1 /*
2 * Copyright (c) 2008, 2011, 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 */
24
25 #include "precompiled.hpp"
26 #include "interpreter/interpreter.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "prims/methodHandles.hpp"
29
30 #define __ _masm->
31
32 #ifdef PRODUCT
33 #define BLOCK_COMMENT(str) /* nothing */
34 #else
35 #define BLOCK_COMMENT(str) __ block_comment(str)
36 #endif
37
38 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
39
40 address MethodHandleEntry::start_compiled_entry(MacroAssembler* _masm,
41 address interpreted_entry) {
42 // Just before the actual machine code entry point, allocate space
43 // for a MethodHandleEntry::Data record, so that we can manage everything
44 // from one base pointer.
45 __ align(wordSize);
46 address target = __ pc() + sizeof(Data);
47 while (__ pc() < target) {
48 __ nop();
49 __ align(wordSize);
50 }
51
52 MethodHandleEntry* me = (MethodHandleEntry*) __ pc();
53 me->set_end_address(__ pc()); // set a temporary end_address
54 me->set_from_interpreted_entry(interpreted_entry);
55 me->set_type_checking_entry(NULL);
56
57 return (address) me;
58 }
59
60 MethodHandleEntry* MethodHandleEntry::finish_compiled_entry(MacroAssembler* _masm,
61 address start_addr) {
62 MethodHandleEntry* me = (MethodHandleEntry*) start_addr;
63 assert(me->end_address() == start_addr, "valid ME");
64
65 // Fill in the real end_address:
66 __ align(wordSize);
67 me->set_end_address(__ pc());
68
69 return me;
70 }
71
72 // stack walking support
73
74 frame MethodHandles::ricochet_frame_sender(const frame& fr, RegisterMap *map) {
75 //RicochetFrame* f = RicochetFrame::from_frame(fr);
76 // Cf. is_interpreted_frame path of frame::sender
77 intptr_t* younger_sp = fr.sp();
78 intptr_t* sp = fr.sender_sp();
79 map->make_integer_regs_unsaved();
80 map->shift_window(sp, younger_sp);
81 bool this_frame_adjusted_stack = true; // I5_savedSP is live in this RF
82 return frame(sp, younger_sp, this_frame_adjusted_stack);
83 }
84
85 void MethodHandles::ricochet_frame_oops_do(const frame& fr, OopClosure* blk, const RegisterMap* reg_map) {
86 ResourceMark rm;
87 RicochetFrame* f = RicochetFrame::from_frame(fr);
88
89 // pick up the argument type descriptor:
90 Thread* thread = Thread::current();
91 Handle cookie(thread, f->compute_saved_args_layout(true, true));
92
93 // process fixed part
94 blk->do_oop((oop*)f->saved_target_addr());
95 blk->do_oop((oop*)f->saved_args_layout_addr());
96
97 // process variable arguments:
98 if (cookie.is_null()) return; // no arguments to describe
99
100 // the cookie is actually the invokeExact method for my target
101 // his argument signature is what I'm interested in
102 assert(cookie->is_method(), "");
103 methodHandle invoker(thread, methodOop(cookie()));
104 assert(invoker->name() == vmSymbols::invokeExact_name(), "must be this kind of method");
105 assert(!invoker->is_static(), "must have MH argument");
106 int slot_count = invoker->size_of_parameters();
107 assert(slot_count >= 1, "must include 'this'");
108 intptr_t* base = f->saved_args_base();
109 intptr_t* retval = NULL;
110 if (f->has_return_value_slot())
111 retval = f->return_value_slot_addr();
112 int slot_num = slot_count - 1;
113 intptr_t* loc = &base[slot_num];
114 //blk->do_oop((oop*) loc); // original target, which is irrelevant
115 int arg_num = 0;
116 for (SignatureStream ss(invoker->signature()); !ss.is_done(); ss.next()) {
117 if (ss.at_return_type()) continue;
118 BasicType ptype = ss.type();
119 if (ptype == T_ARRAY) ptype = T_OBJECT; // fold all refs to T_OBJECT
120 assert(ptype >= T_BOOLEAN && ptype <= T_OBJECT, "not array or void");
121 slot_num -= type2size[ptype];
122 loc = &base[slot_num];
123 bool is_oop = (ptype == T_OBJECT && loc != retval);
124 if (is_oop) blk->do_oop((oop*)loc);
125 arg_num += 1;
126 }
127 assert(slot_num == 0, "must have processed all the arguments");
128 }
129
130 // Ricochet Frames
131 const Register MethodHandles::RicochetFrame::L1_continuation = L1;
132 const Register MethodHandles::RicochetFrame::L2_saved_target = L2;
133 const Register MethodHandles::RicochetFrame::L3_saved_args_layout = L3;
134 const Register MethodHandles::RicochetFrame::L4_saved_args_base = L4; // cf. Gargs = G4
135 const Register MethodHandles::RicochetFrame::L5_conversion = L5;
136 #ifdef ASSERT
137 const Register MethodHandles::RicochetFrame::L0_magic_number_1 = L0;
138 #endif //ASSERT
139
140 oop MethodHandles::RicochetFrame::compute_saved_args_layout(bool read_cache, bool write_cache) {
141 if (read_cache) {
142 oop cookie = saved_args_layout();
143 if (cookie != NULL) return cookie;
144 }
145 oop target = saved_target();
146 oop mtype = java_lang_invoke_MethodHandle::type(target);
147 oop mtform = java_lang_invoke_MethodType::form(mtype);
148 oop cookie = java_lang_invoke_MethodTypeForm::vmlayout(mtform);
149 if (write_cache) {
150 (*saved_args_layout_addr()) = cookie;
151 }
152 return cookie;
153 }
154
155 void MethodHandles::RicochetFrame::generate_ricochet_blob(MacroAssembler* _masm,
156 // output params:
157 int* bounce_offset,
158 int* exception_offset,
159 int* frame_size_in_words) {
160 (*frame_size_in_words) = RicochetFrame::frame_size_in_bytes() / wordSize;
161
162 address start = __ pc();
163
164 #ifdef ASSERT
165 __ illtrap(0); __ illtrap(0); __ illtrap(0);
166 // here's a hint of something special:
167 __ set(MAGIC_NUMBER_1, G0);
168 __ set(MAGIC_NUMBER_2, G0);
169 #endif //ASSERT
170 __ illtrap(0); // not reached
171
172 // Return values are in registers.
173 // L1_continuation contains a cleanup continuation we must return
174 // to.
175
176 (*bounce_offset) = __ pc() - start;
177 BLOCK_COMMENT("ricochet_blob.bounce");
178
179 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
180 trace_method_handle(_masm, "ricochet_blob.bounce");
181
182 __ JMP(L1_continuation, 0);
183 __ delayed()->nop();
184 __ illtrap(0);
185
186 DEBUG_ONLY(__ set(MAGIC_NUMBER_2, G0));
187
188 (*exception_offset) = __ pc() - start;
189 BLOCK_COMMENT("ricochet_blob.exception");
190
191 // compare this to Interpreter::rethrow_exception_entry, which is parallel code
192 // for example, see TemplateInterpreterGenerator::generate_throw_exception
193 // Live registers in:
194 // Oexception (O0): exception
195 // Oissuing_pc (O1): return address/pc that threw exception (ignored, always equal to bounce addr)
196 __ verify_oop(Oexception);
197
198 // Take down the frame.
199
200 // Cf. InterpreterMacroAssembler::remove_activation.
201 leave_ricochet_frame(_masm, /*recv_reg=*/ noreg, I5_savedSP, I7);
202
203 // We are done with this activation frame; find out where to go next.
204 // The continuation point will be an exception handler, which expects
205 // the following registers set up:
206 //
207 // Oexception: exception
208 // Oissuing_pc: the local call that threw exception
209 // Other On: garbage
210 // In/Ln: the contents of the caller's register window
211 //
212 // We do the required restore at the last possible moment, because we
213 // need to preserve some state across a runtime call.
214 // (Remember that the caller activation is unknown--it might not be
215 // interpreted, so things like Lscratch are useless in the caller.)
216 __ mov(Oexception, Oexception ->after_save()); // get exception in I0 so it will be on O0 after restore
217 __ add(I7, frame::pc_return_offset, Oissuing_pc->after_save()); // likewise set I1 to a value local to the caller
218 __ call_VM_leaf(L7_thread_cache,
219 CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
220 G2_thread, Oissuing_pc->after_save());
221
222 // The caller's SP was adjusted upon method entry to accomodate
223 // the callee's non-argument locals. Undo that adjustment.
224 __ JMP(O0, 0); // return exception handler in caller
225 __ delayed()->restore(I5_savedSP, G0, SP);
226
227 // (same old exception object is already in Oexception; see above)
228 // Note that an "issuing PC" is actually the next PC after the call
229 }
230
231 void MethodHandles::RicochetFrame::enter_ricochet_frame(MacroAssembler* _masm,
232 Register recv_reg,
233 Register argv_reg,
234 address return_handler) {
235 // does not include the __ save()
236 assert(argv_reg == Gargs, "");
237 Address G3_mh_vmtarget( recv_reg, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes());
238 Address G3_amh_conversion(recv_reg, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
239
240 // Create the RicochetFrame.
241 // Unlike on x86 we can store all required information in local
242 // registers.
243 BLOCK_COMMENT("push RicochetFrame {");
244 __ set(ExternalAddress(return_handler), L1_continuation);
245 __ load_heap_oop(G3_mh_vmtarget, L2_saved_target);
246 __ mov(G0, L3_saved_args_layout);
247 __ mov(Gargs, L4_saved_args_base);
248 __ lduw(G3_amh_conversion, L5_conversion); // 32-bit field
249 // I5, I6, I7 are already set up
250 DEBUG_ONLY(__ set((int32_t) MAGIC_NUMBER_1, L0_magic_number_1));
251 BLOCK_COMMENT("} RicochetFrame");
252 }
253
254 void MethodHandles::RicochetFrame::leave_ricochet_frame(MacroAssembler* _masm,
255 Register recv_reg,
256 Register new_sp_reg,
257 Register sender_pc_reg) {
258 assert(new_sp_reg == I5_savedSP, "exact_sender_sp already in place");
259 assert(sender_pc_reg == I7, "in a fixed place");
260 // does not include the __ ret() & __ restore()
261 assert_different_registers(recv_reg, new_sp_reg, sender_pc_reg);
262 // Take down the frame.
263 // Cf. InterpreterMacroAssembler::remove_activation.
264 BLOCK_COMMENT("end_ricochet_frame {");
265 if (recv_reg->is_valid())
266 __ mov(L2_saved_target, recv_reg);
267 BLOCK_COMMENT("} end_ricochet_frame");
268 }
269
270 // Emit code to verify that FP is pointing at a valid ricochet frame.
271 #ifdef ASSERT
272 enum {
273 ARG_LIMIT = 255, SLOP = 45,
274 // use this parameter for checking for garbage stack movements:
275 UNREASONABLE_STACK_MOVE = (ARG_LIMIT + SLOP)
276 // the slop defends against false alarms due to fencepost errors
277 };
278
279 void MethodHandles::RicochetFrame::verify_clean(MacroAssembler* _masm) {
280 // The stack should look like this:
281 // ... keep1 | dest=42 | keep2 | magic | handler | magic | recursive args | [RF]
282 // Check various invariants.
