Thu, 15 Aug 2013 20:04:10 -0400
8003424: Enable Class Data Sharing for CompressedOops
8016729: ObjectAlignmentInBytes=16 now forces the use of heap based compressed oops
8005933: The -Xshare:auto option is ignored for -server
Summary: Move klass metaspace above the heap and support CDS with compressed klass ptrs.
Reviewed-by: coleenp, kvn, mgerdin, tschatzl, stefank
1 /*
2 * Copyright (c) 2003, 2012, 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.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/icBuffer.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "oops/compiledICHolder.hpp"
33 #include "prims/jvmtiRedefineClassesTrace.hpp"
34 #include "runtime/sharedRuntime.hpp"
35 #include "runtime/vframeArray.hpp"
36 #include "vmreg_x86.inline.hpp"
37 #ifdef COMPILER1
38 #include "c1/c1_Runtime1.hpp"
39 #endif
40 #ifdef COMPILER2
41 #include "opto/runtime.hpp"
42 #endif
44 #define __ masm->
46 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
48 class RegisterSaver {
49 // Capture info about frame layout
50 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off
51 enum layout {
52 fpu_state_off = 0,
53 fpu_state_end = fpu_state_off+FPUStateSizeInWords,
54 st0_off, st0H_off,
55 st1_off, st1H_off,
56 st2_off, st2H_off,
57 st3_off, st3H_off,
58 st4_off, st4H_off,
59 st5_off, st5H_off,
60 st6_off, st6H_off,
61 st7_off, st7H_off,
62 xmm_off,
63 DEF_XMM_OFFS(0),
64 DEF_XMM_OFFS(1),
65 DEF_XMM_OFFS(2),
66 DEF_XMM_OFFS(3),
67 DEF_XMM_OFFS(4),
68 DEF_XMM_OFFS(5),
69 DEF_XMM_OFFS(6),
70 DEF_XMM_OFFS(7),
71 flags_off = xmm7_off + 16/BytesPerInt + 1, // 16-byte stack alignment fill word
72 rdi_off,
73 rsi_off,
74 ignore_off, // extra copy of rbp,
75 rsp_off,
76 rbx_off,
77 rdx_off,
78 rcx_off,
79 rax_off,
80 // The frame sender code expects that rbp will be in the "natural" place and
81 // will override any oopMap setting for it. We must therefore force the layout
82 // so that it agrees with the frame sender code.
83 rbp_off,
84 return_off, // slot for return address
85 reg_save_size };
86 enum { FPU_regs_live = flags_off - fpu_state_end };
88 public:
90 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words,
91 int* total_frame_words, bool verify_fpu = true, bool save_vectors = false);
92 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false);
94 static int rax_offset() { return rax_off; }
95 static int rbx_offset() { return rbx_off; }
97 // Offsets into the register save area
98 // Used by deoptimization when it is managing result register
99 // values on its own
101 static int raxOffset(void) { return rax_off; }
102 static int rdxOffset(void) { return rdx_off; }
103 static int rbxOffset(void) { return rbx_off; }
104 static int xmm0Offset(void) { return xmm0_off; }
105 // This really returns a slot in the fp save area, which one is not important
106 static int fpResultOffset(void) { return st0_off; }
108 // During deoptimization only the result register need to be restored
109 // all the other values have already been extracted.
111 static void restore_result_registers(MacroAssembler* masm);
113 };
115 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words,
116 int* total_frame_words, bool verify_fpu, bool save_vectors) {
117 int vect_words = 0;
118 #ifdef COMPILER2
119 if (save_vectors) {
120 assert(UseAVX > 0, "256bit vectors are supported only with AVX");
121 assert(MaxVectorSize == 32, "only 256bit vectors are supported now");
122 // Save upper half of YMM registes
123 vect_words = 8 * 16 / wordSize;
124 additional_frame_words += vect_words;
125 }
126 #else
127 assert(!save_vectors, "vectors are generated only by C2");
128 #endif
129 int frame_size_in_bytes = (reg_save_size + additional_frame_words) * wordSize;
130 int frame_words = frame_size_in_bytes / wordSize;
131 *total_frame_words = frame_words;
133 assert(FPUStateSizeInWords == 27, "update stack layout");
135 // save registers, fpu state, and flags
136 // We assume caller has already has return address slot on the stack
137 // We push epb twice in this sequence because we want the real rbp,
138 // to be under the return like a normal enter and we want to use pusha
139 // We push by hand instead of pusing push
140 __ enter();
141 __ pusha();
142 __ pushf();
143 __ subptr(rsp,FPU_regs_live*wordSize); // Push FPU registers space
144 __ push_FPU_state(); // Save FPU state & init
146 if (verify_fpu) {
147 // Some stubs may have non standard FPU control word settings so
148 // only check and reset the value when it required to be the
149 // standard value. The safepoint blob in particular can be used
150 // in methods which are using the 24 bit control word for
151 // optimized float math.
153 #ifdef ASSERT
154 // Make sure the control word has the expected value
155 Label ok;
156 __ cmpw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std());
157 __ jccb(Assembler::equal, ok);
158 __ stop("corrupted control word detected");
159 __ bind(ok);
160 #endif
162 // Reset the control word to guard against exceptions being unmasked
163 // since fstp_d can cause FPU stack underflow exceptions. Write it
164 // into the on stack copy and then reload that to make sure that the
165 // current and future values are correct.
166 __ movw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std());
167 }
169 __ frstor(Address(rsp, 0));
170 if (!verify_fpu) {
171 // Set the control word so that exceptions are masked for the
172 // following code.
173 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
174 }
176 // Save the FPU registers in de-opt-able form
178 __ fstp_d(Address(rsp, st0_off*wordSize)); // st(0)
179 __ fstp_d(Address(rsp, st1_off*wordSize)); // st(1)
180 __ fstp_d(Address(rsp, st2_off*wordSize)); // st(2)
181 __ fstp_d(Address(rsp, st3_off*wordSize)); // st(3)
182 __ fstp_d(Address(rsp, st4_off*wordSize)); // st(4)
183 __ fstp_d(Address(rsp, st5_off*wordSize)); // st(5)
184 __ fstp_d(Address(rsp, st6_off*wordSize)); // st(6)
185 __ fstp_d(Address(rsp, st7_off*wordSize)); // st(7)
187 if( UseSSE == 1 ) { // Save the XMM state
188 __ movflt(Address(rsp,xmm0_off*wordSize),xmm0);
189 __ movflt(Address(rsp,xmm1_off*wordSize),xmm1);
190 __ movflt(Address(rsp,xmm2_off*wordSize),xmm2);
191 __ movflt(Address(rsp,xmm3_off*wordSize),xmm3);
192 __ movflt(Address(rsp,xmm4_off*wordSize),xmm4);
193 __ movflt(Address(rsp,xmm5_off*wordSize),xmm5);
194 __ movflt(Address(rsp,xmm6_off*wordSize),xmm6);
195 __ movflt(Address(rsp,xmm7_off*wordSize),xmm7);
196 } else if( UseSSE >= 2 ) {
197 // Save whole 128bit (16 bytes) XMM regiters
198 __ movdqu(Address(rsp,xmm0_off*wordSize),xmm0);
199 __ movdqu(Address(rsp,xmm1_off*wordSize),xmm1);
200 __ movdqu(Address(rsp,xmm2_off*wordSize),xmm2);
201 __ movdqu(Address(rsp,xmm3_off*wordSize),xmm3);
202 __ movdqu(Address(rsp,xmm4_off*wordSize),xmm4);
203 __ movdqu(Address(rsp,xmm5_off*wordSize),xmm5);
204 __ movdqu(Address(rsp,xmm6_off*wordSize),xmm6);
205 __ movdqu(Address(rsp,xmm7_off*wordSize),xmm7);
206 }
208 if (vect_words > 0) {
209 assert(vect_words*wordSize == 128, "");
210 __ subptr(rsp, 128); // Save upper half of YMM registes
211 __ vextractf128h(Address(rsp, 0),xmm0);
212 __ vextractf128h(Address(rsp, 16),xmm1);
213 __ vextractf128h(Address(rsp, 32),xmm2);
214 __ vextractf128h(Address(rsp, 48),xmm3);
215 __ vextractf128h(Address(rsp, 64),xmm4);
216 __ vextractf128h(Address(rsp, 80),xmm5);
217 __ vextractf128h(Address(rsp, 96),xmm6);
218 __ vextractf128h(Address(rsp,112),xmm7);
219 }
221 // Set an oopmap for the call site. This oopmap will map all
222 // oop-registers and debug-info registers as callee-saved. This
223 // will allow deoptimization at this safepoint to find all possible
224 // debug-info recordings, as well as let GC find all oops.
226 OopMapSet *oop_maps = new OopMapSet();
227 OopMap* map = new OopMap( frame_words, 0 );
229 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x) + additional_frame_words)
231 map->set_callee_saved(STACK_OFFSET( rax_off), rax->as_VMReg());
232 map->set_callee_saved(STACK_OFFSET( rcx_off), rcx->as_VMReg());
233 map->set_callee_saved(STACK_OFFSET( rdx_off), rdx->as_VMReg());
234 map->set_callee_saved(STACK_OFFSET( rbx_off), rbx->as_VMReg());
235 // rbp, location is known implicitly, no oopMap
236 map->set_callee_saved(STACK_OFFSET( rsi_off), rsi->as_VMReg());
237 map->set_callee_saved(STACK_OFFSET( rdi_off), rdi->as_VMReg());
238 map->set_callee_saved(STACK_OFFSET(st0_off), as_FloatRegister(0)->as_VMReg());
239 map->set_callee_saved(STACK_OFFSET(st1_off), as_FloatRegister(1)->as_VMReg());
240 map->set_callee_saved(STACK_OFFSET(st2_off), as_FloatRegister(2)->as_VMReg());
241 map->set_callee_saved(STACK_OFFSET(st3_off), as_FloatRegister(3)->as_VMReg());
242 map->set_callee_saved(STACK_OFFSET(st4_off), as_FloatRegister(4)->as_VMReg());
243 map->set_callee_saved(STACK_OFFSET(st5_off), as_FloatRegister(5)->as_VMReg());
244 map->set_callee_saved(STACK_OFFSET(st6_off), as_FloatRegister(6)->as_VMReg());
245 map->set_callee_saved(STACK_OFFSET(st7_off), as_FloatRegister(7)->as_VMReg());
246 map->set_callee_saved(STACK_OFFSET(xmm0_off), xmm0->as_VMReg());
247 map->set_callee_saved(STACK_OFFSET(xmm1_off), xmm1->as_VMReg());
248 map->set_callee_saved(STACK_OFFSET(xmm2_off), xmm2->as_VMReg());
249 map->set_callee_saved(STACK_OFFSET(xmm3_off), xmm3->as_VMReg());
250 map->set_callee_saved(STACK_OFFSET(xmm4_off), xmm4->as_VMReg());
251 map->set_callee_saved(STACK_OFFSET(xmm5_off), xmm5->as_VMReg());
252 map->set_callee_saved(STACK_OFFSET(xmm6_off), xmm6->as_VMReg());
253 map->set_callee_saved(STACK_OFFSET(xmm7_off), xmm7->as_VMReg());
254 // %%% This is really a waste but we'll keep things as they were for now
255 if (true) {
256 #define NEXTREG(x) (x)->as_VMReg()->next()
257 map->set_callee_saved(STACK_OFFSET(st0H_off), NEXTREG(as_FloatRegister(0)));
258 map->set_callee_saved(STACK_OFFSET(st1H_off), NEXTREG(as_FloatRegister(1)));
259 map->set_callee_saved(STACK_OFFSET(st2H_off), NEXTREG(as_FloatRegister(2)));
260 map->set_callee_saved(STACK_OFFSET(st3H_off), NEXTREG(as_FloatRegister(3)));
261 map->set_callee_saved(STACK_OFFSET(st4H_off), NEXTREG(as_FloatRegister(4)));
262 map->set_callee_saved(STACK_OFFSET(st5H_off), NEXTREG(as_FloatRegister(5)));
263 map->set_callee_saved(STACK_OFFSET(st6H_off), NEXTREG(as_FloatRegister(6)));
264 map->set_callee_saved(STACK_OFFSET(st7H_off), NEXTREG(as_FloatRegister(7)));
265 map->set_callee_saved(STACK_OFFSET(xmm0H_off), NEXTREG(xmm0));
266 map->set_callee_saved(STACK_OFFSET(xmm1H_off), NEXTREG(xmm1));
267 map->set_callee_saved(STACK_OFFSET(xmm2H_off), NEXTREG(xmm2));
268 map->set_callee_saved(STACK_OFFSET(xmm3H_off), NEXTREG(xmm3));
269 map->set_callee_saved(STACK_OFFSET(xmm4H_off), NEXTREG(xmm4));
270 map->set_callee_saved(STACK_OFFSET(xmm5H_off), NEXTREG(xmm5));
271 map->set_callee_saved(STACK_OFFSET(xmm6H_off), NEXTREG(xmm6));
272 map->set_callee_saved(STACK_OFFSET(xmm7H_off), NEXTREG(xmm7));
273 #undef NEXTREG
274 #undef STACK_OFFSET
275 }
277 return map;
279 }
281 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) {
282 // Recover XMM & FPU state
283 int additional_frame_bytes = 0;
284 #ifdef COMPILER2
285 if (restore_vectors) {
286 assert(UseAVX > 0, "256bit vectors are supported only with AVX");
287 assert(MaxVectorSize == 32, "only 256bit vectors are supported now");
288 additional_frame_bytes = 128;
289 }
290 #else
291 assert(!restore_vectors, "vectors are generated only by C2");
292 #endif
293 if (UseSSE == 1) {
294 assert(additional_frame_bytes == 0, "");
295 __ movflt(xmm0,Address(rsp,xmm0_off*wordSize));
296 __ movflt(xmm1,Address(rsp,xmm1_off*wordSize));
297 __ movflt(xmm2,Address(rsp,xmm2_off*wordSize));
298 __ movflt(xmm3,Address(rsp,xmm3_off*wordSize));
299 __ movflt(xmm4,Address(rsp,xmm4_off*wordSize));
300 __ movflt(xmm5,Address(rsp,xmm5_off*wordSize));
301 __ movflt(xmm6,Address(rsp,xmm6_off*wordSize));
302 __ movflt(xmm7,Address(rsp,xmm7_off*wordSize));
303 } else if (UseSSE >= 2) {
304 #define STACK_ADDRESS(x) Address(rsp,(x)*wordSize + additional_frame_bytes)
305 __ movdqu(xmm0,STACK_ADDRESS(xmm0_off));
306 __ movdqu(xmm1,STACK_ADDRESS(xmm1_off));
307 __ movdqu(xmm2,STACK_ADDRESS(xmm2_off));
308 __ movdqu(xmm3,STACK_ADDRESS(xmm3_off));
309 __ movdqu(xmm4,STACK_ADDRESS(xmm4_off));
310 __ movdqu(xmm5,STACK_ADDRESS(xmm5_off));
311 __ movdqu(xmm6,STACK_ADDRESS(xmm6_off));
312 __ movdqu(xmm7,STACK_ADDRESS(xmm7_off));
313 #undef STACK_ADDRESS
314 }
315 if (restore_vectors) {
316 // Restore upper half of YMM registes.
317 assert(additional_frame_bytes == 128, "");
318 __ vinsertf128h(xmm0, Address(rsp, 0));
319 __ vinsertf128h(xmm1, Address(rsp, 16));
320 __ vinsertf128h(xmm2, Address(rsp, 32));
321 __ vinsertf128h(xmm3, Address(rsp, 48));
322 __ vinsertf128h(xmm4, Address(rsp, 64));
323 __ vinsertf128h(xmm5, Address(rsp, 80));
324 __ vinsertf128h(xmm6, Address(rsp, 96));
325 __ vinsertf128h(xmm7, Address(rsp,112));
326 __ addptr(rsp, additional_frame_bytes);
327 }
328 __ pop_FPU_state();
329 __ addptr(rsp, FPU_regs_live*wordSize); // Pop FPU registers
331 __ popf();
332 __ popa();
333 // Get the rbp, described implicitly by the frame sender code (no oopMap)
334 __ pop(rbp);
336 }
338 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
340 // Just restore result register. Only used by deoptimization. By
341 // now any callee save register that needs to be restore to a c2
342 // caller of the deoptee has been extracted into the vframeArray
343 // and will be stuffed into the c2i adapter we create for later
344 // restoration so only result registers need to be restored here.
