Wed, 11 Dec 2013 00:06:11 +0100
8029940: PPC64 (part 122): C2 compiler port
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * Copyright 2012, 2013 SAP AG. 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 */
26 #include "precompiled.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/icBuffer.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "oops/compiledICHolder.hpp"
33 #include "prims/jvmtiRedefineClassesTrace.hpp"
34 #include "runtime/sharedRuntime.hpp"
35 #include "runtime/vframeArray.hpp"
36 #include "vmreg_ppc.inline.hpp"
37 #ifdef COMPILER1
38 #include "c1/c1_Runtime1.hpp"
39 #endif
40 #ifdef COMPILER2
41 #include "opto/runtime.hpp"
42 #endif
44 #define __ masm->
46 #ifdef PRODUCT
47 #define BLOCK_COMMENT(str) // nothing
48 #else
49 #define BLOCK_COMMENT(str) __ block_comment(str)
50 #endif
52 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
55 // Used by generate_deopt_blob. Defined in .ad file.
56 extern uint size_deopt_handler();
59 class RegisterSaver {
60 // Used for saving volatile registers.
61 public:
63 // Support different return pc locations.
64 enum ReturnPCLocation {
65 return_pc_is_lr,
66 return_pc_is_r4,
67 return_pc_is_thread_saved_exception_pc
68 };
70 static OopMap* push_frame_abi112_and_save_live_registers(MacroAssembler* masm,
71 int* out_frame_size_in_bytes,
72 bool generate_oop_map,
73 int return_pc_adjustment,
74 ReturnPCLocation return_pc_location);
75 static void restore_live_registers_and_pop_frame(MacroAssembler* masm,
76 int frame_size_in_bytes,
77 bool restore_ctr);
79 static void push_frame_and_save_argument_registers(MacroAssembler* masm,
80 Register r_temp,
81 int frame_size,
82 int total_args,
83 const VMRegPair *regs, const VMRegPair *regs2 = NULL);
84 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm,
85 int frame_size,
86 int total_args,
87 const VMRegPair *regs, const VMRegPair *regs2 = NULL);
89 // During deoptimization only the result registers need to be restored
90 // all the other values have already been extracted.
91 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes);
93 // Constants and data structures:
95 typedef enum {
96 int_reg = 0,
97 float_reg = 1,
98 special_reg = 2
99 } RegisterType;
101 typedef enum {
102 reg_size = 8,
103 half_reg_size = reg_size / 2,
104 } RegisterConstants;
106 typedef struct {
107 RegisterType reg_type;
108 int reg_num;
109 VMReg vmreg;
110 } LiveRegType;
111 };
114 #define RegisterSaver_LiveSpecialReg(regname) \
115 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() }
117 #define RegisterSaver_LiveIntReg(regname) \
118 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() }
120 #define RegisterSaver_LiveFloatReg(regname) \
121 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() }
123 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
124 // Live registers which get spilled to the stack. Register
125 // positions in this array correspond directly to the stack layout.
127 //
128 // live special registers:
129 //
130 RegisterSaver_LiveSpecialReg(SR_CTR),
131 //
132 // live float registers:
133 //
134 RegisterSaver_LiveFloatReg( F0 ),
135 RegisterSaver_LiveFloatReg( F1 ),
136 RegisterSaver_LiveFloatReg( F2 ),
137 RegisterSaver_LiveFloatReg( F3 ),
138 RegisterSaver_LiveFloatReg( F4 ),
139 RegisterSaver_LiveFloatReg( F5 ),
140 RegisterSaver_LiveFloatReg( F6 ),
141 RegisterSaver_LiveFloatReg( F7 ),
142 RegisterSaver_LiveFloatReg( F8 ),
143 RegisterSaver_LiveFloatReg( F9 ),
144 RegisterSaver_LiveFloatReg( F10 ),
145 RegisterSaver_LiveFloatReg( F11 ),
146 RegisterSaver_LiveFloatReg( F12 ),
147 RegisterSaver_LiveFloatReg( F13 ),
148 RegisterSaver_LiveFloatReg( F14 ),
149 RegisterSaver_LiveFloatReg( F15 ),
150 RegisterSaver_LiveFloatReg( F16 ),
151 RegisterSaver_LiveFloatReg( F17 ),
152 RegisterSaver_LiveFloatReg( F18 ),
153 RegisterSaver_LiveFloatReg( F19 ),
154 RegisterSaver_LiveFloatReg( F20 ),
155 RegisterSaver_LiveFloatReg( F21 ),
156 RegisterSaver_LiveFloatReg( F22 ),
157 RegisterSaver_LiveFloatReg( F23 ),
158 RegisterSaver_LiveFloatReg( F24 ),
159 RegisterSaver_LiveFloatReg( F25 ),
160 RegisterSaver_LiveFloatReg( F26 ),
161 RegisterSaver_LiveFloatReg( F27 ),
162 RegisterSaver_LiveFloatReg( F28 ),
163 RegisterSaver_LiveFloatReg( F29 ),
164 RegisterSaver_LiveFloatReg( F30 ),
165 RegisterSaver_LiveFloatReg( F31 ),
166 //
167 // live integer registers:
168 //
169 RegisterSaver_LiveIntReg( R0 ),
170 //RegisterSaver_LiveIntReg( R1 ), // stack pointer
171 RegisterSaver_LiveIntReg( R2 ),
172 RegisterSaver_LiveIntReg( R3 ),
173 RegisterSaver_LiveIntReg( R4 ),
174 RegisterSaver_LiveIntReg( R5 ),
175 RegisterSaver_LiveIntReg( R6 ),
176 RegisterSaver_LiveIntReg( R7 ),
177 RegisterSaver_LiveIntReg( R8 ),
178 RegisterSaver_LiveIntReg( R9 ),
179 RegisterSaver_LiveIntReg( R10 ),
180 RegisterSaver_LiveIntReg( R11 ),
181 RegisterSaver_LiveIntReg( R12 ),
182 //RegisterSaver_LiveIntReg( R13 ), // system thread id
183 RegisterSaver_LiveIntReg( R14 ),
184 RegisterSaver_LiveIntReg( R15 ),
185 RegisterSaver_LiveIntReg( R16 ),
186 RegisterSaver_LiveIntReg( R17 ),
187 RegisterSaver_LiveIntReg( R18 ),
188 RegisterSaver_LiveIntReg( R19 ),
189 RegisterSaver_LiveIntReg( R20 ),
190 RegisterSaver_LiveIntReg( R21 ),
191 RegisterSaver_LiveIntReg( R22 ),
192 RegisterSaver_LiveIntReg( R23 ),
193 RegisterSaver_LiveIntReg( R24 ),
194 RegisterSaver_LiveIntReg( R25 ),
195 RegisterSaver_LiveIntReg( R26 ),
196 RegisterSaver_LiveIntReg( R27 ),
197 RegisterSaver_LiveIntReg( R28 ),
198 RegisterSaver_LiveIntReg( R29 ),
199 RegisterSaver_LiveIntReg( R31 ),
200 RegisterSaver_LiveIntReg( R30 ), // r30 must be the last register
201 };
203 OopMap* RegisterSaver::push_frame_abi112_and_save_live_registers(MacroAssembler* masm,
204 int* out_frame_size_in_bytes,
205 bool generate_oop_map,
206 int return_pc_adjustment,
207 ReturnPCLocation return_pc_location) {
208 // Push an abi112-frame and store all registers which may be live.
209 // If requested, create an OopMap: Record volatile registers as
210 // callee-save values in an OopMap so their save locations will be
211 // propagated to the RegisterMap of the caller frame during
212 // StackFrameStream construction (needed for deoptimization; see
213 // compiledVFrame::create_stack_value).
214 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment.
216 int i;
217 int offset;
219 // calcualte frame size
220 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
221 sizeof(RegisterSaver::LiveRegType);
222 const int register_save_size = regstosave_num * reg_size;
223 const int frame_size_in_bytes = round_to(register_save_size, frame::alignment_in_bytes)
224 + frame::abi_112_size;
225 *out_frame_size_in_bytes = frame_size_in_bytes;
226 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
227 const int register_save_offset = frame_size_in_bytes - register_save_size;
229 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
230 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : NULL;
232 BLOCK_COMMENT("push_frame_abi112_and_save_live_registers {");
234 // Save r30 in the last slot of the not yet pushed frame so that we
235 // can use it as scratch reg.
236 __ std(R30, -reg_size, R1_SP);
237 assert(-reg_size == register_save_offset - frame_size_in_bytes + ((regstosave_num-1)*reg_size),
238 "consistency check");
240 // save the flags
241 // Do the save_LR_CR by hand and adjust the return pc if requested.
242 __ mfcr(R30);
243 __ std(R30, _abi(cr), R1_SP);
244 switch (return_pc_location) {
245 case return_pc_is_lr: __ mflr(R30); break;
246 case return_pc_is_r4: __ mr(R30, R4); break;
247 case return_pc_is_thread_saved_exception_pc:
248 __ ld(R30, thread_(saved_exception_pc)); break;
249 default: ShouldNotReachHere();
250 }
251 if (return_pc_adjustment != 0)
252 __ addi(R30, R30, return_pc_adjustment);
253 __ std(R30, _abi(lr), R1_SP);
255 // push a new frame
256 __ push_frame(frame_size_in_bytes, R30);
258 // save all registers (ints and floats)
259 offset = register_save_offset;
260 for (int i = 0; i < regstosave_num; i++) {
261 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
262 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
264 switch (reg_type) {
265 case RegisterSaver::int_reg: {
266 if (reg_num != 30) { // We spilled R30 right at the beginning.
267 __ std(as_Register(reg_num), offset, R1_SP);
268 }
269 break;
270 }
271 case RegisterSaver::float_reg: {
272 __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
273 break;
274 }
275 case RegisterSaver::special_reg: {
276 if (reg_num == SR_CTR_SpecialRegisterEnumValue) {
277 __ mfctr(R30);
278 __ std(R30, offset, R1_SP);
279 } else {
280 Unimplemented();
281 }
282 break;
283 }
284 default:
285 ShouldNotReachHere();
286 }
288 if (generate_oop_map) {
289 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
290 RegisterSaver_LiveRegs[i].vmreg);
291 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2),
292 RegisterSaver_LiveRegs[i].vmreg->next());
293 }
294 offset += reg_size;
295 }
297 BLOCK_COMMENT("} push_frame_abi112_and_save_live_registers");
299 // And we're done.
300 return map;
301 }
304 // Pop the current frame and restore all the registers that we
305 // saved.
306 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
307 int frame_size_in_bytes,
308 bool restore_ctr) {
309 int i;
310 int offset;
311 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
312 sizeof(RegisterSaver::LiveRegType);
313 const int register_save_size = regstosave_num * reg_size;
314 const int register_save_offset = frame_size_in_bytes - register_save_size;
316 BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
318 // restore all registers (ints and floats)
319 offset = register_save_offset;
320 for (int i = 0; i < regstosave_num; i++) {
321 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
322 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
324 switch (reg_type) {
325 case RegisterSaver::int_reg: {
326 if (reg_num != 30) // R30 restored at the end, it's the tmp reg!
327 __ ld(as_Register(reg_num), offset, R1_SP);
328 break;
329 }
330 case RegisterSaver::float_reg: {
331 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
332 break;
333 }
334 case RegisterSaver::special_reg: {
335 if (reg_num == SR_CTR_SpecialRegisterEnumValue) {
336 if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
337 __ ld(R30, offset, R1_SP);
338 __ mtctr(R30);
339 }
340 } else {
341 Unimplemented();
342 }
343 break;
344 }
345 default:
346 ShouldNotReachHere();
347 }
348 offset += reg_size;
349 }
351 // pop the frame
352 __ pop_frame();
354 // restore the flags
355 __ restore_LR_CR(R30);
357 // restore scratch register's value
358 __ ld(R30, -reg_size, R1_SP);
360 BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
361 }
363 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
364 int frame_size,int total_args, const VMRegPair *regs,
365 const VMRegPair *regs2) {
366 __ push_frame(frame_size, r_temp);
367 int st_off = frame_size - wordSize;
368 for (int i = 0; i < total_args; i++) {
369 VMReg r_1 = regs[i].first();
370 VMReg r_2 = regs[i].second();
371 if (!r_1->is_valid()) {
372 assert(!r_2->is_valid(), "");
373 continue;
374 }
375 if (r_1->is_Register()) {
376 Register r = r_1->as_Register();
377 __ std(r, st_off, R1_SP);
378 st_off -= wordSize;
379 } else if (r_1->is_FloatRegister()) {
380 FloatRegister f = r_1->as_FloatRegister();
381 __ stfd(f, st_off, R1_SP);
382 st_off -= wordSize;
383 }
384 }
385 if (regs2 != NULL) {
386 for (int i = 0; i < total_args; i++) {
387 VMReg r_1 = regs2[i].first();
388 VMReg r_2 = regs2[i].second();
389 if (!r_1->is_valid()) {
390 assert(!r_2->is_valid(), "");
391 continue;
392 }
393 if (r_1->is_Register()) {
394 Register r = r_1->as_Register();
395 __ std(r, st_off, R1_SP);
396 st_off -= wordSize;
397 } else if (r_1->is_FloatRegister()) {
398 FloatRegister f = r_1->as_FloatRegister();
399 __ stfd(f, st_off, R1_SP);
400 st_off -= wordSize;
401 }
402 }
403 }
404 }
406 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
407 int total_args, const VMRegPair *regs,
408 const VMRegPair *regs2) {
409 int st_off = frame_size - wordSize;
410 for (int i = 0; i < total_args; i++) {
411 VMReg r_1 = regs[i].first();
412 VMReg r_2 = regs[i].second();
413 if (r_1->is_Register()) {
414 Register r = r_1->as_Register();
415 __ ld(r, st_off, R1_SP);
416 st_off -= wordSize;
417 } else if (r_1->is_FloatRegister()) {
418 FloatRegister f = r_1->as_FloatRegister();
419 __ lfd(f, st_off, R1_SP);
420 st_off -= wordSize;
421 }
422 }
423 if (regs2 != NULL)
424 for (int i = 0; i < total_args; i++) {
425 VMReg r_1 = regs2[i].first();
426 VMReg r_2 = regs2[i].second();
427 if (r_1->is_Register()) {
428 Register r = r_1->as_Register();
429 __ ld(r, st_off, R1_SP);
430 st_off -= wordSize;
431 } else if (r_1->is_FloatRegister()) {
432 FloatRegister f = r_1->as_FloatRegister();
433 __ lfd(f, st_off, R1_SP);
434 st_off -= wordSize;
435 }
436 }
437 __ pop_frame();
438 }
440 // Restore the registers that might be holding a result.
