Thu, 03 Apr 2014 12:37:53 +0200
8039146: Fix 64-bit store to int JNIHandleBlock::_top
Reviewed-by: coleenp, kvn
1 /*
2 * Copyright (c) 2003, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "code/debugInfoRec.hpp"
28 #include "code/icBuffer.hpp"
29 #include "code/vtableStubs.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "oops/compiledICHolder.hpp"
32 #include "prims/jvmtiRedefineClassesTrace.hpp"
33 #include "runtime/sharedRuntime.hpp"
34 #include "runtime/vframeArray.hpp"
35 #include "vmreg_sparc.inline.hpp"
36 #ifdef COMPILER1
37 #include "c1/c1_Runtime1.hpp"
38 #endif
39 #ifdef COMPILER2
40 #include "opto/runtime.hpp"
41 #endif
42 #ifdef SHARK
43 #include "compiler/compileBroker.hpp"
44 #include "shark/sharkCompiler.hpp"
45 #endif
47 #define __ masm->
50 class RegisterSaver {
52 // Used for saving volatile registers. This is Gregs, Fregs, I/L/O.
53 // The Oregs are problematic. In the 32bit build the compiler can
54 // have O registers live with 64 bit quantities. A window save will
55 // cut the heads off of the registers. We have to do a very extensive
56 // stack dance to save and restore these properly.
58 // Note that the Oregs problem only exists if we block at either a polling
59 // page exception a compiled code safepoint that was not originally a call
60 // or deoptimize following one of these kinds of safepoints.
62 // Lots of registers to save. For all builds, a window save will preserve
63 // the %i and %l registers. For the 32-bit longs-in-two entries and 64-bit
64 // builds a window-save will preserve the %o registers. In the LION build
65 // we need to save the 64-bit %o registers which requires we save them
66 // before the window-save (as then they become %i registers and get their
67 // heads chopped off on interrupt). We have to save some %g registers here
68 // as well.
69 enum {
70 // This frame's save area. Includes extra space for the native call:
71 // vararg's layout space and the like. Briefly holds the caller's
72 // register save area.
73 call_args_area = frame::register_save_words_sp_offset +
74 frame::memory_parameter_word_sp_offset*wordSize,
75 // Make sure save locations are always 8 byte aligned.
76 // can't use round_to because it doesn't produce compile time constant
77 start_of_extra_save_area = ((call_args_area + 7) & ~7),
78 g1_offset = start_of_extra_save_area, // g-regs needing saving
79 g3_offset = g1_offset+8,
80 g4_offset = g3_offset+8,
81 g5_offset = g4_offset+8,
82 o0_offset = g5_offset+8,
83 o1_offset = o0_offset+8,
84 o2_offset = o1_offset+8,
85 o3_offset = o2_offset+8,
86 o4_offset = o3_offset+8,
87 o5_offset = o4_offset+8,
88 start_of_flags_save_area = o5_offset+8,
89 ccr_offset = start_of_flags_save_area,
90 fsr_offset = ccr_offset + 8,
91 d00_offset = fsr_offset+8, // Start of float save area
92 register_save_size = d00_offset+8*32
93 };
96 public:
98 static int Oexception_offset() { return o0_offset; };
99 static int G3_offset() { return g3_offset; };
100 static int G5_offset() { return g5_offset; };
101 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
102 static void restore_live_registers(MacroAssembler* masm);
104 // During deoptimization only the result register need to be restored
105 // all the other values have already been extracted.
107 static void restore_result_registers(MacroAssembler* masm);
108 };
110 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
111 // Record volatile registers as callee-save values in an OopMap so their save locations will be
112 // propagated to the caller frame's RegisterMap during StackFrameStream construction (needed for
113 // deoptimization; see compiledVFrame::create_stack_value). The caller's I, L and O registers
114 // are saved in register windows - I's and L's in the caller's frame and O's in the stub frame
115 // (as the stub's I's) when the runtime routine called by the stub creates its frame.
116 int i;
117 // Always make the frame size 16 byte aligned.
118 int frame_size = round_to(additional_frame_words + register_save_size, 16);
119 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words
120 int frame_size_in_slots = frame_size / sizeof(jint);
121 // CodeBlob frame size is in words.
122 *total_frame_words = frame_size / wordSize;
123 // OopMap* map = new OopMap(*total_frame_words, 0);
124 OopMap* map = new OopMap(frame_size_in_slots, 0);
126 #if !defined(_LP64)
128 // Save 64-bit O registers; they will get their heads chopped off on a 'save'.
129 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
130 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
131 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
132 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
133 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
134 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
135 #endif /* _LP64 */
137 __ save(SP, -frame_size, SP);
139 #ifndef _LP64
140 // Reload the 64 bit Oregs. Although they are now Iregs we load them
141 // to Oregs here to avoid interrupts cutting off their heads
143 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
144 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
145 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
146 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
147 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
148 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
150 __ stx(O0, SP, o0_offset+STACK_BIAS);
151 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset + 4)>>2), O0->as_VMReg());
153 __ stx(O1, SP, o1_offset+STACK_BIAS);
155 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset + 4)>>2), O1->as_VMReg());
157 __ stx(O2, SP, o2_offset+STACK_BIAS);
158 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset + 4)>>2), O2->as_VMReg());
160 __ stx(O3, SP, o3_offset+STACK_BIAS);
161 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset + 4)>>2), O3->as_VMReg());
163 __ stx(O4, SP, o4_offset+STACK_BIAS);
164 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset + 4)>>2), O4->as_VMReg());
166 __ stx(O5, SP, o5_offset+STACK_BIAS);
167 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset + 4)>>2), O5->as_VMReg());
168 #endif /* _LP64 */
171 #ifdef _LP64
172 int debug_offset = 0;
173 #else
174 int debug_offset = 4;
175 #endif
176 // Save the G's
177 __ stx(G1, SP, g1_offset+STACK_BIAS);
178 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset + debug_offset)>>2), G1->as_VMReg());
180 __ stx(G3, SP, g3_offset+STACK_BIAS);
181 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset + debug_offset)>>2), G3->as_VMReg());
183 __ stx(G4, SP, g4_offset+STACK_BIAS);
184 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset + debug_offset)>>2), G4->as_VMReg());
186 __ stx(G5, SP, g5_offset+STACK_BIAS);
187 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset + debug_offset)>>2), G5->as_VMReg());
189 // This is really a waste but we'll keep things as they were for now
190 if (true) {
191 #ifndef _LP64
192 map->set_callee_saved(VMRegImpl::stack2reg((o0_offset)>>2), O0->as_VMReg()->next());
193 map->set_callee_saved(VMRegImpl::stack2reg((o1_offset)>>2), O1->as_VMReg()->next());
194 map->set_callee_saved(VMRegImpl::stack2reg((o2_offset)>>2), O2->as_VMReg()->next());
195 map->set_callee_saved(VMRegImpl::stack2reg((o3_offset)>>2), O3->as_VMReg()->next());
196 map->set_callee_saved(VMRegImpl::stack2reg((o4_offset)>>2), O4->as_VMReg()->next());
197 map->set_callee_saved(VMRegImpl::stack2reg((o5_offset)>>2), O5->as_VMReg()->next());
198 map->set_callee_saved(VMRegImpl::stack2reg((g1_offset)>>2), G1->as_VMReg()->next());
199 map->set_callee_saved(VMRegImpl::stack2reg((g3_offset)>>2), G3->as_VMReg()->next());
200 map->set_callee_saved(VMRegImpl::stack2reg((g4_offset)>>2), G4->as_VMReg()->next());
201 map->set_callee_saved(VMRegImpl::stack2reg((g5_offset)>>2), G5->as_VMReg()->next());
202 #endif /* _LP64 */
203 }
206 // Save the flags
207 __ rdccr( G5 );
208 __ stx(G5, SP, ccr_offset+STACK_BIAS);
209 __ stxfsr(SP, fsr_offset+STACK_BIAS);
211 // Save all the FP registers: 32 doubles (32 floats correspond to the 2 halves of the first 16 doubles)
212 int offset = d00_offset;
213 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
214 FloatRegister f = as_FloatRegister(i);
215 __ stf(FloatRegisterImpl::D, f, SP, offset+STACK_BIAS);
216 // Record as callee saved both halves of double registers (2 float registers).
217 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2), f->as_VMReg());
218 map->set_callee_saved(VMRegImpl::stack2reg((offset + sizeof(float))>>2), f->as_VMReg()->next());
219 offset += sizeof(double);
220 }
222 // And we're done.
224 return map;
225 }
228 // Pop the current frame and restore all the registers that we
229 // saved.
230 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
232 // Restore all the FP registers
233 for( int i=0; i<FloatRegisterImpl::number_of_registers; i+=2 ) {
234 __ ldf(FloatRegisterImpl::D, SP, d00_offset+i*sizeof(float)+STACK_BIAS, as_FloatRegister(i));
235 }
237 __ ldx(SP, ccr_offset+STACK_BIAS, G1);
238 __ wrccr (G1) ;
240 // Restore the G's
241 // Note that G2 (AKA GThread) must be saved and restored separately.
242 // TODO-FIXME: save and restore some of the other ASRs, viz., %asi and %gsr.
244 __ ldx(SP, g1_offset+STACK_BIAS, G1);
245 __ ldx(SP, g3_offset+STACK_BIAS, G3);
246 __ ldx(SP, g4_offset+STACK_BIAS, G4);
247 __ ldx(SP, g5_offset+STACK_BIAS, G5);
250 #if !defined(_LP64)
251 // Restore the 64-bit O's.
252 __ ldx(SP, o0_offset+STACK_BIAS, O0);
253 __ ldx(SP, o1_offset+STACK_BIAS, O1);
254 __ ldx(SP, o2_offset+STACK_BIAS, O2);
255 __ ldx(SP, o3_offset+STACK_BIAS, O3);
256 __ ldx(SP, o4_offset+STACK_BIAS, O4);
257 __ ldx(SP, o5_offset+STACK_BIAS, O5);
259 // And temporarily place them in TLS
261 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
262 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
263 __ stx(O2, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8);
264 __ stx(O3, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8);
265 __ stx(O4, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8);
266 __ stx(O5, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8);
267 #endif /* _LP64 */
269 // Restore flags
271 __ ldxfsr(SP, fsr_offset+STACK_BIAS);
273 __ restore();
275 #if !defined(_LP64)
276 // Now reload the 64bit Oregs after we've restore the window.
277 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
278 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
279 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+2*8, O2);
280 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+3*8, O3);
281 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+4*8, O4);
282 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+5*8, O5);
283 #endif /* _LP64 */
285 }
287 // Pop the current frame and restore the registers that might be holding
288 // a result.
289 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
291 #if !defined(_LP64)
292 // 32bit build returns longs in G1
293 __ ldx(SP, g1_offset+STACK_BIAS, G1);
295 // Retrieve the 64-bit O's.
296 __ ldx(SP, o0_offset+STACK_BIAS, O0);
297 __ ldx(SP, o1_offset+STACK_BIAS, O1);
298 // and save to TLS
299 __ stx(O0, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8);
300 __ stx(O1, G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8);
301 #endif /* _LP64 */
303 __ ldf(FloatRegisterImpl::D, SP, d00_offset+STACK_BIAS, as_FloatRegister(0));
305 __ restore();
307 #if !defined(_LP64)
308 // Now reload the 64bit Oregs after we've restore the window.
309 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+0*8, O0);
310 __ ldx(G2_thread, JavaThread::o_reg_temps_offset_in_bytes()+1*8, O1);
311 #endif /* _LP64 */
313 }
315 // Is vector's size (in bytes) bigger than a size saved by default?
316 // 8 bytes FP registers are saved by default on SPARC.
317 bool SharedRuntime::is_wide_vector(int size) {
318 // Note, MaxVectorSize == 8 on SPARC.
319 assert(size <= 8, err_msg_res("%d bytes vectors are not supported", size));
320 return size > 8;
321 }
323 // The java_calling_convention describes stack locations as ideal slots on
324 // a frame with no abi restrictions. Since we must observe abi restrictions
325 // (like the placement of the register window) the slots must be biased by
326 // the following value.
327 static int reg2offset(VMReg r) {
328 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
329 }
331 static VMRegPair reg64_to_VMRegPair(Register r) {
332 VMRegPair ret;
333 if (wordSize == 8) {
334 ret.set2(r->as_VMReg());
335 } else {
336 ret.set_pair(r->successor()->as_VMReg(), r->as_VMReg());
337 }
338 return ret;
339 }
341 // ---------------------------------------------------------------------------
342 // Read the array of BasicTypes from a signature, and compute where the
343 // arguments should go. Values in the VMRegPair regs array refer to 4-byte (VMRegImpl::stack_slot_size)
344 // quantities. Values less than VMRegImpl::stack0 are registers, those above
345 // refer to 4-byte stack slots. All stack slots are based off of the window
346 // top. VMRegImpl::stack0 refers to the first slot past the 16-word window,
347 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
348 // values 0-63 (up to RegisterImpl::number_of_registers) are the 64-bit
349 // integer registers. Values 64-95 are the (32-bit only) float registers.
350 // Each 32-bit quantity is given its own number, so the integer registers
351 // (in either 32- or 64-bit builds) use 2 numbers. For example, there is
352 // an O0-low and an O0-high. Essentially, all int register numbers are doubled.
354 // Register results are passed in O0-O5, for outgoing call arguments. To
355 // convert to incoming arguments, convert all O's to I's. The regs array
356 // refer to the low and hi 32-bit words of 64-bit registers or stack slots.
357 // If the regs[].second() field is set to VMRegImpl::Bad(), it means it's unused (a
358 // 32-bit value was passed). If both are VMRegImpl::Bad(), it means no value was
359 // passed (used as a placeholder for the other half of longs and doubles in
360 // the 64-bit build). regs[].second() is either VMRegImpl::Bad() or regs[].second() is
361 // regs[].first()+1 (regs[].first() may be misaligned in the C calling convention).
362 // Sparc never passes a value in regs[].second() but not regs[].first() (regs[].first()
363 // == VMRegImpl::Bad() && regs[].second() != VMRegImpl::Bad()) nor unrelated values in the
364 // same VMRegPair.
366 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
367 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
368 // units regardless of build.
371 // ---------------------------------------------------------------------------
372 // The compiled Java calling convention. The Java convention always passes
373 // 64-bit values in adjacent aligned locations (either registers or stack),
374 // floats in float registers and doubles in aligned float pairs. There is
375 // no backing varargs store for values in registers.
376 // In the 32-bit build, longs are passed on the stack (cannot be
377 // passed in I's, because longs in I's get their heads chopped off at
378 // interrupt).
379 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
380 VMRegPair *regs,
381 int total_args_passed,
382 int is_outgoing) {
383 assert(F31->as_VMReg()->is_reg(), "overlapping stack/register numbers");
385 const int int_reg_max = SPARC_ARGS_IN_REGS_NUM;
386 const int flt_reg_max = 8;
388 int int_reg = 0;
389 int flt_reg = 0;
390 int slot = 0;
392 for (int i = 0; i < total_args_passed; i++) {
393 switch (sig_bt[i]) {
394 case T_INT:
395 case T_SHORT:
396 case T_CHAR:
397 case T_BYTE:
398 case T_BOOLEAN:
399 #ifndef _LP64
400 case T_OBJECT:
401 case T_ARRAY:
402 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
403 #endif // _LP64
404 if (int_reg < int_reg_max) {
405 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
406 regs[i].set1(r->as_VMReg());
407 } else {
408 regs[i].set1(VMRegImpl::stack2reg(slot++));
409 }
410 break;
412 #ifdef _LP64
413 case T_LONG:
414 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
415 // fall-through
416 case T_OBJECT:
417 case T_ARRAY:
418 case T_ADDRESS: // Used, e.g., in slow-path locking for the lock's stack address
419 if (int_reg < int_reg_max) {
420 Register r = is_outgoing ? as_oRegister(int_reg++) : as_iRegister(int_reg++);
421 regs[i].set2(r->as_VMReg());
422 } else {
423 slot = round_to(slot, 2); // align
424 regs[i].set2(VMRegImpl::stack2reg(slot));
425 slot += 2;
426 }
427 break;
428 #else
429 case T_LONG:
430 assert(sig_bt[i+1] == T_VOID, "expecting VOID in other half");
431 // On 32-bit SPARC put longs always on the stack to keep the pressure off
432 // integer argument registers. They should be used for oops.
433 slot = round_to(slot, 2); // align
434 regs[i].set2(VMRegImpl::stack2reg(slot));
435 slot += 2;
436 #endif
437 break;
439 case T_FLOAT:
440 if (flt_reg < flt_reg_max) {
441 FloatRegister r = as_FloatRegister(flt_reg++);
442 regs[i].set1(r->as_VMReg());
443 } else {
444 regs[i].set1(VMRegImpl::stack2reg(slot++));
445 }
446 break;
448 case T_DOUBLE:
449 assert(sig_bt[i+1] == T_VOID, "expecting half");
450 if (round_to(flt_reg, 2) + 1 < flt_reg_max) {
451 flt_reg = round_to(flt_reg, 2); // align
452 FloatRegister r = as_FloatRegister(flt_reg);
453 regs[i].set2(r->as_VMReg());
454 flt_reg += 2;
455 } else {
456 slot = round_to(slot, 2); // align
457 regs[i].set2(VMRegImpl::stack2reg(slot));
458 slot += 2;
459 }
460 break;
462 case T_VOID:
463 regs[i].set_bad(); // Halves of longs & doubles
464 break;
466 default:
467 fatal(err_msg_res("unknown basic type %d", sig_bt[i]));
468 break;
469 }
470 }
472 // retun the amount of stack space these arguments will need.
473 return slot;
474 }
476 // Helper class mostly to avoid passing masm everywhere, and handle
477 // store displacement overflow logic.
478 class AdapterGenerator {
479 MacroAssembler *masm;
480 Register Rdisp;
481 void set_Rdisp(Register r) { Rdisp = r; }
483 void patch_callers_callsite();
485 // base+st_off points to top of argument
486 int arg_offset(const int st_off) { return st_off; }
487 int next_arg_offset(const int st_off) {
488 return st_off - Interpreter::stackElementSize;
489 }
491 // Argument slot values may be loaded first into a register because
492 // they might not fit into displacement.