283
284 Register O7_temp = O7, O5_temp = O5;
285
286 Label L_ok_1, L_ok_2, L_ok_3, L_ok_4;
287 BLOCK_COMMENT("verify_clean {");
288 // Magic numbers must check out:
289 __ set((int32_t) MAGIC_NUMBER_1, O7_temp);
290 __ cmp(O7_temp, L0_magic_number_1);
291 __ br(Assembler::equal, false, Assembler::pt, L_ok_1);
292 __ delayed()->nop();
293 __ stop("damaged ricochet frame: MAGIC_NUMBER_1 not found");
294
295 __ BIND(L_ok_1);
296
297 // Arguments pointer must look reasonable:
298 #ifdef _LP64
299 Register FP_temp = O5_temp;
300 __ add(FP, STACK_BIAS, FP_temp);
301 #else
302 Register FP_temp = FP;
303 #endif
304 __ cmp(L4_saved_args_base, FP_temp);
305 __ br(Assembler::greaterEqualUnsigned, false, Assembler::pt, L_ok_2);
306 __ delayed()->nop();
307 __ stop("damaged ricochet frame: L4 < FP");
308
309 __ BIND(L_ok_2);
310 // Disable until we decide on it's fate
311 // __ sub(L4_saved_args_base, UNREASONABLE_STACK_MOVE * Interpreter::stackElementSize, O7_temp);
312 // __ cmp(O7_temp, FP_temp);
313 // __ br(Assembler::lessEqualUnsigned, false, Assembler::pt, L_ok_3);
314 // __ delayed()->nop();
315 // __ stop("damaged ricochet frame: (L4 - UNREASONABLE_STACK_MOVE) > FP");
316
317 __ BIND(L_ok_3);
318 extract_conversion_dest_type(_masm, L5_conversion, O7_temp);
319 __ cmp(O7_temp, T_VOID);
320 __ br(Assembler::equal, false, Assembler::pt, L_ok_4);
321 __ delayed()->nop();
322 extract_conversion_vminfo(_masm, L5_conversion, O5_temp);
323 __ ld_ptr(L4_saved_args_base, __ argument_offset(O5_temp, O5_temp), O7_temp);
324 assert(__ is_simm13(RETURN_VALUE_PLACEHOLDER), "must be simm13");
325 __ cmp(O7_temp, (int32_t) RETURN_VALUE_PLACEHOLDER);
326 __ brx(Assembler::equal, false, Assembler::pt, L_ok_4);
327 __ delayed()->nop();
328 __ stop("damaged ricochet frame: RETURN_VALUE_PLACEHOLDER not found");
329 __ BIND(L_ok_4);
330 BLOCK_COMMENT("} verify_clean");
331 }
332 #endif //ASSERT
333
334 void MethodHandles::load_klass_from_Class(MacroAssembler* _masm, Register klass_reg, Register temp_reg, Register temp2_reg) {
335 if (VerifyMethodHandles)
336 verify_klass(_masm, klass_reg, SystemDictionaryHandles::Class_klass(), temp_reg, temp2_reg,
337 "AMH argument is a Class");
338 __ load_heap_oop(Address(klass_reg, java_lang_Class::klass_offset_in_bytes()), klass_reg);
339 }
340
341 void MethodHandles::load_conversion_vminfo(MacroAssembler* _masm, Address conversion_field_addr, Register reg) {
342 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
343 assert(CONV_VMINFO_MASK == right_n_bits(BitsPerByte), "else change type of following load");
344 __ ldub(conversion_field_addr.plus_disp(BytesPerInt - 1), reg);
345 }
346
347 void MethodHandles::extract_conversion_vminfo(MacroAssembler* _masm, Register conversion_field_reg, Register reg) {
348 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
349 __ and3(conversion_field_reg, CONV_VMINFO_MASK, reg);
350 }
351
352 void MethodHandles::extract_conversion_dest_type(MacroAssembler* _masm, Register conversion_field_reg, Register reg) {
353 __ srl(conversion_field_reg, CONV_DEST_TYPE_SHIFT, reg);
354 __ and3(reg, 0x0F, reg);
355 }
356
357 void MethodHandles::load_stack_move(MacroAssembler* _masm,
358 Address G3_amh_conversion,
359 Register stack_move_reg) {
360 BLOCK_COMMENT("load_stack_move {");
361 __ ldsw(G3_amh_conversion, stack_move_reg);
362 __ sra(stack_move_reg, CONV_STACK_MOVE_SHIFT, stack_move_reg);
363 if (VerifyMethodHandles) {
364 Label L_ok, L_bad;
365 int32_t stack_move_limit = 0x0800; // extra-large
366 __ cmp(stack_move_reg, stack_move_limit);
367 __ br(Assembler::greaterEqual, false, Assembler::pn, L_bad);
368 __ delayed()->nop();
369 __ cmp(stack_move_reg, -stack_move_limit);
370 __ br(Assembler::greater, false, Assembler::pt, L_ok);
371 __ delayed()->nop();
372 __ BIND(L_bad);
373 __ stop("load_stack_move of garbage value");
374 __ BIND(L_ok);
375 }
376 BLOCK_COMMENT("} load_stack_move");
377 }
378
379 #ifdef ASSERT
380 void MethodHandles::RicochetFrame::verify() const {
381 assert(magic_number_1() == MAGIC_NUMBER_1, "");
382 if (!Universe::heap()->is_gc_active()) {
383 if (saved_args_layout() != NULL) {
384 assert(saved_args_layout()->is_method(), "must be valid oop");
385 }
386 if (saved_target() != NULL) {
387 assert(java_lang_invoke_MethodHandle::is_instance(saved_target()), "checking frame value");
388 }
389 }
390 int conv_op = adapter_conversion_op(conversion());
391 assert(conv_op == java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS ||
392 conv_op == java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS ||
393 conv_op == java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF,
394 "must be a sane conversion");
395 if (has_return_value_slot()) {
396 assert(*return_value_slot_addr() == RETURN_VALUE_PLACEHOLDER, "");
397 }
398 }
399
400 void MethodHandles::verify_argslot(MacroAssembler* _masm, Register argslot_reg, Register temp_reg, const char* error_message) {
401 // Verify that argslot lies within (Gargs, FP].
402 Label L_ok, L_bad;
403 BLOCK_COMMENT("verify_argslot {");
404 __ add(FP, STACK_BIAS, temp_reg); // STACK_BIAS is zero on !_LP64
405 __ cmp(argslot_reg, temp_reg);
406 __ brx(Assembler::greaterUnsigned, false, Assembler::pn, L_bad);
407 __ delayed()->nop();
408 __ cmp(Gargs, argslot_reg);
409 __ brx(Assembler::lessEqualUnsigned, false, Assembler::pt, L_ok);
410 __ delayed()->nop();
411 __ BIND(L_bad);
412 __ stop(error_message);
413 __ BIND(L_ok);
414 BLOCK_COMMENT("} verify_argslot");
415 }
416
417 void MethodHandles::verify_argslots(MacroAssembler* _masm,
418 RegisterOrConstant arg_slots,
419 Register arg_slot_base_reg,
420 Register temp_reg,
421 Register temp2_reg,
422 bool negate_argslots,
423 const char* error_message) {
424 // Verify that [argslot..argslot+size) lies within (Gargs, FP).
425 Label L_ok, L_bad;
426 BLOCK_COMMENT("verify_argslots {");
427 if (negate_argslots) {
428 if (arg_slots.is_constant()) {
429 arg_slots = -1 * arg_slots.as_constant();
430 } else {
431 __ neg(arg_slots.as_register(), temp_reg);
432 arg_slots = temp_reg;
433 }
434 }
435 __ add(arg_slot_base_reg, __ argument_offset(arg_slots, temp_reg), temp_reg);
436 __ add(FP, STACK_BIAS, temp2_reg); // STACK_BIAS is zero on !_LP64
437 __ cmp(temp_reg, temp2_reg);
438 __ brx(Assembler::greaterUnsigned, false, Assembler::pn, L_bad);
439 __ delayed()->nop();
440 // Gargs points to the first word so adjust by BytesPerWord
441 __ add(arg_slot_base_reg, BytesPerWord, temp_reg);
442 __ cmp(Gargs, temp_reg);
443 __ brx(Assembler::lessEqualUnsigned, false, Assembler::pt, L_ok);
444 __ delayed()->nop();
445 __ BIND(L_bad);
446 __ stop(error_message);
447 __ BIND(L_ok);
448 BLOCK_COMMENT("} verify_argslots");
449 }
450
451 // Make sure that arg_slots has the same sign as the given direction.
452 // If (and only if) arg_slots is a assembly-time constant, also allow it to be zero.
453 void MethodHandles::verify_stack_move(MacroAssembler* _masm,
454 RegisterOrConstant arg_slots, int direction) {
455 enum { UNREASONABLE_STACK_MOVE = 256 * 4 }; // limit of 255 arguments
456 bool allow_zero = arg_slots.is_constant();
457 if (direction == 0) { direction = +1; allow_zero = true; }
458 assert(stack_move_unit() == -1, "else add extra checks here");
459 if (arg_slots.is_register()) {
460 Label L_ok, L_bad;
461 BLOCK_COMMENT("verify_stack_move {");
462 // __ btst(-stack_move_unit() - 1, arg_slots.as_register()); // no need
463 // __ br(Assembler::notZero, false, Assembler::pn, L_bad);
464 // __ delayed()->nop();
465 __ cmp(arg_slots.as_register(), (int32_t) NULL_WORD);
466 if (direction > 0) {
467 __ br(allow_zero ? Assembler::less : Assembler::lessEqual, false, Assembler::pn, L_bad);
468 __ delayed()->nop();
469 __ cmp(arg_slots.as_register(), (int32_t) UNREASONABLE_STACK_MOVE);
470 __ br(Assembler::less, false, Assembler::pn, L_ok);
471 __ delayed()->nop();
472 } else {
473 __ br(allow_zero ? Assembler::greater : Assembler::greaterEqual, false, Assembler::pn, L_bad);
474 __ delayed()->nop();
475 __ cmp(arg_slots.as_register(), (int32_t) -UNREASONABLE_STACK_MOVE);
476 __ br(Assembler::greater, false, Assembler::pn, L_ok);
477 __ delayed()->nop();
478 }
479 __ BIND(L_bad);
480 if (direction > 0)
481 __ stop("assert arg_slots > 0");
482 else
483 __ stop("assert arg_slots < 0");
484 __ BIND(L_ok);
485 BLOCK_COMMENT("} verify_stack_move");
486 } else {
487 intptr_t size = arg_slots.as_constant();
488 if (direction < 0) size = -size;
489 assert(size >= 0, "correct direction of constant move");
490 assert(size < UNREASONABLE_STACK_MOVE, "reasonable size of constant move");
491 }
492 }
493
494 void MethodHandles::verify_klass(MacroAssembler* _masm,
495 Register obj_reg, KlassHandle klass,
496 Register temp_reg, Register temp2_reg,
497 const char* error_message) {
498 oop* klass_addr = klass.raw_value();
499 assert(klass_addr >= SystemDictionaryHandles::Object_klass().raw_value() &&
500 klass_addr <= SystemDictionaryHandles::Long_klass().raw_value(),
501 "must be one of the SystemDictionaryHandles");
502 Label L_ok, L_bad;
503 BLOCK_COMMENT("verify_klass {");
504 __ verify_oop(obj_reg);
505 __ br_null(obj_reg, false, Assembler::pn, L_bad);
506 __ delayed()->nop();
507 __ load_klass(obj_reg, temp_reg);
508 __ set(ExternalAddress(klass_addr), temp2_reg);
509 __ ld_ptr(Address(temp2_reg, 0), temp2_reg);
510 __ cmp(temp_reg, temp2_reg);
511 __ brx(Assembler::equal, false, Assembler::pt, L_ok);
512 __ delayed()->nop();
513 intptr_t super_check_offset = klass->super_check_offset();
514 __ ld_ptr(Address(temp_reg, super_check_offset), temp_reg);
515 __ set(ExternalAddress(klass_addr), temp2_reg);
516 __ ld_ptr(Address(temp2_reg, 0), temp2_reg);
517 __ cmp(temp_reg, temp2_reg);
518 __ brx(Assembler::equal, false, Assembler::pt, L_ok);
519 __ delayed()->nop();
520 __ BIND(L_bad);
521 __ stop(error_message);
522 __ BIND(L_ok);
523 BLOCK_COMMENT("} verify_klass");
524 }
525 #endif // ASSERT
526
527 // Code generation
528 address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm) {
529 // I5_savedSP/O5_savedSP: sender SP (must preserve)
530 // G4 (Gargs): incoming argument list (must preserve)
531 // G5_method: invoke methodOop
532 // G3_method_handle: receiver method handle (must load from sp[MethodTypeForm.vmslots])
533 // O0, O1, O2, O3, O4: garbage temps, blown away
534 Register O0_mtype = O0;
535 Register O1_scratch = O1;
536 Register O2_scratch = O2;
537 Register O3_scratch = O3;
538 Register O4_argslot = O4;
539 Register O4_argbase = O4;
540
541 // emit WrongMethodType path first, to enable back-branch from main path
542 Label wrong_method_type;
543 __ bind(wrong_method_type);
544 Label invoke_generic_slow_path;
545 assert(methodOopDesc::intrinsic_id_size_in_bytes() == sizeof(u1), "");;
546 __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
547 __ cmp(O1_scratch, (int) vmIntrinsics::_invokeExact);
548 __ brx(Assembler::notEqual, false, Assembler::pt, invoke_generic_slow_path);
549 __ delayed()->nop();
550 __ mov(O0_mtype, G5_method_type); // required by throw_WrongMethodType
551 __ mov(G3_method_handle, G3_method_handle); // already in this register
552 // O0 will be filled in with JavaThread in stub
553 __ jump_to(AddressLiteral(StubRoutines::throw_WrongMethodTypeException_entry()), O3_scratch);
554 __ delayed()->nop();
555
556 // here's where control starts out:
557 __ align(CodeEntryAlignment);
558 address entry_point = __ pc();
559
560 // fetch the MethodType from the method handle
561 // FIXME: Interpreter should transmit pre-popped stack pointer, to locate base of arg list.
562 // This would simplify several touchy bits of code.
563 // See 6984712: JSR 292 method handle calls need a clean argument base pointer
564 {
565 Register tem = G5_method;
566 for (jint* pchase = methodOopDesc::method_type_offsets_chain(); (*pchase) != -1; pchase++) {
567 __ ld_ptr(Address(tem, *pchase), O0_mtype);
568 tem = O0_mtype; // in case there is another indirection
569 }
570 }
571
572 // given the MethodType, find out where the MH argument is buried
573 __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, O1_scratch)), O4_argslot);
574 __ ldsw( Address(O4_argslot, __ delayed_value(java_lang_invoke_MethodTypeForm::vmslots_offset_in_bytes, O1_scratch)), O4_argslot);
575 __ add(__ argument_address(O4_argslot, O4_argslot, 1), O4_argbase);
576 // Note: argument_address uses its input as a scratch register!