441 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) {
442 int i;
443 int offset;
444 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
445 sizeof(RegisterSaver::LiveRegType);
446 const int register_save_size = regstosave_num * reg_size;
447 const int register_save_offset = frame_size_in_bytes - register_save_size;
449 // restore all result registers (ints and floats)
450 offset = register_save_offset;
451 for (int i = 0; i < regstosave_num; i++) {
452 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
453 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
454 switch (reg_type) {
455 case RegisterSaver::int_reg: {
456 if (as_Register(reg_num)==R3_RET) // int result_reg
457 __ ld(as_Register(reg_num), offset, R1_SP);
458 break;
459 }
460 case RegisterSaver::float_reg: {
461 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
462 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
463 break;
464 }
465 case RegisterSaver::special_reg: {
466 // Special registers don't hold a result.
467 break;
468 }
469 default:
470 ShouldNotReachHere();
471 }
472 offset += reg_size;
473 }
474 }
476 // Is vector's size (in bytes) bigger than a size saved by default?
477 bool SharedRuntime::is_wide_vector(int size) {
478 ResourceMark rm;
479 // Note, MaxVectorSize == 8 on PPC64.
480 assert(size <= 8, err_msg_res("%d bytes vectors are not supported", size));
481 return size > 8;
482 }
483 #ifdef COMPILER2
484 static int reg2slot(VMReg r) {
485 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
486 }
488 static int reg2offset(VMReg r) {
489 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
490 }
491 #endif
493 // ---------------------------------------------------------------------------
494 // Read the array of BasicTypes from a signature, and compute where the
495 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
496 // quantities. Values less than VMRegImpl::stack0 are registers, those above
497 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
498 // as framesizes are fixed.
499 // VMRegImpl::stack0 refers to the first slot 0(sp).
500 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
501 // up to RegisterImpl::number_of_registers) are the 64-bit
502 // integer registers.
504 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
505 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
506 // units regardless of build. Of course for i486 there is no 64 bit build
508 // The Java calling convention is a "shifted" version of the C ABI.
509 // By skipping the first C ABI register we can call non-static jni methods
510 // with small numbers of arguments without having to shuffle the arguments
511 // at all. Since we control the java ABI we ought to at least get some
512 // advantage out of it.
514 const VMReg java_iarg_reg[8] = {
515 R3->as_VMReg(),
516 R4->as_VMReg(),
517 R5->as_VMReg(),
518 R6->as_VMReg(),
519 R7->as_VMReg(),
520 R8->as_VMReg(),
521 R9->as_VMReg(),
522 R10->as_VMReg()
523 };
525 const VMReg java_farg_reg[13] = {
526 F1->as_VMReg(),
527 F2->as_VMReg(),
528 F3->as_VMReg(),
529 F4->as_VMReg(),
530 F5->as_VMReg(),
531 F6->as_VMReg(),
532 F7->as_VMReg(),
533 F8->as_VMReg(),
534 F9->as_VMReg(),
535 F10->as_VMReg(),
536 F11->as_VMReg(),
537 F12->as_VMReg(),
538 F13->as_VMReg()
539 };
541 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
542 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
544 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
545 VMRegPair *regs,
546 int total_args_passed,
547 int is_outgoing) {
548 // C2c calling conventions for compiled-compiled calls.
549 // Put 8 ints/longs into registers _AND_ 13 float/doubles into
550 // registers _AND_ put the rest on the stack.
552 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats
553 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
555 int i;
556 VMReg reg;
557 int stk = 0;
558 int ireg = 0;
559 int freg = 0;
561 // We put the first 8 arguments into registers and the rest on the
562 // stack, float arguments are already in their argument registers
563 // due to c2c calling conventions (see calling_convention).
564 for (int i = 0; i < total_args_passed; ++i) {
565 switch(sig_bt[i]) {
566 case T_BOOLEAN:
567 case T_CHAR:
568 case T_BYTE:
569 case T_SHORT:
570 case T_INT:
571 if (ireg < num_java_iarg_registers) {
572 // Put int/ptr in register
573 reg = java_iarg_reg[ireg];
574 ++ireg;
575 } else {
576 // Put int/ptr on stack.
577 reg = VMRegImpl::stack2reg(stk);
578 stk += inc_stk_for_intfloat;
579 }
580 regs[i].set1(reg);
581 break;
582 case T_LONG:
583 assert(sig_bt[i+1] == T_VOID, "expecting half");
584 if (ireg < num_java_iarg_registers) {
585 // Put long in register.
586 reg = java_iarg_reg[ireg];
587 ++ireg;
588 } else {
589 // Put long on stack. They must be aligned to 2 slots.
590 if (stk & 0x1) ++stk;
591 reg = VMRegImpl::stack2reg(stk);
592 stk += inc_stk_for_longdouble;
593 }
594 regs[i].set2(reg);
595 break;
596 case T_OBJECT:
597 case T_ARRAY:
598 case T_ADDRESS:
599 if (ireg < num_java_iarg_registers) {
600 // Put ptr in register.
601 reg = java_iarg_reg[ireg];
602 ++ireg;
603 } else {
604 // Put ptr on stack. Objects must be aligned to 2 slots too,
605 // because "64-bit pointers record oop-ishness on 2 aligned
606 // adjacent registers." (see OopFlow::build_oop_map).
607 if (stk & 0x1) ++stk;
608 reg = VMRegImpl::stack2reg(stk);
609 stk += inc_stk_for_longdouble;
610 }
611 regs[i].set2(reg);
612 break;
613 case T_FLOAT:
614 if (freg < num_java_farg_registers) {
615 // Put float in register.
616 reg = java_farg_reg[freg];
617 ++freg;
618 } else {
619 // Put float on stack.
620 reg = VMRegImpl::stack2reg(stk);
621 stk += inc_stk_for_intfloat;
622 }
623 regs[i].set1(reg);
624 break;
625 case T_DOUBLE:
626 assert(sig_bt[i+1] == T_VOID, "expecting half");
627 if (freg < num_java_farg_registers) {
628 // Put double in register.
629 reg = java_farg_reg[freg];
630 ++freg;
631 } else {
632 // Put double on stack. They must be aligned to 2 slots.
633 if (stk & 0x1) ++stk;
634 reg = VMRegImpl::stack2reg(stk);
635 stk += inc_stk_for_longdouble;
636 }
637 regs[i].set2(reg);
638 break;
639 case T_VOID:
640 // Do not count halves.
641 regs[i].set_bad();
642 break;
643 default:
644 ShouldNotReachHere();
645 }
646 }
647 return round_to(stk, 2);
648 }
650 #ifdef COMPILER2
651 // Calling convention for calling C code.
652 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
653 VMRegPair *regs,
654 VMRegPair *regs2,
655 int total_args_passed) {
656 // Calling conventions for C runtime calls and calls to JNI native methods.
657 //
658 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
659 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
660 // the first 13 flt/dbl's in the first 13 fp regs but additionally
661 // copy flt/dbl to the stack if they are beyond the 8th argument.
663 const VMReg iarg_reg[8] = {
664 R3->as_VMReg(),
665 R4->as_VMReg(),
666 R5->as_VMReg(),
667 R6->as_VMReg(),
668 R7->as_VMReg(),
669 R8->as_VMReg(),
670 R9->as_VMReg(),
671 R10->as_VMReg()
672 };
674 const VMReg farg_reg[13] = {
675 F1->as_VMReg(),
676 F2->as_VMReg(),
677 F3->as_VMReg(),
678 F4->as_VMReg(),
679 F5->as_VMReg(),
680 F6->as_VMReg(),
681 F7->as_VMReg(),
682 F8->as_VMReg(),
683 F9->as_VMReg(),
684 F10->as_VMReg(),
685 F11->as_VMReg(),
686 F12->as_VMReg(),
687 F13->as_VMReg()
688 };
690 // Check calling conventions consistency.
691 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
692 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
693 "consistency");
695 // `Stk' counts stack slots. Due to alignment, 32 bit values occupy
696 // 2 such slots, like 64 bit values do.
697 const int inc_stk_for_intfloat = 2; // 2 slots for ints and floats
698 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
700 int i;
701 VMReg reg;
702 // Leave room for C-compatible ABI_112.
703 int stk = (frame::abi_112_size - frame::jit_out_preserve_size) / VMRegImpl::stack_slot_size;
704 int arg = 0;
705 int freg = 0;
707 // Avoid passing C arguments in the wrong stack slots.
708 assert((SharedRuntime::out_preserve_stack_slots() + stk) * VMRegImpl::stack_slot_size == 112,
709 "passing C arguments in wrong stack slots");
711 // We fill-out regs AND regs2 if an argument must be passed in a
712 // register AND in a stack slot. If regs2 is NULL in such a
713 // situation, we bail-out with a fatal error.
714 for (int i = 0; i < total_args_passed; ++i, ++arg) {
715 // Initialize regs2 to BAD.
716 if (regs2 != NULL) regs2[i].set_bad();
718 switch(sig_bt[i]) {
720 //
721 // If arguments 0-7 are integers, they are passed in integer registers.
722 // Argument i is placed in iarg_reg[i].
723 //
724 case T_BOOLEAN:
725 case T_CHAR:
726 case T_BYTE:
727 case T_SHORT:
728 case T_INT:
729 // We must cast ints to longs and use full 64 bit stack slots
730 // here. We do the cast in GraphKit::gen_stub() and just guard
731 // here against loosing that change.
732 assert(CCallingConventionRequiresIntsAsLongs,
733 "argument of type int should be promoted to type long");
734 guarantee(i > 0 && sig_bt[i-1] == T_LONG,
735 "argument of type (bt) should have been promoted to type (T_LONG,bt) for bt in "
736 "{T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}");
737 // Do not count halves.
738 regs[i].set_bad();
739 --arg;
740 break;
741 case T_LONG:
742 guarantee(sig_bt[i+1] == T_VOID ||
743 sig_bt[i+1] == T_BOOLEAN || sig_bt[i+1] == T_CHAR ||
744 sig_bt[i+1] == T_BYTE || sig_bt[i+1] == T_SHORT ||
745 sig_bt[i+1] == T_INT,
746 "expecting type (T_LONG,half) or type (T_LONG,bt) with bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}");
747 case T_OBJECT:
748 case T_ARRAY:
749 case T_ADDRESS:
750 case T_METADATA:
751 // Oops are already boxed if required (JNI).
752 if (arg < Argument::n_int_register_parameters_c) {
753 reg = iarg_reg[arg];
754 } else {
755 reg = VMRegImpl::stack2reg(stk);
756 stk += inc_stk_for_longdouble;
757 }
758 regs[i].set2(reg);
759 break;
761 //
762 // Floats are treated differently from int regs: The first 13 float arguments
763 // are passed in registers (not the float args among the first 13 args).
764 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed
765 // in farg_reg[j] if argument i is the j-th float argument of this call.
766 //
767 case T_FLOAT:
768 if (freg < Argument::n_float_register_parameters_c) {
769 // Put float in register ...
770 reg = farg_reg[freg];
771 ++freg;
773 // Argument i for i > 8 is placed on the stack even if it's
774 // placed in a register (if it's a float arg). Aix disassembly
775 // shows that xlC places these float args on the stack AND in
776 // a register. This is not documented, but we follow this
777 // convention, too.
778 if (arg >= Argument::n_regs_not_on_stack_c) {
779 // ... and on the stack.
780 guarantee(regs2 != NULL, "must pass float in register and stack slot");
781 VMReg reg2 = VMRegImpl::stack2reg(stk LINUX_ONLY(+1));
782 regs2[i].set1(reg2);
783 stk += inc_stk_for_intfloat;
784 }
786 } else {
787 // Put float on stack.
788 reg = VMRegImpl::stack2reg(stk LINUX_ONLY(+1));
789 stk += inc_stk_for_intfloat;
790 }
791 regs[i].set1(reg);
792 break;
793 case T_DOUBLE:
794 assert(sig_bt[i+1] == T_VOID, "expecting half");
795 if (freg < Argument::n_float_register_parameters_c) {
796 // Put double in register ...
797 reg = farg_reg[freg];
798 ++freg;
800 // Argument i for i > 8 is placed on the stack even if it's
801 // placed in a register (if it's a double arg). Aix disassembly
802 // shows that xlC places these float args on the stack AND in
803 // a register. This is not documented, but we follow this
804 // convention, too.
805 if (arg >= Argument::n_regs_not_on_stack_c) {
806 // ... and on the stack.
807 guarantee(regs2 != NULL, "must pass float in register and stack slot");
808 VMReg reg2 = VMRegImpl::stack2reg(stk);
809 regs2[i].set2(reg2);
810 stk += inc_stk_for_longdouble;
811 }
812 } else {
813 // Put double on stack.
