1 /*
2 * Copyright (c) 2003, 2024, 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 "precompiled.hpp"
28 #include "asm/macroAssembler.hpp"
29 #include "asm/macroAssembler.inline.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 #if INCLUDE_JVMCI
64 #include "jvmci/jvmciJavaClasses.hpp"
65 #endif
66
67 #define __ masm->
68
69 #ifdef PRODUCT
70 #define BLOCK_COMMENT(str) /* nothing */
71 #else
72 #define BLOCK_COMMENT(str) __ block_comment(str)
73 #endif
74
75 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
76
77 class RegisterSaver {
78 const bool _save_vectors;
79 public:
80 RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
81 ~RegisterSaver() {}
82 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
83 void restore_live_registers(MacroAssembler* masm);
84
85 // Offsets into the register save area
86 // Used by deoptimization when it is managing result register
87 // values on its own
88 // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
89 // |---v0---|<---SP
90 // |---v1---|save vectors only in generate_handler_blob
91 // |-- .. --|
92 // |---v31--|-----
93 // |---f0---|
94 // |---f1---|
95 // | .. |
96 // |---f31--|
97 // |---reserved slot for stack alignment---|
98 // |---x5---|
99 // | x6 |
100 // |---.. --|
101 // |---x31--|
102 // |---fp---|
103 // |---ra---|
104 int v0_offset_in_bytes(void) { return 0; }
105 int f0_offset_in_bytes(void) {
106 int f0_offset = 0;
107 #ifdef COMPILER2
108 if (_save_vectors) {
109 f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
110 BytesPerInt;
111 }
112 #endif
113 return f0_offset;
114 }
115 int reserved_slot_offset_in_bytes(void) {
116 return f0_offset_in_bytes() +
117 FloatRegister::max_slots_per_register *
118 FloatRegister::number_of_registers *
119 BytesPerInt;
120 }
121
122 int reg_offset_in_bytes(Register r) {
123 assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
124 return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
125 }
126
127 int freg_offset_in_bytes(FloatRegister f) {
128 return f0_offset_in_bytes() + f->encoding() * wordSize;
129 }
130
131 int ra_offset_in_bytes(void) {
132 return reserved_slot_offset_in_bytes() +
133 (Register::number_of_registers - 3) *
134 Register::max_slots_per_register *
135 BytesPerInt;
136 }
137 };
138
139 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
140 int vector_size_in_bytes = 0;
141 int vector_size_in_slots = 0;
142 #ifdef COMPILER2
143 if (_save_vectors) {
144 vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
145 vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
146 }
147 #endif
148
149 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
150 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
151 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
152 // The caller will allocate additional_frame_words
153 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
154 // CodeBlob frame size is in words.
155 int frame_size_in_words = frame_size_in_bytes / wordSize;
156 *total_frame_words = frame_size_in_words;
157
158 // Save Integer, Float and Vector registers.
159 __ enter();
160 __ push_CPU_state(_save_vectors, vector_size_in_bytes);
161
162 // Set an oopmap for the call site. This oopmap will map all
163 // oop-registers and debug-info registers as callee-saved. This
164 // will allow deoptimization at this safepoint to find all possible
165 // debug-info recordings, as well as let GC find all oops.
166
167 OopMapSet *oop_maps = new OopMapSet();
168 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
169 assert_cond(oop_maps != nullptr && oop_map != nullptr);
170
171 int sp_offset_in_slots = 0;
172 int step_in_slots = 0;
173 if (_save_vectors) {
174 step_in_slots = vector_size_in_slots;
175 for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
176 VectorRegister r = as_VectorRegister(i);
177 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
178 }
179 }
180
181 step_in_slots = FloatRegister::max_slots_per_register;
182 for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
183 FloatRegister r = as_FloatRegister(i);
184 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
185 }
186
187 step_in_slots = Register::max_slots_per_register;
188 // skip the slot reserved for alignment, see MacroAssembler::push_reg;
189 // also skip x5 ~ x6 on the stack because they are caller-saved registers.
190 sp_offset_in_slots += Register::max_slots_per_register * 3;
191 // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
192 for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
193 Register r = as_Register(i);
194 if (r != xthread) {
195 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
196 }
197 }
198
199 return oop_map;
200 }
201
202 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
203 #ifdef COMPILER2
204 __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
205 #else
206 #if !INCLUDE_JVMCI
207 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
208 #endif
209 __ pop_CPU_state(_save_vectors);
210 #endif
211 __ leave();
212 }
213
214 // Is vector's size (in bytes) bigger than a size saved by default?
215 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
216 bool SharedRuntime::is_wide_vector(int size) {
217 return UseRVV;
218 }
219
220 // ---------------------------------------------------------------------------
221 // Read the array of BasicTypes from a signature, and compute where the
222 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
223 // quantities. Values less than VMRegImpl::stack0 are registers, those above
224 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
225 // as framesizes are fixed.
226 // VMRegImpl::stack0 refers to the first slot 0(sp).
227 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
228 // Register up to Register::number_of_registers) are the 64-bit
229 // integer registers.
230
231 // Note: the INPUTS in sig_bt are in units of Java argument words,
232 // which are 64-bit. The OUTPUTS are in 32-bit units.
233
234 // The Java calling convention is a "shifted" version of the C ABI.
235 // By skipping the first C ABI register we can call non-static jni
236 // methods with small numbers of arguments without having to shuffle
237 // the arguments at all. Since we control the java ABI we ought to at
238 // least get some advantage out of it.
239
240 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
241 VMRegPair *regs,
242 int total_args_passed) {
243 // Create the mapping between argument positions and
244 // registers.
245 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
246 j_rarg0, j_rarg1, j_rarg2, j_rarg3,
247 j_rarg4, j_rarg5, j_rarg6, j_rarg7
248 };
249 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
250 j_farg0, j_farg1, j_farg2, j_farg3,
251 j_farg4, j_farg5, j_farg6, j_farg7
252 };
253
254 uint int_args = 0;
255 uint fp_args = 0;
256 uint stk_args = 0;
257
258 for (int i = 0; i < total_args_passed; i++) {
259 switch (sig_bt[i]) {
260 case T_BOOLEAN: // fall through
261 case T_CHAR: // fall through
262 case T_BYTE: // fall through
263 case T_SHORT: // fall through
264 case T_INT:
265 if (int_args < Argument::n_int_register_parameters_j) {
266 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
267 } else {
268 stk_args = align_up(stk_args, 2);
269 regs[i].set1(VMRegImpl::stack2reg(stk_args));
270 stk_args += 1;
271 }
272 break;
273 case T_VOID:
274 // halves of T_LONG or T_DOUBLE
275 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
276 regs[i].set_bad();
277 break;
278 case T_LONG: // fall through
279 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
280 case T_OBJECT: // fall through
281 case T_ARRAY: // fall through
282 case T_ADDRESS:
283 if (int_args < Argument::n_int_register_parameters_j) {
284 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
285 } else {
286 stk_args = align_up(stk_args, 2);
287 regs[i].set2(VMRegImpl::stack2reg(stk_args));
288 stk_args += 2;
289 }
290 break;
291 case T_FLOAT:
292 if (fp_args < Argument::n_float_register_parameters_j) {
293 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
294 } else {
295 stk_args = align_up(stk_args, 2);
296 regs[i].set1(VMRegImpl::stack2reg(stk_args));
297 stk_args += 1;
298 }
299 break;
300 case T_DOUBLE:
301 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
302 if (fp_args < Argument::n_float_register_parameters_j) {
303 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
304 } else {
305 stk_args = align_up(stk_args, 2);
306 regs[i].set2(VMRegImpl::stack2reg(stk_args));
307 stk_args += 2;
308 }
309 break;
310 default:
311 ShouldNotReachHere();
312 }
313 }
314
315 return stk_args;
316 }
317
318 // Patch the callers callsite with entry to compiled code if it exists.
319 static void patch_callers_callsite(MacroAssembler *masm) {
320 Label L;
321 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
322 __ beqz(t0, L);
323
324 __ enter();
325 __ push_CPU_state();
326
327 // VM needs caller's callsite
328 // VM needs target method
329 // This needs to be a long call since we will relocate this adapter to
330 // the codeBuffer and it may not reach
331
332 #ifndef PRODUCT
333 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
334 #endif
335
336 __ mv(c_rarg0, xmethod);
337 __ mv(c_rarg1, ra);
338 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
339
340 __ pop_CPU_state();
341 // restore sp
342 __ leave();
343 __ bind(L);
344 }
345
346 static void gen_c2i_adapter(MacroAssembler *masm,
347 int total_args_passed,
348 int comp_args_on_stack,
349 const BasicType *sig_bt,
350 const VMRegPair *regs,
351 Label& skip_fixup) {
352 // Before we get into the guts of the C2I adapter, see if we should be here
353 // at all. We've come from compiled code and are attempting to jump to the
354 // interpreter, which means the caller made a static call to get here
355 // (vcalls always get a compiled target if there is one). Check for a
356 // compiled target. If there is one, we need to patch the caller's call.
357 patch_callers_callsite(masm);
358
359 __ bind(skip_fixup);
360
361 int words_pushed = 0;
362
363 // Since all args are passed on the stack, total_args_passed *
364 // Interpreter::stackElementSize is the space we need.
365
366 int extraspace = total_args_passed * Interpreter::stackElementSize;
367
368 __ mv(x19_sender_sp, sp);
369
370 // stack is aligned, keep it that way
371 extraspace = align_up(extraspace, 2 * wordSize);
372
373 if (extraspace) {
374 __ sub(sp, sp, extraspace);
375 }
376
377 // Now write the args into the outgoing interpreter space
378 for (int i = 0; i < total_args_passed; i++) {
379 if (sig_bt[i] == T_VOID) {
380 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
381 continue;
382 }
383
384 // offset to start parameters
385 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
386 int next_off = st_off - Interpreter::stackElementSize;
387
388 // Say 4 args:
389 // i st_off
390 // 0 32 T_LONG
391 // 1 24 T_VOID
392 // 2 16 T_OBJECT
393 // 3 8 T_BOOL
394 // - 0 return address
395 //
396 // However to make thing extra confusing. Because we can fit a Java long/double in
397 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
398 // leaves one slot empty and only stores to a single slot. In this case the
399 // slot that is occupied is the T_VOID slot. See I said it was confusing.
