1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2012, 2024 SAP SE. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/compiledIC.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "frame_ppc.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/shared/gcLocker.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "interpreter/interp_masm.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/klass.inline.hpp"
38 #include "prims/methodHandles.hpp"
39 #include "runtime/continuation.hpp"
40 #include "runtime/continuationEntry.inline.hpp"
41 #include "runtime/jniHandles.hpp"
42 #include "runtime/os.inline.hpp"
43 #include "runtime/safepointMechanism.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/signature.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/timerTrace.hpp"
48 #include "runtime/vframeArray.hpp"
49 #include "utilities/align.hpp"
50 #include "utilities/macros.hpp"
51 #include "vmreg_ppc.inline.hpp"
52 #ifdef COMPILER1
53 #include "c1/c1_Runtime1.hpp"
54 #endif
55 #ifdef COMPILER2
56 #include "opto/ad.hpp"
57 #include "opto/runtime.hpp"
58 #endif
59
60 #include <alloca.h>
61
62 #define __ masm->
63
64 #ifdef PRODUCT
65 #define BLOCK_COMMENT(str) // nothing
66 #else
67 #define BLOCK_COMMENT(str) __ block_comment(str)
68 #endif
69
70 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
71
72
73 class RegisterSaver {
74 // Used for saving volatile registers.
75 public:
76
77 // Support different return pc locations.
78 enum ReturnPCLocation {
79 return_pc_is_lr,
80 return_pc_is_pre_saved,
81 return_pc_is_thread_saved_exception_pc
82 };
83
84 static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
85 int* out_frame_size_in_bytes,
86 bool generate_oop_map,
87 int return_pc_adjustment,
88 ReturnPCLocation return_pc_location,
89 bool save_vectors = false);
90 static void restore_live_registers_and_pop_frame(MacroAssembler* masm,
91 int frame_size_in_bytes,
92 bool restore_ctr,
93 bool save_vectors = false);
94
95 static void push_frame_and_save_argument_registers(MacroAssembler* masm,
96 Register r_temp,
97 int frame_size,
98 int total_args,
99 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
100 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm,
101 int frame_size,
102 int total_args,
103 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
104
105 // During deoptimization only the result registers need to be restored
106 // all the other values have already been extracted.
107 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes);
108
109 // Constants and data structures:
110
111 typedef enum {
112 int_reg,
113 float_reg,
114 special_reg,
115 vs_reg
116 } RegisterType;
117
118 typedef enum {
119 reg_size = 8,
120 half_reg_size = reg_size / 2,
121 vs_reg_size = 16
122 } RegisterConstants;
123
124 typedef struct {
125 RegisterType reg_type;
126 int reg_num;
127 VMReg vmreg;
128 } LiveRegType;
129 };
130
131
132 #define RegisterSaver_LiveIntReg(regname) \
133 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() }
134
135 #define RegisterSaver_LiveFloatReg(regname) \
136 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() }
137
138 #define RegisterSaver_LiveSpecialReg(regname) \
139 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() }
140
141 #define RegisterSaver_LiveVSReg(regname) \
142 { RegisterSaver::vs_reg, regname->encoding(), regname->as_VMReg() }
143
144 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
145 // Live registers which get spilled to the stack. Register
146 // positions in this array correspond directly to the stack layout.
147
148 //
149 // live special registers:
150 //
151 RegisterSaver_LiveSpecialReg(SR_CTR),
152 //
153 // live float registers:
154 //
155 RegisterSaver_LiveFloatReg( F0 ),
156 RegisterSaver_LiveFloatReg( F1 ),
157 RegisterSaver_LiveFloatReg( F2 ),
158 RegisterSaver_LiveFloatReg( F3 ),
159 RegisterSaver_LiveFloatReg( F4 ),
160 RegisterSaver_LiveFloatReg( F5 ),
161 RegisterSaver_LiveFloatReg( F6 ),
162 RegisterSaver_LiveFloatReg( F7 ),
163 RegisterSaver_LiveFloatReg( F8 ),
164 RegisterSaver_LiveFloatReg( F9 ),
165 RegisterSaver_LiveFloatReg( F10 ),
166 RegisterSaver_LiveFloatReg( F11 ),
167 RegisterSaver_LiveFloatReg( F12 ),
168 RegisterSaver_LiveFloatReg( F13 ),
169 RegisterSaver_LiveFloatReg( F14 ),
170 RegisterSaver_LiveFloatReg( F15 ),
171 RegisterSaver_LiveFloatReg( F16 ),
172 RegisterSaver_LiveFloatReg( F17 ),
173 RegisterSaver_LiveFloatReg( F18 ),
174 RegisterSaver_LiveFloatReg( F19 ),
175 RegisterSaver_LiveFloatReg( F20 ),
176 RegisterSaver_LiveFloatReg( F21 ),
177 RegisterSaver_LiveFloatReg( F22 ),
178 RegisterSaver_LiveFloatReg( F23 ),
179 RegisterSaver_LiveFloatReg( F24 ),
180 RegisterSaver_LiveFloatReg( F25 ),
181 RegisterSaver_LiveFloatReg( F26 ),
182 RegisterSaver_LiveFloatReg( F27 ),
183 RegisterSaver_LiveFloatReg( F28 ),
184 RegisterSaver_LiveFloatReg( F29 ),
185 RegisterSaver_LiveFloatReg( F30 ),
186 RegisterSaver_LiveFloatReg( F31 ),
187 //
188 // live integer registers:
189 //
190 RegisterSaver_LiveIntReg( R0 ),
191 //RegisterSaver_LiveIntReg( R1 ), // stack pointer
192 RegisterSaver_LiveIntReg( R2 ),
193 RegisterSaver_LiveIntReg( R3 ),
194 RegisterSaver_LiveIntReg( R4 ),
195 RegisterSaver_LiveIntReg( R5 ),
196 RegisterSaver_LiveIntReg( R6 ),
197 RegisterSaver_LiveIntReg( R7 ),
198 RegisterSaver_LiveIntReg( R8 ),
199 RegisterSaver_LiveIntReg( R9 ),
200 RegisterSaver_LiveIntReg( R10 ),
201 RegisterSaver_LiveIntReg( R11 ),
202 RegisterSaver_LiveIntReg( R12 ),
203 //RegisterSaver_LiveIntReg( R13 ), // system thread id
204 RegisterSaver_LiveIntReg( R14 ),
205 RegisterSaver_LiveIntReg( R15 ),
206 RegisterSaver_LiveIntReg( R16 ),
207 RegisterSaver_LiveIntReg( R17 ),
208 RegisterSaver_LiveIntReg( R18 ),
209 RegisterSaver_LiveIntReg( R19 ),
210 RegisterSaver_LiveIntReg( R20 ),
211 RegisterSaver_LiveIntReg( R21 ),
212 RegisterSaver_LiveIntReg( R22 ),
213 RegisterSaver_LiveIntReg( R23 ),
214 RegisterSaver_LiveIntReg( R24 ),
215 RegisterSaver_LiveIntReg( R25 ),
216 RegisterSaver_LiveIntReg( R26 ),
217 RegisterSaver_LiveIntReg( R27 ),
218 RegisterSaver_LiveIntReg( R28 ),
219 RegisterSaver_LiveIntReg( R29 ),
220 RegisterSaver_LiveIntReg( R30 ),
221 RegisterSaver_LiveIntReg( R31 ) // must be the last register (see save/restore functions below)
222 };
223
224 static const RegisterSaver::LiveRegType RegisterSaver_LiveVSRegs[] = {
225 //
226 // live vector scalar registers (optional, only these ones are used by C2):
227 //
228 RegisterSaver_LiveVSReg( VSR32 ),
229 RegisterSaver_LiveVSReg( VSR33 ),
230 RegisterSaver_LiveVSReg( VSR34 ),
231 RegisterSaver_LiveVSReg( VSR35 ),
232 RegisterSaver_LiveVSReg( VSR36 ),
233 RegisterSaver_LiveVSReg( VSR37 ),
234 RegisterSaver_LiveVSReg( VSR38 ),
235 RegisterSaver_LiveVSReg( VSR39 ),
236 RegisterSaver_LiveVSReg( VSR40 ),
237 RegisterSaver_LiveVSReg( VSR41 ),
238 RegisterSaver_LiveVSReg( VSR42 ),
239 RegisterSaver_LiveVSReg( VSR43 ),
240 RegisterSaver_LiveVSReg( VSR44 ),
241 RegisterSaver_LiveVSReg( VSR45 ),
242 RegisterSaver_LiveVSReg( VSR46 ),
243 RegisterSaver_LiveVSReg( VSR47 ),
244 RegisterSaver_LiveVSReg( VSR48 ),
245 RegisterSaver_LiveVSReg( VSR49 ),
246 RegisterSaver_LiveVSReg( VSR50 ),
247 RegisterSaver_LiveVSReg( VSR51 )
248 };
249
250
251 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
252 int* out_frame_size_in_bytes,
253 bool generate_oop_map,
254 int return_pc_adjustment,
255 ReturnPCLocation return_pc_location,
256 bool save_vectors) {
257 // Push an abi_reg_args-frame and store all registers which may be live.
258 // If requested, create an OopMap: Record volatile registers as
259 // callee-save values in an OopMap so their save locations will be
260 // propagated to the RegisterMap of the caller frame during
261 // StackFrameStream construction (needed for deoptimization; see
262 // compiledVFrame::create_stack_value).
263 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment.
264 // Updated return pc is returned in R31 (if not return_pc_is_pre_saved).
265
266 // calculate frame size
267 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
268 sizeof(RegisterSaver::LiveRegType);
269 const int vsregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVSRegs) /
270 sizeof(RegisterSaver::LiveRegType))
271 : 0;
272 const int register_save_size = regstosave_num * reg_size + vsregstosave_num * vs_reg_size;
273 const int frame_size_in_bytes = align_up(register_save_size, frame::alignment_in_bytes)
274 + frame::native_abi_reg_args_size;
275
276 *out_frame_size_in_bytes = frame_size_in_bytes;
277 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
278 const int register_save_offset = frame_size_in_bytes - register_save_size;
279
280 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
281 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : nullptr;
282
283 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {");
284
285 // push a new frame
286 __ push_frame(frame_size_in_bytes, noreg);
287
288 // Save some registers in the last (non-vector) slots of the new frame so we
289 // can use them as scratch regs or to determine the return pc.
290 __ std(R31, frame_size_in_bytes - reg_size - vsregstosave_num * vs_reg_size, R1_SP);
291 __ std(R30, frame_size_in_bytes - 2*reg_size - vsregstosave_num * vs_reg_size, R1_SP);
292
293 // save the flags
294 // Do the save_LR by hand and adjust the return pc if requested.
295 switch (return_pc_location) {
296 case return_pc_is_lr: __ mflr(R31); break;
297 case return_pc_is_pre_saved: assert(return_pc_adjustment == 0, "unsupported"); break;
298 case return_pc_is_thread_saved_exception_pc: __ ld(R31, thread_(saved_exception_pc)); break;
299 default: ShouldNotReachHere();
300 }
301 if (return_pc_location != return_pc_is_pre_saved) {
302 if (return_pc_adjustment != 0) {
303 __ addi(R31, R31, return_pc_adjustment);
304 }
305 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
306 }
307
308 // save all registers (ints and floats)
309 int offset = register_save_offset;
310
311 for (int i = 0; i < regstosave_num; i++) {
312 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
313 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
314
315 switch (reg_type) {
316 case RegisterSaver::int_reg: {
317 if (reg_num < 30) { // We spilled R30-31 right at the beginning.
318 __ std(as_Register(reg_num), offset, R1_SP);
319 }
320 break;
321 }
322 case RegisterSaver::float_reg: {
323 __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
324 break;
325 }
326 case RegisterSaver::special_reg: {
327 if (reg_num == SR_CTR.encoding()) {
328 __ mfctr(R30);
329 __ std(R30, offset, R1_SP);
330 } else {
331 Unimplemented();
332 }
333 break;
334 }
335 default:
336 ShouldNotReachHere();
337 }
338
339 if (generate_oop_map) {
340 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
341 RegisterSaver_LiveRegs[i].vmreg);
342 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2),
343 RegisterSaver_LiveRegs[i].vmreg->next());
344 }
345 offset += reg_size;
346 }
347
348 for (int i = 0; i < vsregstosave_num; i++) {
349 int reg_num = RegisterSaver_LiveVSRegs[i].reg_num;
350 int reg_type = RegisterSaver_LiveVSRegs[i].reg_type;
351
352 __ li(R30, offset);
353 __ stxvd2x(as_VectorSRegister(reg_num), R30, R1_SP);
354
355 if (generate_oop_map) {
356 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
357 RegisterSaver_LiveVSRegs[i].vmreg);
358 }
359 offset += vs_reg_size;
360 }
361
362 assert(offset == frame_size_in_bytes, "consistency check");
363
364 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers");
365
366 // And we're done.
367 return map;
368 }
369
370
371 // Pop the current frame and restore all the registers that we
372 // saved.
373 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
374 int frame_size_in_bytes,
375 bool restore_ctr,
376 bool save_vectors) {
377 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
378 sizeof(RegisterSaver::LiveRegType);
379 const int vsregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVSRegs) /
380 sizeof(RegisterSaver::LiveRegType))
381 : 0;
382 const int register_save_size = regstosave_num * reg_size + vsregstosave_num * vs_reg_size;
383
384 const int register_save_offset = frame_size_in_bytes - register_save_size;
385
386 BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
387
388 // restore all registers (ints and floats)
389 int offset = register_save_offset;
390
391 for (int i = 0; i < regstosave_num; i++) {
392 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
393 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
394
395 switch (reg_type) {
396 case RegisterSaver::int_reg: {
397 if (reg_num != 31) // R31 restored at the end, it's the tmp reg!
398 __ ld(as_Register(reg_num), offset, R1_SP);
399 break;
400 }
401 case RegisterSaver::float_reg: {
402 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
403 break;
404 }
405 case RegisterSaver::special_reg: {
406 if (reg_num == SR_CTR.encoding()) {
407 if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
408 __ ld(R31, offset, R1_SP);
409 __ mtctr(R31);
410 }
411 } else {
412 Unimplemented();
413 }
414 break;
415 }
416 default:
417 ShouldNotReachHere();
418 }
419 offset += reg_size;
420 }
421
422 for (int i = 0; i < vsregstosave_num; i++) {
423 int reg_num = RegisterSaver_LiveVSRegs[i].reg_num;
424 int reg_type = RegisterSaver_LiveVSRegs[i].reg_type;
425
426 __ li(R31, offset);
427 __ lxvd2x(as_VectorSRegister(reg_num), R31, R1_SP);
428
429 offset += vs_reg_size;
430 }
431
432 assert(offset == frame_size_in_bytes, "consistency check");
433
434 // restore link and the flags
435 __ ld(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
436 __ mtlr(R31);
437
438 // restore scratch register's value
439 __ ld(R31, frame_size_in_bytes - reg_size - vsregstosave_num * vs_reg_size, R1_SP);
440
441 // pop the frame
442 __ addi(R1_SP, R1_SP, frame_size_in_bytes);
443
444 BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
445 }
446
447 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
448 int frame_size,int total_args, const VMRegPair *regs,
449 const VMRegPair *regs2) {
450 __ push_frame(frame_size, r_temp);
451 int st_off = frame_size - wordSize;
452 for (int i = 0; i < total_args; i++) {
453 VMReg r_1 = regs[i].first();
454 VMReg r_2 = regs[i].second();
455 if (!r_1->is_valid()) {
456 assert(!r_2->is_valid(), "");
457 continue;
458 }
459 if (r_1->is_Register()) {
460 Register r = r_1->as_Register();
461 __ std(r, st_off, R1_SP);
462 st_off -= wordSize;
463 } else if (r_1->is_FloatRegister()) {
464 FloatRegister f = r_1->as_FloatRegister();
465 __ stfd(f, st_off, R1_SP);
466 st_off -= wordSize;
467 }
468 }
469 if (regs2 != nullptr) {
470 for (int i = 0; i < total_args; i++) {
471 VMReg r_1 = regs2[i].first();
472 VMReg r_2 = regs2[i].second();
473 if (!r_1->is_valid()) {
474 assert(!r_2->is_valid(), "");
475 continue;
476 }
477 if (r_1->is_Register()) {
478 Register r = r_1->as_Register();
479 __ std(r, st_off, R1_SP);
480 st_off -= wordSize;
481 } else if (r_1->is_FloatRegister()) {
482 FloatRegister f = r_1->as_FloatRegister();
483 __ stfd(f, st_off, R1_SP);
484 st_off -= wordSize;
485 }
486 }
487 }
488 }
489
490 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
491 int total_args, const VMRegPair *regs,
492 const VMRegPair *regs2) {
493 int st_off = frame_size - wordSize;
494 for (int i = 0; i < total_args; i++) {
495 VMReg r_1 = regs[i].first();
496 VMReg r_2 = regs[i].second();
497 if (r_1->is_Register()) {
498 Register r = r_1->as_Register();
499 __ ld(r, st_off, R1_SP);
500 st_off -= wordSize;
501 } else if (r_1->is_FloatRegister()) {
502 FloatRegister f = r_1->as_FloatRegister();
503 __ lfd(f, st_off, R1_SP);
504 st_off -= wordSize;
505 }
506 }
507 if (regs2 != nullptr)
508 for (int i = 0; i < total_args; i++) {
509 VMReg r_1 = regs2[i].first();
510 VMReg r_2 = regs2[i].second();
511 if (r_1->is_Register()) {
512 Register r = r_1->as_Register();
513 __ ld(r, st_off, R1_SP);
514 st_off -= wordSize;
515 } else if (r_1->is_FloatRegister()) {
516 FloatRegister f = r_1->as_FloatRegister();
517 __ lfd(f, st_off, R1_SP);
518 st_off -= wordSize;
519 }
520 }
521 __ pop_frame();
522 }
523
524 // Restore the registers that might be holding a result.
