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
528 void MethodHandles::jump_from_method_handle(MacroAssembler* _masm, Register method, Register target, Register temp) {
529 assert(method == G5_method, "interpreter calling convention");
530 __ verify_oop(method);
531 __ ld_ptr(G5_method, in_bytes(methodOopDesc::from_interpreted_offset()), target);
532 if (JvmtiExport::can_post_interpreter_events()) {
533 // JVMTI events, such as single-stepping, are implemented partly by avoiding running
534 // compiled code in threads for which the event is enabled. Check here for
535 // interp_only_mode if these events CAN be enabled.
536 __ verify_thread();
537 Label skip_compiled_code;
538
539 const Address interp_only(G2_thread, JavaThread::interp_only_mode_offset());
540 __ ld(interp_only, temp);
541 __ tst(temp);
542 __ br(Assembler::notZero, true, Assembler::pn, skip_compiled_code);
543 __ delayed()->ld_ptr(G5_method, in_bytes(methodOopDesc::interpreter_entry_offset()), target);
544 __ bind(skip_compiled_code);
545 }
546 __ jmp(target, 0);
547 __ delayed()->nop();
548 }
549
550
551 // Code generation
552 address MethodHandles::generate_method_handle_interpreter_entry(MacroAssembler* _masm) {
553 // I5_savedSP/O5_savedSP: sender SP (must preserve)
554 // G4 (Gargs): incoming argument list (must preserve)
555 // G5_method: invoke methodOop
556 // G3_method_handle: receiver method handle (must load from sp[MethodTypeForm.vmslots])
557 // O0, O1, O2, O3, O4: garbage temps, blown away
558 Register O0_mtype = O0;
559 Register O1_scratch = O1;
560 Register O2_scratch = O2;
561 Register O3_scratch = O3;
562 Register O4_argslot = O4;
563 Register O4_argbase = O4;
564
565 // emit WrongMethodType path first, to enable back-branch from main path
566 Label wrong_method_type;
567 __ bind(wrong_method_type);
568 Label invoke_generic_slow_path;
569 assert(methodOopDesc::intrinsic_id_size_in_bytes() == sizeof(u1), "");;
570 __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
571 __ cmp(O1_scratch, (int) vmIntrinsics::_invokeExact);
572 __ brx(Assembler::notEqual, false, Assembler::pt, invoke_generic_slow_path);
573 __ delayed()->nop();
574 __ mov(O0_mtype, G5_method_type); // required by throw_WrongMethodType
575 __ mov(G3_method_handle, G3_method_handle); // already in this register
576 // O0 will be filled in with JavaThread in stub
577 __ jump_to(AddressLiteral(StubRoutines::throw_WrongMethodTypeException_entry()), O3_scratch);
578 __ delayed()->nop();
579
580 // here's where control starts out:
581 __ align(CodeEntryAlignment);
582 address entry_point = __ pc();
583
584 // fetch the MethodType from the method handle
585 // FIXME: Interpreter should transmit pre-popped stack pointer, to locate base of arg list.
586 // This would simplify several touchy bits of code.
587 // See 6984712: JSR 292 method handle calls need a clean argument base pointer
588 {
589 Register tem = G5_method;
590 for (jint* pchase = methodOopDesc::method_type_offsets_chain(); (*pchase) != -1; pchase++) {
591 __ ld_ptr(Address(tem, *pchase), O0_mtype);
592 tem = O0_mtype; // in case there is another indirection
593 }
594 }
595
596 // given the MethodType, find out where the MH argument is buried
597 __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, O1_scratch)), O4_argslot);
598 __ ldsw( Address(O4_argslot, __ delayed_value(java_lang_invoke_MethodTypeForm::vmslots_offset_in_bytes, O1_scratch)), O4_argslot);
599 __ add(__ argument_address(O4_argslot, O4_argslot, 1), O4_argbase);
600 // Note: argument_address uses its input as a scratch register!
601 Address mh_receiver_slot_addr(O4_argbase, -Interpreter::stackElementSize);
602 __ ld_ptr(mh_receiver_slot_addr, G3_method_handle);
603
604 trace_method_handle(_masm, "invokeExact");
605
606 __ check_method_handle_type(O0_mtype, G3_method_handle, O1_scratch, wrong_method_type);
607
608 // Nobody uses the MH receiver slot after this. Make sure.
609 DEBUG_ONLY(__ set((int32_t) 0x999999, O1_scratch); __ st_ptr(O1_scratch, mh_receiver_slot_addr));
610
611 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
612
613 // for invokeGeneric (only), apply argument and result conversions on the fly
614 __ bind(invoke_generic_slow_path);
615 #ifdef ASSERT
616 if (VerifyMethodHandles) {
617 Label L;
618 __ ldub(Address(G5_method, methodOopDesc::intrinsic_id_offset_in_bytes()), O1_scratch);
619 __ cmp(O1_scratch, (int) vmIntrinsics::_invokeGeneric);
620 __ brx(Assembler::equal, false, Assembler::pt, L);
621 __ delayed()->nop();
622 __ stop("bad methodOop::intrinsic_id");
623 __ bind(L);
624 }
625 #endif //ASSERT
626
627 // make room on the stack for another pointer:
628 insert_arg_slots(_masm, 2 * stack_move_unit(), O4_argbase, O1_scratch, O2_scratch, O3_scratch);
629 // load up an adapter from the calling type (Java weaves this)
630 Register O2_form = O2_scratch;
631 Register O3_adapter = O3_scratch;
632 __ load_heap_oop(Address(O0_mtype, __ delayed_value(java_lang_invoke_MethodType::form_offset_in_bytes, O1_scratch)), O2_form);
633 __ load_heap_oop(Address(O2_form, __ delayed_value(java_lang_invoke_MethodTypeForm::genericInvoker_offset_in_bytes, O1_scratch)), O3_adapter);
634 __ verify_oop(O3_adapter);
635 __ st_ptr(O3_adapter, Address(O4_argbase, 1 * Interpreter::stackElementSize));
636 // As a trusted first argument, pass the type being called, so the adapter knows
637 // the actual types of the arguments and return values.
638 // (Generic invokers are shared among form-families of method-type.)
639 __ st_ptr(O0_mtype, Address(O4_argbase, 0 * Interpreter::stackElementSize));
640 // FIXME: assert that O3_adapter is of the right method-type.
641 __ mov(O3_adapter, G3_method_handle);
642 trace_method_handle(_masm, "invokeGeneric");
643 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
644
645 return entry_point;
646 }
647
648 // Workaround for C++ overloading nastiness on '0' for RegisterOrConstant.
649 static RegisterOrConstant constant(int value) {
650 return RegisterOrConstant(value);
651 }
652
653 static void load_vmargslot(MacroAssembler* _masm, Address vmargslot_addr, Register result) {
654 __ ldsw(vmargslot_addr, result);
655 }
656
657 static RegisterOrConstant adjust_SP_and_Gargs_down_by_slots(MacroAssembler* _masm,
658 RegisterOrConstant arg_slots,
659 Register temp_reg, Register temp2_reg) {
660 // Keep the stack pointer 2*wordSize aligned.
661 const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
662 if (arg_slots.is_constant()) {
663 const int offset = arg_slots.as_constant() << LogBytesPerWord;
664 const int masked_offset = round_to(offset, 2 * BytesPerWord);
665 const int masked_offset2 = (offset + 1*BytesPerWord) & ~TwoWordAlignmentMask;
666 assert(masked_offset == masked_offset2, "must agree");
667 __ sub(Gargs, offset, Gargs);
668 __ sub(SP, masked_offset, SP );
669 return offset;
670 } else {
671 #ifdef ASSERT
672 {
673 Label L_ok;
674 __ cmp(arg_slots.as_register(), 0);
675 __ br(Assembler::greaterEqual, false, Assembler::pt, L_ok);
676 __ delayed()->nop();
677 __ stop("negative arg_slots");
678 __ bind(L_ok);
679 }
680 #endif
681 __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
682 __ add( temp_reg, 1*BytesPerWord, temp2_reg);
683 __ andn(temp2_reg, TwoWordAlignmentMask, temp2_reg);
684 __ sub(Gargs, temp_reg, Gargs);
685 __ sub(SP, temp2_reg, SP );
686 return temp_reg;
687 }
688 }
689
690 static RegisterOrConstant adjust_SP_and_Gargs_up_by_slots(MacroAssembler* _masm,
691 RegisterOrConstant arg_slots,
692 Register temp_reg, Register temp2_reg) {
693 // Keep the stack pointer 2*wordSize aligned.
694 const int TwoWordAlignmentMask = right_n_bits(LogBytesPerWord + 1);
695 if (arg_slots.is_constant()) {
696 const int offset = arg_slots.as_constant() << LogBytesPerWord;
697 const int masked_offset = offset & ~TwoWordAlignmentMask;
698 __ add(Gargs, offset, Gargs);
699 __ add(SP, masked_offset, SP );
700 return offset;
701 } else {
702 __ sll_ptr(arg_slots.as_register(), LogBytesPerWord, temp_reg);
703 __ andn(temp_reg, TwoWordAlignmentMask, temp2_reg);
704 __ add(Gargs, temp_reg, Gargs);
705 __ add(SP, temp2_reg, SP );
706 return temp_reg;
707 }
708 }
709
710 // Helper to insert argument slots into the stack.
711 // arg_slots must be a multiple of stack_move_unit() and < 0
712 // argslot_reg is decremented to point to the new (shifted) location of the argslot
713 // But, temp_reg ends up holding the original value of argslot_reg.
714 void MethodHandles::insert_arg_slots(MacroAssembler* _masm,
715 RegisterOrConstant arg_slots,
716 Register argslot_reg,
717 Register temp_reg, Register temp2_reg, Register temp3_reg) {
718 // allow constant zero
719 if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
720 return;
721
722 assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
723 (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
724
725 BLOCK_COMMENT("insert_arg_slots {");
726 if (VerifyMethodHandles)
727 verify_argslot(_masm, argslot_reg, temp_reg, "insertion point must fall within current frame");
728 if (VerifyMethodHandles)
729 verify_stack_move(_masm, arg_slots, -1);
730
731 // Make space on the stack for the inserted argument(s).
732 // Then pull down everything shallower than argslot_reg.
733 // The stacked return address gets pulled down with everything else.
734 // That is, copy [sp, argslot) downward by -size words. In pseudo-code:
735 // sp -= size;
736 // for (temp = sp + size; temp < argslot; temp++)
737 // temp[-size] = temp[0]
738 // argslot -= size;
739
740 // offset is temp3_reg in case of arg_slots being a register.