493 RegisterOrConstant arg_slot(const int st_off);
494 RegisterOrConstant next_arg_slot(const int st_off);
496 // Stores long into offset pointed to by base
497 void store_c2i_long(Register r, Register base,
498 const int st_off, bool is_stack);
499 void store_c2i_object(Register r, Register base,
500 const int st_off);
501 void store_c2i_int(Register r, Register base,
502 const int st_off);
503 void store_c2i_double(VMReg r_2,
504 VMReg r_1, Register base, const int st_off);
505 void store_c2i_float(FloatRegister f, Register base,
506 const int st_off);
508 public:
509 void gen_c2i_adapter(int total_args_passed,
510 // VMReg max_arg,
511 int comp_args_on_stack, // VMRegStackSlots
512 const BasicType *sig_bt,
513 const VMRegPair *regs,
514 Label& skip_fixup);
515 void gen_i2c_adapter(int total_args_passed,
516 // VMReg max_arg,
517 int comp_args_on_stack, // VMRegStackSlots
518 const BasicType *sig_bt,
519 const VMRegPair *regs);
521 AdapterGenerator(MacroAssembler *_masm) : masm(_masm) {}
522 };
525 // Patch the callers callsite with entry to compiled code if it exists.
526 void AdapterGenerator::patch_callers_callsite() {
527 Label L;
528 __ ld_ptr(G5_method, in_bytes(Method::code_offset()), G3_scratch);
529 __ br_null(G3_scratch, false, Assembler::pt, L);
530 __ delayed()->nop();
531 // Call into the VM to patch the caller, then jump to compiled callee
532 __ save_frame(4); // Args in compiled layout; do not blow them
534 // Must save all the live Gregs the list is:
535 // G1: 1st Long arg (32bit build)
536 // G2: global allocated to TLS
537 // G3: used in inline cache check (scratch)
538 // G4: 2nd Long arg (32bit build);
539 // G5: used in inline cache check (Method*)
541 // The longs must go to the stack by hand since in the 32 bit build they can be trashed by window ops.
543 #ifdef _LP64
544 // mov(s,d)
545 __ mov(G1, L1);
546 __ mov(G4, L4);
547 __ mov(G5_method, L5);
548 __ mov(G5_method, O0); // VM needs target method
549 __ mov(I7, O1); // VM needs caller's callsite
550 // Must be a leaf call...
551 // can be very far once the blob has been relocated
552 AddressLiteral dest(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
553 __ relocate(relocInfo::runtime_call_type);
554 __ jumpl_to(dest, O7, O7);
555 __ delayed()->mov(G2_thread, L7_thread_cache);
556 __ mov(L7_thread_cache, G2_thread);
557 __ mov(L1, G1);
558 __ mov(L4, G4);
559 __ mov(L5, G5_method);
560 #else
561 __ stx(G1, FP, -8 + STACK_BIAS);
562 __ stx(G4, FP, -16 + STACK_BIAS);
563 __ mov(G5_method, L5);
564 __ mov(G5_method, O0); // VM needs target method
565 __ mov(I7, O1); // VM needs caller's callsite
566 // Must be a leaf call...
567 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), relocInfo::runtime_call_type);
568 __ delayed()->mov(G2_thread, L7_thread_cache);
569 __ mov(L7_thread_cache, G2_thread);
570 __ ldx(FP, -8 + STACK_BIAS, G1);
571 __ ldx(FP, -16 + STACK_BIAS, G4);
572 __ mov(L5, G5_method);
573 #endif /* _LP64 */
575 __ restore(); // Restore args
576 __ bind(L);
577 }
580 RegisterOrConstant AdapterGenerator::arg_slot(const int st_off) {
581 RegisterOrConstant roc(arg_offset(st_off));
582 return __ ensure_simm13_or_reg(roc, Rdisp);
583 }
585 RegisterOrConstant AdapterGenerator::next_arg_slot(const int st_off) {
586 RegisterOrConstant roc(next_arg_offset(st_off));
587 return __ ensure_simm13_or_reg(roc, Rdisp);
588 }
591 // Stores long into offset pointed to by base
592 void AdapterGenerator::store_c2i_long(Register r, Register base,
593 const int st_off, bool is_stack) {
594 #ifdef _LP64
595 // In V9, longs are given 2 64-bit slots in the interpreter, but the
596 // data is passed in only 1 slot.
597 __ stx(r, base, next_arg_slot(st_off));
598 #else
599 #ifdef COMPILER2
600 // Misaligned store of 64-bit data
601 __ stw(r, base, arg_slot(st_off)); // lo bits
602 __ srlx(r, 32, r);
603 __ stw(r, base, next_arg_slot(st_off)); // hi bits
604 #else
605 if (is_stack) {
606 // Misaligned store of 64-bit data
607 __ stw(r, base, arg_slot(st_off)); // lo bits
608 __ srlx(r, 32, r);
609 __ stw(r, base, next_arg_slot(st_off)); // hi bits
610 } else {
611 __ stw(r->successor(), base, arg_slot(st_off) ); // lo bits
612 __ stw(r , base, next_arg_slot(st_off)); // hi bits
613 }
614 #endif // COMPILER2
615 #endif // _LP64
616 }
618 void AdapterGenerator::store_c2i_object(Register r, Register base,
619 const int st_off) {
620 __ st_ptr (r, base, arg_slot(st_off));
621 }
623 void AdapterGenerator::store_c2i_int(Register r, Register base,
624 const int st_off) {
625 __ st (r, base, arg_slot(st_off));
626 }
628 // Stores into offset pointed to by base
629 void AdapterGenerator::store_c2i_double(VMReg r_2,
630 VMReg r_1, Register base, const int st_off) {
631 #ifdef _LP64
632 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
633 // data is passed in only 1 slot.
634 __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
635 #else
636 // Need to marshal 64-bit value from misaligned Lesp loads
637 __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), base, next_arg_slot(st_off));
638 __ stf(FloatRegisterImpl::S, r_2->as_FloatRegister(), base, arg_slot(st_off) );
639 #endif
640 }
642 void AdapterGenerator::store_c2i_float(FloatRegister f, Register base,
643 const int st_off) {
644 __ stf(FloatRegisterImpl::S, f, base, arg_slot(st_off));
645 }
647 void AdapterGenerator::gen_c2i_adapter(
648 int total_args_passed,
649 // VMReg max_arg,
650 int comp_args_on_stack, // VMRegStackSlots
651 const BasicType *sig_bt,
652 const VMRegPair *regs,
653 Label& L_skip_fixup) {
655 // Before we get into the guts of the C2I adapter, see if we should be here
656 // at all. We've come from compiled code and are attempting to jump to the
657 // interpreter, which means the caller made a static call to get here
658 // (vcalls always get a compiled target if there is one). Check for a
659 // compiled target. If there is one, we need to patch the caller's call.
660 // However we will run interpreted if we come thru here. The next pass
661 // thru the call site will run compiled. If we ran compiled here then
662 // we can (theorectically) do endless i2c->c2i->i2c transitions during
663 // deopt/uncommon trap cycles. If we always go interpreted here then
664 // we can have at most one and don't need to play any tricks to keep
665 // from endlessly growing the stack.
666 //
667 // Actually if we detected that we had an i2c->c2i transition here we
668 // ought to be able to reset the world back to the state of the interpreted
669 // call and not bother building another interpreter arg area. We don't
670 // do that at this point.
672 patch_callers_callsite();
674 __ bind(L_skip_fixup);
676 // Since all args are passed on the stack, total_args_passed*wordSize is the
677 // space we need. Add in varargs area needed by the interpreter. Round up
678 // to stack alignment.
679 const int arg_size = total_args_passed * Interpreter::stackElementSize;
680 const int varargs_area =
681 (frame::varargs_offset - frame::register_save_words)*wordSize;
682 const int extraspace = round_to(arg_size + varargs_area, 2*wordSize);
684 const int bias = STACK_BIAS;
685 const int interp_arg_offset = frame::varargs_offset*wordSize +
686 (total_args_passed-1)*Interpreter::stackElementSize;
688 const Register base = SP;
690 // Make some extra space on the stack.
691 __ sub(SP, __ ensure_simm13_or_reg(extraspace, G3_scratch), SP);
692 set_Rdisp(G3_scratch);
694 // Write the args into the outgoing interpreter space.
695 for (int i = 0; i < total_args_passed; i++) {
696 const int st_off = interp_arg_offset - (i*Interpreter::stackElementSize) + bias;
697 VMReg r_1 = regs[i].first();
698 VMReg r_2 = regs[i].second();
699 if (!r_1->is_valid()) {
700 assert(!r_2->is_valid(), "");
701 continue;
702 }
703 if (r_1->is_stack()) { // Pretend stack targets are loaded into G1
704 RegisterOrConstant ld_off = reg2offset(r_1) + extraspace + bias;
705 ld_off = __ ensure_simm13_or_reg(ld_off, Rdisp);
706 r_1 = G1_scratch->as_VMReg();// as part of the load/store shuffle
707 if (!r_2->is_valid()) __ ld (base, ld_off, G1_scratch);
708 else __ ldx(base, ld_off, G1_scratch);
709 }
711 if (r_1->is_Register()) {
712 Register r = r_1->as_Register()->after_restore();
713 if (sig_bt[i] == T_OBJECT || sig_bt[i] == T_ARRAY) {
714 store_c2i_object(r, base, st_off);
715 } else if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
716 store_c2i_long(r, base, st_off, r_2->is_stack());
717 } else {
718 store_c2i_int(r, base, st_off);
719 }
720 } else {
721 assert(r_1->is_FloatRegister(), "");
722 if (sig_bt[i] == T_FLOAT) {
723 store_c2i_float(r_1->as_FloatRegister(), base, st_off);
724 } else {
725 assert(sig_bt[i] == T_DOUBLE, "wrong type");
726 store_c2i_double(r_2, r_1, base, st_off);
727 }
728 }
729 }
731 // Load the interpreter entry point.
732 __ ld_ptr(G5_method, in_bytes(Method::interpreter_entry_offset()), G3_scratch);
734 // Pass O5_savedSP as an argument to the interpreter.
735 // The interpreter will restore SP to this value before returning.
736 __ add(SP, __ ensure_simm13_or_reg(extraspace, G1), O5_savedSP);
738 __ mov((frame::varargs_offset)*wordSize -
739 1*Interpreter::stackElementSize+bias+BytesPerWord, G1);
740 // Jump to the interpreter just as if interpreter was doing it.
741 __ jmpl(G3_scratch, 0, G0);
742 // Setup Lesp for the call. Cannot actually set Lesp as the current Lesp
743 // (really L0) is in use by the compiled frame as a generic temp. However,
744 // the interpreter does not know where its args are without some kind of
745 // arg pointer being passed in. Pass it in Gargs.
746 __ delayed()->add(SP, G1, Gargs);
747 }
749 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg, Register temp2_reg,
750 address code_start, address code_end,
751 Label& L_ok) {
752 Label L_fail;
753 __ set(ExternalAddress(code_start), temp_reg);
754 __ set(pointer_delta(code_end, code_start, 1), temp2_reg);
755 __ cmp(pc_reg, temp_reg);
756 __ brx(Assembler::lessEqualUnsigned, false, Assembler::pn, L_fail);
757 __ delayed()->add(temp_reg, temp2_reg, temp_reg);
758 __ cmp(pc_reg, temp_reg);
759 __ cmp_and_brx_short(pc_reg, temp_reg, Assembler::lessUnsigned, Assembler::pt, L_ok);
760 __ bind(L_fail);
761 }
763 void AdapterGenerator::gen_i2c_adapter(
764 int total_args_passed,
765 // VMReg max_arg,
766 int comp_args_on_stack, // VMRegStackSlots
767 const BasicType *sig_bt,
768 const VMRegPair *regs) {
770 // Generate an I2C adapter: adjust the I-frame to make space for the C-frame
771 // layout. Lesp was saved by the calling I-frame and will be restored on
772 // return. Meanwhile, outgoing arg space is all owned by the callee
773 // C-frame, so we can mangle it at will. After adjusting the frame size,
774 // hoist register arguments and repack other args according to the compiled
775 // code convention. Finally, end in a jump to the compiled code. The entry
776 // point address is the start of the buffer.
778 // We will only enter here from an interpreted frame and never from after
779 // passing thru a c2i. Azul allowed this but we do not. If we lose the
780 // race and use a c2i we will remain interpreted for the race loser(s).
781 // This removes all sorts of headaches on the x86 side and also eliminates
782 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
784 // More detail:
785 // Adapters can be frameless because they do not require the caller
786 // to perform additional cleanup work, such as correcting the stack pointer.
787 // An i2c adapter is frameless because the *caller* frame, which is interpreted,
788 // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
789 // even if a callee has modified the stack pointer.
790 // A c2i adapter is frameless because the *callee* frame, which is interpreted,
791 // routinely repairs its caller's stack pointer (from sender_sp, which is set
792 // up via the senderSP register).
793 // In other words, if *either* the caller or callee is interpreted, we can
794 // get the stack pointer repaired after a call.
795 // This is why c2i and i2c adapters cannot be indefinitely composed.
796 // In particular, if a c2i adapter were to somehow call an i2c adapter,
797 // both caller and callee would be compiled methods, and neither would
798 // clean up the stack pointer changes performed by the two adapters.
799 // If this happens, control eventually transfers back to the compiled
800 // caller, but with an uncorrected stack, causing delayed havoc.
802 if (VerifyAdapterCalls &&
803 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) {
804 // So, let's test for cascading c2i/i2c adapters right now.
805 // assert(Interpreter::contains($return_addr) ||
806 // StubRoutines::contains($return_addr),
807 // "i2c adapter must return to an interpreter frame");
808 __ block_comment("verify_i2c { ");
809 Label L_ok;
810 if (Interpreter::code() != NULL)
811 range_check(masm, O7, O0, O1,
812 Interpreter::code()->code_start(), Interpreter::code()->code_end(),
813 L_ok);
814 if (StubRoutines::code1() != NULL)
815 range_check(masm, O7, O0, O1,
816 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(),
817 L_ok);
818 if (StubRoutines::code2() != NULL)
819 range_check(masm, O7, O0, O1,
820 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(),
821 L_ok);
822 const char* msg = "i2c adapter must return to an interpreter frame";
823 __ block_comment(msg);
824 __ stop(msg);
825 __ bind(L_ok);
826 __ block_comment("} verify_i2ce ");
827 }
829 // As you can see from the list of inputs & outputs there are not a lot
830 // of temp registers to work with: mostly G1, G3 & G4.
832 // Inputs:
833 // G2_thread - TLS
834 // G5_method - Method oop
835 // G4 (Gargs) - Pointer to interpreter's args
836 // O0..O4 - free for scratch
837 // O5_savedSP - Caller's saved SP, to be restored if needed
838 // O6 - Current SP!
839 // O7 - Valid return address
840 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
842 // Outputs:
843 // G2_thread - TLS
844 // O0-O5 - Outgoing args in compiled layout
845 // O6 - Adjusted or restored SP
846 // O7 - Valid return address
847 // L0-L7, I0-I7 - Caller's temps (no frame pushed yet)
848 // F0-F7 - more outgoing args
851 // Gargs is the incoming argument base, and also an outgoing argument.
852 __ sub(Gargs, BytesPerWord, Gargs);
854 // ON ENTRY TO THE CODE WE ARE MAKING, WE HAVE AN INTERPRETED FRAME
855 // WITH O7 HOLDING A VALID RETURN PC
856 //
857 // | |
858 // : java stack :
859 // | |
860 // +--------------+ <--- start of outgoing args
861 // | receiver | |
862 // : rest of args : |---size is java-arg-words
863 // | | |
864 // +--------------+ <--- O4_args (misaligned) and Lesp if prior is not C2I
865 // | | |
866 // : unused : |---Space for max Java stack, plus stack alignment
867 // | | |
868 // +--------------+ <--- SP + 16*wordsize
869 // | |
870 // : window :
871 // | |
872 // +--------------+ <--- SP
874 // WE REPACK THE STACK. We use the common calling convention layout as
875 // discovered by calling SharedRuntime::calling_convention. We assume it
876 // causes an arbitrary shuffle of memory, which may require some register
877 // temps to do the shuffle. We hope for (and optimize for) the case where
878 // temps are not needed. We may have to resize the stack slightly, in case
879 // we need alignment padding (32-bit interpreter can pass longs & doubles
880 // misaligned, but the compilers expect them aligned).
881 //
882 // | |
883 // : java stack :
884 // | |
885 // +--------------+ <--- start of outgoing args
886 // | pad, align | |
887 // +--------------+ |
888 // | ints, longs, | |
889 // | floats, | |---Outgoing stack args.
890 // : doubles : | First few args in registers.
891 // | | |
892 // +--------------+ <--- SP' + 16*wordsize
893 // | |
894 // : window :
895 // | |
896 // +--------------+ <--- SP'
898 // ON EXIT FROM THE CODE WE ARE MAKING, WE STILL HAVE AN INTERPRETED FRAME
899 // WITH O7 HOLDING A VALID RETURN PC - ITS JUST THAT THE ARGS ARE NOW SETUP
900 // FOR COMPILED CODE AND THE FRAME SLIGHTLY GROWN.
902 // Cut-out for having no stack args. Since up to 6 args are passed
903 // in registers, we will commonly have no stack args.
904 if (comp_args_on_stack > 0) {
905 // Convert VMReg stack slots to words.
906 int comp_words_on_stack = round_to(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
907 // Round up to miminum stack alignment, in wordSize
908 comp_words_on_stack = round_to(comp_words_on_stack, 2);
909 // Now compute the distance from Lesp to SP. This calculation does not
910 // include the space for total_args_passed because Lesp has not yet popped
911 // the arguments.
912 __ sub(SP, (comp_words_on_stack)*wordSize, SP);
913 }
915 // Now generate the shuffle code. Pick up all register args and move the
916 // rest through G1_scratch.
917 for (int i = 0; i < total_args_passed; i++) {
918 if (sig_bt[i] == T_VOID) {
919 // Longs and doubles are passed in native word order, but misaligned
920 // in the 32-bit build.
921 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
922 continue;
923 }
925 // Pick up 0, 1 or 2 words from Lesp+offset. Assume mis-aligned in the
926 // 32-bit build and aligned in the 64-bit build. Look for the obvious
927 // ldx/lddf optimizations.
929 // Load in argument order going down.
930 const int ld_off = (total_args_passed-i)*Interpreter::stackElementSize;
931 set_Rdisp(G1_scratch);
933 VMReg r_1 = regs[i].first();
934 VMReg r_2 = regs[i].second();
935 if (!r_1->is_valid()) {
936 assert(!r_2->is_valid(), "");
937 continue;
938 }
939 if (r_1->is_stack()) { // Pretend stack targets are loaded into F8/F9
940 r_1 = F8->as_VMReg(); // as part of the load/store shuffle
941 if (r_2->is_valid()) r_2 = r_1->next();
942 }
943 if (r_1->is_Register()) { // Register argument
944 Register r = r_1->as_Register()->after_restore();
945 if (!r_2->is_valid()) {
946 __ ld(Gargs, arg_slot(ld_off), r);
947 } else {
948 #ifdef _LP64
949 // In V9, longs are given 2 64-bit slots in the interpreter, but the
950 // data is passed in only 1 slot.