400
401 VMReg r_1 = regs[i].first();
402 VMReg r_2 = regs[i].second();
403 if (!r_1->is_valid()) {
404 assert(!r_2->is_valid(), "");
405 continue;
406 }
407 if (r_1->is_stack()) {
408 // memory to memory use t0
409 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
410 + extraspace
411 + words_pushed * wordSize);
412 if (!r_2->is_valid()) {
413 __ lwu(t0, Address(sp, ld_off));
414 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
415 } else {
416 __ ld(t0, Address(sp, ld_off), /*temp register*/esp);
417
418 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
419 // T_DOUBLE and T_LONG use two slots in the interpreter
420 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
421 // ld_off == LSW, ld_off+wordSize == MSW
422 // st_off == MSW, next_off == LSW
423 __ sd(t0, Address(sp, next_off), /*temp register*/esp);
424 #ifdef ASSERT
425 // Overwrite the unused slot with known junk
426 __ mv(t0, 0xdeadffffdeadaaaaul);
427 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
428 #endif /* ASSERT */
429 } else {
430 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
431 }
432 }
433 } else if (r_1->is_Register()) {
434 Register r = r_1->as_Register();
435 if (!r_2->is_valid()) {
436 // must be only an int (or less ) so move only 32bits to slot
437 __ sd(r, Address(sp, st_off));
438 } else {
439 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
440 // T_DOUBLE and T_LONG use two slots in the interpreter
441 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
442 // long/double in gpr
443 #ifdef ASSERT
444 // Overwrite the unused slot with known junk
445 __ mv(t0, 0xdeadffffdeadaaabul);
446 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
447 #endif /* ASSERT */
448 __ sd(r, Address(sp, next_off));
449 } else {
450 __ sd(r, Address(sp, st_off));
451 }
452 }
453 } else {
454 assert(r_1->is_FloatRegister(), "");
455 if (!r_2->is_valid()) {
456 // only a float use just part of the slot
457 __ fsw(r_1->as_FloatRegister(), Address(sp, st_off));
458 } else {
459 #ifdef ASSERT
460 // Overwrite the unused slot with known junk
461 __ mv(t0, 0xdeadffffdeadaaacul);
462 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
463 #endif /* ASSERT */
464 __ fsd(r_1->as_FloatRegister(), Address(sp, next_off));
465 }
466 }
467 }
468
469 __ mv(esp, sp); // Interp expects args on caller's expression stack
470
471 __ ld(t0, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
472 __ jr(t0);
473 }
474
475 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
476 int total_args_passed,
477 int comp_args_on_stack,
478 const BasicType *sig_bt,
479 const VMRegPair *regs) {
480 // Note: x19_sender_sp contains the senderSP on entry. We must
481 // preserve it since we may do a i2c -> c2i transition if we lose a
482 // race where compiled code goes non-entrant while we get args
483 // ready.
484
485 // Cut-out for having no stack args.
486 int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
487 if (comp_args_on_stack != 0) {
488 __ sub(t0, sp, comp_words_on_stack * wordSize);
489 __ andi(sp, t0, -16);
490 }
491
492 // Will jump to the compiled code just as if compiled code was doing it.
493 // Pre-load the register-jump target early, to schedule it better.
494 __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset())));
495
496 #if INCLUDE_JVMCI
497 if (EnableJVMCI) {
498 // check if this call should be routed towards a specific entry point
499 __ ld(t0, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
500 Label no_alternative_target;
501 __ beqz(t0, no_alternative_target);
502 __ mv(t1, t0);
503 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
504 __ bind(no_alternative_target);
505 }
506 #endif // INCLUDE_JVMCI
507
508 // Now generate the shuffle code.
509 for (int i = 0; i < total_args_passed; i++) {
510 if (sig_bt[i] == T_VOID) {
511 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
512 continue;
513 }
514
515 // Pick up 0, 1 or 2 words from SP+offset.
516
517 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
518 "scrambled load targets?");
519 // Load in argument order going down.
520 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
521 // Point to interpreter value (vs. tag)
522 int next_off = ld_off - Interpreter::stackElementSize;
523
524 VMReg r_1 = regs[i].first();
525 VMReg r_2 = regs[i].second();
526 if (!r_1->is_valid()) {
527 assert(!r_2->is_valid(), "");
528 continue;
529 }
530 if (r_1->is_stack()) {
531 // Convert stack slot to an SP offset (+ wordSize to account for return address )
532 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
533 if (!r_2->is_valid()) {
534 __ lw(t0, Address(esp, ld_off));
535 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
536 } else {
537 //
538 // We are using two optoregs. This can be either T_OBJECT,
539 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
540 // two slots but only uses one for thr T_LONG or T_DOUBLE case
541 // So we must adjust where to pick up the data to match the
542 // interpreter.
543 //
544 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
545 // are accessed as negative so LSW is at LOW address
546
547 // ld_off is MSW so get LSW
548 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
549 next_off : ld_off;
550 __ ld(t0, Address(esp, offset));
551 // st_off is LSW (i.e. reg.first())
552 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
553 }
554 } else if (r_1->is_Register()) { // Register argument
555 Register r = r_1->as_Register();
556 if (r_2->is_valid()) {
557 //
558 // We are using two VMRegs. This can be either T_OBJECT,
559 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
560 // two slots but only uses one for thr T_LONG or T_DOUBLE case
561 // So we must adjust where to pick up the data to match the
562 // interpreter.
563
564 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
565 next_off : ld_off;
566
567 // this can be a misaligned move
568 __ ld(r, Address(esp, offset));
569 } else {
570 // sign extend and use a full word?
571 __ lw(r, Address(esp, ld_off));
572 }
573 } else {
574 if (!r_2->is_valid()) {
575 __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
576 } else {
577 __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
578 }
579 }
580 }
581
582 __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
583
584 // 6243940 We might end up in handle_wrong_method if
585 // the callee is deoptimized as we race thru here. If that
586 // happens we don't want to take a safepoint because the
587 // caller frame will look interpreted and arguments are now
588 // "compiled" so it is much better to make this transition
589 // invisible to the stack walking code. Unfortunately if
590 // we try and find the callee by normal means a safepoint
591 // is possible. So we stash the desired callee in the thread
592 // and the vm will find there should this case occur.
593
594 __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
595
596 __ jr(t1);
597 }
598
599 // ---------------------------------------------------------------
600 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
601 int total_args_passed,
602 int comp_args_on_stack,
603 const BasicType *sig_bt,
604 const VMRegPair *regs,
605 AdapterFingerPrint* fingerprint) {
606 address i2c_entry = __ pc();
607 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
608
609 address c2i_unverified_entry = __ pc();
610 Label skip_fixup;
611
612 const Register receiver = j_rarg0;
613 const Register data = t1;
614 const Register tmp = t2; // A call-clobbered register not used for arg passing
615
616 // -------------------------------------------------------------------------
617 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls
618 // to the interpreter. The args start out packed in the compiled layout. They
619 // need to be unpacked into the interpreter layout. This will almost always
620 // require some stack space. We grow the current (compiled) stack, then repack
621 // the args. We finally end in a jump to the generic interpreter entry point.
622 // On exit from the interpreter, the interpreter will restore our SP (lest the
623 // compiled code, which relies solely on SP and not FP, get sick).
624
625 {
626 __ block_comment("c2i_unverified_entry {");
627
628 __ ic_check();
629 __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
630
631 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
632 __ beqz(t0, skip_fixup);
633 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
634 __ block_comment("} c2i_unverified_entry");
635 }
636
637 address c2i_entry = __ pc();
638
639 // Class initialization barrier for static methods
640 address c2i_no_clinit_check_entry = nullptr;
641 if (VM_Version::supports_fast_class_init_checks()) {
642 Label L_skip_barrier;
643
644 { // Bypass the barrier for non-static methods
645 __ lwu(t0, Address(xmethod, Method::access_flags_offset()));
646 __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
647 __ beqz(t1, L_skip_barrier); // non-static
648 }
649
650 __ load_method_holder(t1, xmethod);
651 __ clinit_barrier(t1, t0, &L_skip_barrier);
652 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
653
654 __ bind(L_skip_barrier);
655 c2i_no_clinit_check_entry = __ pc();
656 }
657
658 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
659 bs->c2i_entry_barrier(masm);
660
661 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
662
663 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
664 }
665
666 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
667 uint num_bits,
668 uint total_args_passed) {
669 assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
670 assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
671
672 // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
673 static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
674 v8, v9, v10, v11, v12, v13, v14, v15,
675 v16, v17, v18, v19, v20, v21, v22, v23
676 };
677
678 const int next_reg_val = 3;
679 for (uint i = 0; i < total_args_passed; i++) {
680 VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
681 regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
682 }
683 return 0;
684 }
685
686 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
687 VMRegPair *regs,
688 int total_args_passed) {
689
690 // We return the amount of VMRegImpl stack slots we need to reserve for all
691 // the arguments NOT counting out_preserve_stack_slots.