525 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) {
526 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
527 sizeof(RegisterSaver::LiveRegType);
528 const int register_save_size = regstosave_num * reg_size; // VS registers not relevant here.
529 const int register_save_offset = frame_size_in_bytes - register_save_size;
530
531 // restore all result registers (ints and floats)
532 int offset = register_save_offset;
533 for (int i = 0; i < regstosave_num; i++) {
534 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
535 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
536 switch (reg_type) {
537 case RegisterSaver::int_reg: {
538 if (as_Register(reg_num)==R3_RET) // int result_reg
539 __ ld(as_Register(reg_num), offset, R1_SP);
540 break;
541 }
542 case RegisterSaver::float_reg: {
543 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
544 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
545 break;
546 }
547 case RegisterSaver::special_reg: {
548 // Special registers don't hold a result.
549 break;
550 }
551 default:
552 ShouldNotReachHere();
553 }
554 offset += reg_size;
555 }
556
557 assert(offset == frame_size_in_bytes, "consistency check");
558 }
559
560 // Is vector's size (in bytes) bigger than a size saved by default?
561 bool SharedRuntime::is_wide_vector(int size) {
562 // Note, MaxVectorSize == 8/16 on PPC64.
563 assert(size <= (SuperwordUseVSX ? 16 : 8), "%d bytes vectors are not supported", size);
564 return size > 8;
565 }
566
567 static int reg2slot(VMReg r) {
568 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
569 }
570
571 static int reg2offset(VMReg r) {
572 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
573 }
574
575 // ---------------------------------------------------------------------------
576 // Read the array of BasicTypes from a signature, and compute where the
577 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
578 // quantities. Values less than VMRegImpl::stack0 are registers, those above
579 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
580 // as framesizes are fixed.
581 // VMRegImpl::stack0 refers to the first slot 0(sp).
582 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
583 // up to Register::number_of_registers) are the 64-bit
584 // integer registers.
585
586 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
587 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
588 // units regardless of build. Of course for i486 there is no 64 bit build
589
590 // The Java calling convention is a "shifted" version of the C ABI.
591 // By skipping the first C ABI register we can call non-static jni methods
592 // with small numbers of arguments without having to shuffle the arguments
593 // at all. Since we control the java ABI we ought to at least get some
594 // advantage out of it.
595
596 const VMReg java_iarg_reg[8] = {
597 R3->as_VMReg(),
598 R4->as_VMReg(),
599 R5->as_VMReg(),
600 R6->as_VMReg(),
601 R7->as_VMReg(),
602 R8->as_VMReg(),
603 R9->as_VMReg(),
604 R10->as_VMReg()
605 };
606
607 const VMReg java_farg_reg[13] = {
608 F1->as_VMReg(),
609 F2->as_VMReg(),
610 F3->as_VMReg(),
611 F4->as_VMReg(),
612 F5->as_VMReg(),
613 F6->as_VMReg(),
614 F7->as_VMReg(),
615 F8->as_VMReg(),
616 F9->as_VMReg(),
617 F10->as_VMReg(),
618 F11->as_VMReg(),
619 F12->as_VMReg(),
620 F13->as_VMReg()
621 };
622
623 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
624 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
625
626 STATIC_ASSERT(num_java_iarg_registers == Argument::n_int_register_parameters_j);
627 STATIC_ASSERT(num_java_farg_registers == Argument::n_float_register_parameters_j);
628
629 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
630 VMRegPair *regs,
631 int total_args_passed) {
632 // C2c calling conventions for compiled-compiled calls.
633 // Put 8 ints/longs into registers _AND_ 13 float/doubles into
634 // registers _AND_ put the rest on the stack.
635
636 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats
637 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
638
639 int i;
640 VMReg reg;
641 int stk = 0;
642 int ireg = 0;
643 int freg = 0;
644
645 // We put the first 8 arguments into registers and the rest on the
646 // stack, float arguments are already in their argument registers
647 // due to c2c calling conventions (see calling_convention).
648 for (int i = 0; i < total_args_passed; ++i) {
649 switch(sig_bt[i]) {
650 case T_BOOLEAN:
651 case T_CHAR:
652 case T_BYTE:
653 case T_SHORT:
654 case T_INT:
655 if (ireg < num_java_iarg_registers) {
656 // Put int/ptr in register
657 reg = java_iarg_reg[ireg];
658 ++ireg;
659 } else {
660 // Put int/ptr on stack.
661 reg = VMRegImpl::stack2reg(stk);
662 stk += inc_stk_for_intfloat;
663 }
664 regs[i].set1(reg);
665 break;
666 case T_LONG:
667 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
668 if (ireg < num_java_iarg_registers) {
669 // Put long in register.
670 reg = java_iarg_reg[ireg];
671 ++ireg;
672 } else {
673 // Put long on stack. They must be aligned to 2 slots.
674 if (stk & 0x1) ++stk;
675 reg = VMRegImpl::stack2reg(stk);
676 stk += inc_stk_for_longdouble;
677 }
678 regs[i].set2(reg);
679 break;
680 case T_OBJECT:
681 case T_ARRAY:
682 case T_ADDRESS:
683 if (ireg < num_java_iarg_registers) {
684 // Put ptr in register.
685 reg = java_iarg_reg[ireg];
686 ++ireg;
687 } else {
688 // Put ptr on stack. Objects must be aligned to 2 slots too,
689 // because "64-bit pointers record oop-ishness on 2 aligned
690 // adjacent registers." (see OopFlow::build_oop_map).
691 if (stk & 0x1) ++stk;
692 reg = VMRegImpl::stack2reg(stk);
693 stk += inc_stk_for_longdouble;
694 }
695 regs[i].set2(reg);
696 break;
697 case T_FLOAT:
698 if (freg < num_java_farg_registers) {
699 // Put float in register.
700 reg = java_farg_reg[freg];
701 ++freg;
702 } else {
703 // Put float on stack.
704 reg = VMRegImpl::stack2reg(stk);
705 stk += inc_stk_for_intfloat;
706 }
707 regs[i].set1(reg);
708 break;
709 case T_DOUBLE:
710 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
711 if (freg < num_java_farg_registers) {
712 // Put double in register.
713 reg = java_farg_reg[freg];
714 ++freg;
715 } else {
716 // Put double on stack. They must be aligned to 2 slots.
717 if (stk & 0x1) ++stk;
718 reg = VMRegImpl::stack2reg(stk);
719 stk += inc_stk_for_longdouble;
720 }
721 regs[i].set2(reg);
722 break;
723 case T_VOID:
724 // Do not count halves.
725 regs[i].set_bad();
726 break;
727 default:
728 ShouldNotReachHere();
729 }
730 }
731 return stk;
732 }
733
734 #if defined(COMPILER1) || defined(COMPILER2)
735 // Calling convention for calling C code.
736 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
737 VMRegPair *regs,
738 int total_args_passed) {
739 // Calling conventions for C runtime calls and calls to JNI native methods.
740 //
741 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
742 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
743 // the first 13 flt/dbl's in the first 13 fp regs but additionally
744 // copy flt/dbl to the stack if they are beyond the 8th argument.
745
746 const VMReg iarg_reg[8] = {
747 R3->as_VMReg(),
748 R4->as_VMReg(),
749 R5->as_VMReg(),
750 R6->as_VMReg(),
751 R7->as_VMReg(),
752 R8->as_VMReg(),
753 R9->as_VMReg(),
754 R10->as_VMReg()
755 };
756
757 const VMReg farg_reg[13] = {
758 F1->as_VMReg(),
759 F2->as_VMReg(),
760 F3->as_VMReg(),
761 F4->as_VMReg(),
762 F5->as_VMReg(),
763 F6->as_VMReg(),
764 F7->as_VMReg(),
765 F8->as_VMReg(),
766 F9->as_VMReg(),
767 F10->as_VMReg(),
768 F11->as_VMReg(),
769 F12->as_VMReg(),
770 F13->as_VMReg()
771 };
772
773 // Check calling conventions consistency.
774 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
775 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
776 "consistency");
777
778 const int additional_frame_header_slots = ((frame::native_abi_minframe_size - frame::jit_out_preserve_size)
779 / VMRegImpl::stack_slot_size);
780 const int float_offset_in_slots = Argument::float_on_stack_offset_in_bytes_c / VMRegImpl::stack_slot_size;
781
782 VMReg reg;
783 int arg = 0;
784 int freg = 0;
785 bool stack_used = false;
786
787 for (int i = 0; i < total_args_passed; ++i, ++arg) {
788 // Each argument corresponds to a slot in the Parameter Save Area (if not omitted)
789 int stk = (arg * 2) + additional_frame_header_slots;
790
791 switch(sig_bt[i]) {
792 //
793 // If arguments 0-7 are integers, they are passed in integer registers.
794 // Argument i is placed in iarg_reg[i].
795 //
796 case T_BOOLEAN:
797 case T_CHAR:
798 case T_BYTE:
799 case T_SHORT:
800 case T_INT:
801 // We must cast ints to longs and use full 64 bit stack slots
802 // here. Thus fall through, handle as long.
803 case T_LONG:
804 case T_OBJECT:
805 case T_ARRAY:
806 case T_ADDRESS:
807 case T_METADATA:
808 // Oops are already boxed if required (JNI).
809 if (arg < Argument::n_int_register_parameters_c) {
810 reg = iarg_reg[arg];
811 } else {
812 reg = VMRegImpl::stack2reg(stk);
813 stack_used = true;
814 }
815 regs[i].set2(reg);
816 break;
817
818 //
819 // Floats are treated differently from int regs: The first 13 float arguments
820 // are passed in registers (not the float args among the first 13 args).
821 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed
822 // in farg_reg[j] if argument i is the j-th float argument of this call.
823 //
824 case T_FLOAT:
825 if (freg < Argument::n_float_register_parameters_c) {
826 // Put float in register ...
827 reg = farg_reg[freg];
828 ++freg;
829 } else {
830 // Put float on stack.
831 reg = VMRegImpl::stack2reg(stk + float_offset_in_slots);
832 stack_used = true;
833 }
834 regs[i].set1(reg);
835 break;
836 case T_DOUBLE:
837 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
838 if (freg < Argument::n_float_register_parameters_c) {
839 // Put double in register ...
840 reg = farg_reg[freg];
841 ++freg;
842 } else {
843 // Put double on stack.
844 reg = VMRegImpl::stack2reg(stk);
845 stack_used = true;
846 }
847 regs[i].set2(reg);
848 break;
849
850 case T_VOID:
851 // Do not count halves.
852 regs[i].set_bad();
853 --arg;
854 break;
855 default:
856 ShouldNotReachHere();
857 }
858 }
859
860 // Return size of the stack frame excluding the jit_out_preserve part in single-word slots.
861 #if defined(ABI_ELFv2)
862 assert(additional_frame_header_slots == 0, "ABIv2 shouldn't use extra slots");
863 // ABIv2 allows omitting the Parameter Save Area if the callee's prototype
864 // indicates that all parameters can be passed in registers.
865 return stack_used ? (arg * 2) : 0;
866 #else
867 // The Parameter Save Area needs to be at least 8 double-word slots for ABIv1.
868 // We have to add extra slots because ABIv1 uses a larger header.
869 return MAX2(arg, 8) * 2 + additional_frame_header_slots;
870 #endif
871 }
872 #endif // COMPILER2
873
874 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
875 uint num_bits,
876 uint total_args_passed) {
877 Unimplemented();
878 return 0;
879 }
880
881 static address gen_c2i_adapter(MacroAssembler *masm,
882 int total_args_passed,
883 int comp_args_on_stack,
884 const BasicType *sig_bt,
885 const VMRegPair *regs,
886 Label& call_interpreter,
887 const Register& ientry) {
888
889 address c2i_entrypoint;
890
891 const Register sender_SP = R21_sender_SP; // == R21_tmp1
892 const Register code = R22_tmp2;
893 //const Register ientry = R23_tmp3;
894 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
895 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
896 int value_regs_index = 0;
897
898 const Register return_pc = R27_tmp7;
899 const Register tmp = R28_tmp8;
900
901 assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
902
903 // Adapter needs TOP_IJAVA_FRAME_ABI.
904 const int adapter_size = frame::top_ijava_frame_abi_size +
905 align_up(total_args_passed * wordSize, frame::alignment_in_bytes);
906
907 // regular (verified) c2i entry point
908 c2i_entrypoint = __ pc();
909
910 // Does compiled code exists? If yes, patch the caller's callsite.
911 __ ld(code, method_(code));
912 __ cmpdi(CCR0, code, 0);
913 __ ld(ientry, method_(interpreter_entry)); // preloaded
914 __ beq(CCR0, call_interpreter);
915
916
917 // Patch caller's callsite, method_(code) was not null which means that
918 // compiled code exists.
919 __ mflr(return_pc);
920 __ std(return_pc, _abi0(lr), R1_SP);
921 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
922
923 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
924
925 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
926 __ ld(return_pc, _abi0(lr), R1_SP);
927 __ ld(ientry, method_(interpreter_entry)); // preloaded
928 __ mtlr(return_pc);
929
930
931 // Call the interpreter.
932 __ BIND(call_interpreter);
933 __ mtctr(ientry);
934
935 // Get a copy of the current SP for loading caller's arguments.
936 __ mr(sender_SP, R1_SP);
937
938 // Add space for the adapter.
939 __ resize_frame(-adapter_size, R12_scratch2);
940
941 int st_off = adapter_size - wordSize;
942
943 // Write the args into the outgoing interpreter space.
944 for (int i = 0; i < total_args_passed; i++) {
945 VMReg r_1 = regs[i].first();
946 VMReg r_2 = regs[i].second();
947 if (!r_1->is_valid()) {
948 assert(!r_2->is_valid(), "");
949 continue;
950 }
951 if (r_1->is_stack()) {
952 Register tmp_reg = value_regs[value_regs_index];
953 value_regs_index = (value_regs_index + 1) % num_value_regs;
954 // The calling convention produces OptoRegs that ignore the out
955 // preserve area (JIT's ABI). We must account for it here.
956 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
957 if (!r_2->is_valid()) {
958 __ lwz(tmp_reg, ld_off, sender_SP);
959 } else {
960 __ ld(tmp_reg, ld_off, sender_SP);
961 }
962 // Pretend stack targets were loaded into tmp_reg.
963 r_1 = tmp_reg->as_VMReg();
964 }
965
966 if (r_1->is_Register()) {
967 Register r = r_1->as_Register();
968 if (!r_2->is_valid()) {
969 __ stw(r, st_off, R1_SP);
970 st_off-=wordSize;
971 } else {
972 // Longs are given 2 64-bit slots in the interpreter, but the
973 // data is passed in only 1 slot.
974 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
975 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
976 st_off-=wordSize;
977 }
978 __ std(r, st_off, R1_SP);
979 st_off-=wordSize;
980 }
981 } else {
982 assert(r_1->is_FloatRegister(), "");
983 FloatRegister f = r_1->as_FloatRegister();
984 if (!r_2->is_valid()) {
985 __ stfs(f, st_off, R1_SP);
986 st_off-=wordSize;
987 } else {
988 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
989 // data is passed in only 1 slot.
990 // One of these should get known junk...
991 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
992 st_off-=wordSize;
993 __ stfd(f, st_off, R1_SP);
994 st_off-=wordSize;
995 }
996 }
997 }
998
999 // Jump to the interpreter just as if interpreter was doing it.