814 reg = VMRegImpl::stack2reg(stk);
815 stk += inc_stk_for_longdouble;
816 }
817 regs[i].set2(reg);
818 break;
820 case T_VOID:
821 // Do not count halves.
822 regs[i].set_bad();
823 --arg;
824 break;
825 default:
826 ShouldNotReachHere();
827 }
828 }
830 return round_to(stk, 2);
831 }
832 #endif // COMPILER2
834 static address gen_c2i_adapter(MacroAssembler *masm,
835 int total_args_passed,
836 int comp_args_on_stack,
837 const BasicType *sig_bt,
838 const VMRegPair *regs,
839 Label& call_interpreter,
840 const Register& ientry) {
842 address c2i_entrypoint;
844 const Register sender_SP = R21_sender_SP; // == R21_tmp1
845 const Register code = R22_tmp2;
846 //const Register ientry = R23_tmp3;
847 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
848 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
849 int value_regs_index = 0;
851 const Register return_pc = R27_tmp7;
852 const Register tmp = R28_tmp8;
854 assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
856 // Adapter needs TOP_IJAVA_FRAME_ABI.
857 const int adapter_size = frame::top_ijava_frame_abi_size +
858 round_to(total_args_passed * wordSize, frame::alignment_in_bytes);
860 // regular (verified) c2i entry point
861 c2i_entrypoint = __ pc();
863 // Does compiled code exists? If yes, patch the caller's callsite.
864 __ ld(code, method_(code));
865 __ cmpdi(CCR0, code, 0);
866 __ ld(ientry, method_(interpreter_entry)); // preloaded
867 __ beq(CCR0, call_interpreter);
870 // Patch caller's callsite, method_(code) was not NULL which means that
871 // compiled code exists.
872 __ mflr(return_pc);
873 __ std(return_pc, _abi(lr), R1_SP);
874 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
876 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
878 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
879 __ ld(return_pc, _abi(lr), R1_SP);
880 __ ld(ientry, method_(interpreter_entry)); // preloaded
881 __ mtlr(return_pc);
884 // Call the interpreter.
885 __ BIND(call_interpreter);
886 __ mtctr(ientry);
888 // Get a copy of the current SP for loading caller's arguments.
889 __ mr(sender_SP, R1_SP);
891 // Add space for the adapter.
892 __ resize_frame(-adapter_size, R12_scratch2);
894 int st_off = adapter_size - wordSize;
896 // Write the args into the outgoing interpreter space.
897 for (int i = 0; i < total_args_passed; i++) {
898 VMReg r_1 = regs[i].first();
899 VMReg r_2 = regs[i].second();
900 if (!r_1->is_valid()) {
901 assert(!r_2->is_valid(), "");
902 continue;
903 }
904 if (r_1->is_stack()) {
905 Register tmp_reg = value_regs[value_regs_index];
906 value_regs_index = (value_regs_index + 1) % num_value_regs;
907 // The calling convention produces OptoRegs that ignore the out
908 // preserve area (JIT's ABI). We must account for it here.
909 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
910 if (!r_2->is_valid()) {
911 __ lwz(tmp_reg, ld_off, sender_SP);
912 } else {
913 __ ld(tmp_reg, ld_off, sender_SP);
914 }
915 // Pretend stack targets were loaded into tmp_reg.
916 r_1 = tmp_reg->as_VMReg();
917 }
919 if (r_1->is_Register()) {
920 Register r = r_1->as_Register();
921 if (!r_2->is_valid()) {
922 __ stw(r, st_off, R1_SP);
923 st_off-=wordSize;
924 } else {
925 // Longs are given 2 64-bit slots in the interpreter, but the
926 // data is passed in only 1 slot.
927 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
928 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
929 st_off-=wordSize;
930 }
931 __ std(r, st_off, R1_SP);
932 st_off-=wordSize;
933 }
934 } else {
935 assert(r_1->is_FloatRegister(), "");
936 FloatRegister f = r_1->as_FloatRegister();
937 if (!r_2->is_valid()) {
938 __ stfs(f, st_off, R1_SP);
939 st_off-=wordSize;
940 } else {
941 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
942 // data is passed in only 1 slot.
943 // One of these should get known junk...
944 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
945 st_off-=wordSize;
946 __ stfd(f, st_off, R1_SP);
947 st_off-=wordSize;
948 }
949 }
950 }
952 // Jump to the interpreter just as if interpreter was doing it.
954 #ifdef CC_INTERP
955 const Register tos = R17_tos;
956 #endif
958 // load TOS
959 __ addi(tos, R1_SP, st_off);
961 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
962 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
963 __ bctr();
965 return c2i_entrypoint;
966 }
968 static void gen_i2c_adapter(MacroAssembler *masm,
969 int total_args_passed,
970 int comp_args_on_stack,
971 const BasicType *sig_bt,
972 const VMRegPair *regs) {
974 // Load method's entry-point from methodOop.
975 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
976 __ mtctr(R12_scratch2);
978 // We will only enter here from an interpreted frame and never from after
979 // passing thru a c2i. Azul allowed this but we do not. If we lose the
980 // race and use a c2i we will remain interpreted for the race loser(s).
981 // This removes all sorts of headaches on the x86 side and also eliminates
982 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
984 // Note: r13 contains the senderSP on entry. We must preserve it since
985 // we may do a i2c -> c2i transition if we lose a race where compiled
986 // code goes non-entrant while we get args ready.
987 // In addition we use r13 to locate all the interpreter args as
988 // we must align the stack to 16 bytes on an i2c entry else we
989 // lose alignment we expect in all compiled code and register
990 // save code can segv when fxsave instructions find improperly
991 // aligned stack pointer.
993 #ifdef CC_INTERP
994 const Register ld_ptr = R17_tos;
995 #endif
996 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
997 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
998 int value_regs_index = 0;
1000 int ld_offset = total_args_passed*wordSize;
1002 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1003 // in registers, we will occasionally have no stack args.
1004 int comp_words_on_stack = 0;
1005 if (comp_args_on_stack) {
1006 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1007 // registers are below. By subtracting stack0, we either get a negative
1008 // number (all values in registers) or the maximum stack slot accessed.
1010 // Convert 4-byte c2 stack slots to words.
1011 comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1012 // Round up to miminum stack alignment, in wordSize.
1013 comp_words_on_stack = round_to(comp_words_on_stack, 2);
1014 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1015 }
1017 // Now generate the shuffle code. Pick up all register args and move the
1018 // rest through register value=Z_R12.
1019 BLOCK_COMMENT("Shuffle arguments");
1020 for (int i = 0; i < total_args_passed; i++) {
1021 if (sig_bt[i] == T_VOID) {
1022 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1023 continue;
1024 }
1026 // Pick up 0, 1 or 2 words from ld_ptr.
1027 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1028 "scrambled load targets?");
1029 VMReg r_1 = regs[i].first();
1030 VMReg r_2 = regs[i].second();
1031 if (!r_1->is_valid()) {
1032 assert(!r_2->is_valid(), "");
1033 continue;
1034 }
1035 if (r_1->is_FloatRegister()) {
1036 if (!r_2->is_valid()) {
1037 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1038 ld_offset-=wordSize;
1039 } else {
1040 // Skip the unused interpreter slot.
1041 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1042 ld_offset-=2*wordSize;
1043 }
1044 } else {
1045 Register r;
1046 if (r_1->is_stack()) {
1047 // Must do a memory to memory move thru "value".
1048 r = value_regs[value_regs_index];
1049 value_regs_index = (value_regs_index + 1) % num_value_regs;
1050 } else {
1051 r = r_1->as_Register();
1052 }
1053 if (!r_2->is_valid()) {
1054 // Not sure we need to do this but it shouldn't hurt.
1055 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ADDRESS || sig_bt[i] == T_ARRAY) {
1056 __ ld(r, ld_offset, ld_ptr);
1057 ld_offset-=wordSize;
1058 } else {
1059 __ lwz(r, ld_offset, ld_ptr);
1060 ld_offset-=wordSize;
1061 }
1062 } else {
1063 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1064 // data is passed in only 1 slot.
1065 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1066 ld_offset-=wordSize;
1067 }
1068 __ ld(r, ld_offset, ld_ptr);
1069 ld_offset-=wordSize;
1070 }
1072 if (r_1->is_stack()) {
1073 // Now store value where the compiler expects it
1074 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1076 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN ||
1077 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) {
1078 __ stw(r, st_off, R1_SP);
1079 } else {
1080 __ std(r, st_off, R1_SP);
1081 }
1082 }
1083 }
1084 }
1086 BLOCK_COMMENT("Store method oop");
1087 // Store method oop into thread->callee_target.
1088 // We might end up in handle_wrong_method if the callee is
1089 // deoptimized as we race thru here. If that happens we don't want
1090 // to take a safepoint because the caller frame will look
1091 // interpreted and arguments are now "compiled" so it is much better
1092 // to make this transition invisible to the stack walking
1093 // code. Unfortunately if we try and find the callee by normal means
1094 // a safepoint is possible. So we stash the desired callee in the
1095 // thread and the vm will find there should this case occur.
1096 __ std(R19_method, thread_(callee_target));
1098 // Jump to the compiled code just as if compiled code was doing it.
1099 __ bctr();
1100 }
1102 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1103 int total_args_passed,
1104 int comp_args_on_stack,
1105 const BasicType *sig_bt,
1106 const VMRegPair *regs,
1107 AdapterFingerPrint* fingerprint) {
1108 address i2c_entry;
1109 address c2i_unverified_entry;
1110 address c2i_entry;
1113 // entry: i2c
1115 __ align(CodeEntryAlignment);
1116 i2c_entry = __ pc();
1117 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1120 // entry: c2i unverified
1122 __ align(CodeEntryAlignment);
1123 BLOCK_COMMENT("c2i unverified entry");
1124 c2i_unverified_entry = __ pc();
1126 // inline_cache contains a compiledICHolder
1127 const Register ic = R19_method;
1128 const Register ic_klass = R11_scratch1;
1129 const Register receiver_klass = R12_scratch2;
1130 const Register code = R21_tmp1;
1131 const Register ientry = R23_tmp3;
1133 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1134 assert(R11_scratch1 == R11, "need prologue scratch register");
1136 Label call_interpreter;
1138 assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()),
1139 "klass offset should reach into any page");
1140 // Check for NULL argument if we don't have implicit null checks.
1141 if (!ImplicitNullChecks || !os::zero_page_read_protected()) {
1142 if (TrapBasedNullChecks) {
1143 __ trap_null_check(R3_ARG1);
1144 } else {
1145 Label valid;
1146 __ cmpdi(CCR0, R3_ARG1, 0);
1147 __ bne_predict_taken(CCR0, valid);
1148 // We have a null argument, branch to ic_miss_stub.
1149 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(),
1150 relocInfo::runtime_call_type);
1151 __ BIND(valid);
1152 }
1153 }
1154 // Assume argument is not NULL, load klass from receiver.
1155 __ load_klass(receiver_klass, R3_ARG1);
1157 __ ld(ic_klass, CompiledICHolder::holder_klass_offset(), ic);
1159 if (TrapBasedICMissChecks) {
1160 __ trap_ic_miss_check(receiver_klass, ic_klass);
1161 } else {
1162 Label valid;
1163 __ cmpd(CCR0, receiver_klass, ic_klass);
1164 __ beq_predict_taken(CCR0, valid);
1165 // We have an unexpected klass, branch to ic_miss_stub.
1166 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(),
1167 relocInfo::runtime_call_type);
1168 __ BIND(valid);
1169 }
1171 // Argument is valid and klass is as expected, continue.
1173 // Extract method from inline cache, verified entry point needs it.
1174 __ ld(R19_method, CompiledICHolder::holder_method_offset(), ic);
1175 assert(R19_method == ic, "the inline cache register is dead here");
1177 __ ld(code, method_(code));
1178 __ cmpdi(CCR0, code, 0);
1179 __ ld(ientry, method_(interpreter_entry)); // preloaded
1180 __ beq_predict_taken(CCR0, call_interpreter);
1182 // Branch to ic_miss_stub.
1183 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1185 // entry: c2i
1187 c2i_entry = gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry);
1189 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
1190 }
1192 #ifdef COMPILER2
1193 // An oop arg. Must pass a handle not the oop itself.
1194 static void object_move(MacroAssembler* masm,
1195 int frame_size_in_slots,
1196 OopMap* oop_map, int oop_handle_offset,
1197 bool is_receiver, int* receiver_offset,
1198 VMRegPair src, VMRegPair dst,
1199 Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1200 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1201 "receiver has already been moved");
1203 // We must pass a handle. First figure out the location we use as a handle.
1205 if (src.first()->is_stack()) {
1206 // stack to stack or reg
1208 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1209 Label skip;
1210 const int oop_slot_in_callers_frame = reg2slot(src.first());
1212 guarantee(!is_receiver, "expecting receiver in register");
1213 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1215 __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1216 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp);
1217 __ cmpdi(CCR0, r_temp_2, 0);
1218 __ bne(CCR0, skip);
1219 // Use a NULL handle if oop is NULL.
1220 __ li(r_handle, 0);
1221 __ bind(skip);
1223 if (dst.first()->is_stack()) {
1224 // stack to stack
1225 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1226 } else {
1227 // stack to reg
1228 // Nothing to do, r_handle is already the dst register.
