 // codeBlob framesize is in words (not VMRegImpl::slot_size)
2752 RuntimeStub* stub =
2753 RuntimeStub::new_runtime_stub(name,
2754 &code,
2755 frame_complete,
2756 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2757 oop_maps, false);
2758 assert(stub != nullptr, "create runtime stub fail!");
2759 return stub;
2760 }
2761
2762 #if INCLUDE_JFR
2763
2764 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2765 __ set_last_Java_frame(sp, fp, the_pc, t0);
2766 __ mv(c_rarg0, thread);
2767 }
2768
2769 static void jfr_epilogue(MacroAssembler* masm) {
2770 __ reset_last_Java_frame(true);
2771 }
2772 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2773 // It returns a jobject handle to the event writer.
2774 // The handle is dereferenced and the return value is the event writer oop.
2775 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2776 enum layout {
2777 fp_off,
2778 fp_off2,
2779 return_off,
2780 return_off2,
2781 framesize // inclusive of return address
2782 };
2783
2784 int insts_size = 1024;
2785 int locs_size = 64;
2786 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
2787 CodeBuffer code(name, insts_size, locs_size);
2788 OopMapSet* oop_maps = new OopMapSet();
2789 MacroAssembler* masm = new MacroAssembler(&code);
2790
2791 address start = __ pc();
2792 __ enter();
2793 int frame_complete = __ pc() - start;
2794 address the_pc = __ pc();
2795 jfr_prologue(the_pc, masm, xthread);
2796 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2797
2798 jfr_epilogue(masm);
2799 __ resolve_global_jobject(x10, t0, t1);
2800 __ leave();
2801 __ ret();
2802
2803 OopMap* map = new OopMap(framesize, 1);
2804 oop_maps->add_gc_map(the_pc - start, map);
2805
2806 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2807 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2808 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2809 oop_maps, false);
2810 return stub;
2811 }
2812
2813 // For c2: call to return a leased buffer.
2814 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2815 enum layout {
2816 fp_off,
2817 fp_off2,
2818 return_off,
2819 return_off2,
2820 framesize // inclusive of return address
2821 };
2822
2823 int insts_size = 1024;
2824 int locs_size = 64;
2825 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
2826 CodeBuffer code(name, insts_size, locs_size);
2827 OopMapSet* oop_maps = new OopMapSet();
2828 MacroAssembler* masm = new MacroAssembler(&code);
2829
2830 address start = __ pc();
2831 __ enter();
2832 int frame_complete = __ pc() - start;
2833 address the_pc = __ pc();
2834 jfr_prologue(the_pc, masm, xthread);
2835 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2836
2837 jfr_epilogue(masm);
2838 __ leave();
2839 __ ret();
2840
2841 OopMap* map = new OopMap(framesize, 1);
2842 oop_maps->add_gc_map(the_pc - start, map);
2843
2844 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2845 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2846 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2847 oop_maps, false);
2848 return stub;
2849 }
2850
2851 #endif // INCLUDE_JFR