1000
1001 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1002
1003 // load TOS
1004 __ addi(R15_esp, R1_SP, st_off);
1005
1006 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
1007 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
1008 __ bctr();
1009
1010 return c2i_entrypoint;
1011 }
1012
1013 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
1014 int total_args_passed,
1015 int comp_args_on_stack,
1016 const BasicType *sig_bt,
1017 const VMRegPair *regs) {
1018
1019 // Load method's entry-point from method.
1020 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
1021 __ mtctr(R12_scratch2);
1022
1023 // We will only enter here from an interpreted frame and never from after
1024 // passing thru a c2i. Azul allowed this but we do not. If we lose the
1025 // race and use a c2i we will remain interpreted for the race loser(s).
1026 // This removes all sorts of headaches on the x86 side and also eliminates
1027 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
1028
1029 // Note: r13 contains the senderSP on entry. We must preserve it since
1030 // we may do a i2c -> c2i transition if we lose a race where compiled
1031 // code goes non-entrant while we get args ready.
1032 // In addition we use r13 to locate all the interpreter args as
1033 // we must align the stack to 16 bytes on an i2c entry else we
1034 // lose alignment we expect in all compiled code and register
1035 // save code can segv when fxsave instructions find improperly
1036 // aligned stack pointer.
1037
1038 const Register ld_ptr = R15_esp;
1039 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
1040 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
1041 int value_regs_index = 0;
1042
1043 int ld_offset = total_args_passed*wordSize;
1044
1045 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1046 // in registers, we will occasionally have no stack args.
1047 int comp_words_on_stack = 0;
1048 if (comp_args_on_stack) {
1049 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1050 // registers are below. By subtracting stack0, we either get a negative
1051 // number (all values in registers) or the maximum stack slot accessed.
1052
1053 // Convert 4-byte c2 stack slots to words.
1054 comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1055 // Round up to miminum stack alignment, in wordSize.
1056 comp_words_on_stack = align_up(comp_words_on_stack, 2);
1057 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1058 }
1059
1060 // Now generate the shuffle code. Pick up all register args and move the
1061 // rest through register value=Z_R12.
1062 BLOCK_COMMENT("Shuffle arguments");
1063 for (int i = 0; i < total_args_passed; i++) {
1064 if (sig_bt[i] == T_VOID) {
1065 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1066 continue;
1067 }
1068
1069 // Pick up 0, 1 or 2 words from ld_ptr.
1070 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1071 "scrambled load targets?");
1072 VMReg r_1 = regs[i].first();
1073 VMReg r_2 = regs[i].second();
1074 if (!r_1->is_valid()) {
1075 assert(!r_2->is_valid(), "");
1076 continue;
1077 }
1078 if (r_1->is_FloatRegister()) {
1079 if (!r_2->is_valid()) {
1080 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1081 ld_offset-=wordSize;
1082 } else {
1083 // Skip the unused interpreter slot.
1084 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1085 ld_offset-=2*wordSize;
1086 }
1087 } else {
1088 Register r;
1089 if (r_1->is_stack()) {
1090 // Must do a memory to memory move thru "value".
1091 r = value_regs[value_regs_index];
1092 value_regs_index = (value_regs_index + 1) % num_value_regs;
1093 } else {
1094 r = r_1->as_Register();
1095 }
1096 if (!r_2->is_valid()) {
1097 // Not sure we need to do this but it shouldn't hurt.
1098 if (is_reference_type(sig_bt[i]) || sig_bt[i] == T_ADDRESS) {
1099 __ ld(r, ld_offset, ld_ptr);
1100 ld_offset-=wordSize;
1101 } else {
1102 __ lwz(r, ld_offset, ld_ptr);
1103 ld_offset-=wordSize;
1104 }
1105 } else {
1106 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1107 // data is passed in only 1 slot.
1108 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1109 ld_offset-=wordSize;
1110 }
1111 __ ld(r, ld_offset, ld_ptr);
1112 ld_offset-=wordSize;
1113 }
1114
1115 if (r_1->is_stack()) {
1116 // Now store value where the compiler expects it
1117 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1118
1119 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN ||
1120 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) {
1121 __ stw(r, st_off, R1_SP);
1122 } else {
1123 __ std(r, st_off, R1_SP);
1124 }
1125 }
1126 }
1127 }
1128
1129 __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1130
1131 BLOCK_COMMENT("Store method");
1132 // Store method into thread->callee_target.
1133 // We might end up in handle_wrong_method if the callee is
1134 // deoptimized as we race thru here. If that happens we don't want
1135 // to take a safepoint because the caller frame will look
1136 // interpreted and arguments are now "compiled" so it is much better
1137 // to make this transition invisible to the stack walking
1138 // code. Unfortunately if we try and find the callee by normal means
1139 // a safepoint is possible. So we stash the desired callee in the
1140 // thread and the vm will find there should this case occur.
1141 __ std(R19_method, thread_(callee_target));
1142
1143 // Jump to the compiled code just as if compiled code was doing it.
1144 __ bctr();
1145 }
1146
1147 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1148 int total_args_passed,
1149 int comp_args_on_stack,
1150 const BasicType *sig_bt,
1151 const VMRegPair *regs,
1152 AdapterFingerPrint* fingerprint) {
1153 address i2c_entry;
1154 address c2i_unverified_entry;
1155 address c2i_entry;
1156
1157
1158 // entry: i2c
1159
1160 __ align(CodeEntryAlignment);
1161 i2c_entry = __ pc();
1162 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1163
1164
1165 // entry: c2i unverified
1166
1167 __ align(CodeEntryAlignment);
1168 BLOCK_COMMENT("c2i unverified entry");
1169 c2i_unverified_entry = __ pc();
1170
1171 // inline_cache contains a CompiledICData
1172 const Register ic = R19_inline_cache_reg;
1173 const Register ic_klass = R11_scratch1;
1174 const Register receiver_klass = R12_scratch2;
1175 const Register code = R21_tmp1;
1176 const Register ientry = R23_tmp3;
1177
1178 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1179 assert(R11_scratch1 == R11, "need prologue scratch register");
1180
1181 Label call_interpreter;
1182
1183 __ ic_check(4 /* end_alignment */);
1184 __ ld(R19_method, CompiledICData::speculated_method_offset(), ic);
1185 // Argument is valid and klass is as expected, continue.
1186
1187 __ ld(code, method_(code));
1188 __ cmpdi(CCR0, code, 0);
1189 __ ld(ientry, method_(interpreter_entry)); // preloaded
1190 __ beq_predict_taken(CCR0, call_interpreter);
1191
1192 // Branch to ic_miss_stub.
1193 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1194
1195 // entry: c2i
1196
1197 c2i_entry = __ pc();
1198
1199 // Class initialization barrier for static methods
1200 address c2i_no_clinit_check_entry = nullptr;
1201 if (VM_Version::supports_fast_class_init_checks()) {
1202 Label L_skip_barrier;
1203
1204 { // Bypass the barrier for non-static methods
1205 __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1206 __ andi_(R0, R0, JVM_ACC_STATIC);
1207 __ beq(CCR0, L_skip_barrier); // non-static
1208 }
1209
1210 Register klass = R11_scratch1;
1211 __ load_method_holder(klass, R19_method);
1212 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
1213
1214 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
1215 __ mtctr(klass);
1216 __ bctr();
1217
1218 __ bind(L_skip_barrier);
1219 c2i_no_clinit_check_entry = __ pc();
1220 }
1221
1222 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1223 bs->c2i_entry_barrier(masm, /* tmp register*/ ic_klass, /* tmp register*/ receiver_klass, /* tmp register*/ code);
1224
1225 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry);
1226
1227 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry,
1228 c2i_no_clinit_check_entry);
1229 }
1230
1231 // An oop arg. Must pass a handle not the oop itself.
1232 static void object_move(MacroAssembler* masm,
1233 int frame_size_in_slots,
1234 OopMap* oop_map, int oop_handle_offset,
1235 bool is_receiver, int* receiver_offset,
1236 VMRegPair src, VMRegPair dst,
1237 Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1238 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1239 "receiver has already been moved");
1240
1241 // We must pass a handle. First figure out the location we use as a handle.
1242
1243 if (src.first()->is_stack()) {
1244 // stack to stack or reg
1245
1246 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1247 Label skip;
1248 const int oop_slot_in_callers_frame = reg2slot(src.first());
1249
1250 guarantee(!is_receiver, "expecting receiver in register");
1251 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1252
1253 __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1254 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp);
1255 __ cmpdi(CCR0, r_temp_2, 0);
1256 __ bne(CCR0, skip);
1257 // Use a null handle if oop is null.
1258 __ li(r_handle, 0);
1259 __ bind(skip);
1260
1261 if (dst.first()->is_stack()) {
1262 // stack to stack
1263 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1264 } else {
1265 // stack to reg
1266 // Nothing to do, r_handle is already the dst register.
1267 }
1268 } else {
1269 // reg to stack or reg
1270 const Register r_oop = src.first()->as_Register();
1271 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1272 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1273 + oop_handle_offset; // in slots
1274 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1275 Label skip;
1276
1277 if (is_receiver) {
1278 *receiver_offset = oop_offset;
1279 }
1280 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1281
1282 __ std( r_oop, oop_offset, R1_SP);
1283 __ addi(r_handle, R1_SP, oop_offset);
1284
1285 __ cmpdi(CCR0, r_oop, 0);
1286 __ bne(CCR0, skip);
1287 // Use a null handle if oop is null.
1288 __ li(r_handle, 0);
1289 __ bind(skip);
1290
1291 if (dst.first()->is_stack()) {
1292 // reg to stack
1293 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1294 } else {
1295 // reg to reg
1296 // Nothing to do, r_handle is already the dst register.
1297 }
1298 }
1299 }
1300
1301 static void int_move(MacroAssembler*masm,
1302 VMRegPair src, VMRegPair dst,
1303 Register r_caller_sp, Register r_temp) {
1304 assert(src.first()->is_valid(), "incoming must be int");
1305 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1306
1307 if (src.first()->is_stack()) {
1308 if (dst.first()->is_stack()) {
1309 // stack to stack
1310 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1311 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1312 } else {
1313 // stack to reg
1314 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1315 }
1316 } else if (dst.first()->is_stack()) {
1317 // reg to stack
1318 __ extsw(r_temp, src.first()->as_Register());
1319 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1320 } else {
1321 // reg to reg
1322 __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1323 }
1324 }
1325
1326 static void long_move(MacroAssembler*masm,
1327 VMRegPair src, VMRegPair dst,
1328 Register r_caller_sp, Register r_temp) {
1329 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1330 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1331
1332 if (src.first()->is_stack()) {
1333 if (dst.first()->is_stack()) {
1334 // stack to stack
1335 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1336 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1337 } else {
1338 // stack to reg
1339 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1340 }
1341 } else if (dst.first()->is_stack()) {
1342 // reg to stack
1343 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1344 } else {
1345 // reg to reg
1346 if (dst.first()->as_Register() != src.first()->as_Register())
1347 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1348 }
1349 }
1350
1351 static void float_move(MacroAssembler*masm,
1352 VMRegPair src, VMRegPair dst,
1353 Register r_caller_sp, Register r_temp) {
1354 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1355 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1356
1357 if (src.first()->is_stack()) {
1358 if (dst.first()->is_stack()) {
1359 // stack to stack
1360 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1361 __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1362 } else {
1363 // stack to reg
1364 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1365 }
1366 } else if (dst.first()->is_stack()) {
1367 // reg to stack
1368 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1369 } else {
1370 // reg to reg
1371 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1372 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1373 }
1374 }
1375
1376 static void double_move(MacroAssembler*masm,
1377 VMRegPair src, VMRegPair dst,
1378 Register r_caller_sp, Register r_temp) {
1379 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1380 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1381
1382 if (src.first()->is_stack()) {
1383 if (dst.first()->is_stack()) {
1384 // stack to stack
1385 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1386 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1387 } else {
1388 // stack to reg
1389 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1390 }
1391 } else if (dst.first()->is_stack()) {
1392 // reg to stack
1393 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1394 } else {
1395 // reg to reg
1396 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1397 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1398 }
1399 }
1400
1401 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1402 switch (ret_type) {
1403 case T_BOOLEAN:
1404 case T_CHAR:
1405 case T_BYTE:
1406 case T_SHORT:
1407 case T_INT:
1408 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1409 break;
1410 case T_ARRAY:
1411 case T_OBJECT:
1412 case T_LONG:
1413 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1414 break;
1415 case T_FLOAT:
1416 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1417 break;
1418 case T_DOUBLE:
1419 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1420 break;
1421 case T_VOID:
1422 break;
1423 default:
1424 ShouldNotReachHere();
1425 break;
1426 }
1427 }
1428
1429 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1430 switch (ret_type) {
1431 case T_BOOLEAN:
1432 case T_CHAR:
1433 case T_BYTE:
1434 case T_SHORT:
1435 case T_INT:
1436 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1437 break;
1438 case T_ARRAY:
1439 case T_OBJECT:
1440 case T_LONG:
1441 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1442 break;
1443 case T_FLOAT:
1444 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1445 break;
1446 case T_DOUBLE:
1447 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1448 break;
1449 case T_VOID:
1450 break;
1451 default:
1452 ShouldNotReachHere();
1453 break;
1454 }
1455 }
1456
1457 static void verify_oop_args(MacroAssembler* masm,
1458 const methodHandle& method,
1459 const BasicType* sig_bt,
1460 const VMRegPair* regs) {
1461 Register temp_reg = R19_method; // not part of any compiled calling seq
1462 if (VerifyOops) {
1463 for (int i = 0; i < method->size_of_parameters(); i++) {
1464 if (is_reference_type(sig_bt[i])) {
1465 VMReg r = regs[i].first();
1466 assert(r->is_valid(), "bad oop arg");
1467 if (r->is_stack()) {
1468 __ ld(temp_reg, reg2offset(r), R1_SP);
1469 __ verify_oop(temp_reg, FILE_AND_LINE);
1470 } else {
1471 __ verify_oop(r->as_Register(), FILE_AND_LINE);
1472 }
1473 }
1474 }
1475 }
1476 }
1477
1478 static void gen_special_dispatch(MacroAssembler* masm,
1479 const methodHandle& method,
1480 const BasicType* sig_bt,
1481 const VMRegPair* regs) {
1482 verify_oop_args(masm, method, sig_bt, regs);
1483 vmIntrinsics::ID iid = method->intrinsic_id();
1484
1485 // Now write the args into the outgoing interpreter space
1486 bool has_receiver = false;
1487 Register receiver_reg = noreg;
1488 int member_arg_pos = -1;
1489 Register member_reg = noreg;
1490 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1491 if (ref_kind != 0) {
1492 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1493 member_reg = R19_method; // known to be free at this point
1494 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1495 } else if (iid == vmIntrinsics::_invokeBasic) {
1496 has_receiver = true;
1497 } else if (iid == vmIntrinsics::_linkToNative) {
1498 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1499 member_reg = R19_method; // known to be free at this point
1500 } else {
1501 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1502 }
1503
1504 if (member_reg != noreg) {
1505 // Load the member_arg into register, if necessary.
1506 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1507 VMReg r = regs[member_arg_pos].first();
1508 if (r->is_stack()) {
1509 __ ld(member_reg, reg2offset(r), R1_SP);
1510 } else {
1511 // no data motion is needed
1512 member_reg = r->as_Register();
1513 }
1514 }
1515
1516 if (has_receiver) {
1517 // Make sure the receiver is loaded into a register.
1518 assert(method->size_of_parameters() > 0, "oob");
1519 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1520 VMReg r = regs[0].first();
1521 assert(r->is_valid(), "bad receiver arg");
1522 if (r->is_stack()) {
1523 // Porting note: This assumes that compiled calling conventions always
1524 // pass the receiver oop in a register. If this is not true on some
1525 // platform, pick a temp and load the receiver from stack.
1526 fatal("receiver always in a register");
1527 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point?
1528 __ ld(receiver_reg, reg2offset(r), R1_SP);
1529 } else {
1530 // no data motion is needed
1531 receiver_reg = r->as_Register();
1532 }
1533 }
1534
1535 // Figure out which address we are really jumping to:
1536 MethodHandles::generate_method_handle_dispatch(masm, iid,
1537 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1538 }
1539
1540 //---------------------------- continuation_enter_setup ---------------------------
1541 //
1542 // Frame setup.
1543 //
1544 // Arguments:
1545 // None.
1546 //
1547 // Results:
1548 // R1_SP: pointer to blank ContinuationEntry in the pushed frame.