692
693 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
694 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
695 c_rarg4, c_rarg5, c_rarg6, c_rarg7
696 };
697 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
698 c_farg0, c_farg1, c_farg2, c_farg3,
699 c_farg4, c_farg5, c_farg6, c_farg7
700 };
701
702 uint int_args = 0;
703 uint fp_args = 0;
704 uint stk_args = 0; // inc by 2 each time
705
706 for (int i = 0; i < total_args_passed; i++) {
707 switch (sig_bt[i]) {
708 case T_BOOLEAN: // fall through
709 case T_CHAR: // fall through
710 case T_BYTE: // fall through
711 case T_SHORT: // fall through
712 case T_INT:
713 if (int_args < Argument::n_int_register_parameters_c) {
714 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
715 } else {
716 regs[i].set1(VMRegImpl::stack2reg(stk_args));
717 stk_args += 2;
718 }
719 break;
720 case T_LONG: // fall through
721 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
722 case T_OBJECT: // fall through
723 case T_ARRAY: // fall through
724 case T_ADDRESS: // fall through
725 case T_METADATA:
726 if (int_args < Argument::n_int_register_parameters_c) {
727 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
728 } else {
729 regs[i].set2(VMRegImpl::stack2reg(stk_args));
730 stk_args += 2;
731 }
732 break;
733 case T_FLOAT:
734 if (fp_args < Argument::n_float_register_parameters_c) {
735 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
736 } else if (int_args < Argument::n_int_register_parameters_c) {
737 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
738 } else {
739 regs[i].set1(VMRegImpl::stack2reg(stk_args));
740 stk_args += 2;
741 }
742 break;
743 case T_DOUBLE:
744 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
745 if (fp_args < Argument::n_float_register_parameters_c) {
746 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
747 } else if (int_args < Argument::n_int_register_parameters_c) {
748 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
749 } else {
750 regs[i].set2(VMRegImpl::stack2reg(stk_args));
751 stk_args += 2;
752 }
753 break;
754 case T_VOID: // Halves of longs and doubles
755 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
756 regs[i].set_bad();
757 break;
758 default:
759 ShouldNotReachHere();
760 }
761 }
762
763 return stk_args;
764 }
765
766 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
767 // We always ignore the frame_slots arg and just use the space just below frame pointer
768 // which by this time is free to use
769 switch (ret_type) {
770 case T_FLOAT:
771 __ fsw(f10, Address(fp, -3 * wordSize));
772 break;
773 case T_DOUBLE:
774 __ fsd(f10, Address(fp, -3 * wordSize));
775 break;
776 case T_VOID: break;
777 default: {
778 __ sd(x10, Address(fp, -3 * wordSize));
779 }
780 }
781 }
782
783 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
784 // We always ignore the frame_slots arg and just use the space just below frame pointer
785 // which by this time is free to use
786 switch (ret_type) {
787 case T_FLOAT:
788 __ flw(f10, Address(fp, -3 * wordSize));
789 break;
790 case T_DOUBLE:
791 __ fld(f10, Address(fp, -3 * wordSize));
792 break;
793 case T_VOID: break;
794 default: {
795 __ ld(x10, Address(fp, -3 * wordSize));
796 }
797 }
798 }
799
800 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
801 RegSet x;
802 for ( int i = first_arg ; i < arg_count ; i++ ) {
803 if (args[i].first()->is_Register()) {
804 x = x + args[i].first()->as_Register();
805 } else if (args[i].first()->is_FloatRegister()) {
806 __ addi(sp, sp, -2 * wordSize);
807 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
808 }
809 }
810 __ push_reg(x, sp);
811 }
812
813 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
814 RegSet x;
815 for ( int i = first_arg ; i < arg_count ; i++ ) {
816 if (args[i].first()->is_Register()) {
817 x = x + args[i].first()->as_Register();
818 } else {
819 ;
820 }
821 }
822 __ pop_reg(x, sp);
823 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
824 if (args[i].first()->is_Register()) {
825 ;
826 } else if (args[i].first()->is_FloatRegister()) {
827 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
828 __ add(sp, sp, 2 * wordSize);
829 }
830 }
831 }
832
833 static void verify_oop_args(MacroAssembler* masm,
834 const methodHandle& method,
835 const BasicType* sig_bt,
836 const VMRegPair* regs) {
837 const Register temp_reg = x9; // not part of any compiled calling seq
838 if (VerifyOops) {
839 for (int i = 0; i < method->size_of_parameters(); i++) {
840 if (sig_bt[i] == T_OBJECT ||
841 sig_bt[i] == T_ARRAY) {
842 VMReg r = regs[i].first();
843 assert(r->is_valid(), "bad oop arg");
844 if (r->is_stack()) {
845 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
846 __ verify_oop(temp_reg);
847 } else {
848 __ verify_oop(r->as_Register());
849 }
850 }
851 }
852 }
853 }
854
855 // on exit, sp points to the ContinuationEntry
856 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
857 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
858 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
859 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
860
861 stack_slots += (int)ContinuationEntry::size() / wordSize;
862 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
863
864 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
865
866 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
867 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
868 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
869
870 return map;
871 }
872
873 // on entry c_rarg1 points to the continuation
874 // sp points to ContinuationEntry
875 // c_rarg3 -- isVirtualThread
876 static void fill_continuation_entry(MacroAssembler* masm) {
877 #ifdef ASSERT
878 __ mv(t0, ContinuationEntry::cookie_value());
879 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
880 #endif
881
882 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
883 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
884 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
885 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
886 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
887
888 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
889 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
890 __ ld(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
891 __ sd(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
892
893 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
894 __ sd(zr, Address(xthread, JavaThread::held_monitor_count_offset()));
895 }
896
897 // on entry, sp points to the ContinuationEntry
898 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
899 static void continuation_enter_cleanup(MacroAssembler* masm) {
900 #ifndef PRODUCT
901 Label OK;
902 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
903 __ beq(sp, t0, OK);
904 __ stop("incorrect sp");
905 __ bind(OK);
906 #endif
907
908 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
909 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
910
911 if (CheckJNICalls) {
912 // Check if this is a virtual thread continuation
913 Label L_skip_vthread_code;
914 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
915 __ beqz(t0, L_skip_vthread_code);
916
917 // If the held monitor count is > 0 and this vthread is terminating then
918 // it failed to release a JNI monitor. So we issue the same log message
919 // that JavaThread::exit does.
920 __ ld(t0, Address(xthread, JavaThread::jni_monitor_count_offset()));
921 __ beqz(t0, L_skip_vthread_code);
922
923 // Save return value potentially containing the exception oop in callee-saved x9
924 __ mv(x9, x10);
925 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
926 // Restore potential return value
927 __ mv(x10, x9);
928
929 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
930 // on termination. The held count is implicitly zeroed below when we restore from
931 // the parent held count (which has to be zero).
932 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
933
934 __ bind(L_skip_vthread_code);
935 }
936 #ifdef ASSERT
937 else {
938 // Check if this is a virtual thread continuation
939 Label L_skip_vthread_code;
940 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
941 __ beqz(t0, L_skip_vthread_code);
942
943 // See comment just above. If not checking JNI calls the JNI count is only
944 // needed for assertion checking.
945 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
946
947 __ bind(L_skip_vthread_code);
948 }
949 #endif
950
951 __ ld(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
952 __ sd(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
953
954 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
955 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
956 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
957 }
958
959 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
960 // On entry: c_rarg1 -- the continuation object
961 // c_rarg2 -- isContinue
962 // c_rarg3 -- isVirtualThread
963 static void gen_continuation_enter(MacroAssembler* masm,
964 const methodHandle& method,
965 const BasicType* sig_bt,
966 const VMRegPair* regs,
967 int& exception_offset,
968 OopMapSet*oop_maps,
969 int& frame_complete,
970 int& stack_slots,
971 int& interpreted_entry_offset,
972 int& compiled_entry_offset) {
973 // verify_oop_args(masm, method, sig_bt, regs);
974 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
975
976 address start = __ pc();
977
978 Label call_thaw, exit;
979
980 // i2i entry used at interp_only_mode only
981 interpreted_entry_offset = __ pc() - start;
982 {
983 #ifdef ASSERT
984 Label is_interp_only;
985 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
986 __ bnez(t0, is_interp_only);
987 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
988 __ bind(is_interp_only);
989 #endif
990
991 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
992 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
993 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
994 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
995 __ push_cont_fastpath(xthread);
996
997 __ enter();
998 stack_slots = 2; // will be adjusted in setup
999 OopMap* map = continuation_enter_setup(masm, stack_slots);
1000 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1001 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1002
1003 fill_continuation_entry(masm);
1004
1005 __ bnez(c_rarg2, call_thaw);
1006
1007 // Make sure the call is patchable
1008 __ align(NativeInstruction::instruction_size);
1009
1010 const address tr_call = __ reloc_call(resolve);
1011 if (tr_call == nullptr) {
1012 fatal("CodeCache is full at gen_continuation_enter");
1013 }
1014
1015 oop_maps->add_gc_map(__ pc() - start, map);
1016 __ post_call_nop();
1017
1018 __ j(exit);
1019
1020 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1021 if (stub == nullptr) {
1022 fatal("CodeCache is full at gen_continuation_enter");
1023 }
1024 }
1025
1026 // compiled entry
1027 __ align(CodeEntryAlignment);
1028 compiled_entry_offset = __ pc() - start;
1029
1030 __ enter();
1031 stack_slots = 2; // will be adjusted in setup
1032 OopMap* map = continuation_enter_setup(masm, stack_slots);
1033 frame_complete = __ pc() - start;
1034
1035 fill_continuation_entry(masm);
1036
1037 __ bnez(c_rarg2, call_thaw);
1038
1039 // Make sure the call is patchable
1040 __ align(NativeInstruction::instruction_size);
1041
1042 const address tr_call = __ reloc_call(resolve);
1043 if (tr_call == nullptr) {
1044 fatal("CodeCache is full at gen_continuation_enter");
1045 }
1046
1047 oop_maps->add_gc_map(__ pc() - start, map);
1048 __ post_call_nop();
1049
1050 __ j(exit);
1051
1052 __ bind(call_thaw);
1053
1054 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1055 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1056 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1057 ContinuationEntry::_return_pc_offset = __ pc() - start;
1058 __ post_call_nop();
1059
1060 __ bind(exit);
1061 ContinuationEntry::_cleanup_offset = __ pc() - start;
1062 continuation_enter_cleanup(masm);
1063 __ leave();
1064 __ ret();
1065
1066 // exception handling
1067 exception_offset = __ pc() - start;
1068 {
1069 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1070
1071 continuation_enter_cleanup(masm);
1072
1073 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1074 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1075
1076 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1077
1078 __ mv(x11, x10); // the exception handler
1079 __ mv(x10, x9); // restore return value contaning the exception oop
1080 __ verify_oop(x10);
1081
1082 __ leave();
1083 __ mv(x13, ra);
1084 __ jr(x11); // the exception handler
1085 }
1086
1087 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1088 if (stub == nullptr) {
1089 fatal("CodeCache is full at gen_continuation_enter");
1090 }
1091 }
1092
1093 static void gen_continuation_yield(MacroAssembler* masm,
1094 const methodHandle& method,
1095 const BasicType* sig_bt,
1096 const VMRegPair* regs,
1097 OopMapSet* oop_maps,
1098 int& frame_complete,
1099 int& stack_slots,
1100 int& compiled_entry_offset) {
1101 enum layout {
1102 fp_off,
1103 fp_off2,
1104 return_off,
1105 return_off2,
1106 framesize // inclusive of return address
1107 };
1108 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1109
1110 stack_slots = framesize / VMRegImpl::slots_per_word;
1111 assert(stack_slots == 2, "recheck layout");
1112
1113 address start = __ pc();
1114
1115 compiled_entry_offset = __ pc() - start;
1116 __ enter();
1117
1118 __ mv(c_rarg1, sp);
1119
1120 frame_complete = __ pc() - start;
1121 address the_pc = __ pc();
1122
1123 __ 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
1124
1125 __ mv(c_rarg0, xthread);
1126 __ set_last_Java_frame(sp, fp, the_pc, t0);
1127 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1128 __ reset_last_Java_frame(true);
1129
1130 Label pinned;
1131
1132 __ bnez(x10, pinned);
1133
1134 // We've succeeded, set sp to the ContinuationEntry
1135 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1136 continuation_enter_cleanup(masm);
1137
1138 __ bind(pinned); // pinned -- return to caller
1139
1140 // handle pending exception thrown by freeze
1141 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1142 Label ok;
1143 __ beqz(t0, ok);
1144 __ leave();
1145 __ la(t0, RuntimeAddress(StubRoutines::forward_exception_entry()));
1146 __ jr(t0);
1147 __ bind(ok);
1148
1149 __ leave();
1150 __ ret();
1151
1152 OopMap* map = new OopMap(framesize, 1);
1153 oop_maps->add_gc_map(the_pc - start, map);
1154 }
1155
1156 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1157 ::continuation_enter_cleanup(masm);
1158 }
1159
1160 static void gen_special_dispatch(MacroAssembler* masm,
1161 const methodHandle& method,
1162 const BasicType* sig_bt,
1163 const VMRegPair* regs) {
1164 verify_oop_args(masm, method, sig_bt, regs);
1165 vmIntrinsics::ID iid = method->intrinsic_id();
1166
1167 // Now write the args into the outgoing interpreter space
1168 bool has_receiver = false;
1169 Register receiver_reg = noreg;
1170 int member_arg_pos = -1;
1171 Register member_reg = noreg;
1172 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1173 if (ref_kind != 0) {
1174 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1175 member_reg = x9; // known to be free at this point
1176 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1177 } else if (iid == vmIntrinsics::_invokeBasic) {
1178 has_receiver = true;
1179 } else if (iid == vmIntrinsics::_linkToNative) {
1180 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1181 member_reg = x9; // known to be free at this point
1182 } else {
1183 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1184 }
1185
1186 if (member_reg != noreg) {
1187 // Load the member_arg into register, if necessary.