345 //
347 __ frstor(Address(rsp, 0)); // Restore fpu state
349 // Recover XMM & FPU state
350 if( UseSSE == 1 ) {
351 __ movflt(xmm0, Address(rsp, xmm0_off*wordSize));
352 } else if( UseSSE >= 2 ) {
353 __ movdbl(xmm0, Address(rsp, xmm0_off*wordSize));
354 }
355 __ movptr(rax, Address(rsp, rax_off*wordSize));
356 __ movptr(rdx, Address(rsp, rdx_off*wordSize));
357 // Pop all of the register save are off the stack except the return address
358 __ addptr(rsp, return_off * wordSize);
359 }
361 // Is vector's size (in bytes) bigger than a size saved by default?
362 // 16 bytes XMM registers are saved by default using SSE2 movdqu instructions.
363 // Note, MaxVectorSize == 0 with UseSSE < 2 and vectors are not generated.
364 bool SharedRuntime::is_wide_vector(int size) {
365 return size > 16;
366 }
368 // The java_calling_convention describes stack locations as ideal slots on
369 // a frame with no abi restrictions. Since we must observe abi restrictions
370 // (like the placement of the register window) the slots must be biased by
371 // the following value.
372 static int reg2offset_in(VMReg r) {
373 // Account for saved rbp, and return address
374 // This should really be in_preserve_stack_slots
375 return (r->reg2stack() + 2) * VMRegImpl::stack_slot_size;
376 }
378 static int reg2offset_out(VMReg r) {
379 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
380 }
382 // ---------------------------------------------------------------------------
383 // Read the array of BasicTypes from a signature, and compute where the
384 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
385 // quantities. Values less than SharedInfo::stack0 are registers, those above
386 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
387 // as framesizes are fixed.
388 // VMRegImpl::stack0 refers to the first slot 0(sp).
389 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
390 // up to RegisterImpl::number_of_registers) are the 32-bit
391 // integer registers.
393 // Pass first two oop/int args in registers ECX and EDX.
394 // Pass first two float/double args in registers XMM0 and XMM1.
395 // Doubles have precedence, so if you pass a mix of floats and doubles
396 // the doubles will grab the registers before the floats will.
398 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
399 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
400 // units regardless of build. Of course for i486 there is no 64 bit build
403 // ---------------------------------------------------------------------------
404 // The compiled Java calling convention.
405 // Pass first two oop/int args in registers ECX and EDX.
406 // Pass first two float/double args in registers XMM0 and XMM1.
407 // Doubles have precedence, so if you pass a mix of floats and doubles
408 // the doubles will grab the registers before the floats will.
409 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
410 VMRegPair *regs,
411 int total_args_passed,
412 int is_outgoing) {
413 uint stack = 0; // Starting stack position for args on stack
416 // Pass first two oop/int args in registers ECX and EDX.
417 uint reg_arg0 = 9999;
418 uint reg_arg1 = 9999;
420 // Pass first two float/double args in registers XMM0 and XMM1.
421 // Doubles have precedence, so if you pass a mix of floats and doubles
422 // the doubles will grab the registers before the floats will.
423 // CNC - TURNED OFF FOR non-SSE.
424 // On Intel we have to round all doubles (and most floats) at
425 // call sites by storing to the stack in any case.
426 // UseSSE=0 ==> Don't Use ==> 9999+0
427 // UseSSE=1 ==> Floats only ==> 9999+1
428 // UseSSE>=2 ==> Floats or doubles ==> 9999+2
429 enum { fltarg_dontuse = 9999+0, fltarg_float_only = 9999+1, fltarg_flt_dbl = 9999+2 };
430 uint fargs = (UseSSE>=2) ? 2 : UseSSE;
431 uint freg_arg0 = 9999+fargs;
432 uint freg_arg1 = 9999+fargs;
434 // Pass doubles & longs aligned on the stack. First count stack slots for doubles
435 int i;
436 for( i = 0; i < total_args_passed; i++) {
437 if( sig_bt[i] == T_DOUBLE ) {
438 // first 2 doubles go in registers
439 if( freg_arg0 == fltarg_flt_dbl ) freg_arg0 = i;
440 else if( freg_arg1 == fltarg_flt_dbl ) freg_arg1 = i;
441 else // Else double is passed low on the stack to be aligned.
442 stack += 2;
443 } else if( sig_bt[i] == T_LONG ) {
444 stack += 2;
445 }
446 }
447 int dstack = 0; // Separate counter for placing doubles
449 // Now pick where all else goes.
450 for( i = 0; i < total_args_passed; i++) {
451 // From the type and the argument number (count) compute the location
452 switch( sig_bt[i] ) {
453 case T_SHORT:
454 case T_CHAR:
455 case T_BYTE:
456 case T_BOOLEAN:
457 case T_INT:
458 case T_ARRAY:
459 case T_OBJECT:
460 case T_ADDRESS:
461 if( reg_arg0 == 9999 ) {
462 reg_arg0 = i;
463 regs[i].set1(rcx->as_VMReg());
464 } else if( reg_arg1 == 9999 ) {
465 reg_arg1 = i;
466 regs[i].set1(rdx->as_VMReg());
467 } else {
468 regs[i].set1(VMRegImpl::stack2reg(stack++));
469 }
470 break;
471 case T_FLOAT:
472 if( freg_arg0 == fltarg_flt_dbl || freg_arg0 == fltarg_float_only ) {
473 freg_arg0 = i;
474 regs[i].set1(xmm0->as_VMReg());
475 } else if( freg_arg1 == fltarg_flt_dbl || freg_arg1 == fltarg_float_only ) {
476 freg_arg1 = i;
477 regs[i].set1(xmm1->as_VMReg());
478 } else {
479 regs[i].set1(VMRegImpl::stack2reg(stack++));
480 }
481 break;
482 case T_LONG:
483 assert(sig_bt[i+1] == T_VOID, "missing Half" );
484 regs[i].set2(VMRegImpl::stack2reg(dstack));
485 dstack += 2;
486 break;
487 case T_DOUBLE:
488 assert(sig_bt[i+1] == T_VOID, "missing Half" );
489 if( freg_arg0 == (uint)i ) {
490 regs[i].set2(xmm0->as_VMReg());
491 } else if( freg_arg1 == (uint)i ) {
492 regs[i].set2(xmm1->as_VMReg());
493 } else {
494 regs[i].set2(VMRegImpl::stack2reg(dstack));
495 dstack += 2;
496 }
497 break;
498 case T_VOID: regs[i].set_bad(); break;
499 break;
500 default:
501 ShouldNotReachHere();
502 break;
503 }
504 }
506 // return value can be odd number of VMRegImpl stack slots make multiple of 2
507 return round_to(stack, 2);
508 }
510 // Patch the callers callsite with entry to compiled code if it exists.
511 static void patch_callers_callsite(MacroAssembler *masm) {
512 Label L;
513 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD);
514 __ jcc(Assembler::equal, L);
515 // Schedule the branch target address early.
516 // Call into the VM to patch the caller, then jump to compiled callee
517 // rax, isn't live so capture return address while we easily can
518 __ movptr(rax, Address(rsp, 0));
519 __ pusha();
520 __ pushf();
522 if (UseSSE == 1) {
523 __ subptr(rsp, 2*wordSize);
524 __ movflt(Address(rsp, 0), xmm0);
525 __ movflt(Address(rsp, wordSize), xmm1);
526 }
527 if (UseSSE >= 2) {
528 __ subptr(rsp, 4*wordSize);
529 __ movdbl(Address(rsp, 0), xmm0);
530 __ movdbl(Address(rsp, 2*wordSize), xmm1);
531 }
532 #ifdef COMPILER2
533 // C2 may leave the stack dirty if not in SSE2+ mode
534 if (UseSSE >= 2) {
535 __ verify_FPU(0, "c2i transition should have clean FPU stack");
536 } else {
537 __ empty_FPU_stack();
538 }
539 #endif /* COMPILER2 */
541 // VM needs caller's callsite
542 __ push(rax);
543 // VM needs target method
544 __ push(rbx);
545 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
546 __ addptr(rsp, 2*wordSize);
548 if (UseSSE == 1) {
549 __ movflt(xmm0, Address(rsp, 0));
550 __ movflt(xmm1, Address(rsp, wordSize));
551 __ addptr(rsp, 2*wordSize);
552 }
553 if (UseSSE >= 2) {
554 __ movdbl(xmm0, Address(rsp, 0));
555 __ movdbl(xmm1, Address(rsp, 2*wordSize));
556 __ addptr(rsp, 4*wordSize);
557 }
559 __ popf();
560 __ popa();
561 __ bind(L);
562 }
565 static void move_c2i_double(MacroAssembler *masm, XMMRegister r, int st_off) {
566 int next_off = st_off - Interpreter::stackElementSize;
567 __ movdbl(Address(rsp, next_off), r);
568 }
570 static void gen_c2i_adapter(MacroAssembler *masm,
571 int total_args_passed,
572 int comp_args_on_stack,
573 const BasicType *sig_bt,
574 const VMRegPair *regs,
575 Label& skip_fixup) {
576 // Before we get into the guts of the C2I adapter, see if we should be here
577 // at all. We've come from compiled code and are attempting to jump to the
578 // interpreter, which means the caller made a static call to get here
579 // (vcalls always get a compiled target if there is one). Check for a
580 // compiled target. If there is one, we need to patch the caller's call.
581 patch_callers_callsite(masm);
583 __ bind(skip_fixup);
585 #ifdef COMPILER2
586 // C2 may leave the stack dirty if not in SSE2+ mode
587 if (UseSSE >= 2) {
588 __ verify_FPU(0, "c2i transition should have clean FPU stack");
589 } else {
590 __ empty_FPU_stack();
591 }
592 #endif /* COMPILER2 */
594 // Since all args are passed on the stack, total_args_passed * interpreter_
595 // stack_element_size is the
596 // space we need.
597 int extraspace = total_args_passed * Interpreter::stackElementSize;
599 // Get return address
600 __ pop(rax);
602 // set senderSP value
603 __ movptr(rsi, rsp);
605 __ subptr(rsp, extraspace);
607 // Now write the args into the outgoing interpreter space
608 for (int i = 0; i < total_args_passed; i++) {
609 if (sig_bt[i] == T_VOID) {
610 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
611 continue;
612 }
614 // st_off points to lowest address on stack.
615 int st_off = ((total_args_passed - 1) - i) * Interpreter::stackElementSize;
616 int next_off = st_off - Interpreter::stackElementSize;
618 // Say 4 args:
619 // i st_off
620 // 0 12 T_LONG
621 // 1 8 T_VOID
622 // 2 4 T_OBJECT
623 // 3 0 T_BOOL
624 VMReg r_1 = regs[i].first();
625 VMReg r_2 = regs[i].second();
626 if (!r_1->is_valid()) {
627 assert(!r_2->is_valid(), "");
628 continue;
629 }
631 if (r_1->is_stack()) {
632 // memory to memory use fpu stack top
633 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
635 if (!r_2->is_valid()) {
636 __ movl(rdi, Address(rsp, ld_off));
637 __ movptr(Address(rsp, st_off), rdi);
638 } else {
640 // ld_off == LSW, ld_off+VMRegImpl::stack_slot_size == MSW
641 // st_off == MSW, st_off-wordSize == LSW
643 __ movptr(rdi, Address(rsp, ld_off));
644 __ movptr(Address(rsp, next_off), rdi);
645 #ifndef _LP64
646 __ movptr(rdi, Address(rsp, ld_off + wordSize));
647 __ movptr(Address(rsp, st_off), rdi);
648 #else
649 #ifdef ASSERT
650 // Overwrite the unused slot with known junk
651 __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
652 __ movptr(Address(rsp, st_off), rax);
653 #endif /* ASSERT */
654 #endif // _LP64
655 }
656 } else if (r_1->is_Register()) {
657 Register r = r_1->as_Register();
658 if (!r_2->is_valid()) {
659 __ movl(Address(rsp, st_off), r);
660 } else {
661 // long/double in gpr
662 NOT_LP64(ShouldNotReachHere());
663 // Two VMRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
664 // T_DOUBLE and T_LONG use two slots in the interpreter
665 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
666 // long/double in gpr
667 #ifdef ASSERT
668 // Overwrite the unused slot with known junk
669 LP64_ONLY(__ mov64(rax, CONST64(0xdeadffffdeadaaab)));
670 __ movptr(Address(rsp, st_off), rax);
671 #endif /* ASSERT */
672 __ movptr(Address(rsp, next_off), r);
673 } else {
674 __ movptr(Address(rsp, st_off), r);
675 }
676 }
677 } else {
678 assert(r_1->is_XMMRegister(), "");
679 if (!r_2->is_valid()) {
680 __ movflt(Address(rsp, st_off), r_1->as_XMMRegister());
681 } else {
682 assert(sig_bt[i] == T_DOUBLE || sig_bt[i] == T_LONG, "wrong type");
683 move_c2i_double(masm, r_1->as_XMMRegister(), st_off);
684 }
685 }
686 }
688 // Schedule the branch target address early.
689 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
690 // And repush original return address
691 __ push(rax);
692 __ jmp(rcx);
693 }
696 static void move_i2c_double(MacroAssembler *masm, XMMRegister r, Register saved_sp, int ld_off) {
697 int next_val_off = ld_off - Interpreter::stackElementSize;
698 __ movdbl(r, Address(saved_sp, next_val_off));
699 }
701 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg,
702 address code_start, address code_end,
703 Label& L_ok) {
704 Label L_fail;
705 __ lea(temp_reg, ExternalAddress(code_start));
706 __ cmpptr(pc_reg, temp_reg);
707 __ jcc(Assembler::belowEqual, L_fail);
708 __ lea(temp_reg, ExternalAddress(code_end));
709 __ cmpptr(pc_reg, temp_reg);
710 __ jcc(Assembler::below, L_ok);
711 __ bind(L_fail);
712 }
714 static void gen_i2c_adapter(MacroAssembler *masm,
715 int total_args_passed,
716 int comp_args_on_stack,
717 const BasicType *sig_bt,
718 const VMRegPair *regs) {
720 // Note: rsi contains the senderSP on entry. We must preserve it since
721 // we may do a i2c -> c2i transition if we lose a race where compiled
722 // code goes non-entrant while we get args ready.
724 // Adapters can be frameless because they do not require the caller
725 // to perform additional cleanup work, such as correcting the stack pointer.
726 // An i2c adapter is frameless because the *caller* frame, which is interpreted,
727 // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
728 // even if a callee has modified the stack pointer.
729 // A c2i adapter is frameless because the *callee* frame, which is interpreted,
730 // routinely repairs its caller's stack pointer (from sender_sp, which is set
731 // up via the senderSP register).
732 // In other words, if *either* the caller or callee is interpreted, we can
733 // get the stack pointer repaired after a call.
734 // This is why c2i and i2c adapters cannot be indefinitely composed.
735 // In particular, if a c2i adapter were to somehow call an i2c adapter,
736 // both caller and callee would be compiled methods, and neither would
737 // clean up the stack pointer changes performed by the two adapters.
738 // If this happens, control eventually transfers back to the compiled
739 // caller, but with an uncorrected stack, causing delayed havoc.
741 // Pick up the return address
742 __ movptr(rax, Address(rsp, 0));
744 if (VerifyAdapterCalls &&
745 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) {
746 // So, let's test for cascading c2i/i2c adapters right now.