577 Address mh_receiver_slot_addr(O4_argbase, -Interpreter::stackElementSize);
578 __ ld_ptr(mh_receiver_slot_addr, G3_method_handle);
579
580 trace_method_handle(_masm, "invokeExact");
581
582 __ check_method_handle_type(O0_mtype, G3_method_handle, O1_scratch, wrong_method_type);
583
584 // Nobody uses the MH receiver slot after this. Make sure.
585 DEBUG_ONLY(__ set((int32_t) 0x999999, O1_scratch); __ st_ptr(O1_scratch, mh_receiver_slot_addr));
586
587 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
588
589 // for invokeGeneric (only), apply argument and result conversions on the fly
590 __ bind(invoke_generic_slow_path);
591 #ifdef ASSERT
592 if (VerifyMethodHandles) {
593 Label L;
594 __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
595 __ cmp(O1_scratch, (int) vmIntrinsics::_invokeGeneric);
596 __ brx(Assembler::equal, false, Assembler::pt, L);
597 __ delayed()->nop();
598 __ stop("bad methodOop::intrinsic_id");
599 __ bind(L);
600 }
601 #endif //ASSERT
602
603 // make room on the stack for another pointer:
604 insert_arg_slots(_masm, 2 * stack_move_unit(), O4_argbase, O1_scratch, O2_scratch, O3_scratch);
605 // load up an adapter from the calling type (Java weaves this)
606 Register O2_form = O2_scratch;
607 Register O3_adapter = O3_scratch;
608 __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, O1_scratch)), O2_form);
609 __ load_heap_oop(Address(O2_form, __ delayed_value(java_lang_invoke_MethodTypeForm::genericInvoker_offset_in_bytes, O1_scratch)), O3_adapter);
610 __ verify_oop(O3_adapter);
611 __ st_ptr(O3_adapter, Address(O4_argbase, 1 * Interpreter::stackElementSize));
612 // As a trusted first argument, pass the type being called, so the adapter knows
613 // the actual types of the arguments and return values.
614 // (Generic invokers are shared among form-families of method-type.)
615 __ st_ptr(O0_mtype, Address(O4_argbase, 0 * Interpreter::stackElementSize));
616 // FIXME: assert that O3_adapter is of the right method-type.
617 __ mov(O3_adapter, G3_method_handle);
618 trace_method_handle(_masm, "invokeGeneric");
619 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
620
621 return entry_point;
622 }
623
624 // Workaround for C++ overloading nastiness on '0' for RegisterOrConstant.
625 static RegisterOrConstant constant(int value) {
626 return RegisterOrConstant(value);
627 }
628
629 static void load_vmargslot(MacroAssembler* _masm, Address vmargslot_addr, Register result) {
630 __ ldsw(vmargslot_addr, result);
631 }
632
633 static RegisterOrConstant adjust_SP_and_Gargs_down_by_slots(MacroAssembler* _masm,
634 RegisterOrConstant arg_slots,
635 Register temp_reg, Register temp2_reg) {
636 // Keep the stack pointer 2*wordSize aligned.
637 const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
638 if (arg_slots.is_constant()) {
639 const int offset = arg_slots.as_constant() << LogBytesPerWord;
640 const int masked_offset = round_to(offset, 2 * BytesPerWord);
641 const int masked_offset2 = (offset + 1*BytesPerWord) & ~TwoWordAlignmentMask;
642 assert(masked_offset == masked_offset2, "must agree");
643 __ sub(Gargs, offset, Gargs);
644 __ sub(SP, masked_offset, SP );
645 return offset;
646 } else {
647 #ifdef ASSERT
648 {
649 Label L_ok;
650 __ cmp(arg_slots.as_register(), 0);
651 __ br(Assembler::greaterEqual, false, Assembler::pt, L_ok);
652 __ delayed()->nop();
653 __ stop("negative arg_slots");
654 __ bind(L_ok);
655 }
656 #endif
657 __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
658 __ add( temp_reg, 1*BytesPerWord, temp2_reg);
659 __ andn(temp2_reg, TwoWordAlignmentMask, temp2_reg);
660 __ sub(Gargs, temp_reg, Gargs);
661 __ sub(SP, temp2_reg, SP );
662 return temp_reg;
663 }
664 }
665
666 static RegisterOrConstant adjust_SP_and_Gargs_up_by_slots(MacroAssembler* _masm,
667 RegisterOrConstant arg_slots,
668 Register temp_reg, Register temp2_reg) {
669 // Keep the stack pointer 2*wordSize aligned.
670 const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
671 if (arg_slots.is_constant()) {
672 const int offset = arg_slots.as_constant() << LogBytesPerWord;
673 const int masked_offset = offset & ~TwoWordAlignmentMask;
674 __ add(Gargs, offset, Gargs);
675 __ add(SP, masked_offset, SP );
676 return offset;
677 } else {
678 __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
679 __ andn(temp_reg, TwoWordAlignmentMask, temp2_reg);
680 __ add(Gargs, temp_reg, Gargs);
681 __ add(SP, temp2_reg, SP );
682 return temp_reg;
683 }
684 }
685
686 // Helper to insert argument slots into the stack.
687 // arg_slots must be a multiple of stack_move_unit() and < 0
688 // argslot_reg is decremented to point to the new (shifted) location of the argslot
689 // But, temp_reg ends up holding the original value of argslot_reg.
690 void MethodHandles::insert_arg_slots(MacroAssembler* _masm,
691 RegisterOrConstant arg_slots,
692 Register argslot_reg,
693 Register temp_reg, Register temp2_reg, Register temp3_reg) {
694 // allow constant zero
695 if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
696 return;
697
698 assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
699 (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
700
701 BLOCK_COMMENT("insert_arg_slots {");
702 if (VerifyMethodHandles)
703 verify_argslot(_masm, argslot_reg, temp_reg, "insertion point must fall within current frame");
704 if (VerifyMethodHandles)
705 verify_stack_move(_masm, arg_slots, -1);
706
707 // Make space on the stack for the inserted argument(s).
708 // Then pull down everything shallower than argslot_reg.
709 // The stacked return address gets pulled down with everything else.
710 // That is, copy [sp, argslot) downward by -size words. In pseudo-code:
711 // sp -= size;
712 // for (temp = sp + size; temp < argslot; temp++)
713 // temp[-size] = temp[0]
714 // argslot -= size;
715
716 // offset is temp3_reg in case of arg_slots being a register.
717 RegisterOrConstant offset = adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
718 __ sub(Gargs, offset, temp_reg); // source pointer for copy
719
720 {
721 Label loop;
722 __ BIND(loop);
723 // pull one word down each time through the loop
724 __ ld_ptr( Address(temp_reg, 0 ), temp2_reg);
725 __ st_ptr(temp2_reg, Address(temp_reg, offset) );
726 __ add(temp_reg, wordSize, temp_reg);
727 __ cmp(temp_reg, argslot_reg);
728 __ brx(Assembler::lessUnsigned, false, Assembler::pt, loop);
729 __ delayed()->nop(); // FILLME
730 }
731
732 // Now move the argslot down, to point to the opened-up space.
733 __ add(argslot_reg, offset, argslot_reg);
734 BLOCK_COMMENT("} insert_arg_slots");
735 }
736
737
738 // Helper to remove argument slots from the stack.
739 // arg_slots must be a multiple of stack_move_unit() and > 0
740 void MethodHandles::remove_arg_slots(MacroAssembler* _masm,
741 RegisterOrConstant arg_slots,
742 Register argslot_reg,
743 Register temp_reg, Register temp2_reg, Register temp3_reg) {
744 // allow constant zero
745 if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
746 return;
747 assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
748 (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
749
750 BLOCK_COMMENT("remove_arg_slots {");
751 if (VerifyMethodHandles)
752 verify_argslots(_masm, arg_slots, argslot_reg, temp_reg, temp2_reg, false,
753 "deleted argument(s) must fall within current frame");
754 if (VerifyMethodHandles)
755 verify_stack_move(_masm, arg_slots, +1);
756
757 // Pull up everything shallower than argslot.
758 // Then remove the excess space on the stack.
759 // The stacked return address gets pulled up with everything else.
760 // That is, copy [sp, argslot) upward by size words. In pseudo-code:
761 // for (temp = argslot-1; temp >= sp; --temp)
762 // temp[size] = temp[0]
763 // argslot += size;
764 // sp += size;
765
766 RegisterOrConstant offset = __ regcon_sll_ptr(arg_slots, LogBytesPerWord, temp3_reg);
767 __ sub(argslot_reg, wordSize, temp_reg); // source pointer for copy
768
769 {
770 Label L_loop;
771 __ BIND(L_loop);
772 // pull one word up each time through the loop
773 __ ld_ptr( Address(temp_reg, 0 ), temp2_reg);
774 __ st_ptr(temp2_reg, Address(temp_reg, offset) );
775 __ sub(temp_reg, wordSize, temp_reg);
776 __ cmp(temp_reg, Gargs);
777 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, L_loop);
778 __ delayed()->nop(); // FILLME
779 }
780
781 // And adjust the argslot address to point at the deletion point.
782 __ add(argslot_reg, offset, argslot_reg);
783
784 // We don't need the offset at this point anymore, just adjust SP and Gargs.
785 (void) adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
786
787 BLOCK_COMMENT("} remove_arg_slots");
788 }
789
790 // Helper to copy argument slots to the top of the stack.
791 // The sequence starts with argslot_reg and is counted by slot_count
792 // slot_count must be a multiple of stack_move_unit() and >= 0
793 // This function blows the temps but does not change argslot_reg.
794 void MethodHandles::push_arg_slots(MacroAssembler* _masm,
795 Register argslot_reg,
796 RegisterOrConstant slot_count,
797 Register temp_reg, Register temp2_reg) {
798 // allow constant zero
799 if (slot_count.is_constant() && slot_count.as_constant() == 0)
800 return;
801 assert_different_registers(argslot_reg, temp_reg, temp2_reg,
802 (!slot_count.is_register() ? Gargs : slot_count.as_register()),
803 SP);
804 assert(Interpreter::stackElementSize == wordSize, "else change this code");
805
806 BLOCK_COMMENT("push_arg_slots {");
807 if (VerifyMethodHandles)
808 verify_stack_move(_masm, slot_count, 0);
809
810 RegisterOrConstant offset = adjust_SP_and_Gargs_down_by_slots(_masm, slot_count, temp2_reg, temp_reg);
811
812 if (slot_count.is_constant()) {
813 for (int i = slot_count.as_constant() - 1; i >= 0; i--) {
814 __ ld_ptr( Address(argslot_reg, i * wordSize), temp_reg);
815 __ st_ptr(temp_reg, Address(Gargs, i * wordSize));
816 }
817 } else {
818 Label L_plural, L_loop, L_break;
819 // Emit code to dynamically check for the common cases, zero and one slot.