951 RegisterOrConstant slot = (sig_bt[i] == T_LONG) ?
952 next_arg_slot(ld_off) : arg_slot(ld_off);
953 __ ldx(Gargs, slot, r);
954 #else
955 fatal("longs should be on stack");
956 #endif
957 }
958 } else {
959 assert(r_1->is_FloatRegister(), "");
960 if (!r_2->is_valid()) {
961 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_1->as_FloatRegister());
962 } else {
963 #ifdef _LP64
964 // In V9, doubles are given 2 64-bit slots in the interpreter, but the
965 // data is passed in only 1 slot. This code also handles longs that
966 // are passed on the stack, but need a stack-to-stack move through a
967 // spare float register.
968 RegisterOrConstant slot = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
969 next_arg_slot(ld_off) : arg_slot(ld_off);
970 __ ldf(FloatRegisterImpl::D, Gargs, slot, r_1->as_FloatRegister());
971 #else
972 // Need to marshal 64-bit value from misaligned Lesp loads
973 __ ldf(FloatRegisterImpl::S, Gargs, next_arg_slot(ld_off), r_1->as_FloatRegister());
974 __ ldf(FloatRegisterImpl::S, Gargs, arg_slot(ld_off), r_2->as_FloatRegister());
975 #endif
976 }
977 }
978 // Was the argument really intended to be on the stack, but was loaded
979 // into F8/F9?
980 if (regs[i].first()->is_stack()) {
981 assert(r_1->as_FloatRegister() == F8, "fix this code");
982 // Convert stack slot to an SP offset
983 int st_off = reg2offset(regs[i].first()) + STACK_BIAS;
984 // Store down the shuffled stack word. Target address _is_ aligned.
985 RegisterOrConstant slot = __ ensure_simm13_or_reg(st_off, Rdisp);
986 if (!r_2->is_valid()) __ stf(FloatRegisterImpl::S, r_1->as_FloatRegister(), SP, slot);
987 else __ stf(FloatRegisterImpl::D, r_1->as_FloatRegister(), SP, slot);
988 }
989 }
991 // Jump to the compiled code just as if compiled code was doing it.
992 __ ld_ptr(G5_method, in_bytes(Method::from_compiled_offset()), G3);
994 // 6243940 We might end up in handle_wrong_method if
995 // the callee is deoptimized as we race thru here. If that
996 // happens we don't want to take a safepoint because the
997 // caller frame will look interpreted and arguments are now
998 // "compiled" so it is much better to make this transition
999 // invisible to the stack walking code. Unfortunately if
1000 // we try and find the callee by normal means a safepoint
1001 // is possible. So we stash the desired callee in the thread
1002 // and the vm will find there should this case occur.
1003 Address callee_target_addr(G2_thread, JavaThread::callee_target_offset());
1004 __ st_ptr(G5_method, callee_target_addr);
1005 __ jmpl(G3, 0, G0);
1006 __ delayed()->nop();
1007 }
1009 // ---------------------------------------------------------------
1010 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1011 int total_args_passed,
1012 // VMReg max_arg,
1013 int comp_args_on_stack, // VMRegStackSlots
1014 const BasicType *sig_bt,
1015 const VMRegPair *regs,
1016 AdapterFingerPrint* fingerprint) {
1017 address i2c_entry = __ pc();
1019 AdapterGenerator agen(masm);
1021 agen.gen_i2c_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs);
1024 // -------------------------------------------------------------------------
1025 // Generate a C2I adapter. On entry we know G5 holds the Method*. The
1026 // args start out packed in the compiled layout. They need to be unpacked
1027 // into the interpreter layout. This will almost always require some stack
1028 // space. We grow the current (compiled) stack, then repack the args. We
1029 // finally end in a jump to the generic interpreter entry point. On exit
1030 // from the interpreter, the interpreter will restore our SP (lest the
1031 // compiled code, which relys solely on SP and not FP, get sick).
1033 address c2i_unverified_entry = __ pc();
1034 Label L_skip_fixup;
1035 {
1036 Register R_temp = G1; // another scratch register
1038 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
1040 __ verify_oop(O0);
1041 __ load_klass(O0, G3_scratch);
1043 __ ld_ptr(G5_method, CompiledICHolder::holder_klass_offset(), R_temp);
1044 __ cmp(G3_scratch, R_temp);
1046 Label ok, ok2;
1047 __ brx(Assembler::equal, false, Assembler::pt, ok);
1048 __ delayed()->ld_ptr(G5_method, CompiledICHolder::holder_method_offset(), G5_method);
1049 __ jump_to(ic_miss, G3_scratch);
1050 __ delayed()->nop();
1052 __ bind(ok);
1053 // Method might have been compiled since the call site was patched to
1054 // interpreted if that is the case treat it as a miss so we can get
1055 // the call site corrected.
1056 __ ld_ptr(G5_method, in_bytes(Method::code_offset()), G3_scratch);
1057 __ bind(ok2);
1058 __ br_null(G3_scratch, false, Assembler::pt, L_skip_fixup);
1059 __ delayed()->nop();
1060 __ jump_to(ic_miss, G3_scratch);
1061 __ delayed()->nop();
1063 }
1065 address c2i_entry = __ pc();
1067 agen.gen_c2i_adapter(total_args_passed, comp_args_on_stack, sig_bt, regs, L_skip_fixup);
1069 __ flush();
1070 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
1072 }
1074 // Helper function for native calling conventions
1075 static VMReg int_stk_helper( int i ) {
1076 // Bias any stack based VMReg we get by ignoring the window area
1077 // but not the register parameter save area.
1078 //
1079 // This is strange for the following reasons. We'd normally expect
1080 // the calling convention to return an VMReg for a stack slot
1081 // completely ignoring any abi reserved area. C2 thinks of that
1082 // abi area as only out_preserve_stack_slots. This does not include
1083 // the area allocated by the C abi to store down integer arguments
1084 // because the java calling convention does not use it. So
1085 // since c2 assumes that there are only out_preserve_stack_slots
1086 // to bias the optoregs (which impacts VMRegs) when actually referencing any actual stack
1087 // location the c calling convention must add in this bias amount
1088 // to make up for the fact that the out_preserve_stack_slots is
1089 // insufficient for C calls. What a mess. I sure hope those 6
1090 // stack words were worth it on every java call!
1092 // Another way of cleaning this up would be for out_preserve_stack_slots
1093 // to take a parameter to say whether it was C or java calling conventions.
1094 // Then things might look a little better (but not much).
1096 int mem_parm_offset = i - SPARC_ARGS_IN_REGS_NUM;
1097 if( mem_parm_offset < 0 ) {
1098 return as_oRegister(i)->as_VMReg();
1099 } else {
1100 int actual_offset = (mem_parm_offset + frame::memory_parameter_word_sp_offset) * VMRegImpl::slots_per_word;
1101 // Now return a biased offset that will be correct when out_preserve_slots is added back in
1102 return VMRegImpl::stack2reg(actual_offset - SharedRuntime::out_preserve_stack_slots());
1103 }
1104 }
1107 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
1108 VMRegPair *regs,
1109 VMRegPair *regs2,
1110 int total_args_passed) {
1111 assert(regs2 == NULL, "not needed on sparc");
1113 // Return the number of VMReg stack_slots needed for the args.
1114 // This value does not include an abi space (like register window
1115 // save area).
1117 // The native convention is V8 if !LP64
1118 // The LP64 convention is the V9 convention which is slightly more sane.
1120 // We return the amount of VMReg stack slots we need to reserve for all
1121 // the arguments NOT counting out_preserve_stack_slots. Since we always
1122 // have space for storing at least 6 registers to memory we start with that.
1123 // See int_stk_helper for a further discussion.
1124 int max_stack_slots = (frame::varargs_offset * VMRegImpl::slots_per_word) - SharedRuntime::out_preserve_stack_slots();
1126 #ifdef _LP64
1127 // V9 convention: All things "as-if" on double-wide stack slots.
1128 // Hoist any int/ptr/long's in the first 6 to int regs.
1129 // Hoist any flt/dbl's in the first 16 dbl regs.
1130 int j = 0; // Count of actual args, not HALVES
1131 for( int i=0; i<total_args_passed; i++, j++ ) {
1132 switch( sig_bt[i] ) {
1133 case T_BOOLEAN:
1134 case T_BYTE:
1135 case T_CHAR:
1136 case T_INT:
1137 case T_SHORT:
1138 regs[i].set1( int_stk_helper( j ) ); break;
1139 case T_LONG:
1140 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1141 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1142 case T_ARRAY:
1143 case T_OBJECT:
1144 case T_METADATA:
1145 regs[i].set2( int_stk_helper( j ) );
1146 break;
1147 case T_FLOAT:
1148 if ( j < 16 ) {
1149 // V9ism: floats go in ODD registers
1150 regs[i].set1(as_FloatRegister(1 + (j<<1))->as_VMReg());
1151 } else {
1152 // V9ism: floats go in ODD stack slot
1153 regs[i].set1(VMRegImpl::stack2reg(1 + (j<<1)));
1154 }
1155 break;
1156 case T_DOUBLE:
1157 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1158 if ( j < 16 ) {
1159 // V9ism: doubles go in EVEN/ODD regs
1160 regs[i].set2(as_FloatRegister(j<<1)->as_VMReg());
1161 } else {
1162 // V9ism: doubles go in EVEN/ODD stack slots
1163 regs[i].set2(VMRegImpl::stack2reg(j<<1));
1164 }
1165 break;
1166 case T_VOID: regs[i].set_bad(); j--; break; // Do not count HALVES
1167 default:
1168 ShouldNotReachHere();
1169 }
1170 if (regs[i].first()->is_stack()) {
1171 int off = regs[i].first()->reg2stack();
1172 if (off > max_stack_slots) max_stack_slots = off;
1173 }
1174 if (regs[i].second()->is_stack()) {
1175 int off = regs[i].second()->reg2stack();
1176 if (off > max_stack_slots) max_stack_slots = off;
1177 }
1178 }
1180 #else // _LP64
1181 // V8 convention: first 6 things in O-regs, rest on stack.
1182 // Alignment is willy-nilly.
1183 for( int i=0; i<total_args_passed; i++ ) {
1184 switch( sig_bt[i] ) {
1185 case T_ADDRESS: // raw pointers, like current thread, for VM calls
1186 case T_ARRAY:
1187 case T_BOOLEAN:
1188 case T_BYTE:
1189 case T_CHAR:
1190 case T_FLOAT:
1191 case T_INT:
1192 case T_OBJECT:
1193 case T_METADATA:
1194 case T_SHORT:
1195 regs[i].set1( int_stk_helper( i ) );
1196 break;
1197 case T_DOUBLE:
1198 case T_LONG:
1199 assert( sig_bt[i+1] == T_VOID, "expecting half" );
1200 regs[i].set_pair( int_stk_helper( i+1 ), int_stk_helper( i ) );
1201 break;
1202 case T_VOID: regs[i].set_bad(); break;
1203 default:
1204 ShouldNotReachHere();
1205 }
1206 if (regs[i].first()->is_stack()) {
1207 int off = regs[i].first()->reg2stack();
1208 if (off > max_stack_slots) max_stack_slots = off;
1209 }
1210 if (regs[i].second()->is_stack()) {
1211 int off = regs[i].second()->reg2stack();
1212 if (off > max_stack_slots) max_stack_slots = off;
1213 }
1214 }
1215 #endif // _LP64
1217 return round_to(max_stack_slots + 1, 2);
1219 }
1222 // ---------------------------------------------------------------------------
1223 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1224 switch (ret_type) {
1225 case T_FLOAT:
1226 __ stf(FloatRegisterImpl::S, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS);
1227 break;
1228 case T_DOUBLE:
1229 __ stf(FloatRegisterImpl::D, F0, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS);
1230 break;
1231 }
1232 }
1234 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1235 switch (ret_type) {
1236 case T_FLOAT:
1237 __ ldf(FloatRegisterImpl::S, SP, frame_slots*VMRegImpl::stack_slot_size - 4+STACK_BIAS, F0);
1238 break;
1239 case T_DOUBLE:
1240 __ ldf(FloatRegisterImpl::D, SP, frame_slots*VMRegImpl::stack_slot_size - 8+STACK_BIAS, F0);
1241 break;
1242 }
1243 }
1245 // Check and forward and pending exception. Thread is stored in
1246 // L7_thread_cache and possibly NOT in G2_thread. Since this is a native call, there
1247 // is no exception handler. We merely pop this frame off and throw the
1248 // exception in the caller's frame.
1249 static void check_forward_pending_exception(MacroAssembler *masm, Register Rex_oop) {
1250 Label L;
1251 __ br_null(Rex_oop, false, Assembler::pt, L);
1252 __ delayed()->mov(L7_thread_cache, G2_thread); // restore in case we have exception
1253 // Since this is a native call, we *know* the proper exception handler
1254 // without calling into the VM: it's the empty function. Just pop this
1255 // frame and then jump to forward_exception_entry; O7 will contain the
1256 // native caller's return PC.
1257 AddressLiteral exception_entry(StubRoutines::forward_exception_entry());
1258 __ jump_to(exception_entry, G3_scratch);
1259 __ delayed()->restore(); // Pop this frame off.
1260 __ bind(L);
1261 }
1263 // A simple move of integer like type
1264 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1265 if (src.first()->is_stack()) {
1266 if (dst.first()->is_stack()) {
1267 // stack to stack
1268 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1269 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1270 } else {
1271 // stack to reg
1272 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1273 }
1274 } else if (dst.first()->is_stack()) {
1275 // reg to stack
1276 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1277 } else {
1278 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1279 }
1280 }
1282 // On 64 bit we will store integer like items to the stack as
1283 // 64 bits items (sparc abi) even though java would only store
1284 // 32bits for a parameter. On 32bit it will simply be 32 bits
1285 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
1286 static void move32_64(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1287 if (src.first()->is_stack()) {
1288 if (dst.first()->is_stack()) {
1289 // stack to stack
1290 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1291 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1292 } else {
1293 // stack to reg
1294 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1295 }
1296 } else if (dst.first()->is_stack()) {
1297 // reg to stack
1298 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1299 } else {
1300 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1301 }
1302 }
1305 static void move_ptr(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1306 if (src.first()->is_stack()) {
1307 if (dst.first()->is_stack()) {
1308 // stack to stack
1309 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1310 __ st_ptr(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1311 } else {
1312 // stack to reg
1313 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1314 }
1315 } else if (dst.first()->is_stack()) {
1316 // reg to stack
1317 __ st_ptr(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1318 } else {
1319 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1320 }
1321 }
1324 // An oop arg. Must pass a handle not the oop itself
1325 static void object_move(MacroAssembler* masm,
1326 OopMap* map,
1327 int oop_handle_offset,
1328 int framesize_in_slots,
1329 VMRegPair src,
1330 VMRegPair dst,
1331 bool is_receiver,
1332 int* receiver_offset) {
1334 // must pass a handle. First figure out the location we use as a handle
1336 if (src.first()->is_stack()) {
1337 // Oop is already on the stack
1338 Register rHandle = dst.first()->is_stack() ? L5 : dst.first()->as_Register();
1339 __ add(FP, reg2offset(src.first()) + STACK_BIAS, rHandle);
1340 __ ld_ptr(rHandle, 0, L4);
1341 #ifdef _LP64
1342 __ movr( Assembler::rc_z, L4, G0, rHandle );
1343 #else
1344 __ tst( L4 );
1345 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1346 #endif
1347 if (dst.first()->is_stack()) {
1348 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1349 }
1350 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1351 if (is_receiver) {
1352 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1353 }
1354 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1355 } else {
1356 // Oop is in an input register pass we must flush it to the stack
1357 const Register rOop = src.first()->as_Register();
1358 const Register rHandle = L5;
1359 int oop_slot = rOop->input_number() * VMRegImpl::slots_per_word + oop_handle_offset;
1360 int offset = oop_slot*VMRegImpl::stack_slot_size;
1361 Label skip;
1362 __ st_ptr(rOop, SP, offset + STACK_BIAS);
1363 if (is_receiver) {
1364 *receiver_offset = oop_slot * VMRegImpl::stack_slot_size;
1365 }
1366 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1367 __ add(SP, offset + STACK_BIAS, rHandle);
1368 #ifdef _LP64
1369 __ movr( Assembler::rc_z, rOop, G0, rHandle );
1370 #else
1371 __ tst( rOop );
1372 __ movcc( Assembler::zero, false, Assembler::icc, G0, rHandle );
1373 #endif
1375 if (dst.first()->is_stack()) {
1376 __ st_ptr(rHandle, SP, reg2offset(dst.first()) + STACK_BIAS);
1377 } else {
1378 __ mov(rHandle, dst.first()->as_Register());
1379 }
1380 }
1381 }
1383 // A float arg may have to do float reg int reg conversion
1384 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1385 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1387 if (src.first()->is_stack()) {
1388 if (dst.first()->is_stack()) {
1389 // stack to stack the easiest of the bunch
1390 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1391 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1392 } else {
1393 // stack to reg
1394 if (dst.first()->is_Register()) {
1395 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1396 } else {
1397 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1398 }
1399 }
1400 } else if (dst.first()->is_stack()) {
1401 // reg to stack
1402 if (src.first()->is_Register()) {
1403 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1404 } else {
1405 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1406 }
1407 } else {
1408 // reg to reg
1409 if (src.first()->is_Register()) {
1410 if (dst.first()->is_Register()) {
1411 // gpr -> gpr
1412 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1413 } else {
1414 // gpr -> fpr
1415 __ st(src.first()->as_Register(), FP, -4 + STACK_BIAS);
1416 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.first()->as_FloatRegister());
1417 }
1418 } else if (dst.first()->is_Register()) {
1419 // fpr -> gpr
1420 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), FP, -4 + STACK_BIAS);
1421 __ ld(FP, -4 + STACK_BIAS, dst.first()->as_Register());
1422 } else {
1423 // fpr -> fpr
1424 // In theory these overlap but the ordering is such that this is likely a nop
1425 if ( src.first() != dst.first()) {
1426 __ fmov(FloatRegisterImpl::S, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1427 }
1428 }
1429 }
1430 }
1432 static void split_long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1433 VMRegPair src_lo(src.first());
1434 VMRegPair src_hi(src.second());
1435 VMRegPair dst_lo(dst.first());
1436 VMRegPair dst_hi(dst.second());
1437 simple_move32(masm, src_lo, dst_lo);
1438 simple_move32(masm, src_hi, dst_hi);
1439 }
1441 // A long move
1442 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1444 // Do the simple ones here else do two int moves
1445 if (src.is_single_phys_reg() ) {
1446 if (dst.is_single_phys_reg()) {
1447 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1448 } else {
1449 // split src into two separate registers
1450 // Remember hi means hi address or lsw on sparc
1451 // Move msw to lsw
1452 if (dst.second()->is_reg()) {
1453 // MSW -> MSW
1454 __ srax(src.first()->as_Register(), 32, dst.first()->as_Register());
1455 // Now LSW -> LSW
1456 // this will only move lo -> lo and ignore hi
1457 VMRegPair split(dst.second());
1458 simple_move32(masm, src, split);
1459 } else {
1460 VMRegPair split(src.first(), L4->as_VMReg());
1461 // MSW -> MSW (lo ie. first word)
1462 __ srax(src.first()->as_Register(), 32, L4);
1463 split_long_move(masm, split, dst);
1464 }
1465 }
1466 } else if (dst.is_single_phys_reg()) {
1467 if (src.is_adjacent_aligned_on_stack(2)) {
1468 __ ldx(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1469 } else {
1470 // dst is a single reg.