741 RegisterOrConstant offset = adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
742 __ sub(Gargs, offset, temp_reg); // source pointer for copy
743
744 {
745 Label loop;
746 __ BIND(loop);
747 // pull one word down each time through the loop
748 __ ld_ptr( Address(temp_reg, 0 ), temp2_reg);
749 __ st_ptr(temp2_reg, Address(temp_reg, offset) );
750 __ add(temp_reg, wordSize, temp_reg);
751 __ cmp(temp_reg, argslot_reg);
752 __ brx(Assembler::lessUnsigned, false, Assembler::pt, loop);
753 __ delayed()->nop(); // FILLME
754 }
755
756 // Now move the argslot down, to point to the opened-up space.
757 __ add(argslot_reg, offset, argslot_reg);
758 BLOCK_COMMENT("} insert_arg_slots");
759 }
760
761
762 // Helper to remove argument slots from the stack.
763 // arg_slots must be a multiple of stack_move_unit() and > 0
764 void MethodHandles::remove_arg_slots(MacroAssembler* _masm,
765 RegisterOrConstant arg_slots,
766 Register argslot_reg,
767 Register temp_reg, Register temp2_reg, Register temp3_reg) {
768 // allow constant zero
769 if (arg_slots.is_constant() && arg_slots.as_constant() == 0)
770 return;
771 assert_different_registers(argslot_reg, temp_reg, temp2_reg, temp3_reg,
772 (!arg_slots.is_register() ? Gargs : arg_slots.as_register()));
773
774 BLOCK_COMMENT("remove_arg_slots {");
775 if (VerifyMethodHandles)
776 verify_argslots(_masm, arg_slots, argslot_reg, temp_reg, temp2_reg, false,
777 "deleted argument(s) must fall within current frame");
778 if (VerifyMethodHandles)
779 verify_stack_move(_masm, arg_slots, +1);
780
781 // Pull up everything shallower than argslot.
782 // Then remove the excess space on the stack.
783 // The stacked return address gets pulled up with everything else.
784 // That is, copy [sp, argslot) upward by size words. In pseudo-code:
785 // for (temp = argslot-1; temp >= sp; --temp)
786 // temp[size] = temp[0]
787 // argslot += size;
788 // sp += size;
789
790 RegisterOrConstant offset = __ regcon_sll_ptr(arg_slots, LogBytesPerWord, temp3_reg);
791 __ sub(argslot_reg, wordSize, temp_reg); // source pointer for copy
792
793 {
794 Label L_loop;
795 __ BIND(L_loop);
796 // pull one word up each time through the loop
797 __ ld_ptr( Address(temp_reg, 0 ), temp2_reg);
798 __ st_ptr(temp2_reg, Address(temp_reg, offset) );
799 __ sub(temp_reg, wordSize, temp_reg);
800 __ cmp(temp_reg, Gargs);
801 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pt, L_loop);
802 __ delayed()->nop(); // FILLME
803 }
804
805 // And adjust the argslot address to point at the deletion point.
806 __ add(argslot_reg, offset, argslot_reg);
807
808 // We don't need the offset at this point anymore, just adjust SP and Gargs.
809 (void) adjust_SP_and_Gargs_up_by_slots(_masm, arg_slots, temp3_reg, temp_reg);
810
811 BLOCK_COMMENT("} remove_arg_slots");
812 }
813
814 // Helper to copy argument slots to the top of the stack.
815 // The sequence starts with argslot_reg and is counted by slot_count
816 // slot_count must be a multiple of stack_move_unit() and >= 0
817 // This function blows the temps but does not change argslot_reg.
818 void MethodHandles::push_arg_slots(MacroAssembler* _masm,
819 Register argslot_reg,
820 RegisterOrConstant slot_count,
821 Register temp_reg, Register temp2_reg) {
822 // allow constant zero
823 if (slot_count.is_constant() && slot_count.as_constant() == 0)
824 return;
825 assert_different_registers(argslot_reg, temp_reg, temp2_reg,
826 (!slot_count.is_register() ? Gargs : slot_count.as_register()),
827 SP);
828 assert(Interpreter::stackElementSize == wordSize, "else change this code");
829
830 BLOCK_COMMENT("push_arg_slots {");
831 if (VerifyMethodHandles)
832 verify_stack_move(_masm, slot_count, 0);
833
834 RegisterOrConstant offset = adjust_SP_and_Gargs_down_by_slots(_masm, slot_count, temp2_reg, temp_reg);
835
836 if (slot_count.is_constant()) {
837 for (int i = slot_count.as_constant() - 1; i >= 0; i--) {
838 __ ld_ptr( Address(argslot_reg, i * wordSize), temp_reg);
839 __ st_ptr(temp_reg, Address(Gargs, i * wordSize));
840 }
841 } else {
842 Label L_plural, L_loop, L_break;
843 // Emit code to dynamically check for the common cases, zero and one slot.
844 __ cmp(slot_count.as_register(), (int32_t) 1);
845 __ br(Assembler::greater, false, Assembler::pn, L_plural);
846 __ delayed()->nop();
847 __ br(Assembler::less, false, Assembler::pn, L_break);
848 __ delayed()->nop();
849 __ ld_ptr( Address(argslot_reg, 0), temp_reg);
850 __ st_ptr(temp_reg, Address(Gargs, 0));
851 __ ba(false, L_break);
852 __ delayed()->nop(); // FILLME
853 __ BIND(L_plural);
854
855 // Loop for 2 or more:
856 // top = &argslot[slot_count]
857 // while (top > argslot) *(--Gargs) = *(--top)
858 Register top_reg = temp_reg;
859 __ add(argslot_reg, offset, top_reg);
860 __ add(Gargs, offset, Gargs ); // move back up again so we can go down
861 __ BIND(L_loop);
862 __ sub(top_reg, wordSize, top_reg);
863 __ sub(Gargs, wordSize, Gargs );
864 __ ld_ptr( Address(top_reg, 0), temp2_reg);
865 __ st_ptr(temp2_reg, Address(Gargs, 0));
866 __ cmp(top_reg, argslot_reg);
867 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, L_loop);
868 __ delayed()->nop(); // FILLME
869 __ BIND(L_break);
870 }
871 BLOCK_COMMENT("} push_arg_slots");
872 }
873
874 // in-place movement; no change to Gargs
875 // blows temp_reg, temp2_reg
876 void MethodHandles::move_arg_slots_up(MacroAssembler* _masm,
877 Register bottom_reg, // invariant
878 Address top_addr, // can use temp_reg
879 RegisterOrConstant positive_distance_in_slots, // destroyed if register
880 Register temp_reg, Register temp2_reg) {
881 assert_different_registers(bottom_reg,
882 temp_reg, temp2_reg,
883 positive_distance_in_slots.register_or_noreg());
884 BLOCK_COMMENT("move_arg_slots_up {");
885 Label L_loop, L_break;
886 Register top_reg = temp_reg;
887 if (!top_addr.is_same_address(Address(top_reg, 0))) {
888 __ add(top_addr, top_reg);
889 }
890 // Detect empty (or broken) loop:
891 #ifdef ASSERT
892 if (VerifyMethodHandles) {
893 // Verify that &bottom < &top (non-empty interval)
894 Label L_ok, L_bad;
895 if (positive_distance_in_slots.is_register()) {
896 __ cmp(positive_distance_in_slots.as_register(), (int32_t) 0);
897 __ br(Assembler::lessEqual, false, Assembler::pn, L_bad);
898 __ delayed()->nop();
899 }
900 __ cmp(bottom_reg, top_reg);
901 __ brx(Assembler::lessUnsigned, false, Assembler::pt, L_ok);
902 __ delayed()->nop();
903 __ BIND(L_bad);
904 __ stop("valid bounds (copy up)");
905 __ BIND(L_ok);
906 }
907 #endif
908 __ cmp(bottom_reg, top_reg);
909 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pn, L_break);
910 __ delayed()->nop();
911 // work top down to bottom, copying contiguous data upwards
912 // In pseudo-code:
913 // while (--top >= bottom) *(top + distance) = *(top + 0);
914 RegisterOrConstant offset = __ argument_offset(positive_distance_in_slots, positive_distance_in_slots.register_or_noreg());
915 __ BIND(L_loop);
916 __ sub(top_reg, wordSize, top_reg);
917 __ ld_ptr( Address(top_reg, 0 ), temp2_reg);
918 __ st_ptr(temp2_reg, Address(top_reg, offset) );
919 __ cmp(top_reg, bottom_reg);
920 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, L_loop);
921 __ delayed()->nop(); // FILLME
922 assert(Interpreter::stackElementSize == wordSize, "else change loop");
923 __ BIND(L_break);
924 BLOCK_COMMENT("} move_arg_slots_up");
925 }
926
927 // in-place movement; no change to rsp
928 // blows temp_reg, temp2_reg
929 void MethodHandles::move_arg_slots_down(MacroAssembler* _masm,
930 Address bottom_addr, // can use temp_reg
931 Register top_reg, // invariant
932 RegisterOrConstant negative_distance_in_slots, // destroyed if register
933 Register temp_reg, Register temp2_reg) {
934 assert_different_registers(top_reg,
935 negative_distance_in_slots.register_or_noreg(),
936 temp_reg, temp2_reg);
937 BLOCK_COMMENT("move_arg_slots_down {");
938 Label L_loop, L_break;
939 Register bottom_reg = temp_reg;
940 if (!bottom_addr.is_same_address(Address(bottom_reg, 0))) {
941 __ add(bottom_addr, bottom_reg);
942 }
943 // Detect empty (or broken) loop:
944 #ifdef ASSERT
945 assert(!negative_distance_in_slots.is_constant() || negative_distance_in_slots.as_constant() < 0, "");
946 if (VerifyMethodHandles) {
947 // Verify that &bottom < &top (non-empty interval)
948 Label L_ok, L_bad;
949 if (negative_distance_in_slots.is_register()) {
950 __ cmp(negative_distance_in_slots.as_register(), (int32_t) 0);
951 __ br(Assembler::greaterEqual, false, Assembler::pn, L_bad);
952 __ delayed()->nop();
953 }
954 __ cmp(bottom_reg, top_reg);
955 __ brx(Assembler::lessUnsigned, false, Assembler::pt, L_ok);
956 __ delayed()->nop();
957 __ BIND(L_bad);
958 __ stop("valid bounds (copy down)");
959 __ BIND(L_ok);
960 }
961 #endif
962 __ cmp(bottom_reg, top_reg);
963 __ brx(Assembler::greaterEqualUnsigned, false, Assembler::pn, L_break);
964 __ delayed()->nop();
965 // work bottom up to top, copying contiguous data downwards
966 // In pseudo-code:
967 // while (bottom < top) *(bottom - distance) = *(bottom + 0), bottom++;
968 RegisterOrConstant offset = __ argument_offset(negative_distance_in_slots, negative_distance_in_slots.register_or_noreg());
969 __ BIND(L_loop);
970 __ ld_ptr( Address(bottom_reg, 0 ), temp2_reg);
971 __ st_ptr(temp2_reg, Address(bottom_reg, offset) );
972 __ add(bottom_reg, wordSize, bottom_reg);
973 __ cmp(bottom_reg, top_reg);
974 __ brx(Assembler::lessUnsigned, false, Assembler::pt, L_loop);
975 __ delayed()->nop(); // FILLME
976 assert(Interpreter::stackElementSize == wordSize, "else change loop");
977 __ BIND(L_break);
978 BLOCK_COMMENT("} move_arg_slots_down");
979 }
980
981 // Copy from a field or array element to a stacked argument slot.