1188 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1189 VMReg r = regs[member_arg_pos].first();
1190 if (r->is_stack()) {
1191 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1192 } else {
1193 // no data motion is needed
1194 member_reg = r->as_Register();
1195 }
1196 }
1197
1198 if (has_receiver) {
1199 // Make sure the receiver is loaded into a register.
1200 assert(method->size_of_parameters() > 0, "oob");
1201 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1202 VMReg r = regs[0].first();
1203 assert(r->is_valid(), "bad receiver arg");
1204 if (r->is_stack()) {
1205 // Porting note: This assumes that compiled calling conventions always
1206 // pass the receiver oop in a register. If this is not true on some
1207 // platform, pick a temp and load the receiver from stack.
1208 fatal("receiver always in a register");
1209 receiver_reg = x12; // known to be free at this point
1210 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1211 } else {
1212 // no data motion is needed
1213 receiver_reg = r->as_Register();
1214 }
1215 }
1216
1217 // Figure out which address we are really jumping to:
1218 MethodHandles::generate_method_handle_dispatch(masm, iid,
1219 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1220 }
1221
1222 // ---------------------------------------------------------------------------
1223 // Generate a native wrapper for a given method. The method takes arguments
1224 // in the Java compiled code convention, marshals them to the native
1225 // convention (handlizes oops, etc), transitions to native, makes the call,
1226 // returns to java state (possibly blocking), unhandlizes any result and
1227 // returns.
1228 //
1229 // Critical native functions are a shorthand for the use of
1230 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1231 // functions. The wrapper is expected to unpack the arguments before
1232 // passing them to the callee and perform checks before and after the
1233 // native call to ensure that they GCLocker
1234 // lock_critical/unlock_critical semantics are followed. Some other
1235 // parts of JNI setup are skipped like the tear down of the JNI handle
1236 // block and the check for pending exceptions it's impossible for them
1237 // to be thrown.
1238 //
1239 // They are roughly structured like this:
1240 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1241 // tranistion to thread_in_native
1242 // unpack array arguments and call native entry point
1243 // check for safepoint in progress
1244 // check if any thread suspend flags are set
1245 // call into JVM and possible unlock the JNI critical
1246 // if a GC was suppressed while in the critical native.
1247 // transition back to thread_in_Java
1248 // return to caller
1249 //
1250 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1251 const methodHandle& method,
1252 int compile_id,
1253 BasicType* in_sig_bt,
1254 VMRegPair* in_regs,
1255 BasicType ret_type) {
1256 if (method->is_continuation_native_intrinsic()) {
1257 int exception_offset = -1;
1258 OopMapSet* oop_maps = new OopMapSet();
1259 int frame_complete = -1;
1260 int stack_slots = -1;
1261 int interpreted_entry_offset = -1;
1262 int vep_offset = -1;
1263 if (method->is_continuation_enter_intrinsic()) {
1264 gen_continuation_enter(masm,
1265 method,
1266 in_sig_bt,
1267 in_regs,
1268 exception_offset,
1269 oop_maps,
1270 frame_complete,
1271 stack_slots,
1272 interpreted_entry_offset,
1273 vep_offset);
1274 } else if (method->is_continuation_yield_intrinsic()) {
1275 gen_continuation_yield(masm,
1276 method,
1277 in_sig_bt,
1278 in_regs,
1279 oop_maps,
1280 frame_complete,
1281 stack_slots,
1282 vep_offset);
1283 } else {
1284 guarantee(false, "Unknown Continuation native intrinsic");
1285 }
1286
1287 #ifdef ASSERT
1288 if (method->is_continuation_enter_intrinsic()) {
1289 assert(interpreted_entry_offset != -1, "Must be set");
1290 assert(exception_offset != -1, "Must be set");
1291 } else {
1292 assert(interpreted_entry_offset == -1, "Must be unset");
1293 assert(exception_offset == -1, "Must be unset");
1294 }
1295 assert(frame_complete != -1, "Must be set");
1296 assert(stack_slots != -1, "Must be set");
1297 assert(vep_offset != -1, "Must be set");
1298 #endif
1299
1300 __ flush();
1301 nmethod* nm = nmethod::new_native_nmethod(method,
1302 compile_id,
1303 masm->code(),
1304 vep_offset,
1305 frame_complete,
1306 stack_slots,
1307 in_ByteSize(-1),
1308 in_ByteSize(-1),
1309 oop_maps,
1310 exception_offset);
1311 if (nm == nullptr) return nm;
1312 if (method->is_continuation_enter_intrinsic()) {
1313 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1314 } else if (method->is_continuation_yield_intrinsic()) {
1315 _cont_doYield_stub = nm;
1316 } else {
1317 guarantee(false, "Unknown Continuation native intrinsic");
1318 }
1319 return nm;
1320 }
1321
1322 if (method->is_method_handle_intrinsic()) {
1323 vmIntrinsics::ID iid = method->intrinsic_id();
1324 intptr_t start = (intptr_t)__ pc();
1325 int vep_offset = ((intptr_t)__ pc()) - start;
1326
1327 // First instruction must be a nop as it may need to be patched on deoptimisation
1328 {
1329 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching.
1330 MacroAssembler::assert_alignment(__ pc());
1331 __ nop(); // 4 bytes
1332 }
1333 gen_special_dispatch(masm,
1334 method,
1335 in_sig_bt,
1336 in_regs);
1337 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1338 __ flush();
1339 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1340 return nmethod::new_native_nmethod(method,
1341 compile_id,
1342 masm->code(),
1343 vep_offset,
1344 frame_complete,
1345 stack_slots / VMRegImpl::slots_per_word,
1346 in_ByteSize(-1),
1347 in_ByteSize(-1),
1348 (OopMapSet*)nullptr);
1349 }
1350 address native_func = method->native_function();
1351 assert(native_func != nullptr, "must have function");
1352
1353 // An OopMap for lock (and class if static)
1354 OopMapSet *oop_maps = new OopMapSet();
1355 assert_cond(oop_maps != nullptr);
1356 intptr_t start = (intptr_t)__ pc();
1357
1358 // We have received a description of where all the java arg are located
1359 // on entry to the wrapper. We need to convert these args to where
1360 // the jni function will expect them. To figure out where they go
1361 // we convert the java signature to a C signature by inserting
1362 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1363
1364 const int total_in_args = method->size_of_parameters();
1365 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1366
1367 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1368 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1369 BasicType* in_elem_bt = nullptr;
1370
1371 int argc = 0;
1372 out_sig_bt[argc++] = T_ADDRESS;
1373 if (method->is_static()) {
1374 out_sig_bt[argc++] = T_OBJECT;
1375 }
1376
1377 for (int i = 0; i < total_in_args ; i++) {
1378 out_sig_bt[argc++] = in_sig_bt[i];
1379 }
1380
1381 // Now figure out where the args must be stored and how much stack space
1382 // they require.
1383 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1384
1385 // Compute framesize for the wrapper. We need to handlize all oops in
1386 // incoming registers
1387
1388 // Calculate the total number of stack slots we will need.
1389
1390 // First count the abi requirement plus all of the outgoing args
1391 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1392
1393 // Now the space for the inbound oop handle area
1394 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1395
1396 int oop_handle_offset = stack_slots;
1397 stack_slots += total_save_slots;
1398
1399 // Now any space we need for handlizing a klass if static method
1400
1401 int klass_slot_offset = 0;
1402 int klass_offset = -1;
1403 int lock_slot_offset = 0;
1404 bool is_static = false;
1405
1406 if (method->is_static()) {
1407 klass_slot_offset = stack_slots;
1408 stack_slots += VMRegImpl::slots_per_word;
1409 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1410 is_static = true;
1411 }
1412
1413 // Plus a lock if needed
1414
1415 if (method->is_synchronized()) {
1416 lock_slot_offset = stack_slots;
1417 stack_slots += VMRegImpl::slots_per_word;
1418 }
1419
1420 // Now a place (+2) to save return values or temp during shuffling
1421 // + 4 for return address (which we own) and saved fp
1422 stack_slots += 6;
1423
1424 // Ok The space we have allocated will look like:
1425 //
1426 //
1427 // FP-> | |
1428 // | 2 slots (ra) |
1429 // | 2 slots (fp) |
1430 // |---------------------|
1431 // | 2 slots for moves |
1432 // |---------------------|
1433 // | lock box (if sync) |
1434 // |---------------------| <- lock_slot_offset
1435 // | klass (if static) |
1436 // |---------------------| <- klass_slot_offset
1437 // | oopHandle area |
1438 // |---------------------| <- oop_handle_offset (8 java arg registers)
1439 // | outbound memory |
1440 // | based arguments |
1441 // | |
1442 // |---------------------|
1443 // | |
1444 // SP-> | out_preserved_slots |
1445 //
1446 //
1447
1448
1449 // Now compute actual number of stack words we need rounding to make
1450 // stack properly aligned.
1451 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1452
1453 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1454
1455 // First thing make an ic check to see if we should even be here
1456
1457 // We are free to use all registers as temps without saving them and
1458 // restoring them except fp. fp is the only callee save register
1459 // as far as the interpreter and the compiler(s) are concerned.
1460
1461
1462 const Register ic_reg = t1;
1463 const Register receiver = j_rarg0;
1464
1465 __ verify_oop(receiver);
1466 assert_different_registers(receiver, t0, t1);
1467
1468 __ ic_check();
1469
1470 int vep_offset = ((intptr_t)__ pc()) - start;
1471
1472 // If we have to make this method not-entrant we'll overwrite its
1473 // first instruction with a jump.
1474 {
1475 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching.
1476 MacroAssembler::assert_alignment(__ pc());
1477 __ nop(); // 4 bytes
1478 }
1479
1480 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1481 Label L_skip_barrier;
1482 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1483 __ clinit_barrier(t1, t0, &L_skip_barrier);
1484 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1485
1486 __ bind(L_skip_barrier);
1487 }
1488
1489 // Generate stack overflow check
1490 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1491
1492 // Generate a new frame for the wrapper.
1493 __ enter();
1494 // -2 because return address is already present and so is saved fp
1495 __ sub(sp, sp, stack_size - 2 * wordSize);
1496
1497 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1498 assert_cond(bs != nullptr);
1499 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1500
1501 // Frame is now completed as far as size and linkage.
1502 int frame_complete = ((intptr_t)__ pc()) - start;
1503
1504 // We use x18 as the oop handle for the receiver/klass
1505 // It is callee save so it survives the call to native
1506
1507 const Register oop_handle_reg = x18;
1508
1509 //
1510 // We immediately shuffle the arguments so that any vm call we have to
1511 // make from here on out (sync slow path, jvmti, etc.) we will have
1512 // captured the oops from our caller and have a valid oopMap for
1513 // them.