1471 // Remember lo is low address not msb for stack slots
1472 // and lo is the "real" register for registers
1473 // src is
1475 VMRegPair split;
1477 if (src.first()->is_reg()) {
1478 // src.lo (msw) is a reg, src.hi is stk/reg
1479 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> src.lo [the MSW is in the LSW of the reg]
1480 split.set_pair(dst.first(), src.first());
1481 } else {
1482 // msw is stack move to L5
1483 // lsw is stack move to dst.lo (real reg)
1484 // we will move: src.hi (LSW) -> dst.lo, src.lo (MSW) -> L5
1485 split.set_pair(dst.first(), L5->as_VMReg());
1486 }
1488 // src.lo -> src.lo/L5, src.hi -> dst.lo (the real reg)
1489 // msw -> src.lo/L5, lsw -> dst.lo
1490 split_long_move(masm, src, split);
1492 // So dst now has the low order correct position the
1493 // msw half
1494 __ sllx(split.first()->as_Register(), 32, L5);
1496 const Register d = dst.first()->as_Register();
1497 __ or3(L5, d, d);
1498 }
1499 } else {
1500 // For LP64 we can probably do better.
1501 split_long_move(masm, src, dst);
1502 }
1503 }
1505 // A double move
1506 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1508 // The painful thing here is that like long_move a VMRegPair might be
1509 // 1: a single physical register
1510 // 2: two physical registers (v8)
1511 // 3: a physical reg [lo] and a stack slot [hi] (v8)
1512 // 4: two stack slots
1514 // Since src is always a java calling convention we know that the src pair
1515 // is always either all registers or all stack (and aligned?)
1517 // in a register [lo] and a stack slot [hi]
1518 if (src.first()->is_stack()) {
1519 if (dst.first()->is_stack()) {
1520 // stack to stack the easiest of the bunch
1521 // ought to be a way to do this where if alignment is ok we use ldd/std when possible
1522 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1523 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1524 __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1525 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1526 } else {
1527 // stack to reg
1528 if (dst.second()->is_stack()) {
1529 // stack -> reg, stack -> stack
1530 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1531 if (dst.first()->is_Register()) {
1532 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1533 } else {
1534 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1535 }
1536 // This was missing. (very rare case)
1537 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1538 } else {
1539 // stack -> reg
1540 // Eventually optimize for alignment QQQ
1541 if (dst.first()->is_Register()) {
1542 __ ld(FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_Register());
1543 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_Register());
1544 } else {
1545 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.first()) + STACK_BIAS, dst.first()->as_FloatRegister());
1546 __ ldf(FloatRegisterImpl::S, FP, reg2offset(src.second()) + STACK_BIAS, dst.second()->as_FloatRegister());
1547 }
1548 }
1549 }
1550 } else if (dst.first()->is_stack()) {
1551 // reg to stack
1552 if (src.first()->is_Register()) {
1553 // Eventually optimize for alignment QQQ
1554 __ st(src.first()->as_Register(), SP, reg2offset(dst.first()) + STACK_BIAS);
1555 if (src.second()->is_stack()) {
1556 __ ld(FP, reg2offset(src.second()) + STACK_BIAS, L4);
1557 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1558 } else {
1559 __ st(src.second()->as_Register(), SP, reg2offset(dst.second()) + STACK_BIAS);
1560 }
1561 } else {
1562 // fpr to stack
1563 if (src.second()->is_stack()) {
1564 ShouldNotReachHere();
1565 } else {
1566 // Is the stack aligned?
1567 if (reg2offset(dst.first()) & 0x7) {
1568 // No do as pairs
1569 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1570 __ stf(FloatRegisterImpl::S, src.second()->as_FloatRegister(), SP, reg2offset(dst.second()) + STACK_BIAS);
1571 } else {
1572 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), SP, reg2offset(dst.first()) + STACK_BIAS);
1573 }
1574 }
1575 }
1576 } else {
1577 // reg to reg
1578 if (src.first()->is_Register()) {
1579 if (dst.first()->is_Register()) {
1580 // gpr -> gpr
1581 __ mov(src.first()->as_Register(), dst.first()->as_Register());
1582 __ mov(src.second()->as_Register(), dst.second()->as_Register());
1583 } else {
1584 // gpr -> fpr
1585 // ought to be able to do a single store
1586 __ stx(src.first()->as_Register(), FP, -8 + STACK_BIAS);
1587 __ stx(src.second()->as_Register(), FP, -4 + STACK_BIAS);
1588 // ought to be able to do a single load
1589 __ ldf(FloatRegisterImpl::S, FP, -8 + STACK_BIAS, dst.first()->as_FloatRegister());
1590 __ ldf(FloatRegisterImpl::S, FP, -4 + STACK_BIAS, dst.second()->as_FloatRegister());
1591 }
1592 } else if (dst.first()->is_Register()) {
1593 // fpr -> gpr
1594 // ought to be able to do a single store
1595 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(), FP, -8 + STACK_BIAS);
1596 // ought to be able to do a single load
1597 // REMEMBER first() is low address not LSB
1598 __ ld(FP, -8 + STACK_BIAS, dst.first()->as_Register());
1599 if (dst.second()->is_Register()) {
1600 __ ld(FP, -4 + STACK_BIAS, dst.second()->as_Register());
1601 } else {
1602 __ ld(FP, -4 + STACK_BIAS, L4);
1603 __ st(L4, SP, reg2offset(dst.second()) + STACK_BIAS);
1604 }
1605 } else {
1606 // fpr -> fpr
1607 // In theory these overlap but the ordering is such that this is likely a nop
1608 if ( src.first() != dst.first()) {
1609 __ fmov(FloatRegisterImpl::D, src.first()->as_FloatRegister(), dst.first()->as_FloatRegister());
1610 }
1611 }
1612 }
1613 }
1615 // Creates an inner frame if one hasn't already been created, and
1616 // saves a copy of the thread in L7_thread_cache
1617 static void create_inner_frame(MacroAssembler* masm, bool* already_created) {
1618 if (!*already_created) {
1619 __ save_frame(0);
1620 // Save thread in L7 (INNER FRAME); it crosses a bunch of VM calls below
1621 // Don't use save_thread because it smashes G2 and we merely want to save a
1622 // copy
1623 __ mov(G2_thread, L7_thread_cache);
1624 *already_created = true;
1625 }
1626 }
1629 static void save_or_restore_arguments(MacroAssembler* masm,
1630 const int stack_slots,
1631 const int total_in_args,
1632 const int arg_save_area,
1633 OopMap* map,
1634 VMRegPair* in_regs,
1635 BasicType* in_sig_bt) {
1636 // if map is non-NULL then the code should store the values,
1637 // otherwise it should load them.
1638 if (map != NULL) {
1639 // Fill in the map
1640 for (int i = 0; i < total_in_args; i++) {
1641 if (in_sig_bt[i] == T_ARRAY) {
1642 if (in_regs[i].first()->is_stack()) {
1643 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1644 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots));
1645 } else if (in_regs[i].first()->is_Register()) {
1646 map->set_oop(in_regs[i].first());
1647 } else {
1648 ShouldNotReachHere();
1649 }
1650 }
1651 }
1652 }
1654 // Save or restore double word values
1655 int handle_index = 0;
1656 for (int i = 0; i < total_in_args; i++) {
1657 int slot = handle_index + arg_save_area;
1658 int offset = slot * VMRegImpl::stack_slot_size;
1659 if (in_sig_bt[i] == T_LONG && in_regs[i].first()->is_Register()) {
1660 const Register reg = in_regs[i].first()->as_Register();
1661 if (reg->is_global()) {
1662 handle_index += 2;
1663 assert(handle_index <= stack_slots, "overflow");
1664 if (map != NULL) {
1665 __ stx(reg, SP, offset + STACK_BIAS);
1666 } else {
1667 __ ldx(SP, offset + STACK_BIAS, reg);
1668 }
1669 }
1670 } else if (in_sig_bt[i] == T_DOUBLE && in_regs[i].first()->is_FloatRegister()) {
1671 handle_index += 2;
1672 assert(handle_index <= stack_slots, "overflow");
1673 if (map != NULL) {
1674 __ stf(FloatRegisterImpl::D, in_regs[i].first()->as_FloatRegister(), SP, offset + STACK_BIAS);
1675 } else {
1676 __ ldf(FloatRegisterImpl::D, SP, offset + STACK_BIAS, in_regs[i].first()->as_FloatRegister());
1677 }
1678 }
1679 }
1680 // Save floats
1681 for (int i = 0; i < total_in_args; i++) {
1682 int slot = handle_index + arg_save_area;
1683 int offset = slot * VMRegImpl::stack_slot_size;
1684 if (in_sig_bt[i] == T_FLOAT && in_regs[i].first()->is_FloatRegister()) {
1685 handle_index++;
1686 assert(handle_index <= stack_slots, "overflow");
1687 if (map != NULL) {
1688 __ stf(FloatRegisterImpl::S, in_regs[i].first()->as_FloatRegister(), SP, offset + STACK_BIAS);
1689 } else {
1690 __ ldf(FloatRegisterImpl::S, SP, offset + STACK_BIAS, in_regs[i].first()->as_FloatRegister());
1691 }
1692 }
1693 }
1695 }
1698 // Check GC_locker::needs_gc and enter the runtime if it's true. This
1699 // keeps a new JNI critical region from starting until a GC has been
1700 // forced. Save down any oops in registers and describe them in an
1701 // OopMap.
1702 static void check_needs_gc_for_critical_native(MacroAssembler* masm,
1703 const int stack_slots,
1704 const int total_in_args,
1705 const int arg_save_area,
1706 OopMapSet* oop_maps,
1707 VMRegPair* in_regs,
1708 BasicType* in_sig_bt) {
1709 __ block_comment("check GC_locker::needs_gc");
1710 Label cont;
1711 AddressLiteral sync_state(GC_locker::needs_gc_address());
1712 __ load_bool_contents(sync_state, G3_scratch);
1713 __ cmp_zero_and_br(Assembler::equal, G3_scratch, cont);
1714 __ delayed()->nop();
1716 // Save down any values that are live in registers and call into the
1717 // runtime to halt for a GC
1718 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1719 save_or_restore_arguments(masm, stack_slots, total_in_args,
1720 arg_save_area, map, in_regs, in_sig_bt);
1722 __ mov(G2_thread, L7_thread_cache);
1724 __ set_last_Java_frame(SP, noreg);
1726 __ block_comment("block_for_jni_critical");
1727 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical), relocInfo::runtime_call_type);
1728 __ delayed()->mov(L7_thread_cache, O0);
1729 oop_maps->add_gc_map( __ offset(), map);
1731 __ restore_thread(L7_thread_cache); // restore G2_thread
1732 __ reset_last_Java_frame();
1734 // Reload all the register arguments
1735 save_or_restore_arguments(masm, stack_slots, total_in_args,
1736 arg_save_area, NULL, in_regs, in_sig_bt);
1738 __ bind(cont);
1739 #ifdef ASSERT
1740 if (StressCriticalJNINatives) {
1741 // Stress register saving
1742 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1743 save_or_restore_arguments(masm, stack_slots, total_in_args,
1744 arg_save_area, map, in_regs, in_sig_bt);
1745 // Destroy argument registers
1746 for (int i = 0; i < total_in_args; i++) {
1747 if (in_regs[i].first()->is_Register()) {
1748 const Register reg = in_regs[i].first()->as_Register();
1749 if (reg->is_global()) {
1750 __ mov(G0, reg);
1751 }
1752 } else if (in_regs[i].first()->is_FloatRegister()) {
1753 __ fneg(FloatRegisterImpl::D, in_regs[i].first()->as_FloatRegister(), in_regs[i].first()->as_FloatRegister());
1754 }
1755 }
1757 save_or_restore_arguments(masm, stack_slots, total_in_args,
1758 arg_save_area, NULL, in_regs, in_sig_bt);
1759 }
1760 #endif
1761 }
1763 // Unpack an array argument into a pointer to the body and the length
1764 // if the array is non-null, otherwise pass 0 for both.
1765 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) {
1766 // Pass the length, ptr pair
1767 Label is_null, done;
1768 if (reg.first()->is_stack()) {
1769 VMRegPair tmp = reg64_to_VMRegPair(L2);
1770 // Load the arg up from the stack
1771 move_ptr(masm, reg, tmp);
1772 reg = tmp;
1773 }
1774 __ cmp(reg.first()->as_Register(), G0);
1775 __ brx(Assembler::equal, false, Assembler::pt, is_null);
1776 __ delayed()->add(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type), L4);
1777 move_ptr(masm, reg64_to_VMRegPair(L4), body_arg);
1778 __ ld(reg.first()->as_Register(), arrayOopDesc::length_offset_in_bytes(), L4);
1779 move32_64(masm, reg64_to_VMRegPair(L4), length_arg);
1780 __ ba_short(done);
1781 __ bind(is_null);
1782 // Pass zeros
1783 move_ptr(masm, reg64_to_VMRegPair(G0), body_arg);
1784 move32_64(masm, reg64_to_VMRegPair(G0), length_arg);
1785 __ bind(done);
1786 }
1788 static void verify_oop_args(MacroAssembler* masm,
1789 methodHandle method,
1790 const BasicType* sig_bt,
1791 const VMRegPair* regs) {
1792 Register temp_reg = G5_method; // not part of any compiled calling seq
1793 if (VerifyOops) {
1794 for (int i = 0; i < method->size_of_parameters(); i++) {
1795 if (sig_bt[i] == T_OBJECT ||
1796 sig_bt[i] == T_ARRAY) {
1797 VMReg r = regs[i].first();
1798 assert(r->is_valid(), "bad oop arg");
1799 if (r->is_stack()) {
1800 RegisterOrConstant ld_off = reg2offset(r) + STACK_BIAS;
1801 ld_off = __ ensure_simm13_or_reg(ld_off, temp_reg);
1802 __ ld_ptr(SP, ld_off, temp_reg);
1803 __ verify_oop(temp_reg);
1804 } else {
1805 __ verify_oop(r->as_Register());
1806 }
1807 }
1808 }
1809 }
1810 }
1812 static void gen_special_dispatch(MacroAssembler* masm,
1813 methodHandle method,
1814 const BasicType* sig_bt,
1815 const VMRegPair* regs) {
1816 verify_oop_args(masm, method, sig_bt, regs);
1817 vmIntrinsics::ID iid = method->intrinsic_id();
1819 // Now write the args into the outgoing interpreter space
1820 bool has_receiver = false;
1821 Register receiver_reg = noreg;
1822 int member_arg_pos = -1;
1823 Register member_reg = noreg;
1824 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1825 if (ref_kind != 0) {
1826 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1827 member_reg = G5_method; // known to be free at this point
1828 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1829 } else if (iid == vmIntrinsics::_invokeBasic) {
1830 has_receiver = true;
1831 } else {
1832 fatal(err_msg_res("unexpected intrinsic id %d", iid));
1833 }
1835 if (member_reg != noreg) {
1836 // Load the member_arg into register, if necessary.
1837 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1838 VMReg r = regs[member_arg_pos].first();
1839 if (r->is_stack()) {
1840 RegisterOrConstant ld_off = reg2offset(r) + STACK_BIAS;
1841 ld_off = __ ensure_simm13_or_reg(ld_off, member_reg);
1842 __ ld_ptr(SP, ld_off, member_reg);
1843 } else {
1844 // no data motion is needed
1845 member_reg = r->as_Register();
1846 }
1847 }
1849 if (has_receiver) {
1850 // Make sure the receiver is loaded into a register.
1851 assert(method->size_of_parameters() > 0, "oob");
1852 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1853 VMReg r = regs[0].first();
1854 assert(r->is_valid(), "bad receiver arg");
1855 if (r->is_stack()) {
1856 // Porting note: This assumes that compiled calling conventions always
1857 // pass the receiver oop in a register. If this is not true on some
1858 // platform, pick a temp and load the receiver from stack.
1859 fatal("receiver always in a register");
1860 receiver_reg = G3_scratch; // known to be free at this point
1861 RegisterOrConstant ld_off = reg2offset(r) + STACK_BIAS;
1862 ld_off = __ ensure_simm13_or_reg(ld_off, member_reg);
1863 __ ld_ptr(SP, ld_off, receiver_reg);
1864 } else {
1865 // no data motion is needed
1866 receiver_reg = r->as_Register();
1867 }
1868 }
1870 // Figure out which address we are really jumping to:
1871 MethodHandles::generate_method_handle_dispatch(masm, iid,
1872 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1873 }
1875 // ---------------------------------------------------------------------------
1876 // Generate a native wrapper for a given method. The method takes arguments
1877 // in the Java compiled code convention, marshals them to the native
1878 // convention (handlizes oops, etc), transitions to native, makes the call,
1879 // returns to java state (possibly blocking), unhandlizes any result and
1880 // returns.
1881 //
1882 // Critical native functions are a shorthand for the use of
1883 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1884 // functions. The wrapper is expected to unpack the arguments before
1885 // passing them to the callee and perform checks before and after the
1886 // native call to ensure that they GC_locker
1887 // lock_critical/unlock_critical semantics are followed. Some other
1888 // parts of JNI setup are skipped like the tear down of the JNI handle
1889 // block and the check for pending exceptions it's impossible for them
1890 // to be thrown.
1891 //
1892 // They are roughly structured like this:
1893 // if (GC_locker::needs_gc())
1894 // SharedRuntime::block_for_jni_critical();
1895 // tranistion to thread_in_native
1896 // unpack arrray arguments and call native entry point
1897 // check for safepoint in progress
1898 // check if any thread suspend flags are set
1899 // call into JVM and possible unlock the JNI critical
1900 // if a GC was suppressed while in the critical native.