1229 }
1230 } else {
1231 // reg to stack or reg
1232 const Register r_oop = src.first()->as_Register();
1233 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1234 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1235 + oop_handle_offset; // in slots
1236 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1237 Label skip;
1239 if (is_receiver) {
1240 *receiver_offset = oop_offset;
1241 }
1242 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1244 __ std( r_oop, oop_offset, R1_SP);
1245 __ addi(r_handle, R1_SP, oop_offset);
1247 __ cmpdi(CCR0, r_oop, 0);
1248 __ bne(CCR0, skip);
1249 // Use a NULL handle if oop is NULL.
1250 __ li(r_handle, 0);
1251 __ bind(skip);
1253 if (dst.first()->is_stack()) {
1254 // reg to stack
1255 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1256 } else {
1257 // reg to reg
1258 // Nothing to do, r_handle is already the dst register.
1259 }
1260 }
1261 }
1263 static void int_move(MacroAssembler*masm,
1264 VMRegPair src, VMRegPair dst,
1265 Register r_caller_sp, Register r_temp) {
1266 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long-int");
1267 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1269 if (src.first()->is_stack()) {
1270 if (dst.first()->is_stack()) {
1271 // stack to stack
1272 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1273 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1274 } else {
1275 // stack to reg
1276 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1277 }
1278 } else if (dst.first()->is_stack()) {
1279 // reg to stack
1280 __ extsw(r_temp, src.first()->as_Register());
1281 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1282 } else {
1283 // reg to reg
1284 __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1285 }
1286 }
1288 static void long_move(MacroAssembler*masm,
1289 VMRegPair src, VMRegPair dst,
1290 Register r_caller_sp, Register r_temp) {
1291 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1292 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1294 if (src.first()->is_stack()) {
1295 if (dst.first()->is_stack()) {
1296 // stack to stack
1297 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1298 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1299 } else {
1300 // stack to reg
1301 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1302 }
1303 } else if (dst.first()->is_stack()) {
1304 // reg to stack
1305 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1306 } else {
1307 // reg to reg
1308 if (dst.first()->as_Register() != src.first()->as_Register())
1309 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1310 }
1311 }
1313 static void float_move(MacroAssembler*masm,
1314 VMRegPair src, VMRegPair dst,
1315 Register r_caller_sp, Register r_temp) {
1316 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1317 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1319 if (src.first()->is_stack()) {
1320 if (dst.first()->is_stack()) {
1321 // stack to stack
1322 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1323 __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1324 } else {
1325 // stack to reg
1326 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1327 }
1328 } else if (dst.first()->is_stack()) {
1329 // reg to stack
1330 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1331 } else {
1332 // reg to reg
1333 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1334 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1335 }
1336 }
1338 static void double_move(MacroAssembler*masm,
1339 VMRegPair src, VMRegPair dst,
1340 Register r_caller_sp, Register r_temp) {
1341 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1342 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1344 if (src.first()->is_stack()) {
1345 if (dst.first()->is_stack()) {
1346 // stack to stack
1347 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1348 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1349 } else {
1350 // stack to reg
1351 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1352 }
1353 } else if (dst.first()->is_stack()) {
1354 // reg to stack
1355 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1356 } else {
1357 // reg to reg
1358 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1359 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1360 }
1361 }
1363 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1364 switch (ret_type) {
1365 case T_BOOLEAN:
1366 case T_CHAR:
1367 case T_BYTE:
1368 case T_SHORT:
1369 case T_INT:
1370 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1371 break;
1372 case T_ARRAY:
1373 case T_OBJECT:
1374 case T_LONG:
1375 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1376 break;
1377 case T_FLOAT:
1378 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1379 break;
1380 case T_DOUBLE:
1381 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1382 break;
1383 case T_VOID:
1384 break;
1385 default:
1386 ShouldNotReachHere();
1387 break;
1388 }
1389 }
1391 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1392 switch (ret_type) {
1393 case T_BOOLEAN:
1394 case T_CHAR:
1395 case T_BYTE:
1396 case T_SHORT:
1397 case T_INT:
1398 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1399 break;
1400 case T_ARRAY:
1401 case T_OBJECT:
1402 case T_LONG:
1403 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1404 break;
1405 case T_FLOAT:
1406 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1407 break;
1408 case T_DOUBLE:
1409 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1410 break;
1411 case T_VOID:
1412 break;
1413 default:
1414 ShouldNotReachHere();
1415 break;
1416 }
1417 }
1419 static void save_or_restore_arguments(MacroAssembler* masm,
1420 const int stack_slots,
1421 const int total_in_args,
1422 const int arg_save_area,
1423 OopMap* map,
1424 VMRegPair* in_regs,
1425 BasicType* in_sig_bt) {
1426 // If map is non-NULL then the code should store the values,
1427 // otherwise it should load them.
1428 int slot = arg_save_area;
1429 // Save down double word first.
1430 for (int i = 0; i < total_in_args; i++) {
1431 if (in_regs[i].first()->is_FloatRegister() && in_sig_bt[i] == T_DOUBLE) {
1432 int offset = slot * VMRegImpl::stack_slot_size;
1433 slot += VMRegImpl::slots_per_word;
1434 assert(slot <= stack_slots, "overflow (after DOUBLE stack slot)");
1435 if (map != NULL) {
1436 __ stfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP);
1437 } else {
1438 __ lfd(in_regs[i].first()->as_FloatRegister(), offset, R1_SP);
1439 }
1440 } else if (in_regs[i].first()->is_Register() &&
1441 (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_ARRAY)) {
1442 int offset = slot * VMRegImpl::stack_slot_size;
1443 if (map != NULL) {
1444 __ std(in_regs[i].first()->as_Register(), offset, R1_SP);
1445 if (in_sig_bt[i] == T_ARRAY) {
1446 map->set_oop(VMRegImpl::stack2reg(slot));
1447 }
1448 } else {
1449 __ ld(in_regs[i].first()->as_Register(), offset, R1_SP);
1450 }
1451 slot += VMRegImpl::slots_per_word;
1452 assert(slot <= stack_slots, "overflow (after LONG/ARRAY stack slot)");
1453 }
1454 }
1455 // Save or restore single word registers.
1456 for (int i = 0; i < total_in_args; i++) {
1457 // PPC64: pass ints as longs: must only deal with floats here.
1458 if (in_regs[i].first()->is_FloatRegister()) {
1459 if (in_sig_bt[i] == T_FLOAT) {
1460 int offset = slot * VMRegImpl::stack_slot_size;
1461 slot++;
1462 assert(slot <= stack_slots, "overflow (after FLOAT stack slot)");
1463 if (map != NULL) {
1464 __ stfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP);
1465 } else {
1466 __ lfs(in_regs[i].first()->as_FloatRegister(), offset, R1_SP);
1467 }
1468 }
1469 } else if (in_regs[i].first()->is_stack()) {
1470 if (in_sig_bt[i] == T_ARRAY && map != NULL) {
1471 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1472 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots));
1473 }
1474 }
1475 }
1476 }
1478 // Check GC_locker::needs_gc and enter the runtime if it's true. This
1479 // keeps a new JNI critical region from starting until a GC has been
1480 // forced. Save down any oops in registers and describe them in an
1481 // OopMap.
1482 static void check_needs_gc_for_critical_native(MacroAssembler* masm,
1483 const int stack_slots,
1484 const int total_in_args,
1485 const int arg_save_area,
1486 OopMapSet* oop_maps,
1487 VMRegPair* in_regs,
1488 BasicType* in_sig_bt,
1489 Register tmp_reg ) {
1490 __ block_comment("check GC_locker::needs_gc");
1491 Label cont;
1492 __ lbz(tmp_reg, (RegisterOrConstant)(intptr_t)GC_locker::needs_gc_address());
1493 __ cmplwi(CCR0, tmp_reg, 0);
1494 __ beq(CCR0, cont);
1496 // Save down any values that are live in registers and call into the
1497 // runtime to halt for a GC.
1498 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1499 save_or_restore_arguments(masm, stack_slots, total_in_args,
1500 arg_save_area, map, in_regs, in_sig_bt);
1502 __ mr(R3_ARG1, R16_thread);
1503 __ set_last_Java_frame(R1_SP, noreg);
1505 __ block_comment("block_for_jni_critical");
1506 address entry_point = CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical);
1507 __ call_c(CAST_FROM_FN_PTR(FunctionDescriptor*, entry_point), relocInfo::runtime_call_type);
1508 address start = __ pc() - __ offset(),
1509 calls_return_pc = __ last_calls_return_pc();
1510 oop_maps->add_gc_map(calls_return_pc - start, map);
1512 __ reset_last_Java_frame();
1514 // Reload all the register arguments.
1515 save_or_restore_arguments(masm, stack_slots, total_in_args,
1516 arg_save_area, NULL, in_regs, in_sig_bt);
1518 __ BIND(cont);
1520 #ifdef ASSERT
1521 if (StressCriticalJNINatives) {
1522 // Stress register saving.
1523 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1524 save_or_restore_arguments(masm, stack_slots, total_in_args,
1525 arg_save_area, map, in_regs, in_sig_bt);
1526 // Destroy argument registers.
1527 for (int i = 0; i < total_in_args; i++) {
1528 if (in_regs[i].first()->is_Register()) {
1529 const Register reg = in_regs[i].first()->as_Register();
1530 __ neg(reg, reg);
1531 } else if (in_regs[i].first()->is_FloatRegister()) {
1532 __ fneg(in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister());
1533 }
1534 }
1536 save_or_restore_arguments(masm, stack_slots, total_in_args,
1537 arg_save_area, NULL, in_regs, in_sig_bt);
1538 }
1539 #endif
1540 }
1542 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst, Register r_caller_sp, Register r_temp) {
1543 if (src.first()->is_stack()) {
1544 if (dst.first()->is_stack()) {
1545 // stack to stack
1546 __ ld(r_temp, reg2offset(src.first()), r_caller_sp);
1547 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1548 } else {
1549 // stack to reg
1550 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1551 }
1552 } else if (dst.first()->is_stack()) {
1553 // reg to stack
1554 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1555 } else {
1556 if (dst.first() != src.first()) {
1557 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1558 }
1559 }
1560 }
1562 // Unpack an array argument into a pointer to the body and the length
1563 // if the array is non-null, otherwise pass 0 for both.
1564 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type,
1565 VMRegPair body_arg, VMRegPair length_arg, Register r_caller_sp,
1566 Register tmp_reg, Register tmp2_reg) {
1567 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg,
1568 "possible collision");
1569 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg,
1570 "possible collision");
1572 // Pass the length, ptr pair.
1573 Label set_out_args;
1574 VMRegPair tmp, tmp2;
1575 tmp.set_ptr(tmp_reg->as_VMReg());
1576 tmp2.set_ptr(tmp2_reg->as_VMReg());
1577 if (reg.first()->is_stack()) {
1578 // Load the arg up from the stack.
1579 move_ptr(masm, reg, tmp, r_caller_sp, /*unused*/ R0);
1580 reg = tmp;
1581 }
1582 __ li(tmp2_reg, 0); // Pass zeros if Array=null.
1583 if (tmp_reg != reg.first()->as_Register()) __ li(tmp_reg, 0);
1584 __ cmpdi(CCR0, reg.first()->as_Register(), 0);
1585 __ beq(CCR0, set_out_args);
1586 __ lwa(tmp2_reg, arrayOopDesc::length_offset_in_bytes(), reg.first()->as_Register());
1587 __ addi(tmp_reg, reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type));
1588 __ bind(set_out_args);
1589 move_ptr(masm, tmp, body_arg, r_caller_sp, /*unused*/ R0);
1590 move_ptr(masm, tmp2, length_arg, r_caller_sp, /*unused*/ R0); // Same as move32_64 on PPC64.
1591 }
1593 static void verify_oop_args(MacroAssembler* masm,
1594 methodHandle method,
1595 const BasicType* sig_bt,
1596 const VMRegPair* regs) {
1597 Register temp_reg = R19_method; // not part of any compiled calling seq
1598 if (VerifyOops) {
1599 for (int i = 0; i < method->size_of_parameters(); i++) {
1600 if (sig_bt[i] == T_OBJECT ||
1601 sig_bt[i] == T_ARRAY) {
1602 VMReg r = regs[i].first();
1603 assert(r->is_valid(), "bad oop arg");
1604 if (r->is_stack()) {
1605 __ ld(temp_reg, reg2offset(r), R1_SP);
1606 __ verify_oop(temp_reg);
1607 } else {
1608 __ verify_oop(r->as_Register());
1609 }
1610 }
1611 }
1612 }
1613 }
1615 static void gen_special_dispatch(MacroAssembler* masm,
1616 methodHandle method,
1617 const BasicType* sig_bt,
1618 const VMRegPair* regs) {
1619 verify_oop_args(masm, method, sig_bt, regs);
1620 vmIntrinsics::ID iid = method->intrinsic_id();
1622 // Now write the args into the outgoing interpreter space
1623 bool has_receiver = false;
1624 Register receiver_reg = noreg;
1625 int member_arg_pos = -1;
1626 Register member_reg = noreg;
1627 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1628 if (ref_kind != 0) {
1629 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1630 member_reg = R19_method; // known to be free at this point
1631 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1632 } else if (iid == vmIntrinsics::_invokeBasic) {
1633 has_receiver = true;
1634 } else {
1635 fatal(err_msg_res("unexpected intrinsic id %d", iid));
1636 }
1638 if (member_reg != noreg) {
1639 // Load the member_arg into register, if necessary.
1640 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1641 VMReg r = regs[member_arg_pos].first();
1642 if (r->is_stack()) {
1643 __ ld(member_reg, reg2offset(r), R1_SP);
1644 } else {
1645 // no data motion is needed
1646 member_reg = r->as_Register();
1647 }
1648 }
1650 if (has_receiver) {
1651 // Make sure the receiver is loaded into a register.