820 __ cmp(slot_count.as_register(), (int32_t) 1);
821 __ br(Assembler::greater, false, Assembler::pn, L_plural);
822 __ delayed()->nop();
823 __ br(Assembler::less, false, Assembler::pn, L_break);
824 __ delayed()->nop();
825 __ ld_ptr( Address(argslot_reg, 0), temp_reg);
826 __ st_ptr(temp_reg, Address(Gargs, 0));
827 __ ba(false, L_break);
828 __ delayed()->nop(); // FILLME
829 __ BIND(L_plural);
830
831 // Loop for 2 or more:
832 // top = &argslot[slot_count]
833 // while (top > argslot) *(--Gargs) = *(--top)
834 Register top_reg = temp_reg;
835 __ add(argslot_reg, offset, top_reg);
836 __ add(Gargs, offset, Gargs ); // move back up again so we can go down
837 __ BIND(L_loop);
838 __ sub(top_reg, wordSize, top_reg);
839 __ sub(Gargs, wordSize, Gargs );
840 __ ld_ptr( Address(top_reg, 0), temp2_reg);
841 __ st_ptr(temp2_reg, Address(Gargs, 0));
842 __ cmp(top_reg, argslot_reg);
843 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, L_loop);
844 __ delayed()->nop(); // FILLME
845 __ BIND(L_break);
846 }
847 BLOCK_COMMENT("} push_arg_slots");
848 }
849
850 // in-place movement; no change to Gargs
851 // blows temp_reg, temp2_reg
852 void MethodHandles::move_arg_slots_up(MacroAssembler* _masm,
853 Register bottom_reg, // invariant
854 Address top_addr, // can use temp_reg
855 RegisterOrConstant positive_distance_in_slots, // destroyed if register
856 Register temp_reg, Register temp2_reg) {
857 assert_different_registers(bottom_reg,
858 temp_reg, temp2_reg,
859 positive_distance_in_slots.register_or_noreg());
860 BLOCK_COMMENT("move_arg_slots_up {");
861 Label L_loop, L_break;
862 Register top_reg = temp_reg;
863 if (!top_addr.is_same_address(Address(top_reg, 0))) {
864 __ add(top_addr, top_reg);
865 }
866 // Detect empty (or broken) loop:
867 #ifdef ASSERT
868 if (VerifyMethodHandles) {
869 // Verify that &bottom < &top (non-empty interval)
870 Label L_ok, L_bad;
871 if (positive_distance_in_slots.is_register()) {
872 __ cmp(positive_distance_in_slots.as_register(), (int32_t) 0);
873 __ br(Assembler::lessEqual, false, Assembler::pn, L_bad);
874 __ delayed()->nop();
875 }
876 __ cmp(bottom_reg, top_reg);
877 __ brx(Assembler::lessUnsigned, false, Assembler::pt, L_ok);
878 __ delayed()->nop();
879 __ BIND(L_bad);
880 __ stop("valid bounds (copy up)");
881 __ BIND(L_ok);
882 }
883 #endif
884 __ cmp(bottom_reg, top_reg);
885 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pn, L_break);
886 __ delayed()->nop();
887 // work top down to bottom, copying contiguous data upwards
888 // In pseudo-code:
889 // while (--top >= bottom) *(top + distance) = *(top + 0);
890 RegisterOrConstant offset = __ argument_offset(positive_distance_in_slots, positive_distance_in_slots.register_or_noreg());
891 __ BIND(L_loop);
892 __ sub(top_reg, wordSize, top_reg);
893 __ ld_ptr( Address(top_reg, 0 ), temp2_reg);
894 __ st_ptr(temp2_reg, Address(top_reg, offset) );
895 __ cmp(top_reg, bottom_reg);
896 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, L_loop);
897 __ delayed()->nop(); // FILLME
898 assert(Interpreter::stackElementSize == wordSize, "else change loop");
899 __ BIND(L_break);
900 BLOCK_COMMENT("} move_arg_slots_up");
901 }
902
903 // in-place movement; no change to rsp
904 // blows temp_reg, temp2_reg
905 void MethodHandles::move_arg_slots_down(MacroAssembler* _masm,
906 Address bottom_addr, // can use temp_reg
907 Register top_reg, // invariant
908 RegisterOrConstant negative_distance_in_slots, // destroyed if register
909 Register temp_reg, Register temp2_reg) {
910 assert_different_registers(top_reg,
911 negative_distance_in_slots.register_or_noreg(),
912 temp_reg, temp2_reg);
913 BLOCK_COMMENT("move_arg_slots_down {");
914 Label L_loop, L_break;
915 Register bottom_reg = temp_reg;
916 if (!bottom_addr.is_same_address(Address(bottom_reg, 0))) {
917 __ add(bottom_addr, bottom_reg);
918 }
919 // Detect empty (or broken) loop:
920 #ifdef ASSERT
921 assert(!negative_distance_in_slots.is_constant() || negative_distance_in_slots.as_constant() < 0, "");
922 if (VerifyMethodHandles) {
923 // Verify that &bottom < &top (non-empty interval)
924 Label L_ok, L_bad;
925 if (negative_distance_in_slots.is_register()) {
926 __ cmp(negative_distance_in_slots.as_register(), (int32_t) 0);
927 __ br(Assembler::greaterEqual, false, Assembler::pn, L_bad);
928 __ delayed()->nop();
929 }
930 __ cmp(bottom_reg, top_reg);
931 __ brx(Assembler::lessUnsigned, false, Assembler::pt, L_ok);
932 __ delayed()->nop();
933 __ BIND(L_bad);
934 __ stop("valid bounds (copy down)");
935 __ BIND(L_ok);
936 }
937 #endif
938 __ cmp(bottom_reg, top_reg);
939 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pn, L_break);
940 __ delayed()->nop();
941 // work bottom up to top, copying contiguous data downwards
942 // In pseudo-code:
943 // while (bottom < top) *(bottom - distance) = *(bottom + 0), bottom++;
944 RegisterOrConstant offset = __ argument_offset(negative_distance_in_slots, negative_distance_in_slots.register_or_noreg());
945 __ BIND(L_loop);
946 __ ld_ptr( Address(bottom_reg, 0 ), temp2_reg);
947 __ st_ptr(temp2_reg, Address(bottom_reg, offset) );
948 __ add(bottom_reg, wordSize, bottom_reg);
949 __ cmp(bottom_reg, top_reg);
950 __ brx(Assembler::lessUnsigned, false, Assembler::pt, L_loop);
951 __ delayed()->nop(); // FILLME
952 assert(Interpreter::stackElementSize == wordSize, "else change loop");
953 __ BIND(L_break);
954 BLOCK_COMMENT("} move_arg_slots_down");
955 }
956
957 // Copy from a field or array element to a stacked argument slot.
958 // is_element (ignored) says whether caller is loading an array element instead of an instance field.
959 void MethodHandles::move_typed_arg(MacroAssembler* _masm,
960 BasicType type, bool is_element,
961 Address value_src, Address slot_dest,
962 Register temp_reg) {
963 assert(!slot_dest.uses(temp_reg), "must be different register");
964 BLOCK_COMMENT(!is_element ? "move_typed_arg {" : "move_typed_arg { (array element)");
965 if (type == T_OBJECT || type == T_ARRAY) {
966 __ load_heap_oop(value_src, temp_reg);
967 __ verify_oop(temp_reg);
968 __ st_ptr(temp_reg, slot_dest);
969 } else if (type != T_VOID) {
970 int arg_size = type2aelembytes(type);
971 bool arg_is_signed = is_signed_subword_type(type);
972 int slot_size = is_subword_type(type) ? type2aelembytes(T_INT) : arg_size; // store int sub-words as int
973 __ load_sized_value( value_src, temp_reg, arg_size, arg_is_signed);
974 __ store_sized_value(temp_reg, slot_dest, slot_size );
975 }
976 BLOCK_COMMENT("} move_typed_arg");
977 }
978
979 // Cf. TemplateInterpreterGenerator::generate_return_entry_for and
980 // InterpreterMacroAssembler::save_return_value
981 void MethodHandles::move_return_value(MacroAssembler* _masm, BasicType type,
982 Address return_slot) {
983 BLOCK_COMMENT("move_return_value {");
984 // Look at the type and pull the value out of the corresponding register.
985 if (type == T_VOID) {
986 // nothing to do
987 } else if (type == T_OBJECT) {
988 __ verify_oop(O0);
989 __ st_ptr(O0, return_slot);
990 } else if (type == T_INT || is_subword_type(type)) {
991 int type_size = type2aelembytes(T_INT);
992 __ store_sized_value(O0, return_slot, type_size);
993 } else if (type == T_LONG) {
994 // store the value by parts
995 // Note: We assume longs are continguous (if misaligned) on the interpreter stack.
996 #if !defined(_LP64) && defined(COMPILER2)
997 __ stx(G1, return_slot);
998 #else
999 #ifdef _LP64
1000 __ stx(O0, return_slot);
1001 #else
1002 if (return_slot.has_disp()) {
1003 // The displacement is a constant
1004 __ st(O0, return_slot);
1005 __ st(O1, return_slot.plus_disp(Interpreter::stackElementSize));
1006 } else {
1007 __ std(O0, return_slot);
1008 }
1009 #endif
1010 #endif
1011 } else if (type == T_FLOAT) {
1012 __ stf(FloatRegisterImpl::S, Ftos_f, return_slot);
1013 } else if (type == T_DOUBLE) {
1014 __ stf(FloatRegisterImpl::D, Ftos_f, return_slot);
1015 } else {
1016 ShouldNotReachHere();
1017 }
1018 BLOCK_COMMENT("} move_return_value");
1019 }
1020
1021 #ifndef PRODUCT
1022 extern "C" void print_method_handle(oop mh);
1023 void trace_method_handle_stub(const char* adaptername,
1024 oopDesc* mh,
1025 intptr_t* saved_sp) {
1026 bool has_mh = (strstr(adaptername, "return/") == NULL); // return adapters don't have mh
1027 tty->print_cr("MH %s mh="INTPTR_FORMAT " saved_sp=" INTPTR_FORMAT, adaptername, (intptr_t) mh, saved_sp);
1028 if (has_mh)
1029 print_method_handle(mh);
1030 }
1031 void MethodHandles::trace_method_handle(MacroAssembler* _masm, const char* adaptername) {
1032 if (!TraceMethodHandles) return;
1033 BLOCK_COMMENT("trace_method_handle {");
1034 // save: Gargs, O5_savedSP
1035 __ save_frame(16);
1036 __ set((intptr_t) adaptername, O0);
1037 __ mov(G3_method_handle, O1);
1038 __ mov(I5_savedSP, O2);
1039 __ mov(G3_method_handle, L3);
1040 __ mov(Gargs, L4);
1041 __ mov(G5_method_type, L5);
1042 __ call_VM_leaf(L7, CAST_FROM_FN_PTR(address, trace_method_handle_stub));
1043
1044 __ mov(L3, G3_method_handle);
1045 __ mov(L4, Gargs);
1046 __ mov(L5, G5_method_type);
1047 __ restore();
1048 BLOCK_COMMENT("} trace_method_handle");
1049 }
1050 #endif // PRODUCT
1051
1052 // which conversion op types are implemented here?
1053 int MethodHandles::adapter_conversion_ops_supported_mask() {
1054 return ((1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_ONLY)
1055 |(1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_RAW)
1056 |(1<<java_lang_invoke_AdapterMethodHandle::OP_CHECK_CAST)
1057 |(1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_PRIM)
1058 |(1<<java_lang_invoke_AdapterMethodHandle::OP_REF_TO_PRIM)
1059 // OP_PRIM_TO_REF is below...
1060 |(1<<java_lang_invoke_AdapterMethodHandle::OP_SWAP_ARGS)
1061 |(1<<java_lang_invoke_AdapterMethodHandle::OP_ROT_ARGS)
1062 |(1<<java_lang_invoke_AdapterMethodHandle::OP_DUP_ARGS)
1063 |(1<<java_lang_invoke_AdapterMethodHandle::OP_DROP_ARGS)
1064 // OP_COLLECT_ARGS is below...
1065 |(1<<java_lang_invoke_AdapterMethodHandle::OP_SPREAD_ARGS)
1066 |(!UseRicochetFrames ? 0 :
1067 java_lang_invoke_MethodTypeForm::vmlayout_offset_in_bytes() <= 0 ? 0 :
1068 ((1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF)
1069 |(1<<java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS)
1070 |(1<<java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS)
1071 )
1072 )
1073 );
1074 }
1075
1076 //------------------------------------------------------------------------------
1077 // MethodHandles::generate_method_handle_stub
1078 //
1079 // Generate an "entry" field for a method handle.
1080 // This determines how the method handle will respond to calls.
1081 void MethodHandles::generate_method_handle_stub(MacroAssembler* _masm, MethodHandles::EntryKind ek) {
1082 MethodHandles::EntryKind ek_orig = ek_original_kind(ek);
1083
1084 // Here is the register state during an interpreted call,
1085 // as set up by generate_method_handle_interpreter_entry():
1086 // - G5: garbage temp (was MethodHandle.invoke methodOop, unused)
1087 // - G3: receiver method handle
1088 // - O5_savedSP: sender SP (must preserve)
1089
1090 const Register O0_scratch = O0;
1091 const Register O1_scratch = O1;
1092 const Register O2_scratch = O2;
1093 const Register O3_scratch = O3;
1094 const Register O4_scratch = O4;
1095 const Register G5_scratch = G5;
1096
1097 // Often used names:
1098 const Register O0_argslot = O0;
1099
1100 // Argument registers for _raise_exception:
1101 const Register O0_code = O0;
1102 const Register O1_actual = O1;
1103 const Register O2_required = O2;
1104
1105 guarantee(java_lang_invoke_MethodHandle::vmentry_offset_in_bytes() != 0, "must have offsets");
1106
1107 // Some handy addresses:
1108 Address G5_method_fie( G5_method, in_bytes(methodOopDesc::from_interpreted_offset()));
1109 Address G5_method_fce( G5_method, in_bytes(methodOopDesc::from_compiled_offset()));
1110
1111 Address G3_mh_vmtarget( G3_method_handle, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes());
1112
1113 Address G3_dmh_vmindex( G3_method_handle, java_lang_invoke_DirectMethodHandle::vmindex_offset_in_bytes());
1114
1115 Address G3_bmh_vmargslot( G3_method_handle, java_lang_invoke_BoundMethodHandle::vmargslot_offset_in_bytes());
1116 Address G3_bmh_argument( G3_method_handle, java_lang_invoke_BoundMethodHandle::argument_offset_in_bytes());
1117
1118 Address G3_amh_vmargslot( G3_method_handle, java_lang_invoke_AdapterMethodHandle::vmargslot_offset_in_bytes());
1119 Address G3_amh_argument ( G3_method_handle, java_lang_invoke_AdapterMethodHandle::argument_offset_in_bytes());
1120 Address G3_amh_conversion(G3_method_handle, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
1121
1122 const int java_mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1123
1124 if (have_entry(ek)) {
1125 __ nop(); // empty stubs make SG sick
1126 return;
1127 }
1128
1129 address interp_entry = __ pc();
1130
1131 trace_method_handle(_masm, entry_name(ek));
1132
1133 BLOCK_COMMENT(err_msg("Entry %s {", entry_name(ek)));
1134
1135 switch ((int) ek) {
1136 case _raise_exception:
1137 {
1138 // Not a real MH entry, but rather shared code for raising an
1139 // exception. Since we use the compiled entry, arguments are
1140 // expected in compiler argument registers.