1514
1515 // -----------------
1516 // The Grand Shuffle
1517
1518 // The Java calling convention is either equal (linux) or denser (win64) than the
1519 // c calling convention. However the because of the jni_env argument the c calling
1520 // convention always has at least one more (and two for static) arguments than Java.
1521 // Therefore if we move the args from java -> c backwards then we will never have
1522 // a register->register conflict and we don't have to build a dependency graph
1523 // and figure out how to break any cycles.
1524 //
1525
1526 // Record esp-based slot for receiver on stack for non-static methods
1527 int receiver_offset = -1;
1528
1529 // This is a trick. We double the stack slots so we can claim
1530 // the oops in the caller's frame. Since we are sure to have
1531 // more args than the caller doubling is enough to make
1532 // sure we can capture all the incoming oop args from the
1533 // caller.
1534 //
1535 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1536 assert_cond(map != nullptr);
1537
1538 int float_args = 0;
1539 int int_args = 0;
1540
1541 #ifdef ASSERT
1542 bool reg_destroyed[Register::number_of_registers];
1543 bool freg_destroyed[FloatRegister::number_of_registers];
1544 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1545 reg_destroyed[r] = false;
1546 }
1547 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1548 freg_destroyed[f] = false;
1549 }
1550
1551 #endif /* ASSERT */
1552
1553 // For JNI natives the incoming and outgoing registers are offset upwards.
1554 GrowableArray<int> arg_order(2 * total_in_args);
1555 VMRegPair tmp_vmreg;
1556 tmp_vmreg.set2(x9->as_VMReg());
1557
1558 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1559 arg_order.push(i);
1560 arg_order.push(c_arg);
1561 }
1562
1563 int temploc = -1;
1564 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1565 int i = arg_order.at(ai);
1566 int c_arg = arg_order.at(ai + 1);
1567 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1568 assert(c_arg != -1 && i != -1, "wrong order");
1569 #ifdef ASSERT
1570 if (in_regs[i].first()->is_Register()) {
1571 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1572 } else if (in_regs[i].first()->is_FloatRegister()) {
1573 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1574 }
1575 if (out_regs[c_arg].first()->is_Register()) {
1576 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1577 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1578 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1579 }
1580 #endif /* ASSERT */
1581 switch (in_sig_bt[i]) {
1582 case T_ARRAY:
1583 case T_OBJECT:
1584 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1585 ((i == 0) && (!is_static)),
1586 &receiver_offset);
1587 int_args++;
1588 break;
1589 case T_VOID:
1590 break;
1591
1592 case T_FLOAT:
1593 __ float_move(in_regs[i], out_regs[c_arg]);
1594 float_args++;
1595 break;
1596
1597 case T_DOUBLE:
1598 assert( i + 1 < total_in_args &&
1599 in_sig_bt[i + 1] == T_VOID &&
1600 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1601 __ double_move(in_regs[i], out_regs[c_arg]);
1602 float_args++;
1603 break;
1604
1605 case T_LONG :
1606 __ long_move(in_regs[i], out_regs[c_arg]);
1607 int_args++;
1608 break;
1609
1610 case T_ADDRESS:
1611 assert(false, "found T_ADDRESS in java args");
1612 break;
1613
1614 default:
1615 __ move32_64(in_regs[i], out_regs[c_arg]);
1616 int_args++;
1617 }
1618 }
1619
1620 // point c_arg at the first arg that is already loaded in case we
1621 // need to spill before we call out
1622 int c_arg = total_c_args - total_in_args;
1623
1624 // Pre-load a static method's oop into c_rarg1.
1625 if (method->is_static()) {
1626
1627 // load oop into a register
1628 __ movoop(c_rarg1,
1629 JNIHandles::make_local(method->method_holder()->java_mirror()));
1630
1631 // Now handlize the static class mirror it's known not-null.
1632 __ sd(c_rarg1, Address(sp, klass_offset));
1633 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1634
1635 // Now get the handle
1636 __ la(c_rarg1, Address(sp, klass_offset));
1637 // and protect the arg if we must spill
1638 c_arg--;
1639 }
1640
1641 // Change state to native (we save the return address in the thread, since it might not
1642 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1643 // points into the right code segment. It does not have to be the correct return pc.
1644 // We use the same pc/oopMap repeatedly when we call out.
1645
1646 Label native_return;
1647 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
1648 // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1649 __ set_last_Java_frame(sp, noreg, native_return, t0);
1650 } else {
1651 intptr_t the_pc = (intptr_t) __ pc();
1652 oop_maps->add_gc_map(the_pc - start, map);
1653
1654 __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1655 }
1656
1657 Label dtrace_method_entry, dtrace_method_entry_done;
1658 if (DTraceMethodProbes) {
1659 __ j(dtrace_method_entry);
1660 __ bind(dtrace_method_entry_done);
1661 }
1662
1663 // RedefineClasses() tracing support for obsolete method entry
1664 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1665 // protect the args we've loaded
1666 save_args(masm, total_c_args, c_arg, out_regs);
1667 __ mov_metadata(c_rarg1, method());
1668 __ call_VM_leaf(
1669 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1670 xthread, c_rarg1);
1671 restore_args(masm, total_c_args, c_arg, out_regs);
1672 }
1673
1674 // Lock a synchronized method
1675
1676 // Register definitions used by locking and unlocking
1677
1678 const Register swap_reg = x10;
1679 const Register obj_reg = x9; // Will contain the oop
1680 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1681 const Register old_hdr = x30; // value of old header at unlock time
1682 const Register lock_tmp = x31; // Temporary used by lightweight_lock/unlock
1683 const Register tmp = ra;
1684
1685 Label slow_path_lock;
1686 Label lock_done;
1687
1688 if (method->is_synchronized()) {
1689 Label count;
1690
1691 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1692
1693 // Get the handle (the 2nd argument)
1694 __ mv(oop_handle_reg, c_rarg1);
1695
1696 // Get address of the box
1697
1698 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1699
1700 // Load the oop from the handle
1701 __ ld(obj_reg, Address(oop_handle_reg, 0));
1702
1703 if (LockingMode == LM_MONITOR) {
1704 __ j(slow_path_lock);
1705 } else if (LockingMode == LM_LEGACY) {
1706 // Load (object->mark() | 1) into swap_reg % x10
1707 __ ld(t0, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1708 __ ori(swap_reg, t0, 1);
1709
1710 // Save (object->mark() | 1) into BasicLock's displaced header
1711 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1712
1713 // src -> dest if dest == x10 else x10 <- dest
1714 __ cmpxchg_obj_header(x10, lock_reg, obj_reg, lock_tmp, count, /*fallthrough*/nullptr);
1715
1716 // Test if the oopMark is an obvious stack pointer, i.e.,
1717 // 1) (mark & 3) == 0, and
1718 // 2) sp <= mark < mark + os::pagesize()
1719 // These 3 tests can be done by evaluating the following
1720 // expression: ((mark - sp) & (3 - os::vm_page_size())),
1721 // assuming both stack pointer and pagesize have their
1722 // least significant 2 bits clear.
1723 // NOTE: the oopMark is in swap_reg % 10 as the result of cmpxchg
1724
1725 __ sub(swap_reg, swap_reg, sp);
1726 __ andi(swap_reg, swap_reg, 3 - (int)os::vm_page_size());
1727
1728 // Save the test result, for recursive case, the result is zero
1729 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1730 __ bnez(swap_reg, slow_path_lock);
1731
1732 __ bind(count);
1733 __ inc_held_monitor_count();
1734 } else {
1735 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
1736 __ lightweight_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1737 }
1738
1739 // Slow path will re-enter here
1740 __ bind(lock_done);
1741 }
1742
1743
1744 // Finally just about ready to make the JNI call
1745
1746 // get JNIEnv* which is first argument to native
1747 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1748
1749 // Now set thread in native
1750 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1751 __ mv(t0, _thread_in_native);
1752 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1753 __ sw(t0, Address(t1));
1754
1755 __ rt_call(native_func);
1756
1757 // Verify or restore cpu control state after JNI call
1758 __ restore_cpu_control_state_after_jni(t0);
1759
1760 // Unpack native results.
1761 if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1762 __ cast_primitive_type(ret_type, x10);
1763 }
1764
1765 Label safepoint_in_progress, safepoint_in_progress_done;
1766 Label after_transition;
1767
1768 // Switch thread to "native transition" state before reading the synchronization state.
1769 // This additional state is necessary because reading and testing the synchronization
1770 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1771 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1772 // VM thread changes sync state to synchronizing and suspends threads for GC.
1773 // Thread A is resumed to finish this native method, but doesn't block here since it
1774 // didn't see any synchronization is progress, and escapes.
1775 __ mv(t0, _thread_in_native_trans);
1776
1777 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1778
1779 // Force this write out before the read below
1780 if (!UseSystemMemoryBarrier) {
1781 __ membar(MacroAssembler::AnyAny);
1782 }
1783
1784 // check for safepoint operation in progress and/or pending suspend requests
1785 {
1786 // We need an acquire here to ensure that any subsequent load of the
1787 // global SafepointSynchronize::_state flag is ordered after this load
1788 // of the thread-local polling word. We don't want this poll to
1789 // return false (i.e. not safepointing) and a later poll of the global
1790 // SafepointSynchronize::_state spuriously to return true.
1791 // This is to avoid a race when we're in a native->Java transition
1792 // racing the code which wakes up from a safepoint.
1793
1794 __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */);
1795 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1796 __ bnez(t0, safepoint_in_progress);
1797 __ bind(safepoint_in_progress_done);
1798 }
1799
1800 // change thread state
1801 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1802 __ mv(t0, _thread_in_Java);
1803 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1804 __ sw(t0, Address(t1));
1805 __ bind(after_transition);
1806
1807 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
1808 // Check preemption for Object.wait()
1809 __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1810 __ beqz(t1, native_return);
1811 __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1812 __ jr(t1);
1813 __ bind(native_return);
1814
1815 intptr_t the_pc = (intptr_t) __ pc();
1816 oop_maps->add_gc_map(the_pc - start, map);
1817 }
1818
1819 Label reguard;
1820 Label reguard_done;
1821 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1822 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1823 __ beq(t0, t1, reguard);
1824 __ bind(reguard_done);
1825
1826 // native result if any is live
1827
1828 // Unlock
1829 Label unlock_done;
1830 Label slow_path_unlock;
1831 if (method->is_synchronized()) {
1832
1833 // Get locked oop from the handle we passed to jni
1834 __ ld(obj_reg, Address(oop_handle_reg, 0));
1835
1836 Label done, not_recursive;
1837
1838 if (LockingMode == LM_LEGACY) {
1839 // Simple recursive lock?