1549 //
1550 // Kills:
1551 // R0, R20
1552 //
1553 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& framesize_words) {
1554 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1555 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1556 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1557
1558 const int frame_size_in_bytes = (int)ContinuationEntry::size();
1559 assert(is_aligned(frame_size_in_bytes, frame::alignment_in_bytes), "alignment error");
1560
1561 framesize_words = frame_size_in_bytes / wordSize;
1562
1563 DEBUG_ONLY(__ block_comment("setup {"));
1564 // Save return pc and push entry frame
1565 const Register return_pc = R20;
1566 __ mflr(return_pc);
1567 __ std(return_pc, _abi0(lr), R1_SP); // SP->lr = return_pc
1568 __ push_frame(frame_size_in_bytes , R0); // SP -= frame_size_in_bytes
1569
1570 OopMap* map = new OopMap((int)frame_size_in_bytes / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1571
1572 __ ld_ptr(R0, JavaThread::cont_entry_offset(), R16_thread);
1573 __ st_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1574 __ st_ptr(R0, ContinuationEntry::parent_offset(), R1_SP);
1575 DEBUG_ONLY(__ block_comment("} setup"));
1576
1577 return map;
1578 }
1579
1580 //---------------------------- fill_continuation_entry ---------------------------
1581 //
1582 // Initialize the new ContinuationEntry.
1583 //
1584 // Arguments:
1585 // R1_SP: pointer to blank Continuation entry
1586 // reg_cont_obj: pointer to the continuation
1587 // reg_flags: flags
1588 //
1589 // Results:
1590 // R1_SP: pointer to filled out ContinuationEntry
1591 //
1592 // Kills:
1593 // R8_ARG6, R9_ARG7, R10_ARG8
1594 //
1595 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1596 assert_different_registers(reg_cont_obj, reg_flags);
1597 Register zero = R8_ARG6;
1598 Register tmp2 = R9_ARG7;
1599 Register tmp3 = R10_ARG8;
1600
1601 DEBUG_ONLY(__ block_comment("fill {"));
1602 #ifdef ASSERT
1603 __ load_const_optimized(tmp2, ContinuationEntry::cookie_value());
1604 __ stw(tmp2, in_bytes(ContinuationEntry::cookie_offset()), R1_SP);
1605 __ std(tmp2, _abi0(cr), R1_SP);
1606 #endif //ASSERT
1607
1608 __ li(zero, 0);
1609 __ st_ptr(reg_cont_obj, ContinuationEntry::cont_offset(), R1_SP);
1610 __ stw(reg_flags, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1611 __ st_ptr(zero, ContinuationEntry::chunk_offset(), R1_SP);
1612 __ stw(zero, in_bytes(ContinuationEntry::argsize_offset()), R1_SP);
1613 __ stw(zero, in_bytes(ContinuationEntry::pin_count_offset()), R1_SP);
1614
1615 __ ld_ptr(tmp2, JavaThread::cont_fastpath_offset(), R16_thread);
1616 __ ld(tmp3, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1617 __ st_ptr(tmp2, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1618 __ std(tmp3, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1619
1620 __ st_ptr(zero, JavaThread::cont_fastpath_offset(), R16_thread);
1621 __ std(zero, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1622 DEBUG_ONLY(__ block_comment("} fill"));
1623 }
1624
1625 //---------------------------- continuation_enter_cleanup ---------------------------
1626 //
1627 // Copy corresponding attributes from the top ContinuationEntry to the JavaThread
1628 // before deleting it.
1629 //
1630 // Arguments:
1631 // R1_SP: pointer to the ContinuationEntry
1632 //
1633 // Results:
1634 // None.
1635 //
1636 // Kills:
1637 // R8_ARG6, R9_ARG7, R10_ARG8, R15_esp
1638 //
1639 static void continuation_enter_cleanup(MacroAssembler* masm) {
1640 Register tmp1 = R8_ARG6;
1641 Register tmp2 = R9_ARG7;
1642 Register tmp3 = R10_ARG8;
1643
1644 #ifdef ASSERT
1645 __ block_comment("clean {");
1646 __ ld_ptr(tmp1, JavaThread::cont_entry_offset(), R16_thread);
1647 __ cmpd(CCR0, R1_SP, tmp1);
1648 __ asm_assert_eq(FILE_AND_LINE ": incorrect R1_SP");
1649 __ load_const_optimized(tmp1, ContinuationEntry::cookie_value());
1650 __ ld(tmp2, _abi0(cr), R1_SP);
1651 __ cmpd(CCR0, tmp1, tmp2);
1652 __ asm_assert_eq(FILE_AND_LINE ": cookie not found");
1653 #endif
1654
1655 __ ld_ptr(tmp1, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1656 __ st_ptr(tmp1, JavaThread::cont_fastpath_offset(), R16_thread);
1657
1658 if (CheckJNICalls) {
1659 // Check if this is a virtual thread continuation
1660 Label L_skip_vthread_code;
1661 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1662 __ cmpwi(CCR0, R0, 0);
1663 __ beq(CCR0, L_skip_vthread_code);
1664
1665 // If the held monitor count is > 0 and this vthread is terminating then
1666 // it failed to release a JNI monitor. So we issue the same log message
1667 // that JavaThread::exit does.
1668 __ ld(R0, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1669 __ cmpdi(CCR0, R0, 0);
1670 __ beq(CCR0, L_skip_vthread_code);
1671
1672 // Save return value potentially containing the exception oop
1673 Register ex_oop = R15_esp; // nonvolatile register
1674 __ mr(ex_oop, R3_RET);
1675 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1676 // Restore potental return value
1677 __ mr(R3_RET, ex_oop);
1678
1679 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1680 // on termination. The held count is implicitly zeroed below when we restore from
1681 // the parent held count (which has to be zero).
1682 __ li(tmp1, 0);
1683 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1684
1685 __ bind(L_skip_vthread_code);
1686 }
1687 #ifdef ASSERT
1688 else {
1689 // Check if this is a virtual thread continuation
1690 Label L_skip_vthread_code;
1691 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1692 __ cmpwi(CCR0, R0, 0);
1693 __ beq(CCR0, L_skip_vthread_code);
1694
1695 // See comment just above. If not checking JNI calls the JNI count is only
1696 // needed for assertion checking.
1697 __ li(tmp1, 0);
1698 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1699
1700 __ bind(L_skip_vthread_code);
1701 }
1702 #endif
1703
1704 __ ld(tmp2, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1705 __ ld_ptr(tmp3, ContinuationEntry::parent_offset(), R1_SP);
1706 __ std(tmp2, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1707 __ st_ptr(tmp3, JavaThread::cont_entry_offset(), R16_thread);
1708 DEBUG_ONLY(__ block_comment("} clean"));
1709 }
1710
1711 static void check_continuation_enter_argument(VMReg actual_vmreg,
1712 Register expected_reg,
1713 const char* name) {
1714 assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1715 assert(actual_vmreg->as_Register() == expected_reg,
1716 "%s is in unexpected register: %s instead of %s",
1717 name, actual_vmreg->as_Register()->name(), expected_reg->name());
1718 }
1719
1720 static void gen_continuation_enter(MacroAssembler* masm,
1721 const VMRegPair* regs,
1722 int& exception_offset,
1723 OopMapSet* oop_maps,
1724 int& frame_complete,
1725 int& framesize_words,
1726 int& interpreted_entry_offset,
1727 int& compiled_entry_offset) {
1728
1729 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1730 int pos_cont_obj = 0;
1731 int pos_is_cont = 1;
1732 int pos_is_virtual = 2;
1733
1734 // The platform-specific calling convention may present the arguments in various registers.
1735 // To simplify the rest of the code, we expect the arguments to reside at these known
1736 // registers, and we additionally check the placement here in case calling convention ever
1737 // changes.
1738 Register reg_cont_obj = R3_ARG1;
1739 Register reg_is_cont = R4_ARG2;
1740 Register reg_is_virtual = R5_ARG3;
1741
1742 check_continuation_enter_argument(regs[pos_cont_obj].first(), reg_cont_obj, "Continuation object");
1743 check_continuation_enter_argument(regs[pos_is_cont].first(), reg_is_cont, "isContinue");
1744 check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1745
1746 address resolve_static_call = SharedRuntime::get_resolve_static_call_stub();
1747
1748 address start = __ pc();
1749
1750 Label L_thaw, L_exit;
1751
1752 // i2i entry used at interp_only_mode only
1753 interpreted_entry_offset = __ pc() - start;
1754 {
1755 #ifdef ASSERT
1756 Label is_interp_only;
1757 __ lwz(R0, in_bytes(JavaThread::interp_only_mode_offset()), R16_thread);
1758 __ cmpwi(CCR0, R0, 0);
1759 __ bne(CCR0, is_interp_only);
1760 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1761 __ bind(is_interp_only);
1762 #endif
1763
1764 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1765 __ ld(reg_cont_obj, Interpreter::stackElementSize*3, R15_esp);
1766 __ lwz(reg_is_cont, Interpreter::stackElementSize*2, R15_esp);
1767 __ lwz(reg_is_virtual, Interpreter::stackElementSize*1, R15_esp);
1768
1769 __ push_cont_fastpath();
1770
1771 OopMap* map = continuation_enter_setup(masm, framesize_words);
1772
1773 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1774 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1775
1776 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1777
1778 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1779 __ cmpwi(CCR0, reg_is_cont, 0);
1780 __ bne(CCR0, L_thaw);
1781
1782 // --- call Continuation.enter(Continuation c, boolean isContinue)
1783
1784 // Emit compiled static call. The call will be always resolved to the c2i
1785 // entry of Continuation.enter(Continuation c, boolean isContinue).
1786 // There are special cases in SharedRuntime::resolve_static_call_C() and
1787 // SharedRuntime::resolve_sub_helper_internal() to achieve this
1788 // See also corresponding call below.
1789 address c2i_call_pc = __ pc();
1790 int start_offset = __ offset();
1791 // Put the entry point as a constant into the constant pool.
1792 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1793 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1794 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1795
1796 // Emit the trampoline stub which will be related to the branch-and-link below.
1797 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1798 guarantee(stub != nullptr, "no space for trampoline stub");
1799
1800 __ relocate(relocInfo::static_call_type);
1801 // Note: At this point we do not have the address of the trampoline
1802 // stub, and the entry point might be too far away for bl, so __ pc()
1803 // serves as dummy and the bl will be patched later.
1804 __ bl(__ pc());
1805 oop_maps->add_gc_map(__ pc() - start, map);
1806 __ post_call_nop();
1807
1808 __ b(L_exit);
1809
1810 // static stub for the call above
1811 stub = CompiledDirectCall::emit_to_interp_stub(masm, c2i_call_pc);
1812 guarantee(stub != nullptr, "no space for static stub");
1813 }
1814
1815 // compiled entry
1816 __ align(CodeEntryAlignment);
1817 compiled_entry_offset = __ pc() - start;
1818
1819 OopMap* map = continuation_enter_setup(masm, framesize_words);
1820
1821 // Frame is now completed as far as size and linkage.
1822 frame_complete =__ pc() - start;
1823
1824 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1825
1826 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1827 __ cmpwi(CCR0, reg_is_cont, 0);
1828 __ bne(CCR0, L_thaw);
1829
1830 // --- call Continuation.enter(Continuation c, boolean isContinue)
1831
1832 // Emit compiled static call
1833 // The call needs to be resolved. There's a special case for this in
1834 // SharedRuntime::find_callee_info_helper() which calls
1835 // LinkResolver::resolve_continuation_enter() which resolves the call to
1836 // Continuation.enter(Continuation c, boolean isContinue).
1837 address call_pc = __ pc();
1838 int start_offset = __ offset();
1839 // Put the entry point as a constant into the constant pool.
1840 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1841 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1842 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1843
1844 // Emit the trampoline stub which will be related to the branch-and-link below.
1845 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1846 guarantee(stub != nullptr, "no space for trampoline stub");
1847
1848 __ relocate(relocInfo::static_call_type);
1849 // Note: At this point we do not have the address of the trampoline
1850 // stub, and the entry point might be too far away for bl, so __ pc()
1851 // serves as dummy and the bl will be patched later.
1852 __ bl(__ pc());
1853 oop_maps->add_gc_map(__ pc() - start, map);
1854 __ post_call_nop();
1855
1856 __ b(L_exit);
1857
1858 // --- Thawing path
1859
1860 __ bind(L_thaw);
1861 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1862 __ add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(StubRoutines::cont_thaw()));
1863 __ mtctr(R0);
1864 __ bctrl();
1865 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1866 ContinuationEntry::_return_pc_offset = __ pc() - start;
1867 __ post_call_nop();
1868
1869 // --- Normal exit (resolve/thawing)
1870
1871 __ bind(L_exit);
1872 ContinuationEntry::_cleanup_offset = __ pc() - start;
1873 continuation_enter_cleanup(masm);
1874
1875 // Pop frame and return
1876 DEBUG_ONLY(__ ld_ptr(R0, 0, R1_SP));
1877 __ addi(R1_SP, R1_SP, framesize_words*wordSize);
1878 DEBUG_ONLY(__ cmpd(CCR0, R0, R1_SP));
1879 __ asm_assert_eq(FILE_AND_LINE ": inconsistent frame size");
1880 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1881 __ mtlr(R0);
1882 __ blr();
1883
1884 // --- Exception handling path
1885
1886 exception_offset = __ pc() - start;
1887
1888 continuation_enter_cleanup(masm);
1889 Register ex_pc = R17_tos; // nonvolatile register
1890 Register ex_oop = R15_esp; // nonvolatile register
1891 __ ld(ex_pc, _abi0(callers_sp), R1_SP); // Load caller's return pc
1892 __ ld(ex_pc, _abi0(lr), ex_pc);
1893 __ mr(ex_oop, R3_RET); // save return value containing the exception oop
1894 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, ex_pc);
1895 __ mtlr(R3_RET); // the exception handler
1896 __ ld(R1_SP, _abi0(callers_sp), R1_SP); // remove enterSpecial frame
1897
1898 // Continue at exception handler
1899 // See OptoRuntime::generate_exception_blob for register arguments
1900 __ mr(R3_ARG1, ex_oop); // pass exception oop
1901 __ mr(R4_ARG2, ex_pc); // pass exception pc
1902 __ blr();
1903
1904 // static stub for the call above
1905 stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1906 guarantee(stub != nullptr, "no space for static stub");
1907 }
1908
1909 static void gen_continuation_yield(MacroAssembler* masm,
1910 const VMRegPair* regs,
1911 OopMapSet* oop_maps,
1912 int& frame_complete,
1913 int& framesize_words,
1914 int& compiled_entry_offset) {
1915 Register tmp = R10_ARG8;
1916
1917 const int framesize_bytes = (int)align_up((int)frame::native_abi_reg_args_size, frame::alignment_in_bytes);
1918 framesize_words = framesize_bytes / wordSize;
1919
1920 address start = __ pc();
1921 compiled_entry_offset = __ pc() - start;
1922
1923 // Save return pc and push entry frame
1924 __ mflr(tmp);
1925 __ std(tmp, _abi0(lr), R1_SP); // SP->lr = return_pc
1926 __ push_frame(framesize_bytes , R0); // SP -= frame_size_in_bytes
1927
1928 DEBUG_ONLY(__ block_comment("Frame Complete"));
1929 frame_complete = __ pc() - start;
1930 address last_java_pc = __ pc();
1931
1932 // This nop must be exactly at the PC we push into the frame info.
1933 // We use this nop for fast CodeBlob lookup, associate the OopMap
1934 // with it right away.