747 // assert(Interpreter::contains($return_addr) ||
748 // StubRoutines::contains($return_addr),
749 // "i2c adapter must return to an interpreter frame");
750 __ block_comment("verify_i2c { ");
751 Label L_ok;
752 if (Interpreter::code() != NULL)
753 range_check(masm, rax, rdi,
754 Interpreter::code()->code_start(), Interpreter::code()->code_end(),
755 L_ok);
756 if (StubRoutines::code1() != NULL)
757 range_check(masm, rax, rdi,
758 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(),
759 L_ok);
760 if (StubRoutines::code2() != NULL)
761 range_check(masm, rax, rdi,
762 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(),
763 L_ok);
764 const char* msg = "i2c adapter must return to an interpreter frame";
765 __ block_comment(msg);
766 __ stop(msg);
767 __ bind(L_ok);
768 __ block_comment("} verify_i2ce ");
769 }
771 // Must preserve original SP for loading incoming arguments because
772 // we need to align the outgoing SP for compiled code.
773 __ movptr(rdi, rsp);
775 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
776 // in registers, we will occasionally have no stack args.
777 int comp_words_on_stack = 0;
778 if (comp_args_on_stack) {
779 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
780 // registers are below. By subtracting stack0, we either get a negative
781 // number (all values in registers) or the maximum stack slot accessed.
782 // int comp_args_on_stack = VMRegImpl::reg2stack(max_arg);
783 // Convert 4-byte stack slots to words.
784 comp_words_on_stack = round_to(comp_args_on_stack*4, wordSize)>>LogBytesPerWord;
785 // Round up to miminum stack alignment, in wordSize
786 comp_words_on_stack = round_to(comp_words_on_stack, 2);
787 __ subptr(rsp, comp_words_on_stack * wordSize);
788 }
790 // Align the outgoing SP
791 __ andptr(rsp, -(StackAlignmentInBytes));
793 // push the return address on the stack (note that pushing, rather
794 // than storing it, yields the correct frame alignment for the callee)
795 __ push(rax);
797 // Put saved SP in another register
798 const Register saved_sp = rax;
799 __ movptr(saved_sp, rdi);
802 // Will jump to the compiled code just as if compiled code was doing it.
803 // Pre-load the register-jump target early, to schedule it better.
804 __ movptr(rdi, Address(rbx, in_bytes(Method::from_compiled_offset())));
806 // Now generate the shuffle code. Pick up all register args and move the
807 // rest through the floating point stack top.
808 for (int i = 0; i < total_args_passed; i++) {
809 if (sig_bt[i] == T_VOID) {
810 // Longs and doubles are passed in native word order, but misaligned
811 // in the 32-bit build.
812 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
813 continue;
814 }
816 // Pick up 0, 1 or 2 words from SP+offset.
818 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
819 "scrambled load targets?");
820 // Load in argument order going down.
821 int ld_off = (total_args_passed - i) * Interpreter::stackElementSize;
822 // Point to interpreter value (vs. tag)
823 int next_off = ld_off - Interpreter::stackElementSize;
824 //
825 //
826 //
827 VMReg r_1 = regs[i].first();
828 VMReg r_2 = regs[i].second();
829 if (!r_1->is_valid()) {
830 assert(!r_2->is_valid(), "");
831 continue;
832 }
833 if (r_1->is_stack()) {
834 // Convert stack slot to an SP offset (+ wordSize to account for return address )
835 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
837 // We can use rsi as a temp here because compiled code doesn't need rsi as an input
838 // and if we end up going thru a c2i because of a miss a reasonable value of rsi
839 // we be generated.
840 if (!r_2->is_valid()) {
841 // __ fld_s(Address(saved_sp, ld_off));
842 // __ fstp_s(Address(rsp, st_off));
843 __ movl(rsi, Address(saved_sp, ld_off));
844 __ movptr(Address(rsp, st_off), rsi);
845 } else {
846 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
847 // are accessed as negative so LSW is at LOW address
849 // ld_off is MSW so get LSW
850 // st_off is LSW (i.e. reg.first())
851 // __ fld_d(Address(saved_sp, next_off));
852 // __ fstp_d(Address(rsp, st_off));
853 //
854 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
855 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
856 // So we must adjust where to pick up the data to match the interpreter.
857 //
858 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
859 // are accessed as negative so LSW is at LOW address
861 // ld_off is MSW so get LSW
862 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
863 next_off : ld_off;
864 __ movptr(rsi, Address(saved_sp, offset));
865 __ movptr(Address(rsp, st_off), rsi);
866 #ifndef _LP64
867 __ movptr(rsi, Address(saved_sp, ld_off));
868 __ movptr(Address(rsp, st_off + wordSize), rsi);
869 #endif // _LP64
870 }
871 } else if (r_1->is_Register()) { // Register argument
872 Register r = r_1->as_Register();
873 assert(r != rax, "must be different");
874 if (r_2->is_valid()) {
875 //
876 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
877 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
878 // So we must adjust where to pick up the data to match the interpreter.
880 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
881 next_off : ld_off;
883 // this can be a misaligned move
884 __ movptr(r, Address(saved_sp, offset));
885 #ifndef _LP64
886 assert(r_2->as_Register() != rax, "need another temporary register");
887 // Remember r_1 is low address (and LSB on x86)
888 // So r_2 gets loaded from high address regardless of the platform
889 __ movptr(r_2->as_Register(), Address(saved_sp, ld_off));
890 #endif // _LP64
891 } else {
892 __ movl(r, Address(saved_sp, ld_off));
893 }
894 } else {
895 assert(r_1->is_XMMRegister(), "");
896 if (!r_2->is_valid()) {
897 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
898 } else {
899 move_i2c_double(masm, r_1->as_XMMRegister(), saved_sp, ld_off);
900 }
901 }
902 }
904 // 6243940 We might end up in handle_wrong_method if
905 // the callee is deoptimized as we race thru here. If that
906 // happens we don't want to take a safepoint because the
907 // caller frame will look interpreted and arguments are now
908 // "compiled" so it is much better to make this transition
909 // invisible to the stack walking code. Unfortunately if
910 // we try and find the callee by normal means a safepoint
911 // is possible. So we stash the desired callee in the thread
912 // and the vm will find there should this case occur.
914 __ get_thread(rax);
915 __ movptr(Address(rax, JavaThread::callee_target_offset()), rbx);
917 // move Method* to rax, in case we end up in an c2i adapter.
918 // the c2i adapters expect Method* in rax, (c2) because c2's
919 // resolve stubs return the result (the method) in rax,.
920 // I'd love to fix this.
921 __ mov(rax, rbx);
923 __ jmp(rdi);
924 }
926 // ---------------------------------------------------------------
927 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
928 int total_args_passed,
929 int comp_args_on_stack,
930 const BasicType *sig_bt,
931 const VMRegPair *regs,
932 AdapterFingerPrint* fingerprint) {
933 address i2c_entry = __ pc();
935 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
937 // -------------------------------------------------------------------------
938 // Generate a C2I adapter. On entry we know rbx, holds the Method* during calls
939 // to the interpreter. The args start out packed in the compiled layout. They
940 // need to be unpacked into the interpreter layout. This will almost always
941 // require some stack space. We grow the current (compiled) stack, then repack
942 // the args. We finally end in a jump to the generic interpreter entry point.
943 // On exit from the interpreter, the interpreter will restore our SP (lest the
944 // compiled code, which relys solely on SP and not EBP, get sick).
946 address c2i_unverified_entry = __ pc();
947 Label skip_fixup;
949 Register holder = rax;
950 Register receiver = rcx;
951 Register temp = rbx;
953 {
955 Label missed;
956 __ movptr(temp, Address(receiver, oopDesc::klass_offset_in_bytes()));
957 __ cmpptr(temp, Address(holder, CompiledICHolder::holder_klass_offset()));
958 __ movptr(rbx, Address(holder, CompiledICHolder::holder_method_offset()));
959 __ jcc(Assembler::notEqual, missed);
960 // Method might have been compiled since the call site was patched to
961 // interpreted if that is the case treat it as a miss so we can get
962 // the call site corrected.
963 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD);
964 __ jcc(Assembler::equal, skip_fixup);
966 __ bind(missed);
967 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
968 }
970 address c2i_entry = __ pc();
972 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
974 __ flush();
975 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
976 }
978 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
979 VMRegPair *regs,
980 int total_args_passed) {
981 // We return the amount of VMRegImpl stack slots we need to reserve for all
982 // the arguments NOT counting out_preserve_stack_slots.
984 uint stack = 0; // All arguments on stack
986 for( int i = 0; i < total_args_passed; i++) {
987 // From the type and the argument number (count) compute the location
988 switch( sig_bt[i] ) {
989 case T_BOOLEAN:
990 case T_CHAR:
991 case T_FLOAT:
992 case T_BYTE:
993 case T_SHORT:
994 case T_INT:
995 case T_OBJECT:
996 case T_ARRAY:
997 case T_ADDRESS:
998 case T_METADATA:
999 regs[i].set1(VMRegImpl::stack2reg(stack++));
1000 break;
1001 case T_LONG:
1002 case T_DOUBLE: // The stack numbering is reversed from Java
1003 // Since C arguments do not get reversed, the ordering for
1004 // doubles on the stack must be opposite the Java convention
1005 assert(sig_bt[i+1] == T_VOID, "missing Half" );
1006 regs[i].set2(VMRegImpl::stack2reg(stack));
1007 stack += 2;
1008 break;
1009 case T_VOID: regs[i].set_bad(); break;
1010 default:
1011 ShouldNotReachHere();
1012 break;
1013 }
1014 }
1015 return stack;
1016 }
1018 // A simple move of integer like type
1019 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1020 if (src.first()->is_stack()) {
1021 if (dst.first()->is_stack()) {
1022 // stack to stack
1023 // __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1024 // __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1025 __ movl2ptr(rax, Address(rbp, reg2offset_in(src.first())));
1026 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1027 } else {
1028 // stack to reg
1029 __ movl2ptr(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first())));
1030 }
1031 } else if (dst.first()->is_stack()) {
1032 // reg to stack
1033 // no need to sign extend on 64bit
1034 __ movptr(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register());
1035 } else {
1036 if (dst.first() != src.first()) {
1037 __ mov(dst.first()->as_Register(), src.first()->as_Register());
1038 }
1039 }
1040 }
1042 // An oop arg. Must pass a handle not the oop itself
1043 static void object_move(MacroAssembler* masm,
1044 OopMap* map,
1045 int oop_handle_offset,
1046 int framesize_in_slots,
1047 VMRegPair src,
1048 VMRegPair dst,
1049 bool is_receiver,
1050 int* receiver_offset) {
1052 // Because of the calling conventions we know that src can be a
1053 // register or a stack location. dst can only be a stack location.
1055 assert(dst.first()->is_stack(), "must be stack");
1056 // must pass a handle. First figure out the location we use as a handle
1058 if (src.first()->is_stack()) {
1059 // Oop is already on the stack as an argument
1060 Register rHandle = rax;
1061 Label nil;
1062 __ xorptr(rHandle, rHandle);
1063 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD);
1064 __ jcc(Assembler::equal, nil);
1065 __ lea(rHandle, Address(rbp, reg2offset_in(src.first())));
1066 __ bind(nil);
1067 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle);
1069 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1070 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1071 if (is_receiver) {
1072 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1073 }
1074 } else {
1075 // Oop is in an a register we must store it to the space we reserve
1076 // on the stack for oop_handles
1077 const Register rOop = src.first()->as_Register();
1078 const Register rHandle = rax;
1079 int oop_slot = (rOop == rcx ? 0 : 1) * VMRegImpl::slots_per_word + oop_handle_offset;
1080 int offset = oop_slot*VMRegImpl::stack_slot_size;
1081 Label skip;
1082 __ movptr(Address(rsp, offset), rOop);
1083 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1084 __ xorptr(rHandle, rHandle);
1085 __ cmpptr(rOop, (int32_t)NULL_WORD);
1086 __ jcc(Assembler::equal, skip);
1087 __ lea(rHandle, Address(rsp, offset));
1088 __ bind(skip);
1089 // Store the handle parameter
1090 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle);
1091 if (is_receiver) {
1092 *receiver_offset = offset;
1093 }
1094 }
1095 }
1097 // A float arg may have to do float reg int reg conversion
1098 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1099 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1101 // Because of the calling convention we know that src is either a stack location
1102 // or an xmm register. dst can only be a stack location.
1104 assert(dst.first()->is_stack() && ( src.first()->is_stack() || src.first()->is_XMMRegister()), "bad parameters");
1106 if (src.first()->is_stack()) {
1107 __ movl(rax, Address(rbp, reg2offset_in(src.first())));
1108 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1109 } else {
1110 // reg to stack
1111 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister());
1112 }
1113 }
1115 // A long move
1116 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1118 // The only legal possibility for a long_move VMRegPair is:
1119 // 1: two stack slots (possibly unaligned)
1120 // as neither the java or C calling convention will use registers
1121 // for longs.
1123 if (src.first()->is_stack() && dst.first()->is_stack()) {
1124 assert(src.second()->is_stack() && dst.second()->is_stack(), "must be all stack");
1125 __ movptr(rax, Address(rbp, reg2offset_in(src.first())));
1126 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second()))));
1127 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1128 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx));
1129 } else {
1130 ShouldNotReachHere();
1131 }
1132 }
1134 // A double move
1135 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1137 // The only legal possibilities for a double_move VMRegPair are:
1138 // The painful thing here is that like long_move a VMRegPair might be
1140 // Because of the calling convention we know that src is either
1141 // 1: a single physical register (xmm registers only)
1142 // 2: two stack slots (possibly unaligned)
1143 // dst can only be a pair of stack slots.
1145 assert(dst.first()->is_stack() && (src.first()->is_XMMRegister() || src.first()->is_stack()), "bad args");
1147 if (src.first()->is_stack()) {
1148 // source is all stack
1149 __ movptr(rax, Address(rbp, reg2offset_in(src.first())));
1150 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second()))));
1151 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1152 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx));
1153 } else {
1154 // reg to stack
1155 // No worries about stack alignment
1156 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister());
1157 }
1158 }
1161 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1162 // We always ignore the frame_slots arg and just use the space just below frame pointer
1163 // which by this time is free to use
1164 switch (ret_type) {
1165 case T_FLOAT:
1166 __ fstp_s(Address(rbp, -wordSize));
1167 break;
1168 case T_DOUBLE:
1169 __ fstp_d(Address(rbp, -2*wordSize));
1170 break;
1171 case T_VOID: break;
1172 case T_LONG:
1173 __ movptr(Address(rbp, -wordSize), rax);
1174 NOT_LP64(__ movptr(Address(rbp, -2*wordSize), rdx));
1175 break;
1176 default: {
1177 __ movptr(Address(rbp, -wordSize), rax);
1178 }
1179 }
1180 }
1182 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1183 // We always ignore the frame_slots arg and just use the space just below frame pointer
1184 // which by this time is free to use
1185 switch (ret_type) {
1186 case T_FLOAT:
1187 __ fld_s(Address(rbp, -wordSize));
1188 break;
1189 case T_DOUBLE:
1190 __ fld_d(Address(rbp, -2*wordSize));
1191 break;
1192 case T_LONG:
1193 __ movptr(rax, Address(rbp, -wordSize));
1194 NOT_LP64(__ movptr(rdx, Address(rbp, -2*wordSize)));
1195 break;
1196 case T_VOID: break;
1197 default: {
1198 __ movptr(rax, Address(rbp, -wordSize));
1199 }
1200 }
1201 }
1204 static void save_or_restore_arguments(MacroAssembler* masm,
1205 const int stack_slots,
1206 const int total_in_args,
1207 const int arg_save_area,
1208 OopMap* map,
1209 VMRegPair* in_regs,
1210 BasicType* in_sig_bt) {
1211 // if map is non-NULL then the code should store the values,
1212 // otherwise it should load them.