1141 assert(raise_exception_method(), "must be set");
1142 assert(raise_exception_method()->from_compiled_entry(), "method must be linked");
1143
1144 __ mov(O5_savedSP, SP); // Cut the stack back to where the caller started.
1145
1146 Label L_no_method;
1147 // FIXME: fill in _raise_exception_method with a suitable java.lang.invoke method
1148 __ set(AddressLiteral((address) &_raise_exception_method), G5_method);
1149 __ ld_ptr(Address(G5_method, 0), G5_method);
1150
1151 const int jobject_oop_offset = 0;
1152 __ ld_ptr(Address(G5_method, jobject_oop_offset), G5_method);
1153
1154 __ verify_oop(G5_method);
1155 __ jump_indirect_to(G5_method_fce, O3_scratch); // jump to compiled entry
1156 __ delayed()->nop();
1157 }
1158 break;
1159
1160 case _invokestatic_mh:
1161 case _invokespecial_mh:
1162 {
1163 __ load_heap_oop(G3_mh_vmtarget, G5_method); // target is a methodOop
1164 __ verify_oop(G5_method);
1165 // Same as TemplateTable::invokestatic or invokespecial,
1166 // minus the CP setup and profiling:
1167 if (ek == _invokespecial_mh) {
1168 // Must load & check the first argument before entering the target method.
1169 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1170 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1171 __ null_check(G3_method_handle);
1172 __ verify_oop(G3_method_handle);
1173 }
1174 __ jump_indirect_to(G5_method_fie, O1_scratch);
1175 __ delayed()->nop();
1176 }
1177 break;
1178
1179 case _invokevirtual_mh:
1180 {
1181 // Same as TemplateTable::invokevirtual,
1182 // minus the CP setup and profiling:
1183
1184 // Pick out the vtable index and receiver offset from the MH,
1185 // and then we can discard it:
1186 Register O2_index = O2_scratch;
1187 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1188 __ ldsw(G3_dmh_vmindex, O2_index);
1189 // Note: The verifier allows us to ignore G3_mh_vmtarget.
1190 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1191 __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1192
1193 // Get receiver klass:
1194 Register O0_klass = O0_argslot;
1195 __ load_klass(G3_method_handle, O0_klass);
1196 __ verify_oop(O0_klass);
1197
1198 // Get target methodOop & entry point:
1199 const int base = instanceKlass::vtable_start_offset() * wordSize;
1200 assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1201
1202 __ sll_ptr(O2_index, LogBytesPerWord, O2_index);
1203 __ add(O0_klass, O2_index, O0_klass);
1204 Address vtable_entry_addr(O0_klass, base + vtableEntry::method_offset_in_bytes());
1205 __ ld_ptr(vtable_entry_addr, G5_method);
1206
1207 __ verify_oop(G5_method);
1208 __ jump_indirect_to(G5_method_fie, O1_scratch);
1209 __ delayed()->nop();
1210 }
1211 break;
1212
1213 case _invokeinterface_mh:
1214 {
1215 // Same as TemplateTable::invokeinterface,
1216 // minus the CP setup and profiling:
1217 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1218 Register O1_intf = O1_scratch;
1219 Register G5_index = G5_scratch;
1220 __ load_heap_oop(G3_mh_vmtarget, O1_intf);
1221 __ ldsw(G3_dmh_vmindex, G5_index);
1222 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1223 __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1224
1225 // Get receiver klass:
1226 Register O0_klass = O0_argslot;
1227 __ load_klass(G3_method_handle, O0_klass);
1228 __ verify_oop(O0_klass);
1229
1230 // Get interface:
1231 Label no_such_interface;
1232 __ verify_oop(O1_intf);
1233 __ lookup_interface_method(O0_klass, O1_intf,
1234 // Note: next two args must be the same:
1235 G5_index, G5_method,
1236 O2_scratch,
1237 O3_scratch,
1238 no_such_interface);
1239
1240 __ verify_oop(G5_method);
1241 __ jump_indirect_to(G5_method_fie, O1_scratch);
1242 __ delayed()->nop();
1243
1244 __ bind(no_such_interface);
1245 // Throw an exception.
1246 // For historical reasons, it will be IncompatibleClassChangeError.
1247 __ unimplemented("not tested yet");
1248 __ ld_ptr(Address(O1_intf, java_mirror_offset), O2_required); // required interface
1249 __ mov( O0_klass, O1_actual); // bad receiver
1250 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1251 __ delayed()->mov(Bytecodes::_invokeinterface, O0_code); // who is complaining?
1252 }
1253 break;
1254
1255 case _bound_ref_mh:
1256 case _bound_int_mh:
1257 case _bound_long_mh:
1258 case _bound_ref_direct_mh:
1259 case _bound_int_direct_mh:
1260 case _bound_long_direct_mh:
1261 {
1262 const bool direct_to_method = (ek >= _bound_ref_direct_mh);
1263 BasicType arg_type = ek_bound_mh_arg_type(ek);
1264 int arg_slots = type2size[arg_type];
1265
1266 // Make room for the new argument:
1267 load_vmargslot(_masm, G3_bmh_vmargslot, O0_argslot);
1268 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1269
1270 insert_arg_slots(_masm, arg_slots * stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1271
1272 // Store bound argument into the new stack slot:
1273 __ load_heap_oop(G3_bmh_argument, O1_scratch);
1274 if (arg_type == T_OBJECT) {
1275 __ st_ptr(O1_scratch, Address(O0_argslot, 0));
1276 } else {
1277 Address prim_value_addr(O1_scratch, java_lang_boxing_object::value_offset_in_bytes(arg_type));
1278 move_typed_arg(_masm, arg_type, false,
1279 prim_value_addr,
1280 Address(O0_argslot, 0),
1281 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
1282 }
1283
1284 if (direct_to_method) {
1285 __ load_heap_oop(G3_mh_vmtarget, G5_method); // target is a methodOop
1286 __ verify_oop(G5_method);
1287 __ jump_indirect_to(G5_method_fie, O1_scratch);
1288 __ delayed()->nop();
1289 } else {
1290 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle); // target is a methodOop
1291 __ verify_oop(G3_method_handle);
1292 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1293 }
1294 }
1295 break;
1296
1297 case _adapter_retype_only:
1298 case _adapter_retype_raw:
1299 // Immediately jump to the next MH layer:
1300 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1301 __ verify_oop(G3_method_handle);
1302 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1303 // This is OK when all parameter types widen.
1304 // It is also OK when a return type narrows.
1305 break;
1306
1307 case _adapter_check_cast:
1308 {
1309 // Check a reference argument before jumping to the next layer of MH:
1310 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1311 Address vmarg = __ argument_address(O0_argslot, O0_argslot);
1312
1313 // What class are we casting to?
1314 Register O1_klass = O1_scratch; // Interesting AMH data.
1315 __ load_heap_oop(G3_amh_argument, O1_klass); // This is a Class object!
1316 load_klass_from_Class(_masm, O1_klass, O2_scratch, O3_scratch);
1317
1318 Label L_done;
1319 __ ld_ptr(vmarg, O2_scratch);
1320 __ tst(O2_scratch);
1321 __ brx(Assembler::zero, false, Assembler::pn, L_done); // No cast if null.
1322 __ delayed()->nop();
1323 __ load_klass(O2_scratch, O2_scratch);
1324
1325 // Live at this point:
1326 // - O0_argslot : argslot index in vmarg; may be required in the failing path
1327 // - O1_klass : klass required by the target method
1328 // - O2_scratch : argument klass to test
1329 // - G3_method_handle: adapter method handle
1330 __ check_klass_subtype(O2_scratch, O1_klass, O3_scratch, O4_scratch, L_done);
1331
1332 // If we get here, the type check failed!
1333 __ load_heap_oop(G3_amh_argument, O2_required); // required class
1334 __ ld_ptr( vmarg, O1_actual); // bad object
1335 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1336 __ delayed()->mov(Bytecodes::_checkcast, O0_code); // who is complaining?
1337
1338 __ BIND(L_done);
1339 // Get the new MH:
1340 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1341 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1342 }
1343 break;
1344
1345 case _adapter_prim_to_prim:
1346 case _adapter_ref_to_prim:
1347 // Handled completely by optimized cases.
1348 __ stop("init_AdapterMethodHandle should not issue this");
1349 break;
1350
1351 case _adapter_opt_i2i: // optimized subcase of adapt_prim_to_prim
1352 //case _adapter_opt_f2i: // optimized subcase of adapt_prim_to_prim
1353 case _adapter_opt_l2i: // optimized subcase of adapt_prim_to_prim
1354 case _adapter_opt_unboxi: // optimized subcase of adapt_ref_to_prim
1355 {
1356 // Perform an in-place conversion to int or an int subword.
1357 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1358 Address value;
1359 Address vmarg;
1360 bool value_left_justified = false;
1361
1362 switch (ek) {
1363 case _adapter_opt_i2i:
1364 value = vmarg = __ argument_address(O0_argslot, O0_argslot);
1365 break;
1366 case _adapter_opt_l2i:
1367 {
1368 // just delete the extra slot
1369 #ifdef _LP64
1370 // In V9, longs are given 2 64-bit slots in the interpreter, but the
1371 // data is passed in only 1 slot.
1372 // Keep the second slot.
1373 __ add(__ argument_address(O0_argslot, O0_argslot, -1), O0_argslot);
1374 remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1375 value = Address(O0_argslot, 4); // Get least-significant 32-bit of 64-bit value.
1376 vmarg = Address(O0_argslot, Interpreter::stackElementSize);
1377 #else
1378 // Keep the first slot.
1379 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1380 remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1381 value = Address(O0_argslot, 0);
1382 vmarg = value;
1383 #endif
1384 }
1385 break;
1386 case _adapter_opt_unboxi:
1387 {
1388 vmarg = __ argument_address(O0_argslot, O0_argslot);
1389 // Load the value up from the heap.
1390 __ ld_ptr(vmarg, O1_scratch);
1391 int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_INT);
1392 #ifdef ASSERT
1393 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1394 if (is_subword_type(BasicType(bt)))
1395 assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(BasicType(bt)), "");
1396 }
1397 #endif
1398 __ null_check(O1_scratch, value_offset);
1399 value = Address(O1_scratch, value_offset);
1400 #ifdef _BIG_ENDIAN
1401 // Values stored in objects are packed.
1402 value_left_justified = true;
1403 #endif
1404 }
1405 break;
1406 default:
1407 ShouldNotReachHere();
1408 }
1409
1410 // This check is required on _BIG_ENDIAN
1411 Register G5_vminfo = G5_scratch;
1412 __ ldsw(G3_amh_conversion, G5_vminfo);
1413 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
1414
1415 // Original 32-bit vmdata word must be of this form:
1416 // | MBZ:6 | signBitCount:8 | srcDstTypes:8 | conversionOp:8 |
1417 __ lduw(value, O1_scratch);
1418 if (!value_left_justified)
1419 __ sll(O1_scratch, G5_vminfo, O1_scratch);
1420 Label zero_extend, done;
1421 __ btst(CONV_VMINFO_SIGN_FLAG, G5_vminfo);
1422 __ br(Assembler::zero, false, Assembler::pn, zero_extend);
1423 __ delayed()->nop();
1424
1425 // this path is taken for int->byte, int->short
1426 __ sra(O1_scratch, G5_vminfo, O1_scratch);
1427 __ ba(false, done);
1428 __ delayed()->nop();
1429
1430 __ bind(zero_extend);
1431 // this is taken for int->char
1432 __ srl(O1_scratch, G5_vminfo, O1_scratch);
1433
1434 __ bind(done);
1435 __ st(O1_scratch, vmarg);
1436
1437 // Get the new MH:
1438 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1439 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1440 }
1441 break;
1442
1443 case _adapter_opt_i2l: // optimized subcase of adapt_prim_to_prim
1444 case _adapter_opt_unboxl: // optimized subcase of adapt_ref_to_prim
1445 {
1446 // Perform an in-place int-to-long or ref-to-long conversion.
1447 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1448
1449 // On big-endian machine we duplicate the slot and store the MSW
1450 // in the first slot.
1451 __ add(__ argument_address(O0_argslot, O0_argslot, 1), O0_argslot);
1452
1453 insert_arg_slots(_masm, stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1454
1455 Address arg_lsw(O0_argslot, 0);
1456 Address arg_msw(O0_argslot, -Interpreter::stackElementSize);
1457
1458 switch (ek) {
1459 case _adapter_opt_i2l:
1460 {
1461 #ifdef _LP64
1462 __ ldsw(arg_lsw, O2_scratch); // Load LSW sign-extended
1463 #else
1464 __ ldsw(arg_lsw, O3_scratch); // Load LSW sign-extended
1465 __ srlx(O3_scratch, BitsPerInt, O2_scratch); // Move MSW value to lower 32-bits for std
1466 #endif
1467 __ st_long(O2_scratch, arg_msw); // Uses O2/O3 on !_LP64
1468 }
1469 break;
1470 case _adapter_opt_unboxl:
1471 {
1472 // Load the value up from the heap.