1935 __ post_call_nop();
1936 OopMap* map = new OopMap(framesize_bytes / VMRegImpl::stack_slot_size, 1);
1937 oop_maps->add_gc_map(last_java_pc - start, map);
1938
1939 __ calculate_address_from_global_toc(tmp, last_java_pc); // will be relocated
1940 __ set_last_Java_frame(R1_SP, tmp);
1941 __ call_VM_leaf(Continuation::freeze_entry(), R16_thread, R1_SP);
1942 __ reset_last_Java_frame();
1943
1944 Label L_pinned;
1945
1946 __ cmpwi(CCR0, R3_RET, 0);
1947 __ bne(CCR0, L_pinned);
1948
1949 // yield succeeded
1950
1951 // Pop frames of continuation including this stub's frame
1952 __ ld_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1953 // The frame pushed by gen_continuation_enter is on top now again
1954 continuation_enter_cleanup(masm);
1955
1956 // Pop frame and return
1957 Label L_return;
1958 __ bind(L_return);
1959 __ pop_frame();
1960 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1961 __ mtlr(R0);
1962 __ blr();
1963
1964 // yield failed - continuation is pinned
1965
1966 __ bind(L_pinned);
1967
1968 // handle pending exception thrown by freeze
1969 __ ld(tmp, in_bytes(JavaThread::pending_exception_offset()), R16_thread);
1970 __ cmpdi(CCR0, tmp, 0);
1971 __ beq(CCR0, L_return); // return if no exception is pending
1972 __ pop_frame();
1973 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1974 __ mtlr(R0);
1975 __ load_const_optimized(tmp, StubRoutines::forward_exception_entry(), R0);
1976 __ mtctr(tmp);
1977 __ bctr();
1978 }
1979
1980 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1981 ::continuation_enter_cleanup(masm);
1982 }
1983
1984 // ---------------------------------------------------------------------------
1985 // Generate a native wrapper for a given method. The method takes arguments
1986 // in the Java compiled code convention, marshals them to the native
1987 // convention (handlizes oops, etc), transitions to native, makes the call,
1988 // returns to java state (possibly blocking), unhandlizes any result and
1989 // returns.
1990 //
1991 // Critical native functions are a shorthand for the use of
1992 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1993 // functions. The wrapper is expected to unpack the arguments before
1994 // passing them to the callee. Critical native functions leave the state _in_Java,
1995 // since they cannot stop for GC.
1996 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1997 // block and the check for pending exceptions it's impossible for them
1998 // to be thrown.
1999 //
2000 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
2001 const methodHandle& method,
2002 int compile_id,
2003 BasicType *in_sig_bt,
2004 VMRegPair *in_regs,
2005 BasicType ret_type) {
2006 if (method->is_continuation_native_intrinsic()) {
2007 int exception_offset = -1;
2008 OopMapSet* oop_maps = new OopMapSet();
2009 int frame_complete = -1;
2010 int stack_slots = -1;
2011 int interpreted_entry_offset = -1;
2012 int vep_offset = -1;
2013 if (method->is_continuation_enter_intrinsic()) {
2014 gen_continuation_enter(masm,
2015 in_regs,
2016 exception_offset,
2017 oop_maps,
2018 frame_complete,
2019 stack_slots,
2020 interpreted_entry_offset,
2021 vep_offset);
2022 } else if (method->is_continuation_yield_intrinsic()) {
2023 gen_continuation_yield(masm,
2024 in_regs,
2025 oop_maps,
2026 frame_complete,
2027 stack_slots,
2028 vep_offset);
2029 } else {
2030 guarantee(false, "Unknown Continuation native intrinsic");
2031 }
2032
2033 #ifdef ASSERT
2034 if (method->is_continuation_enter_intrinsic()) {
2035 assert(interpreted_entry_offset != -1, "Must be set");
2036 assert(exception_offset != -1, "Must be set");
2037 } else {
2038 assert(interpreted_entry_offset == -1, "Must be unset");
2039 assert(exception_offset == -1, "Must be unset");
2040 }
2041 assert(frame_complete != -1, "Must be set");
2042 assert(stack_slots != -1, "Must be set");
2043 assert(vep_offset != -1, "Must be set");
2044 #endif
2045
2046 __ flush();
2047 nmethod* nm = nmethod::new_native_nmethod(method,
2048 compile_id,
2049 masm->code(),
2050 vep_offset,
2051 frame_complete,
2052 stack_slots,
2053 in_ByteSize(-1),
2054 in_ByteSize(-1),
2055 oop_maps,
2056 exception_offset);
2057 if (nm == nullptr) return nm;
2058 if (method->is_continuation_enter_intrinsic()) {
2059 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2060 } else if (method->is_continuation_yield_intrinsic()) {
2061 _cont_doYield_stub = nm;
2062 }
2063 return nm;
2064 }
2065
2066 if (method->is_method_handle_intrinsic()) {
2067 vmIntrinsics::ID iid = method->intrinsic_id();
2068 intptr_t start = (intptr_t)__ pc();
2069 int vep_offset = ((intptr_t)__ pc()) - start;
2070 gen_special_dispatch(masm,
2071 method,
2072 in_sig_bt,
2073 in_regs);
2074 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
2075 __ flush();
2076 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
2077 return nmethod::new_native_nmethod(method,
2078 compile_id,
2079 masm->code(),
2080 vep_offset,
2081 frame_complete,
2082 stack_slots / VMRegImpl::slots_per_word,
2083 in_ByteSize(-1),
2084 in_ByteSize(-1),
2085 (OopMapSet*)nullptr);
2086 }
2087
2088 address native_func = method->native_function();
2089 assert(native_func != nullptr, "must have function");
2090
2091 // First, create signature for outgoing C call
2092 // --------------------------------------------------------------------------
2093
2094 int total_in_args = method->size_of_parameters();
2095 // We have received a description of where all the java args are located
2096 // on entry to the wrapper. We need to convert these args to where
2097 // the jni function will expect them. To figure out where they go
2098 // we convert the java signature to a C signature by inserting
2099 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2100
2101 // Calculate the total number of C arguments and create arrays for the
2102 // signature and the outgoing registers.
2103 // On ppc64, we have two arrays for the outgoing registers, because
2104 // some floating-point arguments must be passed in registers _and_
2105 // in stack locations.
2106 bool method_is_static = method->is_static();
2107 int total_c_args = total_in_args + (method_is_static ? 2 : 1);
2108
2109 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2110 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2111 BasicType* in_elem_bt = nullptr;
2112
2113 // Create the signature for the C call:
2114 // 1) add the JNIEnv*
2115 // 2) add the class if the method is static
2116 // 3) copy the rest of the incoming signature (shifted by the number of
2117 // hidden arguments).
2118
2119 int argc = 0;
2120 out_sig_bt[argc++] = T_ADDRESS;
2121 if (method->is_static()) {
2122 out_sig_bt[argc++] = T_OBJECT;
2123 }
2124
2125 for (int i = 0; i < total_in_args ; i++ ) {
2126 out_sig_bt[argc++] = in_sig_bt[i];
2127 }
2128
2129
2130 // Compute the wrapper's frame size.
2131 // --------------------------------------------------------------------------
2132
2133 // Now figure out where the args must be stored and how much stack space
2134 // they require.
2135 //
2136 // Compute framesize for the wrapper. We need to handlize all oops in
2137 // incoming registers.
2138 //
2139 // Calculate the total number of stack slots we will need:
2140 // 1) abi requirements
2141 // 2) outgoing arguments
2142 // 3) space for inbound oop handle area
2143 // 4) space for handlizing a klass if static method
2144 // 5) space for a lock if synchronized method
2145 // 6) workspace for saving return values, int <-> float reg moves, etc.
2146 // 7) alignment
2147 //
2148 // Layout of the native wrapper frame:
2149 // (stack grows upwards, memory grows downwards)
2150 //
2151 // NW [ABI_REG_ARGS] <-- 1) R1_SP
2152 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset
2153 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset
2154 // klass <-- 4) R1_SP + klass_offset
2155 // lock <-- 5) R1_SP + lock_offset
2156 // [workspace] <-- 6) R1_SP + workspace_offset
2157 // [alignment] (optional) <-- 7)
2158 // caller [JIT_TOP_ABI_48] <-- r_callers_sp
2159 //
2160 // - *_slot_offset Indicates offset from SP in number of stack slots.
2161 // - *_offset Indicates offset from SP in bytes.
2162
2163 int stack_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args) + // 1+2)
2164 SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
2165
2166 // Now the space for the inbound oop handle area.
2167 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
2168
2169 int oop_handle_slot_offset = stack_slots;
2170 stack_slots += total_save_slots; // 3)
2171
2172 int klass_slot_offset = 0;
2173 int klass_offset = -1;
2174 if (method_is_static) { // 4)
2175 klass_slot_offset = stack_slots;
2176 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2177 stack_slots += VMRegImpl::slots_per_word;
2178 }
2179
2180 int lock_slot_offset = 0;
2181 int lock_offset = -1;
2182 if (method->is_synchronized()) { // 5)
2183 lock_slot_offset = stack_slots;
2184 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
2185 stack_slots += VMRegImpl::slots_per_word;
2186 }
2187
2188 int workspace_slot_offset = stack_slots; // 6)
2189 stack_slots += 2;
2190
2191 // Now compute actual number of stack words we need.
2192 // Rounding to make stack properly aligned.
2193 stack_slots = align_up(stack_slots, // 7)
2194 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
2195 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
2196
2197
2198 // Now we can start generating code.
2199 // --------------------------------------------------------------------------
2200
2201 intptr_t start_pc = (intptr_t)__ pc();
2202 intptr_t vep_start_pc;
2203 intptr_t frame_done_pc;
2204
2205 Label handle_pending_exception;
2206 Label last_java_pc;
2207
2208 Register r_callers_sp = R21;
2209 Register r_temp_1 = R22;
2210 Register r_temp_2 = R23;
2211 Register r_temp_3 = R24;
2212 Register r_temp_4 = R25;
2213 Register r_temp_5 = R26;
2214 Register r_temp_6 = R27;
2215 Register r_last_java_pc = R28;
2216
2217 Register r_carg1_jnienv = noreg;
2218 Register r_carg2_classorobject = noreg;
2219 r_carg1_jnienv = out_regs[0].first()->as_Register();
2220 r_carg2_classorobject = out_regs[1].first()->as_Register();
2221
2222
2223 // Generate the Unverified Entry Point (UEP).
2224 // --------------------------------------------------------------------------
2225 assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
2226
2227 // Check ic: object class == cached class?
2228 if (!method_is_static) {
2229 __ ic_check(4 /* end_alignment */);
2230 }
2231
2232 // Generate the Verified Entry Point (VEP).
2233 // --------------------------------------------------------------------------
2234 vep_start_pc = (intptr_t)__ pc();
2235
2236 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
2237 Label L_skip_barrier;
2238 Register klass = r_temp_1;
2239 // Notify OOP recorder (don't need the relocation)
2240 AddressLiteral md = __ constant_metadata_address(method->method_holder());
2241 __ load_const_optimized(klass, md.value(), R0);
2242 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
2243
2244 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
2245 __ mtctr(klass);
2246 __ bctr();
2247
2248 __ bind(L_skip_barrier);
2249 }
2250
2251 __ save_LR(r_temp_1);
2252 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2253 __ mr(r_callers_sp, R1_SP); // Remember frame pointer.
2254 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame.
2255
2256 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2257 bs->nmethod_entry_barrier(masm, r_temp_1);
2258
2259 frame_done_pc = (intptr_t)__ pc();
2260
2261 // Native nmethod wrappers never take possession of the oop arguments.
2262 // So the caller will gc the arguments.
2263 // The only thing we need an oopMap for is if the call is static.
2264 //
2265 // An OopMap for lock (and class if static), and one for the VM call itself.
2266 OopMapSet *oop_maps = new OopMapSet();
2267 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2268
2269 // Move arguments from register/stack to register/stack.
2270 // --------------------------------------------------------------------------
2271 //
2272 // We immediately shuffle the arguments so that for any vm call we have
2273 // to make from here on out (sync slow path, jvmti, etc.) we will have
2274 // captured the oops from our caller and have a valid oopMap for them.
2275 //
2276 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2277 // (derived from JavaThread* which is in R16_thread) and, if static,
2278 // the class mirror instead of a receiver. This pretty much guarantees that
2279 // register layout will not match. We ignore these extra arguments during
2280 // the shuffle. The shuffle is described by the two calling convention
2281 // vectors we have in our possession. We simply walk the java vector to
2282 // get the source locations and the c vector to get the destinations.
2283
2284 // Record sp-based slot for receiver on stack for non-static methods.
2285 int receiver_offset = -1;
2286
2287 // We move the arguments backward because the floating point registers
2288 // destination will always be to a register with a greater or equal
2289 // register number or the stack.
2290 // in is the index of the incoming Java arguments
2291 // out is the index of the outgoing C arguments
2292
2293 #ifdef ASSERT
2294 bool reg_destroyed[Register::number_of_registers];
2295 bool freg_destroyed[FloatRegister::number_of_registers];
2296 for (int r = 0 ; r < Register::number_of_registers ; r++) {
2297 reg_destroyed[r] = false;
2298 }
2299 for (int f = 0 ; f < FloatRegister::number_of_registers ; f++) {
2300 freg_destroyed[f] = false;
2301 }
2302 #endif // ASSERT
2303
2304 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2305
2306 #ifdef ASSERT
2307 if (in_regs[in].first()->is_Register()) {
2308 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2309 } else if (in_regs[in].first()->is_FloatRegister()) {
2310 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2311 }
2312 if (out_regs[out].first()->is_Register()) {
2313 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2314 } else if (out_regs[out].first()->is_FloatRegister()) {
2315 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2316 }
2317 #endif // ASSERT
2318
2319 switch (in_sig_bt[in]) {
2320 case T_BOOLEAN:
2321 case T_CHAR:
2322 case T_BYTE:
2323 case T_SHORT:
2324 case T_INT:
2325 // Move int and do sign extension.
2326 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2327 break;
2328 case T_LONG:
2329 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2330 break;
2331 case T_ARRAY:
2332 case T_OBJECT:
2333 object_move(masm, stack_slots,
2334 oop_map, oop_handle_slot_offset,
2335 ((in == 0) && (!method_is_static)), &receiver_offset,
2336 in_regs[in], out_regs[out],
2337 r_callers_sp, r_temp_1, r_temp_2);
2338 break;
2339 case T_VOID:
2340 break;
2341 case T_FLOAT:
2342 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2343 break;
2344 case T_DOUBLE:
2345 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2346 break;
2347 case T_ADDRESS:
2348 fatal("found type (T_ADDRESS) in java args");
2349 break;
2350 default:
2351 ShouldNotReachHere();
2352 break;
2353 }
2354 }
2355
2356 // Pre-load a static method's oop into ARG2.
2357 // Used both by locking code and the normal JNI call code.
2358 if (method_is_static) {
2359 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2360 r_carg2_classorobject);
2361
2362 // Now handlize the static class mirror in carg2. It's known not-null.
2363 __ std(r_carg2_classorobject, klass_offset, R1_SP);
2364 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2365 __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2366 }
2367
2368 // Get JNIEnv* which is first argument to native.
2369 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2370
2371 // NOTE:
2372 //
2373 // We have all of the arguments setup at this point.
2374 // We MUST NOT touch any outgoing regs from this point on.
2375 // So if we must call out we must push a new frame.
2376
2377 // The last java pc will also be used as resume pc if this is the wrapper for wait0.
2378 // For this purpose the precise location matters but not for oopmap lookup.
2379 __ calculate_address_from_global_toc(r_last_java_pc, last_java_pc, true, true, true, true);
2380
2381 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2382 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2383
2384 # if 0
2385 // DTrace method entry
2386 # endif
2387
2388 // Lock a synchronized method.
2389 // --------------------------------------------------------------------------
2390
2391 if (method->is_synchronized()) {
2392 Register r_oop = r_temp_4;
2393 const Register r_box = r_temp_5;
2394 Label done, locked;
2395
2396 // Load the oop for the object or class. r_carg2_classorobject contains
2397 // either the handlized oop from the incoming arguments or the handlized
2398 // class mirror (if the method is static).
2399 __ ld(r_oop, 0, r_carg2_classorobject);
2400
2401 // Get the lock box slot's address.
2402 __ addi(r_box, R1_SP, lock_offset);
2403
2404 // Try fastpath for locking.
2405 if (LockingMode == LM_LIGHTWEIGHT) {
2406 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2407 Register r_temp_3_or_noreg = UseObjectMonitorTable ? r_temp_3 : noreg;
2408 __ compiler_fast_lock_lightweight_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3_or_noreg);
2409 } else {
2410 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2411 __ compiler_fast_lock_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2412 }
2413 __ beq(CCR0, locked);
2414
2415 // None of the above fast optimizations worked so we have to get into the
2416 // slow case of monitor enter. Inline a special case of call_VM that
2417 // disallows any pending_exception.
2418
2419 // Save argument registers and leave room for C-compatible ABI_REG_ARGS.
2420 int frame_size = frame::native_abi_reg_args_size + align_up(total_c_args * wordSize, frame::alignment_in_bytes);
2421 __ mr(R11_scratch1, R1_SP);
2422 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs);
2423
2424 // Do the call.
2425 __ set_last_Java_frame(R11_scratch1, r_last_java_pc);
2426 assert(r_last_java_pc->is_nonvolatile(), "r_last_java_pc needs to be preserved accross complete_monitor_locking_C call");
2427 // The following call will not be preempted.
2428 // push_cont_fastpath forces freeze slow path in case we try to preempt where we will pin the
2429 // vthread to the carrier (see FreezeBase::recurse_freeze_native_frame()).