1652 assert(method->size_of_parameters() > 0, "oob");
1653 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1654 VMReg r = regs[0].first();
1655 assert(r->is_valid(), "bad receiver arg");
1656 if (r->is_stack()) {
1657 // Porting note: This assumes that compiled calling conventions always
1658 // pass the receiver oop in a register. If this is not true on some
1659 // platform, pick a temp and load the receiver from stack.
1660 fatal("receiver always in a register");
1661 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point?
1662 __ ld(receiver_reg, reg2offset(r), R1_SP);
1663 } else {
1664 // no data motion is needed
1665 receiver_reg = r->as_Register();
1666 }
1667 }
1669 // Figure out which address we are really jumping to:
1670 MethodHandles::generate_method_handle_dispatch(masm, iid,
1671 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1672 }
1674 #endif // COMPILER2
1676 // ---------------------------------------------------------------------------
1677 // Generate a native wrapper for a given method. The method takes arguments
1678 // in the Java compiled code convention, marshals them to the native
1679 // convention (handlizes oops, etc), transitions to native, makes the call,
1680 // returns to java state (possibly blocking), unhandlizes any result and
1681 // returns.
1682 //
1683 // Critical native functions are a shorthand for the use of
1684 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1685 // functions. The wrapper is expected to unpack the arguments before
1686 // passing them to the callee and perform checks before and after the
1687 // native call to ensure that they GC_locker
1688 // lock_critical/unlock_critical semantics are followed. Some other
1689 // parts of JNI setup are skipped like the tear down of the JNI handle
1690 // block and the check for pending exceptions it's impossible for them
1691 // to be thrown.
1692 //
1693 // They are roughly structured like this:
1694 // if (GC_locker::needs_gc())
1695 // SharedRuntime::block_for_jni_critical();
1696 // tranistion to thread_in_native
1697 // unpack arrray arguments and call native entry point
1698 // check for safepoint in progress
1699 // check if any thread suspend flags are set
1700 // call into JVM and possible unlock the JNI critical
1701 // if a GC was suppressed while in the critical native.
1702 // transition back to thread_in_Java
1703 // return to caller
1704 //
1705 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
1706 methodHandle method,
1707 int compile_id,
1708 BasicType *in_sig_bt,
1709 VMRegPair *in_regs,
1710 BasicType ret_type) {
1711 #ifdef COMPILER2
1712 if (method->is_method_handle_intrinsic()) {
1713 vmIntrinsics::ID iid = method->intrinsic_id();
1714 intptr_t start = (intptr_t)__ pc();
1715 int vep_offset = ((intptr_t)__ pc()) - start;
1716 gen_special_dispatch(masm,
1717 method,
1718 in_sig_bt,
1719 in_regs);
1720 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1721 __ flush();
1722 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1723 return nmethod::new_native_nmethod(method,
1724 compile_id,
1725 masm->code(),
1726 vep_offset,
1727 frame_complete,
1728 stack_slots / VMRegImpl::slots_per_word,
1729 in_ByteSize(-1),
1730 in_ByteSize(-1),
1731 (OopMapSet*)NULL);
1732 }
1734 bool is_critical_native = true;
1735 address native_func = method->critical_native_function();
1736 if (native_func == NULL) {
1737 native_func = method->native_function();
1738 is_critical_native = false;
1739 }
1740 assert(native_func != NULL, "must have function");
1742 // First, create signature for outgoing C call
1743 // --------------------------------------------------------------------------
1745 int total_in_args = method->size_of_parameters();
1746 // We have received a description of where all the java args are located
1747 // on entry to the wrapper. We need to convert these args to where
1748 // the jni function will expect them. To figure out where they go
1749 // we convert the java signature to a C signature by inserting
1750 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1751 //
1752 // Additionally, on ppc64 we must convert integers to longs in the C
1753 // signature. We do this in advance in order to have no trouble with
1754 // indexes into the bt-arrays.
1755 // So convert the signature and registers now, and adjust the total number
1756 // of in-arguments accordingly.
1757 int i2l_argcnt = convert_ints_to_longints_argcnt(total_in_args, in_sig_bt); // PPC64: pass ints as longs.
1759 // Calculate the total number of C arguments and create arrays for the
1760 // signature and the outgoing registers.
1761 // On ppc64, we have two arrays for the outgoing registers, because
1762 // some floating-point arguments must be passed in registers _and_
1763 // in stack locations.
1764 bool method_is_static = method->is_static();
1765 int total_c_args = i2l_argcnt;
1767 if (!is_critical_native) {
1768 int n_hidden_args = method_is_static ? 2 : 1;
1769 total_c_args += n_hidden_args;
1770 } else {
1771 // No JNIEnv*, no this*, but unpacked arrays (base+length).
1772 for (int i = 0; i < total_in_args; i++) {
1773 if (in_sig_bt[i] == T_ARRAY) {
1774 total_c_args += 2; // PPC64: T_LONG, T_INT, T_ADDRESS (see convert_ints_to_longints and c_calling_convention)
1775 }
1776 }
1777 }
1779 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1780 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1781 VMRegPair *out_regs2 = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1782 BasicType* in_elem_bt = NULL;
1784 // Create the signature for the C call:
1785 // 1) add the JNIEnv*
1786 // 2) add the class if the method is static
1787 // 3) copy the rest of the incoming signature (shifted by the number of
1788 // hidden arguments).
1790 int argc = 0;
1791 if (!is_critical_native) {
1792 convert_ints_to_longints(i2l_argcnt, total_in_args, in_sig_bt, in_regs); // PPC64: pass ints as longs.
1794 out_sig_bt[argc++] = T_ADDRESS;
1795 if (method->is_static()) {
1796 out_sig_bt[argc++] = T_OBJECT;
1797 }
1799 for (int i = 0; i < total_in_args ; i++ ) {
1800 out_sig_bt[argc++] = in_sig_bt[i];
1801 }
1802 } else {
1803 Thread* THREAD = Thread::current();
1804 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, i2l_argcnt);
1805 SignatureStream ss(method->signature());
1806 int o = 0;
1807 for (int i = 0; i < total_in_args ; i++, o++) {
1808 if (in_sig_bt[i] == T_ARRAY) {
1809 // Arrays are passed as int, elem* pair
1810 Symbol* atype = ss.as_symbol(CHECK_NULL);
1811 const char* at = atype->as_C_string();
1812 if (strlen(at) == 2) {
1813 assert(at[0] == '[', "must be");
1814 switch (at[1]) {
1815 case 'B': in_elem_bt[o] = T_BYTE; break;
1816 case 'C': in_elem_bt[o] = T_CHAR; break;
1817 case 'D': in_elem_bt[o] = T_DOUBLE; break;
1818 case 'F': in_elem_bt[o] = T_FLOAT; break;
1819 case 'I': in_elem_bt[o] = T_INT; break;
1820 case 'J': in_elem_bt[o] = T_LONG; break;
1821 case 'S': in_elem_bt[o] = T_SHORT; break;
1822 case 'Z': in_elem_bt[o] = T_BOOLEAN; break;
1823 default: ShouldNotReachHere();
1824 }
1825 }
1826 } else {
1827 in_elem_bt[o] = T_VOID;
1828 switch(in_sig_bt[i]) { // PPC64: pass ints as longs.
1829 case T_BOOLEAN:
1830 case T_CHAR:
1831 case T_BYTE:
1832 case T_SHORT:
1833 case T_INT: in_elem_bt[++o] = T_VOID; break;
1834 default: break;
1835 }
1836 }
1837 if (in_sig_bt[i] != T_VOID) {
1838 assert(in_sig_bt[i] == ss.type(), "must match");
1839 ss.next();
1840 }
1841 }
1842 assert(i2l_argcnt==o, "must match");
1844 convert_ints_to_longints(i2l_argcnt, total_in_args, in_sig_bt, in_regs); // PPC64: pass ints as longs.
1846 for (int i = 0; i < total_in_args ; i++ ) {
1847 if (in_sig_bt[i] == T_ARRAY) {
1848 // Arrays are passed as int, elem* pair.
1849 out_sig_bt[argc++] = T_LONG; // PPC64: pass ints as longs.
1850 out_sig_bt[argc++] = T_INT;
1851 out_sig_bt[argc++] = T_ADDRESS;
1852 } else {
1853 out_sig_bt[argc++] = in_sig_bt[i];
1854 }
1855 }
1856 }
1859 // Compute the wrapper's frame size.
1860 // --------------------------------------------------------------------------
1862 // Now figure out where the args must be stored and how much stack space
1863 // they require.
1864 //
1865 // Compute framesize for the wrapper. We need to handlize all oops in
1866 // incoming registers.
1867 //
1868 // Calculate the total number of stack slots we will need:
1869 // 1) abi requirements
1870 // 2) outgoing arguments
1871 // 3) space for inbound oop handle area
1872 // 4) space for handlizing a klass if static method
1873 // 5) space for a lock if synchronized method
1874 // 6) workspace for saving return values, int <-> float reg moves, etc.
1875 // 7) alignment
1876 //
1877 // Layout of the native wrapper frame:
1878 // (stack grows upwards, memory grows downwards)
1879 //
1880 // NW [ABI_112] <-- 1) R1_SP
1881 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset
1882 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset (save area for critical natives)
1883 // klass <-- 4) R1_SP + klass_offset
1884 // lock <-- 5) R1_SP + lock_offset
1885 // [workspace] <-- 6) R1_SP + workspace_offset
1886 // [alignment] (optional) <-- 7)
1887 // caller [JIT_TOP_ABI_48] <-- r_callers_sp
1888 //
1889 // - *_slot_offset Indicates offset from SP in number of stack slots.
1890 // - *_offset Indicates offset from SP in bytes.
1892 int stack_slots = c_calling_convention(out_sig_bt, out_regs, out_regs2, total_c_args) // 1+2)
1893 + SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
1895 // Now the space for the inbound oop handle area.
1896 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
1897 if (is_critical_native) {
1898 // Critical natives may have to call out so they need a save area
1899 // for register arguments.
1900 int double_slots = 0;
1901 int single_slots = 0;
1902 for (int i = 0; i < total_in_args; i++) {
1903 if (in_regs[i].first()->is_Register()) {
1904 const Register reg = in_regs[i].first()->as_Register();
1905 switch (in_sig_bt[i]) {
1906 case T_BOOLEAN:
1907 case T_BYTE:
1908 case T_SHORT:
1909 case T_CHAR:
1910 case T_INT: /*single_slots++;*/ break; // PPC64: pass ints as longs.
1911 case T_ARRAY:
1912 case T_LONG: double_slots++; break;
1913 default: ShouldNotReachHere();
1914 }
1915 } else if (in_regs[i].first()->is_FloatRegister()) {
1916 switch (in_sig_bt[i]) {
1917 case T_FLOAT: single_slots++; break;
1918 case T_DOUBLE: double_slots++; break;
1919 default: ShouldNotReachHere();
1920 }
1921 }
1922 }
1923 total_save_slots = double_slots * 2 + round_to(single_slots, 2); // round to even
1924 }
1926 int oop_handle_slot_offset = stack_slots;
1927 stack_slots += total_save_slots; // 3)
1929 int klass_slot_offset = 0;
1930 int klass_offset = -1;
1931 if (method_is_static && !is_critical_native) { // 4)
1932 klass_slot_offset = stack_slots;
1933 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1934 stack_slots += VMRegImpl::slots_per_word;
1935 }
1937 int lock_slot_offset = 0;
1938 int lock_offset = -1;
1939 if (method->is_synchronized()) { // 5)
1940 lock_slot_offset = stack_slots;
1941 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
1942 stack_slots += VMRegImpl::slots_per_word;
1943 }
1945 int workspace_slot_offset = stack_slots; // 6)
1946 stack_slots += 2;
1948 // Now compute actual number of stack words we need.
1949 // Rounding to make stack properly aligned.
1950 stack_slots = round_to(stack_slots, // 7)
1951 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
1952 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
1955 // Now we can start generating code.
1956 // --------------------------------------------------------------------------
1958 intptr_t start_pc = (intptr_t)__ pc();
1959 intptr_t vep_start_pc;
1960 intptr_t frame_done_pc;
1961 intptr_t oopmap_pc;
1963 Label ic_miss;
1964 Label handle_pending_exception;
1966 Register r_callers_sp = R21;
1967 Register r_temp_1 = R22;
1968 Register r_temp_2 = R23;
1969 Register r_temp_3 = R24;
1970 Register r_temp_4 = R25;
1971 Register r_temp_5 = R26;
1972 Register r_temp_6 = R27;
1973 Register r_return_pc = R28;
1975 Register r_carg1_jnienv = noreg;
1976 Register r_carg2_classorobject = noreg;
1977 if (!is_critical_native) {
1978 r_carg1_jnienv = out_regs[0].first()->as_Register();
1979 r_carg2_classorobject = out_regs[1].first()->as_Register();
1980 }
1983 // Generate the Unverified Entry Point (UEP).
1984 // --------------------------------------------------------------------------
1985 assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
1987 // Check ic: object class == cached class?
1988 if (!method_is_static) {
1989 Register ic = as_Register(Matcher::inline_cache_reg_encode());
1990 Register receiver_klass = r_temp_1;
1992 __ cmpdi(CCR0, R3_ARG1, 0);
1993 __ beq(CCR0, ic_miss);
1994 __ verify_oop(R3_ARG1);
1995 __ load_klass(receiver_klass, R3_ARG1);
1997 __ cmpd(CCR0, receiver_klass, ic);
1998 __ bne(CCR0, ic_miss);
1999 }
2002 // Generate the Verified Entry Point (VEP).