982 // is_element (ignored) says whether caller is loading an array element instead of an instance field.
983 void MethodHandles::move_typed_arg(MacroAssembler* _masm,
984 BasicType type, bool is_element,
985 Address value_src, Address slot_dest,
986 Register temp_reg) {
987 assert(!slot_dest.uses(temp_reg), "must be different register");
988 BLOCK_COMMENT(!is_element ? "move_typed_arg {" : "move_typed_arg { (array element)");
989 if (type == T_OBJECT || type == T_ARRAY) {
990 __ load_heap_oop(value_src, temp_reg);
991 __ verify_oop(temp_reg);
992 __ st_ptr(temp_reg, slot_dest);
993 } else if (type != T_VOID) {
994 int arg_size = type2aelembytes(type);
995 bool arg_is_signed = is_signed_subword_type(type);
996 int slot_size = is_subword_type(type) ? type2aelembytes(T_INT) : arg_size; // store int sub-words as int
997 __ load_sized_value( value_src, temp_reg, arg_size, arg_is_signed);
998 __ store_sized_value(temp_reg, slot_dest, slot_size );
999 }
1000 BLOCK_COMMENT("} move_typed_arg");
1001 }
1002
1003 // Cf. TemplateInterpreterGenerator::generate_return_entry_for and
1004 // InterpreterMacroAssembler::save_return_value
1005 void MethodHandles::move_return_value(MacroAssembler* _masm, BasicType type,
1006 Address return_slot) {
1007 BLOCK_COMMENT("move_return_value {");
1008 // Look at the type and pull the value out of the corresponding register.
1009 if (type == T_VOID) {
1010 // nothing to do
1011 } else if (type == T_OBJECT) {
1012 __ verify_oop(O0);
1013 __ st_ptr(O0, return_slot);
1014 } else if (type == T_INT || is_subword_type(type)) {
1015 int type_size = type2aelembytes(T_INT);
1016 __ store_sized_value(O0, return_slot, type_size);
1017 } else if (type == T_LONG) {
1018 // store the value by parts
1019 // Note: We assume longs are continguous (if misaligned) on the interpreter stack.
1020 #if !defined(_LP64) && defined(COMPILER2)
1021 __ stx(G1, return_slot);
1022 #else
1023 #ifdef _LP64
1024 __ stx(O0, return_slot);
1025 #else
1026 if (return_slot.has_disp()) {
1027 // The displacement is a constant
1028 __ st(O0, return_slot);
1029 __ st(O1, return_slot.plus_disp(Interpreter::stackElementSize));
1030 } else {
1031 __ std(O0, return_slot);
1032 }
1033 #endif
1034 #endif
1035 } else if (type == T_FLOAT) {
1036 __ stf(FloatRegisterImpl::S, Ftos_f, return_slot);
1037 } else if (type == T_DOUBLE) {
1038 __ stf(FloatRegisterImpl::D, Ftos_f, return_slot);
1039 } else {
1040 ShouldNotReachHere();
1041 }
1042 BLOCK_COMMENT("} move_return_value");
1043 }
1044
1045 #ifndef PRODUCT
1046 extern "C" void print_method_handle(oop mh);
1047 void trace_method_handle_stub(const char* adaptername,
1048 oopDesc* mh,
1049 intptr_t* saved_sp) {
1050 bool has_mh = (strstr(adaptername, "return/") == NULL); // return adapters don't have mh
1051 tty->print_cr("MH %s mh="INTPTR_FORMAT " saved_sp=" INTPTR_FORMAT, adaptername, (intptr_t) mh, saved_sp);
1052 if (has_mh)
1053 print_method_handle(mh);
1054 }
1055 void MethodHandles::trace_method_handle(MacroAssembler* _masm, const char* adaptername) {
1056 if (!TraceMethodHandles) return;
1057 BLOCK_COMMENT("trace_method_handle {");
1058 // save: Gargs, O5_savedSP
1059 __ save_frame(16);
1060 __ set((intptr_t) adaptername, O0);
1061 __ mov(G3_method_handle, O1);
1062 __ mov(I5_savedSP, O2);
1063 __ mov(G3_method_handle, L3);
1064 __ mov(Gargs, L4);
1065 __ mov(G5_method_type, L5);
1066 __ call_VM_leaf(L7, CAST_FROM_FN_PTR(address, trace_method_handle_stub));
1067
1068 __ mov(L3, G3_method_handle);
1069 __ mov(L4, Gargs);
1070 __ mov(L5, G5_method_type);
1071 __ restore();
1072 BLOCK_COMMENT("} trace_method_handle");
1073 }
1074 #endif // PRODUCT
1075
1076 // which conversion op types are implemented here?
1077 int MethodHandles::adapter_conversion_ops_supported_mask() {
1078 return ((1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_ONLY)
1079 |(1<<java_lang_invoke_AdapterMethodHandle::OP_RETYPE_RAW)
1080 |(1<<java_lang_invoke_AdapterMethodHandle::OP_CHECK_CAST)
1081 |(1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_PRIM)
1082 |(1<<java_lang_invoke_AdapterMethodHandle::OP_REF_TO_PRIM)
1083 // OP_PRIM_TO_REF is below...
1084 |(1<<java_lang_invoke_AdapterMethodHandle::OP_SWAP_ARGS)
1085 |(1<<java_lang_invoke_AdapterMethodHandle::OP_ROT_ARGS)
1086 |(1<<java_lang_invoke_AdapterMethodHandle::OP_DUP_ARGS)
1087 |(1<<java_lang_invoke_AdapterMethodHandle::OP_DROP_ARGS)
1088 // OP_COLLECT_ARGS is below...
1089 |(1<<java_lang_invoke_AdapterMethodHandle::OP_SPREAD_ARGS)
1090 |(!UseRicochetFrames ? 0 :
1091 java_lang_invoke_MethodTypeForm::vmlayout_offset_in_bytes() <= 0 ? 0 :
1092 ((1<<java_lang_invoke_AdapterMethodHandle::OP_PRIM_TO_REF)
1093 |(1<<java_lang_invoke_AdapterMethodHandle::OP_COLLECT_ARGS)
1094 |(1<<java_lang_invoke_AdapterMethodHandle::OP_FOLD_ARGS)
1095 )
1096 )
1097 );
1098 }
1099
1100 //------------------------------------------------------------------------------
1101 // MethodHandles::generate_method_handle_stub
1102 //
1103 // Generate an "entry" field for a method handle.
1104 // This determines how the method handle will respond to calls.
1105 void MethodHandles::generate_method_handle_stub(MacroAssembler* _masm, MethodHandles::EntryKind ek) {
1106 MethodHandles::EntryKind ek_orig = ek_original_kind(ek);
1107
1108 // Here is the register state during an interpreted call,
1109 // as set up by generate_method_handle_interpreter_entry():
1110 // - G5: garbage temp (was MethodHandle.invoke methodOop, unused)
1111 // - G3: receiver method handle
1112 // - O5_savedSP: sender SP (must preserve)
1113
1114 const Register O0_scratch = O0;
1115 const Register O1_scratch = O1;
1116 const Register O2_scratch = O2;
1117 const Register O3_scratch = O3;
1118 const Register O4_scratch = O4;
1119 const Register G5_scratch = G5;
1120
1121 // Often used names:
1122 const Register O0_argslot = O0;
1123
1124 // Argument registers for _raise_exception:
1125 const Register O0_code = O0;
1126 const Register O1_actual = O1;
1127 const Register O2_required = O2;
1128
1129 guarantee(java_lang_invoke_MethodHandle::vmentry_offset_in_bytes() != 0, "must have offsets");
1130
1131 // Some handy addresses:
1132 Address G3_mh_vmtarget( G3_method_handle, java_lang_invoke_MethodHandle::vmtarget_offset_in_bytes());
1133
1134 Address G3_dmh_vmindex( G3_method_handle, java_lang_invoke_DirectMethodHandle::vmindex_offset_in_bytes());
1135
1136 Address G3_bmh_vmargslot( G3_method_handle, java_lang_invoke_BoundMethodHandle::vmargslot_offset_in_bytes());
1137 Address G3_bmh_argument( G3_method_handle, java_lang_invoke_BoundMethodHandle::argument_offset_in_bytes());
1138
1139 Address G3_amh_vmargslot( G3_method_handle, java_lang_invoke_AdapterMethodHandle::vmargslot_offset_in_bytes());
1140 Address G3_amh_argument ( G3_method_handle, java_lang_invoke_AdapterMethodHandle::argument_offset_in_bytes());
1141 Address G3_amh_conversion(G3_method_handle, java_lang_invoke_AdapterMethodHandle::conversion_offset_in_bytes());
1142
1143 const int java_mirror_offset = klassOopDesc::klass_part_offset_in_bytes() + Klass::java_mirror_offset_in_bytes();
1144
1145 if (have_entry(ek)) {
1146 __ nop(); // empty stubs make SG sick
1147 return;
1148 }
1149
1150 address interp_entry = __ pc();
1151
1152 trace_method_handle(_masm, entry_name(ek));
1153
1154 BLOCK_COMMENT(err_msg("Entry %s {", entry_name(ek)));
1155
1156 switch ((int) ek) {
1157 case _raise_exception:
1158 {
1159 // Not a real MH entry, but rather shared code for raising an
1160 // exception. For sharing purposes the arguments are passed into registers
1161 // and then placed in the intepreter calling convention here.