2430 __ push_cont_fastpath();
2431 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2432 __ pop_cont_fastpath();
2433 __ reset_last_Java_frame();
2434
2435 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs);
2436
2437 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2438 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C");
2439
2440 __ bind(locked);
2441 }
2442
2443 __ set_last_Java_frame(R1_SP, r_last_java_pc);
2444
2445 // Publish thread state
2446 // --------------------------------------------------------------------------
2447
2448 // Transition from _thread_in_Java to _thread_in_native.
2449 __ li(R0, _thread_in_native);
2450 __ release();
2451 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2452 __ stw(R0, thread_(thread_state));
2453
2454
2455 // The JNI call
2456 // --------------------------------------------------------------------------
2457 __ call_c(native_func, relocInfo::runtime_call_type);
2458
2459
2460 // Now, we are back from the native code.
2461
2462
2463 // Unpack the native result.
2464 // --------------------------------------------------------------------------
2465
2466 // For int-types, we do any needed sign-extension required.
2467 // Care must be taken that the return values (R3_RET and F1_RET)
2468 // will survive any VM calls for blocking or unlocking.
2469 // An OOP result (handle) is done specially in the slow-path code.
2470
2471 switch (ret_type) {
2472 case T_VOID: break; // Nothing to do!
2473 case T_FLOAT: break; // Got it where we want it (unless slow-path).
2474 case T_DOUBLE: break; // Got it where we want it (unless slow-path).
2475 case T_LONG: break; // Got it where we want it (unless slow-path).
2476 case T_OBJECT: break; // Really a handle.
2477 // Cannot de-handlize until after reclaiming jvm_lock.
2478 case T_ARRAY: break;
2479
2480 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1)
2481 __ normalize_bool(R3_RET);
2482 break;
2483 }
2484 case T_BYTE: { // sign extension
2485 __ extsb(R3_RET, R3_RET);
2486 break;
2487 }
2488 case T_CHAR: { // unsigned result
2489 __ andi(R3_RET, R3_RET, 0xffff);
2490 break;
2491 }
2492 case T_SHORT: { // sign extension
2493 __ extsh(R3_RET, R3_RET);
2494 break;
2495 }
2496 case T_INT: // nothing to do
2497 break;
2498 default:
2499 ShouldNotReachHere();
2500 break;
2501 }
2502
2503 // Publish thread state
2504 // --------------------------------------------------------------------------
2505
2506 // Switch thread to "native transition" state before reading the
2507 // synchronization state. This additional state is necessary because reading
2508 // and testing the synchronization state is not atomic w.r.t. GC, as this
2509 // scenario demonstrates:
2510 // - Java thread A, in _thread_in_native state, loads _not_synchronized
2511 // and is preempted.
2512 // - VM thread changes sync state to synchronizing and suspends threads
2513 // for GC.
2514 // - Thread A is resumed to finish this native method, but doesn't block
2515 // here since it didn't see any synchronization in progress, and escapes.
2516
2517 // Transition from _thread_in_native to _thread_in_native_trans.
2518 __ li(R0, _thread_in_native_trans);
2519 __ release();
2520 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2521 __ stw(R0, thread_(thread_state));
2522
2523
2524 // Must we block?
2525 // --------------------------------------------------------------------------
2526
2527 // Block, if necessary, before resuming in _thread_in_Java state.
2528 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2529 {
2530 Label no_block, sync;
2531
2532 // Force this write out before the read below.
2533 if (!UseSystemMemoryBarrier) {
2534 __ fence();
2535 }
2536
2537 Register sync_state_addr = r_temp_4;
2538 Register sync_state = r_temp_5;
2539 Register suspend_flags = r_temp_6;
2540
2541 // No synchronization in progress nor yet synchronized
2542 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
2543 __ safepoint_poll(sync, sync_state, true /* at_return */, false /* in_nmethod */);
2544
2545 // Not suspended.
2546 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2547 __ lwz(suspend_flags, thread_(suspend_flags));
2548 __ cmpwi(CCR1, suspend_flags, 0);
2549 __ beq(CCR1, no_block);
2550
2551 // Block. Save any potential method result value before the operation and
2552 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2553 // lets us share the oopMap we used when we went native rather than create
2554 // a distinct one for this pc.
2555 __ bind(sync);
2556 __ isync();
2557
2558 address entry_point =
2559 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2560 save_native_result(masm, ret_type, workspace_slot_offset);
2561 __ call_VM_leaf(entry_point, R16_thread);
2562 restore_native_result(masm, ret_type, workspace_slot_offset);
2563
2564 __ bind(no_block);
2565
2566 // Publish thread state.
2567 // --------------------------------------------------------------------------
2568
2569 // Thread state is thread_in_native_trans. Any safepoint blocking has
2570 // already happened so we can now change state to _thread_in_Java.
2571
2572 // Transition from _thread_in_native_trans to _thread_in_Java.
2573 __ li(R0, _thread_in_Java);
2574 __ lwsync(); // Acquire safepoint and suspend state, release thread state.
2575 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2576 __ stw(R0, thread_(thread_state));
2577
2578 // Check preemption for Object.wait()
2579 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
2580 Label not_preempted;
2581 __ ld(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2582 __ cmpdi(CCR0, R0, 0);
2583 __ beq(CCR0, not_preempted);
2584 __ mtlr(R0);
2585 __ li(R0, 0);
2586 __ std(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2587 __ blr();
2588 __ bind(not_preempted);
2589 }
2590 __ bind(last_java_pc);
2591 // We use the same pc/oopMap repeatedly when we call out above.
2592 intptr_t oopmap_pc = (intptr_t) __ pc();
2593 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2594 }
2595
2596 // Reguard any pages if necessary.
2597 // --------------------------------------------------------------------------
2598
2599 Label no_reguard;
2600 __ lwz(r_temp_1, thread_(stack_guard_state));
2601 __ cmpwi(CCR0, r_temp_1, StackOverflow::stack_guard_yellow_reserved_disabled);
2602 __ bne(CCR0, no_reguard);
2603
2604 save_native_result(masm, ret_type, workspace_slot_offset);
2605 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2606 restore_native_result(masm, ret_type, workspace_slot_offset);
2607
2608 __ bind(no_reguard);
2609
2610
2611 // Unlock
2612 // --------------------------------------------------------------------------
2613
2614 if (method->is_synchronized()) {
2615 const Register r_oop = r_temp_4;
2616 const Register r_box = r_temp_5;
2617 const Register r_exception = r_temp_6;
2618 Label done;
2619
2620 // Get oop and address of lock object box.
2621 if (method_is_static) {
2622 assert(klass_offset != -1, "");
2623 __ ld(r_oop, klass_offset, R1_SP);
2624 } else {
2625 assert(receiver_offset != -1, "");
2626 __ ld(r_oop, receiver_offset, R1_SP);
2627 }
2628 __ addi(r_box, R1_SP, lock_offset);
2629
2630 // Try fastpath for unlocking.
2631 if (LockingMode == LM_LIGHTWEIGHT) {
2632 __ compiler_fast_unlock_lightweight_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2633 } else {
2634 __ compiler_fast_unlock_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2635 }
2636 __ beq(CCR0, done);
2637
2638 // Save and restore any potential method result value around the unlocking operation.
2639 save_native_result(masm, ret_type, workspace_slot_offset);
2640
2641 // Must save pending exception around the slow-path VM call. Since it's a
2642 // leaf call, the pending exception (if any) can be kept in a register.
2643 __ ld(r_exception, thread_(pending_exception));
2644 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2645 __ li(R0, 0);
2646 __ std(R0, thread_(pending_exception));
2647
2648 // Slow case of monitor enter.
2649 // Inline a special case of call_VM that disallows any pending_exception.
2650 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread).
2651 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread);
2652
2653 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2654 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C");
2655
2656 restore_native_result(masm, ret_type, workspace_slot_offset);
2657
2658 // Check_forward_pending_exception jump to forward_exception if any pending
2659 // exception is set. The forward_exception routine expects to see the
2660 // exception in pending_exception and not in a register. Kind of clumsy,
2661 // since all folks who branch to forward_exception must have tested
2662 // pending_exception first and hence have it in a register already.
2663 __ std(r_exception, thread_(pending_exception));
2664
2665 __ bind(done);
2666 }
2667
2668 # if 0
2669 // DTrace method exit
2670 # endif
2671
2672 // Clear "last Java frame" SP and PC.
2673 // --------------------------------------------------------------------------
2674
2675 // Last java frame won't be set if we're resuming after preemption
2676 bool maybe_preempted = LockingMode != LM_LEGACY && method->is_object_wait0();
2677 __ reset_last_Java_frame(!maybe_preempted /* check_last_java_sp */);
2678
2679 // Unbox oop result, e.g. JNIHandles::resolve value.
2680 // --------------------------------------------------------------------------
2681
2682 if (is_reference_type(ret_type)) {
2683 __ resolve_jobject(R3_RET, r_temp_1, r_temp_2, MacroAssembler::PRESERVATION_NONE);
2684 }
2685
2686 if (CheckJNICalls) {
2687 // clear_pending_jni_exception_check
2688 __ load_const_optimized(R0, 0L);
2689 __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
2690 }
2691
2692 // Reset handle block.
2693 // --------------------------------------------------------------------------
2694 __ ld(r_temp_1, thread_(active_handles));
2695 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2696 __ li(r_temp_2, 0);
2697 __ stw(r_temp_2, in_bytes(JNIHandleBlock::top_offset()), r_temp_1);
2698
2699
2700 // Check for pending exceptions.
2701 // --------------------------------------------------------------------------
2702 __ ld(r_temp_2, thread_(pending_exception));
2703 __ cmpdi(CCR0, r_temp_2, 0);
2704 __ bne(CCR0, handle_pending_exception);
2705
2706 // Return
2707 // --------------------------------------------------------------------------
2708
2709 __ pop_frame();
2710 __ restore_LR(R11);
2711 __ blr();
2712
2713
2714 // Handler for pending exceptions (out-of-line).
2715 // --------------------------------------------------------------------------
2716 // Since this is a native call, we know the proper exception handler
2717 // is the empty function. We just pop this frame and then jump to
2718 // forward_exception_entry.
2719 __ bind(handle_pending_exception);
2720
2721 __ pop_frame();
2722 __ restore_LR(R11);
2723 __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2724 relocInfo::runtime_call_type);
2725
2726 // Done.
2727 // --------------------------------------------------------------------------
2728
2729 __ flush();
2730
2731 nmethod *nm = nmethod::new_native_nmethod(method,
2732 compile_id,
2733 masm->code(),
2734 vep_start_pc-start_pc,
2735 frame_done_pc-start_pc,
2736 stack_slots / VMRegImpl::slots_per_word,
2737 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2738 in_ByteSize(lock_offset),
2739 oop_maps);
2740
2741 return nm;
2742 }
2743
2744 // This function returns the adjust size (in number of words) to a c2i adapter
2745 // activation for use during deoptimization.
2746 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2747 return align_up((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::frame_alignment_in_words);
2748 }
2749
2750 uint SharedRuntime::in_preserve_stack_slots() {
2751 return frame::jit_in_preserve_size / VMRegImpl::stack_slot_size;
2752 }
2753
2754 uint SharedRuntime::out_preserve_stack_slots() {
2755 #if defined(COMPILER1) || defined(COMPILER2)
2756 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2757 #else
2758 return 0;
2759 #endif
2760 }
2761
2762 VMReg SharedRuntime::thread_register() {
2763 // On PPC virtual threads don't save the JavaThread* in their context (e.g. C1 stub frames).
2764 ShouldNotCallThis();
2765 return nullptr;
2766 }
2767
2768 #if defined(COMPILER1) || defined(COMPILER2)
2769 // Frame generation for deopt and uncommon trap blobs.
2770 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2771 /* Read */
2772 Register unroll_block_reg,
2773 /* Update */
2774 Register frame_sizes_reg,
2775 Register number_of_frames_reg,
2776 Register pcs_reg,
2777 /* Invalidate */
2778 Register frame_size_reg,
2779 Register pc_reg) {
2780
2781 __ ld(pc_reg, 0, pcs_reg);
2782 __ ld(frame_size_reg, 0, frame_sizes_reg);
2783 __ std(pc_reg, _abi0(lr), R1_SP);
2784 __ push_frame(frame_size_reg, R0/*tmp*/);
2785 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP);
2786 __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2787 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2788 __ addi(pcs_reg, pcs_reg, wordSize);
2789 }
2790
2791 // Loop through the UnrollBlock info and create new frames.
2792 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2793 /* read */
2794 Register unroll_block_reg,
2795 /* invalidate */
2796 Register frame_sizes_reg,
2797 Register number_of_frames_reg,
2798 Register pcs_reg,
2799 Register frame_size_reg,
2800 Register pc_reg) {
2801 Label loop;
2802
2803 // _number_of_frames is of type int (deoptimization.hpp)
2804 __ lwa(number_of_frames_reg,
2805 in_bytes(Deoptimization::UnrollBlock::number_of_frames_offset()),
2806 unroll_block_reg);
2807 __ ld(pcs_reg,
2808 in_bytes(Deoptimization::UnrollBlock::frame_pcs_offset()),
2809 unroll_block_reg);
2810 __ ld(frame_sizes_reg,
2811 in_bytes(Deoptimization::UnrollBlock::frame_sizes_offset()),
2812 unroll_block_reg);
2813
2814 // stack: (caller_of_deoptee, ...).
2815
2816 // At this point we either have an interpreter frame or a compiled
2817 // frame on top of stack. If it is a compiled frame we push a new c2i
2818 // adapter here
2819
2820 // Memorize top-frame stack-pointer.
2821 __ mr(frame_size_reg/*old_sp*/, R1_SP);
2822
2823 // Resize interpreter top frame OR C2I adapter.
2824
2825 // At this moment, the top frame (which is the caller of the deoptee) is
2826 // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2827 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2828 // outgoing arguments.
2829 //
2830 // In order to push the interpreter frame for the deoptee, we need to
2831 // resize the top frame such that we are able to place the deoptee's
2832 // locals in the frame.
2833 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2834 // into a valid PARENT_IJAVA_FRAME_ABI.
2835
2836 __ lwa(R11_scratch1,
2837 in_bytes(Deoptimization::UnrollBlock::caller_adjustment_offset()),
2838 unroll_block_reg);
2839 __ neg(R11_scratch1, R11_scratch1);
2840
2841 // R11_scratch1 contains size of locals for frame resizing.
2842 // R12_scratch2 contains top frame's lr.
2843
2844 // Resize frame by complete frame size prevents TOC from being
2845 // overwritten by locals. A more stack space saving way would be
2846 // to copy the TOC to its location in the new abi.
2847 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2848
2849 // now, resize the frame
2850 __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2851
2852 // In the case where we have resized a c2i frame above, the optional
2853 // alignment below the locals has size 32 (why?).
2854 __ std(R12_scratch2, _abi0(lr), R1_SP);
2855
2856 // Initialize initial_caller_sp.
2857 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP);
2858
2859 #ifdef ASSERT
2860 // Make sure that there is at least one entry in the array.
2861 __ cmpdi(CCR0, number_of_frames_reg, 0);
2862 __ asm_assert_ne("array_size must be > 0");
2863 #endif
2864
2865 // Now push the new interpreter frames.
2866 //
2867 __ bind(loop);
2868 // Allocate a new frame, fill in the pc.
2869 push_skeleton_frame(masm, deopt,
2870 unroll_block_reg,
2871 frame_sizes_reg,
2872 number_of_frames_reg,
2873 pcs_reg,
2874 frame_size_reg,
2875 pc_reg);
2876 __ cmpdi(CCR0, number_of_frames_reg, 0);
2877 __ bne(CCR0, loop);
2878
2879 // Get the return address pointing into the frame manager.
2880 __ ld(R0, 0, pcs_reg);
2881 // Store it in the top interpreter frame.
2882 __ std(R0, _abi0(lr), R1_SP);
2883 // Initialize frame_manager_lr of interpreter top frame.
2884 }
2885 #endif
2886
2887 void SharedRuntime::generate_deopt_blob() {
2888 // Allocate space for the code
2889 ResourceMark rm;
2890 // Setup code generation tools
2891 const char* name = SharedRuntime::stub_name(SharedStubId::deopt_id);
2892 CodeBuffer buffer(name, 2048, 1024);
2893 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2894 Label exec_mode_initialized;
2895 int frame_size_in_words;
2896 OopMap* map = nullptr;
2897 OopMapSet *oop_maps = new OopMapSet();
2898
2899 // size of ABI112 plus spill slots for R3_RET and F1_RET.