2003 // --------------------------------------------------------------------------
2004 vep_start_pc = (intptr_t)__ pc();
2006 __ save_LR_CR(r_temp_1);
2007 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2008 __ mr(r_callers_sp, R1_SP); // Remember frame pointer.
2009 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame.
2010 frame_done_pc = (intptr_t)__ pc();
2012 // Native nmethod wrappers never take possesion of the oop arguments.
2013 // So the caller will gc the arguments.
2014 // The only thing we need an oopMap for is if the call is static.
2015 //
2016 // An OopMap for lock (and class if static), and one for the VM call itself.
2017 OopMapSet *oop_maps = new OopMapSet();
2018 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2020 if (is_critical_native) {
2021 check_needs_gc_for_critical_native(masm, stack_slots, total_in_args, oop_handle_slot_offset, oop_maps, in_regs, in_sig_bt, r_temp_1);
2022 }
2024 // Move arguments from register/stack to register/stack.
2025 // --------------------------------------------------------------------------
2026 //
2027 // We immediately shuffle the arguments so that for any vm call we have
2028 // to make from here on out (sync slow path, jvmti, etc.) we will have
2029 // captured the oops from our caller and have a valid oopMap for them.
2030 //
2031 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2032 // (derived from JavaThread* which is in R16_thread) and, if static,
2033 // the class mirror instead of a receiver. This pretty much guarantees that
2034 // register layout will not match. We ignore these extra arguments during
2035 // the shuffle. The shuffle is described by the two calling convention
2036 // vectors we have in our possession. We simply walk the java vector to
2037 // get the source locations and the c vector to get the destinations.
2039 // Record sp-based slot for receiver on stack for non-static methods.
2040 int receiver_offset = -1;
2042 // We move the arguments backward because the floating point registers
2043 // destination will always be to a register with a greater or equal
2044 // register number or the stack.
2045 // in is the index of the incoming Java arguments
2046 // out is the index of the outgoing C arguments
2048 #ifdef ASSERT
2049 bool reg_destroyed[RegisterImpl::number_of_registers];
2050 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2051 for (int r = 0 ; r < RegisterImpl::number_of_registers ; r++) {
2052 reg_destroyed[r] = false;
2053 }
2054 for (int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++) {
2055 freg_destroyed[f] = false;
2056 }
2057 #endif // ASSERT
2059 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2061 #ifdef ASSERT
2062 if (in_regs[in].first()->is_Register()) {
2063 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2064 } else if (in_regs[in].first()->is_FloatRegister()) {
2065 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2066 }
2067 if (out_regs[out].first()->is_Register()) {
2068 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2069 } else if (out_regs[out].first()->is_FloatRegister()) {
2070 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2071 }
2072 if (out_regs2[out].first()->is_Register()) {
2073 reg_destroyed[out_regs2[out].first()->as_Register()->encoding()] = true;
2074 } else if (out_regs2[out].first()->is_FloatRegister()) {
2075 freg_destroyed[out_regs2[out].first()->as_FloatRegister()->encoding()] = true;
2076 }
2077 #endif // ASSERT
2079 switch (in_sig_bt[in]) {
2080 case T_BOOLEAN:
2081 case T_CHAR:
2082 case T_BYTE:
2083 case T_SHORT:
2084 case T_INT:
2085 guarantee(in > 0 && in_sig_bt[in-1] == T_LONG,
2086 "expecting type (T_LONG,bt) for bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}");
2087 break;
2088 case T_LONG:
2089 if (in_sig_bt[in+1] == T_VOID) {
2090 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2091 } else {
2092 guarantee(in_sig_bt[in+1] == T_BOOLEAN || in_sig_bt[in+1] == T_CHAR ||
2093 in_sig_bt[in+1] == T_BYTE || in_sig_bt[in+1] == T_SHORT ||
2094 in_sig_bt[in+1] == T_INT,
2095 "expecting type (T_LONG,bt) for bt in {T_BOOLEAN, T_CHAR, T_BYTE, T_SHORT, T_INT}");
2096 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2097 }
2098 break;
2099 case T_ARRAY:
2100 if (is_critical_native) {
2101 int body_arg = out;
2102 out -= 2; // Point to length arg. PPC64: pass ints as longs.
2103 unpack_array_argument(masm, in_regs[in], in_elem_bt[in], out_regs[body_arg], out_regs[out],
2104 r_callers_sp, r_temp_1, r_temp_2);
2105 break;
2106 }
2107 case T_OBJECT:
2108 assert(!is_critical_native, "no oop arguments");
2109 object_move(masm, stack_slots,
2110 oop_map, oop_handle_slot_offset,
2111 ((in == 0) && (!method_is_static)), &receiver_offset,
2112 in_regs[in], out_regs[out],
2113 r_callers_sp, r_temp_1, r_temp_2);
2114 break;
2115 case T_VOID:
2116 break;
2117 case T_FLOAT:
2118 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2119 if (out_regs2[out].first()->is_valid()) {
2120 float_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1);
2121 }
2122 break;
2123 case T_DOUBLE:
2124 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2125 if (out_regs2[out].first()->is_valid()) {
2126 double_move(masm, in_regs[in], out_regs2[out], r_callers_sp, r_temp_1);
2127 }
2128 break;
2129 case T_ADDRESS:
2130 fatal("found type (T_ADDRESS) in java args");
2131 break;
2132 default:
2133 ShouldNotReachHere();
2134 break;
2135 }
2136 }
2138 // Pre-load a static method's oop into ARG2.
2139 // Used both by locking code and the normal JNI call code.
2140 if (method_is_static && !is_critical_native) {
2141 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2142 r_carg2_classorobject);
2144 // Now handlize the static class mirror in carg2. It's known not-null.
2145 __ std(r_carg2_classorobject, klass_offset, R1_SP);
2146 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2147 __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2148 }
2150 // Get JNIEnv* which is first argument to native.
2151 if (!is_critical_native) {
2152 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2153 }
2155 // NOTE:
2156 //
2157 // We have all of the arguments setup at this point.
2158 // We MUST NOT touch any outgoing regs from this point on.
2159 // So if we must call out we must push a new frame.
2161 // Get current pc for oopmap, and load it patchable relative to global toc.
2162 oopmap_pc = (intptr_t) __ pc();
2163 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true);
2165 // We use the same pc/oopMap repeatedly when we call out.
2166 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2168 // r_return_pc now has the pc loaded that we will use when we finally call
2169 // to native.
2171 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2172 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2175 # if 0
2176 // DTrace method entry
2177 # endif
2179 // Lock a synchronized method.
2180 // --------------------------------------------------------------------------
2182 if (method->is_synchronized()) {
2183 assert(!is_critical_native, "unhandled");
2184 ConditionRegister r_flag = CCR1;
2185 Register r_oop = r_temp_4;
2186 const Register r_box = r_temp_5;
2187 Label done, locked;
2189 // Load the oop for the object or class. r_carg2_classorobject contains
2190 // either the handlized oop from the incoming arguments or the handlized
2191 // class mirror (if the method is static).
2192 __ ld(r_oop, 0, r_carg2_classorobject);
2194 // Get the lock box slot's address.
2195 __ addi(r_box, R1_SP, lock_offset);
2197 # ifdef ASSERT
2198 if (UseBiasedLocking) {
2199 // Making the box point to itself will make it clear it went unused
2200 // but also be obviously invalid.
2201 __ std(r_box, 0, r_box);
2202 }
2203 # endif // ASSERT
2205 // Try fastpath for locking.
2206 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2207 __ compiler_fast_lock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2208 __ beq(r_flag, locked);
2210 // None of the above fast optimizations worked so we have to get into the
2211 // slow case of monitor enter. Inline a special case of call_VM that
2212 // disallows any pending_exception.
2214 // Save argument registers and leave room for C-compatible ABI_112.
2215 int frame_size = frame::abi_112_size +
2216 round_to(total_c_args * wordSize, frame::alignment_in_bytes);
2217 __ mr(R11_scratch1, R1_SP);
2218 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs, out_regs2);
2220 // Do the call.
2221 __ set_last_Java_frame(R11_scratch1, r_return_pc);
2222 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register");
2223 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2224 __ reset_last_Java_frame();
2226 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs, out_regs2);
2228 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2229 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C", 0);
2231 __ bind(locked);
2232 }
2235 // Publish thread state
2236 // --------------------------------------------------------------------------
2238 // Use that pc we placed in r_return_pc a while back as the current frame anchor.
2239 __ set_last_Java_frame(R1_SP, r_return_pc);
2241 // Transition from _thread_in_Java to _thread_in_native.
2242 __ li(R0, _thread_in_native);
2243 __ release();
2244 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2245 __ stw(R0, thread_(thread_state));
2246 if (UseMembar) {
2247 __ fence();
2248 }
2251 // The JNI call
2252 // --------------------------------------------------------------------------
2254 FunctionDescriptor* fd_native_method = (FunctionDescriptor*) native_func;
2255 __ call_c(fd_native_method, relocInfo::runtime_call_type);
2258 // Now, we are back from the native code.
2261 // Unpack the native result.
2262 // --------------------------------------------------------------------------
2264 // For int-types, we do any needed sign-extension required.
2265 // Care must be taken that the return values (R3_RET and F1_RET)
2266 // will survive any VM calls for blocking or unlocking.
2267 // An OOP result (handle) is done specially in the slow-path code.
2269 switch (ret_type) {
2270 case T_VOID: break; // Nothing to do!
2271 case T_FLOAT: break; // Got it where we want it (unless slow-path).
2272 case T_DOUBLE: break; // Got it where we want it (unless slow-path).
2273 case T_LONG: break; // Got it where we want it (unless slow-path).
2274 case T_OBJECT: break; // Really a handle.
2275 // Cannot de-handlize until after reclaiming jvm_lock.
2276 case T_ARRAY: break;
2278 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1)
2279 Label skip_modify;
2280 __ cmpwi(CCR0, R3_RET, 0);
2281 __ beq(CCR0, skip_modify);
2282 __ li(R3_RET, 1);
2283 __ bind(skip_modify);
2284 break;
2285 }
2286 case T_BYTE: { // sign extension
2287 __ extsb(R3_RET, R3_RET);
2288 break;
2289 }
2290 case T_CHAR: { // unsigned result
2291 __ andi(R3_RET, R3_RET, 0xffff);
2292 break;
2293 }
2294 case T_SHORT: { // sign extension
2295 __ extsh(R3_RET, R3_RET);
2296 break;
2297 }
2298 case T_INT: // nothing to do
2299 break;
2300 default:
2301 ShouldNotReachHere();
2302 break;
2303 }
2306 // Publish thread state
2307 // --------------------------------------------------------------------------
2309 // Switch thread to "native transition" state before reading the
2310 // synchronization state. This additional state is necessary because reading
2311 // and testing the synchronization state is not atomic w.r.t. GC, as this
2312 // scenario demonstrates:
2313 // - Java thread A, in _thread_in_native state, loads _not_synchronized
2314 // and is preempted.
2315 // - VM thread changes sync state to synchronizing and suspends threads
2316 // for GC.
2317 // - Thread A is resumed to finish this native method, but doesn't block
2318 // here since it didn't see any synchronization in progress, and escapes.
2320 // Transition from _thread_in_native to _thread_in_native_trans.
2321 __ li(R0, _thread_in_native_trans);
2322 __ release();
2323 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2324 __ stw(R0, thread_(thread_state));
2327 // Must we block?
2328 // --------------------------------------------------------------------------
2330 // Block, if necessary, before resuming in _thread_in_Java state.
2331 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2332 Label after_transition;
2333 {
2334 Label no_block, sync;
2336 if (os::is_MP()) {
2337 if (UseMembar) {
2338 // Force this write out before the read below.
2339 __ fence();
2340 } else {
2341 // Write serialization page so VM thread can do a pseudo remote membar.
2342 // We use the current thread pointer to calculate a thread specific
2343 // offset to write to within the page. This minimizes bus traffic
2344 // due to cache line collision.
2345 __ serialize_memory(R16_thread, r_temp_4, r_temp_5);
2346 }
2347 }
2349 Register sync_state_addr = r_temp_4;
2350 Register sync_state = r_temp_5;
2351 Register suspend_flags = r_temp_6;
2353 __ load_const(sync_state_addr, SafepointSynchronize::address_of_state(), /*temp*/ sync_state);
2355 // TODO: PPC port assert(4 == SafepointSynchronize::sz_state(), "unexpected field size");
2356 __ lwz(sync_state, 0, sync_state_addr);
2358 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2359 __ lwz(suspend_flags, thread_(suspend_flags));
2361 __ acquire();
2363 Label do_safepoint;
2364 // No synchronization in progress nor yet synchronized.
2365 __ cmpwi(CCR0, sync_state, SafepointSynchronize::_not_synchronized);
2366 // Not suspended.
2367 __ cmpwi(CCR1, suspend_flags, 0);
2369 __ bne(CCR0, sync);
2370 __ beq(CCR1, no_block);
2372 // Block. Save any potential method result value before the operation and
2373 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2374 // lets us share the oopMap we used when we went native rather than create
2375 // a distinct one for this pc.
2376 __ bind(sync);
2378 address entry_point = is_critical_native
2379 ? CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition)
2380 : CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2381 save_native_result(masm, ret_type, workspace_slot_offset);
2382 __ call_VM_leaf(entry_point, R16_thread);
2383 restore_native_result(masm, ret_type, workspace_slot_offset);
2385 if (is_critical_native) {
2386 __ b(after_transition); // No thread state transition here.
2387 }
2388 __ bind(no_block);
2389 }
2391 // Publish thread state.
2392 // --------------------------------------------------------------------------
2394 // Thread state is thread_in_native_trans. Any safepoint blocking has
2395 // already happened so we can now change state to _thread_in_Java.