1213 int handle_index = 0;
1214 // Save down double word first
1215 for ( int i = 0; i < total_in_args; i++) {
1216 if (in_regs[i].first()->is_XMMRegister() && in_sig_bt[i] == T_DOUBLE) {
1217 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area;
1218 int offset = slot * VMRegImpl::stack_slot_size;
1219 handle_index += 2;
1220 assert(handle_index <= stack_slots, "overflow");
1221 if (map != NULL) {
1222 __ movdbl(Address(rsp, offset), in_regs[i].first()->as_XMMRegister());
1223 } else {
1224 __ movdbl(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset));
1225 }
1226 }
1227 if (in_regs[i].first()->is_Register() && in_sig_bt[i] == T_LONG) {
1228 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area;
1229 int offset = slot * VMRegImpl::stack_slot_size;
1230 handle_index += 2;
1231 assert(handle_index <= stack_slots, "overflow");
1232 if (map != NULL) {
1233 __ movl(Address(rsp, offset), in_regs[i].first()->as_Register());
1234 if (in_regs[i].second()->is_Register()) {
1235 __ movl(Address(rsp, offset + 4), in_regs[i].second()->as_Register());
1236 }
1237 } else {
1238 __ movl(in_regs[i].first()->as_Register(), Address(rsp, offset));
1239 if (in_regs[i].second()->is_Register()) {
1240 __ movl(in_regs[i].second()->as_Register(), Address(rsp, offset + 4));
1241 }
1242 }
1243 }
1244 }
1245 // Save or restore single word registers
1246 for ( int i = 0; i < total_in_args; i++) {
1247 if (in_regs[i].first()->is_Register()) {
1248 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area;
1249 int offset = slot * VMRegImpl::stack_slot_size;
1250 assert(handle_index <= stack_slots, "overflow");
1251 if (in_sig_bt[i] == T_ARRAY && map != NULL) {
1252 map->set_oop(VMRegImpl::stack2reg(slot));;
1253 }
1255 // Value is in an input register pass we must flush it to the stack
1256 const Register reg = in_regs[i].first()->as_Register();
1257 switch (in_sig_bt[i]) {
1258 case T_ARRAY:
1259 if (map != NULL) {
1260 __ movptr(Address(rsp, offset), reg);
1261 } else {
1262 __ movptr(reg, Address(rsp, offset));
1263 }
1264 break;
1265 case T_BOOLEAN:
1266 case T_CHAR:
1267 case T_BYTE:
1268 case T_SHORT:
1269 case T_INT:
1270 if (map != NULL) {
1271 __ movl(Address(rsp, offset), reg);
1272 } else {
1273 __ movl(reg, Address(rsp, offset));
1274 }
1275 break;
1276 case T_OBJECT:
1277 default: ShouldNotReachHere();
1278 }
1279 } else if (in_regs[i].first()->is_XMMRegister()) {
1280 if (in_sig_bt[i] == T_FLOAT) {
1281 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area;
1282 int offset = slot * VMRegImpl::stack_slot_size;
1283 assert(handle_index <= stack_slots, "overflow");
1284 if (map != NULL) {
1285 __ movflt(Address(rsp, offset), in_regs[i].first()->as_XMMRegister());
1286 } else {
1287 __ movflt(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset));
1288 }
1289 }
1290 } else if (in_regs[i].first()->is_stack()) {
1291 if (in_sig_bt[i] == T_ARRAY && map != NULL) {
1292 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1293 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots));
1294 }
1295 }
1296 }
1297 }
1299 // Check GC_locker::needs_gc and enter the runtime if it's true. This
1300 // keeps a new JNI critical region from starting until a GC has been
1301 // forced. Save down any oops in registers and describe them in an
1302 // OopMap.
1303 static void check_needs_gc_for_critical_native(MacroAssembler* masm,
1304 Register thread,
1305 int stack_slots,
1306 int total_c_args,
1307 int total_in_args,
1308 int arg_save_area,
1309 OopMapSet* oop_maps,
1310 VMRegPair* in_regs,
1311 BasicType* in_sig_bt) {
1312 __ block_comment("check GC_locker::needs_gc");
1313 Label cont;
1314 __ cmp8(ExternalAddress((address)GC_locker::needs_gc_address()), false);
1315 __ jcc(Assembler::equal, cont);
1317 // Save down any incoming oops and call into the runtime to halt for a GC
1319 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1321 save_or_restore_arguments(masm, stack_slots, total_in_args,
1322 arg_save_area, map, in_regs, in_sig_bt);
1324 address the_pc = __ pc();
1325 oop_maps->add_gc_map( __ offset(), map);
1326 __ set_last_Java_frame(thread, rsp, noreg, the_pc);
1328 __ block_comment("block_for_jni_critical");
1329 __ push(thread);
1330 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical)));
1331 __ increment(rsp, wordSize);
1333 __ get_thread(thread);
1334 __ reset_last_Java_frame(thread, false, true);
1336 save_or_restore_arguments(masm, stack_slots, total_in_args,
1337 arg_save_area, NULL, in_regs, in_sig_bt);
1339 __ bind(cont);
1340 #ifdef ASSERT
1341 if (StressCriticalJNINatives) {
1342 // Stress register saving
1343 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1344 save_or_restore_arguments(masm, stack_slots, total_in_args,
1345 arg_save_area, map, in_regs, in_sig_bt);
1346 // Destroy argument registers
1347 for (int i = 0; i < total_in_args - 1; i++) {
1348 if (in_regs[i].first()->is_Register()) {
1349 const Register reg = in_regs[i].first()->as_Register();
1350 __ xorptr(reg, reg);
1351 } else if (in_regs[i].first()->is_XMMRegister()) {
1352 __ xorpd(in_regs[i].first()->as_XMMRegister(), in_regs[i].first()->as_XMMRegister());
1353 } else if (in_regs[i].first()->is_FloatRegister()) {
1354 ShouldNotReachHere();
1355 } else if (in_regs[i].first()->is_stack()) {
1356 // Nothing to do
1357 } else {
1358 ShouldNotReachHere();
1359 }
1360 if (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_DOUBLE) {
1361 i++;
1362 }
1363 }
1365 save_or_restore_arguments(masm, stack_slots, total_in_args,
1366 arg_save_area, NULL, in_regs, in_sig_bt);
1367 }
1368 #endif
1369 }
1371 // Unpack an array argument into a pointer to the body and the length
1372 // if the array is non-null, otherwise pass 0 for both.
1373 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) {
1374 Register tmp_reg = rax;
1375 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg,
1376 "possible collision");
1377 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg,
1378 "possible collision");
1380 // Pass the length, ptr pair
1381 Label is_null, done;
1382 VMRegPair tmp(tmp_reg->as_VMReg());
1383 if (reg.first()->is_stack()) {
1384 // Load the arg up from the stack
1385 simple_move32(masm, reg, tmp);
1386 reg = tmp;
1387 }
1388 __ testptr(reg.first()->as_Register(), reg.first()->as_Register());
1389 __ jccb(Assembler::equal, is_null);
1390 __ lea(tmp_reg, Address(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type)));
1391 simple_move32(masm, tmp, body_arg);
1392 // load the length relative to the body.
1393 __ movl(tmp_reg, Address(tmp_reg, arrayOopDesc::length_offset_in_bytes() -
1394 arrayOopDesc::base_offset_in_bytes(in_elem_type)));
1395 simple_move32(masm, tmp, length_arg);
1396 __ jmpb(done);
1397 __ bind(is_null);
1398 // Pass zeros
1399 __ xorptr(tmp_reg, tmp_reg);
1400 simple_move32(masm, tmp, body_arg);
1401 simple_move32(masm, tmp, length_arg);
1402 __ bind(done);
1403 }
1405 static void verify_oop_args(MacroAssembler* masm,
1406 methodHandle method,
1407 const BasicType* sig_bt,
1408 const VMRegPair* regs) {
1409 Register temp_reg = rbx; // not part of any compiled calling seq
1410 if (VerifyOops) {
1411 for (int i = 0; i < method->size_of_parameters(); i++) {
1412 if (sig_bt[i] == T_OBJECT ||
1413 sig_bt[i] == T_ARRAY) {
1414 VMReg r = regs[i].first();
1415 assert(r->is_valid(), "bad oop arg");
1416 if (r->is_stack()) {
1417 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1418 __ verify_oop(temp_reg);
1419 } else {
1420 __ verify_oop(r->as_Register());
1421 }
1422 }
1423 }
1424 }
1425 }
1427 static void gen_special_dispatch(MacroAssembler* masm,
1428 methodHandle method,
1429 const BasicType* sig_bt,
1430 const VMRegPair* regs) {
1431 verify_oop_args(masm, method, sig_bt, regs);
1432 vmIntrinsics::ID iid = method->intrinsic_id();
1434 // Now write the args into the outgoing interpreter space
1435 bool has_receiver = false;
1436 Register receiver_reg = noreg;
1437 int member_arg_pos = -1;
1438 Register member_reg = noreg;
1439 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1440 if (ref_kind != 0) {
1441 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1442 member_reg = rbx; // known to be free at this point
1443 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1444 } else if (iid == vmIntrinsics::_invokeBasic) {
1445 has_receiver = true;
1446 } else {
1447 fatal(err_msg_res("unexpected intrinsic id %d", iid));
1448 }
1450 if (member_reg != noreg) {
1451 // Load the member_arg into register, if necessary.
1452 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1453 VMReg r = regs[member_arg_pos].first();
1454 if (r->is_stack()) {
1455 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1456 } else {
1457 // no data motion is needed
1458 member_reg = r->as_Register();
1459 }
1460 }
1462 if (has_receiver) {
1463 // Make sure the receiver is loaded into a register.
1464 assert(method->size_of_parameters() > 0, "oob");
1465 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1466 VMReg r = regs[0].first();
1467 assert(r->is_valid(), "bad receiver arg");
1468 if (r->is_stack()) {
1469 // Porting note: This assumes that compiled calling conventions always
1470 // pass the receiver oop in a register. If this is not true on some
1471 // platform, pick a temp and load the receiver from stack.
1472 fatal("receiver always in a register");
1473 receiver_reg = rcx; // known to be free at this point
1474 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1475 } else {
1476 // no data motion is needed
1477 receiver_reg = r->as_Register();
1478 }
1479 }
1481 // Figure out which address we are really jumping to:
1482 MethodHandles::generate_method_handle_dispatch(masm, iid,
1483 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1484 }
1486 // ---------------------------------------------------------------------------
1487 // Generate a native wrapper for a given method. The method takes arguments
1488 // in the Java compiled code convention, marshals them to the native
1489 // convention (handlizes oops, etc), transitions to native, makes the call,
1490 // returns to java state (possibly blocking), unhandlizes any result and
1491 // returns.
1492 //
1493 // Critical native functions are a shorthand for the use of
1494 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1495 // functions. The wrapper is expected to unpack the arguments before
1496 // passing them to the callee and perform checks before and after the
1497 // native call to ensure that they GC_locker
1498 // lock_critical/unlock_critical semantics are followed. Some other
1499 // parts of JNI setup are skipped like the tear down of the JNI handle
1500 // block and the check for pending exceptions it's impossible for them
1501 // to be thrown.
1502 //
1503 // They are roughly structured like this:
1504 // if (GC_locker::needs_gc())
1505 // SharedRuntime::block_for_jni_critical();
1506 // tranistion to thread_in_native
1507 // unpack arrray arguments and call native entry point
1508 // check for safepoint in progress
1509 // check if any thread suspend flags are set
1510 // call into JVM and possible unlock the JNI critical
1511 // if a GC was suppressed while in the critical native.
1512 // transition back to thread_in_Java
1513 // return to caller
1514 //
1515 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1516 methodHandle method,
1517 int compile_id,
1518 BasicType* in_sig_bt,
1519 VMRegPair* in_regs,
1520 BasicType ret_type) {
1521 if (method->is_method_handle_intrinsic()) {
1522 vmIntrinsics::ID iid = method->intrinsic_id();
1523 intptr_t start = (intptr_t)__ pc();
1524 int vep_offset = ((intptr_t)__ pc()) - start;
1525 gen_special_dispatch(masm,
1526 method,
1527 in_sig_bt,
1528 in_regs);
1529 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1530 __ flush();
1531 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1532 return nmethod::new_native_nmethod(method,
1533 compile_id,
1534 masm->code(),
1535 vep_offset,
1536 frame_complete,
1537 stack_slots / VMRegImpl::slots_per_word,
1538 in_ByteSize(-1),
1539 in_ByteSize(-1),
1540 (OopMapSet*)NULL);
1541 }
1542 bool is_critical_native = true;
1543 address native_func = method->critical_native_function();
1544 if (native_func == NULL) {
1545 native_func = method->native_function();
1546 is_critical_native = false;
1547 }
1548 assert(native_func != NULL, "must have function");
1550 // An OopMap for lock (and class if static)
1551 OopMapSet *oop_maps = new OopMapSet();
1553 // We have received a description of where all the java arg are located
1554 // on entry to the wrapper. We need to convert these args to where
1555 // the jni function will expect them. To figure out where they go
1556 // we convert the java signature to a C signature by inserting
1557 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1559 const int total_in_args = method->size_of_parameters();
1560 int total_c_args = total_in_args;
1561 if (!is_critical_native) {
1562 total_c_args += 1;
1563 if (method->is_static()) {
1564 total_c_args++;
1565 }
1566 } else {
1567 for (int i = 0; i < total_in_args; i++) {
1568 if (in_sig_bt[i] == T_ARRAY) {
1569 total_c_args++;
1570 }
1571 }
1572 }
1574 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1575 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1576 BasicType* in_elem_bt = NULL;
1578 int argc = 0;
1579 if (!is_critical_native) {
1580 out_sig_bt[argc++] = T_ADDRESS;
1581 if (method->is_static()) {
1582 out_sig_bt[argc++] = T_OBJECT;
1583 }
1585 for (int i = 0; i < total_in_args ; i++ ) {
1586 out_sig_bt[argc++] = in_sig_bt[i];
1587 }
1588 } else {
1589 Thread* THREAD = Thread::current();
1590 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args);
1591 SignatureStream ss(method->signature());
1592 for (int i = 0; i < total_in_args ; i++ ) {
1593 if (in_sig_bt[i] == T_ARRAY) {
1594 // Arrays are passed as int, elem* pair
1595 out_sig_bt[argc++] = T_INT;
1596 out_sig_bt[argc++] = T_ADDRESS;
1597 Symbol* atype = ss.as_symbol(CHECK_NULL);
1598 const char* at = atype->as_C_string();
1599 if (strlen(at) == 2) {
1600 assert(at[0] == '[', "must be");
1601 switch (at[1]) {
1602 case 'B': in_elem_bt[i] = T_BYTE; break;
1603 case 'C': in_elem_bt[i] = T_CHAR; break;
1604 case 'D': in_elem_bt[i] = T_DOUBLE; break;
1605 case 'F': in_elem_bt[i] = T_FLOAT; break;
1606 case 'I': in_elem_bt[i] = T_INT; break;
1607 case 'J': in_elem_bt[i] = T_LONG; break;
1608 case 'S': in_elem_bt[i] = T_SHORT; break;
1609 case 'Z': in_elem_bt[i] = T_BOOLEAN; break;
1610 default: ShouldNotReachHere();
1611 }
1612 }
1613 } else {
1614 out_sig_bt[argc++] = in_sig_bt[i];
1615 in_elem_bt[i] = T_VOID;
1616 }
1617 if (in_sig_bt[i] != T_VOID) {
1618 assert(in_sig_bt[i] == ss.type(), "must match");
1619 ss.next();
1620 }
1621 }
1622 }
1624 // Now figure out where the args must be stored and how much stack space
1625 // they require.
1626 int out_arg_slots;
1627 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1629 // Compute framesize for the wrapper. We need to handlize all oops in
1630 // registers a max of 2 on x86.
1632 // Calculate the total number of stack slots we will need.
1634 // First count the abi requirement plus all of the outgoing args
1635 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1637 // Now the space for the inbound oop handle area
1638 int total_save_slots = 2 * VMRegImpl::slots_per_word; // 2 arguments passed in registers
1639 if (is_critical_native) {
1640 // Critical natives may have to call out so they need a save area
1641 // for register arguments.