1162 assert(raise_exception_method(), "must be set");
1163 assert(raise_exception_method()->from_compiled_entry(), "method must be linked");
1164
1165 __ set(AddressLiteral((address) &_raise_exception_method), G5_method);
1166 __ ld_ptr(Address(G5_method, 0), G5_method);
1167
1168 const int jobject_oop_offset = 0;
1169 __ ld_ptr(Address(G5_method, jobject_oop_offset), G5_method);
1170
1171 adjust_SP_and_Gargs_down_by_slots(_masm, 3, noreg, noreg);
1172
1173 __ st_ptr(O0_code, __ argument_address(constant(2), noreg, 0));
1174 __ st_ptr(O1_actual, __ argument_address(constant(1), noreg, 0));
1175 __ st_ptr(O2_required, __ argument_address(constant(0), noreg, 0));
1176 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1177 }
1178 break;
1179
1180 case _invokestatic_mh:
1181 case _invokespecial_mh:
1182 {
1183 __ load_heap_oop(G3_mh_vmtarget, G5_method); // target is a methodOop
1184 // Same as TemplateTable::invokestatic or invokespecial,
1185 // minus the CP setup and profiling:
1186 if (ek == _invokespecial_mh) {
1187 // Must load & check the first argument before entering the target method.
1188 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1189 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1190 __ null_check(G3_method_handle);
1191 __ verify_oop(G3_method_handle);
1192 }
1193 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1194 }
1195 break;
1196
1197 case _invokevirtual_mh:
1198 {
1199 // Same as TemplateTable::invokevirtual,
1200 // minus the CP setup and profiling:
1201
1202 // Pick out the vtable index and receiver offset from the MH,
1203 // and then we can discard it:
1204 Register O2_index = O2_scratch;
1205 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1206 __ ldsw(G3_dmh_vmindex, O2_index);
1207 // Note: The verifier allows us to ignore G3_mh_vmtarget.
1208 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1209 __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1210
1211 // Get receiver klass:
1212 Register O0_klass = O0_argslot;
1213 __ load_klass(G3_method_handle, O0_klass);
1214 __ verify_oop(O0_klass);
1215
1216 // Get target methodOop & entry point:
1217 const int base = instanceKlass::vtable_start_offset() * wordSize;
1218 assert(vtableEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1219
1220 __ sll_ptr(O2_index, LogBytesPerWord, O2_index);
1221 __ add(O0_klass, O2_index, O0_klass);
1222 Address vtable_entry_addr(O0_klass, base + vtableEntry::method_offset_in_bytes());
1223 __ ld_ptr(vtable_entry_addr, G5_method);
1224
1225 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1226 }
1227 break;
1228
1229 case _invokeinterface_mh:
1230 {
1231 // Same as TemplateTable::invokeinterface,
1232 // minus the CP setup and profiling:
1233 __ load_method_handle_vmslots(O0_argslot, G3_method_handle, O1_scratch);
1234 Register O1_intf = O1_scratch;
1235 Register G5_index = G5_scratch;
1236 __ load_heap_oop(G3_mh_vmtarget, O1_intf);
1237 __ ldsw(G3_dmh_vmindex, G5_index);
1238 __ ld_ptr(__ argument_address(O0_argslot, O0_argslot, -1), G3_method_handle);
1239 __ null_check(G3_method_handle, oopDesc::klass_offset_in_bytes());
1240
1241 // Get receiver klass:
1242 Register O0_klass = O0_argslot;
1243 __ load_klass(G3_method_handle, O0_klass);
1244 __ verify_oop(O0_klass);
1245
1246 // Get interface:
1247 Label no_such_interface;
1248 __ verify_oop(O1_intf);
1249 __ lookup_interface_method(O0_klass, O1_intf,
1250 // Note: next two args must be the same:
1251 G5_index, G5_method,
1252 O2_scratch,
1253 O3_scratch,
1254 no_such_interface);
1255
1256 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1257
1258 __ bind(no_such_interface);
1259 // Throw an exception.
1260 // For historical reasons, it will be IncompatibleClassChangeError.
1261 __ unimplemented("not tested yet");
1262 __ ld_ptr(Address(O1_intf, java_mirror_offset), O2_required); // required interface
1263 __ mov( O0_klass, O1_actual); // bad receiver
1264 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1265 __ delayed()->mov(Bytecodes::_invokeinterface, O0_code); // who is complaining?
1266 }
1267 break;
1268
1269 case _bound_ref_mh:
1270 case _bound_int_mh:
1271 case _bound_long_mh:
1272 case _bound_ref_direct_mh:
1273 case _bound_int_direct_mh:
1274 case _bound_long_direct_mh:
1275 {
1276 const bool direct_to_method = (ek >= _bound_ref_direct_mh);
1277 BasicType arg_type = ek_bound_mh_arg_type(ek);
1278 int arg_slots = type2size[arg_type];
1279
1280 // Make room for the new argument:
1281 load_vmargslot(_masm, G3_bmh_vmargslot, O0_argslot);
1282 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1283
1284 insert_arg_slots(_masm, arg_slots * stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1285
1286 // Store bound argument into the new stack slot:
1287 __ load_heap_oop(G3_bmh_argument, O1_scratch);
1288 if (arg_type == T_OBJECT) {
1289 __ st_ptr(O1_scratch, Address(O0_argslot, 0));
1290 } else {
1291 Address prim_value_addr(O1_scratch, java_lang_boxing_object::value_offset_in_bytes(arg_type));
1292 move_typed_arg(_masm, arg_type, false,
1293 prim_value_addr,
1294 Address(O0_argslot, 0),
1295 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
1296 }
1297
1298 if (direct_to_method) {
1299 __ load_heap_oop(G3_mh_vmtarget, G5_method); // target is a methodOop
1300 jump_from_method_handle(_masm, G5_method, O1_scratch, O2_scratch);
1301 } else {
1302 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle); // target is a methodOop
1303 __ verify_oop(G3_method_handle);
1304 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1305 }
1306 }
1307 break;
1308
1309 case _adapter_retype_only:
1310 case _adapter_retype_raw:
1311 // Immediately jump to the next MH layer:
1312 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1313 __ verify_oop(G3_method_handle);
1314 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1315 // This is OK when all parameter types widen.
1316 // It is also OK when a return type narrows.
1317 break;
1318
1319 case _adapter_check_cast:
1320 {
1321 // Check a reference argument before jumping to the next layer of MH:
1322 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1323 Address vmarg = __ argument_address(O0_argslot, O0_argslot);
1324
1325 // What class are we casting to?
1326 Register O1_klass = O1_scratch; // Interesting AMH data.
1327 __ load_heap_oop(G3_amh_argument, O1_klass); // This is a Class object!
1328 load_klass_from_Class(_masm, O1_klass, O2_scratch, O3_scratch);
1329
1330 Label L_done;
1331 __ ld_ptr(vmarg, O2_scratch);
1332 __ tst(O2_scratch);
1333 __ brx(Assembler::zero, false, Assembler::pn, L_done); // No cast if null.
1334 __ delayed()->nop();
1335 __ load_klass(O2_scratch, O2_scratch);
1336
1337 // Live at this point:
1338 // - O0_argslot : argslot index in vmarg; may be required in the failing path
1339 // - O1_klass : klass required by the target method
1340 // - O2_scratch : argument klass to test
1341 // - G3_method_handle: adapter method handle
1342 __ check_klass_subtype(O2_scratch, O1_klass, O3_scratch, O4_scratch, L_done);
1343
1344 // If we get here, the type check failed!
1345 __ load_heap_oop(G3_amh_argument, O2_required); // required class
1346 __ ld_ptr( vmarg, O1_actual); // bad object
1347 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
1348 __ delayed()->mov(Bytecodes::_checkcast, O0_code); // who is complaining?
1349
1350 __ BIND(L_done);
1351 // Get the new MH:
1352 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1353 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1354 }
1355 break;
1356
1357 case _adapter_prim_to_prim:
1358 case _adapter_ref_to_prim:
1359 // Handled completely by optimized cases.
1360 __ stop("init_AdapterMethodHandle should not issue this");
1361 break;
1362
1363 case _adapter_opt_i2i: // optimized subcase of adapt_prim_to_prim
1364 //case _adapter_opt_f2i: // optimized subcase of adapt_prim_to_prim
1365 case _adapter_opt_l2i: // optimized subcase of adapt_prim_to_prim
1366 case _adapter_opt_unboxi: // optimized subcase of adapt_ref_to_prim
1367 {
1368 // Perform an in-place conversion to int or an int subword.
1369 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1370 Address value;
1371 Address vmarg;
1372 bool value_left_justified = false;
1373
1374 switch (ek) {
1375 case _adapter_opt_i2i:
1376 value = vmarg = __ argument_address(O0_argslot, O0_argslot);
1377 break;
1378 case _adapter_opt_l2i:
1379 {
1380 // just delete the extra slot
1381 #ifdef _LP64
1382 // In V9, longs are given 2 64-bit slots in the interpreter, but the
1383 // data is passed in only 1 slot.
1384 // Keep the second slot.
1385 __ add(__ argument_address(O0_argslot, O0_argslot, -1), O0_argslot);
1386 remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1387 value = Address(O0_argslot, 4); // Get least-significant 32-bit of 64-bit value.
1388 vmarg = Address(O0_argslot, Interpreter::stackElementSize);
1389 #else
1390 // Keep the first slot.
1391 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1392 remove_arg_slots(_masm, -stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1393 value = Address(O0_argslot, 0);
1394 vmarg = value;
1395 #endif
1396 }
1397 break;
1398 case _adapter_opt_unboxi:
1399 {
1400 vmarg = __ argument_address(O0_argslot, O0_argslot);
1401 // Load the value up from the heap.
1402 __ ld_ptr(vmarg, O1_scratch);
1403 int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_INT);
1404 #ifdef ASSERT
1405 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1406 if (is_subword_type(BasicType(bt)))
1407 assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(BasicType(bt)), "");
1408 }
1409 #endif
1410 __ null_check(O1_scratch, value_offset);
1411 value = Address(O1_scratch, value_offset);
1412 #ifdef _BIG_ENDIAN
1413 // Values stored in objects are packed.
1414 value_left_justified = true;
1415 #endif
1416 }
1417 break;
1418 default:
1419 ShouldNotReachHere();
1420 }
1421
1422 // This check is required on _BIG_ENDIAN
1423 Register G5_vminfo = G5_scratch;
1424 __ ldsw(G3_amh_conversion, G5_vminfo);
1425 assert(CONV_VMINFO_SHIFT == 0, "preshifted");
1426
1427 // Original 32-bit vmdata word must be of this form:
1428 // | MBZ:6 | signBitCount:8 | srcDstTypes:8 | conversionOp:8 |
1429 __ lduw(value, O1_scratch);
1430 if (!value_left_justified)
1431 __ sll(O1_scratch, G5_vminfo, O1_scratch);
1432 Label zero_extend, done;
1433 __ btst(CONV_VMINFO_SIGN_FLAG, G5_vminfo);
1434 __ br(Assembler::zero, false, Assembler::pn, zero_extend);
1435 __ delayed()->nop();
1436
1437 // this path is taken for int->byte, int->short
1438 __ sra(O1_scratch, G5_vminfo, O1_scratch);
1439 __ ba(false, done);
1440 __ delayed()->nop();
1441
1442 __ bind(zero_extend);
1443 // this is taken for int->char
1444 __ srl(O1_scratch, G5_vminfo, O1_scratch);
1445
1446 __ bind(done);
1447 __ st(O1_scratch, vmarg);
1448
1449 // Get the new MH:
1450 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1451 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1452 }
1453 break;
1454
1455 case _adapter_opt_i2l: // optimized subcase of adapt_prim_to_prim
1456 case _adapter_opt_unboxl: // optimized subcase of adapt_ref_to_prim
1457 {
1458 // Perform an in-place int-to-long or ref-to-long conversion.