1901 // transition back to thread_in_Java
1902 // return to caller
1903 //
1904 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1905 methodHandle method,
1906 int compile_id,
1907 BasicType* in_sig_bt,
1908 VMRegPair* in_regs,
1909 BasicType ret_type) {
1910 if (method->is_method_handle_intrinsic()) {
1911 vmIntrinsics::ID iid = method->intrinsic_id();
1912 intptr_t start = (intptr_t)__ pc();
1913 int vep_offset = ((intptr_t)__ pc()) - start;
1914 gen_special_dispatch(masm,
1915 method,
1916 in_sig_bt,
1917 in_regs);
1918 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1919 __ flush();
1920 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1921 return nmethod::new_native_nmethod(method,
1922 compile_id,
1923 masm->code(),
1924 vep_offset,
1925 frame_complete,
1926 stack_slots / VMRegImpl::slots_per_word,
1927 in_ByteSize(-1),
1928 in_ByteSize(-1),
1929 (OopMapSet*)NULL);
1930 }
1931 bool is_critical_native = true;
1932 address native_func = method->critical_native_function();
1933 if (native_func == NULL) {
1934 native_func = method->native_function();
1935 is_critical_native = false;
1936 }
1937 assert(native_func != NULL, "must have function");
1939 // Native nmethod wrappers never take possesion of the oop arguments.
1940 // So the caller will gc the arguments. The only thing we need an
1941 // oopMap for is if the call is static
1942 //
1943 // An OopMap for lock (and class if static), and one for the VM call itself
1944 OopMapSet *oop_maps = new OopMapSet();
1945 intptr_t start = (intptr_t)__ pc();
1947 // First thing make an ic check to see if we should even be here
1948 {
1949 Label L;
1950 const Register temp_reg = G3_scratch;
1951 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
1952 __ verify_oop(O0);
1953 __ load_klass(O0, temp_reg);
1954 __ cmp_and_brx_short(temp_reg, G5_inline_cache_reg, Assembler::equal, Assembler::pt, L);
1956 __ jump_to(ic_miss, temp_reg);
1957 __ delayed()->nop();
1958 __ align(CodeEntryAlignment);
1959 __ bind(L);
1960 }
1962 int vep_offset = ((intptr_t)__ pc()) - start;
1964 #ifdef COMPILER1
1965 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
1966 // Object.hashCode can pull the hashCode from the header word
1967 // instead of doing a full VM transition once it's been computed.
1968 // Since hashCode is usually polymorphic at call sites we can't do
1969 // this optimization at the call site without a lot of work.
1970 Label slowCase;
1971 Register receiver = O0;
1972 Register result = O0;
1973 Register header = G3_scratch;
1974 Register hash = G3_scratch; // overwrite header value with hash value
1975 Register mask = G1; // to get hash field from header
1977 // Read the header and build a mask to get its hash field. Give up if the object is not unlocked.
1978 // We depend on hash_mask being at most 32 bits and avoid the use of
1979 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
1980 // vm: see markOop.hpp.
1981 __ ld_ptr(receiver, oopDesc::mark_offset_in_bytes(), header);
1982 __ sethi(markOopDesc::hash_mask, mask);
1983 __ btst(markOopDesc::unlocked_value, header);
1984 __ br(Assembler::zero, false, Assembler::pn, slowCase);
1985 if (UseBiasedLocking) {
1986 // Check if biased and fall through to runtime if so
1987 __ delayed()->nop();
1988 __ btst(markOopDesc::biased_lock_bit_in_place, header);
1989 __ br(Assembler::notZero, false, Assembler::pn, slowCase);
1990 }
1991 __ delayed()->or3(mask, markOopDesc::hash_mask & 0x3ff, mask);
1993 // Check for a valid (non-zero) hash code and get its value.
1994 #ifdef _LP64
1995 __ srlx(header, markOopDesc::hash_shift, hash);
1996 #else
1997 __ srl(header, markOopDesc::hash_shift, hash);
1998 #endif
1999 __ andcc(hash, mask, hash);
2000 __ br(Assembler::equal, false, Assembler::pn, slowCase);
2001 __ delayed()->nop();
2003 // leaf return.
2004 __ retl();
2005 __ delayed()->mov(hash, result);
2006 __ bind(slowCase);
2007 }
2008 #endif // COMPILER1
2011 // We have received a description of where all the java arg are located
2012 // on entry to the wrapper. We need to convert these args to where
2013 // the jni function will expect them. To figure out where they go
2014 // we convert the java signature to a C signature by inserting
2015 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2017 const int total_in_args = method->size_of_parameters();
2018 int total_c_args = total_in_args;
2019 int total_save_slots = 6 * VMRegImpl::slots_per_word;
2020 if (!is_critical_native) {
2021 total_c_args += 1;
2022 if (method->is_static()) {
2023 total_c_args++;
2024 }
2025 } else {
2026 for (int i = 0; i < total_in_args; i++) {
2027 if (in_sig_bt[i] == T_ARRAY) {
2028 // These have to be saved and restored across the safepoint
2029 total_c_args++;
2030 }
2031 }
2032 }
2034 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2035 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2036 BasicType* in_elem_bt = NULL;
2038 int argc = 0;
2039 if (!is_critical_native) {
2040 out_sig_bt[argc++] = T_ADDRESS;
2041 if (method->is_static()) {
2042 out_sig_bt[argc++] = T_OBJECT;
2043 }
2045 for (int i = 0; i < total_in_args ; i++ ) {
2046 out_sig_bt[argc++] = in_sig_bt[i];
2047 }
2048 } else {
2049 Thread* THREAD = Thread::current();
2050 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args);
2051 SignatureStream ss(method->signature());
2052 for (int i = 0; i < total_in_args ; i++ ) {
2053 if (in_sig_bt[i] == T_ARRAY) {
2054 // Arrays are passed as int, elem* pair
2055 out_sig_bt[argc++] = T_INT;
2056 out_sig_bt[argc++] = T_ADDRESS;
2057 Symbol* atype = ss.as_symbol(CHECK_NULL);
2058 const char* at = atype->as_C_string();
2059 if (strlen(at) == 2) {
2060 assert(at[0] == '[', "must be");
2061 switch (at[1]) {
2062 case 'B': in_elem_bt[i] = T_BYTE; break;
2063 case 'C': in_elem_bt[i] = T_CHAR; break;
2064 case 'D': in_elem_bt[i] = T_DOUBLE; break;
2065 case 'F': in_elem_bt[i] = T_FLOAT; break;
2066 case 'I': in_elem_bt[i] = T_INT; break;
2067 case 'J': in_elem_bt[i] = T_LONG; break;
2068 case 'S': in_elem_bt[i] = T_SHORT; break;
2069 case 'Z': in_elem_bt[i] = T_BOOLEAN; break;
2070 default: ShouldNotReachHere();
2071 }
2072 }
2073 } else {
2074 out_sig_bt[argc++] = in_sig_bt[i];
2075 in_elem_bt[i] = T_VOID;
2076 }
2077 if (in_sig_bt[i] != T_VOID) {
2078 assert(in_sig_bt[i] == ss.type(), "must match");
2079 ss.next();
2080 }
2081 }
2082 }
2084 // Now figure out where the args must be stored and how much stack space
2085 // they require (neglecting out_preserve_stack_slots but space for storing
2086 // the 1st six register arguments). It's weird see int_stk_helper.
2087 //
2088 int out_arg_slots;
2089 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args);
2091 if (is_critical_native) {
2092 // Critical natives may have to call out so they need a save area
2093 // for register arguments.
2094 int double_slots = 0;
2095 int single_slots = 0;
2096 for ( int i = 0; i < total_in_args; i++) {
2097 if (in_regs[i].first()->is_Register()) {
2098 const Register reg = in_regs[i].first()->as_Register();
2099 switch (in_sig_bt[i]) {
2100 case T_ARRAY:
2101 case T_BOOLEAN:
2102 case T_BYTE:
2103 case T_SHORT:
2104 case T_CHAR:
2105 case T_INT: assert(reg->is_in(), "don't need to save these"); break;
2106 case T_LONG: if (reg->is_global()) double_slots++; break;
2107 default: ShouldNotReachHere();
2108 }
2109 } else if (in_regs[i].first()->is_FloatRegister()) {
2110 switch (in_sig_bt[i]) {
2111 case T_FLOAT: single_slots++; break;
2112 case T_DOUBLE: double_slots++; break;
2113 default: ShouldNotReachHere();
2114 }
2115 }
2116 }
2117 total_save_slots = double_slots * 2 + single_slots;
2118 }
2120 // Compute framesize for the wrapper. We need to handlize all oops in
2121 // registers. We must create space for them here that is disjoint from
2122 // the windowed save area because we have no control over when we might
2123 // flush the window again and overwrite values that gc has since modified.
2124 // (The live window race)
2125 //
2126 // We always just allocate 6 word for storing down these object. This allow
2127 // us to simply record the base and use the Ireg number to decide which
2128 // slot to use. (Note that the reg number is the inbound number not the
2129 // outbound number).
2130 // We must shuffle args to match the native convention, and include var-args space.
2132 // Calculate the total number of stack slots we will need.
2134 // First count the abi requirement plus all of the outgoing args
2135 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2137 // Now the space for the inbound oop handle area
2139 int oop_handle_offset = round_to(stack_slots, 2);
2140 stack_slots += total_save_slots;
2142 // Now any space we need for handlizing a klass if static method
2144 int klass_slot_offset = 0;
2145 int klass_offset = -1;
2146 int lock_slot_offset = 0;
2147 bool is_static = false;
2149 if (method->is_static()) {
2150 klass_slot_offset = stack_slots;
2151 stack_slots += VMRegImpl::slots_per_word;
2152 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2153 is_static = true;
2154 }
2156 // Plus a lock if needed
2158 if (method->is_synchronized()) {
2159 lock_slot_offset = stack_slots;
2160 stack_slots += VMRegImpl::slots_per_word;
2161 }
2163 // Now a place to save return value or as a temporary for any gpr -> fpr moves
2164 stack_slots += 2;
2166 // Ok The space we have allocated will look like:
2167 //
2168 //
2169 // FP-> | |
2170 // |---------------------|
2171 // | 2 slots for moves |
2172 // |---------------------|
2173 // | lock box (if sync) |
2174 // |---------------------| <- lock_slot_offset
2175 // | klass (if static) |
2176 // |---------------------| <- klass_slot_offset
2177 // | oopHandle area |
2178 // |---------------------| <- oop_handle_offset
2179 // | outbound memory |
2180 // | based arguments |
2181 // | |
2182 // |---------------------|
2183 // | vararg area |
2184 // |---------------------|
2185 // | |
2186 // SP-> | out_preserved_slots |
2187 //
2188 //
2191 // Now compute actual number of stack words we need rounding to make
2192 // stack properly aligned.
2193 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
2195 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2197 // Generate stack overflow check before creating frame
2198 __ generate_stack_overflow_check(stack_size);
2200 // Generate a new frame for the wrapper.
2201 __ save(SP, -stack_size, SP);
2203 int frame_complete = ((intptr_t)__ pc()) - start;
2205 __ verify_thread();
2207 if (is_critical_native) {
2208 check_needs_gc_for_critical_native(masm, stack_slots, total_in_args,
2209 oop_handle_offset, oop_maps, in_regs, in_sig_bt);
2210 }
2212 //
2213 // We immediately shuffle the arguments so that any vm call we have to
2214 // make from here on out (sync slow path, jvmti, etc.) we will have
2215 // captured the oops from our caller and have a valid oopMap for
2216 // them.
2218 // -----------------
2219 // The Grand Shuffle
2220 //
2221 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2222 // (derived from JavaThread* which is in L7_thread_cache) and, if static,
2223 // the class mirror instead of a receiver. This pretty much guarantees that
2224 // register layout will not match. We ignore these extra arguments during
2225 // the shuffle. The shuffle is described by the two calling convention
2226 // vectors we have in our possession. We simply walk the java vector to
2227 // get the source locations and the c vector to get the destinations.
2228 // Because we have a new window and the argument registers are completely
2229 // disjoint ( I0 -> O1, I1 -> O2, ...) we have nothing to worry about
2230 // here.
2232 // This is a trick. We double the stack slots so we can claim
2233 // the oops in the caller's frame. Since we are sure to have
2234 // more args than the caller doubling is enough to make
2235 // sure we can capture all the incoming oop args from the
2236 // caller.
2237 //
2238 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2239 // Record sp-based slot for receiver on stack for non-static methods
2240 int receiver_offset = -1;
2242 // We move the arguments backward because the floating point registers
2243 // destination will always be to a register with a greater or equal register
2244 // number or the stack.
2246 #ifdef ASSERT
2247 bool reg_destroyed[RegisterImpl::number_of_registers];
2248 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2249 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2250 reg_destroyed[r] = false;
2251 }
2252 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2253 freg_destroyed[f] = false;
2254 }
2256 #endif /* ASSERT */
2258 for ( int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0 ; i--, c_arg-- ) {
2260 #ifdef ASSERT
2261 if (in_regs[i].first()->is_Register()) {
2262 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "ack!");
2263 } else if (in_regs[i].first()->is_FloatRegister()) {
2264 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)], "ack!");
2265 }
2266 if (out_regs[c_arg].first()->is_Register()) {
2267 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
2268 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
2269 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding(FloatRegisterImpl::S)] = true;
2270 }
2271 #endif /* ASSERT */
2273 switch (in_sig_bt[i]) {
2274 case T_ARRAY:
2275 if (is_critical_native) {
2276 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg], out_regs[c_arg - 1]);
2277 c_arg--;
2278 break;
2279 }
2280 case T_OBJECT:
2281 assert(!is_critical_native, "no oop arguments");
2282 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
2283 ((i == 0) && (!is_static)),
2284 &receiver_offset);
2285 break;
2286 case T_VOID:
2287 break;
2289 case T_FLOAT:
2290 float_move(masm, in_regs[i], out_regs[c_arg]);
2291 break;
2293 case T_DOUBLE:
2294 assert( i + 1 < total_in_args &&
2295 in_sig_bt[i + 1] == T_VOID &&
2296 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
2297 double_move(masm, in_regs[i], out_regs[c_arg]);
2298 break;
2300 case T_LONG :
2301 long_move(masm, in_regs[i], out_regs[c_arg]);
2302 break;
2304 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2306 default:
2307 move32_64(masm, in_regs[i], out_regs[c_arg]);
2308 }
2309 }
2311 // Pre-load a static method's oop into O1. Used both by locking code and
2312 // the normal JNI call code.
2313 if (method->is_static() && !is_critical_native) {
2314 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()), O1);
2316 // Now handlize the static class mirror in O1. It's known not-null.
2317 __ st_ptr(O1, SP, klass_offset + STACK_BIAS);
2318 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2319 __ add(SP, klass_offset + STACK_BIAS, O1);
2320 }
2323 const Register L6_handle = L6;
2325 if (method->is_synchronized()) {
2326 assert(!is_critical_native, "unhandled");
2327 __ mov(O1, L6_handle);
2328 }
2330 // We have all of the arguments setup at this point. We MUST NOT touch any Oregs
2331 // except O6/O7. So if we must call out we must push a new frame. We immediately
2332 // push a new frame and flush the windows.
2333 #ifdef _LP64
2334 intptr_t thepc = (intptr_t) __ pc();
2335 {
2336 address here = __ pc();
2337 // Call the next instruction
2338 __ call(here + 8, relocInfo::none);
2339 __ delayed()->nop();
2340 }
2341 #else
2342 intptr_t thepc = __ load_pc_address(O7, 0);
2343 #endif /* _LP64 */
2345 // We use the same pc/oopMap repeatedly when we call out
2346 oop_maps->add_gc_map(thepc - start, map);
2348 // O7 now has the pc loaded that we will use when we finally call to native.
2350 // Save thread in L7; it crosses a bunch of VM calls below
2351 // Don't use save_thread because it smashes G2 and we merely
2352 // want to save a copy
2353 __ mov(G2_thread, L7_thread_cache);
2356 // If we create an inner frame once is plenty
2357 // when we create it we must also save G2_thread
2358 bool inner_frame_created = false;
2360 // dtrace method entry support
2361 {
2362 SkipIfEqual skip_if(
2363 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2364 // create inner frame
2365 __ save_frame(0);
2366 __ mov(G2_thread, L7_thread_cache);
2367 __ set_metadata_constant(method(), O1);
2368 __ call_VM_leaf(L7_thread_cache,
2369 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2370 G2_thread, O1);
2371 __ restore();
2372 }
2374 // RedefineClasses() tracing support for obsolete method entry
2375 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
2376 // create inner frame
2377 __ save_frame(0);
2378 __ mov(G2_thread, L7_thread_cache);
2379 __ set_metadata_constant(method(), O1);
2380 __ call_VM_leaf(L7_thread_cache,
2381 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
2382 G2_thread, O1);
2383 __ restore();
2384 }
2386 // We are in the jni frame unless saved_frame is true in which case
2387 // we are in one frame deeper (the "inner" frame). If we are in the
2388 // "inner" frames the args are in the Iregs and if the jni frame then
2389 // they are in the Oregs.
2390 // If we ever need to go to the VM (for locking, jvmti) then
2391 // we will always be in the "inner" frame.
2393 // Lock a synchronized method
2394 int lock_offset = -1; // Set if locked
2395 if (method->is_synchronized()) {
2396 Register Roop = O1;
2397 const Register L3_box = L3;
2399 create_inner_frame(masm, &inner_frame_created);
2401 __ ld_ptr(I1, 0, O1);
2402 Label done;
2404 lock_offset = (lock_slot_offset * VMRegImpl::stack_slot_size);
2405 __ add(FP, lock_offset+STACK_BIAS, L3_box);
2406 #ifdef ASSERT
2407 if (UseBiasedLocking) {
2408 // making the box point to itself will make it clear it went unused
2409 // but also be obviously invalid
2410 __ st_ptr(L3_box, L3_box, 0);
2411 }
2412 #endif // ASSERT
2413 //
2414 // Compiler_lock_object (Roop, Rmark, Rbox, Rscratch) -- kills Rmark, Rbox, Rscratch
2415 //
2416 __ compiler_lock_object(Roop, L1, L3_box, L2);
2417 __ br(Assembler::equal, false, Assembler::pt, done);
2418 __ delayed() -> add(FP, lock_offset+STACK_BIAS, L3_box);
2421 // None of the above fast optimizations worked so we have to get into the
2422 // slow case of monitor enter. Inline a special case of call_VM that
2423 // disallows any pending_exception.