1840 __ ld(t0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1841 __ bnez(t0, not_recursive);
1842 __ dec_held_monitor_count();
1843 __ j(done);
1844 }
1845
1846 __ bind(not_recursive);
1847
1848 // Must save x10 if if it is live now because cmpxchg must use it
1849 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1850 save_native_result(masm, ret_type, stack_slots);
1851 }
1852
1853 if (LockingMode == LM_MONITOR) {
1854 __ j(slow_path_unlock);
1855 } else if (LockingMode == LM_LEGACY) {
1856 // get address of the stack lock
1857 __ la(x10, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1858 // get old displaced header
1859 __ ld(old_hdr, Address(x10, 0));
1860
1861 // Atomic swap old header if oop still contains the stack lock
1862 Label count;
1863 __ cmpxchg_obj_header(x10, old_hdr, obj_reg, lock_tmp, count, &slow_path_unlock);
1864 __ bind(count);
1865 __ dec_held_monitor_count();
1866 } else {
1867 assert(LockingMode == LM_LIGHTWEIGHT, "");
1868 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1869 }
1870
1871 // slow path re-enters here
1872 __ bind(unlock_done);
1873 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1874 restore_native_result(masm, ret_type, stack_slots);
1875 }
1876
1877 __ bind(done);
1878 }
1879
1880 Label dtrace_method_exit, dtrace_method_exit_done;
1881 if (DTraceMethodProbes) {
1882 __ j(dtrace_method_exit);
1883 __ bind(dtrace_method_exit_done);
1884 }
1885
1886 __ reset_last_Java_frame(false);
1887
1888 // Unbox oop result, e.g. JNIHandles::resolve result.
1889 if (is_reference_type(ret_type)) {
1890 __ resolve_jobject(x10, x11, x12);
1891 }
1892
1893 if (CheckJNICalls) {
1894 // clear_pending_jni_exception_check
1895 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1896 }
1897
1898 // reset handle block
1899 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1900 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1901
1902 __ leave();
1903
1904 // Any exception pending?
1905 Label exception_pending;
1906 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1907 __ bnez(t0, exception_pending);
1908
1909 // We're done
1910 __ ret();
1911
1912 // Unexpected paths are out of line and go here
1913
1914 // forward the exception
1915 __ bind(exception_pending);
1916
1917 // and forward the exception
1918 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1919
1920 // Slow path locking & unlocking
1921 if (method->is_synchronized()) {
1922
1923 __ block_comment("Slow path lock {");
1924 __ bind(slow_path_lock);
1925
1926 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1927 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1928
1929 // protect the args we've loaded
1930 save_args(masm, total_c_args, c_arg, out_regs);
1931
1932 __ mv(c_rarg0, obj_reg);
1933 __ mv(c_rarg1, lock_reg);
1934 __ mv(c_rarg2, xthread);
1935
1936 // Not a leaf but we have last_Java_frame setup as we want
1937 // Force freeze slow path in case we try to preempt. We will pin the
1938 // vthread to the carrier (see FreezeBase::recurse_freeze_native_frame()).
1939 __ push_cont_fastpath();
1940 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1941 __ pop_cont_fastpath();
1942 restore_args(masm, total_c_args, c_arg, out_regs);
1943
1944 #ifdef ASSERT
1945 { Label L;
1946 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1947 __ beqz(t0, L);
1948 __ stop("no pending exception allowed on exit from monitorenter");
1949 __ bind(L);
1950 }
1951 #endif
1952 __ j(lock_done);
1953
1954 __ block_comment("} Slow path lock");
1955
1956 __ block_comment("Slow path unlock {");
1957 __ bind(slow_path_unlock);
1958
1959 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1960 save_native_result(masm, ret_type, stack_slots);
1961 }
1962
1963 __ mv(c_rarg2, xthread);
1964 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1965 __ mv(c_rarg0, obj_reg);
1966
1967 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1968 // NOTE that obj_reg == x9 currently
1969 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1970 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1971
1972 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1973
1974 #ifdef ASSERT
1975 {
1976 Label L;
1977 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1978 __ beqz(t0, L);
1979 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1980 __ bind(L);
1981 }
1982 #endif /* ASSERT */
1983
1984 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1985
1986 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1987 restore_native_result(masm, ret_type, stack_slots);
1988 }
1989 __ j(unlock_done);
1990
1991 __ block_comment("} Slow path unlock");
1992
1993 } // synchronized
1994
1995 // SLOW PATH Reguard the stack if needed
1996
1997 __ bind(reguard);
1998 save_native_result(masm, ret_type, stack_slots);
1999 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2000 restore_native_result(masm, ret_type, stack_slots);
2001 // and continue
2002 __ j(reguard_done);
2003
2004 // SLOW PATH safepoint
2005 {
2006 __ block_comment("safepoint {");
2007 __ bind(safepoint_in_progress);
2008
2009 // Don't use call_VM as it will see a possible pending exception and forward it
2010 // and never return here preventing us from clearing _last_native_pc down below.
2011 //
2012 save_native_result(masm, ret_type, stack_slots);
2013 __ mv(c_rarg0, xthread);
2014 #ifndef PRODUCT
2015 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2016 #endif
2017 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
2018
2019 // Restore any method result value
2020 restore_native_result(masm, ret_type, stack_slots);
2021
2022 __ j(safepoint_in_progress_done);
2023 __ block_comment("} safepoint");
2024 }
2025
2026 // SLOW PATH dtrace support
2027 if (DTraceMethodProbes) {
2028 {
2029 __ block_comment("dtrace entry {");
2030 __ bind(dtrace_method_entry);
2031
2032 // We have all of the arguments setup at this point. We must not touch any register
2033 // argument registers at this point (what if we save/restore them there are no oop?
2034
2035 save_args(masm, total_c_args, c_arg, out_regs);
2036 __ mov_metadata(c_rarg1, method());
2037 __ call_VM_leaf(
2038 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2039 xthread, c_rarg1);
2040 restore_args(masm, total_c_args, c_arg, out_regs);
2041 __ j(dtrace_method_entry_done);
2042 __ block_comment("} dtrace entry");
2043 }
2044
2045 {
2046 __ block_comment("dtrace exit {");
2047 __ bind(dtrace_method_exit);
2048 save_native_result(masm, ret_type, stack_slots);
2049 __ mov_metadata(c_rarg1, method());
2050 __ call_VM_leaf(
2051 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2052 xthread, c_rarg1);
2053 restore_native_result(masm, ret_type, stack_slots);
2054 __ j(dtrace_method_exit_done);
2055 __ block_comment("} dtrace exit");
2056 }
2057 }
2058
2059 __ flush();
2060
2061 nmethod *nm = nmethod::new_native_nmethod(method,
2062 compile_id,
2063 masm->code(),
2064 vep_offset,
2065 frame_complete,
2066 stack_slots / VMRegImpl::slots_per_word,
2067 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2068 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2069 oop_maps);
2070 assert(nm != nullptr, "create native nmethod fail!");
2071 return nm;
2072 }
2073
2074 // this function returns the adjust size (in number of words) to a c2i adapter
2075 // activation for use during deoptimization
2076 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2077 assert(callee_locals >= callee_parameters,
2078 "test and remove; got more parms than locals");
2079 if (callee_locals < callee_parameters) {
2080 return 0; // No adjustment for negative locals
2081 }
2082 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2083 // diff is counted in stack words
2084 return align_up(diff, 2);
2085 }
2086
2087 //------------------------------generate_deopt_blob----------------------------
2088 void SharedRuntime::generate_deopt_blob() {
2089 // Allocate space for the code
2090 ResourceMark rm;
2091 // Setup code generation tools
2092 int pad = 0;
2093 #if INCLUDE_JVMCI
2094 if (EnableJVMCI) {
2095 pad += 512; // Increase the buffer size when compiling for JVMCI
2096 }
2097 #endif
2098 const char* name = SharedRuntime::stub_name(SharedStubId::deopt_id);
2099 CodeBuffer buffer(name, 2048 + pad, 1024);
2100 MacroAssembler* masm = new MacroAssembler(&buffer);
2101 int frame_size_in_words = -1;
2102 OopMap* map = nullptr;
2103 OopMapSet *oop_maps = new OopMapSet();
2104 assert_cond(masm != nullptr && oop_maps != nullptr);
2105 RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0);
2106
2107 // -------------
2108 // This code enters when returning to a de-optimized nmethod. A return
2109 // address has been pushed on the stack, and return values are in
2110 // registers.
2111 // If we are doing a normal deopt then we were called from the patched
2112 // nmethod from the point we returned to the nmethod. So the return
2113 // address on the stack is wrong by NativeCall::instruction_size
2114 // We will adjust the value so it looks like we have the original return
2115 // address on the stack (like when we eagerly deoptimized).
2116 // In the case of an exception pending when deoptimizing, we enter
2117 // with a return address on the stack that points after the call we patched
2118 // into the exception handler. We have the following register state from,
2119 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2120 // x10: exception oop
2121 // x9: exception handler
2122 // x13: throwing pc
2123 // So in this case we simply jam x13 into the useless return address and
2124 // the stack looks just like we want.
2125 //
2126 // At this point we need to de-opt. We save the argument return
2127 // registers. We call the first C routine, fetch_unroll_info(). This
2128 // routine captures the return values and returns a structure which
2129 // describes the current frame size and the sizes of all replacement frames.
2130 // The current frame is compiled code and may contain many inlined
2131 // functions, each with their own JVM state. We pop the current frame, then
2132 // push all the new frames. Then we call the C routine unpack_frames() to
2133 // populate these frames. Finally unpack_frames() returns us the new target
2134 // address. Notice that callee-save registers are BLOWN here; they have
2135 // already been captured in the vframeArray at the time the return PC was
2136 // patched.
2137 address start = __ pc();
2138 Label cont;
2139
2140 // Prolog for non exception case!
2141
2142 // Save everything in sight.
2143 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2144
2145 // Normal deoptimization. Save exec mode for unpack_frames.
2146 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2147 __ j(cont);
2148
2149 int reexecute_offset = __ pc() - start;
2150 #if INCLUDE_JVMCI && !defined(COMPILER1)
2151 if (UseJVMCICompiler) {
2152 // JVMCI does not use this kind of deoptimization
2153 __ should_not_reach_here();
2154 }
2155 #endif
2156
2157 // Reexecute case
2158 // return address is the pc describes what bci to do re-execute at
2159
2160 // No need to update map as each call to save_live_registers will produce identical oopmap
2161 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2162
2163 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2164 __ j(cont);
2165
2166 #if INCLUDE_JVMCI
2167 Label after_fetch_unroll_info_call;
2168 int implicit_exception_uncommon_trap_offset = 0;
2169 int uncommon_trap_offset = 0;
2170
2171 if (EnableJVMCI) {
2172 implicit_exception_uncommon_trap_offset = __ pc() - start;
2173
2174 __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2175 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2176
2177 uncommon_trap_offset = __ pc() - start;
2178
2179 // Save everything in sight.
2180 reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2181 // fetch_unroll_info needs to call last_java_frame()
2182 Label retaddr;
2183 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2184
2185 __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2186 __ mv(t0, -1);
2187 __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2188
2189 __ mv(xcpool, Deoptimization::Unpack_reexecute);
2190 __ mv(c_rarg0, xthread);
2191 __ orrw(c_rarg2, zr, xcpool); // exec mode
2192 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2193 __ bind(retaddr);
2194 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2195
2196 __ reset_last_Java_frame(false);
2197
2198 __ j(after_fetch_unroll_info_call);
2199 } // EnableJVMCI
2200 #endif // INCLUDE_JVMCI
2201
2202 int exception_offset = __ pc() - start;
2203
2204 // Prolog for exception case
2205
2206 // all registers are dead at this entry point, except for x10, and
2207 // x13 which contain the exception oop and exception pc
2208 // respectively. Set them in TLS and fall thru to the
2209 // unpack_with_exception_in_tls entry point.