1642 int double_slots = 0;
1643 int single_slots = 0;
1644 for ( int i = 0; i < total_in_args; i++) {
1645 if (in_regs[i].first()->is_Register()) {
1646 const Register reg = in_regs[i].first()->as_Register();
1647 switch (in_sig_bt[i]) {
1648 case T_ARRAY: // critical array (uses 2 slots on LP64)
1649 case T_BOOLEAN:
1650 case T_BYTE:
1651 case T_SHORT:
1652 case T_CHAR:
1653 case T_INT: single_slots++; break;
1654 case T_LONG: double_slots++; break;
1655 default: ShouldNotReachHere();
1656 }
1657 } else if (in_regs[i].first()->is_XMMRegister()) {
1658 switch (in_sig_bt[i]) {
1659 case T_FLOAT: single_slots++; break;
1660 case T_DOUBLE: double_slots++; break;
1661 default: ShouldNotReachHere();
1662 }
1663 } else if (in_regs[i].first()->is_FloatRegister()) {
1664 ShouldNotReachHere();
1665 }
1666 }
1667 total_save_slots = double_slots * 2 + single_slots;
1668 // align the save area
1669 if (double_slots != 0) {
1670 stack_slots = round_to(stack_slots, 2);
1671 }
1672 }
1674 int oop_handle_offset = stack_slots;
1675 stack_slots += total_save_slots;
1677 // Now any space we need for handlizing a klass if static method
1679 int klass_slot_offset = 0;
1680 int klass_offset = -1;
1681 int lock_slot_offset = 0;
1682 bool is_static = false;
1684 if (method->is_static()) {
1685 klass_slot_offset = stack_slots;
1686 stack_slots += VMRegImpl::slots_per_word;
1687 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1688 is_static = true;
1689 }
1691 // Plus a lock if needed
1693 if (method->is_synchronized()) {
1694 lock_slot_offset = stack_slots;
1695 stack_slots += VMRegImpl::slots_per_word;
1696 }
1698 // Now a place (+2) to save return values or temp during shuffling
1699 // + 2 for return address (which we own) and saved rbp,
1700 stack_slots += 4;
1702 // Ok The space we have allocated will look like:
1703 //
1704 //
1705 // FP-> | |
1706 // |---------------------|
1707 // | 2 slots for moves |
1708 // |---------------------|
1709 // | lock box (if sync) |
1710 // |---------------------| <- lock_slot_offset (-lock_slot_rbp_offset)
1711 // | klass (if static) |
1712 // |---------------------| <- klass_slot_offset
1713 // | oopHandle area |
1714 // |---------------------| <- oop_handle_offset (a max of 2 registers)
1715 // | outbound memory |
1716 // | based arguments |
1717 // | |
1718 // |---------------------|
1719 // | |
1720 // SP-> | out_preserved_slots |
1721 //
1722 //
1723 // ****************************************************************************
1724 // WARNING - on Windows Java Natives use pascal calling convention and pop the
1725 // arguments off of the stack after the jni call. Before the call we can use
1726 // instructions that are SP relative. After the jni call we switch to FP
1727 // relative instructions instead of re-adjusting the stack on windows.
1728 // ****************************************************************************
1731 // Now compute actual number of stack words we need rounding to make
1732 // stack properly aligned.
1733 stack_slots = round_to(stack_slots, StackAlignmentInSlots);
1735 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1737 intptr_t start = (intptr_t)__ pc();
1739 // First thing make an ic check to see if we should even be here
1741 // We are free to use all registers as temps without saving them and
1742 // restoring them except rbp. rbp is the only callee save register
1743 // as far as the interpreter and the compiler(s) are concerned.
1746 const Register ic_reg = rax;
1747 const Register receiver = rcx;
1748 Label hit;
1749 Label exception_pending;
1751 __ verify_oop(receiver);
1752 __ cmpptr(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes()));
1753 __ jcc(Assembler::equal, hit);
1755 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1757 // verified entry must be aligned for code patching.
1758 // and the first 5 bytes must be in the same cache line
1759 // if we align at 8 then we will be sure 5 bytes are in the same line
1760 __ align(8);
1762 __ bind(hit);
1764 int vep_offset = ((intptr_t)__ pc()) - start;
1766 #ifdef COMPILER1
1767 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
1768 // Object.hashCode can pull the hashCode from the header word
1769 // instead of doing a full VM transition once it's been computed.
1770 // Since hashCode is usually polymorphic at call sites we can't do
1771 // this optimization at the call site without a lot of work.
1772 Label slowCase;
1773 Register receiver = rcx;
1774 Register result = rax;
1775 __ movptr(result, Address(receiver, oopDesc::mark_offset_in_bytes()));
1777 // check if locked
1778 __ testptr(result, markOopDesc::unlocked_value);
1779 __ jcc (Assembler::zero, slowCase);
1781 if (UseBiasedLocking) {
1782 // Check if biased and fall through to runtime if so
1783 __ testptr(result, markOopDesc::biased_lock_bit_in_place);
1784 __ jcc (Assembler::notZero, slowCase);
1785 }
1787 // get hash
1788 __ andptr(result, markOopDesc::hash_mask_in_place);
1789 // test if hashCode exists
1790 __ jcc (Assembler::zero, slowCase);
1791 __ shrptr(result, markOopDesc::hash_shift);
1792 __ ret(0);
1793 __ bind (slowCase);
1794 }
1795 #endif // COMPILER1
1797 // The instruction at the verified entry point must be 5 bytes or longer
1798 // because it can be patched on the fly by make_non_entrant. The stack bang
1799 // instruction fits that requirement.
1801 // Generate stack overflow check
1803 if (UseStackBanging) {
1804 __ bang_stack_with_offset(StackShadowPages*os::vm_page_size());
1805 } else {
1806 // need a 5 byte instruction to allow MT safe patching to non-entrant
1807 __ fat_nop();
1808 }
1810 // Generate a new frame for the wrapper.
1811 __ enter();
1812 // -2 because return address is already present and so is saved rbp
1813 __ subptr(rsp, stack_size - 2*wordSize);
1815 // Frame is now completed as far as size and linkage.
1816 int frame_complete = ((intptr_t)__ pc()) - start;
1818 // Calculate the difference between rsp and rbp,. We need to know it
1819 // after the native call because on windows Java Natives will pop
1820 // the arguments and it is painful to do rsp relative addressing
1821 // in a platform independent way. So after the call we switch to
1822 // rbp, relative addressing.
1824 int fp_adjustment = stack_size - 2*wordSize;
1826 #ifdef COMPILER2
1827 // C2 may leave the stack dirty if not in SSE2+ mode
1828 if (UseSSE >= 2) {
1829 __ verify_FPU(0, "c2i transition should have clean FPU stack");
1830 } else {
1831 __ empty_FPU_stack();
1832 }
1833 #endif /* COMPILER2 */
1835 // Compute the rbp, offset for any slots used after the jni call
1837 int lock_slot_rbp_offset = (lock_slot_offset*VMRegImpl::stack_slot_size) - fp_adjustment;
1839 // We use rdi as a thread pointer because it is callee save and
1840 // if we load it once it is usable thru the entire wrapper
1841 const Register thread = rdi;
1843 // We use rsi as the oop handle for the receiver/klass
1844 // It is callee save so it survives the call to native
1846 const Register oop_handle_reg = rsi;
1848 __ get_thread(thread);
1850 if (is_critical_native) {
1851 check_needs_gc_for_critical_native(masm, thread, stack_slots, total_c_args, total_in_args,
1852 oop_handle_offset, oop_maps, in_regs, in_sig_bt);
1853 }
1855 //
1856 // We immediately shuffle the arguments so that any vm call we have to
1857 // make from here on out (sync slow path, jvmti, etc.) we will have
1858 // captured the oops from our caller and have a valid oopMap for
1859 // them.
1861 // -----------------
1862 // The Grand Shuffle
1863 //
1864 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
1865 // and, if static, the class mirror instead of a receiver. This pretty much
1866 // guarantees that register layout will not match (and x86 doesn't use reg
1867 // parms though amd does). Since the native abi doesn't use register args
1868 // and the java conventions does we don't have to worry about collisions.
1869 // All of our moved are reg->stack or stack->stack.
1870 // We ignore the extra arguments during the shuffle and handle them at the
1871 // last moment. The shuffle is described by the two calling convention
1872 // vectors we have in our possession. We simply walk the java vector to
1873 // get the source locations and the c vector to get the destinations.
1875 int c_arg = is_critical_native ? 0 : (method->is_static() ? 2 : 1 );
1877 // Record rsp-based slot for receiver on stack for non-static methods
1878 int receiver_offset = -1;
1880 // This is a trick. We double the stack slots so we can claim
1881 // the oops in the caller's frame. Since we are sure to have
1882 // more args than the caller doubling is enough to make
1883 // sure we can capture all the incoming oop args from the
1884 // caller.
1885 //
1886 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1888 // Mark location of rbp,
1889 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, rbp->as_VMReg());
1891 // We know that we only have args in at most two integer registers (rcx, rdx). So rax, rbx
1892 // Are free to temporaries if we have to do stack to steck moves.
1893 // All inbound args are referenced based on rbp, and all outbound args via rsp.
1895 for (int i = 0; i < total_in_args ; i++, c_arg++ ) {
1896 switch (in_sig_bt[i]) {
1897 case T_ARRAY:
1898 if (is_critical_native) {
1899 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]);
1900 c_arg++;
1901 break;
1902 }
1903 case T_OBJECT:
1904 assert(!is_critical_native, "no oop arguments");
1905 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1906 ((i == 0) && (!is_static)),
1907 &receiver_offset);
1908 break;
1909 case T_VOID:
1910 break;
1912 case T_FLOAT:
1913 float_move(masm, in_regs[i], out_regs[c_arg]);
1914 break;
1916 case T_DOUBLE:
1917 assert( i + 1 < total_in_args &&
1918 in_sig_bt[i + 1] == T_VOID &&
1919 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1920 double_move(masm, in_regs[i], out_regs[c_arg]);
1921 break;
1923 case T_LONG :
1924 long_move(masm, in_regs[i], out_regs[c_arg]);
1925 break;
1927 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1929 default:
1930 simple_move32(masm, in_regs[i], out_regs[c_arg]);
1931 }
1932 }
1934 // Pre-load a static method's oop into rsi. Used both by locking code and
1935 // the normal JNI call code.
1936 if (method->is_static() && !is_critical_native) {
1938 // load opp into a register
1939 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror()));
1941 // Now handlize the static class mirror it's known not-null.
1942 __ movptr(Address(rsp, klass_offset), oop_handle_reg);
1943 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1945 // Now get the handle
1946 __ lea(oop_handle_reg, Address(rsp, klass_offset));
1947 // store the klass handle as second argument
1948 __ movptr(Address(rsp, wordSize), oop_handle_reg);
1949 }
1951 // Change state to native (we save the return address in the thread, since it might not
1952 // be pushed on the stack when we do a a stack traversal). It is enough that the pc()
1953 // points into the right code segment. It does not have to be the correct return pc.
1954 // We use the same pc/oopMap repeatedly when we call out
1956 intptr_t the_pc = (intptr_t) __ pc();
1957 oop_maps->add_gc_map(the_pc - start, map);
1959 __ set_last_Java_frame(thread, rsp, noreg, (address)the_pc);
1962 // We have all of the arguments setup at this point. We must not touch any register
1963 // argument registers at this point (what if we save/restore them there are no oop?
1965 {
1966 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0);
1967 __ mov_metadata(rax, method());
1968 __ call_VM_leaf(
1969 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1970 thread, rax);
1971 }
1973 // RedefineClasses() tracing support for obsolete method entry
1974 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
1975 __ mov_metadata(rax, method());
1976 __ call_VM_leaf(
1977 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1978 thread, rax);
1979 }
1981 // These are register definitions we need for locking/unlocking
1982 const Register swap_reg = rax; // Must use rax, for cmpxchg instruction
1983 const Register obj_reg = rcx; // Will contain the oop
1984 const Register lock_reg = rdx; // Address of compiler lock object (BasicLock)
1986 Label slow_path_lock;
1987 Label lock_done;
1989 // Lock a synchronized method
1990 if (method->is_synchronized()) {
1991 assert(!is_critical_native, "unhandled");
1994 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1996 // Get the handle (the 2nd argument)
1997 __ movptr(oop_handle_reg, Address(rsp, wordSize));
1999 // Get address of the box
2001 __ lea(lock_reg, Address(rbp, lock_slot_rbp_offset));
2003 // Load the oop from the handle
2004 __ movptr(obj_reg, Address(oop_handle_reg, 0));
2006 if (UseBiasedLocking) {
2007 // Note that oop_handle_reg is trashed during this call
2008 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, oop_handle_reg, false, lock_done, &slow_path_lock);
2009 }
2011 // Load immediate 1 into swap_reg %rax,
2012 __ movptr(swap_reg, 1);
2014 // Load (object->mark() | 1) into swap_reg %rax,
2015 __ orptr(swap_reg, Address(obj_reg, 0));
2017 // Save (object->mark() | 1) into BasicLock's displaced header
2018 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2020 if (os::is_MP()) {
2021 __ lock();
2022 }
2024 // src -> dest iff dest == rax, else rax, <- dest
2025 // *obj_reg = lock_reg iff *obj_reg == rax, else rax, = *(obj_reg)
2026 __ cmpxchgptr(lock_reg, Address(obj_reg, 0));
2027 __ jcc(Assembler::equal, lock_done);
2029 // Test if the oopMark is an obvious stack pointer, i.e.,
2030 // 1) (mark & 3) == 0, and
2031 // 2) rsp <= mark < mark + os::pagesize()
2032 // These 3 tests can be done by evaluating the following
2033 // expression: ((mark - rsp) & (3 - os::vm_page_size())),
2034 // assuming both stack pointer and pagesize have their
2035 // least significant 2 bits clear.
2036 // NOTE: the oopMark is in swap_reg %rax, as the result of cmpxchg
2038 __ subptr(swap_reg, rsp);
2039 __ andptr(swap_reg, 3 - os::vm_page_size());
2041 // Save the test result, for recursive case, the result is zero
2042 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2043 __ jcc(Assembler::notEqual, slow_path_lock);
2044 // Slow path will re-enter here
2045 __ bind(lock_done);
2047 if (UseBiasedLocking) {
2048 // Re-fetch oop_handle_reg as we trashed it above
2049 __ movptr(oop_handle_reg, Address(rsp, wordSize));
2050 }
2051 }
2054 // Finally just about ready to make the JNI call
2057 // get JNIEnv* which is first argument to native
2058 if (!is_critical_native) {
2059 __ lea(rdx, Address(thread, in_bytes(JavaThread::jni_environment_offset())));
2060 __ movptr(Address(rsp, 0), rdx);
2061 }
2063 // Now set thread in native
2064 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native);
2066 __ call(RuntimeAddress(native_func));
2068 // Verify or restore cpu control state after JNI call
2069 __ restore_cpu_control_state_after_jni();
2071 // WARNING - on Windows Java Natives use pascal calling convention and pop the
2072 // arguments off of the stack. We could just re-adjust the stack pointer here
2073 // and continue to do SP relative addressing but we instead switch to FP
2074 // relative addressing.
2076 // Unpack native results.
2077 switch (ret_type) {
2078 case T_BOOLEAN: __ c2bool(rax); break;
2079 case T_CHAR : __ andptr(rax, 0xFFFF); break;
2080 case T_BYTE : __ sign_extend_byte (rax); break;
2081 case T_SHORT : __ sign_extend_short(rax); break;
2082 case T_INT : /* nothing to do */ break;
2083 case T_DOUBLE :
2084 case T_FLOAT :
2085 // Result is in st0 we'll save as needed
2086 break;
2087 case T_ARRAY: // Really a handle
2088 case T_OBJECT: // Really a handle
2089 break; // can't de-handlize until after safepoint check
2090 case T_VOID: break;
2091 case T_LONG: break;
2092 default : ShouldNotReachHere();
2093 }
2095 // Switch thread to "native transition" state before reading the synchronization state.