2424 __ mov(Roop, O0); // Need oop in O0
2425 __ mov(L3_box, O1);
2427 // Record last_Java_sp, in case the VM code releases the JVM lock.
2429 __ set_last_Java_frame(FP, I7);
2431 // do the call
2432 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), relocInfo::runtime_call_type);
2433 __ delayed()->mov(L7_thread_cache, O2);
2435 __ restore_thread(L7_thread_cache); // restore G2_thread
2436 __ reset_last_Java_frame();
2438 #ifdef ASSERT
2439 { Label L;
2440 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2441 __ br_null_short(O0, Assembler::pt, L);
2442 __ stop("no pending exception allowed on exit from IR::monitorenter");
2443 __ bind(L);
2444 }
2445 #endif
2446 __ bind(done);
2447 }
2450 // Finally just about ready to make the JNI call
2452 __ flushw();
2453 if (inner_frame_created) {
2454 __ restore();
2455 } else {
2456 // Store only what we need from this frame
2457 // QQQ I think that non-v9 (like we care) we don't need these saves
2458 // either as the flush traps and the current window goes too.
2459 __ st_ptr(FP, SP, FP->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2460 __ st_ptr(I7, SP, I7->sp_offset_in_saved_window()*wordSize + STACK_BIAS);
2461 }
2463 // get JNIEnv* which is first argument to native
2464 if (!is_critical_native) {
2465 __ add(G2_thread, in_bytes(JavaThread::jni_environment_offset()), O0);
2466 }
2468 // Use that pc we placed in O7 a while back as the current frame anchor
2469 __ set_last_Java_frame(SP, O7);
2471 // We flushed the windows ages ago now mark them as flushed before transitioning.
2472 __ set(JavaFrameAnchor::flushed, G3_scratch);
2473 __ st(G3_scratch, G2_thread, JavaThread::frame_anchor_offset() + JavaFrameAnchor::flags_offset());
2475 // Transition from _thread_in_Java to _thread_in_native.
2476 __ set(_thread_in_native, G3_scratch);
2478 #ifdef _LP64
2479 AddressLiteral dest(native_func);
2480 __ relocate(relocInfo::runtime_call_type);
2481 __ jumpl_to(dest, O7, O7);
2482 #else
2483 __ call(native_func, relocInfo::runtime_call_type);
2484 #endif
2485 __ delayed()->st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2487 __ restore_thread(L7_thread_cache); // restore G2_thread
2489 // Unpack native results. For int-types, we do any needed sign-extension
2490 // and move things into I0. The return value there will survive any VM
2491 // calls for blocking or unlocking. An FP or OOP result (handle) is done
2492 // specially in the slow-path code.
2493 switch (ret_type) {
2494 case T_VOID: break; // Nothing to do!
2495 case T_FLOAT: break; // Got it where we want it (unless slow-path)
2496 case T_DOUBLE: break; // Got it where we want it (unless slow-path)
2497 // In 64 bits build result is in O0, in O0, O1 in 32bit build
2498 case T_LONG:
2499 #ifndef _LP64
2500 __ mov(O1, I1);
2501 #endif
2502 // Fall thru
2503 case T_OBJECT: // Really a handle
2504 case T_ARRAY:
2505 case T_INT:
2506 __ mov(O0, I0);
2507 break;
2508 case T_BOOLEAN: __ subcc(G0, O0, G0); __ addc(G0, 0, I0); break; // !0 => true; 0 => false
2509 case T_BYTE : __ sll(O0, 24, O0); __ sra(O0, 24, I0); break;
2510 case T_CHAR : __ sll(O0, 16, O0); __ srl(O0, 16, I0); break; // cannot use and3, 0xFFFF too big as immediate value!
2511 case T_SHORT : __ sll(O0, 16, O0); __ sra(O0, 16, I0); break;
2512 break; // Cannot de-handlize until after reclaiming jvm_lock
2513 default:
2514 ShouldNotReachHere();
2515 }
2517 Label after_transition;
2518 // must we block?
2520 // Block, if necessary, before resuming in _thread_in_Java state.
2521 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2522 { Label no_block;
2523 AddressLiteral sync_state(SafepointSynchronize::address_of_state());
2525 // Switch thread to "native transition" state before reading the synchronization state.
2526 // This additional state is necessary because reading and testing the synchronization
2527 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2528 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2529 // VM thread changes sync state to synchronizing and suspends threads for GC.
2530 // Thread A is resumed to finish this native method, but doesn't block here since it
2531 // didn't see any synchronization is progress, and escapes.
2532 __ set(_thread_in_native_trans, G3_scratch);
2533 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2534 if(os::is_MP()) {
2535 if (UseMembar) {
2536 // Force this write out before the read below
2537 __ membar(Assembler::StoreLoad);
2538 } else {
2539 // Write serialization page so VM thread can do a pseudo remote membar.
2540 // We use the current thread pointer to calculate a thread specific
2541 // offset to write to within the page. This minimizes bus traffic
2542 // due to cache line collision.
2543 __ serialize_memory(G2_thread, G1_scratch, G3_scratch);
2544 }
2545 }
2546 __ load_contents(sync_state, G3_scratch);
2547 __ cmp(G3_scratch, SafepointSynchronize::_not_synchronized);
2549 Label L;
2550 Address suspend_state(G2_thread, JavaThread::suspend_flags_offset());
2551 __ br(Assembler::notEqual, false, Assembler::pn, L);
2552 __ delayed()->ld(suspend_state, G3_scratch);
2553 __ cmp_and_br_short(G3_scratch, 0, Assembler::equal, Assembler::pt, no_block);
2554 __ bind(L);
2556 // Block. Save any potential method result value before the operation and
2557 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2558 // lets us share the oopMap we used when we went native rather the create
2559 // a distinct one for this pc
2560 //
2561 save_native_result(masm, ret_type, stack_slots);
2562 if (!is_critical_native) {
2563 __ call_VM_leaf(L7_thread_cache,
2564 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans),
2565 G2_thread);
2566 } else {
2567 __ call_VM_leaf(L7_thread_cache,
2568 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans_and_transition),
2569 G2_thread);
2570 }
2572 // Restore any method result value
2573 restore_native_result(masm, ret_type, stack_slots);
2575 if (is_critical_native) {
2576 // The call above performed the transition to thread_in_Java so
2577 // skip the transition logic below.
2578 __ ba(after_transition);
2579 __ delayed()->nop();
2580 }
2582 __ bind(no_block);
2583 }
2585 // thread state is thread_in_native_trans. Any safepoint blocking has already
2586 // happened so we can now change state to _thread_in_Java.
2587 __ set(_thread_in_Java, G3_scratch);
2588 __ st(G3_scratch, G2_thread, JavaThread::thread_state_offset());
2589 __ bind(after_transition);
2591 Label no_reguard;
2592 __ ld(G2_thread, JavaThread::stack_guard_state_offset(), G3_scratch);
2593 __ cmp_and_br_short(G3_scratch, JavaThread::stack_guard_yellow_disabled, Assembler::notEqual, Assembler::pt, no_reguard);
2595 save_native_result(masm, ret_type, stack_slots);
2596 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2597 __ delayed()->nop();
2599 __ restore_thread(L7_thread_cache); // restore G2_thread
2600 restore_native_result(masm, ret_type, stack_slots);
2602 __ bind(no_reguard);
2604 // Handle possible exception (will unlock if necessary)
2606 // native result if any is live in freg or I0 (and I1 if long and 32bit vm)
2608 // Unlock
2609 if (method->is_synchronized()) {
2610 Label done;
2611 Register I2_ex_oop = I2;
2612 const Register L3_box = L3;
2613 // Get locked oop from the handle we passed to jni
2614 __ ld_ptr(L6_handle, 0, L4);
2615 __ add(SP, lock_offset+STACK_BIAS, L3_box);
2616 // Must save pending exception around the slow-path VM call. Since it's a
2617 // leaf call, the pending exception (if any) can be kept in a register.
2618 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), I2_ex_oop);
2619 // Now unlock
2620 // (Roop, Rmark, Rbox, Rscratch)
2621 __ compiler_unlock_object(L4, L1, L3_box, L2);
2622 __ br(Assembler::equal, false, Assembler::pt, done);
2623 __ delayed()-> add(SP, lock_offset+STACK_BIAS, L3_box);
2625 // save and restore any potential method result value around the unlocking
2626 // operation. Will save in I0 (or stack for FP returns).
2627 save_native_result(masm, ret_type, stack_slots);
2629 // Must clear pending-exception before re-entering the VM. Since this is
2630 // a leaf call, pending-exception-oop can be safely kept in a register.
2631 __ st_ptr(G0, G2_thread, in_bytes(Thread::pending_exception_offset()));
2633 // slow case of monitor enter. Inline a special case of call_VM that
2634 // disallows any pending_exception.
2635 __ mov(L3_box, O1);
2637 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), relocInfo::runtime_call_type);
2638 __ delayed()->mov(L4, O0); // Need oop in O0
2640 __ restore_thread(L7_thread_cache); // restore G2_thread
2642 #ifdef ASSERT
2643 { Label L;
2644 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O0);
2645 __ br_null_short(O0, Assembler::pt, L);
2646 __ stop("no pending exception allowed on exit from IR::monitorexit");
2647 __ bind(L);
2648 }
2649 #endif
2650 restore_native_result(masm, ret_type, stack_slots);
2651 // check_forward_pending_exception jump to forward_exception if any pending
2652 // exception is set. The forward_exception routine expects to see the
2653 // exception in pending_exception and not in a register. Kind of clumsy,
2654 // since all folks who branch to forward_exception must have tested
2655 // pending_exception first and hence have it in a register already.
2656 __ st_ptr(I2_ex_oop, G2_thread, in_bytes(Thread::pending_exception_offset()));
2657 __ bind(done);
2658 }
2660 // Tell dtrace about this method exit
2661 {
2662 SkipIfEqual skip_if(
2663 masm, G3_scratch, &DTraceMethodProbes, Assembler::zero);
2664 save_native_result(masm, ret_type, stack_slots);
2665 __ set_metadata_constant(method(), O1);
2666 __ call_VM_leaf(L7_thread_cache,
2667 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2668 G2_thread, O1);
2669 restore_native_result(masm, ret_type, stack_slots);
2670 }
2672 // Clear "last Java frame" SP and PC.
2673 __ verify_thread(); // G2_thread must be correct
2674 __ reset_last_Java_frame();
2676 // Unpack oop result
2677 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2678 Label L;
2679 __ addcc(G0, I0, G0);
2680 __ brx(Assembler::notZero, true, Assembler::pt, L);
2681 __ delayed()->ld_ptr(I0, 0, I0);
2682 __ mov(G0, I0);
2683 __ bind(L);
2684 __ verify_oop(I0);
2685 }
2687 if (!is_critical_native) {
2688 // reset handle block
2689 __ ld_ptr(G2_thread, in_bytes(JavaThread::active_handles_offset()), L5);
2690 __ st(G0, L5, JNIHandleBlock::top_offset_in_bytes());
2692 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), G3_scratch);
2693 check_forward_pending_exception(masm, G3_scratch);
2694 }
2697 // Return
2699 #ifndef _LP64
2700 if (ret_type == T_LONG) {
2702 // Must leave proper result in O0,O1 and G1 (c2/tiered only)
2703 __ sllx(I0, 32, G1); // Shift bits into high G1
2704 __ srl (I1, 0, I1); // Zero extend O1 (harmless?)
2705 __ or3 (I1, G1, G1); // OR 64 bits into G1
2706 }
2707 #endif
2709 __ ret();
2710 __ delayed()->restore();
2712 __ flush();
2714 nmethod *nm = nmethod::new_native_nmethod(method,
2715 compile_id,
2716 masm->code(),
2717 vep_offset,
2718 frame_complete,
2719 stack_slots / VMRegImpl::slots_per_word,
2720 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2721 in_ByteSize(lock_offset),
2722 oop_maps);
2724 if (is_critical_native) {
2725 nm->set_lazy_critical_native(true);
2726 }
2727 return nm;
2729 }
2731 #ifdef HAVE_DTRACE_H
2732 // ---------------------------------------------------------------------------
2733 // Generate a dtrace nmethod for a given signature. The method takes arguments
2734 // in the Java compiled code convention, marshals them to the native
2735 // abi and then leaves nops at the position you would expect to call a native
2736 // function. When the probe is enabled the nops are replaced with a trap
2737 // instruction that dtrace inserts and the trace will cause a notification
2738 // to dtrace.
2739 //
2740 // The probes are only able to take primitive types and java/lang/String as
2741 // arguments. No other java types are allowed. Strings are converted to utf8
2742 // strings so that from dtrace point of view java strings are converted to C
2743 // strings. There is an arbitrary fixed limit on the total space that a method
2744 // can use for converting the strings. (256 chars per string in the signature).
2745 // So any java string larger then this is truncated.
2747 static int fp_offset[ConcreteRegisterImpl::number_of_registers] = { 0 };
2748 static bool offsets_initialized = false;
2750 nmethod *SharedRuntime::generate_dtrace_nmethod(
2751 MacroAssembler *masm, methodHandle method) {
2754 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
2755 // be single threaded in this method.
2756 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
2758 // Fill in the signature array, for the calling-convention call.
2759 int total_args_passed = method->size_of_parameters();
2761 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
2762 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
2764 // The signature we are going to use for the trap that dtrace will see
2765 // java/lang/String is converted. We drop "this" and any other object
2766 // is converted to NULL. (A one-slot java/lang/Long object reference
2767 // is converted to a two-slot long, which is why we double the allocation).
2768 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
2769 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
2771 int i=0;
2772 int total_strings = 0;
2773 int first_arg_to_pass = 0;
2774 int total_c_args = 0;
2776 // Skip the receiver as dtrace doesn't want to see it
2777 if( !method->is_static() ) {
2778 in_sig_bt[i++] = T_OBJECT;
2779 first_arg_to_pass = 1;
2780 }
2782 SignatureStream ss(method->signature());
2783 for ( ; !ss.at_return_type(); ss.next()) {
2784 BasicType bt = ss.type();
2785 in_sig_bt[i++] = bt; // Collect remaining bits of signature
2786 out_sig_bt[total_c_args++] = bt;
2787 if( bt == T_OBJECT) {
2788 Symbol* s = ss.as_symbol_or_null();
2789 if (s == vmSymbols::java_lang_String()) {
2790 total_strings++;
2791 out_sig_bt[total_c_args-1] = T_ADDRESS;
2792 } else if (s == vmSymbols::java_lang_Boolean() ||
2793 s == vmSymbols::java_lang_Byte()) {
2794 out_sig_bt[total_c_args-1] = T_BYTE;
2795 } else if (s == vmSymbols::java_lang_Character() ||
2796 s == vmSymbols::java_lang_Short()) {
2797 out_sig_bt[total_c_args-1] = T_SHORT;
2798 } else if (s == vmSymbols::java_lang_Integer() ||
2799 s == vmSymbols::java_lang_Float()) {
2800 out_sig_bt[total_c_args-1] = T_INT;
2801 } else if (s == vmSymbols::java_lang_Long() ||
2802 s == vmSymbols::java_lang_Double()) {
2803 out_sig_bt[total_c_args-1] = T_LONG;
2804 out_sig_bt[total_c_args++] = T_VOID;
2805 }
2806 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
2807 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
2808 // We convert double to long
2809 out_sig_bt[total_c_args-1] = T_LONG;
2810 out_sig_bt[total_c_args++] = T_VOID;
2811 } else if ( bt == T_FLOAT) {
2812 // We convert float to int
2813 out_sig_bt[total_c_args-1] = T_INT;
2814 }
2815 }
2817 assert(i==total_args_passed, "validly parsed signature");
2819 // Now get the compiled-Java layout as input arguments
2820 int comp_args_on_stack;
2821 comp_args_on_stack = SharedRuntime::java_calling_convention(
2822 in_sig_bt, in_regs, total_args_passed, false);
2824 // We have received a description of where all the java arg are located
2825 // on entry to the wrapper. We need to convert these args to where
2826 // the a native (non-jni) function would expect them. To figure out
2827 // where they go we convert the java signature to a C signature and remove
2828 // T_VOID for any long/double we might have received.
2831 // Now figure out where the args must be stored and how much stack space
2832 // they require (neglecting out_preserve_stack_slots but space for storing
2833 // the 1st six register arguments). It's weird see int_stk_helper.
2834 //
2835 int out_arg_slots;
2836 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args);
2838 // Calculate the total number of stack slots we will need.
2840 // First count the abi requirement plus all of the outgoing args
2841 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2843 // Plus a temp for possible converion of float/double/long register args
2845 int conversion_temp = stack_slots;
2846 stack_slots += 2;
2849 // Now space for the string(s) we must convert
2851 int string_locs = stack_slots;
2852 stack_slots += total_strings *
2853 (max_dtrace_string_size / VMRegImpl::stack_slot_size);
2855 // Ok The space we have allocated will look like:
2856 //
2857 //
2858 // FP-> | |
2859 // |---------------------|
2860 // | string[n] |
2861 // |---------------------| <- string_locs[n]
2862 // | string[n-1] |
2863 // |---------------------| <- string_locs[n-1]
2864 // | ... |
2865 // | ... |
2866 // |---------------------| <- string_locs[1]
2867 // | string[0] |
2868 // |---------------------| <- string_locs[0]
2869 // | temp |
2870 // |---------------------| <- conversion_temp
2871 // | outbound memory |
2872 // | based arguments |
2873 // | |
2874 // |---------------------|
2875 // | |
2876 // SP-> | out_preserved_slots |
2877 //
2878 //
2880 // Now compute actual number of stack words we need rounding to make
2881 // stack properly aligned.
2882 stack_slots = round_to(stack_slots, 4 * VMRegImpl::slots_per_word);
2884 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2886 intptr_t start = (intptr_t)__ pc();
2888 // First thing make an ic check to see if we should even be here
2890 {
2891 Label L;
2892 const Register temp_reg = G3_scratch;
2893 AddressLiteral ic_miss(SharedRuntime::get_ic_miss_stub());
2894 __ verify_oop(O0);
2895 __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), temp_reg);
2896 __ cmp_and_brx_short(temp_reg, G5_inline_cache_reg, Assembler::equal, Assembler::pt, L);
2898 __ jump_to(ic_miss, temp_reg);
2899 __ delayed()->nop();
2900 __ align(CodeEntryAlignment);
2901 __ bind(L);
2902 }
2904 int vep_offset = ((intptr_t)__ pc()) - start;
2907 // The instruction at the verified entry point must be 5 bytes or longer
2908 // because it can be patched on the fly by make_non_entrant. The stack bang
2909 // instruction fits that requirement.