2210
2211 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2212 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2213
2214 int exception_in_tls_offset = __ pc() - start;
2215
2216 // new implementation because exception oop is now passed in JavaThread
2217
2218 // Prolog for exception case
2219 // All registers must be preserved because they might be used by LinearScan
2220 // Exceptiop oop and throwing PC are passed in JavaThread
2221 // tos: stack at point of call to method that threw the exception (i.e. only
2222 // args are on the stack, no return address)
2223
2224 // The return address pushed by save_live_registers will be patched
2225 // later with the throwing pc. The correct value is not available
2226 // now because loading it from memory would destroy registers.
2227
2228 // NB: The SP at this point must be the SP of the method that is
2229 // being deoptimized. Deoptimization assumes that the frame created
2230 // here by save_live_registers is immediately below the method's SP.
2231 // This is a somewhat fragile mechanism.
2232
2233 // Save everything in sight.
2234 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2235
2236 // Now it is safe to overwrite any register
2237
2238 // Deopt during an exception. Save exec mode for unpack_frames.
2239 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2240
2241 // load throwing pc from JavaThread and patch it as the return address
2242 // of the current frame. Then clear the field in JavaThread
2243
2244 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2245 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2246 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2247
2248 #ifdef ASSERT
2249 // verify that there is really an exception oop in JavaThread
2250 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2251 __ verify_oop(x10);
2252
2253 // verify that there is no pending exception
2254 Label no_pending_exception;
2255 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2256 __ beqz(t0, no_pending_exception);
2257 __ stop("must not have pending exception here");
2258 __ bind(no_pending_exception);
2259 #endif
2260
2261 __ bind(cont);
2262
2263 // Call C code. Need thread and this frame, but NOT official VM entry
2264 // crud. We cannot block on this call, no GC can happen.
2265 //
2266 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2267
2268 // fetch_unroll_info needs to call last_java_frame().
2269
2270 Label retaddr;
2271 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2272 #ifdef ASSERT
2273 {
2274 Label L;
2275 __ ld(t0, Address(xthread,
2276 JavaThread::last_Java_fp_offset()));
2277 __ beqz(t0, L);
2278 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2279 __ bind(L);
2280 }
2281 #endif // ASSERT
2282 __ mv(c_rarg0, xthread);
2283 __ mv(c_rarg1, xcpool);
2284 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2285 __ bind(retaddr);
2286
2287 // Need to have an oopmap that tells fetch_unroll_info where to
2288 // find any register it might need.
2289 oop_maps->add_gc_map(__ pc() - start, map);
2290
2291 __ reset_last_Java_frame(false);
2292
2293 #if INCLUDE_JVMCI
2294 if (EnableJVMCI) {
2295 __ bind(after_fetch_unroll_info_call);
2296 }
2297 #endif
2298
2299 // Load UnrollBlock* into x15
2300 __ mv(x15, x10);
2301
2302 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2303 Label noException;
2304 __ mv(t0, Deoptimization::Unpack_exception);
2305 __ bne(xcpool, t0, noException); // Was exception pending?
2306 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2307 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2308 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2309 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2310
2311 __ verify_oop(x10);
2312
2313 // Overwrite the result registers with the exception results.
2314 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2315
2316 __ bind(noException);
2317
2318 // Only register save data is on the stack.
2319 // Now restore the result registers. Everything else is either dead
2320 // or captured in the vframeArray.
2321
2322 // Restore fp result register
2323 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2324 // Restore integer result register
2325 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2326
2327 // Pop all of the register save area off the stack
2328 __ add(sp, sp, frame_size_in_words * wordSize);
2329
2330 // All of the register save area has been popped of the stack. Only the
2331 // return address remains.
2332
2333 // Pop all the frames we must move/replace.
2334 //
2335 // Frame picture (youngest to oldest)
2336 // 1: self-frame (no frame link)
2337 // 2: deopting frame (no frame link)
2338 // 3: caller of deopting frame (could be compiled/interpreted).
2339 //
2340 // Note: by leaving the return address of self-frame on the stack
2341 // and using the size of frame 2 to adjust the stack
2342 // when we are done the return to frame 3 will still be on the stack.
2343
2344 // Pop deoptimized frame
2345 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2346 __ sub(x12, x12, 2 * wordSize);
2347 __ add(sp, sp, x12);
2348 __ ld(fp, Address(sp, 0));
2349 __ ld(ra, Address(sp, wordSize));
2350 __ addi(sp, sp, 2 * wordSize);
2351 // RA should now be the return address to the caller (3)
2352
2353 #ifdef ASSERT
2354 // Compilers generate code that bang the stack by as much as the
2355 // interpreter would need. So this stack banging should never
2356 // trigger a fault. Verify that it does not on non product builds.
2357 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2358 __ bang_stack_size(x9, x12);
2359 #endif
2360 // Load address of array of frame pcs into x12
2361 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2362
2363 // Load address of array of frame sizes into x14
2364 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2365
2366 // Load counter into x13
2367 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2368
2369 // Now adjust the caller's stack to make up for the extra locals
2370 // but record the original sp so that we can save it in the skeletal interpreter
2371 // frame and the stack walking of interpreter_sender will get the unextended sp
2372 // value and not the "real" sp value.
2373
2374 const Register sender_sp = x16;
2375
2376 __ mv(sender_sp, sp);
2377 __ lwu(x9, Address(x15,
2378 Deoptimization::UnrollBlock::
2379 caller_adjustment_offset()));
2380 __ sub(sp, sp, x9);
2381
2382 // Push interpreter frames in a loop
2383 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2384 __ mv(t1, t0);
2385 Label loop;
2386 __ bind(loop);
2387 __ ld(x9, Address(x14, 0)); // Load frame size
2388 __ addi(x14, x14, wordSize);
2389 __ sub(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2390 __ ld(ra, Address(x12, 0)); // Load pc
2391 __ addi(x12, x12, wordSize);
2392 __ enter(); // Save old & set new fp
2393 __ sub(sp, sp, x9); // Prolog
2394 // This value is corrected by layout_activation_impl
2395 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2396 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2397 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2398 __ addi(x13, x13, -1); // Decrement counter
2399 __ bnez(x13, loop);
2400
2401 // Re-push self-frame
2402 __ ld(ra, Address(x12));
2403 __ enter();
2404
2405 // Allocate a full sized register save area. We subtract 2 because
2406 // enter() just pushed 2 words
2407 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2408
2409 // Restore frame locals after moving the frame
2410 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2411 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2412
2413 // Call C code. Need thread but NOT official VM entry
2414 // crud. We cannot block on this call, no GC can happen. Call should
2415 // restore return values to their stack-slots with the new SP.
2416 //
2417 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2418
2419 // Use fp because the frames look interpreted now
2420 // Don't need the precise return PC here, just precise enough to point into this code blob.
2421 address the_pc = __ pc();
2422 __ set_last_Java_frame(sp, fp, the_pc, t0);
2423
2424 __ mv(c_rarg0, xthread);
2425 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2426 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2427
2428 // Set an oopmap for the call site
2429 // Use the same PC we used for the last java frame
2430 oop_maps->add_gc_map(the_pc - start,
2431 new OopMap(frame_size_in_words, 0));
2432
2433 // Clear fp AND pc
2434 __ reset_last_Java_frame(true);
2435
2436 // Collect return values
2437 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2438 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2439
2440 // Pop self-frame.
2441 __ leave(); // Epilog
2442
2443 // Jump to interpreter
2444 __ ret();
2445
2446 // Make sure all code is generated
2447 masm->flush();
2448
2449 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2450 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2451 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2452 #if INCLUDE_JVMCI
2453 if (EnableJVMCI) {
2454 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2455 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2456 }
2457 #endif
2458 }
2459
2460 // Number of stack slots between incoming argument block and the start of
2461 // a new frame. The PROLOG must add this many slots to the stack. The
2462 // EPILOG must remove this many slots.
2463 // RISCV needs two words for RA (return address) and FP (frame pointer).
2464 uint SharedRuntime::in_preserve_stack_slots() {
2465 return 2 * VMRegImpl::slots_per_word;
2466 }
2467
2468 uint SharedRuntime::out_preserve_stack_slots() {
2469 return 0;
2470 }
2471
2472 VMReg SharedRuntime::thread_register() {
2473 return xthread->as_VMReg();
2474 }
2475
2476 //------------------------------generate_handler_blob------
2477 //
2478 // Generate a special Compile2Runtime blob that saves all registers,
2479 // and setup oopmap.
2480 //
2481 SafepointBlob* SharedRuntime::generate_handler_blob(SharedStubId id, address call_ptr) {
2482 assert(is_polling_page_id(id), "expected a polling page stub id");
2483
2484 ResourceMark rm;
2485 OopMapSet *oop_maps = new OopMapSet();
2486 assert_cond(oop_maps != nullptr);
2487 OopMap* map = nullptr;
2488
2489 // Allocate space for the code. Setup code generation tools.
2490 const char* name = SharedRuntime::stub_name(id);
2491 CodeBuffer buffer(name, 2048, 1024);
2492 MacroAssembler* masm = new MacroAssembler(&buffer);
2493 assert_cond(masm != nullptr);
2494
2495 address start = __ pc();
2496 address call_pc = nullptr;
2497 int frame_size_in_words = -1;
2498 bool cause_return = (id == SharedStubId::polling_page_return_handler_id);
2499 RegisterSaver reg_saver(id == SharedStubId::polling_page_vectors_safepoint_handler_id /* save_vectors */);
2500
2501 // Save Integer and Float registers.
2502 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2503
2504 // The following is basically a call_VM. However, we need the precise
2505 // address of the call in order to generate an oopmap. Hence, we do all the
2506 // work ourselves.
2507
2508 Label retaddr;
2509 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2510
2511 // The return address must always be correct so that frame constructor never
2512 // sees an invalid pc.
2513
2514 if (!cause_return) {
2515 // overwrite the return address pushed by save_live_registers
2516 // Additionally, x18 is a callee-saved register so we can look at
2517 // it later to determine if someone changed the return address for
2518 // us!
2519 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2520 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2521 }
2522
2523 // Do the call
2524 __ mv(c_rarg0, xthread);
2525 __ rt_call(call_ptr);
2526 __ bind(retaddr);
2527
2528 // Set an oopmap for the call site. This oopmap will map all
2529 // oop-registers and debug-info registers as callee-saved. This
2530 // will allow deoptimization at this safepoint to find all possible
2531 // debug-info recordings, as well as let GC find all oops.