1473 __ ld_ptr(arg_lsw, O1_scratch);
1474 int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_LONG);
1475 assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(T_DOUBLE), "");
1476 __ null_check(O1_scratch, value_offset);
1477 __ ld_long(Address(O1_scratch, value_offset), O2_scratch); // Uses O2/O3 on !_LP64
1478 __ st_long(O2_scratch, arg_msw);
1479 }
1480 break;
1481 default:
1482 ShouldNotReachHere();
1483 }
1484
1485 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1486 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1487 }
1488 break;
1489
1490 case _adapter_opt_f2d: // optimized subcase of adapt_prim_to_prim
1491 case _adapter_opt_d2f: // optimized subcase of adapt_prim_to_prim
1492 {
1493 // perform an in-place floating primitive conversion
1494 __ unimplemented(entry_name(ek));
1495 }
1496 break;
1497
1498 case _adapter_prim_to_ref:
1499 __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
1500 break;
1501
1502 case _adapter_swap_args:
1503 case _adapter_rot_args:
1504 // handled completely by optimized cases
1505 __ stop("init_AdapterMethodHandle should not issue this");
1506 break;
1507
1508 case _adapter_opt_swap_1:
1509 case _adapter_opt_swap_2:
1510 case _adapter_opt_rot_1_up:
1511 case _adapter_opt_rot_1_down:
1512 case _adapter_opt_rot_2_up:
1513 case _adapter_opt_rot_2_down:
1514 {
1515 int swap_slots = ek_adapter_opt_swap_slots(ek);
1516 int rotate = ek_adapter_opt_swap_mode(ek);
1517
1518 // 'argslot' is the position of the first argument to swap.
1519 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1520 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1521 if (VerifyMethodHandles)
1522 verify_argslot(_masm, O0_argslot, O2_scratch, "swap point must fall within current frame");
1523
1524 // 'vminfo' is the second.
1525 Register O1_destslot = O1_scratch;
1526 load_conversion_vminfo(_masm, G3_amh_conversion, O1_destslot);
1527 __ add(__ argument_address(O1_destslot, O1_destslot), O1_destslot);
1528 if (VerifyMethodHandles)
1529 verify_argslot(_masm, O1_destslot, O2_scratch, "swap point must fall within current frame");
1530
1531 assert(Interpreter::stackElementSize == wordSize, "else rethink use of wordSize here");
1532 if (!rotate) {
1533 // simple swap
1534 for (int i = 0; i < swap_slots; i++) {
1535 __ ld_ptr( Address(O0_argslot, i * wordSize), O2_scratch);
1536 __ ld_ptr( Address(O1_destslot, i * wordSize), O3_scratch);
1537 __ st_ptr(O3_scratch, Address(O0_argslot, i * wordSize));
1538 __ st_ptr(O2_scratch, Address(O1_destslot, i * wordSize));
1539 }
1540 } else {
1541 // A rotate is actually pair of moves, with an "odd slot" (or pair)
1542 // changing place with a series of other slots.
1543 // First, push the "odd slot", which is going to get overwritten
1544 switch (swap_slots) {
1545 case 2 : __ ld_ptr(Address(O0_argslot, 1 * wordSize), O4_scratch); // fall-thru
1546 case 1 : __ ld_ptr(Address(O0_argslot, 0 * wordSize), O3_scratch); break;
1547 default: ShouldNotReachHere();
1548 }
1549 if (rotate > 0) {
1550 // Here is rotate > 0:
1551 // (low mem) (high mem)
1552 // | dest: more_slots... | arg: odd_slot :arg+1 |
1553 // =>
1554 // | dest: odd_slot | dest+1: more_slots... :arg+1 |
1555 // work argslot down to destslot, copying contiguous data upwards
1556 // pseudo-code:
1557 // argslot = src_addr - swap_bytes
1558 // destslot = dest_addr
1559 // while (argslot >= destslot) *(argslot + swap_bytes) = *(argslot + 0), argslot--;
1560 move_arg_slots_up(_masm,
1561 O1_destslot,
1562 Address(O0_argslot, 0),
1563 swap_slots,
1564 O0_argslot, O2_scratch);
1565 } else {
1566 // Here is the other direction, rotate < 0:
1567 // (low mem) (high mem)
1568 // | arg: odd_slot | arg+1: more_slots... :dest+1 |
1569 // =>
1570 // | arg: more_slots... | dest: odd_slot :dest+1 |
1571 // work argslot up to destslot, copying contiguous data downwards
1572 // pseudo-code:
1573 // argslot = src_addr + swap_bytes
1574 // destslot = dest_addr
1575 // while (argslot <= destslot) *(argslot - swap_bytes) = *(argslot + 0), argslot++;
1576 // dest_slot denotes an exclusive upper limit
1577 int limit_bias = OP_ROT_ARGS_DOWN_LIMIT_BIAS;
1578 if (limit_bias != 0)
1579 __ add(O1_destslot, - limit_bias * wordSize, O1_destslot);
1580 move_arg_slots_down(_masm,
1581 Address(O0_argslot, swap_slots * wordSize),
1582 O1_destslot,
1583 -swap_slots,
1584 O0_argslot, O2_scratch);
1585
1586 __ sub(O1_destslot, swap_slots * wordSize, O1_destslot);
1587 }
1588 // pop the original first chunk into the destination slot, now free
1589 switch (swap_slots) {
1590 case 2 : __ st_ptr(O4_scratch, Address(O1_destslot, 1 * wordSize)); // fall-thru
1591 case 1 : __ st_ptr(O3_scratch, Address(O1_destslot, 0 * wordSize)); break;
1592 default: ShouldNotReachHere();
1593 }
1594 }
1595
1596 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1597 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1598 }
1599 break;
1600
1601 case _adapter_dup_args:
1602 {
1603 // 'argslot' is the position of the first argument to duplicate.
1604 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1605 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1606
1607 // 'stack_move' is negative number of words to duplicate.
1608 Register O1_stack_move = O1_scratch;
1609 load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1610
1611 if (VerifyMethodHandles) {
1612 verify_argslots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, true,
1613 "copied argument(s) must fall within current frame");
1614 }
1615
1616 // insert location is always the bottom of the argument list:
1617 __ neg(O1_stack_move);
1618 push_arg_slots(_masm, O0_argslot, O1_stack_move, O2_scratch, O3_scratch);
1619
1620 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1621 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1622 }
1623 break;
1624
1625 case _adapter_drop_args:
1626 {
1627 // 'argslot' is the position of the first argument to nuke.
1628 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1629 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1630
1631 // 'stack_move' is number of words to drop.
1632 Register O1_stack_move = O1_scratch;
1633 load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1634
1635 remove_arg_slots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, O4_scratch);
1636
1637 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1638 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1639 }
1640 break;
1641
1642 case _adapter_collect_args:
1643 case _adapter_fold_args:
1644 case _adapter_spread_args:
1645 // Handled completely by optimized cases.
1646 __ stop("init_AdapterMethodHandle should not issue this");
1647 break;
1648
1649 case _adapter_opt_collect_ref:
1650 case _adapter_opt_collect_int:
1651 case _adapter_opt_collect_long:
1652 case _adapter_opt_collect_float:
1653 case _adapter_opt_collect_double:
1654 case _adapter_opt_collect_void:
1655 case _adapter_opt_collect_0_ref:
1656 case _adapter_opt_collect_1_ref:
1657 case _adapter_opt_collect_2_ref:
1658 case _adapter_opt_collect_3_ref:
1659 case _adapter_opt_collect_4_ref:
1660 case _adapter_opt_collect_5_ref:
1661 case _adapter_opt_filter_S0_ref:
1662 case _adapter_opt_filter_S1_ref:
1663 case _adapter_opt_filter_S2_ref:
1664 case _adapter_opt_filter_S3_ref:
1665 case _adapter_opt_filter_S4_ref:
1666 case _adapter_opt_filter_S5_ref:
1667 case _adapter_opt_collect_2_S0_ref:
1668 case _adapter_opt_collect_2_S1_ref:
1669 case _adapter_opt_collect_2_S2_ref:
1670 case _adapter_opt_collect_2_S3_ref:
1671 case _adapter_opt_collect_2_S4_ref:
1672 case _adapter_opt_collect_2_S5_ref:
1673 case _adapter_opt_fold_ref:
1674 case _adapter_opt_fold_int:
1675 case _adapter_opt_fold_long:
1676 case _adapter_opt_fold_float:
1677 case _adapter_opt_fold_double:
1678 case _adapter_opt_fold_void:
1679 case _adapter_opt_fold_1_ref:
1680 case _adapter_opt_fold_2_ref:
1681 case _adapter_opt_fold_3_ref:
1682 case _adapter_opt_fold_4_ref:
1683 case _adapter_opt_fold_5_ref:
1684 {
1685 // Given a fresh incoming stack frame, build a new ricochet frame.
1686 // On entry, TOS points at a return PC, and FP is the callers frame ptr.
1687 // RSI/R13 has the caller's exact stack pointer, which we must also preserve.
1688 // RCX contains an AdapterMethodHandle of the indicated kind.
1689
1690 // Relevant AMH fields:
1691 // amh.vmargslot:
1692 // points to the trailing edge of the arguments
1693 // to filter, collect, or fold. For a boxing operation,
1694 // it points just after the single primitive value.
1695 // amh.argument:
1696 // recursively called MH, on |collect| arguments
1697 // amh.vmtarget:
1698 // final destination MH, on return value, etc.
1699 // amh.conversion.dest:
1700 // tells what is the type of the return value
1701 // (not needed here, since dest is also derived from ek)
1702 // amh.conversion.vminfo:
1703 // points to the trailing edge of the return value
1704 // when the vmtarget is to be called; this is
1705 // equal to vmargslot + (retained ? |collect| : 0)
1706
1707 // Pass 0 or more argument slots to the recursive target.
1708 int collect_count_constant = ek_adapter_opt_collect_count(ek);
1709
1710 // The collected arguments are copied from the saved argument list:
1711 int collect_slot_constant = ek_adapter_opt_collect_slot(ek);
1712
1713 assert(ek_orig == _adapter_collect_args ||
1714 ek_orig == _adapter_fold_args, "");
1715 bool retain_original_args = (ek_orig == _adapter_fold_args);
1716
1717 // The return value is replaced (or inserted) at the 'vminfo' argslot.
1718 // Sometimes we can compute this statically.
1719 int dest_slot_constant = -1;
1720 if (!retain_original_args)
1721 dest_slot_constant = collect_slot_constant;
1722 else if (collect_slot_constant >= 0 && collect_count_constant >= 0)
1723 // We are preserving all the arguments, and the return value is prepended,
1724 // so the return slot is to the left (above) the |collect| sequence.
1725 dest_slot_constant = collect_slot_constant + collect_count_constant;
1726
1727 // Replace all those slots by the result of the recursive call.
1728 // The result type can be one of ref, int, long, float, double, void.
1729 // In the case of void, nothing is pushed on the stack after return.
1730 BasicType dest = ek_adapter_opt_collect_type(ek);
1731 assert(dest == type2wfield[dest], "dest is a stack slot type");
1732 int dest_count = type2size[dest];
1733 assert(dest_count == 1 || dest_count == 2 || (dest_count == 0 && dest == T_VOID), "dest has a size");
1734
1735 // Choose a return continuation.
1736 EntryKind ek_ret = _adapter_opt_return_any;
1737 if (dest != T_CONFLICT && OptimizeMethodHandles) {
1738 switch (dest) {
1739 case T_INT : ek_ret = _adapter_opt_return_int; break;
1740 case T_LONG : ek_ret = _adapter_opt_return_long; break;
1741 case T_FLOAT : ek_ret = _adapter_opt_return_float; break;
1742 case T_DOUBLE : ek_ret = _adapter_opt_return_double; break;
1743 case T_OBJECT : ek_ret = _adapter_opt_return_ref; break;
1744 case T_VOID : ek_ret = _adapter_opt_return_void; break;
1745 default : ShouldNotReachHere();
1746 }
1747 if (dest == T_OBJECT && dest_slot_constant >= 0) {
1748 EntryKind ek_try = EntryKind(_adapter_opt_return_S0_ref + dest_slot_constant);
1749 if (ek_try <= _adapter_opt_return_LAST &&
1750 ek_adapter_opt_return_slot(ek_try) == dest_slot_constant) {
1751 ek_ret = ek_try;
1752 }
1753 }
1754 assert(ek_adapter_opt_return_type(ek_ret) == dest, "");
1755 }
1756
1757 // Already pushed: ... keep1 | collect | keep2 |
1758
1759 // Push a few extra argument words, if we need them to store the return value.
1760 {
1761 int extra_slots = 0;
1762 if (retain_original_args) {
1763 extra_slots = dest_count;
1764 } else if (collect_count_constant == -1) {
1765 extra_slots = dest_count; // collect_count might be zero; be generous
1766 } else if (dest_count > collect_count_constant) {
1767 extra_slots = (dest_count - collect_count_constant);
1768 } else {
1769 // else we know we have enough dead space in |collect| to repurpose for return values
1770 }
1771 if (extra_slots != 0) {
1772 __ sub(SP, round_to(extra_slots, 2) * Interpreter::stackElementSize, SP);
1773 }
1774 }
1775
1776 // Set up Ricochet Frame.