2911 // Generate stack overflow check before creating frame
2912 __ generate_stack_overflow_check(stack_size);
2914 assert(((intptr_t)__ pc() - start - vep_offset) >= 5,
2915 "valid size for make_non_entrant");
2917 // Generate a new frame for the wrapper.
2918 __ save(SP, -stack_size, SP);
2920 // Frame is now completed as far a size and linkage.
2922 int frame_complete = ((intptr_t)__ pc()) - start;
2924 #ifdef ASSERT
2925 bool reg_destroyed[RegisterImpl::number_of_registers];
2926 bool freg_destroyed[FloatRegisterImpl::number_of_registers];
2927 for ( int r = 0 ; r < RegisterImpl::number_of_registers ; r++ ) {
2928 reg_destroyed[r] = false;
2929 }
2930 for ( int f = 0 ; f < FloatRegisterImpl::number_of_registers ; f++ ) {
2931 freg_destroyed[f] = false;
2932 }
2934 #endif /* ASSERT */
2936 VMRegPair zero;
2937 const Register g0 = G0; // without this we get a compiler warning (why??)
2938 zero.set2(g0->as_VMReg());
2940 int c_arg, j_arg;
2942 Register conversion_off = noreg;
2944 for (j_arg = first_arg_to_pass, c_arg = 0 ;
2945 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
2947 VMRegPair src = in_regs[j_arg];
2948 VMRegPair dst = out_regs[c_arg];
2950 #ifdef ASSERT
2951 if (src.first()->is_Register()) {
2952 assert(!reg_destroyed[src.first()->as_Register()->encoding()], "ack!");
2953 } else if (src.first()->is_FloatRegister()) {
2954 assert(!freg_destroyed[src.first()->as_FloatRegister()->encoding(
2955 FloatRegisterImpl::S)], "ack!");
2956 }
2957 if (dst.first()->is_Register()) {
2958 reg_destroyed[dst.first()->as_Register()->encoding()] = true;
2959 } else if (dst.first()->is_FloatRegister()) {
2960 freg_destroyed[dst.first()->as_FloatRegister()->encoding(
2961 FloatRegisterImpl::S)] = true;
2962 }
2963 #endif /* ASSERT */
2965 switch (in_sig_bt[j_arg]) {
2966 case T_ARRAY:
2967 case T_OBJECT:
2968 {
2969 if (out_sig_bt[c_arg] == T_BYTE || out_sig_bt[c_arg] == T_SHORT ||
2970 out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
2971 // need to unbox a one-slot value
2972 Register in_reg = L0;
2973 Register tmp = L2;
2974 if ( src.first()->is_reg() ) {
2975 in_reg = src.first()->as_Register();
2976 } else {
2977 assert(Assembler::is_simm13(reg2offset(src.first()) + STACK_BIAS),
2978 "must be");
2979 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, in_reg);
2980 }
2981 // If the final destination is an acceptable register
2982 if ( dst.first()->is_reg() ) {
2983 if ( dst.is_single_phys_reg() || out_sig_bt[c_arg] != T_LONG ) {
2984 tmp = dst.first()->as_Register();
2985 }
2986 }
2988 Label skipUnbox;
2989 if ( wordSize == 4 && out_sig_bt[c_arg] == T_LONG ) {
2990 __ mov(G0, tmp->successor());
2991 }
2992 __ br_null(in_reg, true, Assembler::pn, skipUnbox);
2993 __ delayed()->mov(G0, tmp);
2995 BasicType bt = out_sig_bt[c_arg];
2996 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
2997 switch (bt) {
2998 case T_BYTE:
2999 __ ldub(in_reg, box_offset, tmp); break;
3000 case T_SHORT:
3001 __ lduh(in_reg, box_offset, tmp); break;
3002 case T_INT:
3003 __ ld(in_reg, box_offset, tmp); break;
3004 case T_LONG:
3005 __ ld_long(in_reg, box_offset, tmp); break;
3006 default: ShouldNotReachHere();
3007 }
3009 __ bind(skipUnbox);
3010 // If tmp wasn't final destination copy to final destination
3011 if (tmp == L2) {
3012 VMRegPair tmp_as_VM = reg64_to_VMRegPair(L2);
3013 if (out_sig_bt[c_arg] == T_LONG) {
3014 long_move(masm, tmp_as_VM, dst);
3015 } else {
3016 move32_64(masm, tmp_as_VM, out_regs[c_arg]);
3017 }
3018 }
3019 if (out_sig_bt[c_arg] == T_LONG) {
3020 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
3021 ++c_arg; // move over the T_VOID to keep the loop indices in sync
3022 }
3023 } else if (out_sig_bt[c_arg] == T_ADDRESS) {
3024 Register s =
3025 src.first()->is_reg() ? src.first()->as_Register() : L2;
3026 Register d =
3027 dst.first()->is_reg() ? dst.first()->as_Register() : L2;
3029 // We store the oop now so that the conversion pass can reach
3030 // while in the inner frame. This will be the only store if
3031 // the oop is NULL.
3032 if (s != L2) {
3033 // src is register
3034 if (d != L2) {
3035 // dst is register
3036 __ mov(s, d);
3037 } else {
3038 assert(Assembler::is_simm13(reg2offset(dst.first()) +
3039 STACK_BIAS), "must be");
3040 __ st_ptr(s, SP, reg2offset(dst.first()) + STACK_BIAS);
3041 }
3042 } else {
3043 // src not a register
3044 assert(Assembler::is_simm13(reg2offset(src.first()) +
3045 STACK_BIAS), "must be");
3046 __ ld_ptr(FP, reg2offset(src.first()) + STACK_BIAS, d);
3047 if (d == L2) {
3048 assert(Assembler::is_simm13(reg2offset(dst.first()) +
3049 STACK_BIAS), "must be");
3050 __ st_ptr(d, SP, reg2offset(dst.first()) + STACK_BIAS);
3051 }
3052 }
3053 } else if (out_sig_bt[c_arg] != T_VOID) {
3054 // Convert the arg to NULL
3055 if (dst.first()->is_reg()) {
3056 __ mov(G0, dst.first()->as_Register());
3057 } else {
3058 assert(Assembler::is_simm13(reg2offset(dst.first()) +
3059 STACK_BIAS), "must be");
3060 __ st_ptr(G0, SP, reg2offset(dst.first()) + STACK_BIAS);
3061 }
3062 }
3063 }
3064 break;
3065 case T_VOID:
3066 break;
3068 case T_FLOAT:
3069 if (src.first()->is_stack()) {
3070 // Stack to stack/reg is simple
3071 move32_64(masm, src, dst);
3072 } else {
3073 if (dst.first()->is_reg()) {
3074 // freg -> reg
3075 int off =
3076 STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
3077 Register d = dst.first()->as_Register();
3078 if (Assembler::is_simm13(off)) {
3079 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3080 SP, off);
3081 __ ld(SP, off, d);
3082 } else {
3083 if (conversion_off == noreg) {
3084 __ set(off, L6);
3085 conversion_off = L6;
3086 }
3087 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3088 SP, conversion_off);
3089 __ ld(SP, conversion_off , d);
3090 }
3091 } else {
3092 // freg -> mem
3093 int off = STACK_BIAS + reg2offset(dst.first());
3094 if (Assembler::is_simm13(off)) {
3095 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3096 SP, off);
3097 } else {
3098 if (conversion_off == noreg) {
3099 __ set(off, L6);
3100 conversion_off = L6;
3101 }
3102 __ stf(FloatRegisterImpl::S, src.first()->as_FloatRegister(),
3103 SP, conversion_off);
3104 }
3105 }
3106 }
3107 break;
3109 case T_DOUBLE:
3110 assert( j_arg + 1 < total_args_passed &&
3111 in_sig_bt[j_arg + 1] == T_VOID &&
3112 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
3113 if (src.first()->is_stack()) {
3114 // Stack to stack/reg is simple
3115 long_move(masm, src, dst);
3116 } else {
3117 Register d = dst.first()->is_reg() ? dst.first()->as_Register() : L2;
3119 // Destination could be an odd reg on 32bit in which case
3120 // we can't load direct to the destination.
3122 if (!d->is_even() && wordSize == 4) {
3123 d = L2;
3124 }
3125 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
3126 if (Assembler::is_simm13(off)) {
3127 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
3128 SP, off);
3129 __ ld_long(SP, off, d);
3130 } else {
3131 if (conversion_off == noreg) {
3132 __ set(off, L6);
3133 conversion_off = L6;
3134 }
3135 __ stf(FloatRegisterImpl::D, src.first()->as_FloatRegister(),
3136 SP, conversion_off);
3137 __ ld_long(SP, conversion_off, d);
3138 }
3139 if (d == L2) {
3140 long_move(masm, reg64_to_VMRegPair(L2), dst);
3141 }
3142 }
3143 break;
3145 case T_LONG :
3146 // 32bit can't do a split move of something like g1 -> O0, O1
3147 // so use a memory temp
3148 if (src.is_single_phys_reg() && wordSize == 4) {
3149 Register tmp = L2;
3150 if (dst.first()->is_reg() &&
3151 (wordSize == 8 || dst.first()->as_Register()->is_even())) {
3152 tmp = dst.first()->as_Register();
3153 }
3155 int off = STACK_BIAS + conversion_temp * VMRegImpl::stack_slot_size;
3156 if (Assembler::is_simm13(off)) {
3157 __ stx(src.first()->as_Register(), SP, off);
3158 __ ld_long(SP, off, tmp);
3159 } else {
3160 if (conversion_off == noreg) {
3161 __ set(off, L6);
3162 conversion_off = L6;
3163 }
3164 __ stx(src.first()->as_Register(), SP, conversion_off);
3165 __ ld_long(SP, conversion_off, tmp);
3166 }
3168 if (tmp == L2) {
3169 long_move(masm, reg64_to_VMRegPair(L2), dst);
3170 }
3171 } else {
3172 long_move(masm, src, dst);
3173 }
3174 break;
3176 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
3178 default:
3179 move32_64(masm, src, dst);
3180 }
3181 }
3184 // If we have any strings we must store any register based arg to the stack
3185 // This includes any still live xmm registers too.
3187 if (total_strings > 0 ) {
3189 // protect all the arg registers
3190 __ save_frame(0);
3191 __ mov(G2_thread, L7_thread_cache);
3192 const Register L2_string_off = L2;
3194 // Get first string offset
3195 __ set(string_locs * VMRegImpl::stack_slot_size, L2_string_off);
3197 for (c_arg = 0 ; c_arg < total_c_args ; c_arg++ ) {
3198 if (out_sig_bt[c_arg] == T_ADDRESS) {
3200 VMRegPair dst = out_regs[c_arg];
3201 const Register d = dst.first()->is_reg() ?
3202 dst.first()->as_Register()->after_save() : noreg;
3204 // It's a string the oop and it was already copied to the out arg
3205 // position
3206 if (d != noreg) {
3207 __ mov(d, O0);
3208 } else {
3209 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
3210 "must be");
3211 __ ld_ptr(FP, reg2offset(dst.first()) + STACK_BIAS, O0);
3212 }
3213 Label skip;
3215 __ br_null(O0, false, Assembler::pn, skip);
3216 __ delayed()->add(FP, L2_string_off, O1);
3218 if (d != noreg) {
3219 __ mov(O1, d);
3220 } else {
3221 assert(Assembler::is_simm13(reg2offset(dst.first()) + STACK_BIAS),
3222 "must be");
3223 __ st_ptr(O1, FP, reg2offset(dst.first()) + STACK_BIAS);
3224 }
3226 __ call(CAST_FROM_FN_PTR(address, SharedRuntime::get_utf),
3227 relocInfo::runtime_call_type);
3228 __ delayed()->add(L2_string_off, max_dtrace_string_size, L2_string_off);
3230 __ bind(skip);
3232 }
3234 }
3235 __ mov(L7_thread_cache, G2_thread);
3236 __ restore();
3238 }
3241 // Ok now we are done. Need to place the nop that dtrace wants in order to
3242 // patch in the trap
3244 int patch_offset = ((intptr_t)__ pc()) - start;
3246 __ nop();
3249 // Return
3251 __ ret();
3252 __ delayed()->restore();
3254 __ flush();
3256 nmethod *nm = nmethod::new_dtrace_nmethod(
3257 method, masm->code(), vep_offset, patch_offset, frame_complete,
3258 stack_slots / VMRegImpl::slots_per_word);
3259 return nm;
3261 }
3263 #endif // HAVE_DTRACE_H
3265 // this function returns the adjust size (in number of words) to a c2i adapter
3266 // activation for use during deoptimization
3267 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
3268 assert(callee_locals >= callee_parameters,
3269 "test and remove; got more parms than locals");
3270 if (callee_locals < callee_parameters)
3271 return 0; // No adjustment for negative locals
3272 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
3273 return round_to(diff, WordsPerLong);
3274 }
3276 // "Top of Stack" slots that may be unused by the calling convention but must
3277 // otherwise be preserved.
3278 // On Intel these are not necessary and the value can be zero.
3279 // On Sparc this describes the words reserved for storing a register window
3280 // when an interrupt occurs.
3281 uint SharedRuntime::out_preserve_stack_slots() {
3282 return frame::register_save_words * VMRegImpl::slots_per_word;
3283 }
3285 static void gen_new_frame(MacroAssembler* masm, bool deopt) {
3286 //
3287 // Common out the new frame generation for deopt and uncommon trap
3288 //
3289 Register G3pcs = G3_scratch; // Array of new pcs (input)
3290 Register Oreturn0 = O0;
3291 Register Oreturn1 = O1;
3292 Register O2UnrollBlock = O2;
3293 Register O3array = O3; // Array of frame sizes (input)
3294 Register O4array_size = O4; // number of frames (input)
3295 Register O7frame_size = O7; // number of frames (input)
3297 __ ld_ptr(O3array, 0, O7frame_size);
3298 __ sub(G0, O7frame_size, O7frame_size);
3299 __ save(SP, O7frame_size, SP);
3300 __ ld_ptr(G3pcs, 0, I7); // load frame's new pc
3302 #ifdef ASSERT
3303 // make sure that the frames are aligned properly
3304 #ifndef _LP64
3305 __ btst(wordSize*2-1, SP);
3306 __ breakpoint_trap(Assembler::notZero, Assembler::ptr_cc);
3307 #endif
3308 #endif
3310 // Deopt needs to pass some extra live values from frame to frame
3312 if (deopt) {
3313 __ mov(Oreturn0->after_save(), Oreturn0);
3314 __ mov(Oreturn1->after_save(), Oreturn1);
3315 }
3317 __ mov(O4array_size->after_save(), O4array_size);
3318 __ sub(O4array_size, 1, O4array_size);
3319 __ mov(O3array->after_save(), O3array);
3320 __ mov(O2UnrollBlock->after_save(), O2UnrollBlock);
3321 __ add(G3pcs, wordSize, G3pcs); // point to next pc value
3323 #ifdef ASSERT
3324 // trash registers to show a clear pattern in backtraces
3325 __ set(0xDEAD0000, I0);
3326 __ add(I0, 2, I1);
3327 __ add(I0, 4, I2);
3328 __ add(I0, 6, I3);
3329 __ add(I0, 8, I4);
3330 // Don't touch I5 could have valuable savedSP
3331 __ set(0xDEADBEEF, L0);
3332 __ mov(L0, L1);
3333 __ mov(L0, L2);
3334 __ mov(L0, L3);
3335 __ mov(L0, L4);
3336 __ mov(L0, L5);
3338 // trash the return value as there is nothing to return yet
3339 __ set(0xDEAD0001, O7);
3340 #endif
3342 __ mov(SP, O5_savedSP);
3343 }
3346 static void make_new_frames(MacroAssembler* masm, bool deopt) {
3347 //
3348 // loop through the UnrollBlock info and create new frames
3349 //
3350 Register G3pcs = G3_scratch;
3351 Register Oreturn0 = O0;
3352 Register Oreturn1 = O1;
3353 Register O2UnrollBlock = O2;
3354 Register O3array = O3;
3355 Register O4array_size = O4;
3356 Label loop;
3358 // Before we make new frames, check to see if stack is available.
3359 // Do this after the caller's return address is on top of stack
3360 if (UseStackBanging) {
3361 // Get total frame size for interpreted frames
3362 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes(), O4);
3363 __ bang_stack_size(O4, O3, G3_scratch);
3364 }
3366 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes(), O4array_size);
3367 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes(), G3pcs);
3368 __ ld_ptr(O2UnrollBlock, Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes(), O3array);
3370 // Adjust old interpreter frame to make space for new frame's extra java locals
3371 //
3372 // We capture the original sp for the transition frame only because it is needed in
3373 // order to properly calculate interpreter_sp_adjustment. Even though in real life
3374 // every interpreter frame captures a savedSP it is only needed at the transition
3375 // (fortunately). If we had to have it correct everywhere then we would need to
3376 // be told the sp_adjustment for each frame we create. If the frame size array
3377 // were to have twice the frame count entries then we could have pairs [sp_adjustment, frame_size]
3378 // for each frame we create and keep up the illusion every where.
3379 //
3381 __ ld(O2UnrollBlock, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes(), O7);
3382 __ mov(SP, O5_savedSP); // remember initial sender's original sp before adjustment
3383 __ sub(SP, O7, SP);
3385 #ifdef ASSERT
3386 // make sure that there is at least one entry in the array
3387 __ tst(O4array_size);
3388 __ breakpoint_trap(Assembler::zero, Assembler::icc);
3389 #endif
3391 // Now push the new interpreter frames
3392 __ bind(loop);
3394 // allocate a new frame, filling the registers
3396 gen_new_frame(masm, deopt); // allocate an interpreter frame
3398 __ cmp_zero_and_br(Assembler::notZero, O4array_size, loop);
3399 __ delayed()->add(O3array, wordSize, O3array);
3400 __ ld_ptr(G3pcs, 0, O7); // load final frame new pc
3402 }
3404 //------------------------------generate_deopt_blob----------------------------
3405 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3406 // instead.