2900 const int frame_size_in_bytes = frame::native_abi_reg_args_spill_size;
2901 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2902 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2903
2904 const Register exec_mode_reg = R21_tmp1;
2905
2906 const address start = __ pc();
2907
2908 #if defined(COMPILER1) || defined(COMPILER2)
2909 // --------------------------------------------------------------------------
2910 // Prolog for non exception case!
2911
2912 // We have been called from the deopt handler of the deoptee.
2913 //
2914 // deoptee:
2915 // ...
2916 // call X
2917 // ...
2918 // deopt_handler: call_deopt_stub
2919 // cur. return pc --> ...
2920 //
2921 // So currently SR_LR points behind the call in the deopt handler.
2922 // We adjust it such that it points to the start of the deopt handler.
2923 // The return_pc has been stored in the frame of the deoptee and
2924 // will replace the address of the deopt_handler in the call
2925 // to Deoptimization::fetch_unroll_info below.
2926 // We can't grab a free register here, because all registers may
2927 // contain live values, so let the RegisterSaver do the adjustment
2928 // of the return pc.
2929 const int return_pc_adjustment_no_exception = -MacroAssembler::bl64_patchable_size;
2930
2931 // Push the "unpack frame"
2932 // Save everything in sight.
2933 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2934 &first_frame_size_in_bytes,
2935 /*generate_oop_map=*/ true,
2936 return_pc_adjustment_no_exception,
2937 RegisterSaver::return_pc_is_lr);
2938 assert(map != nullptr, "OopMap must have been created");
2939
2940 __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
2941 // Save exec mode for unpack_frames.
2942 __ b(exec_mode_initialized);
2943
2944 // --------------------------------------------------------------------------
2945 // Prolog for exception case
2946
2947 // An exception is pending.
2948 // We have been called with a return (interpreter) or a jump (exception blob).
2949 //
2950 // - R3_ARG1: exception oop
2951 // - R4_ARG2: exception pc
2952
2953 int exception_offset = __ pc() - start;
2954
2955 BLOCK_COMMENT("Prolog for exception case");
2956
2957 // Store exception oop and pc in thread (location known to GC).
2958 // This is needed since the call to "fetch_unroll_info()" may safepoint.
2959 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2960 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2961 __ std(R4_ARG2, _abi0(lr), R1_SP);
2962
2963 // Vanilla deoptimization with an exception pending in exception_oop.
2964 int exception_in_tls_offset = __ pc() - start;
2965
2966 // Push the "unpack frame".
2967 // Save everything in sight.
2968 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2969 &first_frame_size_in_bytes,
2970 /*generate_oop_map=*/ false,
2971 /*return_pc_adjustment_exception=*/ 0,
2972 RegisterSaver::return_pc_is_pre_saved);
2973
2974 // Deopt during an exception. Save exec mode for unpack_frames.
2975 __ li(exec_mode_reg, Deoptimization::Unpack_exception);
2976
2977 // fall through
2978
2979 int reexecute_offset = 0;
2980 #ifdef COMPILER1
2981 __ b(exec_mode_initialized);
2982
2983 // Reexecute entry, similar to c2 uncommon trap
2984 reexecute_offset = __ pc() - start;
2985
2986 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2987 &first_frame_size_in_bytes,
2988 /*generate_oop_map=*/ false,
2989 /*return_pc_adjustment_reexecute=*/ 0,
2990 RegisterSaver::return_pc_is_pre_saved);
2991 __ li(exec_mode_reg, Deoptimization::Unpack_reexecute);
2992 #endif
2993
2994 // --------------------------------------------------------------------------
2995 __ BIND(exec_mode_initialized);
2996
2997 const Register unroll_block_reg = R22_tmp2;
2998
2999 // We need to set `last_Java_frame' because `fetch_unroll_info' will
3000 // call `last_Java_frame()'. The value of the pc in the frame is not
3001 // particularly important. It just needs to identify this blob.
3002 __ set_last_Java_frame(R1_SP, noreg);
3003
3004 // With EscapeAnalysis turned on, this call may safepoint!
3005 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg);
3006 address calls_return_pc = __ last_calls_return_pc();
3007 // Set an oopmap for the call site that describes all our saved registers.
3008 oop_maps->add_gc_map(calls_return_pc - start, map);
3009
3010 __ reset_last_Java_frame();
3011 // Save the return value.
3012 __ mr(unroll_block_reg, R3_RET);
3013
3014 // Restore only the result registers that have been saved
3015 // by save_volatile_registers(...).
3016 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes);
3017
3018 // reload the exec mode from the UnrollBlock (it might have changed)
3019 __ lwz(exec_mode_reg, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3020 // In excp_deopt_mode, restore and clear exception oop which we
3021 // stored in the thread during exception entry above. The exception
3022 // oop will be the return value of this stub.
3023 Label skip_restore_excp;
3024 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception);
3025 __ bne(CCR0, skip_restore_excp);
3026 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3027 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3028 __ li(R0, 0);
3029 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3030 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3031 __ BIND(skip_restore_excp);
3032
3033 __ pop_frame();
3034
3035 // stack: (deoptee, optional i2c, caller of deoptee, ...).
3036
3037 // pop the deoptee's frame
3038 __ pop_frame();
3039
3040 // stack: (caller_of_deoptee, ...).
3041
3042 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3043 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3044 // and the frame is effectively not resized.
3045 Register caller_sp = R23_tmp3;
3046 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3047 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3048
3049 // Loop through the `UnrollBlock' info and create interpreter frames.
3050 push_skeleton_frames(masm, true/*deopt*/,
3051 unroll_block_reg,
3052 R23_tmp3,
3053 R24_tmp4,
3054 R25_tmp5,
3055 R26_tmp6,
3056 R27_tmp7);
3057
3058 // stack: (skeletal interpreter frame, ..., optional skeletal
3059 // interpreter frame, optional c2i, caller of deoptee, ...).
3060
3061 // push an `unpack_frame' taking care of float / int return values.
3062 __ push_frame(frame_size_in_bytes, R0/*tmp*/);
3063
3064 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3065 // skeletal interpreter frame, optional c2i, caller of deoptee,
3066 // ...).
3067
3068 // Spill live volatile registers since we'll do a call.
3069 __ std( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3070 __ stfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3071
3072 // Let the unpacker layout information in the skeletal frames just
3073 // allocated.
3074 __ calculate_address_from_global_toc(R3_RET, calls_return_pc, true, true, true, true);
3075 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
3076 // This is a call to a LEAF method, so no oop map is required.
3077 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3078 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3079 __ reset_last_Java_frame();
3080
3081 // Restore the volatiles saved above.
3082 __ ld( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3083 __ lfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3084
3085 // Pop the unpack frame.
3086 __ pop_frame();
3087 __ restore_LR(R0);
3088
3089 // stack: (top interpreter frame, ..., optional interpreter frame,
3090 // optional c2i, caller of deoptee, ...).
3091
3092 // Initialize R14_state.
3093 __ restore_interpreter_state(R11_scratch1);
3094 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3095
3096 // Return to the interpreter entry point.
3097 __ blr();
3098 __ flush();
3099 #else // COMPILER2
3100 __ unimplemented("deopt blob needed only with compiler");
3101 int exception_offset = __ pc() - start;
3102 #endif // COMPILER2
3103
3104 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset,
3105 reexecute_offset, first_frame_size_in_bytes / wordSize);
3106 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3107 }
3108
3109 #ifdef COMPILER2
3110 void OptoRuntime::generate_uncommon_trap_blob() {
3111 // Allocate space for the code.
3112 ResourceMark rm;
3113 // Setup code generation tools.
3114 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
3115 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3116 address start = __ pc();
3117
3118 Register unroll_block_reg = R21_tmp1;
3119 Register klass_index_reg = R22_tmp2;
3120 Register unc_trap_reg = R23_tmp3;
3121 Register r_return_pc = R27_tmp7;
3122
3123 OopMapSet* oop_maps = new OopMapSet();
3124 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3125 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3126
3127 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3128
3129 // Push a dummy `unpack_frame' and call
3130 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3131 // vframe array and return the `UnrollBlock' information.
3132
3133 // Save LR to compiled frame.
3134 __ save_LR(R11_scratch1);
3135
3136 // Push an "uncommon_trap" frame.
3137 __ push_frame_reg_args(0, R11_scratch1);
3138
3139 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
3140
3141 // Set the `unpack_frame' as last_Java_frame.
3142 // `Deoptimization::uncommon_trap' expects it and considers its
3143 // sender frame as the deoptee frame.
3144 // Remember the offset of the instruction whose address will be
3145 // moved to R11_scratch1.
3146 address gc_map_pc = __ pc();
3147 __ calculate_address_from_global_toc(r_return_pc, gc_map_pc, true, true, true, true);
3148 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3149
3150 __ mr(klass_index_reg, R3);
3151 __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap);
3152 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
3153 R16_thread, klass_index_reg, R5_ARG3);
3154
3155 // Set an oopmap for the call site.
3156 oop_maps->add_gc_map(gc_map_pc - start, map);
3157
3158 __ reset_last_Java_frame();
3159
3160 // Pop the `unpack frame'.
3161 __ pop_frame();
3162
3163 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3164
3165 // Save the return value.
3166 __ mr(unroll_block_reg, R3_RET);
3167
3168 // Pop the uncommon_trap frame.
3169 __ pop_frame();
3170
3171 // stack: (caller_of_deoptee, ...).
3172
3173 #ifdef ASSERT
3174 __ lwz(R22_tmp2, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3175 __ cmpdi(CCR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap);
3176 __ asm_assert_eq("OptoRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
3177 #endif
3178
3179 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3180 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3181 // and the frame is effectively not resized.
3182 Register caller_sp = R23_tmp3;
3183 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3184 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3185
3186 // Allocate new interpreter frame(s) and possibly a c2i adapter
3187 // frame.
3188 push_skeleton_frames(masm, false/*deopt*/,
3189 unroll_block_reg,
3190 R22_tmp2,
3191 R23_tmp3,
3192 R24_tmp4,
3193 R25_tmp5,
3194 R26_tmp6);
3195
3196 // stack: (skeletal interpreter frame, ..., optional skeletal
3197 // interpreter frame, optional c2i, caller of deoptee, ...).
3198
3199 // Push a dummy `unpack_frame' taking care of float return values.
3200 // Call `Deoptimization::unpack_frames' to layout information in the
3201 // interpreter frames just created.
3202
3203 // Push a simple "unpack frame" here.
3204 __ push_frame_reg_args(0, R11_scratch1);
3205
3206 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3207 // skeletal interpreter frame, optional c2i, caller of deoptee,
3208 // ...).
3209
3210 // Set the "unpack_frame" as last_Java_frame.
3211 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3212
3213 // Indicate it is the uncommon trap case.
3214 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3215 // Let the unpacker layout information in the skeletal frames just
3216 // allocated.
3217 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3218 R16_thread, unc_trap_reg);
3219
3220 __ reset_last_Java_frame();
3221 // Pop the `unpack frame'.
3222 __ pop_frame();
3223 // Restore LR from top interpreter frame.
3224 __ restore_LR(R11_scratch1);
3225
3226 // stack: (top interpreter frame, ..., optional interpreter frame,
3227 // optional c2i, caller of deoptee, ...).
3228
3229 __ restore_interpreter_state(R11_scratch1);
3230 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3231
3232 // Return to the interpreter entry point.
3233 __ blr();
3234
3235 masm->flush();
3236
3237 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3238 }
3239 #endif // COMPILER2
3240
3241 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3242 SafepointBlob* SharedRuntime::generate_handler_blob(SharedStubId id, address call_ptr) {
3243 assert(StubRoutines::forward_exception_entry() != nullptr,
3244 "must be generated before");
3245 assert(is_polling_page_id(id), "expected a polling page stub id");
3246
3247 ResourceMark rm;
3248 OopMapSet *oop_maps = new OopMapSet();
3249 OopMap* map;
3250
3251 // Allocate space for the code. Setup code generation tools.
3252 const char* name = SharedRuntime::stub_name(id);
3253 CodeBuffer buffer(name, 2048, 1024);
3254 MacroAssembler* masm = new MacroAssembler(&buffer);
3255
3256 address start = __ pc();
3257 int frame_size_in_bytes = 0;
3258
3259 RegisterSaver::ReturnPCLocation return_pc_location;
3260 bool cause_return = (id == SharedStubId::polling_page_return_handler_id);
3261 if (cause_return) {
3262 // Nothing to do here. The frame has already been popped in MachEpilogNode.
3263 // Register LR already contains the return pc.
3264 return_pc_location = RegisterSaver::return_pc_is_pre_saved;
3265 } else {
3266 // Use thread()->saved_exception_pc() as return pc.
3267 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3268 }
3269
3270 bool save_vectors = (id == SharedStubId::polling_page_vectors_safepoint_handler_id);
3271
3272 // Save registers, fpu state, and flags. Set R31 = return pc.
3273 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3274 &frame_size_in_bytes,
3275 /*generate_oop_map=*/ true,
3276 /*return_pc_adjustment=*/0,
3277 return_pc_location, save_vectors);
3278
3279 // The following is basically a call_VM. However, we need the precise
3280 // address of the call in order to generate an oopmap. Hence, we do all the
3281 // work ourselves.
3282 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3283
3284 // The return address must always be correct so that the frame constructor
3285 // never sees an invalid pc.
3286
3287 // Do the call
3288 __ call_VM_leaf(call_ptr, R16_thread);
3289 address calls_return_pc = __ last_calls_return_pc();
3290
3291 // Set an oopmap for the call site. This oopmap will map all
3292 // oop-registers and debug-info registers as callee-saved. This
3293 // will allow deoptimization at this safepoint to find all possible
3294 // debug-info recordings, as well as let GC find all oops.
3295 oop_maps->add_gc_map(calls_return_pc - start, map);
3296
3297 Label noException;
3298
3299 // Clear the last Java frame.
3300 __ reset_last_Java_frame();
3301
3302 BLOCK_COMMENT(" Check pending exception.");
3303 const Register pending_exception = R0;
3304 __ ld(pending_exception, thread_(pending_exception));
3305 __ cmpdi(CCR0, pending_exception, 0);
3306 __ beq(CCR0, noException);
3307
3308 // Exception pending
3309 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3310 frame_size_in_bytes,
3311 /*restore_ctr=*/true, save_vectors);
3312
3313 BLOCK_COMMENT(" Jump to forward_exception_entry.");
3314 // Jump to forward_exception_entry, with the issuing PC in LR
3315 // so it looks like the original nmethod called forward_exception_entry.
3316 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3317
3318 // No exception case.
3319 __ BIND(noException);
3320
3321 if (!cause_return) {
3322 Label no_adjust;
3323 // If our stashed return pc was modified by the runtime we avoid touching it
3324 __ ld(R0, frame_size_in_bytes + _abi0(lr), R1_SP);
3325 __ cmpd(CCR0, R0, R31);
3326 __ bne(CCR0, no_adjust);
3327
3328 // Adjust return pc forward to step over the safepoint poll instruction
3329 __ addi(R31, R31, 4);
3330 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
3331
3332 __ bind(no_adjust);
3333 }
3334
3335 // Normal exit, restore registers and exit.
3336 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3337 frame_size_in_bytes,
3338 /*restore_ctr=*/true, save_vectors);
3339
3340 __ blr();
3341
3342 // Make sure all code is generated
3343 masm->flush();
3344
3345 // Fill-out other meta info
3346 // CodeBlob frame size is in words.
3347 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3348 }
3349
3350 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3351 //
3352 // Generate a stub that calls into the vm to find out the proper destination
3353 // of a java call. All the argument registers are live at this point
3354 // but since this is generic code we don't know what they are and the caller
3355 // must do any gc of the args.
3356 //
3357 RuntimeStub* SharedRuntime::generate_resolve_blob(SharedStubId id, address destination) {
3358 assert(is_resolve_id(id), "expected a resolve stub id");
3359
3360 // allocate space for the code
3361 ResourceMark rm;
3362
3363 const char* name = SharedRuntime::stub_name(id);
3364 CodeBuffer buffer(name, 1000, 512);
3365 MacroAssembler* masm = new MacroAssembler(&buffer);
3366
3367 int frame_size_in_bytes;
3368
3369 OopMapSet *oop_maps = new OopMapSet();
3370 OopMap* map = nullptr;
3371
3372 address start = __ pc();
3373
3374 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3375 &frame_size_in_bytes,
3376 /*generate_oop_map*/ true,
3377 /*return_pc_adjustment*/ 0,
3378 RegisterSaver::return_pc_is_lr);
3379
3380 // Use noreg as last_Java_pc, the return pc will be reconstructed
3381 // from the physical frame.