2397 // Transition from _thread_in_native_trans to _thread_in_Java.
2398 __ li(R0, _thread_in_Java);
2399 __ release();
2400 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2401 __ stw(R0, thread_(thread_state));
2402 if (UseMembar) {
2403 __ fence();
2404 }
2405 __ bind(after_transition);
2407 // Reguard any pages if necessary.
2408 // --------------------------------------------------------------------------
2410 Label no_reguard;
2411 __ lwz(r_temp_1, thread_(stack_guard_state));
2412 __ cmpwi(CCR0, r_temp_1, JavaThread::stack_guard_yellow_disabled);
2413 __ bne(CCR0, no_reguard);
2415 save_native_result(masm, ret_type, workspace_slot_offset);
2416 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2417 restore_native_result(masm, ret_type, workspace_slot_offset);
2419 __ bind(no_reguard);
2422 // Unlock
2423 // --------------------------------------------------------------------------
2425 if (method->is_synchronized()) {
2427 ConditionRegister r_flag = CCR1;
2428 const Register r_oop = r_temp_4;
2429 const Register r_box = r_temp_5;
2430 const Register r_exception = r_temp_6;
2431 Label done;
2433 // Get oop and address of lock object box.
2434 if (method_is_static) {
2435 assert(klass_offset != -1, "");
2436 __ ld(r_oop, klass_offset, R1_SP);
2437 } else {
2438 assert(receiver_offset != -1, "");
2439 __ ld(r_oop, receiver_offset, R1_SP);
2440 }
2441 __ addi(r_box, R1_SP, lock_offset);
2443 // Try fastpath for unlocking.
2444 __ compiler_fast_unlock_object(r_flag, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2445 __ beq(r_flag, done);
2447 // Save and restore any potential method result value around the unlocking operation.
2448 save_native_result(masm, ret_type, workspace_slot_offset);
2450 // Must save pending exception around the slow-path VM call. Since it's a
2451 // leaf call, the pending exception (if any) can be kept in a register.
2452 __ ld(r_exception, thread_(pending_exception));
2453 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2454 __ li(R0, 0);
2455 __ std(R0, thread_(pending_exception));
2457 // Slow case of monitor enter.
2458 // Inline a special case of call_VM that disallows any pending_exception.
2459 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box);
2461 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2462 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C", 0);
2464 restore_native_result(masm, ret_type, workspace_slot_offset);
2466 // Check_forward_pending_exception jump to forward_exception if any pending
2467 // exception is set. The forward_exception routine expects to see the
2468 // exception in pending_exception and not in a register. Kind of clumsy,
2469 // since all folks who branch to forward_exception must have tested
2470 // pending_exception first and hence have it in a register already.
2471 __ std(r_exception, thread_(pending_exception));
2473 __ bind(done);
2474 }
2476 # if 0
2477 // DTrace method exit
2478 # endif
2480 // Clear "last Java frame" SP and PC.
2481 // --------------------------------------------------------------------------
2483 __ reset_last_Java_frame();
2485 // Unpack oop result.
2486 // --------------------------------------------------------------------------
2488 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2489 Label skip_unboxing;
2490 __ cmpdi(CCR0, R3_RET, 0);
2491 __ beq(CCR0, skip_unboxing);
2492 __ ld(R3_RET, 0, R3_RET);
2493 __ bind(skip_unboxing);
2494 __ verify_oop(R3_RET);
2495 }
2498 // Reset handle block.
2499 // --------------------------------------------------------------------------
2500 if (!is_critical_native) {
2501 __ ld(r_temp_1, thread_(active_handles));
2502 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2503 __ li(r_temp_2, 0);
2504 __ stw(r_temp_2, JNIHandleBlock::top_offset_in_bytes(), r_temp_1);
2507 // Check for pending exceptions.
2508 // --------------------------------------------------------------------------
2509 __ ld(r_temp_2, thread_(pending_exception));
2510 __ cmpdi(CCR0, r_temp_2, 0);
2511 __ bne(CCR0, handle_pending_exception);
2512 }
2514 // Return
2515 // --------------------------------------------------------------------------
2517 __ pop_frame();
2518 __ restore_LR_CR(R11);
2519 __ blr();
2522 // Handler for pending exceptions (out-of-line).
2523 // --------------------------------------------------------------------------
2525 // Since this is a native call, we know the proper exception handler
2526 // is the empty function. We just pop this frame and then jump to
2527 // forward_exception_entry.
2528 if (!is_critical_native) {
2529 __ align(InteriorEntryAlignment);
2530 __ bind(handle_pending_exception);
2532 __ pop_frame();
2533 __ restore_LR_CR(R11);
2534 __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2535 relocInfo::runtime_call_type);
2536 }
2538 // Handler for a cache miss (out-of-line).
2539 // --------------------------------------------------------------------------
2541 if (!method_is_static) {
2542 __ align(InteriorEntryAlignment);
2543 __ bind(ic_miss);
2545 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(),
2546 relocInfo::runtime_call_type);
2547 }
2549 // Done.
2550 // --------------------------------------------------------------------------
2552 __ flush();
2554 nmethod *nm = nmethod::new_native_nmethod(method,
2555 compile_id,
2556 masm->code(),
2557 vep_start_pc-start_pc,
2558 frame_done_pc-start_pc,
2559 stack_slots / VMRegImpl::slots_per_word,
2560 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2561 in_ByteSize(lock_offset),
2562 oop_maps);
2564 if (is_critical_native) {
2565 nm->set_lazy_critical_native(true);
2566 }
2568 return nm;
2569 #else
2570 ShouldNotReachHere();
2571 return NULL;
2572 #endif // COMPILER2
2573 }
2575 // This function returns the adjust size (in number of words) to a c2i adapter
2576 // activation for use during deoptimization.
2577 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2578 return round_to((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::alignment_in_bytes);
2579 }
2581 uint SharedRuntime::out_preserve_stack_slots() {
2582 #ifdef COMPILER2
2583 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2584 #else
2585 return 0;
2586 #endif
2587 }
2589 #ifdef COMPILER2
2590 // Frame generation for deopt and uncommon trap blobs.
2591 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2592 /* Read */
2593 Register unroll_block_reg,
2594 /* Update */
2595 Register frame_sizes_reg,
2596 Register number_of_frames_reg,
2597 Register pcs_reg,
2598 /* Invalidate */
2599 Register frame_size_reg,
2600 Register pc_reg) {
2602 __ ld(pc_reg, 0, pcs_reg);
2603 __ ld(frame_size_reg, 0, frame_sizes_reg);
2604 __ std(pc_reg, _abi(lr), R1_SP);
2605 __ push_frame(frame_size_reg, R0/*tmp*/);
2606 #ifdef CC_INTERP
2607 __ std(R1_SP, _parent_ijava_frame_abi(initial_caller_sp), R1_SP);
2608 #else
2609 Unimplemented();
2610 #endif
2611 __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2612 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2613 __ addi(pcs_reg, pcs_reg, wordSize);
2614 }
2616 // Loop through the UnrollBlock info and create new frames.
2617 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2618 /* read */
2619 Register unroll_block_reg,
2620 /* invalidate */
2621 Register frame_sizes_reg,
2622 Register number_of_frames_reg,
2623 Register pcs_reg,
2624 Register frame_size_reg,
2625 Register pc_reg) {
2626 Label loop;
2628 // _number_of_frames is of type int (deoptimization.hpp)
2629 __ lwa(number_of_frames_reg,
2630 Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(),
2631 unroll_block_reg);
2632 __ ld(pcs_reg,
2633 Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(),
2634 unroll_block_reg);
2635 __ ld(frame_sizes_reg,
2636 Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(),
2637 unroll_block_reg);
2639 // stack: (caller_of_deoptee, ...).
2641 // At this point we either have an interpreter frame or a compiled
2642 // frame on top of stack. If it is a compiled frame we push a new c2i
2643 // adapter here
2645 // Memorize top-frame stack-pointer.
2646 __ mr(frame_size_reg/*old_sp*/, R1_SP);
2648 // Resize interpreter top frame OR C2I adapter.
2650 // At this moment, the top frame (which is the caller of the deoptee) is
2651 // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2652 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2653 // outgoing arguments.
2654 //
2655 // In order to push the interpreter frame for the deoptee, we need to
2656 // resize the top frame such that we are able to place the deoptee's
2657 // locals in the frame.
2658 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2659 // into a valid PARENT_IJAVA_FRAME_ABI.
2661 __ lwa(R11_scratch1,
2662 Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(),
2663 unroll_block_reg);
2664 __ neg(R11_scratch1, R11_scratch1);
2666 // R11_scratch1 contains size of locals for frame resizing.
2667 // R12_scratch2 contains top frame's lr.
2669 // Resize frame by complete frame size prevents TOC from being
2670 // overwritten by locals. A more stack space saving way would be
2671 // to copy the TOC to its location in the new abi.
2672 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2674 // now, resize the frame
2675 __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2677 // In the case where we have resized a c2i frame above, the optional
2678 // alignment below the locals has size 32 (why?).
2679 __ std(R12_scratch2, _abi(lr), R1_SP);
2681 // Initialize initial_caller_sp.
2682 __ std(frame_size_reg/*old_sp*/, _parent_ijava_frame_abi(initial_caller_sp), R1_SP);
2684 #ifdef ASSERT
2685 // Make sure that there is at least one entry in the array.
2686 __ cmpdi(CCR0, number_of_frames_reg, 0);
2687 __ asm_assert_ne("array_size must be > 0", 0x205);
2688 #endif
2690 // Now push the new interpreter frames.
2691 //
2692 __ bind(loop);
2693 // Allocate a new frame, fill in the pc.
2694 push_skeleton_frame(masm, deopt,
2695 unroll_block_reg,
2696 frame_sizes_reg,
2697 number_of_frames_reg,
2698 pcs_reg,
2699 frame_size_reg,
2700 pc_reg);
2701 __ cmpdi(CCR0, number_of_frames_reg, 0);
2702 __ bne(CCR0, loop);
2704 // Get the return address pointing into the frame manager.
2705 __ ld(R0, 0, pcs_reg);
2706 // Store it in the top interpreter frame.
2707 __ std(R0, _abi(lr), R1_SP);
2708 // Initialize frame_manager_lr of interpreter top frame.
2709 #ifdef CC_INTERP
2710 __ std(R0, _top_ijava_frame_abi(frame_manager_lr), R1_SP);
2711 #endif
2712 }
2713 #endif
2715 void SharedRuntime::generate_deopt_blob() {
2716 // Allocate space for the code
2717 ResourceMark rm;
2718 // Setup code generation tools
2719 CodeBuffer buffer("deopt_blob", 2048, 1024);
2720 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2721 Label exec_mode_initialized;
2722 int frame_size_in_words;
2723 OopMap* map = NULL;
2724 OopMapSet *oop_maps = new OopMapSet();
2726 // size of ABI112 plus spill slots for R3_RET and F1_RET.
2727 const int frame_size_in_bytes = frame::abi_112_spill_size;
2728 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2729 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2731 const Register exec_mode_reg = R21_tmp1;
2733 const address start = __ pc();
2735 #ifdef COMPILER2
2736 // --------------------------------------------------------------------------
2737 // Prolog for non exception case!
2739 // We have been called from the deopt handler of the deoptee.
2740 //
2741 // deoptee:
2742 // ...
2743 // call X
2744 // ...
2745 // deopt_handler: call_deopt_stub
2746 // cur. return pc --> ...
2747 //
2748 // So currently SR_LR points behind the call in the deopt handler.
2749 // We adjust it such that it points to the start of the deopt handler.
2750 // The return_pc has been stored in the frame of the deoptee and
2751 // will replace the address of the deopt_handler in the call
2752 // to Deoptimization::fetch_unroll_info below.
2753 // We can't grab a free register here, because all registers may
2754 // contain live values, so let the RegisterSaver do the adjustment
2755 // of the return pc.
2756 const int return_pc_adjustment_no_exception = -size_deopt_handler();
2758 // Push the "unpack frame"
2759 // Save everything in sight.
2760 map = RegisterSaver::push_frame_abi112_and_save_live_registers(masm,
2761 &first_frame_size_in_bytes,
2762 /*generate_oop_map=*/ true,
2763 return_pc_adjustment_no_exception,
2764 RegisterSaver::return_pc_is_lr);
2765 assert(map != NULL, "OopMap must have been created");
2767 __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
2768 // Save exec mode for unpack_frames.
2769 __ b(exec_mode_initialized);
2771 // --------------------------------------------------------------------------
2772 // Prolog for exception case
2774 // An exception is pending.
2775 // We have been called with a return (interpreter) or a jump (exception blob).
2776 //
2777 // - R3_ARG1: exception oop
2778 // - R4_ARG2: exception pc
2780 int exception_offset = __ pc() - start;
2782 BLOCK_COMMENT("Prolog for exception case");
2784 // The RegisterSaves doesn't need to adjust the return pc for this situation.
2785 const int return_pc_adjustment_exception = 0;
2787 // Push the "unpack frame".
2788 // Save everything in sight.
2789 assert(R4 == R4_ARG2, "exception pc must be in r4");
2790 RegisterSaver::push_frame_abi112_and_save_live_registers(masm,
2791 &first_frame_size_in_bytes,
2792 /*generate_oop_map=*/ false,
2793 return_pc_adjustment_exception,
2794 RegisterSaver::return_pc_is_r4);
2796 // Deopt during an exception. Save exec mode for unpack_frames.
2797 __ li(exec_mode_reg, Deoptimization::Unpack_exception);
2799 // Store exception oop and pc in thread (location known to GC).