3407 void SharedRuntime::generate_deopt_blob() {
3408 // allocate space for the code
3409 ResourceMark rm;
3410 // setup code generation tools
3411 int pad = VerifyThread ? 512 : 0;// Extra slop space for more verify code
3412 if (UseStackBanging) {
3413 pad += StackShadowPages*16 + 32;
3414 }
3415 #ifdef _LP64
3416 CodeBuffer buffer("deopt_blob", 2100+pad, 512);
3417 #else
3418 // Measured 8/7/03 at 1212 in 32bit debug build (no VerifyThread)
3419 // Measured 8/7/03 at 1396 in 32bit debug build (VerifyThread)
3420 CodeBuffer buffer("deopt_blob", 1600+pad, 512);
3421 #endif /* _LP64 */
3422 MacroAssembler* masm = new MacroAssembler(&buffer);
3423 FloatRegister Freturn0 = F0;
3424 Register Greturn1 = G1;
3425 Register Oreturn0 = O0;
3426 Register Oreturn1 = O1;
3427 Register O2UnrollBlock = O2;
3428 Register L0deopt_mode = L0;
3429 Register G4deopt_mode = G4_scratch;
3430 int frame_size_words;
3431 Address saved_Freturn0_addr(FP, -sizeof(double) + STACK_BIAS);
3432 #if !defined(_LP64) && defined(COMPILER2)
3433 Address saved_Greturn1_addr(FP, -sizeof(double) -sizeof(jlong) + STACK_BIAS);
3434 #endif
3435 Label cont;
3437 OopMapSet *oop_maps = new OopMapSet();
3439 //
3440 // This is the entry point for code which is returning to a de-optimized
3441 // frame.
3442 // The steps taken by this frame are as follows:
3443 // - push a dummy "register_save" and save the return values (O0, O1, F0/F1, G1)
3444 // and all potentially live registers (at a pollpoint many registers can be live).
3445 //
3446 // - call the C routine: Deoptimization::fetch_unroll_info (this function
3447 // returns information about the number and size of interpreter frames
3448 // which are equivalent to the frame which is being deoptimized)
3449 // - deallocate the unpack frame, restoring only results values. Other
3450 // volatile registers will now be captured in the vframeArray as needed.
3451 // - deallocate the deoptimization frame
3452 // - in a loop using the information returned in the previous step
3453 // push new interpreter frames (take care to propagate the return
3454 // values through each new frame pushed)
3455 // - create a dummy "unpack_frame" and save the return values (O0, O1, F0)
3456 // - call the C routine: Deoptimization::unpack_frames (this function
3457 // lays out values on the interpreter frame which was just created)
3458 // - deallocate the dummy unpack_frame
3459 // - ensure that all the return values are correctly set and then do
3460 // a return to the interpreter entry point
3461 //
3462 // Refer to the following methods for more information:
3463 // - Deoptimization::fetch_unroll_info
3464 // - Deoptimization::unpack_frames
3466 OopMap* map = NULL;
3468 int start = __ offset();
3470 // restore G2, the trampoline destroyed it
3471 __ get_thread();
3473 // On entry we have been called by the deoptimized nmethod with a call that
3474 // replaced the original call (or safepoint polling location) so the deoptimizing
3475 // pc is now in O7. Return values are still in the expected places
3477 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3478 __ ba(cont);
3479 __ delayed()->mov(Deoptimization::Unpack_deopt, L0deopt_mode);
3481 int exception_offset = __ offset() - start;
3483 // restore G2, the trampoline destroyed it
3484 __ get_thread();
3486 // On entry we have been jumped to by the exception handler (or exception_blob
3487 // for server). O0 contains the exception oop and O7 contains the original
3488 // exception pc. So if we push a frame here it will look to the
3489 // stack walking code (fetch_unroll_info) just like a normal call so
3490 // state will be extracted normally.
3492 // save exception oop in JavaThread and fall through into the
3493 // exception_in_tls case since they are handled in same way except
3494 // for where the pending exception is kept.
3495 __ st_ptr(Oexception, G2_thread, JavaThread::exception_oop_offset());
3497 //
3498 // Vanilla deoptimization with an exception pending in exception_oop
3499 //
3500 int exception_in_tls_offset = __ offset() - start;
3502 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3503 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3505 // Restore G2_thread
3506 __ get_thread();
3508 #ifdef ASSERT
3509 {
3510 // verify that there is really an exception oop in exception_oop
3511 Label has_exception;
3512 __ ld_ptr(G2_thread, JavaThread::exception_oop_offset(), Oexception);
3513 __ br_notnull_short(Oexception, Assembler::pt, has_exception);
3514 __ stop("no exception in thread");
3515 __ bind(has_exception);
3517 // verify that there is no pending exception
3518 Label no_pending_exception;
3519 Address exception_addr(G2_thread, Thread::pending_exception_offset());
3520 __ ld_ptr(exception_addr, Oexception);
3521 __ br_null_short(Oexception, Assembler::pt, no_pending_exception);
3522 __ stop("must not have pending exception here");
3523 __ bind(no_pending_exception);
3524 }
3525 #endif
3527 __ ba(cont);
3528 __ delayed()->mov(Deoptimization::Unpack_exception, L0deopt_mode);;
3530 //
3531 // Reexecute entry, similar to c2 uncommon trap
3532 //
3533 int reexecute_offset = __ offset() - start;
3535 // No need to update oop_map as each call to save_live_registers will produce identical oopmap
3536 (void) RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3538 __ mov(Deoptimization::Unpack_reexecute, L0deopt_mode);
3540 __ bind(cont);
3542 __ set_last_Java_frame(SP, noreg);
3544 // do the call by hand so we can get the oopmap
3546 __ mov(G2_thread, L7_thread_cache);
3547 __ call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), relocInfo::runtime_call_type);
3548 __ delayed()->mov(G2_thread, O0);
3550 // Set an oopmap for the call site this describes all our saved volatile registers
3552 oop_maps->add_gc_map( __ offset()-start, map);
3554 __ mov(L7_thread_cache, G2_thread);
3556 __ reset_last_Java_frame();
3558 // NOTE: we know that only O0/O1 will be reloaded by restore_result_registers
3559 // so this move will survive
3561 __ mov(L0deopt_mode, G4deopt_mode);
3563 __ mov(O0, O2UnrollBlock->after_save());
3565 RegisterSaver::restore_result_registers(masm);
3567 Label noException;
3568 __ cmp_and_br_short(G4deopt_mode, Deoptimization::Unpack_exception, Assembler::notEqual, Assembler::pt, noException);
3570 // Move the pending exception from exception_oop to Oexception so
3571 // the pending exception will be picked up the interpreter.
3572 __ ld_ptr(G2_thread, in_bytes(JavaThread::exception_oop_offset()), Oexception);
3573 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_oop_offset()));
3574 __ st_ptr(G0, G2_thread, in_bytes(JavaThread::exception_pc_offset()));
3575 __ bind(noException);
3577 // deallocate the deoptimization frame taking care to preserve the return values
3578 __ mov(Oreturn0, Oreturn0->after_save());
3579 __ mov(Oreturn1, Oreturn1->after_save());
3580 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3581 __ restore();
3583 // Allocate new interpreter frame(s) and possible c2i adapter frame
3585 make_new_frames(masm, true);
3587 // push a dummy "unpack_frame" taking care of float return values and
3588 // call Deoptimization::unpack_frames to have the unpacker layout
3589 // information in the interpreter frames just created and then return
3590 // to the interpreter entry point
3591 __ save(SP, -frame_size_words*wordSize, SP);
3592 __ stf(FloatRegisterImpl::D, Freturn0, saved_Freturn0_addr);
3593 #if !defined(_LP64)
3594 #if defined(COMPILER2)
3595 // 32-bit 1-register longs return longs in G1
3596 __ stx(Greturn1, saved_Greturn1_addr);
3597 #endif
3598 __ set_last_Java_frame(SP, noreg);
3599 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, G4deopt_mode);
3600 #else
3601 // LP64 uses g4 in set_last_Java_frame
3602 __ mov(G4deopt_mode, O1);
3603 __ set_last_Java_frame(SP, G0);
3604 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O1);
3605 #endif
3606 __ reset_last_Java_frame();
3607 __ ldf(FloatRegisterImpl::D, saved_Freturn0_addr, Freturn0);
3609 #if !defined(_LP64) && defined(COMPILER2)
3610 // In 32 bit, C2 returns longs in G1 so restore the saved G1 into
3611 // I0/I1 if the return value is long.
3612 Label not_long;
3613 __ cmp_and_br_short(O0,T_LONG, Assembler::notEqual, Assembler::pt, not_long);
3614 __ ldd(saved_Greturn1_addr,I0);
3615 __ bind(not_long);
3616 #endif
3617 __ ret();
3618 __ delayed()->restore();
3620 masm->flush();
3621 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_words);
3622 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3623 }
3625 #ifdef COMPILER2
3627 //------------------------------generate_uncommon_trap_blob--------------------
3628 // Ought to generate an ideal graph & compile, but here's some SPARC ASM
3629 // instead.
3630 void SharedRuntime::generate_uncommon_trap_blob() {
3631 // allocate space for the code
3632 ResourceMark rm;
3633 // setup code generation tools
3634 int pad = VerifyThread ? 512 : 0;
3635 if (UseStackBanging) {
3636 pad += StackShadowPages*16 + 32;
3637 }
3638 #ifdef _LP64
3639 CodeBuffer buffer("uncommon_trap_blob", 2700+pad, 512);
3640 #else
3641 // Measured 8/7/03 at 660 in 32bit debug build (no VerifyThread)
3642 // Measured 8/7/03 at 1028 in 32bit debug build (VerifyThread)
3643 CodeBuffer buffer("uncommon_trap_blob", 2000+pad, 512);
3644 #endif
3645 MacroAssembler* masm = new MacroAssembler(&buffer);
3646 Register O2UnrollBlock = O2;
3647 Register O2klass_index = O2;
3649 //
3650 // This is the entry point for all traps the compiler takes when it thinks
3651 // it cannot handle further execution of compilation code. The frame is
3652 // deoptimized in these cases and converted into interpreter frames for
3653 // execution
3654 // The steps taken by this frame are as follows:
3655 // - push a fake "unpack_frame"
3656 // - call the C routine Deoptimization::uncommon_trap (this function
3657 // packs the current compiled frame into vframe arrays and returns
3658 // information about the number and size of interpreter frames which
3659 // are equivalent to the frame which is being deoptimized)
3660 // - deallocate the "unpack_frame"
3661 // - deallocate the deoptimization frame
3662 // - in a loop using the information returned in the previous step
3663 // push interpreter frames;
3664 // - create a dummy "unpack_frame"
3665 // - call the C routine: Deoptimization::unpack_frames (this function
3666 // lays out values on the interpreter frame which was just created)
3667 // - deallocate the dummy unpack_frame
3668 // - return to the interpreter entry point
3669 //
3670 // Refer to the following methods for more information:
3671 // - Deoptimization::uncommon_trap
3672 // - Deoptimization::unpack_frame
3674 // the unloaded class index is in O0 (first parameter to this blob)
3676 // push a dummy "unpack_frame"
3677 // and call Deoptimization::uncommon_trap to pack the compiled frame into
3678 // vframe array and return the UnrollBlock information
3679 __ save_frame(0);
3680 __ set_last_Java_frame(SP, noreg);
3681 __ mov(I0, O2klass_index);
3682 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap), G2_thread, O2klass_index);
3683 __ reset_last_Java_frame();
3684 __ mov(O0, O2UnrollBlock->after_save());
3685 __ restore();
3687 // deallocate the deoptimized frame taking care to preserve the return values
3688 __ mov(O2UnrollBlock, O2UnrollBlock->after_save());
3689 __ restore();
3691 // Allocate new interpreter frame(s) and possible c2i adapter frame
3693 make_new_frames(masm, false);
3695 // push a dummy "unpack_frame" taking care of float return values and
3696 // call Deoptimization::unpack_frames to have the unpacker layout
3697 // information in the interpreter frames just created and then return
3698 // to the interpreter entry point
3699 __ save_frame(0);
3700 __ set_last_Java_frame(SP, noreg);
3701 __ mov(Deoptimization::Unpack_uncommon_trap, O3); // indicate it is the uncommon trap case
3702 __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames), G2_thread, O3);
3703 __ reset_last_Java_frame();
3704 __ ret();
3705 __ delayed()->restore();
3707 masm->flush();
3708 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, NULL, __ total_frame_size_in_bytes(0)/wordSize);
3709 }
3711 #endif // COMPILER2
3713 //------------------------------generate_handler_blob-------------------
3714 //
3715 // Generate a special Compile2Runtime blob that saves all registers, and sets
3716 // up an OopMap.
3717 //
3718 // This blob is jumped to (via a breakpoint and the signal handler) from a
3719 // safepoint in compiled code. On entry to this blob, O7 contains the
3720 // address in the original nmethod at which we should resume normal execution.
3721 // Thus, this blob looks like a subroutine which must preserve lots of
3722 // registers and return normally. Note that O7 is never register-allocated,
3723 // so it is guaranteed to be free here.
3724 //
3726 // The hardest part of what this blob must do is to save the 64-bit %o
3727 // registers in the 32-bit build. A simple 'save' turn the %o's to %i's and
3728 // an interrupt will chop off their heads. Making space in the caller's frame
3729 // first will let us save the 64-bit %o's before save'ing, but we cannot hand
3730 // the adjusted FP off to the GC stack-crawler: this will modify the caller's
3731 // SP and mess up HIS OopMaps. So we first adjust the caller's SP, then save
3732 // the 64-bit %o's, then do a save, then fixup the caller's SP (our FP).
3733 // Tricky, tricky, tricky...
3735 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3736 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3738 // allocate space for the code
3739 ResourceMark rm;
3740 // setup code generation tools
3741 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3742 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3743 // even larger with TraceJumps
3744 int pad = TraceJumps ? 512 : 0;
3745 CodeBuffer buffer("handler_blob", 1600 + pad, 512);
3746 MacroAssembler* masm = new MacroAssembler(&buffer);
3747 int frame_size_words;
3748 OopMapSet *oop_maps = new OopMapSet();
3749 OopMap* map = NULL;
3751 int start = __ offset();
3753 bool cause_return = (poll_type == POLL_AT_RETURN);
3754 // If this causes a return before the processing, then do a "restore"
3755 if (cause_return) {
3756 __ restore();
3757 } else {
3758 // Make it look like we were called via the poll
3759 // so that frame constructor always sees a valid return address
3760 __ ld_ptr(G2_thread, in_bytes(JavaThread::saved_exception_pc_offset()), O7);
3761 __ sub(O7, frame::pc_return_offset, O7);
3762 }
3764 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3766 // setup last_Java_sp (blows G4)
3767 __ set_last_Java_frame(SP, noreg);
3769 // call into the runtime to handle illegal instructions exception
3770 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3771 __ mov(G2_thread, O0);
3772 __ save_thread(L7_thread_cache);
3773 __ call(call_ptr);
3774 __ delayed()->nop();
3776 // Set an oopmap for the call site.
3777 // We need this not only for callee-saved registers, but also for volatile
3778 // registers that the compiler might be keeping live across a safepoint.
3780 oop_maps->add_gc_map( __ offset() - start, map);
3782 __ restore_thread(L7_thread_cache);
3783 // clear last_Java_sp
3784 __ reset_last_Java_frame();
3786 // Check for exceptions
3787 Label pending;
3789 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3790 __ br_notnull_short(O1, Assembler::pn, pending);
3792 RegisterSaver::restore_live_registers(masm);
3794 // We are back the the original state on entry and ready to go.
3796 __ retl();
3797 __ delayed()->nop();
3799 // Pending exception after the safepoint
3801 __ bind(pending);
3803 RegisterSaver::restore_live_registers(masm);
3805 // We are back the the original state on entry.
3807 // Tail-call forward_exception_entry, with the issuing PC in O7,
3808 // so it looks like the original nmethod called forward_exception_entry.
3809 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3810 __ JMP(O0, 0);
3811 __ delayed()->nop();
3813 // -------------
3814 // make sure all code is generated
3815 masm->flush();
3817 // return exception blob
3818 return SafepointBlob::create(&buffer, oop_maps, frame_size_words);
3819 }
3821 //
3822 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3823 //
3824 // Generate a stub that calls into vm to find out the proper destination
3825 // of a java call. All the argument registers are live at this point
3826 // but since this is generic code we don't know what they are and the caller
3827 // must do any gc of the args.
3828 //
3829 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3830 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3832 // allocate space for the code
3833 ResourceMark rm;
3834 // setup code generation tools
3835 // Measured 8/7/03 at 896 in 32bit debug build (no VerifyThread)
3836 // Measured 8/7/03 at 1080 in 32bit debug build (VerifyThread)
3837 // even larger with TraceJumps
3838 int pad = TraceJumps ? 512 : 0;
3839 CodeBuffer buffer(name, 1600 + pad, 512);
3840 MacroAssembler* masm = new MacroAssembler(&buffer);
3841 int frame_size_words;
3842 OopMapSet *oop_maps = new OopMapSet();
3843 OopMap* map = NULL;
3845 int start = __ offset();
3847 map = RegisterSaver::save_live_registers(masm, 0, &frame_size_words);
3849 int frame_complete = __ offset();
3851 // setup last_Java_sp (blows G4)
3852 __ set_last_Java_frame(SP, noreg);
3854 // call into the runtime to handle illegal instructions exception
3855 // Do not use call_VM_leaf, because we need to make a GC map at this call site.
3856 __ mov(G2_thread, O0);
3857 __ save_thread(L7_thread_cache);
3858 __ call(destination, relocInfo::runtime_call_type);
3859 __ delayed()->nop();
3861 // O0 contains the address we are going to jump to assuming no exception got installed
3863 // Set an oopmap for the call site.
3864 // We need this not only for callee-saved registers, but also for volatile
3865 // registers that the compiler might be keeping live across a safepoint.
3867 oop_maps->add_gc_map( __ offset() - start, map);
3869 __ restore_thread(L7_thread_cache);
3870 // clear last_Java_sp
3871 __ reset_last_Java_frame();
3873 // Check for exceptions
3874 Label pending;
3876 __ ld_ptr(G2_thread, in_bytes(Thread::pending_exception_offset()), O1);
3877 __ br_notnull_short(O1, Assembler::pn, pending);
3879 // get the returned Method*
3881 __ get_vm_result_2(G5_method);
3882 __ stx(G5_method, SP, RegisterSaver::G5_offset()+STACK_BIAS);
3884 // O0 is where we want to jump, overwrite G3 which is saved and scratch
3886 __ stx(O0, SP, RegisterSaver::G3_offset()+STACK_BIAS);
3888 RegisterSaver::restore_live_registers(masm);
3890 // We are back the the original state on entry and ready to go.
3892 __ JMP(G3, 0);
3893 __ delayed()->nop();
3895 // Pending exception after the safepoint
3897 __ bind(pending);
3899 RegisterSaver::restore_live_registers(masm);
3901 // We are back the the original state on entry.
3903 // Tail-call forward_exception_entry, with the issuing PC in O7,
3904 // so it looks like the original nmethod called forward_exception_entry.
3905 __ set((intptr_t)StubRoutines::forward_exception_entry(), O0);
3906 __ JMP(O0, 0);
3907 __ delayed()->nop();
3909 // -------------
3910 // make sure all code is generated
3911 masm->flush();
3913 // return the blob
3914 // frame_size_words or bytes??
3915 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3916 }