1459 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1460
1461 // On big-endian machine we duplicate the slot and store the MSW
1462 // in the first slot.
1463 __ add(__ argument_address(O0_argslot, O0_argslot, 1), O0_argslot);
1464
1465 insert_arg_slots(_masm, stack_move_unit(), O0_argslot, O1_scratch, O2_scratch, O3_scratch);
1466
1467 Address arg_lsw(O0_argslot, 0);
1468 Address arg_msw(O0_argslot, -Interpreter::stackElementSize);
1469
1470 switch (ek) {
1471 case _adapter_opt_i2l:
1472 {
1473 #ifdef _LP64
1474 __ ldsw(arg_lsw, O2_scratch); // Load LSW sign-extended
1475 #else
1476 __ ldsw(arg_lsw, O3_scratch); // Load LSW sign-extended
1477 __ srlx(O3_scratch, BitsPerInt, O2_scratch); // Move MSW value to lower 32-bits for std
1478 #endif
1479 __ st_long(O2_scratch, arg_msw); // Uses O2/O3 on !_LP64
1480 }
1481 break;
1482 case _adapter_opt_unboxl:
1483 {
1484 // Load the value up from the heap.
1485 __ ld_ptr(arg_lsw, O1_scratch);
1486 int value_offset = java_lang_boxing_object::value_offset_in_bytes(T_LONG);
1487 assert(value_offset == java_lang_boxing_object::value_offset_in_bytes(T_DOUBLE), "");
1488 __ null_check(O1_scratch, value_offset);
1489 __ ld_long(Address(O1_scratch, value_offset), O2_scratch); // Uses O2/O3 on !_LP64
1490 __ st_long(O2_scratch, arg_msw);
1491 }
1492 break;
1493 default:
1494 ShouldNotReachHere();
1495 }
1496
1497 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1498 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1499 }
1500 break;
1501
1502 case _adapter_opt_f2d: // optimized subcase of adapt_prim_to_prim
1503 case _adapter_opt_d2f: // optimized subcase of adapt_prim_to_prim
1504 {
1505 // perform an in-place floating primitive conversion
1506 __ unimplemented(entry_name(ek));
1507 }
1508 break;
1509
1510 case _adapter_prim_to_ref:
1511 __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
1512 break;
1513
1514 case _adapter_swap_args:
1515 case _adapter_rot_args:
1516 // handled completely by optimized cases
1517 __ stop("init_AdapterMethodHandle should not issue this");
1518 break;
1519
1520 case _adapter_opt_swap_1:
1521 case _adapter_opt_swap_2:
1522 case _adapter_opt_rot_1_up:
1523 case _adapter_opt_rot_1_down:
1524 case _adapter_opt_rot_2_up:
1525 case _adapter_opt_rot_2_down:
1526 {
1527 int swap_slots = ek_adapter_opt_swap_slots(ek);
1528 int rotate = ek_adapter_opt_swap_mode(ek);
1529
1530 // 'argslot' is the position of the first argument to swap.
1531 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1532 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1533 if (VerifyMethodHandles)
1534 verify_argslot(_masm, O0_argslot, O2_scratch, "swap point must fall within current frame");
1535
1536 // 'vminfo' is the second.
1537 Register O1_destslot = O1_scratch;
1538 load_conversion_vminfo(_masm, G3_amh_conversion, O1_destslot);
1539 __ add(__ argument_address(O1_destslot, O1_destslot), O1_destslot);
1540 if (VerifyMethodHandles)
1541 verify_argslot(_masm, O1_destslot, O2_scratch, "swap point must fall within current frame");
1542
1543 assert(Interpreter::stackElementSize == wordSize, "else rethink use of wordSize here");
1544 if (!rotate) {
1545 // simple swap
1546 for (int i = 0; i < swap_slots; i++) {
1547 __ ld_ptr( Address(O0_argslot, i * wordSize), O2_scratch);
1548 __ ld_ptr( Address(O1_destslot, i * wordSize), O3_scratch);
1549 __ st_ptr(O3_scratch, Address(O0_argslot, i * wordSize));
1550 __ st_ptr(O2_scratch, Address(O1_destslot, i * wordSize));
1551 }
1552 } else {
1553 // A rotate is actually pair of moves, with an "odd slot" (or pair)
1554 // changing place with a series of other slots.
1555 // First, push the "odd slot", which is going to get overwritten
1556 switch (swap_slots) {
1557 case 2 : __ ld_ptr(Address(O0_argslot, 1 * wordSize), O4_scratch); // fall-thru
1558 case 1 : __ ld_ptr(Address(O0_argslot, 0 * wordSize), O3_scratch); break;
1559 default: ShouldNotReachHere();
1560 }
1561 if (rotate > 0) {
1562 // Here is rotate > 0:
1563 // (low mem) (high mem)
1564 // | dest: more_slots... | arg: odd_slot :arg+1 |
1565 // =>
1566 // | dest: odd_slot | dest+1: more_slots... :arg+1 |
1567 // work argslot down to destslot, copying contiguous data upwards
1568 // pseudo-code:
1569 // argslot = src_addr - swap_bytes
1570 // destslot = dest_addr
1571 // while (argslot >= destslot) *(argslot + swap_bytes) = *(argslot + 0), argslot--;
1572 move_arg_slots_up(_masm,
1573 O1_destslot,
1574 Address(O0_argslot, 0),
1575 swap_slots,
1576 O0_argslot, O2_scratch);
1577 } else {
1578 // Here is the other direction, rotate < 0:
1579 // (low mem) (high mem)
1580 // | arg: odd_slot | arg+1: more_slots... :dest+1 |
1581 // =>
1582 // | arg: more_slots... | dest: odd_slot :dest+1 |
1583 // work argslot up to destslot, copying contiguous data downwards
1584 // pseudo-code:
1585 // argslot = src_addr + swap_bytes
1586 // destslot = dest_addr
1587 // while (argslot <= destslot) *(argslot - swap_bytes) = *(argslot + 0), argslot++;
1588 // dest_slot denotes an exclusive upper limit
1589 int limit_bias = OP_ROT_ARGS_DOWN_LIMIT_BIAS;
1590 if (limit_bias != 0)
1591 __ add(O1_destslot, - limit_bias * wordSize, O1_destslot);
1592 move_arg_slots_down(_masm,
1593 Address(O0_argslot, swap_slots * wordSize),
1594 O1_destslot,
1595 -swap_slots,
1596 O0_argslot, O2_scratch);
1597
1598 __ sub(O1_destslot, swap_slots * wordSize, O1_destslot);
1599 }
1600 // pop the original first chunk into the destination slot, now free
1601 switch (swap_slots) {
1602 case 2 : __ st_ptr(O4_scratch, Address(O1_destslot, 1 * wordSize)); // fall-thru
1603 case 1 : __ st_ptr(O3_scratch, Address(O1_destslot, 0 * wordSize)); break;
1604 default: ShouldNotReachHere();
1605 }
1606 }
1607
1608 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1609 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1610 }
1611 break;
1612
1613 case _adapter_dup_args:
1614 {
1615 // 'argslot' is the position of the first argument to duplicate.
1616 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1617 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1618
1619 // 'stack_move' is negative number of words to duplicate.
1620 Register O1_stack_move = O1_scratch;
1621 load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1622
1623 if (VerifyMethodHandles) {
1624 verify_argslots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, true,
1625 "copied argument(s) must fall within current frame");
1626 }
1627
1628 // insert location is always the bottom of the argument list:
1629 __ neg(O1_stack_move);
1630 push_arg_slots(_masm, O0_argslot, O1_stack_move, O2_scratch, O3_scratch);
1631
1632 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1633 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1634 }
1635 break;
1636
1637 case _adapter_drop_args:
1638 {
1639 // 'argslot' is the position of the first argument to nuke.
1640 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
1641 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
1642
1643 // 'stack_move' is number of words to drop.
1644 Register O1_stack_move = O1_scratch;
1645 load_stack_move(_masm, G3_amh_conversion, O1_stack_move);
1646
1647 remove_arg_slots(_masm, O1_stack_move, O0_argslot, O2_scratch, O3_scratch, O4_scratch);
1648
1649 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
1650 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
1651 }
1652 break;
1653
1654 case _adapter_collect_args:
1655 case _adapter_fold_args:
1656 case _adapter_spread_args:
1657 // Handled completely by optimized cases.
1658 __ stop("init_AdapterMethodHandle should not issue this");
1659 break;
1660
1661 case _adapter_opt_collect_ref:
1662 case _adapter_opt_collect_int:
1663 case _adapter_opt_collect_long:
1664 case _adapter_opt_collect_float:
1665 case _adapter_opt_collect_double:
1666 case _adapter_opt_collect_void:
1667 case _adapter_opt_collect_0_ref:
1668 case _adapter_opt_collect_1_ref:
1669 case _adapter_opt_collect_2_ref:
1670 case _adapter_opt_collect_3_ref:
1671 case _adapter_opt_collect_4_ref:
1672 case _adapter_opt_collect_5_ref:
1673 case _adapter_opt_filter_S0_ref:
1674 case _adapter_opt_filter_S1_ref:
1675 case _adapter_opt_filter_S2_ref:
1676 case _adapter_opt_filter_S3_ref:
1677 case _adapter_opt_filter_S4_ref:
1678 case _adapter_opt_filter_S5_ref:
1679 case _adapter_opt_collect_2_S0_ref:
1680 case _adapter_opt_collect_2_S1_ref:
1681 case _adapter_opt_collect_2_S2_ref:
1682 case _adapter_opt_collect_2_S3_ref:
1683 case _adapter_opt_collect_2_S4_ref:
1684 case _adapter_opt_collect_2_S5_ref:
1685 case _adapter_opt_fold_ref:
1686 case _adapter_opt_fold_int:
1687 case _adapter_opt_fold_long:
1688 case _adapter_opt_fold_float:
1689 case _adapter_opt_fold_double:
1690 case _adapter_opt_fold_void:
1691 case _adapter_opt_fold_1_ref:
1692 case _adapter_opt_fold_2_ref:
1693 case _adapter_opt_fold_3_ref:
1694 case _adapter_opt_fold_4_ref:
1695 case _adapter_opt_fold_5_ref:
1696 {
1697 // Given a fresh incoming stack frame, build a new ricochet frame.