2800 // This is needed since the call to "fetch_unroll_info()" may safepoint.
2801 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2802 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2804 // fall through
2806 // --------------------------------------------------------------------------
2807 __ BIND(exec_mode_initialized);
2809 {
2810 const Register unroll_block_reg = R22_tmp2;
2812 // We need to set `last_Java_frame' because `fetch_unroll_info' will
2813 // call `last_Java_frame()'. The value of the pc in the frame is not
2814 // particularly important. It just needs to identify this blob.
2815 __ set_last_Java_frame(R1_SP, noreg);
2817 // With EscapeAnalysis turned on, this call may safepoint!
2818 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread);
2819 address calls_return_pc = __ last_calls_return_pc();
2820 // Set an oopmap for the call site that describes all our saved registers.
2821 oop_maps->add_gc_map(calls_return_pc - start, map);
2823 __ reset_last_Java_frame();
2824 // Save the return value.
2825 __ mr(unroll_block_reg, R3_RET);
2827 // Restore only the result registers that have been saved
2828 // by save_volatile_registers(...).
2829 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes);
2831 // In excp_deopt_mode, restore and clear exception oop which we
2832 // stored in the thread during exception entry above. The exception
2833 // oop will be the return value of this stub.
2834 Label skip_restore_excp;
2835 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception);
2836 __ bne(CCR0, skip_restore_excp);
2837 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2838 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2839 __ li(R0, 0);
2840 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2841 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2842 __ BIND(skip_restore_excp);
2844 // reload narrro_oop_base
2845 if (UseCompressedOops && Universe::narrow_oop_base() != 0) {
2846 __ load_const_optimized(R30, Universe::narrow_oop_base());
2847 }
2849 __ pop_frame();
2851 // stack: (deoptee, optional i2c, caller of deoptee, ...).
2853 // pop the deoptee's frame
2854 __ pop_frame();
2856 // stack: (caller_of_deoptee, ...).
2858 // Loop through the `UnrollBlock' info and create interpreter frames.
2859 push_skeleton_frames(masm, true/*deopt*/,
2860 unroll_block_reg,
2861 R23_tmp3,
2862 R24_tmp4,
2863 R25_tmp5,
2864 R26_tmp6,
2865 R27_tmp7);
2867 // stack: (skeletal interpreter frame, ..., optional skeletal
2868 // interpreter frame, optional c2i, caller of deoptee, ...).
2869 }
2871 // push an `unpack_frame' taking care of float / int return values.
2872 __ push_frame(frame_size_in_bytes, R0/*tmp*/);
2874 // stack: (unpack frame, skeletal interpreter frame, ..., optional
2875 // skeletal interpreter frame, optional c2i, caller of deoptee,
2876 // ...).
2878 // Spill live volatile registers since we'll do a call.
2879 __ std( R3_RET, _abi_112_spill(spill_ret), R1_SP);
2880 __ stfd(F1_RET, _abi_112_spill(spill_fret), R1_SP);
2882 // Let the unpacker layout information in the skeletal frames just
2883 // allocated.
2884 __ get_PC_trash_LR(R3_RET);
2885 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
2886 // This is a call to a LEAF method, so no oop map is required.
2887 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
2888 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
2889 __ reset_last_Java_frame();
2891 // Restore the volatiles saved above.
2892 __ ld( R3_RET, _abi_112_spill(spill_ret), R1_SP);
2893 __ lfd(F1_RET, _abi_112_spill(spill_fret), R1_SP);
2895 // Pop the unpack frame.
2896 __ pop_frame();
2897 __ restore_LR_CR(R0);
2899 // stack: (top interpreter frame, ..., optional interpreter frame,
2900 // optional c2i, caller of deoptee, ...).
2902 // Initialize R14_state.
2903 __ ld(R14_state, 0, R1_SP);
2904 __ addi(R14_state, R14_state, -frame::interpreter_frame_cinterpreterstate_size_in_bytes());
2905 // Also inititialize R15_prev_state.
2906 __ restore_prev_state();
2908 // Return to the interpreter entry point.
2909 __ blr();
2910 __ flush();
2911 #else // COMPILER2
2912 __ unimplemented("deopt blob needed only with compiler");
2913 int exception_offset = __ pc() - start;
2914 #endif // COMPILER2
2916 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, 0, first_frame_size_in_bytes / wordSize);
2917 }
2919 #ifdef COMPILER2
2920 void SharedRuntime::generate_uncommon_trap_blob() {
2921 // Allocate space for the code.
2922 ResourceMark rm;
2923 // Setup code generation tools.
2924 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
2925 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2926 address start = __ pc();
2928 Register unroll_block_reg = R21_tmp1;
2929 Register klass_index_reg = R22_tmp2;
2930 Register unc_trap_reg = R23_tmp3;
2932 OopMapSet* oop_maps = new OopMapSet();
2933 int frame_size_in_bytes = frame::abi_112_size;
2934 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
2936 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
2938 // Push a dummy `unpack_frame' and call
2939 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
2940 // vframe array and return the `UnrollBlock' information.
2942 // Save LR to compiled frame.
2943 __ save_LR_CR(R11_scratch1);
2945 // Push an "uncommon_trap" frame.
2946 __ push_frame_abi112(0, R11_scratch1);
2948 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
2950 // Set the `unpack_frame' as last_Java_frame.
2951 // `Deoptimization::uncommon_trap' expects it and considers its
2952 // sender frame as the deoptee frame.
2953 // Remember the offset of the instruction whose address will be
2954 // moved to R11_scratch1.
2955 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
2957 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
2959 __ mr(klass_index_reg, R3);
2960 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
2961 R16_thread, klass_index_reg);
2963 // Set an oopmap for the call site.
2964 oop_maps->add_gc_map(gc_map_pc - start, map);
2966 __ reset_last_Java_frame();
2968 // Pop the `unpack frame'.
2969 __ pop_frame();
2971 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
2973 // Save the return value.
2974 __ mr(unroll_block_reg, R3_RET);
2976 // Pop the uncommon_trap frame.
2977 __ pop_frame();
2979 // stack: (caller_of_deoptee, ...).
2981 // Allocate new interpreter frame(s) and possibly a c2i adapter
2982 // frame.
2983 push_skeleton_frames(masm, false/*deopt*/,
2984 unroll_block_reg,
2985 R22_tmp2,
2986 R23_tmp3,
2987 R24_tmp4,
2988 R25_tmp5,
2989 R26_tmp6);
2991 // stack: (skeletal interpreter frame, ..., optional skeletal
2992 // interpreter frame, optional c2i, caller of deoptee, ...).
2994 // Push a dummy `unpack_frame' taking care of float return values.
2995 // Call `Deoptimization::unpack_frames' to layout information in the
2996 // interpreter frames just created.
2998 // Push a simple "unpack frame" here.
2999 __ push_frame_abi112(0, R11_scratch1);
3001 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3002 // skeletal interpreter frame, optional c2i, caller of deoptee,
3003 // ...).
3005 // Set the "unpack_frame" as last_Java_frame.
3006 __ get_PC_trash_LR(R11_scratch1);
3007 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
3009 // Indicate it is the uncommon trap case.
3010 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3011 // Let the unpacker layout information in the skeletal frames just
3012 // allocated.
3013 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3014 R16_thread, unc_trap_reg);
3016 __ reset_last_Java_frame();
3017 // Pop the `unpack frame'.
3018 __ pop_frame();
3019 // Restore LR from top interpreter frame.
3020 __ restore_LR_CR(R11_scratch1);
3022 // stack: (top interpreter frame, ..., optional interpreter frame,
3023 // optional c2i, caller of deoptee, ...).
3025 // Initialize R14_state, ...
3026 __ ld(R11_scratch1, 0, R1_SP);
3027 __ addi(R14_state, R11_scratch1, -frame::interpreter_frame_cinterpreterstate_size_in_bytes());
3028 // also initialize R15_prev_state.
3029 __ restore_prev_state();
3030 // Return to the interpreter entry point.
3031 __ blr();
3033 masm->flush();
3035 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3036 }
3037 #endif // COMPILER2
3039 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3040 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3041 assert(StubRoutines::forward_exception_entry() != NULL,
3042 "must be generated before");
3044 ResourceMark rm;
3045 OopMapSet *oop_maps = new OopMapSet();
3046 OopMap* map;
3048 // Allocate space for the code. Setup code generation tools.
3049 CodeBuffer buffer("handler_blob", 2048, 1024);
3050 MacroAssembler* masm = new MacroAssembler(&buffer);
3052 address start = __ pc();
3053 int frame_size_in_bytes = 0;
3055 RegisterSaver::ReturnPCLocation return_pc_location;
3056 bool cause_return = (poll_type == POLL_AT_RETURN);
3057 if (cause_return) {
3058 // Nothing to do here. The frame has already been popped in MachEpilogNode.
3059 // Register LR already contains the return pc.
3060 return_pc_location = RegisterSaver::return_pc_is_lr;
3061 } else {
3062 // Use thread()->saved_exception_pc() as return pc.
3063 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3064 }
3066 // Save registers, fpu state, and flags.
3067 map = RegisterSaver::push_frame_abi112_and_save_live_registers(masm,
3068 &frame_size_in_bytes,
3069 /*generate_oop_map=*/ true,
3070 /*return_pc_adjustment=*/0,
3071 return_pc_location);
3073 // The following is basically a call_VM. However, we need the precise
3074 // address of the call in order to generate an oopmap. Hence, we do all the
3075 // work outselves.
3076 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3078 // The return address must always be correct so that the frame constructor
3079 // never sees an invalid pc.
3081 // Do the call
3082 __ call_VM_leaf(call_ptr, R16_thread);
3083 address calls_return_pc = __ last_calls_return_pc();
3085 // Set an oopmap for the call site. This oopmap will map all
3086 // oop-registers and debug-info registers as callee-saved. This
3087 // will allow deoptimization at this safepoint to find all possible
3088 // debug-info recordings, as well as let GC find all oops.
3089 oop_maps->add_gc_map(calls_return_pc - start, map);
3091 Label noException;
3093 // Clear the last Java frame.
3094 __ reset_last_Java_frame();
3096 BLOCK_COMMENT(" Check pending exception.");
3097 const Register pending_exception = R0;
3098 __ ld(pending_exception, thread_(pending_exception));
3099 __ cmpdi(CCR0, pending_exception, 0);
3100 __ beq(CCR0, noException);
3102 // Exception pending
3103 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3104 frame_size_in_bytes,
3105 /*restore_ctr=*/true);
3108 BLOCK_COMMENT(" Jump to forward_exception_entry.");
3109 // Jump to forward_exception_entry, with the issuing PC in LR
3110 // so it looks like the original nmethod called forward_exception_entry.
3111 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3113 // No exception case.
3114 __ BIND(noException);
3117 // Normal exit, restore registers and exit.
3118 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3119 frame_size_in_bytes,
3120 /*restore_ctr=*/true);
3122 __ blr();
3124 // Make sure all code is generated
3125 masm->flush();
3127 // Fill-out other meta info
3128 // CodeBlob frame size is in words.
3129 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3130 }
3132 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3133 //
3134 // Generate a stub that calls into the vm to find out the proper destination
3135 // of a java call. All the argument registers are live at this point
3136 // but since this is generic code we don't know what they are and the caller
3137 // must do any gc of the args.
3138 //
3139 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3141 // allocate space for the code
3142 ResourceMark rm;
3144 CodeBuffer buffer(name, 1000, 512);
3145 MacroAssembler* masm = new MacroAssembler(&buffer);
3147 int frame_size_in_bytes;
3149 OopMapSet *oop_maps = new OopMapSet();
3150 OopMap* map = NULL;
3152 address start = __ pc();
3154 map = RegisterSaver::push_frame_abi112_and_save_live_registers(masm,
3155 &frame_size_in_bytes,
3156 /*generate_oop_map*/ true,
3157 /*return_pc_adjustment*/ 0,
3158 RegisterSaver::return_pc_is_lr);
3160 // Use noreg as last_Java_pc, the return pc will be reconstructed
3161 // from the physical frame.
3162 __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3164 int frame_complete = __ offset();
3166 // Pass R19_method as 2nd (optional) argument, used by
3167 // counter_overflow_stub.
3168 __ call_VM_leaf(destination, R16_thread, R19_method);
3169 address calls_return_pc = __ last_calls_return_pc();
3170 // Set an oopmap for the call site.
3171 // We need this not only for callee-saved registers, but also for volatile
3172 // registers that the compiler might be keeping live across a safepoint.
3173 // Create the oopmap for the call's return pc.
3174 oop_maps->add_gc_map(calls_return_pc - start, map);
3176 // R3_RET contains the address we are going to jump to assuming no exception got installed.
3178 // clear last_Java_sp
3179 __ reset_last_Java_frame();
3181 // Check for pending exceptions.
3182 BLOCK_COMMENT("Check for pending exceptions.");
3183 Label pending;
3184 __ ld(R11_scratch1, thread_(pending_exception));
3185 __ cmpdi(CCR0, R11_scratch1, 0);
3186 __ bne(CCR0, pending);
3188 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3190 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3192 // Get the returned methodOop.
3193 __ get_vm_result_2(R19_method);
3195 __ bctr();
3198 // Pending exception after the safepoint.
3199 __ BIND(pending);
3201 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3203 // exception pending => remove activation and forward to exception handler
3205 __ li(R11_scratch1, 0);
3206 __ ld(R3_ARG1, thread_(pending_exception));
3207 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread);
3208 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3210 // -------------
3211 // Make sure all code is generated.
3212 masm->flush();
3214 // return the blob
3215 // frame_size_words or bytes??
3216 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3217 oop_maps, true);
3218 }