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