1698 // On entry, TOS points at a return PC, and FP is the callers frame ptr.
1699 // RSI/R13 has the caller's exact stack pointer, which we must also preserve.
1700 // RCX contains an AdapterMethodHandle of the indicated kind.
1701
1702 // Relevant AMH fields:
1703 // amh.vmargslot:
1704 // points to the trailing edge of the arguments
1705 // to filter, collect, or fold. For a boxing operation,
1706 // it points just after the single primitive value.
1707 // amh.argument:
1708 // recursively called MH, on |collect| arguments
1709 // amh.vmtarget:
1710 // final destination MH, on return value, etc.
1711 // amh.conversion.dest:
1712 // tells what is the type of the return value
1713 // (not needed here, since dest is also derived from ek)
1714 // amh.conversion.vminfo:
1715 // points to the trailing edge of the return value
1716 // when the vmtarget is to be called; this is
1717 // equal to vmargslot + (retained ? |collect| : 0)
1718
1719 // Pass 0 or more argument slots to the recursive target.
1720 int collect_count_constant = ek_adapter_opt_collect_count(ek);
1721
1722 // The collected arguments are copied from the saved argument list:
1723 int collect_slot_constant = ek_adapter_opt_collect_slot(ek);
1724
1725 assert(ek_orig == _adapter_collect_args ||
1726 ek_orig == _adapter_fold_args, "");
1727 bool retain_original_args = (ek_orig == _adapter_fold_args);
1728
1729 // The return value is replaced (or inserted) at the 'vminfo' argslot.
1730 // Sometimes we can compute this statically.
1731 int dest_slot_constant = -1;
1732 if (!retain_original_args)
1733 dest_slot_constant = collect_slot_constant;
1734 else if (collect_slot_constant >= 0 && collect_count_constant >= 0)
1735 // We are preserving all the arguments, and the return value is prepended,
1736 // so the return slot is to the left (above) the |collect| sequence.
1737 dest_slot_constant = collect_slot_constant + collect_count_constant;
1738
1739 // Replace all those slots by the result of the recursive call.
1740 // The result type can be one of ref, int, long, float, double, void.
1741 // In the case of void, nothing is pushed on the stack after return.
1742 BasicType dest = ek_adapter_opt_collect_type(ek);
1743 assert(dest == type2wfield[dest], "dest is a stack slot type");
1744 int dest_count = type2size[dest];
1745 assert(dest_count == 1 || dest_count == 2 || (dest_count == 0 && dest == T_VOID), "dest has a size");
1746
1747 // Choose a return continuation.
1748 EntryKind ek_ret = _adapter_opt_return_any;
1749 if (dest != T_CONFLICT && OptimizeMethodHandles) {
1750 switch (dest) {
1751 case T_INT : ek_ret = _adapter_opt_return_int; break;
1752 case T_LONG : ek_ret = _adapter_opt_return_long; break;
1753 case T_FLOAT : ek_ret = _adapter_opt_return_float; break;
1754 case T_DOUBLE : ek_ret = _adapter_opt_return_double; break;
1755 case T_OBJECT : ek_ret = _adapter_opt_return_ref; break;
1756 case T_VOID : ek_ret = _adapter_opt_return_void; break;
1757 default : ShouldNotReachHere();
1758 }
1759 if (dest == T_OBJECT && dest_slot_constant >= 0) {
1760 EntryKind ek_try = EntryKind(_adapter_opt_return_S0_ref + dest_slot_constant);
1761 if (ek_try <= _adapter_opt_return_LAST &&
1762 ek_adapter_opt_return_slot(ek_try) == dest_slot_constant) {
1763 ek_ret = ek_try;
1764 }
1765 }
1766 assert(ek_adapter_opt_return_type(ek_ret) == dest, "");
1767 }
1768
1769 // Already pushed: ... keep1 | collect | keep2 |
1770
1771 // Push a few extra argument words, if we need them to store the return value.
1772 {
1773 int extra_slots = 0;
1774 if (retain_original_args) {
1775 extra_slots = dest_count;
1776 } else if (collect_count_constant == -1) {
1777 extra_slots = dest_count; // collect_count might be zero; be generous
1778 } else if (dest_count > collect_count_constant) {
1779 extra_slots = (dest_count - collect_count_constant);
1780 } else {
1781 // else we know we have enough dead space in |collect| to repurpose for return values
1782 }
1783 if (extra_slots != 0) {
1784 __ sub(SP, round_to(extra_slots, 2) * Interpreter::stackElementSize, SP);
1785 }
1786 }
1787
1788 // Set up Ricochet Frame.
1789 __ mov(SP, O5_savedSP); // record SP for the callee
1790
1791 // One extra (empty) slot for outgoing target MH (see Gargs computation below).
1792 __ save_frame(2); // Note: we need to add 2 slots since frame::memory_parameter_word_sp_offset is 23.
1793
1794 // Note: Gargs is live throughout the following, until we make our recursive call.
1795 // And the RF saves a copy in L4_saved_args_base.
1796
1797 RicochetFrame::enter_ricochet_frame(_masm, G3_method_handle, Gargs,
1798 entry(ek_ret)->from_interpreted_entry());
1799
1800 // Compute argument base:
1801 // Set up Gargs for current frame, extra (empty) slot is for outgoing target MH (space reserved by save_frame above).
1802 __ add(FP, STACK_BIAS - (1 * Interpreter::stackElementSize), Gargs);
1803
1804 // Now pushed: ... keep1 | collect | keep2 | extra | [RF]
1805
1806 #ifdef ASSERT
1807 if (VerifyMethodHandles && dest != T_CONFLICT) {
1808 BLOCK_COMMENT("verify AMH.conv.dest {");
1809 extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O1_scratch);
1810 Label L_dest_ok;
1811 __ cmp(O1_scratch, (int) dest);
1812 __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1813 __ delayed()->nop();
1814 if (dest == T_INT) {
1815 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
1816 if (is_subword_type(BasicType(bt))) {
1817 __ cmp(O1_scratch, (int) bt);
1818 __ br(Assembler::equal, false, Assembler::pt, L_dest_ok);
1819 __ delayed()->nop();
1820 }
1821 }
1822 }
1823 __ stop("bad dest in AMH.conv");
1824 __ BIND(L_dest_ok);
1825 BLOCK_COMMENT("} verify AMH.conv.dest");
1826 }
1827 #endif //ASSERT
1828
1829 // Find out where the original copy of the recursive argument sequence begins.
1830 Register O0_coll = O0_scratch;
1831 {
1832 RegisterOrConstant collect_slot = collect_slot_constant;
1833 if (collect_slot_constant == -1) {
1834 load_vmargslot(_masm, G3_amh_vmargslot, O1_scratch);
1835 collect_slot = O1_scratch;
1836 }
1837 // collect_slot might be 0, but we need the move anyway.
1838 __ add(RicochetFrame::L4_saved_args_base, __ argument_offset(collect_slot, collect_slot.register_or_noreg()), O0_coll);
1839 // O0_coll now points at the trailing edge of |collect| and leading edge of |keep2|
1840 }
1841
1842 // Replace the old AMH with the recursive MH. (No going back now.)
1843 // In the case of a boxing call, the recursive call is to a 'boxer' method,
1844 // such as Integer.valueOf or Long.valueOf. In the case of a filter
1845 // or collect call, it will take one or more arguments, transform them,
1846 // and return some result, to store back into argument_base[vminfo].
1847 __ load_heap_oop(G3_amh_argument, G3_method_handle);
1848 if (VerifyMethodHandles) verify_method_handle(_masm, G3_method_handle, O1_scratch, O2_scratch);
1849
1850 // Calculate |collect|, the number of arguments we are collecting.
1851 Register O1_collect_count = O1_scratch;
1852 RegisterOrConstant collect_count;
1853 if (collect_count_constant < 0) {
1854 __ load_method_handle_vmslots(O1_collect_count, G3_method_handle, O2_scratch);
1855 collect_count = O1_collect_count;
1856 } else {
1857 collect_count = collect_count_constant;
1858 #ifdef ASSERT
1859 if (VerifyMethodHandles) {
1860 BLOCK_COMMENT("verify collect_count_constant {");
1861 __ load_method_handle_vmslots(O3_scratch, G3_method_handle, O2_scratch);
1862 Label L_count_ok;
1863 __ cmp(O3_scratch, collect_count_constant);
1864 __ br(Assembler::equal, false, Assembler::pt, L_count_ok);
1865 __ delayed()->nop();
1866 __ stop("bad vminfo in AMH.conv");
1867 __ BIND(L_count_ok);
1868 BLOCK_COMMENT("} verify collect_count_constant");
1869 }
1870 #endif //ASSERT
1871 }
1872
1873 // copy |collect| slots directly to TOS:
1874 push_arg_slots(_masm, O0_coll, collect_count, O2_scratch, O3_scratch);
1875 // Now pushed: ... keep1 | collect | keep2 | RF... | collect |
1876 // O0_coll still points at the trailing edge of |collect| and leading edge of |keep2|
1877
1878 // If necessary, adjust the saved arguments to make room for the eventual return value.
1879 // Normal adjustment: ... keep1 | +dest+ | -collect- | keep2 | RF... | collect |
1880 // If retaining args: ... keep1 | +dest+ | collect | keep2 | RF... | collect |
1881 // In the non-retaining case, this might move keep2 either up or down.
1882 // We don't have to copy the whole | RF... collect | complex,
1883 // but we must adjust RF.saved_args_base.
1884 // Also, from now on, we will forget about the original copy of |collect|.
1885 // If we are retaining it, we will treat it as part of |keep2|.
1886 // For clarity we will define |keep3| = |collect|keep2| or |keep2|.
1887
1888 BLOCK_COMMENT("adjust trailing arguments {");
1889 // Compare the sizes of |+dest+| and |-collect-|, which are opposed opening and closing movements.
1890 int open_count = dest_count;
1891 RegisterOrConstant close_count = collect_count_constant;
1892 Register O1_close_count = O1_collect_count;
1893 if (retain_original_args) {
1894 close_count = constant(0);
1895 } else if (collect_count_constant == -1) {
1896 close_count = O1_collect_count;
1897 }
1898
1899 // How many slots need moving? This is simply dest_slot (0 => no |keep3|).