1777 __ mov(SP, O5_savedSP); // record SP for the callee
1778
1779 // One extra (empty) slot for outgoing target MH (see Gargs computation below).
1780 __ save_frame(2); // Note: we need to add 2 slots since frame::memory_parameter_word_sp_offset is 23.
1781
1782 // Note: Gargs is live throughout the following, until we make our recursive call.
1783 // And the RF saves a copy in L4_saved_args_base.
1784
1785 RicochetFrame::enter_ricochet_frame(_masm, G3_method_handle, Gargs,
1786 entry(ek_ret)->from_interpreted_entry());
1787
1788 // Compute argument base:
1789 // Set up Gargs for current frame, extra (empty) slot is for outgoing target MH (space reserved by save_frame above).
1790 __ add(FP, STACK_BIAS - (1 * Interpreter::stackElementSize), Gargs);
1791
1792 // Now pushed: ... keep1 | collect | keep2 | extra | [RF]
1793
1794 #ifdef ASSERT
1795 if (VerifyMethodHandles && dest != T_CONFLICT) {
1796 BLOCK_COMMENT("verify AMH.conv.dest {");
1797 extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O1_scratch);
1798 Label L_dest_ok;
1799 __ cmp(O1_scratch, (int) dest);
1800 __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1801 __ delayed()->nop();
1802 if (dest == T_INT) {
1803 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1804 if (is_subword_type(BasicType(bt))) {
1805 __ cmp(O1_scratch, (int) bt);
1806 __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1807 __ delayed()->nop();
1808 }
1809 }
1810 }
1811 __ stop("bad dest in AMH.conv");
1812 __ BIND(L_dest_ok);
1813 BLOCK_COMMENT("} verify AMH.conv.dest");
1814 }
1815 #endif //ASSERT
1816
1817 // Find out where the original copy of the recursive argument sequence begins.
1818 Register O0_coll = O0_scratch;
1819 {
1820 RegisterOrConstant collect_slot = collect_slot_constant;
1821 if (collect_slot_constant == -1) {
1822 load_vmargslot(_masm, G3_amh_vmargslot, O1_scratch);
1823 collect_slot = O1_scratch;
1824 }
1825 // collect_slot might be 0, but we need the move anyway.
1826 __ add(RicochetFrame::L4_saved_args_base, __ argument_offset(collect_slot, collect_slot.register_or_noreg()), O0_coll);
1827 // O0_coll now points at the trailing edge of |collect| and leading edge of |keep2|
1828 }
1829
1830 // Replace the old AMH with the recursive MH. (No going back now.)
1831 // In the case of a boxing call, the recursive call is to a 'boxer' method,
1832 // such as Integer.valueOf or Long.valueOf. In the case of a filter
1833 // or collect call, it will take one or more arguments, transform them,
1834 // and return some result, to store back into argument_base[vminfo].
1835 __ load_heap_oop(G3_amh_argument, G3_method_handle);
1836 if (VerifyMethodHandles) verify_method_handle(_masm, G3_method_handle, O1_scratch, O2_scratch);
1837
1838 // Calculate |collect|, the number of arguments we are collecting.
1839 Register O1_collect_count = O1_scratch;
1840 RegisterOrConstant collect_count;
1841 if (collect_count_constant < 0) {
1842 __ load_method_handle_vmslots(O1_collect_count, G3_method_handle, O2_scratch);
1843 collect_count = O1_collect_count;
1844 } else {
1845 collect_count = collect_count_constant;
1846 #ifdef ASSERT
1847 if (VerifyMethodHandles) {
1848 BLOCK_COMMENT("verify collect_count_constant {");
1849 __ load_method_handle_vmslots(O3_scratch, G3_method_handle, O2_scratch);
1850 Label L_count_ok;
1851 __ cmp(O3_scratch, collect_count_constant);
1852 __ br(Assembler::equal, false, Assembler::pt, L_count_ok);
1853 __ delayed()->nop();
1854 __ stop("bad vminfo in AMH.conv");
1855 __ BIND(L_count_ok);
1856 BLOCK_COMMENT("} verify collect_count_constant");
1857 }
1858 #endif //ASSERT
1859 }
1860
1861 // copy |collect| slots directly to TOS:
1862 push_arg_slots(_masm, O0_coll, collect_count, O2_scratch, O3_scratch);
1863 // Now pushed: ... keep1 | collect | keep2 | RF... | collect |
1864 // O0_coll still points at the trailing edge of |collect| and leading edge of |keep2|
1865
1866 // If necessary, adjust the saved arguments to make room for the eventual return value.
1867 // Normal adjustment: ... keep1 | +dest+ | -collect- | keep2 | RF... | collect |
1868 // If retaining args: ... keep1 | +dest+ | collect | keep2 | RF... | collect |
1869 // In the non-retaining case, this might move keep2 either up or down.
1870 // We don't have to copy the whole | RF... collect | complex,
1871 // but we must adjust RF.saved_args_base.
1872 // Also, from now on, we will forget about the original copy of |collect|.
1873 // If we are retaining it, we will treat it as part of |keep2|.
1874 // For clarity we will define |keep3| = |collect|keep2| or |keep2|.
1875
1876 BLOCK_COMMENT("adjust trailing arguments {");
1877 // Compare the sizes of |+dest+| and |-collect-|, which are opposed opening and closing movements.
1878 int open_count = dest_count;
1879 RegisterOrConstant close_count = collect_count_constant;
1880 Register O1_close_count = O1_collect_count;
1881 if (retain_original_args) {
1882 close_count = constant(0);
1883 } else if (collect_count_constant == -1) {
1884 close_count = O1_collect_count;
1885 }
1886
1887 // How many slots need moving? This is simply dest_slot (0 => no |keep3|).
1888 RegisterOrConstant keep3_count;
1889 Register O2_keep3_count = O2_scratch;
1890 if (dest_slot_constant < 0) {
1891 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O2_keep3_count);
1892 keep3_count = O2_keep3_count;
1893 } else {
1894 keep3_count = dest_slot_constant;
1895 #ifdef ASSERT
1896 if (VerifyMethodHandles && dest_slot_constant < 0) {
1897 BLOCK_COMMENT("verify dest_slot_constant {");
1898 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O3_scratch);
1899 Label L_vminfo_ok;
1900 __ cmp(O3_scratch, dest_slot_constant);
1901 __ br(Assembler::equal, false, Assembler::pt, L_vminfo_ok);
1902 __ delayed()->nop();
1903 __ stop("bad vminfo in AMH.conv");
1904 __ BIND(L_vminfo_ok);
1905 BLOCK_COMMENT("} verify dest_slot_constant");
1906 }
1907 #endif //ASSERT
1908 }
1909
1910 // tasks remaining:
1911 bool move_keep3 = (!keep3_count.is_constant() || keep3_count.as_constant() != 0);
1912 bool stomp_dest = (NOT_DEBUG(dest == T_OBJECT) DEBUG_ONLY(dest_count != 0));
1913 bool fix_arg_base = (!close_count.is_constant() || open_count != close_count.as_constant());
1914
1915 // Old and new argument locations (based at slot 0).
1916 // Net shift (&new_argv - &old_argv) is (close_count - open_count).
1917 bool zero_open_count = (open_count == 0); // remember this bit of info
1918 if (move_keep3 && fix_arg_base) {
1919 // It will be easier to have everything in one register:
1920 if (close_count.is_register()) {
1921 // Deduct open_count from close_count register to get a clean +/- value.
1922 __ sub(close_count.as_register(), open_count, close_count.as_register());
1923 } else {
1924 close_count = close_count.as_constant() - open_count;
1925 }
1926 open_count = 0;
1927 }
1928 Register L4_old_argv = RicochetFrame::L4_saved_args_base;
1929 Register O3_new_argv = O3_scratch;
1930 if (fix_arg_base) {
1931 __ add(L4_old_argv, __ argument_offset(close_count, O4_scratch), O3_new_argv,
1932 -(open_count * Interpreter::stackElementSize));
1933 }
1934
1935 // First decide if any actual data are to be moved.
1936 // We can skip if (a) |keep3| is empty, or (b) the argument list size didn't change.
1937 // (As it happens, all movements involve an argument list size change.)
1938
1939 // If there are variable parameters, use dynamic checks to skip around the whole mess.
1940 Label L_done;
1941 if (keep3_count.is_register()) {
1942 __ tst(keep3_count.as_register());
1943 __ br(Assembler::zero, false, Assembler::pn, L_done);
1944 __ delayed()->nop();
1945 }
1946 if (close_count.is_register()) {
1947 __ cmp(close_count.as_register(), open_count);
1948 __ br(Assembler::equal, false, Assembler::pn, L_done);
1949 __ delayed()->nop();
1950 }
1951
1952 if (move_keep3 && fix_arg_base) {
1953 bool emit_move_down = false, emit_move_up = false, emit_guard = false;
1954 if (!close_count.is_constant()) {
1955 emit_move_down = emit_guard = !zero_open_count;
1956 emit_move_up = true;
1957 } else if (open_count != close_count.as_constant()) {
1958 emit_move_down = (open_count > close_count.as_constant());
1959 emit_move_up = !emit_move_down;
1960 }
1961 Label L_move_up;
1962 if (emit_guard) {
1963 __ cmp(close_count.as_register(), open_count);
1964 __ br(Assembler::greater, false, Assembler::pn, L_move_up);
1965 __ delayed()->nop();
1966 }
1967
1968 if (emit_move_down) {
1969 // Move arguments down if |+dest+| > |-collect-|
1970 // (This is rare, except when arguments are retained.)
1971 // This opens space for the return value.
1972 if (keep3_count.is_constant()) {
1973 for (int i = 0; i < keep3_count.as_constant(); i++) {
1974 __ ld_ptr( Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
1975 __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize) );
1976 }
1977 } else {
1978 // Live: O1_close_count, O2_keep3_count, O3_new_argv
1979 Register argv_top = O0_scratch;
1980 __ add(L4_old_argv, __ argument_offset(keep3_count, O4_scratch), argv_top);
1981 move_arg_slots_down(_masm,
1982 Address(L4_old_argv, 0), // beginning of old argv
1983 argv_top, // end of old argv
1984 close_count, // distance to move down (must be negative)
1985 O4_scratch, G5_scratch);
1986 }
1987 }
1988
1989 if (emit_guard) {
1990 __ ba(false, L_done); // assumes emit_move_up is true also
1991 __ delayed()->nop();
1992 __ BIND(L_move_up);
1993 }
1994
1995 if (emit_move_up) {
1996 // Move arguments up if |+dest+| < |-collect-|
1997 // (This is usual, except when |keep3| is empty.)
1998 // This closes up the space occupied by the now-deleted collect values.
1999 if (keep3_count.is_constant()) {
2000 for (int i = keep3_count.as_constant() - 1; i >= 0; i--) {
2001 __ ld_ptr( Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
2002 __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize) );
2003 }
2004 } else {
2005 Address argv_top(L4_old_argv, __ argument_offset(keep3_count, O4_scratch));
2006 // Live: O1_close_count, O2_keep3_count, O3_new_argv
2007 move_arg_slots_up(_masm,
2008 L4_old_argv, // beginning of old argv
2009 argv_top, // end of old argv
2010 close_count, // distance to move up (must be positive)
2011 O4_scratch, G5_scratch);
2012 }
2013 }
2014 }
2015 __ BIND(L_done);
2016
2017 if (fix_arg_base) {
2018 // adjust RF.saved_args_base
2019 __ mov(O3_new_argv, RicochetFrame::L4_saved_args_base);
2020 }
2021
2022 if (stomp_dest) {
2023 // Stomp the return slot, so it doesn't hold garbage.
2024 // This isn't strictly necessary, but it may help detect bugs.
2025 __ set(RicochetFrame::RETURN_VALUE_PLACEHOLDER, O4_scratch);
2026 __ st_ptr(O4_scratch, Address(RicochetFrame::L4_saved_args_base,
2027 __ argument_offset(keep3_count, keep3_count.register_or_noreg()))); // uses O2_keep3_count
2028 }
2029 BLOCK_COMMENT("} adjust trailing arguments");
2030
2031 BLOCK_COMMENT("do_recursive_call");
2032 __ mov(SP, O5_savedSP); // record SP for the callee
2033 __ set(ExternalAddress(SharedRuntime::ricochet_blob()->bounce_addr() - frame::pc_return_offset), O7);
2034 // The globally unique bounce address has two purposes:
2035 // 1. It helps the JVM recognize this frame (frame::is_ricochet_frame).
2036 // 2. When returned to, it cuts back the stack and redirects control flow
2037 // to the return handler.
2038 // The return handler will further cut back the stack when it takes
2039 // down the RF. Perhaps there is a way to streamline this further.