2096 // This additional state is necessary because reading and testing the synchronization
2097 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2098 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2099 // VM thread changes sync state to synchronizing and suspends threads for GC.
2100 // Thread A is resumed to finish this native method, but doesn't block here since it
2101 // didn't see any synchronization is progress, and escapes.
2102 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans);
2104 if(os::is_MP()) {
2105 if (UseMembar) {
2106 // Force this write out before the read below
2107 __ membar(Assembler::Membar_mask_bits(
2108 Assembler::LoadLoad | Assembler::LoadStore |
2109 Assembler::StoreLoad | Assembler::StoreStore));
2110 } else {
2111 // Write serialization page so VM thread can do a pseudo remote membar.
2112 // We use the current thread pointer to calculate a thread specific
2113 // offset to write to within the page. This minimizes bus traffic
2114 // due to cache line collision.
2115 __ serialize_memory(thread, rcx);
2116 }
2117 }
2119 if (AlwaysRestoreFPU) {
2120 // Make sure the control word is correct.
2121 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
2122 }
2124 Label after_transition;
2126 // check for safepoint operation in progress and/or pending suspend requests
2127 { Label Continue;
2129 __ cmp32(ExternalAddress((address)SafepointSynchronize::address_of_state()),
2130 SafepointSynchronize::_not_synchronized);
2132 Label L;
2133 __ jcc(Assembler::notEqual, L);
2134 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0);
2135 __ jcc(Assembler::equal, Continue);
2136 __ bind(L);
2138 // Don't use call_VM as it will see a possible pending exception and forward it
2139 // and never return here preventing us from clearing _last_native_pc down below.
2140 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are
2141 // preserved and correspond to the bcp/locals pointers. So we do a runtime call
2142 // by hand.
2143 //
2144 save_native_result(masm, ret_type, stack_slots);
2145 __ push(thread);
2146 if (!is_critical_native) {
2147 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address,
2148 JavaThread::check_special_condition_for_native_trans)));
2149 } else {
2150 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address,
2151 JavaThread::check_special_condition_for_native_trans_and_transition)));
2152 }
2153 __ increment(rsp, wordSize);
2154 // Restore any method result value
2155 restore_native_result(masm, ret_type, stack_slots);
2157 if (is_critical_native) {
2158 // The call above performed the transition to thread_in_Java so
2159 // skip the transition logic below.
2160 __ jmpb(after_transition);
2161 }
2163 __ bind(Continue);
2164 }
2166 // change thread state
2167 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java);
2168 __ bind(after_transition);
2170 Label reguard;
2171 Label reguard_done;
2172 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_disabled);
2173 __ jcc(Assembler::equal, reguard);
2175 // slow path reguard re-enters here
2176 __ bind(reguard_done);
2178 // Handle possible exception (will unlock if necessary)
2180 // native result if any is live
2182 // Unlock
2183 Label slow_path_unlock;
2184 Label unlock_done;
2185 if (method->is_synchronized()) {
2187 Label done;
2189 // Get locked oop from the handle we passed to jni
2190 __ movptr(obj_reg, Address(oop_handle_reg, 0));
2192 if (UseBiasedLocking) {
2193 __ biased_locking_exit(obj_reg, rbx, done);
2194 }
2196 // Simple recursive lock?
2198 __ cmpptr(Address(rbp, lock_slot_rbp_offset), (int32_t)NULL_WORD);
2199 __ jcc(Assembler::equal, done);
2201 // Must save rax, if if it is live now because cmpxchg must use it
2202 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2203 save_native_result(masm, ret_type, stack_slots);
2204 }
2206 // get old displaced header
2207 __ movptr(rbx, Address(rbp, lock_slot_rbp_offset));
2209 // get address of the stack lock
2210 __ lea(rax, Address(rbp, lock_slot_rbp_offset));
2212 // Atomic swap old header if oop still contains the stack lock
2213 if (os::is_MP()) {
2214 __ lock();
2215 }
2217 // src -> dest iff dest == rax, else rax, <- dest
2218 // *obj_reg = rbx, iff *obj_reg == rax, else rax, = *(obj_reg)
2219 __ cmpxchgptr(rbx, Address(obj_reg, 0));
2220 __ jcc(Assembler::notEqual, slow_path_unlock);
2222 // slow path re-enters here
2223 __ bind(unlock_done);
2224 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2225 restore_native_result(masm, ret_type, stack_slots);
2226 }
2228 __ bind(done);
2230 }
2232 {
2233 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0);
2234 // Tell dtrace about this method exit
2235 save_native_result(masm, ret_type, stack_slots);
2236 __ mov_metadata(rax, method());
2237 __ call_VM_leaf(
2238 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2239 thread, rax);
2240 restore_native_result(masm, ret_type, stack_slots);
2241 }
2243 // We can finally stop using that last_Java_frame we setup ages ago
2245 __ reset_last_Java_frame(thread, false, true);
2247 // Unpack oop result
2248 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2249 Label L;
2250 __ cmpptr(rax, (int32_t)NULL_WORD);
2251 __ jcc(Assembler::equal, L);
2252 __ movptr(rax, Address(rax, 0));
2253 __ bind(L);
2254 __ verify_oop(rax);
2255 }
2257 if (!is_critical_native) {
2258 // reset handle block
2259 __ movptr(rcx, Address(thread, JavaThread::active_handles_offset()));
2260 __ movptr(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), NULL_WORD);
2262 // Any exception pending?
2263 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
2264 __ jcc(Assembler::notEqual, exception_pending);
2265 }
2267 // no exception, we're almost done
2269 // check that only result value is on FPU stack
2270 __ verify_FPU(ret_type == T_FLOAT || ret_type == T_DOUBLE ? 1 : 0, "native_wrapper normal exit");
2272 // Fixup floating pointer results so that result looks like a return from a compiled method
2273 if (ret_type == T_FLOAT) {
2274 if (UseSSE >= 1) {
2275 // Pop st0 and store as float and reload into xmm register
2276 __ fstp_s(Address(rbp, -4));
2277 __ movflt(xmm0, Address(rbp, -4));
2278 }
2279 } else if (ret_type == T_DOUBLE) {
2280 if (UseSSE >= 2) {
2281 // Pop st0 and store as double and reload into xmm register
2282 __ fstp_d(Address(rbp, -8));
2283 __ movdbl(xmm0, Address(rbp, -8));
2284 }
2285 }
2287 // Return
2289 __ leave();
2290 __ ret(0);
2292 // Unexpected paths are out of line and go here
2294 // Slow path locking & unlocking
2295 if (method->is_synchronized()) {
2297 // BEGIN Slow path lock
2299 __ bind(slow_path_lock);
2301 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2302 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2303 __ push(thread);
2304 __ push(lock_reg);
2305 __ push(obj_reg);
2306 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C)));
2307 __ addptr(rsp, 3*wordSize);
2309 #ifdef ASSERT
2310 { Label L;
2311 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD);
2312 __ jcc(Assembler::equal, L);
2313 __ stop("no pending exception allowed on exit from monitorenter");
2314 __ bind(L);
2315 }
2316 #endif
2317 __ jmp(lock_done);
2319 // END Slow path lock
2321 // BEGIN Slow path unlock
2322 __ bind(slow_path_unlock);
2324 // Slow path unlock
2326 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2327 save_native_result(masm, ret_type, stack_slots);
2328 }
2329 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2331 __ pushptr(Address(thread, in_bytes(Thread::pending_exception_offset())));
2332 __ movptr(Address(thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD);
2335 // should be a peal
2336 // +wordSize because of the push above
2337 __ lea(rax, Address(rbp, lock_slot_rbp_offset));
2338 __ push(rax);
2340 __ push(obj_reg);
2341 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)));
2342 __ addptr(rsp, 2*wordSize);
2343 #ifdef ASSERT
2344 {
2345 Label L;
2346 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
2347 __ jcc(Assembler::equal, L);
2348 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2349 __ bind(L);
2350 }
2351 #endif /* ASSERT */
2353 __ popptr(Address(thread, in_bytes(Thread::pending_exception_offset())));
2355 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2356 restore_native_result(masm, ret_type, stack_slots);
2357 }
2358 __ jmp(unlock_done);
2359 // END Slow path unlock
2361 }
2363 // SLOW PATH Reguard the stack if needed
2365 __ bind(reguard);
2366 save_native_result(masm, ret_type, stack_slots);
2367 {
2368 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)));
2369 }
2370 restore_native_result(masm, ret_type, stack_slots);
2371 __ jmp(reguard_done);
2374 // BEGIN EXCEPTION PROCESSING
2376 if (!is_critical_native) {
2377 // Forward the exception
2378 __ bind(exception_pending);
2380 // remove possible return value from FPU register stack
2381 __ empty_FPU_stack();
2383 // pop our frame
2384 __ leave();
2385 // and forward the exception
2386 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2387 }
2389 __ flush();
2391 nmethod *nm = nmethod::new_native_nmethod(method,
2392 compile_id,
2393 masm->code(),
2394 vep_offset,
2395 frame_complete,
2396 stack_slots / VMRegImpl::slots_per_word,
2397 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2398 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2399 oop_maps);
2401 if (is_critical_native) {
2402 nm->set_lazy_critical_native(true);
2403 }
2405 return nm;
2407 }
2409 #ifdef HAVE_DTRACE_H
2410 // ---------------------------------------------------------------------------
2411 // Generate a dtrace nmethod for a given signature. The method takes arguments
2412 // in the Java compiled code convention, marshals them to the native
2413 // abi and then leaves nops at the position you would expect to call a native
2414 // function. When the probe is enabled the nops are replaced with a trap
2415 // instruction that dtrace inserts and the trace will cause a notification
2416 // to dtrace.
2417 //
2418 // The probes are only able to take primitive types and java/lang/String as
2419 // arguments. No other java types are allowed. Strings are converted to utf8
2420 // strings so that from dtrace point of view java strings are converted to C
2421 // strings. There is an arbitrary fixed limit on the total space that a method
2422 // can use for converting the strings. (256 chars per string in the signature).
2423 // So any java string larger then this is truncated.
2425 nmethod *SharedRuntime::generate_dtrace_nmethod(
2426 MacroAssembler *masm, methodHandle method) {
2428 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
2429 // be single threaded in this method.
2430 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
2432 // Fill in the signature array, for the calling-convention call.
2433 int total_args_passed = method->size_of_parameters();
2435 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
2436 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
2438 // The signature we are going to use for the trap that dtrace will see
2439 // java/lang/String is converted. We drop "this" and any other object
2440 // is converted to NULL. (A one-slot java/lang/Long object reference
2441 // is converted to a two-slot long, which is why we double the allocation).
2442 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
2443 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
2445 int i=0;
2446 int total_strings = 0;
2447 int first_arg_to_pass = 0;
2448 int total_c_args = 0;
2450 if( !method->is_static() ) { // Pass in receiver first
2451 in_sig_bt[i++] = T_OBJECT;
2452 first_arg_to_pass = 1;
2453 }
2455 // We need to convert the java args to where a native (non-jni) function
2456 // would expect them. To figure out where they go we convert the java
2457 // signature to a C signature.
2459 SignatureStream ss(method->signature());
2460 for ( ; !ss.at_return_type(); ss.next()) {
2461 BasicType bt = ss.type();
2462 in_sig_bt[i++] = bt; // Collect remaining bits of signature
2463 out_sig_bt[total_c_args++] = bt;
2464 if( bt == T_OBJECT) {
2465 Symbol* s = ss.as_symbol_or_null(); // symbol is created
2466 if (s == vmSymbols::java_lang_String()) {
2467 total_strings++;
2468 out_sig_bt[total_c_args-1] = T_ADDRESS;
2469 } else if (s == vmSymbols::java_lang_Boolean() ||
2470 s == vmSymbols::java_lang_Character() ||
2471 s == vmSymbols::java_lang_Byte() ||
2472 s == vmSymbols::java_lang_Short() ||
2473 s == vmSymbols::java_lang_Integer() ||
2474 s == vmSymbols::java_lang_Float()) {
2475 out_sig_bt[total_c_args-1] = T_INT;
2476 } else if (s == vmSymbols::java_lang_Long() ||
2477 s == vmSymbols::java_lang_Double()) {
2478 out_sig_bt[total_c_args-1] = T_LONG;
2479 out_sig_bt[total_c_args++] = T_VOID;
2480 }
2481 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
2482 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
2483 out_sig_bt[total_c_args++] = T_VOID;
2484 }
2485 }
2487 assert(i==total_args_passed, "validly parsed signature");
2489 // Now get the compiled-Java layout as input arguments
2490 int comp_args_on_stack;
2491 comp_args_on_stack = SharedRuntime::java_calling_convention(
2492 in_sig_bt, in_regs, total_args_passed, false);
2494 // Now figure out where the args must be stored and how much stack space
2495 // they require (neglecting out_preserve_stack_slots).
2497 int out_arg_slots;
2498 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2500 // Calculate the total number of stack slots we will need.
2502 // First count the abi requirement plus all of the outgoing args
2503 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2505 // Now space for the string(s) we must convert
2507 int* string_locs = NEW_RESOURCE_ARRAY(int, total_strings + 1);
2508 for (i = 0; i < total_strings ; i++) {
2509 string_locs[i] = stack_slots;
2510 stack_slots += max_dtrace_string_size / VMRegImpl::stack_slot_size;
2511 }
2513 // + 2 for return address (which we own) and saved rbp,
2515 stack_slots += 2;
2517 // Ok The space we have allocated will look like:
2518 //
2519 //
2520 // FP-> | |
2521 // |---------------------|
2522 // | string[n] |
2523 // |---------------------| <- string_locs[n]
2524 // | string[n-1] |
2525 // |---------------------| <- string_locs[n-1]
2526 // | ... |
2527 // | ... |
2528 // |---------------------| <- string_locs[1]
2529 // | string[0] |
2530 // |---------------------| <- string_locs[0]
2531 // | outbound memory |
2532 // | based arguments |
2533 // | |
2534 // |---------------------|
2535 // | |
2536 // SP-> | out_preserved_slots |
2537 //
2538 //
2540 // Now compute actual number of stack words we need rounding to make
2541 // stack properly aligned.
2542 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
2544 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2546 intptr_t start = (intptr_t)__ pc();
2548 // First thing make an ic check to see if we should even be here
2550 // We are free to use all registers as temps without saving them and
2551 // restoring them except rbp. rbp, is the only callee save register
2552 // as far as the interpreter and the compiler(s) are concerned.
2554 const Register ic_reg = rax;
2555 const Register receiver = rcx;
2556 Label hit;
2557 Label exception_pending;
2560 __ verify_oop(receiver);
2561 __ cmpl(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes()));
2562 __ jcc(Assembler::equal, hit);
2564 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
2566 // verified entry must be aligned for code patching.
2567 // and the first 5 bytes must be in the same cache line
2568 // if we align at 8 then we will be sure 5 bytes are in the same line
2569 __ align(8);
2571 __ bind(hit);
2573 int vep_offset = ((intptr_t)__ pc()) - start;
2576 // The instruction at the verified entry point must be 5 bytes or longer
2577 // because it can be patched on the fly by make_non_entrant. The stack bang
2578 // instruction fits that requirement.
2580 // Generate stack overflow check
2583 if (UseStackBanging) {
2584 if (stack_size <= StackShadowPages*os::vm_page_size()) {
2585 __ bang_stack_with_offset(StackShadowPages*os::vm_page_size());
2586 } else {
2587 __ movl(rax, stack_size);
2588 __ bang_stack_size(rax, rbx);
2589 }
2590 } else {
2591 // need a 5 byte instruction to allow MT safe patching to non-entrant
2592 __ fat_nop();
2593 }
2595 assert(((int)__ pc() - start - vep_offset) >= 5,
2596 "valid size for make_non_entrant");
2598 // Generate a new frame for the wrapper.