1900 RegisterOrConstant keep3_count;
1901 Register O2_keep3_count = O2_scratch;
1902 if (dest_slot_constant < 0) {
1903 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O2_keep3_count);
1904 keep3_count = O2_keep3_count;
1905 } else {
1906 keep3_count = dest_slot_constant;
1907 #ifdef ASSERT
1908 if (VerifyMethodHandles && dest_slot_constant < 0) {
1909 BLOCK_COMMENT("verify dest_slot_constant {");
1910 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O3_scratch);
1911 Label L_vminfo_ok;
1912 __ cmp(O3_scratch, dest_slot_constant);
1913 __ br(Assembler::equal, false, Assembler::pt, L_vminfo_ok);
1914 __ delayed()->nop();
1915 __ stop("bad vminfo in AMH.conv");
1916 __ BIND(L_vminfo_ok);
1917 BLOCK_COMMENT("} verify dest_slot_constant");
1918 }
1919 #endif //ASSERT
1920 }
1921
1922 // tasks remaining:
1923 bool move_keep3 = (!keep3_count.is_constant() || keep3_count.as_constant() != 0);
1924 bool stomp_dest = (NOT_DEBUG(dest == T_OBJECT) DEBUG_ONLY(dest_count != 0));
1925 bool fix_arg_base = (!close_count.is_constant() || open_count != close_count.as_constant());
1926
1927 // Old and new argument locations (based at slot 0).
1928 // Net shift (&new_argv - &old_argv) is (close_count - open_count).
1929 bool zero_open_count = (open_count == 0); // remember this bit of info
1930 if (move_keep3 && fix_arg_base) {
1931 // It will be easier to have everything in one register:
1932 if (close_count.is_register()) {
1933 // Deduct open_count from close_count register to get a clean +/- value.
1934 __ sub(close_count.as_register(), open_count, close_count.as_register());
1935 } else {
1936 close_count = close_count.as_constant() - open_count;
1937 }
1938 open_count = 0;
1939 }
1940 Register L4_old_argv = RicochetFrame::L4_saved_args_base;
1941 Register O3_new_argv = O3_scratch;
1942 if (fix_arg_base) {
1943 __ add(L4_old_argv, __ argument_offset(close_count, O4_scratch), O3_new_argv,
1944 -(open_count * Interpreter::stackElementSize));
1945 }
1946
1947 // First decide if any actual data are to be moved.
1948 // We can skip if (a) |keep3| is empty, or (b) the argument list size didn't change.
1949 // (As it happens, all movements involve an argument list size change.)
1950
1951 // If there are variable parameters, use dynamic checks to skip around the whole mess.
1952 Label L_done;
1953 if (keep3_count.is_register()) {
1954 __ tst(keep3_count.as_register());
1955 __ br(Assembler::zero, false, Assembler::pn, L_done);
1956 __ delayed()->nop();
1957 }
1958 if (close_count.is_register()) {
1959 __ cmp(close_count.as_register(), open_count);
1960 __ br(Assembler::equal, false, Assembler::pn, L_done);
1961 __ delayed()->nop();
1962 }
1963
1964 if (move_keep3 && fix_arg_base) {
1965 bool emit_move_down = false, emit_move_up = false, emit_guard = false;
1966 if (!close_count.is_constant()) {
1967 emit_move_down = emit_guard = !zero_open_count;
1968 emit_move_up = true;
1969 } else if (open_count != close_count.as_constant()) {
1970 emit_move_down = (open_count > close_count.as_constant());
1971 emit_move_up = !emit_move_down;
1972 }
1973 Label L_move_up;
1974 if (emit_guard) {
1975 __ cmp(close_count.as_register(), open_count);
1976 __ br(Assembler::greater, false, Assembler::pn, L_move_up);
1977 __ delayed()->nop();
1978 }
1979
1980 if (emit_move_down) {
1981 // Move arguments down if |+dest+| > |-collect-|
1982 // (This is rare, except when arguments are retained.)
1983 // This opens space for the return value.
1984 if (keep3_count.is_constant()) {
1985 for (int i = 0; i < keep3_count.as_constant(); i++) {
1986 __ ld_ptr( Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
1987 __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize) );
1988 }
1989 } else {
1990 // Live: O1_close_count, O2_keep3_count, O3_new_argv
1991 Register argv_top = O0_scratch;
1992 __ add(L4_old_argv, __ argument_offset(keep3_count, O4_scratch), argv_top);
1993 move_arg_slots_down(_masm,
1994 Address(L4_old_argv, 0), // beginning of old argv
1995 argv_top, // end of old argv
1996 close_count, // distance to move down (must be negative)
1997 O4_scratch, G5_scratch);
1998 }
1999 }
2000
2001 if (emit_guard) {
2002 __ ba(false, L_done); // assumes emit_move_up is true also
2003 __ delayed()->nop();
2004 __ BIND(L_move_up);
2005 }
2006
2007 if (emit_move_up) {
2008 // Move arguments up if |+dest+| < |-collect-|
2009 // (This is usual, except when |keep3| is empty.)
2010 // This closes up the space occupied by the now-deleted collect values.
2011 if (keep3_count.is_constant()) {
2012 for (int i = keep3_count.as_constant() - 1; i >= 0; i--) {
2013 __ ld_ptr( Address(L4_old_argv, i * Interpreter::stackElementSize), O4_scratch);
2014 __ st_ptr(O4_scratch, Address(O3_new_argv, i * Interpreter::stackElementSize) );
2015 }
2016 } else {
2017 Address argv_top(L4_old_argv, __ argument_offset(keep3_count, O4_scratch));
2018 // Live: O1_close_count, O2_keep3_count, O3_new_argv
2019 move_arg_slots_up(_masm,
2020 L4_old_argv, // beginning of old argv
2021 argv_top, // end of old argv
2022 close_count, // distance to move up (must be positive)
2023 O4_scratch, G5_scratch);
2024 }
2025 }
2026 }
2027 __ BIND(L_done);
2028
2029 if (fix_arg_base) {
2030 // adjust RF.saved_args_base
2031 __ mov(O3_new_argv, RicochetFrame::L4_saved_args_base);
2032 }
2033
2034 if (stomp_dest) {
2035 // Stomp the return slot, so it doesn't hold garbage.
2036 // This isn't strictly necessary, but it may help detect bugs.
2037 __ set(RicochetFrame::RETURN_VALUE_PLACEHOLDER, O4_scratch);
2038 __ st_ptr(O4_scratch, Address(RicochetFrame::L4_saved_args_base,
2039 __ argument_offset(keep3_count, keep3_count.register_or_noreg()))); // uses O2_keep3_count
2040 }
2041 BLOCK_COMMENT("} adjust trailing arguments");
2042
2043 BLOCK_COMMENT("do_recursive_call");
2044 __ mov(SP, O5_savedSP); // record SP for the callee
2045 __ set(ExternalAddress(SharedRuntime::ricochet_blob()->bounce_addr() - frame::pc_return_offset), O7);
2046 // The globally unique bounce address has two purposes:
2047 // 1. It helps the JVM recognize this frame (frame::is_ricochet_frame).
2048 // 2. When returned to, it cuts back the stack and redirects control flow
2049 // to the return handler.
2050 // The return handler will further cut back the stack when it takes
2051 // down the RF. Perhaps there is a way to streamline this further.
2052
2053 // State during recursive call:
2054 // ... keep1 | dest | dest=42 | keep3 | RF... | collect | bounce_pc |
2055 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2056 }
2057 break;
2058
2059 case _adapter_opt_return_ref:
2060 case _adapter_opt_return_int:
2061 case _adapter_opt_return_long:
2062 case _adapter_opt_return_float:
2063 case _adapter_opt_return_double:
2064 case _adapter_opt_return_void:
2065 case _adapter_opt_return_S0_ref:
2066 case _adapter_opt_return_S1_ref:
2067 case _adapter_opt_return_S2_ref:
2068 case _adapter_opt_return_S3_ref:
2069 case _adapter_opt_return_S4_ref:
2070 case _adapter_opt_return_S5_ref:
2071 {
2072 BasicType dest_type_constant = ek_adapter_opt_return_type(ek);
2073 int dest_slot_constant = ek_adapter_opt_return_slot(ek);
2074
2075 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
2076
2077 if (dest_slot_constant == -1) {
2078 // The current stub is a general handler for this dest_type.
2079 // It can be called from _adapter_opt_return_any below.
2080 // Stash the address in a little table.
2081 assert((dest_type_constant & CONV_TYPE_MASK) == dest_type_constant, "oob");
2082 address return_handler = __ pc();
2083 _adapter_return_handlers[dest_type_constant] = return_handler;
2084 if (dest_type_constant == T_INT) {
2085 // do the subword types too
2086 for (int bt = T_BOOLEAN; bt < T_INT; bt++) {
2087 if (is_subword_type(BasicType(bt)) &&
2088 _adapter_return_handlers[bt] == NULL) {
2089 _adapter_return_handlers[bt] = return_handler;
2090 }
2091 }
2092 }
2093 }
2094
2095 // On entry to this continuation handler, make Gargs live again.
2096 __ mov(RicochetFrame::L4_saved_args_base, Gargs);
2097
2098 Register O7_temp = O7;
2099 Register O5_vminfo = O5;
2100
2101 RegisterOrConstant dest_slot = dest_slot_constant;
2102 if (dest_slot_constant == -1) {
2103 extract_conversion_vminfo(_masm, RicochetFrame::L5_conversion, O5_vminfo);
2104 dest_slot = O5_vminfo;
2105 }
2106 // Store the result back into the argslot.
2107 // This code uses the interpreter calling sequence, in which the return value
2108 // is usually left in the TOS register, as defined by InterpreterMacroAssembler::pop.
2109 // There are certain irregularities with floating point values, which can be seen
2110 // in TemplateInterpreterGenerator::generate_return_entry_for.
2111 move_return_value(_masm, dest_type_constant, __ argument_address(dest_slot, O7_temp));
2112
2113 RicochetFrame::leave_ricochet_frame(_masm, G3_method_handle, I5_savedSP, I7);
2114
2115 // Load the final target and go.