2532
2533 oop_maps->add_gc_map( __ pc() - start, map);
2534
2535 Label noException;
2536
2537 __ reset_last_Java_frame(false);
2538
2539 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2540
2541 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2542 __ beqz(t0, noException);
2543
2544 // Exception pending
2545
2546 reg_saver.restore_live_registers(masm);
2547
2548 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2549
2550 // No exception case
2551 __ bind(noException);
2552
2553 Label no_adjust, bail;
2554 if (!cause_return) {
2555 // If our stashed return pc was modified by the runtime we avoid touching it
2556 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2557 __ bne(x18, t0, no_adjust);
2558
2559 #ifdef ASSERT
2560 // Verify the correct encoding of the poll we're about to skip.
2561 // See NativeInstruction::is_lwu_to_zr()
2562 __ lwu(t0, Address(x18));
2563 __ andi(t1, t0, 0b0000011);
2564 __ mv(t2, 0b0000011);
2565 __ bne(t1, t2, bail); // 0-6:0b0000011
2566 __ srli(t1, t0, 7);
2567 __ andi(t1, t1, 0b00000);
2568 __ bnez(t1, bail); // 7-11:0b00000
2569 __ srli(t1, t0, 12);
2570 __ andi(t1, t1, 0b110);
2571 __ mv(t2, 0b110);
2572 __ bne(t1, t2, bail); // 12-14:0b110
2573 #endif
2574 // Adjust return pc forward to step over the safepoint poll instruction
2575 __ add(x18, x18, NativeInstruction::instruction_size);
2576 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2577 }
2578
2579 __ bind(no_adjust);
2580 // Normal exit, restore registers and exit.
2581
2582 reg_saver.restore_live_registers(masm);
2583 __ ret();
2584
2585 #ifdef ASSERT
2586 __ bind(bail);
2587 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2588 #endif
2589
2590 // Make sure all code is generated
2591 masm->flush();
2592
2593 // Fill-out other meta info
2594 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2595 }
2596
2597 //
2598 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2599 //
2600 // Generate a stub that calls into vm to find out the proper destination
2601 // of a java call. All the argument registers are live at this point
2602 // but since this is generic code we don't know what they are and the caller
2603 // must do any gc of the args.
2604 //
2605 RuntimeStub* SharedRuntime::generate_resolve_blob(SharedStubId id, address destination) {
2606 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2607 assert(is_resolve_id(id), "expected a resolve stub id");
2608
2609 // allocate space for the code
2610 ResourceMark rm;
2611
2612 const char* name = SharedRuntime::stub_name(id);
2613 CodeBuffer buffer(name, 1000, 512);
2614 MacroAssembler* masm = new MacroAssembler(&buffer);
2615 assert_cond(masm != nullptr);
2616
2617 int frame_size_in_words = -1;
2618 RegisterSaver reg_saver(false /* save_vectors */);
2619
2620 OopMapSet *oop_maps = new OopMapSet();
2621 assert_cond(oop_maps != nullptr);
2622 OopMap* map = nullptr;
2623
2624 int start = __ offset();
2625
2626 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2627
2628 int frame_complete = __ offset();
2629
2630 {
2631 Label retaddr;
2632 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2633
2634 __ mv(c_rarg0, xthread);
2635 __ rt_call(destination);
2636 __ bind(retaddr);
2637 }
2638
2639 // Set an oopmap for the call site.
2640 // We need this not only for callee-saved registers, but also for volatile
2641 // registers that the compiler might be keeping live across a safepoint.
2642
2643 oop_maps->add_gc_map( __ offset() - start, map);
2644
2645 // x10 contains the address we are going to jump to assuming no exception got installed
2646
2647 // clear last_Java_sp
2648 __ reset_last_Java_frame(false);
2649 // check for pending exceptions
2650 Label pending;
2651 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2652 __ bnez(t0, pending);
2653
2654 // get the returned Method*
2655 __ get_vm_result_2(xmethod, xthread);
2656 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2657
2658 // x10 is where we want to jump, overwrite t0 which is saved and temporary
2659 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t0)));
2660 reg_saver.restore_live_registers(masm);
2661
2662 // We are back to the original state on entry and ready to go.
2663
2664 __ jr(t0);
2665
2666 // Pending exception after the safepoint
2667
2668 __ bind(pending);
2669
2670 reg_saver.restore_live_registers(masm);
2671
2672 // exception pending => remove activation and forward to exception handler
2673
2674 __ sd(zr, Address(xthread, JavaThread::vm_result_offset()));
2675
2676 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2677 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2678
2679 // -------------
2680 // make sure all code is generated
2681 masm->flush();
2682
2683 // return the blob
2684 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2685 }
2686
2687 // Continuation point for throwing of implicit exceptions that are
2688 // not handled in the current activation. Fabricates an exception
2689 // oop and initiates normal exception dispatching in this
2690 // frame. Since we need to preserve callee-saved values (currently
2691 // only for C2, but done for C1 as well) we need a callee-saved oop
2692 // map and therefore have to make these stubs into RuntimeStubs
2693 // rather than BufferBlobs. If the compiler needs all registers to
2694 // be preserved between the fault point and the exception handler
2695 // then it must assume responsibility for that in
2696 // AbstractCompiler::continuation_for_implicit_null_exception or
2697 // continuation_for_implicit_division_by_zero_exception. All other
2698 // implicit exceptions (e.g., NullPointerException or
2699 // AbstractMethodError on entry) are either at call sites or
2700 // otherwise assume that stack unwinding will be initiated, so
2701 // caller saved registers were assumed volatile in the compiler.
2702
2703 RuntimeStub* SharedRuntime::generate_throw_exception(SharedStubId id, address runtime_entry) {
2704 assert(is_throw_id(id), "expected a throw stub id");
2705
2706 const char* name = SharedRuntime::stub_name(id);
2707
2708 // Information about frame layout at time of blocking runtime call.
2709 // Note that we only have to preserve callee-saved registers since
2710 // the compilers are responsible for supplying a continuation point
2711 // if they expect all registers to be preserved.
2712 // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2713 assert_cond(runtime_entry != nullptr);
2714 enum layout {
2715 fp_off = 0,
2716 fp_off2,
2717 return_off,
2718 return_off2,
2719 framesize // inclusive of return address
2720 };
2721
2722 const int insts_size = 1024;
2723 const int locs_size = 64;
2724
2725 ResourceMark rm;
2726 const char* timer_msg = "SharedRuntime generate_throw_exception";
2727 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2728
2729 CodeBuffer code(name, insts_size, locs_size);
2730 OopMapSet* oop_maps = new OopMapSet();
2731 MacroAssembler* masm = new MacroAssembler(&code);
2732 assert_cond(oop_maps != nullptr && masm != nullptr);
2733
2734 address start = __ pc();
2735
2736 // This is an inlined and slightly modified version of call_VM
2737 // which has the ability to fetch the return PC out of
2738 // thread-local storage and also sets up last_Java_sp slightly
2739 // differently than the real call_VM
2740
2741 __ enter(); // Save FP and RA before call
2742
2743 assert(is_even(framesize / 2), "sp not 16-byte aligned");
2744
2745 // ra and fp are already in place
2746 __ addi(sp, fp, 0 - ((unsigned)framesize << LogBytesPerInt)); // prolog
2747
2748 int frame_complete = __ pc() - start;
2749
2750 // Set up last_Java_sp and last_Java_fp
2751 address the_pc = __ pc();
2752 __ set_last_Java_frame(sp, fp, the_pc, t0);
2753
2754 // Call runtime
2755 __ mv(c_rarg0, xthread);
2756 BLOCK_COMMENT("call runtime_entry");
2757 __ rt_call(runtime_entry);
2758
2759 // Generate oop map
2760 OopMap* map = new OopMap(framesize, 0);
2761 assert_cond(map != nullptr);
2762
2763 oop_maps->add_gc_map(the_pc - start, map);
2764
2765 __ reset_last_Java_frame(true);
2766
2767 __ leave();
2768
2769 // check for pending exceptions
2770 #ifdef ASSERT
2771 Label L;
2772 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2773 __ bnez(t0, L);
2774 __ should_not_reach_here();
2775 __ bind(L);
2776 #endif // ASSERT
2777 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2778
2779 // codeBlob framesize is in words (not VMRegImpl::slot_size)
2780 RuntimeStub* stub =
2781 RuntimeStub::new_runtime_stub(name,
2782 &code,
2783 frame_complete,
2784 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2785 oop_maps, false);
2786 assert(stub != nullptr, "create runtime stub fail!");
2787 return stub;
2788 }
2789
2790 #if INCLUDE_JFR
2791
2792 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2793 __ set_last_Java_frame(sp, fp, the_pc, t0);
2794 __ mv(c_rarg0, thread);
2795 }
2796
2797 static void jfr_epilogue(MacroAssembler* masm) {
2798 __ reset_last_Java_frame(true);
2799 }
2800 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2801 // It returns a jobject handle to the event writer.
2802 // The handle is dereferenced and the return value is the event writer oop.
2803 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2804 enum layout {
2805 fp_off,
2806 fp_off2,
2807 return_off,
2808 return_off2,
2809 framesize // inclusive of return address
2810 };
2811
2812 int insts_size = 1024;
2813 int locs_size = 64;
2814 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
2815 CodeBuffer code(name, insts_size, locs_size);
2816 OopMapSet* oop_maps = new OopMapSet();
2817 MacroAssembler* masm = new MacroAssembler(&code);
2818
2819 address start = __ pc();
2820 __ enter();
2821 int frame_complete = __ pc() - start;
2822 address the_pc = __ pc();
2823 jfr_prologue(the_pc, masm, xthread);
2824 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2825
2826 jfr_epilogue(masm);
2827 __ resolve_global_jobject(x10, t0, t1);
2828 __ leave();
2829 __ ret();
2830
2831 OopMap* map = new OopMap(framesize, 1);
2832 oop_maps->add_gc_map(the_pc - start, map);
2833
2834 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2835 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2836 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2837 oop_maps, false);
2838 return stub;
2839 }
2840
2841 // For c2: call to return a leased buffer.
2842 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2843 enum layout {
2844 fp_off,
2845 fp_off2,
2846 return_off,
2847 return_off2,
2848 framesize // inclusive of return address
2849 };
2850
2851 int insts_size = 1024;
2852 int locs_size = 64;
2853 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_return_lease_id);
2854 CodeBuffer code(name, insts_size, locs_size);
2855 OopMapSet* oop_maps = new OopMapSet();
2856 MacroAssembler* masm = new MacroAssembler(&code);
2857
2858 address start = __ pc();
2859 __ enter();
2860 int frame_complete = __ pc() - start;
2861 address the_pc = __ pc();
2862 jfr_prologue(the_pc, masm, xthread);
2863 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2864
2865 jfr_epilogue(masm);
2866 __ leave();
2867 __ ret();
2868
2869 OopMap* map = new OopMap(framesize, 1);
2870 oop_maps->add_gc_map(the_pc - start, map);
2871
2872 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2873 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2874 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2875 oop_maps, false);
2876 return stub;
2877 }
2878
2879 #endif // INCLUDE_JFR