2599 __ enter();
2601 // -2 because return address is already present and so is saved rbp,
2602 if (stack_size - 2*wordSize != 0) {
2603 __ subl(rsp, stack_size - 2*wordSize);
2604 }
2606 // Frame is now completed as far a size and linkage.
2608 int frame_complete = ((intptr_t)__ pc()) - start;
2610 // First thing we do store all the args as if we are doing the call.
2611 // Since the C calling convention is stack based that ensures that
2612 // all the Java register args are stored before we need to convert any
2613 // string we might have.
2615 int sid = 0;
2616 int c_arg, j_arg;
2617 int string_reg = 0;
2619 for (j_arg = first_arg_to_pass, c_arg = 0 ;
2620 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
2622 VMRegPair src = in_regs[j_arg];
2623 VMRegPair dst = out_regs[c_arg];
2624 assert(dst.first()->is_stack() || in_sig_bt[j_arg] == T_VOID,
2625 "stack based abi assumed");
2627 switch (in_sig_bt[j_arg]) {
2629 case T_ARRAY:
2630 case T_OBJECT:
2631 if (out_sig_bt[c_arg] == T_ADDRESS) {
2632 // Any register based arg for a java string after the first
2633 // will be destroyed by the call to get_utf so we store
2634 // the original value in the location the utf string address
2635 // will eventually be stored.
2636 if (src.first()->is_reg()) {
2637 if (string_reg++ != 0) {
2638 simple_move32(masm, src, dst);
2639 }
2640 }
2641 } else if (out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
2642 // need to unbox a one-word value
2643 Register in_reg = rax;
2644 if ( src.first()->is_reg() ) {
2645 in_reg = src.first()->as_Register();
2646 } else {
2647 simple_move32(masm, src, in_reg->as_VMReg());
2648 }
2649 Label skipUnbox;
2650 __ movl(Address(rsp, reg2offset_out(dst.first())), NULL_WORD);
2651 if ( out_sig_bt[c_arg] == T_LONG ) {
2652 __ movl(Address(rsp, reg2offset_out(dst.second())), NULL_WORD);
2653 }
2654 __ testl(in_reg, in_reg);
2655 __ jcc(Assembler::zero, skipUnbox);
2656 assert(dst.first()->is_stack() &&
2657 (!dst.second()->is_valid() || dst.second()->is_stack()),
2658 "value(s) must go into stack slots");
2660 BasicType bt = out_sig_bt[c_arg];
2661 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
2662 if ( bt == T_LONG ) {
2663 __ movl(rbx, Address(in_reg,
2664 box_offset + VMRegImpl::stack_slot_size));
2665 __ movl(Address(rsp, reg2offset_out(dst.second())), rbx);
2666 }
2667 __ movl(in_reg, Address(in_reg, box_offset));
2668 __ movl(Address(rsp, reg2offset_out(dst.first())), in_reg);
2669 __ bind(skipUnbox);
2670 } else {
2671 // Convert the arg to NULL
2672 __ movl(Address(rsp, reg2offset_out(dst.first())), NULL_WORD);
2673 }
2674 if (out_sig_bt[c_arg] == T_LONG) {
2675 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
2676 ++c_arg; // Move over the T_VOID To keep the loop indices in sync
2677 }
2678 break;
2680 case T_VOID:
2681 break;
2683 case T_FLOAT:
2684 float_move(masm, src, dst);
2685 break;
2687 case T_DOUBLE:
2688 assert( j_arg + 1 < total_args_passed &&
2689 in_sig_bt[j_arg + 1] == T_VOID, "bad arg list");
2690 double_move(masm, src, dst);
2691 break;
2693 case T_LONG :
2694 long_move(masm, src, dst);
2695 break;
2697 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2699 default:
2700 simple_move32(masm, src, dst);
2701 }
2702 }
2704 // Now we must convert any string we have to utf8
2705 //
2707 for (sid = 0, j_arg = first_arg_to_pass, c_arg = 0 ;
2708 sid < total_strings ; j_arg++, c_arg++ ) {
2710 if (out_sig_bt[c_arg] == T_ADDRESS) {
2712 Address utf8_addr = Address(
2713 rsp, string_locs[sid++] * VMRegImpl::stack_slot_size);
2714 __ leal(rax, utf8_addr);
2716 // The first string we find might still be in the original java arg
2717 // register
2718 VMReg orig_loc = in_regs[j_arg].first();
2719 Register string_oop;
2721 // This is where the argument will eventually reside
2722 Address dest = Address(rsp, reg2offset_out(out_regs[c_arg].first()));
2724 if (sid == 1 && orig_loc->is_reg()) {
2725 string_oop = orig_loc->as_Register();
2726 assert(string_oop != rax, "smashed arg");
2727 } else {
2729 if (orig_loc->is_reg()) {
2730 // Get the copy of the jls object
2731 __ movl(rcx, dest);
2732 } else {
2733 // arg is still in the original location
2734 __ movl(rcx, Address(rbp, reg2offset_in(orig_loc)));
2735 }
2736 string_oop = rcx;
2738 }
2739 Label nullString;
2740 __ movl(dest, NULL_WORD);
2741 __ testl(string_oop, string_oop);
2742 __ jcc(Assembler::zero, nullString);
2744 // Now we can store the address of the utf string as the argument
2745 __ movl(dest, rax);
2747 // And do the conversion
2748 __ call_VM_leaf(CAST_FROM_FN_PTR(
2749 address, SharedRuntime::get_utf), string_oop, rax);
2750 __ bind(nullString);
2751 }
2753 if (in_sig_bt[j_arg] == T_OBJECT && out_sig_bt[c_arg] == T_LONG) {
2754 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
2755 ++c_arg; // Move over the T_VOID To keep the loop indices in sync
2756 }
2757 }
2760 // Ok now we are done. Need to place the nop that dtrace wants in order to
2761 // patch in the trap
2763 int patch_offset = ((intptr_t)__ pc()) - start;
2765 __ nop();
2768 // Return
2770 __ leave();
2771 __ ret(0);
2773 __ flush();
2775 nmethod *nm = nmethod::new_dtrace_nmethod(
2776 method, masm->code(), vep_offset, patch_offset, frame_complete,
2777 stack_slots / VMRegImpl::slots_per_word);
2778 return nm;
2780 }
2782 #endif // HAVE_DTRACE_H
2784 // this function returns the adjust size (in number of words) to a c2i adapter
2785 // activation for use during deoptimization
2786 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) {
2787 return (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2788 }
2791 uint SharedRuntime::out_preserve_stack_slots() {
2792 return 0;
2793 }
2795 //------------------------------generate_deopt_blob----------------------------
2796 void SharedRuntime::generate_deopt_blob() {
2797 // allocate space for the code
2798 ResourceMark rm;
2799 // setup code generation tools
2800 CodeBuffer buffer("deopt_blob", 1024, 1024);
2801 MacroAssembler* masm = new MacroAssembler(&buffer);
2802 int frame_size_in_words;
2803 OopMap* map = NULL;
2804 // Account for the extra args we place on the stack
2805 // by the time we call fetch_unroll_info
2806 const int additional_words = 2; // deopt kind, thread
2808 OopMapSet *oop_maps = new OopMapSet();
2810 // -------------
2811 // This code enters when returning to a de-optimized nmethod. A return
2812 // address has been pushed on the the stack, and return values are in
2813 // registers.
2814 // If we are doing a normal deopt then we were called from the patched
2815 // nmethod from the point we returned to the nmethod. So the return
2816 // address on the stack is wrong by NativeCall::instruction_size
2817 // We will adjust the value to it looks like we have the original return
2818 // address on the stack (like when we eagerly deoptimized).
2819 // In the case of an exception pending with deoptimized then we enter
2820 // with a return address on the stack that points after the call we patched
2821 // into the exception handler. We have the following register state:
2822 // rax,: exception
2823 // rbx,: exception handler
2824 // rdx: throwing pc
2825 // So in this case we simply jam rdx into the useless return address and
2826 // the stack looks just like we want.
2827 //
2828 // At this point we need to de-opt. We save the argument return
2829 // registers. We call the first C routine, fetch_unroll_info(). This
2830 // routine captures the return values and returns a structure which
2831 // describes the current frame size and the sizes of all replacement frames.
2832 // The current frame is compiled code and may contain many inlined
2833 // functions, each with their own JVM state. We pop the current frame, then
2834 // push all the new frames. Then we call the C routine unpack_frames() to
2835 // populate these frames. Finally unpack_frames() returns us the new target
2836 // address. Notice that callee-save registers are BLOWN here; they have
2837 // already been captured in the vframeArray at the time the return PC was
2838 // patched.
2839 address start = __ pc();
2840 Label cont;
2842 // Prolog for non exception case!
2844 // Save everything in sight.
2846 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2847 // Normal deoptimization
2848 __ push(Deoptimization::Unpack_deopt);
2849 __ jmp(cont);
2851 int reexecute_offset = __ pc() - start;
2853 // Reexecute case
2854 // return address is the pc describes what bci to do re-execute at
2856 // No need to update map as each call to save_live_registers will produce identical oopmap
2857 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2859 __ push(Deoptimization::Unpack_reexecute);
2860 __ jmp(cont);
2862 int exception_offset = __ pc() - start;
2864 // Prolog for exception case
2866 // all registers are dead at this entry point, except for rax, and
2867 // rdx which contain the exception oop and exception pc
2868 // respectively. Set them in TLS and fall thru to the
2869 // unpack_with_exception_in_tls entry point.
2871 __ get_thread(rdi);
2872 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), rdx);
2873 __ movptr(Address(rdi, JavaThread::exception_oop_offset()), rax);
2875 int exception_in_tls_offset = __ pc() - start;
2877 // new implementation because exception oop is now passed in JavaThread
2879 // Prolog for exception case
2880 // All registers must be preserved because they might be used by LinearScan
2881 // Exceptiop oop and throwing PC are passed in JavaThread
2882 // tos: stack at point of call to method that threw the exception (i.e. only
2883 // args are on the stack, no return address)
2885 // make room on stack for the return address
2886 // It will be patched later with the throwing pc. The correct value is not
2887 // available now because loading it from memory would destroy registers.
2888 __ push(0);
2890 // Save everything in sight.
2892 // No need to update map as each call to save_live_registers will produce identical oopmap
2893 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2895 // Now it is safe to overwrite any register
2897 // store the correct deoptimization type
2898 __ push(Deoptimization::Unpack_exception);
2900 // load throwing pc from JavaThread and patch it as the return address
2901 // of the current frame. Then clear the field in JavaThread
2902 __ get_thread(rdi);
2903 __ movptr(rdx, Address(rdi, JavaThread::exception_pc_offset()));
2904 __ movptr(Address(rbp, wordSize), rdx);
2905 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), NULL_WORD);
2907 #ifdef ASSERT
2908 // verify that there is really an exception oop in JavaThread
2909 __ movptr(rax, Address(rdi, JavaThread::exception_oop_offset()));
2910 __ verify_oop(rax);
2912 // verify that there is no pending exception
2913 Label no_pending_exception;
2914 __ movptr(rax, Address(rdi, Thread::pending_exception_offset()));
2915 __ testptr(rax, rax);
2916 __ jcc(Assembler::zero, no_pending_exception);
2917 __ stop("must not have pending exception here");
2918 __ bind(no_pending_exception);
2919 #endif
2921 __ bind(cont);
2923 // Compiled code leaves the floating point stack dirty, empty it.
2924 __ empty_FPU_stack();
2927 // Call C code. Need thread and this frame, but NOT official VM entry
2928 // crud. We cannot block on this call, no GC can happen.
2929 __ get_thread(rcx);
2930 __ push(rcx);
2931 // fetch_unroll_info needs to call last_java_frame()
2932 __ set_last_Java_frame(rcx, noreg, noreg, NULL);
2934 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2936 // Need to have an oopmap that tells fetch_unroll_info where to
2937 // find any register it might need.
2939 oop_maps->add_gc_map( __ pc()-start, map);
2941 // Discard arg to fetch_unroll_info
2942 __ pop(rcx);
2944 __ get_thread(rcx);
2945 __ reset_last_Java_frame(rcx, false, false);
2947 // Load UnrollBlock into EDI
2948 __ mov(rdi, rax);
2950 // Move the unpack kind to a safe place in the UnrollBlock because
2951 // we are very short of registers
2953 Address unpack_kind(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes());
2954 // retrieve the deopt kind from where we left it.
2955 __ pop(rax);
2956 __ movl(unpack_kind, rax); // save the unpack_kind value
2958 Label noException;
2959 __ cmpl(rax, Deoptimization::Unpack_exception); // Was exception pending?
2960 __ jcc(Assembler::notEqual, noException);
2961 __ movptr(rax, Address(rcx, JavaThread::exception_oop_offset()));
2962 __ movptr(rdx, Address(rcx, JavaThread::exception_pc_offset()));
2963 __ movptr(Address(rcx, JavaThread::exception_oop_offset()), NULL_WORD);
2964 __ movptr(Address(rcx, JavaThread::exception_pc_offset()), NULL_WORD);
2966 __ verify_oop(rax);
2968 // Overwrite the result registers with the exception results.
2969 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax);
2970 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx);
2972 __ bind(noException);
2974 // Stack is back to only having register save data on the stack.
2975 // Now restore the result registers. Everything else is either dead or captured
2976 // in the vframeArray.
2978 RegisterSaver::restore_result_registers(masm);
2980 // Non standard control word may be leaked out through a safepoint blob, and we can
2981 // deopt at a poll point with the non standard control word. However, we should make
2982 // sure the control word is correct after restore_result_registers.
2983 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
2985 // All of the register save area has been popped of the stack. Only the
2986 // return address remains.
2988 // Pop all the frames we must move/replace.
2989 //
2990 // Frame picture (youngest to oldest)
2991 // 1: self-frame (no frame link)
2992 // 2: deopting frame (no frame link)
2993 // 3: caller of deopting frame (could be compiled/interpreted).
2994 //
2995 // Note: by leaving the return address of self-frame on the stack
2996 // and using the size of frame 2 to adjust the stack
2997 // when we are done the return to frame 3 will still be on the stack.
2999 // Pop deoptimized frame
3000 __ addptr(rsp, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes()));
3002 // sp should be pointing at the return address to the caller (3)
3004 // Stack bang to make sure there's enough room for these interpreter frames.
3005 if (UseStackBanging) {
3006 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()));
3007 __ bang_stack_size(rbx, rcx);
3008 }
3010 // Load array of frame pcs into ECX
3011 __ movptr(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
3013 __ pop(rsi); // trash the old pc
3015 // Load array of frame sizes into ESI
3016 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
3018 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes());
3020 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
3021 __ movl(counter, rbx);
3023 // Pick up the initial fp we should save
3024 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes()));
3026 // Now adjust the caller's stack to make up for the extra locals
3027 // but record the original sp so that we can save it in the skeletal interpreter
3028 // frame and the stack walking of interpreter_sender will get the unextended sp
3029 // value and not the "real" sp value.
3031 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes());
3032 __ movptr(sp_temp, rsp);
3033 __ movl2ptr(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
3034 __ subptr(rsp, rbx);
3036 // Push interpreter frames in a loop
3037 Label loop;
3038 __ bind(loop);
3039 __ movptr(rbx, Address(rsi, 0)); // Load frame size
3040 #ifdef CC_INTERP
3041 __ subptr(rbx, 4*wordSize); // we'll push pc and ebp by hand and
3042 #ifdef ASSERT
3043 __ push(0xDEADDEAD); // Make a recognizable pattern
3044 __ push(0xDEADDEAD);
3045 #else /* ASSERT */
3046 __ subptr(rsp, 2*wordSize); // skip the "static long no_param"
3047 #endif /* ASSERT */
3048 #else /* CC_INTERP */
3049 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand
3050 #endif /* CC_INTERP */
3051 __ pushptr(Address(rcx, 0)); // save return address
3052 __ enter(); // save old & set new rbp,
3053 __ subptr(rsp, rbx); // Prolog!