2116 if (VerifyMethodHandles) verify_method_handle(_masm, G3_method_handle, O0_scratch, O1_scratch);
2117 __ restore(I5_savedSP, G0, SP);
2118 __ jump_to_method_handle_entry(G3_method_handle, O0_scratch);
2119 __ illtrap(0);
2120 }
2121 break;
2122
2123 case _adapter_opt_return_any:
2124 {
2125 Register O7_temp = O7;
2126 Register O5_dest_type = O5;
2127
2128 if (VerifyMethodHandles) RicochetFrame::verify_clean(_masm);
2129 extract_conversion_dest_type(_masm, RicochetFrame::L5_conversion, O5_dest_type);
2130 __ set(ExternalAddress((address) &_adapter_return_handlers[0]), O7_temp);
2131 __ sll_ptr(O5_dest_type, LogBytesPerWord, O5_dest_type);
2132 __ ld_ptr(O7_temp, O5_dest_type, O7_temp);
2133
2134 #ifdef ASSERT
2135 { Label L_ok;
2136 __ br_notnull(O7_temp, false, Assembler::pt, L_ok);
2137 __ delayed()->nop();
2138 __ stop("bad method handle return");
2139 __ BIND(L_ok);
2140 }
2141 #endif //ASSERT
2142 __ JMP(O7_temp, 0);
2143 __ delayed()->nop();
2144 }
2145 break;
2146
2147 case _adapter_opt_spread_0:
2148 case _adapter_opt_spread_1_ref:
2149 case _adapter_opt_spread_2_ref:
2150 case _adapter_opt_spread_3_ref:
2151 case _adapter_opt_spread_4_ref:
2152 case _adapter_opt_spread_5_ref:
2153 case _adapter_opt_spread_ref:
2154 case _adapter_opt_spread_byte:
2155 case _adapter_opt_spread_char:
2156 case _adapter_opt_spread_short:
2157 case _adapter_opt_spread_int:
2158 case _adapter_opt_spread_long:
2159 case _adapter_opt_spread_float:
2160 case _adapter_opt_spread_double:
2161 {
2162 // spread an array out into a group of arguments
2163 int length_constant = ek_adapter_opt_spread_count(ek);
2164 bool length_can_be_zero = (length_constant == 0);
2165 if (length_constant < 0) {
2166 // some adapters with variable length must handle the zero case
2167 if (!OptimizeMethodHandles ||
2168 ek_adapter_opt_spread_type(ek) != T_OBJECT)
2169 length_can_be_zero = true;
2170 }
2171
2172 // find the address of the array argument
2173 load_vmargslot(_masm, G3_amh_vmargslot, O0_argslot);
2174 __ add(__ argument_address(O0_argslot, O0_argslot), O0_argslot);
2175
2176 // O0_argslot points both to the array and to the first output arg
2177 Address vmarg = Address(O0_argslot, 0);
2178
2179 // Get the array value.
2180 Register O1_array = O1_scratch;
2181 Register O2_array_klass = O2_scratch;
2182 BasicType elem_type = ek_adapter_opt_spread_type(ek);
2183 int elem_slots = type2size[elem_type]; // 1 or 2
2184 int array_slots = 1; // array is always a T_OBJECT
2185 int length_offset = arrayOopDesc::length_offset_in_bytes();
2186 int elem0_offset = arrayOopDesc::base_offset_in_bytes(elem_type);
2187 __ ld_ptr(vmarg, O1_array);
2188
2189 Label L_array_is_empty, L_insert_arg_space, L_copy_args, L_args_done;
2190 if (length_can_be_zero) {
2191 // handle the null pointer case, if zero is allowed
2192 Label L_skip;
2193 if (length_constant < 0) {
2194 load_conversion_vminfo(_masm, G3_amh_conversion, O3_scratch);
2195 __ br_zero(Assembler::notZero, false, Assembler::pn, O3_scratch, L_skip);
2196 __ delayed()->nop();
2197 }
2198 __ br_null(O1_array, false, Assembler::pn, L_array_is_empty);
2199 __ delayed()->nop();
2200 __ BIND(L_skip);
2201 }
2202 __ null_check(O1_array, oopDesc::klass_offset_in_bytes());
2203 __ load_klass(O1_array, O2_array_klass);
2204
2205 // Check the array type.
2206 Register O3_klass = O3_scratch;
2207 __ load_heap_oop(G3_amh_argument, O3_klass); // this is a Class object!
2208 load_klass_from_Class(_masm, O3_klass, O4_scratch, G5_scratch);
2209
2210 Label L_ok_array_klass, L_bad_array_klass, L_bad_array_length;
2211 __ check_klass_subtype(O2_array_klass, O3_klass, O4_scratch, G5_scratch, L_ok_array_klass);
2212 // If we get here, the type check failed!
2213 __ ba(false, L_bad_array_klass);
2214 __ delayed()->nop();
2215 __ BIND(L_ok_array_klass);
2216
2217 // Check length.
2218 if (length_constant >= 0) {
2219 __ ldsw(Address(O1_array, length_offset), O4_scratch);
2220 __ cmp(O4_scratch, length_constant);
2221 } else {
2222 Register O3_vminfo = O3_scratch;
2223 load_conversion_vminfo(_masm, G3_amh_conversion, O3_vminfo);
2224 __ ldsw(Address(O1_array, length_offset), O4_scratch);
2225 __ cmp(O3_vminfo, O4_scratch);
2226 }
2227 __ br(Assembler::notEqual, false, Assembler::pn, L_bad_array_length);
2228 __ delayed()->nop();
2229
2230 Register O2_argslot_limit = O2_scratch;
2231
2232 // Array length checks out. Now insert any required stack slots.
2233 if (length_constant == -1) {
2234 // Form a pointer to the end of the affected region.
2235 __ add(O0_argslot, Interpreter::stackElementSize, O2_argslot_limit);
2236 // 'stack_move' is negative number of words to insert
2237 // This number already accounts for elem_slots.
2238 Register O3_stack_move = O3_scratch;
2239 load_stack_move(_masm, G3_amh_conversion, O3_stack_move);
2240 __ cmp(O3_stack_move, 0);
2241 assert(stack_move_unit() < 0, "else change this comparison");
2242 __ br(Assembler::less, false, Assembler::pn, L_insert_arg_space);
2243 __ delayed()->nop();
2244 __ br(Assembler::equal, false, Assembler::pn, L_copy_args);
2245 __ delayed()->nop();
2246 // single argument case, with no array movement
2247 __ BIND(L_array_is_empty);
2248 remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2249 O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2250 __ ba(false, L_args_done); // no spreading to do
2251 __ delayed()->nop();
2252 __ BIND(L_insert_arg_space);
2253 // come here in the usual case, stack_move < 0 (2 or more spread arguments)
2254 // Live: O1_array, O2_argslot_limit, O3_stack_move
2255 insert_arg_slots(_masm, O3_stack_move,
2256 O0_argslot, O4_scratch, G5_scratch, O1_scratch);
2257 // reload from rdx_argslot_limit since rax_argslot is now decremented
2258 __ ld_ptr(Address(O2_argslot_limit, -Interpreter::stackElementSize), O1_array);
2259 } else if (length_constant >= 1) {
2260 int new_slots = (length_constant * elem_slots) - array_slots;
2261 insert_arg_slots(_masm, new_slots * stack_move_unit(),
2262 O0_argslot, O2_scratch, O3_scratch, O4_scratch);
2263 } else if (length_constant == 0) {
2264 __ BIND(L_array_is_empty);
2265 remove_arg_slots(_masm, -stack_move_unit() * array_slots,
2266 O0_argslot, O1_scratch, O2_scratch, O3_scratch);
2267 } else {
2268 ShouldNotReachHere();
2269 }
2270
2271 // Copy from the array to the new slots.
2272 // Note: Stack change code preserves integrity of O0_argslot pointer.
2273 // So even after slot insertions, O0_argslot still points to first argument.
2274 // Beware: Arguments that are shallow on the stack are deep in the array,
2275 // and vice versa. So a downward-growing stack (the usual) has to be copied
2276 // elementwise in reverse order from the source array.
2277 __ BIND(L_copy_args);
2278 if (length_constant == -1) {
2279 // [O0_argslot, O2_argslot_limit) is the area we are inserting into.
2280 // Array element [0] goes at O0_argslot_limit[-wordSize].
2281 Register O1_source = O1_array;
2282 __ add(Address(O1_array, elem0_offset), O1_source);
2283 Register O4_fill_ptr = O4_scratch;
2284 __ mov(O2_argslot_limit, O4_fill_ptr);
2285 Label L_loop;
2286 __ BIND(L_loop);
2287 __ add(O4_fill_ptr, -Interpreter::stackElementSize * elem_slots, O4_fill_ptr);
2288 move_typed_arg(_masm, elem_type, true,
2289 Address(O1_source, 0), Address(O4_fill_ptr, 0),
2290 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
2291 __ add(O1_source, type2aelembytes(elem_type), O1_source);
2292 __ cmp(O4_fill_ptr, O0_argslot);
2293 __ brx(Assembler::greaterUnsigned, false, Assembler::pt, L_loop);
2294 __ delayed()->nop(); // FILLME
2295 } else if (length_constant == 0) {
2296 // nothing to copy
2297 } else {
2298 int elem_offset = elem0_offset;
2299 int slot_offset = length_constant * Interpreter::stackElementSize;
2300 for (int index = 0; index < length_constant; index++) {
2301 slot_offset -= Interpreter::stackElementSize * elem_slots; // fill backward
2302 move_typed_arg(_masm, elem_type, true,
2303 Address(O1_array, elem_offset), Address(O0_argslot, slot_offset),
2304 O2_scratch); // must be an even register for !_LP64 long moves (uses O2/O3)
2305 elem_offset += type2aelembytes(elem_type);
2306 }
2307 }
2308 __ BIND(L_args_done);
2309
2310 // Arguments are spread. Move to next method handle.
2311 __ load_heap_oop(G3_mh_vmtarget, G3_method_handle);
2312 __ jump_to_method_handle_entry(G3_method_handle, O1_scratch);
2313
2314 __ BIND(L_bad_array_klass);
2315 assert(!vmarg.uses(O2_required), "must be different registers");
2316 __ load_heap_oop(Address(O2_array_klass, java_mirror_offset), O2_required); // required class
2317 __ ld_ptr( vmarg, O1_actual); // bad object
2318 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2319 __ delayed()->mov(Bytecodes::_aaload, O0_code); // who is complaining?
2320
2321 __ bind(L_bad_array_length);
2322 assert(!vmarg.uses(O2_required), "must be different registers");
2323 __ mov( G3_method_handle, O2_required); // required class
2324 __ ld_ptr(vmarg, O1_actual); // bad object
2325 __ jump_to(AddressLiteral(from_interpreted_entry(_raise_exception)), O3_scratch);
2326 __ delayed()->mov(Bytecodes::_arraylength, O0_code); // who is complaining?
2327 }
2328 break;
2329
2330 default:
2331 DEBUG_ONLY(tty->print_cr("bad ek=%d (%s)", (int)ek, entry_name(ek)));
2332 ShouldNotReachHere();
2333 }
2334 BLOCK_COMMENT(err_msg("} Entry %s", entry_name(ek)));
2335
2336 address me_cookie = MethodHandleEntry::start_compiled_entry(_masm, interp_entry);
2337 __ unimplemented(entry_name(ek)); // %%% FIXME: NYI
2338
2339 init_entry(ek, MethodHandleEntry::finish_compiled_entry(_masm, me_cookie));
2340 }