2040
2041 // State during recursive call:
2042 // ... keep1 | dest | dest=42 | keep3 | RF... | collect | bounce_pc |
2043 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2044 }
2045 break;
2046
2047 case _adapter_opt_return_ref:
2048 case _adapter_opt_return_int:
2049 case _adapter_opt_return_long:
2050 case _adapter_opt_return_float:
2051 case _adapter_opt_return_double:
2052 case _adapter_opt_return_void:
2053 case _adapter_opt_return_S0_ref:
2054 case _adapter_opt_return_S1_ref:
2055 case _adapter_opt_return_S2_ref:
2056 case _adapter_opt_return_S3_ref:
2057 case _adapter_opt_return_S4_ref:
2058 case _adapter_opt_return_S5_ref:
2059 {
2060 BasicType dest_type_constant = ek_adapter_opt_return_type(ek);
2061 int dest_slot_constant = ek_adapter_opt_return_slot(ek);
2062
2063 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
2064
2065 if (dest_slot_constant == -1) {
2066 // The current stub is a general handler for this dest_type.
2067 // It can be called from _adapter_opt_return_any below.
2068 // Stash the address in a little table.
2069 assert((dest_type_constant & CONV_TYPE_MASK) == dest_type_constant, "oob");
2070 address return_handler = __ pc();
2071 _adapter_return_handlers[dest_type_constant] = return_handler;
2072 if (dest_type_constant == T_INT) {
2073 // do the subword types too
2074 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
2075 if (is_subword_type(BasicType(bt)) &&
2076 _adapter_return_handlers[bt] == NULL) {
2077 _adapter_return_handlers[bt] = return_handler;
2078 }
2079 }
2080 }
2081 }
2082
2083 // On entry to this continuation handler, make Gargs live again.
2084 __ mov(RicochetFrame::L4_saved_args_base, Gargs);
2085
2086 Register O7_temp = O7;
2087 Register O5_vminfo = O5;
2088
2089 RegisterOrConstant dest_slot = dest_slot_constant;
2090 if (dest_slot_constant == -1) {
2091 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O5_vminfo);
2092 dest_slot = O5_vminfo;
2093 }
2094 // Store the result back into the argslot.
2095 // This code uses the interpreter calling sequence, in which the return value
2096 // is usually left in the TOS register, as defined by InterpreterMacroAssembler::pop.
2097 // There are certain irregularities with floating point values, which can be seen
2098 // in TemplateInterpreterGenerator::generate_return_entry_for.
2099 move_return_value(_masm, dest_type_constant, __ argument_address(dest_slot, O7_temp));
2100
2101 RicochetFrame::leave_ricochet_frame(_masm, G3_method_handle, I5_savedSP, I7);
2102
2103 // Load the final target and go.
2104 if (VerifyMethodHandles) verify_method_handle(_masm, G3_method_handle, O0_scratch, O1_scratch);
2105 __ restore(I5_savedSP, G0, SP);
2106 __ jump_to_method_handle_entry(G3_method_handle, O0_scratch);
2107 __ illtrap(0);
2108 }
2109 break;
2110
2111 case _adapter_opt_return_any:
2112 {
2113 Register O7_temp = O7;
2114 Register O5_dest_type = O5;
2115
2116 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
2117 extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O5_dest_type);
2118 __ set(ExternalAddress((address) &_adapter_return_handlers[0]), O7_temp);
2119 __ sll_ptr(O5_dest_type, LogBytesPerWord, O5_dest_type);
2120 __ ld_ptr(O7_temp, O5_dest_type, O7_temp);
2121
2122 #ifdef ASSERT
2123 { Label L_ok;
2124 __ br_notnull(O7_temp, false, Assembler::pt, L_ok);
2125 __ delayed()->nop();
2126 __ stop("bad method handle return");
2127 __ BIND(L_ok);
2128 }
2129 #endif //ASSERT
2130 __ JMP(O7_temp, 0);
2131 __ delayed()->nop();
2132 }
2133 break;
2134
2135 case _adapter_opt_spread_0:
2136 case _adapter_opt_spread_1_ref:
2137 case _adapter_opt_spread_2_ref:
2138 case _adapter_opt_spread_3_ref:
2139 case _adapter_opt_spread_4_ref:
2140 case _adapter_opt_spread_5_ref:
2141 case _adapter_opt_spread_ref:
2142 case _adapter_opt_spread_byte:
2143 case _adapter_opt_spread_char:
2144 case _adapter_opt_spread_short:
2145 case _adapter_opt_spread_int:
2146 case _adapter_opt_spread_long:
2147 case _adapter_opt_spread_float:
2148 case _adapter_opt_spread_double:
2149 {
2150 // spread an array out into a group of arguments
2151 int length_constant = ek_adapter_opt_spread_count(ek);
2152 bool length_can_be_zero = (length_constant == 0);
2153 if (length_constant < 0) {
2154 // some adapters with variable length must handle the zero case
2155 if (!OptimizeMethodHandles ||
2156 ek_adapter_opt_spread_type(ek) != T_OBJECT)
2157 length_can_be_zero = true;
2158 }
2159
2160 // find the address of the array argument
2161 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
2162 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
2163
2164 // O0_argslot points both to the array and to the first output arg
2165 Address vmarg = Address(O0_argslot, 0);
2166
2167 // Get the array value.
2168 Register O1_array = O1_scratch;
2169 Register O2_array_klass = O2_scratch;
2170 BasicType elem_type = ek_adapter_opt_spread_type(ek);
2171 int elem_slots = type2size[elem_type]; // 1 or 2
2172 int array_slots = 1; // array is always a T_OBJECT
2173 int length_offset = arrayOopDesc::length_offset_in_bytes();
2174 int elem0_offset = arrayOopDesc::base_offset_in_bytes(elem_type);
2175 __ ld_ptr(vmarg, O1_array);
2176
2177 Label L_array_is_empty, L_insert_arg_space, L_copy_args, L_args_done;
2178 if (length_can_be_zero) {
2179 // handle the null pointer case, if zero is allowed
2180 Label L_skip;
2181 if (length_constant < 0) {
2182 load_conversion_vminfo(_masm, G3_amh_conversion, O3_scratch);
2183 __ br_zero(Assembler::notZero, false, Assembler::pn, O3_scratch, L_skip);
2184 __ delayed()->nop();
2185 }
2186 __ br_null(O1_array, false, Assembler::pn, L_array_is_empty);
2187 __ delayed()->nop();
2188 __ BIND(L_skip);
2189 }
2190 __ null_check(O1_array, oopDesc::klass_offset_in_bytes());
2191 __ load_klass(O1_array, O2_array_klass);
2192
2193 // Check the array type.
2194 Register O3_klass = O3_scratch;
2195 __ load_heap_oop(G3_amh_argument, O3_klass); // this is a Class object!
2196 load_klass_from_Class(_masm, O3_klass, O4_scratch, G5_scratch);
2197
2198 Label L_ok_array_klass, L_bad_array_klass, L_bad_array_length;
2199 __ check_klass_subtype(O2_array_klass, O3_klass, O4_scratch, G5_scratch, L_ok_array_klass);
2200 // If we get here, the type check failed!
2201 __ ba(false, L_bad_array_klass);
2202 __ delayed()->nop();
2203 __ BIND(L_ok_array_klass);
2204
2205 // Check length.
2206 if (length_constant >= 0) {
2207 __ ldsw(Address(O1_array, length_offset), O4_scratch);
2208 __ cmp(O4_scratch, length_constant);
2209 } else {
2210 Register O3_vminfo = O3_scratch;
2211 load_conversion_vminfo(_masm, G3_amh_conversion, O3_vminfo);
2212 __ ldsw(Address(O1_array, length_offset), O4_scratch);
2213 __ cmp(O3_vminfo, O4_scratch);
2214 }
2215 __ br(Assembler::notEqual, false, Assembler::pn, L_bad_array_length);
2216 __ delayed()->nop();
2217
2218 Register O2_argslot_limit = O2_scratch;
2219
2220 // Array length checks out. Now insert any required stack slots.
2221 if (length_constant == -1) {
2222 // Form a pointer to the end of the affected region.
2223 __ add(O0_argslot, Interpreter::stackElementSize, O2_argslot_limit);
2224 // 'stack_move' is negative number of words to insert
2225 // This number already accounts for elem_slots.
2226 Register O3_stack_move = O3_scratch;
2227 load_stack_move(_masm, G3_amh_conversion, O3_stack_move);
2228 __ cmp(O3_stack_move, 0);
2229 assert(stack_move_unit() < 0, "else change this comparison");
2230 __ br(Assembler::less, false, Assembler::pn, L_insert_arg_space);
2231 __ delayed()->nop();
2232 __ br(Assembler::equal, false, Assembler::pn, L_copy_args);
2233 __ delayed()->nop();
2234 // single argument case, with no array movement
2235 __ BIND(L_array_is_empty);
2236 remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2237 O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2238 __ ba(false, L_args_done); // no spreading to do
2239 __ delayed()->nop();
2240 __ BIND(L_insert_arg_space);
2241 // come here in the usual case, stack_move < 0 (2 or more spread arguments)
2242 // Live: O1_array, O2_argslot_limit, O3_stack_move
2243 insert_arg_slots(_masm, O3_stack_move,
2244 O0_argslot, O4_scratch, G5_scratch, O1_scratch);
2245 // reload from rdx_argslot_limit since rax_argslot is now decremented
2246 __ ld_ptr(Address(O2_argslot_limit, -Interpreter::stackElementSize), O1_array);
2247 } else if (length_constant >= 1) {
2248 int new_slots = (length_constant * elem_slots) - array_slots;
2249 insert_arg_slots(_masm, new_slots * stack_move_unit(),
2250 O0_argslot, O2_scratch, O3_scratch, O4_scratch);
2251 } else if (length_constant == 0) {
2252 __ BIND(L_array_is_empty);
2253 remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2254 O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2255 } else {
2256 ShouldNotReachHere();
2257 }
2258
2259 // Copy from the array to the new slots.
2260 // Note: Stack change code preserves integrity of O0_argslot pointer.
2261 // So even after slot insertions, O0_argslot still points to first argument.
2262 // Beware: Arguments that are shallow on the stack are deep in the array,
2263 // and vice versa. So a downward-growing stack (the usual) has to be copied
2264 // elementwise in reverse order from the source array.
2265 __ BIND(L_copy_args);
2266 if (length_constant == -1) {
2267 // [O0_argslot, O2_argslot_limit) is the area we are inserting into.
2268 // Array element [0] goes at O0_argslot_limit[-wordSize].
2269 Register O1_source = O1_array;
2270 __ add(Address(O1_array, elem0_offset), O1_source);
2271 Register O4_fill_ptr = O4_scratch;
2272 __ mov(O2_argslot_limit, O4_fill_ptr);
2273 Label L_loop;
2274 __ BIND(L_loop);
2275 __ add(O4_fill_ptr, -Interpreter::stackElementSize * elem_slots, O4_fill_ptr);
2276 move_typed_arg(_masm, elem_type, true,
2277 Address(O1_source, 0), Address(O4_fill_ptr, 0),
2278 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
2279 __ add(O1_source, type2aelembytes(elem_type), O1_source);
2280 __ cmp(O4_fill_ptr, O0_argslot);
2281 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, L_loop);
2282 __ delayed()->nop(); // FILLME
2283 } else if (length_constant == 0) {
2284 // nothing to copy
2285 } else {
2286 int elem_offset = elem0_offset;
2287 int slot_offset = length_constant * Interpreter::stackElementSize;
2288 for (int index = 0; index < length_constant; index++) {
2289 slot_offset -= Interpreter::stackElementSize * elem_slots; // fill backward
2290 move_typed_arg(_masm, elem_type, true,
2291 Address(O1_array, elem_offset), Address(O0_argslot, slot_offset),
2292 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
2293 elem_offset += type2aelembytes(elem_type);
2294 }
2295 }
2296 __ BIND(L_args_done);
2297
2298 // Arguments are spread. Move to next method handle.
2299 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
2300 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2301
2302 __ BIND(L_bad_array_klass);
2303 assert(!vmarg.uses(O2_required), "must be different registers");
2304 __ load_heap_oop(Address(O2_array_klass, java_mirror_offset), O2_required); // required class
2305 __ ld_ptr( vmarg, O1_actual); // bad object
2306 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2307 __ delayed()->mov(Bytecodes::_aaload, O0_code); // who is complaining?
2308
2309 __ bind(L_bad_array_length);
2310 assert(!vmarg.uses(O2_required), "must be different registers");
2311 __ mov( G3_method_handle, O2_required); // required class
2312 __ ld_ptr(vmarg, O1_actual); // bad object
2313 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2314 __ delayed()->mov(Bytecodes::_arraylength, O0_code); // who is complaining?
2315 }
2316 break;
2317
2318 default:
2319 DEBUG_ONLY(tty->print_cr("bad ek=%d (%s)", (int)ek, entry_name(ek)));
2320 ShouldNotReachHere();
2321 }
2322 BLOCK_COMMENT(err_msg("} Entry %s", entry_name(ek)));
2323
2324 address me_cookie = MethodHandleEntry::start_compiled_entry(_masm, interp_entry);
2325 __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
2326
2327 init_entry(ek, MethodHandleEntry::finish_compiled_entry(_masm, me_cookie));
2328 }