3054 __ movptr(rbx, sp_temp); // sender's sp
3055 #ifdef CC_INTERP
3056 __ movptr(Address(rbp,
3057 -(sizeof(BytecodeInterpreter)) + in_bytes(byte_offset_of(BytecodeInterpreter, _sender_sp))),
3058 rbx); // Make it walkable
3059 #else /* CC_INTERP */
3060 // This value is corrected by layout_activation_impl
3061 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD);
3062 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable
3063 #endif /* CC_INTERP */
3064 __ movptr(sp_temp, rsp); // pass to next frame
3065 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
3066 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
3067 __ decrementl(counter); // decrement counter
3068 __ jcc(Assembler::notZero, loop);
3069 __ pushptr(Address(rcx, 0)); // save final return address
3071 // Re-push self-frame
3072 __ enter(); // save old & set new rbp,
3074 // Return address and rbp, are in place
3075 // We'll push additional args later. Just allocate a full sized
3076 // register save area
3077 __ subptr(rsp, (frame_size_in_words-additional_words - 2) * wordSize);
3079 // Restore frame locals after moving the frame
3080 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax);
3081 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx);
3082 __ fstp_d(Address(rsp, RegisterSaver::fpResultOffset()*wordSize)); // Pop float stack and store in local
3083 if( UseSSE>=2 ) __ movdbl(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0);
3084 if( UseSSE==1 ) __ movflt(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0);
3086 // Set up the args to unpack_frame
3088 __ pushl(unpack_kind); // get the unpack_kind value
3089 __ get_thread(rcx);
3090 __ push(rcx);
3092 // set last_Java_sp, last_Java_fp
3093 __ set_last_Java_frame(rcx, noreg, rbp, NULL);
3095 // Call C code. Need thread but NOT official VM entry
3096 // crud. We cannot block on this call, no GC can happen. Call should
3097 // restore return values to their stack-slots with the new SP.
3098 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
3099 // Set an oopmap for the call site
3100 oop_maps->add_gc_map( __ pc()-start, new OopMap( frame_size_in_words, 0 ));
3102 // rax, contains the return result type
3103 __ push(rax);
3105 __ get_thread(rcx);
3106 __ reset_last_Java_frame(rcx, false, false);
3108 // Collect return values
3109 __ movptr(rax,Address(rsp, (RegisterSaver::raxOffset() + additional_words + 1)*wordSize));
3110 __ movptr(rdx,Address(rsp, (RegisterSaver::rdxOffset() + additional_words + 1)*wordSize));
3112 // Clear floating point stack before returning to interpreter
3113 __ empty_FPU_stack();
3115 // Check if we should push the float or double return value.
3116 Label results_done, yes_double_value;
3117 __ cmpl(Address(rsp, 0), T_DOUBLE);
3118 __ jcc (Assembler::zero, yes_double_value);
3119 __ cmpl(Address(rsp, 0), T_FLOAT);
3120 __ jcc (Assembler::notZero, results_done);
3122 // return float value as expected by interpreter
3123 if( UseSSE>=1 ) __ movflt(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize));
3124 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize));
3125 __ jmp(results_done);
3127 // return double value as expected by interpreter
3128 __ bind(yes_double_value);
3129 if( UseSSE>=2 ) __ movdbl(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize));
3130 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize));
3132 __ bind(results_done);
3134 // Pop self-frame.
3135 __ leave(); // Epilog!
3137 // Jump to interpreter
3138 __ ret(0);
3140 // -------------
3141 // make sure all code is generated
3142 masm->flush();
3144 _deopt_blob = DeoptimizationBlob::create( &buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
3145 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3146 }
3149 #ifdef COMPILER2
3150 //------------------------------generate_uncommon_trap_blob--------------------
3151 void SharedRuntime::generate_uncommon_trap_blob() {
3152 // allocate space for the code
3153 ResourceMark rm;
3154 // setup code generation tools
3155 CodeBuffer buffer("uncommon_trap_blob", 512, 512);
3156 MacroAssembler* masm = new MacroAssembler(&buffer);
3158 enum frame_layout {
3159 arg0_off, // thread sp + 0 // Arg location for
3160 arg1_off, // unloaded_class_index sp + 1 // calling C
3161 // The frame sender code expects that rbp will be in the "natural" place and
3162 // will override any oopMap setting for it. We must therefore force the layout
3163 // so that it agrees with the frame sender code.
3164 rbp_off, // callee saved register sp + 2
3165 return_off, // slot for return address sp + 3
3166 framesize
3167 };
3169 address start = __ pc();
3170 // Push self-frame.
3171 __ subptr(rsp, return_off*wordSize); // Epilog!
3173 // rbp, is an implicitly saved callee saved register (i.e. the calling
3174 // convention will save restore it in prolog/epilog) Other than that
3175 // there are no callee save registers no that adapter frames are gone.
3176 __ movptr(Address(rsp, rbp_off*wordSize), rbp);
3178 // Clear the floating point exception stack
3179 __ empty_FPU_stack();
3181 // set last_Java_sp
3182 __ get_thread(rdx);
3183 __ set_last_Java_frame(rdx, noreg, noreg, NULL);
3185 // Call C code. Need thread but NOT official VM entry
3186 // crud. We cannot block on this call, no GC can happen. Call should
3187 // capture callee-saved registers as well as return values.
3188 __ movptr(Address(rsp, arg0_off*wordSize), rdx);
3189 // argument already in ECX
3190 __ movl(Address(rsp, arg1_off*wordSize),rcx);
3191 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap)));
3193 // Set an oopmap for the call site
3194 OopMapSet *oop_maps = new OopMapSet();
3195 OopMap* map = new OopMap( framesize, 0 );
3196 // No oopMap for rbp, it is known implicitly
3198 oop_maps->add_gc_map( __ pc()-start, map);
3200 __ get_thread(rcx);
3202 __ reset_last_Java_frame(rcx, false, false);
3204 // Load UnrollBlock into EDI
3205 __ movptr(rdi, rax);
3207 // Pop all the frames we must move/replace.
3208 //
3209 // Frame picture (youngest to oldest)
3210 // 1: self-frame (no frame link)
3211 // 2: deopting frame (no frame link)
3212 // 3: caller of deopting frame (could be compiled/interpreted).
3214 // Pop self-frame. We have no frame, and must rely only on EAX and ESP.
3215 __ addptr(rsp,(framesize-1)*wordSize); // Epilog!
3217 // Pop deoptimized frame
3218 __ movl2ptr(rcx, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes()));
3219 __ addptr(rsp, rcx);
3221 // sp should be pointing at the return address to the caller (3)
3223 // Stack bang to make sure there's enough room for these interpreter frames.
3224 if (UseStackBanging) {
3225 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()));
3226 __ bang_stack_size(rbx, rcx);
3227 }
3230 // Load array of frame pcs into ECX
3231 __ movl(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
3233 __ pop(rsi); // trash the pc
3235 // Load array of frame sizes into ESI
3236 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
3238 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes());
3240 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
3241 __ movl(counter, rbx);
3243 // Pick up the initial fp we should save
3244 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes()));
3246 // Now adjust the caller's stack to make up for the extra locals
3247 // but record the original sp so that we can save it in the skeletal interpreter
3248 // frame and the stack walking of interpreter_sender will get the unextended sp
3249 // value and not the "real" sp value.
3251 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes());
3252 __ movptr(sp_temp, rsp);
3253 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
3254 __ subptr(rsp, rbx);
3256 // Push interpreter frames in a loop
3257 Label loop;
3258 __ bind(loop);
3259 __ movptr(rbx, Address(rsi, 0)); // Load frame size
3260 #ifdef CC_INTERP
3261 __ subptr(rbx, 4*wordSize); // we'll push pc and ebp by hand and
3262 #ifdef ASSERT
3263 __ push(0xDEADDEAD); // Make a recognizable pattern
3264 __ push(0xDEADDEAD); // (parm to RecursiveInterpreter...)
3265 #else /* ASSERT */
3266 __ subptr(rsp, 2*wordSize); // skip the "static long no_param"
3267 #endif /* ASSERT */
3268 #else /* CC_INTERP */
3269 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand
3270 #endif /* CC_INTERP */
3271 __ pushptr(Address(rcx, 0)); // save return address
3272 __ enter(); // save old & set new rbp,
3273 __ subptr(rsp, rbx); // Prolog!
3274 __ movptr(rbx, sp_temp); // sender's sp
3275 #ifdef CC_INTERP
3276 __ movptr(Address(rbp,
3277 -(sizeof(BytecodeInterpreter)) + in_bytes(byte_offset_of(BytecodeInterpreter, _sender_sp))),
3278 rbx); // Make it walkable
3279 #else /* CC_INTERP */
3280 // This value is corrected by layout_activation_impl
3281 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD );
3282 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable
3283 #endif /* CC_INTERP */
3284 __ movptr(sp_temp, rsp); // pass to next frame
3285 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
3286 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
3287 __ decrementl(counter); // decrement counter
3288 __ jcc(Assembler::notZero, loop);
3289 __ pushptr(Address(rcx, 0)); // save final return address
3291 // Re-push self-frame
3292 __ enter(); // save old & set new rbp,
3293 __ subptr(rsp, (framesize-2) * wordSize); // Prolog!
3296 // set last_Java_sp, last_Java_fp
3297 __ get_thread(rdi);
3298 __ set_last_Java_frame(rdi, noreg, rbp, NULL);
3300 // Call C code. Need thread but NOT official VM entry
3301 // crud. We cannot block on this call, no GC can happen. Call should
3302 // restore return values to their stack-slots with the new SP.
3303 __ movptr(Address(rsp,arg0_off*wordSize),rdi);
3304 __ movl(Address(rsp,arg1_off*wordSize), Deoptimization::Unpack_uncommon_trap);
3305 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
3306 // Set an oopmap for the call site
3307 oop_maps->add_gc_map( __ pc()-start, new OopMap( framesize, 0 ) );
3309 __ get_thread(rdi);
3310 __ reset_last_Java_frame(rdi, true, false);
3312 // Pop self-frame.
3313 __ leave(); // Epilog!
3315 // Jump to interpreter
3316 __ ret(0);
3318 // -------------
3319 // make sure all code is generated
3320 masm->flush();
3322 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, framesize);
3323 }
3324 #endif // COMPILER2
3326 //------------------------------generate_handler_blob------
3327 //
3328 // Generate a special Compile2Runtime blob that saves all registers,
3329 // setup oopmap, and calls safepoint code to stop the compiled code for
3330 // a safepoint.
3331 //
3332 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3334 // Account for thread arg in our frame
3335 const int additional_words = 1;
3336 int frame_size_in_words;
3338 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3340 ResourceMark rm;
3341 OopMapSet *oop_maps = new OopMapSet();
3342 OopMap* map;
3344 // allocate space for the code
3345 // setup code generation tools
3346 CodeBuffer buffer("handler_blob", 1024, 512);
3347 MacroAssembler* masm = new MacroAssembler(&buffer);
3349 const Register java_thread = rdi; // callee-saved for VC++
3350 address start = __ pc();
3351 address call_pc = NULL;
3352 bool cause_return = (poll_type == POLL_AT_RETURN);
3353 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP);
3354 // If cause_return is true we are at a poll_return and there is
3355 // the return address on the stack to the caller on the nmethod
3356 // that is safepoint. We can leave this return on the stack and
3357 // effectively complete the return and safepoint in the caller.
3358 // Otherwise we push space for a return address that the safepoint
3359 // handler will install later to make the stack walking sensible.
3360 if (!cause_return)
3361 __ push(rbx); // Make room for return address (or push it again)
3363 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false, save_vectors);
3365 // The following is basically a call_VM. However, we need the precise
3366 // address of the call in order to generate an oopmap. Hence, we do all the
3367 // work ourselves.
3369 // Push thread argument and setup last_Java_sp
3370 __ get_thread(java_thread);
3371 __ push(java_thread);
3372 __ set_last_Java_frame(java_thread, noreg, noreg, NULL);
3374 // if this was not a poll_return then we need to correct the return address now.
3375 if (!cause_return) {
3376 __ movptr(rax, Address(java_thread, JavaThread::saved_exception_pc_offset()));
3377 __ movptr(Address(rbp, wordSize), rax);
3378 }
3380 // do the call
3381 __ call(RuntimeAddress(call_ptr));
3383 // Set an oopmap for the call site. This oopmap will map all
3384 // oop-registers and debug-info registers as callee-saved. This
3385 // will allow deoptimization at this safepoint to find all possible
3386 // debug-info recordings, as well as let GC find all oops.
3388 oop_maps->add_gc_map( __ pc() - start, map);
3390 // Discard arg
3391 __ pop(rcx);
3393 Label noException;
3395 // Clear last_Java_sp again
3396 __ get_thread(java_thread);
3397 __ reset_last_Java_frame(java_thread, false, false);
3399 __ cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
3400 __ jcc(Assembler::equal, noException);
3402 // Exception pending
3403 RegisterSaver::restore_live_registers(masm, save_vectors);
3405 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3407 __ bind(noException);
3409 // Normal exit, register restoring and exit
3410 RegisterSaver::restore_live_registers(masm, save_vectors);
3412 __ ret(0);
3414 // make sure all code is generated
3415 masm->flush();
3417 // Fill-out other meta info
3418 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
3419 }
3421 //
3422 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3423 //
3424 // Generate a stub that calls into vm to find out the proper destination
3425 // of a java call. All the argument registers are live at this point
3426 // but since this is generic code we don't know what they are and the caller
3427 // must do any gc of the args.
3428 //
3429 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3430 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3432 // allocate space for the code
3433 ResourceMark rm;
3435 CodeBuffer buffer(name, 1000, 512);
3436 MacroAssembler* masm = new MacroAssembler(&buffer);
3438 int frame_size_words;
3439 enum frame_layout {
3440 thread_off,
3441 extra_words };
3443 OopMapSet *oop_maps = new OopMapSet();
3444 OopMap* map = NULL;
3446 int start = __ offset();
3448 map = RegisterSaver::save_live_registers(masm, extra_words, &frame_size_words);
3450 int frame_complete = __ offset();
3452 const Register thread = rdi;
3453 __ get_thread(rdi);
3455 __ push(thread);
3456 __ set_last_Java_frame(thread, noreg, rbp, NULL);
3458 __ call(RuntimeAddress(destination));
3461 // Set an oopmap for the call site.
3462 // We need this not only for callee-saved registers, but also for volatile
3463 // registers that the compiler might be keeping live across a safepoint.
3465 oop_maps->add_gc_map( __ offset() - start, map);
3467 // rax, contains the address we are going to jump to assuming no exception got installed
3469 __ addptr(rsp, wordSize);
3471 // clear last_Java_sp
3472 __ reset_last_Java_frame(thread, true, false);
3473 // check for pending exceptions
3474 Label pending;
3475 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
3476 __ jcc(Assembler::notEqual, pending);
3478 // get the returned Method*
3479 __ get_vm_result_2(rbx, thread);
3480 __ movptr(Address(rsp, RegisterSaver::rbx_offset() * wordSize), rbx);
3482 __ movptr(Address(rsp, RegisterSaver::rax_offset() * wordSize), rax);
3484 RegisterSaver::restore_live_registers(masm);
3486 // We are back the the original state on entry and ready to go.
3488 __ jmp(rax);
3490 // Pending exception after the safepoint
3492 __ bind(pending);
3494 RegisterSaver::restore_live_registers(masm);
3496 // exception pending => remove activation and forward to exception handler
3498 __ get_thread(thread);
3499 __ movptr(Address(thread, JavaThread::vm_result_offset()), NULL_WORD);
3500 __ movptr(rax, Address(thread, Thread::pending_exception_offset()));
3501 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3503 // -------------
3504 // make sure all code is generated
3505 masm->flush();
3507 // return the blob
3508 // frame_size_words or bytes??
3509 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3510 }