3382 __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3383
3384 int frame_complete = __ offset();
3385
3386 // Pass R19_method as 2nd (optional) argument, used by
3387 // counter_overflow_stub.
3388 __ call_VM_leaf(destination, R16_thread, R19_method);
3389 address calls_return_pc = __ last_calls_return_pc();
3390 // Set an oopmap for the call site.
3391 // We need this not only for callee-saved registers, but also for volatile
3392 // registers that the compiler might be keeping live across a safepoint.
3393 // Create the oopmap for the call's return pc.
3394 oop_maps->add_gc_map(calls_return_pc - start, map);
3395
3396 // R3_RET contains the address we are going to jump to assuming no exception got installed.
3397
3398 // clear last_Java_sp
3399 __ reset_last_Java_frame();
3400
3401 // Check for pending exceptions.
3402 BLOCK_COMMENT("Check for pending exceptions.");
3403 Label pending;
3404 __ ld(R11_scratch1, thread_(pending_exception));
3405 __ cmpdi(CCR0, R11_scratch1, 0);
3406 __ bne(CCR0, pending);
3407
3408 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3409
3410 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3411
3412 // Get the returned method.
3413 __ get_vm_result_2(R19_method);
3414
3415 __ bctr();
3416
3417
3418 // Pending exception after the safepoint.
3419 __ BIND(pending);
3420
3421 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3422
3423 // exception pending => remove activation and forward to exception handler
3424
3425 __ li(R11_scratch1, 0);
3426 __ ld(R3_ARG1, thread_(pending_exception));
3427 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread);
3428 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3429
3430 // -------------
3431 // Make sure all code is generated.
3432 masm->flush();
3433
3434 // return the blob
3435 // frame_size_words or bytes??
3436 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3437 oop_maps, true);
3438 }
3439
3440 // Continuation point for throwing of implicit exceptions that are
3441 // not handled in the current activation. Fabricates an exception
3442 // oop and initiates normal exception dispatching in this
3443 // frame. Only callee-saved registers are preserved (through the
3444 // normal register window / RegisterMap handling). If the compiler
3445 // needs all registers to be preserved between the fault point and
3446 // the exception handler then it must assume responsibility for that
3447 // in AbstractCompiler::continuation_for_implicit_null_exception or
3448 // continuation_for_implicit_division_by_zero_exception. All other
3449 // implicit exceptions (e.g., NullPointerException or
3450 // AbstractMethodError on entry) are either at call sites or
3451 // otherwise assume that stack unwinding will be initiated, so
3452 // caller saved registers were assumed volatile in the compiler.
3453 //
3454 // Note that we generate only this stub into a RuntimeStub, because
3455 // it needs to be properly traversed and ignored during GC, so we
3456 // change the meaning of the "__" macro within this method.
3457 //
3458 // Note: the routine set_pc_not_at_call_for_caller in
3459 // SharedRuntime.cpp requires that this code be generated into a
3460 // RuntimeStub.
3461 RuntimeStub* SharedRuntime::generate_throw_exception(SharedStubId id, address runtime_entry) {
3462 assert(is_throw_id(id), "expected a throw stub id");
3463
3464 const char* name = SharedRuntime::stub_name(id);
3465
3466 ResourceMark rm;
3467 const char* timer_msg = "SharedRuntime generate_throw_exception";
3468 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3469
3470 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
3471 MacroAssembler* masm = new MacroAssembler(&code);
3472
3473 OopMapSet* oop_maps = new OopMapSet();
3474 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3475 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3476
3477 address start = __ pc();
3478
3479 __ save_LR(R11_scratch1);
3480
3481 // Push a frame.
3482 __ push_frame_reg_args(0, R11_scratch1);
3483
3484 address frame_complete_pc = __ pc();
3485
3486 // Note that we always have a runtime stub frame on the top of
3487 // stack by this point. Remember the offset of the instruction
3488 // whose address will be moved to R11_scratch1.
3489 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
3490
3491 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
3492
3493 __ mr(R3_ARG1, R16_thread);
3494 __ call_c(runtime_entry);
3495
3496 // Set an oopmap for the call site.
3497 oop_maps->add_gc_map((int)(gc_map_pc - start), map);
3498
3499 __ reset_last_Java_frame();
3500
3501 #ifdef ASSERT
3502 // Make sure that this code is only executed if there is a pending
3503 // exception.
3504 {
3505 Label L;
3506 __ ld(R0,
3507 in_bytes(Thread::pending_exception_offset()),
3508 R16_thread);
3509 __ cmpdi(CCR0, R0, 0);
3510 __ bne(CCR0, L);
3511 __ stop("SharedRuntime::throw_exception: no pending exception");
3512 __ bind(L);
3513 }
3514 #endif
3515
3516 // Pop frame.
3517 __ pop_frame();
3518
3519 __ restore_LR(R11_scratch1);
3520
3521 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
3522 __ mtctr(R11_scratch1);
3523 __ bctr();
3524
3525 // Create runtime stub with OopMap.
3526 RuntimeStub* stub =
3527 RuntimeStub::new_runtime_stub(name, &code,
3528 /*frame_complete=*/ (int)(frame_complete_pc - start),
3529 frame_size_in_bytes/wordSize,
3530 oop_maps,
3531 false);
3532 return stub;
3533 }
3534
3535 //------------------------------Montgomery multiplication------------------------
3536 //
3537
3538 // Subtract 0:b from carry:a. Return carry.
3539 static unsigned long
3540 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3541 long i = 0;
3542 unsigned long tmp, tmp2;
3543 __asm__ __volatile__ (
3544 "subfc %[tmp], %[tmp], %[tmp] \n" // pre-set CA
3545 "mtctr %[len] \n"
3546 "0: \n"
3547 "ldx %[tmp], %[i], %[a] \n"
3548 "ldx %[tmp2], %[i], %[b] \n"
3549 "subfe %[tmp], %[tmp2], %[tmp] \n" // subtract extended
3550 "stdx %[tmp], %[i], %[a] \n"
3551 "addi %[i], %[i], 8 \n"
3552 "bdnz 0b \n"
3553 "addme %[tmp], %[carry] \n" // carry + CA - 1
3554 : [i]"+b"(i), [tmp]"=&r"(tmp), [tmp2]"=&r"(tmp2)
3555 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry), [len]"r"(len)
3556 : "ctr", "xer", "memory"
3557 );
3558 return tmp;
3559 }
3560
3561 // Multiply (unsigned) Long A by Long B, accumulating the double-
3562 // length result into the accumulator formed of T0, T1, and T2.
3563 inline void MACC(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3564 unsigned long hi, lo;
3565 __asm__ __volatile__ (
3566 "mulld %[lo], %[A], %[B] \n"
3567 "mulhdu %[hi], %[A], %[B] \n"
3568 "addc %[T0], %[T0], %[lo] \n"
3569 "adde %[T1], %[T1], %[hi] \n"
3570 "addze %[T2], %[T2] \n"
3571 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3572 : [A]"r"(A), [B]"r"(B)
3573 : "xer"
3574 );
3575 }
3576
3577 // As above, but add twice the double-length result into the
3578 // accumulator.
3579 inline void MACC2(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3580 unsigned long hi, lo;
3581 __asm__ __volatile__ (
3582 "mulld %[lo], %[A], %[B] \n"
3583 "mulhdu %[hi], %[A], %[B] \n"
3584 "addc %[T0], %[T0], %[lo] \n"
3585 "adde %[T1], %[T1], %[hi] \n"
3586 "addze %[T2], %[T2] \n"
3587 "addc %[T0], %[T0], %[lo] \n"
3588 "adde %[T1], %[T1], %[hi] \n"
3589 "addze %[T2], %[T2] \n"
3590 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3591 : [A]"r"(A), [B]"r"(B)
3592 : "xer"
3593 );
3594 }
3595
3596 // Fast Montgomery multiplication. The derivation of the algorithm is
3597 // in "A Cryptographic Library for the Motorola DSP56000,
3598 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3599 static void
3600 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3601 unsigned long m[], unsigned long inv, int len) {
3602 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3603 int i;
3604
3605 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3606
3607 for (i = 0; i < len; i++) {
3608 int j;
3609 for (j = 0; j < i; j++) {
3610 MACC(a[j], b[i-j], t0, t1, t2);
3611 MACC(m[j], n[i-j], t0, t1, t2);
3612 }
3613 MACC(a[i], b[0], t0, t1, t2);
3614 m[i] = t0 * inv;
3615 MACC(m[i], n[0], t0, t1, t2);
3616
3617 assert(t0 == 0, "broken Montgomery multiply");
3618
3619 t0 = t1; t1 = t2; t2 = 0;
3620 }
3621
3622 for (i = len; i < 2*len; i++) {
3623 int j;
3624 for (j = i-len+1; j < len; j++) {
3625 MACC(a[j], b[i-j], t0, t1, t2);
3626 MACC(m[j], n[i-j], t0, t1, t2);
3627 }
3628 m[i-len] = t0;
3629 t0 = t1; t1 = t2; t2 = 0;
3630 }
3631
3632 while (t0) {
3633 t0 = sub(m, n, t0, len);
3634 }
3635 }
3636
3637 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3638 // multiplies so it should be up to 25% faster than Montgomery
3639 // multiplication. However, its loop control is more complex and it
3640 // may actually run slower on some machines.
3641 static void
3642 montgomery_square(unsigned long a[], unsigned long n[],
3643 unsigned long m[], unsigned long inv, int len) {
3644 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3645 int i;
3646
3647 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3648
3649 for (i = 0; i < len; i++) {
3650 int j;
3651 int end = (i+1)/2;
3652 for (j = 0; j < end; j++) {
3653 MACC2(a[j], a[i-j], t0, t1, t2);
3654 MACC(m[j], n[i-j], t0, t1, t2);
3655 }
3656 if ((i & 1) == 0) {
3657 MACC(a[j], a[j], t0, t1, t2);
3658 }
3659 for (; j < i; j++) {
3660 MACC(m[j], n[i-j], t0, t1, t2);
3661 }
3662 m[i] = t0 * inv;
3663 MACC(m[i], n[0], t0, t1, t2);
3664
3665 assert(t0 == 0, "broken Montgomery square");
3666
3667 t0 = t1; t1 = t2; t2 = 0;
3668 }
3669
3670 for (i = len; i < 2*len; i++) {
3671 int start = i-len+1;
3672 int end = start + (len - start)/2;
3673 int j;
3674 for (j = start; j < end; j++) {
3675 MACC2(a[j], a[i-j], t0, t1, t2);
3676 MACC(m[j], n[i-j], t0, t1, t2);
3677 }
3678 if ((i & 1) == 0) {
3679 MACC(a[j], a[j], t0, t1, t2);
3680 }
3681 for (; j < len; j++) {
3682 MACC(m[j], n[i-j], t0, t1, t2);
3683 }
3684 m[i-len] = t0;
3685 t0 = t1; t1 = t2; t2 = 0;
3686 }
3687
3688 while (t0) {
3689 t0 = sub(m, n, t0, len);
3690 }
3691 }
3692
3693 // The threshold at which squaring is advantageous was determined
3694 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3695 // Doesn't seem to be relevant for Power8 so we use the same value.
3696 #define MONTGOMERY_SQUARING_THRESHOLD 64
3697
3698 // Copy len longwords from s to d, word-swapping as we go. The
3699 // destination array is reversed.
3700 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3701 d += len;
3702 while(len-- > 0) {
3703 d--;
3704 unsigned long s_val = *s;
3705 // Swap words in a longword on little endian machines.
3706 #ifdef VM_LITTLE_ENDIAN
3707 s_val = (s_val << 32) | (s_val >> 32);
3708 #endif
3709 *d = s_val;
3710 s++;
3711 }
3712 }
3713
3714 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3715 jint len, jlong inv,
3716 jint *m_ints) {
3717 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3718 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3719 int longwords = len/2;
3720
3721 // Make very sure we don't use so much space that the stack might
3722 // overflow. 512 jints corresponds to an 16384-bit integer and
3723 // will use here a total of 8k bytes of stack space.
3724 int divisor = sizeof(unsigned long) * 4;
3725 guarantee(longwords <= 8192 / divisor, "must be");
3726 int total_allocation = longwords * sizeof (unsigned long) * 4;
3727 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3728
3729 // Local scratch arrays
3730 unsigned long
3731 *a = scratch + 0 * longwords,
3732 *b = scratch + 1 * longwords,
3733 *n = scratch + 2 * longwords,
3734 *m = scratch + 3 * longwords;
3735
3736 reverse_words((unsigned long *)a_ints, a, longwords);
3737 reverse_words((unsigned long *)b_ints, b, longwords);
3738 reverse_words((unsigned long *)n_ints, n, longwords);
3739
3740 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3741
3742 reverse_words(m, (unsigned long *)m_ints, longwords);
3743 }
3744
3745 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3746 jint len, jlong inv,
3747 jint *m_ints) {
3748 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3749 assert(len % 2 == 0, "array length in montgomery_square must be even");
3750 int longwords = len/2;
3751
3752 // Make very sure we don't use so much space that the stack might
3753 // overflow. 512 jints corresponds to an 16384-bit integer and
3754 // will use here a total of 6k bytes of stack space.
3755 int divisor = sizeof(unsigned long) * 3;
3756 guarantee(longwords <= (8192 / divisor), "must be");
3757 int total_allocation = longwords * sizeof (unsigned long) * 3;
3758 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3759
3760 // Local scratch arrays
3761 unsigned long
3762 *a = scratch + 0 * longwords,
3763 *n = scratch + 1 * longwords,
3764 *m = scratch + 2 * longwords;
3765
3766 reverse_words((unsigned long *)a_ints, a, longwords);
3767 reverse_words((unsigned long *)n_ints, n, longwords);
3768
3769 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3770 ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3771 } else {
3772 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3773 }
3774
3775 reverse_words(m, (unsigned long *)m_ints, longwords);
3776 }
3777
3778 #if INCLUDE_JFR
3779
3780 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3781 // It returns a jobject handle to the event writer.
3782 // The handle is dereferenced and the return value is the event writer oop.
3783 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3784 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
3785 CodeBuffer code(name, 512, 64);
3786 MacroAssembler* masm = new MacroAssembler(&code);
3787
3788 Register tmp1 = R10_ARG8;
3789 Register tmp2 = R9_ARG7;
3790
3791 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3792 address start = __ pc();
3793 __ mflr(tmp1);
3794 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3795 __ push_frame_reg_args(0, tmp1);
3796 int frame_complete = __ pc() - start;
3797 __ set_last_Java_frame(R1_SP, noreg);
3798 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), R16_thread);
3799 address calls_return_pc = __ last_calls_return_pc();
3800 __ reset_last_Java_frame();
3801 // The handle is dereferenced through a load barrier.
3802 __ resolve_global_jobject(R3_RET, tmp1, tmp2, MacroAssembler::PRESERVATION_NONE);
3803 __ pop_frame();
3804 __ ld(tmp1, _abi0(lr), R1_SP);
3805 __ mtlr(tmp1);
3806 __ blr();
3807
3808 OopMapSet* oop_maps = new OopMapSet();
3809 OopMap* map = new OopMap(framesize, 0);
3810 oop_maps->add_gc_map(calls_return_pc - start, map);
3811
3812 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3813 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3814 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3815 oop_maps, false);
3816 return stub;
3817 }
3818
3819 // For c2: call to return a leased buffer.
3820 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3821 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_return_lease_id);
3822 CodeBuffer code(name, 512, 64);
3823 MacroAssembler* masm = new MacroAssembler(&code);
3824
3825 Register tmp1 = R10_ARG8;
3826 Register tmp2 = R9_ARG7;
3827
3828 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3829 address start = __ pc();
3830 __ mflr(tmp1);
3831 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3832 __ push_frame_reg_args(0, tmp1);
3833 int frame_complete = __ pc() - start;
3834 __ set_last_Java_frame(R1_SP, noreg);
3835 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), R16_thread);
3836 address calls_return_pc = __ last_calls_return_pc();
3837 __ reset_last_Java_frame();
3838 __ pop_frame();
3839 __ ld(tmp1, _abi0(lr), R1_SP);
3840 __ mtlr(tmp1);
3841 __ blr();
3842
3843 OopMapSet* oop_maps = new OopMapSet();
3844 OopMap* map = new OopMap(framesize, 0);
3845 oop_maps->add_gc_map(calls_return_pc - start, map);
3846
3847 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3848 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3849 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3850 oop_maps, false);
3851 return stub;
3852 }
3853
3854 #endif // INCLUDE_JFR